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>
48 #include <asm/irq_regs.h>
50 static struct workqueue_struct *perf_wq;
52 struct remote_function_call {
53 struct task_struct *p;
54 int (*func)(void *info);
59 static void remote_function(void *data)
61 struct remote_function_call *tfc = data;
62 struct task_struct *p = tfc->p;
66 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
70 tfc->ret = tfc->func(tfc->info);
74 * task_function_call - call a function on the cpu on which a task runs
75 * @p: the task to evaluate
76 * @func: the function to be called
77 * @info: the function call argument
79 * Calls the function @func when the task is currently running. This might
80 * be on the current CPU, which just calls the function directly
82 * returns: @func return value, or
83 * -ESRCH - when the process isn't running
84 * -EAGAIN - when the process moved away
87 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
89 struct remote_function_call data = {
93 .ret = -ESRCH, /* No such (running) process */
97 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
103 * cpu_function_call - call a function on the cpu
104 * @func: the function to be called
105 * @info: the function call argument
107 * Calls the function @func on the remote cpu.
109 * returns: @func return value or -ENXIO when the cpu is offline
111 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
113 struct remote_function_call data = {
117 .ret = -ENXIO, /* No such CPU */
120 smp_call_function_single(cpu, remote_function, &data, 1);
125 #define EVENT_OWNER_KERNEL ((void *) -1)
127 static bool is_kernel_event(struct perf_event *event)
129 return event->owner == EVENT_OWNER_KERNEL;
132 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
133 PERF_FLAG_FD_OUTPUT |\
134 PERF_FLAG_PID_CGROUP |\
135 PERF_FLAG_FD_CLOEXEC)
138 * branch priv levels that need permission checks
140 #define PERF_SAMPLE_BRANCH_PERM_PLM \
141 (PERF_SAMPLE_BRANCH_KERNEL |\
142 PERF_SAMPLE_BRANCH_HV)
145 EVENT_FLEXIBLE = 0x1,
147 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
151 * perf_sched_events : >0 events exist
152 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
154 struct static_key_deferred perf_sched_events __read_mostly;
155 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
156 static DEFINE_PER_CPU(int, perf_sched_cb_usages);
158 static atomic_t nr_mmap_events __read_mostly;
159 static atomic_t nr_comm_events __read_mostly;
160 static atomic_t nr_task_events __read_mostly;
161 static atomic_t nr_freq_events __read_mostly;
163 static LIST_HEAD(pmus);
164 static DEFINE_MUTEX(pmus_lock);
165 static struct srcu_struct pmus_srcu;
168 * perf event paranoia level:
169 * -1 - not paranoid at all
170 * 0 - disallow raw tracepoint access for unpriv
171 * 1 - disallow cpu events for unpriv
172 * 2 - disallow kernel profiling for unpriv
174 int sysctl_perf_event_paranoid __read_mostly = 1;
176 /* Minimum for 512 kiB + 1 user control page */
177 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
180 * max perf event sample rate
182 #define DEFAULT_MAX_SAMPLE_RATE 100000
183 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
184 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
186 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
188 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
189 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
191 static int perf_sample_allowed_ns __read_mostly =
192 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
194 void update_perf_cpu_limits(void)
196 u64 tmp = perf_sample_period_ns;
198 tmp *= sysctl_perf_cpu_time_max_percent;
200 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
203 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
205 int perf_proc_update_handler(struct ctl_table *table, int write,
206 void __user *buffer, size_t *lenp,
209 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
214 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
215 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
216 update_perf_cpu_limits();
221 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
223 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
224 void __user *buffer, size_t *lenp,
227 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
232 update_perf_cpu_limits();
238 * perf samples are done in some very critical code paths (NMIs).
239 * If they take too much CPU time, the system can lock up and not
240 * get any real work done. This will drop the sample rate when
241 * we detect that events are taking too long.
243 #define NR_ACCUMULATED_SAMPLES 128
244 static DEFINE_PER_CPU(u64, running_sample_length);
246 static void perf_duration_warn(struct irq_work *w)
248 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
249 u64 avg_local_sample_len;
250 u64 local_samples_len;
252 local_samples_len = __this_cpu_read(running_sample_length);
253 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
255 printk_ratelimited(KERN_WARNING
256 "perf interrupt took too long (%lld > %lld), lowering "
257 "kernel.perf_event_max_sample_rate to %d\n",
258 avg_local_sample_len, allowed_ns >> 1,
259 sysctl_perf_event_sample_rate);
262 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
264 void perf_sample_event_took(u64 sample_len_ns)
266 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
267 u64 avg_local_sample_len;
268 u64 local_samples_len;
273 /* decay the counter by 1 average sample */
274 local_samples_len = __this_cpu_read(running_sample_length);
275 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
276 local_samples_len += sample_len_ns;
277 __this_cpu_write(running_sample_length, local_samples_len);
280 * note: this will be biased artifically low until we have
281 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
282 * from having to maintain a count.
284 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
286 if (avg_local_sample_len <= allowed_ns)
289 if (max_samples_per_tick <= 1)
292 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
293 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
294 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
296 update_perf_cpu_limits();
298 if (!irq_work_queue(&perf_duration_work)) {
299 early_printk("perf interrupt took too long (%lld > %lld), lowering "
300 "kernel.perf_event_max_sample_rate to %d\n",
301 avg_local_sample_len, allowed_ns >> 1,
302 sysctl_perf_event_sample_rate);
306 static atomic64_t perf_event_id;
308 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
309 enum event_type_t event_type);
311 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
312 enum event_type_t event_type,
313 struct task_struct *task);
315 static void update_context_time(struct perf_event_context *ctx);
316 static u64 perf_event_time(struct perf_event *event);
318 void __weak perf_event_print_debug(void) { }
320 extern __weak const char *perf_pmu_name(void)
325 static inline u64 perf_clock(void)
327 return local_clock();
330 static inline struct perf_cpu_context *
331 __get_cpu_context(struct perf_event_context *ctx)
333 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
336 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
337 struct perf_event_context *ctx)
339 raw_spin_lock(&cpuctx->ctx.lock);
341 raw_spin_lock(&ctx->lock);
344 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
345 struct perf_event_context *ctx)
348 raw_spin_unlock(&ctx->lock);
349 raw_spin_unlock(&cpuctx->ctx.lock);
352 #ifdef CONFIG_CGROUP_PERF
355 perf_cgroup_match(struct perf_event *event)
357 struct perf_event_context *ctx = event->ctx;
358 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
360 /* @event doesn't care about cgroup */
364 /* wants specific cgroup scope but @cpuctx isn't associated with any */
369 * Cgroup scoping is recursive. An event enabled for a cgroup is
370 * also enabled for all its descendant cgroups. If @cpuctx's
371 * cgroup is a descendant of @event's (the test covers identity
372 * case), it's a match.
374 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
375 event->cgrp->css.cgroup);
378 static inline void perf_detach_cgroup(struct perf_event *event)
380 css_put(&event->cgrp->css);
384 static inline int is_cgroup_event(struct perf_event *event)
386 return event->cgrp != NULL;
389 static inline u64 perf_cgroup_event_time(struct perf_event *event)
391 struct perf_cgroup_info *t;
393 t = per_cpu_ptr(event->cgrp->info, event->cpu);
397 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
399 struct perf_cgroup_info *info;
404 info = this_cpu_ptr(cgrp->info);
406 info->time += now - info->timestamp;
407 info->timestamp = now;
410 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
412 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
414 __update_cgrp_time(cgrp_out);
417 static inline void update_cgrp_time_from_event(struct perf_event *event)
419 struct perf_cgroup *cgrp;
422 * ensure we access cgroup data only when needed and
423 * when we know the cgroup is pinned (css_get)
425 if (!is_cgroup_event(event))
428 cgrp = perf_cgroup_from_task(current);
430 * Do not update time when cgroup is not active
432 if (cgrp == event->cgrp)
433 __update_cgrp_time(event->cgrp);
437 perf_cgroup_set_timestamp(struct task_struct *task,
438 struct perf_event_context *ctx)
440 struct perf_cgroup *cgrp;
441 struct perf_cgroup_info *info;
444 * ctx->lock held by caller
445 * ensure we do not access cgroup data
446 * unless we have the cgroup pinned (css_get)
448 if (!task || !ctx->nr_cgroups)
451 cgrp = perf_cgroup_from_task(task);
452 info = this_cpu_ptr(cgrp->info);
453 info->timestamp = ctx->timestamp;
456 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
457 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
460 * reschedule events based on the cgroup constraint of task.
462 * mode SWOUT : schedule out everything
463 * mode SWIN : schedule in based on cgroup for next
465 void perf_cgroup_switch(struct task_struct *task, int mode)
467 struct perf_cpu_context *cpuctx;
472 * disable interrupts to avoid geting nr_cgroup
473 * changes via __perf_event_disable(). Also
476 local_irq_save(flags);
479 * we reschedule only in the presence of cgroup
480 * constrained events.
484 list_for_each_entry_rcu(pmu, &pmus, entry) {
485 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
486 if (cpuctx->unique_pmu != pmu)
487 continue; /* ensure we process each cpuctx once */
490 * perf_cgroup_events says at least one
491 * context on this CPU has cgroup events.
493 * ctx->nr_cgroups reports the number of cgroup
494 * events for a context.
496 if (cpuctx->ctx.nr_cgroups > 0) {
497 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
498 perf_pmu_disable(cpuctx->ctx.pmu);
500 if (mode & PERF_CGROUP_SWOUT) {
501 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
503 * must not be done before ctxswout due
504 * to event_filter_match() in event_sched_out()
509 if (mode & PERF_CGROUP_SWIN) {
510 WARN_ON_ONCE(cpuctx->cgrp);
512 * set cgrp before ctxsw in to allow
513 * event_filter_match() to not have to pass
516 cpuctx->cgrp = perf_cgroup_from_task(task);
517 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
519 perf_pmu_enable(cpuctx->ctx.pmu);
520 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
526 local_irq_restore(flags);
529 static inline void perf_cgroup_sched_out(struct task_struct *task,
530 struct task_struct *next)
532 struct perf_cgroup *cgrp1;
533 struct perf_cgroup *cgrp2 = NULL;
536 * we come here when we know perf_cgroup_events > 0
538 cgrp1 = perf_cgroup_from_task(task);
541 * next is NULL when called from perf_event_enable_on_exec()
542 * that will systematically cause a cgroup_switch()
545 cgrp2 = perf_cgroup_from_task(next);
548 * only schedule out current cgroup events if we know
549 * that we are switching to a different cgroup. Otherwise,
550 * do no touch the cgroup events.
553 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
556 static inline void perf_cgroup_sched_in(struct task_struct *prev,
557 struct task_struct *task)
559 struct perf_cgroup *cgrp1;
560 struct perf_cgroup *cgrp2 = NULL;
563 * we come here when we know perf_cgroup_events > 0
565 cgrp1 = perf_cgroup_from_task(task);
567 /* prev can never be NULL */
568 cgrp2 = perf_cgroup_from_task(prev);
571 * only need to schedule in cgroup events if we are changing
572 * cgroup during ctxsw. Cgroup events were not scheduled
573 * out of ctxsw out if that was not the case.
576 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
579 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
580 struct perf_event_attr *attr,
581 struct perf_event *group_leader)
583 struct perf_cgroup *cgrp;
584 struct cgroup_subsys_state *css;
585 struct fd f = fdget(fd);
591 css = css_tryget_online_from_dir(f.file->f_path.dentry,
592 &perf_event_cgrp_subsys);
598 cgrp = container_of(css, struct perf_cgroup, css);
602 * all events in a group must monitor
603 * the same cgroup because a task belongs
604 * to only one perf cgroup at a time
606 if (group_leader && group_leader->cgrp != cgrp) {
607 perf_detach_cgroup(event);
616 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
618 struct perf_cgroup_info *t;
619 t = per_cpu_ptr(event->cgrp->info, event->cpu);
620 event->shadow_ctx_time = now - t->timestamp;
624 perf_cgroup_defer_enabled(struct perf_event *event)
627 * when the current task's perf cgroup does not match
628 * the event's, we need to remember to call the
629 * perf_mark_enable() function the first time a task with
630 * a matching perf cgroup is scheduled in.
632 if (is_cgroup_event(event) && !perf_cgroup_match(event))
633 event->cgrp_defer_enabled = 1;
637 perf_cgroup_mark_enabled(struct perf_event *event,
638 struct perf_event_context *ctx)
640 struct perf_event *sub;
641 u64 tstamp = perf_event_time(event);
643 if (!event->cgrp_defer_enabled)
646 event->cgrp_defer_enabled = 0;
648 event->tstamp_enabled = tstamp - event->total_time_enabled;
649 list_for_each_entry(sub, &event->sibling_list, group_entry) {
650 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
651 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
652 sub->cgrp_defer_enabled = 0;
656 #else /* !CONFIG_CGROUP_PERF */
659 perf_cgroup_match(struct perf_event *event)
664 static inline void perf_detach_cgroup(struct perf_event *event)
667 static inline int is_cgroup_event(struct perf_event *event)
672 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
677 static inline void update_cgrp_time_from_event(struct perf_event *event)
681 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
685 static inline void perf_cgroup_sched_out(struct task_struct *task,
686 struct task_struct *next)
690 static inline void perf_cgroup_sched_in(struct task_struct *prev,
691 struct task_struct *task)
695 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
696 struct perf_event_attr *attr,
697 struct perf_event *group_leader)
703 perf_cgroup_set_timestamp(struct task_struct *task,
704 struct perf_event_context *ctx)
709 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
714 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
718 static inline u64 perf_cgroup_event_time(struct perf_event *event)
724 perf_cgroup_defer_enabled(struct perf_event *event)
729 perf_cgroup_mark_enabled(struct perf_event *event,
730 struct perf_event_context *ctx)
736 * set default to be dependent on timer tick just
739 #define PERF_CPU_HRTIMER (1000 / HZ)
741 * function must be called with interrupts disbled
743 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
745 struct perf_cpu_context *cpuctx;
746 enum hrtimer_restart ret = HRTIMER_NORESTART;
749 WARN_ON(!irqs_disabled());
751 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
753 rotations = perf_rotate_context(cpuctx);
756 * arm timer if needed
759 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
760 ret = HRTIMER_RESTART;
766 /* CPU is going down */
767 void perf_cpu_hrtimer_cancel(int cpu)
769 struct perf_cpu_context *cpuctx;
773 if (WARN_ON(cpu != smp_processor_id()))
776 local_irq_save(flags);
780 list_for_each_entry_rcu(pmu, &pmus, entry) {
781 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
783 if (pmu->task_ctx_nr == perf_sw_context)
786 hrtimer_cancel(&cpuctx->hrtimer);
791 local_irq_restore(flags);
794 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
796 struct hrtimer *hr = &cpuctx->hrtimer;
797 struct pmu *pmu = cpuctx->ctx.pmu;
800 /* no multiplexing needed for SW PMU */
801 if (pmu->task_ctx_nr == perf_sw_context)
805 * check default is sane, if not set then force to
806 * default interval (1/tick)
808 timer = pmu->hrtimer_interval_ms;
810 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
812 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
814 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
815 hr->function = perf_cpu_hrtimer_handler;
818 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
820 struct hrtimer *hr = &cpuctx->hrtimer;
821 struct pmu *pmu = cpuctx->ctx.pmu;
824 if (pmu->task_ctx_nr == perf_sw_context)
827 if (hrtimer_active(hr))
830 if (!hrtimer_callback_running(hr))
831 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
832 0, HRTIMER_MODE_REL_PINNED, 0);
835 void perf_pmu_disable(struct pmu *pmu)
837 int *count = this_cpu_ptr(pmu->pmu_disable_count);
839 pmu->pmu_disable(pmu);
842 void perf_pmu_enable(struct pmu *pmu)
844 int *count = this_cpu_ptr(pmu->pmu_disable_count);
846 pmu->pmu_enable(pmu);
849 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
852 * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
853 * perf_event_task_tick() are fully serialized because they're strictly cpu
854 * affine and perf_event_ctx{activate,deactivate} are called with IRQs
855 * disabled, while perf_event_task_tick is called from IRQ context.
857 static void perf_event_ctx_activate(struct perf_event_context *ctx)
859 struct list_head *head = this_cpu_ptr(&active_ctx_list);
861 WARN_ON(!irqs_disabled());
863 WARN_ON(!list_empty(&ctx->active_ctx_list));
865 list_add(&ctx->active_ctx_list, head);
868 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
870 WARN_ON(!irqs_disabled());
872 WARN_ON(list_empty(&ctx->active_ctx_list));
874 list_del_init(&ctx->active_ctx_list);
877 static void get_ctx(struct perf_event_context *ctx)
879 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
882 static void free_ctx(struct rcu_head *head)
884 struct perf_event_context *ctx;
886 ctx = container_of(head, struct perf_event_context, rcu_head);
887 kfree(ctx->task_ctx_data);
891 static void put_ctx(struct perf_event_context *ctx)
893 if (atomic_dec_and_test(&ctx->refcount)) {
895 put_ctx(ctx->parent_ctx);
897 put_task_struct(ctx->task);
898 call_rcu(&ctx->rcu_head, free_ctx);
903 * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
904 * perf_pmu_migrate_context() we need some magic.
906 * Those places that change perf_event::ctx will hold both
907 * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
909 * Lock ordering is by mutex address. There is one other site where
910 * perf_event_context::mutex nests and that is put_event(). But remember that
911 * that is a parent<->child context relation, and migration does not affect
912 * children, therefore these two orderings should not interact.
914 * The change in perf_event::ctx does not affect children (as claimed above)
915 * because the sys_perf_event_open() case will install a new event and break
916 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
917 * concerned with cpuctx and that doesn't have children.
919 * The places that change perf_event::ctx will issue:
921 * perf_remove_from_context();
923 * perf_install_in_context();
925 * to affect the change. The remove_from_context() + synchronize_rcu() should
926 * quiesce the event, after which we can install it in the new location. This
927 * means that only external vectors (perf_fops, prctl) can perturb the event
928 * while in transit. Therefore all such accessors should also acquire
929 * perf_event_context::mutex to serialize against this.
931 * However; because event->ctx can change while we're waiting to acquire
932 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
936 * task_struct::perf_event_mutex
937 * perf_event_context::mutex
938 * perf_event_context::lock
939 * perf_event::child_mutex;
940 * perf_event::mmap_mutex
943 static struct perf_event_context *
944 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
946 struct perf_event_context *ctx;
950 ctx = ACCESS_ONCE(event->ctx);
951 if (!atomic_inc_not_zero(&ctx->refcount)) {
957 mutex_lock_nested(&ctx->mutex, nesting);
958 if (event->ctx != ctx) {
959 mutex_unlock(&ctx->mutex);
967 static inline struct perf_event_context *
968 perf_event_ctx_lock(struct perf_event *event)
970 return perf_event_ctx_lock_nested(event, 0);
973 static void perf_event_ctx_unlock(struct perf_event *event,
974 struct perf_event_context *ctx)
976 mutex_unlock(&ctx->mutex);
981 * This must be done under the ctx->lock, such as to serialize against
982 * context_equiv(), therefore we cannot call put_ctx() since that might end up
983 * calling scheduler related locks and ctx->lock nests inside those.
985 static __must_check struct perf_event_context *
986 unclone_ctx(struct perf_event_context *ctx)
988 struct perf_event_context *parent_ctx = ctx->parent_ctx;
990 lockdep_assert_held(&ctx->lock);
993 ctx->parent_ctx = NULL;
999 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1002 * only top level events have the pid namespace they were created in
1005 event = event->parent;
1007 return task_tgid_nr_ns(p, event->ns);
1010 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1013 * only top level events have the pid namespace they were created in
1016 event = event->parent;
1018 return task_pid_nr_ns(p, event->ns);
1022 * If we inherit events we want to return the parent event id
1025 static u64 primary_event_id(struct perf_event *event)
1030 id = event->parent->id;
1036 * Get the perf_event_context for a task and lock it.
1037 * This has to cope with with the fact that until it is locked,
1038 * the context could get moved to another task.
1040 static struct perf_event_context *
1041 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1043 struct perf_event_context *ctx;
1047 * One of the few rules of preemptible RCU is that one cannot do
1048 * rcu_read_unlock() while holding a scheduler (or nested) lock when
1049 * part of the read side critical section was preemptible -- see
1050 * rcu_read_unlock_special().
1052 * Since ctx->lock nests under rq->lock we must ensure the entire read
1053 * side critical section is non-preemptible.
1057 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1060 * If this context is a clone of another, it might
1061 * get swapped for another underneath us by
1062 * perf_event_task_sched_out, though the
1063 * rcu_read_lock() protects us from any context
1064 * getting freed. Lock the context and check if it
1065 * got swapped before we could get the lock, and retry
1066 * if so. If we locked the right context, then it
1067 * can't get swapped on us any more.
1069 raw_spin_lock_irqsave(&ctx->lock, *flags);
1070 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1071 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1077 if (!atomic_inc_not_zero(&ctx->refcount)) {
1078 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1088 * Get the context for a task and increment its pin_count so it
1089 * can't get swapped to another task. This also increments its
1090 * reference count so that the context can't get freed.
1092 static struct perf_event_context *
1093 perf_pin_task_context(struct task_struct *task, int ctxn)
1095 struct perf_event_context *ctx;
1096 unsigned long flags;
1098 ctx = perf_lock_task_context(task, ctxn, &flags);
1101 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1106 static void perf_unpin_context(struct perf_event_context *ctx)
1108 unsigned long flags;
1110 raw_spin_lock_irqsave(&ctx->lock, flags);
1112 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1116 * Update the record of the current time in a context.
1118 static void update_context_time(struct perf_event_context *ctx)
1120 u64 now = perf_clock();
1122 ctx->time += now - ctx->timestamp;
1123 ctx->timestamp = now;
1126 static u64 perf_event_time(struct perf_event *event)
1128 struct perf_event_context *ctx = event->ctx;
1130 if (is_cgroup_event(event))
1131 return perf_cgroup_event_time(event);
1133 return ctx ? ctx->time : 0;
1137 * Update the total_time_enabled and total_time_running fields for a event.
1138 * The caller of this function needs to hold the ctx->lock.
1140 static void update_event_times(struct perf_event *event)
1142 struct perf_event_context *ctx = event->ctx;
1145 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1146 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1149 * in cgroup mode, time_enabled represents
1150 * the time the event was enabled AND active
1151 * tasks were in the monitored cgroup. This is
1152 * independent of the activity of the context as
1153 * there may be a mix of cgroup and non-cgroup events.
1155 * That is why we treat cgroup events differently
1158 if (is_cgroup_event(event))
1159 run_end = perf_cgroup_event_time(event);
1160 else if (ctx->is_active)
1161 run_end = ctx->time;
1163 run_end = event->tstamp_stopped;
1165 event->total_time_enabled = run_end - event->tstamp_enabled;
1167 if (event->state == PERF_EVENT_STATE_INACTIVE)
1168 run_end = event->tstamp_stopped;
1170 run_end = perf_event_time(event);
1172 event->total_time_running = run_end - event->tstamp_running;
1177 * Update total_time_enabled and total_time_running for all events in a group.
1179 static void update_group_times(struct perf_event *leader)
1181 struct perf_event *event;
1183 update_event_times(leader);
1184 list_for_each_entry(event, &leader->sibling_list, group_entry)
1185 update_event_times(event);
1188 static struct list_head *
1189 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1191 if (event->attr.pinned)
1192 return &ctx->pinned_groups;
1194 return &ctx->flexible_groups;
1198 * Add a event from the lists for its context.
1199 * Must be called with ctx->mutex and ctx->lock held.
1202 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1204 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1205 event->attach_state |= PERF_ATTACH_CONTEXT;
1208 * If we're a stand alone event or group leader, we go to the context
1209 * list, group events are kept attached to the group so that
1210 * perf_group_detach can, at all times, locate all siblings.
1212 if (event->group_leader == event) {
1213 struct list_head *list;
1215 if (is_software_event(event))
1216 event->group_flags |= PERF_GROUP_SOFTWARE;
1218 list = ctx_group_list(event, ctx);
1219 list_add_tail(&event->group_entry, list);
1222 if (is_cgroup_event(event))
1225 list_add_rcu(&event->event_entry, &ctx->event_list);
1227 if (event->attr.inherit_stat)
1234 * Initialize event state based on the perf_event_attr::disabled.
1236 static inline void perf_event__state_init(struct perf_event *event)
1238 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1239 PERF_EVENT_STATE_INACTIVE;
1243 * Called at perf_event creation and when events are attached/detached from a
1246 static void perf_event__read_size(struct perf_event *event)
1248 int entry = sizeof(u64); /* value */
1252 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1253 size += sizeof(u64);
1255 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1256 size += sizeof(u64);
1258 if (event->attr.read_format & PERF_FORMAT_ID)
1259 entry += sizeof(u64);
1261 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1262 nr += event->group_leader->nr_siblings;
1263 size += sizeof(u64);
1267 event->read_size = size;
1270 static void perf_event__header_size(struct perf_event *event)
1272 struct perf_sample_data *data;
1273 u64 sample_type = event->attr.sample_type;
1276 perf_event__read_size(event);
1278 if (sample_type & PERF_SAMPLE_IP)
1279 size += sizeof(data->ip);
1281 if (sample_type & PERF_SAMPLE_ADDR)
1282 size += sizeof(data->addr);
1284 if (sample_type & PERF_SAMPLE_PERIOD)
1285 size += sizeof(data->period);
1287 if (sample_type & PERF_SAMPLE_WEIGHT)
1288 size += sizeof(data->weight);
1290 if (sample_type & PERF_SAMPLE_READ)
1291 size += event->read_size;
1293 if (sample_type & PERF_SAMPLE_DATA_SRC)
1294 size += sizeof(data->data_src.val);
1296 if (sample_type & PERF_SAMPLE_TRANSACTION)
1297 size += sizeof(data->txn);
1299 event->header_size = size;
1302 static void perf_event__id_header_size(struct perf_event *event)
1304 struct perf_sample_data *data;
1305 u64 sample_type = event->attr.sample_type;
1308 if (sample_type & PERF_SAMPLE_TID)
1309 size += sizeof(data->tid_entry);
1311 if (sample_type & PERF_SAMPLE_TIME)
1312 size += sizeof(data->time);
1314 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1315 size += sizeof(data->id);
1317 if (sample_type & PERF_SAMPLE_ID)
1318 size += sizeof(data->id);
1320 if (sample_type & PERF_SAMPLE_STREAM_ID)
1321 size += sizeof(data->stream_id);
1323 if (sample_type & PERF_SAMPLE_CPU)
1324 size += sizeof(data->cpu_entry);
1326 event->id_header_size = size;
1329 static void perf_group_attach(struct perf_event *event)
1331 struct perf_event *group_leader = event->group_leader, *pos;
1334 * We can have double attach due to group movement in perf_event_open.
1336 if (event->attach_state & PERF_ATTACH_GROUP)
1339 event->attach_state |= PERF_ATTACH_GROUP;
1341 if (group_leader == event)
1344 WARN_ON_ONCE(group_leader->ctx != event->ctx);
1346 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1347 !is_software_event(event))
1348 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1350 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1351 group_leader->nr_siblings++;
1353 perf_event__header_size(group_leader);
1355 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1356 perf_event__header_size(pos);
1360 * Remove a event from the lists for its context.
1361 * Must be called with ctx->mutex and ctx->lock held.
1364 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1366 struct perf_cpu_context *cpuctx;
1368 WARN_ON_ONCE(event->ctx != ctx);
1369 lockdep_assert_held(&ctx->lock);
1372 * We can have double detach due to exit/hot-unplug + close.
1374 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1377 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1379 if (is_cgroup_event(event)) {
1381 cpuctx = __get_cpu_context(ctx);
1383 * if there are no more cgroup events
1384 * then cler cgrp to avoid stale pointer
1385 * in update_cgrp_time_from_cpuctx()
1387 if (!ctx->nr_cgroups)
1388 cpuctx->cgrp = NULL;
1392 if (event->attr.inherit_stat)
1395 list_del_rcu(&event->event_entry);
1397 if (event->group_leader == event)
1398 list_del_init(&event->group_entry);
1400 update_group_times(event);
1403 * If event was in error state, then keep it
1404 * that way, otherwise bogus counts will be
1405 * returned on read(). The only way to get out
1406 * of error state is by explicit re-enabling
1409 if (event->state > PERF_EVENT_STATE_OFF)
1410 event->state = PERF_EVENT_STATE_OFF;
1415 static void perf_group_detach(struct perf_event *event)
1417 struct perf_event *sibling, *tmp;
1418 struct list_head *list = NULL;
1421 * We can have double detach due to exit/hot-unplug + close.
1423 if (!(event->attach_state & PERF_ATTACH_GROUP))
1426 event->attach_state &= ~PERF_ATTACH_GROUP;
1429 * If this is a sibling, remove it from its group.
1431 if (event->group_leader != event) {
1432 list_del_init(&event->group_entry);
1433 event->group_leader->nr_siblings--;
1437 if (!list_empty(&event->group_entry))
1438 list = &event->group_entry;
1441 * If this was a group event with sibling events then
1442 * upgrade the siblings to singleton events by adding them
1443 * to whatever list we are on.
1445 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1447 list_move_tail(&sibling->group_entry, list);
1448 sibling->group_leader = sibling;
1450 /* Inherit group flags from the previous leader */
1451 sibling->group_flags = event->group_flags;
1453 WARN_ON_ONCE(sibling->ctx != event->ctx);
1457 perf_event__header_size(event->group_leader);
1459 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1460 perf_event__header_size(tmp);
1464 * User event without the task.
1466 static bool is_orphaned_event(struct perf_event *event)
1468 return event && !is_kernel_event(event) && !event->owner;
1472 * Event has a parent but parent's task finished and it's
1473 * alive only because of children holding refference.
1475 static bool is_orphaned_child(struct perf_event *event)
1477 return is_orphaned_event(event->parent);
1480 static void orphans_remove_work(struct work_struct *work);
1482 static void schedule_orphans_remove(struct perf_event_context *ctx)
1484 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1487 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1489 ctx->orphans_remove_sched = true;
1493 static int __init perf_workqueue_init(void)
1495 perf_wq = create_singlethread_workqueue("perf");
1496 WARN(!perf_wq, "failed to create perf workqueue\n");
1497 return perf_wq ? 0 : -1;
1500 core_initcall(perf_workqueue_init);
1503 event_filter_match(struct perf_event *event)
1505 return (event->cpu == -1 || event->cpu == smp_processor_id())
1506 && perf_cgroup_match(event);
1510 event_sched_out(struct perf_event *event,
1511 struct perf_cpu_context *cpuctx,
1512 struct perf_event_context *ctx)
1514 u64 tstamp = perf_event_time(event);
1517 WARN_ON_ONCE(event->ctx != ctx);
1518 lockdep_assert_held(&ctx->lock);
1521 * An event which could not be activated because of
1522 * filter mismatch still needs to have its timings
1523 * maintained, otherwise bogus information is return
1524 * via read() for time_enabled, time_running:
1526 if (event->state == PERF_EVENT_STATE_INACTIVE
1527 && !event_filter_match(event)) {
1528 delta = tstamp - event->tstamp_stopped;
1529 event->tstamp_running += delta;
1530 event->tstamp_stopped = tstamp;
1533 if (event->state != PERF_EVENT_STATE_ACTIVE)
1536 perf_pmu_disable(event->pmu);
1538 event->state = PERF_EVENT_STATE_INACTIVE;
1539 if (event->pending_disable) {
1540 event->pending_disable = 0;
1541 event->state = PERF_EVENT_STATE_OFF;
1543 event->tstamp_stopped = tstamp;
1544 event->pmu->del(event, 0);
1547 if (!is_software_event(event))
1548 cpuctx->active_oncpu--;
1549 if (!--ctx->nr_active)
1550 perf_event_ctx_deactivate(ctx);
1551 if (event->attr.freq && event->attr.sample_freq)
1553 if (event->attr.exclusive || !cpuctx->active_oncpu)
1554 cpuctx->exclusive = 0;
1556 if (is_orphaned_child(event))
1557 schedule_orphans_remove(ctx);
1559 perf_pmu_enable(event->pmu);
1563 group_sched_out(struct perf_event *group_event,
1564 struct perf_cpu_context *cpuctx,
1565 struct perf_event_context *ctx)
1567 struct perf_event *event;
1568 int state = group_event->state;
1570 event_sched_out(group_event, cpuctx, ctx);
1573 * Schedule out siblings (if any):
1575 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1576 event_sched_out(event, cpuctx, ctx);
1578 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1579 cpuctx->exclusive = 0;
1582 struct remove_event {
1583 struct perf_event *event;
1588 * Cross CPU call to remove a performance event
1590 * We disable the event on the hardware level first. After that we
1591 * remove it from the context list.
1593 static int __perf_remove_from_context(void *info)
1595 struct remove_event *re = info;
1596 struct perf_event *event = re->event;
1597 struct perf_event_context *ctx = event->ctx;
1598 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1600 raw_spin_lock(&ctx->lock);
1601 event_sched_out(event, cpuctx, ctx);
1602 if (re->detach_group)
1603 perf_group_detach(event);
1604 list_del_event(event, ctx);
1605 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1607 cpuctx->task_ctx = NULL;
1609 raw_spin_unlock(&ctx->lock);
1616 * Remove the event from a task's (or a CPU's) list of events.
1618 * CPU events are removed with a smp call. For task events we only
1619 * call when the task is on a CPU.
1621 * If event->ctx is a cloned context, callers must make sure that
1622 * every task struct that event->ctx->task could possibly point to
1623 * remains valid. This is OK when called from perf_release since
1624 * that only calls us on the top-level context, which can't be a clone.
1625 * When called from perf_event_exit_task, it's OK because the
1626 * context has been detached from its task.
1628 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1630 struct perf_event_context *ctx = event->ctx;
1631 struct task_struct *task = ctx->task;
1632 struct remove_event re = {
1634 .detach_group = detach_group,
1637 lockdep_assert_held(&ctx->mutex);
1641 * Per cpu events are removed via an smp call. The removal can
1642 * fail if the CPU is currently offline, but in that case we
1643 * already called __perf_remove_from_context from
1644 * perf_event_exit_cpu.
1646 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1651 if (!task_function_call(task, __perf_remove_from_context, &re))
1654 raw_spin_lock_irq(&ctx->lock);
1656 * If we failed to find a running task, but find the context active now
1657 * that we've acquired the ctx->lock, retry.
1659 if (ctx->is_active) {
1660 raw_spin_unlock_irq(&ctx->lock);
1662 * Reload the task pointer, it might have been changed by
1663 * a concurrent perf_event_context_sched_out().
1670 * Since the task isn't running, its safe to remove the event, us
1671 * holding the ctx->lock ensures the task won't get scheduled in.
1674 perf_group_detach(event);
1675 list_del_event(event, ctx);
1676 raw_spin_unlock_irq(&ctx->lock);
1680 * Cross CPU call to disable a performance event
1682 int __perf_event_disable(void *info)
1684 struct perf_event *event = info;
1685 struct perf_event_context *ctx = event->ctx;
1686 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1689 * If this is a per-task event, need to check whether this
1690 * event's task is the current task on this cpu.
1692 * Can trigger due to concurrent perf_event_context_sched_out()
1693 * flipping contexts around.
1695 if (ctx->task && cpuctx->task_ctx != ctx)
1698 raw_spin_lock(&ctx->lock);
1701 * If the event is on, turn it off.
1702 * If it is in error state, leave it in error state.
1704 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1705 update_context_time(ctx);
1706 update_cgrp_time_from_event(event);
1707 update_group_times(event);
1708 if (event == event->group_leader)
1709 group_sched_out(event, cpuctx, ctx);
1711 event_sched_out(event, cpuctx, ctx);
1712 event->state = PERF_EVENT_STATE_OFF;
1715 raw_spin_unlock(&ctx->lock);
1723 * If event->ctx is a cloned context, callers must make sure that
1724 * every task struct that event->ctx->task could possibly point to
1725 * remains valid. This condition is satisifed when called through
1726 * perf_event_for_each_child or perf_event_for_each because they
1727 * hold the top-level event's child_mutex, so any descendant that
1728 * goes to exit will block in sync_child_event.
1729 * When called from perf_pending_event it's OK because event->ctx
1730 * is the current context on this CPU and preemption is disabled,
1731 * hence we can't get into perf_event_task_sched_out for this context.
1733 static void _perf_event_disable(struct perf_event *event)
1735 struct perf_event_context *ctx = event->ctx;
1736 struct task_struct *task = ctx->task;
1740 * Disable the event on the cpu that it's on
1742 cpu_function_call(event->cpu, __perf_event_disable, event);
1747 if (!task_function_call(task, __perf_event_disable, event))
1750 raw_spin_lock_irq(&ctx->lock);
1752 * If the event is still active, we need to retry the cross-call.
1754 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1755 raw_spin_unlock_irq(&ctx->lock);
1757 * Reload the task pointer, it might have been changed by
1758 * a concurrent perf_event_context_sched_out().
1765 * Since we have the lock this context can't be scheduled
1766 * in, so we can change the state safely.
1768 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1769 update_group_times(event);
1770 event->state = PERF_EVENT_STATE_OFF;
1772 raw_spin_unlock_irq(&ctx->lock);
1776 * Strictly speaking kernel users cannot create groups and therefore this
1777 * interface does not need the perf_event_ctx_lock() magic.
1779 void perf_event_disable(struct perf_event *event)
1781 struct perf_event_context *ctx;
1783 ctx = perf_event_ctx_lock(event);
1784 _perf_event_disable(event);
1785 perf_event_ctx_unlock(event, ctx);
1787 EXPORT_SYMBOL_GPL(perf_event_disable);
1789 static void perf_set_shadow_time(struct perf_event *event,
1790 struct perf_event_context *ctx,
1794 * use the correct time source for the time snapshot
1796 * We could get by without this by leveraging the
1797 * fact that to get to this function, the caller
1798 * has most likely already called update_context_time()
1799 * and update_cgrp_time_xx() and thus both timestamp
1800 * are identical (or very close). Given that tstamp is,
1801 * already adjusted for cgroup, we could say that:
1802 * tstamp - ctx->timestamp
1804 * tstamp - cgrp->timestamp.
1806 * Then, in perf_output_read(), the calculation would
1807 * work with no changes because:
1808 * - event is guaranteed scheduled in
1809 * - no scheduled out in between
1810 * - thus the timestamp would be the same
1812 * But this is a bit hairy.
1814 * So instead, we have an explicit cgroup call to remain
1815 * within the time time source all along. We believe it
1816 * is cleaner and simpler to understand.
1818 if (is_cgroup_event(event))
1819 perf_cgroup_set_shadow_time(event, tstamp);
1821 event->shadow_ctx_time = tstamp - ctx->timestamp;
1824 #define MAX_INTERRUPTS (~0ULL)
1826 static void perf_log_throttle(struct perf_event *event, int enable);
1829 event_sched_in(struct perf_event *event,
1830 struct perf_cpu_context *cpuctx,
1831 struct perf_event_context *ctx)
1833 u64 tstamp = perf_event_time(event);
1836 lockdep_assert_held(&ctx->lock);
1838 if (event->state <= PERF_EVENT_STATE_OFF)
1841 event->state = PERF_EVENT_STATE_ACTIVE;
1842 event->oncpu = smp_processor_id();
1845 * Unthrottle events, since we scheduled we might have missed several
1846 * ticks already, also for a heavily scheduling task there is little
1847 * guarantee it'll get a tick in a timely manner.
1849 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1850 perf_log_throttle(event, 1);
1851 event->hw.interrupts = 0;
1855 * The new state must be visible before we turn it on in the hardware:
1859 perf_pmu_disable(event->pmu);
1861 event->tstamp_running += tstamp - event->tstamp_stopped;
1863 perf_set_shadow_time(event, ctx, tstamp);
1865 if (event->pmu->add(event, PERF_EF_START)) {
1866 event->state = PERF_EVENT_STATE_INACTIVE;
1872 if (!is_software_event(event))
1873 cpuctx->active_oncpu++;
1874 if (!ctx->nr_active++)
1875 perf_event_ctx_activate(ctx);
1876 if (event->attr.freq && event->attr.sample_freq)
1879 if (event->attr.exclusive)
1880 cpuctx->exclusive = 1;
1882 if (is_orphaned_child(event))
1883 schedule_orphans_remove(ctx);
1886 perf_pmu_enable(event->pmu);
1892 group_sched_in(struct perf_event *group_event,
1893 struct perf_cpu_context *cpuctx,
1894 struct perf_event_context *ctx)
1896 struct perf_event *event, *partial_group = NULL;
1897 struct pmu *pmu = ctx->pmu;
1898 u64 now = ctx->time;
1899 bool simulate = false;
1901 if (group_event->state == PERF_EVENT_STATE_OFF)
1904 pmu->start_txn(pmu);
1906 if (event_sched_in(group_event, cpuctx, ctx)) {
1907 pmu->cancel_txn(pmu);
1908 perf_cpu_hrtimer_restart(cpuctx);
1913 * Schedule in siblings as one group (if any):
1915 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1916 if (event_sched_in(event, cpuctx, ctx)) {
1917 partial_group = event;
1922 if (!pmu->commit_txn(pmu))
1927 * Groups can be scheduled in as one unit only, so undo any
1928 * partial group before returning:
1929 * The events up to the failed event are scheduled out normally,
1930 * tstamp_stopped will be updated.
1932 * The failed events and the remaining siblings need to have
1933 * their timings updated as if they had gone thru event_sched_in()
1934 * and event_sched_out(). This is required to get consistent timings
1935 * across the group. This also takes care of the case where the group
1936 * could never be scheduled by ensuring tstamp_stopped is set to mark
1937 * the time the event was actually stopped, such that time delta
1938 * calculation in update_event_times() is correct.
1940 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1941 if (event == partial_group)
1945 event->tstamp_running += now - event->tstamp_stopped;
1946 event->tstamp_stopped = now;
1948 event_sched_out(event, cpuctx, ctx);
1951 event_sched_out(group_event, cpuctx, ctx);
1953 pmu->cancel_txn(pmu);
1955 perf_cpu_hrtimer_restart(cpuctx);
1961 * Work out whether we can put this event group on the CPU now.
1963 static int group_can_go_on(struct perf_event *event,
1964 struct perf_cpu_context *cpuctx,
1968 * Groups consisting entirely of software events can always go on.
1970 if (event->group_flags & PERF_GROUP_SOFTWARE)
1973 * If an exclusive group is already on, no other hardware
1976 if (cpuctx->exclusive)
1979 * If this group is exclusive and there are already
1980 * events on the CPU, it can't go on.
1982 if (event->attr.exclusive && cpuctx->active_oncpu)
1985 * Otherwise, try to add it if all previous groups were able
1991 static void add_event_to_ctx(struct perf_event *event,
1992 struct perf_event_context *ctx)
1994 u64 tstamp = perf_event_time(event);
1996 list_add_event(event, ctx);
1997 perf_group_attach(event);
1998 event->tstamp_enabled = tstamp;
1999 event->tstamp_running = tstamp;
2000 event->tstamp_stopped = tstamp;
2003 static void task_ctx_sched_out(struct perf_event_context *ctx);
2005 ctx_sched_in(struct perf_event_context *ctx,
2006 struct perf_cpu_context *cpuctx,
2007 enum event_type_t event_type,
2008 struct task_struct *task);
2010 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2011 struct perf_event_context *ctx,
2012 struct task_struct *task)
2014 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2016 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2017 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2019 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2023 * Cross CPU call to install and enable a performance event
2025 * Must be called with ctx->mutex held
2027 static int __perf_install_in_context(void *info)
2029 struct perf_event *event = info;
2030 struct perf_event_context *ctx = event->ctx;
2031 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2032 struct perf_event_context *task_ctx = cpuctx->task_ctx;
2033 struct task_struct *task = current;
2035 perf_ctx_lock(cpuctx, task_ctx);
2036 perf_pmu_disable(cpuctx->ctx.pmu);
2039 * If there was an active task_ctx schedule it out.
2042 task_ctx_sched_out(task_ctx);
2045 * If the context we're installing events in is not the
2046 * active task_ctx, flip them.
2048 if (ctx->task && task_ctx != ctx) {
2050 raw_spin_unlock(&task_ctx->lock);
2051 raw_spin_lock(&ctx->lock);
2056 cpuctx->task_ctx = task_ctx;
2057 task = task_ctx->task;
2060 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
2062 update_context_time(ctx);
2064 * update cgrp time only if current cgrp
2065 * matches event->cgrp. Must be done before
2066 * calling add_event_to_ctx()
2068 update_cgrp_time_from_event(event);
2070 add_event_to_ctx(event, ctx);
2073 * Schedule everything back in
2075 perf_event_sched_in(cpuctx, task_ctx, task);
2077 perf_pmu_enable(cpuctx->ctx.pmu);
2078 perf_ctx_unlock(cpuctx, task_ctx);
2084 * Attach a performance event to a context
2086 * First we add the event to the list with the hardware enable bit
2087 * in event->hw_config cleared.
2089 * If the event is attached to a task which is on a CPU we use a smp
2090 * call to enable it in the task context. The task might have been
2091 * scheduled away, but we check this in the smp call again.
2094 perf_install_in_context(struct perf_event_context *ctx,
2095 struct perf_event *event,
2098 struct task_struct *task = ctx->task;
2100 lockdep_assert_held(&ctx->mutex);
2103 if (event->cpu != -1)
2108 * Per cpu events are installed via an smp call and
2109 * the install is always successful.
2111 cpu_function_call(cpu, __perf_install_in_context, event);
2116 if (!task_function_call(task, __perf_install_in_context, event))
2119 raw_spin_lock_irq(&ctx->lock);
2121 * If we failed to find a running task, but find the context active now
2122 * that we've acquired the ctx->lock, retry.
2124 if (ctx->is_active) {
2125 raw_spin_unlock_irq(&ctx->lock);
2127 * Reload the task pointer, it might have been changed by
2128 * a concurrent perf_event_context_sched_out().
2135 * Since the task isn't running, its safe to add the event, us holding
2136 * the ctx->lock ensures the task won't get scheduled in.
2138 add_event_to_ctx(event, ctx);
2139 raw_spin_unlock_irq(&ctx->lock);
2143 * Put a event into inactive state and update time fields.
2144 * Enabling the leader of a group effectively enables all
2145 * the group members that aren't explicitly disabled, so we
2146 * have to update their ->tstamp_enabled also.
2147 * Note: this works for group members as well as group leaders
2148 * since the non-leader members' sibling_lists will be empty.
2150 static void __perf_event_mark_enabled(struct perf_event *event)
2152 struct perf_event *sub;
2153 u64 tstamp = perf_event_time(event);
2155 event->state = PERF_EVENT_STATE_INACTIVE;
2156 event->tstamp_enabled = tstamp - event->total_time_enabled;
2157 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2158 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2159 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2164 * Cross CPU call to enable a performance event
2166 static int __perf_event_enable(void *info)
2168 struct perf_event *event = info;
2169 struct perf_event_context *ctx = event->ctx;
2170 struct perf_event *leader = event->group_leader;
2171 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2175 * There's a time window between 'ctx->is_active' check
2176 * in perf_event_enable function and this place having:
2178 * - ctx->lock unlocked
2180 * where the task could be killed and 'ctx' deactivated
2181 * by perf_event_exit_task.
2183 if (!ctx->is_active)
2186 raw_spin_lock(&ctx->lock);
2187 update_context_time(ctx);
2189 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2193 * set current task's cgroup time reference point
2195 perf_cgroup_set_timestamp(current, ctx);
2197 __perf_event_mark_enabled(event);
2199 if (!event_filter_match(event)) {
2200 if (is_cgroup_event(event))
2201 perf_cgroup_defer_enabled(event);
2206 * If the event is in a group and isn't the group leader,
2207 * then don't put it on unless the group is on.
2209 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2212 if (!group_can_go_on(event, cpuctx, 1)) {
2215 if (event == leader)
2216 err = group_sched_in(event, cpuctx, ctx);
2218 err = event_sched_in(event, cpuctx, ctx);
2223 * If this event can't go on and it's part of a
2224 * group, then the whole group has to come off.
2226 if (leader != event) {
2227 group_sched_out(leader, cpuctx, ctx);
2228 perf_cpu_hrtimer_restart(cpuctx);
2230 if (leader->attr.pinned) {
2231 update_group_times(leader);
2232 leader->state = PERF_EVENT_STATE_ERROR;
2237 raw_spin_unlock(&ctx->lock);
2245 * If event->ctx is a cloned context, callers must make sure that
2246 * every task struct that event->ctx->task could possibly point to
2247 * remains valid. This condition is satisfied when called through
2248 * perf_event_for_each_child or perf_event_for_each as described
2249 * for perf_event_disable.
2251 static void _perf_event_enable(struct perf_event *event)
2253 struct perf_event_context *ctx = event->ctx;
2254 struct task_struct *task = ctx->task;
2258 * Enable the event on the cpu that it's on
2260 cpu_function_call(event->cpu, __perf_event_enable, event);
2264 raw_spin_lock_irq(&ctx->lock);
2265 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2269 * If the event is in error state, clear that first.
2270 * That way, if we see the event in error state below, we
2271 * know that it has gone back into error state, as distinct
2272 * from the task having been scheduled away before the
2273 * cross-call arrived.
2275 if (event->state == PERF_EVENT_STATE_ERROR)
2276 event->state = PERF_EVENT_STATE_OFF;
2279 if (!ctx->is_active) {
2280 __perf_event_mark_enabled(event);
2284 raw_spin_unlock_irq(&ctx->lock);
2286 if (!task_function_call(task, __perf_event_enable, event))
2289 raw_spin_lock_irq(&ctx->lock);
2292 * If the context is active and the event is still off,
2293 * we need to retry the cross-call.
2295 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2297 * task could have been flipped by a concurrent
2298 * perf_event_context_sched_out()
2305 raw_spin_unlock_irq(&ctx->lock);
2309 * See perf_event_disable();
2311 void perf_event_enable(struct perf_event *event)
2313 struct perf_event_context *ctx;
2315 ctx = perf_event_ctx_lock(event);
2316 _perf_event_enable(event);
2317 perf_event_ctx_unlock(event, ctx);
2319 EXPORT_SYMBOL_GPL(perf_event_enable);
2321 static int _perf_event_refresh(struct perf_event *event, int refresh)
2324 * not supported on inherited events
2326 if (event->attr.inherit || !is_sampling_event(event))
2329 atomic_add(refresh, &event->event_limit);
2330 _perf_event_enable(event);
2336 * See perf_event_disable()
2338 int perf_event_refresh(struct perf_event *event, int refresh)
2340 struct perf_event_context *ctx;
2343 ctx = perf_event_ctx_lock(event);
2344 ret = _perf_event_refresh(event, refresh);
2345 perf_event_ctx_unlock(event, ctx);
2349 EXPORT_SYMBOL_GPL(perf_event_refresh);
2351 static void ctx_sched_out(struct perf_event_context *ctx,
2352 struct perf_cpu_context *cpuctx,
2353 enum event_type_t event_type)
2355 struct perf_event *event;
2356 int is_active = ctx->is_active;
2358 ctx->is_active &= ~event_type;
2359 if (likely(!ctx->nr_events))
2362 update_context_time(ctx);
2363 update_cgrp_time_from_cpuctx(cpuctx);
2364 if (!ctx->nr_active)
2367 perf_pmu_disable(ctx->pmu);
2368 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2369 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2370 group_sched_out(event, cpuctx, ctx);
2373 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2374 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2375 group_sched_out(event, cpuctx, ctx);
2377 perf_pmu_enable(ctx->pmu);
2381 * Test whether two contexts are equivalent, i.e. whether they have both been
2382 * cloned from the same version of the same context.
2384 * Equivalence is measured using a generation number in the context that is
2385 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2386 * and list_del_event().
2388 static int context_equiv(struct perf_event_context *ctx1,
2389 struct perf_event_context *ctx2)
2391 lockdep_assert_held(&ctx1->lock);
2392 lockdep_assert_held(&ctx2->lock);
2394 /* Pinning disables the swap optimization */
2395 if (ctx1->pin_count || ctx2->pin_count)
2398 /* If ctx1 is the parent of ctx2 */
2399 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2402 /* If ctx2 is the parent of ctx1 */
2403 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2407 * If ctx1 and ctx2 have the same parent; we flatten the parent
2408 * hierarchy, see perf_event_init_context().
2410 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2411 ctx1->parent_gen == ctx2->parent_gen)
2418 static void __perf_event_sync_stat(struct perf_event *event,
2419 struct perf_event *next_event)
2423 if (!event->attr.inherit_stat)
2427 * Update the event value, we cannot use perf_event_read()
2428 * because we're in the middle of a context switch and have IRQs
2429 * disabled, which upsets smp_call_function_single(), however
2430 * we know the event must be on the current CPU, therefore we
2431 * don't need to use it.
2433 switch (event->state) {
2434 case PERF_EVENT_STATE_ACTIVE:
2435 event->pmu->read(event);
2438 case PERF_EVENT_STATE_INACTIVE:
2439 update_event_times(event);
2447 * In order to keep per-task stats reliable we need to flip the event
2448 * values when we flip the contexts.
2450 value = local64_read(&next_event->count);
2451 value = local64_xchg(&event->count, value);
2452 local64_set(&next_event->count, value);
2454 swap(event->total_time_enabled, next_event->total_time_enabled);
2455 swap(event->total_time_running, next_event->total_time_running);
2458 * Since we swizzled the values, update the user visible data too.
2460 perf_event_update_userpage(event);
2461 perf_event_update_userpage(next_event);
2464 static void perf_event_sync_stat(struct perf_event_context *ctx,
2465 struct perf_event_context *next_ctx)
2467 struct perf_event *event, *next_event;
2472 update_context_time(ctx);
2474 event = list_first_entry(&ctx->event_list,
2475 struct perf_event, event_entry);
2477 next_event = list_first_entry(&next_ctx->event_list,
2478 struct perf_event, event_entry);
2480 while (&event->event_entry != &ctx->event_list &&
2481 &next_event->event_entry != &next_ctx->event_list) {
2483 __perf_event_sync_stat(event, next_event);
2485 event = list_next_entry(event, event_entry);
2486 next_event = list_next_entry(next_event, event_entry);
2490 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2491 struct task_struct *next)
2493 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2494 struct perf_event_context *next_ctx;
2495 struct perf_event_context *parent, *next_parent;
2496 struct perf_cpu_context *cpuctx;
2502 cpuctx = __get_cpu_context(ctx);
2503 if (!cpuctx->task_ctx)
2507 next_ctx = next->perf_event_ctxp[ctxn];
2511 parent = rcu_dereference(ctx->parent_ctx);
2512 next_parent = rcu_dereference(next_ctx->parent_ctx);
2514 /* If neither context have a parent context; they cannot be clones. */
2515 if (!parent && !next_parent)
2518 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2520 * Looks like the two contexts are clones, so we might be
2521 * able to optimize the context switch. We lock both
2522 * contexts and check that they are clones under the
2523 * lock (including re-checking that neither has been
2524 * uncloned in the meantime). It doesn't matter which
2525 * order we take the locks because no other cpu could
2526 * be trying to lock both of these tasks.
2528 raw_spin_lock(&ctx->lock);
2529 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2530 if (context_equiv(ctx, next_ctx)) {
2532 * XXX do we need a memory barrier of sorts
2533 * wrt to rcu_dereference() of perf_event_ctxp
2535 task->perf_event_ctxp[ctxn] = next_ctx;
2536 next->perf_event_ctxp[ctxn] = ctx;
2538 next_ctx->task = task;
2540 swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
2544 perf_event_sync_stat(ctx, next_ctx);
2546 raw_spin_unlock(&next_ctx->lock);
2547 raw_spin_unlock(&ctx->lock);
2553 raw_spin_lock(&ctx->lock);
2554 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2555 cpuctx->task_ctx = NULL;
2556 raw_spin_unlock(&ctx->lock);
2560 void perf_sched_cb_dec(struct pmu *pmu)
2562 this_cpu_dec(perf_sched_cb_usages);
2565 void perf_sched_cb_inc(struct pmu *pmu)
2567 this_cpu_inc(perf_sched_cb_usages);
2571 * This function provides the context switch callback to the lower code
2572 * layer. It is invoked ONLY when the context switch callback is enabled.
2574 static void perf_pmu_sched_task(struct task_struct *prev,
2575 struct task_struct *next,
2578 struct perf_cpu_context *cpuctx;
2580 unsigned long flags;
2585 local_irq_save(flags);
2589 list_for_each_entry_rcu(pmu, &pmus, entry) {
2590 if (pmu->sched_task) {
2591 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2593 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2595 perf_pmu_disable(pmu);
2597 pmu->sched_task(cpuctx->task_ctx, sched_in);
2599 perf_pmu_enable(pmu);
2601 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2607 local_irq_restore(flags);
2610 #define for_each_task_context_nr(ctxn) \
2611 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2614 * Called from scheduler to remove the events of the current task,
2615 * with interrupts disabled.
2617 * We stop each event and update the event value in event->count.
2619 * This does not protect us against NMI, but disable()
2620 * sets the disabled bit in the control field of event _before_
2621 * accessing the event control register. If a NMI hits, then it will
2622 * not restart the event.
2624 void __perf_event_task_sched_out(struct task_struct *task,
2625 struct task_struct *next)
2629 if (__this_cpu_read(perf_sched_cb_usages))
2630 perf_pmu_sched_task(task, next, false);
2632 for_each_task_context_nr(ctxn)
2633 perf_event_context_sched_out(task, ctxn, next);
2636 * if cgroup events exist on this CPU, then we need
2637 * to check if we have to switch out PMU state.
2638 * cgroup event are system-wide mode only
2640 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2641 perf_cgroup_sched_out(task, next);
2644 static void task_ctx_sched_out(struct perf_event_context *ctx)
2646 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2648 if (!cpuctx->task_ctx)
2651 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2654 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2655 cpuctx->task_ctx = NULL;
2659 * Called with IRQs disabled
2661 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2662 enum event_type_t event_type)
2664 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2668 ctx_pinned_sched_in(struct perf_event_context *ctx,
2669 struct perf_cpu_context *cpuctx)
2671 struct perf_event *event;
2673 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2674 if (event->state <= PERF_EVENT_STATE_OFF)
2676 if (!event_filter_match(event))
2679 /* may need to reset tstamp_enabled */
2680 if (is_cgroup_event(event))
2681 perf_cgroup_mark_enabled(event, ctx);
2683 if (group_can_go_on(event, cpuctx, 1))
2684 group_sched_in(event, cpuctx, ctx);
2687 * If this pinned group hasn't been scheduled,
2688 * put it in error state.
2690 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2691 update_group_times(event);
2692 event->state = PERF_EVENT_STATE_ERROR;
2698 ctx_flexible_sched_in(struct perf_event_context *ctx,
2699 struct perf_cpu_context *cpuctx)
2701 struct perf_event *event;
2704 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2705 /* Ignore events in OFF or ERROR state */
2706 if (event->state <= PERF_EVENT_STATE_OFF)
2709 * Listen to the 'cpu' scheduling filter constraint
2712 if (!event_filter_match(event))
2715 /* may need to reset tstamp_enabled */
2716 if (is_cgroup_event(event))
2717 perf_cgroup_mark_enabled(event, ctx);
2719 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2720 if (group_sched_in(event, cpuctx, ctx))
2727 ctx_sched_in(struct perf_event_context *ctx,
2728 struct perf_cpu_context *cpuctx,
2729 enum event_type_t event_type,
2730 struct task_struct *task)
2733 int is_active = ctx->is_active;
2735 ctx->is_active |= event_type;
2736 if (likely(!ctx->nr_events))
2740 ctx->timestamp = now;
2741 perf_cgroup_set_timestamp(task, ctx);
2743 * First go through the list and put on any pinned groups
2744 * in order to give them the best chance of going on.
2746 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2747 ctx_pinned_sched_in(ctx, cpuctx);
2749 /* Then walk through the lower prio flexible groups */
2750 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2751 ctx_flexible_sched_in(ctx, cpuctx);
2754 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2755 enum event_type_t event_type,
2756 struct task_struct *task)
2758 struct perf_event_context *ctx = &cpuctx->ctx;
2760 ctx_sched_in(ctx, cpuctx, event_type, task);
2763 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2764 struct task_struct *task)
2766 struct perf_cpu_context *cpuctx;
2768 cpuctx = __get_cpu_context(ctx);
2769 if (cpuctx->task_ctx == ctx)
2772 perf_ctx_lock(cpuctx, ctx);
2773 perf_pmu_disable(ctx->pmu);
2775 * We want to keep the following priority order:
2776 * cpu pinned (that don't need to move), task pinned,
2777 * cpu flexible, task flexible.
2779 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2782 cpuctx->task_ctx = ctx;
2784 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2786 perf_pmu_enable(ctx->pmu);
2787 perf_ctx_unlock(cpuctx, ctx);
2791 * Called from scheduler to add the events of the current task
2792 * with interrupts disabled.
2794 * We restore the event value and then enable it.
2796 * This does not protect us against NMI, but enable()
2797 * sets the enabled bit in the control field of event _before_
2798 * accessing the event control register. If a NMI hits, then it will
2799 * keep the event running.
2801 void __perf_event_task_sched_in(struct task_struct *prev,
2802 struct task_struct *task)
2804 struct perf_event_context *ctx;
2807 for_each_task_context_nr(ctxn) {
2808 ctx = task->perf_event_ctxp[ctxn];
2812 perf_event_context_sched_in(ctx, task);
2815 * if cgroup events exist on this CPU, then we need
2816 * to check if we have to switch in PMU state.
2817 * cgroup event are system-wide mode only
2819 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2820 perf_cgroup_sched_in(prev, task);
2822 if (__this_cpu_read(perf_sched_cb_usages))
2823 perf_pmu_sched_task(prev, task, true);
2826 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2828 u64 frequency = event->attr.sample_freq;
2829 u64 sec = NSEC_PER_SEC;
2830 u64 divisor, dividend;
2832 int count_fls, nsec_fls, frequency_fls, sec_fls;
2834 count_fls = fls64(count);
2835 nsec_fls = fls64(nsec);
2836 frequency_fls = fls64(frequency);
2840 * We got @count in @nsec, with a target of sample_freq HZ
2841 * the target period becomes:
2844 * period = -------------------
2845 * @nsec * sample_freq
2850 * Reduce accuracy by one bit such that @a and @b converge
2851 * to a similar magnitude.
2853 #define REDUCE_FLS(a, b) \
2855 if (a##_fls > b##_fls) { \
2865 * Reduce accuracy until either term fits in a u64, then proceed with
2866 * the other, so that finally we can do a u64/u64 division.
2868 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2869 REDUCE_FLS(nsec, frequency);
2870 REDUCE_FLS(sec, count);
2873 if (count_fls + sec_fls > 64) {
2874 divisor = nsec * frequency;
2876 while (count_fls + sec_fls > 64) {
2877 REDUCE_FLS(count, sec);
2881 dividend = count * sec;
2883 dividend = count * sec;
2885 while (nsec_fls + frequency_fls > 64) {
2886 REDUCE_FLS(nsec, frequency);
2890 divisor = nsec * frequency;
2896 return div64_u64(dividend, divisor);
2899 static DEFINE_PER_CPU(int, perf_throttled_count);
2900 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2902 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2904 struct hw_perf_event *hwc = &event->hw;
2905 s64 period, sample_period;
2908 period = perf_calculate_period(event, nsec, count);
2910 delta = (s64)(period - hwc->sample_period);
2911 delta = (delta + 7) / 8; /* low pass filter */
2913 sample_period = hwc->sample_period + delta;
2918 hwc->sample_period = sample_period;
2920 if (local64_read(&hwc->period_left) > 8*sample_period) {
2922 event->pmu->stop(event, PERF_EF_UPDATE);
2924 local64_set(&hwc->period_left, 0);
2927 event->pmu->start(event, PERF_EF_RELOAD);
2932 * combine freq adjustment with unthrottling to avoid two passes over the
2933 * events. At the same time, make sure, having freq events does not change
2934 * the rate of unthrottling as that would introduce bias.
2936 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2939 struct perf_event *event;
2940 struct hw_perf_event *hwc;
2941 u64 now, period = TICK_NSEC;
2945 * only need to iterate over all events iff:
2946 * - context have events in frequency mode (needs freq adjust)
2947 * - there are events to unthrottle on this cpu
2949 if (!(ctx->nr_freq || needs_unthr))
2952 raw_spin_lock(&ctx->lock);
2953 perf_pmu_disable(ctx->pmu);
2955 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2956 if (event->state != PERF_EVENT_STATE_ACTIVE)
2959 if (!event_filter_match(event))
2962 perf_pmu_disable(event->pmu);
2966 if (hwc->interrupts == MAX_INTERRUPTS) {
2967 hwc->interrupts = 0;
2968 perf_log_throttle(event, 1);
2969 event->pmu->start(event, 0);
2972 if (!event->attr.freq || !event->attr.sample_freq)
2976 * stop the event and update event->count
2978 event->pmu->stop(event, PERF_EF_UPDATE);
2980 now = local64_read(&event->count);
2981 delta = now - hwc->freq_count_stamp;
2982 hwc->freq_count_stamp = now;
2986 * reload only if value has changed
2987 * we have stopped the event so tell that
2988 * to perf_adjust_period() to avoid stopping it
2992 perf_adjust_period(event, period, delta, false);
2994 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2996 perf_pmu_enable(event->pmu);
2999 perf_pmu_enable(ctx->pmu);
3000 raw_spin_unlock(&ctx->lock);
3004 * Round-robin a context's events:
3006 static void rotate_ctx(struct perf_event_context *ctx)
3009 * Rotate the first entry last of non-pinned groups. Rotation might be
3010 * disabled by the inheritance code.
3012 if (!ctx->rotate_disable)
3013 list_rotate_left(&ctx->flexible_groups);
3016 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
3018 struct perf_event_context *ctx = NULL;
3021 if (cpuctx->ctx.nr_events) {
3022 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3026 ctx = cpuctx->task_ctx;
3027 if (ctx && ctx->nr_events) {
3028 if (ctx->nr_events != ctx->nr_active)
3035 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3036 perf_pmu_disable(cpuctx->ctx.pmu);
3038 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3040 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3042 rotate_ctx(&cpuctx->ctx);
3046 perf_event_sched_in(cpuctx, ctx, current);
3048 perf_pmu_enable(cpuctx->ctx.pmu);
3049 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3055 #ifdef CONFIG_NO_HZ_FULL
3056 bool perf_event_can_stop_tick(void)
3058 if (atomic_read(&nr_freq_events) ||
3059 __this_cpu_read(perf_throttled_count))
3066 void perf_event_task_tick(void)
3068 struct list_head *head = this_cpu_ptr(&active_ctx_list);
3069 struct perf_event_context *ctx, *tmp;
3072 WARN_ON(!irqs_disabled());
3074 __this_cpu_inc(perf_throttled_seq);
3075 throttled = __this_cpu_xchg(perf_throttled_count, 0);
3077 list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3078 perf_adjust_freq_unthr_context(ctx, throttled);
3081 static int event_enable_on_exec(struct perf_event *event,
3082 struct perf_event_context *ctx)
3084 if (!event->attr.enable_on_exec)
3087 event->attr.enable_on_exec = 0;
3088 if (event->state >= PERF_EVENT_STATE_INACTIVE)
3091 __perf_event_mark_enabled(event);
3097 * Enable all of a task's events that have been marked enable-on-exec.
3098 * This expects task == current.
3100 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
3102 struct perf_event_context *clone_ctx = NULL;
3103 struct perf_event *event;
3104 unsigned long flags;
3108 local_irq_save(flags);
3109 if (!ctx || !ctx->nr_events)
3113 * We must ctxsw out cgroup events to avoid conflict
3114 * when invoking perf_task_event_sched_in() later on
3115 * in this function. Otherwise we end up trying to
3116 * ctxswin cgroup events which are already scheduled
3119 perf_cgroup_sched_out(current, NULL);
3121 raw_spin_lock(&ctx->lock);
3122 task_ctx_sched_out(ctx);
3124 list_for_each_entry(event, &ctx->event_list, event_entry) {
3125 ret = event_enable_on_exec(event, ctx);
3131 * Unclone this context if we enabled any event.
3134 clone_ctx = unclone_ctx(ctx);
3136 raw_spin_unlock(&ctx->lock);
3139 * Also calls ctxswin for cgroup events, if any:
3141 perf_event_context_sched_in(ctx, ctx->task);
3143 local_irq_restore(flags);
3149 void perf_event_exec(void)
3151 struct perf_event_context *ctx;
3155 for_each_task_context_nr(ctxn) {
3156 ctx = current->perf_event_ctxp[ctxn];
3160 perf_event_enable_on_exec(ctx);
3166 * Cross CPU call to read the hardware event
3168 static void __perf_event_read(void *info)
3170 struct perf_event *event = info;
3171 struct perf_event_context *ctx = event->ctx;
3172 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3175 * If this is a task context, we need to check whether it is
3176 * the current task context of this cpu. If not it has been
3177 * scheduled out before the smp call arrived. In that case
3178 * event->count would have been updated to a recent sample
3179 * when the event was scheduled out.
3181 if (ctx->task && cpuctx->task_ctx != ctx)
3184 raw_spin_lock(&ctx->lock);
3185 if (ctx->is_active) {
3186 update_context_time(ctx);
3187 update_cgrp_time_from_event(event);
3189 update_event_times(event);
3190 if (event->state == PERF_EVENT_STATE_ACTIVE)
3191 event->pmu->read(event);
3192 raw_spin_unlock(&ctx->lock);
3195 static inline u64 perf_event_count(struct perf_event *event)
3197 return local64_read(&event->count) + atomic64_read(&event->child_count);
3200 static u64 perf_event_read(struct perf_event *event)
3203 * If event is enabled and currently active on a CPU, update the
3204 * value in the event structure:
3206 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3207 smp_call_function_single(event->oncpu,
3208 __perf_event_read, event, 1);
3209 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3210 struct perf_event_context *ctx = event->ctx;
3211 unsigned long flags;
3213 raw_spin_lock_irqsave(&ctx->lock, flags);
3215 * may read while context is not active
3216 * (e.g., thread is blocked), in that case
3217 * we cannot update context time
3219 if (ctx->is_active) {
3220 update_context_time(ctx);
3221 update_cgrp_time_from_event(event);
3223 update_event_times(event);
3224 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3227 return perf_event_count(event);
3231 * Initialize the perf_event context in a task_struct:
3233 static void __perf_event_init_context(struct perf_event_context *ctx)
3235 raw_spin_lock_init(&ctx->lock);
3236 mutex_init(&ctx->mutex);
3237 INIT_LIST_HEAD(&ctx->active_ctx_list);
3238 INIT_LIST_HEAD(&ctx->pinned_groups);
3239 INIT_LIST_HEAD(&ctx->flexible_groups);
3240 INIT_LIST_HEAD(&ctx->event_list);
3241 atomic_set(&ctx->refcount, 1);
3242 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3245 static struct perf_event_context *
3246 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3248 struct perf_event_context *ctx;
3250 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3254 __perf_event_init_context(ctx);
3257 get_task_struct(task);
3264 static struct task_struct *
3265 find_lively_task_by_vpid(pid_t vpid)
3267 struct task_struct *task;
3274 task = find_task_by_vpid(vpid);
3276 get_task_struct(task);
3280 return ERR_PTR(-ESRCH);
3282 /* Reuse ptrace permission checks for now. */
3284 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3289 put_task_struct(task);
3290 return ERR_PTR(err);
3295 * Returns a matching context with refcount and pincount.
3297 static struct perf_event_context *
3298 find_get_context(struct pmu *pmu, struct task_struct *task,
3299 struct perf_event *event)
3301 struct perf_event_context *ctx, *clone_ctx = NULL;
3302 struct perf_cpu_context *cpuctx;
3303 void *task_ctx_data = NULL;
3304 unsigned long flags;
3306 int cpu = event->cpu;
3309 /* Must be root to operate on a CPU event: */
3310 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3311 return ERR_PTR(-EACCES);
3314 * We could be clever and allow to attach a event to an
3315 * offline CPU and activate it when the CPU comes up, but
3318 if (!cpu_online(cpu))
3319 return ERR_PTR(-ENODEV);
3321 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3330 ctxn = pmu->task_ctx_nr;
3334 if (event->attach_state & PERF_ATTACH_TASK_DATA) {
3335 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
3336 if (!task_ctx_data) {
3343 ctx = perf_lock_task_context(task, ctxn, &flags);
3345 clone_ctx = unclone_ctx(ctx);
3348 if (task_ctx_data && !ctx->task_ctx_data) {
3349 ctx->task_ctx_data = task_ctx_data;
3350 task_ctx_data = NULL;
3352 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3357 ctx = alloc_perf_context(pmu, task);
3362 if (task_ctx_data) {
3363 ctx->task_ctx_data = task_ctx_data;
3364 task_ctx_data = NULL;
3368 mutex_lock(&task->perf_event_mutex);
3370 * If it has already passed perf_event_exit_task().
3371 * we must see PF_EXITING, it takes this mutex too.
3373 if (task->flags & PF_EXITING)
3375 else if (task->perf_event_ctxp[ctxn])
3380 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3382 mutex_unlock(&task->perf_event_mutex);
3384 if (unlikely(err)) {
3393 kfree(task_ctx_data);
3397 kfree(task_ctx_data);
3398 return ERR_PTR(err);
3401 static void perf_event_free_filter(struct perf_event *event);
3403 static void free_event_rcu(struct rcu_head *head)
3405 struct perf_event *event;
3407 event = container_of(head, struct perf_event, rcu_head);
3409 put_pid_ns(event->ns);
3410 perf_event_free_filter(event);
3414 static void ring_buffer_put(struct ring_buffer *rb);
3415 static void ring_buffer_attach(struct perf_event *event,
3416 struct ring_buffer *rb);
3418 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3423 if (is_cgroup_event(event))
3424 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3427 static void unaccount_event(struct perf_event *event)
3432 if (event->attach_state & PERF_ATTACH_TASK)
3433 static_key_slow_dec_deferred(&perf_sched_events);
3434 if (event->attr.mmap || event->attr.mmap_data)
3435 atomic_dec(&nr_mmap_events);
3436 if (event->attr.comm)
3437 atomic_dec(&nr_comm_events);
3438 if (event->attr.task)
3439 atomic_dec(&nr_task_events);
3440 if (event->attr.freq)
3441 atomic_dec(&nr_freq_events);
3442 if (is_cgroup_event(event))
3443 static_key_slow_dec_deferred(&perf_sched_events);
3444 if (has_branch_stack(event))
3445 static_key_slow_dec_deferred(&perf_sched_events);
3447 unaccount_event_cpu(event, event->cpu);
3450 static void __free_event(struct perf_event *event)
3452 if (!event->parent) {
3453 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3454 put_callchain_buffers();
3458 event->destroy(event);
3461 put_ctx(event->ctx);
3464 module_put(event->pmu->module);
3466 call_rcu(&event->rcu_head, free_event_rcu);
3469 static void _free_event(struct perf_event *event)
3471 irq_work_sync(&event->pending);
3473 unaccount_event(event);
3477 * Can happen when we close an event with re-directed output.
3479 * Since we have a 0 refcount, perf_mmap_close() will skip
3480 * over us; possibly making our ring_buffer_put() the last.
3482 mutex_lock(&event->mmap_mutex);
3483 ring_buffer_attach(event, NULL);
3484 mutex_unlock(&event->mmap_mutex);
3487 if (is_cgroup_event(event))
3488 perf_detach_cgroup(event);
3490 __free_event(event);
3494 * Used to free events which have a known refcount of 1, such as in error paths
3495 * where the event isn't exposed yet and inherited events.
3497 static void free_event(struct perf_event *event)
3499 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3500 "unexpected event refcount: %ld; ptr=%p\n",
3501 atomic_long_read(&event->refcount), event)) {
3502 /* leak to avoid use-after-free */
3510 * Remove user event from the owner task.
3512 static void perf_remove_from_owner(struct perf_event *event)
3514 struct task_struct *owner;
3517 owner = ACCESS_ONCE(event->owner);
3519 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3520 * !owner it means the list deletion is complete and we can indeed
3521 * free this event, otherwise we need to serialize on
3522 * owner->perf_event_mutex.
3524 smp_read_barrier_depends();
3527 * Since delayed_put_task_struct() also drops the last
3528 * task reference we can safely take a new reference
3529 * while holding the rcu_read_lock().
3531 get_task_struct(owner);
3537 * If we're here through perf_event_exit_task() we're already
3538 * holding ctx->mutex which would be an inversion wrt. the
3539 * normal lock order.
3541 * However we can safely take this lock because its the child
3544 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
3547 * We have to re-check the event->owner field, if it is cleared
3548 * we raced with perf_event_exit_task(), acquiring the mutex
3549 * ensured they're done, and we can proceed with freeing the
3553 list_del_init(&event->owner_entry);
3554 mutex_unlock(&owner->perf_event_mutex);
3555 put_task_struct(owner);
3560 * Called when the last reference to the file is gone.
3562 static void put_event(struct perf_event *event)
3564 struct perf_event_context *ctx;
3566 if (!atomic_long_dec_and_test(&event->refcount))
3569 if (!is_kernel_event(event))
3570 perf_remove_from_owner(event);
3573 * There are two ways this annotation is useful:
3575 * 1) there is a lock recursion from perf_event_exit_task
3576 * see the comment there.
3578 * 2) there is a lock-inversion with mmap_sem through
3579 * perf_event_read_group(), which takes faults while
3580 * holding ctx->mutex, however this is called after
3581 * the last filedesc died, so there is no possibility
3582 * to trigger the AB-BA case.
3584 ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING);
3585 WARN_ON_ONCE(ctx->parent_ctx);
3586 perf_remove_from_context(event, true);
3587 mutex_unlock(&ctx->mutex);
3592 int perf_event_release_kernel(struct perf_event *event)
3597 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3599 static int perf_release(struct inode *inode, struct file *file)
3601 put_event(file->private_data);
3606 * Remove all orphanes events from the context.
3608 static void orphans_remove_work(struct work_struct *work)
3610 struct perf_event_context *ctx;
3611 struct perf_event *event, *tmp;
3613 ctx = container_of(work, struct perf_event_context,
3614 orphans_remove.work);
3616 mutex_lock(&ctx->mutex);
3617 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3618 struct perf_event *parent_event = event->parent;
3620 if (!is_orphaned_child(event))
3623 perf_remove_from_context(event, true);
3625 mutex_lock(&parent_event->child_mutex);
3626 list_del_init(&event->child_list);
3627 mutex_unlock(&parent_event->child_mutex);
3630 put_event(parent_event);
3633 raw_spin_lock_irq(&ctx->lock);
3634 ctx->orphans_remove_sched = false;
3635 raw_spin_unlock_irq(&ctx->lock);
3636 mutex_unlock(&ctx->mutex);
3641 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3643 struct perf_event *child;
3649 mutex_lock(&event->child_mutex);
3650 total += perf_event_read(event);
3651 *enabled += event->total_time_enabled +
3652 atomic64_read(&event->child_total_time_enabled);
3653 *running += event->total_time_running +
3654 atomic64_read(&event->child_total_time_running);
3656 list_for_each_entry(child, &event->child_list, child_list) {
3657 total += perf_event_read(child);
3658 *enabled += child->total_time_enabled;
3659 *running += child->total_time_running;
3661 mutex_unlock(&event->child_mutex);
3665 EXPORT_SYMBOL_GPL(perf_event_read_value);
3667 static int perf_event_read_group(struct perf_event *event,
3668 u64 read_format, char __user *buf)
3670 struct perf_event *leader = event->group_leader, *sub;
3671 struct perf_event_context *ctx = leader->ctx;
3672 int n = 0, size = 0, ret;
3673 u64 count, enabled, running;
3676 lockdep_assert_held(&ctx->mutex);
3678 count = perf_event_read_value(leader, &enabled, &running);
3680 values[n++] = 1 + leader->nr_siblings;
3681 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3682 values[n++] = enabled;
3683 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3684 values[n++] = running;
3685 values[n++] = count;
3686 if (read_format & PERF_FORMAT_ID)
3687 values[n++] = primary_event_id(leader);
3689 size = n * sizeof(u64);
3691 if (copy_to_user(buf, values, size))
3696 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3699 values[n++] = perf_event_read_value(sub, &enabled, &running);
3700 if (read_format & PERF_FORMAT_ID)
3701 values[n++] = primary_event_id(sub);
3703 size = n * sizeof(u64);
3705 if (copy_to_user(buf + ret, values, size)) {
3715 static int perf_event_read_one(struct perf_event *event,
3716 u64 read_format, char __user *buf)
3718 u64 enabled, running;
3722 values[n++] = perf_event_read_value(event, &enabled, &running);
3723 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3724 values[n++] = enabled;
3725 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3726 values[n++] = running;
3727 if (read_format & PERF_FORMAT_ID)
3728 values[n++] = primary_event_id(event);
3730 if (copy_to_user(buf, values, n * sizeof(u64)))
3733 return n * sizeof(u64);
3736 static bool is_event_hup(struct perf_event *event)
3740 if (event->state != PERF_EVENT_STATE_EXIT)
3743 mutex_lock(&event->child_mutex);
3744 no_children = list_empty(&event->child_list);
3745 mutex_unlock(&event->child_mutex);
3750 * Read the performance event - simple non blocking version for now
3753 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3755 u64 read_format = event->attr.read_format;
3759 * Return end-of-file for a read on a event that is in
3760 * error state (i.e. because it was pinned but it couldn't be
3761 * scheduled on to the CPU at some point).
3763 if (event->state == PERF_EVENT_STATE_ERROR)
3766 if (count < event->read_size)
3769 WARN_ON_ONCE(event->ctx->parent_ctx);
3770 if (read_format & PERF_FORMAT_GROUP)
3771 ret = perf_event_read_group(event, read_format, buf);
3773 ret = perf_event_read_one(event, read_format, buf);
3779 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3781 struct perf_event *event = file->private_data;
3782 struct perf_event_context *ctx;
3785 ctx = perf_event_ctx_lock(event);
3786 ret = perf_read_hw(event, buf, count);
3787 perf_event_ctx_unlock(event, ctx);
3792 static unsigned int perf_poll(struct file *file, poll_table *wait)
3794 struct perf_event *event = file->private_data;
3795 struct ring_buffer *rb;
3796 unsigned int events = POLLHUP;
3798 poll_wait(file, &event->waitq, wait);
3800 if (is_event_hup(event))
3804 * Pin the event->rb by taking event->mmap_mutex; otherwise
3805 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3807 mutex_lock(&event->mmap_mutex);
3810 events = atomic_xchg(&rb->poll, 0);
3811 mutex_unlock(&event->mmap_mutex);
3815 static void _perf_event_reset(struct perf_event *event)
3817 (void)perf_event_read(event);
3818 local64_set(&event->count, 0);
3819 perf_event_update_userpage(event);
3823 * Holding the top-level event's child_mutex means that any
3824 * descendant process that has inherited this event will block
3825 * in sync_child_event if it goes to exit, thus satisfying the
3826 * task existence requirements of perf_event_enable/disable.
3828 static void perf_event_for_each_child(struct perf_event *event,
3829 void (*func)(struct perf_event *))
3831 struct perf_event *child;
3833 WARN_ON_ONCE(event->ctx->parent_ctx);
3835 mutex_lock(&event->child_mutex);
3837 list_for_each_entry(child, &event->child_list, child_list)
3839 mutex_unlock(&event->child_mutex);
3842 static void perf_event_for_each(struct perf_event *event,
3843 void (*func)(struct perf_event *))
3845 struct perf_event_context *ctx = event->ctx;
3846 struct perf_event *sibling;
3848 lockdep_assert_held(&ctx->mutex);
3850 event = event->group_leader;
3852 perf_event_for_each_child(event, func);
3853 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3854 perf_event_for_each_child(sibling, func);
3857 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3859 struct perf_event_context *ctx = event->ctx;
3860 int ret = 0, active;
3863 if (!is_sampling_event(event))
3866 if (copy_from_user(&value, arg, sizeof(value)))
3872 raw_spin_lock_irq(&ctx->lock);
3873 if (event->attr.freq) {
3874 if (value > sysctl_perf_event_sample_rate) {
3879 event->attr.sample_freq = value;
3881 event->attr.sample_period = value;
3882 event->hw.sample_period = value;
3885 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3887 perf_pmu_disable(ctx->pmu);
3888 event->pmu->stop(event, PERF_EF_UPDATE);
3891 local64_set(&event->hw.period_left, 0);
3894 event->pmu->start(event, PERF_EF_RELOAD);
3895 perf_pmu_enable(ctx->pmu);
3899 raw_spin_unlock_irq(&ctx->lock);
3904 static const struct file_operations perf_fops;
3906 static inline int perf_fget_light(int fd, struct fd *p)
3908 struct fd f = fdget(fd);
3912 if (f.file->f_op != &perf_fops) {
3920 static int perf_event_set_output(struct perf_event *event,
3921 struct perf_event *output_event);
3922 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3924 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
3926 void (*func)(struct perf_event *);
3930 case PERF_EVENT_IOC_ENABLE:
3931 func = _perf_event_enable;
3933 case PERF_EVENT_IOC_DISABLE:
3934 func = _perf_event_disable;
3936 case PERF_EVENT_IOC_RESET:
3937 func = _perf_event_reset;
3940 case PERF_EVENT_IOC_REFRESH:
3941 return _perf_event_refresh(event, arg);
3943 case PERF_EVENT_IOC_PERIOD:
3944 return perf_event_period(event, (u64 __user *)arg);
3946 case PERF_EVENT_IOC_ID:
3948 u64 id = primary_event_id(event);
3950 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3955 case PERF_EVENT_IOC_SET_OUTPUT:
3959 struct perf_event *output_event;
3961 ret = perf_fget_light(arg, &output);
3964 output_event = output.file->private_data;
3965 ret = perf_event_set_output(event, output_event);
3968 ret = perf_event_set_output(event, NULL);
3973 case PERF_EVENT_IOC_SET_FILTER:
3974 return perf_event_set_filter(event, (void __user *)arg);
3980 if (flags & PERF_IOC_FLAG_GROUP)
3981 perf_event_for_each(event, func);
3983 perf_event_for_each_child(event, func);
3988 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3990 struct perf_event *event = file->private_data;
3991 struct perf_event_context *ctx;
3994 ctx = perf_event_ctx_lock(event);
3995 ret = _perf_ioctl(event, cmd, arg);
3996 perf_event_ctx_unlock(event, ctx);
4001 #ifdef CONFIG_COMPAT
4002 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
4005 switch (_IOC_NR(cmd)) {
4006 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
4007 case _IOC_NR(PERF_EVENT_IOC_ID):
4008 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4009 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
4010 cmd &= ~IOCSIZE_MASK;
4011 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
4015 return perf_ioctl(file, cmd, arg);
4018 # define perf_compat_ioctl NULL
4021 int perf_event_task_enable(void)
4023 struct perf_event_context *ctx;
4024 struct perf_event *event;
4026 mutex_lock(¤t->perf_event_mutex);
4027 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4028 ctx = perf_event_ctx_lock(event);
4029 perf_event_for_each_child(event, _perf_event_enable);
4030 perf_event_ctx_unlock(event, ctx);
4032 mutex_unlock(¤t->perf_event_mutex);
4037 int perf_event_task_disable(void)
4039 struct perf_event_context *ctx;
4040 struct perf_event *event;
4042 mutex_lock(¤t->perf_event_mutex);
4043 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4044 ctx = perf_event_ctx_lock(event);
4045 perf_event_for_each_child(event, _perf_event_disable);
4046 perf_event_ctx_unlock(event, ctx);
4048 mutex_unlock(¤t->perf_event_mutex);
4053 static int perf_event_index(struct perf_event *event)
4055 if (event->hw.state & PERF_HES_STOPPED)
4058 if (event->state != PERF_EVENT_STATE_ACTIVE)
4061 return event->pmu->event_idx(event);
4064 static void calc_timer_values(struct perf_event *event,
4071 *now = perf_clock();
4072 ctx_time = event->shadow_ctx_time + *now;
4073 *enabled = ctx_time - event->tstamp_enabled;
4074 *running = ctx_time - event->tstamp_running;
4077 static void perf_event_init_userpage(struct perf_event *event)
4079 struct perf_event_mmap_page *userpg;
4080 struct ring_buffer *rb;
4083 rb = rcu_dereference(event->rb);
4087 userpg = rb->user_page;
4089 /* Allow new userspace to detect that bit 0 is deprecated */
4090 userpg->cap_bit0_is_deprecated = 1;
4091 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
4097 void __weak arch_perf_update_userpage(
4098 struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
4103 * Callers need to ensure there can be no nesting of this function, otherwise
4104 * the seqlock logic goes bad. We can not serialize this because the arch
4105 * code calls this from NMI context.
4107 void perf_event_update_userpage(struct perf_event *event)
4109 struct perf_event_mmap_page *userpg;
4110 struct ring_buffer *rb;
4111 u64 enabled, running, now;
4114 rb = rcu_dereference(event->rb);
4119 * compute total_time_enabled, total_time_running
4120 * based on snapshot values taken when the event
4121 * was last scheduled in.
4123 * we cannot simply called update_context_time()
4124 * because of locking issue as we can be called in
4127 calc_timer_values(event, &now, &enabled, &running);
4129 userpg = rb->user_page;
4131 * Disable preemption so as to not let the corresponding user-space
4132 * spin too long if we get preempted.
4137 userpg->index = perf_event_index(event);
4138 userpg->offset = perf_event_count(event);
4140 userpg->offset -= local64_read(&event->hw.prev_count);
4142 userpg->time_enabled = enabled +
4143 atomic64_read(&event->child_total_time_enabled);
4145 userpg->time_running = running +
4146 atomic64_read(&event->child_total_time_running);
4148 arch_perf_update_userpage(event, userpg, now);
4157 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
4159 struct perf_event *event = vma->vm_file->private_data;
4160 struct ring_buffer *rb;
4161 int ret = VM_FAULT_SIGBUS;
4163 if (vmf->flags & FAULT_FLAG_MKWRITE) {
4164 if (vmf->pgoff == 0)
4170 rb = rcu_dereference(event->rb);
4174 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4177 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4181 get_page(vmf->page);
4182 vmf->page->mapping = vma->vm_file->f_mapping;
4183 vmf->page->index = vmf->pgoff;
4192 static void ring_buffer_attach(struct perf_event *event,
4193 struct ring_buffer *rb)
4195 struct ring_buffer *old_rb = NULL;
4196 unsigned long flags;
4200 * Should be impossible, we set this when removing
4201 * event->rb_entry and wait/clear when adding event->rb_entry.
4203 WARN_ON_ONCE(event->rcu_pending);
4206 event->rcu_batches = get_state_synchronize_rcu();
4207 event->rcu_pending = 1;
4209 spin_lock_irqsave(&old_rb->event_lock, flags);
4210 list_del_rcu(&event->rb_entry);
4211 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4214 if (event->rcu_pending && rb) {
4215 cond_synchronize_rcu(event->rcu_batches);
4216 event->rcu_pending = 0;
4220 spin_lock_irqsave(&rb->event_lock, flags);
4221 list_add_rcu(&event->rb_entry, &rb->event_list);
4222 spin_unlock_irqrestore(&rb->event_lock, flags);
4225 rcu_assign_pointer(event->rb, rb);
4228 ring_buffer_put(old_rb);
4230 * Since we detached before setting the new rb, so that we
4231 * could attach the new rb, we could have missed a wakeup.
4234 wake_up_all(&event->waitq);
4238 static void ring_buffer_wakeup(struct perf_event *event)
4240 struct ring_buffer *rb;
4243 rb = rcu_dereference(event->rb);
4245 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4246 wake_up_all(&event->waitq);
4251 static void rb_free_rcu(struct rcu_head *rcu_head)
4253 struct ring_buffer *rb;
4255 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
4259 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
4261 struct ring_buffer *rb;
4264 rb = rcu_dereference(event->rb);
4266 if (!atomic_inc_not_zero(&rb->refcount))
4274 static void ring_buffer_put(struct ring_buffer *rb)
4276 if (!atomic_dec_and_test(&rb->refcount))
4279 WARN_ON_ONCE(!list_empty(&rb->event_list));
4281 call_rcu(&rb->rcu_head, rb_free_rcu);
4284 static void perf_mmap_open(struct vm_area_struct *vma)
4286 struct perf_event *event = vma->vm_file->private_data;
4288 atomic_inc(&event->mmap_count);
4289 atomic_inc(&event->rb->mmap_count);
4291 if (event->pmu->event_mapped)
4292 event->pmu->event_mapped(event);
4296 * A buffer can be mmap()ed multiple times; either directly through the same
4297 * event, or through other events by use of perf_event_set_output().
4299 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4300 * the buffer here, where we still have a VM context. This means we need
4301 * to detach all events redirecting to us.
4303 static void perf_mmap_close(struct vm_area_struct *vma)
4305 struct perf_event *event = vma->vm_file->private_data;
4307 struct ring_buffer *rb = ring_buffer_get(event);
4308 struct user_struct *mmap_user = rb->mmap_user;
4309 int mmap_locked = rb->mmap_locked;
4310 unsigned long size = perf_data_size(rb);
4312 if (event->pmu->event_unmapped)
4313 event->pmu->event_unmapped(event);
4315 atomic_dec(&rb->mmap_count);
4317 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4320 ring_buffer_attach(event, NULL);
4321 mutex_unlock(&event->mmap_mutex);
4323 /* If there's still other mmap()s of this buffer, we're done. */
4324 if (atomic_read(&rb->mmap_count))
4328 * No other mmap()s, detach from all other events that might redirect
4329 * into the now unreachable buffer. Somewhat complicated by the
4330 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4334 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4335 if (!atomic_long_inc_not_zero(&event->refcount)) {
4337 * This event is en-route to free_event() which will
4338 * detach it and remove it from the list.
4344 mutex_lock(&event->mmap_mutex);
4346 * Check we didn't race with perf_event_set_output() which can
4347 * swizzle the rb from under us while we were waiting to
4348 * acquire mmap_mutex.
4350 * If we find a different rb; ignore this event, a next
4351 * iteration will no longer find it on the list. We have to
4352 * still restart the iteration to make sure we're not now
4353 * iterating the wrong list.
4355 if (event->rb == rb)
4356 ring_buffer_attach(event, NULL);
4358 mutex_unlock(&event->mmap_mutex);
4362 * Restart the iteration; either we're on the wrong list or
4363 * destroyed its integrity by doing a deletion.
4370 * It could be there's still a few 0-ref events on the list; they'll
4371 * get cleaned up by free_event() -- they'll also still have their
4372 * ref on the rb and will free it whenever they are done with it.
4374 * Aside from that, this buffer is 'fully' detached and unmapped,
4375 * undo the VM accounting.
4378 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4379 vma->vm_mm->pinned_vm -= mmap_locked;
4380 free_uid(mmap_user);
4383 ring_buffer_put(rb); /* could be last */
4386 static const struct vm_operations_struct perf_mmap_vmops = {
4387 .open = perf_mmap_open,
4388 .close = perf_mmap_close,
4389 .fault = perf_mmap_fault,
4390 .page_mkwrite = perf_mmap_fault,
4393 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4395 struct perf_event *event = file->private_data;
4396 unsigned long user_locked, user_lock_limit;
4397 struct user_struct *user = current_user();
4398 unsigned long locked, lock_limit;
4399 struct ring_buffer *rb;
4400 unsigned long vma_size;
4401 unsigned long nr_pages;
4402 long user_extra, extra;
4403 int ret = 0, flags = 0;
4406 * Don't allow mmap() of inherited per-task counters. This would
4407 * create a performance issue due to all children writing to the
4410 if (event->cpu == -1 && event->attr.inherit)
4413 if (!(vma->vm_flags & VM_SHARED))
4416 vma_size = vma->vm_end - vma->vm_start;
4417 nr_pages = (vma_size / PAGE_SIZE) - 1;
4420 * If we have rb pages ensure they're a power-of-two number, so we
4421 * can do bitmasks instead of modulo.
4423 if (!is_power_of_2(nr_pages))
4426 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4429 if (vma->vm_pgoff != 0)
4432 WARN_ON_ONCE(event->ctx->parent_ctx);
4434 mutex_lock(&event->mmap_mutex);
4436 if (event->rb->nr_pages != nr_pages) {
4441 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4443 * Raced against perf_mmap_close() through
4444 * perf_event_set_output(). Try again, hope for better
4447 mutex_unlock(&event->mmap_mutex);
4454 user_extra = nr_pages + 1;
4455 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4458 * Increase the limit linearly with more CPUs:
4460 user_lock_limit *= num_online_cpus();
4462 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4465 if (user_locked > user_lock_limit)
4466 extra = user_locked - user_lock_limit;
4468 lock_limit = rlimit(RLIMIT_MEMLOCK);
4469 lock_limit >>= PAGE_SHIFT;
4470 locked = vma->vm_mm->pinned_vm + extra;
4472 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4473 !capable(CAP_IPC_LOCK)) {
4480 if (vma->vm_flags & VM_WRITE)
4481 flags |= RING_BUFFER_WRITABLE;
4483 rb = rb_alloc(nr_pages,
4484 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4492 atomic_set(&rb->mmap_count, 1);
4493 rb->mmap_locked = extra;
4494 rb->mmap_user = get_current_user();
4496 atomic_long_add(user_extra, &user->locked_vm);
4497 vma->vm_mm->pinned_vm += extra;
4499 ring_buffer_attach(event, rb);
4501 perf_event_init_userpage(event);
4502 perf_event_update_userpage(event);
4506 atomic_inc(&event->mmap_count);
4507 mutex_unlock(&event->mmap_mutex);
4510 * Since pinned accounting is per vm we cannot allow fork() to copy our
4513 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4514 vma->vm_ops = &perf_mmap_vmops;
4516 if (event->pmu->event_mapped)
4517 event->pmu->event_mapped(event);
4522 static int perf_fasync(int fd, struct file *filp, int on)
4524 struct inode *inode = file_inode(filp);
4525 struct perf_event *event = filp->private_data;
4528 mutex_lock(&inode->i_mutex);
4529 retval = fasync_helper(fd, filp, on, &event->fasync);
4530 mutex_unlock(&inode->i_mutex);
4538 static const struct file_operations perf_fops = {
4539 .llseek = no_llseek,
4540 .release = perf_release,
4543 .unlocked_ioctl = perf_ioctl,
4544 .compat_ioctl = perf_compat_ioctl,
4546 .fasync = perf_fasync,
4552 * If there's data, ensure we set the poll() state and publish everything
4553 * to user-space before waking everybody up.
4556 void perf_event_wakeup(struct perf_event *event)
4558 ring_buffer_wakeup(event);
4560 if (event->pending_kill) {
4561 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4562 event->pending_kill = 0;
4566 static void perf_pending_event(struct irq_work *entry)
4568 struct perf_event *event = container_of(entry,
4569 struct perf_event, pending);
4571 if (event->pending_disable) {
4572 event->pending_disable = 0;
4573 __perf_event_disable(event);
4576 if (event->pending_wakeup) {
4577 event->pending_wakeup = 0;
4578 perf_event_wakeup(event);
4583 * We assume there is only KVM supporting the callbacks.
4584 * Later on, we might change it to a list if there is
4585 * another virtualization implementation supporting the callbacks.
4587 struct perf_guest_info_callbacks *perf_guest_cbs;
4589 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4591 perf_guest_cbs = cbs;
4594 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4596 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4598 perf_guest_cbs = NULL;
4601 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4604 perf_output_sample_regs(struct perf_output_handle *handle,
4605 struct pt_regs *regs, u64 mask)
4609 for_each_set_bit(bit, (const unsigned long *) &mask,
4610 sizeof(mask) * BITS_PER_BYTE) {
4613 val = perf_reg_value(regs, bit);
4614 perf_output_put(handle, val);
4618 static void perf_sample_regs_user(struct perf_regs *regs_user,
4619 struct pt_regs *regs,
4620 struct pt_regs *regs_user_copy)
4622 if (user_mode(regs)) {
4623 regs_user->abi = perf_reg_abi(current);
4624 regs_user->regs = regs;
4625 } else if (current->mm) {
4626 perf_get_regs_user(regs_user, regs, regs_user_copy);
4628 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
4629 regs_user->regs = NULL;
4633 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
4634 struct pt_regs *regs)
4636 regs_intr->regs = regs;
4637 regs_intr->abi = perf_reg_abi(current);
4642 * Get remaining task size from user stack pointer.
4644 * It'd be better to take stack vma map and limit this more
4645 * precisly, but there's no way to get it safely under interrupt,
4646 * so using TASK_SIZE as limit.
4648 static u64 perf_ustack_task_size(struct pt_regs *regs)
4650 unsigned long addr = perf_user_stack_pointer(regs);
4652 if (!addr || addr >= TASK_SIZE)
4655 return TASK_SIZE - addr;
4659 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4660 struct pt_regs *regs)
4664 /* No regs, no stack pointer, no dump. */
4669 * Check if we fit in with the requested stack size into the:
4671 * If we don't, we limit the size to the TASK_SIZE.
4673 * - remaining sample size
4674 * If we don't, we customize the stack size to
4675 * fit in to the remaining sample size.
4678 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4679 stack_size = min(stack_size, (u16) task_size);
4681 /* Current header size plus static size and dynamic size. */
4682 header_size += 2 * sizeof(u64);
4684 /* Do we fit in with the current stack dump size? */
4685 if ((u16) (header_size + stack_size) < header_size) {
4687 * If we overflow the maximum size for the sample,
4688 * we customize the stack dump size to fit in.
4690 stack_size = USHRT_MAX - header_size - sizeof(u64);
4691 stack_size = round_up(stack_size, sizeof(u64));
4698 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4699 struct pt_regs *regs)
4701 /* Case of a kernel thread, nothing to dump */
4704 perf_output_put(handle, size);
4713 * - the size requested by user or the best one we can fit
4714 * in to the sample max size
4716 * - user stack dump data
4718 * - the actual dumped size
4722 perf_output_put(handle, dump_size);
4725 sp = perf_user_stack_pointer(regs);
4726 rem = __output_copy_user(handle, (void *) sp, dump_size);
4727 dyn_size = dump_size - rem;
4729 perf_output_skip(handle, rem);
4732 perf_output_put(handle, dyn_size);
4736 static void __perf_event_header__init_id(struct perf_event_header *header,
4737 struct perf_sample_data *data,
4738 struct perf_event *event)
4740 u64 sample_type = event->attr.sample_type;
4742 data->type = sample_type;
4743 header->size += event->id_header_size;
4745 if (sample_type & PERF_SAMPLE_TID) {
4746 /* namespace issues */
4747 data->tid_entry.pid = perf_event_pid(event, current);
4748 data->tid_entry.tid = perf_event_tid(event, current);
4751 if (sample_type & PERF_SAMPLE_TIME)
4752 data->time = perf_clock();
4754 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4755 data->id = primary_event_id(event);
4757 if (sample_type & PERF_SAMPLE_STREAM_ID)
4758 data->stream_id = event->id;
4760 if (sample_type & PERF_SAMPLE_CPU) {
4761 data->cpu_entry.cpu = raw_smp_processor_id();
4762 data->cpu_entry.reserved = 0;
4766 void perf_event_header__init_id(struct perf_event_header *header,
4767 struct perf_sample_data *data,
4768 struct perf_event *event)
4770 if (event->attr.sample_id_all)
4771 __perf_event_header__init_id(header, data, event);
4774 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4775 struct perf_sample_data *data)
4777 u64 sample_type = data->type;
4779 if (sample_type & PERF_SAMPLE_TID)
4780 perf_output_put(handle, data->tid_entry);
4782 if (sample_type & PERF_SAMPLE_TIME)
4783 perf_output_put(handle, data->time);
4785 if (sample_type & PERF_SAMPLE_ID)
4786 perf_output_put(handle, data->id);
4788 if (sample_type & PERF_SAMPLE_STREAM_ID)
4789 perf_output_put(handle, data->stream_id);
4791 if (sample_type & PERF_SAMPLE_CPU)
4792 perf_output_put(handle, data->cpu_entry);
4794 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4795 perf_output_put(handle, data->id);
4798 void perf_event__output_id_sample(struct perf_event *event,
4799 struct perf_output_handle *handle,
4800 struct perf_sample_data *sample)
4802 if (event->attr.sample_id_all)
4803 __perf_event__output_id_sample(handle, sample);
4806 static void perf_output_read_one(struct perf_output_handle *handle,
4807 struct perf_event *event,
4808 u64 enabled, u64 running)
4810 u64 read_format = event->attr.read_format;
4814 values[n++] = perf_event_count(event);
4815 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4816 values[n++] = enabled +
4817 atomic64_read(&event->child_total_time_enabled);
4819 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4820 values[n++] = running +
4821 atomic64_read(&event->child_total_time_running);
4823 if (read_format & PERF_FORMAT_ID)
4824 values[n++] = primary_event_id(event);
4826 __output_copy(handle, values, n * sizeof(u64));
4830 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4832 static void perf_output_read_group(struct perf_output_handle *handle,
4833 struct perf_event *event,
4834 u64 enabled, u64 running)
4836 struct perf_event *leader = event->group_leader, *sub;
4837 u64 read_format = event->attr.read_format;
4841 values[n++] = 1 + leader->nr_siblings;
4843 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4844 values[n++] = enabled;
4846 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4847 values[n++] = running;
4849 if (leader != event)
4850 leader->pmu->read(leader);
4852 values[n++] = perf_event_count(leader);
4853 if (read_format & PERF_FORMAT_ID)
4854 values[n++] = primary_event_id(leader);
4856 __output_copy(handle, values, n * sizeof(u64));
4858 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4861 if ((sub != event) &&
4862 (sub->state == PERF_EVENT_STATE_ACTIVE))
4863 sub->pmu->read(sub);
4865 values[n++] = perf_event_count(sub);
4866 if (read_format & PERF_FORMAT_ID)
4867 values[n++] = primary_event_id(sub);
4869 __output_copy(handle, values, n * sizeof(u64));
4873 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4874 PERF_FORMAT_TOTAL_TIME_RUNNING)
4876 static void perf_output_read(struct perf_output_handle *handle,
4877 struct perf_event *event)
4879 u64 enabled = 0, running = 0, now;
4880 u64 read_format = event->attr.read_format;
4883 * compute total_time_enabled, total_time_running
4884 * based on snapshot values taken when the event
4885 * was last scheduled in.
4887 * we cannot simply called update_context_time()
4888 * because of locking issue as we are called in
4891 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4892 calc_timer_values(event, &now, &enabled, &running);
4894 if (event->attr.read_format & PERF_FORMAT_GROUP)
4895 perf_output_read_group(handle, event, enabled, running);
4897 perf_output_read_one(handle, event, enabled, running);
4900 void perf_output_sample(struct perf_output_handle *handle,
4901 struct perf_event_header *header,
4902 struct perf_sample_data *data,
4903 struct perf_event *event)
4905 u64 sample_type = data->type;
4907 perf_output_put(handle, *header);
4909 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4910 perf_output_put(handle, data->id);
4912 if (sample_type & PERF_SAMPLE_IP)
4913 perf_output_put(handle, data->ip);
4915 if (sample_type & PERF_SAMPLE_TID)
4916 perf_output_put(handle, data->tid_entry);
4918 if (sample_type & PERF_SAMPLE_TIME)
4919 perf_output_put(handle, data->time);
4921 if (sample_type & PERF_SAMPLE_ADDR)
4922 perf_output_put(handle, data->addr);
4924 if (sample_type & PERF_SAMPLE_ID)
4925 perf_output_put(handle, data->id);
4927 if (sample_type & PERF_SAMPLE_STREAM_ID)
4928 perf_output_put(handle, data->stream_id);
4930 if (sample_type & PERF_SAMPLE_CPU)
4931 perf_output_put(handle, data->cpu_entry);
4933 if (sample_type & PERF_SAMPLE_PERIOD)
4934 perf_output_put(handle, data->period);
4936 if (sample_type & PERF_SAMPLE_READ)
4937 perf_output_read(handle, event);
4939 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4940 if (data->callchain) {
4943 if (data->callchain)
4944 size += data->callchain->nr;
4946 size *= sizeof(u64);
4948 __output_copy(handle, data->callchain, size);
4951 perf_output_put(handle, nr);
4955 if (sample_type & PERF_SAMPLE_RAW) {
4957 perf_output_put(handle, data->raw->size);
4958 __output_copy(handle, data->raw->data,
4965 .size = sizeof(u32),
4968 perf_output_put(handle, raw);
4972 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4973 if (data->br_stack) {
4976 size = data->br_stack->nr
4977 * sizeof(struct perf_branch_entry);
4979 perf_output_put(handle, data->br_stack->nr);
4980 perf_output_copy(handle, data->br_stack->entries, size);
4983 * we always store at least the value of nr
4986 perf_output_put(handle, nr);
4990 if (sample_type & PERF_SAMPLE_REGS_USER) {
4991 u64 abi = data->regs_user.abi;
4994 * If there are no regs to dump, notice it through
4995 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4997 perf_output_put(handle, abi);
5000 u64 mask = event->attr.sample_regs_user;
5001 perf_output_sample_regs(handle,
5002 data->regs_user.regs,
5007 if (sample_type & PERF_SAMPLE_STACK_USER) {
5008 perf_output_sample_ustack(handle,
5009 data->stack_user_size,
5010 data->regs_user.regs);
5013 if (sample_type & PERF_SAMPLE_WEIGHT)
5014 perf_output_put(handle, data->weight);
5016 if (sample_type & PERF_SAMPLE_DATA_SRC)
5017 perf_output_put(handle, data->data_src.val);
5019 if (sample_type & PERF_SAMPLE_TRANSACTION)
5020 perf_output_put(handle, data->txn);
5022 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5023 u64 abi = data->regs_intr.abi;
5025 * If there are no regs to dump, notice it through
5026 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5028 perf_output_put(handle, abi);
5031 u64 mask = event->attr.sample_regs_intr;
5033 perf_output_sample_regs(handle,
5034 data->regs_intr.regs,
5039 if (!event->attr.watermark) {
5040 int wakeup_events = event->attr.wakeup_events;
5042 if (wakeup_events) {
5043 struct ring_buffer *rb = handle->rb;
5044 int events = local_inc_return(&rb->events);
5046 if (events >= wakeup_events) {
5047 local_sub(wakeup_events, &rb->events);
5048 local_inc(&rb->wakeup);
5054 void perf_prepare_sample(struct perf_event_header *header,
5055 struct perf_sample_data *data,
5056 struct perf_event *event,
5057 struct pt_regs *regs)
5059 u64 sample_type = event->attr.sample_type;
5061 header->type = PERF_RECORD_SAMPLE;
5062 header->size = sizeof(*header) + event->header_size;
5065 header->misc |= perf_misc_flags(regs);
5067 __perf_event_header__init_id(header, data, event);
5069 if (sample_type & PERF_SAMPLE_IP)
5070 data->ip = perf_instruction_pointer(regs);
5072 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5075 data->callchain = perf_callchain(event, regs);
5077 if (data->callchain)
5078 size += data->callchain->nr;
5080 header->size += size * sizeof(u64);
5083 if (sample_type & PERF_SAMPLE_RAW) {
5084 int size = sizeof(u32);
5087 size += data->raw->size;
5089 size += sizeof(u32);
5091 WARN_ON_ONCE(size & (sizeof(u64)-1));
5092 header->size += size;
5095 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5096 int size = sizeof(u64); /* nr */
5097 if (data->br_stack) {
5098 size += data->br_stack->nr
5099 * sizeof(struct perf_branch_entry);
5101 header->size += size;
5104 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
5105 perf_sample_regs_user(&data->regs_user, regs,
5106 &data->regs_user_copy);
5108 if (sample_type & PERF_SAMPLE_REGS_USER) {
5109 /* regs dump ABI info */
5110 int size = sizeof(u64);
5112 if (data->regs_user.regs) {
5113 u64 mask = event->attr.sample_regs_user;
5114 size += hweight64(mask) * sizeof(u64);
5117 header->size += size;
5120 if (sample_type & PERF_SAMPLE_STACK_USER) {
5122 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5123 * processed as the last one or have additional check added
5124 * in case new sample type is added, because we could eat
5125 * up the rest of the sample size.
5127 u16 stack_size = event->attr.sample_stack_user;
5128 u16 size = sizeof(u64);
5130 stack_size = perf_sample_ustack_size(stack_size, header->size,
5131 data->regs_user.regs);
5134 * If there is something to dump, add space for the dump
5135 * itself and for the field that tells the dynamic size,
5136 * which is how many have been actually dumped.
5139 size += sizeof(u64) + stack_size;
5141 data->stack_user_size = stack_size;
5142 header->size += size;
5145 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5146 /* regs dump ABI info */
5147 int size = sizeof(u64);
5149 perf_sample_regs_intr(&data->regs_intr, regs);
5151 if (data->regs_intr.regs) {
5152 u64 mask = event->attr.sample_regs_intr;
5154 size += hweight64(mask) * sizeof(u64);
5157 header->size += size;
5161 static void perf_event_output(struct perf_event *event,
5162 struct perf_sample_data *data,
5163 struct pt_regs *regs)
5165 struct perf_output_handle handle;
5166 struct perf_event_header header;
5168 /* protect the callchain buffers */
5171 perf_prepare_sample(&header, data, event, regs);
5173 if (perf_output_begin(&handle, event, header.size))
5176 perf_output_sample(&handle, &header, data, event);
5178 perf_output_end(&handle);
5188 struct perf_read_event {
5189 struct perf_event_header header;
5196 perf_event_read_event(struct perf_event *event,
5197 struct task_struct *task)
5199 struct perf_output_handle handle;
5200 struct perf_sample_data sample;
5201 struct perf_read_event read_event = {
5203 .type = PERF_RECORD_READ,
5205 .size = sizeof(read_event) + event->read_size,
5207 .pid = perf_event_pid(event, task),
5208 .tid = perf_event_tid(event, task),
5212 perf_event_header__init_id(&read_event.header, &sample, event);
5213 ret = perf_output_begin(&handle, event, read_event.header.size);
5217 perf_output_put(&handle, read_event);
5218 perf_output_read(&handle, event);
5219 perf_event__output_id_sample(event, &handle, &sample);
5221 perf_output_end(&handle);
5224 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5227 perf_event_aux_ctx(struct perf_event_context *ctx,
5228 perf_event_aux_output_cb output,
5231 struct perf_event *event;
5233 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5234 if (event->state < PERF_EVENT_STATE_INACTIVE)
5236 if (!event_filter_match(event))
5238 output(event, data);
5243 perf_event_aux(perf_event_aux_output_cb output, void *data,
5244 struct perf_event_context *task_ctx)
5246 struct perf_cpu_context *cpuctx;
5247 struct perf_event_context *ctx;
5252 list_for_each_entry_rcu(pmu, &pmus, entry) {
5253 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5254 if (cpuctx->unique_pmu != pmu)
5256 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5259 ctxn = pmu->task_ctx_nr;
5262 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5264 perf_event_aux_ctx(ctx, output, data);
5266 put_cpu_ptr(pmu->pmu_cpu_context);
5271 perf_event_aux_ctx(task_ctx, output, data);
5278 * task tracking -- fork/exit
5280 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5283 struct perf_task_event {
5284 struct task_struct *task;
5285 struct perf_event_context *task_ctx;
5288 struct perf_event_header header;
5298 static int perf_event_task_match(struct perf_event *event)
5300 return event->attr.comm || event->attr.mmap ||
5301 event->attr.mmap2 || event->attr.mmap_data ||
5305 static void perf_event_task_output(struct perf_event *event,
5308 struct perf_task_event *task_event = data;
5309 struct perf_output_handle handle;
5310 struct perf_sample_data sample;
5311 struct task_struct *task = task_event->task;
5312 int ret, size = task_event->event_id.header.size;
5314 if (!perf_event_task_match(event))
5317 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5319 ret = perf_output_begin(&handle, event,
5320 task_event->event_id.header.size);
5324 task_event->event_id.pid = perf_event_pid(event, task);
5325 task_event->event_id.ppid = perf_event_pid(event, current);
5327 task_event->event_id.tid = perf_event_tid(event, task);
5328 task_event->event_id.ptid = perf_event_tid(event, current);
5330 perf_output_put(&handle, task_event->event_id);
5332 perf_event__output_id_sample(event, &handle, &sample);
5334 perf_output_end(&handle);
5336 task_event->event_id.header.size = size;
5339 static void perf_event_task(struct task_struct *task,
5340 struct perf_event_context *task_ctx,
5343 struct perf_task_event task_event;
5345 if (!atomic_read(&nr_comm_events) &&
5346 !atomic_read(&nr_mmap_events) &&
5347 !atomic_read(&nr_task_events))
5350 task_event = (struct perf_task_event){
5352 .task_ctx = task_ctx,
5355 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5357 .size = sizeof(task_event.event_id),
5363 .time = perf_clock(),
5367 perf_event_aux(perf_event_task_output,
5372 void perf_event_fork(struct task_struct *task)
5374 perf_event_task(task, NULL, 1);
5381 struct perf_comm_event {
5382 struct task_struct *task;
5387 struct perf_event_header header;
5394 static int perf_event_comm_match(struct perf_event *event)
5396 return event->attr.comm;
5399 static void perf_event_comm_output(struct perf_event *event,
5402 struct perf_comm_event *comm_event = data;
5403 struct perf_output_handle handle;
5404 struct perf_sample_data sample;
5405 int size = comm_event->event_id.header.size;
5408 if (!perf_event_comm_match(event))
5411 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5412 ret = perf_output_begin(&handle, event,
5413 comm_event->event_id.header.size);
5418 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5419 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5421 perf_output_put(&handle, comm_event->event_id);
5422 __output_copy(&handle, comm_event->comm,
5423 comm_event->comm_size);
5425 perf_event__output_id_sample(event, &handle, &sample);
5427 perf_output_end(&handle);
5429 comm_event->event_id.header.size = size;
5432 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5434 char comm[TASK_COMM_LEN];
5437 memset(comm, 0, sizeof(comm));
5438 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5439 size = ALIGN(strlen(comm)+1, sizeof(u64));
5441 comm_event->comm = comm;
5442 comm_event->comm_size = size;
5444 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5446 perf_event_aux(perf_event_comm_output,
5451 void perf_event_comm(struct task_struct *task, bool exec)
5453 struct perf_comm_event comm_event;
5455 if (!atomic_read(&nr_comm_events))
5458 comm_event = (struct perf_comm_event){
5464 .type = PERF_RECORD_COMM,
5465 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5473 perf_event_comm_event(&comm_event);
5480 struct perf_mmap_event {
5481 struct vm_area_struct *vma;
5483 const char *file_name;
5491 struct perf_event_header header;
5501 static int perf_event_mmap_match(struct perf_event *event,
5504 struct perf_mmap_event *mmap_event = data;
5505 struct vm_area_struct *vma = mmap_event->vma;
5506 int executable = vma->vm_flags & VM_EXEC;
5508 return (!executable && event->attr.mmap_data) ||
5509 (executable && (event->attr.mmap || event->attr.mmap2));
5512 static void perf_event_mmap_output(struct perf_event *event,
5515 struct perf_mmap_event *mmap_event = data;
5516 struct perf_output_handle handle;
5517 struct perf_sample_data sample;
5518 int size = mmap_event->event_id.header.size;
5521 if (!perf_event_mmap_match(event, data))
5524 if (event->attr.mmap2) {
5525 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5526 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5527 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5528 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5529 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5530 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5531 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5534 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5535 ret = perf_output_begin(&handle, event,
5536 mmap_event->event_id.header.size);
5540 mmap_event->event_id.pid = perf_event_pid(event, current);
5541 mmap_event->event_id.tid = perf_event_tid(event, current);
5543 perf_output_put(&handle, mmap_event->event_id);
5545 if (event->attr.mmap2) {
5546 perf_output_put(&handle, mmap_event->maj);
5547 perf_output_put(&handle, mmap_event->min);
5548 perf_output_put(&handle, mmap_event->ino);
5549 perf_output_put(&handle, mmap_event->ino_generation);
5550 perf_output_put(&handle, mmap_event->prot);
5551 perf_output_put(&handle, mmap_event->flags);
5554 __output_copy(&handle, mmap_event->file_name,
5555 mmap_event->file_size);
5557 perf_event__output_id_sample(event, &handle, &sample);
5559 perf_output_end(&handle);
5561 mmap_event->event_id.header.size = size;
5564 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5566 struct vm_area_struct *vma = mmap_event->vma;
5567 struct file *file = vma->vm_file;
5568 int maj = 0, min = 0;
5569 u64 ino = 0, gen = 0;
5570 u32 prot = 0, flags = 0;
5577 struct inode *inode;
5580 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5586 * d_path() works from the end of the rb backwards, so we
5587 * need to add enough zero bytes after the string to handle
5588 * the 64bit alignment we do later.
5590 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5595 inode = file_inode(vma->vm_file);
5596 dev = inode->i_sb->s_dev;
5598 gen = inode->i_generation;
5602 if (vma->vm_flags & VM_READ)
5604 if (vma->vm_flags & VM_WRITE)
5606 if (vma->vm_flags & VM_EXEC)
5609 if (vma->vm_flags & VM_MAYSHARE)
5612 flags = MAP_PRIVATE;
5614 if (vma->vm_flags & VM_DENYWRITE)
5615 flags |= MAP_DENYWRITE;
5616 if (vma->vm_flags & VM_MAYEXEC)
5617 flags |= MAP_EXECUTABLE;
5618 if (vma->vm_flags & VM_LOCKED)
5619 flags |= MAP_LOCKED;
5620 if (vma->vm_flags & VM_HUGETLB)
5621 flags |= MAP_HUGETLB;
5625 if (vma->vm_ops && vma->vm_ops->name) {
5626 name = (char *) vma->vm_ops->name(vma);
5631 name = (char *)arch_vma_name(vma);
5635 if (vma->vm_start <= vma->vm_mm->start_brk &&
5636 vma->vm_end >= vma->vm_mm->brk) {
5640 if (vma->vm_start <= vma->vm_mm->start_stack &&
5641 vma->vm_end >= vma->vm_mm->start_stack) {
5651 strlcpy(tmp, name, sizeof(tmp));
5655 * Since our buffer works in 8 byte units we need to align our string
5656 * size to a multiple of 8. However, we must guarantee the tail end is
5657 * zero'd out to avoid leaking random bits to userspace.
5659 size = strlen(name)+1;
5660 while (!IS_ALIGNED(size, sizeof(u64)))
5661 name[size++] = '\0';
5663 mmap_event->file_name = name;
5664 mmap_event->file_size = size;
5665 mmap_event->maj = maj;
5666 mmap_event->min = min;
5667 mmap_event->ino = ino;
5668 mmap_event->ino_generation = gen;
5669 mmap_event->prot = prot;
5670 mmap_event->flags = flags;
5672 if (!(vma->vm_flags & VM_EXEC))
5673 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5675 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5677 perf_event_aux(perf_event_mmap_output,
5684 void perf_event_mmap(struct vm_area_struct *vma)
5686 struct perf_mmap_event mmap_event;
5688 if (!atomic_read(&nr_mmap_events))
5691 mmap_event = (struct perf_mmap_event){
5697 .type = PERF_RECORD_MMAP,
5698 .misc = PERF_RECORD_MISC_USER,
5703 .start = vma->vm_start,
5704 .len = vma->vm_end - vma->vm_start,
5705 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5707 /* .maj (attr_mmap2 only) */
5708 /* .min (attr_mmap2 only) */
5709 /* .ino (attr_mmap2 only) */
5710 /* .ino_generation (attr_mmap2 only) */
5711 /* .prot (attr_mmap2 only) */
5712 /* .flags (attr_mmap2 only) */
5715 perf_event_mmap_event(&mmap_event);
5719 * IRQ throttle logging
5722 static void perf_log_throttle(struct perf_event *event, int enable)
5724 struct perf_output_handle handle;
5725 struct perf_sample_data sample;
5729 struct perf_event_header header;
5733 } throttle_event = {
5735 .type = PERF_RECORD_THROTTLE,
5737 .size = sizeof(throttle_event),
5739 .time = perf_clock(),
5740 .id = primary_event_id(event),
5741 .stream_id = event->id,
5745 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5747 perf_event_header__init_id(&throttle_event.header, &sample, event);
5749 ret = perf_output_begin(&handle, event,
5750 throttle_event.header.size);
5754 perf_output_put(&handle, throttle_event);
5755 perf_event__output_id_sample(event, &handle, &sample);
5756 perf_output_end(&handle);
5760 * Generic event overflow handling, sampling.
5763 static int __perf_event_overflow(struct perf_event *event,
5764 int throttle, struct perf_sample_data *data,
5765 struct pt_regs *regs)
5767 int events = atomic_read(&event->event_limit);
5768 struct hw_perf_event *hwc = &event->hw;
5773 * Non-sampling counters might still use the PMI to fold short
5774 * hardware counters, ignore those.
5776 if (unlikely(!is_sampling_event(event)))
5779 seq = __this_cpu_read(perf_throttled_seq);
5780 if (seq != hwc->interrupts_seq) {
5781 hwc->interrupts_seq = seq;
5782 hwc->interrupts = 1;
5785 if (unlikely(throttle
5786 && hwc->interrupts >= max_samples_per_tick)) {
5787 __this_cpu_inc(perf_throttled_count);
5788 hwc->interrupts = MAX_INTERRUPTS;
5789 perf_log_throttle(event, 0);
5790 tick_nohz_full_kick();
5795 if (event->attr.freq) {
5796 u64 now = perf_clock();
5797 s64 delta = now - hwc->freq_time_stamp;
5799 hwc->freq_time_stamp = now;
5801 if (delta > 0 && delta < 2*TICK_NSEC)
5802 perf_adjust_period(event, delta, hwc->last_period, true);
5806 * XXX event_limit might not quite work as expected on inherited
5810 event->pending_kill = POLL_IN;
5811 if (events && atomic_dec_and_test(&event->event_limit)) {
5813 event->pending_kill = POLL_HUP;
5814 event->pending_disable = 1;
5815 irq_work_queue(&event->pending);
5818 if (event->overflow_handler)
5819 event->overflow_handler(event, data, regs);
5821 perf_event_output(event, data, regs);
5823 if (event->fasync && event->pending_kill) {
5824 event->pending_wakeup = 1;
5825 irq_work_queue(&event->pending);
5831 int perf_event_overflow(struct perf_event *event,
5832 struct perf_sample_data *data,
5833 struct pt_regs *regs)
5835 return __perf_event_overflow(event, 1, data, regs);
5839 * Generic software event infrastructure
5842 struct swevent_htable {
5843 struct swevent_hlist *swevent_hlist;
5844 struct mutex hlist_mutex;
5847 /* Recursion avoidance in each contexts */
5848 int recursion[PERF_NR_CONTEXTS];
5850 /* Keeps track of cpu being initialized/exited */
5854 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5857 * We directly increment event->count and keep a second value in
5858 * event->hw.period_left to count intervals. This period event
5859 * is kept in the range [-sample_period, 0] so that we can use the
5863 u64 perf_swevent_set_period(struct perf_event *event)
5865 struct hw_perf_event *hwc = &event->hw;
5866 u64 period = hwc->last_period;
5870 hwc->last_period = hwc->sample_period;
5873 old = val = local64_read(&hwc->period_left);
5877 nr = div64_u64(period + val, period);
5878 offset = nr * period;
5880 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5886 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5887 struct perf_sample_data *data,
5888 struct pt_regs *regs)
5890 struct hw_perf_event *hwc = &event->hw;
5894 overflow = perf_swevent_set_period(event);
5896 if (hwc->interrupts == MAX_INTERRUPTS)
5899 for (; overflow; overflow--) {
5900 if (__perf_event_overflow(event, throttle,
5903 * We inhibit the overflow from happening when
5904 * hwc->interrupts == MAX_INTERRUPTS.
5912 static void perf_swevent_event(struct perf_event *event, u64 nr,
5913 struct perf_sample_data *data,
5914 struct pt_regs *regs)
5916 struct hw_perf_event *hwc = &event->hw;
5918 local64_add(nr, &event->count);
5923 if (!is_sampling_event(event))
5926 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5928 return perf_swevent_overflow(event, 1, data, regs);
5930 data->period = event->hw.last_period;
5932 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5933 return perf_swevent_overflow(event, 1, data, regs);
5935 if (local64_add_negative(nr, &hwc->period_left))
5938 perf_swevent_overflow(event, 0, data, regs);
5941 static int perf_exclude_event(struct perf_event *event,
5942 struct pt_regs *regs)
5944 if (event->hw.state & PERF_HES_STOPPED)
5948 if (event->attr.exclude_user && user_mode(regs))
5951 if (event->attr.exclude_kernel && !user_mode(regs))
5958 static int perf_swevent_match(struct perf_event *event,
5959 enum perf_type_id type,
5961 struct perf_sample_data *data,
5962 struct pt_regs *regs)
5964 if (event->attr.type != type)
5967 if (event->attr.config != event_id)
5970 if (perf_exclude_event(event, regs))
5976 static inline u64 swevent_hash(u64 type, u32 event_id)
5978 u64 val = event_id | (type << 32);
5980 return hash_64(val, SWEVENT_HLIST_BITS);
5983 static inline struct hlist_head *
5984 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5986 u64 hash = swevent_hash(type, event_id);
5988 return &hlist->heads[hash];
5991 /* For the read side: events when they trigger */
5992 static inline struct hlist_head *
5993 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5995 struct swevent_hlist *hlist;
5997 hlist = rcu_dereference(swhash->swevent_hlist);
6001 return __find_swevent_head(hlist, type, event_id);
6004 /* For the event head insertion and removal in the hlist */
6005 static inline struct hlist_head *
6006 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
6008 struct swevent_hlist *hlist;
6009 u32 event_id = event->attr.config;
6010 u64 type = event->attr.type;
6013 * Event scheduling is always serialized against hlist allocation
6014 * and release. Which makes the protected version suitable here.
6015 * The context lock guarantees that.
6017 hlist = rcu_dereference_protected(swhash->swevent_hlist,
6018 lockdep_is_held(&event->ctx->lock));
6022 return __find_swevent_head(hlist, type, event_id);
6025 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
6027 struct perf_sample_data *data,
6028 struct pt_regs *regs)
6030 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6031 struct perf_event *event;
6032 struct hlist_head *head;
6035 head = find_swevent_head_rcu(swhash, type, event_id);
6039 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6040 if (perf_swevent_match(event, type, event_id, data, regs))
6041 perf_swevent_event(event, nr, data, regs);
6047 DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]);
6049 int perf_swevent_get_recursion_context(void)
6051 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6053 return get_recursion_context(swhash->recursion);
6055 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
6057 inline void perf_swevent_put_recursion_context(int rctx)
6059 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6061 put_recursion_context(swhash->recursion, rctx);
6064 void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6066 struct perf_sample_data data;
6068 if (WARN_ON_ONCE(!regs))
6071 perf_sample_data_init(&data, addr, 0);
6072 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
6075 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6079 preempt_disable_notrace();
6080 rctx = perf_swevent_get_recursion_context();
6081 if (unlikely(rctx < 0))
6084 ___perf_sw_event(event_id, nr, regs, addr);
6086 perf_swevent_put_recursion_context(rctx);
6088 preempt_enable_notrace();
6091 static void perf_swevent_read(struct perf_event *event)
6095 static int perf_swevent_add(struct perf_event *event, int flags)
6097 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6098 struct hw_perf_event *hwc = &event->hw;
6099 struct hlist_head *head;
6101 if (is_sampling_event(event)) {
6102 hwc->last_period = hwc->sample_period;
6103 perf_swevent_set_period(event);
6106 hwc->state = !(flags & PERF_EF_START);
6108 head = find_swevent_head(swhash, event);
6111 * We can race with cpu hotplug code. Do not
6112 * WARN if the cpu just got unplugged.
6114 WARN_ON_ONCE(swhash->online);
6118 hlist_add_head_rcu(&event->hlist_entry, head);
6119 perf_event_update_userpage(event);
6124 static void perf_swevent_del(struct perf_event *event, int flags)
6126 hlist_del_rcu(&event->hlist_entry);
6129 static void perf_swevent_start(struct perf_event *event, int flags)
6131 event->hw.state = 0;
6134 static void perf_swevent_stop(struct perf_event *event, int flags)
6136 event->hw.state = PERF_HES_STOPPED;
6139 /* Deref the hlist from the update side */
6140 static inline struct swevent_hlist *
6141 swevent_hlist_deref(struct swevent_htable *swhash)
6143 return rcu_dereference_protected(swhash->swevent_hlist,
6144 lockdep_is_held(&swhash->hlist_mutex));
6147 static void swevent_hlist_release(struct swevent_htable *swhash)
6149 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
6154 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
6155 kfree_rcu(hlist, rcu_head);
6158 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
6160 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6162 mutex_lock(&swhash->hlist_mutex);
6164 if (!--swhash->hlist_refcount)
6165 swevent_hlist_release(swhash);
6167 mutex_unlock(&swhash->hlist_mutex);
6170 static void swevent_hlist_put(struct perf_event *event)
6174 for_each_possible_cpu(cpu)
6175 swevent_hlist_put_cpu(event, cpu);
6178 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
6180 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6183 mutex_lock(&swhash->hlist_mutex);
6185 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6186 struct swevent_hlist *hlist;
6188 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
6193 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6195 swhash->hlist_refcount++;
6197 mutex_unlock(&swhash->hlist_mutex);
6202 static int swevent_hlist_get(struct perf_event *event)
6205 int cpu, failed_cpu;
6208 for_each_possible_cpu(cpu) {
6209 err = swevent_hlist_get_cpu(event, cpu);
6219 for_each_possible_cpu(cpu) {
6220 if (cpu == failed_cpu)
6222 swevent_hlist_put_cpu(event, cpu);
6229 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6231 static void sw_perf_event_destroy(struct perf_event *event)
6233 u64 event_id = event->attr.config;
6235 WARN_ON(event->parent);
6237 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6238 swevent_hlist_put(event);
6241 static int perf_swevent_init(struct perf_event *event)
6243 u64 event_id = event->attr.config;
6245 if (event->attr.type != PERF_TYPE_SOFTWARE)
6249 * no branch sampling for software events
6251 if (has_branch_stack(event))
6255 case PERF_COUNT_SW_CPU_CLOCK:
6256 case PERF_COUNT_SW_TASK_CLOCK:
6263 if (event_id >= PERF_COUNT_SW_MAX)
6266 if (!event->parent) {
6269 err = swevent_hlist_get(event);
6273 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6274 event->destroy = sw_perf_event_destroy;
6280 static struct pmu perf_swevent = {
6281 .task_ctx_nr = perf_sw_context,
6283 .event_init = perf_swevent_init,
6284 .add = perf_swevent_add,
6285 .del = perf_swevent_del,
6286 .start = perf_swevent_start,
6287 .stop = perf_swevent_stop,
6288 .read = perf_swevent_read,
6291 #ifdef CONFIG_EVENT_TRACING
6293 static int perf_tp_filter_match(struct perf_event *event,
6294 struct perf_sample_data *data)
6296 void *record = data->raw->data;
6298 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6303 static int perf_tp_event_match(struct perf_event *event,
6304 struct perf_sample_data *data,
6305 struct pt_regs *regs)
6307 if (event->hw.state & PERF_HES_STOPPED)
6310 * All tracepoints are from kernel-space.
6312 if (event->attr.exclude_kernel)
6315 if (!perf_tp_filter_match(event, data))
6321 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6322 struct pt_regs *regs, struct hlist_head *head, int rctx,
6323 struct task_struct *task)
6325 struct perf_sample_data data;
6326 struct perf_event *event;
6328 struct perf_raw_record raw = {
6333 perf_sample_data_init(&data, addr, 0);
6336 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6337 if (perf_tp_event_match(event, &data, regs))
6338 perf_swevent_event(event, count, &data, regs);
6342 * If we got specified a target task, also iterate its context and
6343 * deliver this event there too.
6345 if (task && task != current) {
6346 struct perf_event_context *ctx;
6347 struct trace_entry *entry = record;
6350 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6354 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6355 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6357 if (event->attr.config != entry->type)
6359 if (perf_tp_event_match(event, &data, regs))
6360 perf_swevent_event(event, count, &data, regs);
6366 perf_swevent_put_recursion_context(rctx);
6368 EXPORT_SYMBOL_GPL(perf_tp_event);
6370 static void tp_perf_event_destroy(struct perf_event *event)
6372 perf_trace_destroy(event);
6375 static int perf_tp_event_init(struct perf_event *event)
6379 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6383 * no branch sampling for tracepoint events
6385 if (has_branch_stack(event))
6388 err = perf_trace_init(event);
6392 event->destroy = tp_perf_event_destroy;
6397 static struct pmu perf_tracepoint = {
6398 .task_ctx_nr = perf_sw_context,
6400 .event_init = perf_tp_event_init,
6401 .add = perf_trace_add,
6402 .del = perf_trace_del,
6403 .start = perf_swevent_start,
6404 .stop = perf_swevent_stop,
6405 .read = perf_swevent_read,
6408 static inline void perf_tp_register(void)
6410 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6413 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6418 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6421 filter_str = strndup_user(arg, PAGE_SIZE);
6422 if (IS_ERR(filter_str))
6423 return PTR_ERR(filter_str);
6425 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6431 static void perf_event_free_filter(struct perf_event *event)
6433 ftrace_profile_free_filter(event);
6438 static inline void perf_tp_register(void)
6442 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6447 static void perf_event_free_filter(struct perf_event *event)
6451 #endif /* CONFIG_EVENT_TRACING */
6453 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6454 void perf_bp_event(struct perf_event *bp, void *data)
6456 struct perf_sample_data sample;
6457 struct pt_regs *regs = data;
6459 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6461 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6462 perf_swevent_event(bp, 1, &sample, regs);
6467 * hrtimer based swevent callback
6470 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6472 enum hrtimer_restart ret = HRTIMER_RESTART;
6473 struct perf_sample_data data;
6474 struct pt_regs *regs;
6475 struct perf_event *event;
6478 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6480 if (event->state != PERF_EVENT_STATE_ACTIVE)
6481 return HRTIMER_NORESTART;
6483 event->pmu->read(event);
6485 perf_sample_data_init(&data, 0, event->hw.last_period);
6486 regs = get_irq_regs();
6488 if (regs && !perf_exclude_event(event, regs)) {
6489 if (!(event->attr.exclude_idle && is_idle_task(current)))
6490 if (__perf_event_overflow(event, 1, &data, regs))
6491 ret = HRTIMER_NORESTART;
6494 period = max_t(u64, 10000, event->hw.sample_period);
6495 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6500 static void perf_swevent_start_hrtimer(struct perf_event *event)
6502 struct hw_perf_event *hwc = &event->hw;
6505 if (!is_sampling_event(event))
6508 period = local64_read(&hwc->period_left);
6513 local64_set(&hwc->period_left, 0);
6515 period = max_t(u64, 10000, hwc->sample_period);
6517 __hrtimer_start_range_ns(&hwc->hrtimer,
6518 ns_to_ktime(period), 0,
6519 HRTIMER_MODE_REL_PINNED, 0);
6522 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6524 struct hw_perf_event *hwc = &event->hw;
6526 if (is_sampling_event(event)) {
6527 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6528 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6530 hrtimer_cancel(&hwc->hrtimer);
6534 static void perf_swevent_init_hrtimer(struct perf_event *event)
6536 struct hw_perf_event *hwc = &event->hw;
6538 if (!is_sampling_event(event))
6541 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6542 hwc->hrtimer.function = perf_swevent_hrtimer;
6545 * Since hrtimers have a fixed rate, we can do a static freq->period
6546 * mapping and avoid the whole period adjust feedback stuff.
6548 if (event->attr.freq) {
6549 long freq = event->attr.sample_freq;
6551 event->attr.sample_period = NSEC_PER_SEC / freq;
6552 hwc->sample_period = event->attr.sample_period;
6553 local64_set(&hwc->period_left, hwc->sample_period);
6554 hwc->last_period = hwc->sample_period;
6555 event->attr.freq = 0;
6560 * Software event: cpu wall time clock
6563 static void cpu_clock_event_update(struct perf_event *event)
6568 now = local_clock();
6569 prev = local64_xchg(&event->hw.prev_count, now);
6570 local64_add(now - prev, &event->count);
6573 static void cpu_clock_event_start(struct perf_event *event, int flags)
6575 local64_set(&event->hw.prev_count, local_clock());
6576 perf_swevent_start_hrtimer(event);
6579 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6581 perf_swevent_cancel_hrtimer(event);
6582 cpu_clock_event_update(event);
6585 static int cpu_clock_event_add(struct perf_event *event, int flags)
6587 if (flags & PERF_EF_START)
6588 cpu_clock_event_start(event, flags);
6589 perf_event_update_userpage(event);
6594 static void cpu_clock_event_del(struct perf_event *event, int flags)
6596 cpu_clock_event_stop(event, flags);
6599 static void cpu_clock_event_read(struct perf_event *event)
6601 cpu_clock_event_update(event);
6604 static int cpu_clock_event_init(struct perf_event *event)
6606 if (event->attr.type != PERF_TYPE_SOFTWARE)
6609 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6613 * no branch sampling for software events
6615 if (has_branch_stack(event))
6618 perf_swevent_init_hrtimer(event);
6623 static struct pmu perf_cpu_clock = {
6624 .task_ctx_nr = perf_sw_context,
6626 .event_init = cpu_clock_event_init,
6627 .add = cpu_clock_event_add,
6628 .del = cpu_clock_event_del,
6629 .start = cpu_clock_event_start,
6630 .stop = cpu_clock_event_stop,
6631 .read = cpu_clock_event_read,
6635 * Software event: task time clock
6638 static void task_clock_event_update(struct perf_event *event, u64 now)
6643 prev = local64_xchg(&event->hw.prev_count, now);
6645 local64_add(delta, &event->count);
6648 static void task_clock_event_start(struct perf_event *event, int flags)
6650 local64_set(&event->hw.prev_count, event->ctx->time);
6651 perf_swevent_start_hrtimer(event);
6654 static void task_clock_event_stop(struct perf_event *event, int flags)
6656 perf_swevent_cancel_hrtimer(event);
6657 task_clock_event_update(event, event->ctx->time);
6660 static int task_clock_event_add(struct perf_event *event, int flags)
6662 if (flags & PERF_EF_START)
6663 task_clock_event_start(event, flags);
6664 perf_event_update_userpage(event);
6669 static void task_clock_event_del(struct perf_event *event, int flags)
6671 task_clock_event_stop(event, PERF_EF_UPDATE);
6674 static void task_clock_event_read(struct perf_event *event)
6676 u64 now = perf_clock();
6677 u64 delta = now - event->ctx->timestamp;
6678 u64 time = event->ctx->time + delta;
6680 task_clock_event_update(event, time);
6683 static int task_clock_event_init(struct perf_event *event)
6685 if (event->attr.type != PERF_TYPE_SOFTWARE)
6688 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6692 * no branch sampling for software events
6694 if (has_branch_stack(event))
6697 perf_swevent_init_hrtimer(event);
6702 static struct pmu perf_task_clock = {
6703 .task_ctx_nr = perf_sw_context,
6705 .event_init = task_clock_event_init,
6706 .add = task_clock_event_add,
6707 .del = task_clock_event_del,
6708 .start = task_clock_event_start,
6709 .stop = task_clock_event_stop,
6710 .read = task_clock_event_read,
6713 static void perf_pmu_nop_void(struct pmu *pmu)
6717 static int perf_pmu_nop_int(struct pmu *pmu)
6722 static void perf_pmu_start_txn(struct pmu *pmu)
6724 perf_pmu_disable(pmu);
6727 static int perf_pmu_commit_txn(struct pmu *pmu)
6729 perf_pmu_enable(pmu);
6733 static void perf_pmu_cancel_txn(struct pmu *pmu)
6735 perf_pmu_enable(pmu);
6738 static int perf_event_idx_default(struct perf_event *event)
6744 * Ensures all contexts with the same task_ctx_nr have the same
6745 * pmu_cpu_context too.
6747 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6754 list_for_each_entry(pmu, &pmus, entry) {
6755 if (pmu->task_ctx_nr == ctxn)
6756 return pmu->pmu_cpu_context;
6762 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6766 for_each_possible_cpu(cpu) {
6767 struct perf_cpu_context *cpuctx;
6769 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6771 if (cpuctx->unique_pmu == old_pmu)
6772 cpuctx->unique_pmu = pmu;
6776 static void free_pmu_context(struct pmu *pmu)
6780 mutex_lock(&pmus_lock);
6782 * Like a real lame refcount.
6784 list_for_each_entry(i, &pmus, entry) {
6785 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6786 update_pmu_context(i, pmu);
6791 free_percpu(pmu->pmu_cpu_context);
6793 mutex_unlock(&pmus_lock);
6795 static struct idr pmu_idr;
6798 type_show(struct device *dev, struct device_attribute *attr, char *page)
6800 struct pmu *pmu = dev_get_drvdata(dev);
6802 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6804 static DEVICE_ATTR_RO(type);
6807 perf_event_mux_interval_ms_show(struct device *dev,
6808 struct device_attribute *attr,
6811 struct pmu *pmu = dev_get_drvdata(dev);
6813 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6817 perf_event_mux_interval_ms_store(struct device *dev,
6818 struct device_attribute *attr,
6819 const char *buf, size_t count)
6821 struct pmu *pmu = dev_get_drvdata(dev);
6822 int timer, cpu, ret;
6824 ret = kstrtoint(buf, 0, &timer);
6831 /* same value, noting to do */
6832 if (timer == pmu->hrtimer_interval_ms)
6835 pmu->hrtimer_interval_ms = timer;
6837 /* update all cpuctx for this PMU */
6838 for_each_possible_cpu(cpu) {
6839 struct perf_cpu_context *cpuctx;
6840 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6841 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6843 if (hrtimer_active(&cpuctx->hrtimer))
6844 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6849 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6851 static struct attribute *pmu_dev_attrs[] = {
6852 &dev_attr_type.attr,
6853 &dev_attr_perf_event_mux_interval_ms.attr,
6856 ATTRIBUTE_GROUPS(pmu_dev);
6858 static int pmu_bus_running;
6859 static struct bus_type pmu_bus = {
6860 .name = "event_source",
6861 .dev_groups = pmu_dev_groups,
6864 static void pmu_dev_release(struct device *dev)
6869 static int pmu_dev_alloc(struct pmu *pmu)
6873 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6877 pmu->dev->groups = pmu->attr_groups;
6878 device_initialize(pmu->dev);
6879 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6883 dev_set_drvdata(pmu->dev, pmu);
6884 pmu->dev->bus = &pmu_bus;
6885 pmu->dev->release = pmu_dev_release;
6886 ret = device_add(pmu->dev);
6894 put_device(pmu->dev);
6898 static struct lock_class_key cpuctx_mutex;
6899 static struct lock_class_key cpuctx_lock;
6901 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6905 mutex_lock(&pmus_lock);
6907 pmu->pmu_disable_count = alloc_percpu(int);
6908 if (!pmu->pmu_disable_count)
6917 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6925 if (pmu_bus_running) {
6926 ret = pmu_dev_alloc(pmu);
6932 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6933 if (pmu->pmu_cpu_context)
6934 goto got_cpu_context;
6937 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6938 if (!pmu->pmu_cpu_context)
6941 for_each_possible_cpu(cpu) {
6942 struct perf_cpu_context *cpuctx;
6944 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6945 __perf_event_init_context(&cpuctx->ctx);
6946 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6947 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6948 cpuctx->ctx.pmu = pmu;
6950 __perf_cpu_hrtimer_init(cpuctx, cpu);
6952 cpuctx->unique_pmu = pmu;
6956 if (!pmu->start_txn) {
6957 if (pmu->pmu_enable) {
6959 * If we have pmu_enable/pmu_disable calls, install
6960 * transaction stubs that use that to try and batch
6961 * hardware accesses.
6963 pmu->start_txn = perf_pmu_start_txn;
6964 pmu->commit_txn = perf_pmu_commit_txn;
6965 pmu->cancel_txn = perf_pmu_cancel_txn;
6967 pmu->start_txn = perf_pmu_nop_void;
6968 pmu->commit_txn = perf_pmu_nop_int;
6969 pmu->cancel_txn = perf_pmu_nop_void;
6973 if (!pmu->pmu_enable) {
6974 pmu->pmu_enable = perf_pmu_nop_void;
6975 pmu->pmu_disable = perf_pmu_nop_void;
6978 if (!pmu->event_idx)
6979 pmu->event_idx = perf_event_idx_default;
6981 list_add_rcu(&pmu->entry, &pmus);
6984 mutex_unlock(&pmus_lock);
6989 device_del(pmu->dev);
6990 put_device(pmu->dev);
6993 if (pmu->type >= PERF_TYPE_MAX)
6994 idr_remove(&pmu_idr, pmu->type);
6997 free_percpu(pmu->pmu_disable_count);
7000 EXPORT_SYMBOL_GPL(perf_pmu_register);
7002 void perf_pmu_unregister(struct pmu *pmu)
7004 mutex_lock(&pmus_lock);
7005 list_del_rcu(&pmu->entry);
7006 mutex_unlock(&pmus_lock);
7009 * We dereference the pmu list under both SRCU and regular RCU, so
7010 * synchronize against both of those.
7012 synchronize_srcu(&pmus_srcu);
7015 free_percpu(pmu->pmu_disable_count);
7016 if (pmu->type >= PERF_TYPE_MAX)
7017 idr_remove(&pmu_idr, pmu->type);
7018 device_del(pmu->dev);
7019 put_device(pmu->dev);
7020 free_pmu_context(pmu);
7022 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
7024 static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)
7028 if (!try_module_get(pmu->module))
7031 ret = pmu->event_init(event);
7033 module_put(pmu->module);
7038 struct pmu *perf_init_event(struct perf_event *event)
7040 struct pmu *pmu = NULL;
7044 idx = srcu_read_lock(&pmus_srcu);
7047 pmu = idr_find(&pmu_idr, event->attr.type);
7050 ret = perf_try_init_event(pmu, event);
7056 list_for_each_entry_rcu(pmu, &pmus, entry) {
7057 ret = perf_try_init_event(pmu, event);
7061 if (ret != -ENOENT) {
7066 pmu = ERR_PTR(-ENOENT);
7068 srcu_read_unlock(&pmus_srcu, idx);
7073 static void account_event_cpu(struct perf_event *event, int cpu)
7078 if (is_cgroup_event(event))
7079 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
7082 static void account_event(struct perf_event *event)
7087 if (event->attach_state & PERF_ATTACH_TASK)
7088 static_key_slow_inc(&perf_sched_events.key);
7089 if (event->attr.mmap || event->attr.mmap_data)
7090 atomic_inc(&nr_mmap_events);
7091 if (event->attr.comm)
7092 atomic_inc(&nr_comm_events);
7093 if (event->attr.task)
7094 atomic_inc(&nr_task_events);
7095 if (event->attr.freq) {
7096 if (atomic_inc_return(&nr_freq_events) == 1)
7097 tick_nohz_full_kick_all();
7099 if (has_branch_stack(event))
7100 static_key_slow_inc(&perf_sched_events.key);
7101 if (is_cgroup_event(event))
7102 static_key_slow_inc(&perf_sched_events.key);
7104 account_event_cpu(event, event->cpu);
7108 * Allocate and initialize a event structure
7110 static struct perf_event *
7111 perf_event_alloc(struct perf_event_attr *attr, int cpu,
7112 struct task_struct *task,
7113 struct perf_event *group_leader,
7114 struct perf_event *parent_event,
7115 perf_overflow_handler_t overflow_handler,
7119 struct perf_event *event;
7120 struct hw_perf_event *hwc;
7123 if ((unsigned)cpu >= nr_cpu_ids) {
7124 if (!task || cpu != -1)
7125 return ERR_PTR(-EINVAL);
7128 event = kzalloc(sizeof(*event), GFP_KERNEL);
7130 return ERR_PTR(-ENOMEM);
7133 * Single events are their own group leaders, with an
7134 * empty sibling list:
7137 group_leader = event;
7139 mutex_init(&event->child_mutex);
7140 INIT_LIST_HEAD(&event->child_list);
7142 INIT_LIST_HEAD(&event->group_entry);
7143 INIT_LIST_HEAD(&event->event_entry);
7144 INIT_LIST_HEAD(&event->sibling_list);
7145 INIT_LIST_HEAD(&event->rb_entry);
7146 INIT_LIST_HEAD(&event->active_entry);
7147 INIT_HLIST_NODE(&event->hlist_entry);
7150 init_waitqueue_head(&event->waitq);
7151 init_irq_work(&event->pending, perf_pending_event);
7153 mutex_init(&event->mmap_mutex);
7155 atomic_long_set(&event->refcount, 1);
7157 event->attr = *attr;
7158 event->group_leader = group_leader;
7162 event->parent = parent_event;
7164 event->ns = get_pid_ns(task_active_pid_ns(current));
7165 event->id = atomic64_inc_return(&perf_event_id);
7167 event->state = PERF_EVENT_STATE_INACTIVE;
7170 event->attach_state = PERF_ATTACH_TASK;
7172 if (attr->type == PERF_TYPE_TRACEPOINT)
7173 event->hw.tp_target = task;
7174 #ifdef CONFIG_HAVE_HW_BREAKPOINT
7176 * hw_breakpoint is a bit difficult here..
7178 else if (attr->type == PERF_TYPE_BREAKPOINT)
7179 event->hw.bp_target = task;
7183 if (!overflow_handler && parent_event) {
7184 overflow_handler = parent_event->overflow_handler;
7185 context = parent_event->overflow_handler_context;
7188 event->overflow_handler = overflow_handler;
7189 event->overflow_handler_context = context;
7191 perf_event__state_init(event);
7196 hwc->sample_period = attr->sample_period;
7197 if (attr->freq && attr->sample_freq)
7198 hwc->sample_period = 1;
7199 hwc->last_period = hwc->sample_period;
7201 local64_set(&hwc->period_left, hwc->sample_period);
7204 * we currently do not support PERF_FORMAT_GROUP on inherited events
7206 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7209 if (!has_branch_stack(event))
7210 event->attr.branch_sample_type = 0;
7212 pmu = perf_init_event(event);
7215 else if (IS_ERR(pmu)) {
7220 if (!event->parent) {
7221 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7222 err = get_callchain_buffers();
7232 event->destroy(event);
7233 module_put(pmu->module);
7236 put_pid_ns(event->ns);
7239 return ERR_PTR(err);
7242 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7243 struct perf_event_attr *attr)
7248 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7252 * zero the full structure, so that a short copy will be nice.
7254 memset(attr, 0, sizeof(*attr));
7256 ret = get_user(size, &uattr->size);
7260 if (size > PAGE_SIZE) /* silly large */
7263 if (!size) /* abi compat */
7264 size = PERF_ATTR_SIZE_VER0;
7266 if (size < PERF_ATTR_SIZE_VER0)
7270 * If we're handed a bigger struct than we know of,
7271 * ensure all the unknown bits are 0 - i.e. new
7272 * user-space does not rely on any kernel feature
7273 * extensions we dont know about yet.
7275 if (size > sizeof(*attr)) {
7276 unsigned char __user *addr;
7277 unsigned char __user *end;
7280 addr = (void __user *)uattr + sizeof(*attr);
7281 end = (void __user *)uattr + size;
7283 for (; addr < end; addr++) {
7284 ret = get_user(val, addr);
7290 size = sizeof(*attr);
7293 ret = copy_from_user(attr, uattr, size);
7297 if (attr->__reserved_1)
7300 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7303 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7306 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7307 u64 mask = attr->branch_sample_type;
7309 /* only using defined bits */
7310 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7313 /* at least one branch bit must be set */
7314 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7317 /* propagate priv level, when not set for branch */
7318 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7320 /* exclude_kernel checked on syscall entry */
7321 if (!attr->exclude_kernel)
7322 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7324 if (!attr->exclude_user)
7325 mask |= PERF_SAMPLE_BRANCH_USER;
7327 if (!attr->exclude_hv)
7328 mask |= PERF_SAMPLE_BRANCH_HV;
7330 * adjust user setting (for HW filter setup)
7332 attr->branch_sample_type = mask;
7334 /* privileged levels capture (kernel, hv): check permissions */
7335 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7336 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7340 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7341 ret = perf_reg_validate(attr->sample_regs_user);
7346 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7347 if (!arch_perf_have_user_stack_dump())
7351 * We have __u32 type for the size, but so far
7352 * we can only use __u16 as maximum due to the
7353 * __u16 sample size limit.
7355 if (attr->sample_stack_user >= USHRT_MAX)
7357 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7361 if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
7362 ret = perf_reg_validate(attr->sample_regs_intr);
7367 put_user(sizeof(*attr), &uattr->size);
7373 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7375 struct ring_buffer *rb = NULL;
7381 /* don't allow circular references */
7382 if (event == output_event)
7386 * Don't allow cross-cpu buffers
7388 if (output_event->cpu != event->cpu)
7392 * If its not a per-cpu rb, it must be the same task.
7394 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7398 mutex_lock(&event->mmap_mutex);
7399 /* Can't redirect output if we've got an active mmap() */
7400 if (atomic_read(&event->mmap_count))
7404 /* get the rb we want to redirect to */
7405 rb = ring_buffer_get(output_event);
7410 ring_buffer_attach(event, rb);
7414 mutex_unlock(&event->mmap_mutex);
7420 static void mutex_lock_double(struct mutex *a, struct mutex *b)
7426 mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
7430 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7432 * @attr_uptr: event_id type attributes for monitoring/sampling
7435 * @group_fd: group leader event fd
7437 SYSCALL_DEFINE5(perf_event_open,
7438 struct perf_event_attr __user *, attr_uptr,
7439 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7441 struct perf_event *group_leader = NULL, *output_event = NULL;
7442 struct perf_event *event, *sibling;
7443 struct perf_event_attr attr;
7444 struct perf_event_context *ctx, *uninitialized_var(gctx);
7445 struct file *event_file = NULL;
7446 struct fd group = {NULL, 0};
7447 struct task_struct *task = NULL;
7452 int f_flags = O_RDWR;
7454 /* for future expandability... */
7455 if (flags & ~PERF_FLAG_ALL)
7458 err = perf_copy_attr(attr_uptr, &attr);
7462 if (!attr.exclude_kernel) {
7463 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7468 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7471 if (attr.sample_period & (1ULL << 63))
7476 * In cgroup mode, the pid argument is used to pass the fd
7477 * opened to the cgroup directory in cgroupfs. The cpu argument
7478 * designates the cpu on which to monitor threads from that
7481 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7484 if (flags & PERF_FLAG_FD_CLOEXEC)
7485 f_flags |= O_CLOEXEC;
7487 event_fd = get_unused_fd_flags(f_flags);
7491 if (group_fd != -1) {
7492 err = perf_fget_light(group_fd, &group);
7495 group_leader = group.file->private_data;
7496 if (flags & PERF_FLAG_FD_OUTPUT)
7497 output_event = group_leader;
7498 if (flags & PERF_FLAG_FD_NO_GROUP)
7499 group_leader = NULL;
7502 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7503 task = find_lively_task_by_vpid(pid);
7505 err = PTR_ERR(task);
7510 if (task && group_leader &&
7511 group_leader->attr.inherit != attr.inherit) {
7518 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7520 if (IS_ERR(event)) {
7521 err = PTR_ERR(event);
7525 if (flags & PERF_FLAG_PID_CGROUP) {
7526 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7528 __free_event(event);
7533 if (is_sampling_event(event)) {
7534 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7540 account_event(event);
7543 * Special case software events and allow them to be part of
7544 * any hardware group.
7549 (is_software_event(event) != is_software_event(group_leader))) {
7550 if (is_software_event(event)) {
7552 * If event and group_leader are not both a software
7553 * event, and event is, then group leader is not.
7555 * Allow the addition of software events to !software
7556 * groups, this is safe because software events never
7559 pmu = group_leader->pmu;
7560 } else if (is_software_event(group_leader) &&
7561 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7563 * In case the group is a pure software group, and we
7564 * try to add a hardware event, move the whole group to
7565 * the hardware context.
7572 * Get the target context (task or percpu):
7574 ctx = find_get_context(pmu, task, event);
7581 put_task_struct(task);
7586 * Look up the group leader (we will attach this event to it):
7592 * Do not allow a recursive hierarchy (this new sibling
7593 * becoming part of another group-sibling):
7595 if (group_leader->group_leader != group_leader)
7598 * Do not allow to attach to a group in a different
7599 * task or CPU context:
7603 * Make sure we're both on the same task, or both
7606 if (group_leader->ctx->task != ctx->task)
7610 * Make sure we're both events for the same CPU;
7611 * grouping events for different CPUs is broken; since
7612 * you can never concurrently schedule them anyhow.
7614 if (group_leader->cpu != event->cpu)
7617 if (group_leader->ctx != ctx)
7622 * Only a group leader can be exclusive or pinned
7624 if (attr.exclusive || attr.pinned)
7629 err = perf_event_set_output(event, output_event);
7634 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7636 if (IS_ERR(event_file)) {
7637 err = PTR_ERR(event_file);
7642 gctx = group_leader->ctx;
7645 * See perf_event_ctx_lock() for comments on the details
7646 * of swizzling perf_event::ctx.
7648 mutex_lock_double(&gctx->mutex, &ctx->mutex);
7650 perf_remove_from_context(group_leader, false);
7652 list_for_each_entry(sibling, &group_leader->sibling_list,
7654 perf_remove_from_context(sibling, false);
7658 mutex_lock(&ctx->mutex);
7661 WARN_ON_ONCE(ctx->parent_ctx);
7665 * Wait for everybody to stop referencing the events through
7666 * the old lists, before installing it on new lists.
7671 * Install the group siblings before the group leader.
7673 * Because a group leader will try and install the entire group
7674 * (through the sibling list, which is still in-tact), we can
7675 * end up with siblings installed in the wrong context.
7677 * By installing siblings first we NO-OP because they're not
7678 * reachable through the group lists.
7680 list_for_each_entry(sibling, &group_leader->sibling_list,
7682 perf_event__state_init(sibling);
7683 perf_install_in_context(ctx, sibling, sibling->cpu);
7688 * Removing from the context ends up with disabled
7689 * event. What we want here is event in the initial
7690 * startup state, ready to be add into new context.
7692 perf_event__state_init(group_leader);
7693 perf_install_in_context(ctx, group_leader, group_leader->cpu);
7697 perf_install_in_context(ctx, event, event->cpu);
7698 perf_unpin_context(ctx);
7701 mutex_unlock(&gctx->mutex);
7704 mutex_unlock(&ctx->mutex);
7708 event->owner = current;
7710 mutex_lock(¤t->perf_event_mutex);
7711 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7712 mutex_unlock(¤t->perf_event_mutex);
7715 * Precalculate sample_data sizes
7717 perf_event__header_size(event);
7718 perf_event__id_header_size(event);
7721 * Drop the reference on the group_event after placing the
7722 * new event on the sibling_list. This ensures destruction
7723 * of the group leader will find the pointer to itself in
7724 * perf_group_detach().
7727 fd_install(event_fd, event_file);
7731 perf_unpin_context(ctx);
7739 put_task_struct(task);
7743 put_unused_fd(event_fd);
7748 * perf_event_create_kernel_counter
7750 * @attr: attributes of the counter to create
7751 * @cpu: cpu in which the counter is bound
7752 * @task: task to profile (NULL for percpu)
7755 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7756 struct task_struct *task,
7757 perf_overflow_handler_t overflow_handler,
7760 struct perf_event_context *ctx;
7761 struct perf_event *event;
7765 * Get the target context (task or percpu):
7768 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7769 overflow_handler, context);
7770 if (IS_ERR(event)) {
7771 err = PTR_ERR(event);
7775 /* Mark owner so we could distinguish it from user events. */
7776 event->owner = EVENT_OWNER_KERNEL;
7778 account_event(event);
7780 ctx = find_get_context(event->pmu, task, event);
7786 WARN_ON_ONCE(ctx->parent_ctx);
7787 mutex_lock(&ctx->mutex);
7788 perf_install_in_context(ctx, event, cpu);
7789 perf_unpin_context(ctx);
7790 mutex_unlock(&ctx->mutex);
7797 return ERR_PTR(err);
7799 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7801 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7803 struct perf_event_context *src_ctx;
7804 struct perf_event_context *dst_ctx;
7805 struct perf_event *event, *tmp;
7808 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7809 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7812 * See perf_event_ctx_lock() for comments on the details
7813 * of swizzling perf_event::ctx.
7815 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex);
7816 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7818 perf_remove_from_context(event, false);
7819 unaccount_event_cpu(event, src_cpu);
7821 list_add(&event->migrate_entry, &events);
7825 * Wait for the events to quiesce before re-instating them.
7830 * Re-instate events in 2 passes.
7832 * Skip over group leaders and only install siblings on this first
7833 * pass, siblings will not get enabled without a leader, however a
7834 * leader will enable its siblings, even if those are still on the old
7837 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7838 if (event->group_leader == event)
7841 list_del(&event->migrate_entry);
7842 if (event->state >= PERF_EVENT_STATE_OFF)
7843 event->state = PERF_EVENT_STATE_INACTIVE;
7844 account_event_cpu(event, dst_cpu);
7845 perf_install_in_context(dst_ctx, event, dst_cpu);
7850 * Once all the siblings are setup properly, install the group leaders
7853 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7854 list_del(&event->migrate_entry);
7855 if (event->state >= PERF_EVENT_STATE_OFF)
7856 event->state = PERF_EVENT_STATE_INACTIVE;
7857 account_event_cpu(event, dst_cpu);
7858 perf_install_in_context(dst_ctx, event, dst_cpu);
7861 mutex_unlock(&dst_ctx->mutex);
7862 mutex_unlock(&src_ctx->mutex);
7864 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7866 static void sync_child_event(struct perf_event *child_event,
7867 struct task_struct *child)
7869 struct perf_event *parent_event = child_event->parent;
7872 if (child_event->attr.inherit_stat)
7873 perf_event_read_event(child_event, child);
7875 child_val = perf_event_count(child_event);
7878 * Add back the child's count to the parent's count:
7880 atomic64_add(child_val, &parent_event->child_count);
7881 atomic64_add(child_event->total_time_enabled,
7882 &parent_event->child_total_time_enabled);
7883 atomic64_add(child_event->total_time_running,
7884 &parent_event->child_total_time_running);
7887 * Remove this event from the parent's list
7889 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7890 mutex_lock(&parent_event->child_mutex);
7891 list_del_init(&child_event->child_list);
7892 mutex_unlock(&parent_event->child_mutex);
7895 * Make sure user/parent get notified, that we just
7898 perf_event_wakeup(parent_event);
7901 * Release the parent event, if this was the last
7904 put_event(parent_event);
7908 __perf_event_exit_task(struct perf_event *child_event,
7909 struct perf_event_context *child_ctx,
7910 struct task_struct *child)
7913 * Do not destroy the 'original' grouping; because of the context
7914 * switch optimization the original events could've ended up in a
7915 * random child task.
7917 * If we were to destroy the original group, all group related
7918 * operations would cease to function properly after this random
7921 * Do destroy all inherited groups, we don't care about those
7922 * and being thorough is better.
7924 perf_remove_from_context(child_event, !!child_event->parent);
7927 * It can happen that the parent exits first, and has events
7928 * that are still around due to the child reference. These
7929 * events need to be zapped.
7931 if (child_event->parent) {
7932 sync_child_event(child_event, child);
7933 free_event(child_event);
7935 child_event->state = PERF_EVENT_STATE_EXIT;
7936 perf_event_wakeup(child_event);
7940 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7942 struct perf_event *child_event, *next;
7943 struct perf_event_context *child_ctx, *clone_ctx = NULL;
7944 unsigned long flags;
7946 if (likely(!child->perf_event_ctxp[ctxn])) {
7947 perf_event_task(child, NULL, 0);
7951 local_irq_save(flags);
7953 * We can't reschedule here because interrupts are disabled,
7954 * and either child is current or it is a task that can't be
7955 * scheduled, so we are now safe from rescheduling changing
7958 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7961 * Take the context lock here so that if find_get_context is
7962 * reading child->perf_event_ctxp, we wait until it has
7963 * incremented the context's refcount before we do put_ctx below.
7965 raw_spin_lock(&child_ctx->lock);
7966 task_ctx_sched_out(child_ctx);
7967 child->perf_event_ctxp[ctxn] = NULL;
7970 * If this context is a clone; unclone it so it can't get
7971 * swapped to another process while we're removing all
7972 * the events from it.
7974 clone_ctx = unclone_ctx(child_ctx);
7975 update_context_time(child_ctx);
7976 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7982 * Report the task dead after unscheduling the events so that we
7983 * won't get any samples after PERF_RECORD_EXIT. We can however still
7984 * get a few PERF_RECORD_READ events.
7986 perf_event_task(child, child_ctx, 0);
7989 * We can recurse on the same lock type through:
7991 * __perf_event_exit_task()
7992 * sync_child_event()
7994 * mutex_lock(&ctx->mutex)
7996 * But since its the parent context it won't be the same instance.
7998 mutex_lock(&child_ctx->mutex);
8000 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
8001 __perf_event_exit_task(child_event, child_ctx, child);
8003 mutex_unlock(&child_ctx->mutex);
8009 * When a child task exits, feed back event values to parent events.
8011 void perf_event_exit_task(struct task_struct *child)
8013 struct perf_event *event, *tmp;
8016 mutex_lock(&child->perf_event_mutex);
8017 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
8019 list_del_init(&event->owner_entry);
8022 * Ensure the list deletion is visible before we clear
8023 * the owner, closes a race against perf_release() where
8024 * we need to serialize on the owner->perf_event_mutex.
8027 event->owner = NULL;
8029 mutex_unlock(&child->perf_event_mutex);
8031 for_each_task_context_nr(ctxn)
8032 perf_event_exit_task_context(child, ctxn);
8035 static void perf_free_event(struct perf_event *event,
8036 struct perf_event_context *ctx)
8038 struct perf_event *parent = event->parent;
8040 if (WARN_ON_ONCE(!parent))
8043 mutex_lock(&parent->child_mutex);
8044 list_del_init(&event->child_list);
8045 mutex_unlock(&parent->child_mutex);
8049 raw_spin_lock_irq(&ctx->lock);
8050 perf_group_detach(event);
8051 list_del_event(event, ctx);
8052 raw_spin_unlock_irq(&ctx->lock);
8057 * Free an unexposed, unused context as created by inheritance by
8058 * perf_event_init_task below, used by fork() in case of fail.
8060 * Not all locks are strictly required, but take them anyway to be nice and
8061 * help out with the lockdep assertions.
8063 void perf_event_free_task(struct task_struct *task)
8065 struct perf_event_context *ctx;
8066 struct perf_event *event, *tmp;
8069 for_each_task_context_nr(ctxn) {
8070 ctx = task->perf_event_ctxp[ctxn];
8074 mutex_lock(&ctx->mutex);
8076 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
8078 perf_free_event(event, ctx);
8080 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
8082 perf_free_event(event, ctx);
8084 if (!list_empty(&ctx->pinned_groups) ||
8085 !list_empty(&ctx->flexible_groups))
8088 mutex_unlock(&ctx->mutex);
8094 void perf_event_delayed_put(struct task_struct *task)
8098 for_each_task_context_nr(ctxn)
8099 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
8103 * inherit a event from parent task to child task:
8105 static struct perf_event *
8106 inherit_event(struct perf_event *parent_event,
8107 struct task_struct *parent,
8108 struct perf_event_context *parent_ctx,
8109 struct task_struct *child,
8110 struct perf_event *group_leader,
8111 struct perf_event_context *child_ctx)
8113 enum perf_event_active_state parent_state = parent_event->state;
8114 struct perf_event *child_event;
8115 unsigned long flags;
8118 * Instead of creating recursive hierarchies of events,
8119 * we link inherited events back to the original parent,
8120 * which has a filp for sure, which we use as the reference
8123 if (parent_event->parent)
8124 parent_event = parent_event->parent;
8126 child_event = perf_event_alloc(&parent_event->attr,
8129 group_leader, parent_event,
8131 if (IS_ERR(child_event))
8134 if (is_orphaned_event(parent_event) ||
8135 !atomic_long_inc_not_zero(&parent_event->refcount)) {
8136 free_event(child_event);
8143 * Make the child state follow the state of the parent event,
8144 * not its attr.disabled bit. We hold the parent's mutex,
8145 * so we won't race with perf_event_{en, dis}able_family.
8147 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
8148 child_event->state = PERF_EVENT_STATE_INACTIVE;
8150 child_event->state = PERF_EVENT_STATE_OFF;
8152 if (parent_event->attr.freq) {
8153 u64 sample_period = parent_event->hw.sample_period;
8154 struct hw_perf_event *hwc = &child_event->hw;
8156 hwc->sample_period = sample_period;
8157 hwc->last_period = sample_period;
8159 local64_set(&hwc->period_left, sample_period);
8162 child_event->ctx = child_ctx;
8163 child_event->overflow_handler = parent_event->overflow_handler;
8164 child_event->overflow_handler_context
8165 = parent_event->overflow_handler_context;
8168 * Precalculate sample_data sizes
8170 perf_event__header_size(child_event);
8171 perf_event__id_header_size(child_event);
8174 * Link it up in the child's context:
8176 raw_spin_lock_irqsave(&child_ctx->lock, flags);
8177 add_event_to_ctx(child_event, child_ctx);
8178 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8181 * Link this into the parent event's child list
8183 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8184 mutex_lock(&parent_event->child_mutex);
8185 list_add_tail(&child_event->child_list, &parent_event->child_list);
8186 mutex_unlock(&parent_event->child_mutex);
8191 static int inherit_group(struct perf_event *parent_event,
8192 struct task_struct *parent,
8193 struct perf_event_context *parent_ctx,
8194 struct task_struct *child,
8195 struct perf_event_context *child_ctx)
8197 struct perf_event *leader;
8198 struct perf_event *sub;
8199 struct perf_event *child_ctr;
8201 leader = inherit_event(parent_event, parent, parent_ctx,
8202 child, NULL, child_ctx);
8204 return PTR_ERR(leader);
8205 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
8206 child_ctr = inherit_event(sub, parent, parent_ctx,
8207 child, leader, child_ctx);
8208 if (IS_ERR(child_ctr))
8209 return PTR_ERR(child_ctr);
8215 inherit_task_group(struct perf_event *event, struct task_struct *parent,
8216 struct perf_event_context *parent_ctx,
8217 struct task_struct *child, int ctxn,
8221 struct perf_event_context *child_ctx;
8223 if (!event->attr.inherit) {
8228 child_ctx = child->perf_event_ctxp[ctxn];
8231 * This is executed from the parent task context, so
8232 * inherit events that have been marked for cloning.
8233 * First allocate and initialize a context for the
8237 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
8241 child->perf_event_ctxp[ctxn] = child_ctx;
8244 ret = inherit_group(event, parent, parent_ctx,
8254 * Initialize the perf_event context in task_struct
8256 static int perf_event_init_context(struct task_struct *child, int ctxn)
8258 struct perf_event_context *child_ctx, *parent_ctx;
8259 struct perf_event_context *cloned_ctx;
8260 struct perf_event *event;
8261 struct task_struct *parent = current;
8262 int inherited_all = 1;
8263 unsigned long flags;
8266 if (likely(!parent->perf_event_ctxp[ctxn]))
8270 * If the parent's context is a clone, pin it so it won't get
8273 parent_ctx = perf_pin_task_context(parent, ctxn);
8278 * No need to check if parent_ctx != NULL here; since we saw
8279 * it non-NULL earlier, the only reason for it to become NULL
8280 * is if we exit, and since we're currently in the middle of
8281 * a fork we can't be exiting at the same time.
8285 * Lock the parent list. No need to lock the child - not PID
8286 * hashed yet and not running, so nobody can access it.
8288 mutex_lock(&parent_ctx->mutex);
8291 * We dont have to disable NMIs - we are only looking at
8292 * the list, not manipulating it:
8294 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8295 ret = inherit_task_group(event, parent, parent_ctx,
8296 child, ctxn, &inherited_all);
8302 * We can't hold ctx->lock when iterating the ->flexible_group list due
8303 * to allocations, but we need to prevent rotation because
8304 * rotate_ctx() will change the list from interrupt context.
8306 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8307 parent_ctx->rotate_disable = 1;
8308 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8310 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8311 ret = inherit_task_group(event, parent, parent_ctx,
8312 child, ctxn, &inherited_all);
8317 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8318 parent_ctx->rotate_disable = 0;
8320 child_ctx = child->perf_event_ctxp[ctxn];
8322 if (child_ctx && inherited_all) {
8324 * Mark the child context as a clone of the parent
8325 * context, or of whatever the parent is a clone of.
8327 * Note that if the parent is a clone, the holding of
8328 * parent_ctx->lock avoids it from being uncloned.
8330 cloned_ctx = parent_ctx->parent_ctx;
8332 child_ctx->parent_ctx = cloned_ctx;
8333 child_ctx->parent_gen = parent_ctx->parent_gen;
8335 child_ctx->parent_ctx = parent_ctx;
8336 child_ctx->parent_gen = parent_ctx->generation;
8338 get_ctx(child_ctx->parent_ctx);
8341 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8342 mutex_unlock(&parent_ctx->mutex);
8344 perf_unpin_context(parent_ctx);
8345 put_ctx(parent_ctx);
8351 * Initialize the perf_event context in task_struct
8353 int perf_event_init_task(struct task_struct *child)
8357 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8358 mutex_init(&child->perf_event_mutex);
8359 INIT_LIST_HEAD(&child->perf_event_list);
8361 for_each_task_context_nr(ctxn) {
8362 ret = perf_event_init_context(child, ctxn);
8364 perf_event_free_task(child);
8372 static void __init perf_event_init_all_cpus(void)
8374 struct swevent_htable *swhash;
8377 for_each_possible_cpu(cpu) {
8378 swhash = &per_cpu(swevent_htable, cpu);
8379 mutex_init(&swhash->hlist_mutex);
8380 INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu));
8384 static void perf_event_init_cpu(int cpu)
8386 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8388 mutex_lock(&swhash->hlist_mutex);
8389 swhash->online = true;
8390 if (swhash->hlist_refcount > 0) {
8391 struct swevent_hlist *hlist;
8393 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8395 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8397 mutex_unlock(&swhash->hlist_mutex);
8400 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8401 static void __perf_event_exit_context(void *__info)
8403 struct remove_event re = { .detach_group = true };
8404 struct perf_event_context *ctx = __info;
8407 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8408 __perf_remove_from_context(&re);
8412 static void perf_event_exit_cpu_context(int cpu)
8414 struct perf_event_context *ctx;
8418 idx = srcu_read_lock(&pmus_srcu);
8419 list_for_each_entry_rcu(pmu, &pmus, entry) {
8420 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8422 mutex_lock(&ctx->mutex);
8423 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8424 mutex_unlock(&ctx->mutex);
8426 srcu_read_unlock(&pmus_srcu, idx);
8429 static void perf_event_exit_cpu(int cpu)
8431 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8433 perf_event_exit_cpu_context(cpu);
8435 mutex_lock(&swhash->hlist_mutex);
8436 swhash->online = false;
8437 swevent_hlist_release(swhash);
8438 mutex_unlock(&swhash->hlist_mutex);
8441 static inline void perf_event_exit_cpu(int cpu) { }
8445 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8449 for_each_online_cpu(cpu)
8450 perf_event_exit_cpu(cpu);
8456 * Run the perf reboot notifier at the very last possible moment so that
8457 * the generic watchdog code runs as long as possible.
8459 static struct notifier_block perf_reboot_notifier = {
8460 .notifier_call = perf_reboot,
8461 .priority = INT_MIN,
8465 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8467 unsigned int cpu = (long)hcpu;
8469 switch (action & ~CPU_TASKS_FROZEN) {
8471 case CPU_UP_PREPARE:
8472 case CPU_DOWN_FAILED:
8473 perf_event_init_cpu(cpu);
8476 case CPU_UP_CANCELED:
8477 case CPU_DOWN_PREPARE:
8478 perf_event_exit_cpu(cpu);
8487 void __init perf_event_init(void)
8493 perf_event_init_all_cpus();
8494 init_srcu_struct(&pmus_srcu);
8495 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8496 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8497 perf_pmu_register(&perf_task_clock, NULL, -1);
8499 perf_cpu_notifier(perf_cpu_notify);
8500 register_reboot_notifier(&perf_reboot_notifier);
8502 ret = init_hw_breakpoint();
8503 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8505 /* do not patch jump label more than once per second */
8506 jump_label_rate_limit(&perf_sched_events, HZ);
8509 * Build time assertion that we keep the data_head at the intended
8510 * location. IOW, validation we got the __reserved[] size right.
8512 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8516 static int __init perf_event_sysfs_init(void)
8521 mutex_lock(&pmus_lock);
8523 ret = bus_register(&pmu_bus);
8527 list_for_each_entry(pmu, &pmus, entry) {
8528 if (!pmu->name || pmu->type < 0)
8531 ret = pmu_dev_alloc(pmu);
8532 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8534 pmu_bus_running = 1;
8538 mutex_unlock(&pmus_lock);
8542 device_initcall(perf_event_sysfs_init);
8544 #ifdef CONFIG_CGROUP_PERF
8545 static struct cgroup_subsys_state *
8546 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8548 struct perf_cgroup *jc;
8550 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8552 return ERR_PTR(-ENOMEM);
8554 jc->info = alloc_percpu(struct perf_cgroup_info);
8557 return ERR_PTR(-ENOMEM);
8563 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8565 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8567 free_percpu(jc->info);
8571 static int __perf_cgroup_move(void *info)
8573 struct task_struct *task = info;
8574 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8578 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8579 struct cgroup_taskset *tset)
8581 struct task_struct *task;
8583 cgroup_taskset_for_each(task, tset)
8584 task_function_call(task, __perf_cgroup_move, task);
8587 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8588 struct cgroup_subsys_state *old_css,
8589 struct task_struct *task)
8592 * cgroup_exit() is called in the copy_process() failure path.
8593 * Ignore this case since the task hasn't ran yet, this avoids
8594 * trying to poke a half freed task state from generic code.
8596 if (!(task->flags & PF_EXITING))
8599 task_function_call(task, __perf_cgroup_move, task);
8602 struct cgroup_subsys perf_event_cgrp_subsys = {
8603 .css_alloc = perf_cgroup_css_alloc,
8604 .css_free = perf_cgroup_css_free,
8605 .exit = perf_cgroup_exit,
8606 .attach = perf_cgroup_attach,
8608 #endif /* CONFIG_CGROUP_PERF */