Re: [PATCH 3/6] ubifs: Use 64bit readdir cookies

[J. Bruce Fields]

On Thu, Dec 29, 2016 at 04:49:54PM +0100, Richard Weinberger wrote:
> Bruce,
> 
> On 29.12.2016 16:34, J. Bruce Fields wrote:
> >> That way UBIFS can provide a 64bit readdir() cookie which is required for NFS3.
> > 
> > Sounds good.  And if a matching entry isn't found (as in the case of a
> > concurrent unlink), what happens?  The answer must be the same as for
> > ext4, but I've forgotten the details....  I guess it must find the next
> > highest cookie (thinking of the cookie as a 64-bit integer of some kind)
> > that exists in the directory.  And that must be the same order that
> > readdir normally returns entries in.
> 
> If a 64bit cookie is not found, the lookup function returns -ENOENT.
> In UBIFS we cannot just select a higher or lower key (cookie in this case),
> since it is the B-tree key and would point to a completely different
> entry.
> 
> So, in case of a concurrent unlink() one would succeed and one fail with
> -ENOENT. Unless I miss something that seems okay to me.

Unlink takes (parent directory, name), not a directory cookie.

The problem is concurrent unlink and nfs readdir.  So:

	NFS server returns readdir result with cookie X

	Somebody unlinks the entry at X.

	NFS server gets readdir request with cookie X.

Then the NFS client will get a spurious -ENOENT.

I'm not sure how best to reproduce that.... Maybe:

	Create a directory on an nfs-exported filesystem with lots of
	entries.

	Start a loop (or loops?) renaming directory entries within the
	directory as fast as possible (or deleting and creating entries;
	I assume it's the same thing for our purposes).

	read the directory from an nfs client.

I'm not sure how many entries is "lots".... Ideally you want a single
read of the directory to require the client to make lots of READDIR
requests to the server.  You could help by running:

	echo 1024 >/proc/fs/nfsd/max_block_size

before starting knfsd.  That should force it to return no more than 1K
of data in each READDIR reply.

--b.