kernel/irq_work.c

   1 /*
   2  * Copyright (C) 2010 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
   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 #ifdef CONFIG_SMP
  60 /*
  61  * Enqueue the irq_work @work on @cpu unless it's already pending
  62  * somewhere.
  63  *
  64  * Can be re-enqueued while the callback is still in progress.
  65  */
  66 bool irq_work_queue_on(struct irq_work *work, int cpu)
  67 {
  68         /* All work should have been flushed before going offline */
  69         WARN_ON_ONCE(cpu_is_offline(cpu));
  70
  71         /* Arch remote IPI send/receive backend aren't NMI safe */
  72         WARN_ON_ONCE(in_nmi());
  73
  74         /* Only queue if not already pending */
  75         if (!irq_work_claim(work))
  76                 return false;
  77
  78         if (llist_add(&work->llnode, &per_cpu(raised_list, cpu)))
  79                 arch_send_call_function_single_ipi(cpu);
  80
  81         return true;
  82 }
  83 EXPORT_SYMBOL_GPL(irq_work_queue_on);
  84 #endif
  85
  86 /* Enqueue the irq work @work on the current CPU */
  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;
  92
  93         /* Queue the entry and raise the IPI if needed. */
  94         preempt_disable();
  95
  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, &__get_cpu_var(lazy_list)) &&
  99                     tick_nohz_tick_stopped())
 100                         arch_irq_work_raise();
 101         } else {
 102                 if (llist_add(&work->llnode, &__get_cpu_var(raised_list)))
 103                         arch_irq_work_raise();
 104         }
 105
 106         preempt_enable();
 107
 108         return true;
 109 }
 110 EXPORT_SYMBOL_GPL(irq_work_queue);
 111
 112 bool irq_work_needs_cpu(void)
 113 {
 114         struct llist_head *raised, *lazy;
 115
 116         raised = &__get_cpu_var(raised_list);
 117         lazy = &__get_cpu_var(lazy_list);
 118
 119         if (llist_empty(raised) || arch_irq_work_has_interrupt())
 120                 if (llist_empty(lazy))
 121                         return false;
 122
 123         /* All work should have been flushed before going offline */
 124         WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
 125
 126         return true;
 127 }
 128
 129 static void irq_work_run_list(struct llist_head *list)
 130 {
 131         unsigned long flags;
 132         struct irq_work *work;
 133         struct llist_node *llnode;
 134
 135         BUG_ON(!irqs_disabled());
 136
 137         if (llist_empty(list))
 138                 return;
 139
 140         llnode = llist_del_all(list);
 141         while (llnode != NULL) {
 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         }
 163 }
 164
 165 /*
 166  * hotplug calls this through:
 167  *  hotplug_cfd() -> flush_smp_call_function_queue()
 168  */
 169 void irq_work_run(void)
 170 {
 171         irq_work_run_list(&__get_cpu_var(raised_list));
 172         irq_work_run_list(&__get_cpu_var(lazy_list));
 173 }
 174 EXPORT_SYMBOL_GPL(irq_work_run);
 175
 176 void irq_work_tick(void)
 177 {
 178         struct llist_head *raised = &__get_cpu_var(raised_list);
 179
 180         if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
 181                 irq_work_run_list(raised);
 182         irq_work_run_list(&__get_cpu_var(lazy_list));
 183 }
 184
 185 /*
 186  * Synchronize against the irq_work @entry, ensures the entry is not
 187  * currently in use.
 188  */
 189 void irq_work_sync(struct irq_work *work)
 190 {
 191         WARN_ON_ONCE(irqs_disabled());
 192
 193         while (work->flags & IRQ_WORK_BUSY)
 194                 cpu_relax();
 195 }
 196 EXPORT_SYMBOL_GPL(irq_work_sync);