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 unsigned int blk_iopoll_budget __read_mostly = 256;
  21
  22 static DEFINE_PER_CPU(struct list_head, blk_cpu_iopoll);
  23
  24 /**
  25  * blk_iopoll_sched - Schedule a run of the iopoll handler
  26  * @iop:      The parent iopoll structure
  27  *
  28  * Description:
  29  *     Add this blk_iopoll structure to the pending poll list and trigger the
  30  *     raise of the blk iopoll softirq. The driver must already have gotten a
  31  *     succesful return from blk_iopoll_sched_prep() before calling this.
  32  **/
  33 void blk_iopoll_sched(struct blk_iopoll *iop)
  34 {
  35         unsigned long flags;
  36
  37         local_irq_save(flags);
  38         list_add_tail(&iop->list, &__get_cpu_var(blk_cpu_iopoll));
  39         __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
  40         local_irq_restore(flags);
  41 }
  42 EXPORT_SYMBOL(blk_iopoll_sched);
  43
  44 /**
  45  * __blk_iopoll_complete - Mark this @iop as un-polled again
  46  * @iop:      The parent iopoll structure
  47  *
  48  * Description:
  49  *     See blk_iopoll_complete(). This function must be called with interrupts
  50  *     disabled.
  51  **/
  52 void __blk_iopoll_complete(struct blk_iopoll *iop)
  53 {
  54         list_del(&iop->list);
  55         smp_mb__before_clear_bit();
  56         clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
  57 }
  58 EXPORT_SYMBOL(__blk_iopoll_complete);
  59
  60 /**
  61  * blk_iopoll_complete - Mark this @iop as un-polled again
  62  * @iop:      The parent iopoll structure
  63  *
  64  * Description:
  65  *     If a driver consumes less than the assigned budget in its run of the
  66  *     iopoll handler, it'll end the polled mode by calling this function. The
  67  *     iopoll handler will not be invoked again before blk_iopoll_sched_prep()
  68  *     is called.
  69  **/
  70 void blk_iopoll_complete(struct blk_iopoll *iopoll)
  71 {
  72         unsigned long flags;
  73
  74         local_irq_save(flags);
  75         __blk_iopoll_complete(iopoll);
  76         local_irq_restore(flags);
  77 }
  78 EXPORT_SYMBOL(blk_iopoll_complete);
  79
  80 static void blk_iopoll_softirq(struct softirq_action *h)
  81 {
  82         struct list_head *list = &__get_cpu_var(blk_cpu_iopoll);
  83         int rearm = 0, budget = blk_iopoll_budget;
  84         unsigned long start_time = jiffies;
  85
  86         local_irq_disable();
  87
  88         while (!list_empty(list)) {
  89                 struct blk_iopoll *iop;
  90                 int work, weight;
  91
  92                 /*
  93                  * If softirq window is exhausted then punt.
  94                  */
  95                 if (budget <= 0 || time_after(jiffies, start_time)) {
  96                         rearm = 1;
  97                         break;
  98                 }
  99
 100                 local_irq_enable();
 101
 102                 /* Even though interrupts have been re-enabled, this
 103                  * access is safe because interrupts can only add new
 104                  * entries to the tail of this list, and only ->poll()
 105                  * calls can remove this head entry from the list.
 106                  */
 107                 iop = list_entry(list->next, struct blk_iopoll, list);
 108
 109                 weight = iop->weight;
 110                 work = 0;
 111                 if (test_bit(IOPOLL_F_SCHED, &iop->state))
 112                         work = iop->poll(iop, weight);
 113
 114                 budget -= work;
 115
 116                 local_irq_disable();
 117
 118                 /* Drivers must not modify the NAPI state if they
 119                  * consume the entire weight.  In such cases this code
 120                  * still "owns" the NAPI instance and therefore can
 121                  * move the instance around on the list at-will.
 122                  */
 123                 if (work >= weight) {
 124                         if (blk_iopoll_disable_pending(iop))
 125                                 __blk_iopoll_complete(iop);
 126                         else
 127                                 list_move_tail(&iop->list, list);
 128                 }
 129         }
 130
 131         if (rearm)
 132                 __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
 133
 134         local_irq_enable();
 135 }
 136
 137 /**
 138  * blk_iopoll_disable - Disable iopoll on this @iop
 139  * @iop:      The parent iopoll structure
 140  *
 141  * Description:
 142  *     Disable io polling and wait for any pending callbacks to have completed.
 143  **/
 144 void blk_iopoll_disable(struct blk_iopoll *iop)
 145 {
 146         set_bit(IOPOLL_F_DISABLE, &iop->state);
 147         while (test_and_set_bit(IOPOLL_F_SCHED, &iop->state))
 148                 msleep(1);
 149         clear_bit(IOPOLL_F_DISABLE, &iop->state);
 150 }
 151 EXPORT_SYMBOL(blk_iopoll_disable);
 152
 153 /**
 154  * blk_iopoll_enable - Enable iopoll on this @iop
 155  * @iop:      The parent iopoll structure
 156  *
 157  * Description:
 158  *     Enable iopoll on this @iop. Note that the handler run will not be
 159  *     scheduled, it will only mark it as active.
 160  **/
 161 void blk_iopoll_enable(struct blk_iopoll *iop)
 162 {
 163         BUG_ON(!test_bit(IOPOLL_F_SCHED, &iop->state));
 164         smp_mb__before_clear_bit();
 165         clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
 166 }
 167 EXPORT_SYMBOL(blk_iopoll_enable);
 168
 169 /**
 170  * blk_iopoll_init - Initialize this @iop
 171  * @iop:      The parent iopoll structure
 172  * @weight:   The default weight (or command completion budget)
 173  * @poll_fn:  The handler to invoke
 174  *
 175  * Description:
 176  *     Initialize this blk_iopoll structure. Before being actively used, the
 177  *     driver must call blk_iopoll_enable().
 178  **/
 179 void blk_iopoll_init(struct blk_iopoll *iop, int weight, blk_iopoll_fn *poll_fn)
 180 {
 181         memset(iop, 0, sizeof(*iop));
 182         INIT_LIST_HEAD(&iop->list);
 183         iop->weight = weight;
 184         iop->poll = poll_fn;
 185         set_bit(IOPOLL_F_SCHED, &iop->state);
 186 }
 187 EXPORT_SYMBOL(blk_iopoll_init);
 188
 189 static int __cpuinit blk_iopoll_cpu_notify(struct notifier_block *self,
 190                                           unsigned long action, void *hcpu)
 191 {
 192         /*
 193          * If a CPU goes away, splice its entries to the current CPU
 194          * and trigger a run of the softirq
 195          */
 196         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 197                 int cpu = (unsigned long) hcpu;
 198
 199                 local_irq_disable();
 200                 list_splice_init(&per_cpu(blk_cpu_iopoll, cpu),
 201                                  &__get_cpu_var(blk_cpu_iopoll));
 202                 raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
 203                 local_irq_enable();
 204         }
 205
 206         return NOTIFY_OK;
 207 }
 208
 209 static struct notifier_block __cpuinitdata blk_iopoll_cpu_notifier = {
 210         .notifier_call  = blk_iopoll_cpu_notify,
 211 };
 212
 213 static __init int blk_iopoll_setup(void)
 214 {
 215         int i;
 216
 217         for_each_possible_cpu(i)
 218                 INIT_LIST_HEAD(&per_cpu(blk_cpu_iopoll, i));
 219
 220         open_softirq(BLOCK_IOPOLL_SOFTIRQ, blk_iopoll_softirq);
 221         register_hotcpu_notifier(&blk_iopoll_cpu_notifier);
 222         return 0;
 223 }
 224 subsys_initcall(blk_iopoll_setup);