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