ldiskfs/kernel_patches/patches/ext3-mballoc-pa_free-mismatch.patch

   1 Index: linux-stage/fs/ext3/mballoc.c
   2 ===================================================================
   3 --- linux-stage.orig/fs/ext3/mballoc.c  2010-01-26 23:36:25.000000000 +0800
   4 +++ linux-stage/fs/ext3/mballoc.c       2010-01-26 23:42:57.000000000 +0800
   5 @@ -35,6 +35,7 @@
   6  #include <linux/pagemap.h>
   7  #include <linux/seq_file.h>
   8  #include <linux/version.h>
   9 +#include <linux/genhd.h>
  10
  11  #include "group.h"
  12
  13 @@ -360,6 +361,7 @@
  14         unsigned short          pa_free;        /* how many blocks are free */
  15         unsigned short          pa_linear;      /* consumed in one direction
  16                                                  * strictly, for group prealloc */
  17 +       unsigned short          pa_error;
  18         spinlock_t              *pa_obj_lock;
  19         struct inode            *pa_inode;      /* hack, for history only */
  20  };
  21 @@ -3558,6 +3560,7 @@
  22         spin_lock_init(&pa->pa_lock);
  23         pa->pa_deleted = 0;
  24         pa->pa_linear = 0;
  25 +       pa->pa_error = 0;
  26
  27         mb_debug("new inode pa %p: %lu/%lu for %lu\n", pa,
  28                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
  29 @@ -3615,6 +3618,7 @@
  30         spin_lock_init(&pa->pa_lock);
  31         pa->pa_deleted = 0;
  32         pa->pa_linear = 1;
  33 +       pa->pa_error = 0;
  34
  35         mb_debug("new group pa %p: %lu/%lu for %lu\n", pa,
  36                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
  37 @@ -3671,7 +3675,10 @@
  38         sector_t start;
  39         int err = 0, free = 0;
  40
  41 +       BUG_ON(!ext3_is_group_locked(sb, e3b->bd_group));
  42         BUG_ON(pa->pa_deleted == 0);
  43 +       BUG_ON(pa->pa_linear != 0);
  44 +       BUG_ON(pa->pa_inode == NULL);
  45         ext3_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
  46         BUG_ON(group != e3b->bd_group && pa->pa_len != 0);
  47         end = bit + pa->pa_len;
  48 @@ -3704,14 +3711,19 @@
  49                 mb_free_blocks(pa->pa_inode, e3b, bit, next - bit);
  50                 bit = next + 1;
  51         }
  52 -       if (free != pa->pa_free) {
  53 -               printk("pa %p: logic %lu, phys. %lu, len %lu\n",
  54 -                       pa, (unsigned long) pa->pa_lstart,
  55 -                       (unsigned long) pa->pa_pstart,
  56 -                       (unsigned long) pa->pa_len);
  57 -               printk("free %u, pa_free %u\n", free, pa->pa_free);
  58 -       }
  59 -       BUG_ON(free != pa->pa_free);
  60 +
  61 +       /* "free < pa->pa_free" means we maybe double alloc the same blocks,
  62 +        * otherwise maybe leave some free blocks unavailable, no need to BUG.*/
  63 +       if ((free > pa->pa_free && !pa->pa_error) || (free < pa->pa_free))
  64 +               ext3_error(sb, __FUNCTION__, "pa free mismatch: [pa %p] "
  65 +                             "[phy %lu] [logic %lu] [len %u] [free %u] "
  66 +                             "[error %u] [inode %lu] [freed %u]", pa,
  67 +                             (unsigned long)pa->pa_pstart,
  68 +                             (unsigned long)pa->pa_lstart,
  69 +                             (unsigned)pa->pa_len, (unsigned)pa->pa_free,
  70 +                             (unsigned)pa->pa_error, pa->pa_inode->i_ino,
  71 +                             free);
  72 +       BUG_ON(pa->pa_free != free);
  73         atomic_add(free, &sbi->s_mb_discarded);
  74
  75         return err;
  76 @@ -4189,6 +4201,19 @@
  77                 *errp = -EDQUOT;
  78                 return 0;
  79         }
  80 +
  81 +       if (dev_check_rdonly(sb->s_bdev)) {
  82 +               struct block_device *bdev = sb->s_bdev;
  83 +
  84 +               printk(KERN_WARNING "Alloc from readonly device %s (%#x): "
  85 +                      "[inode %lu] [logic %lu] [goal %lu] [ll %lu] [pl %lu] "
  86 +                      "[lr %lu] [pr %lu] [len %lu] [flags %lu]\n",
  87 +                      bdev->bd_disk ? bdev->bd_disk->disk_name : "",
  88 +                      bdev->bd_dev, ar->inode->i_ino, ar->logical, ar->goal,
  89 +                      ar->lleft, ar->pleft, ar->lright, ar->pright, ar->len,
  90 +                      ar->flags);
  91 +       }
  92 +
  93         inquota = ar->len;
  94
  95         ext3_mb_poll_new_transaction(sb, handle);
  96 @@ -4217,10 +4242,34 @@
  97         }
  98
  99         if (likely(ac.ac_status == AC_STATUS_FOUND)) {
 100 -               ext3_mb_mark_diskspace_used(&ac, handle);
 101 -               *errp = 0;
 102 -               block = ext3_grp_offs_to_block(sb, &ac.ac_b_ex);
 103 -               ar->len = ac.ac_b_ex.fe_len;
 104 +               *errp = ext3_mb_mark_diskspace_used(&ac, handle);
 105 +               if (!*errp) {
 106 +                       block = ext3_grp_offs_to_block(sb, &ac.ac_b_ex);
 107 +                       ar->len = ac.ac_b_ex.fe_len;
 108 +               } else {
 109 +                        ac.ac_b_ex.fe_len = 0;
 110 +                        ar->len = 0;
 111 +                        ext3_mb_show_ac(&ac);
 112 +                        if (ac.ac_pa) {
 113 +                                struct ext3_prealloc_space *pa = ac.ac_pa;
 114 +
 115 +                               /* We can not make sure whether the bitmap has
 116 +                                * been updated or not when fail case. So can
 117 +                                * not revert pa_free back, just mark pa_error*/
 118 +                               pa->pa_error++;
 119 +                                ext3_error(sb, __FUNCTION__,
 120 +                                             "Updating bitmap error: [err %d] "
 121 +                                             "[pa %p] [phy %lu] [logic %lu] "
 122 +                                             "[len %u] [free %u] [error %u] "
 123 +                                             "[inode %lu]", *errp, pa,
 124 +                                             (unsigned long)pa->pa_pstart,
 125 +                                             (unsigned long)pa->pa_lstart,
 126 +                                             (unsigned)pa->pa_len,
 127 +                                             (unsigned)pa->pa_free,
 128 +                                             (unsigned)pa->pa_error,
 129 +                                       pa->pa_inode ? pa->pa_inode->i_ino : 0);
 130 +                        }
 131 +               }
 132         } else {
 133                 freed  = ext3_mb_discard_preallocations(sb, ac.ac_o_ex.fe_len);
 134                 if (freed)
 135 @@ -4388,6 +4437,15 @@
 136                 goto error_return;
 137         }
 138
 139 +       if (dev_check_rdonly(sb->s_bdev)) {
 140 +               struct block_device *bdev = sb->s_bdev;
 141 +
 142 +               printk(KERN_WARNING "Release to readonly device %s (%#x): "
 143 +                      "[inode %lu] [block %lu] [count %lu] [is_meta %d]\n",
 144 +                      bdev->bd_disk ? bdev->bd_disk->disk_name : "",
 145 +                      bdev->bd_dev, inode->i_ino, block, count, metadata);
 146 +       }
 147 +
 148         ext3_debug("freeing block %lu\n", block);
 149
 150         ac.ac_op = EXT3_MB_HISTORY_FREE;