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