1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2010 Red Hat, Inc., Peter Zijlstra
4 *
5 * Provides a framework for enqueueing and running callbacks from hardirq
6 * context. The enqueueing is NMI-safe.
7 */
8
9 #include <linux/bug.h>
10 #include <linux/kernel.h>
11 #include <linux/export.h>
12 #include <linux/irq_work.h>
13 #include <linux/percpu.h>
14 #include <linux/hardirq.h>
15 #include <linux/irqflags.h>
16 #include <linux/sched.h>
17 #include <linux/tick.h>
18 #include <linux/cpu.h>
19 #include <linux/notifier.h>
20 #include <linux/smp.h>
21 #include <asm/processor.h>
22
23
24 static DEFINE_PER_CPU(struct llist_head, raised_list);
25 static DEFINE_PER_CPU(struct llist_head, lazy_list);
26
27 /*
28 * Claim the entry so that no one else will poke at it.
29 */
irq_work_claim(struct irq_work * work)30 static bool irq_work_claim(struct irq_work *work)
31 {
32 unsigned long flags, oflags, nflags;
33
34 /*
35 * Start with our best wish as a premise but only trust any
36 * flag value after cmpxchg() result.
37 */
38 flags = work->flags & ~IRQ_WORK_PENDING;
39 for (;;) {
40 nflags = flags | IRQ_WORK_CLAIMED;
41 oflags = cmpxchg(&work->flags, flags, nflags);
42 if (oflags == flags)
43 break;
44 if (oflags & IRQ_WORK_PENDING)
45 return false;
46 flags = oflags;
47 cpu_relax();
48 }
49
50 return true;
51 }
52
arch_irq_work_raise(void)53 void __weak arch_irq_work_raise(void)
54 {
55 /*
56 * Lame architectures will get the timer tick callback
57 */
58 }
59
60 /* Enqueue on current CPU, work must already be claimed and preempt disabled */
__irq_work_queue_local(struct irq_work * work)61 static void __irq_work_queue_local(struct irq_work *work)
62 {
63 /* If the work is "lazy", handle it from next tick if any */
64 if (work->flags & IRQ_WORK_LAZY) {
65 if (llist_add(&work->llnode, this_cpu_ptr(&lazy_list)) &&
66 tick_nohz_tick_stopped())
67 arch_irq_work_raise();
68 } else {
69 if (llist_add(&work->llnode, this_cpu_ptr(&raised_list)))
70 arch_irq_work_raise();
71 }
72 }
73
74 /* Enqueue the irq work @work on the current CPU */
irq_work_queue(struct irq_work * work)75 bool irq_work_queue(struct irq_work *work)
76 {
77 /* Only queue if not already pending */
78 if (!irq_work_claim(work))
79 return false;
80
81 /* Queue the entry and raise the IPI if needed. */
82 preempt_disable();
83 __irq_work_queue_local(work);
84 preempt_enable();
85
86 return true;
87 }
88 EXPORT_SYMBOL_GPL(irq_work_queue);
89
90 /*
91 * Enqueue the irq_work @work on @cpu unless it's already pending
92 * somewhere.
93 *
94 * Can be re-enqueued while the callback is still in progress.
95 */
irq_work_queue_on(struct irq_work * work,int cpu)96 bool irq_work_queue_on(struct irq_work *work, int cpu)
97 {
98 #ifndef CONFIG_SMP
99 return irq_work_queue(work);
100
101 #else /* CONFIG_SMP: */
102 /* All work should have been flushed before going offline */
103 WARN_ON_ONCE(cpu_is_offline(cpu));
104
105 /* Only queue if not already pending */
106 if (!irq_work_claim(work))
107 return false;
108
109 preempt_disable();
110 if (cpu != smp_processor_id()) {
111 /* Arch remote IPI send/receive backend aren't NMI safe */
112 WARN_ON_ONCE(in_nmi());
113 if (llist_add(&work->llnode, &per_cpu(raised_list, cpu)))
114 arch_send_call_function_single_ipi(cpu);
115 } else {
116 __irq_work_queue_local(work);
117 }
118 preempt_enable();
119
120 return true;
121 #endif /* CONFIG_SMP */
122 }
123
124
irq_work_needs_cpu(void)125 bool irq_work_needs_cpu(void)
126 {
127 struct llist_head *raised, *lazy;
128
129 raised = this_cpu_ptr(&raised_list);
130 lazy = this_cpu_ptr(&lazy_list);
131
132 if (llist_empty(raised) || arch_irq_work_has_interrupt())
133 if (llist_empty(lazy))
134 return false;
135
136 /* All work should have been flushed before going offline */
137 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
138
139 return true;
140 }
141
irq_work_run_list(struct llist_head * list)142 static void irq_work_run_list(struct llist_head *list)
143 {
144 struct irq_work *work, *tmp;
145 struct llist_node *llnode;
146 unsigned long flags;
147
148 BUG_ON(!irqs_disabled());
149
150 if (llist_empty(list))
151 return;
152
153 llnode = llist_del_all(list);
154 llist_for_each_entry_safe(work, tmp, llnode, llnode) {
155 /*
156 * Clear the PENDING bit, after this point the @work
157 * can be re-used.
158 * Make it immediately visible so that other CPUs trying
159 * to claim that work don't rely on us to handle their data
160 * while we are in the middle of the func.
161 */
162 flags = work->flags & ~IRQ_WORK_PENDING;
163 xchg(&work->flags, flags);
164
165 work->func(work);
166 /*
167 * Clear the BUSY bit and return to the free state if
168 * no-one else claimed it meanwhile.
169 */
170 (void)cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
171 }
172 }
173
174 /*
175 * hotplug calls this through:
176 * hotplug_cfd() -> flush_smp_call_function_queue()
177 */
irq_work_run(void)178 void irq_work_run(void)
179 {
180 irq_work_run_list(this_cpu_ptr(&raised_list));
181 irq_work_run_list(this_cpu_ptr(&lazy_list));
182 }
183 EXPORT_SYMBOL_GPL(irq_work_run);
184
irq_work_tick(void)185 void irq_work_tick(void)
186 {
187 struct llist_head *raised = this_cpu_ptr(&raised_list);
188
189 if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
190 irq_work_run_list(raised);
191 irq_work_run_list(this_cpu_ptr(&lazy_list));
192 }
193
194 /*
195 * Synchronize against the irq_work @entry, ensures the entry is not
196 * currently in use.
197 */
irq_work_sync(struct irq_work * work)198 void irq_work_sync(struct irq_work *work)
199 {
200 lockdep_assert_irqs_enabled();
201
202 while (work->flags & IRQ_WORK_BUSY)
203 cpu_relax();
204 }
205 EXPORT_SYMBOL_GPL(irq_work_sync);
206