2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
45 #include <asm/irq_regs.h>
47 struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
54 static void remote_function(void *data)
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
65 tfc->ret = tfc->func(tfc->info);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
84 struct remote_function_call data = {
88 .ret = -ESRCH, /* No such (running) process */
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
108 struct remote_function_call data = {
112 .ret = -ENXIO, /* No such CPU */
115 smp_call_function_single(cpu, remote_function, &data, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP |\
123 PERF_FLAG_FD_CLOEXEC)
126 * branch priv levels that need permission checks
128 #define PERF_SAMPLE_BRANCH_PERM_PLM \
129 (PERF_SAMPLE_BRANCH_KERNEL |\
130 PERF_SAMPLE_BRANCH_HV)
133 EVENT_FLEXIBLE = 0x1,
135 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
139 * perf_sched_events : >0 events exist
140 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
142 struct static_key_deferred perf_sched_events __read_mostly;
143 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
144 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
146 static atomic_t nr_mmap_events __read_mostly;
147 static atomic_t nr_comm_events __read_mostly;
148 static atomic_t nr_task_events __read_mostly;
149 static atomic_t nr_freq_events __read_mostly;
151 static LIST_HEAD(pmus);
152 static DEFINE_MUTEX(pmus_lock);
153 static struct srcu_struct pmus_srcu;
156 * perf event paranoia level:
157 * -1 - not paranoid at all
158 * 0 - disallow raw tracepoint access for unpriv
159 * 1 - disallow cpu events for unpriv
160 * 2 - disallow kernel profiling for unpriv
162 int sysctl_perf_event_paranoid __read_mostly = 1;
164 /* Minimum for 512 kiB + 1 user control page */
165 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
168 * max perf event sample rate
170 #define DEFAULT_MAX_SAMPLE_RATE 100000
171 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
172 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
174 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
176 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
177 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
179 static int perf_sample_allowed_ns __read_mostly =
180 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
182 void update_perf_cpu_limits(void)
184 u64 tmp = perf_sample_period_ns;
186 tmp *= sysctl_perf_cpu_time_max_percent;
188 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
191 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
193 int perf_proc_update_handler(struct ctl_table *table, int write,
194 void __user *buffer, size_t *lenp,
197 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
202 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
203 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
204 update_perf_cpu_limits();
209 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
211 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
212 void __user *buffer, size_t *lenp,
215 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
220 update_perf_cpu_limits();
226 * perf samples are done in some very critical code paths (NMIs).
227 * If they take too much CPU time, the system can lock up and not
228 * get any real work done. This will drop the sample rate when
229 * we detect that events are taking too long.
231 #define NR_ACCUMULATED_SAMPLES 128
232 static DEFINE_PER_CPU(u64, running_sample_length);
234 void perf_sample_event_took(u64 sample_len_ns)
236 u64 avg_local_sample_len;
237 u64 local_samples_len;
238 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
243 /* decay the counter by 1 average sample */
244 local_samples_len = __get_cpu_var(running_sample_length);
245 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
246 local_samples_len += sample_len_ns;
247 __get_cpu_var(running_sample_length) = local_samples_len;
250 * note: this will be biased artifically low until we have
251 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
252 * from having to maintain a count.
254 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
256 if (avg_local_sample_len <= allowed_ns)
259 if (max_samples_per_tick <= 1)
262 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
263 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
264 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
266 printk_ratelimited(KERN_WARNING
267 "perf samples too long (%lld > %lld), lowering "
268 "kernel.perf_event_max_sample_rate to %d\n",
269 avg_local_sample_len, allowed_ns,
270 sysctl_perf_event_sample_rate);
272 update_perf_cpu_limits();
275 static atomic64_t perf_event_id;
277 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
278 enum event_type_t event_type);
280 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
281 enum event_type_t event_type,
282 struct task_struct *task);
284 static void update_context_time(struct perf_event_context *ctx);
285 static u64 perf_event_time(struct perf_event *event);
287 void __weak perf_event_print_debug(void) { }
289 extern __weak const char *perf_pmu_name(void)
294 static inline u64 perf_clock(void)
296 return local_clock();
299 static inline struct perf_cpu_context *
300 __get_cpu_context(struct perf_event_context *ctx)
302 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
305 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
306 struct perf_event_context *ctx)
308 raw_spin_lock(&cpuctx->ctx.lock);
310 raw_spin_lock(&ctx->lock);
313 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
314 struct perf_event_context *ctx)
317 raw_spin_unlock(&ctx->lock);
318 raw_spin_unlock(&cpuctx->ctx.lock);
321 #ifdef CONFIG_CGROUP_PERF
324 * perf_cgroup_info keeps track of time_enabled for a cgroup.
325 * This is a per-cpu dynamically allocated data structure.
327 struct perf_cgroup_info {
333 struct cgroup_subsys_state css;
334 struct perf_cgroup_info __percpu *info;
338 * Must ensure cgroup is pinned (css_get) before calling
339 * this function. In other words, we cannot call this function
340 * if there is no cgroup event for the current CPU context.
342 static inline struct perf_cgroup *
343 perf_cgroup_from_task(struct task_struct *task)
345 return container_of(task_css(task, perf_subsys_id),
346 struct perf_cgroup, css);
350 perf_cgroup_match(struct perf_event *event)
352 struct perf_event_context *ctx = event->ctx;
353 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
355 /* @event doesn't care about cgroup */
359 /* wants specific cgroup scope but @cpuctx isn't associated with any */
364 * Cgroup scoping is recursive. An event enabled for a cgroup is
365 * also enabled for all its descendant cgroups. If @cpuctx's
366 * cgroup is a descendant of @event's (the test covers identity
367 * case), it's a match.
369 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
370 event->cgrp->css.cgroup);
373 static inline bool perf_tryget_cgroup(struct perf_event *event)
375 return css_tryget(&event->cgrp->css);
378 static inline void perf_put_cgroup(struct perf_event *event)
380 css_put(&event->cgrp->css);
383 static inline void perf_detach_cgroup(struct perf_event *event)
385 perf_put_cgroup(event);
389 static inline int is_cgroup_event(struct perf_event *event)
391 return event->cgrp != NULL;
394 static inline u64 perf_cgroup_event_time(struct perf_event *event)
396 struct perf_cgroup_info *t;
398 t = per_cpu_ptr(event->cgrp->info, event->cpu);
402 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
404 struct perf_cgroup_info *info;
409 info = this_cpu_ptr(cgrp->info);
411 info->time += now - info->timestamp;
412 info->timestamp = now;
415 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
417 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
419 __update_cgrp_time(cgrp_out);
422 static inline void update_cgrp_time_from_event(struct perf_event *event)
424 struct perf_cgroup *cgrp;
427 * ensure we access cgroup data only when needed and
428 * when we know the cgroup is pinned (css_get)
430 if (!is_cgroup_event(event))
433 cgrp = perf_cgroup_from_task(current);
435 * Do not update time when cgroup is not active
437 if (cgrp == event->cgrp)
438 __update_cgrp_time(event->cgrp);
442 perf_cgroup_set_timestamp(struct task_struct *task,
443 struct perf_event_context *ctx)
445 struct perf_cgroup *cgrp;
446 struct perf_cgroup_info *info;
449 * ctx->lock held by caller
450 * ensure we do not access cgroup data
451 * unless we have the cgroup pinned (css_get)
453 if (!task || !ctx->nr_cgroups)
456 cgrp = perf_cgroup_from_task(task);
457 info = this_cpu_ptr(cgrp->info);
458 info->timestamp = ctx->timestamp;
461 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
462 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
465 * reschedule events based on the cgroup constraint of task.
467 * mode SWOUT : schedule out everything
468 * mode SWIN : schedule in based on cgroup for next
470 void perf_cgroup_switch(struct task_struct *task, int mode)
472 struct perf_cpu_context *cpuctx;
477 * disable interrupts to avoid geting nr_cgroup
478 * changes via __perf_event_disable(). Also
481 local_irq_save(flags);
484 * we reschedule only in the presence of cgroup
485 * constrained events.
489 list_for_each_entry_rcu(pmu, &pmus, entry) {
490 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
491 if (cpuctx->unique_pmu != pmu)
492 continue; /* ensure we process each cpuctx once */
495 * perf_cgroup_events says at least one
496 * context on this CPU has cgroup events.
498 * ctx->nr_cgroups reports the number of cgroup
499 * events for a context.
501 if (cpuctx->ctx.nr_cgroups > 0) {
502 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
503 perf_pmu_disable(cpuctx->ctx.pmu);
505 if (mode & PERF_CGROUP_SWOUT) {
506 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
508 * must not be done before ctxswout due
509 * to event_filter_match() in event_sched_out()
514 if (mode & PERF_CGROUP_SWIN) {
515 WARN_ON_ONCE(cpuctx->cgrp);
517 * set cgrp before ctxsw in to allow
518 * event_filter_match() to not have to pass
521 cpuctx->cgrp = perf_cgroup_from_task(task);
522 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
524 perf_pmu_enable(cpuctx->ctx.pmu);
525 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
531 local_irq_restore(flags);
534 static inline void perf_cgroup_sched_out(struct task_struct *task,
535 struct task_struct *next)
537 struct perf_cgroup *cgrp1;
538 struct perf_cgroup *cgrp2 = NULL;
541 * we come here when we know perf_cgroup_events > 0
543 cgrp1 = perf_cgroup_from_task(task);
546 * next is NULL when called from perf_event_enable_on_exec()
547 * that will systematically cause a cgroup_switch()
550 cgrp2 = perf_cgroup_from_task(next);
553 * only schedule out current cgroup events if we know
554 * that we are switching to a different cgroup. Otherwise,
555 * do no touch the cgroup events.
558 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
561 static inline void perf_cgroup_sched_in(struct task_struct *prev,
562 struct task_struct *task)
564 struct perf_cgroup *cgrp1;
565 struct perf_cgroup *cgrp2 = NULL;
568 * we come here when we know perf_cgroup_events > 0
570 cgrp1 = perf_cgroup_from_task(task);
572 /* prev can never be NULL */
573 cgrp2 = perf_cgroup_from_task(prev);
576 * only need to schedule in cgroup events if we are changing
577 * cgroup during ctxsw. Cgroup events were not scheduled
578 * out of ctxsw out if that was not the case.
581 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
584 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
585 struct perf_event_attr *attr,
586 struct perf_event *group_leader)
588 struct perf_cgroup *cgrp;
589 struct cgroup_subsys_state *css;
590 struct fd f = fdget(fd);
598 css = css_from_dir(f.file->f_dentry, &perf_subsys);
604 cgrp = container_of(css, struct perf_cgroup, css);
607 /* must be done before we fput() the file */
608 if (!perf_tryget_cgroup(event)) {
615 * all events in a group must monitor
616 * the same cgroup because a task belongs
617 * to only one perf cgroup at a time
619 if (group_leader && group_leader->cgrp != cgrp) {
620 perf_detach_cgroup(event);
630 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
632 struct perf_cgroup_info *t;
633 t = per_cpu_ptr(event->cgrp->info, event->cpu);
634 event->shadow_ctx_time = now - t->timestamp;
638 perf_cgroup_defer_enabled(struct perf_event *event)
641 * when the current task's perf cgroup does not match
642 * the event's, we need to remember to call the
643 * perf_mark_enable() function the first time a task with
644 * a matching perf cgroup is scheduled in.
646 if (is_cgroup_event(event) && !perf_cgroup_match(event))
647 event->cgrp_defer_enabled = 1;
651 perf_cgroup_mark_enabled(struct perf_event *event,
652 struct perf_event_context *ctx)
654 struct perf_event *sub;
655 u64 tstamp = perf_event_time(event);
657 if (!event->cgrp_defer_enabled)
660 event->cgrp_defer_enabled = 0;
662 event->tstamp_enabled = tstamp - event->total_time_enabled;
663 list_for_each_entry(sub, &event->sibling_list, group_entry) {
664 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
665 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
666 sub->cgrp_defer_enabled = 0;
670 #else /* !CONFIG_CGROUP_PERF */
673 perf_cgroup_match(struct perf_event *event)
678 static inline void perf_detach_cgroup(struct perf_event *event)
681 static inline int is_cgroup_event(struct perf_event *event)
686 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
691 static inline void update_cgrp_time_from_event(struct perf_event *event)
695 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
699 static inline void perf_cgroup_sched_out(struct task_struct *task,
700 struct task_struct *next)
704 static inline void perf_cgroup_sched_in(struct task_struct *prev,
705 struct task_struct *task)
709 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
710 struct perf_event_attr *attr,
711 struct perf_event *group_leader)
717 perf_cgroup_set_timestamp(struct task_struct *task,
718 struct perf_event_context *ctx)
723 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
728 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
732 static inline u64 perf_cgroup_event_time(struct perf_event *event)
738 perf_cgroup_defer_enabled(struct perf_event *event)
743 perf_cgroup_mark_enabled(struct perf_event *event,
744 struct perf_event_context *ctx)
750 * set default to be dependent on timer tick just
753 #define PERF_CPU_HRTIMER (1000 / HZ)
755 * function must be called with interrupts disbled
757 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
759 struct perf_cpu_context *cpuctx;
760 enum hrtimer_restart ret = HRTIMER_NORESTART;
763 WARN_ON(!irqs_disabled());
765 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
767 rotations = perf_rotate_context(cpuctx);
770 * arm timer if needed
773 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
774 ret = HRTIMER_RESTART;
780 /* CPU is going down */
781 void perf_cpu_hrtimer_cancel(int cpu)
783 struct perf_cpu_context *cpuctx;
787 if (WARN_ON(cpu != smp_processor_id()))
790 local_irq_save(flags);
794 list_for_each_entry_rcu(pmu, &pmus, entry) {
795 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
797 if (pmu->task_ctx_nr == perf_sw_context)
800 hrtimer_cancel(&cpuctx->hrtimer);
805 local_irq_restore(flags);
808 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
810 struct hrtimer *hr = &cpuctx->hrtimer;
811 struct pmu *pmu = cpuctx->ctx.pmu;
814 /* no multiplexing needed for SW PMU */
815 if (pmu->task_ctx_nr == perf_sw_context)
819 * check default is sane, if not set then force to
820 * default interval (1/tick)
822 timer = pmu->hrtimer_interval_ms;
824 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
826 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
828 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
829 hr->function = perf_cpu_hrtimer_handler;
832 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
834 struct hrtimer *hr = &cpuctx->hrtimer;
835 struct pmu *pmu = cpuctx->ctx.pmu;
838 if (pmu->task_ctx_nr == perf_sw_context)
841 if (hrtimer_active(hr))
844 if (!hrtimer_callback_running(hr))
845 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
846 0, HRTIMER_MODE_REL_PINNED, 0);
849 void perf_pmu_disable(struct pmu *pmu)
851 int *count = this_cpu_ptr(pmu->pmu_disable_count);
853 pmu->pmu_disable(pmu);
856 void perf_pmu_enable(struct pmu *pmu)
858 int *count = this_cpu_ptr(pmu->pmu_disable_count);
860 pmu->pmu_enable(pmu);
863 static DEFINE_PER_CPU(struct list_head, rotation_list);
866 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
867 * because they're strictly cpu affine and rotate_start is called with IRQs
868 * disabled, while rotate_context is called from IRQ context.
870 static void perf_pmu_rotate_start(struct pmu *pmu)
872 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
873 struct list_head *head = &__get_cpu_var(rotation_list);
875 WARN_ON(!irqs_disabled());
877 if (list_empty(&cpuctx->rotation_list))
878 list_add(&cpuctx->rotation_list, head);
881 static void get_ctx(struct perf_event_context *ctx)
883 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
886 static void put_ctx(struct perf_event_context *ctx)
888 if (atomic_dec_and_test(&ctx->refcount)) {
890 put_ctx(ctx->parent_ctx);
892 put_task_struct(ctx->task);
893 kfree_rcu(ctx, rcu_head);
897 static void unclone_ctx(struct perf_event_context *ctx)
899 if (ctx->parent_ctx) {
900 put_ctx(ctx->parent_ctx);
901 ctx->parent_ctx = NULL;
906 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
909 * only top level events have the pid namespace they were created in
912 event = event->parent;
914 return task_tgid_nr_ns(p, event->ns);
917 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
920 * only top level events have the pid namespace they were created in
923 event = event->parent;
925 return task_pid_nr_ns(p, event->ns);
929 * If we inherit events we want to return the parent event id
932 static u64 primary_event_id(struct perf_event *event)
937 id = event->parent->id;
943 * Get the perf_event_context for a task and lock it.
944 * This has to cope with with the fact that until it is locked,
945 * the context could get moved to another task.
947 static struct perf_event_context *
948 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
950 struct perf_event_context *ctx;
954 * One of the few rules of preemptible RCU is that one cannot do
955 * rcu_read_unlock() while holding a scheduler (or nested) lock when
956 * part of the read side critical section was preemptible -- see
957 * rcu_read_unlock_special().
959 * Since ctx->lock nests under rq->lock we must ensure the entire read
960 * side critical section is non-preemptible.
964 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
967 * If this context is a clone of another, it might
968 * get swapped for another underneath us by
969 * perf_event_task_sched_out, though the
970 * rcu_read_lock() protects us from any context
971 * getting freed. Lock the context and check if it
972 * got swapped before we could get the lock, and retry
973 * if so. If we locked the right context, then it
974 * can't get swapped on us any more.
976 raw_spin_lock_irqsave(&ctx->lock, *flags);
977 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
978 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
984 if (!atomic_inc_not_zero(&ctx->refcount)) {
985 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
995 * Get the context for a task and increment its pin_count so it
996 * can't get swapped to another task. This also increments its
997 * reference count so that the context can't get freed.
999 static struct perf_event_context *
1000 perf_pin_task_context(struct task_struct *task, int ctxn)
1002 struct perf_event_context *ctx;
1003 unsigned long flags;
1005 ctx = perf_lock_task_context(task, ctxn, &flags);
1008 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1013 static void perf_unpin_context(struct perf_event_context *ctx)
1015 unsigned long flags;
1017 raw_spin_lock_irqsave(&ctx->lock, flags);
1019 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1023 * Update the record of the current time in a context.
1025 static void update_context_time(struct perf_event_context *ctx)
1027 u64 now = perf_clock();
1029 ctx->time += now - ctx->timestamp;
1030 ctx->timestamp = now;
1033 static u64 perf_event_time(struct perf_event *event)
1035 struct perf_event_context *ctx = event->ctx;
1037 if (is_cgroup_event(event))
1038 return perf_cgroup_event_time(event);
1040 return ctx ? ctx->time : 0;
1044 * Update the total_time_enabled and total_time_running fields for a event.
1045 * The caller of this function needs to hold the ctx->lock.
1047 static void update_event_times(struct perf_event *event)
1049 struct perf_event_context *ctx = event->ctx;
1052 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1053 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1056 * in cgroup mode, time_enabled represents
1057 * the time the event was enabled AND active
1058 * tasks were in the monitored cgroup. This is
1059 * independent of the activity of the context as
1060 * there may be a mix of cgroup and non-cgroup events.
1062 * That is why we treat cgroup events differently
1065 if (is_cgroup_event(event))
1066 run_end = perf_cgroup_event_time(event);
1067 else if (ctx->is_active)
1068 run_end = ctx->time;
1070 run_end = event->tstamp_stopped;
1072 event->total_time_enabled = run_end - event->tstamp_enabled;
1074 if (event->state == PERF_EVENT_STATE_INACTIVE)
1075 run_end = event->tstamp_stopped;
1077 run_end = perf_event_time(event);
1079 event->total_time_running = run_end - event->tstamp_running;
1084 * Update total_time_enabled and total_time_running for all events in a group.
1086 static void update_group_times(struct perf_event *leader)
1088 struct perf_event *event;
1090 update_event_times(leader);
1091 list_for_each_entry(event, &leader->sibling_list, group_entry)
1092 update_event_times(event);
1095 static struct list_head *
1096 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1098 if (event->attr.pinned)
1099 return &ctx->pinned_groups;
1101 return &ctx->flexible_groups;
1105 * Add a event from the lists for its context.
1106 * Must be called with ctx->mutex and ctx->lock held.
1109 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1111 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1112 event->attach_state |= PERF_ATTACH_CONTEXT;
1115 * If we're a stand alone event or group leader, we go to the context
1116 * list, group events are kept attached to the group so that
1117 * perf_group_detach can, at all times, locate all siblings.
1119 if (event->group_leader == event) {
1120 struct list_head *list;
1122 if (is_software_event(event))
1123 event->group_flags |= PERF_GROUP_SOFTWARE;
1125 list = ctx_group_list(event, ctx);
1126 list_add_tail(&event->group_entry, list);
1129 if (is_cgroup_event(event))
1132 if (has_branch_stack(event))
1133 ctx->nr_branch_stack++;
1135 list_add_rcu(&event->event_entry, &ctx->event_list);
1136 if (!ctx->nr_events)
1137 perf_pmu_rotate_start(ctx->pmu);
1139 if (event->attr.inherit_stat)
1146 * Initialize event state based on the perf_event_attr::disabled.
1148 static inline void perf_event__state_init(struct perf_event *event)
1150 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1151 PERF_EVENT_STATE_INACTIVE;
1155 * Called at perf_event creation and when events are attached/detached from a
1158 static void perf_event__read_size(struct perf_event *event)
1160 int entry = sizeof(u64); /* value */
1164 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1165 size += sizeof(u64);
1167 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1168 size += sizeof(u64);
1170 if (event->attr.read_format & PERF_FORMAT_ID)
1171 entry += sizeof(u64);
1173 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1174 nr += event->group_leader->nr_siblings;
1175 size += sizeof(u64);
1179 event->read_size = size;
1182 static void perf_event__header_size(struct perf_event *event)
1184 struct perf_sample_data *data;
1185 u64 sample_type = event->attr.sample_type;
1188 perf_event__read_size(event);
1190 if (sample_type & PERF_SAMPLE_IP)
1191 size += sizeof(data->ip);
1193 if (sample_type & PERF_SAMPLE_ADDR)
1194 size += sizeof(data->addr);
1196 if (sample_type & PERF_SAMPLE_PERIOD)
1197 size += sizeof(data->period);
1199 if (sample_type & PERF_SAMPLE_WEIGHT)
1200 size += sizeof(data->weight);
1202 if (sample_type & PERF_SAMPLE_READ)
1203 size += event->read_size;
1205 if (sample_type & PERF_SAMPLE_DATA_SRC)
1206 size += sizeof(data->data_src.val);
1208 if (sample_type & PERF_SAMPLE_TRANSACTION)
1209 size += sizeof(data->txn);
1211 event->header_size = size;
1214 static void perf_event__id_header_size(struct perf_event *event)
1216 struct perf_sample_data *data;
1217 u64 sample_type = event->attr.sample_type;
1220 if (sample_type & PERF_SAMPLE_TID)
1221 size += sizeof(data->tid_entry);
1223 if (sample_type & PERF_SAMPLE_TIME)
1224 size += sizeof(data->time);
1226 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1227 size += sizeof(data->id);
1229 if (sample_type & PERF_SAMPLE_ID)
1230 size += sizeof(data->id);
1232 if (sample_type & PERF_SAMPLE_STREAM_ID)
1233 size += sizeof(data->stream_id);
1235 if (sample_type & PERF_SAMPLE_CPU)
1236 size += sizeof(data->cpu_entry);
1238 event->id_header_size = size;
1241 static void perf_group_attach(struct perf_event *event)
1243 struct perf_event *group_leader = event->group_leader, *pos;
1246 * We can have double attach due to group movement in perf_event_open.
1248 if (event->attach_state & PERF_ATTACH_GROUP)
1251 event->attach_state |= PERF_ATTACH_GROUP;
1253 if (group_leader == event)
1256 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1257 !is_software_event(event))
1258 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1260 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1261 group_leader->nr_siblings++;
1263 perf_event__header_size(group_leader);
1265 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1266 perf_event__header_size(pos);
1270 * Remove a event from the lists for its context.
1271 * Must be called with ctx->mutex and ctx->lock held.
1274 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1276 struct perf_cpu_context *cpuctx;
1278 * We can have double detach due to exit/hot-unplug + close.
1280 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1283 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1285 if (is_cgroup_event(event)) {
1287 cpuctx = __get_cpu_context(ctx);
1289 * if there are no more cgroup events
1290 * then cler cgrp to avoid stale pointer
1291 * in update_cgrp_time_from_cpuctx()
1293 if (!ctx->nr_cgroups)
1294 cpuctx->cgrp = NULL;
1297 if (has_branch_stack(event))
1298 ctx->nr_branch_stack--;
1301 if (event->attr.inherit_stat)
1304 list_del_rcu(&event->event_entry);
1306 if (event->group_leader == event)
1307 list_del_init(&event->group_entry);
1309 update_group_times(event);
1312 * If event was in error state, then keep it
1313 * that way, otherwise bogus counts will be
1314 * returned on read(). The only way to get out
1315 * of error state is by explicit re-enabling
1318 if (event->state > PERF_EVENT_STATE_OFF)
1319 event->state = PERF_EVENT_STATE_OFF;
1324 static void perf_group_detach(struct perf_event *event)
1326 struct perf_event *sibling, *tmp;
1327 struct list_head *list = NULL;
1330 * We can have double detach due to exit/hot-unplug + close.
1332 if (!(event->attach_state & PERF_ATTACH_GROUP))
1335 event->attach_state &= ~PERF_ATTACH_GROUP;
1338 * If this is a sibling, remove it from its group.
1340 if (event->group_leader != event) {
1341 list_del_init(&event->group_entry);
1342 event->group_leader->nr_siblings--;
1346 if (!list_empty(&event->group_entry))
1347 list = &event->group_entry;
1350 * If this was a group event with sibling events then
1351 * upgrade the siblings to singleton events by adding them
1352 * to whatever list we are on.
1354 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1356 list_move_tail(&sibling->group_entry, list);
1357 sibling->group_leader = sibling;
1359 /* Inherit group flags from the previous leader */
1360 sibling->group_flags = event->group_flags;
1364 perf_event__header_size(event->group_leader);
1366 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1367 perf_event__header_size(tmp);
1371 event_filter_match(struct perf_event *event)
1373 return (event->cpu == -1 || event->cpu == smp_processor_id())
1374 && perf_cgroup_match(event);
1378 event_sched_out(struct perf_event *event,
1379 struct perf_cpu_context *cpuctx,
1380 struct perf_event_context *ctx)
1382 u64 tstamp = perf_event_time(event);
1385 * An event which could not be activated because of
1386 * filter mismatch still needs to have its timings
1387 * maintained, otherwise bogus information is return
1388 * via read() for time_enabled, time_running:
1390 if (event->state == PERF_EVENT_STATE_INACTIVE
1391 && !event_filter_match(event)) {
1392 delta = tstamp - event->tstamp_stopped;
1393 event->tstamp_running += delta;
1394 event->tstamp_stopped = tstamp;
1397 if (event->state != PERF_EVENT_STATE_ACTIVE)
1400 perf_pmu_disable(event->pmu);
1402 event->state = PERF_EVENT_STATE_INACTIVE;
1403 if (event->pending_disable) {
1404 event->pending_disable = 0;
1405 event->state = PERF_EVENT_STATE_OFF;
1407 event->tstamp_stopped = tstamp;
1408 event->pmu->del(event, 0);
1411 if (!is_software_event(event))
1412 cpuctx->active_oncpu--;
1414 if (event->attr.freq && event->attr.sample_freq)
1416 if (event->attr.exclusive || !cpuctx->active_oncpu)
1417 cpuctx->exclusive = 0;
1419 perf_pmu_enable(event->pmu);
1423 group_sched_out(struct perf_event *group_event,
1424 struct perf_cpu_context *cpuctx,
1425 struct perf_event_context *ctx)
1427 struct perf_event *event;
1428 int state = group_event->state;
1430 event_sched_out(group_event, cpuctx, ctx);
1433 * Schedule out siblings (if any):
1435 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1436 event_sched_out(event, cpuctx, ctx);
1438 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1439 cpuctx->exclusive = 0;
1442 struct remove_event {
1443 struct perf_event *event;
1448 * Cross CPU call to remove a performance event
1450 * We disable the event on the hardware level first. After that we
1451 * remove it from the context list.
1453 static int __perf_remove_from_context(void *info)
1455 struct remove_event *re = info;
1456 struct perf_event *event = re->event;
1457 struct perf_event_context *ctx = event->ctx;
1458 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1460 raw_spin_lock(&ctx->lock);
1461 event_sched_out(event, cpuctx, ctx);
1462 if (re->detach_group)
1463 perf_group_detach(event);
1464 list_del_event(event, ctx);
1465 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1467 cpuctx->task_ctx = NULL;
1469 raw_spin_unlock(&ctx->lock);
1476 * Remove the event from a task's (or a CPU's) list of events.
1478 * CPU events are removed with a smp call. For task events we only
1479 * call when the task is on a CPU.
1481 * If event->ctx is a cloned context, callers must make sure that
1482 * every task struct that event->ctx->task could possibly point to
1483 * remains valid. This is OK when called from perf_release since
1484 * that only calls us on the top-level context, which can't be a clone.
1485 * When called from perf_event_exit_task, it's OK because the
1486 * context has been detached from its task.
1488 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1490 struct perf_event_context *ctx = event->ctx;
1491 struct task_struct *task = ctx->task;
1492 struct remove_event re = {
1494 .detach_group = detach_group,
1497 lockdep_assert_held(&ctx->mutex);
1501 * Per cpu events are removed via an smp call and
1502 * the removal is always successful.
1504 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1509 if (!task_function_call(task, __perf_remove_from_context, &re))
1512 raw_spin_lock_irq(&ctx->lock);
1514 * If we failed to find a running task, but find the context active now
1515 * that we've acquired the ctx->lock, retry.
1517 if (ctx->is_active) {
1518 raw_spin_unlock_irq(&ctx->lock);
1523 * Since the task isn't running, its safe to remove the event, us
1524 * holding the ctx->lock ensures the task won't get scheduled in.
1527 perf_group_detach(event);
1528 list_del_event(event, ctx);
1529 raw_spin_unlock_irq(&ctx->lock);
1533 * Cross CPU call to disable a performance event
1535 int __perf_event_disable(void *info)
1537 struct perf_event *event = info;
1538 struct perf_event_context *ctx = event->ctx;
1539 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1542 * If this is a per-task event, need to check whether this
1543 * event's task is the current task on this cpu.
1545 * Can trigger due to concurrent perf_event_context_sched_out()
1546 * flipping contexts around.
1548 if (ctx->task && cpuctx->task_ctx != ctx)
1551 raw_spin_lock(&ctx->lock);
1554 * If the event is on, turn it off.
1555 * If it is in error state, leave it in error state.
1557 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1558 update_context_time(ctx);
1559 update_cgrp_time_from_event(event);
1560 update_group_times(event);
1561 if (event == event->group_leader)
1562 group_sched_out(event, cpuctx, ctx);
1564 event_sched_out(event, cpuctx, ctx);
1565 event->state = PERF_EVENT_STATE_OFF;
1568 raw_spin_unlock(&ctx->lock);
1576 * If event->ctx is a cloned context, callers must make sure that
1577 * every task struct that event->ctx->task could possibly point to
1578 * remains valid. This condition is satisifed when called through
1579 * perf_event_for_each_child or perf_event_for_each because they
1580 * hold the top-level event's child_mutex, so any descendant that
1581 * goes to exit will block in sync_child_event.
1582 * When called from perf_pending_event it's OK because event->ctx
1583 * is the current context on this CPU and preemption is disabled,
1584 * hence we can't get into perf_event_task_sched_out for this context.
1586 void perf_event_disable(struct perf_event *event)
1588 struct perf_event_context *ctx = event->ctx;
1589 struct task_struct *task = ctx->task;
1593 * Disable the event on the cpu that it's on
1595 cpu_function_call(event->cpu, __perf_event_disable, event);
1600 if (!task_function_call(task, __perf_event_disable, event))
1603 raw_spin_lock_irq(&ctx->lock);
1605 * If the event is still active, we need to retry the cross-call.
1607 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1608 raw_spin_unlock_irq(&ctx->lock);
1610 * Reload the task pointer, it might have been changed by
1611 * a concurrent perf_event_context_sched_out().
1618 * Since we have the lock this context can't be scheduled
1619 * in, so we can change the state safely.
1621 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1622 update_group_times(event);
1623 event->state = PERF_EVENT_STATE_OFF;
1625 raw_spin_unlock_irq(&ctx->lock);
1627 EXPORT_SYMBOL_GPL(perf_event_disable);
1629 static void perf_set_shadow_time(struct perf_event *event,
1630 struct perf_event_context *ctx,
1634 * use the correct time source for the time snapshot
1636 * We could get by without this by leveraging the
1637 * fact that to get to this function, the caller
1638 * has most likely already called update_context_time()
1639 * and update_cgrp_time_xx() and thus both timestamp
1640 * are identical (or very close). Given that tstamp is,
1641 * already adjusted for cgroup, we could say that:
1642 * tstamp - ctx->timestamp
1644 * tstamp - cgrp->timestamp.
1646 * Then, in perf_output_read(), the calculation would
1647 * work with no changes because:
1648 * - event is guaranteed scheduled in
1649 * - no scheduled out in between
1650 * - thus the timestamp would be the same
1652 * But this is a bit hairy.
1654 * So instead, we have an explicit cgroup call to remain
1655 * within the time time source all along. We believe it
1656 * is cleaner and simpler to understand.
1658 if (is_cgroup_event(event))
1659 perf_cgroup_set_shadow_time(event, tstamp);
1661 event->shadow_ctx_time = tstamp - ctx->timestamp;
1664 #define MAX_INTERRUPTS (~0ULL)
1666 static void perf_log_throttle(struct perf_event *event, int enable);
1669 event_sched_in(struct perf_event *event,
1670 struct perf_cpu_context *cpuctx,
1671 struct perf_event_context *ctx)
1673 u64 tstamp = perf_event_time(event);
1676 if (event->state <= PERF_EVENT_STATE_OFF)
1679 event->state = PERF_EVENT_STATE_ACTIVE;
1680 event->oncpu = smp_processor_id();
1683 * Unthrottle events, since we scheduled we might have missed several
1684 * ticks already, also for a heavily scheduling task there is little
1685 * guarantee it'll get a tick in a timely manner.
1687 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1688 perf_log_throttle(event, 1);
1689 event->hw.interrupts = 0;
1693 * The new state must be visible before we turn it on in the hardware:
1697 perf_pmu_disable(event->pmu);
1699 if (event->pmu->add(event, PERF_EF_START)) {
1700 event->state = PERF_EVENT_STATE_INACTIVE;
1706 event->tstamp_running += tstamp - event->tstamp_stopped;
1708 perf_set_shadow_time(event, ctx, tstamp);
1710 if (!is_software_event(event))
1711 cpuctx->active_oncpu++;
1713 if (event->attr.freq && event->attr.sample_freq)
1716 if (event->attr.exclusive)
1717 cpuctx->exclusive = 1;
1720 perf_pmu_enable(event->pmu);
1726 group_sched_in(struct perf_event *group_event,
1727 struct perf_cpu_context *cpuctx,
1728 struct perf_event_context *ctx)
1730 struct perf_event *event, *partial_group = NULL;
1731 struct pmu *pmu = group_event->pmu;
1732 u64 now = ctx->time;
1733 bool simulate = false;
1735 if (group_event->state == PERF_EVENT_STATE_OFF)
1738 pmu->start_txn(pmu);
1740 if (event_sched_in(group_event, cpuctx, ctx)) {
1741 pmu->cancel_txn(pmu);
1742 perf_cpu_hrtimer_restart(cpuctx);
1747 * Schedule in siblings as one group (if any):
1749 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1750 if (event_sched_in(event, cpuctx, ctx)) {
1751 partial_group = event;
1756 if (!pmu->commit_txn(pmu))
1761 * Groups can be scheduled in as one unit only, so undo any
1762 * partial group before returning:
1763 * The events up to the failed event are scheduled out normally,
1764 * tstamp_stopped will be updated.
1766 * The failed events and the remaining siblings need to have
1767 * their timings updated as if they had gone thru event_sched_in()
1768 * and event_sched_out(). This is required to get consistent timings
1769 * across the group. This also takes care of the case where the group
1770 * could never be scheduled by ensuring tstamp_stopped is set to mark
1771 * the time the event was actually stopped, such that time delta
1772 * calculation in update_event_times() is correct.
1774 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1775 if (event == partial_group)
1779 event->tstamp_running += now - event->tstamp_stopped;
1780 event->tstamp_stopped = now;
1782 event_sched_out(event, cpuctx, ctx);
1785 event_sched_out(group_event, cpuctx, ctx);
1787 pmu->cancel_txn(pmu);
1789 perf_cpu_hrtimer_restart(cpuctx);
1795 * Work out whether we can put this event group on the CPU now.
1797 static int group_can_go_on(struct perf_event *event,
1798 struct perf_cpu_context *cpuctx,
1802 * Groups consisting entirely of software events can always go on.
1804 if (event->group_flags & PERF_GROUP_SOFTWARE)
1807 * If an exclusive group is already on, no other hardware
1810 if (cpuctx->exclusive)
1813 * If this group is exclusive and there are already
1814 * events on the CPU, it can't go on.
1816 if (event->attr.exclusive && cpuctx->active_oncpu)
1819 * Otherwise, try to add it if all previous groups were able
1825 static void add_event_to_ctx(struct perf_event *event,
1826 struct perf_event_context *ctx)
1828 u64 tstamp = perf_event_time(event);
1830 list_add_event(event, ctx);
1831 perf_group_attach(event);
1832 event->tstamp_enabled = tstamp;
1833 event->tstamp_running = tstamp;
1834 event->tstamp_stopped = tstamp;
1837 static void task_ctx_sched_out(struct perf_event_context *ctx);
1839 ctx_sched_in(struct perf_event_context *ctx,
1840 struct perf_cpu_context *cpuctx,
1841 enum event_type_t event_type,
1842 struct task_struct *task);
1844 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1845 struct perf_event_context *ctx,
1846 struct task_struct *task)
1848 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1850 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1851 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1853 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1857 * Cross CPU call to install and enable a performance event
1859 * Must be called with ctx->mutex held
1861 static int __perf_install_in_context(void *info)
1863 struct perf_event *event = info;
1864 struct perf_event_context *ctx = event->ctx;
1865 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1866 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1867 struct task_struct *task = current;
1869 perf_ctx_lock(cpuctx, task_ctx);
1870 perf_pmu_disable(cpuctx->ctx.pmu);
1873 * If there was an active task_ctx schedule it out.
1876 task_ctx_sched_out(task_ctx);
1879 * If the context we're installing events in is not the
1880 * active task_ctx, flip them.
1882 if (ctx->task && task_ctx != ctx) {
1884 raw_spin_unlock(&task_ctx->lock);
1885 raw_spin_lock(&ctx->lock);
1890 cpuctx->task_ctx = task_ctx;
1891 task = task_ctx->task;
1894 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1896 update_context_time(ctx);
1898 * update cgrp time only if current cgrp
1899 * matches event->cgrp. Must be done before
1900 * calling add_event_to_ctx()
1902 update_cgrp_time_from_event(event);
1904 add_event_to_ctx(event, ctx);
1907 * Schedule everything back in
1909 perf_event_sched_in(cpuctx, task_ctx, task);
1911 perf_pmu_enable(cpuctx->ctx.pmu);
1912 perf_ctx_unlock(cpuctx, task_ctx);
1918 * Attach a performance event to a context
1920 * First we add the event to the list with the hardware enable bit
1921 * in event->hw_config cleared.
1923 * If the event is attached to a task which is on a CPU we use a smp
1924 * call to enable it in the task context. The task might have been
1925 * scheduled away, but we check this in the smp call again.
1928 perf_install_in_context(struct perf_event_context *ctx,
1929 struct perf_event *event,
1932 struct task_struct *task = ctx->task;
1934 lockdep_assert_held(&ctx->mutex);
1937 if (event->cpu != -1)
1942 * Per cpu events are installed via an smp call and
1943 * the install is always successful.
1945 cpu_function_call(cpu, __perf_install_in_context, event);
1950 if (!task_function_call(task, __perf_install_in_context, event))
1953 raw_spin_lock_irq(&ctx->lock);
1955 * If we failed to find a running task, but find the context active now
1956 * that we've acquired the ctx->lock, retry.
1958 if (ctx->is_active) {
1959 raw_spin_unlock_irq(&ctx->lock);
1964 * Since the task isn't running, its safe to add the event, us holding
1965 * the ctx->lock ensures the task won't get scheduled in.
1967 add_event_to_ctx(event, ctx);
1968 raw_spin_unlock_irq(&ctx->lock);
1972 * Put a event into inactive state and update time fields.
1973 * Enabling the leader of a group effectively enables all
1974 * the group members that aren't explicitly disabled, so we
1975 * have to update their ->tstamp_enabled also.
1976 * Note: this works for group members as well as group leaders
1977 * since the non-leader members' sibling_lists will be empty.
1979 static void __perf_event_mark_enabled(struct perf_event *event)
1981 struct perf_event *sub;
1982 u64 tstamp = perf_event_time(event);
1984 event->state = PERF_EVENT_STATE_INACTIVE;
1985 event->tstamp_enabled = tstamp - event->total_time_enabled;
1986 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1987 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1988 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1993 * Cross CPU call to enable a performance event
1995 static int __perf_event_enable(void *info)
1997 struct perf_event *event = info;
1998 struct perf_event_context *ctx = event->ctx;
1999 struct perf_event *leader = event->group_leader;
2000 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2004 * There's a time window between 'ctx->is_active' check
2005 * in perf_event_enable function and this place having:
2007 * - ctx->lock unlocked
2009 * where the task could be killed and 'ctx' deactivated
2010 * by perf_event_exit_task.
2012 if (!ctx->is_active)
2015 raw_spin_lock(&ctx->lock);
2016 update_context_time(ctx);
2018 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2022 * set current task's cgroup time reference point
2024 perf_cgroup_set_timestamp(current, ctx);
2026 __perf_event_mark_enabled(event);
2028 if (!event_filter_match(event)) {
2029 if (is_cgroup_event(event))
2030 perf_cgroup_defer_enabled(event);
2035 * If the event is in a group and isn't the group leader,
2036 * then don't put it on unless the group is on.
2038 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2041 if (!group_can_go_on(event, cpuctx, 1)) {
2044 if (event == leader)
2045 err = group_sched_in(event, cpuctx, ctx);
2047 err = event_sched_in(event, cpuctx, ctx);
2052 * If this event can't go on and it's part of a
2053 * group, then the whole group has to come off.
2055 if (leader != event) {
2056 group_sched_out(leader, cpuctx, ctx);
2057 perf_cpu_hrtimer_restart(cpuctx);
2059 if (leader->attr.pinned) {
2060 update_group_times(leader);
2061 leader->state = PERF_EVENT_STATE_ERROR;
2066 raw_spin_unlock(&ctx->lock);
2074 * If event->ctx is a cloned context, callers must make sure that
2075 * every task struct that event->ctx->task could possibly point to
2076 * remains valid. This condition is satisfied when called through
2077 * perf_event_for_each_child or perf_event_for_each as described
2078 * for perf_event_disable.
2080 void perf_event_enable(struct perf_event *event)
2082 struct perf_event_context *ctx = event->ctx;
2083 struct task_struct *task = ctx->task;
2087 * Enable the event on the cpu that it's on
2089 cpu_function_call(event->cpu, __perf_event_enable, event);
2093 raw_spin_lock_irq(&ctx->lock);
2094 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2098 * If the event is in error state, clear that first.
2099 * That way, if we see the event in error state below, we
2100 * know that it has gone back into error state, as distinct
2101 * from the task having been scheduled away before the
2102 * cross-call arrived.
2104 if (event->state == PERF_EVENT_STATE_ERROR)
2105 event->state = PERF_EVENT_STATE_OFF;
2108 if (!ctx->is_active) {
2109 __perf_event_mark_enabled(event);
2113 raw_spin_unlock_irq(&ctx->lock);
2115 if (!task_function_call(task, __perf_event_enable, event))
2118 raw_spin_lock_irq(&ctx->lock);
2121 * If the context is active and the event is still off,
2122 * we need to retry the cross-call.
2124 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2126 * task could have been flipped by a concurrent
2127 * perf_event_context_sched_out()
2134 raw_spin_unlock_irq(&ctx->lock);
2136 EXPORT_SYMBOL_GPL(perf_event_enable);
2138 int perf_event_refresh(struct perf_event *event, int refresh)
2141 * not supported on inherited events
2143 if (event->attr.inherit || !is_sampling_event(event))
2146 atomic_add(refresh, &event->event_limit);
2147 perf_event_enable(event);
2151 EXPORT_SYMBOL_GPL(perf_event_refresh);
2153 static void ctx_sched_out(struct perf_event_context *ctx,
2154 struct perf_cpu_context *cpuctx,
2155 enum event_type_t event_type)
2157 struct perf_event *event;
2158 int is_active = ctx->is_active;
2160 ctx->is_active &= ~event_type;
2161 if (likely(!ctx->nr_events))
2164 update_context_time(ctx);
2165 update_cgrp_time_from_cpuctx(cpuctx);
2166 if (!ctx->nr_active)
2169 perf_pmu_disable(ctx->pmu);
2170 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2171 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2172 group_sched_out(event, cpuctx, ctx);
2175 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2176 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2177 group_sched_out(event, cpuctx, ctx);
2179 perf_pmu_enable(ctx->pmu);
2183 * Test whether two contexts are equivalent, i.e. whether they have both been
2184 * cloned from the same version of the same context.
2186 * Equivalence is measured using a generation number in the context that is
2187 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2188 * and list_del_event().
2190 static int context_equiv(struct perf_event_context *ctx1,
2191 struct perf_event_context *ctx2)
2193 /* Pinning disables the swap optimization */
2194 if (ctx1->pin_count || ctx2->pin_count)
2197 /* If ctx1 is the parent of ctx2 */
2198 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2201 /* If ctx2 is the parent of ctx1 */
2202 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2206 * If ctx1 and ctx2 have the same parent; we flatten the parent
2207 * hierarchy, see perf_event_init_context().
2209 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2210 ctx1->parent_gen == ctx2->parent_gen)
2217 static void __perf_event_sync_stat(struct perf_event *event,
2218 struct perf_event *next_event)
2222 if (!event->attr.inherit_stat)
2226 * Update the event value, we cannot use perf_event_read()
2227 * because we're in the middle of a context switch and have IRQs
2228 * disabled, which upsets smp_call_function_single(), however
2229 * we know the event must be on the current CPU, therefore we
2230 * don't need to use it.
2232 switch (event->state) {
2233 case PERF_EVENT_STATE_ACTIVE:
2234 event->pmu->read(event);
2237 case PERF_EVENT_STATE_INACTIVE:
2238 update_event_times(event);
2246 * In order to keep per-task stats reliable we need to flip the event
2247 * values when we flip the contexts.
2249 value = local64_read(&next_event->count);
2250 value = local64_xchg(&event->count, value);
2251 local64_set(&next_event->count, value);
2253 swap(event->total_time_enabled, next_event->total_time_enabled);
2254 swap(event->total_time_running, next_event->total_time_running);
2257 * Since we swizzled the values, update the user visible data too.
2259 perf_event_update_userpage(event);
2260 perf_event_update_userpage(next_event);
2263 static void perf_event_sync_stat(struct perf_event_context *ctx,
2264 struct perf_event_context *next_ctx)
2266 struct perf_event *event, *next_event;
2271 update_context_time(ctx);
2273 event = list_first_entry(&ctx->event_list,
2274 struct perf_event, event_entry);
2276 next_event = list_first_entry(&next_ctx->event_list,
2277 struct perf_event, event_entry);
2279 while (&event->event_entry != &ctx->event_list &&
2280 &next_event->event_entry != &next_ctx->event_list) {
2282 __perf_event_sync_stat(event, next_event);
2284 event = list_next_entry(event, event_entry);
2285 next_event = list_next_entry(next_event, event_entry);
2289 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2290 struct task_struct *next)
2292 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2293 struct perf_event_context *next_ctx;
2294 struct perf_event_context *parent, *next_parent;
2295 struct perf_cpu_context *cpuctx;
2301 cpuctx = __get_cpu_context(ctx);
2302 if (!cpuctx->task_ctx)
2306 next_ctx = next->perf_event_ctxp[ctxn];
2310 parent = rcu_dereference(ctx->parent_ctx);
2311 next_parent = rcu_dereference(next_ctx->parent_ctx);
2313 /* If neither context have a parent context; they cannot be clones. */
2314 if (!parent && !next_parent)
2317 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2319 * Looks like the two contexts are clones, so we might be
2320 * able to optimize the context switch. We lock both
2321 * contexts and check that they are clones under the
2322 * lock (including re-checking that neither has been
2323 * uncloned in the meantime). It doesn't matter which
2324 * order we take the locks because no other cpu could
2325 * be trying to lock both of these tasks.
2327 raw_spin_lock(&ctx->lock);
2328 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2329 if (context_equiv(ctx, next_ctx)) {
2331 * XXX do we need a memory barrier of sorts
2332 * wrt to rcu_dereference() of perf_event_ctxp
2334 task->perf_event_ctxp[ctxn] = next_ctx;
2335 next->perf_event_ctxp[ctxn] = ctx;
2337 next_ctx->task = task;
2340 perf_event_sync_stat(ctx, next_ctx);
2342 raw_spin_unlock(&next_ctx->lock);
2343 raw_spin_unlock(&ctx->lock);
2349 raw_spin_lock(&ctx->lock);
2350 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2351 cpuctx->task_ctx = NULL;
2352 raw_spin_unlock(&ctx->lock);
2356 #define for_each_task_context_nr(ctxn) \
2357 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2360 * Called from scheduler to remove the events of the current task,
2361 * with interrupts disabled.
2363 * We stop each event and update the event value in event->count.
2365 * This does not protect us against NMI, but disable()
2366 * sets the disabled bit in the control field of event _before_
2367 * accessing the event control register. If a NMI hits, then it will
2368 * not restart the event.
2370 void __perf_event_task_sched_out(struct task_struct *task,
2371 struct task_struct *next)
2375 for_each_task_context_nr(ctxn)
2376 perf_event_context_sched_out(task, ctxn, next);
2379 * if cgroup events exist on this CPU, then we need
2380 * to check if we have to switch out PMU state.
2381 * cgroup event are system-wide mode only
2383 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2384 perf_cgroup_sched_out(task, next);
2387 static void task_ctx_sched_out(struct perf_event_context *ctx)
2389 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2391 if (!cpuctx->task_ctx)
2394 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2397 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2398 cpuctx->task_ctx = NULL;
2402 * Called with IRQs disabled
2404 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2405 enum event_type_t event_type)
2407 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2411 ctx_pinned_sched_in(struct perf_event_context *ctx,
2412 struct perf_cpu_context *cpuctx)
2414 struct perf_event *event;
2416 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2417 if (event->state <= PERF_EVENT_STATE_OFF)
2419 if (!event_filter_match(event))
2422 /* may need to reset tstamp_enabled */
2423 if (is_cgroup_event(event))
2424 perf_cgroup_mark_enabled(event, ctx);
2426 if (group_can_go_on(event, cpuctx, 1))
2427 group_sched_in(event, cpuctx, ctx);
2430 * If this pinned group hasn't been scheduled,
2431 * put it in error state.
2433 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2434 update_group_times(event);
2435 event->state = PERF_EVENT_STATE_ERROR;
2441 ctx_flexible_sched_in(struct perf_event_context *ctx,
2442 struct perf_cpu_context *cpuctx)
2444 struct perf_event *event;
2447 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2448 /* Ignore events in OFF or ERROR state */
2449 if (event->state <= PERF_EVENT_STATE_OFF)
2452 * Listen to the 'cpu' scheduling filter constraint
2455 if (!event_filter_match(event))
2458 /* may need to reset tstamp_enabled */
2459 if (is_cgroup_event(event))
2460 perf_cgroup_mark_enabled(event, ctx);
2462 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2463 if (group_sched_in(event, cpuctx, ctx))
2470 ctx_sched_in(struct perf_event_context *ctx,
2471 struct perf_cpu_context *cpuctx,
2472 enum event_type_t event_type,
2473 struct task_struct *task)
2476 int is_active = ctx->is_active;
2478 ctx->is_active |= event_type;
2479 if (likely(!ctx->nr_events))
2483 ctx->timestamp = now;
2484 perf_cgroup_set_timestamp(task, ctx);
2486 * First go through the list and put on any pinned groups
2487 * in order to give them the best chance of going on.
2489 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2490 ctx_pinned_sched_in(ctx, cpuctx);
2492 /* Then walk through the lower prio flexible groups */
2493 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2494 ctx_flexible_sched_in(ctx, cpuctx);
2497 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2498 enum event_type_t event_type,
2499 struct task_struct *task)
2501 struct perf_event_context *ctx = &cpuctx->ctx;
2503 ctx_sched_in(ctx, cpuctx, event_type, task);
2506 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2507 struct task_struct *task)
2509 struct perf_cpu_context *cpuctx;
2511 cpuctx = __get_cpu_context(ctx);
2512 if (cpuctx->task_ctx == ctx)
2515 perf_ctx_lock(cpuctx, ctx);
2516 perf_pmu_disable(ctx->pmu);
2518 * We want to keep the following priority order:
2519 * cpu pinned (that don't need to move), task pinned,
2520 * cpu flexible, task flexible.
2522 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2525 cpuctx->task_ctx = ctx;
2527 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2529 perf_pmu_enable(ctx->pmu);
2530 perf_ctx_unlock(cpuctx, ctx);
2533 * Since these rotations are per-cpu, we need to ensure the
2534 * cpu-context we got scheduled on is actually rotating.
2536 perf_pmu_rotate_start(ctx->pmu);
2540 * When sampling the branck stack in system-wide, it may be necessary
2541 * to flush the stack on context switch. This happens when the branch
2542 * stack does not tag its entries with the pid of the current task.
2543 * Otherwise it becomes impossible to associate a branch entry with a
2544 * task. This ambiguity is more likely to appear when the branch stack
2545 * supports priv level filtering and the user sets it to monitor only
2546 * at the user level (which could be a useful measurement in system-wide
2547 * mode). In that case, the risk is high of having a branch stack with
2548 * branch from multiple tasks. Flushing may mean dropping the existing
2549 * entries or stashing them somewhere in the PMU specific code layer.
2551 * This function provides the context switch callback to the lower code
2552 * layer. It is invoked ONLY when there is at least one system-wide context
2553 * with at least one active event using taken branch sampling.
2555 static void perf_branch_stack_sched_in(struct task_struct *prev,
2556 struct task_struct *task)
2558 struct perf_cpu_context *cpuctx;
2560 unsigned long flags;
2562 /* no need to flush branch stack if not changing task */
2566 local_irq_save(flags);
2570 list_for_each_entry_rcu(pmu, &pmus, entry) {
2571 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2574 * check if the context has at least one
2575 * event using PERF_SAMPLE_BRANCH_STACK
2577 if (cpuctx->ctx.nr_branch_stack > 0
2578 && pmu->flush_branch_stack) {
2580 pmu = cpuctx->ctx.pmu;
2582 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2584 perf_pmu_disable(pmu);
2586 pmu->flush_branch_stack();
2588 perf_pmu_enable(pmu);
2590 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2596 local_irq_restore(flags);
2600 * Called from scheduler to add the events of the current task
2601 * with interrupts disabled.
2603 * We restore the event value and then enable it.
2605 * This does not protect us against NMI, but enable()
2606 * sets the enabled bit in the control field of event _before_
2607 * accessing the event control register. If a NMI hits, then it will
2608 * keep the event running.
2610 void __perf_event_task_sched_in(struct task_struct *prev,
2611 struct task_struct *task)
2613 struct perf_event_context *ctx;
2616 for_each_task_context_nr(ctxn) {
2617 ctx = task->perf_event_ctxp[ctxn];
2621 perf_event_context_sched_in(ctx, task);
2624 * if cgroup events exist on this CPU, then we need
2625 * to check if we have to switch in PMU state.
2626 * cgroup event are system-wide mode only
2628 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2629 perf_cgroup_sched_in(prev, task);
2631 /* check for system-wide branch_stack events */
2632 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2633 perf_branch_stack_sched_in(prev, task);
2636 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2638 u64 frequency = event->attr.sample_freq;
2639 u64 sec = NSEC_PER_SEC;
2640 u64 divisor, dividend;
2642 int count_fls, nsec_fls, frequency_fls, sec_fls;
2644 count_fls = fls64(count);
2645 nsec_fls = fls64(nsec);
2646 frequency_fls = fls64(frequency);
2650 * We got @count in @nsec, with a target of sample_freq HZ
2651 * the target period becomes:
2654 * period = -------------------
2655 * @nsec * sample_freq
2660 * Reduce accuracy by one bit such that @a and @b converge
2661 * to a similar magnitude.
2663 #define REDUCE_FLS(a, b) \
2665 if (a##_fls > b##_fls) { \
2675 * Reduce accuracy until either term fits in a u64, then proceed with
2676 * the other, so that finally we can do a u64/u64 division.
2678 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2679 REDUCE_FLS(nsec, frequency);
2680 REDUCE_FLS(sec, count);
2683 if (count_fls + sec_fls > 64) {
2684 divisor = nsec * frequency;
2686 while (count_fls + sec_fls > 64) {
2687 REDUCE_FLS(count, sec);
2691 dividend = count * sec;
2693 dividend = count * sec;
2695 while (nsec_fls + frequency_fls > 64) {
2696 REDUCE_FLS(nsec, frequency);
2700 divisor = nsec * frequency;
2706 return div64_u64(dividend, divisor);
2709 static DEFINE_PER_CPU(int, perf_throttled_count);
2710 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2712 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2714 struct hw_perf_event *hwc = &event->hw;
2715 s64 period, sample_period;
2718 period = perf_calculate_period(event, nsec, count);
2720 delta = (s64)(period - hwc->sample_period);
2721 delta = (delta + 7) / 8; /* low pass filter */
2723 sample_period = hwc->sample_period + delta;
2728 hwc->sample_period = sample_period;
2730 if (local64_read(&hwc->period_left) > 8*sample_period) {
2732 event->pmu->stop(event, PERF_EF_UPDATE);
2734 local64_set(&hwc->period_left, 0);
2737 event->pmu->start(event, PERF_EF_RELOAD);
2742 * combine freq adjustment with unthrottling to avoid two passes over the
2743 * events. At the same time, make sure, having freq events does not change
2744 * the rate of unthrottling as that would introduce bias.
2746 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2749 struct perf_event *event;
2750 struct hw_perf_event *hwc;
2751 u64 now, period = TICK_NSEC;
2755 * only need to iterate over all events iff:
2756 * - context have events in frequency mode (needs freq adjust)
2757 * - there are events to unthrottle on this cpu
2759 if (!(ctx->nr_freq || needs_unthr))
2762 raw_spin_lock(&ctx->lock);
2763 perf_pmu_disable(ctx->pmu);
2765 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2766 if (event->state != PERF_EVENT_STATE_ACTIVE)
2769 if (!event_filter_match(event))
2772 perf_pmu_disable(event->pmu);
2776 if (hwc->interrupts == MAX_INTERRUPTS) {
2777 hwc->interrupts = 0;
2778 perf_log_throttle(event, 1);
2779 event->pmu->start(event, 0);
2782 if (!event->attr.freq || !event->attr.sample_freq)
2786 * stop the event and update event->count
2788 event->pmu->stop(event, PERF_EF_UPDATE);
2790 now = local64_read(&event->count);
2791 delta = now - hwc->freq_count_stamp;
2792 hwc->freq_count_stamp = now;
2796 * reload only if value has changed
2797 * we have stopped the event so tell that
2798 * to perf_adjust_period() to avoid stopping it
2802 perf_adjust_period(event, period, delta, false);
2804 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2806 perf_pmu_enable(event->pmu);
2809 perf_pmu_enable(ctx->pmu);
2810 raw_spin_unlock(&ctx->lock);
2814 * Round-robin a context's events:
2816 static void rotate_ctx(struct perf_event_context *ctx)
2819 * Rotate the first entry last of non-pinned groups. Rotation might be
2820 * disabled by the inheritance code.
2822 if (!ctx->rotate_disable)
2823 list_rotate_left(&ctx->flexible_groups);
2827 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2828 * because they're strictly cpu affine and rotate_start is called with IRQs
2829 * disabled, while rotate_context is called from IRQ context.
2831 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2833 struct perf_event_context *ctx = NULL;
2834 int rotate = 0, remove = 1;
2836 if (cpuctx->ctx.nr_events) {
2838 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2842 ctx = cpuctx->task_ctx;
2843 if (ctx && ctx->nr_events) {
2845 if (ctx->nr_events != ctx->nr_active)
2852 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2853 perf_pmu_disable(cpuctx->ctx.pmu);
2855 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2857 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2859 rotate_ctx(&cpuctx->ctx);
2863 perf_event_sched_in(cpuctx, ctx, current);
2865 perf_pmu_enable(cpuctx->ctx.pmu);
2866 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2869 list_del_init(&cpuctx->rotation_list);
2874 #ifdef CONFIG_NO_HZ_FULL
2875 bool perf_event_can_stop_tick(void)
2877 if (atomic_read(&nr_freq_events) ||
2878 __this_cpu_read(perf_throttled_count))
2885 void perf_event_task_tick(void)
2887 struct list_head *head = &__get_cpu_var(rotation_list);
2888 struct perf_cpu_context *cpuctx, *tmp;
2889 struct perf_event_context *ctx;
2892 WARN_ON(!irqs_disabled());
2894 __this_cpu_inc(perf_throttled_seq);
2895 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2897 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2899 perf_adjust_freq_unthr_context(ctx, throttled);
2901 ctx = cpuctx->task_ctx;
2903 perf_adjust_freq_unthr_context(ctx, throttled);
2907 static int event_enable_on_exec(struct perf_event *event,
2908 struct perf_event_context *ctx)
2910 if (!event->attr.enable_on_exec)
2913 event->attr.enable_on_exec = 0;
2914 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2917 __perf_event_mark_enabled(event);
2923 * Enable all of a task's events that have been marked enable-on-exec.
2924 * This expects task == current.
2926 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2928 struct perf_event *event;
2929 unsigned long flags;
2933 local_irq_save(flags);
2934 if (!ctx || !ctx->nr_events)
2938 * We must ctxsw out cgroup events to avoid conflict
2939 * when invoking perf_task_event_sched_in() later on
2940 * in this function. Otherwise we end up trying to
2941 * ctxswin cgroup events which are already scheduled
2944 perf_cgroup_sched_out(current, NULL);
2946 raw_spin_lock(&ctx->lock);
2947 task_ctx_sched_out(ctx);
2949 list_for_each_entry(event, &ctx->event_list, event_entry) {
2950 ret = event_enable_on_exec(event, ctx);
2956 * Unclone this context if we enabled any event.
2961 raw_spin_unlock(&ctx->lock);
2964 * Also calls ctxswin for cgroup events, if any:
2966 perf_event_context_sched_in(ctx, ctx->task);
2968 local_irq_restore(flags);
2972 * Cross CPU call to read the hardware event
2974 static void __perf_event_read(void *info)
2976 struct perf_event *event = info;
2977 struct perf_event_context *ctx = event->ctx;
2978 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2981 * If this is a task context, we need to check whether it is
2982 * the current task context of this cpu. If not it has been
2983 * scheduled out before the smp call arrived. In that case
2984 * event->count would have been updated to a recent sample
2985 * when the event was scheduled out.
2987 if (ctx->task && cpuctx->task_ctx != ctx)
2990 raw_spin_lock(&ctx->lock);
2991 if (ctx->is_active) {
2992 update_context_time(ctx);
2993 update_cgrp_time_from_event(event);
2995 update_event_times(event);
2996 if (event->state == PERF_EVENT_STATE_ACTIVE)
2997 event->pmu->read(event);
2998 raw_spin_unlock(&ctx->lock);
3001 static inline u64 perf_event_count(struct perf_event *event)
3003 return local64_read(&event->count) + atomic64_read(&event->child_count);
3006 static u64 perf_event_read(struct perf_event *event)
3009 * If event is enabled and currently active on a CPU, update the
3010 * value in the event structure:
3012 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3013 smp_call_function_single(event->oncpu,
3014 __perf_event_read, event, 1);
3015 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3016 struct perf_event_context *ctx = event->ctx;
3017 unsigned long flags;
3019 raw_spin_lock_irqsave(&ctx->lock, flags);
3021 * may read while context is not active
3022 * (e.g., thread is blocked), in that case
3023 * we cannot update context time
3025 if (ctx->is_active) {
3026 update_context_time(ctx);
3027 update_cgrp_time_from_event(event);
3029 update_event_times(event);
3030 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3033 return perf_event_count(event);
3037 * Initialize the perf_event context in a task_struct:
3039 static void __perf_event_init_context(struct perf_event_context *ctx)
3041 raw_spin_lock_init(&ctx->lock);
3042 mutex_init(&ctx->mutex);
3043 INIT_LIST_HEAD(&ctx->pinned_groups);
3044 INIT_LIST_HEAD(&ctx->flexible_groups);
3045 INIT_LIST_HEAD(&ctx->event_list);
3046 atomic_set(&ctx->refcount, 1);
3049 static struct perf_event_context *
3050 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3052 struct perf_event_context *ctx;
3054 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3058 __perf_event_init_context(ctx);
3061 get_task_struct(task);
3068 static struct task_struct *
3069 find_lively_task_by_vpid(pid_t vpid)
3071 struct task_struct *task;
3078 task = find_task_by_vpid(vpid);
3080 get_task_struct(task);
3084 return ERR_PTR(-ESRCH);
3086 /* Reuse ptrace permission checks for now. */
3088 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3093 put_task_struct(task);
3094 return ERR_PTR(err);
3099 * Returns a matching context with refcount and pincount.
3101 static struct perf_event_context *
3102 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3104 struct perf_event_context *ctx;
3105 struct perf_cpu_context *cpuctx;
3106 unsigned long flags;
3110 /* Must be root to operate on a CPU event: */
3111 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3112 return ERR_PTR(-EACCES);
3115 * We could be clever and allow to attach a event to an
3116 * offline CPU and activate it when the CPU comes up, but
3119 if (!cpu_online(cpu))
3120 return ERR_PTR(-ENODEV);
3122 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3131 ctxn = pmu->task_ctx_nr;
3136 ctx = perf_lock_task_context(task, ctxn, &flags);
3140 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3142 ctx = alloc_perf_context(pmu, task);
3148 mutex_lock(&task->perf_event_mutex);
3150 * If it has already passed perf_event_exit_task().
3151 * we must see PF_EXITING, it takes this mutex too.
3153 if (task->flags & PF_EXITING)
3155 else if (task->perf_event_ctxp[ctxn])
3160 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3162 mutex_unlock(&task->perf_event_mutex);
3164 if (unlikely(err)) {
3176 return ERR_PTR(err);
3179 static void perf_event_free_filter(struct perf_event *event);
3181 static void free_event_rcu(struct rcu_head *head)
3183 struct perf_event *event;
3185 event = container_of(head, struct perf_event, rcu_head);
3187 put_pid_ns(event->ns);
3188 perf_event_free_filter(event);
3192 static void ring_buffer_put(struct ring_buffer *rb);
3193 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3195 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3200 if (has_branch_stack(event)) {
3201 if (!(event->attach_state & PERF_ATTACH_TASK))
3202 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3204 if (is_cgroup_event(event))
3205 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3208 static void unaccount_event(struct perf_event *event)
3213 if (event->attach_state & PERF_ATTACH_TASK)
3214 static_key_slow_dec_deferred(&perf_sched_events);
3215 if (event->attr.mmap || event->attr.mmap_data)
3216 atomic_dec(&nr_mmap_events);
3217 if (event->attr.comm)
3218 atomic_dec(&nr_comm_events);
3219 if (event->attr.task)
3220 atomic_dec(&nr_task_events);
3221 if (event->attr.freq)
3222 atomic_dec(&nr_freq_events);
3223 if (is_cgroup_event(event))
3224 static_key_slow_dec_deferred(&perf_sched_events);
3225 if (has_branch_stack(event))
3226 static_key_slow_dec_deferred(&perf_sched_events);
3228 unaccount_event_cpu(event, event->cpu);
3231 static void __free_event(struct perf_event *event)
3233 if (!event->parent) {
3234 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3235 put_callchain_buffers();
3239 event->destroy(event);
3242 put_ctx(event->ctx);
3244 call_rcu(&event->rcu_head, free_event_rcu);
3246 static void free_event(struct perf_event *event)
3248 irq_work_sync(&event->pending);
3250 unaccount_event(event);
3253 struct ring_buffer *rb;
3256 * Can happen when we close an event with re-directed output.
3258 * Since we have a 0 refcount, perf_mmap_close() will skip
3259 * over us; possibly making our ring_buffer_put() the last.
3261 mutex_lock(&event->mmap_mutex);
3264 rcu_assign_pointer(event->rb, NULL);
3265 ring_buffer_detach(event, rb);
3266 ring_buffer_put(rb); /* could be last */
3268 mutex_unlock(&event->mmap_mutex);
3271 if (is_cgroup_event(event))
3272 perf_detach_cgroup(event);
3275 __free_event(event);
3278 int perf_event_release_kernel(struct perf_event *event)
3280 struct perf_event_context *ctx = event->ctx;
3282 WARN_ON_ONCE(ctx->parent_ctx);
3284 * There are two ways this annotation is useful:
3286 * 1) there is a lock recursion from perf_event_exit_task
3287 * see the comment there.
3289 * 2) there is a lock-inversion with mmap_sem through
3290 * perf_event_read_group(), which takes faults while
3291 * holding ctx->mutex, however this is called after
3292 * the last filedesc died, so there is no possibility
3293 * to trigger the AB-BA case.
3295 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3296 perf_remove_from_context(event, true);
3297 mutex_unlock(&ctx->mutex);
3303 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3306 * Called when the last reference to the file is gone.
3308 static void put_event(struct perf_event *event)
3310 struct task_struct *owner;
3312 if (!atomic_long_dec_and_test(&event->refcount))
3316 owner = ACCESS_ONCE(event->owner);
3318 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3319 * !owner it means the list deletion is complete and we can indeed
3320 * free this event, otherwise we need to serialize on
3321 * owner->perf_event_mutex.
3323 smp_read_barrier_depends();
3326 * Since delayed_put_task_struct() also drops the last
3327 * task reference we can safely take a new reference
3328 * while holding the rcu_read_lock().
3330 get_task_struct(owner);
3335 mutex_lock(&owner->perf_event_mutex);
3337 * We have to re-check the event->owner field, if it is cleared
3338 * we raced with perf_event_exit_task(), acquiring the mutex
3339 * ensured they're done, and we can proceed with freeing the
3343 list_del_init(&event->owner_entry);
3344 mutex_unlock(&owner->perf_event_mutex);
3345 put_task_struct(owner);
3348 perf_event_release_kernel(event);
3351 static int perf_release(struct inode *inode, struct file *file)
3353 put_event(file->private_data);
3357 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3359 struct perf_event *child;
3365 mutex_lock(&event->child_mutex);
3366 total += perf_event_read(event);
3367 *enabled += event->total_time_enabled +
3368 atomic64_read(&event->child_total_time_enabled);
3369 *running += event->total_time_running +
3370 atomic64_read(&event->child_total_time_running);
3372 list_for_each_entry(child, &event->child_list, child_list) {
3373 total += perf_event_read(child);
3374 *enabled += child->total_time_enabled;
3375 *running += child->total_time_running;
3377 mutex_unlock(&event->child_mutex);
3381 EXPORT_SYMBOL_GPL(perf_event_read_value);
3383 static int perf_event_read_group(struct perf_event *event,
3384 u64 read_format, char __user *buf)
3386 struct perf_event *leader = event->group_leader, *sub;
3387 int n = 0, size = 0, ret = -EFAULT;
3388 struct perf_event_context *ctx = leader->ctx;
3390 u64 count, enabled, running;
3392 mutex_lock(&ctx->mutex);
3393 count = perf_event_read_value(leader, &enabled, &running);
3395 values[n++] = 1 + leader->nr_siblings;
3396 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3397 values[n++] = enabled;
3398 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3399 values[n++] = running;
3400 values[n++] = count;
3401 if (read_format & PERF_FORMAT_ID)
3402 values[n++] = primary_event_id(leader);
3404 size = n * sizeof(u64);
3406 if (copy_to_user(buf, values, size))
3411 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3414 values[n++] = perf_event_read_value(sub, &enabled, &running);
3415 if (read_format & PERF_FORMAT_ID)
3416 values[n++] = primary_event_id(sub);
3418 size = n * sizeof(u64);
3420 if (copy_to_user(buf + ret, values, size)) {
3428 mutex_unlock(&ctx->mutex);
3433 static int perf_event_read_one(struct perf_event *event,
3434 u64 read_format, char __user *buf)
3436 u64 enabled, running;
3440 values[n++] = perf_event_read_value(event, &enabled, &running);
3441 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3442 values[n++] = enabled;
3443 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3444 values[n++] = running;
3445 if (read_format & PERF_FORMAT_ID)
3446 values[n++] = primary_event_id(event);
3448 if (copy_to_user(buf, values, n * sizeof(u64)))
3451 return n * sizeof(u64);
3455 * Read the performance event - simple non blocking version for now
3458 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3460 u64 read_format = event->attr.read_format;
3464 * Return end-of-file for a read on a event that is in
3465 * error state (i.e. because it was pinned but it couldn't be
3466 * scheduled on to the CPU at some point).
3468 if (event->state == PERF_EVENT_STATE_ERROR)
3471 if (count < event->read_size)
3474 WARN_ON_ONCE(event->ctx->parent_ctx);
3475 if (read_format & PERF_FORMAT_GROUP)
3476 ret = perf_event_read_group(event, read_format, buf);
3478 ret = perf_event_read_one(event, read_format, buf);
3484 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3486 struct perf_event *event = file->private_data;
3488 return perf_read_hw(event, buf, count);
3491 static unsigned int perf_poll(struct file *file, poll_table *wait)
3493 struct perf_event *event = file->private_data;
3494 struct ring_buffer *rb;
3495 unsigned int events = POLL_HUP;
3498 * Pin the event->rb by taking event->mmap_mutex; otherwise
3499 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3501 mutex_lock(&event->mmap_mutex);
3504 events = atomic_xchg(&rb->poll, 0);
3505 mutex_unlock(&event->mmap_mutex);
3507 poll_wait(file, &event->waitq, wait);
3512 static void perf_event_reset(struct perf_event *event)
3514 (void)perf_event_read(event);
3515 local64_set(&event->count, 0);
3516 perf_event_update_userpage(event);
3520 * Holding the top-level event's child_mutex means that any
3521 * descendant process that has inherited this event will block
3522 * in sync_child_event if it goes to exit, thus satisfying the
3523 * task existence requirements of perf_event_enable/disable.
3525 static void perf_event_for_each_child(struct perf_event *event,
3526 void (*func)(struct perf_event *))
3528 struct perf_event *child;
3530 WARN_ON_ONCE(event->ctx->parent_ctx);
3531 mutex_lock(&event->child_mutex);
3533 list_for_each_entry(child, &event->child_list, child_list)
3535 mutex_unlock(&event->child_mutex);
3538 static void perf_event_for_each(struct perf_event *event,
3539 void (*func)(struct perf_event *))
3541 struct perf_event_context *ctx = event->ctx;
3542 struct perf_event *sibling;
3544 WARN_ON_ONCE(ctx->parent_ctx);
3545 mutex_lock(&ctx->mutex);
3546 event = event->group_leader;
3548 perf_event_for_each_child(event, func);
3549 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3550 perf_event_for_each_child(sibling, func);
3551 mutex_unlock(&ctx->mutex);
3554 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3556 struct perf_event_context *ctx = event->ctx;
3557 int ret = 0, active;
3560 if (!is_sampling_event(event))
3563 if (copy_from_user(&value, arg, sizeof(value)))
3569 raw_spin_lock_irq(&ctx->lock);
3570 if (event->attr.freq) {
3571 if (value > sysctl_perf_event_sample_rate) {
3576 event->attr.sample_freq = value;
3578 event->attr.sample_period = value;
3579 event->hw.sample_period = value;
3582 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3584 perf_pmu_disable(ctx->pmu);
3585 event->pmu->stop(event, PERF_EF_UPDATE);
3588 local64_set(&event->hw.period_left, 0);
3591 event->pmu->start(event, PERF_EF_RELOAD);
3592 perf_pmu_enable(ctx->pmu);
3596 raw_spin_unlock_irq(&ctx->lock);
3601 static const struct file_operations perf_fops;
3603 static inline int perf_fget_light(int fd, struct fd *p)
3605 struct fd f = fdget(fd);
3609 if (f.file->f_op != &perf_fops) {
3617 static int perf_event_set_output(struct perf_event *event,
3618 struct perf_event *output_event);
3619 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3621 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3623 struct perf_event *event = file->private_data;
3624 void (*func)(struct perf_event *);
3628 case PERF_EVENT_IOC_ENABLE:
3629 func = perf_event_enable;
3631 case PERF_EVENT_IOC_DISABLE:
3632 func = perf_event_disable;
3634 case PERF_EVENT_IOC_RESET:
3635 func = perf_event_reset;
3638 case PERF_EVENT_IOC_REFRESH:
3639 return perf_event_refresh(event, arg);
3641 case PERF_EVENT_IOC_PERIOD:
3642 return perf_event_period(event, (u64 __user *)arg);
3644 case PERF_EVENT_IOC_ID:
3646 u64 id = primary_event_id(event);
3648 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3653 case PERF_EVENT_IOC_SET_OUTPUT:
3657 struct perf_event *output_event;
3659 ret = perf_fget_light(arg, &output);
3662 output_event = output.file->private_data;
3663 ret = perf_event_set_output(event, output_event);
3666 ret = perf_event_set_output(event, NULL);
3671 case PERF_EVENT_IOC_SET_FILTER:
3672 return perf_event_set_filter(event, (void __user *)arg);
3678 if (flags & PERF_IOC_FLAG_GROUP)
3679 perf_event_for_each(event, func);
3681 perf_event_for_each_child(event, func);
3686 int perf_event_task_enable(void)
3688 struct perf_event *event;
3690 mutex_lock(¤t->perf_event_mutex);
3691 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3692 perf_event_for_each_child(event, perf_event_enable);
3693 mutex_unlock(¤t->perf_event_mutex);
3698 int perf_event_task_disable(void)
3700 struct perf_event *event;
3702 mutex_lock(¤t->perf_event_mutex);
3703 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3704 perf_event_for_each_child(event, perf_event_disable);
3705 mutex_unlock(¤t->perf_event_mutex);
3710 static int perf_event_index(struct perf_event *event)
3712 if (event->hw.state & PERF_HES_STOPPED)
3715 if (event->state != PERF_EVENT_STATE_ACTIVE)
3718 return event->pmu->event_idx(event);
3721 static void calc_timer_values(struct perf_event *event,
3728 *now = perf_clock();
3729 ctx_time = event->shadow_ctx_time + *now;
3730 *enabled = ctx_time - event->tstamp_enabled;
3731 *running = ctx_time - event->tstamp_running;
3734 static void perf_event_init_userpage(struct perf_event *event)
3736 struct perf_event_mmap_page *userpg;
3737 struct ring_buffer *rb;
3740 rb = rcu_dereference(event->rb);
3744 userpg = rb->user_page;
3746 /* Allow new userspace to detect that bit 0 is deprecated */
3747 userpg->cap_bit0_is_deprecated = 1;
3748 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3754 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3759 * Callers need to ensure there can be no nesting of this function, otherwise
3760 * the seqlock logic goes bad. We can not serialize this because the arch
3761 * code calls this from NMI context.
3763 void perf_event_update_userpage(struct perf_event *event)
3765 struct perf_event_mmap_page *userpg;
3766 struct ring_buffer *rb;
3767 u64 enabled, running, now;
3770 rb = rcu_dereference(event->rb);
3775 * compute total_time_enabled, total_time_running
3776 * based on snapshot values taken when the event
3777 * was last scheduled in.
3779 * we cannot simply called update_context_time()
3780 * because of locking issue as we can be called in
3783 calc_timer_values(event, &now, &enabled, &running);
3785 userpg = rb->user_page;
3787 * Disable preemption so as to not let the corresponding user-space
3788 * spin too long if we get preempted.
3793 userpg->index = perf_event_index(event);
3794 userpg->offset = perf_event_count(event);
3796 userpg->offset -= local64_read(&event->hw.prev_count);
3798 userpg->time_enabled = enabled +
3799 atomic64_read(&event->child_total_time_enabled);
3801 userpg->time_running = running +
3802 atomic64_read(&event->child_total_time_running);
3804 arch_perf_update_userpage(userpg, now);
3813 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3815 struct perf_event *event = vma->vm_file->private_data;
3816 struct ring_buffer *rb;
3817 int ret = VM_FAULT_SIGBUS;
3819 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3820 if (vmf->pgoff == 0)
3826 rb = rcu_dereference(event->rb);
3830 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3833 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3837 get_page(vmf->page);
3838 vmf->page->mapping = vma->vm_file->f_mapping;
3839 vmf->page->index = vmf->pgoff;
3848 static void ring_buffer_attach(struct perf_event *event,
3849 struct ring_buffer *rb)
3851 unsigned long flags;
3853 if (!list_empty(&event->rb_entry))
3856 spin_lock_irqsave(&rb->event_lock, flags);
3857 if (list_empty(&event->rb_entry))
3858 list_add(&event->rb_entry, &rb->event_list);
3859 spin_unlock_irqrestore(&rb->event_lock, flags);
3862 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3864 unsigned long flags;
3866 if (list_empty(&event->rb_entry))
3869 spin_lock_irqsave(&rb->event_lock, flags);
3870 list_del_init(&event->rb_entry);
3871 wake_up_all(&event->waitq);
3872 spin_unlock_irqrestore(&rb->event_lock, flags);
3875 static void ring_buffer_wakeup(struct perf_event *event)
3877 struct ring_buffer *rb;
3880 rb = rcu_dereference(event->rb);
3882 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3883 wake_up_all(&event->waitq);
3888 static void rb_free_rcu(struct rcu_head *rcu_head)
3890 struct ring_buffer *rb;
3892 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3896 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3898 struct ring_buffer *rb;
3901 rb = rcu_dereference(event->rb);
3903 if (!atomic_inc_not_zero(&rb->refcount))
3911 static void ring_buffer_put(struct ring_buffer *rb)
3913 if (!atomic_dec_and_test(&rb->refcount))
3916 WARN_ON_ONCE(!list_empty(&rb->event_list));
3918 call_rcu(&rb->rcu_head, rb_free_rcu);
3921 static void perf_mmap_open(struct vm_area_struct *vma)
3923 struct perf_event *event = vma->vm_file->private_data;
3925 atomic_inc(&event->mmap_count);
3926 atomic_inc(&event->rb->mmap_count);
3930 * A buffer can be mmap()ed multiple times; either directly through the same
3931 * event, or through other events by use of perf_event_set_output().
3933 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3934 * the buffer here, where we still have a VM context. This means we need
3935 * to detach all events redirecting to us.
3937 static void perf_mmap_close(struct vm_area_struct *vma)
3939 struct perf_event *event = vma->vm_file->private_data;
3941 struct ring_buffer *rb = event->rb;
3942 struct user_struct *mmap_user = rb->mmap_user;
3943 int mmap_locked = rb->mmap_locked;
3944 unsigned long size = perf_data_size(rb);
3946 atomic_dec(&rb->mmap_count);
3948 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3951 /* Detach current event from the buffer. */
3952 rcu_assign_pointer(event->rb, NULL);
3953 ring_buffer_detach(event, rb);
3954 mutex_unlock(&event->mmap_mutex);
3956 /* If there's still other mmap()s of this buffer, we're done. */
3957 if (atomic_read(&rb->mmap_count)) {
3958 ring_buffer_put(rb); /* can't be last */
3963 * No other mmap()s, detach from all other events that might redirect
3964 * into the now unreachable buffer. Somewhat complicated by the
3965 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3969 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3970 if (!atomic_long_inc_not_zero(&event->refcount)) {
3972 * This event is en-route to free_event() which will
3973 * detach it and remove it from the list.
3979 mutex_lock(&event->mmap_mutex);
3981 * Check we didn't race with perf_event_set_output() which can
3982 * swizzle the rb from under us while we were waiting to
3983 * acquire mmap_mutex.
3985 * If we find a different rb; ignore this event, a next
3986 * iteration will no longer find it on the list. We have to
3987 * still restart the iteration to make sure we're not now
3988 * iterating the wrong list.
3990 if (event->rb == rb) {
3991 rcu_assign_pointer(event->rb, NULL);
3992 ring_buffer_detach(event, rb);
3993 ring_buffer_put(rb); /* can't be last, we still have one */
3995 mutex_unlock(&event->mmap_mutex);
3999 * Restart the iteration; either we're on the wrong list or
4000 * destroyed its integrity by doing a deletion.
4007 * It could be there's still a few 0-ref events on the list; they'll
4008 * get cleaned up by free_event() -- they'll also still have their
4009 * ref on the rb and will free it whenever they are done with it.
4011 * Aside from that, this buffer is 'fully' detached and unmapped,
4012 * undo the VM accounting.
4015 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4016 vma->vm_mm->pinned_vm -= mmap_locked;
4017 free_uid(mmap_user);
4019 ring_buffer_put(rb); /* could be last */
4022 static const struct vm_operations_struct perf_mmap_vmops = {
4023 .open = perf_mmap_open,
4024 .close = perf_mmap_close,
4025 .fault = perf_mmap_fault,
4026 .page_mkwrite = perf_mmap_fault,
4029 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4031 struct perf_event *event = file->private_data;
4032 unsigned long user_locked, user_lock_limit;
4033 struct user_struct *user = current_user();
4034 unsigned long locked, lock_limit;
4035 struct ring_buffer *rb;
4036 unsigned long vma_size;
4037 unsigned long nr_pages;
4038 long user_extra, extra;
4039 int ret = 0, flags = 0;
4042 * Don't allow mmap() of inherited per-task counters. This would
4043 * create a performance issue due to all children writing to the
4046 if (event->cpu == -1 && event->attr.inherit)
4049 if (!(vma->vm_flags & VM_SHARED))
4052 vma_size = vma->vm_end - vma->vm_start;
4053 nr_pages = (vma_size / PAGE_SIZE) - 1;
4056 * If we have rb pages ensure they're a power-of-two number, so we
4057 * can do bitmasks instead of modulo.
4059 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4062 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4065 if (vma->vm_pgoff != 0)
4068 WARN_ON_ONCE(event->ctx->parent_ctx);
4070 mutex_lock(&event->mmap_mutex);
4072 if (event->rb->nr_pages != nr_pages) {
4077 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4079 * Raced against perf_mmap_close() through
4080 * perf_event_set_output(). Try again, hope for better
4083 mutex_unlock(&event->mmap_mutex);
4090 user_extra = nr_pages + 1;
4091 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4094 * Increase the limit linearly with more CPUs:
4096 user_lock_limit *= num_online_cpus();
4098 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4101 if (user_locked > user_lock_limit)
4102 extra = user_locked - user_lock_limit;
4104 lock_limit = rlimit(RLIMIT_MEMLOCK);
4105 lock_limit >>= PAGE_SHIFT;
4106 locked = vma->vm_mm->pinned_vm + extra;
4108 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4109 !capable(CAP_IPC_LOCK)) {
4116 if (vma->vm_flags & VM_WRITE)
4117 flags |= RING_BUFFER_WRITABLE;
4119 rb = rb_alloc(nr_pages,
4120 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4128 atomic_set(&rb->mmap_count, 1);
4129 rb->mmap_locked = extra;
4130 rb->mmap_user = get_current_user();
4132 atomic_long_add(user_extra, &user->locked_vm);
4133 vma->vm_mm->pinned_vm += extra;
4135 ring_buffer_attach(event, rb);
4136 rcu_assign_pointer(event->rb, rb);
4138 perf_event_init_userpage(event);
4139 perf_event_update_userpage(event);
4143 atomic_inc(&event->mmap_count);
4144 mutex_unlock(&event->mmap_mutex);
4147 * Since pinned accounting is per vm we cannot allow fork() to copy our
4150 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4151 vma->vm_ops = &perf_mmap_vmops;
4156 static int perf_fasync(int fd, struct file *filp, int on)
4158 struct inode *inode = file_inode(filp);
4159 struct perf_event *event = filp->private_data;
4162 mutex_lock(&inode->i_mutex);
4163 retval = fasync_helper(fd, filp, on, &event->fasync);
4164 mutex_unlock(&inode->i_mutex);
4172 static const struct file_operations perf_fops = {
4173 .llseek = no_llseek,
4174 .release = perf_release,
4177 .unlocked_ioctl = perf_ioctl,
4178 .compat_ioctl = perf_ioctl,
4180 .fasync = perf_fasync,
4186 * If there's data, ensure we set the poll() state and publish everything
4187 * to user-space before waking everybody up.
4190 void perf_event_wakeup(struct perf_event *event)
4192 ring_buffer_wakeup(event);
4194 if (event->pending_kill) {
4195 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4196 event->pending_kill = 0;
4200 static void perf_pending_event(struct irq_work *entry)
4202 struct perf_event *event = container_of(entry,
4203 struct perf_event, pending);
4205 if (event->pending_disable) {
4206 event->pending_disable = 0;
4207 __perf_event_disable(event);
4210 if (event->pending_wakeup) {
4211 event->pending_wakeup = 0;
4212 perf_event_wakeup(event);
4217 * We assume there is only KVM supporting the callbacks.
4218 * Later on, we might change it to a list if there is
4219 * another virtualization implementation supporting the callbacks.
4221 struct perf_guest_info_callbacks *perf_guest_cbs;
4223 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4225 perf_guest_cbs = cbs;
4228 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4230 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4232 perf_guest_cbs = NULL;
4235 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4238 perf_output_sample_regs(struct perf_output_handle *handle,
4239 struct pt_regs *regs, u64 mask)
4243 for_each_set_bit(bit, (const unsigned long *) &mask,
4244 sizeof(mask) * BITS_PER_BYTE) {
4247 val = perf_reg_value(regs, bit);
4248 perf_output_put(handle, val);
4252 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4253 struct pt_regs *regs)
4255 if (!user_mode(regs)) {
4257 regs = task_pt_regs(current);
4263 regs_user->regs = regs;
4264 regs_user->abi = perf_reg_abi(current);
4269 * Get remaining task size from user stack pointer.
4271 * It'd be better to take stack vma map and limit this more
4272 * precisly, but there's no way to get it safely under interrupt,
4273 * so using TASK_SIZE as limit.
4275 static u64 perf_ustack_task_size(struct pt_regs *regs)
4277 unsigned long addr = perf_user_stack_pointer(regs);
4279 if (!addr || addr >= TASK_SIZE)
4282 return TASK_SIZE - addr;
4286 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4287 struct pt_regs *regs)
4291 /* No regs, no stack pointer, no dump. */
4296 * Check if we fit in with the requested stack size into the:
4298 * If we don't, we limit the size to the TASK_SIZE.
4300 * - remaining sample size
4301 * If we don't, we customize the stack size to
4302 * fit in to the remaining sample size.
4305 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4306 stack_size = min(stack_size, (u16) task_size);
4308 /* Current header size plus static size and dynamic size. */
4309 header_size += 2 * sizeof(u64);
4311 /* Do we fit in with the current stack dump size? */
4312 if ((u16) (header_size + stack_size) < header_size) {
4314 * If we overflow the maximum size for the sample,
4315 * we customize the stack dump size to fit in.
4317 stack_size = USHRT_MAX - header_size - sizeof(u64);
4318 stack_size = round_up(stack_size, sizeof(u64));
4325 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4326 struct pt_regs *regs)
4328 /* Case of a kernel thread, nothing to dump */
4331 perf_output_put(handle, size);
4340 * - the size requested by user or the best one we can fit
4341 * in to the sample max size
4343 * - user stack dump data
4345 * - the actual dumped size
4349 perf_output_put(handle, dump_size);
4352 sp = perf_user_stack_pointer(regs);
4353 rem = __output_copy_user(handle, (void *) sp, dump_size);
4354 dyn_size = dump_size - rem;
4356 perf_output_skip(handle, rem);
4359 perf_output_put(handle, dyn_size);
4363 static void __perf_event_header__init_id(struct perf_event_header *header,
4364 struct perf_sample_data *data,
4365 struct perf_event *event)
4367 u64 sample_type = event->attr.sample_type;
4369 data->type = sample_type;
4370 header->size += event->id_header_size;
4372 if (sample_type & PERF_SAMPLE_TID) {
4373 /* namespace issues */
4374 data->tid_entry.pid = perf_event_pid(event, current);
4375 data->tid_entry.tid = perf_event_tid(event, current);
4378 if (sample_type & PERF_SAMPLE_TIME)
4379 data->time = perf_clock();
4381 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4382 data->id = primary_event_id(event);
4384 if (sample_type & PERF_SAMPLE_STREAM_ID)
4385 data->stream_id = event->id;
4387 if (sample_type & PERF_SAMPLE_CPU) {
4388 data->cpu_entry.cpu = raw_smp_processor_id();
4389 data->cpu_entry.reserved = 0;
4393 void perf_event_header__init_id(struct perf_event_header *header,
4394 struct perf_sample_data *data,
4395 struct perf_event *event)
4397 if (event->attr.sample_id_all)
4398 __perf_event_header__init_id(header, data, event);
4401 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4402 struct perf_sample_data *data)
4404 u64 sample_type = data->type;
4406 if (sample_type & PERF_SAMPLE_TID)
4407 perf_output_put(handle, data->tid_entry);
4409 if (sample_type & PERF_SAMPLE_TIME)
4410 perf_output_put(handle, data->time);
4412 if (sample_type & PERF_SAMPLE_ID)
4413 perf_output_put(handle, data->id);
4415 if (sample_type & PERF_SAMPLE_STREAM_ID)
4416 perf_output_put(handle, data->stream_id);
4418 if (sample_type & PERF_SAMPLE_CPU)
4419 perf_output_put(handle, data->cpu_entry);
4421 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4422 perf_output_put(handle, data->id);
4425 void perf_event__output_id_sample(struct perf_event *event,
4426 struct perf_output_handle *handle,
4427 struct perf_sample_data *sample)
4429 if (event->attr.sample_id_all)
4430 __perf_event__output_id_sample(handle, sample);
4433 static void perf_output_read_one(struct perf_output_handle *handle,
4434 struct perf_event *event,
4435 u64 enabled, u64 running)
4437 u64 read_format = event->attr.read_format;
4441 values[n++] = perf_event_count(event);
4442 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4443 values[n++] = enabled +
4444 atomic64_read(&event->child_total_time_enabled);
4446 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4447 values[n++] = running +
4448 atomic64_read(&event->child_total_time_running);
4450 if (read_format & PERF_FORMAT_ID)
4451 values[n++] = primary_event_id(event);
4453 __output_copy(handle, values, n * sizeof(u64));
4457 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4459 static void perf_output_read_group(struct perf_output_handle *handle,
4460 struct perf_event *event,
4461 u64 enabled, u64 running)
4463 struct perf_event *leader = event->group_leader, *sub;
4464 u64 read_format = event->attr.read_format;
4468 values[n++] = 1 + leader->nr_siblings;
4470 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4471 values[n++] = enabled;
4473 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4474 values[n++] = running;
4476 if (leader != event)
4477 leader->pmu->read(leader);
4479 values[n++] = perf_event_count(leader);
4480 if (read_format & PERF_FORMAT_ID)
4481 values[n++] = primary_event_id(leader);
4483 __output_copy(handle, values, n * sizeof(u64));
4485 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4488 if ((sub != event) &&
4489 (sub->state == PERF_EVENT_STATE_ACTIVE))
4490 sub->pmu->read(sub);
4492 values[n++] = perf_event_count(sub);
4493 if (read_format & PERF_FORMAT_ID)
4494 values[n++] = primary_event_id(sub);
4496 __output_copy(handle, values, n * sizeof(u64));
4500 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4501 PERF_FORMAT_TOTAL_TIME_RUNNING)
4503 static void perf_output_read(struct perf_output_handle *handle,
4504 struct perf_event *event)
4506 u64 enabled = 0, running = 0, now;
4507 u64 read_format = event->attr.read_format;
4510 * compute total_time_enabled, total_time_running
4511 * based on snapshot values taken when the event
4512 * was last scheduled in.
4514 * we cannot simply called update_context_time()
4515 * because of locking issue as we are called in
4518 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4519 calc_timer_values(event, &now, &enabled, &running);
4521 if (event->attr.read_format & PERF_FORMAT_GROUP)
4522 perf_output_read_group(handle, event, enabled, running);
4524 perf_output_read_one(handle, event, enabled, running);
4527 void perf_output_sample(struct perf_output_handle *handle,
4528 struct perf_event_header *header,
4529 struct perf_sample_data *data,
4530 struct perf_event *event)
4532 u64 sample_type = data->type;
4534 perf_output_put(handle, *header);
4536 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4537 perf_output_put(handle, data->id);
4539 if (sample_type & PERF_SAMPLE_IP)
4540 perf_output_put(handle, data->ip);
4542 if (sample_type & PERF_SAMPLE_TID)
4543 perf_output_put(handle, data->tid_entry);
4545 if (sample_type & PERF_SAMPLE_TIME)
4546 perf_output_put(handle, data->time);
4548 if (sample_type & PERF_SAMPLE_ADDR)
4549 perf_output_put(handle, data->addr);
4551 if (sample_type & PERF_SAMPLE_ID)
4552 perf_output_put(handle, data->id);
4554 if (sample_type & PERF_SAMPLE_STREAM_ID)
4555 perf_output_put(handle, data->stream_id);
4557 if (sample_type & PERF_SAMPLE_CPU)
4558 perf_output_put(handle, data->cpu_entry);
4560 if (sample_type & PERF_SAMPLE_PERIOD)
4561 perf_output_put(handle, data->period);
4563 if (sample_type & PERF_SAMPLE_READ)
4564 perf_output_read(handle, event);
4566 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4567 if (data->callchain) {
4570 if (data->callchain)
4571 size += data->callchain->nr;
4573 size *= sizeof(u64);
4575 __output_copy(handle, data->callchain, size);
4578 perf_output_put(handle, nr);
4582 if (sample_type & PERF_SAMPLE_RAW) {
4584 perf_output_put(handle, data->raw->size);
4585 __output_copy(handle, data->raw->data,
4592 .size = sizeof(u32),
4595 perf_output_put(handle, raw);
4599 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4600 if (data->br_stack) {
4603 size = data->br_stack->nr
4604 * sizeof(struct perf_branch_entry);
4606 perf_output_put(handle, data->br_stack->nr);
4607 perf_output_copy(handle, data->br_stack->entries, size);
4610 * we always store at least the value of nr
4613 perf_output_put(handle, nr);
4617 if (sample_type & PERF_SAMPLE_REGS_USER) {
4618 u64 abi = data->regs_user.abi;
4621 * If there are no regs to dump, notice it through
4622 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4624 perf_output_put(handle, abi);
4627 u64 mask = event->attr.sample_regs_user;
4628 perf_output_sample_regs(handle,
4629 data->regs_user.regs,
4634 if (sample_type & PERF_SAMPLE_STACK_USER) {
4635 perf_output_sample_ustack(handle,
4636 data->stack_user_size,
4637 data->regs_user.regs);
4640 if (sample_type & PERF_SAMPLE_WEIGHT)
4641 perf_output_put(handle, data->weight);
4643 if (sample_type & PERF_SAMPLE_DATA_SRC)
4644 perf_output_put(handle, data->data_src.val);
4646 if (sample_type & PERF_SAMPLE_TRANSACTION)
4647 perf_output_put(handle, data->txn);
4649 if (!event->attr.watermark) {
4650 int wakeup_events = event->attr.wakeup_events;
4652 if (wakeup_events) {
4653 struct ring_buffer *rb = handle->rb;
4654 int events = local_inc_return(&rb->events);
4656 if (events >= wakeup_events) {
4657 local_sub(wakeup_events, &rb->events);
4658 local_inc(&rb->wakeup);
4664 void perf_prepare_sample(struct perf_event_header *header,
4665 struct perf_sample_data *data,
4666 struct perf_event *event,
4667 struct pt_regs *regs)
4669 u64 sample_type = event->attr.sample_type;
4671 header->type = PERF_RECORD_SAMPLE;
4672 header->size = sizeof(*header) + event->header_size;
4675 header->misc |= perf_misc_flags(regs);
4677 __perf_event_header__init_id(header, data, event);
4679 if (sample_type & PERF_SAMPLE_IP)
4680 data->ip = perf_instruction_pointer(regs);
4682 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4685 data->callchain = perf_callchain(event, regs);
4687 if (data->callchain)
4688 size += data->callchain->nr;
4690 header->size += size * sizeof(u64);
4693 if (sample_type & PERF_SAMPLE_RAW) {
4694 int size = sizeof(u32);
4697 size += data->raw->size;
4699 size += sizeof(u32);
4701 WARN_ON_ONCE(size & (sizeof(u64)-1));
4702 header->size += size;
4705 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4706 int size = sizeof(u64); /* nr */
4707 if (data->br_stack) {
4708 size += data->br_stack->nr
4709 * sizeof(struct perf_branch_entry);
4711 header->size += size;
4714 if (sample_type & PERF_SAMPLE_REGS_USER) {
4715 /* regs dump ABI info */
4716 int size = sizeof(u64);
4718 perf_sample_regs_user(&data->regs_user, regs);
4720 if (data->regs_user.regs) {
4721 u64 mask = event->attr.sample_regs_user;
4722 size += hweight64(mask) * sizeof(u64);
4725 header->size += size;
4728 if (sample_type & PERF_SAMPLE_STACK_USER) {
4730 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4731 * processed as the last one or have additional check added
4732 * in case new sample type is added, because we could eat
4733 * up the rest of the sample size.
4735 struct perf_regs_user *uregs = &data->regs_user;
4736 u16 stack_size = event->attr.sample_stack_user;
4737 u16 size = sizeof(u64);
4740 perf_sample_regs_user(uregs, regs);
4742 stack_size = perf_sample_ustack_size(stack_size, header->size,
4746 * If there is something to dump, add space for the dump
4747 * itself and for the field that tells the dynamic size,
4748 * which is how many have been actually dumped.
4751 size += sizeof(u64) + stack_size;
4753 data->stack_user_size = stack_size;
4754 header->size += size;
4758 static void perf_event_output(struct perf_event *event,
4759 struct perf_sample_data *data,
4760 struct pt_regs *regs)
4762 struct perf_output_handle handle;
4763 struct perf_event_header header;
4765 /* protect the callchain buffers */
4768 perf_prepare_sample(&header, data, event, regs);
4770 if (perf_output_begin(&handle, event, header.size))
4773 perf_output_sample(&handle, &header, data, event);
4775 perf_output_end(&handle);
4785 struct perf_read_event {
4786 struct perf_event_header header;
4793 perf_event_read_event(struct perf_event *event,
4794 struct task_struct *task)
4796 struct perf_output_handle handle;
4797 struct perf_sample_data sample;
4798 struct perf_read_event read_event = {
4800 .type = PERF_RECORD_READ,
4802 .size = sizeof(read_event) + event->read_size,
4804 .pid = perf_event_pid(event, task),
4805 .tid = perf_event_tid(event, task),
4809 perf_event_header__init_id(&read_event.header, &sample, event);
4810 ret = perf_output_begin(&handle, event, read_event.header.size);
4814 perf_output_put(&handle, read_event);
4815 perf_output_read(&handle, event);
4816 perf_event__output_id_sample(event, &handle, &sample);
4818 perf_output_end(&handle);
4821 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4824 perf_event_aux_ctx(struct perf_event_context *ctx,
4825 perf_event_aux_output_cb output,
4828 struct perf_event *event;
4830 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4831 if (event->state < PERF_EVENT_STATE_INACTIVE)
4833 if (!event_filter_match(event))
4835 output(event, data);
4840 perf_event_aux(perf_event_aux_output_cb output, void *data,
4841 struct perf_event_context *task_ctx)
4843 struct perf_cpu_context *cpuctx;
4844 struct perf_event_context *ctx;
4849 list_for_each_entry_rcu(pmu, &pmus, entry) {
4850 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4851 if (cpuctx->unique_pmu != pmu)
4853 perf_event_aux_ctx(&cpuctx->ctx, output, data);
4856 ctxn = pmu->task_ctx_nr;
4859 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4861 perf_event_aux_ctx(ctx, output, data);
4863 put_cpu_ptr(pmu->pmu_cpu_context);
4868 perf_event_aux_ctx(task_ctx, output, data);
4875 * task tracking -- fork/exit
4877 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4880 struct perf_task_event {
4881 struct task_struct *task;
4882 struct perf_event_context *task_ctx;
4885 struct perf_event_header header;
4895 static int perf_event_task_match(struct perf_event *event)
4897 return event->attr.comm || event->attr.mmap ||
4898 event->attr.mmap2 || event->attr.mmap_data ||
4902 static void perf_event_task_output(struct perf_event *event,
4905 struct perf_task_event *task_event = data;
4906 struct perf_output_handle handle;
4907 struct perf_sample_data sample;
4908 struct task_struct *task = task_event->task;
4909 int ret, size = task_event->event_id.header.size;
4911 if (!perf_event_task_match(event))
4914 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4916 ret = perf_output_begin(&handle, event,
4917 task_event->event_id.header.size);
4921 task_event->event_id.pid = perf_event_pid(event, task);
4922 task_event->event_id.ppid = perf_event_pid(event, current);
4924 task_event->event_id.tid = perf_event_tid(event, task);
4925 task_event->event_id.ptid = perf_event_tid(event, current);
4927 perf_output_put(&handle, task_event->event_id);
4929 perf_event__output_id_sample(event, &handle, &sample);
4931 perf_output_end(&handle);
4933 task_event->event_id.header.size = size;
4936 static void perf_event_task(struct task_struct *task,
4937 struct perf_event_context *task_ctx,
4940 struct perf_task_event task_event;
4942 if (!atomic_read(&nr_comm_events) &&
4943 !atomic_read(&nr_mmap_events) &&
4944 !atomic_read(&nr_task_events))
4947 task_event = (struct perf_task_event){
4949 .task_ctx = task_ctx,
4952 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4954 .size = sizeof(task_event.event_id),
4960 .time = perf_clock(),
4964 perf_event_aux(perf_event_task_output,
4969 void perf_event_fork(struct task_struct *task)
4971 perf_event_task(task, NULL, 1);
4978 struct perf_comm_event {
4979 struct task_struct *task;
4984 struct perf_event_header header;
4991 static int perf_event_comm_match(struct perf_event *event)
4993 return event->attr.comm;
4996 static void perf_event_comm_output(struct perf_event *event,
4999 struct perf_comm_event *comm_event = data;
5000 struct perf_output_handle handle;
5001 struct perf_sample_data sample;
5002 int size = comm_event->event_id.header.size;
5005 if (!perf_event_comm_match(event))
5008 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5009 ret = perf_output_begin(&handle, event,
5010 comm_event->event_id.header.size);
5015 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5016 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5018 perf_output_put(&handle, comm_event->event_id);
5019 __output_copy(&handle, comm_event->comm,
5020 comm_event->comm_size);
5022 perf_event__output_id_sample(event, &handle, &sample);
5024 perf_output_end(&handle);
5026 comm_event->event_id.header.size = size;
5029 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5031 char comm[TASK_COMM_LEN];
5034 memset(comm, 0, sizeof(comm));
5035 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5036 size = ALIGN(strlen(comm)+1, sizeof(u64));
5038 comm_event->comm = comm;
5039 comm_event->comm_size = size;
5041 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5043 perf_event_aux(perf_event_comm_output,
5048 void perf_event_comm(struct task_struct *task)
5050 struct perf_comm_event comm_event;
5051 struct perf_event_context *ctx;
5055 for_each_task_context_nr(ctxn) {
5056 ctx = task->perf_event_ctxp[ctxn];
5060 perf_event_enable_on_exec(ctx);
5064 if (!atomic_read(&nr_comm_events))
5067 comm_event = (struct perf_comm_event){
5073 .type = PERF_RECORD_COMM,
5082 perf_event_comm_event(&comm_event);
5089 struct perf_mmap_event {
5090 struct vm_area_struct *vma;
5092 const char *file_name;
5099 struct perf_event_header header;
5109 static int perf_event_mmap_match(struct perf_event *event,
5112 struct perf_mmap_event *mmap_event = data;
5113 struct vm_area_struct *vma = mmap_event->vma;
5114 int executable = vma->vm_flags & VM_EXEC;
5116 return (!executable && event->attr.mmap_data) ||
5117 (executable && (event->attr.mmap || event->attr.mmap2));
5120 static void perf_event_mmap_output(struct perf_event *event,
5123 struct perf_mmap_event *mmap_event = data;
5124 struct perf_output_handle handle;
5125 struct perf_sample_data sample;
5126 int size = mmap_event->event_id.header.size;
5129 if (!perf_event_mmap_match(event, data))
5132 if (event->attr.mmap2) {
5133 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5134 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5135 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5136 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5137 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5140 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5141 ret = perf_output_begin(&handle, event,
5142 mmap_event->event_id.header.size);
5146 mmap_event->event_id.pid = perf_event_pid(event, current);
5147 mmap_event->event_id.tid = perf_event_tid(event, current);
5149 perf_output_put(&handle, mmap_event->event_id);
5151 if (event->attr.mmap2) {
5152 perf_output_put(&handle, mmap_event->maj);
5153 perf_output_put(&handle, mmap_event->min);
5154 perf_output_put(&handle, mmap_event->ino);
5155 perf_output_put(&handle, mmap_event->ino_generation);
5158 __output_copy(&handle, mmap_event->file_name,
5159 mmap_event->file_size);
5161 perf_event__output_id_sample(event, &handle, &sample);
5163 perf_output_end(&handle);
5165 mmap_event->event_id.header.size = size;
5168 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5170 struct vm_area_struct *vma = mmap_event->vma;
5171 struct file *file = vma->vm_file;
5172 int maj = 0, min = 0;
5173 u64 ino = 0, gen = 0;
5180 struct inode *inode;
5183 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5189 * d_path() works from the end of the rb backwards, so we
5190 * need to add enough zero bytes after the string to handle
5191 * the 64bit alignment we do later.
5193 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5198 inode = file_inode(vma->vm_file);
5199 dev = inode->i_sb->s_dev;
5201 gen = inode->i_generation;
5206 name = (char *)arch_vma_name(vma);
5210 if (vma->vm_start <= vma->vm_mm->start_brk &&
5211 vma->vm_end >= vma->vm_mm->brk) {
5215 if (vma->vm_start <= vma->vm_mm->start_stack &&
5216 vma->vm_end >= vma->vm_mm->start_stack) {
5226 strlcpy(tmp, name, sizeof(tmp));
5230 * Since our buffer works in 8 byte units we need to align our string
5231 * size to a multiple of 8. However, we must guarantee the tail end is
5232 * zero'd out to avoid leaking random bits to userspace.
5234 size = strlen(name)+1;
5235 while (!IS_ALIGNED(size, sizeof(u64)))
5236 name[size++] = '\0';
5238 mmap_event->file_name = name;
5239 mmap_event->file_size = size;
5240 mmap_event->maj = maj;
5241 mmap_event->min = min;
5242 mmap_event->ino = ino;
5243 mmap_event->ino_generation = gen;
5245 if (!(vma->vm_flags & VM_EXEC))
5246 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5248 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5250 perf_event_aux(perf_event_mmap_output,
5257 void perf_event_mmap(struct vm_area_struct *vma)
5259 struct perf_mmap_event mmap_event;
5261 if (!atomic_read(&nr_mmap_events))
5264 mmap_event = (struct perf_mmap_event){
5270 .type = PERF_RECORD_MMAP,
5271 .misc = PERF_RECORD_MISC_USER,
5276 .start = vma->vm_start,
5277 .len = vma->vm_end - vma->vm_start,
5278 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5280 /* .maj (attr_mmap2 only) */
5281 /* .min (attr_mmap2 only) */
5282 /* .ino (attr_mmap2 only) */
5283 /* .ino_generation (attr_mmap2 only) */
5286 perf_event_mmap_event(&mmap_event);
5290 * IRQ throttle logging
5293 static void perf_log_throttle(struct perf_event *event, int enable)
5295 struct perf_output_handle handle;
5296 struct perf_sample_data sample;
5300 struct perf_event_header header;
5304 } throttle_event = {
5306 .type = PERF_RECORD_THROTTLE,
5308 .size = sizeof(throttle_event),
5310 .time = perf_clock(),
5311 .id = primary_event_id(event),
5312 .stream_id = event->id,
5316 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5318 perf_event_header__init_id(&throttle_event.header, &sample, event);
5320 ret = perf_output_begin(&handle, event,
5321 throttle_event.header.size);
5325 perf_output_put(&handle, throttle_event);
5326 perf_event__output_id_sample(event, &handle, &sample);
5327 perf_output_end(&handle);
5331 * Generic event overflow handling, sampling.
5334 static int __perf_event_overflow(struct perf_event *event,
5335 int throttle, struct perf_sample_data *data,
5336 struct pt_regs *regs)
5338 int events = atomic_read(&event->event_limit);
5339 struct hw_perf_event *hwc = &event->hw;
5344 * Non-sampling counters might still use the PMI to fold short
5345 * hardware counters, ignore those.
5347 if (unlikely(!is_sampling_event(event)))
5350 seq = __this_cpu_read(perf_throttled_seq);
5351 if (seq != hwc->interrupts_seq) {
5352 hwc->interrupts_seq = seq;
5353 hwc->interrupts = 1;
5356 if (unlikely(throttle
5357 && hwc->interrupts >= max_samples_per_tick)) {
5358 __this_cpu_inc(perf_throttled_count);
5359 hwc->interrupts = MAX_INTERRUPTS;
5360 perf_log_throttle(event, 0);
5361 tick_nohz_full_kick();
5366 if (event->attr.freq) {
5367 u64 now = perf_clock();
5368 s64 delta = now - hwc->freq_time_stamp;
5370 hwc->freq_time_stamp = now;
5372 if (delta > 0 && delta < 2*TICK_NSEC)
5373 perf_adjust_period(event, delta, hwc->last_period, true);
5377 * XXX event_limit might not quite work as expected on inherited
5381 event->pending_kill = POLL_IN;
5382 if (events && atomic_dec_and_test(&event->event_limit)) {
5384 event->pending_kill = POLL_HUP;
5385 event->pending_disable = 1;
5386 irq_work_queue(&event->pending);
5389 if (event->overflow_handler)
5390 event->overflow_handler(event, data, regs);
5392 perf_event_output(event, data, regs);
5394 if (event->fasync && event->pending_kill) {
5395 event->pending_wakeup = 1;
5396 irq_work_queue(&event->pending);
5402 int perf_event_overflow(struct perf_event *event,
5403 struct perf_sample_data *data,
5404 struct pt_regs *regs)
5406 return __perf_event_overflow(event, 1, data, regs);
5410 * Generic software event infrastructure
5413 struct swevent_htable {
5414 struct swevent_hlist *swevent_hlist;
5415 struct mutex hlist_mutex;
5418 /* Recursion avoidance in each contexts */
5419 int recursion[PERF_NR_CONTEXTS];
5421 /* Keeps track of cpu being initialized/exited */
5425 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5428 * We directly increment event->count and keep a second value in
5429 * event->hw.period_left to count intervals. This period event
5430 * is kept in the range [-sample_period, 0] so that we can use the
5434 u64 perf_swevent_set_period(struct perf_event *event)
5436 struct hw_perf_event *hwc = &event->hw;
5437 u64 period = hwc->last_period;
5441 hwc->last_period = hwc->sample_period;
5444 old = val = local64_read(&hwc->period_left);
5448 nr = div64_u64(period + val, period);
5449 offset = nr * period;
5451 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5457 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5458 struct perf_sample_data *data,
5459 struct pt_regs *regs)
5461 struct hw_perf_event *hwc = &event->hw;
5465 overflow = perf_swevent_set_period(event);
5467 if (hwc->interrupts == MAX_INTERRUPTS)
5470 for (; overflow; overflow--) {
5471 if (__perf_event_overflow(event, throttle,
5474 * We inhibit the overflow from happening when
5475 * hwc->interrupts == MAX_INTERRUPTS.
5483 static void perf_swevent_event(struct perf_event *event, u64 nr,
5484 struct perf_sample_data *data,
5485 struct pt_regs *regs)
5487 struct hw_perf_event *hwc = &event->hw;
5489 local64_add(nr, &event->count);
5494 if (!is_sampling_event(event))
5497 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5499 return perf_swevent_overflow(event, 1, data, regs);
5501 data->period = event->hw.last_period;
5503 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5504 return perf_swevent_overflow(event, 1, data, regs);
5506 if (local64_add_negative(nr, &hwc->period_left))
5509 perf_swevent_overflow(event, 0, data, regs);
5512 static int perf_exclude_event(struct perf_event *event,
5513 struct pt_regs *regs)
5515 if (event->hw.state & PERF_HES_STOPPED)
5519 if (event->attr.exclude_user && user_mode(regs))
5522 if (event->attr.exclude_kernel && !user_mode(regs))
5529 static int perf_swevent_match(struct perf_event *event,
5530 enum perf_type_id type,
5532 struct perf_sample_data *data,
5533 struct pt_regs *regs)
5535 if (event->attr.type != type)
5538 if (event->attr.config != event_id)
5541 if (perf_exclude_event(event, regs))
5547 static inline u64 swevent_hash(u64 type, u32 event_id)
5549 u64 val = event_id | (type << 32);
5551 return hash_64(val, SWEVENT_HLIST_BITS);
5554 static inline struct hlist_head *
5555 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5557 u64 hash = swevent_hash(type, event_id);
5559 return &hlist->heads[hash];
5562 /* For the read side: events when they trigger */
5563 static inline struct hlist_head *
5564 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5566 struct swevent_hlist *hlist;
5568 hlist = rcu_dereference(swhash->swevent_hlist);
5572 return __find_swevent_head(hlist, type, event_id);
5575 /* For the event head insertion and removal in the hlist */
5576 static inline struct hlist_head *
5577 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5579 struct swevent_hlist *hlist;
5580 u32 event_id = event->attr.config;
5581 u64 type = event->attr.type;
5584 * Event scheduling is always serialized against hlist allocation
5585 * and release. Which makes the protected version suitable here.
5586 * The context lock guarantees that.
5588 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5589 lockdep_is_held(&event->ctx->lock));
5593 return __find_swevent_head(hlist, type, event_id);
5596 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5598 struct perf_sample_data *data,
5599 struct pt_regs *regs)
5601 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5602 struct perf_event *event;
5603 struct hlist_head *head;
5606 head = find_swevent_head_rcu(swhash, type, event_id);
5610 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5611 if (perf_swevent_match(event, type, event_id, data, regs))
5612 perf_swevent_event(event, nr, data, regs);
5618 int perf_swevent_get_recursion_context(void)
5620 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5622 return get_recursion_context(swhash->recursion);
5624 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5626 inline void perf_swevent_put_recursion_context(int rctx)
5628 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5630 put_recursion_context(swhash->recursion, rctx);
5633 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5635 struct perf_sample_data data;
5638 preempt_disable_notrace();
5639 rctx = perf_swevent_get_recursion_context();
5643 perf_sample_data_init(&data, addr, 0);
5645 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5647 perf_swevent_put_recursion_context(rctx);
5648 preempt_enable_notrace();
5651 static void perf_swevent_read(struct perf_event *event)
5655 static int perf_swevent_add(struct perf_event *event, int flags)
5657 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5658 struct hw_perf_event *hwc = &event->hw;
5659 struct hlist_head *head;
5661 if (is_sampling_event(event)) {
5662 hwc->last_period = hwc->sample_period;
5663 perf_swevent_set_period(event);
5666 hwc->state = !(flags & PERF_EF_START);
5668 head = find_swevent_head(swhash, event);
5671 * We can race with cpu hotplug code. Do not
5672 * WARN if the cpu just got unplugged.
5674 WARN_ON_ONCE(swhash->online);
5678 hlist_add_head_rcu(&event->hlist_entry, head);
5683 static void perf_swevent_del(struct perf_event *event, int flags)
5685 hlist_del_rcu(&event->hlist_entry);
5688 static void perf_swevent_start(struct perf_event *event, int flags)
5690 event->hw.state = 0;
5693 static void perf_swevent_stop(struct perf_event *event, int flags)
5695 event->hw.state = PERF_HES_STOPPED;
5698 /* Deref the hlist from the update side */
5699 static inline struct swevent_hlist *
5700 swevent_hlist_deref(struct swevent_htable *swhash)
5702 return rcu_dereference_protected(swhash->swevent_hlist,
5703 lockdep_is_held(&swhash->hlist_mutex));
5706 static void swevent_hlist_release(struct swevent_htable *swhash)
5708 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5713 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5714 kfree_rcu(hlist, rcu_head);
5717 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5719 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5721 mutex_lock(&swhash->hlist_mutex);
5723 if (!--swhash->hlist_refcount)
5724 swevent_hlist_release(swhash);
5726 mutex_unlock(&swhash->hlist_mutex);
5729 static void swevent_hlist_put(struct perf_event *event)
5733 for_each_possible_cpu(cpu)
5734 swevent_hlist_put_cpu(event, cpu);
5737 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5739 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5742 mutex_lock(&swhash->hlist_mutex);
5744 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5745 struct swevent_hlist *hlist;
5747 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5752 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5754 swhash->hlist_refcount++;
5756 mutex_unlock(&swhash->hlist_mutex);
5761 static int swevent_hlist_get(struct perf_event *event)
5764 int cpu, failed_cpu;
5767 for_each_possible_cpu(cpu) {
5768 err = swevent_hlist_get_cpu(event, cpu);
5778 for_each_possible_cpu(cpu) {
5779 if (cpu == failed_cpu)
5781 swevent_hlist_put_cpu(event, cpu);
5788 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5790 static void sw_perf_event_destroy(struct perf_event *event)
5792 u64 event_id = event->attr.config;
5794 WARN_ON(event->parent);
5796 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5797 swevent_hlist_put(event);
5800 static int perf_swevent_init(struct perf_event *event)
5802 u64 event_id = event->attr.config;
5804 if (event->attr.type != PERF_TYPE_SOFTWARE)
5808 * no branch sampling for software events
5810 if (has_branch_stack(event))
5814 case PERF_COUNT_SW_CPU_CLOCK:
5815 case PERF_COUNT_SW_TASK_CLOCK:
5822 if (event_id >= PERF_COUNT_SW_MAX)
5825 if (!event->parent) {
5828 err = swevent_hlist_get(event);
5832 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5833 event->destroy = sw_perf_event_destroy;
5839 static int perf_swevent_event_idx(struct perf_event *event)
5844 static struct pmu perf_swevent = {
5845 .task_ctx_nr = perf_sw_context,
5847 .event_init = perf_swevent_init,
5848 .add = perf_swevent_add,
5849 .del = perf_swevent_del,
5850 .start = perf_swevent_start,
5851 .stop = perf_swevent_stop,
5852 .read = perf_swevent_read,
5854 .event_idx = perf_swevent_event_idx,
5857 #ifdef CONFIG_EVENT_TRACING
5859 static int perf_tp_filter_match(struct perf_event *event,
5860 struct perf_sample_data *data)
5862 void *record = data->raw->data;
5864 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5869 static int perf_tp_event_match(struct perf_event *event,
5870 struct perf_sample_data *data,
5871 struct pt_regs *regs)
5873 if (event->hw.state & PERF_HES_STOPPED)
5876 * All tracepoints are from kernel-space.
5878 if (event->attr.exclude_kernel)
5881 if (!perf_tp_filter_match(event, data))
5887 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5888 struct pt_regs *regs, struct hlist_head *head, int rctx,
5889 struct task_struct *task)
5891 struct perf_sample_data data;
5892 struct perf_event *event;
5894 struct perf_raw_record raw = {
5899 perf_sample_data_init(&data, addr, 0);
5902 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5903 if (perf_tp_event_match(event, &data, regs))
5904 perf_swevent_event(event, count, &data, regs);
5908 * If we got specified a target task, also iterate its context and
5909 * deliver this event there too.
5911 if (task && task != current) {
5912 struct perf_event_context *ctx;
5913 struct trace_entry *entry = record;
5916 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5920 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5921 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5923 if (event->attr.config != entry->type)
5925 if (perf_tp_event_match(event, &data, regs))
5926 perf_swevent_event(event, count, &data, regs);
5932 perf_swevent_put_recursion_context(rctx);
5934 EXPORT_SYMBOL_GPL(perf_tp_event);
5936 static void tp_perf_event_destroy(struct perf_event *event)
5938 perf_trace_destroy(event);
5941 static int perf_tp_event_init(struct perf_event *event)
5945 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5949 * no branch sampling for tracepoint events
5951 if (has_branch_stack(event))
5954 err = perf_trace_init(event);
5958 event->destroy = tp_perf_event_destroy;
5963 static struct pmu perf_tracepoint = {
5964 .task_ctx_nr = perf_sw_context,
5966 .event_init = perf_tp_event_init,
5967 .add = perf_trace_add,
5968 .del = perf_trace_del,
5969 .start = perf_swevent_start,
5970 .stop = perf_swevent_stop,
5971 .read = perf_swevent_read,
5973 .event_idx = perf_swevent_event_idx,
5976 static inline void perf_tp_register(void)
5978 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5981 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5986 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5989 filter_str = strndup_user(arg, PAGE_SIZE);
5990 if (IS_ERR(filter_str))
5991 return PTR_ERR(filter_str);
5993 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5999 static void perf_event_free_filter(struct perf_event *event)
6001 ftrace_profile_free_filter(event);
6006 static inline void perf_tp_register(void)
6010 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6015 static void perf_event_free_filter(struct perf_event *event)
6019 #endif /* CONFIG_EVENT_TRACING */
6021 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6022 void perf_bp_event(struct perf_event *bp, void *data)
6024 struct perf_sample_data sample;
6025 struct pt_regs *regs = data;
6027 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6029 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6030 perf_swevent_event(bp, 1, &sample, regs);
6035 * hrtimer based swevent callback
6038 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6040 enum hrtimer_restart ret = HRTIMER_RESTART;
6041 struct perf_sample_data data;
6042 struct pt_regs *regs;
6043 struct perf_event *event;
6046 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6048 if (event->state != PERF_EVENT_STATE_ACTIVE)
6049 return HRTIMER_NORESTART;
6051 event->pmu->read(event);
6053 perf_sample_data_init(&data, 0, event->hw.last_period);
6054 regs = get_irq_regs();
6056 if (regs && !perf_exclude_event(event, regs)) {
6057 if (!(event->attr.exclude_idle && is_idle_task(current)))
6058 if (__perf_event_overflow(event, 1, &data, regs))
6059 ret = HRTIMER_NORESTART;
6062 period = max_t(u64, 10000, event->hw.sample_period);
6063 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6068 static void perf_swevent_start_hrtimer(struct perf_event *event)
6070 struct hw_perf_event *hwc = &event->hw;
6073 if (!is_sampling_event(event))
6076 period = local64_read(&hwc->period_left);
6081 local64_set(&hwc->period_left, 0);
6083 period = max_t(u64, 10000, hwc->sample_period);
6085 __hrtimer_start_range_ns(&hwc->hrtimer,
6086 ns_to_ktime(period), 0,
6087 HRTIMER_MODE_REL_PINNED, 0);
6090 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6092 struct hw_perf_event *hwc = &event->hw;
6094 if (is_sampling_event(event)) {
6095 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6096 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6098 hrtimer_cancel(&hwc->hrtimer);
6102 static void perf_swevent_init_hrtimer(struct perf_event *event)
6104 struct hw_perf_event *hwc = &event->hw;
6106 if (!is_sampling_event(event))
6109 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6110 hwc->hrtimer.function = perf_swevent_hrtimer;
6113 * Since hrtimers have a fixed rate, we can do a static freq->period
6114 * mapping and avoid the whole period adjust feedback stuff.
6116 if (event->attr.freq) {
6117 long freq = event->attr.sample_freq;
6119 event->attr.sample_period = NSEC_PER_SEC / freq;
6120 hwc->sample_period = event->attr.sample_period;
6121 local64_set(&hwc->period_left, hwc->sample_period);
6122 hwc->last_period = hwc->sample_period;
6123 event->attr.freq = 0;
6128 * Software event: cpu wall time clock
6131 static void cpu_clock_event_update(struct perf_event *event)
6136 now = local_clock();
6137 prev = local64_xchg(&event->hw.prev_count, now);
6138 local64_add(now - prev, &event->count);
6141 static void cpu_clock_event_start(struct perf_event *event, int flags)
6143 local64_set(&event->hw.prev_count, local_clock());
6144 perf_swevent_start_hrtimer(event);
6147 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6149 perf_swevent_cancel_hrtimer(event);
6150 cpu_clock_event_update(event);
6153 static int cpu_clock_event_add(struct perf_event *event, int flags)
6155 if (flags & PERF_EF_START)
6156 cpu_clock_event_start(event, flags);
6161 static void cpu_clock_event_del(struct perf_event *event, int flags)
6163 cpu_clock_event_stop(event, flags);
6166 static void cpu_clock_event_read(struct perf_event *event)
6168 cpu_clock_event_update(event);
6171 static int cpu_clock_event_init(struct perf_event *event)
6173 if (event->attr.type != PERF_TYPE_SOFTWARE)
6176 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6180 * no branch sampling for software events
6182 if (has_branch_stack(event))
6185 perf_swevent_init_hrtimer(event);
6190 static struct pmu perf_cpu_clock = {
6191 .task_ctx_nr = perf_sw_context,
6193 .event_init = cpu_clock_event_init,
6194 .add = cpu_clock_event_add,
6195 .del = cpu_clock_event_del,
6196 .start = cpu_clock_event_start,
6197 .stop = cpu_clock_event_stop,
6198 .read = cpu_clock_event_read,
6200 .event_idx = perf_swevent_event_idx,
6204 * Software event: task time clock
6207 static void task_clock_event_update(struct perf_event *event, u64 now)
6212 prev = local64_xchg(&event->hw.prev_count, now);
6214 local64_add(delta, &event->count);
6217 static void task_clock_event_start(struct perf_event *event, int flags)
6219 local64_set(&event->hw.prev_count, event->ctx->time);
6220 perf_swevent_start_hrtimer(event);
6223 static void task_clock_event_stop(struct perf_event *event, int flags)
6225 perf_swevent_cancel_hrtimer(event);
6226 task_clock_event_update(event, event->ctx->time);
6229 static int task_clock_event_add(struct perf_event *event, int flags)
6231 if (flags & PERF_EF_START)
6232 task_clock_event_start(event, flags);
6237 static void task_clock_event_del(struct perf_event *event, int flags)
6239 task_clock_event_stop(event, PERF_EF_UPDATE);
6242 static void task_clock_event_read(struct perf_event *event)
6244 u64 now = perf_clock();
6245 u64 delta = now - event->ctx->timestamp;
6246 u64 time = event->ctx->time + delta;
6248 task_clock_event_update(event, time);
6251 static int task_clock_event_init(struct perf_event *event)
6253 if (event->attr.type != PERF_TYPE_SOFTWARE)
6256 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6260 * no branch sampling for software events
6262 if (has_branch_stack(event))
6265 perf_swevent_init_hrtimer(event);
6270 static struct pmu perf_task_clock = {
6271 .task_ctx_nr = perf_sw_context,
6273 .event_init = task_clock_event_init,
6274 .add = task_clock_event_add,
6275 .del = task_clock_event_del,
6276 .start = task_clock_event_start,
6277 .stop = task_clock_event_stop,
6278 .read = task_clock_event_read,
6280 .event_idx = perf_swevent_event_idx,
6283 static void perf_pmu_nop_void(struct pmu *pmu)
6287 static int perf_pmu_nop_int(struct pmu *pmu)
6292 static void perf_pmu_start_txn(struct pmu *pmu)
6294 perf_pmu_disable(pmu);
6297 static int perf_pmu_commit_txn(struct pmu *pmu)
6299 perf_pmu_enable(pmu);
6303 static void perf_pmu_cancel_txn(struct pmu *pmu)
6305 perf_pmu_enable(pmu);
6308 static int perf_event_idx_default(struct perf_event *event)
6310 return event->hw.idx + 1;
6314 * Ensures all contexts with the same task_ctx_nr have the same
6315 * pmu_cpu_context too.
6317 static void *find_pmu_context(int ctxn)
6324 list_for_each_entry(pmu, &pmus, entry) {
6325 if (pmu->task_ctx_nr == ctxn)
6326 return pmu->pmu_cpu_context;
6332 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6336 for_each_possible_cpu(cpu) {
6337 struct perf_cpu_context *cpuctx;
6339 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6341 if (cpuctx->unique_pmu == old_pmu)
6342 cpuctx->unique_pmu = pmu;
6346 static void free_pmu_context(struct pmu *pmu)
6350 mutex_lock(&pmus_lock);
6352 * Like a real lame refcount.
6354 list_for_each_entry(i, &pmus, entry) {
6355 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6356 update_pmu_context(i, pmu);
6361 free_percpu(pmu->pmu_cpu_context);
6363 mutex_unlock(&pmus_lock);
6365 static struct idr pmu_idr;
6368 type_show(struct device *dev, struct device_attribute *attr, char *page)
6370 struct pmu *pmu = dev_get_drvdata(dev);
6372 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6374 static DEVICE_ATTR_RO(type);
6377 perf_event_mux_interval_ms_show(struct device *dev,
6378 struct device_attribute *attr,
6381 struct pmu *pmu = dev_get_drvdata(dev);
6383 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6387 perf_event_mux_interval_ms_store(struct device *dev,
6388 struct device_attribute *attr,
6389 const char *buf, size_t count)
6391 struct pmu *pmu = dev_get_drvdata(dev);
6392 int timer, cpu, ret;
6394 ret = kstrtoint(buf, 0, &timer);
6401 /* same value, noting to do */
6402 if (timer == pmu->hrtimer_interval_ms)
6405 pmu->hrtimer_interval_ms = timer;
6407 /* update all cpuctx for this PMU */
6408 for_each_possible_cpu(cpu) {
6409 struct perf_cpu_context *cpuctx;
6410 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6411 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6413 if (hrtimer_active(&cpuctx->hrtimer))
6414 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6419 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6421 static struct attribute *pmu_dev_attrs[] = {
6422 &dev_attr_type.attr,
6423 &dev_attr_perf_event_mux_interval_ms.attr,
6426 ATTRIBUTE_GROUPS(pmu_dev);
6428 static int pmu_bus_running;
6429 static struct bus_type pmu_bus = {
6430 .name = "event_source",
6431 .dev_groups = pmu_dev_groups,
6434 static void pmu_dev_release(struct device *dev)
6439 static int pmu_dev_alloc(struct pmu *pmu)
6443 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6447 pmu->dev->groups = pmu->attr_groups;
6448 device_initialize(pmu->dev);
6449 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6453 dev_set_drvdata(pmu->dev, pmu);
6454 pmu->dev->bus = &pmu_bus;
6455 pmu->dev->release = pmu_dev_release;
6456 ret = device_add(pmu->dev);
6464 put_device(pmu->dev);
6468 static struct lock_class_key cpuctx_mutex;
6469 static struct lock_class_key cpuctx_lock;
6471 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6475 mutex_lock(&pmus_lock);
6477 pmu->pmu_disable_count = alloc_percpu(int);
6478 if (!pmu->pmu_disable_count)
6487 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6495 if (pmu_bus_running) {
6496 ret = pmu_dev_alloc(pmu);
6502 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6503 if (pmu->pmu_cpu_context)
6504 goto got_cpu_context;
6507 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6508 if (!pmu->pmu_cpu_context)
6511 for_each_possible_cpu(cpu) {
6512 struct perf_cpu_context *cpuctx;
6514 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6515 __perf_event_init_context(&cpuctx->ctx);
6516 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6517 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6518 cpuctx->ctx.type = cpu_context;
6519 cpuctx->ctx.pmu = pmu;
6521 __perf_cpu_hrtimer_init(cpuctx, cpu);
6523 INIT_LIST_HEAD(&cpuctx->rotation_list);
6524 cpuctx->unique_pmu = pmu;
6528 if (!pmu->start_txn) {
6529 if (pmu->pmu_enable) {
6531 * If we have pmu_enable/pmu_disable calls, install
6532 * transaction stubs that use that to try and batch
6533 * hardware accesses.
6535 pmu->start_txn = perf_pmu_start_txn;
6536 pmu->commit_txn = perf_pmu_commit_txn;
6537 pmu->cancel_txn = perf_pmu_cancel_txn;
6539 pmu->start_txn = perf_pmu_nop_void;
6540 pmu->commit_txn = perf_pmu_nop_int;
6541 pmu->cancel_txn = perf_pmu_nop_void;
6545 if (!pmu->pmu_enable) {
6546 pmu->pmu_enable = perf_pmu_nop_void;
6547 pmu->pmu_disable = perf_pmu_nop_void;
6550 if (!pmu->event_idx)
6551 pmu->event_idx = perf_event_idx_default;
6553 list_add_rcu(&pmu->entry, &pmus);
6556 mutex_unlock(&pmus_lock);
6561 device_del(pmu->dev);
6562 put_device(pmu->dev);
6565 if (pmu->type >= PERF_TYPE_MAX)
6566 idr_remove(&pmu_idr, pmu->type);
6569 free_percpu(pmu->pmu_disable_count);
6573 void perf_pmu_unregister(struct pmu *pmu)
6575 mutex_lock(&pmus_lock);
6576 list_del_rcu(&pmu->entry);
6577 mutex_unlock(&pmus_lock);
6580 * We dereference the pmu list under both SRCU and regular RCU, so
6581 * synchronize against both of those.
6583 synchronize_srcu(&pmus_srcu);
6586 free_percpu(pmu->pmu_disable_count);
6587 if (pmu->type >= PERF_TYPE_MAX)
6588 idr_remove(&pmu_idr, pmu->type);
6589 device_del(pmu->dev);
6590 put_device(pmu->dev);
6591 free_pmu_context(pmu);
6594 struct pmu *perf_init_event(struct perf_event *event)
6596 struct pmu *pmu = NULL;
6600 idx = srcu_read_lock(&pmus_srcu);
6603 pmu = idr_find(&pmu_idr, event->attr.type);
6607 ret = pmu->event_init(event);
6613 list_for_each_entry_rcu(pmu, &pmus, entry) {
6615 ret = pmu->event_init(event);
6619 if (ret != -ENOENT) {
6624 pmu = ERR_PTR(-ENOENT);
6626 srcu_read_unlock(&pmus_srcu, idx);
6631 static void account_event_cpu(struct perf_event *event, int cpu)
6636 if (has_branch_stack(event)) {
6637 if (!(event->attach_state & PERF_ATTACH_TASK))
6638 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6640 if (is_cgroup_event(event))
6641 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6644 static void account_event(struct perf_event *event)
6649 if (event->attach_state & PERF_ATTACH_TASK)
6650 static_key_slow_inc(&perf_sched_events.key);
6651 if (event->attr.mmap || event->attr.mmap_data)
6652 atomic_inc(&nr_mmap_events);
6653 if (event->attr.comm)
6654 atomic_inc(&nr_comm_events);
6655 if (event->attr.task)
6656 atomic_inc(&nr_task_events);
6657 if (event->attr.freq) {
6658 if (atomic_inc_return(&nr_freq_events) == 1)
6659 tick_nohz_full_kick_all();
6661 if (has_branch_stack(event))
6662 static_key_slow_inc(&perf_sched_events.key);
6663 if (is_cgroup_event(event))
6664 static_key_slow_inc(&perf_sched_events.key);
6666 account_event_cpu(event, event->cpu);
6670 * Allocate and initialize a event structure
6672 static struct perf_event *
6673 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6674 struct task_struct *task,
6675 struct perf_event *group_leader,
6676 struct perf_event *parent_event,
6677 perf_overflow_handler_t overflow_handler,
6681 struct perf_event *event;
6682 struct hw_perf_event *hwc;
6685 if ((unsigned)cpu >= nr_cpu_ids) {
6686 if (!task || cpu != -1)
6687 return ERR_PTR(-EINVAL);
6690 event = kzalloc(sizeof(*event), GFP_KERNEL);
6692 return ERR_PTR(-ENOMEM);
6695 * Single events are their own group leaders, with an
6696 * empty sibling list:
6699 group_leader = event;
6701 mutex_init(&event->child_mutex);
6702 INIT_LIST_HEAD(&event->child_list);
6704 INIT_LIST_HEAD(&event->group_entry);
6705 INIT_LIST_HEAD(&event->event_entry);
6706 INIT_LIST_HEAD(&event->sibling_list);
6707 INIT_LIST_HEAD(&event->rb_entry);
6708 INIT_LIST_HEAD(&event->active_entry);
6709 INIT_HLIST_NODE(&event->hlist_entry);
6712 init_waitqueue_head(&event->waitq);
6713 init_irq_work(&event->pending, perf_pending_event);
6715 mutex_init(&event->mmap_mutex);
6717 atomic_long_set(&event->refcount, 1);
6719 event->attr = *attr;
6720 event->group_leader = group_leader;
6724 event->parent = parent_event;
6726 event->ns = get_pid_ns(task_active_pid_ns(current));
6727 event->id = atomic64_inc_return(&perf_event_id);
6729 event->state = PERF_EVENT_STATE_INACTIVE;
6732 event->attach_state = PERF_ATTACH_TASK;
6734 if (attr->type == PERF_TYPE_TRACEPOINT)
6735 event->hw.tp_target = task;
6736 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6738 * hw_breakpoint is a bit difficult here..
6740 else if (attr->type == PERF_TYPE_BREAKPOINT)
6741 event->hw.bp_target = task;
6745 if (!overflow_handler && parent_event) {
6746 overflow_handler = parent_event->overflow_handler;
6747 context = parent_event->overflow_handler_context;
6750 event->overflow_handler = overflow_handler;
6751 event->overflow_handler_context = context;
6753 perf_event__state_init(event);
6758 hwc->sample_period = attr->sample_period;
6759 if (attr->freq && attr->sample_freq)
6760 hwc->sample_period = 1;
6761 hwc->last_period = hwc->sample_period;
6763 local64_set(&hwc->period_left, hwc->sample_period);
6766 * we currently do not support PERF_FORMAT_GROUP on inherited events
6768 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6771 pmu = perf_init_event(event);
6774 else if (IS_ERR(pmu)) {
6779 if (!event->parent) {
6780 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6781 err = get_callchain_buffers();
6791 event->destroy(event);
6794 put_pid_ns(event->ns);
6797 return ERR_PTR(err);
6800 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6801 struct perf_event_attr *attr)
6806 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6810 * zero the full structure, so that a short copy will be nice.
6812 memset(attr, 0, sizeof(*attr));
6814 ret = get_user(size, &uattr->size);
6818 if (size > PAGE_SIZE) /* silly large */
6821 if (!size) /* abi compat */
6822 size = PERF_ATTR_SIZE_VER0;
6824 if (size < PERF_ATTR_SIZE_VER0)
6828 * If we're handed a bigger struct than we know of,
6829 * ensure all the unknown bits are 0 - i.e. new
6830 * user-space does not rely on any kernel feature
6831 * extensions we dont know about yet.
6833 if (size > sizeof(*attr)) {
6834 unsigned char __user *addr;
6835 unsigned char __user *end;
6838 addr = (void __user *)uattr + sizeof(*attr);
6839 end = (void __user *)uattr + size;
6841 for (; addr < end; addr++) {
6842 ret = get_user(val, addr);
6848 size = sizeof(*attr);
6851 ret = copy_from_user(attr, uattr, size);
6855 /* disabled for now */
6859 if (attr->__reserved_1)
6862 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6865 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6868 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6869 u64 mask = attr->branch_sample_type;
6871 /* only using defined bits */
6872 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6875 /* at least one branch bit must be set */
6876 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6879 /* propagate priv level, when not set for branch */
6880 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6882 /* exclude_kernel checked on syscall entry */
6883 if (!attr->exclude_kernel)
6884 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6886 if (!attr->exclude_user)
6887 mask |= PERF_SAMPLE_BRANCH_USER;
6889 if (!attr->exclude_hv)
6890 mask |= PERF_SAMPLE_BRANCH_HV;
6892 * adjust user setting (for HW filter setup)
6894 attr->branch_sample_type = mask;
6896 /* privileged levels capture (kernel, hv): check permissions */
6897 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6898 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6902 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6903 ret = perf_reg_validate(attr->sample_regs_user);
6908 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6909 if (!arch_perf_have_user_stack_dump())
6913 * We have __u32 type for the size, but so far
6914 * we can only use __u16 as maximum due to the
6915 * __u16 sample size limit.
6917 if (attr->sample_stack_user >= USHRT_MAX)
6919 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6927 put_user(sizeof(*attr), &uattr->size);
6933 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6935 struct ring_buffer *rb = NULL, *old_rb = NULL;
6941 /* don't allow circular references */
6942 if (event == output_event)
6946 * Don't allow cross-cpu buffers
6948 if (output_event->cpu != event->cpu)
6952 * If its not a per-cpu rb, it must be the same task.
6954 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6958 mutex_lock(&event->mmap_mutex);
6959 /* Can't redirect output if we've got an active mmap() */
6960 if (atomic_read(&event->mmap_count))
6966 /* get the rb we want to redirect to */
6967 rb = ring_buffer_get(output_event);
6973 ring_buffer_detach(event, old_rb);
6976 ring_buffer_attach(event, rb);
6978 rcu_assign_pointer(event->rb, rb);
6981 ring_buffer_put(old_rb);
6983 * Since we detached before setting the new rb, so that we
6984 * could attach the new rb, we could have missed a wakeup.
6987 wake_up_all(&event->waitq);
6992 mutex_unlock(&event->mmap_mutex);
6999 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7001 * @attr_uptr: event_id type attributes for monitoring/sampling
7004 * @group_fd: group leader event fd
7006 SYSCALL_DEFINE5(perf_event_open,
7007 struct perf_event_attr __user *, attr_uptr,
7008 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7010 struct perf_event *group_leader = NULL, *output_event = NULL;
7011 struct perf_event *event, *sibling;
7012 struct perf_event_attr attr;
7013 struct perf_event_context *ctx;
7014 struct file *event_file = NULL;
7015 struct fd group = {NULL, 0};
7016 struct task_struct *task = NULL;
7021 int f_flags = O_RDWR;
7023 /* for future expandability... */
7024 if (flags & ~PERF_FLAG_ALL)
7027 err = perf_copy_attr(attr_uptr, &attr);
7031 if (!attr.exclude_kernel) {
7032 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7037 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7040 if (attr.sample_period & (1ULL << 63))
7045 * In cgroup mode, the pid argument is used to pass the fd
7046 * opened to the cgroup directory in cgroupfs. The cpu argument
7047 * designates the cpu on which to monitor threads from that
7050 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7053 if (flags & PERF_FLAG_FD_CLOEXEC)
7054 f_flags |= O_CLOEXEC;
7056 event_fd = get_unused_fd_flags(f_flags);
7060 if (group_fd != -1) {
7061 err = perf_fget_light(group_fd, &group);
7064 group_leader = group.file->private_data;
7065 if (flags & PERF_FLAG_FD_OUTPUT)
7066 output_event = group_leader;
7067 if (flags & PERF_FLAG_FD_NO_GROUP)
7068 group_leader = NULL;
7071 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7072 task = find_lively_task_by_vpid(pid);
7074 err = PTR_ERR(task);
7081 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7083 if (IS_ERR(event)) {
7084 err = PTR_ERR(event);
7088 if (flags & PERF_FLAG_PID_CGROUP) {
7089 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7091 __free_event(event);
7096 account_event(event);
7099 * Special case software events and allow them to be part of
7100 * any hardware group.
7105 (is_software_event(event) != is_software_event(group_leader))) {
7106 if (is_software_event(event)) {
7108 * If event and group_leader are not both a software
7109 * event, and event is, then group leader is not.
7111 * Allow the addition of software events to !software
7112 * groups, this is safe because software events never
7115 pmu = group_leader->pmu;
7116 } else if (is_software_event(group_leader) &&
7117 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7119 * In case the group is a pure software group, and we
7120 * try to add a hardware event, move the whole group to
7121 * the hardware context.
7128 * Get the target context (task or percpu):
7130 ctx = find_get_context(pmu, task, event->cpu);
7137 put_task_struct(task);
7142 * Look up the group leader (we will attach this event to it):
7148 * Do not allow a recursive hierarchy (this new sibling
7149 * becoming part of another group-sibling):
7151 if (group_leader->group_leader != group_leader)
7154 * Do not allow to attach to a group in a different
7155 * task or CPU context:
7158 if (group_leader->ctx->type != ctx->type)
7161 if (group_leader->ctx != ctx)
7166 * Only a group leader can be exclusive or pinned
7168 if (attr.exclusive || attr.pinned)
7173 err = perf_event_set_output(event, output_event);
7178 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7180 if (IS_ERR(event_file)) {
7181 err = PTR_ERR(event_file);
7186 struct perf_event_context *gctx = group_leader->ctx;
7188 mutex_lock(&gctx->mutex);
7189 perf_remove_from_context(group_leader, false);
7192 * Removing from the context ends up with disabled
7193 * event. What we want here is event in the initial
7194 * startup state, ready to be add into new context.
7196 perf_event__state_init(group_leader);
7197 list_for_each_entry(sibling, &group_leader->sibling_list,
7199 perf_remove_from_context(sibling, false);
7200 perf_event__state_init(sibling);
7203 mutex_unlock(&gctx->mutex);
7207 WARN_ON_ONCE(ctx->parent_ctx);
7208 mutex_lock(&ctx->mutex);
7212 perf_install_in_context(ctx, group_leader, event->cpu);
7214 list_for_each_entry(sibling, &group_leader->sibling_list,
7216 perf_install_in_context(ctx, sibling, event->cpu);
7221 perf_install_in_context(ctx, event, event->cpu);
7222 perf_unpin_context(ctx);
7223 mutex_unlock(&ctx->mutex);
7227 event->owner = current;
7229 mutex_lock(¤t->perf_event_mutex);
7230 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7231 mutex_unlock(¤t->perf_event_mutex);
7234 * Precalculate sample_data sizes
7236 perf_event__header_size(event);
7237 perf_event__id_header_size(event);
7240 * Drop the reference on the group_event after placing the
7241 * new event on the sibling_list. This ensures destruction
7242 * of the group leader will find the pointer to itself in
7243 * perf_group_detach().
7246 fd_install(event_fd, event_file);
7250 perf_unpin_context(ctx);
7257 put_task_struct(task);
7261 put_unused_fd(event_fd);
7266 * perf_event_create_kernel_counter
7268 * @attr: attributes of the counter to create
7269 * @cpu: cpu in which the counter is bound
7270 * @task: task to profile (NULL for percpu)
7273 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7274 struct task_struct *task,
7275 perf_overflow_handler_t overflow_handler,
7278 struct perf_event_context *ctx;
7279 struct perf_event *event;
7283 * Get the target context (task or percpu):
7286 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7287 overflow_handler, context);
7288 if (IS_ERR(event)) {
7289 err = PTR_ERR(event);
7293 account_event(event);
7295 ctx = find_get_context(event->pmu, task, cpu);
7301 WARN_ON_ONCE(ctx->parent_ctx);
7302 mutex_lock(&ctx->mutex);
7303 perf_install_in_context(ctx, event, cpu);
7304 perf_unpin_context(ctx);
7305 mutex_unlock(&ctx->mutex);
7312 return ERR_PTR(err);
7314 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7316 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7318 struct perf_event_context *src_ctx;
7319 struct perf_event_context *dst_ctx;
7320 struct perf_event *event, *tmp;
7323 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7324 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7326 mutex_lock(&src_ctx->mutex);
7327 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7329 perf_remove_from_context(event, false);
7330 unaccount_event_cpu(event, src_cpu);
7332 list_add(&event->migrate_entry, &events);
7334 mutex_unlock(&src_ctx->mutex);
7338 mutex_lock(&dst_ctx->mutex);
7339 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7340 list_del(&event->migrate_entry);
7341 if (event->state >= PERF_EVENT_STATE_OFF)
7342 event->state = PERF_EVENT_STATE_INACTIVE;
7343 account_event_cpu(event, dst_cpu);
7344 perf_install_in_context(dst_ctx, event, dst_cpu);
7347 mutex_unlock(&dst_ctx->mutex);
7349 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7351 static void sync_child_event(struct perf_event *child_event,
7352 struct task_struct *child)
7354 struct perf_event *parent_event = child_event->parent;
7357 if (child_event->attr.inherit_stat)
7358 perf_event_read_event(child_event, child);
7360 child_val = perf_event_count(child_event);
7363 * Add back the child's count to the parent's count:
7365 atomic64_add(child_val, &parent_event->child_count);
7366 atomic64_add(child_event->total_time_enabled,
7367 &parent_event->child_total_time_enabled);
7368 atomic64_add(child_event->total_time_running,
7369 &parent_event->child_total_time_running);
7372 * Remove this event from the parent's list
7374 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7375 mutex_lock(&parent_event->child_mutex);
7376 list_del_init(&child_event->child_list);
7377 mutex_unlock(&parent_event->child_mutex);
7380 * Release the parent event, if this was the last
7383 put_event(parent_event);
7387 __perf_event_exit_task(struct perf_event *child_event,
7388 struct perf_event_context *child_ctx,
7389 struct task_struct *child)
7391 perf_remove_from_context(child_event, !!child_event->parent);
7394 * It can happen that the parent exits first, and has events
7395 * that are still around due to the child reference. These
7396 * events need to be zapped.
7398 if (child_event->parent) {
7399 sync_child_event(child_event, child);
7400 free_event(child_event);
7404 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7406 struct perf_event *child_event, *tmp;
7407 struct perf_event_context *child_ctx;
7408 unsigned long flags;
7410 if (likely(!child->perf_event_ctxp[ctxn])) {
7411 perf_event_task(child, NULL, 0);
7415 local_irq_save(flags);
7417 * We can't reschedule here because interrupts are disabled,
7418 * and either child is current or it is a task that can't be
7419 * scheduled, so we are now safe from rescheduling changing
7422 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7425 * Take the context lock here so that if find_get_context is
7426 * reading child->perf_event_ctxp, we wait until it has
7427 * incremented the context's refcount before we do put_ctx below.
7429 raw_spin_lock(&child_ctx->lock);
7430 task_ctx_sched_out(child_ctx);
7431 child->perf_event_ctxp[ctxn] = NULL;
7433 * If this context is a clone; unclone it so it can't get
7434 * swapped to another process while we're removing all
7435 * the events from it.
7437 unclone_ctx(child_ctx);
7438 update_context_time(child_ctx);
7439 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7442 * Report the task dead after unscheduling the events so that we
7443 * won't get any samples after PERF_RECORD_EXIT. We can however still
7444 * get a few PERF_RECORD_READ events.
7446 perf_event_task(child, child_ctx, 0);
7449 * We can recurse on the same lock type through:
7451 * __perf_event_exit_task()
7452 * sync_child_event()
7454 * mutex_lock(&ctx->mutex)
7456 * But since its the parent context it won't be the same instance.
7458 mutex_lock(&child_ctx->mutex);
7461 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7463 __perf_event_exit_task(child_event, child_ctx, child);
7465 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7467 __perf_event_exit_task(child_event, child_ctx, child);
7470 * If the last event was a group event, it will have appended all
7471 * its siblings to the list, but we obtained 'tmp' before that which
7472 * will still point to the list head terminating the iteration.
7474 if (!list_empty(&child_ctx->pinned_groups) ||
7475 !list_empty(&child_ctx->flexible_groups))
7478 mutex_unlock(&child_ctx->mutex);
7484 * When a child task exits, feed back event values to parent events.
7486 void perf_event_exit_task(struct task_struct *child)
7488 struct perf_event *event, *tmp;
7491 mutex_lock(&child->perf_event_mutex);
7492 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7494 list_del_init(&event->owner_entry);
7497 * Ensure the list deletion is visible before we clear
7498 * the owner, closes a race against perf_release() where
7499 * we need to serialize on the owner->perf_event_mutex.
7502 event->owner = NULL;
7504 mutex_unlock(&child->perf_event_mutex);
7506 for_each_task_context_nr(ctxn)
7507 perf_event_exit_task_context(child, ctxn);
7510 static void perf_free_event(struct perf_event *event,
7511 struct perf_event_context *ctx)
7513 struct perf_event *parent = event->parent;
7515 if (WARN_ON_ONCE(!parent))
7518 mutex_lock(&parent->child_mutex);
7519 list_del_init(&event->child_list);
7520 mutex_unlock(&parent->child_mutex);
7524 perf_group_detach(event);
7525 list_del_event(event, ctx);
7530 * free an unexposed, unused context as created by inheritance by
7531 * perf_event_init_task below, used by fork() in case of fail.
7533 void perf_event_free_task(struct task_struct *task)
7535 struct perf_event_context *ctx;
7536 struct perf_event *event, *tmp;
7539 for_each_task_context_nr(ctxn) {
7540 ctx = task->perf_event_ctxp[ctxn];
7544 mutex_lock(&ctx->mutex);
7546 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7548 perf_free_event(event, ctx);
7550 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7552 perf_free_event(event, ctx);
7554 if (!list_empty(&ctx->pinned_groups) ||
7555 !list_empty(&ctx->flexible_groups))
7558 mutex_unlock(&ctx->mutex);
7564 void perf_event_delayed_put(struct task_struct *task)
7568 for_each_task_context_nr(ctxn)
7569 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7573 * inherit a event from parent task to child task:
7575 static struct perf_event *
7576 inherit_event(struct perf_event *parent_event,
7577 struct task_struct *parent,
7578 struct perf_event_context *parent_ctx,
7579 struct task_struct *child,
7580 struct perf_event *group_leader,
7581 struct perf_event_context *child_ctx)
7583 struct perf_event *child_event;
7584 unsigned long flags;
7587 * Instead of creating recursive hierarchies of events,
7588 * we link inherited events back to the original parent,
7589 * which has a filp for sure, which we use as the reference
7592 if (parent_event->parent)
7593 parent_event = parent_event->parent;
7595 child_event = perf_event_alloc(&parent_event->attr,
7598 group_leader, parent_event,
7600 if (IS_ERR(child_event))
7603 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7604 free_event(child_event);
7611 * Make the child state follow the state of the parent event,
7612 * not its attr.disabled bit. We hold the parent's mutex,
7613 * so we won't race with perf_event_{en, dis}able_family.
7615 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7616 child_event->state = PERF_EVENT_STATE_INACTIVE;
7618 child_event->state = PERF_EVENT_STATE_OFF;
7620 if (parent_event->attr.freq) {
7621 u64 sample_period = parent_event->hw.sample_period;
7622 struct hw_perf_event *hwc = &child_event->hw;
7624 hwc->sample_period = sample_period;
7625 hwc->last_period = sample_period;
7627 local64_set(&hwc->period_left, sample_period);
7630 child_event->ctx = child_ctx;
7631 child_event->overflow_handler = parent_event->overflow_handler;
7632 child_event->overflow_handler_context
7633 = parent_event->overflow_handler_context;
7636 * Precalculate sample_data sizes
7638 perf_event__header_size(child_event);
7639 perf_event__id_header_size(child_event);
7642 * Link it up in the child's context:
7644 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7645 add_event_to_ctx(child_event, child_ctx);
7646 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7649 * Link this into the parent event's child list
7651 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7652 mutex_lock(&parent_event->child_mutex);
7653 list_add_tail(&child_event->child_list, &parent_event->child_list);
7654 mutex_unlock(&parent_event->child_mutex);
7659 static int inherit_group(struct perf_event *parent_event,
7660 struct task_struct *parent,
7661 struct perf_event_context *parent_ctx,
7662 struct task_struct *child,
7663 struct perf_event_context *child_ctx)
7665 struct perf_event *leader;
7666 struct perf_event *sub;
7667 struct perf_event *child_ctr;
7669 leader = inherit_event(parent_event, parent, parent_ctx,
7670 child, NULL, child_ctx);
7672 return PTR_ERR(leader);
7673 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7674 child_ctr = inherit_event(sub, parent, parent_ctx,
7675 child, leader, child_ctx);
7676 if (IS_ERR(child_ctr))
7677 return PTR_ERR(child_ctr);
7683 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7684 struct perf_event_context *parent_ctx,
7685 struct task_struct *child, int ctxn,
7689 struct perf_event_context *child_ctx;
7691 if (!event->attr.inherit) {
7696 child_ctx = child->perf_event_ctxp[ctxn];
7699 * This is executed from the parent task context, so
7700 * inherit events that have been marked for cloning.
7701 * First allocate and initialize a context for the
7705 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7709 child->perf_event_ctxp[ctxn] = child_ctx;
7712 ret = inherit_group(event, parent, parent_ctx,
7722 * Initialize the perf_event context in task_struct
7724 int perf_event_init_context(struct task_struct *child, int ctxn)
7726 struct perf_event_context *child_ctx, *parent_ctx;
7727 struct perf_event_context *cloned_ctx;
7728 struct perf_event *event;
7729 struct task_struct *parent = current;
7730 int inherited_all = 1;
7731 unsigned long flags;
7734 if (likely(!parent->perf_event_ctxp[ctxn]))
7738 * If the parent's context is a clone, pin it so it won't get
7741 parent_ctx = perf_pin_task_context(parent, ctxn);
7744 * No need to check if parent_ctx != NULL here; since we saw
7745 * it non-NULL earlier, the only reason for it to become NULL
7746 * is if we exit, and since we're currently in the middle of
7747 * a fork we can't be exiting at the same time.
7751 * Lock the parent list. No need to lock the child - not PID
7752 * hashed yet and not running, so nobody can access it.
7754 mutex_lock(&parent_ctx->mutex);
7757 * We dont have to disable NMIs - we are only looking at
7758 * the list, not manipulating it:
7760 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7761 ret = inherit_task_group(event, parent, parent_ctx,
7762 child, ctxn, &inherited_all);
7768 * We can't hold ctx->lock when iterating the ->flexible_group list due
7769 * to allocations, but we need to prevent rotation because
7770 * rotate_ctx() will change the list from interrupt context.
7772 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7773 parent_ctx->rotate_disable = 1;
7774 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7776 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7777 ret = inherit_task_group(event, parent, parent_ctx,
7778 child, ctxn, &inherited_all);
7783 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7784 parent_ctx->rotate_disable = 0;
7786 child_ctx = child->perf_event_ctxp[ctxn];
7788 if (child_ctx && inherited_all) {
7790 * Mark the child context as a clone of the parent
7791 * context, or of whatever the parent is a clone of.
7793 * Note that if the parent is a clone, the holding of
7794 * parent_ctx->lock avoids it from being uncloned.
7796 cloned_ctx = parent_ctx->parent_ctx;
7798 child_ctx->parent_ctx = cloned_ctx;
7799 child_ctx->parent_gen = parent_ctx->parent_gen;
7801 child_ctx->parent_ctx = parent_ctx;
7802 child_ctx->parent_gen = parent_ctx->generation;
7804 get_ctx(child_ctx->parent_ctx);
7807 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7808 mutex_unlock(&parent_ctx->mutex);
7810 perf_unpin_context(parent_ctx);
7811 put_ctx(parent_ctx);
7817 * Initialize the perf_event context in task_struct
7819 int perf_event_init_task(struct task_struct *child)
7823 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7824 mutex_init(&child->perf_event_mutex);
7825 INIT_LIST_HEAD(&child->perf_event_list);
7827 for_each_task_context_nr(ctxn) {
7828 ret = perf_event_init_context(child, ctxn);
7836 static void __init perf_event_init_all_cpus(void)
7838 struct swevent_htable *swhash;
7841 for_each_possible_cpu(cpu) {
7842 swhash = &per_cpu(swevent_htable, cpu);
7843 mutex_init(&swhash->hlist_mutex);
7844 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7848 static void perf_event_init_cpu(int cpu)
7850 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7852 mutex_lock(&swhash->hlist_mutex);
7853 swhash->online = true;
7854 if (swhash->hlist_refcount > 0) {
7855 struct swevent_hlist *hlist;
7857 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7859 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7861 mutex_unlock(&swhash->hlist_mutex);
7864 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7865 static void perf_pmu_rotate_stop(struct pmu *pmu)
7867 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7869 WARN_ON(!irqs_disabled());
7871 list_del_init(&cpuctx->rotation_list);
7874 static void __perf_event_exit_context(void *__info)
7876 struct remove_event re = { .detach_group = false };
7877 struct perf_event_context *ctx = __info;
7879 perf_pmu_rotate_stop(ctx->pmu);
7882 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
7883 __perf_remove_from_context(&re);
7887 static void perf_event_exit_cpu_context(int cpu)
7889 struct perf_event_context *ctx;
7893 idx = srcu_read_lock(&pmus_srcu);
7894 list_for_each_entry_rcu(pmu, &pmus, entry) {
7895 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7897 mutex_lock(&ctx->mutex);
7898 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7899 mutex_unlock(&ctx->mutex);
7901 srcu_read_unlock(&pmus_srcu, idx);
7904 static void perf_event_exit_cpu(int cpu)
7906 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7908 perf_event_exit_cpu_context(cpu);
7910 mutex_lock(&swhash->hlist_mutex);
7911 swhash->online = false;
7912 swevent_hlist_release(swhash);
7913 mutex_unlock(&swhash->hlist_mutex);
7916 static inline void perf_event_exit_cpu(int cpu) { }
7920 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7924 for_each_online_cpu(cpu)
7925 perf_event_exit_cpu(cpu);
7931 * Run the perf reboot notifier at the very last possible moment so that
7932 * the generic watchdog code runs as long as possible.
7934 static struct notifier_block perf_reboot_notifier = {
7935 .notifier_call = perf_reboot,
7936 .priority = INT_MIN,
7940 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7942 unsigned int cpu = (long)hcpu;
7944 switch (action & ~CPU_TASKS_FROZEN) {
7946 case CPU_UP_PREPARE:
7947 case CPU_DOWN_FAILED:
7948 perf_event_init_cpu(cpu);
7951 case CPU_UP_CANCELED:
7952 case CPU_DOWN_PREPARE:
7953 perf_event_exit_cpu(cpu);
7962 void __init perf_event_init(void)
7968 perf_event_init_all_cpus();
7969 init_srcu_struct(&pmus_srcu);
7970 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7971 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7972 perf_pmu_register(&perf_task_clock, NULL, -1);
7974 perf_cpu_notifier(perf_cpu_notify);
7975 register_reboot_notifier(&perf_reboot_notifier);
7977 ret = init_hw_breakpoint();
7978 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7980 /* do not patch jump label more than once per second */
7981 jump_label_rate_limit(&perf_sched_events, HZ);
7984 * Build time assertion that we keep the data_head at the intended
7985 * location. IOW, validation we got the __reserved[] size right.
7987 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7991 static int __init perf_event_sysfs_init(void)
7996 mutex_lock(&pmus_lock);
7998 ret = bus_register(&pmu_bus);
8002 list_for_each_entry(pmu, &pmus, entry) {
8003 if (!pmu->name || pmu->type < 0)
8006 ret = pmu_dev_alloc(pmu);
8007 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8009 pmu_bus_running = 1;
8013 mutex_unlock(&pmus_lock);
8017 device_initcall(perf_event_sysfs_init);
8019 #ifdef CONFIG_CGROUP_PERF
8020 static struct cgroup_subsys_state *
8021 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8023 struct perf_cgroup *jc;
8025 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8027 return ERR_PTR(-ENOMEM);
8029 jc->info = alloc_percpu(struct perf_cgroup_info);
8032 return ERR_PTR(-ENOMEM);
8038 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8040 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8042 free_percpu(jc->info);
8046 static int __perf_cgroup_move(void *info)
8048 struct task_struct *task = info;
8049 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8053 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8054 struct cgroup_taskset *tset)
8056 struct task_struct *task;
8058 cgroup_taskset_for_each(task, css, tset)
8059 task_function_call(task, __perf_cgroup_move, task);
8062 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8063 struct cgroup_subsys_state *old_css,
8064 struct task_struct *task)
8067 * cgroup_exit() is called in the copy_process() failure path.
8068 * Ignore this case since the task hasn't ran yet, this avoids
8069 * trying to poke a half freed task state from generic code.
8071 if (!(task->flags & PF_EXITING))
8074 task_function_call(task, __perf_cgroup_move, task);
8077 struct cgroup_subsys perf_subsys = {
8078 .name = "perf_event",
8079 .subsys_id = perf_subsys_id,
8080 .css_alloc = perf_cgroup_css_alloc,
8081 .css_free = perf_cgroup_css_free,
8082 .exit = perf_cgroup_exit,
8083 .attach = perf_cgroup_attach,
8085 #endif /* CONFIG_CGROUP_PERF */