> On Aug 29, 2020, at 12:17 PM, bfields@xxxxxxxxxxxx wrote:
>
> On Fri, Aug 28, 2020 at 09:13:07PM +0000, Trond Myklebust wrote:
>> On Fri, 2020-08-28 at 15:58 -0400, Chuck Lever wrote:
>>> If a write delegation isn't available, the Linux NFS client uses
>>> a zero-stateid when performing a SETATTR.
>>>
>>> If that client happens to hold a read delegation, the server will
>>> recall it immediately, resulting in a short delay while the
>>> CB_RECALL operation is done. Optimize out this delay by having the
>>> client return any delegation it may hold on a file before issuing a
>>> SETATTR(zero-stateid) on that file.
>>>
>>> Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx>
>>> ---
>>> fs/nfs/nfs4proc.c | 1 +
>>> 1 file changed, 1 insertion(+)
>>>
>>> diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
>>> index dbd01548335b..53a56250cf4b 100644
>>> --- a/fs/nfs/nfs4proc.c
>>> +++ b/fs/nfs/nfs4proc.c
>>> @@ -3314,6 +3314,7 @@ static int _nfs4_do_setattr(struct inode
>>> *inode,
>>> goto zero_stateid;
>>> } else {
>>> zero_stateid:
>>> + nfs4_inode_return_delegation(inode);
>>> nfs4_stateid_copy(&arg->stateid, &zero_stateid);
>>> }
>>> if (delegation_cred)
>>>
>>
>> This should not be needed for NFSv4.1 or greater. Only NFSv4.0 is
>> incapable of identifying the caller and recognising that it is the
>> holder of the delegation.
>
> And the server should be getting this right now in the >=4.1 case, so
> let me know if you see otherwise.
The v5.9-rc1 server appears to be getting this right for v4.1.
Your earlier patch to avoid offering a delegation on OPEN(CREATE) helps
reduce the extra CB_RECALLs for v4.0, but the above patch is needed to
address the non-open-create cases.
--
Chuck Lever
