Loading...
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 */
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
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 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
45void __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 */
53static 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 */
67bool 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}
80EXPORT_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 */
88bool 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 kasan_record_aux_stack(work);
102
103 preempt_disable();
104 if (cpu != smp_processor_id()) {
105 /* Arch remote IPI send/receive backend aren't NMI safe */
106 WARN_ON_ONCE(in_nmi());
107 __smp_call_single_queue(cpu, &work->node.llist);
108 } else {
109 __irq_work_queue_local(work);
110 }
111 preempt_enable();
112
113 return true;
114#endif /* CONFIG_SMP */
115}
116
117
118bool irq_work_needs_cpu(void)
119{
120 struct llist_head *raised, *lazy;
121
122 raised = this_cpu_ptr(&raised_list);
123 lazy = this_cpu_ptr(&lazy_list);
124
125 if (llist_empty(raised) || arch_irq_work_has_interrupt())
126 if (llist_empty(lazy))
127 return false;
128
129 /* All work should have been flushed before going offline */
130 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
131
132 return true;
133}
134
135void irq_work_single(void *arg)
136{
137 struct irq_work *work = arg;
138 int flags;
139
140 /*
141 * Clear the PENDING bit, after this point the @work can be re-used.
142 * The PENDING bit acts as a lock, and we own it, so we can clear it
143 * without atomic ops.
144 */
145 flags = atomic_read(&work->node.a_flags);
146 flags &= ~IRQ_WORK_PENDING;
147 atomic_set(&work->node.a_flags, flags);
148
149 /*
150 * See irq_work_claim().
151 */
152 smp_mb();
153
154 lockdep_irq_work_enter(flags);
155 work->func(work);
156 lockdep_irq_work_exit(flags);
157
158 /*
159 * Clear the BUSY bit, if set, and return to the free state if no-one
160 * else claimed it meanwhile.
161 */
162 (void)atomic_cmpxchg(&work->node.a_flags, flags, flags & ~IRQ_WORK_BUSY);
163}
164
165static void irq_work_run_list(struct llist_head *list)
166{
167 struct irq_work *work, *tmp;
168 struct llist_node *llnode;
169
170 BUG_ON(!irqs_disabled());
171
172 if (llist_empty(list))
173 return;
174
175 llnode = llist_del_all(list);
176 llist_for_each_entry_safe(work, tmp, llnode, node.llist)
177 irq_work_single(work);
178}
179
180/*
181 * hotplug calls this through:
182 * hotplug_cfd() -> flush_smp_call_function_queue()
183 */
184void irq_work_run(void)
185{
186 irq_work_run_list(this_cpu_ptr(&raised_list));
187 irq_work_run_list(this_cpu_ptr(&lazy_list));
188}
189EXPORT_SYMBOL_GPL(irq_work_run);
190
191void irq_work_tick(void)
192{
193 struct llist_head *raised = this_cpu_ptr(&raised_list);
194
195 if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
196 irq_work_run_list(raised);
197 irq_work_run_list(this_cpu_ptr(&lazy_list));
198}
199
200/*
201 * Synchronize against the irq_work @entry, ensures the entry is not
202 * currently in use.
203 */
204void irq_work_sync(struct irq_work *work)
205{
206 lockdep_assert_irqs_enabled();
207
208 while (irq_work_is_busy(work))
209 cpu_relax();
210}
211EXPORT_SYMBOL_GPL(irq_work_sync);