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>
42 #include <linux/module.h>
46 #include <asm/irq_regs.h>
48 struct remote_function_call {
49 struct task_struct *p;
50 int (*func)(void *info);
55 static void remote_function(void *data)
57 struct remote_function_call *tfc = data;
58 struct task_struct *p = tfc->p;
62 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
66 tfc->ret = tfc->func(tfc->info);
70 * task_function_call - call a function on the cpu on which a task runs
71 * @p: the task to evaluate
72 * @func: the function to be called
73 * @info: the function call argument
75 * Calls the function @func when the task is currently running. This might
76 * be on the current CPU, which just calls the function directly
78 * returns: @func return value, or
79 * -ESRCH - when the process isn't running
80 * -EAGAIN - when the process moved away
83 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
85 struct remote_function_call data = {
89 .ret = -ESRCH, /* No such (running) process */
93 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
99 * cpu_function_call - call a function on the cpu
100 * @func: the function to be called
101 * @info: the function call argument
103 * Calls the function @func on the remote cpu.
105 * returns: @func return value or -ENXIO when the cpu is offline
107 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
109 struct remote_function_call data = {
113 .ret = -ENXIO, /* No such CPU */
116 smp_call_function_single(cpu, remote_function, &data, 1);
121 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
122 PERF_FLAG_FD_OUTPUT |\
123 PERF_FLAG_PID_CGROUP |\
124 PERF_FLAG_FD_CLOEXEC)
127 * branch priv levels that need permission checks
129 #define PERF_SAMPLE_BRANCH_PERM_PLM \
130 (PERF_SAMPLE_BRANCH_KERNEL |\
131 PERF_SAMPLE_BRANCH_HV)
134 EVENT_FLEXIBLE = 0x1,
136 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
140 * perf_sched_events : >0 events exist
141 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
143 struct static_key_deferred perf_sched_events __read_mostly;
144 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
145 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
147 static atomic_t nr_mmap_events __read_mostly;
148 static atomic_t nr_comm_events __read_mostly;
149 static atomic_t nr_task_events __read_mostly;
150 static atomic_t nr_freq_events __read_mostly;
152 static LIST_HEAD(pmus);
153 static DEFINE_MUTEX(pmus_lock);
154 static struct srcu_struct pmus_srcu;
157 * perf event paranoia level:
158 * -1 - not paranoid at all
159 * 0 - disallow raw tracepoint access for unpriv
160 * 1 - disallow cpu events for unpriv
161 * 2 - disallow kernel profiling for unpriv
163 int sysctl_perf_event_paranoid __read_mostly = 1;
165 /* Minimum for 512 kiB + 1 user control page */
166 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
169 * max perf event sample rate
171 #define DEFAULT_MAX_SAMPLE_RATE 100000
172 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
173 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
175 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
177 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
178 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
180 static int perf_sample_allowed_ns __read_mostly =
181 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
183 void update_perf_cpu_limits(void)
185 u64 tmp = perf_sample_period_ns;
187 tmp *= sysctl_perf_cpu_time_max_percent;
189 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
192 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
194 int perf_proc_update_handler(struct ctl_table *table, int write,
195 void __user *buffer, size_t *lenp,
198 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
203 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
204 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
205 update_perf_cpu_limits();
210 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
212 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
213 void __user *buffer, size_t *lenp,
216 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
221 update_perf_cpu_limits();
227 * perf samples are done in some very critical code paths (NMIs).
228 * If they take too much CPU time, the system can lock up and not
229 * get any real work done. This will drop the sample rate when
230 * we detect that events are taking too long.
232 #define NR_ACCUMULATED_SAMPLES 128
233 static DEFINE_PER_CPU(u64, running_sample_length);
235 static void perf_duration_warn(struct irq_work *w)
237 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
238 u64 avg_local_sample_len;
239 u64 local_samples_len;
241 local_samples_len = __get_cpu_var(running_sample_length);
242 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
244 printk_ratelimited(KERN_WARNING
245 "perf interrupt took too long (%lld > %lld), lowering "
246 "kernel.perf_event_max_sample_rate to %d\n",
247 avg_local_sample_len, allowed_ns >> 1,
248 sysctl_perf_event_sample_rate);
251 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
253 void perf_sample_event_took(u64 sample_len_ns)
255 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
256 u64 avg_local_sample_len;
257 u64 local_samples_len;
262 /* decay the counter by 1 average sample */
263 local_samples_len = __get_cpu_var(running_sample_length);
264 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
265 local_samples_len += sample_len_ns;
266 __get_cpu_var(running_sample_length) = local_samples_len;
269 * note: this will be biased artifically low until we have
270 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
271 * from having to maintain a count.
273 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
275 if (avg_local_sample_len <= allowed_ns)
278 if (max_samples_per_tick <= 1)
281 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
282 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
283 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
285 update_perf_cpu_limits();
287 if (!irq_work_queue(&perf_duration_work)) {
288 early_printk("perf interrupt took too long (%lld > %lld), lowering "
289 "kernel.perf_event_max_sample_rate to %d\n",
290 avg_local_sample_len, allowed_ns >> 1,
291 sysctl_perf_event_sample_rate);
295 static atomic64_t perf_event_id;
297 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
298 enum event_type_t event_type);
300 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
301 enum event_type_t event_type,
302 struct task_struct *task);
304 static void update_context_time(struct perf_event_context *ctx);
305 static u64 perf_event_time(struct perf_event *event);
307 void __weak perf_event_print_debug(void) { }
309 extern __weak const char *perf_pmu_name(void)
314 static inline u64 perf_clock(void)
316 return local_clock();
319 static inline struct perf_cpu_context *
320 __get_cpu_context(struct perf_event_context *ctx)
322 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
325 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
326 struct perf_event_context *ctx)
328 raw_spin_lock(&cpuctx->ctx.lock);
330 raw_spin_lock(&ctx->lock);
333 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
334 struct perf_event_context *ctx)
337 raw_spin_unlock(&ctx->lock);
338 raw_spin_unlock(&cpuctx->ctx.lock);
341 #ifdef CONFIG_CGROUP_PERF
344 * perf_cgroup_info keeps track of time_enabled for a cgroup.
345 * This is a per-cpu dynamically allocated data structure.
347 struct perf_cgroup_info {
353 struct cgroup_subsys_state css;
354 struct perf_cgroup_info __percpu *info;
358 * Must ensure cgroup is pinned (css_get) before calling
359 * this function. In other words, we cannot call this function
360 * if there is no cgroup event for the current CPU context.
362 static inline struct perf_cgroup *
363 perf_cgroup_from_task(struct task_struct *task)
365 return container_of(task_css(task, perf_event_cgrp_id),
366 struct perf_cgroup, css);
370 perf_cgroup_match(struct perf_event *event)
372 struct perf_event_context *ctx = event->ctx;
373 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
375 /* @event doesn't care about cgroup */
379 /* wants specific cgroup scope but @cpuctx isn't associated with any */
384 * Cgroup scoping is recursive. An event enabled for a cgroup is
385 * also enabled for all its descendant cgroups. If @cpuctx's
386 * cgroup is a descendant of @event's (the test covers identity
387 * case), it's a match.
389 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
390 event->cgrp->css.cgroup);
393 static inline void perf_put_cgroup(struct perf_event *event)
395 css_put(&event->cgrp->css);
398 static inline void perf_detach_cgroup(struct perf_event *event)
400 perf_put_cgroup(event);
404 static inline int is_cgroup_event(struct perf_event *event)
406 return event->cgrp != NULL;
409 static inline u64 perf_cgroup_event_time(struct perf_event *event)
411 struct perf_cgroup_info *t;
413 t = per_cpu_ptr(event->cgrp->info, event->cpu);
417 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
419 struct perf_cgroup_info *info;
424 info = this_cpu_ptr(cgrp->info);
426 info->time += now - info->timestamp;
427 info->timestamp = now;
430 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
432 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
434 __update_cgrp_time(cgrp_out);
437 static inline void update_cgrp_time_from_event(struct perf_event *event)
439 struct perf_cgroup *cgrp;
442 * ensure we access cgroup data only when needed and
443 * when we know the cgroup is pinned (css_get)
445 if (!is_cgroup_event(event))
448 cgrp = perf_cgroup_from_task(current);
450 * Do not update time when cgroup is not active
452 if (cgrp == event->cgrp)
453 __update_cgrp_time(event->cgrp);
457 perf_cgroup_set_timestamp(struct task_struct *task,
458 struct perf_event_context *ctx)
460 struct perf_cgroup *cgrp;
461 struct perf_cgroup_info *info;
464 * ctx->lock held by caller
465 * ensure we do not access cgroup data
466 * unless we have the cgroup pinned (css_get)
468 if (!task || !ctx->nr_cgroups)
471 cgrp = perf_cgroup_from_task(task);
472 info = this_cpu_ptr(cgrp->info);
473 info->timestamp = ctx->timestamp;
476 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
477 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
480 * reschedule events based on the cgroup constraint of task.
482 * mode SWOUT : schedule out everything
483 * mode SWIN : schedule in based on cgroup for next
485 void perf_cgroup_switch(struct task_struct *task, int mode)
487 struct perf_cpu_context *cpuctx;
492 * disable interrupts to avoid geting nr_cgroup
493 * changes via __perf_event_disable(). Also
496 local_irq_save(flags);
499 * we reschedule only in the presence of cgroup
500 * constrained events.
504 list_for_each_entry_rcu(pmu, &pmus, entry) {
505 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
506 if (cpuctx->unique_pmu != pmu)
507 continue; /* ensure we process each cpuctx once */
510 * perf_cgroup_events says at least one
511 * context on this CPU has cgroup events.
513 * ctx->nr_cgroups reports the number of cgroup
514 * events for a context.
516 if (cpuctx->ctx.nr_cgroups > 0) {
517 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
518 perf_pmu_disable(cpuctx->ctx.pmu);
520 if (mode & PERF_CGROUP_SWOUT) {
521 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
523 * must not be done before ctxswout due
524 * to event_filter_match() in event_sched_out()
529 if (mode & PERF_CGROUP_SWIN) {
530 WARN_ON_ONCE(cpuctx->cgrp);
532 * set cgrp before ctxsw in to allow
533 * event_filter_match() to not have to pass
536 cpuctx->cgrp = perf_cgroup_from_task(task);
537 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
539 perf_pmu_enable(cpuctx->ctx.pmu);
540 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
546 local_irq_restore(flags);
549 static inline void perf_cgroup_sched_out(struct task_struct *task,
550 struct task_struct *next)
552 struct perf_cgroup *cgrp1;
553 struct perf_cgroup *cgrp2 = NULL;
556 * we come here when we know perf_cgroup_events > 0
558 cgrp1 = perf_cgroup_from_task(task);
561 * next is NULL when called from perf_event_enable_on_exec()
562 * that will systematically cause a cgroup_switch()
565 cgrp2 = perf_cgroup_from_task(next);
568 * only schedule out current cgroup events if we know
569 * that we are switching to a different cgroup. Otherwise,
570 * do no touch the cgroup events.
573 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
576 static inline void perf_cgroup_sched_in(struct task_struct *prev,
577 struct task_struct *task)
579 struct perf_cgroup *cgrp1;
580 struct perf_cgroup *cgrp2 = NULL;
583 * we come here when we know perf_cgroup_events > 0
585 cgrp1 = perf_cgroup_from_task(task);
587 /* prev can never be NULL */
588 cgrp2 = perf_cgroup_from_task(prev);
591 * only need to schedule in cgroup events if we are changing
592 * cgroup during ctxsw. Cgroup events were not scheduled
593 * out of ctxsw out if that was not the case.
596 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
599 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
600 struct perf_event_attr *attr,
601 struct perf_event *group_leader)
603 struct perf_cgroup *cgrp;
604 struct cgroup_subsys_state *css;
605 struct fd f = fdget(fd);
611 css = css_tryget_from_dir(f.file->f_dentry, &perf_event_cgrp_subsys);
617 cgrp = container_of(css, struct perf_cgroup, css);
621 * all events in a group must monitor
622 * the same cgroup because a task belongs
623 * to only one perf cgroup at a time
625 if (group_leader && group_leader->cgrp != cgrp) {
626 perf_detach_cgroup(event);
635 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
637 struct perf_cgroup_info *t;
638 t = per_cpu_ptr(event->cgrp->info, event->cpu);
639 event->shadow_ctx_time = now - t->timestamp;
643 perf_cgroup_defer_enabled(struct perf_event *event)
646 * when the current task's perf cgroup does not match
647 * the event's, we need to remember to call the
648 * perf_mark_enable() function the first time a task with
649 * a matching perf cgroup is scheduled in.
651 if (is_cgroup_event(event) && !perf_cgroup_match(event))
652 event->cgrp_defer_enabled = 1;
656 perf_cgroup_mark_enabled(struct perf_event *event,
657 struct perf_event_context *ctx)
659 struct perf_event *sub;
660 u64 tstamp = perf_event_time(event);
662 if (!event->cgrp_defer_enabled)
665 event->cgrp_defer_enabled = 0;
667 event->tstamp_enabled = tstamp - event->total_time_enabled;
668 list_for_each_entry(sub, &event->sibling_list, group_entry) {
669 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
670 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
671 sub->cgrp_defer_enabled = 0;
675 #else /* !CONFIG_CGROUP_PERF */
678 perf_cgroup_match(struct perf_event *event)
683 static inline void perf_detach_cgroup(struct perf_event *event)
686 static inline int is_cgroup_event(struct perf_event *event)
691 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
696 static inline void update_cgrp_time_from_event(struct perf_event *event)
700 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
704 static inline void perf_cgroup_sched_out(struct task_struct *task,
705 struct task_struct *next)
709 static inline void perf_cgroup_sched_in(struct task_struct *prev,
710 struct task_struct *task)
714 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
715 struct perf_event_attr *attr,
716 struct perf_event *group_leader)
722 perf_cgroup_set_timestamp(struct task_struct *task,
723 struct perf_event_context *ctx)
728 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
733 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
737 static inline u64 perf_cgroup_event_time(struct perf_event *event)
743 perf_cgroup_defer_enabled(struct perf_event *event)
748 perf_cgroup_mark_enabled(struct perf_event *event,
749 struct perf_event_context *ctx)
755 * set default to be dependent on timer tick just
758 #define PERF_CPU_HRTIMER (1000 / HZ)
760 * function must be called with interrupts disbled
762 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
764 struct perf_cpu_context *cpuctx;
765 enum hrtimer_restart ret = HRTIMER_NORESTART;
768 WARN_ON(!irqs_disabled());
770 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
772 rotations = perf_rotate_context(cpuctx);
775 * arm timer if needed
778 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
779 ret = HRTIMER_RESTART;
785 /* CPU is going down */
786 void perf_cpu_hrtimer_cancel(int cpu)
788 struct perf_cpu_context *cpuctx;
792 if (WARN_ON(cpu != smp_processor_id()))
795 local_irq_save(flags);
799 list_for_each_entry_rcu(pmu, &pmus, entry) {
800 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
802 if (pmu->task_ctx_nr == perf_sw_context)
805 hrtimer_cancel(&cpuctx->hrtimer);
810 local_irq_restore(flags);
813 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
815 struct hrtimer *hr = &cpuctx->hrtimer;
816 struct pmu *pmu = cpuctx->ctx.pmu;
819 /* no multiplexing needed for SW PMU */
820 if (pmu->task_ctx_nr == perf_sw_context)
824 * check default is sane, if not set then force to
825 * default interval (1/tick)
827 timer = pmu->hrtimer_interval_ms;
829 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
831 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
833 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
834 hr->function = perf_cpu_hrtimer_handler;
837 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
839 struct hrtimer *hr = &cpuctx->hrtimer;
840 struct pmu *pmu = cpuctx->ctx.pmu;
843 if (pmu->task_ctx_nr == perf_sw_context)
846 if (hrtimer_active(hr))
849 if (!hrtimer_callback_running(hr))
850 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
851 0, HRTIMER_MODE_REL_PINNED, 0);
854 void perf_pmu_disable(struct pmu *pmu)
856 int *count = this_cpu_ptr(pmu->pmu_disable_count);
858 pmu->pmu_disable(pmu);
861 void perf_pmu_enable(struct pmu *pmu)
863 int *count = this_cpu_ptr(pmu->pmu_disable_count);
865 pmu->pmu_enable(pmu);
868 static DEFINE_PER_CPU(struct list_head, rotation_list);
871 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
872 * because they're strictly cpu affine and rotate_start is called with IRQs
873 * disabled, while rotate_context is called from IRQ context.
875 static void perf_pmu_rotate_start(struct pmu *pmu)
877 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
878 struct list_head *head = &__get_cpu_var(rotation_list);
880 WARN_ON(!irqs_disabled());
882 if (list_empty(&cpuctx->rotation_list))
883 list_add(&cpuctx->rotation_list, head);
886 static void get_ctx(struct perf_event_context *ctx)
888 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
891 static void put_ctx(struct perf_event_context *ctx)
893 if (atomic_dec_and_test(&ctx->refcount)) {
895 put_ctx(ctx->parent_ctx);
897 put_task_struct(ctx->task);
898 kfree_rcu(ctx, rcu_head);
902 static void unclone_ctx(struct perf_event_context *ctx)
904 if (ctx->parent_ctx) {
905 put_ctx(ctx->parent_ctx);
906 ctx->parent_ctx = NULL;
911 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
914 * only top level events have the pid namespace they were created in
917 event = event->parent;
919 return task_tgid_nr_ns(p, event->ns);
922 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
925 * only top level events have the pid namespace they were created in
928 event = event->parent;
930 return task_pid_nr_ns(p, event->ns);
934 * If we inherit events we want to return the parent event id
937 static u64 primary_event_id(struct perf_event *event)
942 id = event->parent->id;
948 * Get the perf_event_context for a task and lock it.
949 * This has to cope with with the fact that until it is locked,
950 * the context could get moved to another task.
952 static struct perf_event_context *
953 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
955 struct perf_event_context *ctx;
959 * One of the few rules of preemptible RCU is that one cannot do
960 * rcu_read_unlock() while holding a scheduler (or nested) lock when
961 * part of the read side critical section was preemptible -- see
962 * rcu_read_unlock_special().
964 * Since ctx->lock nests under rq->lock we must ensure the entire read
965 * side critical section is non-preemptible.
969 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
972 * If this context is a clone of another, it might
973 * get swapped for another underneath us by
974 * perf_event_task_sched_out, though the
975 * rcu_read_lock() protects us from any context
976 * getting freed. Lock the context and check if it
977 * got swapped before we could get the lock, and retry
978 * if so. If we locked the right context, then it
979 * can't get swapped on us any more.
981 raw_spin_lock_irqsave(&ctx->lock, *flags);
982 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
983 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
989 if (!atomic_inc_not_zero(&ctx->refcount)) {
990 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1000 * Get the context for a task and increment its pin_count so it
1001 * can't get swapped to another task. This also increments its
1002 * reference count so that the context can't get freed.
1004 static struct perf_event_context *
1005 perf_pin_task_context(struct task_struct *task, int ctxn)
1007 struct perf_event_context *ctx;
1008 unsigned long flags;
1010 ctx = perf_lock_task_context(task, ctxn, &flags);
1013 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1018 static void perf_unpin_context(struct perf_event_context *ctx)
1020 unsigned long flags;
1022 raw_spin_lock_irqsave(&ctx->lock, flags);
1024 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1028 * Update the record of the current time in a context.
1030 static void update_context_time(struct perf_event_context *ctx)
1032 u64 now = perf_clock();
1034 ctx->time += now - ctx->timestamp;
1035 ctx->timestamp = now;
1038 static u64 perf_event_time(struct perf_event *event)
1040 struct perf_event_context *ctx = event->ctx;
1042 if (is_cgroup_event(event))
1043 return perf_cgroup_event_time(event);
1045 return ctx ? ctx->time : 0;
1049 * Update the total_time_enabled and total_time_running fields for a event.
1050 * The caller of this function needs to hold the ctx->lock.
1052 static void update_event_times(struct perf_event *event)
1054 struct perf_event_context *ctx = event->ctx;
1057 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1058 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1061 * in cgroup mode, time_enabled represents
1062 * the time the event was enabled AND active
1063 * tasks were in the monitored cgroup. This is
1064 * independent of the activity of the context as
1065 * there may be a mix of cgroup and non-cgroup events.
1067 * That is why we treat cgroup events differently
1070 if (is_cgroup_event(event))
1071 run_end = perf_cgroup_event_time(event);
1072 else if (ctx->is_active)
1073 run_end = ctx->time;
1075 run_end = event->tstamp_stopped;
1077 event->total_time_enabled = run_end - event->tstamp_enabled;
1079 if (event->state == PERF_EVENT_STATE_INACTIVE)
1080 run_end = event->tstamp_stopped;
1082 run_end = perf_event_time(event);
1084 event->total_time_running = run_end - event->tstamp_running;
1089 * Update total_time_enabled and total_time_running for all events in a group.
1091 static void update_group_times(struct perf_event *leader)
1093 struct perf_event *event;
1095 update_event_times(leader);
1096 list_for_each_entry(event, &leader->sibling_list, group_entry)
1097 update_event_times(event);
1100 static struct list_head *
1101 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1103 if (event->attr.pinned)
1104 return &ctx->pinned_groups;
1106 return &ctx->flexible_groups;
1110 * Add a event from the lists for its context.
1111 * Must be called with ctx->mutex and ctx->lock held.
1114 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1116 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1117 event->attach_state |= PERF_ATTACH_CONTEXT;
1120 * If we're a stand alone event or group leader, we go to the context
1121 * list, group events are kept attached to the group so that
1122 * perf_group_detach can, at all times, locate all siblings.
1124 if (event->group_leader == event) {
1125 struct list_head *list;
1127 if (is_software_event(event))
1128 event->group_flags |= PERF_GROUP_SOFTWARE;
1130 list = ctx_group_list(event, ctx);
1131 list_add_tail(&event->group_entry, list);
1134 if (is_cgroup_event(event))
1137 if (has_branch_stack(event))
1138 ctx->nr_branch_stack++;
1140 list_add_rcu(&event->event_entry, &ctx->event_list);
1141 if (!ctx->nr_events)
1142 perf_pmu_rotate_start(ctx->pmu);
1144 if (event->attr.inherit_stat)
1151 * Initialize event state based on the perf_event_attr::disabled.
1153 static inline void perf_event__state_init(struct perf_event *event)
1155 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1156 PERF_EVENT_STATE_INACTIVE;
1160 * Called at perf_event creation and when events are attached/detached from a
1163 static void perf_event__read_size(struct perf_event *event)
1165 int entry = sizeof(u64); /* value */
1169 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1170 size += sizeof(u64);
1172 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1173 size += sizeof(u64);
1175 if (event->attr.read_format & PERF_FORMAT_ID)
1176 entry += sizeof(u64);
1178 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1179 nr += event->group_leader->nr_siblings;
1180 size += sizeof(u64);
1184 event->read_size = size;
1187 static void perf_event__header_size(struct perf_event *event)
1189 struct perf_sample_data *data;
1190 u64 sample_type = event->attr.sample_type;
1193 perf_event__read_size(event);
1195 if (sample_type & PERF_SAMPLE_IP)
1196 size += sizeof(data->ip);
1198 if (sample_type & PERF_SAMPLE_ADDR)
1199 size += sizeof(data->addr);
1201 if (sample_type & PERF_SAMPLE_PERIOD)
1202 size += sizeof(data->period);
1204 if (sample_type & PERF_SAMPLE_WEIGHT)
1205 size += sizeof(data->weight);
1207 if (sample_type & PERF_SAMPLE_READ)
1208 size += event->read_size;
1210 if (sample_type & PERF_SAMPLE_DATA_SRC)
1211 size += sizeof(data->data_src.val);
1213 if (sample_type & PERF_SAMPLE_TRANSACTION)
1214 size += sizeof(data->txn);
1216 event->header_size = size;
1219 static void perf_event__id_header_size(struct perf_event *event)
1221 struct perf_sample_data *data;
1222 u64 sample_type = event->attr.sample_type;
1225 if (sample_type & PERF_SAMPLE_TID)
1226 size += sizeof(data->tid_entry);
1228 if (sample_type & PERF_SAMPLE_TIME)
1229 size += sizeof(data->time);
1231 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1232 size += sizeof(data->id);
1234 if (sample_type & PERF_SAMPLE_ID)
1235 size += sizeof(data->id);
1237 if (sample_type & PERF_SAMPLE_STREAM_ID)
1238 size += sizeof(data->stream_id);
1240 if (sample_type & PERF_SAMPLE_CPU)
1241 size += sizeof(data->cpu_entry);
1243 event->id_header_size = size;
1246 static void perf_group_attach(struct perf_event *event)
1248 struct perf_event *group_leader = event->group_leader, *pos;
1251 * We can have double attach due to group movement in perf_event_open.
1253 if (event->attach_state & PERF_ATTACH_GROUP)
1256 event->attach_state |= PERF_ATTACH_GROUP;
1258 if (group_leader == event)
1261 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1262 !is_software_event(event))
1263 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1265 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1266 group_leader->nr_siblings++;
1268 perf_event__header_size(group_leader);
1270 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1271 perf_event__header_size(pos);
1275 * Remove a event from the lists for its context.
1276 * Must be called with ctx->mutex and ctx->lock held.
1279 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1281 struct perf_cpu_context *cpuctx;
1283 * We can have double detach due to exit/hot-unplug + close.
1285 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1288 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1290 if (is_cgroup_event(event)) {
1292 cpuctx = __get_cpu_context(ctx);
1294 * if there are no more cgroup events
1295 * then cler cgrp to avoid stale pointer
1296 * in update_cgrp_time_from_cpuctx()
1298 if (!ctx->nr_cgroups)
1299 cpuctx->cgrp = NULL;
1302 if (has_branch_stack(event))
1303 ctx->nr_branch_stack--;
1306 if (event->attr.inherit_stat)
1309 list_del_rcu(&event->event_entry);
1311 if (event->group_leader == event)
1312 list_del_init(&event->group_entry);
1314 update_group_times(event);
1317 * If event was in error state, then keep it
1318 * that way, otherwise bogus counts will be
1319 * returned on read(). The only way to get out
1320 * of error state is by explicit re-enabling
1323 if (event->state > PERF_EVENT_STATE_OFF)
1324 event->state = PERF_EVENT_STATE_OFF;
1329 static void perf_group_detach(struct perf_event *event)
1331 struct perf_event *sibling, *tmp;
1332 struct list_head *list = NULL;
1335 * We can have double detach due to exit/hot-unplug + close.
1337 if (!(event->attach_state & PERF_ATTACH_GROUP))
1340 event->attach_state &= ~PERF_ATTACH_GROUP;
1343 * If this is a sibling, remove it from its group.
1345 if (event->group_leader != event) {
1346 list_del_init(&event->group_entry);
1347 event->group_leader->nr_siblings--;
1351 if (!list_empty(&event->group_entry))
1352 list = &event->group_entry;
1355 * If this was a group event with sibling events then
1356 * upgrade the siblings to singleton events by adding them
1357 * to whatever list we are on.
1359 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1361 list_move_tail(&sibling->group_entry, list);
1362 sibling->group_leader = sibling;
1364 /* Inherit group flags from the previous leader */
1365 sibling->group_flags = event->group_flags;
1369 perf_event__header_size(event->group_leader);
1371 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1372 perf_event__header_size(tmp);
1376 event_filter_match(struct perf_event *event)
1378 return (event->cpu == -1 || event->cpu == smp_processor_id())
1379 && perf_cgroup_match(event);
1383 event_sched_out(struct perf_event *event,
1384 struct perf_cpu_context *cpuctx,
1385 struct perf_event_context *ctx)
1387 u64 tstamp = perf_event_time(event);
1390 * An event which could not be activated because of
1391 * filter mismatch still needs to have its timings
1392 * maintained, otherwise bogus information is return
1393 * via read() for time_enabled, time_running:
1395 if (event->state == PERF_EVENT_STATE_INACTIVE
1396 && !event_filter_match(event)) {
1397 delta = tstamp - event->tstamp_stopped;
1398 event->tstamp_running += delta;
1399 event->tstamp_stopped = tstamp;
1402 if (event->state != PERF_EVENT_STATE_ACTIVE)
1405 perf_pmu_disable(event->pmu);
1407 event->state = PERF_EVENT_STATE_INACTIVE;
1408 if (event->pending_disable) {
1409 event->pending_disable = 0;
1410 event->state = PERF_EVENT_STATE_OFF;
1412 event->tstamp_stopped = tstamp;
1413 event->pmu->del(event, 0);
1416 if (!is_software_event(event))
1417 cpuctx->active_oncpu--;
1419 if (event->attr.freq && event->attr.sample_freq)
1421 if (event->attr.exclusive || !cpuctx->active_oncpu)
1422 cpuctx->exclusive = 0;
1424 perf_pmu_enable(event->pmu);
1428 group_sched_out(struct perf_event *group_event,
1429 struct perf_cpu_context *cpuctx,
1430 struct perf_event_context *ctx)
1432 struct perf_event *event;
1433 int state = group_event->state;
1435 event_sched_out(group_event, cpuctx, ctx);
1438 * Schedule out siblings (if any):
1440 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1441 event_sched_out(event, cpuctx, ctx);
1443 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1444 cpuctx->exclusive = 0;
1447 struct remove_event {
1448 struct perf_event *event;
1453 * Cross CPU call to remove a performance event
1455 * We disable the event on the hardware level first. After that we
1456 * remove it from the context list.
1458 static int __perf_remove_from_context(void *info)
1460 struct remove_event *re = info;
1461 struct perf_event *event = re->event;
1462 struct perf_event_context *ctx = event->ctx;
1463 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1465 raw_spin_lock(&ctx->lock);
1466 event_sched_out(event, cpuctx, ctx);
1467 if (re->detach_group)
1468 perf_group_detach(event);
1469 list_del_event(event, ctx);
1470 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1472 cpuctx->task_ctx = NULL;
1474 raw_spin_unlock(&ctx->lock);
1481 * Remove the event from a task's (or a CPU's) list of events.
1483 * CPU events are removed with a smp call. For task events we only
1484 * call when the task is on a CPU.
1486 * If event->ctx is a cloned context, callers must make sure that
1487 * every task struct that event->ctx->task could possibly point to
1488 * remains valid. This is OK when called from perf_release since
1489 * that only calls us on the top-level context, which can't be a clone.
1490 * When called from perf_event_exit_task, it's OK because the
1491 * context has been detached from its task.
1493 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1495 struct perf_event_context *ctx = event->ctx;
1496 struct task_struct *task = ctx->task;
1497 struct remove_event re = {
1499 .detach_group = detach_group,
1502 lockdep_assert_held(&ctx->mutex);
1506 * Per cpu events are removed via an smp call and
1507 * the removal is always successful.
1509 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1514 if (!task_function_call(task, __perf_remove_from_context, &re))
1517 raw_spin_lock_irq(&ctx->lock);
1519 * If we failed to find a running task, but find the context active now
1520 * that we've acquired the ctx->lock, retry.
1522 if (ctx->is_active) {
1523 raw_spin_unlock_irq(&ctx->lock);
1528 * Since the task isn't running, its safe to remove the event, us
1529 * holding the ctx->lock ensures the task won't get scheduled in.
1532 perf_group_detach(event);
1533 list_del_event(event, ctx);
1534 raw_spin_unlock_irq(&ctx->lock);
1538 * Cross CPU call to disable a performance event
1540 int __perf_event_disable(void *info)
1542 struct perf_event *event = info;
1543 struct perf_event_context *ctx = event->ctx;
1544 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1547 * If this is a per-task event, need to check whether this
1548 * event's task is the current task on this cpu.
1550 * Can trigger due to concurrent perf_event_context_sched_out()
1551 * flipping contexts around.
1553 if (ctx->task && cpuctx->task_ctx != ctx)
1556 raw_spin_lock(&ctx->lock);
1559 * If the event is on, turn it off.
1560 * If it is in error state, leave it in error state.
1562 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1563 update_context_time(ctx);
1564 update_cgrp_time_from_event(event);
1565 update_group_times(event);
1566 if (event == event->group_leader)
1567 group_sched_out(event, cpuctx, ctx);
1569 event_sched_out(event, cpuctx, ctx);
1570 event->state = PERF_EVENT_STATE_OFF;
1573 raw_spin_unlock(&ctx->lock);
1581 * If event->ctx is a cloned context, callers must make sure that
1582 * every task struct that event->ctx->task could possibly point to
1583 * remains valid. This condition is satisifed when called through
1584 * perf_event_for_each_child or perf_event_for_each because they
1585 * hold the top-level event's child_mutex, so any descendant that
1586 * goes to exit will block in sync_child_event.
1587 * When called from perf_pending_event it's OK because event->ctx
1588 * is the current context on this CPU and preemption is disabled,
1589 * hence we can't get into perf_event_task_sched_out for this context.
1591 void perf_event_disable(struct perf_event *event)
1593 struct perf_event_context *ctx = event->ctx;
1594 struct task_struct *task = ctx->task;
1598 * Disable the event on the cpu that it's on
1600 cpu_function_call(event->cpu, __perf_event_disable, event);
1605 if (!task_function_call(task, __perf_event_disable, event))
1608 raw_spin_lock_irq(&ctx->lock);
1610 * If the event is still active, we need to retry the cross-call.
1612 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1613 raw_spin_unlock_irq(&ctx->lock);
1615 * Reload the task pointer, it might have been changed by
1616 * a concurrent perf_event_context_sched_out().
1623 * Since we have the lock this context can't be scheduled
1624 * in, so we can change the state safely.
1626 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1627 update_group_times(event);
1628 event->state = PERF_EVENT_STATE_OFF;
1630 raw_spin_unlock_irq(&ctx->lock);
1632 EXPORT_SYMBOL_GPL(perf_event_disable);
1634 static void perf_set_shadow_time(struct perf_event *event,
1635 struct perf_event_context *ctx,
1639 * use the correct time source for the time snapshot
1641 * We could get by without this by leveraging the
1642 * fact that to get to this function, the caller
1643 * has most likely already called update_context_time()
1644 * and update_cgrp_time_xx() and thus both timestamp
1645 * are identical (or very close). Given that tstamp is,
1646 * already adjusted for cgroup, we could say that:
1647 * tstamp - ctx->timestamp
1649 * tstamp - cgrp->timestamp.
1651 * Then, in perf_output_read(), the calculation would
1652 * work with no changes because:
1653 * - event is guaranteed scheduled in
1654 * - no scheduled out in between
1655 * - thus the timestamp would be the same
1657 * But this is a bit hairy.
1659 * So instead, we have an explicit cgroup call to remain
1660 * within the time time source all along. We believe it
1661 * is cleaner and simpler to understand.
1663 if (is_cgroup_event(event))
1664 perf_cgroup_set_shadow_time(event, tstamp);
1666 event->shadow_ctx_time = tstamp - ctx->timestamp;
1669 #define MAX_INTERRUPTS (~0ULL)
1671 static void perf_log_throttle(struct perf_event *event, int enable);
1674 event_sched_in(struct perf_event *event,
1675 struct perf_cpu_context *cpuctx,
1676 struct perf_event_context *ctx)
1678 u64 tstamp = perf_event_time(event);
1681 lockdep_assert_held(&ctx->lock);
1683 if (event->state <= PERF_EVENT_STATE_OFF)
1686 event->state = PERF_EVENT_STATE_ACTIVE;
1687 event->oncpu = smp_processor_id();
1690 * Unthrottle events, since we scheduled we might have missed several
1691 * ticks already, also for a heavily scheduling task there is little
1692 * guarantee it'll get a tick in a timely manner.
1694 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1695 perf_log_throttle(event, 1);
1696 event->hw.interrupts = 0;
1700 * The new state must be visible before we turn it on in the hardware:
1704 perf_pmu_disable(event->pmu);
1706 if (event->pmu->add(event, PERF_EF_START)) {
1707 event->state = PERF_EVENT_STATE_INACTIVE;
1713 event->tstamp_running += tstamp - event->tstamp_stopped;
1715 perf_set_shadow_time(event, ctx, tstamp);
1717 if (!is_software_event(event))
1718 cpuctx->active_oncpu++;
1720 if (event->attr.freq && event->attr.sample_freq)
1723 if (event->attr.exclusive)
1724 cpuctx->exclusive = 1;
1727 perf_pmu_enable(event->pmu);
1733 group_sched_in(struct perf_event *group_event,
1734 struct perf_cpu_context *cpuctx,
1735 struct perf_event_context *ctx)
1737 struct perf_event *event, *partial_group = NULL;
1738 struct pmu *pmu = ctx->pmu;
1739 u64 now = ctx->time;
1740 bool simulate = false;
1742 if (group_event->state == PERF_EVENT_STATE_OFF)
1745 pmu->start_txn(pmu);
1747 if (event_sched_in(group_event, cpuctx, ctx)) {
1748 pmu->cancel_txn(pmu);
1749 perf_cpu_hrtimer_restart(cpuctx);
1754 * Schedule in siblings as one group (if any):
1756 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1757 if (event_sched_in(event, cpuctx, ctx)) {
1758 partial_group = event;
1763 if (!pmu->commit_txn(pmu))
1768 * Groups can be scheduled in as one unit only, so undo any
1769 * partial group before returning:
1770 * The events up to the failed event are scheduled out normally,
1771 * tstamp_stopped will be updated.
1773 * The failed events and the remaining siblings need to have
1774 * their timings updated as if they had gone thru event_sched_in()
1775 * and event_sched_out(). This is required to get consistent timings
1776 * across the group. This also takes care of the case where the group
1777 * could never be scheduled by ensuring tstamp_stopped is set to mark
1778 * the time the event was actually stopped, such that time delta
1779 * calculation in update_event_times() is correct.
1781 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1782 if (event == partial_group)
1786 event->tstamp_running += now - event->tstamp_stopped;
1787 event->tstamp_stopped = now;
1789 event_sched_out(event, cpuctx, ctx);
1792 event_sched_out(group_event, cpuctx, ctx);
1794 pmu->cancel_txn(pmu);
1796 perf_cpu_hrtimer_restart(cpuctx);
1802 * Work out whether we can put this event group on the CPU now.
1804 static int group_can_go_on(struct perf_event *event,
1805 struct perf_cpu_context *cpuctx,
1809 * Groups consisting entirely of software events can always go on.
1811 if (event->group_flags & PERF_GROUP_SOFTWARE)
1814 * If an exclusive group is already on, no other hardware
1817 if (cpuctx->exclusive)
1820 * If this group is exclusive and there are already
1821 * events on the CPU, it can't go on.
1823 if (event->attr.exclusive && cpuctx->active_oncpu)
1826 * Otherwise, try to add it if all previous groups were able
1832 static void add_event_to_ctx(struct perf_event *event,
1833 struct perf_event_context *ctx)
1835 u64 tstamp = perf_event_time(event);
1837 list_add_event(event, ctx);
1838 perf_group_attach(event);
1839 event->tstamp_enabled = tstamp;
1840 event->tstamp_running = tstamp;
1841 event->tstamp_stopped = tstamp;
1844 static void task_ctx_sched_out(struct perf_event_context *ctx);
1846 ctx_sched_in(struct perf_event_context *ctx,
1847 struct perf_cpu_context *cpuctx,
1848 enum event_type_t event_type,
1849 struct task_struct *task);
1851 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1852 struct perf_event_context *ctx,
1853 struct task_struct *task)
1855 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1857 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1858 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1860 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1864 * Cross CPU call to install and enable a performance event
1866 * Must be called with ctx->mutex held
1868 static int __perf_install_in_context(void *info)
1870 struct perf_event *event = info;
1871 struct perf_event_context *ctx = event->ctx;
1872 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1873 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1874 struct task_struct *task = current;
1876 perf_ctx_lock(cpuctx, task_ctx);
1877 perf_pmu_disable(cpuctx->ctx.pmu);
1880 * If there was an active task_ctx schedule it out.
1883 task_ctx_sched_out(task_ctx);
1886 * If the context we're installing events in is not the
1887 * active task_ctx, flip them.
1889 if (ctx->task && task_ctx != ctx) {
1891 raw_spin_unlock(&task_ctx->lock);
1892 raw_spin_lock(&ctx->lock);
1897 cpuctx->task_ctx = task_ctx;
1898 task = task_ctx->task;
1901 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1903 update_context_time(ctx);
1905 * update cgrp time only if current cgrp
1906 * matches event->cgrp. Must be done before
1907 * calling add_event_to_ctx()
1909 update_cgrp_time_from_event(event);
1911 add_event_to_ctx(event, ctx);
1914 * Schedule everything back in
1916 perf_event_sched_in(cpuctx, task_ctx, task);
1918 perf_pmu_enable(cpuctx->ctx.pmu);
1919 perf_ctx_unlock(cpuctx, task_ctx);
1925 * Attach a performance event to a context
1927 * First we add the event to the list with the hardware enable bit
1928 * in event->hw_config cleared.
1930 * If the event is attached to a task which is on a CPU we use a smp
1931 * call to enable it in the task context. The task might have been
1932 * scheduled away, but we check this in the smp call again.
1935 perf_install_in_context(struct perf_event_context *ctx,
1936 struct perf_event *event,
1939 struct task_struct *task = ctx->task;
1941 lockdep_assert_held(&ctx->mutex);
1944 if (event->cpu != -1)
1949 * Per cpu events are installed via an smp call and
1950 * the install is always successful.
1952 cpu_function_call(cpu, __perf_install_in_context, event);
1957 if (!task_function_call(task, __perf_install_in_context, event))
1960 raw_spin_lock_irq(&ctx->lock);
1962 * If we failed to find a running task, but find the context active now
1963 * that we've acquired the ctx->lock, retry.
1965 if (ctx->is_active) {
1966 raw_spin_unlock_irq(&ctx->lock);
1971 * Since the task isn't running, its safe to add the event, us holding
1972 * the ctx->lock ensures the task won't get scheduled in.
1974 add_event_to_ctx(event, ctx);
1975 raw_spin_unlock_irq(&ctx->lock);
1979 * Put a event into inactive state and update time fields.
1980 * Enabling the leader of a group effectively enables all
1981 * the group members that aren't explicitly disabled, so we
1982 * have to update their ->tstamp_enabled also.
1983 * Note: this works for group members as well as group leaders
1984 * since the non-leader members' sibling_lists will be empty.
1986 static void __perf_event_mark_enabled(struct perf_event *event)
1988 struct perf_event *sub;
1989 u64 tstamp = perf_event_time(event);
1991 event->state = PERF_EVENT_STATE_INACTIVE;
1992 event->tstamp_enabled = tstamp - event->total_time_enabled;
1993 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1994 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1995 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2000 * Cross CPU call to enable a performance event
2002 static int __perf_event_enable(void *info)
2004 struct perf_event *event = info;
2005 struct perf_event_context *ctx = event->ctx;
2006 struct perf_event *leader = event->group_leader;
2007 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2011 * There's a time window between 'ctx->is_active' check
2012 * in perf_event_enable function and this place having:
2014 * - ctx->lock unlocked
2016 * where the task could be killed and 'ctx' deactivated
2017 * by perf_event_exit_task.
2019 if (!ctx->is_active)
2022 raw_spin_lock(&ctx->lock);
2023 update_context_time(ctx);
2025 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2029 * set current task's cgroup time reference point
2031 perf_cgroup_set_timestamp(current, ctx);
2033 __perf_event_mark_enabled(event);
2035 if (!event_filter_match(event)) {
2036 if (is_cgroup_event(event))
2037 perf_cgroup_defer_enabled(event);
2042 * If the event is in a group and isn't the group leader,
2043 * then don't put it on unless the group is on.
2045 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2048 if (!group_can_go_on(event, cpuctx, 1)) {
2051 if (event == leader)
2052 err = group_sched_in(event, cpuctx, ctx);
2054 err = event_sched_in(event, cpuctx, ctx);
2059 * If this event can't go on and it's part of a
2060 * group, then the whole group has to come off.
2062 if (leader != event) {
2063 group_sched_out(leader, cpuctx, ctx);
2064 perf_cpu_hrtimer_restart(cpuctx);
2066 if (leader->attr.pinned) {
2067 update_group_times(leader);
2068 leader->state = PERF_EVENT_STATE_ERROR;
2073 raw_spin_unlock(&ctx->lock);
2081 * If event->ctx is a cloned context, callers must make sure that
2082 * every task struct that event->ctx->task could possibly point to
2083 * remains valid. This condition is satisfied when called through
2084 * perf_event_for_each_child or perf_event_for_each as described
2085 * for perf_event_disable.
2087 void perf_event_enable(struct perf_event *event)
2089 struct perf_event_context *ctx = event->ctx;
2090 struct task_struct *task = ctx->task;
2094 * Enable the event on the cpu that it's on
2096 cpu_function_call(event->cpu, __perf_event_enable, event);
2100 raw_spin_lock_irq(&ctx->lock);
2101 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2105 * If the event is in error state, clear that first.
2106 * That way, if we see the event in error state below, we
2107 * know that it has gone back into error state, as distinct
2108 * from the task having been scheduled away before the
2109 * cross-call arrived.
2111 if (event->state == PERF_EVENT_STATE_ERROR)
2112 event->state = PERF_EVENT_STATE_OFF;
2115 if (!ctx->is_active) {
2116 __perf_event_mark_enabled(event);
2120 raw_spin_unlock_irq(&ctx->lock);
2122 if (!task_function_call(task, __perf_event_enable, event))
2125 raw_spin_lock_irq(&ctx->lock);
2128 * If the context is active and the event is still off,
2129 * we need to retry the cross-call.
2131 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2133 * task could have been flipped by a concurrent
2134 * perf_event_context_sched_out()
2141 raw_spin_unlock_irq(&ctx->lock);
2143 EXPORT_SYMBOL_GPL(perf_event_enable);
2145 int perf_event_refresh(struct perf_event *event, int refresh)
2148 * not supported on inherited events
2150 if (event->attr.inherit || !is_sampling_event(event))
2153 atomic_add(refresh, &event->event_limit);
2154 perf_event_enable(event);
2158 EXPORT_SYMBOL_GPL(perf_event_refresh);
2160 static void ctx_sched_out(struct perf_event_context *ctx,
2161 struct perf_cpu_context *cpuctx,
2162 enum event_type_t event_type)
2164 struct perf_event *event;
2165 int is_active = ctx->is_active;
2167 ctx->is_active &= ~event_type;
2168 if (likely(!ctx->nr_events))
2171 update_context_time(ctx);
2172 update_cgrp_time_from_cpuctx(cpuctx);
2173 if (!ctx->nr_active)
2176 perf_pmu_disable(ctx->pmu);
2177 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2178 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2179 group_sched_out(event, cpuctx, ctx);
2182 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2183 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2184 group_sched_out(event, cpuctx, ctx);
2186 perf_pmu_enable(ctx->pmu);
2190 * Test whether two contexts are equivalent, i.e. whether they have both been
2191 * cloned from the same version of the same context.
2193 * Equivalence is measured using a generation number in the context that is
2194 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2195 * and list_del_event().
2197 static int context_equiv(struct perf_event_context *ctx1,
2198 struct perf_event_context *ctx2)
2200 /* Pinning disables the swap optimization */
2201 if (ctx1->pin_count || ctx2->pin_count)
2204 /* If ctx1 is the parent of ctx2 */
2205 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2208 /* If ctx2 is the parent of ctx1 */
2209 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2213 * If ctx1 and ctx2 have the same parent; we flatten the parent
2214 * hierarchy, see perf_event_init_context().
2216 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2217 ctx1->parent_gen == ctx2->parent_gen)
2224 static void __perf_event_sync_stat(struct perf_event *event,
2225 struct perf_event *next_event)
2229 if (!event->attr.inherit_stat)
2233 * Update the event value, we cannot use perf_event_read()
2234 * because we're in the middle of a context switch and have IRQs
2235 * disabled, which upsets smp_call_function_single(), however
2236 * we know the event must be on the current CPU, therefore we
2237 * don't need to use it.
2239 switch (event->state) {
2240 case PERF_EVENT_STATE_ACTIVE:
2241 event->pmu->read(event);
2244 case PERF_EVENT_STATE_INACTIVE:
2245 update_event_times(event);
2253 * In order to keep per-task stats reliable we need to flip the event
2254 * values when we flip the contexts.
2256 value = local64_read(&next_event->count);
2257 value = local64_xchg(&event->count, value);
2258 local64_set(&next_event->count, value);
2260 swap(event->total_time_enabled, next_event->total_time_enabled);
2261 swap(event->total_time_running, next_event->total_time_running);
2264 * Since we swizzled the values, update the user visible data too.
2266 perf_event_update_userpage(event);
2267 perf_event_update_userpage(next_event);
2270 static void perf_event_sync_stat(struct perf_event_context *ctx,
2271 struct perf_event_context *next_ctx)
2273 struct perf_event *event, *next_event;
2278 update_context_time(ctx);
2280 event = list_first_entry(&ctx->event_list,
2281 struct perf_event, event_entry);
2283 next_event = list_first_entry(&next_ctx->event_list,
2284 struct perf_event, event_entry);
2286 while (&event->event_entry != &ctx->event_list &&
2287 &next_event->event_entry != &next_ctx->event_list) {
2289 __perf_event_sync_stat(event, next_event);
2291 event = list_next_entry(event, event_entry);
2292 next_event = list_next_entry(next_event, event_entry);
2296 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2297 struct task_struct *next)
2299 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2300 struct perf_event_context *next_ctx;
2301 struct perf_event_context *parent, *next_parent;
2302 struct perf_cpu_context *cpuctx;
2308 cpuctx = __get_cpu_context(ctx);
2309 if (!cpuctx->task_ctx)
2313 next_ctx = next->perf_event_ctxp[ctxn];
2317 parent = rcu_dereference(ctx->parent_ctx);
2318 next_parent = rcu_dereference(next_ctx->parent_ctx);
2320 /* If neither context have a parent context; they cannot be clones. */
2321 if (!parent && !next_parent)
2324 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2326 * Looks like the two contexts are clones, so we might be
2327 * able to optimize the context switch. We lock both
2328 * contexts and check that they are clones under the
2329 * lock (including re-checking that neither has been
2330 * uncloned in the meantime). It doesn't matter which
2331 * order we take the locks because no other cpu could
2332 * be trying to lock both of these tasks.
2334 raw_spin_lock(&ctx->lock);
2335 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2336 if (context_equiv(ctx, next_ctx)) {
2338 * XXX do we need a memory barrier of sorts
2339 * wrt to rcu_dereference() of perf_event_ctxp
2341 task->perf_event_ctxp[ctxn] = next_ctx;
2342 next->perf_event_ctxp[ctxn] = ctx;
2344 next_ctx->task = task;
2347 perf_event_sync_stat(ctx, next_ctx);
2349 raw_spin_unlock(&next_ctx->lock);
2350 raw_spin_unlock(&ctx->lock);
2356 raw_spin_lock(&ctx->lock);
2357 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2358 cpuctx->task_ctx = NULL;
2359 raw_spin_unlock(&ctx->lock);
2363 #define for_each_task_context_nr(ctxn) \
2364 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2367 * Called from scheduler to remove the events of the current task,
2368 * with interrupts disabled.
2370 * We stop each event and update the event value in event->count.
2372 * This does not protect us against NMI, but disable()
2373 * sets the disabled bit in the control field of event _before_
2374 * accessing the event control register. If a NMI hits, then it will
2375 * not restart the event.
2377 void __perf_event_task_sched_out(struct task_struct *task,
2378 struct task_struct *next)
2382 for_each_task_context_nr(ctxn)
2383 perf_event_context_sched_out(task, ctxn, next);
2386 * if cgroup events exist on this CPU, then we need
2387 * to check if we have to switch out PMU state.
2388 * cgroup event are system-wide mode only
2390 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2391 perf_cgroup_sched_out(task, next);
2394 static void task_ctx_sched_out(struct perf_event_context *ctx)
2396 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2398 if (!cpuctx->task_ctx)
2401 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2404 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2405 cpuctx->task_ctx = NULL;
2409 * Called with IRQs disabled
2411 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2412 enum event_type_t event_type)
2414 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2418 ctx_pinned_sched_in(struct perf_event_context *ctx,
2419 struct perf_cpu_context *cpuctx)
2421 struct perf_event *event;
2423 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2424 if (event->state <= PERF_EVENT_STATE_OFF)
2426 if (!event_filter_match(event))
2429 /* may need to reset tstamp_enabled */
2430 if (is_cgroup_event(event))
2431 perf_cgroup_mark_enabled(event, ctx);
2433 if (group_can_go_on(event, cpuctx, 1))
2434 group_sched_in(event, cpuctx, ctx);
2437 * If this pinned group hasn't been scheduled,
2438 * put it in error state.
2440 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2441 update_group_times(event);
2442 event->state = PERF_EVENT_STATE_ERROR;
2448 ctx_flexible_sched_in(struct perf_event_context *ctx,
2449 struct perf_cpu_context *cpuctx)
2451 struct perf_event *event;
2454 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2455 /* Ignore events in OFF or ERROR state */
2456 if (event->state <= PERF_EVENT_STATE_OFF)
2459 * Listen to the 'cpu' scheduling filter constraint
2462 if (!event_filter_match(event))
2465 /* may need to reset tstamp_enabled */
2466 if (is_cgroup_event(event))
2467 perf_cgroup_mark_enabled(event, ctx);
2469 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2470 if (group_sched_in(event, cpuctx, ctx))
2477 ctx_sched_in(struct perf_event_context *ctx,
2478 struct perf_cpu_context *cpuctx,
2479 enum event_type_t event_type,
2480 struct task_struct *task)
2483 int is_active = ctx->is_active;
2485 ctx->is_active |= event_type;
2486 if (likely(!ctx->nr_events))
2490 ctx->timestamp = now;
2491 perf_cgroup_set_timestamp(task, ctx);
2493 * First go through the list and put on any pinned groups
2494 * in order to give them the best chance of going on.
2496 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2497 ctx_pinned_sched_in(ctx, cpuctx);
2499 /* Then walk through the lower prio flexible groups */
2500 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2501 ctx_flexible_sched_in(ctx, cpuctx);
2504 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2505 enum event_type_t event_type,
2506 struct task_struct *task)
2508 struct perf_event_context *ctx = &cpuctx->ctx;
2510 ctx_sched_in(ctx, cpuctx, event_type, task);
2513 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2514 struct task_struct *task)
2516 struct perf_cpu_context *cpuctx;
2518 cpuctx = __get_cpu_context(ctx);
2519 if (cpuctx->task_ctx == ctx)
2522 perf_ctx_lock(cpuctx, ctx);
2523 perf_pmu_disable(ctx->pmu);
2525 * We want to keep the following priority order:
2526 * cpu pinned (that don't need to move), task pinned,
2527 * cpu flexible, task flexible.
2529 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2532 cpuctx->task_ctx = ctx;
2534 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2536 perf_pmu_enable(ctx->pmu);
2537 perf_ctx_unlock(cpuctx, ctx);
2540 * Since these rotations are per-cpu, we need to ensure the
2541 * cpu-context we got scheduled on is actually rotating.
2543 perf_pmu_rotate_start(ctx->pmu);
2547 * When sampling the branck stack in system-wide, it may be necessary
2548 * to flush the stack on context switch. This happens when the branch
2549 * stack does not tag its entries with the pid of the current task.
2550 * Otherwise it becomes impossible to associate a branch entry with a
2551 * task. This ambiguity is more likely to appear when the branch stack
2552 * supports priv level filtering and the user sets it to monitor only
2553 * at the user level (which could be a useful measurement in system-wide
2554 * mode). In that case, the risk is high of having a branch stack with
2555 * branch from multiple tasks. Flushing may mean dropping the existing
2556 * entries or stashing them somewhere in the PMU specific code layer.
2558 * This function provides the context switch callback to the lower code
2559 * layer. It is invoked ONLY when there is at least one system-wide context
2560 * with at least one active event using taken branch sampling.
2562 static void perf_branch_stack_sched_in(struct task_struct *prev,
2563 struct task_struct *task)
2565 struct perf_cpu_context *cpuctx;
2567 unsigned long flags;
2569 /* no need to flush branch stack if not changing task */
2573 local_irq_save(flags);
2577 list_for_each_entry_rcu(pmu, &pmus, entry) {
2578 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2581 * check if the context has at least one
2582 * event using PERF_SAMPLE_BRANCH_STACK
2584 if (cpuctx->ctx.nr_branch_stack > 0
2585 && pmu->flush_branch_stack) {
2587 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2589 perf_pmu_disable(pmu);
2591 pmu->flush_branch_stack();
2593 perf_pmu_enable(pmu);
2595 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2601 local_irq_restore(flags);
2605 * Called from scheduler to add the events of the current task
2606 * with interrupts disabled.
2608 * We restore the event value and then enable it.
2610 * This does not protect us against NMI, but enable()
2611 * sets the enabled bit in the control field of event _before_
2612 * accessing the event control register. If a NMI hits, then it will
2613 * keep the event running.
2615 void __perf_event_task_sched_in(struct task_struct *prev,
2616 struct task_struct *task)
2618 struct perf_event_context *ctx;
2621 for_each_task_context_nr(ctxn) {
2622 ctx = task->perf_event_ctxp[ctxn];
2626 perf_event_context_sched_in(ctx, task);
2629 * if cgroup events exist on this CPU, then we need
2630 * to check if we have to switch in PMU state.
2631 * cgroup event are system-wide mode only
2633 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2634 perf_cgroup_sched_in(prev, task);
2636 /* check for system-wide branch_stack events */
2637 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2638 perf_branch_stack_sched_in(prev, task);
2641 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2643 u64 frequency = event->attr.sample_freq;
2644 u64 sec = NSEC_PER_SEC;
2645 u64 divisor, dividend;
2647 int count_fls, nsec_fls, frequency_fls, sec_fls;
2649 count_fls = fls64(count);
2650 nsec_fls = fls64(nsec);
2651 frequency_fls = fls64(frequency);
2655 * We got @count in @nsec, with a target of sample_freq HZ
2656 * the target period becomes:
2659 * period = -------------------
2660 * @nsec * sample_freq
2665 * Reduce accuracy by one bit such that @a and @b converge
2666 * to a similar magnitude.
2668 #define REDUCE_FLS(a, b) \
2670 if (a##_fls > b##_fls) { \
2680 * Reduce accuracy until either term fits in a u64, then proceed with
2681 * the other, so that finally we can do a u64/u64 division.
2683 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2684 REDUCE_FLS(nsec, frequency);
2685 REDUCE_FLS(sec, count);
2688 if (count_fls + sec_fls > 64) {
2689 divisor = nsec * frequency;
2691 while (count_fls + sec_fls > 64) {
2692 REDUCE_FLS(count, sec);
2696 dividend = count * sec;
2698 dividend = count * sec;
2700 while (nsec_fls + frequency_fls > 64) {
2701 REDUCE_FLS(nsec, frequency);
2705 divisor = nsec * frequency;
2711 return div64_u64(dividend, divisor);
2714 static DEFINE_PER_CPU(int, perf_throttled_count);
2715 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2717 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2719 struct hw_perf_event *hwc = &event->hw;
2720 s64 period, sample_period;
2723 period = perf_calculate_period(event, nsec, count);
2725 delta = (s64)(period - hwc->sample_period);
2726 delta = (delta + 7) / 8; /* low pass filter */
2728 sample_period = hwc->sample_period + delta;
2733 hwc->sample_period = sample_period;
2735 if (local64_read(&hwc->period_left) > 8*sample_period) {
2737 event->pmu->stop(event, PERF_EF_UPDATE);
2739 local64_set(&hwc->period_left, 0);
2742 event->pmu->start(event, PERF_EF_RELOAD);
2747 * combine freq adjustment with unthrottling to avoid two passes over the
2748 * events. At the same time, make sure, having freq events does not change
2749 * the rate of unthrottling as that would introduce bias.
2751 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2754 struct perf_event *event;
2755 struct hw_perf_event *hwc;
2756 u64 now, period = TICK_NSEC;
2760 * only need to iterate over all events iff:
2761 * - context have events in frequency mode (needs freq adjust)
2762 * - there are events to unthrottle on this cpu
2764 if (!(ctx->nr_freq || needs_unthr))
2767 raw_spin_lock(&ctx->lock);
2768 perf_pmu_disable(ctx->pmu);
2770 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2771 if (event->state != PERF_EVENT_STATE_ACTIVE)
2774 if (!event_filter_match(event))
2777 perf_pmu_disable(event->pmu);
2781 if (hwc->interrupts == MAX_INTERRUPTS) {
2782 hwc->interrupts = 0;
2783 perf_log_throttle(event, 1);
2784 event->pmu->start(event, 0);
2787 if (!event->attr.freq || !event->attr.sample_freq)
2791 * stop the event and update event->count
2793 event->pmu->stop(event, PERF_EF_UPDATE);
2795 now = local64_read(&event->count);
2796 delta = now - hwc->freq_count_stamp;
2797 hwc->freq_count_stamp = now;
2801 * reload only if value has changed
2802 * we have stopped the event so tell that
2803 * to perf_adjust_period() to avoid stopping it
2807 perf_adjust_period(event, period, delta, false);
2809 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2811 perf_pmu_enable(event->pmu);
2814 perf_pmu_enable(ctx->pmu);
2815 raw_spin_unlock(&ctx->lock);
2819 * Round-robin a context's events:
2821 static void rotate_ctx(struct perf_event_context *ctx)
2824 * Rotate the first entry last of non-pinned groups. Rotation might be
2825 * disabled by the inheritance code.
2827 if (!ctx->rotate_disable)
2828 list_rotate_left(&ctx->flexible_groups);
2832 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2833 * because they're strictly cpu affine and rotate_start is called with IRQs
2834 * disabled, while rotate_context is called from IRQ context.
2836 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2838 struct perf_event_context *ctx = NULL;
2839 int rotate = 0, remove = 1;
2841 if (cpuctx->ctx.nr_events) {
2843 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2847 ctx = cpuctx->task_ctx;
2848 if (ctx && ctx->nr_events) {
2850 if (ctx->nr_events != ctx->nr_active)
2857 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2858 perf_pmu_disable(cpuctx->ctx.pmu);
2860 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2862 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2864 rotate_ctx(&cpuctx->ctx);
2868 perf_event_sched_in(cpuctx, ctx, current);
2870 perf_pmu_enable(cpuctx->ctx.pmu);
2871 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2874 list_del_init(&cpuctx->rotation_list);
2879 #ifdef CONFIG_NO_HZ_FULL
2880 bool perf_event_can_stop_tick(void)
2882 if (atomic_read(&nr_freq_events) ||
2883 __this_cpu_read(perf_throttled_count))
2890 void perf_event_task_tick(void)
2892 struct list_head *head = &__get_cpu_var(rotation_list);
2893 struct perf_cpu_context *cpuctx, *tmp;
2894 struct perf_event_context *ctx;
2897 WARN_ON(!irqs_disabled());
2899 __this_cpu_inc(perf_throttled_seq);
2900 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2902 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2904 perf_adjust_freq_unthr_context(ctx, throttled);
2906 ctx = cpuctx->task_ctx;
2908 perf_adjust_freq_unthr_context(ctx, throttled);
2912 static int event_enable_on_exec(struct perf_event *event,
2913 struct perf_event_context *ctx)
2915 if (!event->attr.enable_on_exec)
2918 event->attr.enable_on_exec = 0;
2919 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2922 __perf_event_mark_enabled(event);
2928 * Enable all of a task's events that have been marked enable-on-exec.
2929 * This expects task == current.
2931 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2933 struct perf_event *event;
2934 unsigned long flags;
2938 local_irq_save(flags);
2939 if (!ctx || !ctx->nr_events)
2943 * We must ctxsw out cgroup events to avoid conflict
2944 * when invoking perf_task_event_sched_in() later on
2945 * in this function. Otherwise we end up trying to
2946 * ctxswin cgroup events which are already scheduled
2949 perf_cgroup_sched_out(current, NULL);
2951 raw_spin_lock(&ctx->lock);
2952 task_ctx_sched_out(ctx);
2954 list_for_each_entry(event, &ctx->event_list, event_entry) {
2955 ret = event_enable_on_exec(event, ctx);
2961 * Unclone this context if we enabled any event.
2966 raw_spin_unlock(&ctx->lock);
2969 * Also calls ctxswin for cgroup events, if any:
2971 perf_event_context_sched_in(ctx, ctx->task);
2973 local_irq_restore(flags);
2976 void perf_event_exec(void)
2978 struct perf_event_context *ctx;
2982 for_each_task_context_nr(ctxn) {
2983 ctx = current->perf_event_ctxp[ctxn];
2987 perf_event_enable_on_exec(ctx);
2993 * Cross CPU call to read the hardware event
2995 static void __perf_event_read(void *info)
2997 struct perf_event *event = info;
2998 struct perf_event_context *ctx = event->ctx;
2999 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3002 * If this is a task context, we need to check whether it is
3003 * the current task context of this cpu. If not it has been
3004 * scheduled out before the smp call arrived. In that case
3005 * event->count would have been updated to a recent sample
3006 * when the event was scheduled out.
3008 if (ctx->task && cpuctx->task_ctx != ctx)
3011 raw_spin_lock(&ctx->lock);
3012 if (ctx->is_active) {
3013 update_context_time(ctx);
3014 update_cgrp_time_from_event(event);
3016 update_event_times(event);
3017 if (event->state == PERF_EVENT_STATE_ACTIVE)
3018 event->pmu->read(event);
3019 raw_spin_unlock(&ctx->lock);
3022 static inline u64 perf_event_count(struct perf_event *event)
3024 return local64_read(&event->count) + atomic64_read(&event->child_count);
3027 static u64 perf_event_read(struct perf_event *event)
3030 * If event is enabled and currently active on a CPU, update the
3031 * value in the event structure:
3033 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3034 smp_call_function_single(event->oncpu,
3035 __perf_event_read, event, 1);
3036 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3037 struct perf_event_context *ctx = event->ctx;
3038 unsigned long flags;
3040 raw_spin_lock_irqsave(&ctx->lock, flags);
3042 * may read while context is not active
3043 * (e.g., thread is blocked), in that case
3044 * we cannot update context time
3046 if (ctx->is_active) {
3047 update_context_time(ctx);
3048 update_cgrp_time_from_event(event);
3050 update_event_times(event);
3051 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3054 return perf_event_count(event);
3058 * Initialize the perf_event context in a task_struct:
3060 static void __perf_event_init_context(struct perf_event_context *ctx)
3062 raw_spin_lock_init(&ctx->lock);
3063 mutex_init(&ctx->mutex);
3064 INIT_LIST_HEAD(&ctx->pinned_groups);
3065 INIT_LIST_HEAD(&ctx->flexible_groups);
3066 INIT_LIST_HEAD(&ctx->event_list);
3067 atomic_set(&ctx->refcount, 1);
3070 static struct perf_event_context *
3071 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3073 struct perf_event_context *ctx;
3075 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3079 __perf_event_init_context(ctx);
3082 get_task_struct(task);
3089 static struct task_struct *
3090 find_lively_task_by_vpid(pid_t vpid)
3092 struct task_struct *task;
3099 task = find_task_by_vpid(vpid);
3101 get_task_struct(task);
3105 return ERR_PTR(-ESRCH);
3107 /* Reuse ptrace permission checks for now. */
3109 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3114 put_task_struct(task);
3115 return ERR_PTR(err);
3120 * Returns a matching context with refcount and pincount.
3122 static struct perf_event_context *
3123 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3125 struct perf_event_context *ctx;
3126 struct perf_cpu_context *cpuctx;
3127 unsigned long flags;
3131 /* Must be root to operate on a CPU event: */
3132 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3133 return ERR_PTR(-EACCES);
3136 * We could be clever and allow to attach a event to an
3137 * offline CPU and activate it when the CPU comes up, but
3140 if (!cpu_online(cpu))
3141 return ERR_PTR(-ENODEV);
3143 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3152 ctxn = pmu->task_ctx_nr;
3157 ctx = perf_lock_task_context(task, ctxn, &flags);
3161 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3163 ctx = alloc_perf_context(pmu, task);
3169 mutex_lock(&task->perf_event_mutex);
3171 * If it has already passed perf_event_exit_task().
3172 * we must see PF_EXITING, it takes this mutex too.
3174 if (task->flags & PF_EXITING)
3176 else if (task->perf_event_ctxp[ctxn])
3181 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3183 mutex_unlock(&task->perf_event_mutex);
3185 if (unlikely(err)) {
3197 return ERR_PTR(err);
3200 static void perf_event_free_filter(struct perf_event *event);
3202 static void free_event_rcu(struct rcu_head *head)
3204 struct perf_event *event;
3206 event = container_of(head, struct perf_event, rcu_head);
3208 put_pid_ns(event->ns);
3209 perf_event_free_filter(event);
3213 static void ring_buffer_put(struct ring_buffer *rb);
3214 static void ring_buffer_attach(struct perf_event *event,
3215 struct ring_buffer *rb);
3217 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3222 if (has_branch_stack(event)) {
3223 if (!(event->attach_state & PERF_ATTACH_TASK))
3224 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3226 if (is_cgroup_event(event))
3227 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3230 static void unaccount_event(struct perf_event *event)
3235 if (event->attach_state & PERF_ATTACH_TASK)
3236 static_key_slow_dec_deferred(&perf_sched_events);
3237 if (event->attr.mmap || event->attr.mmap_data)
3238 atomic_dec(&nr_mmap_events);
3239 if (event->attr.comm)
3240 atomic_dec(&nr_comm_events);
3241 if (event->attr.task)
3242 atomic_dec(&nr_task_events);
3243 if (event->attr.freq)
3244 atomic_dec(&nr_freq_events);
3245 if (is_cgroup_event(event))
3246 static_key_slow_dec_deferred(&perf_sched_events);
3247 if (has_branch_stack(event))
3248 static_key_slow_dec_deferred(&perf_sched_events);
3250 unaccount_event_cpu(event, event->cpu);
3253 static void __free_event(struct perf_event *event)
3255 if (!event->parent) {
3256 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3257 put_callchain_buffers();
3261 event->destroy(event);
3264 put_ctx(event->ctx);
3267 module_put(event->pmu->module);
3269 call_rcu(&event->rcu_head, free_event_rcu);
3272 static void _free_event(struct perf_event *event)
3274 irq_work_sync(&event->pending);
3276 unaccount_event(event);
3280 * Can happen when we close an event with re-directed output.
3282 * Since we have a 0 refcount, perf_mmap_close() will skip
3283 * over us; possibly making our ring_buffer_put() the last.
3285 mutex_lock(&event->mmap_mutex);
3286 ring_buffer_attach(event, NULL);
3287 mutex_unlock(&event->mmap_mutex);
3290 if (is_cgroup_event(event))
3291 perf_detach_cgroup(event);
3293 __free_event(event);
3297 * Used to free events which have a known refcount of 1, such as in error paths
3298 * where the event isn't exposed yet and inherited events.
3300 static void free_event(struct perf_event *event)
3302 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3303 "unexpected event refcount: %ld; ptr=%p\n",
3304 atomic_long_read(&event->refcount), event)) {
3305 /* leak to avoid use-after-free */
3313 * Called when the last reference to the file is gone.
3315 static void put_event(struct perf_event *event)
3317 struct perf_event_context *ctx = event->ctx;
3318 struct task_struct *owner;
3320 if (!atomic_long_dec_and_test(&event->refcount))
3324 owner = ACCESS_ONCE(event->owner);
3326 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3327 * !owner it means the list deletion is complete and we can indeed
3328 * free this event, otherwise we need to serialize on
3329 * owner->perf_event_mutex.
3331 smp_read_barrier_depends();
3334 * Since delayed_put_task_struct() also drops the last
3335 * task reference we can safely take a new reference
3336 * while holding the rcu_read_lock().
3338 get_task_struct(owner);
3343 mutex_lock(&owner->perf_event_mutex);
3345 * We have to re-check the event->owner field, if it is cleared
3346 * we raced with perf_event_exit_task(), acquiring the mutex
3347 * ensured they're done, and we can proceed with freeing the
3351 list_del_init(&event->owner_entry);
3352 mutex_unlock(&owner->perf_event_mutex);
3353 put_task_struct(owner);
3356 WARN_ON_ONCE(ctx->parent_ctx);
3358 * There are two ways this annotation is useful:
3360 * 1) there is a lock recursion from perf_event_exit_task
3361 * see the comment there.
3363 * 2) there is a lock-inversion with mmap_sem through
3364 * perf_event_read_group(), which takes faults while
3365 * holding ctx->mutex, however this is called after
3366 * the last filedesc died, so there is no possibility
3367 * to trigger the AB-BA case.
3369 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3370 perf_remove_from_context(event, true);
3371 mutex_unlock(&ctx->mutex);
3376 int perf_event_release_kernel(struct perf_event *event)
3381 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3383 static int perf_release(struct inode *inode, struct file *file)
3385 put_event(file->private_data);
3389 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3391 struct perf_event *child;
3397 mutex_lock(&event->child_mutex);
3398 total += perf_event_read(event);
3399 *enabled += event->total_time_enabled +
3400 atomic64_read(&event->child_total_time_enabled);
3401 *running += event->total_time_running +
3402 atomic64_read(&event->child_total_time_running);
3404 list_for_each_entry(child, &event->child_list, child_list) {
3405 total += perf_event_read(child);
3406 *enabled += child->total_time_enabled;
3407 *running += child->total_time_running;
3409 mutex_unlock(&event->child_mutex);
3413 EXPORT_SYMBOL_GPL(perf_event_read_value);
3415 static int perf_event_read_group(struct perf_event *event,
3416 u64 read_format, char __user *buf)
3418 struct perf_event *leader = event->group_leader, *sub;
3419 int n = 0, size = 0, ret = -EFAULT;
3420 struct perf_event_context *ctx = leader->ctx;
3422 u64 count, enabled, running;
3424 mutex_lock(&ctx->mutex);
3425 count = perf_event_read_value(leader, &enabled, &running);
3427 values[n++] = 1 + leader->nr_siblings;
3428 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3429 values[n++] = enabled;
3430 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3431 values[n++] = running;
3432 values[n++] = count;
3433 if (read_format & PERF_FORMAT_ID)
3434 values[n++] = primary_event_id(leader);
3436 size = n * sizeof(u64);
3438 if (copy_to_user(buf, values, size))
3443 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3446 values[n++] = perf_event_read_value(sub, &enabled, &running);
3447 if (read_format & PERF_FORMAT_ID)
3448 values[n++] = primary_event_id(sub);
3450 size = n * sizeof(u64);
3452 if (copy_to_user(buf + ret, values, size)) {
3460 mutex_unlock(&ctx->mutex);
3465 static int perf_event_read_one(struct perf_event *event,
3466 u64 read_format, char __user *buf)
3468 u64 enabled, running;
3472 values[n++] = perf_event_read_value(event, &enabled, &running);
3473 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3474 values[n++] = enabled;
3475 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3476 values[n++] = running;
3477 if (read_format & PERF_FORMAT_ID)
3478 values[n++] = primary_event_id(event);
3480 if (copy_to_user(buf, values, n * sizeof(u64)))
3483 return n * sizeof(u64);
3487 * Read the performance event - simple non blocking version for now
3490 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3492 u64 read_format = event->attr.read_format;
3496 * Return end-of-file for a read on a event that is in
3497 * error state (i.e. because it was pinned but it couldn't be
3498 * scheduled on to the CPU at some point).
3500 if (event->state == PERF_EVENT_STATE_ERROR)
3503 if (count < event->read_size)
3506 WARN_ON_ONCE(event->ctx->parent_ctx);
3507 if (read_format & PERF_FORMAT_GROUP)
3508 ret = perf_event_read_group(event, read_format, buf);
3510 ret = perf_event_read_one(event, read_format, buf);
3516 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3518 struct perf_event *event = file->private_data;
3520 return perf_read_hw(event, buf, count);
3523 static unsigned int perf_poll(struct file *file, poll_table *wait)
3525 struct perf_event *event = file->private_data;
3526 struct ring_buffer *rb;
3527 unsigned int events = POLL_HUP;
3530 * Pin the event->rb by taking event->mmap_mutex; otherwise
3531 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3533 mutex_lock(&event->mmap_mutex);
3536 events = atomic_xchg(&rb->poll, 0);
3537 mutex_unlock(&event->mmap_mutex);
3539 poll_wait(file, &event->waitq, wait);
3544 static void perf_event_reset(struct perf_event *event)
3546 (void)perf_event_read(event);
3547 local64_set(&event->count, 0);
3548 perf_event_update_userpage(event);
3552 * Holding the top-level event's child_mutex means that any
3553 * descendant process that has inherited this event will block
3554 * in sync_child_event if it goes to exit, thus satisfying the
3555 * task existence requirements of perf_event_enable/disable.
3557 static void perf_event_for_each_child(struct perf_event *event,
3558 void (*func)(struct perf_event *))
3560 struct perf_event *child;
3562 WARN_ON_ONCE(event->ctx->parent_ctx);
3563 mutex_lock(&event->child_mutex);
3565 list_for_each_entry(child, &event->child_list, child_list)
3567 mutex_unlock(&event->child_mutex);
3570 static void perf_event_for_each(struct perf_event *event,
3571 void (*func)(struct perf_event *))
3573 struct perf_event_context *ctx = event->ctx;
3574 struct perf_event *sibling;
3576 WARN_ON_ONCE(ctx->parent_ctx);
3577 mutex_lock(&ctx->mutex);
3578 event = event->group_leader;
3580 perf_event_for_each_child(event, func);
3581 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3582 perf_event_for_each_child(sibling, func);
3583 mutex_unlock(&ctx->mutex);
3586 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3588 struct perf_event_context *ctx = event->ctx;
3589 int ret = 0, active;
3592 if (!is_sampling_event(event))
3595 if (copy_from_user(&value, arg, sizeof(value)))
3601 raw_spin_lock_irq(&ctx->lock);
3602 if (event->attr.freq) {
3603 if (value > sysctl_perf_event_sample_rate) {
3608 event->attr.sample_freq = value;
3610 event->attr.sample_period = value;
3611 event->hw.sample_period = value;
3614 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3616 perf_pmu_disable(ctx->pmu);
3617 event->pmu->stop(event, PERF_EF_UPDATE);
3620 local64_set(&event->hw.period_left, 0);
3623 event->pmu->start(event, PERF_EF_RELOAD);
3624 perf_pmu_enable(ctx->pmu);
3628 raw_spin_unlock_irq(&ctx->lock);
3633 static const struct file_operations perf_fops;
3635 static inline int perf_fget_light(int fd, struct fd *p)
3637 struct fd f = fdget(fd);
3641 if (f.file->f_op != &perf_fops) {
3649 static int perf_event_set_output(struct perf_event *event,
3650 struct perf_event *output_event);
3651 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3653 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3655 struct perf_event *event = file->private_data;
3656 void (*func)(struct perf_event *);
3660 case PERF_EVENT_IOC_ENABLE:
3661 func = perf_event_enable;
3663 case PERF_EVENT_IOC_DISABLE:
3664 func = perf_event_disable;
3666 case PERF_EVENT_IOC_RESET:
3667 func = perf_event_reset;
3670 case PERF_EVENT_IOC_REFRESH:
3671 return perf_event_refresh(event, arg);
3673 case PERF_EVENT_IOC_PERIOD:
3674 return perf_event_period(event, (u64 __user *)arg);
3676 case PERF_EVENT_IOC_ID:
3678 u64 id = primary_event_id(event);
3680 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3685 case PERF_EVENT_IOC_SET_OUTPUT:
3689 struct perf_event *output_event;
3691 ret = perf_fget_light(arg, &output);
3694 output_event = output.file->private_data;
3695 ret = perf_event_set_output(event, output_event);
3698 ret = perf_event_set_output(event, NULL);
3703 case PERF_EVENT_IOC_SET_FILTER:
3704 return perf_event_set_filter(event, (void __user *)arg);
3710 if (flags & PERF_IOC_FLAG_GROUP)
3711 perf_event_for_each(event, func);
3713 perf_event_for_each_child(event, func);
3718 int perf_event_task_enable(void)
3720 struct perf_event *event;
3722 mutex_lock(¤t->perf_event_mutex);
3723 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3724 perf_event_for_each_child(event, perf_event_enable);
3725 mutex_unlock(¤t->perf_event_mutex);
3730 int perf_event_task_disable(void)
3732 struct perf_event *event;
3734 mutex_lock(¤t->perf_event_mutex);
3735 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3736 perf_event_for_each_child(event, perf_event_disable);
3737 mutex_unlock(¤t->perf_event_mutex);
3742 static int perf_event_index(struct perf_event *event)
3744 if (event->hw.state & PERF_HES_STOPPED)
3747 if (event->state != PERF_EVENT_STATE_ACTIVE)
3750 return event->pmu->event_idx(event);
3753 static void calc_timer_values(struct perf_event *event,
3760 *now = perf_clock();
3761 ctx_time = event->shadow_ctx_time + *now;
3762 *enabled = ctx_time - event->tstamp_enabled;
3763 *running = ctx_time - event->tstamp_running;
3766 static void perf_event_init_userpage(struct perf_event *event)
3768 struct perf_event_mmap_page *userpg;
3769 struct ring_buffer *rb;
3772 rb = rcu_dereference(event->rb);
3776 userpg = rb->user_page;
3778 /* Allow new userspace to detect that bit 0 is deprecated */
3779 userpg->cap_bit0_is_deprecated = 1;
3780 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3786 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3791 * Callers need to ensure there can be no nesting of this function, otherwise
3792 * the seqlock logic goes bad. We can not serialize this because the arch
3793 * code calls this from NMI context.
3795 void perf_event_update_userpage(struct perf_event *event)
3797 struct perf_event_mmap_page *userpg;
3798 struct ring_buffer *rb;
3799 u64 enabled, running, now;
3802 rb = rcu_dereference(event->rb);
3807 * compute total_time_enabled, total_time_running
3808 * based on snapshot values taken when the event
3809 * was last scheduled in.
3811 * we cannot simply called update_context_time()
3812 * because of locking issue as we can be called in
3815 calc_timer_values(event, &now, &enabled, &running);
3817 userpg = rb->user_page;
3819 * Disable preemption so as to not let the corresponding user-space
3820 * spin too long if we get preempted.
3825 userpg->index = perf_event_index(event);
3826 userpg->offset = perf_event_count(event);
3828 userpg->offset -= local64_read(&event->hw.prev_count);
3830 userpg->time_enabled = enabled +
3831 atomic64_read(&event->child_total_time_enabled);
3833 userpg->time_running = running +
3834 atomic64_read(&event->child_total_time_running);
3836 arch_perf_update_userpage(userpg, now);
3845 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3847 struct perf_event *event = vma->vm_file->private_data;
3848 struct ring_buffer *rb;
3849 int ret = VM_FAULT_SIGBUS;
3851 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3852 if (vmf->pgoff == 0)
3858 rb = rcu_dereference(event->rb);
3862 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3865 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3869 get_page(vmf->page);
3870 vmf->page->mapping = vma->vm_file->f_mapping;
3871 vmf->page->index = vmf->pgoff;
3880 static void ring_buffer_attach(struct perf_event *event,
3881 struct ring_buffer *rb)
3883 struct ring_buffer *old_rb = NULL;
3884 unsigned long flags;
3888 * Should be impossible, we set this when removing
3889 * event->rb_entry and wait/clear when adding event->rb_entry.
3891 WARN_ON_ONCE(event->rcu_pending);
3894 event->rcu_batches = get_state_synchronize_rcu();
3895 event->rcu_pending = 1;
3897 spin_lock_irqsave(&old_rb->event_lock, flags);
3898 list_del_rcu(&event->rb_entry);
3899 spin_unlock_irqrestore(&old_rb->event_lock, flags);
3902 if (event->rcu_pending && rb) {
3903 cond_synchronize_rcu(event->rcu_batches);
3904 event->rcu_pending = 0;
3908 spin_lock_irqsave(&rb->event_lock, flags);
3909 list_add_rcu(&event->rb_entry, &rb->event_list);
3910 spin_unlock_irqrestore(&rb->event_lock, flags);
3913 rcu_assign_pointer(event->rb, rb);
3916 ring_buffer_put(old_rb);
3918 * Since we detached before setting the new rb, so that we
3919 * could attach the new rb, we could have missed a wakeup.
3922 wake_up_all(&event->waitq);
3926 static void ring_buffer_wakeup(struct perf_event *event)
3928 struct ring_buffer *rb;
3931 rb = rcu_dereference(event->rb);
3933 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3934 wake_up_all(&event->waitq);
3939 static void rb_free_rcu(struct rcu_head *rcu_head)
3941 struct ring_buffer *rb;
3943 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3947 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3949 struct ring_buffer *rb;
3952 rb = rcu_dereference(event->rb);
3954 if (!atomic_inc_not_zero(&rb->refcount))
3962 static void ring_buffer_put(struct ring_buffer *rb)
3964 if (!atomic_dec_and_test(&rb->refcount))
3967 WARN_ON_ONCE(!list_empty(&rb->event_list));
3969 call_rcu(&rb->rcu_head, rb_free_rcu);
3972 static void perf_mmap_open(struct vm_area_struct *vma)
3974 struct perf_event *event = vma->vm_file->private_data;
3976 atomic_inc(&event->mmap_count);
3977 atomic_inc(&event->rb->mmap_count);
3981 * A buffer can be mmap()ed multiple times; either directly through the same
3982 * event, or through other events by use of perf_event_set_output().
3984 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3985 * the buffer here, where we still have a VM context. This means we need
3986 * to detach all events redirecting to us.
3988 static void perf_mmap_close(struct vm_area_struct *vma)
3990 struct perf_event *event = vma->vm_file->private_data;
3992 struct ring_buffer *rb = ring_buffer_get(event);
3993 struct user_struct *mmap_user = rb->mmap_user;
3994 int mmap_locked = rb->mmap_locked;
3995 unsigned long size = perf_data_size(rb);
3997 atomic_dec(&rb->mmap_count);
3999 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4002 ring_buffer_attach(event, NULL);
4003 mutex_unlock(&event->mmap_mutex);
4005 /* If there's still other mmap()s of this buffer, we're done. */
4006 if (atomic_read(&rb->mmap_count))
4010 * No other mmap()s, detach from all other events that might redirect
4011 * into the now unreachable buffer. Somewhat complicated by the
4012 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4016 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4017 if (!atomic_long_inc_not_zero(&event->refcount)) {
4019 * This event is en-route to free_event() which will
4020 * detach it and remove it from the list.
4026 mutex_lock(&event->mmap_mutex);
4028 * Check we didn't race with perf_event_set_output() which can
4029 * swizzle the rb from under us while we were waiting to
4030 * acquire mmap_mutex.
4032 * If we find a different rb; ignore this event, a next
4033 * iteration will no longer find it on the list. We have to
4034 * still restart the iteration to make sure we're not now
4035 * iterating the wrong list.
4037 if (event->rb == rb)
4038 ring_buffer_attach(event, NULL);
4040 mutex_unlock(&event->mmap_mutex);
4044 * Restart the iteration; either we're on the wrong list or
4045 * destroyed its integrity by doing a deletion.
4052 * It could be there's still a few 0-ref events on the list; they'll
4053 * get cleaned up by free_event() -- they'll also still have their
4054 * ref on the rb and will free it whenever they are done with it.
4056 * Aside from that, this buffer is 'fully' detached and unmapped,
4057 * undo the VM accounting.
4060 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4061 vma->vm_mm->pinned_vm -= mmap_locked;
4062 free_uid(mmap_user);
4065 ring_buffer_put(rb); /* could be last */
4068 static const struct vm_operations_struct perf_mmap_vmops = {
4069 .open = perf_mmap_open,
4070 .close = perf_mmap_close,
4071 .fault = perf_mmap_fault,
4072 .page_mkwrite = perf_mmap_fault,
4075 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4077 struct perf_event *event = file->private_data;
4078 unsigned long user_locked, user_lock_limit;
4079 struct user_struct *user = current_user();
4080 unsigned long locked, lock_limit;
4081 struct ring_buffer *rb;
4082 unsigned long vma_size;
4083 unsigned long nr_pages;
4084 long user_extra, extra;
4085 int ret = 0, flags = 0;
4088 * Don't allow mmap() of inherited per-task counters. This would
4089 * create a performance issue due to all children writing to the
4092 if (event->cpu == -1 && event->attr.inherit)
4095 if (!(vma->vm_flags & VM_SHARED))
4098 vma_size = vma->vm_end - vma->vm_start;
4099 nr_pages = (vma_size / PAGE_SIZE) - 1;
4102 * If we have rb pages ensure they're a power-of-two number, so we
4103 * can do bitmasks instead of modulo.
4105 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4108 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4111 if (vma->vm_pgoff != 0)
4114 WARN_ON_ONCE(event->ctx->parent_ctx);
4116 mutex_lock(&event->mmap_mutex);
4118 if (event->rb->nr_pages != nr_pages) {
4123 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4125 * Raced against perf_mmap_close() through
4126 * perf_event_set_output(). Try again, hope for better
4129 mutex_unlock(&event->mmap_mutex);
4136 user_extra = nr_pages + 1;
4137 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4140 * Increase the limit linearly with more CPUs:
4142 user_lock_limit *= num_online_cpus();
4144 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4147 if (user_locked > user_lock_limit)
4148 extra = user_locked - user_lock_limit;
4150 lock_limit = rlimit(RLIMIT_MEMLOCK);
4151 lock_limit >>= PAGE_SHIFT;
4152 locked = vma->vm_mm->pinned_vm + extra;
4154 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4155 !capable(CAP_IPC_LOCK)) {
4162 if (vma->vm_flags & VM_WRITE)
4163 flags |= RING_BUFFER_WRITABLE;
4165 rb = rb_alloc(nr_pages,
4166 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4174 atomic_set(&rb->mmap_count, 1);
4175 rb->mmap_locked = extra;
4176 rb->mmap_user = get_current_user();
4178 atomic_long_add(user_extra, &user->locked_vm);
4179 vma->vm_mm->pinned_vm += extra;
4181 ring_buffer_attach(event, rb);
4183 perf_event_init_userpage(event);
4184 perf_event_update_userpage(event);
4188 atomic_inc(&event->mmap_count);
4189 mutex_unlock(&event->mmap_mutex);
4192 * Since pinned accounting is per vm we cannot allow fork() to copy our
4195 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4196 vma->vm_ops = &perf_mmap_vmops;
4201 static int perf_fasync(int fd, struct file *filp, int on)
4203 struct inode *inode = file_inode(filp);
4204 struct perf_event *event = filp->private_data;
4207 mutex_lock(&inode->i_mutex);
4208 retval = fasync_helper(fd, filp, on, &event->fasync);
4209 mutex_unlock(&inode->i_mutex);
4217 static const struct file_operations perf_fops = {
4218 .llseek = no_llseek,
4219 .release = perf_release,
4222 .unlocked_ioctl = perf_ioctl,
4223 .compat_ioctl = perf_ioctl,
4225 .fasync = perf_fasync,
4231 * If there's data, ensure we set the poll() state and publish everything
4232 * to user-space before waking everybody up.
4235 void perf_event_wakeup(struct perf_event *event)
4237 ring_buffer_wakeup(event);
4239 if (event->pending_kill) {
4240 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4241 event->pending_kill = 0;
4245 static void perf_pending_event(struct irq_work *entry)
4247 struct perf_event *event = container_of(entry,
4248 struct perf_event, pending);
4250 if (event->pending_disable) {
4251 event->pending_disable = 0;
4252 __perf_event_disable(event);
4255 if (event->pending_wakeup) {
4256 event->pending_wakeup = 0;
4257 perf_event_wakeup(event);
4262 * We assume there is only KVM supporting the callbacks.
4263 * Later on, we might change it to a list if there is
4264 * another virtualization implementation supporting the callbacks.
4266 struct perf_guest_info_callbacks *perf_guest_cbs;
4268 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4270 perf_guest_cbs = cbs;
4273 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4275 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4277 perf_guest_cbs = NULL;
4280 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4283 perf_output_sample_regs(struct perf_output_handle *handle,
4284 struct pt_regs *regs, u64 mask)
4288 for_each_set_bit(bit, (const unsigned long *) &mask,
4289 sizeof(mask) * BITS_PER_BYTE) {
4292 val = perf_reg_value(regs, bit);
4293 perf_output_put(handle, val);
4297 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4298 struct pt_regs *regs)
4300 if (!user_mode(regs)) {
4302 regs = task_pt_regs(current);
4308 regs_user->regs = regs;
4309 regs_user->abi = perf_reg_abi(current);
4314 * Get remaining task size from user stack pointer.
4316 * It'd be better to take stack vma map and limit this more
4317 * precisly, but there's no way to get it safely under interrupt,
4318 * so using TASK_SIZE as limit.
4320 static u64 perf_ustack_task_size(struct pt_regs *regs)
4322 unsigned long addr = perf_user_stack_pointer(regs);
4324 if (!addr || addr >= TASK_SIZE)
4327 return TASK_SIZE - addr;
4331 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4332 struct pt_regs *regs)
4336 /* No regs, no stack pointer, no dump. */
4341 * Check if we fit in with the requested stack size into the:
4343 * If we don't, we limit the size to the TASK_SIZE.
4345 * - remaining sample size
4346 * If we don't, we customize the stack size to
4347 * fit in to the remaining sample size.
4350 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4351 stack_size = min(stack_size, (u16) task_size);
4353 /* Current header size plus static size and dynamic size. */
4354 header_size += 2 * sizeof(u64);
4356 /* Do we fit in with the current stack dump size? */
4357 if ((u16) (header_size + stack_size) < header_size) {
4359 * If we overflow the maximum size for the sample,
4360 * we customize the stack dump size to fit in.
4362 stack_size = USHRT_MAX - header_size - sizeof(u64);
4363 stack_size = round_up(stack_size, sizeof(u64));
4370 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4371 struct pt_regs *regs)
4373 /* Case of a kernel thread, nothing to dump */
4376 perf_output_put(handle, size);
4385 * - the size requested by user or the best one we can fit
4386 * in to the sample max size
4388 * - user stack dump data
4390 * - the actual dumped size
4394 perf_output_put(handle, dump_size);
4397 sp = perf_user_stack_pointer(regs);
4398 rem = __output_copy_user(handle, (void *) sp, dump_size);
4399 dyn_size = dump_size - rem;
4401 perf_output_skip(handle, rem);
4404 perf_output_put(handle, dyn_size);
4408 static void __perf_event_header__init_id(struct perf_event_header *header,
4409 struct perf_sample_data *data,
4410 struct perf_event *event)
4412 u64 sample_type = event->attr.sample_type;
4414 data->type = sample_type;
4415 header->size += event->id_header_size;
4417 if (sample_type & PERF_SAMPLE_TID) {
4418 /* namespace issues */
4419 data->tid_entry.pid = perf_event_pid(event, current);
4420 data->tid_entry.tid = perf_event_tid(event, current);
4423 if (sample_type & PERF_SAMPLE_TIME)
4424 data->time = perf_clock();
4426 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4427 data->id = primary_event_id(event);
4429 if (sample_type & PERF_SAMPLE_STREAM_ID)
4430 data->stream_id = event->id;
4432 if (sample_type & PERF_SAMPLE_CPU) {
4433 data->cpu_entry.cpu = raw_smp_processor_id();
4434 data->cpu_entry.reserved = 0;
4438 void perf_event_header__init_id(struct perf_event_header *header,
4439 struct perf_sample_data *data,
4440 struct perf_event *event)
4442 if (event->attr.sample_id_all)
4443 __perf_event_header__init_id(header, data, event);
4446 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4447 struct perf_sample_data *data)
4449 u64 sample_type = data->type;
4451 if (sample_type & PERF_SAMPLE_TID)
4452 perf_output_put(handle, data->tid_entry);
4454 if (sample_type & PERF_SAMPLE_TIME)
4455 perf_output_put(handle, data->time);
4457 if (sample_type & PERF_SAMPLE_ID)
4458 perf_output_put(handle, data->id);
4460 if (sample_type & PERF_SAMPLE_STREAM_ID)
4461 perf_output_put(handle, data->stream_id);
4463 if (sample_type & PERF_SAMPLE_CPU)
4464 perf_output_put(handle, data->cpu_entry);
4466 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4467 perf_output_put(handle, data->id);
4470 void perf_event__output_id_sample(struct perf_event *event,
4471 struct perf_output_handle *handle,
4472 struct perf_sample_data *sample)
4474 if (event->attr.sample_id_all)
4475 __perf_event__output_id_sample(handle, sample);
4478 static void perf_output_read_one(struct perf_output_handle *handle,
4479 struct perf_event *event,
4480 u64 enabled, u64 running)
4482 u64 read_format = event->attr.read_format;
4486 values[n++] = perf_event_count(event);
4487 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4488 values[n++] = enabled +
4489 atomic64_read(&event->child_total_time_enabled);
4491 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4492 values[n++] = running +
4493 atomic64_read(&event->child_total_time_running);
4495 if (read_format & PERF_FORMAT_ID)
4496 values[n++] = primary_event_id(event);
4498 __output_copy(handle, values, n * sizeof(u64));
4502 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4504 static void perf_output_read_group(struct perf_output_handle *handle,
4505 struct perf_event *event,
4506 u64 enabled, u64 running)
4508 struct perf_event *leader = event->group_leader, *sub;
4509 u64 read_format = event->attr.read_format;
4513 values[n++] = 1 + leader->nr_siblings;
4515 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4516 values[n++] = enabled;
4518 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4519 values[n++] = running;
4521 if (leader != event)
4522 leader->pmu->read(leader);
4524 values[n++] = perf_event_count(leader);
4525 if (read_format & PERF_FORMAT_ID)
4526 values[n++] = primary_event_id(leader);
4528 __output_copy(handle, values, n * sizeof(u64));
4530 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4533 if ((sub != event) &&
4534 (sub->state == PERF_EVENT_STATE_ACTIVE))
4535 sub->pmu->read(sub);
4537 values[n++] = perf_event_count(sub);
4538 if (read_format & PERF_FORMAT_ID)
4539 values[n++] = primary_event_id(sub);
4541 __output_copy(handle, values, n * sizeof(u64));
4545 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4546 PERF_FORMAT_TOTAL_TIME_RUNNING)
4548 static void perf_output_read(struct perf_output_handle *handle,
4549 struct perf_event *event)
4551 u64 enabled = 0, running = 0, now;
4552 u64 read_format = event->attr.read_format;
4555 * compute total_time_enabled, total_time_running
4556 * based on snapshot values taken when the event
4557 * was last scheduled in.
4559 * we cannot simply called update_context_time()
4560 * because of locking issue as we are called in
4563 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4564 calc_timer_values(event, &now, &enabled, &running);
4566 if (event->attr.read_format & PERF_FORMAT_GROUP)
4567 perf_output_read_group(handle, event, enabled, running);
4569 perf_output_read_one(handle, event, enabled, running);
4572 void perf_output_sample(struct perf_output_handle *handle,
4573 struct perf_event_header *header,
4574 struct perf_sample_data *data,
4575 struct perf_event *event)
4577 u64 sample_type = data->type;
4579 perf_output_put(handle, *header);
4581 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4582 perf_output_put(handle, data->id);
4584 if (sample_type & PERF_SAMPLE_IP)
4585 perf_output_put(handle, data->ip);
4587 if (sample_type & PERF_SAMPLE_TID)
4588 perf_output_put(handle, data->tid_entry);
4590 if (sample_type & PERF_SAMPLE_TIME)
4591 perf_output_put(handle, data->time);
4593 if (sample_type & PERF_SAMPLE_ADDR)
4594 perf_output_put(handle, data->addr);
4596 if (sample_type & PERF_SAMPLE_ID)
4597 perf_output_put(handle, data->id);
4599 if (sample_type & PERF_SAMPLE_STREAM_ID)
4600 perf_output_put(handle, data->stream_id);
4602 if (sample_type & PERF_SAMPLE_CPU)
4603 perf_output_put(handle, data->cpu_entry);
4605 if (sample_type & PERF_SAMPLE_PERIOD)
4606 perf_output_put(handle, data->period);
4608 if (sample_type & PERF_SAMPLE_READ)
4609 perf_output_read(handle, event);
4611 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4612 if (data->callchain) {
4615 if (data->callchain)
4616 size += data->callchain->nr;
4618 size *= sizeof(u64);
4620 __output_copy(handle, data->callchain, size);
4623 perf_output_put(handle, nr);
4627 if (sample_type & PERF_SAMPLE_RAW) {
4629 perf_output_put(handle, data->raw->size);
4630 __output_copy(handle, data->raw->data,
4637 .size = sizeof(u32),
4640 perf_output_put(handle, raw);
4644 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4645 if (data->br_stack) {
4648 size = data->br_stack->nr
4649 * sizeof(struct perf_branch_entry);
4651 perf_output_put(handle, data->br_stack->nr);
4652 perf_output_copy(handle, data->br_stack->entries, size);
4655 * we always store at least the value of nr
4658 perf_output_put(handle, nr);
4662 if (sample_type & PERF_SAMPLE_REGS_USER) {
4663 u64 abi = data->regs_user.abi;
4666 * If there are no regs to dump, notice it through
4667 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4669 perf_output_put(handle, abi);
4672 u64 mask = event->attr.sample_regs_user;
4673 perf_output_sample_regs(handle,
4674 data->regs_user.regs,
4679 if (sample_type & PERF_SAMPLE_STACK_USER) {
4680 perf_output_sample_ustack(handle,
4681 data->stack_user_size,
4682 data->regs_user.regs);
4685 if (sample_type & PERF_SAMPLE_WEIGHT)
4686 perf_output_put(handle, data->weight);
4688 if (sample_type & PERF_SAMPLE_DATA_SRC)
4689 perf_output_put(handle, data->data_src.val);
4691 if (sample_type & PERF_SAMPLE_TRANSACTION)
4692 perf_output_put(handle, data->txn);
4694 if (!event->attr.watermark) {
4695 int wakeup_events = event->attr.wakeup_events;
4697 if (wakeup_events) {
4698 struct ring_buffer *rb = handle->rb;
4699 int events = local_inc_return(&rb->events);
4701 if (events >= wakeup_events) {
4702 local_sub(wakeup_events, &rb->events);
4703 local_inc(&rb->wakeup);
4709 void perf_prepare_sample(struct perf_event_header *header,
4710 struct perf_sample_data *data,
4711 struct perf_event *event,
4712 struct pt_regs *regs)
4714 u64 sample_type = event->attr.sample_type;
4716 header->type = PERF_RECORD_SAMPLE;
4717 header->size = sizeof(*header) + event->header_size;
4720 header->misc |= perf_misc_flags(regs);
4722 __perf_event_header__init_id(header, data, event);
4724 if (sample_type & PERF_SAMPLE_IP)
4725 data->ip = perf_instruction_pointer(regs);
4727 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4730 data->callchain = perf_callchain(event, regs);
4732 if (data->callchain)
4733 size += data->callchain->nr;
4735 header->size += size * sizeof(u64);
4738 if (sample_type & PERF_SAMPLE_RAW) {
4739 int size = sizeof(u32);
4742 size += data->raw->size;
4744 size += sizeof(u32);
4746 WARN_ON_ONCE(size & (sizeof(u64)-1));
4747 header->size += size;
4750 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4751 int size = sizeof(u64); /* nr */
4752 if (data->br_stack) {
4753 size += data->br_stack->nr
4754 * sizeof(struct perf_branch_entry);
4756 header->size += size;
4759 if (sample_type & PERF_SAMPLE_REGS_USER) {
4760 /* regs dump ABI info */
4761 int size = sizeof(u64);
4763 perf_sample_regs_user(&data->regs_user, regs);
4765 if (data->regs_user.regs) {
4766 u64 mask = event->attr.sample_regs_user;
4767 size += hweight64(mask) * sizeof(u64);
4770 header->size += size;
4773 if (sample_type & PERF_SAMPLE_STACK_USER) {
4775 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4776 * processed as the last one or have additional check added
4777 * in case new sample type is added, because we could eat
4778 * up the rest of the sample size.
4780 struct perf_regs_user *uregs = &data->regs_user;
4781 u16 stack_size = event->attr.sample_stack_user;
4782 u16 size = sizeof(u64);
4785 perf_sample_regs_user(uregs, regs);
4787 stack_size = perf_sample_ustack_size(stack_size, header->size,
4791 * If there is something to dump, add space for the dump
4792 * itself and for the field that tells the dynamic size,
4793 * which is how many have been actually dumped.
4796 size += sizeof(u64) + stack_size;
4798 data->stack_user_size = stack_size;
4799 header->size += size;
4803 static void perf_event_output(struct perf_event *event,
4804 struct perf_sample_data *data,
4805 struct pt_regs *regs)
4807 struct perf_output_handle handle;
4808 struct perf_event_header header;
4810 /* protect the callchain buffers */
4813 perf_prepare_sample(&header, data, event, regs);
4815 if (perf_output_begin(&handle, event, header.size))
4818 perf_output_sample(&handle, &header, data, event);
4820 perf_output_end(&handle);
4830 struct perf_read_event {
4831 struct perf_event_header header;
4838 perf_event_read_event(struct perf_event *event,
4839 struct task_struct *task)
4841 struct perf_output_handle handle;
4842 struct perf_sample_data sample;
4843 struct perf_read_event read_event = {
4845 .type = PERF_RECORD_READ,
4847 .size = sizeof(read_event) + event->read_size,
4849 .pid = perf_event_pid(event, task),
4850 .tid = perf_event_tid(event, task),
4854 perf_event_header__init_id(&read_event.header, &sample, event);
4855 ret = perf_output_begin(&handle, event, read_event.header.size);
4859 perf_output_put(&handle, read_event);
4860 perf_output_read(&handle, event);
4861 perf_event__output_id_sample(event, &handle, &sample);
4863 perf_output_end(&handle);
4866 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4869 perf_event_aux_ctx(struct perf_event_context *ctx,
4870 perf_event_aux_output_cb output,
4873 struct perf_event *event;
4875 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4876 if (event->state < PERF_EVENT_STATE_INACTIVE)
4878 if (!event_filter_match(event))
4880 output(event, data);
4885 perf_event_aux(perf_event_aux_output_cb output, void *data,
4886 struct perf_event_context *task_ctx)
4888 struct perf_cpu_context *cpuctx;
4889 struct perf_event_context *ctx;
4894 list_for_each_entry_rcu(pmu, &pmus, entry) {
4895 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4896 if (cpuctx->unique_pmu != pmu)
4898 perf_event_aux_ctx(&cpuctx->ctx, output, data);
4901 ctxn = pmu->task_ctx_nr;
4904 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4906 perf_event_aux_ctx(ctx, output, data);
4908 put_cpu_ptr(pmu->pmu_cpu_context);
4913 perf_event_aux_ctx(task_ctx, output, data);
4920 * task tracking -- fork/exit
4922 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4925 struct perf_task_event {
4926 struct task_struct *task;
4927 struct perf_event_context *task_ctx;
4930 struct perf_event_header header;
4940 static int perf_event_task_match(struct perf_event *event)
4942 return event->attr.comm || event->attr.mmap ||
4943 event->attr.mmap2 || event->attr.mmap_data ||
4947 static void perf_event_task_output(struct perf_event *event,
4950 struct perf_task_event *task_event = data;
4951 struct perf_output_handle handle;
4952 struct perf_sample_data sample;
4953 struct task_struct *task = task_event->task;
4954 int ret, size = task_event->event_id.header.size;
4956 if (!perf_event_task_match(event))
4959 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4961 ret = perf_output_begin(&handle, event,
4962 task_event->event_id.header.size);
4966 task_event->event_id.pid = perf_event_pid(event, task);
4967 task_event->event_id.ppid = perf_event_pid(event, current);
4969 task_event->event_id.tid = perf_event_tid(event, task);
4970 task_event->event_id.ptid = perf_event_tid(event, current);
4972 perf_output_put(&handle, task_event->event_id);
4974 perf_event__output_id_sample(event, &handle, &sample);
4976 perf_output_end(&handle);
4978 task_event->event_id.header.size = size;
4981 static void perf_event_task(struct task_struct *task,
4982 struct perf_event_context *task_ctx,
4985 struct perf_task_event task_event;
4987 if (!atomic_read(&nr_comm_events) &&
4988 !atomic_read(&nr_mmap_events) &&
4989 !atomic_read(&nr_task_events))
4992 task_event = (struct perf_task_event){
4994 .task_ctx = task_ctx,
4997 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4999 .size = sizeof(task_event.event_id),
5005 .time = perf_clock(),
5009 perf_event_aux(perf_event_task_output,
5014 void perf_event_fork(struct task_struct *task)
5016 perf_event_task(task, NULL, 1);
5023 struct perf_comm_event {
5024 struct task_struct *task;
5029 struct perf_event_header header;
5036 static int perf_event_comm_match(struct perf_event *event)
5038 return event->attr.comm;
5041 static void perf_event_comm_output(struct perf_event *event,
5044 struct perf_comm_event *comm_event = data;
5045 struct perf_output_handle handle;
5046 struct perf_sample_data sample;
5047 int size = comm_event->event_id.header.size;
5050 if (!perf_event_comm_match(event))
5053 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5054 ret = perf_output_begin(&handle, event,
5055 comm_event->event_id.header.size);
5060 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5061 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5063 perf_output_put(&handle, comm_event->event_id);
5064 __output_copy(&handle, comm_event->comm,
5065 comm_event->comm_size);
5067 perf_event__output_id_sample(event, &handle, &sample);
5069 perf_output_end(&handle);
5071 comm_event->event_id.header.size = size;
5074 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5076 char comm[TASK_COMM_LEN];
5079 memset(comm, 0, sizeof(comm));
5080 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5081 size = ALIGN(strlen(comm)+1, sizeof(u64));
5083 comm_event->comm = comm;
5084 comm_event->comm_size = size;
5086 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5088 perf_event_aux(perf_event_comm_output,
5093 void perf_event_comm(struct task_struct *task, bool exec)
5095 struct perf_comm_event comm_event;
5097 if (!atomic_read(&nr_comm_events))
5100 comm_event = (struct perf_comm_event){
5106 .type = PERF_RECORD_COMM,
5107 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5115 perf_event_comm_event(&comm_event);
5122 struct perf_mmap_event {
5123 struct vm_area_struct *vma;
5125 const char *file_name;
5132 struct perf_event_header header;
5142 static int perf_event_mmap_match(struct perf_event *event,
5145 struct perf_mmap_event *mmap_event = data;
5146 struct vm_area_struct *vma = mmap_event->vma;
5147 int executable = vma->vm_flags & VM_EXEC;
5149 return (!executable && event->attr.mmap_data) ||
5150 (executable && (event->attr.mmap || event->attr.mmap2));
5153 static void perf_event_mmap_output(struct perf_event *event,
5156 struct perf_mmap_event *mmap_event = data;
5157 struct perf_output_handle handle;
5158 struct perf_sample_data sample;
5159 int size = mmap_event->event_id.header.size;
5162 if (!perf_event_mmap_match(event, data))
5165 if (event->attr.mmap2) {
5166 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5167 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5168 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5169 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5170 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5173 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5174 ret = perf_output_begin(&handle, event,
5175 mmap_event->event_id.header.size);
5179 mmap_event->event_id.pid = perf_event_pid(event, current);
5180 mmap_event->event_id.tid = perf_event_tid(event, current);
5182 perf_output_put(&handle, mmap_event->event_id);
5184 if (event->attr.mmap2) {
5185 perf_output_put(&handle, mmap_event->maj);
5186 perf_output_put(&handle, mmap_event->min);
5187 perf_output_put(&handle, mmap_event->ino);
5188 perf_output_put(&handle, mmap_event->ino_generation);
5191 __output_copy(&handle, mmap_event->file_name,
5192 mmap_event->file_size);
5194 perf_event__output_id_sample(event, &handle, &sample);
5196 perf_output_end(&handle);
5198 mmap_event->event_id.header.size = size;
5201 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5203 struct vm_area_struct *vma = mmap_event->vma;
5204 struct file *file = vma->vm_file;
5205 int maj = 0, min = 0;
5206 u64 ino = 0, gen = 0;
5213 struct inode *inode;
5216 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5222 * d_path() works from the end of the rb backwards, so we
5223 * need to add enough zero bytes after the string to handle
5224 * the 64bit alignment we do later.
5226 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5231 inode = file_inode(vma->vm_file);
5232 dev = inode->i_sb->s_dev;
5234 gen = inode->i_generation;
5239 name = (char *)arch_vma_name(vma);
5243 if (vma->vm_start <= vma->vm_mm->start_brk &&
5244 vma->vm_end >= vma->vm_mm->brk) {
5248 if (vma->vm_start <= vma->vm_mm->start_stack &&
5249 vma->vm_end >= vma->vm_mm->start_stack) {
5259 strlcpy(tmp, name, sizeof(tmp));
5263 * Since our buffer works in 8 byte units we need to align our string
5264 * size to a multiple of 8. However, we must guarantee the tail end is
5265 * zero'd out to avoid leaking random bits to userspace.
5267 size = strlen(name)+1;
5268 while (!IS_ALIGNED(size, sizeof(u64)))
5269 name[size++] = '\0';
5271 mmap_event->file_name = name;
5272 mmap_event->file_size = size;
5273 mmap_event->maj = maj;
5274 mmap_event->min = min;
5275 mmap_event->ino = ino;
5276 mmap_event->ino_generation = gen;
5278 if (!(vma->vm_flags & VM_EXEC))
5279 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5281 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5283 perf_event_aux(perf_event_mmap_output,
5290 void perf_event_mmap(struct vm_area_struct *vma)
5292 struct perf_mmap_event mmap_event;
5294 if (!atomic_read(&nr_mmap_events))
5297 mmap_event = (struct perf_mmap_event){
5303 .type = PERF_RECORD_MMAP,
5304 .misc = PERF_RECORD_MISC_USER,
5309 .start = vma->vm_start,
5310 .len = vma->vm_end - vma->vm_start,
5311 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5313 /* .maj (attr_mmap2 only) */
5314 /* .min (attr_mmap2 only) */
5315 /* .ino (attr_mmap2 only) */
5316 /* .ino_generation (attr_mmap2 only) */
5319 perf_event_mmap_event(&mmap_event);
5323 * IRQ throttle logging
5326 static void perf_log_throttle(struct perf_event *event, int enable)
5328 struct perf_output_handle handle;
5329 struct perf_sample_data sample;
5333 struct perf_event_header header;
5337 } throttle_event = {
5339 .type = PERF_RECORD_THROTTLE,
5341 .size = sizeof(throttle_event),
5343 .time = perf_clock(),
5344 .id = primary_event_id(event),
5345 .stream_id = event->id,
5349 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5351 perf_event_header__init_id(&throttle_event.header, &sample, event);
5353 ret = perf_output_begin(&handle, event,
5354 throttle_event.header.size);
5358 perf_output_put(&handle, throttle_event);
5359 perf_event__output_id_sample(event, &handle, &sample);
5360 perf_output_end(&handle);
5364 * Generic event overflow handling, sampling.
5367 static int __perf_event_overflow(struct perf_event *event,
5368 int throttle, struct perf_sample_data *data,
5369 struct pt_regs *regs)
5371 int events = atomic_read(&event->event_limit);
5372 struct hw_perf_event *hwc = &event->hw;
5377 * Non-sampling counters might still use the PMI to fold short
5378 * hardware counters, ignore those.
5380 if (unlikely(!is_sampling_event(event)))
5383 seq = __this_cpu_read(perf_throttled_seq);
5384 if (seq != hwc->interrupts_seq) {
5385 hwc->interrupts_seq = seq;
5386 hwc->interrupts = 1;
5389 if (unlikely(throttle
5390 && hwc->interrupts >= max_samples_per_tick)) {
5391 __this_cpu_inc(perf_throttled_count);
5392 hwc->interrupts = MAX_INTERRUPTS;
5393 perf_log_throttle(event, 0);
5394 tick_nohz_full_kick();
5399 if (event->attr.freq) {
5400 u64 now = perf_clock();
5401 s64 delta = now - hwc->freq_time_stamp;
5403 hwc->freq_time_stamp = now;
5405 if (delta > 0 && delta < 2*TICK_NSEC)
5406 perf_adjust_period(event, delta, hwc->last_period, true);
5410 * XXX event_limit might not quite work as expected on inherited
5414 event->pending_kill = POLL_IN;
5415 if (events && atomic_dec_and_test(&event->event_limit)) {
5417 event->pending_kill = POLL_HUP;
5418 event->pending_disable = 1;
5419 irq_work_queue(&event->pending);
5422 if (event->overflow_handler)
5423 event->overflow_handler(event, data, regs);
5425 perf_event_output(event, data, regs);
5427 if (event->fasync && event->pending_kill) {
5428 event->pending_wakeup = 1;
5429 irq_work_queue(&event->pending);
5435 int perf_event_overflow(struct perf_event *event,
5436 struct perf_sample_data *data,
5437 struct pt_regs *regs)
5439 return __perf_event_overflow(event, 1, data, regs);
5443 * Generic software event infrastructure
5446 struct swevent_htable {
5447 struct swevent_hlist *swevent_hlist;
5448 struct mutex hlist_mutex;
5451 /* Recursion avoidance in each contexts */
5452 int recursion[PERF_NR_CONTEXTS];
5454 /* Keeps track of cpu being initialized/exited */
5458 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5461 * We directly increment event->count and keep a second value in
5462 * event->hw.period_left to count intervals. This period event
5463 * is kept in the range [-sample_period, 0] so that we can use the
5467 u64 perf_swevent_set_period(struct perf_event *event)
5469 struct hw_perf_event *hwc = &event->hw;
5470 u64 period = hwc->last_period;
5474 hwc->last_period = hwc->sample_period;
5477 old = val = local64_read(&hwc->period_left);
5481 nr = div64_u64(period + val, period);
5482 offset = nr * period;
5484 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5490 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5491 struct perf_sample_data *data,
5492 struct pt_regs *regs)
5494 struct hw_perf_event *hwc = &event->hw;
5498 overflow = perf_swevent_set_period(event);
5500 if (hwc->interrupts == MAX_INTERRUPTS)
5503 for (; overflow; overflow--) {
5504 if (__perf_event_overflow(event, throttle,
5507 * We inhibit the overflow from happening when
5508 * hwc->interrupts == MAX_INTERRUPTS.
5516 static void perf_swevent_event(struct perf_event *event, u64 nr,
5517 struct perf_sample_data *data,
5518 struct pt_regs *regs)
5520 struct hw_perf_event *hwc = &event->hw;
5522 local64_add(nr, &event->count);
5527 if (!is_sampling_event(event))
5530 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5532 return perf_swevent_overflow(event, 1, data, regs);
5534 data->period = event->hw.last_period;
5536 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5537 return perf_swevent_overflow(event, 1, data, regs);
5539 if (local64_add_negative(nr, &hwc->period_left))
5542 perf_swevent_overflow(event, 0, data, regs);
5545 static int perf_exclude_event(struct perf_event *event,
5546 struct pt_regs *regs)
5548 if (event->hw.state & PERF_HES_STOPPED)
5552 if (event->attr.exclude_user && user_mode(regs))
5555 if (event->attr.exclude_kernel && !user_mode(regs))
5562 static int perf_swevent_match(struct perf_event *event,
5563 enum perf_type_id type,
5565 struct perf_sample_data *data,
5566 struct pt_regs *regs)
5568 if (event->attr.type != type)
5571 if (event->attr.config != event_id)
5574 if (perf_exclude_event(event, regs))
5580 static inline u64 swevent_hash(u64 type, u32 event_id)
5582 u64 val = event_id | (type << 32);
5584 return hash_64(val, SWEVENT_HLIST_BITS);
5587 static inline struct hlist_head *
5588 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5590 u64 hash = swevent_hash(type, event_id);
5592 return &hlist->heads[hash];
5595 /* For the read side: events when they trigger */
5596 static inline struct hlist_head *
5597 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5599 struct swevent_hlist *hlist;
5601 hlist = rcu_dereference(swhash->swevent_hlist);
5605 return __find_swevent_head(hlist, type, event_id);
5608 /* For the event head insertion and removal in the hlist */
5609 static inline struct hlist_head *
5610 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5612 struct swevent_hlist *hlist;
5613 u32 event_id = event->attr.config;
5614 u64 type = event->attr.type;
5617 * Event scheduling is always serialized against hlist allocation
5618 * and release. Which makes the protected version suitable here.
5619 * The context lock guarantees that.
5621 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5622 lockdep_is_held(&event->ctx->lock));
5626 return __find_swevent_head(hlist, type, event_id);
5629 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5631 struct perf_sample_data *data,
5632 struct pt_regs *regs)
5634 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5635 struct perf_event *event;
5636 struct hlist_head *head;
5639 head = find_swevent_head_rcu(swhash, type, event_id);
5643 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5644 if (perf_swevent_match(event, type, event_id, data, regs))
5645 perf_swevent_event(event, nr, data, regs);
5651 int perf_swevent_get_recursion_context(void)
5653 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5655 return get_recursion_context(swhash->recursion);
5657 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5659 inline void perf_swevent_put_recursion_context(int rctx)
5661 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5663 put_recursion_context(swhash->recursion, rctx);
5666 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5668 struct perf_sample_data data;
5671 preempt_disable_notrace();
5672 rctx = perf_swevent_get_recursion_context();
5676 perf_sample_data_init(&data, addr, 0);
5678 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5680 perf_swevent_put_recursion_context(rctx);
5681 preempt_enable_notrace();
5684 static void perf_swevent_read(struct perf_event *event)
5688 static int perf_swevent_add(struct perf_event *event, int flags)
5690 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5691 struct hw_perf_event *hwc = &event->hw;
5692 struct hlist_head *head;
5694 if (is_sampling_event(event)) {
5695 hwc->last_period = hwc->sample_period;
5696 perf_swevent_set_period(event);
5699 hwc->state = !(flags & PERF_EF_START);
5701 head = find_swevent_head(swhash, event);
5704 * We can race with cpu hotplug code. Do not
5705 * WARN if the cpu just got unplugged.
5707 WARN_ON_ONCE(swhash->online);
5711 hlist_add_head_rcu(&event->hlist_entry, head);
5716 static void perf_swevent_del(struct perf_event *event, int flags)
5718 hlist_del_rcu(&event->hlist_entry);
5721 static void perf_swevent_start(struct perf_event *event, int flags)
5723 event->hw.state = 0;
5726 static void perf_swevent_stop(struct perf_event *event, int flags)
5728 event->hw.state = PERF_HES_STOPPED;
5731 /* Deref the hlist from the update side */
5732 static inline struct swevent_hlist *
5733 swevent_hlist_deref(struct swevent_htable *swhash)
5735 return rcu_dereference_protected(swhash->swevent_hlist,
5736 lockdep_is_held(&swhash->hlist_mutex));
5739 static void swevent_hlist_release(struct swevent_htable *swhash)
5741 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5746 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5747 kfree_rcu(hlist, rcu_head);
5750 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5752 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5754 mutex_lock(&swhash->hlist_mutex);
5756 if (!--swhash->hlist_refcount)
5757 swevent_hlist_release(swhash);
5759 mutex_unlock(&swhash->hlist_mutex);
5762 static void swevent_hlist_put(struct perf_event *event)
5766 for_each_possible_cpu(cpu)
5767 swevent_hlist_put_cpu(event, cpu);
5770 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5772 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5775 mutex_lock(&swhash->hlist_mutex);
5777 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5778 struct swevent_hlist *hlist;
5780 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5785 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5787 swhash->hlist_refcount++;
5789 mutex_unlock(&swhash->hlist_mutex);
5794 static int swevent_hlist_get(struct perf_event *event)
5797 int cpu, failed_cpu;
5800 for_each_possible_cpu(cpu) {
5801 err = swevent_hlist_get_cpu(event, cpu);
5811 for_each_possible_cpu(cpu) {
5812 if (cpu == failed_cpu)
5814 swevent_hlist_put_cpu(event, cpu);
5821 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5823 static void sw_perf_event_destroy(struct perf_event *event)
5825 u64 event_id = event->attr.config;
5827 WARN_ON(event->parent);
5829 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5830 swevent_hlist_put(event);
5833 static int perf_swevent_init(struct perf_event *event)
5835 u64 event_id = event->attr.config;
5837 if (event->attr.type != PERF_TYPE_SOFTWARE)
5841 * no branch sampling for software events
5843 if (has_branch_stack(event))
5847 case PERF_COUNT_SW_CPU_CLOCK:
5848 case PERF_COUNT_SW_TASK_CLOCK:
5855 if (event_id >= PERF_COUNT_SW_MAX)
5858 if (!event->parent) {
5861 err = swevent_hlist_get(event);
5865 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5866 event->destroy = sw_perf_event_destroy;
5872 static int perf_swevent_event_idx(struct perf_event *event)
5877 static struct pmu perf_swevent = {
5878 .task_ctx_nr = perf_sw_context,
5880 .event_init = perf_swevent_init,
5881 .add = perf_swevent_add,
5882 .del = perf_swevent_del,
5883 .start = perf_swevent_start,
5884 .stop = perf_swevent_stop,
5885 .read = perf_swevent_read,
5887 .event_idx = perf_swevent_event_idx,
5890 #ifdef CONFIG_EVENT_TRACING
5892 static int perf_tp_filter_match(struct perf_event *event,
5893 struct perf_sample_data *data)
5895 void *record = data->raw->data;
5897 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5902 static int perf_tp_event_match(struct perf_event *event,
5903 struct perf_sample_data *data,
5904 struct pt_regs *regs)
5906 if (event->hw.state & PERF_HES_STOPPED)
5909 * All tracepoints are from kernel-space.
5911 if (event->attr.exclude_kernel)
5914 if (!perf_tp_filter_match(event, data))
5920 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5921 struct pt_regs *regs, struct hlist_head *head, int rctx,
5922 struct task_struct *task)
5924 struct perf_sample_data data;
5925 struct perf_event *event;
5927 struct perf_raw_record raw = {
5932 perf_sample_data_init(&data, addr, 0);
5935 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5936 if (perf_tp_event_match(event, &data, regs))
5937 perf_swevent_event(event, count, &data, regs);
5941 * If we got specified a target task, also iterate its context and
5942 * deliver this event there too.
5944 if (task && task != current) {
5945 struct perf_event_context *ctx;
5946 struct trace_entry *entry = record;
5949 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5953 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5954 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5956 if (event->attr.config != entry->type)
5958 if (perf_tp_event_match(event, &data, regs))
5959 perf_swevent_event(event, count, &data, regs);
5965 perf_swevent_put_recursion_context(rctx);
5967 EXPORT_SYMBOL_GPL(perf_tp_event);
5969 static void tp_perf_event_destroy(struct perf_event *event)
5971 perf_trace_destroy(event);
5974 static int perf_tp_event_init(struct perf_event *event)
5978 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5982 * no branch sampling for tracepoint events
5984 if (has_branch_stack(event))
5987 err = perf_trace_init(event);
5991 event->destroy = tp_perf_event_destroy;
5996 static struct pmu perf_tracepoint = {
5997 .task_ctx_nr = perf_sw_context,
5999 .event_init = perf_tp_event_init,
6000 .add = perf_trace_add,
6001 .del = perf_trace_del,
6002 .start = perf_swevent_start,
6003 .stop = perf_swevent_stop,
6004 .read = perf_swevent_read,
6006 .event_idx = perf_swevent_event_idx,
6009 static inline void perf_tp_register(void)
6011 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6014 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6019 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6022 filter_str = strndup_user(arg, PAGE_SIZE);
6023 if (IS_ERR(filter_str))
6024 return PTR_ERR(filter_str);
6026 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6032 static void perf_event_free_filter(struct perf_event *event)
6034 ftrace_profile_free_filter(event);
6039 static inline void perf_tp_register(void)
6043 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6048 static void perf_event_free_filter(struct perf_event *event)
6052 #endif /* CONFIG_EVENT_TRACING */
6054 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6055 void perf_bp_event(struct perf_event *bp, void *data)
6057 struct perf_sample_data sample;
6058 struct pt_regs *regs = data;
6060 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6062 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6063 perf_swevent_event(bp, 1, &sample, regs);
6068 * hrtimer based swevent callback
6071 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6073 enum hrtimer_restart ret = HRTIMER_RESTART;
6074 struct perf_sample_data data;
6075 struct pt_regs *regs;
6076 struct perf_event *event;
6079 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6081 if (event->state != PERF_EVENT_STATE_ACTIVE)
6082 return HRTIMER_NORESTART;
6084 event->pmu->read(event);
6086 perf_sample_data_init(&data, 0, event->hw.last_period);
6087 regs = get_irq_regs();
6089 if (regs && !perf_exclude_event(event, regs)) {
6090 if (!(event->attr.exclude_idle && is_idle_task(current)))
6091 if (__perf_event_overflow(event, 1, &data, regs))
6092 ret = HRTIMER_NORESTART;
6095 period = max_t(u64, 10000, event->hw.sample_period);
6096 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6101 static void perf_swevent_start_hrtimer(struct perf_event *event)
6103 struct hw_perf_event *hwc = &event->hw;
6106 if (!is_sampling_event(event))
6109 period = local64_read(&hwc->period_left);
6114 local64_set(&hwc->period_left, 0);
6116 period = max_t(u64, 10000, hwc->sample_period);
6118 __hrtimer_start_range_ns(&hwc->hrtimer,
6119 ns_to_ktime(period), 0,
6120 HRTIMER_MODE_REL_PINNED, 0);
6123 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6125 struct hw_perf_event *hwc = &event->hw;
6127 if (is_sampling_event(event)) {
6128 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6129 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6131 hrtimer_cancel(&hwc->hrtimer);
6135 static void perf_swevent_init_hrtimer(struct perf_event *event)
6137 struct hw_perf_event *hwc = &event->hw;
6139 if (!is_sampling_event(event))
6142 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6143 hwc->hrtimer.function = perf_swevent_hrtimer;
6146 * Since hrtimers have a fixed rate, we can do a static freq->period
6147 * mapping and avoid the whole period adjust feedback stuff.
6149 if (event->attr.freq) {
6150 long freq = event->attr.sample_freq;
6152 event->attr.sample_period = NSEC_PER_SEC / freq;
6153 hwc->sample_period = event->attr.sample_period;
6154 local64_set(&hwc->period_left, hwc->sample_period);
6155 hwc->last_period = hwc->sample_period;
6156 event->attr.freq = 0;
6161 * Software event: cpu wall time clock
6164 static void cpu_clock_event_update(struct perf_event *event)
6169 now = local_clock();
6170 prev = local64_xchg(&event->hw.prev_count, now);
6171 local64_add(now - prev, &event->count);
6174 static void cpu_clock_event_start(struct perf_event *event, int flags)
6176 local64_set(&event->hw.prev_count, local_clock());
6177 perf_swevent_start_hrtimer(event);
6180 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6182 perf_swevent_cancel_hrtimer(event);
6183 cpu_clock_event_update(event);
6186 static int cpu_clock_event_add(struct perf_event *event, int flags)
6188 if (flags & PERF_EF_START)
6189 cpu_clock_event_start(event, flags);
6194 static void cpu_clock_event_del(struct perf_event *event, int flags)
6196 cpu_clock_event_stop(event, flags);
6199 static void cpu_clock_event_read(struct perf_event *event)
6201 cpu_clock_event_update(event);
6204 static int cpu_clock_event_init(struct perf_event *event)
6206 if (event->attr.type != PERF_TYPE_SOFTWARE)
6209 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6213 * no branch sampling for software events
6215 if (has_branch_stack(event))
6218 perf_swevent_init_hrtimer(event);
6223 static struct pmu perf_cpu_clock = {
6224 .task_ctx_nr = perf_sw_context,
6226 .event_init = cpu_clock_event_init,
6227 .add = cpu_clock_event_add,
6228 .del = cpu_clock_event_del,
6229 .start = cpu_clock_event_start,
6230 .stop = cpu_clock_event_stop,
6231 .read = cpu_clock_event_read,
6233 .event_idx = perf_swevent_event_idx,
6237 * Software event: task time clock
6240 static void task_clock_event_update(struct perf_event *event, u64 now)
6245 prev = local64_xchg(&event->hw.prev_count, now);
6247 local64_add(delta, &event->count);
6250 static void task_clock_event_start(struct perf_event *event, int flags)
6252 local64_set(&event->hw.prev_count, event->ctx->time);
6253 perf_swevent_start_hrtimer(event);
6256 static void task_clock_event_stop(struct perf_event *event, int flags)
6258 perf_swevent_cancel_hrtimer(event);
6259 task_clock_event_update(event, event->ctx->time);
6262 static int task_clock_event_add(struct perf_event *event, int flags)
6264 if (flags & PERF_EF_START)
6265 task_clock_event_start(event, flags);
6270 static void task_clock_event_del(struct perf_event *event, int flags)
6272 task_clock_event_stop(event, PERF_EF_UPDATE);
6275 static void task_clock_event_read(struct perf_event *event)
6277 u64 now = perf_clock();
6278 u64 delta = now - event->ctx->timestamp;
6279 u64 time = event->ctx->time + delta;
6281 task_clock_event_update(event, time);
6284 static int task_clock_event_init(struct perf_event *event)
6286 if (event->attr.type != PERF_TYPE_SOFTWARE)
6289 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6293 * no branch sampling for software events
6295 if (has_branch_stack(event))
6298 perf_swevent_init_hrtimer(event);
6303 static struct pmu perf_task_clock = {
6304 .task_ctx_nr = perf_sw_context,
6306 .event_init = task_clock_event_init,
6307 .add = task_clock_event_add,
6308 .del = task_clock_event_del,
6309 .start = task_clock_event_start,
6310 .stop = task_clock_event_stop,
6311 .read = task_clock_event_read,
6313 .event_idx = perf_swevent_event_idx,
6316 static void perf_pmu_nop_void(struct pmu *pmu)
6320 static int perf_pmu_nop_int(struct pmu *pmu)
6325 static void perf_pmu_start_txn(struct pmu *pmu)
6327 perf_pmu_disable(pmu);
6330 static int perf_pmu_commit_txn(struct pmu *pmu)
6332 perf_pmu_enable(pmu);
6336 static void perf_pmu_cancel_txn(struct pmu *pmu)
6338 perf_pmu_enable(pmu);
6341 static int perf_event_idx_default(struct perf_event *event)
6343 return event->hw.idx + 1;
6347 * Ensures all contexts with the same task_ctx_nr have the same
6348 * pmu_cpu_context too.
6350 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6357 list_for_each_entry(pmu, &pmus, entry) {
6358 if (pmu->task_ctx_nr == ctxn)
6359 return pmu->pmu_cpu_context;
6365 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6369 for_each_possible_cpu(cpu) {
6370 struct perf_cpu_context *cpuctx;
6372 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6374 if (cpuctx->unique_pmu == old_pmu)
6375 cpuctx->unique_pmu = pmu;
6379 static void free_pmu_context(struct pmu *pmu)
6383 mutex_lock(&pmus_lock);
6385 * Like a real lame refcount.
6387 list_for_each_entry(i, &pmus, entry) {
6388 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6389 update_pmu_context(i, pmu);
6394 free_percpu(pmu->pmu_cpu_context);
6396 mutex_unlock(&pmus_lock);
6398 static struct idr pmu_idr;
6401 type_show(struct device *dev, struct device_attribute *attr, char *page)
6403 struct pmu *pmu = dev_get_drvdata(dev);
6405 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6407 static DEVICE_ATTR_RO(type);
6410 perf_event_mux_interval_ms_show(struct device *dev,
6411 struct device_attribute *attr,
6414 struct pmu *pmu = dev_get_drvdata(dev);
6416 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6420 perf_event_mux_interval_ms_store(struct device *dev,
6421 struct device_attribute *attr,
6422 const char *buf, size_t count)
6424 struct pmu *pmu = dev_get_drvdata(dev);
6425 int timer, cpu, ret;
6427 ret = kstrtoint(buf, 0, &timer);
6434 /* same value, noting to do */
6435 if (timer == pmu->hrtimer_interval_ms)
6438 pmu->hrtimer_interval_ms = timer;
6440 /* update all cpuctx for this PMU */
6441 for_each_possible_cpu(cpu) {
6442 struct perf_cpu_context *cpuctx;
6443 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6444 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6446 if (hrtimer_active(&cpuctx->hrtimer))
6447 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6452 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6454 static struct attribute *pmu_dev_attrs[] = {
6455 &dev_attr_type.attr,
6456 &dev_attr_perf_event_mux_interval_ms.attr,
6459 ATTRIBUTE_GROUPS(pmu_dev);
6461 static int pmu_bus_running;
6462 static struct bus_type pmu_bus = {
6463 .name = "event_source",
6464 .dev_groups = pmu_dev_groups,
6467 static void pmu_dev_release(struct device *dev)
6472 static int pmu_dev_alloc(struct pmu *pmu)
6476 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6480 pmu->dev->groups = pmu->attr_groups;
6481 device_initialize(pmu->dev);
6482 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6486 dev_set_drvdata(pmu->dev, pmu);
6487 pmu->dev->bus = &pmu_bus;
6488 pmu->dev->release = pmu_dev_release;
6489 ret = device_add(pmu->dev);
6497 put_device(pmu->dev);
6501 static struct lock_class_key cpuctx_mutex;
6502 static struct lock_class_key cpuctx_lock;
6504 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6508 mutex_lock(&pmus_lock);
6510 pmu->pmu_disable_count = alloc_percpu(int);
6511 if (!pmu->pmu_disable_count)
6520 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6528 if (pmu_bus_running) {
6529 ret = pmu_dev_alloc(pmu);
6535 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6536 if (pmu->pmu_cpu_context)
6537 goto got_cpu_context;
6540 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6541 if (!pmu->pmu_cpu_context)
6544 for_each_possible_cpu(cpu) {
6545 struct perf_cpu_context *cpuctx;
6547 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6548 __perf_event_init_context(&cpuctx->ctx);
6549 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6550 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6551 cpuctx->ctx.type = cpu_context;
6552 cpuctx->ctx.pmu = pmu;
6554 __perf_cpu_hrtimer_init(cpuctx, cpu);
6556 INIT_LIST_HEAD(&cpuctx->rotation_list);
6557 cpuctx->unique_pmu = pmu;
6561 if (!pmu->start_txn) {
6562 if (pmu->pmu_enable) {
6564 * If we have pmu_enable/pmu_disable calls, install
6565 * transaction stubs that use that to try and batch
6566 * hardware accesses.
6568 pmu->start_txn = perf_pmu_start_txn;
6569 pmu->commit_txn = perf_pmu_commit_txn;
6570 pmu->cancel_txn = perf_pmu_cancel_txn;
6572 pmu->start_txn = perf_pmu_nop_void;
6573 pmu->commit_txn = perf_pmu_nop_int;
6574 pmu->cancel_txn = perf_pmu_nop_void;
6578 if (!pmu->pmu_enable) {
6579 pmu->pmu_enable = perf_pmu_nop_void;
6580 pmu->pmu_disable = perf_pmu_nop_void;
6583 if (!pmu->event_idx)
6584 pmu->event_idx = perf_event_idx_default;
6586 list_add_rcu(&pmu->entry, &pmus);
6589 mutex_unlock(&pmus_lock);
6594 device_del(pmu->dev);
6595 put_device(pmu->dev);
6598 if (pmu->type >= PERF_TYPE_MAX)
6599 idr_remove(&pmu_idr, pmu->type);
6602 free_percpu(pmu->pmu_disable_count);
6605 EXPORT_SYMBOL_GPL(perf_pmu_register);
6607 void perf_pmu_unregister(struct pmu *pmu)
6609 mutex_lock(&pmus_lock);
6610 list_del_rcu(&pmu->entry);
6611 mutex_unlock(&pmus_lock);
6614 * We dereference the pmu list under both SRCU and regular RCU, so
6615 * synchronize against both of those.
6617 synchronize_srcu(&pmus_srcu);
6620 free_percpu(pmu->pmu_disable_count);
6621 if (pmu->type >= PERF_TYPE_MAX)
6622 idr_remove(&pmu_idr, pmu->type);
6623 device_del(pmu->dev);
6624 put_device(pmu->dev);
6625 free_pmu_context(pmu);
6627 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
6629 struct pmu *perf_init_event(struct perf_event *event)
6631 struct pmu *pmu = NULL;
6635 idx = srcu_read_lock(&pmus_srcu);
6638 pmu = idr_find(&pmu_idr, event->attr.type);
6641 if (!try_module_get(pmu->module)) {
6642 pmu = ERR_PTR(-ENODEV);
6646 ret = pmu->event_init(event);
6652 list_for_each_entry_rcu(pmu, &pmus, entry) {
6653 if (!try_module_get(pmu->module)) {
6654 pmu = ERR_PTR(-ENODEV);
6658 ret = pmu->event_init(event);
6662 if (ret != -ENOENT) {
6667 pmu = ERR_PTR(-ENOENT);
6669 srcu_read_unlock(&pmus_srcu, idx);
6674 static void account_event_cpu(struct perf_event *event, int cpu)
6679 if (has_branch_stack(event)) {
6680 if (!(event->attach_state & PERF_ATTACH_TASK))
6681 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6683 if (is_cgroup_event(event))
6684 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6687 static void account_event(struct perf_event *event)
6692 if (event->attach_state & PERF_ATTACH_TASK)
6693 static_key_slow_inc(&perf_sched_events.key);
6694 if (event->attr.mmap || event->attr.mmap_data)
6695 atomic_inc(&nr_mmap_events);
6696 if (event->attr.comm)
6697 atomic_inc(&nr_comm_events);
6698 if (event->attr.task)
6699 atomic_inc(&nr_task_events);
6700 if (event->attr.freq) {
6701 if (atomic_inc_return(&nr_freq_events) == 1)
6702 tick_nohz_full_kick_all();
6704 if (has_branch_stack(event))
6705 static_key_slow_inc(&perf_sched_events.key);
6706 if (is_cgroup_event(event))
6707 static_key_slow_inc(&perf_sched_events.key);
6709 account_event_cpu(event, event->cpu);
6713 * Allocate and initialize a event structure
6715 static struct perf_event *
6716 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6717 struct task_struct *task,
6718 struct perf_event *group_leader,
6719 struct perf_event *parent_event,
6720 perf_overflow_handler_t overflow_handler,
6724 struct perf_event *event;
6725 struct hw_perf_event *hwc;
6728 if ((unsigned)cpu >= nr_cpu_ids) {
6729 if (!task || cpu != -1)
6730 return ERR_PTR(-EINVAL);
6733 event = kzalloc(sizeof(*event), GFP_KERNEL);
6735 return ERR_PTR(-ENOMEM);
6738 * Single events are their own group leaders, with an
6739 * empty sibling list:
6742 group_leader = event;
6744 mutex_init(&event->child_mutex);
6745 INIT_LIST_HEAD(&event->child_list);
6747 INIT_LIST_HEAD(&event->group_entry);
6748 INIT_LIST_HEAD(&event->event_entry);
6749 INIT_LIST_HEAD(&event->sibling_list);
6750 INIT_LIST_HEAD(&event->rb_entry);
6751 INIT_LIST_HEAD(&event->active_entry);
6752 INIT_HLIST_NODE(&event->hlist_entry);
6755 init_waitqueue_head(&event->waitq);
6756 init_irq_work(&event->pending, perf_pending_event);
6758 mutex_init(&event->mmap_mutex);
6760 atomic_long_set(&event->refcount, 1);
6762 event->attr = *attr;
6763 event->group_leader = group_leader;
6767 event->parent = parent_event;
6769 event->ns = get_pid_ns(task_active_pid_ns(current));
6770 event->id = atomic64_inc_return(&perf_event_id);
6772 event->state = PERF_EVENT_STATE_INACTIVE;
6775 event->attach_state = PERF_ATTACH_TASK;
6777 if (attr->type == PERF_TYPE_TRACEPOINT)
6778 event->hw.tp_target = task;
6779 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6781 * hw_breakpoint is a bit difficult here..
6783 else if (attr->type == PERF_TYPE_BREAKPOINT)
6784 event->hw.bp_target = task;
6788 if (!overflow_handler && parent_event) {
6789 overflow_handler = parent_event->overflow_handler;
6790 context = parent_event->overflow_handler_context;
6793 event->overflow_handler = overflow_handler;
6794 event->overflow_handler_context = context;
6796 perf_event__state_init(event);
6801 hwc->sample_period = attr->sample_period;
6802 if (attr->freq && attr->sample_freq)
6803 hwc->sample_period = 1;
6804 hwc->last_period = hwc->sample_period;
6806 local64_set(&hwc->period_left, hwc->sample_period);
6809 * we currently do not support PERF_FORMAT_GROUP on inherited events
6811 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6814 pmu = perf_init_event(event);
6817 else if (IS_ERR(pmu)) {
6822 if (!event->parent) {
6823 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6824 err = get_callchain_buffers();
6834 event->destroy(event);
6835 module_put(pmu->module);
6838 put_pid_ns(event->ns);
6841 return ERR_PTR(err);
6844 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6845 struct perf_event_attr *attr)
6850 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6854 * zero the full structure, so that a short copy will be nice.
6856 memset(attr, 0, sizeof(*attr));
6858 ret = get_user(size, &uattr->size);
6862 if (size > PAGE_SIZE) /* silly large */
6865 if (!size) /* abi compat */
6866 size = PERF_ATTR_SIZE_VER0;
6868 if (size < PERF_ATTR_SIZE_VER0)
6872 * If we're handed a bigger struct than we know of,
6873 * ensure all the unknown bits are 0 - i.e. new
6874 * user-space does not rely on any kernel feature
6875 * extensions we dont know about yet.
6877 if (size > sizeof(*attr)) {
6878 unsigned char __user *addr;
6879 unsigned char __user *end;
6882 addr = (void __user *)uattr + sizeof(*attr);
6883 end = (void __user *)uattr + size;
6885 for (; addr < end; addr++) {
6886 ret = get_user(val, addr);
6892 size = sizeof(*attr);
6895 ret = copy_from_user(attr, uattr, size);
6899 /* disabled for now */
6903 if (attr->__reserved_1)
6906 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6909 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6912 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6913 u64 mask = attr->branch_sample_type;
6915 /* only using defined bits */
6916 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6919 /* at least one branch bit must be set */
6920 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6923 /* propagate priv level, when not set for branch */
6924 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6926 /* exclude_kernel checked on syscall entry */
6927 if (!attr->exclude_kernel)
6928 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6930 if (!attr->exclude_user)
6931 mask |= PERF_SAMPLE_BRANCH_USER;
6933 if (!attr->exclude_hv)
6934 mask |= PERF_SAMPLE_BRANCH_HV;
6936 * adjust user setting (for HW filter setup)
6938 attr->branch_sample_type = mask;
6940 /* privileged levels capture (kernel, hv): check permissions */
6941 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6942 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6946 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6947 ret = perf_reg_validate(attr->sample_regs_user);
6952 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6953 if (!arch_perf_have_user_stack_dump())
6957 * We have __u32 type for the size, but so far
6958 * we can only use __u16 as maximum due to the
6959 * __u16 sample size limit.
6961 if (attr->sample_stack_user >= USHRT_MAX)
6963 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6971 put_user(sizeof(*attr), &uattr->size);
6977 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6979 struct ring_buffer *rb = NULL;
6985 /* don't allow circular references */
6986 if (event == output_event)
6990 * Don't allow cross-cpu buffers
6992 if (output_event->cpu != event->cpu)
6996 * If its not a per-cpu rb, it must be the same task.
6998 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7002 mutex_lock(&event->mmap_mutex);
7003 /* Can't redirect output if we've got an active mmap() */
7004 if (atomic_read(&event->mmap_count))
7008 /* get the rb we want to redirect to */
7009 rb = ring_buffer_get(output_event);
7014 ring_buffer_attach(event, rb);
7018 mutex_unlock(&event->mmap_mutex);
7025 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7027 * @attr_uptr: event_id type attributes for monitoring/sampling
7030 * @group_fd: group leader event fd
7032 SYSCALL_DEFINE5(perf_event_open,
7033 struct perf_event_attr __user *, attr_uptr,
7034 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7036 struct perf_event *group_leader = NULL, *output_event = NULL;
7037 struct perf_event *event, *sibling;
7038 struct perf_event_attr attr;
7039 struct perf_event_context *ctx;
7040 struct file *event_file = NULL;
7041 struct fd group = {NULL, 0};
7042 struct task_struct *task = NULL;
7047 int f_flags = O_RDWR;
7049 /* for future expandability... */
7050 if (flags & ~PERF_FLAG_ALL)
7053 err = perf_copy_attr(attr_uptr, &attr);
7057 if (!attr.exclude_kernel) {
7058 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7063 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7066 if (attr.sample_period & (1ULL << 63))
7071 * In cgroup mode, the pid argument is used to pass the fd
7072 * opened to the cgroup directory in cgroupfs. The cpu argument
7073 * designates the cpu on which to monitor threads from that
7076 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7079 if (flags & PERF_FLAG_FD_CLOEXEC)
7080 f_flags |= O_CLOEXEC;
7082 event_fd = get_unused_fd_flags(f_flags);
7086 if (group_fd != -1) {
7087 err = perf_fget_light(group_fd, &group);
7090 group_leader = group.file->private_data;
7091 if (flags & PERF_FLAG_FD_OUTPUT)
7092 output_event = group_leader;
7093 if (flags & PERF_FLAG_FD_NO_GROUP)
7094 group_leader = NULL;
7097 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7098 task = find_lively_task_by_vpid(pid);
7100 err = PTR_ERR(task);
7105 if (task && group_leader &&
7106 group_leader->attr.inherit != attr.inherit) {
7113 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7115 if (IS_ERR(event)) {
7116 err = PTR_ERR(event);
7120 if (flags & PERF_FLAG_PID_CGROUP) {
7121 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7123 __free_event(event);
7128 if (is_sampling_event(event)) {
7129 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7135 account_event(event);
7138 * Special case software events and allow them to be part of
7139 * any hardware group.
7144 (is_software_event(event) != is_software_event(group_leader))) {
7145 if (is_software_event(event)) {
7147 * If event and group_leader are not both a software
7148 * event, and event is, then group leader is not.
7150 * Allow the addition of software events to !software
7151 * groups, this is safe because software events never
7154 pmu = group_leader->pmu;
7155 } else if (is_software_event(group_leader) &&
7156 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7158 * In case the group is a pure software group, and we
7159 * try to add a hardware event, move the whole group to
7160 * the hardware context.
7167 * Get the target context (task or percpu):
7169 ctx = find_get_context(pmu, task, event->cpu);
7176 put_task_struct(task);
7181 * Look up the group leader (we will attach this event to it):
7187 * Do not allow a recursive hierarchy (this new sibling
7188 * becoming part of another group-sibling):
7190 if (group_leader->group_leader != group_leader)
7193 * Do not allow to attach to a group in a different
7194 * task or CPU context:
7197 if (group_leader->ctx->type != ctx->type)
7200 if (group_leader->ctx != ctx)
7205 * Only a group leader can be exclusive or pinned
7207 if (attr.exclusive || attr.pinned)
7212 err = perf_event_set_output(event, output_event);
7217 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7219 if (IS_ERR(event_file)) {
7220 err = PTR_ERR(event_file);
7225 struct perf_event_context *gctx = group_leader->ctx;
7227 mutex_lock(&gctx->mutex);
7228 perf_remove_from_context(group_leader, false);
7231 * Removing from the context ends up with disabled
7232 * event. What we want here is event in the initial
7233 * startup state, ready to be add into new context.
7235 perf_event__state_init(group_leader);
7236 list_for_each_entry(sibling, &group_leader->sibling_list,
7238 perf_remove_from_context(sibling, false);
7239 perf_event__state_init(sibling);
7242 mutex_unlock(&gctx->mutex);
7246 WARN_ON_ONCE(ctx->parent_ctx);
7247 mutex_lock(&ctx->mutex);
7251 perf_install_in_context(ctx, group_leader, event->cpu);
7253 list_for_each_entry(sibling, &group_leader->sibling_list,
7255 perf_install_in_context(ctx, sibling, event->cpu);
7260 perf_install_in_context(ctx, event, event->cpu);
7261 perf_unpin_context(ctx);
7262 mutex_unlock(&ctx->mutex);
7266 event->owner = current;
7268 mutex_lock(¤t->perf_event_mutex);
7269 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7270 mutex_unlock(¤t->perf_event_mutex);
7273 * Precalculate sample_data sizes
7275 perf_event__header_size(event);
7276 perf_event__id_header_size(event);
7279 * Drop the reference on the group_event after placing the
7280 * new event on the sibling_list. This ensures destruction
7281 * of the group leader will find the pointer to itself in
7282 * perf_group_detach().
7285 fd_install(event_fd, event_file);
7289 perf_unpin_context(ctx);
7297 put_task_struct(task);
7301 put_unused_fd(event_fd);
7306 * perf_event_create_kernel_counter
7308 * @attr: attributes of the counter to create
7309 * @cpu: cpu in which the counter is bound
7310 * @task: task to profile (NULL for percpu)
7313 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7314 struct task_struct *task,
7315 perf_overflow_handler_t overflow_handler,
7318 struct perf_event_context *ctx;
7319 struct perf_event *event;
7323 * Get the target context (task or percpu):
7326 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7327 overflow_handler, context);
7328 if (IS_ERR(event)) {
7329 err = PTR_ERR(event);
7333 account_event(event);
7335 ctx = find_get_context(event->pmu, task, cpu);
7341 WARN_ON_ONCE(ctx->parent_ctx);
7342 mutex_lock(&ctx->mutex);
7343 perf_install_in_context(ctx, event, cpu);
7344 perf_unpin_context(ctx);
7345 mutex_unlock(&ctx->mutex);
7352 return ERR_PTR(err);
7354 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7356 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7358 struct perf_event_context *src_ctx;
7359 struct perf_event_context *dst_ctx;
7360 struct perf_event *event, *tmp;
7363 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7364 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7366 mutex_lock(&src_ctx->mutex);
7367 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7369 perf_remove_from_context(event, false);
7370 unaccount_event_cpu(event, src_cpu);
7372 list_add(&event->migrate_entry, &events);
7374 mutex_unlock(&src_ctx->mutex);
7378 mutex_lock(&dst_ctx->mutex);
7379 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7380 list_del(&event->migrate_entry);
7381 if (event->state >= PERF_EVENT_STATE_OFF)
7382 event->state = PERF_EVENT_STATE_INACTIVE;
7383 account_event_cpu(event, dst_cpu);
7384 perf_install_in_context(dst_ctx, event, dst_cpu);
7387 mutex_unlock(&dst_ctx->mutex);
7389 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7391 static void sync_child_event(struct perf_event *child_event,
7392 struct task_struct *child)
7394 struct perf_event *parent_event = child_event->parent;
7397 if (child_event->attr.inherit_stat)
7398 perf_event_read_event(child_event, child);
7400 child_val = perf_event_count(child_event);
7403 * Add back the child's count to the parent's count:
7405 atomic64_add(child_val, &parent_event->child_count);
7406 atomic64_add(child_event->total_time_enabled,
7407 &parent_event->child_total_time_enabled);
7408 atomic64_add(child_event->total_time_running,
7409 &parent_event->child_total_time_running);
7412 * Remove this event from the parent's list
7414 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7415 mutex_lock(&parent_event->child_mutex);
7416 list_del_init(&child_event->child_list);
7417 mutex_unlock(&parent_event->child_mutex);
7420 * Release the parent event, if this was the last
7423 put_event(parent_event);
7427 __perf_event_exit_task(struct perf_event *child_event,
7428 struct perf_event_context *child_ctx,
7429 struct task_struct *child)
7431 perf_remove_from_context(child_event, true);
7434 * It can happen that the parent exits first, and has events
7435 * that are still around due to the child reference. These
7436 * events need to be zapped.
7438 if (child_event->parent) {
7439 sync_child_event(child_event, child);
7440 free_event(child_event);
7444 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7446 struct perf_event *child_event, *next;
7447 struct perf_event_context *child_ctx;
7448 unsigned long flags;
7450 if (likely(!child->perf_event_ctxp[ctxn])) {
7451 perf_event_task(child, NULL, 0);
7455 local_irq_save(flags);
7457 * We can't reschedule here because interrupts are disabled,
7458 * and either child is current or it is a task that can't be
7459 * scheduled, so we are now safe from rescheduling changing
7462 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7465 * Take the context lock here so that if find_get_context is
7466 * reading child->perf_event_ctxp, we wait until it has
7467 * incremented the context's refcount before we do put_ctx below.
7469 raw_spin_lock(&child_ctx->lock);
7470 task_ctx_sched_out(child_ctx);
7471 child->perf_event_ctxp[ctxn] = NULL;
7473 * If this context is a clone; unclone it so it can't get
7474 * swapped to another process while we're removing all
7475 * the events from it.
7477 unclone_ctx(child_ctx);
7478 update_context_time(child_ctx);
7479 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7482 * Report the task dead after unscheduling the events so that we
7483 * won't get any samples after PERF_RECORD_EXIT. We can however still
7484 * get a few PERF_RECORD_READ events.
7486 perf_event_task(child, child_ctx, 0);
7489 * We can recurse on the same lock type through:
7491 * __perf_event_exit_task()
7492 * sync_child_event()
7494 * mutex_lock(&ctx->mutex)
7496 * But since its the parent context it won't be the same instance.
7498 mutex_lock(&child_ctx->mutex);
7500 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7501 __perf_event_exit_task(child_event, child_ctx, child);
7503 mutex_unlock(&child_ctx->mutex);
7509 * When a child task exits, feed back event values to parent events.
7511 void perf_event_exit_task(struct task_struct *child)
7513 struct perf_event *event, *tmp;
7516 mutex_lock(&child->perf_event_mutex);
7517 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7519 list_del_init(&event->owner_entry);
7522 * Ensure the list deletion is visible before we clear
7523 * the owner, closes a race against perf_release() where
7524 * we need to serialize on the owner->perf_event_mutex.
7527 event->owner = NULL;
7529 mutex_unlock(&child->perf_event_mutex);
7531 for_each_task_context_nr(ctxn)
7532 perf_event_exit_task_context(child, ctxn);
7535 static void perf_free_event(struct perf_event *event,
7536 struct perf_event_context *ctx)
7538 struct perf_event *parent = event->parent;
7540 if (WARN_ON_ONCE(!parent))
7543 mutex_lock(&parent->child_mutex);
7544 list_del_init(&event->child_list);
7545 mutex_unlock(&parent->child_mutex);
7549 perf_group_detach(event);
7550 list_del_event(event, ctx);
7555 * free an unexposed, unused context as created by inheritance by
7556 * perf_event_init_task below, used by fork() in case of fail.
7558 void perf_event_free_task(struct task_struct *task)
7560 struct perf_event_context *ctx;
7561 struct perf_event *event, *tmp;
7564 for_each_task_context_nr(ctxn) {
7565 ctx = task->perf_event_ctxp[ctxn];
7569 mutex_lock(&ctx->mutex);
7571 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7573 perf_free_event(event, ctx);
7575 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7577 perf_free_event(event, ctx);
7579 if (!list_empty(&ctx->pinned_groups) ||
7580 !list_empty(&ctx->flexible_groups))
7583 mutex_unlock(&ctx->mutex);
7589 void perf_event_delayed_put(struct task_struct *task)
7593 for_each_task_context_nr(ctxn)
7594 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7598 * inherit a event from parent task to child task:
7600 static struct perf_event *
7601 inherit_event(struct perf_event *parent_event,
7602 struct task_struct *parent,
7603 struct perf_event_context *parent_ctx,
7604 struct task_struct *child,
7605 struct perf_event *group_leader,
7606 struct perf_event_context *child_ctx)
7608 struct perf_event *child_event;
7609 unsigned long flags;
7612 * Instead of creating recursive hierarchies of events,
7613 * we link inherited events back to the original parent,
7614 * which has a filp for sure, which we use as the reference
7617 if (parent_event->parent)
7618 parent_event = parent_event->parent;
7620 child_event = perf_event_alloc(&parent_event->attr,
7623 group_leader, parent_event,
7625 if (IS_ERR(child_event))
7628 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7629 free_event(child_event);
7636 * Make the child state follow the state of the parent event,
7637 * not its attr.disabled bit. We hold the parent's mutex,
7638 * so we won't race with perf_event_{en, dis}able_family.
7640 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7641 child_event->state = PERF_EVENT_STATE_INACTIVE;
7643 child_event->state = PERF_EVENT_STATE_OFF;
7645 if (parent_event->attr.freq) {
7646 u64 sample_period = parent_event->hw.sample_period;
7647 struct hw_perf_event *hwc = &child_event->hw;
7649 hwc->sample_period = sample_period;
7650 hwc->last_period = sample_period;
7652 local64_set(&hwc->period_left, sample_period);
7655 child_event->ctx = child_ctx;
7656 child_event->overflow_handler = parent_event->overflow_handler;
7657 child_event->overflow_handler_context
7658 = parent_event->overflow_handler_context;
7661 * Precalculate sample_data sizes
7663 perf_event__header_size(child_event);
7664 perf_event__id_header_size(child_event);
7667 * Link it up in the child's context:
7669 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7670 add_event_to_ctx(child_event, child_ctx);
7671 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7674 * Link this into the parent event's child list
7676 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7677 mutex_lock(&parent_event->child_mutex);
7678 list_add_tail(&child_event->child_list, &parent_event->child_list);
7679 mutex_unlock(&parent_event->child_mutex);
7684 static int inherit_group(struct perf_event *parent_event,
7685 struct task_struct *parent,
7686 struct perf_event_context *parent_ctx,
7687 struct task_struct *child,
7688 struct perf_event_context *child_ctx)
7690 struct perf_event *leader;
7691 struct perf_event *sub;
7692 struct perf_event *child_ctr;
7694 leader = inherit_event(parent_event, parent, parent_ctx,
7695 child, NULL, child_ctx);
7697 return PTR_ERR(leader);
7698 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7699 child_ctr = inherit_event(sub, parent, parent_ctx,
7700 child, leader, child_ctx);
7701 if (IS_ERR(child_ctr))
7702 return PTR_ERR(child_ctr);
7708 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7709 struct perf_event_context *parent_ctx,
7710 struct task_struct *child, int ctxn,
7714 struct perf_event_context *child_ctx;
7716 if (!event->attr.inherit) {
7721 child_ctx = child->perf_event_ctxp[ctxn];
7724 * This is executed from the parent task context, so
7725 * inherit events that have been marked for cloning.
7726 * First allocate and initialize a context for the
7730 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7734 child->perf_event_ctxp[ctxn] = child_ctx;
7737 ret = inherit_group(event, parent, parent_ctx,
7747 * Initialize the perf_event context in task_struct
7749 int perf_event_init_context(struct task_struct *child, int ctxn)
7751 struct perf_event_context *child_ctx, *parent_ctx;
7752 struct perf_event_context *cloned_ctx;
7753 struct perf_event *event;
7754 struct task_struct *parent = current;
7755 int inherited_all = 1;
7756 unsigned long flags;
7759 if (likely(!parent->perf_event_ctxp[ctxn]))
7763 * If the parent's context is a clone, pin it so it won't get
7766 parent_ctx = perf_pin_task_context(parent, ctxn);
7771 * No need to check if parent_ctx != NULL here; since we saw
7772 * it non-NULL earlier, the only reason for it to become NULL
7773 * is if we exit, and since we're currently in the middle of
7774 * a fork we can't be exiting at the same time.
7778 * Lock the parent list. No need to lock the child - not PID
7779 * hashed yet and not running, so nobody can access it.
7781 mutex_lock(&parent_ctx->mutex);
7784 * We dont have to disable NMIs - we are only looking at
7785 * the list, not manipulating it:
7787 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7788 ret = inherit_task_group(event, parent, parent_ctx,
7789 child, ctxn, &inherited_all);
7795 * We can't hold ctx->lock when iterating the ->flexible_group list due
7796 * to allocations, but we need to prevent rotation because
7797 * rotate_ctx() will change the list from interrupt context.
7799 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7800 parent_ctx->rotate_disable = 1;
7801 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7803 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7804 ret = inherit_task_group(event, parent, parent_ctx,
7805 child, ctxn, &inherited_all);
7810 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7811 parent_ctx->rotate_disable = 0;
7813 child_ctx = child->perf_event_ctxp[ctxn];
7815 if (child_ctx && inherited_all) {
7817 * Mark the child context as a clone of the parent
7818 * context, or of whatever the parent is a clone of.
7820 * Note that if the parent is a clone, the holding of
7821 * parent_ctx->lock avoids it from being uncloned.
7823 cloned_ctx = parent_ctx->parent_ctx;
7825 child_ctx->parent_ctx = cloned_ctx;
7826 child_ctx->parent_gen = parent_ctx->parent_gen;
7828 child_ctx->parent_ctx = parent_ctx;
7829 child_ctx->parent_gen = parent_ctx->generation;
7831 get_ctx(child_ctx->parent_ctx);
7834 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7835 mutex_unlock(&parent_ctx->mutex);
7837 perf_unpin_context(parent_ctx);
7838 put_ctx(parent_ctx);
7844 * Initialize the perf_event context in task_struct
7846 int perf_event_init_task(struct task_struct *child)
7850 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7851 mutex_init(&child->perf_event_mutex);
7852 INIT_LIST_HEAD(&child->perf_event_list);
7854 for_each_task_context_nr(ctxn) {
7855 ret = perf_event_init_context(child, ctxn);
7863 static void __init perf_event_init_all_cpus(void)
7865 struct swevent_htable *swhash;
7868 for_each_possible_cpu(cpu) {
7869 swhash = &per_cpu(swevent_htable, cpu);
7870 mutex_init(&swhash->hlist_mutex);
7871 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7875 static void perf_event_init_cpu(int cpu)
7877 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7879 mutex_lock(&swhash->hlist_mutex);
7880 swhash->online = true;
7881 if (swhash->hlist_refcount > 0) {
7882 struct swevent_hlist *hlist;
7884 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7886 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7888 mutex_unlock(&swhash->hlist_mutex);
7891 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7892 static void perf_pmu_rotate_stop(struct pmu *pmu)
7894 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7896 WARN_ON(!irqs_disabled());
7898 list_del_init(&cpuctx->rotation_list);
7901 static void __perf_event_exit_context(void *__info)
7903 struct remove_event re = { .detach_group = false };
7904 struct perf_event_context *ctx = __info;
7906 perf_pmu_rotate_stop(ctx->pmu);
7909 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
7910 __perf_remove_from_context(&re);
7914 static void perf_event_exit_cpu_context(int cpu)
7916 struct perf_event_context *ctx;
7920 idx = srcu_read_lock(&pmus_srcu);
7921 list_for_each_entry_rcu(pmu, &pmus, entry) {
7922 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7924 mutex_lock(&ctx->mutex);
7925 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7926 mutex_unlock(&ctx->mutex);
7928 srcu_read_unlock(&pmus_srcu, idx);
7931 static void perf_event_exit_cpu(int cpu)
7933 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7935 perf_event_exit_cpu_context(cpu);
7937 mutex_lock(&swhash->hlist_mutex);
7938 swhash->online = false;
7939 swevent_hlist_release(swhash);
7940 mutex_unlock(&swhash->hlist_mutex);
7943 static inline void perf_event_exit_cpu(int cpu) { }
7947 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7951 for_each_online_cpu(cpu)
7952 perf_event_exit_cpu(cpu);
7958 * Run the perf reboot notifier at the very last possible moment so that
7959 * the generic watchdog code runs as long as possible.
7961 static struct notifier_block perf_reboot_notifier = {
7962 .notifier_call = perf_reboot,
7963 .priority = INT_MIN,
7967 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7969 unsigned int cpu = (long)hcpu;
7971 switch (action & ~CPU_TASKS_FROZEN) {
7973 case CPU_UP_PREPARE:
7974 case CPU_DOWN_FAILED:
7975 perf_event_init_cpu(cpu);
7978 case CPU_UP_CANCELED:
7979 case CPU_DOWN_PREPARE:
7980 perf_event_exit_cpu(cpu);
7989 void __init perf_event_init(void)
7995 perf_event_init_all_cpus();
7996 init_srcu_struct(&pmus_srcu);
7997 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7998 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7999 perf_pmu_register(&perf_task_clock, NULL, -1);
8001 perf_cpu_notifier(perf_cpu_notify);
8002 register_reboot_notifier(&perf_reboot_notifier);
8004 ret = init_hw_breakpoint();
8005 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8007 /* do not patch jump label more than once per second */
8008 jump_label_rate_limit(&perf_sched_events, HZ);
8011 * Build time assertion that we keep the data_head at the intended
8012 * location. IOW, validation we got the __reserved[] size right.
8014 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8018 static int __init perf_event_sysfs_init(void)
8023 mutex_lock(&pmus_lock);
8025 ret = bus_register(&pmu_bus);
8029 list_for_each_entry(pmu, &pmus, entry) {
8030 if (!pmu->name || pmu->type < 0)
8033 ret = pmu_dev_alloc(pmu);
8034 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8036 pmu_bus_running = 1;
8040 mutex_unlock(&pmus_lock);
8044 device_initcall(perf_event_sysfs_init);
8046 #ifdef CONFIG_CGROUP_PERF
8047 static struct cgroup_subsys_state *
8048 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8050 struct perf_cgroup *jc;
8052 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8054 return ERR_PTR(-ENOMEM);
8056 jc->info = alloc_percpu(struct perf_cgroup_info);
8059 return ERR_PTR(-ENOMEM);
8065 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8067 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8069 free_percpu(jc->info);
8073 static int __perf_cgroup_move(void *info)
8075 struct task_struct *task = info;
8076 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8080 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8081 struct cgroup_taskset *tset)
8083 struct task_struct *task;
8085 cgroup_taskset_for_each(task, tset)
8086 task_function_call(task, __perf_cgroup_move, task);
8089 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8090 struct cgroup_subsys_state *old_css,
8091 struct task_struct *task)
8094 * cgroup_exit() is called in the copy_process() failure path.
8095 * Ignore this case since the task hasn't ran yet, this avoids
8096 * trying to poke a half freed task state from generic code.
8098 if (!(task->flags & PF_EXITING))
8101 task_function_call(task, __perf_cgroup_move, task);
8104 struct cgroup_subsys perf_event_cgrp_subsys = {
8105 .css_alloc = perf_cgroup_css_alloc,
8106 .css_free = perf_cgroup_css_free,
8107 .exit = perf_cgroup_exit,
8108 .attach = perf_cgroup_attach,
8110 #endif /* CONFIG_CGROUP_PERF */