1 // SPDX-License-Identifier: GPL-2.0-only 1 // SPDX-License-Identifier: GPL-2.0-only
2 /* 2 /*
3 * Copyright (C) 2010 Red Hat, Inc., Peter Zij 3 * Copyright (C) 2010 Red Hat, Inc., Peter Zijlstra
4 * 4 *
5 * Provides a framework for enqueueing and run 5 * Provides a framework for enqueueing and running callbacks from hardirq
6 * context. The enqueueing is NMI-safe. 6 * context. The enqueueing is NMI-safe.
7 */ 7 */
8 8
9 #include <linux/bug.h> 9 #include <linux/bug.h>
10 #include <linux/kernel.h> 10 #include <linux/kernel.h>
11 #include <linux/export.h> 11 #include <linux/export.h>
12 #include <linux/irq_work.h> 12 #include <linux/irq_work.h>
13 #include <linux/percpu.h> 13 #include <linux/percpu.h>
14 #include <linux/hardirq.h> 14 #include <linux/hardirq.h>
15 #include <linux/irqflags.h> 15 #include <linux/irqflags.h>
16 #include <linux/sched.h> 16 #include <linux/sched.h>
17 #include <linux/tick.h> 17 #include <linux/tick.h>
18 #include <linux/cpu.h> 18 #include <linux/cpu.h>
19 #include <linux/notifier.h> 19 #include <linux/notifier.h>
20 #include <linux/smp.h> 20 #include <linux/smp.h>
21 #include <linux/smpboot.h> 21 #include <linux/smpboot.h>
22 #include <asm/processor.h> 22 #include <asm/processor.h>
23 #include <linux/kasan.h> 23 #include <linux/kasan.h>
24 24
25 #include <trace/events/ipi.h> 25 #include <trace/events/ipi.h>
26 26
27 static DEFINE_PER_CPU(struct llist_head, raise 27 static DEFINE_PER_CPU(struct llist_head, raised_list);
28 static DEFINE_PER_CPU(struct llist_head, lazy_ 28 static DEFINE_PER_CPU(struct llist_head, lazy_list);
29 static DEFINE_PER_CPU(struct task_struct *, ir 29 static DEFINE_PER_CPU(struct task_struct *, irq_workd);
30 30
31 static void wake_irq_workd(void) 31 static void wake_irq_workd(void)
32 { 32 {
33 struct task_struct *tsk = __this_cpu_r 33 struct task_struct *tsk = __this_cpu_read(irq_workd);
34 34
35 if (!llist_empty(this_cpu_ptr(&lazy_li 35 if (!llist_empty(this_cpu_ptr(&lazy_list)) && tsk)
36 wake_up_process(tsk); 36 wake_up_process(tsk);
37 } 37 }
38 38
39 #ifdef CONFIG_SMP 39 #ifdef CONFIG_SMP
40 static void irq_work_wake(struct irq_work *ent 40 static void irq_work_wake(struct irq_work *entry)
41 { 41 {
42 wake_irq_workd(); 42 wake_irq_workd();
43 } 43 }
44 44
45 static DEFINE_PER_CPU(struct irq_work, irq_wor 45 static DEFINE_PER_CPU(struct irq_work, irq_work_wakeup) =
46 IRQ_WORK_INIT_HARD(irq_work_wake); 46 IRQ_WORK_INIT_HARD(irq_work_wake);
47 #endif 47 #endif
48 48
49 static int irq_workd_should_run(unsigned int c 49 static int irq_workd_should_run(unsigned int cpu)
50 { 50 {
51 return !llist_empty(this_cpu_ptr(&lazy 51 return !llist_empty(this_cpu_ptr(&lazy_list));
52 } 52 }
53 53
54 /* 54 /*
55 * Claim the entry so that no one else will po 55 * Claim the entry so that no one else will poke at it.
56 */ 56 */
57 static bool irq_work_claim(struct irq_work *wo 57 static bool irq_work_claim(struct irq_work *work)
58 { 58 {
59 int oflags; 59 int oflags;
60 60
61 oflags = atomic_fetch_or(IRQ_WORK_CLAI 61 oflags = atomic_fetch_or(IRQ_WORK_CLAIMED | CSD_TYPE_IRQ_WORK, &work->node.a_flags);
62 /* 62 /*
63 * If the work is already pending, no 63 * If the work is already pending, no need to raise the IPI.
64 * The pairing smp_mb() in irq_work_si 64 * The pairing smp_mb() in irq_work_single() makes sure
65 * everything we did before is visible 65 * everything we did before is visible.
66 */ 66 */
67 if (oflags & IRQ_WORK_PENDING) 67 if (oflags & IRQ_WORK_PENDING)
68 return false; 68 return false;
69 return true; 69 return true;
70 } 70 }
71 71
72 void __weak arch_irq_work_raise(void) 72 void __weak arch_irq_work_raise(void)
73 { 73 {
74 /* 74 /*
75 * Lame architectures will get the tim 75 * Lame architectures will get the timer tick callback
76 */ 76 */
77 } 77 }
78 78
79 static __always_inline void irq_work_raise(str 79 static __always_inline void irq_work_raise(struct irq_work *work)
80 { 80 {
81 if (trace_ipi_send_cpu_enabled() && ar 81 if (trace_ipi_send_cpu_enabled() && arch_irq_work_has_interrupt())
82 trace_ipi_send_cpu(smp_process 82 trace_ipi_send_cpu(smp_processor_id(), _RET_IP_, work->func);
83 83
84 arch_irq_work_raise(); 84 arch_irq_work_raise();
85 } 85 }
86 86
87 /* Enqueue on current CPU, work must already b 87 /* Enqueue on current CPU, work must already be claimed and preempt disabled */
88 static void __irq_work_queue_local(struct irq_ 88 static void __irq_work_queue_local(struct irq_work *work)
89 { 89 {
90 struct llist_head *list; 90 struct llist_head *list;
91 bool rt_lazy_work = false; 91 bool rt_lazy_work = false;
92 bool lazy_work = false; 92 bool lazy_work = false;
93 int work_flags; 93 int work_flags;
94 94
95 work_flags = atomic_read(&work->node.a 95 work_flags = atomic_read(&work->node.a_flags);
96 if (work_flags & IRQ_WORK_LAZY) 96 if (work_flags & IRQ_WORK_LAZY)
97 lazy_work = true; 97 lazy_work = true;
98 else if (IS_ENABLED(CONFIG_PREEMPT_RT) 98 else if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
99 !(work_flags & IRQ_WORK_HARD_ 99 !(work_flags & IRQ_WORK_HARD_IRQ))
100 rt_lazy_work = true; 100 rt_lazy_work = true;
101 101
102 if (lazy_work || rt_lazy_work) 102 if (lazy_work || rt_lazy_work)
103 list = this_cpu_ptr(&lazy_list 103 list = this_cpu_ptr(&lazy_list);
104 else 104 else
105 list = this_cpu_ptr(&raised_li 105 list = this_cpu_ptr(&raised_list);
106 106
107 if (!llist_add(&work->node.llist, list 107 if (!llist_add(&work->node.llist, list))
108 return; 108 return;
109 109
110 /* If the work is "lazy", handle it fr 110 /* If the work is "lazy", handle it from next tick if any */
111 if (!lazy_work || tick_nohz_tick_stopp 111 if (!lazy_work || tick_nohz_tick_stopped())
112 irq_work_raise(work); 112 irq_work_raise(work);
113 } 113 }
114 114
115 /* Enqueue the irq work @work on the current C 115 /* Enqueue the irq work @work on the current CPU */
116 bool irq_work_queue(struct irq_work *work) 116 bool irq_work_queue(struct irq_work *work)
117 { 117 {
118 /* Only queue if not already pending * 118 /* Only queue if not already pending */
119 if (!irq_work_claim(work)) 119 if (!irq_work_claim(work))
120 return false; 120 return false;
121 121
122 /* Queue the entry and raise the IPI i 122 /* Queue the entry and raise the IPI if needed. */
123 preempt_disable(); 123 preempt_disable();
124 __irq_work_queue_local(work); 124 __irq_work_queue_local(work);
125 preempt_enable(); 125 preempt_enable();
126 126
127 return true; 127 return true;
128 } 128 }
129 EXPORT_SYMBOL_GPL(irq_work_queue); 129 EXPORT_SYMBOL_GPL(irq_work_queue);
130 130
131 /* 131 /*
132 * Enqueue the irq_work @work on @cpu unless i 132 * Enqueue the irq_work @work on @cpu unless it's already pending
133 * somewhere. 133 * somewhere.
134 * 134 *
135 * Can be re-enqueued while the callback is st 135 * Can be re-enqueued while the callback is still in progress.
136 */ 136 */
137 bool irq_work_queue_on(struct irq_work *work, 137 bool irq_work_queue_on(struct irq_work *work, int cpu)
138 { 138 {
139 #ifndef CONFIG_SMP 139 #ifndef CONFIG_SMP
140 return irq_work_queue(work); 140 return irq_work_queue(work);
141 141
142 #else /* CONFIG_SMP: */ 142 #else /* CONFIG_SMP: */
143 /* All work should have been flushed b 143 /* All work should have been flushed before going offline */
144 WARN_ON_ONCE(cpu_is_offline(cpu)); 144 WARN_ON_ONCE(cpu_is_offline(cpu));
145 145
146 /* Only queue if not already pending * 146 /* Only queue if not already pending */
147 if (!irq_work_claim(work)) 147 if (!irq_work_claim(work))
148 return false; 148 return false;
149 149
150 kasan_record_aux_stack_noalloc(work); 150 kasan_record_aux_stack_noalloc(work);
151 151
152 preempt_disable(); 152 preempt_disable();
153 if (cpu != smp_processor_id()) { 153 if (cpu != smp_processor_id()) {
154 /* Arch remote IPI send/receiv 154 /* Arch remote IPI send/receive backend aren't NMI safe */
155 WARN_ON_ONCE(in_nmi()); 155 WARN_ON_ONCE(in_nmi());
156 156
157 /* 157 /*
158 * On PREEMPT_RT the items whi 158 * On PREEMPT_RT the items which are not marked as
159 * IRQ_WORK_HARD_IRQ are added 159 * IRQ_WORK_HARD_IRQ are added to the lazy list and a HARD work
160 * item is used on the remote 160 * item is used on the remote CPU to wake the thread.
161 */ 161 */
162 if (IS_ENABLED(CONFIG_PREEMPT_ 162 if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
163 !(atomic_read(&work->node. 163 !(atomic_read(&work->node.a_flags) & IRQ_WORK_HARD_IRQ)) {
164 164
165 if (!llist_add(&work-> 165 if (!llist_add(&work->node.llist, &per_cpu(lazy_list, cpu)))
166 goto out; 166 goto out;
167 167
168 work = &per_cpu(irq_wo 168 work = &per_cpu(irq_work_wakeup, cpu);
169 if (!irq_work_claim(wo 169 if (!irq_work_claim(work))
170 goto out; 170 goto out;
171 } 171 }
172 172
173 __smp_call_single_queue(cpu, & 173 __smp_call_single_queue(cpu, &work->node.llist);
174 } else { 174 } else {
175 __irq_work_queue_local(work); 175 __irq_work_queue_local(work);
176 } 176 }
177 out: 177 out:
178 preempt_enable(); 178 preempt_enable();
179 179
180 return true; 180 return true;
181 #endif /* CONFIG_SMP */ 181 #endif /* CONFIG_SMP */
182 } 182 }
183 183
184 bool irq_work_needs_cpu(void) 184 bool irq_work_needs_cpu(void)
185 { 185 {
186 struct llist_head *raised, *lazy; 186 struct llist_head *raised, *lazy;
187 187
188 raised = this_cpu_ptr(&raised_list); 188 raised = this_cpu_ptr(&raised_list);
189 lazy = this_cpu_ptr(&lazy_list); 189 lazy = this_cpu_ptr(&lazy_list);
190 190
191 if (llist_empty(raised) || arch_irq_wo 191 if (llist_empty(raised) || arch_irq_work_has_interrupt())
192 if (llist_empty(lazy)) 192 if (llist_empty(lazy))
193 return false; 193 return false;
194 194
195 /* All work should have been flushed b 195 /* All work should have been flushed before going offline */
196 WARN_ON_ONCE(cpu_is_offline(smp_proces 196 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
197 197
198 return true; 198 return true;
199 } 199 }
200 200
201 void irq_work_single(void *arg) 201 void irq_work_single(void *arg)
202 { 202 {
203 struct irq_work *work = arg; 203 struct irq_work *work = arg;
204 int flags; 204 int flags;
205 205
206 /* 206 /*
207 * Clear the PENDING bit, after this p 207 * Clear the PENDING bit, after this point the @work can be re-used.
208 * The PENDING bit acts as a lock, and 208 * The PENDING bit acts as a lock, and we own it, so we can clear it
209 * without atomic ops. 209 * without atomic ops.
210 */ 210 */
211 flags = atomic_read(&work->node.a_flag 211 flags = atomic_read(&work->node.a_flags);
212 flags &= ~IRQ_WORK_PENDING; 212 flags &= ~IRQ_WORK_PENDING;
213 atomic_set(&work->node.a_flags, flags) 213 atomic_set(&work->node.a_flags, flags);
214 214
215 /* 215 /*
216 * See irq_work_claim(). 216 * See irq_work_claim().
217 */ 217 */
218 smp_mb(); 218 smp_mb();
219 219
220 lockdep_irq_work_enter(flags); 220 lockdep_irq_work_enter(flags);
221 work->func(work); 221 work->func(work);
222 lockdep_irq_work_exit(flags); 222 lockdep_irq_work_exit(flags);
223 223
224 /* 224 /*
225 * Clear the BUSY bit, if set, and ret 225 * Clear the BUSY bit, if set, and return to the free state if no-one
226 * else claimed it meanwhile. 226 * else claimed it meanwhile.
227 */ 227 */
228 (void)atomic_cmpxchg(&work->node.a_fla 228 (void)atomic_cmpxchg(&work->node.a_flags, flags, flags & ~IRQ_WORK_BUSY);
229 229
230 if ((IS_ENABLED(CONFIG_PREEMPT_RT) && 230 if ((IS_ENABLED(CONFIG_PREEMPT_RT) && !irq_work_is_hard(work)) ||
231 !arch_irq_work_has_interrupt()) 231 !arch_irq_work_has_interrupt())
232 rcuwait_wake_up(&work->irqwait 232 rcuwait_wake_up(&work->irqwait);
233 } 233 }
234 234
235 static void irq_work_run_list(struct llist_hea 235 static void irq_work_run_list(struct llist_head *list)
236 { 236 {
237 struct irq_work *work, *tmp; 237 struct irq_work *work, *tmp;
238 struct llist_node *llnode; 238 struct llist_node *llnode;
239 239
240 /* 240 /*
241 * On PREEMPT_RT IRQ-work which is not 241 * On PREEMPT_RT IRQ-work which is not marked as HARD will be processed
242 * in a per-CPU thread in preemptible 242 * in a per-CPU thread in preemptible context. Only the items which are
243 * marked as IRQ_WORK_HARD_IRQ will be 243 * marked as IRQ_WORK_HARD_IRQ will be processed in hardirq context.
244 */ 244 */
245 BUG_ON(!irqs_disabled() && !IS_ENABLED 245 BUG_ON(!irqs_disabled() && !IS_ENABLED(CONFIG_PREEMPT_RT));
246 246
247 if (llist_empty(list)) 247 if (llist_empty(list))
248 return; 248 return;
249 249
250 llnode = llist_del_all(list); 250 llnode = llist_del_all(list);
251 llist_for_each_entry_safe(work, tmp, l 251 llist_for_each_entry_safe(work, tmp, llnode, node.llist)
252 irq_work_single(work); 252 irq_work_single(work);
253 } 253 }
254 254
255 /* 255 /*
256 * hotplug calls this through: 256 * hotplug calls this through:
257 * hotplug_cfd() -> flush_smp_call_function_q 257 * hotplug_cfd() -> flush_smp_call_function_queue()
258 */ 258 */
259 void irq_work_run(void) 259 void irq_work_run(void)
260 { 260 {
261 irq_work_run_list(this_cpu_ptr(&raised 261 irq_work_run_list(this_cpu_ptr(&raised_list));
262 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 262 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
263 irq_work_run_list(this_cpu_ptr 263 irq_work_run_list(this_cpu_ptr(&lazy_list));
264 else 264 else
265 wake_irq_workd(); 265 wake_irq_workd();
266 } 266 }
267 EXPORT_SYMBOL_GPL(irq_work_run); 267 EXPORT_SYMBOL_GPL(irq_work_run);
268 268
269 void irq_work_tick(void) 269 void irq_work_tick(void)
270 { 270 {
271 struct llist_head *raised = this_cpu_p 271 struct llist_head *raised = this_cpu_ptr(&raised_list);
272 272
273 if (!llist_empty(raised) && !arch_irq_ 273 if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
274 irq_work_run_list(raised); 274 irq_work_run_list(raised);
275 275
276 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 276 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
277 irq_work_run_list(this_cpu_ptr 277 irq_work_run_list(this_cpu_ptr(&lazy_list));
278 else 278 else
279 wake_irq_workd(); 279 wake_irq_workd();
280 } 280 }
281 281
282 /* 282 /*
283 * Synchronize against the irq_work @entry, en 283 * Synchronize against the irq_work @entry, ensures the entry is not
284 * currently in use. 284 * currently in use.
285 */ 285 */
286 void irq_work_sync(struct irq_work *work) 286 void irq_work_sync(struct irq_work *work)
287 { 287 {
288 lockdep_assert_irqs_enabled(); 288 lockdep_assert_irqs_enabled();
289 might_sleep(); 289 might_sleep();
290 290
291 if ((IS_ENABLED(CONFIG_PREEMPT_RT) && 291 if ((IS_ENABLED(CONFIG_PREEMPT_RT) && !irq_work_is_hard(work)) ||
292 !arch_irq_work_has_interrupt()) { 292 !arch_irq_work_has_interrupt()) {
293 rcuwait_wait_event(&work->irqw 293 rcuwait_wait_event(&work->irqwait, !irq_work_is_busy(work),
294 TASK_UNINTE 294 TASK_UNINTERRUPTIBLE);
295 return; 295 return;
296 } 296 }
297 297
298 while (irq_work_is_busy(work)) 298 while (irq_work_is_busy(work))
299 cpu_relax(); 299 cpu_relax();
300 } 300 }
301 EXPORT_SYMBOL_GPL(irq_work_sync); 301 EXPORT_SYMBOL_GPL(irq_work_sync);
302 302
303 static void run_irq_workd(unsigned int cpu) 303 static void run_irq_workd(unsigned int cpu)
304 { 304 {
305 irq_work_run_list(this_cpu_ptr(&lazy_l 305 irq_work_run_list(this_cpu_ptr(&lazy_list));
306 } 306 }
307 307
308 static void irq_workd_setup(unsigned int cpu) 308 static void irq_workd_setup(unsigned int cpu)
309 { 309 {
310 sched_set_fifo_low(current); 310 sched_set_fifo_low(current);
311 } 311 }
312 312
313 static struct smp_hotplug_thread irqwork_threa 313 static struct smp_hotplug_thread irqwork_threads = {
314 .store = &irq_workd, 314 .store = &irq_workd,
315 .setup = irq_workd_se 315 .setup = irq_workd_setup,
316 .thread_should_run = irq_workd_sh 316 .thread_should_run = irq_workd_should_run,
317 .thread_fn = run_irq_work 317 .thread_fn = run_irq_workd,
318 .thread_comm = "irq_work/%u 318 .thread_comm = "irq_work/%u",
319 }; 319 };
320 320
321 static __init int irq_work_init_threads(void) 321 static __init int irq_work_init_threads(void)
322 { 322 {
323 if (IS_ENABLED(CONFIG_PREEMPT_RT)) 323 if (IS_ENABLED(CONFIG_PREEMPT_RT))
324 BUG_ON(smpboot_register_percpu 324 BUG_ON(smpboot_register_percpu_thread(&irqwork_threads));
325 return 0; 325 return 0;
326 } 326 }
327 early_initcall(irq_work_init_threads); 327 early_initcall(irq_work_init_threads);
328 328
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.