[Ocfs2-tools-devel] [PATCH 9/9] fsck.ocfs2: Implement Pass 1D resolution of multiply-claimed clusters.

[Joel Becker]

On Fri, Jul 31, 2009 at 03:13:56PM +0800, Tao Ma wrote:
> 	I just have one qs.
> 	It seems that you cp/paste the whole content of a file regardless
> the number of multiply-claimed clusters. So why not just CoW the
> multiply-claimed ones? If we can do this, it is also helpful for
> libocfs2 support for reflink and I think it is worthwhile.

	Because it's trivially correct.  It involves no complex tree
manipulations and doesn't actually modify the original until we're sure
we have the space for it.  fsck's goal is correctness more than anything
else.
	Imagine I have 1GB of multiply claimed clusters in one extent.
But I have a fragmented filesystem, and I can't allocate a new 1GB
extent.  I now have to do a lot of tree work to insert multiple extents
in a CoW fashion, and I don't have CoW code yet.
	Also there is the problem of intermediate state.  I have to
remove that 1GB extent while I modify the tree, but then fsck can crash
and lose file data.  Here we can't have that happen.
	I will, of course, be adding CoW code for ocfs2_file_write()
when I get back to refcount trees in the tools.  But in my first pass, I
won't be teaching fsck how to fix refcount trees.  In the first, cut,
fsck will validate refcount trees, but invalid ones will be removed and
will drop the inodes into multiple claim resolution.  It's not
efficient, but it gives a correct result.  It's going to be hard enough
making sure I track all the refcounts of valid refcount trees!
	Later work may optimize this, but the first thing to do is have
a correctly working solution.

Joel
-- 

"I always thought the hardest questions were those I could not answer.
 Now I know they are the ones I can never ask."
			- Charlie Watkins

Joel Becker
Principal Software Developer
Oracle
E-mail: joel.becker at oracle.com
Phone: (650) 506-8127