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/ftrace_event.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 struct remote_function_call {
55 struct task_struct *p;
56 int (*func)(void *info);
61 static void remote_function(void *data)
63 struct remote_function_call *tfc = data;
64 struct task_struct *p = tfc->p;
68 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
72 tfc->ret = tfc->func(tfc->info);
76 * task_function_call - call a function on the cpu on which a task runs
77 * @p: the task to evaluate
78 * @func: the function to be called
79 * @info: the function call argument
81 * Calls the function @func when the task is currently running. This might
82 * be on the current CPU, which just calls the function directly
84 * returns: @func return value, or
85 * -ESRCH - when the process isn't running
86 * -EAGAIN - when the process moved away
89 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
91 struct remote_function_call data = {
95 .ret = -ESRCH, /* No such (running) process */
99 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
105 * cpu_function_call - call a function on the cpu
106 * @func: the function to be called
107 * @info: the function call argument
109 * Calls the function @func on the remote cpu.
111 * returns: @func return value or -ENXIO when the cpu is offline
113 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
115 struct remote_function_call data = {
119 .ret = -ENXIO, /* No such CPU */
122 smp_call_function_single(cpu, remote_function, &data, 1);
127 #define EVENT_OWNER_KERNEL ((void *) -1)
129 static bool is_kernel_event(struct perf_event *event)
131 return event->owner == EVENT_OWNER_KERNEL;
134 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
135 PERF_FLAG_FD_OUTPUT |\
136 PERF_FLAG_PID_CGROUP |\
137 PERF_FLAG_FD_CLOEXEC)
140 * branch priv levels that need permission checks
142 #define PERF_SAMPLE_BRANCH_PERM_PLM \
143 (PERF_SAMPLE_BRANCH_KERNEL |\
144 PERF_SAMPLE_BRANCH_HV)
147 EVENT_FLEXIBLE = 0x1,
149 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
153 * perf_sched_events : >0 events exist
154 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
156 struct static_key_deferred perf_sched_events __read_mostly;
157 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
158 static DEFINE_PER_CPU(int, perf_sched_cb_usages);
160 static atomic_t nr_mmap_events __read_mostly;
161 static atomic_t nr_comm_events __read_mostly;
162 static atomic_t nr_task_events __read_mostly;
163 static atomic_t nr_freq_events __read_mostly;
165 static LIST_HEAD(pmus);
166 static DEFINE_MUTEX(pmus_lock);
167 static struct srcu_struct pmus_srcu;
170 * perf event paranoia level:
171 * -1 - not paranoid at all
172 * 0 - disallow raw tracepoint access for unpriv
173 * 1 - disallow cpu events for unpriv
174 * 2 - disallow kernel profiling for unpriv
176 int sysctl_perf_event_paranoid __read_mostly = 1;
178 /* Minimum for 512 kiB + 1 user control page */
179 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
182 * max perf event sample rate
184 #define DEFAULT_MAX_SAMPLE_RATE 100000
185 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
186 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
188 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
190 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
191 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
193 static int perf_sample_allowed_ns __read_mostly =
194 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
196 void update_perf_cpu_limits(void)
198 u64 tmp = perf_sample_period_ns;
200 tmp *= sysctl_perf_cpu_time_max_percent;
202 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
205 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
207 int perf_proc_update_handler(struct ctl_table *table, int write,
208 void __user *buffer, size_t *lenp,
211 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
216 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
217 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
218 update_perf_cpu_limits();
223 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
225 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
226 void __user *buffer, size_t *lenp,
229 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
234 update_perf_cpu_limits();
240 * perf samples are done in some very critical code paths (NMIs).
241 * If they take too much CPU time, the system can lock up and not
242 * get any real work done. This will drop the sample rate when
243 * we detect that events are taking too long.
245 #define NR_ACCUMULATED_SAMPLES 128
246 static DEFINE_PER_CPU(u64, running_sample_length);
248 static void perf_duration_warn(struct irq_work *w)
250 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
251 u64 avg_local_sample_len;
252 u64 local_samples_len;
254 local_samples_len = __this_cpu_read(running_sample_length);
255 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
257 printk_ratelimited(KERN_WARNING
258 "perf interrupt took too long (%lld > %lld), lowering "
259 "kernel.perf_event_max_sample_rate to %d\n",
260 avg_local_sample_len, allowed_ns >> 1,
261 sysctl_perf_event_sample_rate);
264 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
266 void perf_sample_event_took(u64 sample_len_ns)
268 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
269 u64 avg_local_sample_len;
270 u64 local_samples_len;
275 /* decay the counter by 1 average sample */
276 local_samples_len = __this_cpu_read(running_sample_length);
277 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
278 local_samples_len += sample_len_ns;
279 __this_cpu_write(running_sample_length, local_samples_len);
282 * note: this will be biased artifically low until we have
283 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
284 * from having to maintain a count.
286 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
288 if (avg_local_sample_len <= allowed_ns)
291 if (max_samples_per_tick <= 1)
294 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
295 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
296 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
298 update_perf_cpu_limits();
300 if (!irq_work_queue(&perf_duration_work)) {
301 early_printk("perf interrupt took too long (%lld > %lld), lowering "
302 "kernel.perf_event_max_sample_rate to %d\n",
303 avg_local_sample_len, allowed_ns >> 1,
304 sysctl_perf_event_sample_rate);
308 static atomic64_t perf_event_id;
310 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
311 enum event_type_t event_type);
313 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
314 enum event_type_t event_type,
315 struct task_struct *task);
317 static void update_context_time(struct perf_event_context *ctx);
318 static u64 perf_event_time(struct perf_event *event);
320 void __weak perf_event_print_debug(void) { }
322 extern __weak const char *perf_pmu_name(void)
327 static inline u64 perf_clock(void)
329 return local_clock();
332 static inline u64 perf_event_clock(struct perf_event *event)
334 return event->clock();
337 static inline struct perf_cpu_context *
338 __get_cpu_context(struct perf_event_context *ctx)
340 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
343 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
344 struct perf_event_context *ctx)
346 raw_spin_lock(&cpuctx->ctx.lock);
348 raw_spin_lock(&ctx->lock);
351 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
352 struct perf_event_context *ctx)
355 raw_spin_unlock(&ctx->lock);
356 raw_spin_unlock(&cpuctx->ctx.lock);
359 #ifdef CONFIG_CGROUP_PERF
362 perf_cgroup_match(struct perf_event *event)
364 struct perf_event_context *ctx = event->ctx;
365 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
367 /* @event doesn't care about cgroup */
371 /* wants specific cgroup scope but @cpuctx isn't associated with any */
376 * Cgroup scoping is recursive. An event enabled for a cgroup is
377 * also enabled for all its descendant cgroups. If @cpuctx's
378 * cgroup is a descendant of @event's (the test covers identity
379 * case), it's a match.
381 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
382 event->cgrp->css.cgroup);
385 static inline void perf_detach_cgroup(struct perf_event *event)
387 css_put(&event->cgrp->css);
391 static inline int is_cgroup_event(struct perf_event *event)
393 return event->cgrp != NULL;
396 static inline u64 perf_cgroup_event_time(struct perf_event *event)
398 struct perf_cgroup_info *t;
400 t = per_cpu_ptr(event->cgrp->info, event->cpu);
404 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
406 struct perf_cgroup_info *info;
411 info = this_cpu_ptr(cgrp->info);
413 info->time += now - info->timestamp;
414 info->timestamp = now;
417 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
419 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
421 __update_cgrp_time(cgrp_out);
424 static inline void update_cgrp_time_from_event(struct perf_event *event)
426 struct perf_cgroup *cgrp;
429 * ensure we access cgroup data only when needed and
430 * when we know the cgroup is pinned (css_get)
432 if (!is_cgroup_event(event))
435 cgrp = perf_cgroup_from_task(current);
437 * Do not update time when cgroup is not active
439 if (cgrp == event->cgrp)
440 __update_cgrp_time(event->cgrp);
444 perf_cgroup_set_timestamp(struct task_struct *task,
445 struct perf_event_context *ctx)
447 struct perf_cgroup *cgrp;
448 struct perf_cgroup_info *info;
451 * ctx->lock held by caller
452 * ensure we do not access cgroup data
453 * unless we have the cgroup pinned (css_get)
455 if (!task || !ctx->nr_cgroups)
458 cgrp = perf_cgroup_from_task(task);
459 info = this_cpu_ptr(cgrp->info);
460 info->timestamp = ctx->timestamp;
463 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
464 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
467 * reschedule events based on the cgroup constraint of task.
469 * mode SWOUT : schedule out everything
470 * mode SWIN : schedule in based on cgroup for next
472 void perf_cgroup_switch(struct task_struct *task, int mode)
474 struct perf_cpu_context *cpuctx;
479 * disable interrupts to avoid geting nr_cgroup
480 * changes via __perf_event_disable(). Also
483 local_irq_save(flags);
486 * we reschedule only in the presence of cgroup
487 * constrained events.
491 list_for_each_entry_rcu(pmu, &pmus, entry) {
492 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
493 if (cpuctx->unique_pmu != pmu)
494 continue; /* ensure we process each cpuctx once */
497 * perf_cgroup_events says at least one
498 * context on this CPU has cgroup events.
500 * ctx->nr_cgroups reports the number of cgroup
501 * events for a context.
503 if (cpuctx->ctx.nr_cgroups > 0) {
504 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
505 perf_pmu_disable(cpuctx->ctx.pmu);
507 if (mode & PERF_CGROUP_SWOUT) {
508 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
510 * must not be done before ctxswout due
511 * to event_filter_match() in event_sched_out()
516 if (mode & PERF_CGROUP_SWIN) {
517 WARN_ON_ONCE(cpuctx->cgrp);
519 * set cgrp before ctxsw in to allow
520 * event_filter_match() to not have to pass
523 cpuctx->cgrp = perf_cgroup_from_task(task);
524 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
526 perf_pmu_enable(cpuctx->ctx.pmu);
527 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
533 local_irq_restore(flags);
536 static inline void perf_cgroup_sched_out(struct task_struct *task,
537 struct task_struct *next)
539 struct perf_cgroup *cgrp1;
540 struct perf_cgroup *cgrp2 = NULL;
543 * we come here when we know perf_cgroup_events > 0
545 cgrp1 = perf_cgroup_from_task(task);
548 * next is NULL when called from perf_event_enable_on_exec()
549 * that will systematically cause a cgroup_switch()
552 cgrp2 = perf_cgroup_from_task(next);
555 * only schedule out current cgroup events if we know
556 * that we are switching to a different cgroup. Otherwise,
557 * do no touch the cgroup events.
560 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
563 static inline void perf_cgroup_sched_in(struct task_struct *prev,
564 struct task_struct *task)
566 struct perf_cgroup *cgrp1;
567 struct perf_cgroup *cgrp2 = NULL;
570 * we come here when we know perf_cgroup_events > 0
572 cgrp1 = perf_cgroup_from_task(task);
574 /* prev can never be NULL */
575 cgrp2 = perf_cgroup_from_task(prev);
578 * only need to schedule in cgroup events if we are changing
579 * cgroup during ctxsw. Cgroup events were not scheduled
580 * out of ctxsw out if that was not the case.
583 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
586 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
587 struct perf_event_attr *attr,
588 struct perf_event *group_leader)
590 struct perf_cgroup *cgrp;
591 struct cgroup_subsys_state *css;
592 struct fd f = fdget(fd);
598 css = css_tryget_online_from_dir(f.file->f_path.dentry,
599 &perf_event_cgrp_subsys);
605 cgrp = container_of(css, struct perf_cgroup, css);
609 * all events in a group must monitor
610 * the same cgroup because a task belongs
611 * to only one perf cgroup at a time
613 if (group_leader && group_leader->cgrp != cgrp) {
614 perf_detach_cgroup(event);
623 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
625 struct perf_cgroup_info *t;
626 t = per_cpu_ptr(event->cgrp->info, event->cpu);
627 event->shadow_ctx_time = now - t->timestamp;
631 perf_cgroup_defer_enabled(struct perf_event *event)
634 * when the current task's perf cgroup does not match
635 * the event's, we need to remember to call the
636 * perf_mark_enable() function the first time a task with
637 * a matching perf cgroup is scheduled in.
639 if (is_cgroup_event(event) && !perf_cgroup_match(event))
640 event->cgrp_defer_enabled = 1;
644 perf_cgroup_mark_enabled(struct perf_event *event,
645 struct perf_event_context *ctx)
647 struct perf_event *sub;
648 u64 tstamp = perf_event_time(event);
650 if (!event->cgrp_defer_enabled)
653 event->cgrp_defer_enabled = 0;
655 event->tstamp_enabled = tstamp - event->total_time_enabled;
656 list_for_each_entry(sub, &event->sibling_list, group_entry) {
657 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
658 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
659 sub->cgrp_defer_enabled = 0;
663 #else /* !CONFIG_CGROUP_PERF */
666 perf_cgroup_match(struct perf_event *event)
671 static inline void perf_detach_cgroup(struct perf_event *event)
674 static inline int is_cgroup_event(struct perf_event *event)
679 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
684 static inline void update_cgrp_time_from_event(struct perf_event *event)
688 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
692 static inline void perf_cgroup_sched_out(struct task_struct *task,
693 struct task_struct *next)
697 static inline void perf_cgroup_sched_in(struct task_struct *prev,
698 struct task_struct *task)
702 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
703 struct perf_event_attr *attr,
704 struct perf_event *group_leader)
710 perf_cgroup_set_timestamp(struct task_struct *task,
711 struct perf_event_context *ctx)
716 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
721 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
725 static inline u64 perf_cgroup_event_time(struct perf_event *event)
731 perf_cgroup_defer_enabled(struct perf_event *event)
736 perf_cgroup_mark_enabled(struct perf_event *event,
737 struct perf_event_context *ctx)
743 * set default to be dependent on timer tick just
746 #define PERF_CPU_HRTIMER (1000 / HZ)
748 * function must be called with interrupts disbled
750 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
752 struct perf_cpu_context *cpuctx;
753 enum hrtimer_restart ret = HRTIMER_NORESTART;
756 WARN_ON(!irqs_disabled());
758 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
760 rotations = perf_rotate_context(cpuctx);
763 * arm timer if needed
766 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
767 ret = HRTIMER_RESTART;
773 /* CPU is going down */
774 void perf_cpu_hrtimer_cancel(int cpu)
776 struct perf_cpu_context *cpuctx;
780 if (WARN_ON(cpu != smp_processor_id()))
783 local_irq_save(flags);
787 list_for_each_entry_rcu(pmu, &pmus, entry) {
788 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
790 if (pmu->task_ctx_nr == perf_sw_context)
793 hrtimer_cancel(&cpuctx->hrtimer);
798 local_irq_restore(flags);
801 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
803 struct hrtimer *hr = &cpuctx->hrtimer;
804 struct pmu *pmu = cpuctx->ctx.pmu;
807 /* no multiplexing needed for SW PMU */
808 if (pmu->task_ctx_nr == perf_sw_context)
812 * check default is sane, if not set then force to
813 * default interval (1/tick)
815 timer = pmu->hrtimer_interval_ms;
817 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
819 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
821 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
822 hr->function = perf_cpu_hrtimer_handler;
825 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
827 struct hrtimer *hr = &cpuctx->hrtimer;
828 struct pmu *pmu = cpuctx->ctx.pmu;
831 if (pmu->task_ctx_nr == perf_sw_context)
834 if (hrtimer_active(hr))
837 if (!hrtimer_callback_running(hr))
838 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
839 0, HRTIMER_MODE_REL_PINNED, 0);
842 void perf_pmu_disable(struct pmu *pmu)
844 int *count = this_cpu_ptr(pmu->pmu_disable_count);
846 pmu->pmu_disable(pmu);
849 void perf_pmu_enable(struct pmu *pmu)
851 int *count = this_cpu_ptr(pmu->pmu_disable_count);
853 pmu->pmu_enable(pmu);
856 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
859 * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
860 * perf_event_task_tick() are fully serialized because they're strictly cpu
861 * affine and perf_event_ctx{activate,deactivate} are called with IRQs
862 * disabled, while perf_event_task_tick is called from IRQ context.
864 static void perf_event_ctx_activate(struct perf_event_context *ctx)
866 struct list_head *head = this_cpu_ptr(&active_ctx_list);
868 WARN_ON(!irqs_disabled());
870 WARN_ON(!list_empty(&ctx->active_ctx_list));
872 list_add(&ctx->active_ctx_list, head);
875 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
877 WARN_ON(!irqs_disabled());
879 WARN_ON(list_empty(&ctx->active_ctx_list));
881 list_del_init(&ctx->active_ctx_list);
884 static void get_ctx(struct perf_event_context *ctx)
886 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
889 static void free_ctx(struct rcu_head *head)
891 struct perf_event_context *ctx;
893 ctx = container_of(head, struct perf_event_context, rcu_head);
894 kfree(ctx->task_ctx_data);
898 static void put_ctx(struct perf_event_context *ctx)
900 if (atomic_dec_and_test(&ctx->refcount)) {
902 put_ctx(ctx->parent_ctx);
904 put_task_struct(ctx->task);
905 call_rcu(&ctx->rcu_head, free_ctx);
910 * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
911 * perf_pmu_migrate_context() we need some magic.
913 * Those places that change perf_event::ctx will hold both
914 * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
916 * Lock ordering is by mutex address. There is one other site where
917 * perf_event_context::mutex nests and that is put_event(). But remember that
918 * that is a parent<->child context relation, and migration does not affect
919 * children, therefore these two orderings should not interact.
921 * The change in perf_event::ctx does not affect children (as claimed above)
922 * because the sys_perf_event_open() case will install a new event and break
923 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
924 * concerned with cpuctx and that doesn't have children.
926 * The places that change perf_event::ctx will issue:
928 * perf_remove_from_context();
930 * perf_install_in_context();
932 * to affect the change. The remove_from_context() + synchronize_rcu() should
933 * quiesce the event, after which we can install it in the new location. This
934 * means that only external vectors (perf_fops, prctl) can perturb the event
935 * while in transit. Therefore all such accessors should also acquire
936 * perf_event_context::mutex to serialize against this.
938 * However; because event->ctx can change while we're waiting to acquire
939 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
943 * task_struct::perf_event_mutex
944 * perf_event_context::mutex
945 * perf_event_context::lock
946 * perf_event::child_mutex;
947 * perf_event::mmap_mutex
950 static struct perf_event_context *
951 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
953 struct perf_event_context *ctx;
957 ctx = ACCESS_ONCE(event->ctx);
958 if (!atomic_inc_not_zero(&ctx->refcount)) {
964 mutex_lock_nested(&ctx->mutex, nesting);
965 if (event->ctx != ctx) {
966 mutex_unlock(&ctx->mutex);
974 static inline struct perf_event_context *
975 perf_event_ctx_lock(struct perf_event *event)
977 return perf_event_ctx_lock_nested(event, 0);
980 static void perf_event_ctx_unlock(struct perf_event *event,
981 struct perf_event_context *ctx)
983 mutex_unlock(&ctx->mutex);
988 * This must be done under the ctx->lock, such as to serialize against
989 * context_equiv(), therefore we cannot call put_ctx() since that might end up
990 * calling scheduler related locks and ctx->lock nests inside those.
992 static __must_check struct perf_event_context *
993 unclone_ctx(struct perf_event_context *ctx)
995 struct perf_event_context *parent_ctx = ctx->parent_ctx;
997 lockdep_assert_held(&ctx->lock);
1000 ctx->parent_ctx = NULL;
1006 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1009 * only top level events have the pid namespace they were created in
1012 event = event->parent;
1014 return task_tgid_nr_ns(p, event->ns);
1017 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1020 * only top level events have the pid namespace they were created in
1023 event = event->parent;
1025 return task_pid_nr_ns(p, event->ns);
1029 * If we inherit events we want to return the parent event id
1032 static u64 primary_event_id(struct perf_event *event)
1037 id = event->parent->id;
1043 * Get the perf_event_context for a task and lock it.
1044 * This has to cope with with the fact that until it is locked,
1045 * the context could get moved to another task.
1047 static struct perf_event_context *
1048 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1050 struct perf_event_context *ctx;
1054 * One of the few rules of preemptible RCU is that one cannot do
1055 * rcu_read_unlock() while holding a scheduler (or nested) lock when
1056 * part of the read side critical section was preemptible -- see
1057 * rcu_read_unlock_special().
1059 * Since ctx->lock nests under rq->lock we must ensure the entire read
1060 * side critical section is non-preemptible.
1064 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1067 * If this context is a clone of another, it might
1068 * get swapped for another underneath us by
1069 * perf_event_task_sched_out, though the
1070 * rcu_read_lock() protects us from any context
1071 * getting freed. Lock the context and check if it
1072 * got swapped before we could get the lock, and retry
1073 * if so. If we locked the right context, then it
1074 * can't get swapped on us any more.
1076 raw_spin_lock_irqsave(&ctx->lock, *flags);
1077 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1078 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1084 if (!atomic_inc_not_zero(&ctx->refcount)) {
1085 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1095 * Get the context for a task and increment its pin_count so it
1096 * can't get swapped to another task. This also increments its
1097 * reference count so that the context can't get freed.
1099 static struct perf_event_context *
1100 perf_pin_task_context(struct task_struct *task, int ctxn)
1102 struct perf_event_context *ctx;
1103 unsigned long flags;
1105 ctx = perf_lock_task_context(task, ctxn, &flags);
1108 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1113 static void perf_unpin_context(struct perf_event_context *ctx)
1115 unsigned long flags;
1117 raw_spin_lock_irqsave(&ctx->lock, flags);
1119 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1123 * Update the record of the current time in a context.
1125 static void update_context_time(struct perf_event_context *ctx)
1127 u64 now = perf_clock();
1129 ctx->time += now - ctx->timestamp;
1130 ctx->timestamp = now;
1133 static u64 perf_event_time(struct perf_event *event)
1135 struct perf_event_context *ctx = event->ctx;
1137 if (is_cgroup_event(event))
1138 return perf_cgroup_event_time(event);
1140 return ctx ? ctx->time : 0;
1144 * Update the total_time_enabled and total_time_running fields for a event.
1145 * The caller of this function needs to hold the ctx->lock.
1147 static void update_event_times(struct perf_event *event)
1149 struct perf_event_context *ctx = event->ctx;
1152 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1153 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1156 * in cgroup mode, time_enabled represents
1157 * the time the event was enabled AND active
1158 * tasks were in the monitored cgroup. This is
1159 * independent of the activity of the context as
1160 * there may be a mix of cgroup and non-cgroup events.
1162 * That is why we treat cgroup events differently
1165 if (is_cgroup_event(event))
1166 run_end = perf_cgroup_event_time(event);
1167 else if (ctx->is_active)
1168 run_end = ctx->time;
1170 run_end = event->tstamp_stopped;
1172 event->total_time_enabled = run_end - event->tstamp_enabled;
1174 if (event->state == PERF_EVENT_STATE_INACTIVE)
1175 run_end = event->tstamp_stopped;
1177 run_end = perf_event_time(event);
1179 event->total_time_running = run_end - event->tstamp_running;
1184 * Update total_time_enabled and total_time_running for all events in a group.
1186 static void update_group_times(struct perf_event *leader)
1188 struct perf_event *event;
1190 update_event_times(leader);
1191 list_for_each_entry(event, &leader->sibling_list, group_entry)
1192 update_event_times(event);
1195 static struct list_head *
1196 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1198 if (event->attr.pinned)
1199 return &ctx->pinned_groups;
1201 return &ctx->flexible_groups;
1205 * Add a event from the lists for its context.
1206 * Must be called with ctx->mutex and ctx->lock held.
1209 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1211 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1212 event->attach_state |= PERF_ATTACH_CONTEXT;
1215 * If we're a stand alone event or group leader, we go to the context
1216 * list, group events are kept attached to the group so that
1217 * perf_group_detach can, at all times, locate all siblings.
1219 if (event->group_leader == event) {
1220 struct list_head *list;
1222 if (is_software_event(event))
1223 event->group_flags |= PERF_GROUP_SOFTWARE;
1225 list = ctx_group_list(event, ctx);
1226 list_add_tail(&event->group_entry, list);
1229 if (is_cgroup_event(event))
1232 list_add_rcu(&event->event_entry, &ctx->event_list);
1234 if (event->attr.inherit_stat)
1241 * Initialize event state based on the perf_event_attr::disabled.
1243 static inline void perf_event__state_init(struct perf_event *event)
1245 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1246 PERF_EVENT_STATE_INACTIVE;
1250 * Called at perf_event creation and when events are attached/detached from a
1253 static void perf_event__read_size(struct perf_event *event)
1255 int entry = sizeof(u64); /* value */
1259 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1260 size += sizeof(u64);
1262 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1263 size += sizeof(u64);
1265 if (event->attr.read_format & PERF_FORMAT_ID)
1266 entry += sizeof(u64);
1268 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1269 nr += event->group_leader->nr_siblings;
1270 size += sizeof(u64);
1274 event->read_size = size;
1277 static void perf_event__header_size(struct perf_event *event)
1279 struct perf_sample_data *data;
1280 u64 sample_type = event->attr.sample_type;
1283 perf_event__read_size(event);
1285 if (sample_type & PERF_SAMPLE_IP)
1286 size += sizeof(data->ip);
1288 if (sample_type & PERF_SAMPLE_ADDR)
1289 size += sizeof(data->addr);
1291 if (sample_type & PERF_SAMPLE_PERIOD)
1292 size += sizeof(data->period);
1294 if (sample_type & PERF_SAMPLE_WEIGHT)
1295 size += sizeof(data->weight);
1297 if (sample_type & PERF_SAMPLE_READ)
1298 size += event->read_size;
1300 if (sample_type & PERF_SAMPLE_DATA_SRC)
1301 size += sizeof(data->data_src.val);
1303 if (sample_type & PERF_SAMPLE_TRANSACTION)
1304 size += sizeof(data->txn);
1306 event->header_size = size;
1309 static void perf_event__id_header_size(struct perf_event *event)
1311 struct perf_sample_data *data;
1312 u64 sample_type = event->attr.sample_type;
1315 if (sample_type & PERF_SAMPLE_TID)
1316 size += sizeof(data->tid_entry);
1318 if (sample_type & PERF_SAMPLE_TIME)
1319 size += sizeof(data->time);
1321 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1322 size += sizeof(data->id);
1324 if (sample_type & PERF_SAMPLE_ID)
1325 size += sizeof(data->id);
1327 if (sample_type & PERF_SAMPLE_STREAM_ID)
1328 size += sizeof(data->stream_id);
1330 if (sample_type & PERF_SAMPLE_CPU)
1331 size += sizeof(data->cpu_entry);
1333 event->id_header_size = size;
1336 static void perf_group_attach(struct perf_event *event)
1338 struct perf_event *group_leader = event->group_leader, *pos;
1341 * We can have double attach due to group movement in perf_event_open.
1343 if (event->attach_state & PERF_ATTACH_GROUP)
1346 event->attach_state |= PERF_ATTACH_GROUP;
1348 if (group_leader == event)
1351 WARN_ON_ONCE(group_leader->ctx != event->ctx);
1353 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1354 !is_software_event(event))
1355 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1357 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1358 group_leader->nr_siblings++;
1360 perf_event__header_size(group_leader);
1362 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1363 perf_event__header_size(pos);
1367 * Remove a event from the lists for its context.
1368 * Must be called with ctx->mutex and ctx->lock held.
1371 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1373 struct perf_cpu_context *cpuctx;
1375 WARN_ON_ONCE(event->ctx != ctx);
1376 lockdep_assert_held(&ctx->lock);
1379 * We can have double detach due to exit/hot-unplug + close.
1381 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1384 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1386 if (is_cgroup_event(event)) {
1388 cpuctx = __get_cpu_context(ctx);
1390 * if there are no more cgroup events
1391 * then cler cgrp to avoid stale pointer
1392 * in update_cgrp_time_from_cpuctx()
1394 if (!ctx->nr_cgroups)
1395 cpuctx->cgrp = NULL;
1399 if (event->attr.inherit_stat)
1402 list_del_rcu(&event->event_entry);
1404 if (event->group_leader == event)
1405 list_del_init(&event->group_entry);
1407 update_group_times(event);
1410 * If event was in error state, then keep it
1411 * that way, otherwise bogus counts will be
1412 * returned on read(). The only way to get out
1413 * of error state is by explicit re-enabling
1416 if (event->state > PERF_EVENT_STATE_OFF)
1417 event->state = PERF_EVENT_STATE_OFF;
1422 static void perf_group_detach(struct perf_event *event)
1424 struct perf_event *sibling, *tmp;
1425 struct list_head *list = NULL;
1428 * We can have double detach due to exit/hot-unplug + close.
1430 if (!(event->attach_state & PERF_ATTACH_GROUP))
1433 event->attach_state &= ~PERF_ATTACH_GROUP;
1436 * If this is a sibling, remove it from its group.
1438 if (event->group_leader != event) {
1439 list_del_init(&event->group_entry);
1440 event->group_leader->nr_siblings--;
1444 if (!list_empty(&event->group_entry))
1445 list = &event->group_entry;
1448 * If this was a group event with sibling events then
1449 * upgrade the siblings to singleton events by adding them
1450 * to whatever list we are on.
1452 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1454 list_move_tail(&sibling->group_entry, list);
1455 sibling->group_leader = sibling;
1457 /* Inherit group flags from the previous leader */
1458 sibling->group_flags = event->group_flags;
1460 WARN_ON_ONCE(sibling->ctx != event->ctx);
1464 perf_event__header_size(event->group_leader);
1466 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1467 perf_event__header_size(tmp);
1471 * User event without the task.
1473 static bool is_orphaned_event(struct perf_event *event)
1475 return event && !is_kernel_event(event) && !event->owner;
1479 * Event has a parent but parent's task finished and it's
1480 * alive only because of children holding refference.
1482 static bool is_orphaned_child(struct perf_event *event)
1484 return is_orphaned_event(event->parent);
1487 static void orphans_remove_work(struct work_struct *work);
1489 static void schedule_orphans_remove(struct perf_event_context *ctx)
1491 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1494 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1496 ctx->orphans_remove_sched = true;
1500 static int __init perf_workqueue_init(void)
1502 perf_wq = create_singlethread_workqueue("perf");
1503 WARN(!perf_wq, "failed to create perf workqueue\n");
1504 return perf_wq ? 0 : -1;
1507 core_initcall(perf_workqueue_init);
1509 static inline int pmu_filter_match(struct perf_event *event)
1511 struct pmu *pmu = event->pmu;
1512 return pmu->filter_match ? pmu->filter_match(event) : 1;
1516 event_filter_match(struct perf_event *event)
1518 return (event->cpu == -1 || event->cpu == smp_processor_id())
1519 && perf_cgroup_match(event) && pmu_filter_match(event);
1523 event_sched_out(struct perf_event *event,
1524 struct perf_cpu_context *cpuctx,
1525 struct perf_event_context *ctx)
1527 u64 tstamp = perf_event_time(event);
1530 WARN_ON_ONCE(event->ctx != ctx);
1531 lockdep_assert_held(&ctx->lock);
1534 * An event which could not be activated because of
1535 * filter mismatch still needs to have its timings
1536 * maintained, otherwise bogus information is return
1537 * via read() for time_enabled, time_running:
1539 if (event->state == PERF_EVENT_STATE_INACTIVE
1540 && !event_filter_match(event)) {
1541 delta = tstamp - event->tstamp_stopped;
1542 event->tstamp_running += delta;
1543 event->tstamp_stopped = tstamp;
1546 if (event->state != PERF_EVENT_STATE_ACTIVE)
1549 perf_pmu_disable(event->pmu);
1551 event->state = PERF_EVENT_STATE_INACTIVE;
1552 if (event->pending_disable) {
1553 event->pending_disable = 0;
1554 event->state = PERF_EVENT_STATE_OFF;
1556 event->tstamp_stopped = tstamp;
1557 event->pmu->del(event, 0);
1560 if (!is_software_event(event))
1561 cpuctx->active_oncpu--;
1562 if (!--ctx->nr_active)
1563 perf_event_ctx_deactivate(ctx);
1564 if (event->attr.freq && event->attr.sample_freq)
1566 if (event->attr.exclusive || !cpuctx->active_oncpu)
1567 cpuctx->exclusive = 0;
1569 if (is_orphaned_child(event))
1570 schedule_orphans_remove(ctx);
1572 perf_pmu_enable(event->pmu);
1576 group_sched_out(struct perf_event *group_event,
1577 struct perf_cpu_context *cpuctx,
1578 struct perf_event_context *ctx)
1580 struct perf_event *event;
1581 int state = group_event->state;
1583 event_sched_out(group_event, cpuctx, ctx);
1586 * Schedule out siblings (if any):
1588 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1589 event_sched_out(event, cpuctx, ctx);
1591 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1592 cpuctx->exclusive = 0;
1595 struct remove_event {
1596 struct perf_event *event;
1601 * Cross CPU call to remove a performance event
1603 * We disable the event on the hardware level first. After that we
1604 * remove it from the context list.
1606 static int __perf_remove_from_context(void *info)
1608 struct remove_event *re = info;
1609 struct perf_event *event = re->event;
1610 struct perf_event_context *ctx = event->ctx;
1611 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1613 raw_spin_lock(&ctx->lock);
1614 event_sched_out(event, cpuctx, ctx);
1615 if (re->detach_group)
1616 perf_group_detach(event);
1617 list_del_event(event, ctx);
1618 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1620 cpuctx->task_ctx = NULL;
1622 raw_spin_unlock(&ctx->lock);
1629 * Remove the event from a task's (or a CPU's) list of events.
1631 * CPU events are removed with a smp call. For task events we only
1632 * call when the task is on a CPU.
1634 * If event->ctx is a cloned context, callers must make sure that
1635 * every task struct that event->ctx->task could possibly point to
1636 * remains valid. This is OK when called from perf_release since
1637 * that only calls us on the top-level context, which can't be a clone.
1638 * When called from perf_event_exit_task, it's OK because the
1639 * context has been detached from its task.
1641 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1643 struct perf_event_context *ctx = event->ctx;
1644 struct task_struct *task = ctx->task;
1645 struct remove_event re = {
1647 .detach_group = detach_group,
1650 lockdep_assert_held(&ctx->mutex);
1654 * Per cpu events are removed via an smp call. The removal can
1655 * fail if the CPU is currently offline, but in that case we
1656 * already called __perf_remove_from_context from
1657 * perf_event_exit_cpu.
1659 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1664 if (!task_function_call(task, __perf_remove_from_context, &re))
1667 raw_spin_lock_irq(&ctx->lock);
1669 * If we failed to find a running task, but find the context active now
1670 * that we've acquired the ctx->lock, retry.
1672 if (ctx->is_active) {
1673 raw_spin_unlock_irq(&ctx->lock);
1675 * Reload the task pointer, it might have been changed by
1676 * a concurrent perf_event_context_sched_out().
1683 * Since the task isn't running, its safe to remove the event, us
1684 * holding the ctx->lock ensures the task won't get scheduled in.
1687 perf_group_detach(event);
1688 list_del_event(event, ctx);
1689 raw_spin_unlock_irq(&ctx->lock);
1693 * Cross CPU call to disable a performance event
1695 int __perf_event_disable(void *info)
1697 struct perf_event *event = info;
1698 struct perf_event_context *ctx = event->ctx;
1699 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1702 * If this is a per-task event, need to check whether this
1703 * event's task is the current task on this cpu.
1705 * Can trigger due to concurrent perf_event_context_sched_out()
1706 * flipping contexts around.
1708 if (ctx->task && cpuctx->task_ctx != ctx)
1711 raw_spin_lock(&ctx->lock);
1714 * If the event is on, turn it off.
1715 * If it is in error state, leave it in error state.
1717 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1718 update_context_time(ctx);
1719 update_cgrp_time_from_event(event);
1720 update_group_times(event);
1721 if (event == event->group_leader)
1722 group_sched_out(event, cpuctx, ctx);
1724 event_sched_out(event, cpuctx, ctx);
1725 event->state = PERF_EVENT_STATE_OFF;
1728 raw_spin_unlock(&ctx->lock);
1736 * If event->ctx is a cloned context, callers must make sure that
1737 * every task struct that event->ctx->task could possibly point to
1738 * remains valid. This condition is satisifed when called through
1739 * perf_event_for_each_child or perf_event_for_each because they
1740 * hold the top-level event's child_mutex, so any descendant that
1741 * goes to exit will block in sync_child_event.
1742 * When called from perf_pending_event it's OK because event->ctx
1743 * is the current context on this CPU and preemption is disabled,
1744 * hence we can't get into perf_event_task_sched_out for this context.
1746 static void _perf_event_disable(struct perf_event *event)
1748 struct perf_event_context *ctx = event->ctx;
1749 struct task_struct *task = ctx->task;
1753 * Disable the event on the cpu that it's on
1755 cpu_function_call(event->cpu, __perf_event_disable, event);
1760 if (!task_function_call(task, __perf_event_disable, event))
1763 raw_spin_lock_irq(&ctx->lock);
1765 * If the event is still active, we need to retry the cross-call.
1767 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1768 raw_spin_unlock_irq(&ctx->lock);
1770 * Reload the task pointer, it might have been changed by
1771 * a concurrent perf_event_context_sched_out().
1778 * Since we have the lock this context can't be scheduled
1779 * in, so we can change the state safely.
1781 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1782 update_group_times(event);
1783 event->state = PERF_EVENT_STATE_OFF;
1785 raw_spin_unlock_irq(&ctx->lock);
1789 * Strictly speaking kernel users cannot create groups and therefore this
1790 * interface does not need the perf_event_ctx_lock() magic.
1792 void perf_event_disable(struct perf_event *event)
1794 struct perf_event_context *ctx;
1796 ctx = perf_event_ctx_lock(event);
1797 _perf_event_disable(event);
1798 perf_event_ctx_unlock(event, ctx);
1800 EXPORT_SYMBOL_GPL(perf_event_disable);
1802 static void perf_set_shadow_time(struct perf_event *event,
1803 struct perf_event_context *ctx,
1807 * use the correct time source for the time snapshot
1809 * We could get by without this by leveraging the
1810 * fact that to get to this function, the caller
1811 * has most likely already called update_context_time()
1812 * and update_cgrp_time_xx() and thus both timestamp
1813 * are identical (or very close). Given that tstamp is,
1814 * already adjusted for cgroup, we could say that:
1815 * tstamp - ctx->timestamp
1817 * tstamp - cgrp->timestamp.
1819 * Then, in perf_output_read(), the calculation would
1820 * work with no changes because:
1821 * - event is guaranteed scheduled in
1822 * - no scheduled out in between
1823 * - thus the timestamp would be the same
1825 * But this is a bit hairy.
1827 * So instead, we have an explicit cgroup call to remain
1828 * within the time time source all along. We believe it
1829 * is cleaner and simpler to understand.
1831 if (is_cgroup_event(event))
1832 perf_cgroup_set_shadow_time(event, tstamp);
1834 event->shadow_ctx_time = tstamp - ctx->timestamp;
1837 #define MAX_INTERRUPTS (~0ULL)
1839 static void perf_log_throttle(struct perf_event *event, int enable);
1840 static void perf_log_itrace_start(struct perf_event *event);
1843 event_sched_in(struct perf_event *event,
1844 struct perf_cpu_context *cpuctx,
1845 struct perf_event_context *ctx)
1847 u64 tstamp = perf_event_time(event);
1850 lockdep_assert_held(&ctx->lock);
1852 if (event->state <= PERF_EVENT_STATE_OFF)
1855 event->state = PERF_EVENT_STATE_ACTIVE;
1856 event->oncpu = smp_processor_id();
1859 * Unthrottle events, since we scheduled we might have missed several
1860 * ticks already, also for a heavily scheduling task there is little
1861 * guarantee it'll get a tick in a timely manner.
1863 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1864 perf_log_throttle(event, 1);
1865 event->hw.interrupts = 0;
1869 * The new state must be visible before we turn it on in the hardware:
1873 perf_pmu_disable(event->pmu);
1875 event->tstamp_running += tstamp - event->tstamp_stopped;
1877 perf_set_shadow_time(event, ctx, tstamp);
1879 perf_log_itrace_start(event);
1881 if (event->pmu->add(event, PERF_EF_START)) {
1882 event->state = PERF_EVENT_STATE_INACTIVE;
1888 if (!is_software_event(event))
1889 cpuctx->active_oncpu++;
1890 if (!ctx->nr_active++)
1891 perf_event_ctx_activate(ctx);
1892 if (event->attr.freq && event->attr.sample_freq)
1895 if (event->attr.exclusive)
1896 cpuctx->exclusive = 1;
1898 if (is_orphaned_child(event))
1899 schedule_orphans_remove(ctx);
1902 perf_pmu_enable(event->pmu);
1908 group_sched_in(struct perf_event *group_event,
1909 struct perf_cpu_context *cpuctx,
1910 struct perf_event_context *ctx)
1912 struct perf_event *event, *partial_group = NULL;
1913 struct pmu *pmu = ctx->pmu;
1914 u64 now = ctx->time;
1915 bool simulate = false;
1917 if (group_event->state == PERF_EVENT_STATE_OFF)
1920 pmu->start_txn(pmu);
1922 if (event_sched_in(group_event, cpuctx, ctx)) {
1923 pmu->cancel_txn(pmu);
1924 perf_cpu_hrtimer_restart(cpuctx);
1929 * Schedule in siblings as one group (if any):
1931 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1932 if (event_sched_in(event, cpuctx, ctx)) {
1933 partial_group = event;
1938 if (!pmu->commit_txn(pmu))
1943 * Groups can be scheduled in as one unit only, so undo any
1944 * partial group before returning:
1945 * The events up to the failed event are scheduled out normally,
1946 * tstamp_stopped will be updated.
1948 * The failed events and the remaining siblings need to have
1949 * their timings updated as if they had gone thru event_sched_in()
1950 * and event_sched_out(). This is required to get consistent timings
1951 * across the group. This also takes care of the case where the group
1952 * could never be scheduled by ensuring tstamp_stopped is set to mark
1953 * the time the event was actually stopped, such that time delta
1954 * calculation in update_event_times() is correct.
1956 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1957 if (event == partial_group)
1961 event->tstamp_running += now - event->tstamp_stopped;
1962 event->tstamp_stopped = now;
1964 event_sched_out(event, cpuctx, ctx);
1967 event_sched_out(group_event, cpuctx, ctx);
1969 pmu->cancel_txn(pmu);
1971 perf_cpu_hrtimer_restart(cpuctx);
1977 * Work out whether we can put this event group on the CPU now.
1979 static int group_can_go_on(struct perf_event *event,
1980 struct perf_cpu_context *cpuctx,
1984 * Groups consisting entirely of software events can always go on.
1986 if (event->group_flags & PERF_GROUP_SOFTWARE)
1989 * If an exclusive group is already on, no other hardware
1992 if (cpuctx->exclusive)
1995 * If this group is exclusive and there are already
1996 * events on the CPU, it can't go on.
1998 if (event->attr.exclusive && cpuctx->active_oncpu)
2001 * Otherwise, try to add it if all previous groups were able
2007 static void add_event_to_ctx(struct perf_event *event,
2008 struct perf_event_context *ctx)
2010 u64 tstamp = perf_event_time(event);
2012 list_add_event(event, ctx);
2013 perf_group_attach(event);
2014 event->tstamp_enabled = tstamp;
2015 event->tstamp_running = tstamp;
2016 event->tstamp_stopped = tstamp;
2019 static void task_ctx_sched_out(struct perf_event_context *ctx);
2021 ctx_sched_in(struct perf_event_context *ctx,
2022 struct perf_cpu_context *cpuctx,
2023 enum event_type_t event_type,
2024 struct task_struct *task);
2026 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2027 struct perf_event_context *ctx,
2028 struct task_struct *task)
2030 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2032 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2033 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2035 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2039 * Cross CPU call to install and enable a performance event
2041 * Must be called with ctx->mutex held
2043 static int __perf_install_in_context(void *info)
2045 struct perf_event *event = info;
2046 struct perf_event_context *ctx = event->ctx;
2047 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2048 struct perf_event_context *task_ctx = cpuctx->task_ctx;
2049 struct task_struct *task = current;
2051 perf_ctx_lock(cpuctx, task_ctx);
2052 perf_pmu_disable(cpuctx->ctx.pmu);
2055 * If there was an active task_ctx schedule it out.
2058 task_ctx_sched_out(task_ctx);
2061 * If the context we're installing events in is not the
2062 * active task_ctx, flip them.
2064 if (ctx->task && task_ctx != ctx) {
2066 raw_spin_unlock(&task_ctx->lock);
2067 raw_spin_lock(&ctx->lock);
2072 cpuctx->task_ctx = task_ctx;
2073 task = task_ctx->task;
2076 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
2078 update_context_time(ctx);
2080 * update cgrp time only if current cgrp
2081 * matches event->cgrp. Must be done before
2082 * calling add_event_to_ctx()
2084 update_cgrp_time_from_event(event);
2086 add_event_to_ctx(event, ctx);
2089 * Schedule everything back in
2091 perf_event_sched_in(cpuctx, task_ctx, task);
2093 perf_pmu_enable(cpuctx->ctx.pmu);
2094 perf_ctx_unlock(cpuctx, task_ctx);
2100 * Attach a performance event to a context
2102 * First we add the event to the list with the hardware enable bit
2103 * in event->hw_config cleared.
2105 * If the event is attached to a task which is on a CPU we use a smp
2106 * call to enable it in the task context. The task might have been
2107 * scheduled away, but we check this in the smp call again.
2110 perf_install_in_context(struct perf_event_context *ctx,
2111 struct perf_event *event,
2114 struct task_struct *task = ctx->task;
2116 lockdep_assert_held(&ctx->mutex);
2119 if (event->cpu != -1)
2124 * Per cpu events are installed via an smp call and
2125 * the install is always successful.
2127 cpu_function_call(cpu, __perf_install_in_context, event);
2132 if (!task_function_call(task, __perf_install_in_context, event))
2135 raw_spin_lock_irq(&ctx->lock);
2137 * If we failed to find a running task, but find the context active now
2138 * that we've acquired the ctx->lock, retry.
2140 if (ctx->is_active) {
2141 raw_spin_unlock_irq(&ctx->lock);
2143 * Reload the task pointer, it might have been changed by
2144 * a concurrent perf_event_context_sched_out().
2151 * Since the task isn't running, its safe to add the event, us holding
2152 * the ctx->lock ensures the task won't get scheduled in.
2154 add_event_to_ctx(event, ctx);
2155 raw_spin_unlock_irq(&ctx->lock);
2159 * Put a event into inactive state and update time fields.
2160 * Enabling the leader of a group effectively enables all
2161 * the group members that aren't explicitly disabled, so we
2162 * have to update their ->tstamp_enabled also.
2163 * Note: this works for group members as well as group leaders
2164 * since the non-leader members' sibling_lists will be empty.
2166 static void __perf_event_mark_enabled(struct perf_event *event)
2168 struct perf_event *sub;
2169 u64 tstamp = perf_event_time(event);
2171 event->state = PERF_EVENT_STATE_INACTIVE;
2172 event->tstamp_enabled = tstamp - event->total_time_enabled;
2173 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2174 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2175 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2180 * Cross CPU call to enable a performance event
2182 static int __perf_event_enable(void *info)
2184 struct perf_event *event = info;
2185 struct perf_event_context *ctx = event->ctx;
2186 struct perf_event *leader = event->group_leader;
2187 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2191 * There's a time window between 'ctx->is_active' check
2192 * in perf_event_enable function and this place having:
2194 * - ctx->lock unlocked
2196 * where the task could be killed and 'ctx' deactivated
2197 * by perf_event_exit_task.
2199 if (!ctx->is_active)
2202 raw_spin_lock(&ctx->lock);
2203 update_context_time(ctx);
2205 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2209 * set current task's cgroup time reference point
2211 perf_cgroup_set_timestamp(current, ctx);
2213 __perf_event_mark_enabled(event);
2215 if (!event_filter_match(event)) {
2216 if (is_cgroup_event(event))
2217 perf_cgroup_defer_enabled(event);
2222 * If the event is in a group and isn't the group leader,
2223 * then don't put it on unless the group is on.
2225 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2228 if (!group_can_go_on(event, cpuctx, 1)) {
2231 if (event == leader)
2232 err = group_sched_in(event, cpuctx, ctx);
2234 err = event_sched_in(event, cpuctx, ctx);
2239 * If this event can't go on and it's part of a
2240 * group, then the whole group has to come off.
2242 if (leader != event) {
2243 group_sched_out(leader, cpuctx, ctx);
2244 perf_cpu_hrtimer_restart(cpuctx);
2246 if (leader->attr.pinned) {
2247 update_group_times(leader);
2248 leader->state = PERF_EVENT_STATE_ERROR;
2253 raw_spin_unlock(&ctx->lock);
2261 * If event->ctx is a cloned context, callers must make sure that
2262 * every task struct that event->ctx->task could possibly point to
2263 * remains valid. This condition is satisfied when called through
2264 * perf_event_for_each_child or perf_event_for_each as described
2265 * for perf_event_disable.
2267 static void _perf_event_enable(struct perf_event *event)
2269 struct perf_event_context *ctx = event->ctx;
2270 struct task_struct *task = ctx->task;
2274 * Enable the event on the cpu that it's on
2276 cpu_function_call(event->cpu, __perf_event_enable, event);
2280 raw_spin_lock_irq(&ctx->lock);
2281 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2285 * If the event is in error state, clear that first.
2286 * That way, if we see the event in error state below, we
2287 * know that it has gone back into error state, as distinct
2288 * from the task having been scheduled away before the
2289 * cross-call arrived.
2291 if (event->state == PERF_EVENT_STATE_ERROR)
2292 event->state = PERF_EVENT_STATE_OFF;
2295 if (!ctx->is_active) {
2296 __perf_event_mark_enabled(event);
2300 raw_spin_unlock_irq(&ctx->lock);
2302 if (!task_function_call(task, __perf_event_enable, event))
2305 raw_spin_lock_irq(&ctx->lock);
2308 * If the context is active and the event is still off,
2309 * we need to retry the cross-call.
2311 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2313 * task could have been flipped by a concurrent
2314 * perf_event_context_sched_out()
2321 raw_spin_unlock_irq(&ctx->lock);
2325 * See perf_event_disable();
2327 void perf_event_enable(struct perf_event *event)
2329 struct perf_event_context *ctx;
2331 ctx = perf_event_ctx_lock(event);
2332 _perf_event_enable(event);
2333 perf_event_ctx_unlock(event, ctx);
2335 EXPORT_SYMBOL_GPL(perf_event_enable);
2337 static int _perf_event_refresh(struct perf_event *event, int refresh)
2340 * not supported on inherited events
2342 if (event->attr.inherit || !is_sampling_event(event))
2345 atomic_add(refresh, &event->event_limit);
2346 _perf_event_enable(event);
2352 * See perf_event_disable()
2354 int perf_event_refresh(struct perf_event *event, int refresh)
2356 struct perf_event_context *ctx;
2359 ctx = perf_event_ctx_lock(event);
2360 ret = _perf_event_refresh(event, refresh);
2361 perf_event_ctx_unlock(event, ctx);
2365 EXPORT_SYMBOL_GPL(perf_event_refresh);
2367 static void ctx_sched_out(struct perf_event_context *ctx,
2368 struct perf_cpu_context *cpuctx,
2369 enum event_type_t event_type)
2371 struct perf_event *event;
2372 int is_active = ctx->is_active;
2374 ctx->is_active &= ~event_type;
2375 if (likely(!ctx->nr_events))
2378 update_context_time(ctx);
2379 update_cgrp_time_from_cpuctx(cpuctx);
2380 if (!ctx->nr_active)
2383 perf_pmu_disable(ctx->pmu);
2384 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2385 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2386 group_sched_out(event, cpuctx, ctx);
2389 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2390 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2391 group_sched_out(event, cpuctx, ctx);
2393 perf_pmu_enable(ctx->pmu);
2397 * Test whether two contexts are equivalent, i.e. whether they have both been
2398 * cloned from the same version of the same context.
2400 * Equivalence is measured using a generation number in the context that is
2401 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2402 * and list_del_event().
2404 static int context_equiv(struct perf_event_context *ctx1,
2405 struct perf_event_context *ctx2)
2407 lockdep_assert_held(&ctx1->lock);
2408 lockdep_assert_held(&ctx2->lock);
2410 /* Pinning disables the swap optimization */
2411 if (ctx1->pin_count || ctx2->pin_count)
2414 /* If ctx1 is the parent of ctx2 */
2415 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2418 /* If ctx2 is the parent of ctx1 */
2419 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2423 * If ctx1 and ctx2 have the same parent; we flatten the parent
2424 * hierarchy, see perf_event_init_context().
2426 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2427 ctx1->parent_gen == ctx2->parent_gen)
2434 static void __perf_event_sync_stat(struct perf_event *event,
2435 struct perf_event *next_event)
2439 if (!event->attr.inherit_stat)
2443 * Update the event value, we cannot use perf_event_read()
2444 * because we're in the middle of a context switch and have IRQs
2445 * disabled, which upsets smp_call_function_single(), however
2446 * we know the event must be on the current CPU, therefore we
2447 * don't need to use it.
2449 switch (event->state) {
2450 case PERF_EVENT_STATE_ACTIVE:
2451 event->pmu->read(event);
2454 case PERF_EVENT_STATE_INACTIVE:
2455 update_event_times(event);
2463 * In order to keep per-task stats reliable we need to flip the event
2464 * values when we flip the contexts.
2466 value = local64_read(&next_event->count);
2467 value = local64_xchg(&event->count, value);
2468 local64_set(&next_event->count, value);
2470 swap(event->total_time_enabled, next_event->total_time_enabled);
2471 swap(event->total_time_running, next_event->total_time_running);
2474 * Since we swizzled the values, update the user visible data too.
2476 perf_event_update_userpage(event);
2477 perf_event_update_userpage(next_event);
2480 static void perf_event_sync_stat(struct perf_event_context *ctx,
2481 struct perf_event_context *next_ctx)
2483 struct perf_event *event, *next_event;
2488 update_context_time(ctx);
2490 event = list_first_entry(&ctx->event_list,
2491 struct perf_event, event_entry);
2493 next_event = list_first_entry(&next_ctx->event_list,
2494 struct perf_event, event_entry);
2496 while (&event->event_entry != &ctx->event_list &&
2497 &next_event->event_entry != &next_ctx->event_list) {
2499 __perf_event_sync_stat(event, next_event);
2501 event = list_next_entry(event, event_entry);
2502 next_event = list_next_entry(next_event, event_entry);
2506 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2507 struct task_struct *next)
2509 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2510 struct perf_event_context *next_ctx;
2511 struct perf_event_context *parent, *next_parent;
2512 struct perf_cpu_context *cpuctx;
2518 cpuctx = __get_cpu_context(ctx);
2519 if (!cpuctx->task_ctx)
2523 next_ctx = next->perf_event_ctxp[ctxn];
2527 parent = rcu_dereference(ctx->parent_ctx);
2528 next_parent = rcu_dereference(next_ctx->parent_ctx);
2530 /* If neither context have a parent context; they cannot be clones. */
2531 if (!parent && !next_parent)
2534 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2536 * Looks like the two contexts are clones, so we might be
2537 * able to optimize the context switch. We lock both
2538 * contexts and check that they are clones under the
2539 * lock (including re-checking that neither has been
2540 * uncloned in the meantime). It doesn't matter which
2541 * order we take the locks because no other cpu could
2542 * be trying to lock both of these tasks.
2544 raw_spin_lock(&ctx->lock);
2545 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2546 if (context_equiv(ctx, next_ctx)) {
2548 * XXX do we need a memory barrier of sorts
2549 * wrt to rcu_dereference() of perf_event_ctxp
2551 task->perf_event_ctxp[ctxn] = next_ctx;
2552 next->perf_event_ctxp[ctxn] = ctx;
2554 next_ctx->task = task;
2556 swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
2560 perf_event_sync_stat(ctx, next_ctx);
2562 raw_spin_unlock(&next_ctx->lock);
2563 raw_spin_unlock(&ctx->lock);
2569 raw_spin_lock(&ctx->lock);
2570 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2571 cpuctx->task_ctx = NULL;
2572 raw_spin_unlock(&ctx->lock);
2576 void perf_sched_cb_dec(struct pmu *pmu)
2578 this_cpu_dec(perf_sched_cb_usages);
2581 void perf_sched_cb_inc(struct pmu *pmu)
2583 this_cpu_inc(perf_sched_cb_usages);
2587 * This function provides the context switch callback to the lower code
2588 * layer. It is invoked ONLY when the context switch callback is enabled.
2590 static void perf_pmu_sched_task(struct task_struct *prev,
2591 struct task_struct *next,
2594 struct perf_cpu_context *cpuctx;
2596 unsigned long flags;
2601 local_irq_save(flags);
2605 list_for_each_entry_rcu(pmu, &pmus, entry) {
2606 if (pmu->sched_task) {
2607 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2609 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2611 perf_pmu_disable(pmu);
2613 pmu->sched_task(cpuctx->task_ctx, sched_in);
2615 perf_pmu_enable(pmu);
2617 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2623 local_irq_restore(flags);
2626 #define for_each_task_context_nr(ctxn) \
2627 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2630 * Called from scheduler to remove the events of the current task,
2631 * with interrupts disabled.
2633 * We stop each event and update the event value in event->count.
2635 * This does not protect us against NMI, but disable()
2636 * sets the disabled bit in the control field of event _before_
2637 * accessing the event control register. If a NMI hits, then it will
2638 * not restart the event.
2640 void __perf_event_task_sched_out(struct task_struct *task,
2641 struct task_struct *next)
2645 if (__this_cpu_read(perf_sched_cb_usages))
2646 perf_pmu_sched_task(task, next, false);
2648 for_each_task_context_nr(ctxn)
2649 perf_event_context_sched_out(task, ctxn, next);
2652 * if cgroup events exist on this CPU, then we need
2653 * to check if we have to switch out PMU state.
2654 * cgroup event are system-wide mode only
2656 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2657 perf_cgroup_sched_out(task, next);
2660 static void task_ctx_sched_out(struct perf_event_context *ctx)
2662 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2664 if (!cpuctx->task_ctx)
2667 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2670 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2671 cpuctx->task_ctx = NULL;
2675 * Called with IRQs disabled
2677 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2678 enum event_type_t event_type)
2680 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2684 ctx_pinned_sched_in(struct perf_event_context *ctx,
2685 struct perf_cpu_context *cpuctx)
2687 struct perf_event *event;
2689 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2690 if (event->state <= PERF_EVENT_STATE_OFF)
2692 if (!event_filter_match(event))
2695 /* may need to reset tstamp_enabled */
2696 if (is_cgroup_event(event))
2697 perf_cgroup_mark_enabled(event, ctx);
2699 if (group_can_go_on(event, cpuctx, 1))
2700 group_sched_in(event, cpuctx, ctx);
2703 * If this pinned group hasn't been scheduled,
2704 * put it in error state.
2706 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2707 update_group_times(event);
2708 event->state = PERF_EVENT_STATE_ERROR;
2714 ctx_flexible_sched_in(struct perf_event_context *ctx,
2715 struct perf_cpu_context *cpuctx)
2717 struct perf_event *event;
2720 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2721 /* Ignore events in OFF or ERROR state */
2722 if (event->state <= PERF_EVENT_STATE_OFF)
2725 * Listen to the 'cpu' scheduling filter constraint
2728 if (!event_filter_match(event))
2731 /* may need to reset tstamp_enabled */
2732 if (is_cgroup_event(event))
2733 perf_cgroup_mark_enabled(event, ctx);
2735 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2736 if (group_sched_in(event, cpuctx, ctx))
2743 ctx_sched_in(struct perf_event_context *ctx,
2744 struct perf_cpu_context *cpuctx,
2745 enum event_type_t event_type,
2746 struct task_struct *task)
2749 int is_active = ctx->is_active;
2751 ctx->is_active |= event_type;
2752 if (likely(!ctx->nr_events))
2756 ctx->timestamp = now;
2757 perf_cgroup_set_timestamp(task, ctx);
2759 * First go through the list and put on any pinned groups
2760 * in order to give them the best chance of going on.
2762 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2763 ctx_pinned_sched_in(ctx, cpuctx);
2765 /* Then walk through the lower prio flexible groups */
2766 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2767 ctx_flexible_sched_in(ctx, cpuctx);
2770 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2771 enum event_type_t event_type,
2772 struct task_struct *task)
2774 struct perf_event_context *ctx = &cpuctx->ctx;
2776 ctx_sched_in(ctx, cpuctx, event_type, task);
2779 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2780 struct task_struct *task)
2782 struct perf_cpu_context *cpuctx;
2784 cpuctx = __get_cpu_context(ctx);
2785 if (cpuctx->task_ctx == ctx)
2788 perf_ctx_lock(cpuctx, ctx);
2789 perf_pmu_disable(ctx->pmu);
2791 * We want to keep the following priority order:
2792 * cpu pinned (that don't need to move), task pinned,
2793 * cpu flexible, task flexible.
2795 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2798 cpuctx->task_ctx = ctx;
2800 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2802 perf_pmu_enable(ctx->pmu);
2803 perf_ctx_unlock(cpuctx, ctx);
2807 * Called from scheduler to add the events of the current task
2808 * with interrupts disabled.
2810 * We restore the event value and then enable it.
2812 * This does not protect us against NMI, but enable()
2813 * sets the enabled bit in the control field of event _before_
2814 * accessing the event control register. If a NMI hits, then it will
2815 * keep the event running.
2817 void __perf_event_task_sched_in(struct task_struct *prev,
2818 struct task_struct *task)
2820 struct perf_event_context *ctx;
2823 for_each_task_context_nr(ctxn) {
2824 ctx = task->perf_event_ctxp[ctxn];
2828 perf_event_context_sched_in(ctx, task);
2831 * if cgroup events exist on this CPU, then we need
2832 * to check if we have to switch in PMU state.
2833 * cgroup event are system-wide mode only
2835 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2836 perf_cgroup_sched_in(prev, task);
2838 if (__this_cpu_read(perf_sched_cb_usages))
2839 perf_pmu_sched_task(prev, task, true);
2842 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2844 u64 frequency = event->attr.sample_freq;
2845 u64 sec = NSEC_PER_SEC;
2846 u64 divisor, dividend;
2848 int count_fls, nsec_fls, frequency_fls, sec_fls;
2850 count_fls = fls64(count);
2851 nsec_fls = fls64(nsec);
2852 frequency_fls = fls64(frequency);
2856 * We got @count in @nsec, with a target of sample_freq HZ
2857 * the target period becomes:
2860 * period = -------------------
2861 * @nsec * sample_freq
2866 * Reduce accuracy by one bit such that @a and @b converge
2867 * to a similar magnitude.
2869 #define REDUCE_FLS(a, b) \
2871 if (a##_fls > b##_fls) { \
2881 * Reduce accuracy until either term fits in a u64, then proceed with
2882 * the other, so that finally we can do a u64/u64 division.
2884 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2885 REDUCE_FLS(nsec, frequency);
2886 REDUCE_FLS(sec, count);
2889 if (count_fls + sec_fls > 64) {
2890 divisor = nsec * frequency;
2892 while (count_fls + sec_fls > 64) {
2893 REDUCE_FLS(count, sec);
2897 dividend = count * sec;
2899 dividend = count * sec;
2901 while (nsec_fls + frequency_fls > 64) {
2902 REDUCE_FLS(nsec, frequency);
2906 divisor = nsec * frequency;
2912 return div64_u64(dividend, divisor);
2915 static DEFINE_PER_CPU(int, perf_throttled_count);
2916 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2918 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2920 struct hw_perf_event *hwc = &event->hw;
2921 s64 period, sample_period;
2924 period = perf_calculate_period(event, nsec, count);
2926 delta = (s64)(period - hwc->sample_period);
2927 delta = (delta + 7) / 8; /* low pass filter */
2929 sample_period = hwc->sample_period + delta;
2934 hwc->sample_period = sample_period;
2936 if (local64_read(&hwc->period_left) > 8*sample_period) {
2938 event->pmu->stop(event, PERF_EF_UPDATE);
2940 local64_set(&hwc->period_left, 0);
2943 event->pmu->start(event, PERF_EF_RELOAD);
2948 * combine freq adjustment with unthrottling to avoid two passes over the
2949 * events. At the same time, make sure, having freq events does not change
2950 * the rate of unthrottling as that would introduce bias.
2952 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2955 struct perf_event *event;
2956 struct hw_perf_event *hwc;
2957 u64 now, period = TICK_NSEC;
2961 * only need to iterate over all events iff:
2962 * - context have events in frequency mode (needs freq adjust)
2963 * - there are events to unthrottle on this cpu
2965 if (!(ctx->nr_freq || needs_unthr))
2968 raw_spin_lock(&ctx->lock);
2969 perf_pmu_disable(ctx->pmu);
2971 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2972 if (event->state != PERF_EVENT_STATE_ACTIVE)
2975 if (!event_filter_match(event))
2978 perf_pmu_disable(event->pmu);
2982 if (hwc->interrupts == MAX_INTERRUPTS) {
2983 hwc->interrupts = 0;
2984 perf_log_throttle(event, 1);
2985 event->pmu->start(event, 0);
2988 if (!event->attr.freq || !event->attr.sample_freq)
2992 * stop the event and update event->count
2994 event->pmu->stop(event, PERF_EF_UPDATE);
2996 now = local64_read(&event->count);
2997 delta = now - hwc->freq_count_stamp;
2998 hwc->freq_count_stamp = now;
3002 * reload only if value has changed
3003 * we have stopped the event so tell that
3004 * to perf_adjust_period() to avoid stopping it
3008 perf_adjust_period(event, period, delta, false);
3010 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3012 perf_pmu_enable(event->pmu);
3015 perf_pmu_enable(ctx->pmu);
3016 raw_spin_unlock(&ctx->lock);
3020 * Round-robin a context's events:
3022 static void rotate_ctx(struct perf_event_context *ctx)
3025 * Rotate the first entry last of non-pinned groups. Rotation might be
3026 * disabled by the inheritance code.
3028 if (!ctx->rotate_disable)
3029 list_rotate_left(&ctx->flexible_groups);
3032 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
3034 struct perf_event_context *ctx = NULL;
3037 if (cpuctx->ctx.nr_events) {
3038 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3042 ctx = cpuctx->task_ctx;
3043 if (ctx && ctx->nr_events) {
3044 if (ctx->nr_events != ctx->nr_active)
3051 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3052 perf_pmu_disable(cpuctx->ctx.pmu);
3054 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3056 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3058 rotate_ctx(&cpuctx->ctx);
3062 perf_event_sched_in(cpuctx, ctx, current);
3064 perf_pmu_enable(cpuctx->ctx.pmu);
3065 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3071 #ifdef CONFIG_NO_HZ_FULL
3072 bool perf_event_can_stop_tick(void)
3074 if (atomic_read(&nr_freq_events) ||
3075 __this_cpu_read(perf_throttled_count))
3082 void perf_event_task_tick(void)
3084 struct list_head *head = this_cpu_ptr(&active_ctx_list);
3085 struct perf_event_context *ctx, *tmp;
3088 WARN_ON(!irqs_disabled());
3090 __this_cpu_inc(perf_throttled_seq);
3091 throttled = __this_cpu_xchg(perf_throttled_count, 0);
3093 list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3094 perf_adjust_freq_unthr_context(ctx, throttled);
3097 static int event_enable_on_exec(struct perf_event *event,
3098 struct perf_event_context *ctx)
3100 if (!event->attr.enable_on_exec)
3103 event->attr.enable_on_exec = 0;
3104 if (event->state >= PERF_EVENT_STATE_INACTIVE)
3107 __perf_event_mark_enabled(event);
3113 * Enable all of a task's events that have been marked enable-on-exec.
3114 * This expects task == current.
3116 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
3118 struct perf_event_context *clone_ctx = NULL;
3119 struct perf_event *event;
3120 unsigned long flags;
3124 local_irq_save(flags);
3125 if (!ctx || !ctx->nr_events)
3129 * We must ctxsw out cgroup events to avoid conflict
3130 * when invoking perf_task_event_sched_in() later on
3131 * in this function. Otherwise we end up trying to
3132 * ctxswin cgroup events which are already scheduled
3135 perf_cgroup_sched_out(current, NULL);
3137 raw_spin_lock(&ctx->lock);
3138 task_ctx_sched_out(ctx);
3140 list_for_each_entry(event, &ctx->event_list, event_entry) {
3141 ret = event_enable_on_exec(event, ctx);
3147 * Unclone this context if we enabled any event.
3150 clone_ctx = unclone_ctx(ctx);
3152 raw_spin_unlock(&ctx->lock);
3155 * Also calls ctxswin for cgroup events, if any:
3157 perf_event_context_sched_in(ctx, ctx->task);
3159 local_irq_restore(flags);
3165 void perf_event_exec(void)
3167 struct perf_event_context *ctx;
3171 for_each_task_context_nr(ctxn) {
3172 ctx = current->perf_event_ctxp[ctxn];
3176 perf_event_enable_on_exec(ctx);
3182 * Cross CPU call to read the hardware event
3184 static void __perf_event_read(void *info)
3186 struct perf_event *event = info;
3187 struct perf_event_context *ctx = event->ctx;
3188 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3191 * If this is a task context, we need to check whether it is
3192 * the current task context of this cpu. If not it has been
3193 * scheduled out before the smp call arrived. In that case
3194 * event->count would have been updated to a recent sample
3195 * when the event was scheduled out.
3197 if (ctx->task && cpuctx->task_ctx != ctx)
3200 raw_spin_lock(&ctx->lock);
3201 if (ctx->is_active) {
3202 update_context_time(ctx);
3203 update_cgrp_time_from_event(event);
3205 update_event_times(event);
3206 if (event->state == PERF_EVENT_STATE_ACTIVE)
3207 event->pmu->read(event);
3208 raw_spin_unlock(&ctx->lock);
3211 static inline u64 perf_event_count(struct perf_event *event)
3213 if (event->pmu->count)
3214 return event->pmu->count(event);
3216 return __perf_event_count(event);
3219 static u64 perf_event_read(struct perf_event *event)
3222 * If event is enabled and currently active on a CPU, update the
3223 * value in the event structure:
3225 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3226 smp_call_function_single(event->oncpu,
3227 __perf_event_read, event, 1);
3228 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3229 struct perf_event_context *ctx = event->ctx;
3230 unsigned long flags;
3232 raw_spin_lock_irqsave(&ctx->lock, flags);
3234 * may read while context is not active
3235 * (e.g., thread is blocked), in that case
3236 * we cannot update context time
3238 if (ctx->is_active) {
3239 update_context_time(ctx);
3240 update_cgrp_time_from_event(event);
3242 update_event_times(event);
3243 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3246 return perf_event_count(event);
3250 * Initialize the perf_event context in a task_struct:
3252 static void __perf_event_init_context(struct perf_event_context *ctx)
3254 raw_spin_lock_init(&ctx->lock);
3255 mutex_init(&ctx->mutex);
3256 INIT_LIST_HEAD(&ctx->active_ctx_list);
3257 INIT_LIST_HEAD(&ctx->pinned_groups);
3258 INIT_LIST_HEAD(&ctx->flexible_groups);
3259 INIT_LIST_HEAD(&ctx->event_list);
3260 atomic_set(&ctx->refcount, 1);
3261 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3264 static struct perf_event_context *
3265 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3267 struct perf_event_context *ctx;
3269 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3273 __perf_event_init_context(ctx);
3276 get_task_struct(task);
3283 static struct task_struct *
3284 find_lively_task_by_vpid(pid_t vpid)
3286 struct task_struct *task;
3293 task = find_task_by_vpid(vpid);
3295 get_task_struct(task);
3299 return ERR_PTR(-ESRCH);
3301 /* Reuse ptrace permission checks for now. */
3303 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3308 put_task_struct(task);
3309 return ERR_PTR(err);
3314 * Returns a matching context with refcount and pincount.
3316 static struct perf_event_context *
3317 find_get_context(struct pmu *pmu, struct task_struct *task,
3318 struct perf_event *event)
3320 struct perf_event_context *ctx, *clone_ctx = NULL;
3321 struct perf_cpu_context *cpuctx;
3322 void *task_ctx_data = NULL;
3323 unsigned long flags;
3325 int cpu = event->cpu;
3328 /* Must be root to operate on a CPU event: */
3329 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3330 return ERR_PTR(-EACCES);
3333 * We could be clever and allow to attach a event to an
3334 * offline CPU and activate it when the CPU comes up, but
3337 if (!cpu_online(cpu))
3338 return ERR_PTR(-ENODEV);
3340 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3349 ctxn = pmu->task_ctx_nr;
3353 if (event->attach_state & PERF_ATTACH_TASK_DATA) {
3354 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
3355 if (!task_ctx_data) {
3362 ctx = perf_lock_task_context(task, ctxn, &flags);
3364 clone_ctx = unclone_ctx(ctx);
3367 if (task_ctx_data && !ctx->task_ctx_data) {
3368 ctx->task_ctx_data = task_ctx_data;
3369 task_ctx_data = NULL;
3371 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3376 ctx = alloc_perf_context(pmu, task);
3381 if (task_ctx_data) {
3382 ctx->task_ctx_data = task_ctx_data;
3383 task_ctx_data = NULL;
3387 mutex_lock(&task->perf_event_mutex);
3389 * If it has already passed perf_event_exit_task().
3390 * we must see PF_EXITING, it takes this mutex too.
3392 if (task->flags & PF_EXITING)
3394 else if (task->perf_event_ctxp[ctxn])
3399 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3401 mutex_unlock(&task->perf_event_mutex);
3403 if (unlikely(err)) {
3412 kfree(task_ctx_data);
3416 kfree(task_ctx_data);
3417 return ERR_PTR(err);
3420 static void perf_event_free_filter(struct perf_event *event);
3421 static void perf_event_free_bpf_prog(struct perf_event *event);
3423 static void free_event_rcu(struct rcu_head *head)
3425 struct perf_event *event;
3427 event = container_of(head, struct perf_event, rcu_head);
3429 put_pid_ns(event->ns);
3430 perf_event_free_filter(event);
3431 perf_event_free_bpf_prog(event);
3435 static void ring_buffer_attach(struct perf_event *event,
3436 struct ring_buffer *rb);
3438 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3443 if (is_cgroup_event(event))
3444 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3447 static void unaccount_event(struct perf_event *event)
3452 if (event->attach_state & PERF_ATTACH_TASK)
3453 static_key_slow_dec_deferred(&perf_sched_events);
3454 if (event->attr.mmap || event->attr.mmap_data)
3455 atomic_dec(&nr_mmap_events);
3456 if (event->attr.comm)
3457 atomic_dec(&nr_comm_events);
3458 if (event->attr.task)
3459 atomic_dec(&nr_task_events);
3460 if (event->attr.freq)
3461 atomic_dec(&nr_freq_events);
3462 if (is_cgroup_event(event))
3463 static_key_slow_dec_deferred(&perf_sched_events);
3464 if (has_branch_stack(event))
3465 static_key_slow_dec_deferred(&perf_sched_events);
3467 unaccount_event_cpu(event, event->cpu);
3471 * The following implement mutual exclusion of events on "exclusive" pmus
3472 * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
3473 * at a time, so we disallow creating events that might conflict, namely:
3475 * 1) cpu-wide events in the presence of per-task events,
3476 * 2) per-task events in the presence of cpu-wide events,
3477 * 3) two matching events on the same context.
3479 * The former two cases are handled in the allocation path (perf_event_alloc(),
3480 * __free_event()), the latter -- before the first perf_install_in_context().
3482 static int exclusive_event_init(struct perf_event *event)
3484 struct pmu *pmu = event->pmu;
3486 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3490 * Prevent co-existence of per-task and cpu-wide events on the
3491 * same exclusive pmu.
3493 * Negative pmu::exclusive_cnt means there are cpu-wide
3494 * events on this "exclusive" pmu, positive means there are
3497 * Since this is called in perf_event_alloc() path, event::ctx
3498 * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
3499 * to mean "per-task event", because unlike other attach states it
3500 * never gets cleared.
3502 if (event->attach_state & PERF_ATTACH_TASK) {
3503 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
3506 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
3513 static void exclusive_event_destroy(struct perf_event *event)
3515 struct pmu *pmu = event->pmu;
3517 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3520 /* see comment in exclusive_event_init() */
3521 if (event->attach_state & PERF_ATTACH_TASK)
3522 atomic_dec(&pmu->exclusive_cnt);
3524 atomic_inc(&pmu->exclusive_cnt);
3527 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
3529 if ((e1->pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) &&
3530 (e1->cpu == e2->cpu ||
3537 /* Called under the same ctx::mutex as perf_install_in_context() */
3538 static bool exclusive_event_installable(struct perf_event *event,
3539 struct perf_event_context *ctx)
3541 struct perf_event *iter_event;
3542 struct pmu *pmu = event->pmu;
3544 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3547 list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
3548 if (exclusive_event_match(iter_event, event))
3555 static void __free_event(struct perf_event *event)
3557 if (!event->parent) {
3558 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3559 put_callchain_buffers();
3563 event->destroy(event);
3566 put_ctx(event->ctx);
3569 exclusive_event_destroy(event);
3570 module_put(event->pmu->module);
3573 call_rcu(&event->rcu_head, free_event_rcu);
3576 static void _free_event(struct perf_event *event)
3578 irq_work_sync(&event->pending);
3580 unaccount_event(event);
3584 * Can happen when we close an event with re-directed output.
3586 * Since we have a 0 refcount, perf_mmap_close() will skip
3587 * over us; possibly making our ring_buffer_put() the last.
3589 mutex_lock(&event->mmap_mutex);
3590 ring_buffer_attach(event, NULL);
3591 mutex_unlock(&event->mmap_mutex);
3594 if (is_cgroup_event(event))
3595 perf_detach_cgroup(event);
3597 __free_event(event);
3601 * Used to free events which have a known refcount of 1, such as in error paths
3602 * where the event isn't exposed yet and inherited events.
3604 static void free_event(struct perf_event *event)
3606 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3607 "unexpected event refcount: %ld; ptr=%p\n",
3608 atomic_long_read(&event->refcount), event)) {
3609 /* leak to avoid use-after-free */
3617 * Remove user event from the owner task.
3619 static void perf_remove_from_owner(struct perf_event *event)
3621 struct task_struct *owner;
3624 owner = ACCESS_ONCE(event->owner);
3626 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3627 * !owner it means the list deletion is complete and we can indeed
3628 * free this event, otherwise we need to serialize on
3629 * owner->perf_event_mutex.
3631 smp_read_barrier_depends();
3634 * Since delayed_put_task_struct() also drops the last
3635 * task reference we can safely take a new reference
3636 * while holding the rcu_read_lock().
3638 get_task_struct(owner);
3644 * If we're here through perf_event_exit_task() we're already
3645 * holding ctx->mutex which would be an inversion wrt. the
3646 * normal lock order.
3648 * However we can safely take this lock because its the child
3651 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
3654 * We have to re-check the event->owner field, if it is cleared
3655 * we raced with perf_event_exit_task(), acquiring the mutex
3656 * ensured they're done, and we can proceed with freeing the
3660 list_del_init(&event->owner_entry);
3661 mutex_unlock(&owner->perf_event_mutex);
3662 put_task_struct(owner);
3667 * Called when the last reference to the file is gone.
3669 static void put_event(struct perf_event *event)
3671 struct perf_event_context *ctx;
3673 if (!atomic_long_dec_and_test(&event->refcount))
3676 if (!is_kernel_event(event))
3677 perf_remove_from_owner(event);
3680 * There are two ways this annotation is useful:
3682 * 1) there is a lock recursion from perf_event_exit_task
3683 * see the comment there.
3685 * 2) there is a lock-inversion with mmap_sem through
3686 * perf_event_read_group(), which takes faults while
3687 * holding ctx->mutex, however this is called after
3688 * the last filedesc died, so there is no possibility
3689 * to trigger the AB-BA case.
3691 ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING);
3692 WARN_ON_ONCE(ctx->parent_ctx);
3693 perf_remove_from_context(event, true);
3694 perf_event_ctx_unlock(event, ctx);
3699 int perf_event_release_kernel(struct perf_event *event)
3704 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3706 static int perf_release(struct inode *inode, struct file *file)
3708 put_event(file->private_data);
3713 * Remove all orphanes events from the context.
3715 static void orphans_remove_work(struct work_struct *work)
3717 struct perf_event_context *ctx;
3718 struct perf_event *event, *tmp;
3720 ctx = container_of(work, struct perf_event_context,
3721 orphans_remove.work);
3723 mutex_lock(&ctx->mutex);
3724 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3725 struct perf_event *parent_event = event->parent;
3727 if (!is_orphaned_child(event))
3730 perf_remove_from_context(event, true);
3732 mutex_lock(&parent_event->child_mutex);
3733 list_del_init(&event->child_list);
3734 mutex_unlock(&parent_event->child_mutex);
3737 put_event(parent_event);
3740 raw_spin_lock_irq(&ctx->lock);
3741 ctx->orphans_remove_sched = false;
3742 raw_spin_unlock_irq(&ctx->lock);
3743 mutex_unlock(&ctx->mutex);
3748 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3750 struct perf_event *child;
3756 mutex_lock(&event->child_mutex);
3757 total += perf_event_read(event);
3758 *enabled += event->total_time_enabled +
3759 atomic64_read(&event->child_total_time_enabled);
3760 *running += event->total_time_running +
3761 atomic64_read(&event->child_total_time_running);
3763 list_for_each_entry(child, &event->child_list, child_list) {
3764 total += perf_event_read(child);
3765 *enabled += child->total_time_enabled;
3766 *running += child->total_time_running;
3768 mutex_unlock(&event->child_mutex);
3772 EXPORT_SYMBOL_GPL(perf_event_read_value);
3774 static int perf_event_read_group(struct perf_event *event,
3775 u64 read_format, char __user *buf)
3777 struct perf_event *leader = event->group_leader, *sub;
3778 struct perf_event_context *ctx = leader->ctx;
3779 int n = 0, size = 0, ret;
3780 u64 count, enabled, running;
3783 lockdep_assert_held(&ctx->mutex);
3785 count = perf_event_read_value(leader, &enabled, &running);
3787 values[n++] = 1 + leader->nr_siblings;
3788 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3789 values[n++] = enabled;
3790 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3791 values[n++] = running;
3792 values[n++] = count;
3793 if (read_format & PERF_FORMAT_ID)
3794 values[n++] = primary_event_id(leader);
3796 size = n * sizeof(u64);
3798 if (copy_to_user(buf, values, size))
3803 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3806 values[n++] = perf_event_read_value(sub, &enabled, &running);
3807 if (read_format & PERF_FORMAT_ID)
3808 values[n++] = primary_event_id(sub);
3810 size = n * sizeof(u64);
3812 if (copy_to_user(buf + ret, values, size)) {
3822 static int perf_event_read_one(struct perf_event *event,
3823 u64 read_format, char __user *buf)
3825 u64 enabled, running;
3829 values[n++] = perf_event_read_value(event, &enabled, &running);
3830 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3831 values[n++] = enabled;
3832 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3833 values[n++] = running;
3834 if (read_format & PERF_FORMAT_ID)
3835 values[n++] = primary_event_id(event);
3837 if (copy_to_user(buf, values, n * sizeof(u64)))
3840 return n * sizeof(u64);
3843 static bool is_event_hup(struct perf_event *event)
3847 if (event->state != PERF_EVENT_STATE_EXIT)
3850 mutex_lock(&event->child_mutex);
3851 no_children = list_empty(&event->child_list);
3852 mutex_unlock(&event->child_mutex);
3857 * Read the performance event - simple non blocking version for now
3860 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3862 u64 read_format = event->attr.read_format;
3866 * Return end-of-file for a read on a event that is in
3867 * error state (i.e. because it was pinned but it couldn't be
3868 * scheduled on to the CPU at some point).
3870 if (event->state == PERF_EVENT_STATE_ERROR)
3873 if (count < event->read_size)
3876 WARN_ON_ONCE(event->ctx->parent_ctx);
3877 if (read_format & PERF_FORMAT_GROUP)
3878 ret = perf_event_read_group(event, read_format, buf);
3880 ret = perf_event_read_one(event, read_format, buf);
3886 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3888 struct perf_event *event = file->private_data;
3889 struct perf_event_context *ctx;
3892 ctx = perf_event_ctx_lock(event);
3893 ret = perf_read_hw(event, buf, count);
3894 perf_event_ctx_unlock(event, ctx);
3899 static unsigned int perf_poll(struct file *file, poll_table *wait)
3901 struct perf_event *event = file->private_data;
3902 struct ring_buffer *rb;
3903 unsigned int events = POLLHUP;
3905 poll_wait(file, &event->waitq, wait);
3907 if (is_event_hup(event))
3911 * Pin the event->rb by taking event->mmap_mutex; otherwise
3912 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3914 mutex_lock(&event->mmap_mutex);
3917 events = atomic_xchg(&rb->poll, 0);
3918 mutex_unlock(&event->mmap_mutex);
3922 static void _perf_event_reset(struct perf_event *event)
3924 (void)perf_event_read(event);
3925 local64_set(&event->count, 0);
3926 perf_event_update_userpage(event);
3930 * Holding the top-level event's child_mutex means that any
3931 * descendant process that has inherited this event will block
3932 * in sync_child_event if it goes to exit, thus satisfying the
3933 * task existence requirements of perf_event_enable/disable.
3935 static void perf_event_for_each_child(struct perf_event *event,
3936 void (*func)(struct perf_event *))
3938 struct perf_event *child;
3940 WARN_ON_ONCE(event->ctx->parent_ctx);
3942 mutex_lock(&event->child_mutex);
3944 list_for_each_entry(child, &event->child_list, child_list)
3946 mutex_unlock(&event->child_mutex);
3949 static void perf_event_for_each(struct perf_event *event,
3950 void (*func)(struct perf_event *))
3952 struct perf_event_context *ctx = event->ctx;
3953 struct perf_event *sibling;
3955 lockdep_assert_held(&ctx->mutex);
3957 event = event->group_leader;
3959 perf_event_for_each_child(event, func);
3960 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3961 perf_event_for_each_child(sibling, func);
3964 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3966 struct perf_event_context *ctx = event->ctx;
3967 int ret = 0, active;
3970 if (!is_sampling_event(event))
3973 if (copy_from_user(&value, arg, sizeof(value)))
3979 raw_spin_lock_irq(&ctx->lock);
3980 if (event->attr.freq) {
3981 if (value > sysctl_perf_event_sample_rate) {
3986 event->attr.sample_freq = value;
3988 event->attr.sample_period = value;
3989 event->hw.sample_period = value;
3992 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3994 perf_pmu_disable(ctx->pmu);
3995 event->pmu->stop(event, PERF_EF_UPDATE);
3998 local64_set(&event->hw.period_left, 0);
4001 event->pmu->start(event, PERF_EF_RELOAD);
4002 perf_pmu_enable(ctx->pmu);
4006 raw_spin_unlock_irq(&ctx->lock);
4011 static const struct file_operations perf_fops;
4013 static inline int perf_fget_light(int fd, struct fd *p)
4015 struct fd f = fdget(fd);
4019 if (f.file->f_op != &perf_fops) {
4027 static int perf_event_set_output(struct perf_event *event,
4028 struct perf_event *output_event);
4029 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
4030 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
4032 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
4034 void (*func)(struct perf_event *);
4038 case PERF_EVENT_IOC_ENABLE:
4039 func = _perf_event_enable;
4041 case PERF_EVENT_IOC_DISABLE:
4042 func = _perf_event_disable;
4044 case PERF_EVENT_IOC_RESET:
4045 func = _perf_event_reset;
4048 case PERF_EVENT_IOC_REFRESH:
4049 return _perf_event_refresh(event, arg);
4051 case PERF_EVENT_IOC_PERIOD:
4052 return perf_event_period(event, (u64 __user *)arg);
4054 case PERF_EVENT_IOC_ID:
4056 u64 id = primary_event_id(event);
4058 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
4063 case PERF_EVENT_IOC_SET_OUTPUT:
4067 struct perf_event *output_event;
4069 ret = perf_fget_light(arg, &output);
4072 output_event = output.file->private_data;
4073 ret = perf_event_set_output(event, output_event);
4076 ret = perf_event_set_output(event, NULL);
4081 case PERF_EVENT_IOC_SET_FILTER:
4082 return perf_event_set_filter(event, (void __user *)arg);
4084 case PERF_EVENT_IOC_SET_BPF:
4085 return perf_event_set_bpf_prog(event, arg);
4091 if (flags & PERF_IOC_FLAG_GROUP)
4092 perf_event_for_each(event, func);
4094 perf_event_for_each_child(event, func);
4099 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
4101 struct perf_event *event = file->private_data;
4102 struct perf_event_context *ctx;
4105 ctx = perf_event_ctx_lock(event);
4106 ret = _perf_ioctl(event, cmd, arg);
4107 perf_event_ctx_unlock(event, ctx);
4112 #ifdef CONFIG_COMPAT
4113 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
4116 switch (_IOC_NR(cmd)) {
4117 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
4118 case _IOC_NR(PERF_EVENT_IOC_ID):
4119 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4120 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
4121 cmd &= ~IOCSIZE_MASK;
4122 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
4126 return perf_ioctl(file, cmd, arg);
4129 # define perf_compat_ioctl NULL
4132 int perf_event_task_enable(void)
4134 struct perf_event_context *ctx;
4135 struct perf_event *event;
4137 mutex_lock(¤t->perf_event_mutex);
4138 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4139 ctx = perf_event_ctx_lock(event);
4140 perf_event_for_each_child(event, _perf_event_enable);
4141 perf_event_ctx_unlock(event, ctx);
4143 mutex_unlock(¤t->perf_event_mutex);
4148 int perf_event_task_disable(void)
4150 struct perf_event_context *ctx;
4151 struct perf_event *event;
4153 mutex_lock(¤t->perf_event_mutex);
4154 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4155 ctx = perf_event_ctx_lock(event);
4156 perf_event_for_each_child(event, _perf_event_disable);
4157 perf_event_ctx_unlock(event, ctx);
4159 mutex_unlock(¤t->perf_event_mutex);
4164 static int perf_event_index(struct perf_event *event)
4166 if (event->hw.state & PERF_HES_STOPPED)
4169 if (event->state != PERF_EVENT_STATE_ACTIVE)
4172 return event->pmu->event_idx(event);
4175 static void calc_timer_values(struct perf_event *event,
4182 *now = perf_clock();
4183 ctx_time = event->shadow_ctx_time + *now;
4184 *enabled = ctx_time - event->tstamp_enabled;
4185 *running = ctx_time - event->tstamp_running;
4188 static void perf_event_init_userpage(struct perf_event *event)
4190 struct perf_event_mmap_page *userpg;
4191 struct ring_buffer *rb;
4194 rb = rcu_dereference(event->rb);
4198 userpg = rb->user_page;
4200 /* Allow new userspace to detect that bit 0 is deprecated */
4201 userpg->cap_bit0_is_deprecated = 1;
4202 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
4203 userpg->data_offset = PAGE_SIZE;
4204 userpg->data_size = perf_data_size(rb);
4210 void __weak arch_perf_update_userpage(
4211 struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
4216 * Callers need to ensure there can be no nesting of this function, otherwise
4217 * the seqlock logic goes bad. We can not serialize this because the arch
4218 * code calls this from NMI context.
4220 void perf_event_update_userpage(struct perf_event *event)
4222 struct perf_event_mmap_page *userpg;
4223 struct ring_buffer *rb;
4224 u64 enabled, running, now;
4227 rb = rcu_dereference(event->rb);
4232 * compute total_time_enabled, total_time_running
4233 * based on snapshot values taken when the event
4234 * was last scheduled in.
4236 * we cannot simply called update_context_time()
4237 * because of locking issue as we can be called in
4240 calc_timer_values(event, &now, &enabled, &running);
4242 userpg = rb->user_page;
4244 * Disable preemption so as to not let the corresponding user-space
4245 * spin too long if we get preempted.
4250 userpg->index = perf_event_index(event);
4251 userpg->offset = perf_event_count(event);
4253 userpg->offset -= local64_read(&event->hw.prev_count);
4255 userpg->time_enabled = enabled +
4256 atomic64_read(&event->child_total_time_enabled);
4258 userpg->time_running = running +
4259 atomic64_read(&event->child_total_time_running);
4261 arch_perf_update_userpage(event, userpg, now);
4270 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
4272 struct perf_event *event = vma->vm_file->private_data;
4273 struct ring_buffer *rb;
4274 int ret = VM_FAULT_SIGBUS;
4276 if (vmf->flags & FAULT_FLAG_MKWRITE) {
4277 if (vmf->pgoff == 0)
4283 rb = rcu_dereference(event->rb);
4287 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4290 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4294 get_page(vmf->page);
4295 vmf->page->mapping = vma->vm_file->f_mapping;
4296 vmf->page->index = vmf->pgoff;
4305 static void ring_buffer_attach(struct perf_event *event,
4306 struct ring_buffer *rb)
4308 struct ring_buffer *old_rb = NULL;
4309 unsigned long flags;
4313 * Should be impossible, we set this when removing
4314 * event->rb_entry and wait/clear when adding event->rb_entry.
4316 WARN_ON_ONCE(event->rcu_pending);
4319 event->rcu_batches = get_state_synchronize_rcu();
4320 event->rcu_pending = 1;
4322 spin_lock_irqsave(&old_rb->event_lock, flags);
4323 list_del_rcu(&event->rb_entry);
4324 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4327 if (event->rcu_pending && rb) {
4328 cond_synchronize_rcu(event->rcu_batches);
4329 event->rcu_pending = 0;
4333 spin_lock_irqsave(&rb->event_lock, flags);
4334 list_add_rcu(&event->rb_entry, &rb->event_list);
4335 spin_unlock_irqrestore(&rb->event_lock, flags);
4338 rcu_assign_pointer(event->rb, rb);
4341 ring_buffer_put(old_rb);
4343 * Since we detached before setting the new rb, so that we
4344 * could attach the new rb, we could have missed a wakeup.
4347 wake_up_all(&event->waitq);
4351 static void ring_buffer_wakeup(struct perf_event *event)
4353 struct ring_buffer *rb;
4356 rb = rcu_dereference(event->rb);
4358 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4359 wake_up_all(&event->waitq);
4364 static void rb_free_rcu(struct rcu_head *rcu_head)
4366 struct ring_buffer *rb;
4368 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
4372 struct ring_buffer *ring_buffer_get(struct perf_event *event)
4374 struct ring_buffer *rb;
4377 rb = rcu_dereference(event->rb);
4379 if (!atomic_inc_not_zero(&rb->refcount))
4387 void ring_buffer_put(struct ring_buffer *rb)
4389 if (!atomic_dec_and_test(&rb->refcount))
4392 WARN_ON_ONCE(!list_empty(&rb->event_list));
4394 call_rcu(&rb->rcu_head, rb_free_rcu);
4397 static void perf_mmap_open(struct vm_area_struct *vma)
4399 struct perf_event *event = vma->vm_file->private_data;
4401 atomic_inc(&event->mmap_count);
4402 atomic_inc(&event->rb->mmap_count);
4405 atomic_inc(&event->rb->aux_mmap_count);
4407 if (event->pmu->event_mapped)
4408 event->pmu->event_mapped(event);
4412 * A buffer can be mmap()ed multiple times; either directly through the same
4413 * event, or through other events by use of perf_event_set_output().
4415 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4416 * the buffer here, where we still have a VM context. This means we need
4417 * to detach all events redirecting to us.
4419 static void perf_mmap_close(struct vm_area_struct *vma)
4421 struct perf_event *event = vma->vm_file->private_data;
4423 struct ring_buffer *rb = ring_buffer_get(event);
4424 struct user_struct *mmap_user = rb->mmap_user;
4425 int mmap_locked = rb->mmap_locked;
4426 unsigned long size = perf_data_size(rb);
4428 if (event->pmu->event_unmapped)
4429 event->pmu->event_unmapped(event);
4432 * rb->aux_mmap_count will always drop before rb->mmap_count and
4433 * event->mmap_count, so it is ok to use event->mmap_mutex to
4434 * serialize with perf_mmap here.
4436 if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff &&
4437 atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) {
4438 atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm);
4439 vma->vm_mm->pinned_vm -= rb->aux_mmap_locked;
4442 mutex_unlock(&event->mmap_mutex);
4445 atomic_dec(&rb->mmap_count);
4447 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4450 ring_buffer_attach(event, NULL);
4451 mutex_unlock(&event->mmap_mutex);
4453 /* If there's still other mmap()s of this buffer, we're done. */
4454 if (atomic_read(&rb->mmap_count))
4458 * No other mmap()s, detach from all other events that might redirect
4459 * into the now unreachable buffer. Somewhat complicated by the
4460 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4464 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4465 if (!atomic_long_inc_not_zero(&event->refcount)) {
4467 * This event is en-route to free_event() which will
4468 * detach it and remove it from the list.
4474 mutex_lock(&event->mmap_mutex);
4476 * Check we didn't race with perf_event_set_output() which can
4477 * swizzle the rb from under us while we were waiting to
4478 * acquire mmap_mutex.
4480 * If we find a different rb; ignore this event, a next
4481 * iteration will no longer find it on the list. We have to
4482 * still restart the iteration to make sure we're not now
4483 * iterating the wrong list.
4485 if (event->rb == rb)
4486 ring_buffer_attach(event, NULL);
4488 mutex_unlock(&event->mmap_mutex);
4492 * Restart the iteration; either we're on the wrong list or
4493 * destroyed its integrity by doing a deletion.
4500 * It could be there's still a few 0-ref events on the list; they'll
4501 * get cleaned up by free_event() -- they'll also still have their
4502 * ref on the rb and will free it whenever they are done with it.
4504 * Aside from that, this buffer is 'fully' detached and unmapped,
4505 * undo the VM accounting.
4508 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4509 vma->vm_mm->pinned_vm -= mmap_locked;
4510 free_uid(mmap_user);
4513 ring_buffer_put(rb); /* could be last */
4516 static const struct vm_operations_struct perf_mmap_vmops = {
4517 .open = perf_mmap_open,
4518 .close = perf_mmap_close, /* non mergable */
4519 .fault = perf_mmap_fault,
4520 .page_mkwrite = perf_mmap_fault,
4523 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4525 struct perf_event *event = file->private_data;
4526 unsigned long user_locked, user_lock_limit;
4527 struct user_struct *user = current_user();
4528 unsigned long locked, lock_limit;
4529 struct ring_buffer *rb = NULL;
4530 unsigned long vma_size;
4531 unsigned long nr_pages;
4532 long user_extra = 0, extra = 0;
4533 int ret = 0, flags = 0;
4536 * Don't allow mmap() of inherited per-task counters. This would
4537 * create a performance issue due to all children writing to the
4540 if (event->cpu == -1 && event->attr.inherit)
4543 if (!(vma->vm_flags & VM_SHARED))
4546 vma_size = vma->vm_end - vma->vm_start;
4548 if (vma->vm_pgoff == 0) {
4549 nr_pages = (vma_size / PAGE_SIZE) - 1;
4552 * AUX area mapping: if rb->aux_nr_pages != 0, it's already
4553 * mapped, all subsequent mappings should have the same size
4554 * and offset. Must be above the normal perf buffer.
4556 u64 aux_offset, aux_size;
4561 nr_pages = vma_size / PAGE_SIZE;
4563 mutex_lock(&event->mmap_mutex);
4570 aux_offset = ACCESS_ONCE(rb->user_page->aux_offset);
4571 aux_size = ACCESS_ONCE(rb->user_page->aux_size);
4573 if (aux_offset < perf_data_size(rb) + PAGE_SIZE)
4576 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT)
4579 /* already mapped with a different offset */
4580 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff)
4583 if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE)
4586 /* already mapped with a different size */
4587 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages)
4590 if (!is_power_of_2(nr_pages))
4593 if (!atomic_inc_not_zero(&rb->mmap_count))
4596 if (rb_has_aux(rb)) {
4597 atomic_inc(&rb->aux_mmap_count);
4602 atomic_set(&rb->aux_mmap_count, 1);
4603 user_extra = nr_pages;
4609 * If we have rb pages ensure they're a power-of-two number, so we
4610 * can do bitmasks instead of modulo.
4612 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4615 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4618 WARN_ON_ONCE(event->ctx->parent_ctx);
4620 mutex_lock(&event->mmap_mutex);
4622 if (event->rb->nr_pages != nr_pages) {
4627 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4629 * Raced against perf_mmap_close() through
4630 * perf_event_set_output(). Try again, hope for better
4633 mutex_unlock(&event->mmap_mutex);
4640 user_extra = nr_pages + 1;
4643 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4646 * Increase the limit linearly with more CPUs:
4648 user_lock_limit *= num_online_cpus();
4650 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4652 if (user_locked > user_lock_limit)
4653 extra = user_locked - user_lock_limit;
4655 lock_limit = rlimit(RLIMIT_MEMLOCK);
4656 lock_limit >>= PAGE_SHIFT;
4657 locked = vma->vm_mm->pinned_vm + extra;
4659 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4660 !capable(CAP_IPC_LOCK)) {
4665 WARN_ON(!rb && event->rb);
4667 if (vma->vm_flags & VM_WRITE)
4668 flags |= RING_BUFFER_WRITABLE;
4671 rb = rb_alloc(nr_pages,
4672 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4680 atomic_set(&rb->mmap_count, 1);
4681 rb->mmap_user = get_current_user();
4682 rb->mmap_locked = extra;
4684 ring_buffer_attach(event, rb);
4686 perf_event_init_userpage(event);
4687 perf_event_update_userpage(event);
4689 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages,
4690 event->attr.aux_watermark, flags);
4692 rb->aux_mmap_locked = extra;
4697 atomic_long_add(user_extra, &user->locked_vm);
4698 vma->vm_mm->pinned_vm += extra;
4700 atomic_inc(&event->mmap_count);
4702 atomic_dec(&rb->mmap_count);
4705 mutex_unlock(&event->mmap_mutex);
4708 * Since pinned accounting is per vm we cannot allow fork() to copy our
4711 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4712 vma->vm_ops = &perf_mmap_vmops;
4714 if (event->pmu->event_mapped)
4715 event->pmu->event_mapped(event);
4720 static int perf_fasync(int fd, struct file *filp, int on)
4722 struct inode *inode = file_inode(filp);
4723 struct perf_event *event = filp->private_data;
4726 mutex_lock(&inode->i_mutex);
4727 retval = fasync_helper(fd, filp, on, &event->fasync);
4728 mutex_unlock(&inode->i_mutex);
4736 static const struct file_operations perf_fops = {
4737 .llseek = no_llseek,
4738 .release = perf_release,
4741 .unlocked_ioctl = perf_ioctl,
4742 .compat_ioctl = perf_compat_ioctl,
4744 .fasync = perf_fasync,
4750 * If there's data, ensure we set the poll() state and publish everything
4751 * to user-space before waking everybody up.
4754 void perf_event_wakeup(struct perf_event *event)
4756 ring_buffer_wakeup(event);
4758 if (event->pending_kill) {
4759 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4760 event->pending_kill = 0;
4764 static void perf_pending_event(struct irq_work *entry)
4766 struct perf_event *event = container_of(entry,
4767 struct perf_event, pending);
4770 rctx = perf_swevent_get_recursion_context();
4772 * If we 'fail' here, that's OK, it means recursion is already disabled
4773 * and we won't recurse 'further'.
4776 if (event->pending_disable) {
4777 event->pending_disable = 0;
4778 __perf_event_disable(event);
4781 if (event->pending_wakeup) {
4782 event->pending_wakeup = 0;
4783 perf_event_wakeup(event);
4787 perf_swevent_put_recursion_context(rctx);
4791 * We assume there is only KVM supporting the callbacks.
4792 * Later on, we might change it to a list if there is
4793 * another virtualization implementation supporting the callbacks.
4795 struct perf_guest_info_callbacks *perf_guest_cbs;
4797 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4799 perf_guest_cbs = cbs;
4802 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4804 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4806 perf_guest_cbs = NULL;
4809 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4812 perf_output_sample_regs(struct perf_output_handle *handle,
4813 struct pt_regs *regs, u64 mask)
4817 for_each_set_bit(bit, (const unsigned long *) &mask,
4818 sizeof(mask) * BITS_PER_BYTE) {
4821 val = perf_reg_value(regs, bit);
4822 perf_output_put(handle, val);
4826 static void perf_sample_regs_user(struct perf_regs *regs_user,
4827 struct pt_regs *regs,
4828 struct pt_regs *regs_user_copy)
4830 if (user_mode(regs)) {
4831 regs_user->abi = perf_reg_abi(current);
4832 regs_user->regs = regs;
4833 } else if (current->mm) {
4834 perf_get_regs_user(regs_user, regs, regs_user_copy);
4836 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
4837 regs_user->regs = NULL;
4841 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
4842 struct pt_regs *regs)
4844 regs_intr->regs = regs;
4845 regs_intr->abi = perf_reg_abi(current);
4850 * Get remaining task size from user stack pointer.
4852 * It'd be better to take stack vma map and limit this more
4853 * precisly, but there's no way to get it safely under interrupt,
4854 * so using TASK_SIZE as limit.
4856 static u64 perf_ustack_task_size(struct pt_regs *regs)
4858 unsigned long addr = perf_user_stack_pointer(regs);
4860 if (!addr || addr >= TASK_SIZE)
4863 return TASK_SIZE - addr;
4867 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4868 struct pt_regs *regs)
4872 /* No regs, no stack pointer, no dump. */
4877 * Check if we fit in with the requested stack size into the:
4879 * If we don't, we limit the size to the TASK_SIZE.
4881 * - remaining sample size
4882 * If we don't, we customize the stack size to
4883 * fit in to the remaining sample size.
4886 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4887 stack_size = min(stack_size, (u16) task_size);
4889 /* Current header size plus static size and dynamic size. */
4890 header_size += 2 * sizeof(u64);
4892 /* Do we fit in with the current stack dump size? */
4893 if ((u16) (header_size + stack_size) < header_size) {
4895 * If we overflow the maximum size for the sample,
4896 * we customize the stack dump size to fit in.
4898 stack_size = USHRT_MAX - header_size - sizeof(u64);
4899 stack_size = round_up(stack_size, sizeof(u64));
4906 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4907 struct pt_regs *regs)
4909 /* Case of a kernel thread, nothing to dump */
4912 perf_output_put(handle, size);
4921 * - the size requested by user or the best one we can fit
4922 * in to the sample max size
4924 * - user stack dump data
4926 * - the actual dumped size
4930 perf_output_put(handle, dump_size);
4933 sp = perf_user_stack_pointer(regs);
4934 rem = __output_copy_user(handle, (void *) sp, dump_size);
4935 dyn_size = dump_size - rem;
4937 perf_output_skip(handle, rem);
4940 perf_output_put(handle, dyn_size);
4944 static void __perf_event_header__init_id(struct perf_event_header *header,
4945 struct perf_sample_data *data,
4946 struct perf_event *event)
4948 u64 sample_type = event->attr.sample_type;
4950 data->type = sample_type;
4951 header->size += event->id_header_size;
4953 if (sample_type & PERF_SAMPLE_TID) {
4954 /* namespace issues */
4955 data->tid_entry.pid = perf_event_pid(event, current);
4956 data->tid_entry.tid = perf_event_tid(event, current);
4959 if (sample_type & PERF_SAMPLE_TIME)
4960 data->time = perf_event_clock(event);
4962 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4963 data->id = primary_event_id(event);
4965 if (sample_type & PERF_SAMPLE_STREAM_ID)
4966 data->stream_id = event->id;
4968 if (sample_type & PERF_SAMPLE_CPU) {
4969 data->cpu_entry.cpu = raw_smp_processor_id();
4970 data->cpu_entry.reserved = 0;
4974 void perf_event_header__init_id(struct perf_event_header *header,
4975 struct perf_sample_data *data,
4976 struct perf_event *event)
4978 if (event->attr.sample_id_all)
4979 __perf_event_header__init_id(header, data, event);
4982 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4983 struct perf_sample_data *data)
4985 u64 sample_type = data->type;
4987 if (sample_type & PERF_SAMPLE_TID)
4988 perf_output_put(handle, data->tid_entry);
4990 if (sample_type & PERF_SAMPLE_TIME)
4991 perf_output_put(handle, data->time);
4993 if (sample_type & PERF_SAMPLE_ID)
4994 perf_output_put(handle, data->id);
4996 if (sample_type & PERF_SAMPLE_STREAM_ID)
4997 perf_output_put(handle, data->stream_id);
4999 if (sample_type & PERF_SAMPLE_CPU)
5000 perf_output_put(handle, data->cpu_entry);
5002 if (sample_type & PERF_SAMPLE_IDENTIFIER)
5003 perf_output_put(handle, data->id);
5006 void perf_event__output_id_sample(struct perf_event *event,
5007 struct perf_output_handle *handle,
5008 struct perf_sample_data *sample)
5010 if (event->attr.sample_id_all)
5011 __perf_event__output_id_sample(handle, sample);
5014 static void perf_output_read_one(struct perf_output_handle *handle,
5015 struct perf_event *event,
5016 u64 enabled, u64 running)
5018 u64 read_format = event->attr.read_format;
5022 values[n++] = perf_event_count(event);
5023 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
5024 values[n++] = enabled +
5025 atomic64_read(&event->child_total_time_enabled);
5027 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
5028 values[n++] = running +
5029 atomic64_read(&event->child_total_time_running);
5031 if (read_format & PERF_FORMAT_ID)
5032 values[n++] = primary_event_id(event);
5034 __output_copy(handle, values, n * sizeof(u64));
5038 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
5040 static void perf_output_read_group(struct perf_output_handle *handle,
5041 struct perf_event *event,
5042 u64 enabled, u64 running)
5044 struct perf_event *leader = event->group_leader, *sub;
5045 u64 read_format = event->attr.read_format;
5049 values[n++] = 1 + leader->nr_siblings;
5051 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
5052 values[n++] = enabled;
5054 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
5055 values[n++] = running;
5057 if (leader != event)
5058 leader->pmu->read(leader);
5060 values[n++] = perf_event_count(leader);
5061 if (read_format & PERF_FORMAT_ID)
5062 values[n++] = primary_event_id(leader);
5064 __output_copy(handle, values, n * sizeof(u64));
5066 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
5069 if ((sub != event) &&
5070 (sub->state == PERF_EVENT_STATE_ACTIVE))
5071 sub->pmu->read(sub);
5073 values[n++] = perf_event_count(sub);
5074 if (read_format & PERF_FORMAT_ID)
5075 values[n++] = primary_event_id(sub);
5077 __output_copy(handle, values, n * sizeof(u64));
5081 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
5082 PERF_FORMAT_TOTAL_TIME_RUNNING)
5084 static void perf_output_read(struct perf_output_handle *handle,
5085 struct perf_event *event)
5087 u64 enabled = 0, running = 0, now;
5088 u64 read_format = event->attr.read_format;
5091 * compute total_time_enabled, total_time_running
5092 * based on snapshot values taken when the event
5093 * was last scheduled in.
5095 * we cannot simply called update_context_time()
5096 * because of locking issue as we are called in
5099 if (read_format & PERF_FORMAT_TOTAL_TIMES)
5100 calc_timer_values(event, &now, &enabled, &running);
5102 if (event->attr.read_format & PERF_FORMAT_GROUP)
5103 perf_output_read_group(handle, event, enabled, running);
5105 perf_output_read_one(handle, event, enabled, running);
5108 void perf_output_sample(struct perf_output_handle *handle,
5109 struct perf_event_header *header,
5110 struct perf_sample_data *data,
5111 struct perf_event *event)
5113 u64 sample_type = data->type;
5115 perf_output_put(handle, *header);
5117 if (sample_type & PERF_SAMPLE_IDENTIFIER)
5118 perf_output_put(handle, data->id);
5120 if (sample_type & PERF_SAMPLE_IP)
5121 perf_output_put(handle, data->ip);
5123 if (sample_type & PERF_SAMPLE_TID)
5124 perf_output_put(handle, data->tid_entry);
5126 if (sample_type & PERF_SAMPLE_TIME)
5127 perf_output_put(handle, data->time);
5129 if (sample_type & PERF_SAMPLE_ADDR)
5130 perf_output_put(handle, data->addr);
5132 if (sample_type & PERF_SAMPLE_ID)
5133 perf_output_put(handle, data->id);
5135 if (sample_type & PERF_SAMPLE_STREAM_ID)
5136 perf_output_put(handle, data->stream_id);
5138 if (sample_type & PERF_SAMPLE_CPU)
5139 perf_output_put(handle, data->cpu_entry);
5141 if (sample_type & PERF_SAMPLE_PERIOD)
5142 perf_output_put(handle, data->period);
5144 if (sample_type & PERF_SAMPLE_READ)
5145 perf_output_read(handle, event);
5147 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5148 if (data->callchain) {
5151 if (data->callchain)
5152 size += data->callchain->nr;
5154 size *= sizeof(u64);
5156 __output_copy(handle, data->callchain, size);
5159 perf_output_put(handle, nr);
5163 if (sample_type & PERF_SAMPLE_RAW) {
5165 perf_output_put(handle, data->raw->size);
5166 __output_copy(handle, data->raw->data,
5173 .size = sizeof(u32),
5176 perf_output_put(handle, raw);
5180 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5181 if (data->br_stack) {
5184 size = data->br_stack->nr
5185 * sizeof(struct perf_branch_entry);
5187 perf_output_put(handle, data->br_stack->nr);
5188 perf_output_copy(handle, data->br_stack->entries, size);
5191 * we always store at least the value of nr
5194 perf_output_put(handle, nr);
5198 if (sample_type & PERF_SAMPLE_REGS_USER) {
5199 u64 abi = data->regs_user.abi;
5202 * If there are no regs to dump, notice it through
5203 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5205 perf_output_put(handle, abi);
5208 u64 mask = event->attr.sample_regs_user;
5209 perf_output_sample_regs(handle,
5210 data->regs_user.regs,
5215 if (sample_type & PERF_SAMPLE_STACK_USER) {
5216 perf_output_sample_ustack(handle,
5217 data->stack_user_size,
5218 data->regs_user.regs);
5221 if (sample_type & PERF_SAMPLE_WEIGHT)
5222 perf_output_put(handle, data->weight);
5224 if (sample_type & PERF_SAMPLE_DATA_SRC)
5225 perf_output_put(handle, data->data_src.val);
5227 if (sample_type & PERF_SAMPLE_TRANSACTION)
5228 perf_output_put(handle, data->txn);
5230 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5231 u64 abi = data->regs_intr.abi;
5233 * If there are no regs to dump, notice it through
5234 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5236 perf_output_put(handle, abi);
5239 u64 mask = event->attr.sample_regs_intr;
5241 perf_output_sample_regs(handle,
5242 data->regs_intr.regs,
5247 if (!event->attr.watermark) {
5248 int wakeup_events = event->attr.wakeup_events;
5250 if (wakeup_events) {
5251 struct ring_buffer *rb = handle->rb;
5252 int events = local_inc_return(&rb->events);
5254 if (events >= wakeup_events) {
5255 local_sub(wakeup_events, &rb->events);
5256 local_inc(&rb->wakeup);
5262 void perf_prepare_sample(struct perf_event_header *header,
5263 struct perf_sample_data *data,
5264 struct perf_event *event,
5265 struct pt_regs *regs)
5267 u64 sample_type = event->attr.sample_type;
5269 header->type = PERF_RECORD_SAMPLE;
5270 header->size = sizeof(*header) + event->header_size;
5273 header->misc |= perf_misc_flags(regs);
5275 __perf_event_header__init_id(header, data, event);
5277 if (sample_type & PERF_SAMPLE_IP)
5278 data->ip = perf_instruction_pointer(regs);
5280 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5283 data->callchain = perf_callchain(event, regs);
5285 if (data->callchain)
5286 size += data->callchain->nr;
5288 header->size += size * sizeof(u64);
5291 if (sample_type & PERF_SAMPLE_RAW) {
5292 int size = sizeof(u32);
5295 size += data->raw->size;
5297 size += sizeof(u32);
5299 WARN_ON_ONCE(size & (sizeof(u64)-1));
5300 header->size += size;
5303 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5304 int size = sizeof(u64); /* nr */
5305 if (data->br_stack) {
5306 size += data->br_stack->nr
5307 * sizeof(struct perf_branch_entry);
5309 header->size += size;
5312 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
5313 perf_sample_regs_user(&data->regs_user, regs,
5314 &data->regs_user_copy);
5316 if (sample_type & PERF_SAMPLE_REGS_USER) {
5317 /* regs dump ABI info */
5318 int size = sizeof(u64);
5320 if (data->regs_user.regs) {
5321 u64 mask = event->attr.sample_regs_user;
5322 size += hweight64(mask) * sizeof(u64);
5325 header->size += size;
5328 if (sample_type & PERF_SAMPLE_STACK_USER) {
5330 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5331 * processed as the last one or have additional check added
5332 * in case new sample type is added, because we could eat
5333 * up the rest of the sample size.
5335 u16 stack_size = event->attr.sample_stack_user;
5336 u16 size = sizeof(u64);
5338 stack_size = perf_sample_ustack_size(stack_size, header->size,
5339 data->regs_user.regs);
5342 * If there is something to dump, add space for the dump
5343 * itself and for the field that tells the dynamic size,
5344 * which is how many have been actually dumped.
5347 size += sizeof(u64) + stack_size;
5349 data->stack_user_size = stack_size;
5350 header->size += size;
5353 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5354 /* regs dump ABI info */
5355 int size = sizeof(u64);
5357 perf_sample_regs_intr(&data->regs_intr, regs);
5359 if (data->regs_intr.regs) {
5360 u64 mask = event->attr.sample_regs_intr;
5362 size += hweight64(mask) * sizeof(u64);
5365 header->size += size;
5369 static void perf_event_output(struct perf_event *event,
5370 struct perf_sample_data *data,
5371 struct pt_regs *regs)
5373 struct perf_output_handle handle;
5374 struct perf_event_header header;
5376 /* protect the callchain buffers */
5379 perf_prepare_sample(&header, data, event, regs);
5381 if (perf_output_begin(&handle, event, header.size))
5384 perf_output_sample(&handle, &header, data, event);
5386 perf_output_end(&handle);
5396 struct perf_read_event {
5397 struct perf_event_header header;
5404 perf_event_read_event(struct perf_event *event,
5405 struct task_struct *task)
5407 struct perf_output_handle handle;
5408 struct perf_sample_data sample;
5409 struct perf_read_event read_event = {
5411 .type = PERF_RECORD_READ,
5413 .size = sizeof(read_event) + event->read_size,
5415 .pid = perf_event_pid(event, task),
5416 .tid = perf_event_tid(event, task),
5420 perf_event_header__init_id(&read_event.header, &sample, event);
5421 ret = perf_output_begin(&handle, event, read_event.header.size);
5425 perf_output_put(&handle, read_event);
5426 perf_output_read(&handle, event);
5427 perf_event__output_id_sample(event, &handle, &sample);
5429 perf_output_end(&handle);
5432 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5435 perf_event_aux_ctx(struct perf_event_context *ctx,
5436 perf_event_aux_output_cb output,
5439 struct perf_event *event;
5441 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5442 if (event->state < PERF_EVENT_STATE_INACTIVE)
5444 if (!event_filter_match(event))
5446 output(event, data);
5451 perf_event_aux(perf_event_aux_output_cb output, void *data,
5452 struct perf_event_context *task_ctx)
5454 struct perf_cpu_context *cpuctx;
5455 struct perf_event_context *ctx;
5460 list_for_each_entry_rcu(pmu, &pmus, entry) {
5461 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5462 if (cpuctx->unique_pmu != pmu)
5464 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5467 ctxn = pmu->task_ctx_nr;
5470 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5472 perf_event_aux_ctx(ctx, output, data);
5474 put_cpu_ptr(pmu->pmu_cpu_context);
5479 perf_event_aux_ctx(task_ctx, output, data);
5486 * task tracking -- fork/exit
5488 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5491 struct perf_task_event {
5492 struct task_struct *task;
5493 struct perf_event_context *task_ctx;
5496 struct perf_event_header header;
5506 static int perf_event_task_match(struct perf_event *event)
5508 return event->attr.comm || event->attr.mmap ||
5509 event->attr.mmap2 || event->attr.mmap_data ||
5513 static void perf_event_task_output(struct perf_event *event,
5516 struct perf_task_event *task_event = data;
5517 struct perf_output_handle handle;
5518 struct perf_sample_data sample;
5519 struct task_struct *task = task_event->task;
5520 int ret, size = task_event->event_id.header.size;
5522 if (!perf_event_task_match(event))
5525 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5527 ret = perf_output_begin(&handle, event,
5528 task_event->event_id.header.size);
5532 task_event->event_id.pid = perf_event_pid(event, task);
5533 task_event->event_id.ppid = perf_event_pid(event, current);
5535 task_event->event_id.tid = perf_event_tid(event, task);
5536 task_event->event_id.ptid = perf_event_tid(event, current);
5538 task_event->event_id.time = perf_event_clock(event);
5540 perf_output_put(&handle, task_event->event_id);
5542 perf_event__output_id_sample(event, &handle, &sample);
5544 perf_output_end(&handle);
5546 task_event->event_id.header.size = size;
5549 static void perf_event_task(struct task_struct *task,
5550 struct perf_event_context *task_ctx,
5553 struct perf_task_event task_event;
5555 if (!atomic_read(&nr_comm_events) &&
5556 !atomic_read(&nr_mmap_events) &&
5557 !atomic_read(&nr_task_events))
5560 task_event = (struct perf_task_event){
5562 .task_ctx = task_ctx,
5565 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5567 .size = sizeof(task_event.event_id),
5577 perf_event_aux(perf_event_task_output,
5582 void perf_event_fork(struct task_struct *task)
5584 perf_event_task(task, NULL, 1);
5591 struct perf_comm_event {
5592 struct task_struct *task;
5597 struct perf_event_header header;
5604 static int perf_event_comm_match(struct perf_event *event)
5606 return event->attr.comm;
5609 static void perf_event_comm_output(struct perf_event *event,
5612 struct perf_comm_event *comm_event = data;
5613 struct perf_output_handle handle;
5614 struct perf_sample_data sample;
5615 int size = comm_event->event_id.header.size;
5618 if (!perf_event_comm_match(event))
5621 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5622 ret = perf_output_begin(&handle, event,
5623 comm_event->event_id.header.size);
5628 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5629 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5631 perf_output_put(&handle, comm_event->event_id);
5632 __output_copy(&handle, comm_event->comm,
5633 comm_event->comm_size);
5635 perf_event__output_id_sample(event, &handle, &sample);
5637 perf_output_end(&handle);
5639 comm_event->event_id.header.size = size;
5642 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5644 char comm[TASK_COMM_LEN];
5647 memset(comm, 0, sizeof(comm));
5648 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5649 size = ALIGN(strlen(comm)+1, sizeof(u64));
5651 comm_event->comm = comm;
5652 comm_event->comm_size = size;
5654 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5656 perf_event_aux(perf_event_comm_output,
5661 void perf_event_comm(struct task_struct *task, bool exec)
5663 struct perf_comm_event comm_event;
5665 if (!atomic_read(&nr_comm_events))
5668 comm_event = (struct perf_comm_event){
5674 .type = PERF_RECORD_COMM,
5675 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5683 perf_event_comm_event(&comm_event);
5690 struct perf_mmap_event {
5691 struct vm_area_struct *vma;
5693 const char *file_name;
5701 struct perf_event_header header;
5711 static int perf_event_mmap_match(struct perf_event *event,
5714 struct perf_mmap_event *mmap_event = data;
5715 struct vm_area_struct *vma = mmap_event->vma;
5716 int executable = vma->vm_flags & VM_EXEC;
5718 return (!executable && event->attr.mmap_data) ||
5719 (executable && (event->attr.mmap || event->attr.mmap2));
5722 static void perf_event_mmap_output(struct perf_event *event,
5725 struct perf_mmap_event *mmap_event = data;
5726 struct perf_output_handle handle;
5727 struct perf_sample_data sample;
5728 int size = mmap_event->event_id.header.size;
5731 if (!perf_event_mmap_match(event, data))
5734 if (event->attr.mmap2) {
5735 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5736 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5737 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5738 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5739 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5740 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5741 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5744 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5745 ret = perf_output_begin(&handle, event,
5746 mmap_event->event_id.header.size);
5750 mmap_event->event_id.pid = perf_event_pid(event, current);
5751 mmap_event->event_id.tid = perf_event_tid(event, current);
5753 perf_output_put(&handle, mmap_event->event_id);
5755 if (event->attr.mmap2) {
5756 perf_output_put(&handle, mmap_event->maj);
5757 perf_output_put(&handle, mmap_event->min);
5758 perf_output_put(&handle, mmap_event->ino);
5759 perf_output_put(&handle, mmap_event->ino_generation);
5760 perf_output_put(&handle, mmap_event->prot);
5761 perf_output_put(&handle, mmap_event->flags);
5764 __output_copy(&handle, mmap_event->file_name,
5765 mmap_event->file_size);
5767 perf_event__output_id_sample(event, &handle, &sample);
5769 perf_output_end(&handle);
5771 mmap_event->event_id.header.size = size;
5774 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5776 struct vm_area_struct *vma = mmap_event->vma;
5777 struct file *file = vma->vm_file;
5778 int maj = 0, min = 0;
5779 u64 ino = 0, gen = 0;
5780 u32 prot = 0, flags = 0;
5787 struct inode *inode;
5790 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5796 * d_path() works from the end of the rb backwards, so we
5797 * need to add enough zero bytes after the string to handle
5798 * the 64bit alignment we do later.
5800 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5805 inode = file_inode(vma->vm_file);
5806 dev = inode->i_sb->s_dev;
5808 gen = inode->i_generation;
5812 if (vma->vm_flags & VM_READ)
5814 if (vma->vm_flags & VM_WRITE)
5816 if (vma->vm_flags & VM_EXEC)
5819 if (vma->vm_flags & VM_MAYSHARE)
5822 flags = MAP_PRIVATE;
5824 if (vma->vm_flags & VM_DENYWRITE)
5825 flags |= MAP_DENYWRITE;
5826 if (vma->vm_flags & VM_MAYEXEC)
5827 flags |= MAP_EXECUTABLE;
5828 if (vma->vm_flags & VM_LOCKED)
5829 flags |= MAP_LOCKED;
5830 if (vma->vm_flags & VM_HUGETLB)
5831 flags |= MAP_HUGETLB;
5835 if (vma->vm_ops && vma->vm_ops->name) {
5836 name = (char *) vma->vm_ops->name(vma);
5841 name = (char *)arch_vma_name(vma);
5845 if (vma->vm_start <= vma->vm_mm->start_brk &&
5846 vma->vm_end >= vma->vm_mm->brk) {
5850 if (vma->vm_start <= vma->vm_mm->start_stack &&
5851 vma->vm_end >= vma->vm_mm->start_stack) {
5861 strlcpy(tmp, name, sizeof(tmp));
5865 * Since our buffer works in 8 byte units we need to align our string
5866 * size to a multiple of 8. However, we must guarantee the tail end is
5867 * zero'd out to avoid leaking random bits to userspace.
5869 size = strlen(name)+1;
5870 while (!IS_ALIGNED(size, sizeof(u64)))
5871 name[size++] = '\0';
5873 mmap_event->file_name = name;
5874 mmap_event->file_size = size;
5875 mmap_event->maj = maj;
5876 mmap_event->min = min;
5877 mmap_event->ino = ino;
5878 mmap_event->ino_generation = gen;
5879 mmap_event->prot = prot;
5880 mmap_event->flags = flags;
5882 if (!(vma->vm_flags & VM_EXEC))
5883 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5885 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5887 perf_event_aux(perf_event_mmap_output,
5894 void perf_event_mmap(struct vm_area_struct *vma)
5896 struct perf_mmap_event mmap_event;
5898 if (!atomic_read(&nr_mmap_events))
5901 mmap_event = (struct perf_mmap_event){
5907 .type = PERF_RECORD_MMAP,
5908 .misc = PERF_RECORD_MISC_USER,
5913 .start = vma->vm_start,
5914 .len = vma->vm_end - vma->vm_start,
5915 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5917 /* .maj (attr_mmap2 only) */
5918 /* .min (attr_mmap2 only) */
5919 /* .ino (attr_mmap2 only) */
5920 /* .ino_generation (attr_mmap2 only) */
5921 /* .prot (attr_mmap2 only) */
5922 /* .flags (attr_mmap2 only) */
5925 perf_event_mmap_event(&mmap_event);
5928 void perf_event_aux_event(struct perf_event *event, unsigned long head,
5929 unsigned long size, u64 flags)
5931 struct perf_output_handle handle;
5932 struct perf_sample_data sample;
5933 struct perf_aux_event {
5934 struct perf_event_header header;
5940 .type = PERF_RECORD_AUX,
5942 .size = sizeof(rec),
5950 perf_event_header__init_id(&rec.header, &sample, event);
5951 ret = perf_output_begin(&handle, event, rec.header.size);
5956 perf_output_put(&handle, rec);
5957 perf_event__output_id_sample(event, &handle, &sample);
5959 perf_output_end(&handle);
5963 * IRQ throttle logging
5966 static void perf_log_throttle(struct perf_event *event, int enable)
5968 struct perf_output_handle handle;
5969 struct perf_sample_data sample;
5973 struct perf_event_header header;
5977 } throttle_event = {
5979 .type = PERF_RECORD_THROTTLE,
5981 .size = sizeof(throttle_event),
5983 .time = perf_event_clock(event),
5984 .id = primary_event_id(event),
5985 .stream_id = event->id,
5989 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5991 perf_event_header__init_id(&throttle_event.header, &sample, event);
5993 ret = perf_output_begin(&handle, event,
5994 throttle_event.header.size);
5998 perf_output_put(&handle, throttle_event);
5999 perf_event__output_id_sample(event, &handle, &sample);
6000 perf_output_end(&handle);
6003 static void perf_log_itrace_start(struct perf_event *event)
6005 struct perf_output_handle handle;
6006 struct perf_sample_data sample;
6007 struct perf_aux_event {
6008 struct perf_event_header header;
6015 event = event->parent;
6017 if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) ||
6018 event->hw.itrace_started)
6021 event->hw.itrace_started = 1;
6023 rec.header.type = PERF_RECORD_ITRACE_START;
6024 rec.header.misc = 0;
6025 rec.header.size = sizeof(rec);
6026 rec.pid = perf_event_pid(event, current);
6027 rec.tid = perf_event_tid(event, current);
6029 perf_event_header__init_id(&rec.header, &sample, event);
6030 ret = perf_output_begin(&handle, event, rec.header.size);
6035 perf_output_put(&handle, rec);
6036 perf_event__output_id_sample(event, &handle, &sample);
6038 perf_output_end(&handle);
6042 * Generic event overflow handling, sampling.
6045 static int __perf_event_overflow(struct perf_event *event,
6046 int throttle, struct perf_sample_data *data,
6047 struct pt_regs *regs)
6049 int events = atomic_read(&event->event_limit);
6050 struct hw_perf_event *hwc = &event->hw;
6055 * Non-sampling counters might still use the PMI to fold short
6056 * hardware counters, ignore those.
6058 if (unlikely(!is_sampling_event(event)))
6061 seq = __this_cpu_read(perf_throttled_seq);
6062 if (seq != hwc->interrupts_seq) {
6063 hwc->interrupts_seq = seq;
6064 hwc->interrupts = 1;
6067 if (unlikely(throttle
6068 && hwc->interrupts >= max_samples_per_tick)) {
6069 __this_cpu_inc(perf_throttled_count);
6070 hwc->interrupts = MAX_INTERRUPTS;
6071 perf_log_throttle(event, 0);
6072 tick_nohz_full_kick();
6077 if (event->attr.freq) {
6078 u64 now = perf_clock();
6079 s64 delta = now - hwc->freq_time_stamp;
6081 hwc->freq_time_stamp = now;
6083 if (delta > 0 && delta < 2*TICK_NSEC)
6084 perf_adjust_period(event, delta, hwc->last_period, true);
6088 * XXX event_limit might not quite work as expected on inherited
6092 event->pending_kill = POLL_IN;
6093 if (events && atomic_dec_and_test(&event->event_limit)) {
6095 event->pending_kill = POLL_HUP;
6096 event->pending_disable = 1;
6097 irq_work_queue(&event->pending);
6100 if (event->overflow_handler)
6101 event->overflow_handler(event, data, regs);
6103 perf_event_output(event, data, regs);
6105 if (event->fasync && event->pending_kill) {
6106 event->pending_wakeup = 1;
6107 irq_work_queue(&event->pending);
6113 int perf_event_overflow(struct perf_event *event,
6114 struct perf_sample_data *data,
6115 struct pt_regs *regs)
6117 return __perf_event_overflow(event, 1, data, regs);
6121 * Generic software event infrastructure
6124 struct swevent_htable {
6125 struct swevent_hlist *swevent_hlist;
6126 struct mutex hlist_mutex;
6129 /* Recursion avoidance in each contexts */
6130 int recursion[PERF_NR_CONTEXTS];
6132 /* Keeps track of cpu being initialized/exited */
6136 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
6139 * We directly increment event->count and keep a second value in
6140 * event->hw.period_left to count intervals. This period event
6141 * is kept in the range [-sample_period, 0] so that we can use the
6145 u64 perf_swevent_set_period(struct perf_event *event)
6147 struct hw_perf_event *hwc = &event->hw;
6148 u64 period = hwc->last_period;
6152 hwc->last_period = hwc->sample_period;
6155 old = val = local64_read(&hwc->period_left);
6159 nr = div64_u64(period + val, period);
6160 offset = nr * period;
6162 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
6168 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
6169 struct perf_sample_data *data,
6170 struct pt_regs *regs)
6172 struct hw_perf_event *hwc = &event->hw;
6176 overflow = perf_swevent_set_period(event);
6178 if (hwc->interrupts == MAX_INTERRUPTS)
6181 for (; overflow; overflow--) {
6182 if (__perf_event_overflow(event, throttle,
6185 * We inhibit the overflow from happening when
6186 * hwc->interrupts == MAX_INTERRUPTS.
6194 static void perf_swevent_event(struct perf_event *event, u64 nr,
6195 struct perf_sample_data *data,
6196 struct pt_regs *regs)
6198 struct hw_perf_event *hwc = &event->hw;
6200 local64_add(nr, &event->count);
6205 if (!is_sampling_event(event))
6208 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
6210 return perf_swevent_overflow(event, 1, data, regs);
6212 data->period = event->hw.last_period;
6214 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
6215 return perf_swevent_overflow(event, 1, data, regs);
6217 if (local64_add_negative(nr, &hwc->period_left))
6220 perf_swevent_overflow(event, 0, data, regs);
6223 static int perf_exclude_event(struct perf_event *event,
6224 struct pt_regs *regs)
6226 if (event->hw.state & PERF_HES_STOPPED)
6230 if (event->attr.exclude_user && user_mode(regs))
6233 if (event->attr.exclude_kernel && !user_mode(regs))
6240 static int perf_swevent_match(struct perf_event *event,
6241 enum perf_type_id type,
6243 struct perf_sample_data *data,
6244 struct pt_regs *regs)
6246 if (event->attr.type != type)
6249 if (event->attr.config != event_id)
6252 if (perf_exclude_event(event, regs))
6258 static inline u64 swevent_hash(u64 type, u32 event_id)
6260 u64 val = event_id | (type << 32);
6262 return hash_64(val, SWEVENT_HLIST_BITS);
6265 static inline struct hlist_head *
6266 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
6268 u64 hash = swevent_hash(type, event_id);
6270 return &hlist->heads[hash];
6273 /* For the read side: events when they trigger */
6274 static inline struct hlist_head *
6275 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
6277 struct swevent_hlist *hlist;
6279 hlist = rcu_dereference(swhash->swevent_hlist);
6283 return __find_swevent_head(hlist, type, event_id);
6286 /* For the event head insertion and removal in the hlist */
6287 static inline struct hlist_head *
6288 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
6290 struct swevent_hlist *hlist;
6291 u32 event_id = event->attr.config;
6292 u64 type = event->attr.type;
6295 * Event scheduling is always serialized against hlist allocation
6296 * and release. Which makes the protected version suitable here.
6297 * The context lock guarantees that.
6299 hlist = rcu_dereference_protected(swhash->swevent_hlist,
6300 lockdep_is_held(&event->ctx->lock));
6304 return __find_swevent_head(hlist, type, event_id);
6307 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
6309 struct perf_sample_data *data,
6310 struct pt_regs *regs)
6312 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6313 struct perf_event *event;
6314 struct hlist_head *head;
6317 head = find_swevent_head_rcu(swhash, type, event_id);
6321 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6322 if (perf_swevent_match(event, type, event_id, data, regs))
6323 perf_swevent_event(event, nr, data, regs);
6329 DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]);
6331 int perf_swevent_get_recursion_context(void)
6333 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6335 return get_recursion_context(swhash->recursion);
6337 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
6339 inline void perf_swevent_put_recursion_context(int rctx)
6341 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6343 put_recursion_context(swhash->recursion, rctx);
6346 void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6348 struct perf_sample_data data;
6350 if (WARN_ON_ONCE(!regs))
6353 perf_sample_data_init(&data, addr, 0);
6354 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
6357 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6361 preempt_disable_notrace();
6362 rctx = perf_swevent_get_recursion_context();
6363 if (unlikely(rctx < 0))
6366 ___perf_sw_event(event_id, nr, regs, addr);
6368 perf_swevent_put_recursion_context(rctx);
6370 preempt_enable_notrace();
6373 static void perf_swevent_read(struct perf_event *event)
6377 static int perf_swevent_add(struct perf_event *event, int flags)
6379 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6380 struct hw_perf_event *hwc = &event->hw;
6381 struct hlist_head *head;
6383 if (is_sampling_event(event)) {
6384 hwc->last_period = hwc->sample_period;
6385 perf_swevent_set_period(event);
6388 hwc->state = !(flags & PERF_EF_START);
6390 head = find_swevent_head(swhash, event);
6393 * We can race with cpu hotplug code. Do not
6394 * WARN if the cpu just got unplugged.
6396 WARN_ON_ONCE(swhash->online);
6400 hlist_add_head_rcu(&event->hlist_entry, head);
6401 perf_event_update_userpage(event);
6406 static void perf_swevent_del(struct perf_event *event, int flags)
6408 hlist_del_rcu(&event->hlist_entry);
6411 static void perf_swevent_start(struct perf_event *event, int flags)
6413 event->hw.state = 0;
6416 static void perf_swevent_stop(struct perf_event *event, int flags)
6418 event->hw.state = PERF_HES_STOPPED;
6421 /* Deref the hlist from the update side */
6422 static inline struct swevent_hlist *
6423 swevent_hlist_deref(struct swevent_htable *swhash)
6425 return rcu_dereference_protected(swhash->swevent_hlist,
6426 lockdep_is_held(&swhash->hlist_mutex));
6429 static void swevent_hlist_release(struct swevent_htable *swhash)
6431 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
6436 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
6437 kfree_rcu(hlist, rcu_head);
6440 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
6442 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6444 mutex_lock(&swhash->hlist_mutex);
6446 if (!--swhash->hlist_refcount)
6447 swevent_hlist_release(swhash);
6449 mutex_unlock(&swhash->hlist_mutex);
6452 static void swevent_hlist_put(struct perf_event *event)
6456 for_each_possible_cpu(cpu)
6457 swevent_hlist_put_cpu(event, cpu);
6460 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
6462 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6465 mutex_lock(&swhash->hlist_mutex);
6467 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6468 struct swevent_hlist *hlist;
6470 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
6475 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6477 swhash->hlist_refcount++;
6479 mutex_unlock(&swhash->hlist_mutex);
6484 static int swevent_hlist_get(struct perf_event *event)
6487 int cpu, failed_cpu;
6490 for_each_possible_cpu(cpu) {
6491 err = swevent_hlist_get_cpu(event, cpu);
6501 for_each_possible_cpu(cpu) {
6502 if (cpu == failed_cpu)
6504 swevent_hlist_put_cpu(event, cpu);
6511 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6513 static void sw_perf_event_destroy(struct perf_event *event)
6515 u64 event_id = event->attr.config;
6517 WARN_ON(event->parent);
6519 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6520 swevent_hlist_put(event);
6523 static int perf_swevent_init(struct perf_event *event)
6525 u64 event_id = event->attr.config;
6527 if (event->attr.type != PERF_TYPE_SOFTWARE)
6531 * no branch sampling for software events
6533 if (has_branch_stack(event))
6537 case PERF_COUNT_SW_CPU_CLOCK:
6538 case PERF_COUNT_SW_TASK_CLOCK:
6545 if (event_id >= PERF_COUNT_SW_MAX)
6548 if (!event->parent) {
6551 err = swevent_hlist_get(event);
6555 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6556 event->destroy = sw_perf_event_destroy;
6562 static struct pmu perf_swevent = {
6563 .task_ctx_nr = perf_sw_context,
6565 .capabilities = PERF_PMU_CAP_NO_NMI,
6567 .event_init = perf_swevent_init,
6568 .add = perf_swevent_add,
6569 .del = perf_swevent_del,
6570 .start = perf_swevent_start,
6571 .stop = perf_swevent_stop,
6572 .read = perf_swevent_read,
6575 #ifdef CONFIG_EVENT_TRACING
6577 static int perf_tp_filter_match(struct perf_event *event,
6578 struct perf_sample_data *data)
6580 void *record = data->raw->data;
6582 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6587 static int perf_tp_event_match(struct perf_event *event,
6588 struct perf_sample_data *data,
6589 struct pt_regs *regs)
6591 if (event->hw.state & PERF_HES_STOPPED)
6594 * All tracepoints are from kernel-space.
6596 if (event->attr.exclude_kernel)
6599 if (!perf_tp_filter_match(event, data))
6605 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6606 struct pt_regs *regs, struct hlist_head *head, int rctx,
6607 struct task_struct *task)
6609 struct perf_sample_data data;
6610 struct perf_event *event;
6612 struct perf_raw_record raw = {
6617 perf_sample_data_init(&data, addr, 0);
6620 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6621 if (perf_tp_event_match(event, &data, regs))
6622 perf_swevent_event(event, count, &data, regs);
6626 * If we got specified a target task, also iterate its context and
6627 * deliver this event there too.
6629 if (task && task != current) {
6630 struct perf_event_context *ctx;
6631 struct trace_entry *entry = record;
6634 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6638 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6639 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6641 if (event->attr.config != entry->type)
6643 if (perf_tp_event_match(event, &data, regs))
6644 perf_swevent_event(event, count, &data, regs);
6650 perf_swevent_put_recursion_context(rctx);
6652 EXPORT_SYMBOL_GPL(perf_tp_event);
6654 static void tp_perf_event_destroy(struct perf_event *event)
6656 perf_trace_destroy(event);
6659 static int perf_tp_event_init(struct perf_event *event)
6663 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6667 * no branch sampling for tracepoint events
6669 if (has_branch_stack(event))
6672 err = perf_trace_init(event);
6676 event->destroy = tp_perf_event_destroy;
6681 static struct pmu perf_tracepoint = {
6682 .task_ctx_nr = perf_sw_context,
6684 .event_init = perf_tp_event_init,
6685 .add = perf_trace_add,
6686 .del = perf_trace_del,
6687 .start = perf_swevent_start,
6688 .stop = perf_swevent_stop,
6689 .read = perf_swevent_read,
6692 static inline void perf_tp_register(void)
6694 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6697 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6702 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6705 filter_str = strndup_user(arg, PAGE_SIZE);
6706 if (IS_ERR(filter_str))
6707 return PTR_ERR(filter_str);
6709 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6715 static void perf_event_free_filter(struct perf_event *event)
6717 ftrace_profile_free_filter(event);
6720 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
6722 struct bpf_prog *prog;
6724 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6727 if (event->tp_event->prog)
6730 if (!(event->tp_event->flags & TRACE_EVENT_FL_KPROBE))
6731 /* bpf programs can only be attached to kprobes */
6734 prog = bpf_prog_get(prog_fd);
6736 return PTR_ERR(prog);
6738 if (prog->type != BPF_PROG_TYPE_KPROBE) {
6739 /* valid fd, but invalid bpf program type */
6744 event->tp_event->prog = prog;
6749 static void perf_event_free_bpf_prog(struct perf_event *event)
6751 struct bpf_prog *prog;
6753 if (!event->tp_event)
6756 prog = event->tp_event->prog;
6758 event->tp_event->prog = NULL;
6765 static inline void perf_tp_register(void)
6769 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6774 static void perf_event_free_filter(struct perf_event *event)
6778 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
6783 static void perf_event_free_bpf_prog(struct perf_event *event)
6786 #endif /* CONFIG_EVENT_TRACING */
6788 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6789 void perf_bp_event(struct perf_event *bp, void *data)
6791 struct perf_sample_data sample;
6792 struct pt_regs *regs = data;
6794 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6796 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6797 perf_swevent_event(bp, 1, &sample, regs);
6802 * hrtimer based swevent callback
6805 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6807 enum hrtimer_restart ret = HRTIMER_RESTART;
6808 struct perf_sample_data data;
6809 struct pt_regs *regs;
6810 struct perf_event *event;
6813 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6815 if (event->state != PERF_EVENT_STATE_ACTIVE)
6816 return HRTIMER_NORESTART;
6818 event->pmu->read(event);
6820 perf_sample_data_init(&data, 0, event->hw.last_period);
6821 regs = get_irq_regs();
6823 if (regs && !perf_exclude_event(event, regs)) {
6824 if (!(event->attr.exclude_idle && is_idle_task(current)))
6825 if (__perf_event_overflow(event, 1, &data, regs))
6826 ret = HRTIMER_NORESTART;
6829 period = max_t(u64, 10000, event->hw.sample_period);
6830 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6835 static void perf_swevent_start_hrtimer(struct perf_event *event)
6837 struct hw_perf_event *hwc = &event->hw;
6840 if (!is_sampling_event(event))
6843 period = local64_read(&hwc->period_left);
6848 local64_set(&hwc->period_left, 0);
6850 period = max_t(u64, 10000, hwc->sample_period);
6852 __hrtimer_start_range_ns(&hwc->hrtimer,
6853 ns_to_ktime(period), 0,
6854 HRTIMER_MODE_REL_PINNED, 0);
6857 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6859 struct hw_perf_event *hwc = &event->hw;
6861 if (is_sampling_event(event)) {
6862 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6863 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6865 hrtimer_cancel(&hwc->hrtimer);
6869 static void perf_swevent_init_hrtimer(struct perf_event *event)
6871 struct hw_perf_event *hwc = &event->hw;
6873 if (!is_sampling_event(event))
6876 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6877 hwc->hrtimer.function = perf_swevent_hrtimer;
6880 * Since hrtimers have a fixed rate, we can do a static freq->period
6881 * mapping and avoid the whole period adjust feedback stuff.
6883 if (event->attr.freq) {
6884 long freq = event->attr.sample_freq;
6886 event->attr.sample_period = NSEC_PER_SEC / freq;
6887 hwc->sample_period = event->attr.sample_period;
6888 local64_set(&hwc->period_left, hwc->sample_period);
6889 hwc->last_period = hwc->sample_period;
6890 event->attr.freq = 0;
6895 * Software event: cpu wall time clock
6898 static void cpu_clock_event_update(struct perf_event *event)
6903 now = local_clock();
6904 prev = local64_xchg(&event->hw.prev_count, now);
6905 local64_add(now - prev, &event->count);
6908 static void cpu_clock_event_start(struct perf_event *event, int flags)
6910 local64_set(&event->hw.prev_count, local_clock());
6911 perf_swevent_start_hrtimer(event);
6914 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6916 perf_swevent_cancel_hrtimer(event);
6917 cpu_clock_event_update(event);
6920 static int cpu_clock_event_add(struct perf_event *event, int flags)
6922 if (flags & PERF_EF_START)
6923 cpu_clock_event_start(event, flags);
6924 perf_event_update_userpage(event);
6929 static void cpu_clock_event_del(struct perf_event *event, int flags)
6931 cpu_clock_event_stop(event, flags);
6934 static void cpu_clock_event_read(struct perf_event *event)
6936 cpu_clock_event_update(event);
6939 static int cpu_clock_event_init(struct perf_event *event)
6941 if (event->attr.type != PERF_TYPE_SOFTWARE)
6944 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6948 * no branch sampling for software events
6950 if (has_branch_stack(event))
6953 perf_swevent_init_hrtimer(event);
6958 static struct pmu perf_cpu_clock = {
6959 .task_ctx_nr = perf_sw_context,
6961 .capabilities = PERF_PMU_CAP_NO_NMI,
6963 .event_init = cpu_clock_event_init,
6964 .add = cpu_clock_event_add,
6965 .del = cpu_clock_event_del,
6966 .start = cpu_clock_event_start,
6967 .stop = cpu_clock_event_stop,
6968 .read = cpu_clock_event_read,
6972 * Software event: task time clock
6975 static void task_clock_event_update(struct perf_event *event, u64 now)
6980 prev = local64_xchg(&event->hw.prev_count, now);
6982 local64_add(delta, &event->count);
6985 static void task_clock_event_start(struct perf_event *event, int flags)
6987 local64_set(&event->hw.prev_count, event->ctx->time);
6988 perf_swevent_start_hrtimer(event);
6991 static void task_clock_event_stop(struct perf_event *event, int flags)
6993 perf_swevent_cancel_hrtimer(event);
6994 task_clock_event_update(event, event->ctx->time);
6997 static int task_clock_event_add(struct perf_event *event, int flags)
6999 if (flags & PERF_EF_START)
7000 task_clock_event_start(event, flags);
7001 perf_event_update_userpage(event);
7006 static void task_clock_event_del(struct perf_event *event, int flags)
7008 task_clock_event_stop(event, PERF_EF_UPDATE);
7011 static void task_clock_event_read(struct perf_event *event)
7013 u64 now = perf_clock();
7014 u64 delta = now - event->ctx->timestamp;
7015 u64 time = event->ctx->time + delta;
7017 task_clock_event_update(event, time);
7020 static int task_clock_event_init(struct perf_event *event)
7022 if (event->attr.type != PERF_TYPE_SOFTWARE)
7025 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
7029 * no branch sampling for software events
7031 if (has_branch_stack(event))
7034 perf_swevent_init_hrtimer(event);
7039 static struct pmu perf_task_clock = {
7040 .task_ctx_nr = perf_sw_context,
7042 .capabilities = PERF_PMU_CAP_NO_NMI,
7044 .event_init = task_clock_event_init,
7045 .add = task_clock_event_add,
7046 .del = task_clock_event_del,
7047 .start = task_clock_event_start,
7048 .stop = task_clock_event_stop,
7049 .read = task_clock_event_read,
7052 static void perf_pmu_nop_void(struct pmu *pmu)
7056 static int perf_pmu_nop_int(struct pmu *pmu)
7061 static void perf_pmu_start_txn(struct pmu *pmu)
7063 perf_pmu_disable(pmu);
7066 static int perf_pmu_commit_txn(struct pmu *pmu)
7068 perf_pmu_enable(pmu);
7072 static void perf_pmu_cancel_txn(struct pmu *pmu)
7074 perf_pmu_enable(pmu);
7077 static int perf_event_idx_default(struct perf_event *event)
7083 * Ensures all contexts with the same task_ctx_nr have the same
7084 * pmu_cpu_context too.
7086 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
7093 list_for_each_entry(pmu, &pmus, entry) {
7094 if (pmu->task_ctx_nr == ctxn)
7095 return pmu->pmu_cpu_context;
7101 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
7105 for_each_possible_cpu(cpu) {
7106 struct perf_cpu_context *cpuctx;
7108 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7110 if (cpuctx->unique_pmu == old_pmu)
7111 cpuctx->unique_pmu = pmu;
7115 static void free_pmu_context(struct pmu *pmu)
7119 mutex_lock(&pmus_lock);
7121 * Like a real lame refcount.
7123 list_for_each_entry(i, &pmus, entry) {
7124 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
7125 update_pmu_context(i, pmu);
7130 free_percpu(pmu->pmu_cpu_context);
7132 mutex_unlock(&pmus_lock);
7134 static struct idr pmu_idr;
7137 type_show(struct device *dev, struct device_attribute *attr, char *page)
7139 struct pmu *pmu = dev_get_drvdata(dev);
7141 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
7143 static DEVICE_ATTR_RO(type);
7146 perf_event_mux_interval_ms_show(struct device *dev,
7147 struct device_attribute *attr,
7150 struct pmu *pmu = dev_get_drvdata(dev);
7152 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
7156 perf_event_mux_interval_ms_store(struct device *dev,
7157 struct device_attribute *attr,
7158 const char *buf, size_t count)
7160 struct pmu *pmu = dev_get_drvdata(dev);
7161 int timer, cpu, ret;
7163 ret = kstrtoint(buf, 0, &timer);
7170 /* same value, noting to do */
7171 if (timer == pmu->hrtimer_interval_ms)
7174 pmu->hrtimer_interval_ms = timer;
7176 /* update all cpuctx for this PMU */
7177 for_each_possible_cpu(cpu) {
7178 struct perf_cpu_context *cpuctx;
7179 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7180 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
7182 if (hrtimer_active(&cpuctx->hrtimer))
7183 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
7188 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
7190 static struct attribute *pmu_dev_attrs[] = {
7191 &dev_attr_type.attr,
7192 &dev_attr_perf_event_mux_interval_ms.attr,
7195 ATTRIBUTE_GROUPS(pmu_dev);
7197 static int pmu_bus_running;
7198 static struct bus_type pmu_bus = {
7199 .name = "event_source",
7200 .dev_groups = pmu_dev_groups,
7203 static void pmu_dev_release(struct device *dev)
7208 static int pmu_dev_alloc(struct pmu *pmu)
7212 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
7216 pmu->dev->groups = pmu->attr_groups;
7217 device_initialize(pmu->dev);
7218 ret = dev_set_name(pmu->dev, "%s", pmu->name);
7222 dev_set_drvdata(pmu->dev, pmu);
7223 pmu->dev->bus = &pmu_bus;
7224 pmu->dev->release = pmu_dev_release;
7225 ret = device_add(pmu->dev);
7233 put_device(pmu->dev);
7237 static struct lock_class_key cpuctx_mutex;
7238 static struct lock_class_key cpuctx_lock;
7240 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
7244 mutex_lock(&pmus_lock);
7246 pmu->pmu_disable_count = alloc_percpu(int);
7247 if (!pmu->pmu_disable_count)
7256 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
7264 if (pmu_bus_running) {
7265 ret = pmu_dev_alloc(pmu);
7271 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
7272 if (pmu->pmu_cpu_context)
7273 goto got_cpu_context;
7276 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
7277 if (!pmu->pmu_cpu_context)
7280 for_each_possible_cpu(cpu) {
7281 struct perf_cpu_context *cpuctx;
7283 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7284 __perf_event_init_context(&cpuctx->ctx);
7285 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
7286 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
7287 cpuctx->ctx.pmu = pmu;
7289 __perf_cpu_hrtimer_init(cpuctx, cpu);
7291 cpuctx->unique_pmu = pmu;
7295 if (!pmu->start_txn) {
7296 if (pmu->pmu_enable) {
7298 * If we have pmu_enable/pmu_disable calls, install
7299 * transaction stubs that use that to try and batch
7300 * hardware accesses.
7302 pmu->start_txn = perf_pmu_start_txn;
7303 pmu->commit_txn = perf_pmu_commit_txn;
7304 pmu->cancel_txn = perf_pmu_cancel_txn;
7306 pmu->start_txn = perf_pmu_nop_void;
7307 pmu->commit_txn = perf_pmu_nop_int;
7308 pmu->cancel_txn = perf_pmu_nop_void;
7312 if (!pmu->pmu_enable) {
7313 pmu->pmu_enable = perf_pmu_nop_void;
7314 pmu->pmu_disable = perf_pmu_nop_void;
7317 if (!pmu->event_idx)
7318 pmu->event_idx = perf_event_idx_default;
7320 list_add_rcu(&pmu->entry, &pmus);
7321 atomic_set(&pmu->exclusive_cnt, 0);
7324 mutex_unlock(&pmus_lock);
7329 device_del(pmu->dev);
7330 put_device(pmu->dev);
7333 if (pmu->type >= PERF_TYPE_MAX)
7334 idr_remove(&pmu_idr, pmu->type);
7337 free_percpu(pmu->pmu_disable_count);
7340 EXPORT_SYMBOL_GPL(perf_pmu_register);
7342 void perf_pmu_unregister(struct pmu *pmu)
7344 mutex_lock(&pmus_lock);
7345 list_del_rcu(&pmu->entry);
7346 mutex_unlock(&pmus_lock);
7349 * We dereference the pmu list under both SRCU and regular RCU, so
7350 * synchronize against both of those.
7352 synchronize_srcu(&pmus_srcu);
7355 free_percpu(pmu->pmu_disable_count);
7356 if (pmu->type >= PERF_TYPE_MAX)
7357 idr_remove(&pmu_idr, pmu->type);
7358 device_del(pmu->dev);
7359 put_device(pmu->dev);
7360 free_pmu_context(pmu);
7362 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
7364 static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)
7366 struct perf_event_context *ctx = NULL;
7369 if (!try_module_get(pmu->module))
7372 if (event->group_leader != event) {
7373 ctx = perf_event_ctx_lock(event->group_leader);
7378 ret = pmu->event_init(event);
7381 perf_event_ctx_unlock(event->group_leader, ctx);
7384 module_put(pmu->module);
7389 struct pmu *perf_init_event(struct perf_event *event)
7391 struct pmu *pmu = NULL;
7395 idx = srcu_read_lock(&pmus_srcu);
7398 pmu = idr_find(&pmu_idr, event->attr.type);
7401 ret = perf_try_init_event(pmu, event);
7407 list_for_each_entry_rcu(pmu, &pmus, entry) {
7408 ret = perf_try_init_event(pmu, event);
7412 if (ret != -ENOENT) {
7417 pmu = ERR_PTR(-ENOENT);
7419 srcu_read_unlock(&pmus_srcu, idx);
7424 static void account_event_cpu(struct perf_event *event, int cpu)
7429 if (is_cgroup_event(event))
7430 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
7433 static void account_event(struct perf_event *event)
7438 if (event->attach_state & PERF_ATTACH_TASK)
7439 static_key_slow_inc(&perf_sched_events.key);
7440 if (event->attr.mmap || event->attr.mmap_data)
7441 atomic_inc(&nr_mmap_events);
7442 if (event->attr.comm)
7443 atomic_inc(&nr_comm_events);
7444 if (event->attr.task)
7445 atomic_inc(&nr_task_events);
7446 if (event->attr.freq) {
7447 if (atomic_inc_return(&nr_freq_events) == 1)
7448 tick_nohz_full_kick_all();
7450 if (has_branch_stack(event))
7451 static_key_slow_inc(&perf_sched_events.key);
7452 if (is_cgroup_event(event))
7453 static_key_slow_inc(&perf_sched_events.key);
7455 account_event_cpu(event, event->cpu);
7459 * Allocate and initialize a event structure
7461 static struct perf_event *
7462 perf_event_alloc(struct perf_event_attr *attr, int cpu,
7463 struct task_struct *task,
7464 struct perf_event *group_leader,
7465 struct perf_event *parent_event,
7466 perf_overflow_handler_t overflow_handler,
7467 void *context, int cgroup_fd)
7470 struct perf_event *event;
7471 struct hw_perf_event *hwc;
7474 if ((unsigned)cpu >= nr_cpu_ids) {
7475 if (!task || cpu != -1)
7476 return ERR_PTR(-EINVAL);
7479 event = kzalloc(sizeof(*event), GFP_KERNEL);
7481 return ERR_PTR(-ENOMEM);
7484 * Single events are their own group leaders, with an
7485 * empty sibling list:
7488 group_leader = event;
7490 mutex_init(&event->child_mutex);
7491 INIT_LIST_HEAD(&event->child_list);
7493 INIT_LIST_HEAD(&event->group_entry);
7494 INIT_LIST_HEAD(&event->event_entry);
7495 INIT_LIST_HEAD(&event->sibling_list);
7496 INIT_LIST_HEAD(&event->rb_entry);
7497 INIT_LIST_HEAD(&event->active_entry);
7498 INIT_HLIST_NODE(&event->hlist_entry);
7501 init_waitqueue_head(&event->waitq);
7502 init_irq_work(&event->pending, perf_pending_event);
7504 mutex_init(&event->mmap_mutex);
7506 atomic_long_set(&event->refcount, 1);
7508 event->attr = *attr;
7509 event->group_leader = group_leader;
7513 event->parent = parent_event;
7515 event->ns = get_pid_ns(task_active_pid_ns(current));
7516 event->id = atomic64_inc_return(&perf_event_id);
7518 event->state = PERF_EVENT_STATE_INACTIVE;
7521 event->attach_state = PERF_ATTACH_TASK;
7523 * XXX pmu::event_init needs to know what task to account to
7524 * and we cannot use the ctx information because we need the
7525 * pmu before we get a ctx.
7527 event->hw.target = task;
7530 event->clock = &local_clock;
7532 event->clock = parent_event->clock;
7534 if (!overflow_handler && parent_event) {
7535 overflow_handler = parent_event->overflow_handler;
7536 context = parent_event->overflow_handler_context;
7539 event->overflow_handler = overflow_handler;
7540 event->overflow_handler_context = context;
7542 perf_event__state_init(event);
7547 hwc->sample_period = attr->sample_period;
7548 if (attr->freq && attr->sample_freq)
7549 hwc->sample_period = 1;
7550 hwc->last_period = hwc->sample_period;
7552 local64_set(&hwc->period_left, hwc->sample_period);
7555 * we currently do not support PERF_FORMAT_GROUP on inherited events
7557 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7560 if (!has_branch_stack(event))
7561 event->attr.branch_sample_type = 0;
7563 if (cgroup_fd != -1) {
7564 err = perf_cgroup_connect(cgroup_fd, event, attr, group_leader);
7569 pmu = perf_init_event(event);
7572 else if (IS_ERR(pmu)) {
7577 err = exclusive_event_init(event);
7581 if (!event->parent) {
7582 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7583 err = get_callchain_buffers();
7592 exclusive_event_destroy(event);
7596 event->destroy(event);
7597 module_put(pmu->module);
7599 if (is_cgroup_event(event))
7600 perf_detach_cgroup(event);
7602 put_pid_ns(event->ns);
7605 return ERR_PTR(err);
7608 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7609 struct perf_event_attr *attr)
7614 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7618 * zero the full structure, so that a short copy will be nice.
7620 memset(attr, 0, sizeof(*attr));
7622 ret = get_user(size, &uattr->size);
7626 if (size > PAGE_SIZE) /* silly large */
7629 if (!size) /* abi compat */
7630 size = PERF_ATTR_SIZE_VER0;
7632 if (size < PERF_ATTR_SIZE_VER0)
7636 * If we're handed a bigger struct than we know of,
7637 * ensure all the unknown bits are 0 - i.e. new
7638 * user-space does not rely on any kernel feature
7639 * extensions we dont know about yet.
7641 if (size > sizeof(*attr)) {
7642 unsigned char __user *addr;
7643 unsigned char __user *end;
7646 addr = (void __user *)uattr + sizeof(*attr);
7647 end = (void __user *)uattr + size;
7649 for (; addr < end; addr++) {
7650 ret = get_user(val, addr);
7656 size = sizeof(*attr);
7659 ret = copy_from_user(attr, uattr, size);
7663 if (attr->__reserved_1)
7666 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7669 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7672 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7673 u64 mask = attr->branch_sample_type;
7675 /* only using defined bits */
7676 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7679 /* at least one branch bit must be set */
7680 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7683 /* propagate priv level, when not set for branch */
7684 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7686 /* exclude_kernel checked on syscall entry */
7687 if (!attr->exclude_kernel)
7688 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7690 if (!attr->exclude_user)
7691 mask |= PERF_SAMPLE_BRANCH_USER;
7693 if (!attr->exclude_hv)
7694 mask |= PERF_SAMPLE_BRANCH_HV;
7696 * adjust user setting (for HW filter setup)
7698 attr->branch_sample_type = mask;
7700 /* privileged levels capture (kernel, hv): check permissions */
7701 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7702 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7706 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7707 ret = perf_reg_validate(attr->sample_regs_user);
7712 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7713 if (!arch_perf_have_user_stack_dump())
7717 * We have __u32 type for the size, but so far
7718 * we can only use __u16 as maximum due to the
7719 * __u16 sample size limit.
7721 if (attr->sample_stack_user >= USHRT_MAX)
7723 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7727 if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
7728 ret = perf_reg_validate(attr->sample_regs_intr);
7733 put_user(sizeof(*attr), &uattr->size);
7739 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7741 struct ring_buffer *rb = NULL;
7747 /* don't allow circular references */
7748 if (event == output_event)
7752 * Don't allow cross-cpu buffers
7754 if (output_event->cpu != event->cpu)
7758 * If its not a per-cpu rb, it must be the same task.
7760 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7764 * Mixing clocks in the same buffer is trouble you don't need.
7766 if (output_event->clock != event->clock)
7770 * If both events generate aux data, they must be on the same PMU
7772 if (has_aux(event) && has_aux(output_event) &&
7773 event->pmu != output_event->pmu)
7777 mutex_lock(&event->mmap_mutex);
7778 /* Can't redirect output if we've got an active mmap() */
7779 if (atomic_read(&event->mmap_count))
7783 /* get the rb we want to redirect to */
7784 rb = ring_buffer_get(output_event);
7789 ring_buffer_attach(event, rb);
7793 mutex_unlock(&event->mmap_mutex);
7799 static void mutex_lock_double(struct mutex *a, struct mutex *b)
7805 mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
7808 static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id)
7810 bool nmi_safe = false;
7813 case CLOCK_MONOTONIC:
7814 event->clock = &ktime_get_mono_fast_ns;
7818 case CLOCK_MONOTONIC_RAW:
7819 event->clock = &ktime_get_raw_fast_ns;
7823 case CLOCK_REALTIME:
7824 event->clock = &ktime_get_real_ns;
7827 case CLOCK_BOOTTIME:
7828 event->clock = &ktime_get_boot_ns;
7832 event->clock = &ktime_get_tai_ns;
7839 if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI))
7846 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7848 * @attr_uptr: event_id type attributes for monitoring/sampling
7851 * @group_fd: group leader event fd
7853 SYSCALL_DEFINE5(perf_event_open,
7854 struct perf_event_attr __user *, attr_uptr,
7855 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7857 struct perf_event *group_leader = NULL, *output_event = NULL;
7858 struct perf_event *event, *sibling;
7859 struct perf_event_attr attr;
7860 struct perf_event_context *ctx, *uninitialized_var(gctx);
7861 struct file *event_file = NULL;
7862 struct fd group = {NULL, 0};
7863 struct task_struct *task = NULL;
7868 int f_flags = O_RDWR;
7871 /* for future expandability... */
7872 if (flags & ~PERF_FLAG_ALL)
7875 err = perf_copy_attr(attr_uptr, &attr);
7879 if (!attr.exclude_kernel) {
7880 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7885 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7888 if (attr.sample_period & (1ULL << 63))
7893 * In cgroup mode, the pid argument is used to pass the fd
7894 * opened to the cgroup directory in cgroupfs. The cpu argument
7895 * designates the cpu on which to monitor threads from that
7898 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7901 if (flags & PERF_FLAG_FD_CLOEXEC)
7902 f_flags |= O_CLOEXEC;
7904 event_fd = get_unused_fd_flags(f_flags);
7908 if (group_fd != -1) {
7909 err = perf_fget_light(group_fd, &group);
7912 group_leader = group.file->private_data;
7913 if (flags & PERF_FLAG_FD_OUTPUT)
7914 output_event = group_leader;
7915 if (flags & PERF_FLAG_FD_NO_GROUP)
7916 group_leader = NULL;
7919 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7920 task = find_lively_task_by_vpid(pid);
7922 err = PTR_ERR(task);
7927 if (task && group_leader &&
7928 group_leader->attr.inherit != attr.inherit) {
7935 if (flags & PERF_FLAG_PID_CGROUP)
7938 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7939 NULL, NULL, cgroup_fd);
7940 if (IS_ERR(event)) {
7941 err = PTR_ERR(event);
7945 if (is_sampling_event(event)) {
7946 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7952 account_event(event);
7955 * Special case software events and allow them to be part of
7956 * any hardware group.
7960 if (attr.use_clockid) {
7961 err = perf_event_set_clock(event, attr.clockid);
7967 (is_software_event(event) != is_software_event(group_leader))) {
7968 if (is_software_event(event)) {
7970 * If event and group_leader are not both a software
7971 * event, and event is, then group leader is not.
7973 * Allow the addition of software events to !software
7974 * groups, this is safe because software events never
7977 pmu = group_leader->pmu;
7978 } else if (is_software_event(group_leader) &&
7979 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7981 * In case the group is a pure software group, and we
7982 * try to add a hardware event, move the whole group to
7983 * the hardware context.
7990 * Get the target context (task or percpu):
7992 ctx = find_get_context(pmu, task, event);
7998 if ((pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && group_leader) {
8004 put_task_struct(task);
8009 * Look up the group leader (we will attach this event to it):
8015 * Do not allow a recursive hierarchy (this new sibling
8016 * becoming part of another group-sibling):
8018 if (group_leader->group_leader != group_leader)
8021 /* All events in a group should have the same clock */
8022 if (group_leader->clock != event->clock)
8026 * Do not allow to attach to a group in a different
8027 * task or CPU context:
8031 * Make sure we're both on the same task, or both
8034 if (group_leader->ctx->task != ctx->task)
8038 * Make sure we're both events for the same CPU;
8039 * grouping events for different CPUs is broken; since
8040 * you can never concurrently schedule them anyhow.
8042 if (group_leader->cpu != event->cpu)
8045 if (group_leader->ctx != ctx)
8050 * Only a group leader can be exclusive or pinned
8052 if (attr.exclusive || attr.pinned)
8057 err = perf_event_set_output(event, output_event);
8062 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
8064 if (IS_ERR(event_file)) {
8065 err = PTR_ERR(event_file);
8070 gctx = group_leader->ctx;
8073 * See perf_event_ctx_lock() for comments on the details
8074 * of swizzling perf_event::ctx.
8076 mutex_lock_double(&gctx->mutex, &ctx->mutex);
8078 perf_remove_from_context(group_leader, false);
8080 list_for_each_entry(sibling, &group_leader->sibling_list,
8082 perf_remove_from_context(sibling, false);
8086 mutex_lock(&ctx->mutex);
8089 WARN_ON_ONCE(ctx->parent_ctx);
8093 * Wait for everybody to stop referencing the events through
8094 * the old lists, before installing it on new lists.
8099 * Install the group siblings before the group leader.
8101 * Because a group leader will try and install the entire group
8102 * (through the sibling list, which is still in-tact), we can
8103 * end up with siblings installed in the wrong context.
8105 * By installing siblings first we NO-OP because they're not
8106 * reachable through the group lists.
8108 list_for_each_entry(sibling, &group_leader->sibling_list,
8110 perf_event__state_init(sibling);
8111 perf_install_in_context(ctx, sibling, sibling->cpu);
8116 * Removing from the context ends up with disabled
8117 * event. What we want here is event in the initial
8118 * startup state, ready to be add into new context.
8120 perf_event__state_init(group_leader);
8121 perf_install_in_context(ctx, group_leader, group_leader->cpu);
8125 if (!exclusive_event_installable(event, ctx)) {
8127 mutex_unlock(&ctx->mutex);
8132 perf_install_in_context(ctx, event, event->cpu);
8133 perf_unpin_context(ctx);
8136 mutex_unlock(&gctx->mutex);
8139 mutex_unlock(&ctx->mutex);
8143 event->owner = current;
8145 mutex_lock(¤t->perf_event_mutex);
8146 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
8147 mutex_unlock(¤t->perf_event_mutex);
8150 * Precalculate sample_data sizes
8152 perf_event__header_size(event);
8153 perf_event__id_header_size(event);
8156 * Drop the reference on the group_event after placing the
8157 * new event on the sibling_list. This ensures destruction
8158 * of the group leader will find the pointer to itself in
8159 * perf_group_detach().
8162 fd_install(event_fd, event_file);
8166 perf_unpin_context(ctx);
8174 put_task_struct(task);
8178 put_unused_fd(event_fd);
8183 * perf_event_create_kernel_counter
8185 * @attr: attributes of the counter to create
8186 * @cpu: cpu in which the counter is bound
8187 * @task: task to profile (NULL for percpu)
8190 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
8191 struct task_struct *task,
8192 perf_overflow_handler_t overflow_handler,
8195 struct perf_event_context *ctx;
8196 struct perf_event *event;
8200 * Get the target context (task or percpu):
8203 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
8204 overflow_handler, context, -1);
8205 if (IS_ERR(event)) {
8206 err = PTR_ERR(event);
8210 /* Mark owner so we could distinguish it from user events. */
8211 event->owner = EVENT_OWNER_KERNEL;
8213 account_event(event);
8215 ctx = find_get_context(event->pmu, task, event);
8221 WARN_ON_ONCE(ctx->parent_ctx);
8222 mutex_lock(&ctx->mutex);
8223 if (!exclusive_event_installable(event, ctx)) {
8224 mutex_unlock(&ctx->mutex);
8225 perf_unpin_context(ctx);
8231 perf_install_in_context(ctx, event, cpu);
8232 perf_unpin_context(ctx);
8233 mutex_unlock(&ctx->mutex);
8240 return ERR_PTR(err);
8242 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
8244 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
8246 struct perf_event_context *src_ctx;
8247 struct perf_event_context *dst_ctx;
8248 struct perf_event *event, *tmp;
8251 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
8252 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
8255 * See perf_event_ctx_lock() for comments on the details
8256 * of swizzling perf_event::ctx.
8258 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex);
8259 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
8261 perf_remove_from_context(event, false);
8262 unaccount_event_cpu(event, src_cpu);
8264 list_add(&event->migrate_entry, &events);
8268 * Wait for the events to quiesce before re-instating them.
8273 * Re-instate events in 2 passes.
8275 * Skip over group leaders and only install siblings on this first
8276 * pass, siblings will not get enabled without a leader, however a
8277 * leader will enable its siblings, even if those are still on the old
8280 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
8281 if (event->group_leader == event)
8284 list_del(&event->migrate_entry);
8285 if (event->state >= PERF_EVENT_STATE_OFF)
8286 event->state = PERF_EVENT_STATE_INACTIVE;
8287 account_event_cpu(event, dst_cpu);
8288 perf_install_in_context(dst_ctx, event, dst_cpu);
8293 * Once all the siblings are setup properly, install the group leaders
8296 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
8297 list_del(&event->migrate_entry);
8298 if (event->state >= PERF_EVENT_STATE_OFF)
8299 event->state = PERF_EVENT_STATE_INACTIVE;
8300 account_event_cpu(event, dst_cpu);
8301 perf_install_in_context(dst_ctx, event, dst_cpu);
8304 mutex_unlock(&dst_ctx->mutex);
8305 mutex_unlock(&src_ctx->mutex);
8307 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
8309 static void sync_child_event(struct perf_event *child_event,
8310 struct task_struct *child)
8312 struct perf_event *parent_event = child_event->parent;
8315 if (child_event->attr.inherit_stat)
8316 perf_event_read_event(child_event, child);
8318 child_val = perf_event_count(child_event);
8321 * Add back the child's count to the parent's count:
8323 atomic64_add(child_val, &parent_event->child_count);
8324 atomic64_add(child_event->total_time_enabled,
8325 &parent_event->child_total_time_enabled);
8326 atomic64_add(child_event->total_time_running,
8327 &parent_event->child_total_time_running);
8330 * Remove this event from the parent's list
8332 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8333 mutex_lock(&parent_event->child_mutex);
8334 list_del_init(&child_event->child_list);
8335 mutex_unlock(&parent_event->child_mutex);
8338 * Make sure user/parent get notified, that we just
8341 perf_event_wakeup(parent_event);
8344 * Release the parent event, if this was the last
8347 put_event(parent_event);
8351 __perf_event_exit_task(struct perf_event *child_event,
8352 struct perf_event_context *child_ctx,
8353 struct task_struct *child)
8356 * Do not destroy the 'original' grouping; because of the context
8357 * switch optimization the original events could've ended up in a
8358 * random child task.
8360 * If we were to destroy the original group, all group related
8361 * operations would cease to function properly after this random
8364 * Do destroy all inherited groups, we don't care about those
8365 * and being thorough is better.
8367 perf_remove_from_context(child_event, !!child_event->parent);
8370 * It can happen that the parent exits first, and has events
8371 * that are still around due to the child reference. These
8372 * events need to be zapped.
8374 if (child_event->parent) {
8375 sync_child_event(child_event, child);
8376 free_event(child_event);
8378 child_event->state = PERF_EVENT_STATE_EXIT;
8379 perf_event_wakeup(child_event);
8383 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
8385 struct perf_event *child_event, *next;
8386 struct perf_event_context *child_ctx, *clone_ctx = NULL;
8387 unsigned long flags;
8389 if (likely(!child->perf_event_ctxp[ctxn])) {
8390 perf_event_task(child, NULL, 0);
8394 local_irq_save(flags);
8396 * We can't reschedule here because interrupts are disabled,
8397 * and either child is current or it is a task that can't be
8398 * scheduled, so we are now safe from rescheduling changing
8401 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
8404 * Take the context lock here so that if find_get_context is
8405 * reading child->perf_event_ctxp, we wait until it has
8406 * incremented the context's refcount before we do put_ctx below.
8408 raw_spin_lock(&child_ctx->lock);
8409 task_ctx_sched_out(child_ctx);
8410 child->perf_event_ctxp[ctxn] = NULL;
8413 * If this context is a clone; unclone it so it can't get
8414 * swapped to another process while we're removing all
8415 * the events from it.
8417 clone_ctx = unclone_ctx(child_ctx);
8418 update_context_time(child_ctx);
8419 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8425 * Report the task dead after unscheduling the events so that we
8426 * won't get any samples after PERF_RECORD_EXIT. We can however still
8427 * get a few PERF_RECORD_READ events.
8429 perf_event_task(child, child_ctx, 0);
8432 * We can recurse on the same lock type through:
8434 * __perf_event_exit_task()
8435 * sync_child_event()
8437 * mutex_lock(&ctx->mutex)
8439 * But since its the parent context it won't be the same instance.
8441 mutex_lock(&child_ctx->mutex);
8443 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
8444 __perf_event_exit_task(child_event, child_ctx, child);
8446 mutex_unlock(&child_ctx->mutex);
8452 * When a child task exits, feed back event values to parent events.
8454 void perf_event_exit_task(struct task_struct *child)
8456 struct perf_event *event, *tmp;
8459 mutex_lock(&child->perf_event_mutex);
8460 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
8462 list_del_init(&event->owner_entry);
8465 * Ensure the list deletion is visible before we clear
8466 * the owner, closes a race against perf_release() where
8467 * we need to serialize on the owner->perf_event_mutex.
8470 event->owner = NULL;
8472 mutex_unlock(&child->perf_event_mutex);
8474 for_each_task_context_nr(ctxn)
8475 perf_event_exit_task_context(child, ctxn);
8478 static void perf_free_event(struct perf_event *event,
8479 struct perf_event_context *ctx)
8481 struct perf_event *parent = event->parent;
8483 if (WARN_ON_ONCE(!parent))
8486 mutex_lock(&parent->child_mutex);
8487 list_del_init(&event->child_list);
8488 mutex_unlock(&parent->child_mutex);
8492 raw_spin_lock_irq(&ctx->lock);
8493 perf_group_detach(event);
8494 list_del_event(event, ctx);
8495 raw_spin_unlock_irq(&ctx->lock);
8500 * Free an unexposed, unused context as created by inheritance by
8501 * perf_event_init_task below, used by fork() in case of fail.
8503 * Not all locks are strictly required, but take them anyway to be nice and
8504 * help out with the lockdep assertions.
8506 void perf_event_free_task(struct task_struct *task)
8508 struct perf_event_context *ctx;
8509 struct perf_event *event, *tmp;
8512 for_each_task_context_nr(ctxn) {
8513 ctx = task->perf_event_ctxp[ctxn];
8517 mutex_lock(&ctx->mutex);
8519 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
8521 perf_free_event(event, ctx);
8523 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
8525 perf_free_event(event, ctx);
8527 if (!list_empty(&ctx->pinned_groups) ||
8528 !list_empty(&ctx->flexible_groups))
8531 mutex_unlock(&ctx->mutex);
8537 void perf_event_delayed_put(struct task_struct *task)
8541 for_each_task_context_nr(ctxn)
8542 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
8546 * inherit a event from parent task to child task:
8548 static struct perf_event *
8549 inherit_event(struct perf_event *parent_event,
8550 struct task_struct *parent,
8551 struct perf_event_context *parent_ctx,
8552 struct task_struct *child,
8553 struct perf_event *group_leader,
8554 struct perf_event_context *child_ctx)
8556 enum perf_event_active_state parent_state = parent_event->state;
8557 struct perf_event *child_event;
8558 unsigned long flags;
8561 * Instead of creating recursive hierarchies of events,
8562 * we link inherited events back to the original parent,
8563 * which has a filp for sure, which we use as the reference
8566 if (parent_event->parent)
8567 parent_event = parent_event->parent;
8569 child_event = perf_event_alloc(&parent_event->attr,
8572 group_leader, parent_event,
8574 if (IS_ERR(child_event))
8577 if (is_orphaned_event(parent_event) ||
8578 !atomic_long_inc_not_zero(&parent_event->refcount)) {
8579 free_event(child_event);
8586 * Make the child state follow the state of the parent event,
8587 * not its attr.disabled bit. We hold the parent's mutex,
8588 * so we won't race with perf_event_{en, dis}able_family.
8590 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
8591 child_event->state = PERF_EVENT_STATE_INACTIVE;
8593 child_event->state = PERF_EVENT_STATE_OFF;
8595 if (parent_event->attr.freq) {
8596 u64 sample_period = parent_event->hw.sample_period;
8597 struct hw_perf_event *hwc = &child_event->hw;
8599 hwc->sample_period = sample_period;
8600 hwc->last_period = sample_period;
8602 local64_set(&hwc->period_left, sample_period);
8605 child_event->ctx = child_ctx;
8606 child_event->overflow_handler = parent_event->overflow_handler;
8607 child_event->overflow_handler_context
8608 = parent_event->overflow_handler_context;
8611 * Precalculate sample_data sizes
8613 perf_event__header_size(child_event);
8614 perf_event__id_header_size(child_event);
8617 * Link it up in the child's context:
8619 raw_spin_lock_irqsave(&child_ctx->lock, flags);
8620 add_event_to_ctx(child_event, child_ctx);
8621 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8624 * Link this into the parent event's child list
8626 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8627 mutex_lock(&parent_event->child_mutex);
8628 list_add_tail(&child_event->child_list, &parent_event->child_list);
8629 mutex_unlock(&parent_event->child_mutex);
8634 static int inherit_group(struct perf_event *parent_event,
8635 struct task_struct *parent,
8636 struct perf_event_context *parent_ctx,
8637 struct task_struct *child,
8638 struct perf_event_context *child_ctx)
8640 struct perf_event *leader;
8641 struct perf_event *sub;
8642 struct perf_event *child_ctr;
8644 leader = inherit_event(parent_event, parent, parent_ctx,
8645 child, NULL, child_ctx);
8647 return PTR_ERR(leader);
8648 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
8649 child_ctr = inherit_event(sub, parent, parent_ctx,
8650 child, leader, child_ctx);
8651 if (IS_ERR(child_ctr))
8652 return PTR_ERR(child_ctr);
8658 inherit_task_group(struct perf_event *event, struct task_struct *parent,
8659 struct perf_event_context *parent_ctx,
8660 struct task_struct *child, int ctxn,
8664 struct perf_event_context *child_ctx;
8666 if (!event->attr.inherit) {
8671 child_ctx = child->perf_event_ctxp[ctxn];
8674 * This is executed from the parent task context, so
8675 * inherit events that have been marked for cloning.
8676 * First allocate and initialize a context for the
8680 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
8684 child->perf_event_ctxp[ctxn] = child_ctx;
8687 ret = inherit_group(event, parent, parent_ctx,
8697 * Initialize the perf_event context in task_struct
8699 static int perf_event_init_context(struct task_struct *child, int ctxn)
8701 struct perf_event_context *child_ctx, *parent_ctx;
8702 struct perf_event_context *cloned_ctx;
8703 struct perf_event *event;
8704 struct task_struct *parent = current;
8705 int inherited_all = 1;
8706 unsigned long flags;
8709 if (likely(!parent->perf_event_ctxp[ctxn]))
8713 * If the parent's context is a clone, pin it so it won't get
8716 parent_ctx = perf_pin_task_context(parent, ctxn);
8721 * No need to check if parent_ctx != NULL here; since we saw
8722 * it non-NULL earlier, the only reason for it to become NULL
8723 * is if we exit, and since we're currently in the middle of
8724 * a fork we can't be exiting at the same time.
8728 * Lock the parent list. No need to lock the child - not PID
8729 * hashed yet and not running, so nobody can access it.
8731 mutex_lock(&parent_ctx->mutex);
8734 * We dont have to disable NMIs - we are only looking at
8735 * the list, not manipulating it:
8737 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8738 ret = inherit_task_group(event, parent, parent_ctx,
8739 child, ctxn, &inherited_all);
8745 * We can't hold ctx->lock when iterating the ->flexible_group list due
8746 * to allocations, but we need to prevent rotation because
8747 * rotate_ctx() will change the list from interrupt context.
8749 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8750 parent_ctx->rotate_disable = 1;
8751 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8753 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8754 ret = inherit_task_group(event, parent, parent_ctx,
8755 child, ctxn, &inherited_all);
8760 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8761 parent_ctx->rotate_disable = 0;
8763 child_ctx = child->perf_event_ctxp[ctxn];
8765 if (child_ctx && inherited_all) {
8767 * Mark the child context as a clone of the parent
8768 * context, or of whatever the parent is a clone of.
8770 * Note that if the parent is a clone, the holding of
8771 * parent_ctx->lock avoids it from being uncloned.
8773 cloned_ctx = parent_ctx->parent_ctx;
8775 child_ctx->parent_ctx = cloned_ctx;
8776 child_ctx->parent_gen = parent_ctx->parent_gen;
8778 child_ctx->parent_ctx = parent_ctx;
8779 child_ctx->parent_gen = parent_ctx->generation;
8781 get_ctx(child_ctx->parent_ctx);
8784 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8785 mutex_unlock(&parent_ctx->mutex);
8787 perf_unpin_context(parent_ctx);
8788 put_ctx(parent_ctx);
8794 * Initialize the perf_event context in task_struct
8796 int perf_event_init_task(struct task_struct *child)
8800 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8801 mutex_init(&child->perf_event_mutex);
8802 INIT_LIST_HEAD(&child->perf_event_list);
8804 for_each_task_context_nr(ctxn) {
8805 ret = perf_event_init_context(child, ctxn);
8807 perf_event_free_task(child);
8815 static void __init perf_event_init_all_cpus(void)
8817 struct swevent_htable *swhash;
8820 for_each_possible_cpu(cpu) {
8821 swhash = &per_cpu(swevent_htable, cpu);
8822 mutex_init(&swhash->hlist_mutex);
8823 INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu));
8827 static void perf_event_init_cpu(int cpu)
8829 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8831 mutex_lock(&swhash->hlist_mutex);
8832 swhash->online = true;
8833 if (swhash->hlist_refcount > 0) {
8834 struct swevent_hlist *hlist;
8836 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8838 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8840 mutex_unlock(&swhash->hlist_mutex);
8843 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8844 static void __perf_event_exit_context(void *__info)
8846 struct remove_event re = { .detach_group = true };
8847 struct perf_event_context *ctx = __info;
8850 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8851 __perf_remove_from_context(&re);
8855 static void perf_event_exit_cpu_context(int cpu)
8857 struct perf_event_context *ctx;
8861 idx = srcu_read_lock(&pmus_srcu);
8862 list_for_each_entry_rcu(pmu, &pmus, entry) {
8863 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8865 mutex_lock(&ctx->mutex);
8866 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8867 mutex_unlock(&ctx->mutex);
8869 srcu_read_unlock(&pmus_srcu, idx);
8872 static void perf_event_exit_cpu(int cpu)
8874 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8876 perf_event_exit_cpu_context(cpu);
8878 mutex_lock(&swhash->hlist_mutex);
8879 swhash->online = false;
8880 swevent_hlist_release(swhash);
8881 mutex_unlock(&swhash->hlist_mutex);
8884 static inline void perf_event_exit_cpu(int cpu) { }
8888 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8892 for_each_online_cpu(cpu)
8893 perf_event_exit_cpu(cpu);
8899 * Run the perf reboot notifier at the very last possible moment so that
8900 * the generic watchdog code runs as long as possible.
8902 static struct notifier_block perf_reboot_notifier = {
8903 .notifier_call = perf_reboot,
8904 .priority = INT_MIN,
8908 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8910 unsigned int cpu = (long)hcpu;
8912 switch (action & ~CPU_TASKS_FROZEN) {
8914 case CPU_UP_PREPARE:
8915 case CPU_DOWN_FAILED:
8916 perf_event_init_cpu(cpu);
8919 case CPU_UP_CANCELED:
8920 case CPU_DOWN_PREPARE:
8921 perf_event_exit_cpu(cpu);
8930 void __init perf_event_init(void)
8936 perf_event_init_all_cpus();
8937 init_srcu_struct(&pmus_srcu);
8938 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8939 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8940 perf_pmu_register(&perf_task_clock, NULL, -1);
8942 perf_cpu_notifier(perf_cpu_notify);
8943 register_reboot_notifier(&perf_reboot_notifier);
8945 ret = init_hw_breakpoint();
8946 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8948 /* do not patch jump label more than once per second */
8949 jump_label_rate_limit(&perf_sched_events, HZ);
8952 * Build time assertion that we keep the data_head at the intended
8953 * location. IOW, validation we got the __reserved[] size right.
8955 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8959 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
8962 struct perf_pmu_events_attr *pmu_attr =
8963 container_of(attr, struct perf_pmu_events_attr, attr);
8965 if (pmu_attr->event_str)
8966 return sprintf(page, "%s\n", pmu_attr->event_str);
8971 static int __init perf_event_sysfs_init(void)
8976 mutex_lock(&pmus_lock);
8978 ret = bus_register(&pmu_bus);
8982 list_for_each_entry(pmu, &pmus, entry) {
8983 if (!pmu->name || pmu->type < 0)
8986 ret = pmu_dev_alloc(pmu);
8987 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8989 pmu_bus_running = 1;
8993 mutex_unlock(&pmus_lock);
8997 device_initcall(perf_event_sysfs_init);
8999 #ifdef CONFIG_CGROUP_PERF
9000 static struct cgroup_subsys_state *
9001 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9003 struct perf_cgroup *jc;
9005 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
9007 return ERR_PTR(-ENOMEM);
9009 jc->info = alloc_percpu(struct perf_cgroup_info);
9012 return ERR_PTR(-ENOMEM);
9018 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
9020 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
9022 free_percpu(jc->info);
9026 static int __perf_cgroup_move(void *info)
9028 struct task_struct *task = info;
9029 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
9033 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
9034 struct cgroup_taskset *tset)
9036 struct task_struct *task;
9038 cgroup_taskset_for_each(task, tset)
9039 task_function_call(task, __perf_cgroup_move, task);
9042 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
9043 struct cgroup_subsys_state *old_css,
9044 struct task_struct *task)
9047 * cgroup_exit() is called in the copy_process() failure path.
9048 * Ignore this case since the task hasn't ran yet, this avoids
9049 * trying to poke a half freed task state from generic code.
9051 if (!(task->flags & PF_EXITING))
9054 task_function_call(task, __perf_cgroup_move, task);
9057 struct cgroup_subsys perf_event_cgrp_subsys = {
9058 .css_alloc = perf_cgroup_css_alloc,
9059 .css_free = perf_cgroup_css_free,
9060 .exit = perf_cgroup_exit,
9061 .attach = perf_cgroup_attach,
9063 #endif /* CONFIG_CGROUP_PERF */