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 <linux/smpboot.h>
22#include <asm/processor.h>
23#include <linux/kasan.h>
24
25static DEFINE_PER_CPU(struct llist_head, raised_list);
26static DEFINE_PER_CPU(struct llist_head, lazy_list);
27static DEFINE_PER_CPU(struct task_struct *, irq_workd);
28
29static void wake_irq_workd(void)
30{
31 struct task_struct *tsk = __this_cpu_read(irq_workd);
32
33 if (!llist_empty(this_cpu_ptr(&lazy_list)) && tsk)
34 wake_up_process(tsk);
35}
36
37#ifdef CONFIG_SMP
38static void irq_work_wake(struct irq_work *entry)
39{
40 wake_irq_workd();
41}
42
43static DEFINE_PER_CPU(struct irq_work, irq_work_wakeup) =
44 IRQ_WORK_INIT_HARD(irq_work_wake);
45#endif
46
47static int irq_workd_should_run(unsigned int cpu)
48{
49 return !llist_empty(this_cpu_ptr(&lazy_list));
50}
51
52/*
53 * Claim the entry so that no one else will poke at it.
54 */
55static bool irq_work_claim(struct irq_work *work)
56{
57 int oflags;
58
59 oflags = atomic_fetch_or(IRQ_WORK_CLAIMED | CSD_TYPE_IRQ_WORK, &work->node.a_flags);
60 /*
61 * If the work is already pending, no need to raise the IPI.
62 * The pairing smp_mb() in irq_work_single() makes sure
63 * everything we did before is visible.
64 */
65 if (oflags & IRQ_WORK_PENDING)
66 return false;
67 return true;
68}
69
70void __weak arch_irq_work_raise(void)
71{
72 /*
73 * Lame architectures will get the timer tick callback
74 */
75}
76
77/* Enqueue on current CPU, work must already be claimed and preempt disabled */
78static void __irq_work_queue_local(struct irq_work *work)
79{
80 struct llist_head *list;
81 bool rt_lazy_work = false;
82 bool lazy_work = false;
83 int work_flags;
84
85 work_flags = atomic_read(&work->node.a_flags);
86 if (work_flags & IRQ_WORK_LAZY)
87 lazy_work = true;
88 else if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
89 !(work_flags & IRQ_WORK_HARD_IRQ))
90 rt_lazy_work = true;
91
92 if (lazy_work || rt_lazy_work)
93 list = this_cpu_ptr(&lazy_list);
94 else
95 list = this_cpu_ptr(&raised_list);
96
97 if (!llist_add(&work->node.llist, list))
98 return;
99
100 /* If the work is "lazy", handle it from next tick if any */
101 if (!lazy_work || tick_nohz_tick_stopped())
102 arch_irq_work_raise();
103}
104
105/* Enqueue the irq work @work on the current CPU */
106bool irq_work_queue(struct irq_work *work)
107{
108 /* Only queue if not already pending */
109 if (!irq_work_claim(work))
110 return false;
111
112 /* Queue the entry and raise the IPI if needed. */
113 preempt_disable();
114 __irq_work_queue_local(work);
115 preempt_enable();
116
117 return true;
118}
119EXPORT_SYMBOL_GPL(irq_work_queue);
120
121/*
122 * Enqueue the irq_work @work on @cpu unless it's already pending
123 * somewhere.
124 *
125 * Can be re-enqueued while the callback is still in progress.
126 */
127bool irq_work_queue_on(struct irq_work *work, int cpu)
128{
129#ifndef CONFIG_SMP
130 return irq_work_queue(work);
131
132#else /* CONFIG_SMP: */
133 /* All work should have been flushed before going offline */
134 WARN_ON_ONCE(cpu_is_offline(cpu));
135
136 /* Only queue if not already pending */
137 if (!irq_work_claim(work))
138 return false;
139
140 kasan_record_aux_stack_noalloc(work);
141
142 preempt_disable();
143 if (cpu != smp_processor_id()) {
144 /* Arch remote IPI send/receive backend aren't NMI safe */
145 WARN_ON_ONCE(in_nmi());
146
147 /*
148 * On PREEMPT_RT the items which are not marked as
149 * IRQ_WORK_HARD_IRQ are added to the lazy list and a HARD work
150 * item is used on the remote CPU to wake the thread.
151 */
152 if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
153 !(atomic_read(&work->node.a_flags) & IRQ_WORK_HARD_IRQ)) {
154
155 if (!llist_add(&work->node.llist, &per_cpu(lazy_list, cpu)))
156 goto out;
157
158 work = &per_cpu(irq_work_wakeup, cpu);
159 if (!irq_work_claim(work))
160 goto out;
161 }
162
163 __smp_call_single_queue(cpu, &work->node.llist);
164 } else {
165 __irq_work_queue_local(work);
166 }
167out:
168 preempt_enable();
169
170 return true;
171#endif /* CONFIG_SMP */
172}
173
174bool irq_work_needs_cpu(void)
175{
176 struct llist_head *raised, *lazy;
177
178 raised = this_cpu_ptr(&raised_list);
179 lazy = this_cpu_ptr(&lazy_list);
180
181 if (llist_empty(raised) || arch_irq_work_has_interrupt())
182 if (llist_empty(lazy))
183 return false;
184
185 /* All work should have been flushed before going offline */
186 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
187
188 return true;
189}
190
191void irq_work_single(void *arg)
192{
193 struct irq_work *work = arg;
194 int flags;
195
196 /*
197 * Clear the PENDING bit, after this point the @work can be re-used.
198 * The PENDING bit acts as a lock, and we own it, so we can clear it
199 * without atomic ops.
200 */
201 flags = atomic_read(&work->node.a_flags);
202 flags &= ~IRQ_WORK_PENDING;
203 atomic_set(&work->node.a_flags, flags);
204
205 /*
206 * See irq_work_claim().
207 */
208 smp_mb();
209
210 lockdep_irq_work_enter(flags);
211 work->func(work);
212 lockdep_irq_work_exit(flags);
213
214 /*
215 * Clear the BUSY bit, if set, and return to the free state if no-one
216 * else claimed it meanwhile.
217 */
218 (void)atomic_cmpxchg(&work->node.a_flags, flags, flags & ~IRQ_WORK_BUSY);
219
220 if ((IS_ENABLED(CONFIG_PREEMPT_RT) && !irq_work_is_hard(work)) ||
221 !arch_irq_work_has_interrupt())
222 rcuwait_wake_up(&work->irqwait);
223}
224
225static void irq_work_run_list(struct llist_head *list)
226{
227 struct irq_work *work, *tmp;
228 struct llist_node *llnode;
229
230 /*
231 * On PREEMPT_RT IRQ-work which is not marked as HARD will be processed
232 * in a per-CPU thread in preemptible context. Only the items which are
233 * marked as IRQ_WORK_HARD_IRQ will be processed in hardirq context.
234 */
235 BUG_ON(!irqs_disabled() && !IS_ENABLED(CONFIG_PREEMPT_RT));
236
237 if (llist_empty(list))
238 return;
239
240 llnode = llist_del_all(list);
241 llist_for_each_entry_safe(work, tmp, llnode, node.llist)
242 irq_work_single(work);
243}
244
245/*
246 * hotplug calls this through:
247 * hotplug_cfd() -> flush_smp_call_function_queue()
248 */
249void irq_work_run(void)
250{
251 irq_work_run_list(this_cpu_ptr(&raised_list));
252 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
253 irq_work_run_list(this_cpu_ptr(&lazy_list));
254 else
255 wake_irq_workd();
256}
257EXPORT_SYMBOL_GPL(irq_work_run);
258
259void irq_work_tick(void)
260{
261 struct llist_head *raised = this_cpu_ptr(&raised_list);
262
263 if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
264 irq_work_run_list(raised);
265
266 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
267 irq_work_run_list(this_cpu_ptr(&lazy_list));
268 else
269 wake_irq_workd();
270}
271
272/*
273 * Synchronize against the irq_work @entry, ensures the entry is not
274 * currently in use.
275 */
276void irq_work_sync(struct irq_work *work)
277{
278 lockdep_assert_irqs_enabled();
279 might_sleep();
280
281 if ((IS_ENABLED(CONFIG_PREEMPT_RT) && !irq_work_is_hard(work)) ||
282 !arch_irq_work_has_interrupt()) {
283 rcuwait_wait_event(&work->irqwait, !irq_work_is_busy(work),
284 TASK_UNINTERRUPTIBLE);
285 return;
286 }
287
288 while (irq_work_is_busy(work))
289 cpu_relax();
290}
291EXPORT_SYMBOL_GPL(irq_work_sync);
292
293static void run_irq_workd(unsigned int cpu)
294{
295 irq_work_run_list(this_cpu_ptr(&lazy_list));
296}
297
298static void irq_workd_setup(unsigned int cpu)
299{
300 sched_set_fifo_low(current);
301}
302
303static struct smp_hotplug_thread irqwork_threads = {
304 .store = &irq_workd,
305 .setup = irq_workd_setup,
306 .thread_should_run = irq_workd_should_run,
307 .thread_fn = run_irq_workd,
308 .thread_comm = "irq_work/%u",
309};
310
311static __init int irq_work_init_threads(void)
312{
313 if (IS_ENABLED(CONFIG_PREEMPT_RT))
314 BUG_ON(smpboot_register_percpu_thread(&irqwork_threads));
315 return 0;
316}
317early_initcall(irq_work_init_threads);
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);