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
22 static DEFINE_PER_CPU(struct llist_head, irq_work_list);
23 static DEFINE_PER_CPU(int, irq_work_raised);
24
25 /*
26 * Claim the entry so that no one else will poke at it.
27 */
irq_work_claim(struct irq_work * work)28 static 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
arch_irq_work_raise(void)51 void __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 */
irq_work_queue(struct irq_work * work)64 void irq_work_queue(struct irq_work *work)
65 {
66 /* Only queue if not already pending */
67 if (!irq_work_claim(work))
68 return;
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 EXPORT_SYMBOL_GPL(irq_work_queue);
88
irq_work_needs_cpu(void)89 bool irq_work_needs_cpu(void)
90 {
91 struct llist_head *this_list;
92
93 this_list = &__get_cpu_var(irq_work_list);
94 if (llist_empty(this_list))
95 return false;
96
97 /* All work should have been flushed before going offline */
98 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
99
100 return true;
101 }
102
__irq_work_run(void)103 static void __irq_work_run(void)
104 {
105 unsigned long flags;
106 struct irq_work *work;
107 struct llist_head *this_list;
108 struct llist_node *llnode;
109
110
111 /*
112 * Reset the "raised" state right before we check the list because
113 * an NMI may enqueue after we find the list empty from the runner.
114 */
115 __this_cpu_write(irq_work_raised, 0);
116 barrier();
117
118 this_list = &__get_cpu_var(irq_work_list);
119 if (llist_empty(this_list))
120 return;
121
122 BUG_ON(!irqs_disabled());
123
124 llnode = llist_del_all(this_list);
125 while (llnode != NULL) {
126 work = llist_entry(llnode, struct irq_work, llnode);
127
128 llnode = llist_next(llnode);
129
130 /*
131 * Clear the PENDING bit, after this point the @work
132 * can be re-used.
133 * Make it immediately visible so that other CPUs trying
134 * to claim that work don't rely on us to handle their data
135 * while we are in the middle of the func.
136 */
137 flags = work->flags & ~IRQ_WORK_PENDING;
138 xchg(&work->flags, flags);
139
140 work->func(work);
141 /*
142 * Clear the BUSY bit and return to the free state if
143 * no-one else claimed it meanwhile.
144 */
145 (void)cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
146 }
147 }
148
149 /*
150 * Run the irq_work entries on this cpu. Requires to be ran from hardirq
151 * context with local IRQs disabled.
152 */
irq_work_run(void)153 void irq_work_run(void)
154 {
155 BUG_ON(!in_irq());
156 __irq_work_run();
157 }
158 EXPORT_SYMBOL_GPL(irq_work_run);
159
160 /*
161 * Synchronize against the irq_work @entry, ensures the entry is not
162 * currently in use.
163 */
irq_work_sync(struct irq_work * work)164 void irq_work_sync(struct irq_work *work)
165 {
166 WARN_ON_ONCE(irqs_disabled());
167
168 while (work->flags & IRQ_WORK_BUSY)
169 cpu_relax();
170 }
171 EXPORT_SYMBOL_GPL(irq_work_sync);
172
173 #ifdef CONFIG_HOTPLUG_CPU
irq_work_cpu_notify(struct notifier_block * self,unsigned long action,void * hcpu)174 static int irq_work_cpu_notify(struct notifier_block *self,
175 unsigned long action, void *hcpu)
176 {
177 long cpu = (long)hcpu;
178
179 switch (action) {
180 case CPU_DYING:
181 /* Called from stop_machine */
182 if (WARN_ON_ONCE(cpu != smp_processor_id()))
183 break;
184 __irq_work_run();
185 break;
186 default:
187 break;
188 }
189 return NOTIFY_OK;
190 }
191
192 static struct notifier_block cpu_notify;
193
irq_work_init_cpu_notifier(void)194 static __init int irq_work_init_cpu_notifier(void)
195 {
196 cpu_notify.notifier_call = irq_work_cpu_notify;
197 cpu_notify.priority = 0;
198 register_cpu_notifier(&cpu_notify);
199 return 0;
200 }
201 device_initcall(irq_work_init_cpu_notifier);
202
203 #endif /* CONFIG_HOTPLUG_CPU */
204