lustre/kernel_patches/patches/raid5-stripe-by-stripe-handling.patch

   1 Helps to avoid unnesessary reads if request covers full stripe.
   2
   3 Note that reads needed to update parity hurt performance badly
   4
   5 Index: linux-2.6.9/drivers/md/raid5.c
   6 ===================================================================
   7 --- linux-2.6.9.orig/drivers/md/raid5.c 2006-05-22 00:09:56.000000000 +0400
   8 +++ linux-2.6.9/drivers/md/raid5.c      2006-05-22 00:10:01.000000000 +0400
   9 @@ -1412,6 +1412,11 @@ static int make_request (request_queue_t
  10         sector_t new_sector;
  11         sector_t logical_sector, last_sector;
  12         struct stripe_head *sh;
  13 +       sector_t stripe, sectors, block, r_sector, b_sector;
  14 +       int sectors_per_chunk = conf->chunk_size >> 9;
  15 +       int stripes_per_chunk, sectors_per_block;
  16 +       int sectors_per_stripe;
  17 +       int i, j;
  18
  19         atomic_inc(&conf->in_reqs_in_queue);
  20
  21 @@ -1431,30 +1436,66 @@ static int make_request (request_queue_t
  22         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
  23         if ( bio_data_dir(bi) == WRITE )
  24                 md_write_start(mddev);
  25 -       for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
  26 -
  27 -               new_sector = raid5_compute_sector(logical_sector,
  28 -                                                 raid_disks, data_disks, &dd_idx, &pd_idx, conf);
  29 -
  30 -               PRINTK("raid5: make_request, sector %Lu logical %Lu\n",
  31 -                       (unsigned long long)new_sector,
  32 -                       (unsigned long long)logical_sector);
  33
  34 -               sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
  35 +       stripes_per_chunk = conf->chunk_size / STRIPE_SIZE;
  36 +       sectors_per_stripe = STRIPE_SECTORS * data_disks;
  37 +       sectors_per_block = stripes_per_chunk * sectors_per_stripe;
  38 +
  39 +       block = logical_sector & ~((sector_t)sectors_per_block - 1);
  40 +       sector_div(block, sectors_per_block);
  41 +       sectors = bi->bi_size >> 9;
  42 +
  43 +repeat:
  44 +       stripe = block * (sectors_per_block / data_disks);
  45 +       b_sector = stripe * data_disks;
  46 +       /* iterate through all stripes in this block,
  47 +        * where block is a set of internal stripes
  48 +        * which covers chunk */
  49 +       for (i = 0; i < stripes_per_chunk && sectors > 0; i++) {
  50 +               r_sector = b_sector + (i * STRIPE_SECTORS);
  51 +               sh = NULL;
  52 +               /* iterrate through all pages in the stripe */
  53 +               for (j = 0; j < data_disks && sectors > 0; j++) {
  54 +                       if (r_sector + STRIPE_SECTORS <= bi->bi_sector ||
  55 +                                       r_sector >= last_sector) {
  56 +                               r_sector += sectors_per_chunk;
  57 +                               continue;
  58 +                       }
  59 +                       new_sector = raid5_compute_sector(r_sector, raid_disks,
  60 +                                                       data_disks, &dd_idx,
  61 +                                                       &pd_idx, conf);
  62 +                       if (sh == NULL)
  63 +                               sh = get_active_stripe(conf, new_sector, pd_idx,
  64 +                                                       (bi->bi_rw&RWA_MASK));
  65 +                       if (sh) {
  66 +                               add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK));
  67 +                       } else {
  68 +                               /* cannot get stripe for read-ahead, just give-up */
  69 +                               clear_bit(BIO_UPTODATE, &bi->bi_flags);
  70 +                               sectors = 0;
  71 +                               break;
  72 +                       }
  73 +
  74 +                       BUG_ON (new_sector != stripe);
  75 +                       sectors -= STRIPE_SECTORS;
  76 +                       if (bi->bi_sector > r_sector)
  77 +                               sectors += bi->bi_sector - r_sector;
  78 +                       if (r_sector + STRIPE_SECTORS > last_sector)
  79 +                               sectors += r_sector + STRIPE_SECTORS - last_sector;
  80 +                       r_sector += sectors_per_chunk;
  81 +               }
  82                 if (sh) {
  83 -
  84 -                       add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK));
  85 -
  86                         raid5_plug_device(conf);
  87                         handle_stripe(sh);
  88                         release_stripe(sh);
  89 -               } else {
  90 -                       /* cannot get stripe for read-ahead, just give-up */
  91 -                       clear_bit(BIO_UPTODATE, &bi->bi_flags);
  92 -                       break;
  93 +                       sh = NULL;
  94                 }
  95 -
  96 +               stripe += STRIPE_SECTORS;
  97         }
  98 +       block++;
  99 +       if (sectors > 0)
 100 +               goto repeat;
 101 +
 102         spin_lock_irq(&conf->device_lock);
 103         if (--bi->bi_phys_segments == 0) {
 104                 int bytes = bi->bi_size;