File system checking on ext3 after a system crash

[Andreas Dilger]

On Apr 09, 2007  15:34 -0400, Theodore Tso wrote:
> The big thing is that if e2fsck wants to relocate an inode table, you
> almost always want to stop and backup metadata blocks using e2image
> first.  In fact I'm thinking about revamping that logic since right
> now the potential for doing great harm to the filesystem is far too
> high.  So the fact that you might want to say 'n' there is really more
> of a sane of a e2fsck bug, or at least misdesign, more than anything
> else.

That could probably be helped out by the GDT_CSUM feature, so that the
group descriptors can be validated individually instead of en-masse,
and it is much more difficult to accidentally get the wrong bitmap or
itable location written to disk with a good checksum (though bit errors
in memory before calculating the checksum are still possible).

In the "relocate itable" case e2fsck could also read the backup group
descriptors to validate the location of the bitmaps and inode table
blocks even if a given group descriptor doesn't look insane, because of
the risk of a single bit flip only making the block number wrong by a
small amount.

Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.