lustre/kernel_patches/patches/raid6-serialize-ovelapping-reqs.patch

   1 diff -pur linux-2.6.9.orig/drivers/md/raid6main.c linux-2.6.9/drivers/md/raid6main.c
   2 --- linux-2.6.9.orig/drivers/md/raid6main.c     2008-01-10 13:55:37.000000000 +0800
   3 +++ linux-2.6.9/drivers/md/raid6main.c  2008-01-10 13:55:56.000000000 +0800
   4 @@ -749,6 +749,10 @@ static void compute_parity(struct stripe
   5                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
   6                                 chosen = sh->dev[i].towrite;
   7                                 sh->dev[i].towrite = NULL;
   8 +
   9 +                               if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
  10 +                                       wake_up(&conf->wait_for_overlap);
  11 +
  12                                 if (sh->dev[i].written) BUG();
  13                                 sh->dev[i].written = chosen;
  14                         }
  15 @@ -907,7 +911,7 @@ static void compute_block_2(struct strip
  16   * toread/towrite point to the first in a chain.
  17   * The bi_next chain must be in order.
  18   */
  19 -static void add_stripe_bio (struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
  20 +static int add_stripe_bio (struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
  21  {
  22         struct bio **bip;
  23         raid6_conf_t *conf = sh->raid_conf;
  24 @@ -924,10 +928,13 @@ static void add_stripe_bio (struct strip
  25         else
  26                 bip = &sh->dev[dd_idx].toread;
  27         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
  28 -               BUG_ON((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector);
  29 +               if((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
  30 +                       goto overlap;
  31                 bip = & (*bip)->bi_next;
  32         }
  33 -/* FIXME do I need to worry about overlapping bion */
  34 +       if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
  35 +               goto overlap;
  36 +
  37         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
  38                 BUG();
  39         if (*bip)
  40 @@ -954,6 +961,14 @@ static void add_stripe_bio (struct strip
  41                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
  42                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
  43         }
  44 +
  45 +       return 1;
  46 +
  47 +overlap:
  48 +       set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
  49 +       spin_unlock_irq(&conf->device_lock);
  50 +       spin_unlock(&sh->lock);
  51 +       return 0;
  52  }
  53
  54  /*
  55 @@ -1038,6 +1053,9 @@ static void handle_stripe(struct stripe_
  56                         spin_lock_irq(&conf->device_lock);
  57                         rbi = dev->toread;
  58                         dev->toread = NULL;
  59 +
  60 +                       if (test_and_clear_bit(R5_Overlap, &dev->flags))
  61 +                               wake_up(&conf->wait_for_overlap);
  62                         spin_unlock_irq(&conf->device_lock);
  63                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
  64                                 copy_data(0, rbi, dev->page, dev->sector);
  65 @@ -1087,6 +1105,9 @@ static void handle_stripe(struct stripe_
  66                         sh->dev[i].towrite = NULL;
  67                         if (bi) to_write--;
  68
  69 +                       if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
  70 +                               wake_up(&conf->wait_for_overlap);
  71 +
  72                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
  73                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
  74                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
  75 @@ -1115,6 +1136,8 @@ static void handle_stripe(struct stripe_
  76                         if (!test_bit(R5_Insync, &sh->dev[i].flags)) {
  77                                 bi = sh->dev[i].toread;
  78                                 sh->dev[i].toread = NULL;
  79 +                               if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
  80 +                                       wake_up(&conf->wait_for_overlap);
  81                                 if (bi) to_read--;
  82                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
  83                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
  84 @@ -1648,6 +1671,8 @@ static int make_request (request_queue_t
  85                 sh = NULL;
  86                 /* iterrate through all pages in the stripe */
  87                 for (j = 0; j < data_disks && sectors > 0; j++) {
  88 +                       DEFINE_WAIT(w);
  89 +
  90                         if (r_sector + STRIPE_SECTORS <= bi->bi_sector ||
  91                             r_sector >= last_sector) {
  92                                 r_sector += sectors_per_chunk;
  93 @@ -1656,6 +1681,9 @@ static int make_request (request_queue_t
  94                         new_sector = raid6_compute_sector(r_sector, raid_disks,
  95                                                         data_disks, &dd_idx,
  96                                                         &pd_idx, conf);
  97 +
  98 +retry:
  99 +                       prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
 100                         if (sh == NULL) {
 101                                 /* first, try to get stripe w/o blocking
 102                                  * if we can't, then it's time to submit
 103 @@ -1668,10 +1696,18 @@ static int make_request (request_queue_t
 104                                 }
 105                         }
 106                         if (sh) {
 107 -                               add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK));
 108 +                               if(!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
 109 +                                       /* Failed to be added due to overlapped. */
 110 +                                       raid6_unplug_device(mddev->queue);
 111 +                                       release_stripe(sh);
 112 +                                       schedule();
 113 +                                       goto retry;
 114 +                               }
 115 +                               finish_wait(&conf->wait_for_overlap, &w);
 116                         } else {
 117                                 /* cannot get stripe for read-ahead, just give-up */
 118                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
 119 +                               finish_wait(&conf->wait_for_overlap, &w);
 120                                 sectors = 0;
 121                                 break;
 122                         }
 123 @@ -1847,6 +1883,7 @@ static int run (mddev_t *mddev)
 124
 125         conf->device_lock = SPIN_LOCK_UNLOCKED;
 126         init_waitqueue_head(&conf->wait_for_stripe);
 127 +       init_waitqueue_head(&conf->wait_for_overlap);
 128         INIT_LIST_HEAD(&conf->handle_list);
 129         INIT_LIST_HEAD(&conf->delayed_list);
 130         INIT_LIST_HEAD(&conf->inactive_list);
 131 diff -pur linux-2.6.9.orig/include/linux/raid/raid5.h linux-2.6.9/include/linux/raid/raid5.h
 132 --- linux-2.6.9.orig/include/linux/raid/raid5.h 2008-01-10 13:46:05.000000000 +0800
 133 +++ linux-2.6.9/include/linux/raid/raid5.h      2008-01-10 13:55:56.000000000 +0800
 134 @@ -154,6 +154,8 @@ struct stripe_head {
 135  #define        R5_Wantwrite    5
 136  #define        R5_Syncio       6       /* this io need to be accounted as resync io */
 137  #define        R5_Direct       7       /* use page from passed bio to avoid memcpy */
 138 +#define        R5_Overlap      8       /* There is a pending overlapping request
 139 +                                        * on this block */
 140
 141  /*
 142   * Write method
 143 @@ -221,6 +223,7 @@ struct raid5_private_data {
 144         atomic_t                active_stripes;
 145         struct list_head        inactive_list;
 146         wait_queue_head_t       wait_for_stripe;
 147 +       wait_queue_head_t       wait_for_overlap;
 148         int                     inactive_blocked;       /* release of inactive stripes blocked,
 149                                                          * waiting for 25% to be free
 150                                                          */