kernel/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 static unsigned int blk_iopoll_budget __read_mostly = 256;
  18
  19 static DEFINE_PER_CPU(struct list_head, blk_cpu_iopoll);
  20
  21 /**
  22  * blk_iopoll_sched - Schedule a run of the iopoll handler
  23  * @iop:      The parent iopoll structure
  24  *
  25  * Description:
  26  *     Add this blk_iopoll structure to the pending poll list and trigger the
  27  *     raise of the blk iopoll softirq. The driver must already have gotten a
  28  *     successful return from blk_iopoll_sched_prep() before calling this.
  29  **/
  30 void blk_iopoll_sched(struct blk_iopoll *iop)
  31 {
  32         unsigned long flags;
  33
  34         local_irq_save(flags);
  35         list_add_tail(&iop->list, this_cpu_ptr(&blk_cpu_iopoll));
  36         __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
  37         local_irq_restore(flags);
  38         preempt_check_resched_rt();
  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_atomic();
  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 *iop)
  69 {
  70         unsigned long flags;
  71
  72         local_irq_save(flags);
  73         __blk_iopoll_complete(iop);
  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 = this_cpu_ptr(&blk_cpu_iopoll);
  81         int rearm = 0, budget = blk_iopoll_budget;
  82         unsigned long start_time = jiffies;
  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                 /*
 117                  * Drivers must not modify the iopoll state, if they
 118                  * consume their assigned weight (or more, some drivers can't
 119                  * easily just stop processing, they have to complete an
 120                  * entire mask of commands).In such cases this code
 121                  * still "owns" the iopoll instance and therefore can
 122                  * move the instance around on the list at-will.
 123                  */
 124                 if (work >= weight) {
 125                         if (blk_iopoll_disable_pending(iop))
 126                                 __blk_iopoll_complete(iop);
 127                         else
 128                                 list_move_tail(&iop->list, list);
 129                 }
 130         }
 131
 132         if (rearm)
 133                 __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
 134
 135         local_irq_enable();
 136         preempt_check_resched_rt();
 137 }
 138
 139 /**
 140  * blk_iopoll_disable - Disable iopoll on this @iop
 141  * @iop:      The parent iopoll structure
 142  *
 143  * Description:
 144  *     Disable io polling and wait for any pending callbacks to have completed.
 145  **/
 146 void blk_iopoll_disable(struct blk_iopoll *iop)
 147 {
 148         set_bit(IOPOLL_F_DISABLE, &iop->state);
 149         while (test_and_set_bit(IOPOLL_F_SCHED, &iop->state))
 150                 msleep(1);
 151         clear_bit(IOPOLL_F_DISABLE, &iop->state);
 152 }
 153 EXPORT_SYMBOL(blk_iopoll_disable);
 154
 155 /**
 156  * blk_iopoll_enable - Enable iopoll on this @iop
 157  * @iop:      The parent iopoll structure
 158  *
 159  * Description:
 160  *     Enable iopoll on this @iop. Note that the handler run will not be
 161  *     scheduled, it will only mark it as active.
 162  **/
 163 void blk_iopoll_enable(struct blk_iopoll *iop)
 164 {
 165         BUG_ON(!test_bit(IOPOLL_F_SCHED, &iop->state));
 166         smp_mb__before_atomic();
 167         clear_bit_unlock(IOPOLL_F_SCHED, &iop->state);
 168 }
 169 EXPORT_SYMBOL(blk_iopoll_enable);
 170
 171 /**
 172  * blk_iopoll_init - Initialize this @iop
 173  * @iop:      The parent iopoll structure
 174  * @weight:   The default weight (or command completion budget)
 175  * @poll_fn:  The handler to invoke
 176  *
 177  * Description:
 178  *     Initialize this blk_iopoll structure. Before being actively used, the
 179  *     driver must call blk_iopoll_enable().
 180  **/
 181 void blk_iopoll_init(struct blk_iopoll *iop, int weight, blk_iopoll_fn *poll_fn)
 182 {
 183         memset(iop, 0, sizeof(*iop));
 184         INIT_LIST_HEAD(&iop->list);
 185         iop->weight = weight;
 186         iop->poll = poll_fn;
 187         set_bit(IOPOLL_F_SCHED, &iop->state);
 188 }
 189 EXPORT_SYMBOL(blk_iopoll_init);
 190
 191 static int blk_iopoll_cpu_notify(struct notifier_block *self,
 192                                  unsigned long action, void *hcpu)
 193 {
 194         /*
 195          * If a CPU goes away, splice its entries to the current CPU
 196          * and trigger a run of the softirq
 197          */
 198         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 199                 int cpu = (unsigned long) hcpu;
 200
 201                 local_irq_disable();
 202                 list_splice_init(&per_cpu(blk_cpu_iopoll, cpu),
 203                                  this_cpu_ptr(&blk_cpu_iopoll));
 204                 __raise_softirq_irqoff(BLOCK_IOPOLL_SOFTIRQ);
 205                 local_irq_enable();
 206                 preempt_check_resched_rt();
 207         }
 208
 209         return NOTIFY_OK;
 210 }
 211
 212 static struct notifier_block blk_iopoll_cpu_notifier = {
 213         .notifier_call  = blk_iopoll_cpu_notify,
 214 };
 215
 216 static __init int blk_iopoll_setup(void)
 217 {
 218         int i;
 219
 220         for_each_possible_cpu(i)
 221                 INIT_LIST_HEAD(&per_cpu(blk_cpu_iopoll, i));
 222
 223         open_softirq(BLOCK_IOPOLL_SOFTIRQ, blk_iopoll_softirq);
 224         register_hotcpu_notifier(&blk_iopoll_cpu_notifier);
 225         return 0;
 226 }
 227 subsys_initcall(blk_iopoll_setup);