Re: [PATCH 1/1] fix d_revalidate oopsen on NFS exports

[Chris Dunlop <chris@onthe.net.au>]

On Thu, Dec 01, 2011 at 12:50:25AM -0600, Tyler Hicks wrote:
> On 2011-11-29 19:25:01, Chris Dunlop wrote:
>> I haven't seen any response to this patch which fixes an Oops in
>> d_revalidate. I hit this using NFS, but various other file
>> systems look to be likewise vulnerable, hence the broadness of
>> the patch. The sequence leading to the Oops is:
>> 
>> lookup_one_len() [fs/namei.c]
>>    calls __lookup_hash() [fs/namei.c] with nd == NULL,
>>       which can then call the file system specific d_revalidate(), passing in nd == NULL
>>          which will then Oops if nd is used without checking
> 
> Hey Chris - Can you share what you were trying to do when you hit this?
> Were you stacking eCryptfs on top of NFS? Another stacked filesystem on
> top of NFS?
>
> Do you *need* a stacked filesystem to work on top of NFS? If so, we'll
> need to discuss a way forward. Al has previously shown a dislike of
> eCryptfs passing around nameidata (for good reason), but that is what
> NFS currently requires. I looked at doing this a few months back, but
> never got to the implementation stage.

Actually, no, it wasn't eCryptfs or another stacked filesystem.
It seems my dirty little secret must come out: I hit the problem
when trying to use the (necessarily) out-of-tree zfsonlinux
(ZoL) [1], on an NFS root machine.

I don't know exactly what ZoL is using lookup_one_len() for, nor
how to fix it so it isn't, but I've given them the heads up that
it's not supposed to be used outside of original file system [2].

Chris.

[1] http://zfsonlinux.org/
[2] https://github.com/zfsonlinux/zfs/issues/456