kernel/irq_work.c

   1 /*
   2  * Copyright (C) 2010 Red Hat, Inc., Peter Zijlstra
   3  *
   4  * Provides a framework for enqueueing and running callbacks from hardirq
   5  * context. The enqueueing is NMI-safe.
   6  */
   7
   8 #include <linux/bug.h>
   9 #include <linux/kernel.h>
  10 #include <linux/export.h>
  11 #include <linux/irq_work.h>
  12 #include <linux/percpu.h>
  13 #include <linux/hardirq.h>
  14 #include <linux/irqflags.h>
  15 #include <linux/sched.h>
  16 #include <linux/tick.h>
  17 #include <linux/cpu.h>
  18 #include <linux/notifier.h>
  19 #include <linux/smp.h>
  20 #include <asm/processor.h>
  21
  22
  23 static DEFINE_PER_CPU(struct llist_head, raised_list);
  24 static DEFINE_PER_CPU(struct llist_head, lazy_list);
  25
  26 /*
  27  * Claim the entry so that no one else will poke at it.
  28  */
  29 static bool irq_work_claim(struct irq_work *work)
  30 {
  31         unsigned long flags, oflags, nflags;
  32
  33         /*
  34          * Start with our best wish as a premise but only trust any
  35          * flag value after cmpxchg() result.
  36          */
  37         flags = work->flags & ~IRQ_WORK_PENDING;
  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
  49         return true;
  50 }
  51
  52 void __weak arch_irq_work_raise(void)
  53 {
  54         /*
  55          * Lame architectures will get the timer tick callback
  56          */
  57 }
  58
  59 /* Enqueue on current CPU, work must already be claimed and preempt disabled */
  60 static void __irq_work_queue_local(struct irq_work *work)
  61 {
  62         /* If the work is "lazy", handle it from next tick if any */
  63         if (work->flags & IRQ_WORK_LAZY) {
  64                 if (llist_add(&work->llnode, this_cpu_ptr(&lazy_list)) &&
  65                     tick_nohz_tick_stopped())
  66                         arch_irq_work_raise();
  67         } else {
  68                 if (llist_add(&work->llnode, this_cpu_ptr(&raised_list)))
  69                         arch_irq_work_raise();
  70         }
  71 }
  72
  73 /* Enqueue the irq work @work on the current CPU */
  74 bool irq_work_queue(struct irq_work *work)
  75 {
  76         /* Only queue if not already pending */
  77         if (!irq_work_claim(work))
  78                 return false;
  79
  80         /* Queue the entry and raise the IPI if needed. */
  81         preempt_disable();
  82         __irq_work_queue_local(work);
  83         preempt_enable();
  84
  85         return true;
  86 }
  87 EXPORT_SYMBOL_GPL(irq_work_queue);
  88
  89 /*
  90  * Enqueue the irq_work @work on @cpu unless it's already pending
  91  * somewhere.
  92  *
  93  * Can be re-enqueued while the callback is still in progress.
  94  */
  95 bool irq_work_queue_on(struct irq_work *work, int cpu)
  96 {
  97 #ifndef CONFIG_SMP
  98         return irq_work_queue(work);
  99
 100 #else /* CONFIG_SMP: */
 101         /* All work should have been flushed before going offline */
 102         WARN_ON_ONCE(cpu_is_offline(cpu));
 103
 104         /* Only queue if not already pending */
 105         if (!irq_work_claim(work))
 106                 return false;
 107
 108         preempt_disable();
 109         if (cpu != smp_processor_id()) {
 110                 /* Arch remote IPI send/receive backend aren't NMI safe */
 111                 WARN_ON_ONCE(in_nmi());
 112                 if (llist_add(&work->llnode, &per_cpu(raised_list, cpu)))
 113                         arch_send_call_function_single_ipi(cpu);
 114         } else {
 115                 __irq_work_queue_local(work);
 116         }
 117         preempt_enable();
 118
 119         return true;
 120 #endif /* CONFIG_SMP */
 121 }
 122
 123
 124 bool irq_work_needs_cpu(void)
 125 {
 126         struct llist_head *raised, *lazy;
 127
 128         raised = this_cpu_ptr(&raised_list);
 129         lazy = this_cpu_ptr(&lazy_list);
 130
 131         if (llist_empty(raised) || arch_irq_work_has_interrupt())
 132                 if (llist_empty(lazy))
 133                         return false;
 134
 135         /* All work should have been flushed before going offline */
 136         WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
 137
 138         return true;
 139 }
 140
 141 static void irq_work_run_list(struct llist_head *list)
 142 {
 143         struct irq_work *work, *tmp;
 144         struct llist_node *llnode;
 145         unsigned long flags;
 146
 147         BUG_ON(!irqs_disabled());
 148
 149         if (llist_empty(list))
 150                 return;
 151
 152         llnode = llist_del_all(list);
 153         llist_for_each_entry_safe(work, tmp, llnode, llnode) {
 154                 /*
 155                  * Clear the PENDING bit, after this point the @work
 156                  * can be re-used.
 157                  * Make it immediately visible so that other CPUs trying
 158                  * to claim that work don't rely on us to handle their data
 159                  * while we are in the middle of the func.
 160                  */
 161                 flags = work->flags & ~IRQ_WORK_PENDING;
 162                 xchg(&work->flags, flags);
 163
 164                 work->func(work);
 165                 /*
 166                  * Clear the BUSY bit and return to the free state if
 167                  * no-one else claimed it meanwhile.
 168                  */
 169                 (void)cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
 170         }
 171 }
 172
 173 /*
 174  * hotplug calls this through:
 175  *  hotplug_cfd() -> flush_smp_call_function_queue()
 176  */
 177 void irq_work_run(void)
 178 {
 179         irq_work_run_list(this_cpu_ptr(&raised_list));
 180         irq_work_run_list(this_cpu_ptr(&lazy_list));
 181 }
 182 EXPORT_SYMBOL_GPL(irq_work_run);
 183
 184 void irq_work_tick(void)
 185 {
 186         struct llist_head *raised = this_cpu_ptr(&raised_list);
 187
 188         if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
 189                 irq_work_run_list(raised);
 190         irq_work_run_list(this_cpu_ptr(&lazy_list));
 191 }
 192
 193 /*
 194  * Synchronize against the irq_work @entry, ensures the entry is not
 195  * currently in use.
 196  */
 197 void irq_work_sync(struct irq_work *work)
 198 {
 199         lockdep_assert_irqs_enabled();
 200
 201         while (work->flags & IRQ_WORK_BUSY)
 202                 cpu_relax();
 203 }
 204 EXPORT_SYMBOL_GPL(irq_work_sync);