Merge git://git.kernel.org/pub/scm/linux/kernel/git/pablo/nf
[linux-2.6-block.git] / kernel / irq_work.c
CommitLineData
e360adbe
PZ
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
83e3fa6f 8#include <linux/bug.h>
e360adbe 9#include <linux/kernel.h>
9984de1a 10#include <linux/export.h>
e360adbe 11#include <linux/irq_work.h>
967d1f90 12#include <linux/percpu.h>
e360adbe 13#include <linux/hardirq.h>
ef1f0982 14#include <linux/irqflags.h>
bc6679ae
FW
15#include <linux/sched.h>
16#include <linux/tick.h>
c0e980a4
SR
17#include <linux/cpu.h>
18#include <linux/notifier.h>
967d1f90 19#include <asm/processor.h>
e360adbe 20
e360adbe 21
38aaf809 22static DEFINE_PER_CPU(struct llist_head, irq_work_list);
bc6679ae 23static DEFINE_PER_CPU(int, irq_work_raised);
e360adbe
PZ
24
25/*
26 * Claim the entry so that no one else will poke at it.
27 */
38aaf809 28static bool irq_work_claim(struct irq_work *work)
e360adbe 29{
e0bbe2d8 30 unsigned long flags, oflags, nflags;
e360adbe 31
e0bbe2d8
FW
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;
38aaf809 37 for (;;) {
38aaf809 38 nflags = flags | IRQ_WORK_FLAGS;
e0bbe2d8
FW
39 oflags = cmpxchg(&work->flags, flags, nflags);
40 if (oflags == flags)
38aaf809 41 break;
e0bbe2d8
FW
42 if (oflags & IRQ_WORK_PENDING)
43 return false;
44 flags = oflags;
38aaf809
HY
45 cpu_relax();
46 }
e360adbe
PZ
47
48 return true;
49}
50
e360adbe
PZ
51void __weak arch_irq_work_raise(void)
52{
53 /*
54 * Lame architectures will get the timer tick callback
55 */
56}
57
58/*
c02cf5f8 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.
e360adbe 63 */
cd578abb 64bool irq_work_queue(struct irq_work *work)
e360adbe 65{
c02cf5f8 66 /* Only queue if not already pending */
67 if (!irq_work_claim(work))
cd578abb 68 return false;
c02cf5f8 69
70 /* Queue the entry and raise the IPI if needed. */
20b87691 71 preempt_disable();
e360adbe 72
bc6679ae
FW
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 }
e360adbe 84
20b87691 85 preempt_enable();
cd578abb
PZ
86
87 return true;
e360adbe 88}
e360adbe
PZ
89EXPORT_SYMBOL_GPL(irq_work_queue);
90
00b42959
FW
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
8aa2acce
SR
99 /* All work should have been flushed before going offline */
100 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
101
00b42959
FW
102 return true;
103}
104
c0e980a4 105static void __irq_work_run(void)
e360adbe 106{
bc6679ae 107 unsigned long flags;
38aaf809
HY
108 struct irq_work *work;
109 struct llist_head *this_list;
110 struct llist_node *llnode;
e360adbe 111
bc6679ae
FW
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
38aaf809
HY
120 this_list = &__get_cpu_var(irq_work_list);
121 if (llist_empty(this_list))
e360adbe
PZ
122 return;
123
e360adbe
PZ
124 BUG_ON(!irqs_disabled());
125
38aaf809
HY
126 llnode = llist_del_all(this_list);
127 while (llnode != NULL) {
128 work = llist_entry(llnode, struct irq_work, llnode);
e360adbe 129
924f8f5a 130 llnode = llist_next(llnode);
e360adbe
PZ
131
132 /*
38aaf809 133 * Clear the PENDING bit, after this point the @work
e360adbe 134 * can be re-used.
c8446b75
FW
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.
e360adbe 138 */
bc6679ae
FW
139 flags = work->flags & ~IRQ_WORK_PENDING;
140 xchg(&work->flags, flags);
141
38aaf809 142 work->func(work);
e360adbe
PZ
143 /*
144 * Clear the BUSY bit and return to the free state if
145 * no-one else claimed it meanwhile.
146 */
bc6679ae 147 (void)cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
e360adbe
PZ
148 }
149}
c0e980a4
SR
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}
e360adbe
PZ
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 */
38aaf809 166void irq_work_sync(struct irq_work *work)
e360adbe
PZ
167{
168 WARN_ON_ONCE(irqs_disabled());
169
38aaf809 170 while (work->flags & IRQ_WORK_BUSY)
e360adbe
PZ
171 cpu_relax();
172}
173EXPORT_SYMBOL_GPL(irq_work_sync);
c0e980a4
SR
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 */