lustre/kernel_patches/patches/raid5-merge-ios-rhel5.patch

   1 Index: linux-2.6.18-92.1.17/drivers/md/raid5.c
   2 ===================================================================
   3 --- linux-2.6.18-92.1.17.orig/drivers/md/raid5.c
   4 +++ linux-2.6.18-92.1.17/drivers/md/raid5.c
   5 @@ -1277,7 +1277,26 @@ static void compute_block_2(struct strip
   6         }
   7  }
   8
   9 +/*
  10 + * The whole idea is to collect all bio's and then issue them
  11 + * disk by disk to assist merging a bit -bzzz
  12 + */
  13 +static void raid5_flush_bios(raid5_conf_t *conf, struct bio *bios[], int raid_disks)
  14 +{
  15 +       struct bio *bio, *nbio;
  16 +       int i;
  17
  18 +       for (i = 0; i < raid_disks; i++) {
  19 +               bio = bios[i];
  20 +               while (bio) {
  21 +                       nbio = bio->bi_next;
  22 +                       bio->bi_next = NULL;
  23 +                       generic_make_request(bio);
  24 +                       bio = nbio;
  25 +               }
  26 +               bios[i] = NULL;
  27 +       }
  28 +}
  29
  30  /*
  31   * Each stripe/dev can have one or more bion attached.
  32 @@ -1392,7 +1411,7 @@ static int stripe_to_pdidx(sector_t stri
  33   *
  34   */
  35
  36 -static void handle_stripe5(struct stripe_head *sh)
  37 +static void handle_stripe5(struct stripe_head *sh, struct bio *bios[])
  38  {
  39         raid5_conf_t *conf = sh->raid_conf;
  40         int disks = sh->disks;
  41 @@ -1939,7 +1958,11 @@ static void handle_stripe5(struct stripe
  42                             test_bit(R5_ReWrite, &sh->dev[i].flags))
  43                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
  44                         atomic_inc(&conf->out_reqs_in_queue);
  45 -                       generic_make_request(bi);
  46 +                       if (bios) {
  47 +                               bi->bi_next = bios[i];
  48 +                               bios[i] = bi;
  49 +                       } else
  50 +                               generic_make_request(bi);
  51                 } else {
  52                         if (rw == 1)
  53                                 set_bit(STRIPE_DEGRADED, &sh->state);
  54 @@ -1951,7 +1974,7 @@ static void handle_stripe5(struct stripe
  55         }
  56  }
  57
  58 -static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
  59 +static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page, struct bio *bios[])
  60  {
  61         raid6_conf_t *conf = sh->raid_conf;
  62         int disks = conf->raid_disks;
  63 @@ -2499,7 +2522,11 @@ static void handle_stripe6(struct stripe
  64                         if (rw == WRITE &&
  65                             test_bit(R5_ReWrite, &sh->dev[i].flags))
  66                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
  67 -                       generic_make_request(bi);
  68 +                       if (bios) {
  69 +                               bi->bi_next = bios[i];
  70 +                               bios[i] = bi;
  71 +                       } else
  72 +                               generic_make_request(bi);
  73                         atomic_inc(&conf->out_reqs_in_queue);
  74                 } else {
  75                         if (rw == 1)
  76 @@ -2512,12 +2539,12 @@ static void handle_stripe6(struct stripe
  77         }
  78  }
  79
  80 -static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
  81 +static void handle_stripe(struct stripe_head *sh, struct page *tmp_page, struct bio *bios[])
  82  {
  83         if (sh->raid_conf->level == 6)
  84 -               handle_stripe6(sh, tmp_page);
  85 +               handle_stripe6(sh, tmp_page, bios);
  86         else
  87 -               handle_stripe5(sh);
  88 +               handle_stripe5(sh, bios);
  89  }
  90
  91
  92 @@ -2670,6 +2697,7 @@ static int make_request(request_queue_t
  93         int stripes_per_chunk, sectors_per_block;
  94         int sectors_per_stripe;
  95         int i, j;
  96 +       struct bio *bios[MD_SB_DISKS];
  97
  98         DEFINE_WAIT(w);
  99         int disks, data_disks;
 100 @@ -2698,6 +2726,7 @@ static int make_request(request_queue_t
 101         sectors = bi->bi_size >> 9;
 102         stripes_per_chunk = conf->chunk_size / STRIPE_SIZE;
 103
 104 +       memset(&bios, 0, sizeof(bios));
 105  redo_bio:
 106         /* stripe by stripe handle needs a stable raid layout, so if this
 107          * reuqest covers the expanding region, wait it over.
 108 @@ -2756,8 +2785,10 @@ retry:
 109                                          * the raid layout has been changed, we have to redo the
 110                                          * whole bio because we don't which sectors in it has been
 111                                          * done, and which is not done. -jay */
 112 -                                       if (raid5_redo_bio(conf, bi, disks, logical_sector))
 113 +                                       if (raid5_redo_bio(conf, bi, disks, logical_sector)) {
 114 +                                               raid5_flush_bios(conf, bios, disks);
 115                                                 goto redo_bio;
 116 +                                       }
 117
 118                                         if (test_bit(STRIPE_EXPANDING, &sh->state)) {
 119                                                 /* Stripe is busy expanding or
 120 @@ -2766,6 +2797,7 @@ retry:
 121                                                  */
 122                                                 release_stripe(sh);
 123                                                 sh = NULL;
 124 +                                               raid5_flush_bios(conf, bios, disks);
 125                                                 raid5_unplug_device(mddev->queue);
 126                                                 schedule();
 127                                                 goto retry;
 128 @@ -2784,17 +2816,19 @@ retry:
 129                          */
 130                         if (r_sector >= mddev->suspend_lo &&
 131                             r_sector < mddev->suspend_hi) {
 132 -                               handle_stripe(sh, NULL);
 133 +                               handle_stripe(sh, NULL, NULL);
 134                                 release_stripe(sh);
 135                                 sh = NULL;
 136 +                               raid5_flush_bios(conf, bios, disks);
 137                                 schedule();
 138                                 goto retry;
 139                         }
 140
 141                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
 142 -                               handle_stripe(sh, NULL);
 143 +                               handle_stripe(sh, NULL, NULL);
 144                                 release_stripe(sh);
 145                                 sh = NULL;
 146 +                               raid5_flush_bios(conf, bios, disks);
 147                                 raid5_unplug_device(mddev->queue);
 148                                 schedule();
 149                                 goto retry;
 150 @@ -2810,7 +2844,7 @@ retry:
 151                         r_sector += sectors_per_chunk;
 152                 }
 153                 if (sh) {
 154 -                       handle_stripe(sh, NULL);
 155 +                       handle_stripe(sh, NULL, bios);
 156                         release_stripe(sh);
 157                         sh = NULL;
 158                 }
 159 @@ -2820,6 +2854,9 @@ retry:
 160         if (sectors > 0)
 161                 goto repeat;
 162
 163 +       /* flush all of the bios */
 164 +       raid5_flush_bios(conf, bios, disks);
 165 +
 166         spin_lock_irq(&conf->device_lock);
 167         remaining = --bi->bi_phys_segments;
 168         spin_unlock_irq(&conf->device_lock);
 169 @@ -3035,7 +3072,7 @@ static inline sector_t sync_request(mdde
 170         clear_bit(STRIPE_INSYNC, &sh->state);
 171         spin_unlock(&sh->lock);
 172
 173 -       handle_stripe(sh, NULL);
 174 +       handle_stripe(sh, NULL, NULL);
 175         release_stripe(sh);
 176
 177         return STRIPE_SECTORS;
 178 @@ -3091,7 +3128,7 @@ static void raid5d (mddev_t *mddev)
 179
 180                 handled++;
 181                 atomic_inc(&conf->handled_in_raid5d);
 182 -               handle_stripe(sh, conf->spare_page);
 183 +               handle_stripe(sh, conf->spare_page, NULL);
 184                 release_stripe(sh);
 185
 186                 spin_lock_irq(&conf->device_lock);