the-tree/kernel/irq_work.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * Copyright (C) 2010 Red Hat, Inc., Peter Zijlstra
0004  *
0005  * Provides a framework for enqueueing and running callbacks from hardirq
0006  * context. The enqueueing is NMI-safe.
0007  */
0008
0009 #include <linux/bug.h>
0010 #include <linux/kernel.h>
0011 #include <linux/export.h>
0012 #include <linux/irq_work.h>
0013 #include <linux/percpu.h>
0014 #include <linux/hardirq.h>
0015 #include <linux/irqflags.h>
0016 #include <linux/sched.h>
0017 #include <linux/tick.h>
0018 #include <linux/cpu.h>
0019 #include <linux/notifier.h>
0020 #include <linux/smp.h>
0021 #include <linux/smpboot.h>
0022 #include <asm/processor.h>
0023 #include <linux/kasan.h>
0024
0025 static DEFINE_PER_CPU(struct llist_head, raised_list);
0026 static DEFINE_PER_CPU(struct llist_head, lazy_list);
0027 static DEFINE_PER_CPU(struct task_struct *, irq_workd);
0028
0029 static void wake_irq_workd(void)
0030 {
0031     struct task_struct *tsk = __this_cpu_read(irq_workd);
0032
0033     if (!llist_empty(this_cpu_ptr(&lazy_list)) && tsk)
0034         wake_up_process(tsk);
0035 }
0036
0037 #ifdef CONFIG_SMP
0038 static void irq_work_wake(struct irq_work *entry)
0039 {
0040     wake_irq_workd();
0041 }
0042
0043 static DEFINE_PER_CPU(struct irq_work, irq_work_wakeup) =
0044     IRQ_WORK_INIT_HARD(irq_work_wake);
0045 #endif
0046
0047 static int irq_workd_should_run(unsigned int cpu)
0048 {
0049     return !llist_empty(this_cpu_ptr(&lazy_list));
0050 }
0051
0052 /*
0053  * Claim the entry so that no one else will poke at it.
0054  */
0055 static bool irq_work_claim(struct irq_work *work)
0056 {
0057     int oflags;
0058
0059     oflags = atomic_fetch_or(IRQ_WORK_CLAIMED | CSD_TYPE_IRQ_WORK, &work->node.a_flags);
0060     /*
0061      * If the work is already pending, no need to raise the IPI.
0062      * The pairing smp_mb() in irq_work_single() makes sure
0063      * everything we did before is visible.
0064      */
0065     if (oflags & IRQ_WORK_PENDING)
0066         return false;
0067     return true;
0068 }
0069
0070 void __weak arch_irq_work_raise(void)
0071 {
0072     /*
0073      * Lame architectures will get the timer tick callback
0074      */
0075 }
0076
0077 /* Enqueue on current CPU, work must already be claimed and preempt disabled */
0078 static void __irq_work_queue_local(struct irq_work *work)
0079 {
0080     struct llist_head *list;
0081     bool rt_lazy_work = false;
0082     bool lazy_work = false;
0083     int work_flags;
0084
0085     work_flags = atomic_read(&work->node.a_flags);
0086     if (work_flags & IRQ_WORK_LAZY)
0087         lazy_work = true;
0088     else if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
0089          !(work_flags & IRQ_WORK_HARD_IRQ))
0090         rt_lazy_work = true;
0091
0092     if (lazy_work || rt_lazy_work)
0093         list = this_cpu_ptr(&lazy_list);
0094     else
0095         list = this_cpu_ptr(&raised_list);
0096
0097     if (!llist_add(&work->node.llist, list))
0098         return;
0099
0100     /* If the work is "lazy", handle it from next tick if any */
0101     if (!lazy_work || tick_nohz_tick_stopped())
0102         arch_irq_work_raise();
0103 }
0104
0105 /* Enqueue the irq work @work on the current CPU */
0106 bool irq_work_queue(struct irq_work *work)
0107 {
0108     /* Only queue if not already pending */
0109     if (!irq_work_claim(work))
0110         return false;
0111
0112     /* Queue the entry and raise the IPI if needed. */
0113     preempt_disable();
0114     __irq_work_queue_local(work);
0115     preempt_enable();
0116
0117     return true;
0118 }
0119 EXPORT_SYMBOL_GPL(irq_work_queue);
0120
0121 /*
0122  * Enqueue the irq_work @work on @cpu unless it's already pending
0123  * somewhere.
0124  *
0125  * Can be re-enqueued while the callback is still in progress.
0126  */
0127 bool irq_work_queue_on(struct irq_work *work, int cpu)
0128 {
0129 #ifndef CONFIG_SMP
0130     return irq_work_queue(work);
0131
0132 #else /* CONFIG_SMP: */
0133     /* All work should have been flushed before going offline */
0134     WARN_ON_ONCE(cpu_is_offline(cpu));
0135
0136     /* Only queue if not already pending */
0137     if (!irq_work_claim(work))
0138         return false;
0139
0140     kasan_record_aux_stack_noalloc(work);
0141
0142     preempt_disable();
0143     if (cpu != smp_processor_id()) {
0144         /* Arch remote IPI send/receive backend aren't NMI safe */
0145         WARN_ON_ONCE(in_nmi());
0146
0147         /*
0148          * On PREEMPT_RT the items which are not marked as
0149          * IRQ_WORK_HARD_IRQ are added to the lazy list and a HARD work
0150          * item is used on the remote CPU to wake the thread.
0151          */
0152         if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
0153             !(atomic_read(&work->node.a_flags) & IRQ_WORK_HARD_IRQ)) {
0154
0155             if (!llist_add(&work->node.llist, &per_cpu(lazy_list, cpu)))
0156                 goto out;
0157
0158             work = &per_cpu(irq_work_wakeup, cpu);
0159             if (!irq_work_claim(work))
0160                 goto out;
0161         }
0162
0163         __smp_call_single_queue(cpu, &work->node.llist);
0164     } else {
0165         __irq_work_queue_local(work);
0166     }
0167 out:
0168     preempt_enable();
0169
0170     return true;
0171 #endif /* CONFIG_SMP */
0172 }
0173
0174 bool irq_work_needs_cpu(void)
0175 {
0176     struct llist_head *raised, *lazy;
0177
0178     raised = this_cpu_ptr(&raised_list);
0179     lazy = this_cpu_ptr(&lazy_list);
0180
0181     if (llist_empty(raised) || arch_irq_work_has_interrupt())
0182         if (llist_empty(lazy))
0183             return false;
0184
0185     /* All work should have been flushed before going offline */
0186     WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
0187
0188     return true;
0189 }
0190
0191 void irq_work_single(void *arg)
0192 {
0193     struct irq_work *work = arg;
0194     int flags;
0195
0196     /*
0197      * Clear the PENDING bit, after this point the @work can be re-used.
0198      * The PENDING bit acts as a lock, and we own it, so we can clear it
0199      * without atomic ops.
0200      */
0201     flags = atomic_read(&work->node.a_flags);
0202     flags &= ~IRQ_WORK_PENDING;
0203     atomic_set(&work->node.a_flags, flags);
0204
0205     /*
0206      * See irq_work_claim().
0207      */
0208     smp_mb();
0209
0210     lockdep_irq_work_enter(flags);
0211     work->func(work);
0212     lockdep_irq_work_exit(flags);
0213
0214     /*
0215      * Clear the BUSY bit, if set, and return to the free state if no-one
0216      * else claimed it meanwhile.
0217      */
0218     (void)atomic_cmpxchg(&work->node.a_flags, flags, flags & ~IRQ_WORK_BUSY);
0219
0220     if ((IS_ENABLED(CONFIG_PREEMPT_RT) && !irq_work_is_hard(work)) ||
0221         !arch_irq_work_has_interrupt())
0222         rcuwait_wake_up(&work->irqwait);
0223 }
0224
0225 static void irq_work_run_list(struct llist_head *list)
0226 {
0227     struct irq_work *work, *tmp;
0228     struct llist_node *llnode;
0229
0230     /*
0231      * On PREEMPT_RT IRQ-work which is not marked as HARD will be processed
0232      * in a per-CPU thread in preemptible context. Only the items which are
0233      * marked as IRQ_WORK_HARD_IRQ will be processed in hardirq context.
0234      */
0235     BUG_ON(!irqs_disabled() && !IS_ENABLED(CONFIG_PREEMPT_RT));
0236
0237     if (llist_empty(list))
0238         return;
0239
0240     llnode = llist_del_all(list);
0241     llist_for_each_entry_safe(work, tmp, llnode, node.llist)
0242         irq_work_single(work);
0243 }
0244
0245 /*
0246  * hotplug calls this through:
0247  *  hotplug_cfd() -> flush_smp_call_function_queue()
0248  */
0249 void irq_work_run(void)
0250 {
0251     irq_work_run_list(this_cpu_ptr(&raised_list));
0252     if (!IS_ENABLED(CONFIG_PREEMPT_RT))
0253         irq_work_run_list(this_cpu_ptr(&lazy_list));
0254     else
0255         wake_irq_workd();
0256 }
0257 EXPORT_SYMBOL_GPL(irq_work_run);
0258
0259 void irq_work_tick(void)
0260 {
0261     struct llist_head *raised = this_cpu_ptr(&raised_list);
0262
0263     if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
0264         irq_work_run_list(raised);
0265
0266     if (!IS_ENABLED(CONFIG_PREEMPT_RT))
0267         irq_work_run_list(this_cpu_ptr(&lazy_list));
0268     else
0269         wake_irq_workd();
0270 }
0271
0272 /*
0273  * Synchronize against the irq_work @entry, ensures the entry is not
0274  * currently in use.
0275  */
0276 void irq_work_sync(struct irq_work *work)
0277 {
0278     lockdep_assert_irqs_enabled();
0279     might_sleep();
0280
0281     if ((IS_ENABLED(CONFIG_PREEMPT_RT) && !irq_work_is_hard(work)) ||
0282         !arch_irq_work_has_interrupt()) {
0283         rcuwait_wait_event(&work->irqwait, !irq_work_is_busy(work),
0284                    TASK_UNINTERRUPTIBLE);
0285         return;
0286     }
0287
0288     while (irq_work_is_busy(work))
0289         cpu_relax();
0290 }
0291 EXPORT_SYMBOL_GPL(irq_work_sync);
0292
0293 static void run_irq_workd(unsigned int cpu)
0294 {
0295     irq_work_run_list(this_cpu_ptr(&lazy_list));
0296 }
0297
0298 static void irq_workd_setup(unsigned int cpu)
0299 {
0300     sched_set_fifo_low(current);
0301 }
0302
0303 static struct smp_hotplug_thread irqwork_threads = {
0304     .store                  = &irq_workd,
0305     .setup          = irq_workd_setup,
0306     .thread_should_run      = irq_workd_should_run,
0307     .thread_fn              = run_irq_workd,
0308     .thread_comm            = "irq_work/%u",
0309 };
0310
0311 static __init int irq_work_init_threads(void)
0312 {
0313     if (IS_ENABLED(CONFIG_PREEMPT_RT))
0314         BUG_ON(smpboot_register_percpu_thread(&irqwork_threads));
0315     return 0;
0316 }
0317 early_initcall(irq_work_init_threads);