Loading...
1/*
2 * Copyright (C) 2010 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
3 *
4 * Provides a framework for enqueueing and running callbacks from hardirq
5 * context. The enqueueing is NMI-safe.
6 */
7
8#include <linux/bug.h>
9#include <linux/kernel.h>
10#include <linux/export.h>
11#include <linux/irq_work.h>
12#include <linux/percpu.h>
13#include <linux/hardirq.h>
14#include <linux/irqflags.h>
15#include <linux/sched.h>
16#include <linux/tick.h>
17#include <linux/cpu.h>
18#include <linux/notifier.h>
19#include <asm/processor.h>
20
21
22static DEFINE_PER_CPU(struct llist_head, irq_work_list);
23static DEFINE_PER_CPU(int, irq_work_raised);
24
25/*
26 * Claim the entry so that no one else will poke at it.
27 */
28static bool irq_work_claim(struct irq_work *work)
29{
30 unsigned long flags, oflags, nflags;
31
32 /*
33 * Start with our best wish as a premise but only trust any
34 * flag value after cmpxchg() result.
35 */
36 flags = work->flags & ~IRQ_WORK_PENDING;
37 for (;;) {
38 nflags = flags | IRQ_WORK_FLAGS;
39 oflags = cmpxchg(&work->flags, flags, nflags);
40 if (oflags == flags)
41 break;
42 if (oflags & IRQ_WORK_PENDING)
43 return false;
44 flags = oflags;
45 cpu_relax();
46 }
47
48 return true;
49}
50
51void __weak arch_irq_work_raise(void)
52{
53 /*
54 * Lame architectures will get the timer tick callback
55 */
56}
57
58/*
59 * Enqueue the irq_work @entry unless it's already pending
60 * somewhere.
61 *
62 * Can be re-enqueued while the callback is still in progress.
63 */
64bool irq_work_queue(struct irq_work *work)
65{
66 /* Only queue if not already pending */
67 if (!irq_work_claim(work))
68 return false;
69
70 /* Queue the entry and raise the IPI if needed. */
71 preempt_disable();
72
73 llist_add(&work->llnode, &__get_cpu_var(irq_work_list));
74
75 /*
76 * If the work is not "lazy" or the tick is stopped, raise the irq
77 * work interrupt (if supported by the arch), otherwise, just wait
78 * for the next tick.
79 */
80 if (!(work->flags & IRQ_WORK_LAZY) || tick_nohz_tick_stopped()) {
81 if (!this_cpu_cmpxchg(irq_work_raised, 0, 1))
82 arch_irq_work_raise();
83 }
84
85 preempt_enable();
86
87 return true;
88}
89EXPORT_SYMBOL_GPL(irq_work_queue);
90
91bool irq_work_needs_cpu(void)
92{
93 struct llist_head *this_list;
94
95 this_list = &__get_cpu_var(irq_work_list);
96 if (llist_empty(this_list))
97 return false;
98
99 /* All work should have been flushed before going offline */
100 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
101
102 return true;
103}
104
105static void __irq_work_run(void)
106{
107 unsigned long flags;
108 struct irq_work *work;
109 struct llist_head *this_list;
110 struct llist_node *llnode;
111
112
113 /*
114 * Reset the "raised" state right before we check the list because
115 * an NMI may enqueue after we find the list empty from the runner.
116 */
117 __this_cpu_write(irq_work_raised, 0);
118 barrier();
119
120 this_list = &__get_cpu_var(irq_work_list);
121 if (llist_empty(this_list))
122 return;
123
124 BUG_ON(!irqs_disabled());
125
126 llnode = llist_del_all(this_list);
127 while (llnode != NULL) {
128 work = llist_entry(llnode, struct irq_work, llnode);
129
130 llnode = llist_next(llnode);
131
132 /*
133 * Clear the PENDING bit, after this point the @work
134 * can be re-used.
135 * Make it immediately visible so that other CPUs trying
136 * to claim that work don't rely on us to handle their data
137 * while we are in the middle of the func.
138 */
139 flags = work->flags & ~IRQ_WORK_PENDING;
140 xchg(&work->flags, flags);
141
142 work->func(work);
143 /*
144 * Clear the BUSY bit and return to the free state if
145 * no-one else claimed it meanwhile.
146 */
147 (void)cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
148 }
149}
150
151/*
152 * Run the irq_work entries on this cpu. Requires to be ran from hardirq
153 * context with local IRQs disabled.
154 */
155void irq_work_run(void)
156{
157 BUG_ON(!in_irq());
158 __irq_work_run();
159}
160EXPORT_SYMBOL_GPL(irq_work_run);
161
162/*
163 * Synchronize against the irq_work @entry, ensures the entry is not
164 * currently in use.
165 */
166void irq_work_sync(struct irq_work *work)
167{
168 WARN_ON_ONCE(irqs_disabled());
169
170 while (work->flags & IRQ_WORK_BUSY)
171 cpu_relax();
172}
173EXPORT_SYMBOL_GPL(irq_work_sync);
174
175#ifdef CONFIG_HOTPLUG_CPU
176static int irq_work_cpu_notify(struct notifier_block *self,
177 unsigned long action, void *hcpu)
178{
179 long cpu = (long)hcpu;
180
181 switch (action) {
182 case CPU_DYING:
183 /* Called from stop_machine */
184 if (WARN_ON_ONCE(cpu != smp_processor_id()))
185 break;
186 __irq_work_run();
187 break;
188 default:
189 break;
190 }
191 return NOTIFY_OK;
192}
193
194static struct notifier_block cpu_notify;
195
196static __init int irq_work_init_cpu_notifier(void)
197{
198 cpu_notify.notifier_call = irq_work_cpu_notify;
199 cpu_notify.priority = 0;
200 register_cpu_notifier(&cpu_notify);
201 return 0;
202}
203device_initcall(irq_work_init_cpu_notifier);
204
205#endif /* CONFIG_HOTPLUG_CPU */
1/*
2 * Copyright (C) 2010 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
3 *
4 * Provides a framework for enqueueing and running callbacks from hardirq
5 * context. The enqueueing is NMI-safe.
6 */
7
8#include <linux/kernel.h>
9#include <linux/module.h>
10#include <linux/irq_work.h>
11#include <linux/hardirq.h>
12
13/*
14 * An entry can be in one of four states:
15 *
16 * free NULL, 0 -> {claimed} : free to be used
17 * claimed NULL, 3 -> {pending} : claimed to be enqueued
18 * pending next, 3 -> {busy} : queued, pending callback
19 * busy NULL, 2 -> {free, claimed} : callback in progress, can be claimed
20 *
21 * We use the lower two bits of the next pointer to keep PENDING and BUSY
22 * flags.
23 */
24
25#define IRQ_WORK_PENDING 1UL
26#define IRQ_WORK_BUSY 2UL
27#define IRQ_WORK_FLAGS 3UL
28
29static inline bool irq_work_is_set(struct irq_work *entry, int flags)
30{
31 return (unsigned long)entry->next & flags;
32}
33
34static inline struct irq_work *irq_work_next(struct irq_work *entry)
35{
36 unsigned long next = (unsigned long)entry->next;
37 next &= ~IRQ_WORK_FLAGS;
38 return (struct irq_work *)next;
39}
40
41static inline struct irq_work *next_flags(struct irq_work *entry, int flags)
42{
43 unsigned long next = (unsigned long)entry;
44 next |= flags;
45 return (struct irq_work *)next;
46}
47
48static DEFINE_PER_CPU(struct irq_work *, irq_work_list);
49
50/*
51 * Claim the entry so that no one else will poke at it.
52 */
53static bool irq_work_claim(struct irq_work *entry)
54{
55 struct irq_work *next, *nflags;
56
57 do {
58 next = entry->next;
59 if ((unsigned long)next & IRQ_WORK_PENDING)
60 return false;
61 nflags = next_flags(next, IRQ_WORK_FLAGS);
62 } while (cmpxchg(&entry->next, next, nflags) != next);
63
64 return true;
65}
66
67
68void __weak arch_irq_work_raise(void)
69{
70 /*
71 * Lame architectures will get the timer tick callback
72 */
73}
74
75/*
76 * Queue the entry and raise the IPI if needed.
77 */
78static void __irq_work_queue(struct irq_work *entry)
79{
80 struct irq_work *next;
81
82 preempt_disable();
83
84 do {
85 next = __this_cpu_read(irq_work_list);
86 /* Can assign non-atomic because we keep the flags set. */
87 entry->next = next_flags(next, IRQ_WORK_FLAGS);
88 } while (this_cpu_cmpxchg(irq_work_list, next, entry) != next);
89
90 /* The list was empty, raise self-interrupt to start processing. */
91 if (!irq_work_next(entry))
92 arch_irq_work_raise();
93
94 preempt_enable();
95}
96
97/*
98 * Enqueue the irq_work @entry, returns true on success, failure when the
99 * @entry was already enqueued by someone else.
100 *
101 * Can be re-enqueued while the callback is still in progress.
102 */
103bool irq_work_queue(struct irq_work *entry)
104{
105 if (!irq_work_claim(entry)) {
106 /*
107 * Already enqueued, can't do!
108 */
109 return false;
110 }
111
112 __irq_work_queue(entry);
113 return true;
114}
115EXPORT_SYMBOL_GPL(irq_work_queue);
116
117/*
118 * Run the irq_work entries on this cpu. Requires to be ran from hardirq
119 * context with local IRQs disabled.
120 */
121void irq_work_run(void)
122{
123 struct irq_work *list;
124
125 if (this_cpu_read(irq_work_list) == NULL)
126 return;
127
128 BUG_ON(!in_irq());
129 BUG_ON(!irqs_disabled());
130
131 list = this_cpu_xchg(irq_work_list, NULL);
132
133 while (list != NULL) {
134 struct irq_work *entry = list;
135
136 list = irq_work_next(list);
137
138 /*
139 * Clear the PENDING bit, after this point the @entry
140 * can be re-used.
141 */
142 entry->next = next_flags(NULL, IRQ_WORK_BUSY);
143 entry->func(entry);
144 /*
145 * Clear the BUSY bit and return to the free state if
146 * no-one else claimed it meanwhile.
147 */
148 (void)cmpxchg(&entry->next,
149 next_flags(NULL, IRQ_WORK_BUSY),
150 NULL);
151 }
152}
153EXPORT_SYMBOL_GPL(irq_work_run);
154
155/*
156 * Synchronize against the irq_work @entry, ensures the entry is not
157 * currently in use.
158 */
159void irq_work_sync(struct irq_work *entry)
160{
161 WARN_ON_ONCE(irqs_disabled());
162
163 while (irq_work_is_set(entry, IRQ_WORK_BUSY))
164 cpu_relax();
165}
166EXPORT_SYMBOL_GPL(irq_work_sync);