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 int oflags;
33
34 oflags = atomic_fetch_or(IRQ_WORK_CLAIMED | CSD_TYPE_IRQ_WORK, &work->flags);
35 /*
36 * If the work is already pending, no need to raise the IPI.
37 * The pairing atomic_fetch_andnot() in irq_work_run() makes sure
38 * everything we did before is visible.
39 */
40 if (oflags & IRQ_WORK_PENDING)
41 return false;
42 return true;
43 }
44
arch_irq_work_raise(void)45 void __weak arch_irq_work_raise(void)
46 {
47 /*
48 * Lame architectures will get the timer tick callback
49 */
50 }
51
52 /* Enqueue on current CPU, work must already be claimed and preempt disabled */
__irq_work_queue_local(struct irq_work * work)53 static void __irq_work_queue_local(struct irq_work *work)
54 {
55 /* If the work is "lazy", handle it from next tick if any */
56 if (atomic_read(&work->flags) & IRQ_WORK_LAZY) {
57 if (llist_add(&work->llnode, this_cpu_ptr(&lazy_list)) &&
58 tick_nohz_tick_stopped())
59 arch_irq_work_raise();
60 } else {
61 if (llist_add(&work->llnode, this_cpu_ptr(&raised_list)))
62 arch_irq_work_raise();
63 }
64 }
65
66 /* Enqueue the irq work @work on the current CPU */
irq_work_queue(struct irq_work * work)67 bool irq_work_queue(struct irq_work *work)
68 {
69 /* Only queue if not already pending */
70 if (!irq_work_claim(work))
71 return false;
72
73 /* Queue the entry and raise the IPI if needed. */
74 preempt_disable();
75 __irq_work_queue_local(work);
76 preempt_enable();
77
78 return true;
79 }
80 EXPORT_SYMBOL_GPL(irq_work_queue);
81
82 /*
83 * Enqueue the irq_work @work on @cpu unless it's already pending
84 * somewhere.
85 *
86 * Can be re-enqueued while the callback is still in progress.
87 */
irq_work_queue_on(struct irq_work * work,int cpu)88 bool irq_work_queue_on(struct irq_work *work, int cpu)
89 {
90 #ifndef CONFIG_SMP
91 return irq_work_queue(work);
92
93 #else /* CONFIG_SMP: */
94 /* All work should have been flushed before going offline */
95 WARN_ON_ONCE(cpu_is_offline(cpu));
96
97 /* Only queue if not already pending */
98 if (!irq_work_claim(work))
99 return false;
100
101 preempt_disable();
102 if (cpu != smp_processor_id()) {
103 /* Arch remote IPI send/receive backend aren't NMI safe */
104 WARN_ON_ONCE(in_nmi());
105 __smp_call_single_queue(cpu, &work->llnode);
106 } else {
107 __irq_work_queue_local(work);
108 }
109 preempt_enable();
110
111 return true;
112 #endif /* CONFIG_SMP */
113 }
114
115
irq_work_needs_cpu(void)116 bool irq_work_needs_cpu(void)
117 {
118 struct llist_head *raised, *lazy;
119
120 raised = this_cpu_ptr(&raised_list);
121 lazy = this_cpu_ptr(&lazy_list);
122
123 if (llist_empty(raised) || arch_irq_work_has_interrupt())
124 if (llist_empty(lazy))
125 return false;
126
127 /* All work should have been flushed before going offline */
128 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
129
130 return true;
131 }
132
irq_work_single(void * arg)133 void irq_work_single(void *arg)
134 {
135 struct irq_work *work = arg;
136 int flags;
137
138 /*
139 * Clear the PENDING bit, after this point the @work
140 * can be re-used.
141 * Make it immediately visible so that other CPUs trying
142 * to claim that work don't rely on us to handle their data
143 * while we are in the middle of the func.
144 */
145 flags = atomic_fetch_andnot(IRQ_WORK_PENDING, &work->flags);
146
147 lockdep_irq_work_enter(work);
148 work->func(work);
149 lockdep_irq_work_exit(work);
150 /*
151 * Clear the BUSY bit and return to the free state if
152 * no-one else claimed it meanwhile.
153 */
154 flags &= ~IRQ_WORK_PENDING;
155 (void)atomic_cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
156 }
157
irq_work_run_list(struct llist_head * list)158 static void irq_work_run_list(struct llist_head *list)
159 {
160 struct irq_work *work, *tmp;
161 struct llist_node *llnode;
162
163 BUG_ON(!irqs_disabled());
164
165 if (llist_empty(list))
166 return;
167
168 llnode = llist_del_all(list);
169 llist_for_each_entry_safe(work, tmp, llnode, llnode)
170 irq_work_single(work);
171 }
172
173 /*
174 * hotplug calls this through:
175 * hotplug_cfd() -> flush_smp_call_function_queue()
176 */
irq_work_run(void)177 void irq_work_run(void)
178 {
179 irq_work_run_list(this_cpu_ptr(&raised_list));
180 irq_work_run_list(this_cpu_ptr(&lazy_list));
181 }
182 EXPORT_SYMBOL_GPL(irq_work_run);
183
irq_work_tick(void)184 void irq_work_tick(void)
185 {
186 struct llist_head *raised = this_cpu_ptr(&raised_list);
187
188 if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
189 irq_work_run_list(raised);
190 irq_work_run_list(this_cpu_ptr(&lazy_list));
191 }
192
193 /*
194 * Synchronize against the irq_work @entry, ensures the entry is not
195 * currently in use.
196 */
irq_work_sync(struct irq_work * work)197 void irq_work_sync(struct irq_work *work)
198 {
199 lockdep_assert_irqs_enabled();
200
201 while (atomic_read(&work->flags) & IRQ_WORK_BUSY)
202 cpu_relax();
203 }
204 EXPORT_SYMBOL_GPL(irq_work_sync);
205