ldiskfs/kernel_patches/patches/ext4-mballoc-pa_free-mismatch.patch

   1 Index: linux-stage/fs/ext4/mballoc.c
   2 ===================================================================
   3 --- linux-stage.orig/fs/ext4/mballoc.c  2010-01-26 22:50:37.000000000 +0800
   4 +++ linux-stage/fs/ext4/mballoc.c       2010-01-26 22:57:24.000000000 +0800
   5 @@ -3892,6 +3892,7 @@
   6         INIT_LIST_HEAD(&pa->pa_group_list);
   7         pa->pa_deleted = 0;
   8         pa->pa_type = MB_INODE_PA;
   9 +       pa->pa_error = 0;
  10
  11         mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
  12                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
  13 @@ -3956,6 +3957,7 @@
  14         INIT_LIST_HEAD(&pa->pa_group_list);
  15         pa->pa_deleted = 0;
  16         pa->pa_type = MB_GROUP_PA;
  17 +       pa->pa_error = 0;
  18
  19         mb_debug("new group pa %p: %llu/%u for %u\n", pa,
  20                  pa->pa_pstart, pa->pa_len, pa->pa_lstart);
  21 @@ -4019,7 +4021,9 @@
  22         int err = 0;
  23         int free = 0;
  24
  25 +       assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
  26         BUG_ON(pa->pa_deleted == 0);
  27 +       BUG_ON(pa->pa_inode == NULL);
  28         ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
  29         grp_blk_start = pa->pa_pstart - bit;
  30         BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
  31 @@ -4059,11 +4064,18 @@
  32                 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
  33                 bit = next + 1;
  34         }
  35 -       if (free != pa->pa_free) {
  36 -               printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
  37 -                       pa, (unsigned long) pa->pa_lstart,
  38 -                       (unsigned long) pa->pa_pstart,
  39 -                       (unsigned long) pa->pa_len);
  40 +
  41 +        /* "free < pa->pa_free" means we maybe double alloc the same blocks,
  42 +         * otherwise maybe leave some free blocks unavailable, no need to BUG.*/
  43 +        if ((free > pa->pa_free && !pa->pa_error) || (free < pa->pa_free)) {
  44 +                ext4_error(sb,"pa free mismatch: [pa %p] "
  45 +                              "[phy %lu] [logic %lu] [len %u] [free %u] "
  46 +                              "[error %u] [inode %lu] [freed %u]", pa,
  47 +                              (unsigned long)pa->pa_pstart,
  48 +                              (unsigned long)pa->pa_lstart,
  49 +                              (unsigned)pa->pa_len, (unsigned)pa->pa_free,
  50 +                              (unsigned)pa->pa_error, pa->pa_inode->i_ino,
  51 +                              free);
  52                 ext4_grp_locked_error(sb, group,
  53                                         __func__, "free %u, pa_free %u",
  54                                         free, pa->pa_free);
  55 @@ -4072,6 +4084,7 @@
  56                  * from the bitmap and continue.
  57                  */
  58         }
  59 +        BUG_ON(pa->pa_free != free);
  60         atomic_add(free, &sbi->s_mb_discarded);
  61
  62         return err;
  63 @@ -4800,6 +4813,24 @@
  64                 }
  65         }
  66
  67 +
  68 +        if (dev_check_rdonly(sb->s_bdev)) {
  69 +                struct block_device *bdev = sb->s_bdev;
  70 +
  71 +                printk(KERN_WARNING "Alloc from readonly device %s (%#x): "
  72 +                       "[inode %lu] [logic %llu] [goal %llu] [ll %llu] "
  73 +                      "[pl %llu] [lr %llu] [pr %llu] [len %u] [flags %u]\n",
  74 +                       bdev->bd_disk ? bdev->bd_disk->disk_name : "",
  75 +                       bdev->bd_dev, ar->inode->i_ino,
  76 +                      (unsigned long long)ar->logical,
  77 +                      (unsigned long long)ar->goal,
  78 +                       (unsigned long long)ar->lleft,
  79 +                      (unsigned long long)ar->pleft,
  80 +                      (unsigned long long)ar->lright,
  81 +                      (unsigned long long)ar->pright,
  82 +                      ar->len, ar->flags);
  83 +        }
  84 +
  85         ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
  86         if (!ac) {
  87                 ar->len = 0;
  88 @@ -4850,6 +4881,25 @@
  89                         ac->ac_b_ex.fe_len = 0;
  90                         ar->len = 0;
  91                         ext4_mb_show_ac(ac);
  92 +                        if (ac->ac_pa) {
  93 +                               struct ext4_prealloc_space *pa = ac->ac_pa;
  94 +
  95 +                               /* We can not make sure whether the bitmap has
  96 +                                        * been updated or not when fail case. So can
  97 +                                 * not revert pa_free back, just mark pa_error*/
  98 +                                pa->pa_error++;
  99 +                                ext4_error(sb,
 100 +                                            "Updating bitmap error: [err %d] "
 101 +                                            "[pa %p] [phy %lu] [logic %lu] "
 102 +                                            "[len %u] [free %u] [error %u] "
 103 +                                            "[inode %lu]", *errp, pa,
 104 +                                            (unsigned long)pa->pa_pstart,
 105 +                                            (unsigned long)pa->pa_lstart,
 106 +                                            (unsigned)pa->pa_len,
 107 +                                            (unsigned)pa->pa_free,
 108 +                                            (unsigned)pa->pa_error,
 109 +                                       pa->pa_inode ? pa->pa_inode->i_ino : 0);
 110 +                        }
 111                 } else {
 112                         block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
 113                         ar->len = ac->ac_b_ex.fe_len;
 114 @@ -5025,6 +5075,15 @@
 115                 goto error_return;
 116         }
 117
 118 +        if (dev_check_rdonly(sb->s_bdev)) {
 119 +                struct block_device *bdev = sb->s_bdev;
 120 +
 121 +                printk(KERN_WARNING "Release to readonly device %s (%#x): "
 122 +                       "[inode %lu] [block %lu] [count %lu] [is_meta %d]\n",
 123 +                       bdev->bd_disk ? bdev->bd_disk->disk_name : "",
 124 +                       bdev->bd_dev, inode->i_ino, (unsigned long)block, count, metadata);
 125 +        }
 126 +
 127         ext4_debug("freeing block %llu\n", block);
 128         trace_mark(ext4_free_blocks,
 129                    "dev %s block %llu count %lu metadata %d ino %lu",
 130 Index: linux-stage/fs/ext4/mballoc.h
 131 ===================================================================
 132 --- linux-stage.orig/fs/ext4/mballoc.h  2010-01-26 22:50:36.000000000 +0800
 133 +++ linux-stage/fs/ext4/mballoc.h       2010-01-26 22:52:58.000000000 +0800
 134 @@ -21,6 +21,7 @@
 135  #include <linux/blkdev.h>
 136  #include <linux/marker.h>
 137  #include <linux/mutex.h>
 138 +#include <linux/genhd.h>
 139  #include "ext4_jbd2.h"
 140  #include "ext4.h"
 141  #include "group.h"
 142 @@ -134,6 +135,7 @@
 143         ext4_grpblk_t           pa_len;         /* len of preallocated chunk */
 144         ext4_grpblk_t           pa_free;        /* how many blocks are free */
 145         unsigned short          pa_type;        /* pa type. inode or group */
 146 +       unsigned short          pa_error;
 147         spinlock_t              *pa_obj_lock;
 148         struct inode            *pa_inode;      /* hack, for history only */
 149  };