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 #include <linux/kasan.h>
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->node.a_flags);
35 /*
36 * If the work is already pending, no need to raise the IPI.
37 * The pairing smp_mb() in irq_work_single() 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->node.a_flags) & IRQ_WORK_LAZY) {
57 if (llist_add(&work->node.llist, this_cpu_ptr(&lazy_list)) &&
58 tick_nohz_tick_stopped())
59 arch_irq_work_raise();
60 } else {
61 if (llist_add(&work->node.llist, 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 /*record irq_work call stack in order to print it in KASAN reports*/
74 kasan_record_aux_stack(work);
75
76 /* Queue the entry and raise the IPI if needed. */
77 preempt_disable();
78 __irq_work_queue_local(work);
79 preempt_enable();
80
81 return true;
82 }
83 EXPORT_SYMBOL_GPL(irq_work_queue);
84
85 /*
86 * Enqueue the irq_work @work on @cpu unless it's already pending
87 * somewhere.
88 *
89 * Can be re-enqueued while the callback is still in progress.
90 */
irq_work_queue_on(struct irq_work * work,int cpu)91 bool irq_work_queue_on(struct irq_work *work, int cpu)
92 {
93 #ifndef CONFIG_SMP
94 return irq_work_queue(work);
95
96 #else /* CONFIG_SMP: */
97 /* All work should have been flushed before going offline */
98 WARN_ON_ONCE(cpu_is_offline(cpu));
99
100 /* Only queue if not already pending */
101 if (!irq_work_claim(work))
102 return false;
103
104 kasan_record_aux_stack(work);
105
106 preempt_disable();
107 if (cpu != smp_processor_id()) {
108 /* Arch remote IPI send/receive backend aren't NMI safe */
109 WARN_ON_ONCE(in_nmi());
110 __smp_call_single_queue(cpu, &work->node.llist);
111 } else {
112 __irq_work_queue_local(work);
113 }
114 preempt_enable();
115
116 return true;
117 #endif /* CONFIG_SMP */
118 }
119
120
irq_work_needs_cpu(void)121 bool irq_work_needs_cpu(void)
122 {
123 struct llist_head *raised, *lazy;
124
125 raised = this_cpu_ptr(&raised_list);
126 lazy = this_cpu_ptr(&lazy_list);
127
128 if (llist_empty(raised) || arch_irq_work_has_interrupt())
129 if (llist_empty(lazy))
130 return false;
131
132 /* All work should have been flushed before going offline */
133 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
134
135 return true;
136 }
137
irq_work_single(void * arg)138 void irq_work_single(void *arg)
139 {
140 struct irq_work *work = arg;
141 int flags;
142
143 /*
144 * Clear the PENDING bit, after this point the @work can be re-used.
145 * The PENDING bit acts as a lock, and we own it, so we can clear it
146 * without atomic ops.
147 */
148 flags = atomic_read(&work->node.a_flags);
149 flags &= ~IRQ_WORK_PENDING;
150 atomic_set(&work->node.a_flags, flags);
151
152 /*
153 * See irq_work_claim().
154 */
155 smp_mb();
156
157 lockdep_irq_work_enter(flags);
158 work->func(work);
159 lockdep_irq_work_exit(flags);
160
161 /*
162 * Clear the BUSY bit, if set, and return to the free state if no-one
163 * else claimed it meanwhile.
164 */
165 (void)atomic_cmpxchg(&work->node.a_flags, flags, flags & ~IRQ_WORK_BUSY);
166 }
167
irq_work_run_list(struct llist_head * list)168 static void irq_work_run_list(struct llist_head *list)
169 {
170 struct irq_work *work, *tmp;
171 struct llist_node *llnode;
172
173 BUG_ON(!irqs_disabled());
174
175 if (llist_empty(list))
176 return;
177
178 llnode = llist_del_all(list);
179 llist_for_each_entry_safe(work, tmp, llnode, node.llist)
180 irq_work_single(work);
181 }
182
183 /*
184 * hotplug calls this through:
185 * hotplug_cfd() -> flush_smp_call_function_queue()
186 */
irq_work_run(void)187 void irq_work_run(void)
188 {
189 irq_work_run_list(this_cpu_ptr(&raised_list));
190 irq_work_run_list(this_cpu_ptr(&lazy_list));
191 }
192 EXPORT_SYMBOL_GPL(irq_work_run);
193
irq_work_tick(void)194 void irq_work_tick(void)
195 {
196 struct llist_head *raised = this_cpu_ptr(&raised_list);
197
198 if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
199 irq_work_run_list(raised);
200 irq_work_run_list(this_cpu_ptr(&lazy_list));
201 }
202
203 /*
204 * Synchronize against the irq_work @entry, ensures the entry is not
205 * currently in use.
206 */
irq_work_sync(struct irq_work * work)207 void irq_work_sync(struct irq_work *work)
208 {
209 lockdep_assert_irqs_enabled();
210
211 while (irq_work_is_busy(work))
212 cpu_relax();
213 }
214 EXPORT_SYMBOL_GPL(irq_work_sync);
215