On 03/08/2012 12:57 PM, J. Bruce Fields wrote:
> On Thu, Mar 08, 2012 at 08:50:14PM +0000, Myklebust, Trond wrote:
>> On Thu, 2012-03-08 at 15:42 -0500, J. Bruce Fields wrote:
>>> On Thu, Mar 08, 2012 at 03:23:34PM -0500, Olga Kornievskaia wrote:
>>>> On Thu, Mar 8, 2012 at 1:15 PM, Myklebust, Trond
>>>> <Trond.Myklebust@xxxxxxxxxx> wrote:
>>>>> On Thu, 2012-03-08 at 12:52 -0500, Olga Kornievskaia wrote:
>>>>>> wouldn't it be better for you to proactively return a read delegation
>>>>>> then unnecessarily erroring?
>>>>>
>>>>> If nobody else holds a delegation, then the NFS client is actually
>>>>> allowed to keep its read delegation while writing to the file. It does
>>>>> admittedly need to request an OPEN stateid for write in that case...
>>>>> (See section 10.4 of RFC3530bis draft 16)
>>>>
>>>> If we both agree that there has to be a request for an open stateid for
>>>> a write, then instead of returning the read delegation if the client receives
>>>> err_openmode (when it send the request with read delegation stateid
>>>> as you said per 3560bis), can't the client resend the setattr with the open
>>>> stateid? The ordering of the stateid usage is a "should" and not a "must".
>>>>
>>>> In rfc5661, it really doesn't make sense to ever send a setattr with
>>>> a read delegation stateid. According to 9.1.2, the server "MUST" return
>>>> err_open_mode" error in that case.
>>>>
>>>> I gather you are in this case dealing with 4.0 delegations. But I wonder
>>>> if you'll do something else for 4.1 delegation then?
>>>
>>> 3530bis has the same language ("...must verify that the access mode
>>> allows writing and return an NFS4ERR_OPENMODE error if it does not").
>>
>> OK, so we shouldn't send the delegation stateid either for v4 or v4.1.
>> However should we pre-emptively return the delegation? I've been
>> assuming not.
>
> The server's only legal option is to recall it, so it seems odd not to
> pre-emptively return--
Also from the client that sent the setattr? I, as Trond, understood that
the read delegation must be recalled from all clients but the one
doing the setattr/write.
other wise what does it mean:
"allowed to keep its read delegation while writing to the file"
I think the server should filter out it's global recall to dis-include
the caller.
Though I agree that the client must get a writable stateid for
setattr, and should not use it's read-delegation-stateid
> but as you say there's nothing to prevent the
> server from then handing one right back, possibly before you get a
> chance to send the setattr.
>
The above recall filtering will solve that. I know that in layout recalls
we have such facility, and we actually use it in RAID5 exofs.
> (And the linux server might well do that--maybe it should have some
> heuristic not to hand out a delegation that was just returned--not so
> much for this case as just because a return is a sign that the
> delegation isn't useful to that client.)
>
Thanks
Boaz
> --b.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
