Loading...
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
24static DEFINE_PER_CPU(struct llist_head, raised_list);
25static DEFINE_PER_CPU(struct llist_head, lazy_list);
26
27/*
28 * Claim the entry so that no one else will poke at it.
29 */
30static 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
53void __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 */
61static 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 */
75bool 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}
88EXPORT_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 */
96bool 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
125bool 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
142static 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 */
178void 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}
183EXPORT_SYMBOL_GPL(irq_work_run);
184
185void 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 */
198void 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}
205EXPORT_SYMBOL_GPL(irq_work_sync);
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
22static DEFINE_PER_CPU(struct llist_head, irq_work_list);
23static DEFINE_PER_CPU(int, irq_work_raised);
24
25/*
26 * Claim the entry so that no one else will poke at it.
27 */
28static 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
51void __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 */
64bool irq_work_queue(struct irq_work *work)
65{
66 /* Only queue if not already pending */
67 if (!irq_work_claim(work))
68 return false;
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 return true;
88}
89EXPORT_SYMBOL_GPL(irq_work_queue);
90
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
99 /* All work should have been flushed before going offline */
100 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
101
102 return true;
103}
104
105static void __irq_work_run(void)
106{
107 unsigned long flags;
108 struct irq_work *work;
109 struct llist_head *this_list;
110 struct llist_node *llnode;
111
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
120 this_list = &__get_cpu_var(irq_work_list);
121 if (llist_empty(this_list))
122 return;
123
124 BUG_ON(!irqs_disabled());
125
126 llnode = llist_del_all(this_list);
127 while (llnode != NULL) {
128 work = llist_entry(llnode, struct irq_work, llnode);
129
130 llnode = llist_next(llnode);
131
132 /*
133 * Clear the PENDING bit, after this point the @work
134 * can be re-used.
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.
138 */
139 flags = work->flags & ~IRQ_WORK_PENDING;
140 xchg(&work->flags, flags);
141
142 work->func(work);
143 /*
144 * Clear the BUSY bit and return to the free state if
145 * no-one else claimed it meanwhile.
146 */
147 (void)cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
148 }
149}
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}
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 */
166void irq_work_sync(struct irq_work *work)
167{
168 WARN_ON_ONCE(irqs_disabled());
169
170 while (work->flags & IRQ_WORK_BUSY)
171 cpu_relax();
172}
173EXPORT_SYMBOL_GPL(irq_work_sync);
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 */