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

   1 diff -pur b/drivers/md/raid5.c a/drivers/md/raid5.c
   2 --- b/drivers/md/raid5.c        2009-02-20 15:56:36.000000000 +0800
   3 +++ a/drivers/md/raid5.c        2009-02-20 15:57:49.000000000 +0800
   4 @@ -1277,7 +1277,26 @@ static void compute_block_2(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   * Each stripe/dev can have one or more bion attached.
  31 @@ -1392,7 +1411,7 @@ static int stripe_to_pdidx(sector_t stri
  32   *
  33   */
  34
  35 -static void handle_stripe5(struct stripe_head *sh)
  36 +static void handle_stripe5(struct stripe_head *sh, struct bio *bios[])
  37  {
  38         raid5_conf_t *conf = sh->raid_conf;
  39         int disks = sh->disks;
  40 @@ -1939,7 +1958,11 @@ static void handle_stripe5(struct stripe
  41                             test_bit(R5_ReWrite, &sh->dev[i].flags))
  42                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
  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                         if (rw == 1)
  52                                 set_bit(STRIPE_DEGRADED, &sh->state);
  53 @@ -1951,7 +1974,7 @@ static void handle_stripe5(struct stripe
  54         }
  55  }
  56
  57 -static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
  58 +static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page, struct bio *bios[])
  59  {
  60         raid6_conf_t *conf = sh->raid_conf;
  61         int disks = conf->raid_disks;
  62 @@ -2499,7 +2522,11 @@ static void handle_stripe6(struct stripe
  63                         if (rw == WRITE &&
  64                             test_bit(R5_ReWrite, &sh->dev[i].flags))
  65                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
  66 -                       generic_make_request(bi);
  67 +                       if (bios) {
  68 +                               bi->bi_next = bios[i];
  69 +                               bios[i] = bi;
  70 +                       } else
  71 +                               generic_make_request(bi);
  72                         atomic_inc(&conf->out_reqs_in_queue);
  73                 } else {
  74                         if (rw == 1)
  75 @@ -2512,12 +2539,12 @@ static void handle_stripe6(struct stripe
  76         }
  77  }
  78
  79 -static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
  80 +static void handle_stripe(struct stripe_head *sh, struct page *tmp_page, struct bio *bios[])
  81  {
  82         if (sh->raid_conf->level == 6)
  83 -               handle_stripe6(sh, tmp_page);
  84 +               handle_stripe6(sh, tmp_page, bios);
  85         else
  86 -               handle_stripe5(sh);
  87 +               handle_stripe5(sh, bios);
  88  }
  89
  90
  91 @@ -2670,6 +2697,7 @@ static int make_request(request_queue_t
  92         int stripes_per_chunk, sectors_per_block;
  93         int sectors_per_stripe;
  94         int i, j;
  95 +       struct bio *bios[MD_SB_DISKS];
  96
  97         DEFINE_WAIT(w);
  98         int disks, data_disks;
  99 @@ -2698,6 +2726,7 @@ static int make_request(request_queue_t
 100         sectors = bi->bi_size >> 9;
 101         stripes_per_chunk = conf->chunk_size / STRIPE_SIZE;
 102
 103 +       memset(&bios, 0, sizeof(bios));
 104  redo_bio:
 105         /* stripe by stripe handle needs a stable raid layout, so if this
 106          * reuqest covers the expanding region, wait it over.
 107 @@ -2756,8 +2785,10 @@ retry:
 108                                          * the raid layout has been changed, we have to redo the
 109                                          * whole bio because we don't which sectors in it has been
 110                                          * done, and which is not done. -jay */
 111 -                                       if (raid5_redo_bio(conf, bi, disks, logical_sector))
 112 +                                       if (raid5_redo_bio(conf, bi, disks, logical_sector)) {
 113 +                                               raid5_flush_bios(conf, bios, disks);
 114                                                 goto redo_bio;
 115 +                                       }
 116
 117                                         if (test_bit(STRIPE_EXPANDING, &sh->state)) {
 118                                                 /* Stripe is busy expanding or
 119 @@ -2766,6 +2797,7 @@ retry:
 120                                                  */
 121                                                 release_stripe(sh);
 122                                                 sh = NULL;
 123 +                                               raid5_flush_bios(conf, bios, disks);
 124                                                 raid5_unplug_device(mddev->queue);
 125                                                 schedule();
 126                                                 goto retry;
 127 @@ -2784,17 +2816,19 @@ retry:
 128                          */
 129                         if (r_sector >= mddev->suspend_lo &&
 130                             r_sector < mddev->suspend_hi) {
 131 -                               handle_stripe(sh, NULL);
 132 +                               handle_stripe(sh, NULL, NULL);
 133                                 release_stripe(sh);
 134                                 sh = NULL;
 135 +                               raid5_flush_bios(conf, bios, disks);
 136                                 schedule();
 137                                 goto retry;
 138                         }
 139
 140                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
 141 -                               handle_stripe(sh, NULL);
 142 +                               handle_stripe(sh, NULL, NULL);
 143                                 release_stripe(sh);
 144                                 sh = NULL;
 145 +                               raid5_flush_bios(conf, bios, disks);
 146                                 raid5_unplug_device(mddev->queue);
 147                                 schedule();
 148                                 goto retry;
 149 @@ -2810,7 +2844,7 @@ retry:
 150                         r_sector += sectors_per_chunk;
 151                 }
 152                 if (sh) {
 153 -                       handle_stripe(sh, NULL);
 154 +                       handle_stripe(sh, NULL, bios);
 155                         release_stripe(sh);
 156                         sh = NULL;
 157                 }
 158 @@ -2820,6 +2854,9 @@ retry:
 159         if (sectors > 0)
 160                 goto repeat;
 161
 162 +       /* flush all of the bios */
 163 +       raid5_flush_bios(conf, bios, disks);
 164 +
 165         spin_lock_irq(&conf->device_lock);
 166         remaining = --bi->bi_phys_segments;
 167         spin_unlock_irq(&conf->device_lock);
 168 @@ -3035,7 +3072,7 @@ static inline sector_t sync_request(mdde
 169         clear_bit(STRIPE_INSYNC, &sh->state);
 170         spin_unlock(&sh->lock);
 171
 172 -       handle_stripe(sh, NULL);
 173 +       handle_stripe(sh, NULL, NULL);
 174         release_stripe(sh);
 175
 176         return STRIPE_SECTORS;
 177 @@ -3091,9 +3128,11 @@ static void raid5d (mddev_t *mddev)
 178
 179                 handled++;
 180                 atomic_inc(&conf->handled_in_raid5d);
 181 -               handle_stripe(sh, conf->spare_page);
 182 +               handle_stripe(sh, conf->spare_page, NULL);
 183                 release_stripe(sh);
 184
 185 +               cond_resched();
 186 +
 187                 spin_lock_irq(&conf->device_lock);
 188         }
 189         PRINTK("%d stripes handled\n", handled);