projects/linux·blob master · refs

342 lines · 8.1 KiB

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
25#include <trace/events/ipi.h>
26
27static DEFINE_PER_CPU(struct llist_head, raised_list);
28static DEFINE_PER_CPU(struct llist_head, lazy_list);
29static DEFINE_PER_CPU(struct task_struct *, irq_workd);
30
31static void wake_irq_workd(void)
32{
33 struct task_struct *tsk = __this_cpu_read(irq_workd);
34
35 if (!llist_empty(this_cpu_ptr(&lazy_list)) && tsk)
36 wake_up_process(tsk);
37}
38
39#ifdef CONFIG_SMP
40static void irq_work_wake(struct irq_work *entry)
41{
42 wake_irq_workd();
43}
44
45static DEFINE_PER_CPU(struct irq_work, irq_work_wakeup) =
46 IRQ_WORK_INIT_HARD(irq_work_wake);
47#endif
48
49static int irq_workd_should_run(unsigned int cpu)
50{
51 return !llist_empty(this_cpu_ptr(&lazy_list));
52}
53
54/*
55 * Claim the entry so that no one else will poke at it.
56 */
57static bool irq_work_claim(struct irq_work *work)
58{
59 int oflags;
60
61 oflags = atomic_fetch_or(IRQ_WORK_CLAIMED | CSD_TYPE_IRQ_WORK, &work->node.a_flags);
62 /*
63 * If the work is already pending, no need to raise the IPI.
64 * The pairing smp_mb() in irq_work_single() makes sure
65 * everything we did before is visible.
66 */
67 if (oflags & IRQ_WORK_PENDING)
68 return false;
69 return true;
70}
71
72void __weak arch_irq_work_raise(void)
73{
74 /*
75 * Lame architectures will get the timer tick callback
76 */
77}
78
79static __always_inline void irq_work_raise(struct irq_work *work)
80{
81 if (trace_ipi_send_cpu_enabled() && arch_irq_work_has_interrupt())
82 trace_call__ipi_send_cpu(smp_processor_id(), _RET_IP_, work->func);
83
84 arch_irq_work_raise();
85}
86
87/* Enqueue on current CPU, work must already be claimed and preempt disabled */
88static void __irq_work_queue_local(struct irq_work *work)
89{
90 struct llist_head *list;
91 bool rt_lazy_work = false;
92 bool lazy_work = false;
93 int work_flags;
94
95 work_flags = atomic_read(&work->node.a_flags);
96 if (work_flags & IRQ_WORK_LAZY)
97 lazy_work = true;
98 else if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
99 !(work_flags & IRQ_WORK_HARD_IRQ))
100 rt_lazy_work = true;
101
102 if (lazy_work || rt_lazy_work)
103 list = this_cpu_ptr(&lazy_list);
104 else
105 list = this_cpu_ptr(&raised_list);
106
107 if (!llist_add(&work->node.llist, list))
108 return;
109
110 /* If the work is "lazy", handle it from next tick if any */
111 if (!lazy_work || tick_nohz_tick_stopped())
112 irq_work_raise(work);
113}
114
115/* Enqueue the irq work @work on the current CPU */
116bool irq_work_queue(struct irq_work *work)
117{
118 /* Only queue if not already pending */
119 if (!irq_work_claim(work))
120 return false;
121
122 /* Queue the entry and raise the IPI if needed. */
123 preempt_disable();
124 __irq_work_queue_local(work);
125 preempt_enable();
126
127 return true;
128}
129EXPORT_SYMBOL_GPL(irq_work_queue);
130
131/*
132 * Enqueue the irq_work @work on @cpu unless it's already pending
133 * somewhere.
134 *
135 * Can be re-enqueued while the callback is still in progress.
136 */
137bool irq_work_queue_on(struct irq_work *work, int cpu)
138{
139#ifndef CONFIG_SMP
140 return irq_work_queue(work);
141
142#else /* CONFIG_SMP: */
143 /* All work should have been flushed before going offline */
144 WARN_ON_ONCE(cpu_is_offline(cpu));
145
146 /* Only queue if not already pending */
147 if (!irq_work_claim(work))
148 return false;
149
150 kasan_record_aux_stack(work);
151
152 preempt_disable();
153 if (cpu != smp_processor_id()) {
154 /* Arch remote IPI send/receive backend aren't NMI safe */
155 WARN_ON_ONCE(in_nmi());
156
157 /*
158 * On PREEMPT_RT the items which are not marked as
159 * IRQ_WORK_HARD_IRQ are added to the lazy list and a HARD work
160 * item is used on the remote CPU to wake the thread.
161 */
162 if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
163 !(atomic_read(&work->node.a_flags) & IRQ_WORK_HARD_IRQ)) {
164
165 if (!llist_add(&work->node.llist, &per_cpu(lazy_list, cpu)))
166 goto out;
167
168 work = &per_cpu(irq_work_wakeup, cpu);
169 if (!irq_work_claim(work))
170 goto out;
171 }
172
173 __smp_call_single_queue(cpu, &work->node.llist);
174 } else {
175 __irq_work_queue_local(work);
176 }
177out:
178 preempt_enable();
179
180 return true;
181#endif /* CONFIG_SMP */
182}
183
184bool irq_work_needs_cpu(void)
185{
186 struct llist_head *raised, *lazy;
187
188 raised = this_cpu_ptr(&raised_list);
189 lazy = this_cpu_ptr(&lazy_list);
190
191 if (llist_empty(raised) || arch_irq_work_has_interrupt())
192 if (llist_empty(lazy))
193 return false;
194
195 /* All work should have been flushed before going offline */
196 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
197
198 return true;
199}
200
201void irq_work_single(void *arg)
202{
203 struct irq_work *work = arg;
204 int flags;
205
206 /*
207 * Clear the PENDING bit, after this point the @work can be re-used.
208 * The PENDING bit acts as a lock, and we own it, so we can clear it
209 * without atomic ops.
210 */
211 flags = atomic_read(&work->node.a_flags);
212 flags &= ~IRQ_WORK_PENDING;
213 atomic_set(&work->node.a_flags, flags);
214
215 /*
216 * See irq_work_claim().
217 */
218 smp_mb();
219
220 lockdep_irq_work_enter(flags);
221 work->func(work);
222 lockdep_irq_work_exit(flags);
223
224 /*
225 * Clear the BUSY bit, if set, and return to the free state if no-one
226 * else claimed it meanwhile.
227 */
228 (void)atomic_cmpxchg(&work->node.a_flags, flags, flags & ~IRQ_WORK_BUSY);
229
230 if ((IS_ENABLED(CONFIG_PREEMPT_RT) && !irq_work_is_hard(work)) ||
231 !arch_irq_work_has_interrupt())
232 rcuwait_wake_up(&work->irqwait);
233}
234
235static void irq_work_run_list(struct llist_head *list)
236{
237 struct irq_work *work, *tmp;
238 struct llist_node *llnode;
239
240 /*
241 * On PREEMPT_RT IRQ-work which is not marked as HARD will be processed
242 * in a per-CPU thread in preemptible context. Only the items which are
243 * marked as IRQ_WORK_HARD_IRQ will be processed in hardirq context.
244 */
245 BUG_ON(!irqs_disabled() && !IS_ENABLED(CONFIG_PREEMPT_RT));
246
247 if (llist_empty(list))
248 return;
249
250 llnode = llist_del_all(list);
251 llist_for_each_entry_safe(work, tmp, llnode, node.llist)
252 irq_work_single(work);
253}
254
255/* CPU hotplug calls this through smpcfd_dying_cpu() */
256void irq_work_run(void)
257{
258 irq_work_run_list(this_cpu_ptr(&raised_list));
259 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
260 irq_work_run_list(this_cpu_ptr(&lazy_list));
261 else
262 wake_irq_workd();
263}
264EXPORT_SYMBOL_GPL(irq_work_run);
265
266void irq_work_run_cpu(unsigned int cpu)
267{
268 if (WARN_ON_ONCE(!cpumask_test_cpu(cpu, cpu_dying_mask)))
269 return;
270
271 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
272 return;
273
274 irq_work_run_list(per_cpu_ptr(&lazy_list, cpu));
275}
276
277void irq_work_tick(void)
278{
279 struct llist_head *raised = this_cpu_ptr(&raised_list);
280
281 if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
282 irq_work_run_list(raised);
283
284 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
285 irq_work_run_list(this_cpu_ptr(&lazy_list));
286 else
287 wake_irq_workd();
288}
289
290/*
291 * Synchronize against the irq_work @entry, ensures the entry is not
292 * currently in use.
293 */
294void irq_work_sync(struct irq_work *work)
295{
296 lockdep_assert_irqs_enabled();
297 might_sleep();
298
299 if ((IS_ENABLED(CONFIG_PREEMPT_RT) && !irq_work_is_hard(work)) ||
300 !arch_irq_work_has_interrupt()) {
301 rcuwait_wait_event(&work->irqwait, !irq_work_is_busy(work),
302 TASK_UNINTERRUPTIBLE);
303 /*
304 * Ensure irq_work_single() does not access @work
305 * after removing IRQ_WORK_BUSY. It is always
306 * accessed within a RCU-read section.
307 */
308 synchronize_rcu();
309 return;
310 }
311
312 while (irq_work_is_busy(work))
313 cpu_relax();
314}
315EXPORT_SYMBOL_GPL(irq_work_sync);
316
317static void run_irq_workd(unsigned int cpu)
318{
319 guard(rcu)();
320 irq_work_run_list(this_cpu_ptr(&lazy_list));
321}
322
323static void irq_workd_setup(unsigned int cpu)
324{
325 sched_set_fifo_low(current);
326}
327
328static struct smp_hotplug_thread irqwork_threads = {
329 .store = &irq_workd,
330 .setup = irq_workd_setup,
331 .thread_should_run = irq_workd_should_run,
332 .thread_fn = run_irq_workd,
333 .thread_comm = "irq_work/%u",
334};
335
336static __init int irq_work_init_threads(void)
337{
338 if (IS_ENABLED(CONFIG_PREEMPT_RT))
339 BUG_ON(smpboot_register_percpu_thread(&irqwork_threads));
340 return 0;
341}
342early_initcall(irq_work_init_threads);