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_CLAIMED;
  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 /*
  60  * Enqueue the irq_work @work on @cpu unless it's already pending
  61  * somewhere.
  62  *
  63  * Can be re-enqueued while the callback is still in progress.
  64  */
  65 bool irq_work_queue_on(struct irq_work *work, int cpu)
  66 {
  67         /* All work should have been flushed before going offline */
  68         WARN_ON_ONCE(cpu_is_offline(cpu));
  69
  70 #ifdef CONFIG_SMP
  71
  72         /* Arch remote IPI send/receive backend aren't NMI safe */
  73         WARN_ON_ONCE(in_nmi());
  74
  75         /* Only queue if not already pending */
  76         if (!irq_work_claim(work))
  77                 return false;
  78
  79         if (llist_add(&work->llnode, &per_cpu(raised_list, cpu)))
  80                 arch_send_call_function_single_ipi(cpu);
  81
  82 #else /* #ifdef CONFIG_SMP */
  83         irq_work_queue(work);
  84 #endif /* #else #ifdef CONFIG_SMP */
  85
  86         return true;
  87 }
  88
  89 /* Enqueue the irq work @work on the current CPU */
  90 bool irq_work_queue(struct irq_work *work)
  91 {
  92         /* Only queue if not already pending */
  93         if (!irq_work_claim(work))
  94                 return false;
  95
  96         /* Queue the entry and raise the IPI if needed. */
  97         preempt_disable();
  98
  99         /* If the work is "lazy", handle it from next tick if any */
 100         if (work->flags & IRQ_WORK_LAZY) {
 101                 if (llist_add(&work->llnode, this_cpu_ptr(&lazy_list)) &&
 102                     tick_nohz_tick_stopped())
 103                         arch_irq_work_raise();
 104         } else {
 105                 if (llist_add(&work->llnode, this_cpu_ptr(&raised_list)))
 106                         arch_irq_work_raise();
 107         }
 108
 109         preempt_enable();
 110
 111         return true;
 112 }
 113 EXPORT_SYMBOL_GPL(irq_work_queue);
 114
 115 bool irq_work_needs_cpu(void)
 116 {
 117         struct llist_head *raised, *lazy;
 118
 119         raised = this_cpu_ptr(&raised_list);
 120         lazy = this_cpu_ptr(&lazy_list);
 121
 122         if (llist_empty(raised) || arch_irq_work_has_interrupt())
 123                 if (llist_empty(lazy))
 124                         return false;
 125
 126         /* All work should have been flushed before going offline */
 127         WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
 128
 129         return true;
 130 }
 131
 132 static void irq_work_run_list(struct llist_head *list)
 133 {
 134         struct irq_work *work, *tmp;
 135         struct llist_node *llnode;
 136         unsigned long flags;
 137
 138         BUG_ON(!irqs_disabled());
 139
 140         if (llist_empty(list))
 141                 return;
 142
 143         llnode = llist_del_all(list);
 144         llist_for_each_entry_safe(work, tmp, llnode, llnode) {
 145                 /*
 146                  * Clear the PENDING bit, after this point the @work
 147                  * can be re-used.
 148                  * Make it immediately visible so that other CPUs trying
 149                  * to claim that work don't rely on us to handle their data
 150                  * while we are in the middle of the func.
 151                  */
 152                 flags = work->flags & ~IRQ_WORK_PENDING;
 153                 xchg(&work->flags, flags);
 154
 155                 work->func(work);
 156                 /*
 157                  * Clear the BUSY bit and return to the free state if
 158                  * no-one else claimed it meanwhile.
 159                  */
 160                 (void)cmpxchg(&work->flags, flags, flags & ~IRQ_WORK_BUSY);
 161         }
 162 }
 163
 164 /*
 165  * hotplug calls this through:
 166  *  hotplug_cfd() -> flush_smp_call_function_queue()
 167  */
 168 void irq_work_run(void)
 169 {
 170         irq_work_run_list(this_cpu_ptr(&raised_list));
 171         irq_work_run_list(this_cpu_ptr(&lazy_list));
 172 }
 173 EXPORT_SYMBOL_GPL(irq_work_run);
 174
 175 void irq_work_tick(void)
 176 {
 177         struct llist_head *raised = this_cpu_ptr(&raised_list);
 178
 179         if (!llist_empty(raised) && !arch_irq_work_has_interrupt())
 180                 irq_work_run_list(raised);
 181         irq_work_run_list(this_cpu_ptr(&lazy_list));
 182 }
 183
 184 /*
 185  * Synchronize against the irq_work @entry, ensures the entry is not
 186  * currently in use.
 187  */
 188 void irq_work_sync(struct irq_work *work)
 189 {
 190         WARN_ON_ONCE(irqs_disabled());
 191
 192         while (work->flags & IRQ_WORK_BUSY)
 193                 cpu_relax();
 194 }
 195 EXPORT_SYMBOL_GPL(irq_work_sync);