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         if (llist_empty(raised) && llist_empty(lazy))
 119                 return false;
 120
 121         /* All work should have been flushed before going offline */
 122         WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
 123
 124         return true;
 125 }
 126
 127 static void irq_work_run_list(struct llist_head *list)
 128 {
 129         unsigned long flags;
 130         struct irq_work *work;
 131         struct llist_node *llnode;
 132
 133         BUG_ON(!irqs_disabled());
 134
 135         if (llist_empty(list))
 136                 return;
 137
 138         llnode = llist_del_all(list);
 139         while (llnode != NULL) {
 140                 work = llist_entry(llnode, struct irq_work, llnode);
 141
 142                 llnode = llist_next(llnode);
 143
 144                 /*
 145                  * Clear the PENDING bit, after this point the @work
 146                  * can be re-used.
 147                  * Make it immediately visible so that other CPUs trying
 148                  * to claim that work don't rely on us to handle their data
 149                  * while we are in the middle of the func.
 150                  */
 151                 flags = work->flags & ~IRQ_WORK_PENDING;
 152                 xchg(&work->flags, flags);
 153
 154                 work->func(work);
 155                 /*
 156                  * Clear the BUSY bit and return to the free state if
 157                  * no-one else claimed it meanwhile.
 158                  */
 159                 (void)cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
 160         }
 161 }
 162
 163 static void __irq_work_run(void)
 164 {
 165         irq_work_run_list(&__get_cpu_var(raised_list));
 166         irq_work_run_list(&__get_cpu_var(lazy_list));
 167 }
 168
 169 /*
 170  * Run the irq_work entries on this cpu. Requires to be ran from hardirq
 171  * context with local IRQs disabled.
 172  */
 173 void irq_work_run(void)
 174 {
 175         BUG_ON(!in_irq());
 176         __irq_work_run();
 177 }
 178 EXPORT_SYMBOL_GPL(irq_work_run);
 179
 180 /*
 181  * Synchronize against the irq_work @entry, ensures the entry is not
 182  * currently in use.
 183  */
 184 void irq_work_sync(struct irq_work *work)
 185 {
 186         WARN_ON_ONCE(irqs_disabled());
 187
 188         while (work->flags & IRQ_WORK_BUSY)
 189                 cpu_relax();
 190 }
 191 EXPORT_SYMBOL_GPL(irq_work_sync);
 192
 193 #ifdef CONFIG_HOTPLUG_CPU
 194 static int irq_work_cpu_notify(struct notifier_block *self,
 195                                unsigned long action, void *hcpu)
 196 {
 197         long cpu = (long)hcpu;
 198
 199         switch (action) {
 200         case CPU_DYING:
 201                 /* Called from stop_machine */
 202                 if (WARN_ON_ONCE(cpu != smp_processor_id()))
 203                         break;
 204                 __irq_work_run();
 205                 break;
 206         default:
 207                 break;
 208         }
 209         return NOTIFY_OK;
 210 }
 211
 212 static struct notifier_block cpu_notify;
 213
 214 static __init int irq_work_init_cpu_notifier(void)
 215 {
 216         cpu_notify.notifier_call = irq_work_cpu_notify;
 217         cpu_notify.priority = 0;
 218         register_cpu_notifier(&cpu_notify);
 219         return 0;
 220 }
 221 device_initcall(irq_work_init_cpu_notifier);
 222
 223 #endif /* CONFIG_HOTPLUG_CPU */