lustre/kernel_patches/patches/raid5-merge-ios.patch

   1 diff -pur linux-2.6.9-67.orig/drivers/md/raid5.c linux-2.6.9-67/drivers/md/raid5.c
   2 --- linux-2.6.9-67.orig/drivers/md/raid5.c      2009-02-20 16:06:12.000000000 +0800
   3 +++ linux-2.6.9-67/drivers/md/raid5.c   2009-02-20 16:06:37.000000000 +0800
   4 @@ -852,6 +852,26 @@ static void add_stripe_bio (struct strip
   5         }
   6  }
   7
   8 +/*
   9 + * The whole idea is to collect all bio's and then issue them
  10 + * disk by disk to assist merging a bit -bzzz
  11 + */
  12 +static void raid5_flush_bios(raid5_conf_t *conf, struct bio *bios[], int raid_disks)
  13 +{
  14 +       struct bio *bio, *nbio;
  15 +       int i;
  16 +
  17 +       for (i = 0; i < raid_disks; i++) {
  18 +               bio = bios[i];
  19 +               while (bio) {
  20 +                       nbio = bio->bi_next;
  21 +                       bio->bi_next = NULL;
  22 +                       generic_make_request(bio);
  23 +                       bio = nbio;
  24 +               }
  25 +               bios[i] = NULL;
  26 +       }
  27 +}
  28
  29  /*
  30   * handle_stripe - do things to a stripe.
  31 @@ -871,7 +891,7 @@ static void add_stripe_bio (struct strip
  32   *
  33   */
  34
  35 -static void handle_stripe(struct stripe_head *sh)
  36 +static void handle_stripe(struct stripe_head *sh, struct bio *bios[])
  37  {
  38         raid5_conf_t *conf = sh->raid_conf;
  39         int disks = conf->raid_disks;
  40 @@ -1291,7 +1311,11 @@ static void handle_stripe(struct stripe_
  41                         bi->bi_size = STRIPE_SIZE;
  42                         bi->bi_next = NULL;
  43                         atomic_inc(&conf->out_reqs_in_queue);
  44 -                       generic_make_request(bi);
  45 +                       if (bios) {
  46 +                               bi->bi_next = bios[i];
  47 +                               bios[i] = bi;
  48 +                       } else
  49 +                               generic_make_request(bi);
  50                 } else {
  51                         PRINTK("skip op %ld on disc %d for sector %llu\n",
  52                                 bi->bi_rw, i, (unsigned long long)sh->sector);
  53 @@ -1416,6 +1440,7 @@ static int make_request (request_queue_t
  54         int sectors_per_chunk = conf->chunk_size >> 9;
  55         int stripes_per_chunk, sectors_per_block;
  56         int sectors_per_stripe;
  57 +       struct bio *bios[MD_SB_DISKS];
  58         int i, j;
  59
  60         if (unlikely(bio_barrier(bi))) {
  61 @@ -1450,6 +1475,7 @@ static int make_request (request_queue_t
  62         sector_div(block, sectors_per_block);
  63         sectors = bi->bi_size >> 9;
  64
  65 +       memset(&bios, 0, sizeof(bios));
  66  repeat:
  67         stripe = block * (sectors_per_block / data_disks);
  68         b_sector = stripe * data_disks;
  69 @@ -1469,9 +1495,17 @@ repeat:
  70                         new_sector = raid5_compute_sector(r_sector, raid_disks,
  71                                                         data_disks, &dd_idx,
  72                                                         &pd_idx, conf);
  73 -                       if (sh == NULL)
  74 -                               sh = get_active_stripe(conf, new_sector, pd_idx,
  75 -                                                       (bi->bi_rw&RWA_MASK));
  76 +                       if (sh == NULL) {
  77 +                               /* first, try to get stripe w/o blocking
  78 +                                * if we can't, then it's time to submit
  79 +                                * all collected bio's in order to free
  80 +                                * some space in the cache -bzzz */
  81 +                               sh = get_active_stripe(conf, new_sector, pd_idx, 1);
  82 +                               if (!sh && !(bi->bi_rw&RWA_MASK)) {
  83 +                                       raid5_flush_bios(conf, bios, raid_disks);
  84 +                                       sh = get_active_stripe(conf, new_sector, pd_idx, 0);
  85 +                               }
  86 +                       }
  87                         if (sh) {
  88                                 add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK));
  89                         } else {
  90 @@ -1491,7 +1525,7 @@ repeat:
  91                 }
  92                 if (sh) {
  93                         raid5_plug_device(conf);
  94 -                       handle_stripe(sh);
  95 +                       handle_stripe(sh, bios);
  96                         release_stripe(sh);
  97                         sh = NULL;
  98                 }
  99 @@ -1501,6 +1535,9 @@ repeat:
 100         if (sectors > 0)
 101                 goto repeat;
 102
 103 +       /* now flush all bio's */
 104 +       raid5_flush_bios(conf, bios, raid_disks);
 105 +
 106         spin_lock_irq(&conf->device_lock);
 107         if (--bi->bi_phys_segments == 0) {
 108                 int bytes = bi->bi_size;
 109 @@ -1556,7 +1593,7 @@ static int sync_request (mddev_t *mddev,
 110         clear_bit(STRIPE_INSYNC, &sh->state);
 111         spin_unlock(&sh->lock);
 112
 113 -       handle_stripe(sh);
 114 +       handle_stripe(sh, NULL);
 115         release_stripe(sh);
 116
 117         return STRIPE_SECTORS;
 118 @@ -1605,9 +1642,11 @@ static void raid5d (mddev_t *mddev)
 119
 120                 handled++;
 121                 atomic_inc(&conf->handled_in_raid5d);
 122 -               handle_stripe(sh);
 123 +               handle_stripe(sh, NULL);
 124                 release_stripe(sh);
 125
 126 +               cond_resched();
 127 +
 128                 spin_lock_irq(&conf->device_lock);
 129         }
 130         PRINTK("%d stripes handled\n", handled);