Re: [PATCH, RFC] ext4: remove ext4_dir_llseek

[Eric Sandeen <sandeen@redhat.com>]

On 4/25/12 2:23 PM, Eric Sandeen wrote:
> ext4_dir_llseek() was recently added as part of a patch to allow
> returning 64-bit name hashes.  However, reworking _llseek() is not
> necessary to achieve that goal.  One unfortunate side effect of the
> change is that it cut&pasted VFS code back into ext4, after Andi
> had just removed other _llseek cut&paste in ext4 by extending the
> VFS functionality.
> 
> It also re-introduced i_mutex locking in the dir llseek paths,
> un-doing the upstream (mostly) lockless llseek changes from Andi
> in this case.
> 
> Because of the above reasons, and because it introduces new
> EINVAL returns which were not there before, and because SEEK_END+offset
> behaves differently from SEEK_SET w.r.t. offset limits, and because
> it's not clear what problem this is solving, remove it for now.
> 
> (NFS only uses SEEK_SET and SEEK_CUR, so changes to SEEK_END shouldn't
> affect it.)
> 
> If & when a problematic and testable real-world use case is described,
> this can be re-fixed properly, possibly by expanding the VFS _llseek()
> functions to handle custom EOF offsets for cases such as this.

Self-NAK.  Realized this won't work for non-extent dirs, because the
generic llseek will EINVAL for hash "offsets" larger than the max_bitmap_bytes
value.  Will send something a little less drastic.

-Eric