block/blk-iopoll.c

   1 /*
   2  * Functions related to interrupt-poll handling in the block layer. This
   3  * is similar to NAPI for network devices.
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/module.h>
   7 #include <linux/init.h>
   8 #include <linux/bio.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/interrupt.h>
  11 #include <linux/cpu.h>
  12 #include <linux/blk-iopoll.h>
  13 #include <linux/delay.h>
  14
  15 #include "blk.h"
  16
  17 int blk_iopoll_enabled = 1;
  18 EXPORT_SYMBOL(blk_iopoll_enabled);
  19
  20 static DEFINE_PER_CPU(struct list_head, blk_cpu_iopoll);
  21
  22 /**
  23  * blk_iopoll_sched - Schedule a run of the iopoll handler
  24  * @iop:      The parent iopoll structure
  25  *
  26  * Description:
  27  *     Add this blk_iopoll structure to the pending poll list and trigger the
  28  *     raise of the blk iopoll softirq. The driver must already have gotten a
  29  *     succesful return from blk_iopoll_sched_prep() before calling this.
  30  **/
  31 void blk_iopoll_sched(struct blk_iopoll *iop)
  32 {
  33         unsigned long flags;
  34
  35         local_irq_save(flags);
  36         list_add_tail(&iop->list, &__get_cpu_var(blk_cpu_iopoll));
  37         __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
  38         local_irq_restore(flags);
  39 }
  40 EXPORT_SYMBOL(blk_iopoll_sched);
  41
  42 /**
  43  * __blk_iopoll_complete - Mark this @iop as un-polled again
  44  * @iop:      The parent iopoll structure
  45  *
  46  * Description:
  47  *     See blk_iopoll_complete(). This function must be called with interrupts
  48  *     disabled.
  49  **/
  50 void __blk_iopoll_complete(struct blk_iopoll *iop)
  51 {
  52         list_del(&iop->list);
  53         smp_mb__before_clear_bit();
  54         clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
  55 }
  56 EXPORT_SYMBOL(__blk_iopoll_complete);
  57
  58 /**
  59  * blk_iopoll_complete - Mark this @iop as un-polled again
  60  * @iop:      The parent iopoll structure
  61  *
  62  * Description:
  63  *     If a driver consumes less than the assigned budget in its run of the
  64  *     iopoll handler, it'll end the polled mode by calling this function. The
  65  *     iopoll handler will not be invoked again before blk_iopoll_sched_prep()
  66  *     is called.
  67  **/
  68 void blk_iopoll_complete(struct blk_iopoll *iopoll)
  69 {
  70         unsigned long flags;
  71
  72         local_irq_save(flags);
  73         __blk_iopoll_complete(iopoll);
  74         local_irq_restore(flags);
  75 }
  76 EXPORT_SYMBOL(blk_iopoll_complete);
  77
  78 static void blk_iopoll_softirq(struct softirq_action *h)
  79 {
  80         struct list_head *list = &__get_cpu_var(blk_cpu_iopoll);
  81         unsigned long start_time = jiffies;
  82         int rearm = 0, budget = 64;
  83
  84         local_irq_disable();
  85
  86         while (!list_empty(list)) {
  87                 struct blk_iopoll *iop;
  88                 int work, weight;
  89
  90                 /*
  91                  * If softirq window is exhausted then punt.
  92                  */
  93                 if (budget <= 0 || time_after(jiffies, start_time)) {
  94                         rearm = 1;
  95                         break;
  96                 }
  97
  98                 local_irq_enable();
  99
 100                 /* Even though interrupts have been re-enabled, this
 101                  * access is safe because interrupts can only add new
 102                  * entries to the tail of this list, and only ->poll()
 103                  * calls can remove this head entry from the list.
 104                  */
 105                 iop = list_entry(list->next, struct blk_iopoll, list);
 106
 107                 weight = iop->weight;
 108                 work = 0;
 109                 if (test_bit(IOPOLL_F_SCHED, &iop->state))
 110                         work = iop->poll(iop, weight);
 111
 112                 budget -= work;
 113
 114                 local_irq_disable();
 115
 116                 /* Drivers must not modify the NAPI state if they
 117                  * consume the entire weight.  In such cases this code
 118                  * still "owns" the NAPI instance and therefore can
 119                  * move the instance around on the list at-will.
 120                  */
 121                 if (work >= weight) {
 122                         if (blk_iopoll_disable_pending(iop))
 123                                 __blk_iopoll_complete(iop);
 124                         else
 125                                 list_move_tail(&iop->list, list);
 126                 }
 127         }
 128
 129         if (rearm)
 130                 __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
 131
 132         local_irq_enable();
 133 }
 134
 135 /**
 136  * blk_iopoll_disable - Disable iopoll on this @iop
 137  * @iop:      The parent iopoll structure
 138  *
 139  * Description:
 140  *     Disable io polling and wait for any pending callbacks to have completed.
 141  **/
 142 void blk_iopoll_disable(struct blk_iopoll *iop)
 143 {
 144         set_bit(IOPOLL_F_DISABLE, &iop->state);
 145         while (test_and_set_bit(IOPOLL_F_SCHED, &iop->state))
 146                 msleep(1);
 147         clear_bit(IOPOLL_F_DISABLE, &iop->state);
 148 }
 149 EXPORT_SYMBOL(blk_iopoll_disable);
 150
 151 /**
 152  * blk_iopoll_enable - Enable iopoll on this @iop
 153  * @iop:      The parent iopoll structure
 154  *
 155  * Description:
 156  *     Enable iopoll on this @iop. Note that the handler run will not be
 157  *     scheduled, it will only mark it as active.
 158  **/
 159 void blk_iopoll_enable(struct blk_iopoll *iop)
 160 {
 161         BUG_ON(!test_bit(IOPOLL_F_SCHED, &iop->state));
 162         smp_mb__before_clear_bit();
 163         clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
 164 }
 165 EXPORT_SYMBOL(blk_iopoll_enable);
 166
 167 /**
 168  * blk_iopoll_init - Initialize this @iop
 169  * @iop:      The parent iopoll structure
 170  * @weight:   The default weight (or command completion budget)
 171  * @poll_fn:  The handler to invoke
 172  *
 173  * Description:
 174  *     Initialize this blk_iopoll structure. Before being actively used, the
 175  *     driver must call blk_iopoll_enable().
 176  **/
 177 void blk_iopoll_init(struct blk_iopoll *iop, int weight, blk_iopoll_fn *poll_fn)
 178 {
 179         memset(iop, 0, sizeof(*iop));
 180         INIT_LIST_HEAD(&iop->list);
 181         iop->weight = weight;
 182         iop->poll = poll_fn;
 183         set_bit(IOPOLL_F_SCHED, &iop->state);
 184 }
 185 EXPORT_SYMBOL(blk_iopoll_init);
 186
 187 static int __cpuinit blk_iopoll_cpu_notify(struct notifier_block *self,
 188                                           unsigned long action, void *hcpu)
 189 {
 190         /*
 191          * If a CPU goes away, splice its entries to the current CPU
 192          * and trigger a run of the softirq
 193          */
 194         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 195                 int cpu = (unsigned long) hcpu;
 196
 197                 local_irq_disable();
 198                 list_splice_init(&per_cpu(blk_cpu_iopoll, cpu),
 199                                  &__get_cpu_var(blk_cpu_iopoll));
 200                 raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
 201                 local_irq_enable();
 202         }
 203
 204         return NOTIFY_OK;
 205 }
 206
 207 static struct notifier_block __cpuinitdata blk_iopoll_cpu_notifier = {
 208         .notifier_call  = blk_iopoll_cpu_notify,
 209 };
 210
 211 static __init int blk_iopoll_setup(void)
 212 {
 213         int i;
 214
 215         for_each_possible_cpu(i)
 216                 INIT_LIST_HEAD(&per_cpu(blk_cpu_iopoll, i));
 217
 218         open_softirq(BLOCK_IOPOLL_SOFTIRQ, blk_iopoll_softirq);
 219         register_hotcpu_notifier(&blk_iopoll_cpu_notifier);
 220         return 0;
 221 }
 222 subsys_initcall(blk_iopoll_setup);