Line data    Source code

       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 <asm/processor.h>
      22             : 
      23             : 
      24             : static DEFINE_PER_CPU(struct llist_head, raised_list);
      25             : static DEFINE_PER_CPU(struct llist_head, lazy_list);
      26             : 
      27             : /*
      28             :  * Claim the entry so that no one else will poke at it.
      29             :  */
      30          69 : static bool irq_work_claim(struct irq_work *work)
      31             : {
      32          69 :         int oflags;
      33             : 
      34          69 :         oflags = atomic_fetch_or(IRQ_WORK_CLAIMED | CSD_TYPE_IRQ_WORK, &work->node.a_flags);
      35             :         /*
      36             :          * If the work is already pending, no need to raise the IPI.
      37             :          * The pairing smp_mb() in irq_work_single() makes sure
      38             :          * everything we did before is visible.
      39             :          */
      40          69 :         if (oflags & IRQ_WORK_PENDING)
      41             :                 return false;
      42             :         return true;
      43             : }
      44             : 
      45           0 : void __weak arch_irq_work_raise(void)
      46             : {
      47             :         /*
      48             :          * Lame architectures will get the timer tick callback
      49             :          */
      50           0 : }
      51             : 
      52             : /* Enqueue on current CPU, work must already be claimed and preempt disabled */
      53          38 : static void __irq_work_queue_local(struct irq_work *work)
      54             : {
      55             :         /* If the work is "lazy", handle it from next tick if any */
      56          38 :         if (atomic_read(&work->node.a_flags) & IRQ_WORK_LAZY) {
      57          76 :                 if (llist_add(&work->node.llist, this_cpu_ptr(&lazy_list)) &&
      58          38 :                     tick_nohz_tick_stopped())
      59           0 :                         arch_irq_work_raise();
      60             :         } else {
      61           0 :                 if (llist_add(&work->node.llist, this_cpu_ptr(&raised_list)))
      62           0 :                         arch_irq_work_raise();
      63             :         }
      64          38 : }
      65             : 
      66             : /* Enqueue the irq work @work on the current CPU */
      67          69 : bool irq_work_queue(struct irq_work *work)
      68             : {
      69             :         /* Only queue if not already pending */
      70          69 :         if (!irq_work_claim(work))
      71             :                 return false;
      72             : 
      73             :         /* Queue the entry and raise the IPI if needed. */
      74          38 :         preempt_disable();
      75          38 :         __irq_work_queue_local(work);
      76          38 :         preempt_enable();
      77             : 
      78          38 :         return true;
      79             : }
      80             : EXPORT_SYMBOL_GPL(irq_work_queue);
      81             : 
      82             : /*
      83             :  * Enqueue the irq_work @work on @cpu unless it's already pending
      84             :  * somewhere.
      85             :  *
      86             :  * Can be re-enqueued while the callback is still in progress.
      87             :  */
      88           0 : bool irq_work_queue_on(struct irq_work *work, int cpu)
      89             : {
      90             : #ifndef CONFIG_SMP
      91             :         return irq_work_queue(work);
      92             : 
      93             : #else /* CONFIG_SMP: */
      94             :         /* All work should have been flushed before going offline */
      95           0 :         WARN_ON_ONCE(cpu_is_offline(cpu));
      96             : 
      97             :         /* Only queue if not already pending */
      98           0 :         if (!irq_work_claim(work))
      99             :                 return false;
     100             : 
     101           0 :         preempt_disable();
     102           0 :         if (cpu != smp_processor_id()) {
     103             :                 /* Arch remote IPI send/receive backend aren't NMI safe */
     104           0 :                 WARN_ON_ONCE(in_nmi());
     105           0 :                 __smp_call_single_queue(cpu, &work->node.llist);
     106             :         } else {
     107           0 :                 __irq_work_queue_local(work);
     108             :         }
     109           0 :         preempt_enable();
     110             : 
     111           0 :         return true;
     112             : #endif /* CONFIG_SMP */
     113             : }
     114             : 
     115             : 
     116        1458 : bool irq_work_needs_cpu(void)
     117             : {
     118        1458 :         struct llist_head *raised, *lazy;
     119             : 
     120        1458 :         raised = this_cpu_ptr(&raised_list);
     121        1459 :         lazy = this_cpu_ptr(&lazy_list);
     122             : 
     123        1460 :         if (llist_empty(raised) || arch_irq_work_has_interrupt())
     124        1460 :                 if (llist_empty(lazy))
     125             :                         return false;
     126             : 
     127             :         /* All work should have been flushed before going offline */
     128           0 :         WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
     129             : 
     130             :         return true;
     131             : }
     132             : 
     133          38 : void irq_work_single(void *arg)
     134             : {
     135          38 :         struct irq_work *work = arg;
     136          38 :         int flags;
     137             : 
     138             :         /*
     139             :          * Clear the PENDING bit, after this point the @work can be re-used.
     140             :          * The PENDING bit acts as a lock, and we own it, so we can clear it
     141             :          * without atomic ops.
     142             :          */
     143          38 :         flags = atomic_read(&work->node.a_flags);
     144          38 :         flags &= ~IRQ_WORK_PENDING;
     145          38 :         atomic_set(&work->node.a_flags, flags);
     146             : 
     147             :         /*
     148             :          * See irq_work_claim().
     149             :          */
     150          38 :         smp_mb();
     151             : 
     152          38 :         lockdep_irq_work_enter(flags);
     153          38 :         work->func(work);
     154          38 :         lockdep_irq_work_exit(flags);
     155             : 
     156             :         /*
     157             :          * Clear the BUSY bit, if set, and return to the free state if no-one
     158             :          * else claimed it meanwhile.
     159             :          */
     160          38 :         (void)atomic_cmpxchg(&work->node.a_flags, flags, flags & ~IRQ_WORK_BUSY);
     161          38 : }
     162             : 
     163       31867 : static void irq_work_run_list(struct llist_head *list)
     164             : {
     165       31867 :         struct irq_work *work, *tmp;
     166       31867 :         struct llist_node *llnode;
     167             : 
     168       31867 :         BUG_ON(!irqs_disabled());
     169             : 
     170       31913 :         if (llist_empty(list))
     171             :                 return;
     172             : 
     173          38 :         llnode = llist_del_all(list);
     174         114 :         llist_for_each_entry_safe(work, tmp, llnode, node.llist)
     175          38 :                 irq_work_single(work);
     176             : }
     177             : 
     178             : /*
     179             :  * hotplug calls this through:
     180             :  *  hotplug_cfd() -> flush_smp_call_function_queue()
     181             :  */
     182           0 : void irq_work_run(void)
     183             : {
     184           0 :         irq_work_run_list(this_cpu_ptr(&raised_list));
     185           0 :         irq_work_run_list(this_cpu_ptr(&lazy_list));
     186           0 : }
     187             : EXPORT_SYMBOL_GPL(irq_work_run);
     188             : 
     189       31420 : void irq_work_tick(void)
     190             : {
     191       31420 :         struct llist_head *raised = this_cpu_ptr(&raised_list);
     192             : 
     193       32066 :         if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
     194           0 :                 irq_work_run_list(raised);
     195       32066 :         irq_work_run_list(this_cpu_ptr(&lazy_list));
     196       31952 : }
     197             : 
     198             : /*
     199             :  * Synchronize against the irq_work @entry, ensures the entry is not
     200             :  * currently in use.
     201             :  */
     202           0 : void irq_work_sync(struct irq_work *work)
     203             : {
     204           0 :         lockdep_assert_irqs_enabled();
     205             : 
     206           0 :         while (irq_work_is_busy(work))
     207           0 :                 cpu_relax();
     208           0 : }
     209             : EXPORT_SYMBOL_GPL(irq_work_sync);