[Ocfs2-tools-devel] How to deal with file metadata corruption?

[Sunil Mushran]

On 05/12/2011 03:47 PM, Goldwyn Rodrigues wrote:
> Hi,
>
> I have a case of a corrupted file whose metadata is corrupted. The file
> is large and the corrupted part of extent block looks like this:
>
>          SubAlloc Bit: 576   SubAlloc Slot: 0
>          Blknum: 513089   Next Leaf: 0
>          Tree Depth: 1   Count: 252   Next Free Rec: 159
>          ## Offset        Clusters       Block#
>
> <snipped>
>
>          142 36821655      252            512689
>          143 36821907      674            512692
>          144 36822581      364            512694
>          145 36822945      4294966509     512697
>          146 36822158      4294967295     512699
>          147 36822157      1617           512702
>          148 36823774      4294966096     512724
>          149 36822574      4294967148     512770
>          150 36822426      0              145678608
>          151 36822426      149            145678613
>          152 36822575      4294966876     512841
>          153 36822155      0              512843
>          154 36822155      289            512847
>          155 36822444      290            512849
>          156 36822734      4294967185     512851
>          157 36822623      73266          512853
>          158 36895889      852847         512856
>
> Clearly, number of clusters can't be 0 and even 4294967185 seems to be
> a pretty high number for a file which is so fragmented.
>
> Is ocfs2-tools dealing with such an corruption? If not, what would be the
> best path to deal with it? The parent record contains the number of
> clusters the leaves contain and it can be validated that way.
> However, do you discard the extra blocks?
>
> The problem is: Trying to delete this file invokes a BUG_ON() which
> checks the sanity extent records.

It's not just the cluster count. Even the offset and the block
numbers are whacked. offsets are decreasing.

Forget fsck, one will need more info even to fix it by hand.

If the file can be fully deleted, then best to free the inode bit
in the correct inode_alloc, init the dentry in the directory.
Maybe add a dtime in the inode.

Should not be too hard via libocfs2.

Mark?