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/kernel.h>
9#include <linux/module.h>
10#include <linux/irq_work.h>
11#include <linux/hardirq.h>
12
13/*
14 * An entry can be in one of four states:
15 *
16 * free NULL, 0 -> {claimed} : free to be used
17 * claimed NULL, 3 -> {pending} : claimed to be enqueued
18 * pending next, 3 -> {busy} : queued, pending callback
19 * busy NULL, 2 -> {free, claimed} : callback in progress, can be claimed
20 *
21 * We use the lower two bits of the next pointer to keep PENDING and BUSY
22 * flags.
23 */
24
25#define IRQ_WORK_PENDING 1UL
26#define IRQ_WORK_BUSY 2UL
27#define IRQ_WORK_FLAGS 3UL
28
29static inline bool irq_work_is_set(struct irq_work *entry, int flags)
30{
31 return (unsigned long)entry->next & flags;
32}
33
34static inline struct irq_work *irq_work_next(struct irq_work *entry)
35{
36 unsigned long next = (unsigned long)entry->next;
37 next &= ~IRQ_WORK_FLAGS;
38 return (struct irq_work *)next;
39}
40
41static inline struct irq_work *next_flags(struct irq_work *entry, int flags)
42{
43 unsigned long next = (unsigned long)entry;
44 next |= flags;
45 return (struct irq_work *)next;
46}
47
48static DEFINE_PER_CPU(struct irq_work *, irq_work_list);
49
50/*
51 * Claim the entry so that no one else will poke at it.
52 */
53static bool irq_work_claim(struct irq_work *entry)
54{
55 struct irq_work *next, *nflags;
56
57 do {
58 next = entry->next;
59 if ((unsigned long)next & IRQ_WORK_PENDING)
60 return false;
61 nflags = next_flags(next, IRQ_WORK_FLAGS);
62 } while (cmpxchg(&entry->next, next, nflags) != next);
63
64 return true;
65}
66
67
68void __weak arch_irq_work_raise(void)
69{
70 /*
71 * Lame architectures will get the timer tick callback
72 */
73}
74
75/*
76 * Queue the entry and raise the IPI if needed.
77 */
78static void __irq_work_queue(struct irq_work *entry)
79{
80 struct irq_work *next;
81
82 preempt_disable();
83
84 do {
85 next = __this_cpu_read(irq_work_list);
86 /* Can assign non-atomic because we keep the flags set. */
87 entry->next = next_flags(next, IRQ_WORK_FLAGS);
88 } while (this_cpu_cmpxchg(irq_work_list, next, entry) != next);
89
90 /* The list was empty, raise self-interrupt to start processing. */
91 if (!irq_work_next(entry))
92 arch_irq_work_raise();
93
94 preempt_enable();
95}
96
97/*
98 * Enqueue the irq_work @entry, returns true on success, failure when the
99 * @entry was already enqueued by someone else.
100 *
101 * Can be re-enqueued while the callback is still in progress.
102 */
103bool irq_work_queue(struct irq_work *entry)
104{
105 if (!irq_work_claim(entry)) {
106 /*
107 * Already enqueued, can't do!
108 */
109 return false;
110 }
111
112 __irq_work_queue(entry);
113 return true;
114}
115EXPORT_SYMBOL_GPL(irq_work_queue);
116
117/*
118 * Run the irq_work entries on this cpu. Requires to be ran from hardirq
119 * context with local IRQs disabled.
120 */
121void irq_work_run(void)
122{
123 struct irq_work *list;
124
125 if (this_cpu_read(irq_work_list) == NULL)
126 return;
127
128 BUG_ON(!in_irq());
129 BUG_ON(!irqs_disabled());
130
131 list = this_cpu_xchg(irq_work_list, NULL);
132
133 while (list != NULL) {
134 struct irq_work *entry = list;
135
136 list = irq_work_next(list);
137
138 /*
139 * Clear the PENDING bit, after this point the @entry
140 * can be re-used.
141 */
142 entry->next = next_flags(NULL, IRQ_WORK_BUSY);
143 entry->func(entry);
144 /*
145 * Clear the BUSY bit and return to the free state if
146 * no-one else claimed it meanwhile.
147 */
148 (void)cmpxchg(&entry->next,
149 next_flags(NULL, IRQ_WORK_BUSY),
150 NULL);
151 }
152}
153EXPORT_SYMBOL_GPL(irq_work_run);
154
155/*
156 * Synchronize against the irq_work @entry, ensures the entry is not
157 * currently in use.
158 */
159void irq_work_sync(struct irq_work *entry)
160{
161 WARN_ON_ONCE(irqs_disabled());
162
163 while (irq_work_is_set(entry, IRQ_WORK_BUSY))
164 cpu_relax();
165}
166EXPORT_SYMBOL_GPL(irq_work_sync);
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);