Re: [PATCH] NFSD: Fill in WCC data for REMOVE, RMDIR, MKNOD, and MKDIR

[Chuck Lever <chuck.lever@oracle.com>]

On 07/ 7/10 05:10 PM, J. Bruce Fields wrote:
> On Tue, Jul 06, 2010 at 04:53:34PM -0400, Chuck Lever wrote:
>> Some well-known NFSv3 clients drop their directory entry caches when
>> they receive replies with no WCC data.
>
> Can we include any more details?  (And/or a simple test case that
> demonstrates the difference?)

I have a version of this patch with a much more comprehensive 
description which I can post later today.

As for a test case... among other things, I untarred a large tarball on 
a Mac OS 10.6 NFS client, and, once the server was fixed, it went as 
quickly as you would expect.  Before, it took much longer.

>> Without this data, they
>> employ extra READ, LOOKUP, and GETATTR requests to ensure their
>> directory entry caches are up to date, causing performance to suffer
>> needlessly.
>>
>> In order to return WCC data, our server has to have both the pre-op
>> and the post-op attribute data on hand when a reply is XDR encoded.
>> The pre-op data is filled in when the incoming fh is locked, and the
>> post-op data is filled in when the fh is unlocked.
>>
>> Unfortunately, for REMOVE, RMDIR, MKNOD, and MKDIR, the directory fh
>> is not unlocked until well after the reply has been XDR encoded.
>
> So it wasn't happening until an fh_put() was done in .pc_release()?

Correct.