Re: [PATCH] NFSv4: Use exponential backoff delay for NFS4_ERRDELAY

["Myklebust, Trond" <Trond.Myklebust@netapp.com>]

On Apr 25, 2013, at 2:46 PM, "bfields@fieldses.org" <bfields@fieldses.org>
 wrote:

> On Thu, Apr 25, 2013 at 02:40:11PM -0400, Chuck Lever wrote:
>> 
>> On Apr 25, 2013, at 2:19 PM, "bfields@fieldses.org" <bfields@fieldses.org> wrote:
>> 
>>> On Thu, Apr 25, 2013 at 02:10:36PM +0000, Myklebust, Trond wrote:
>>>> On Thu, 2013-04-25 at 09:49 -0400, bfields@fieldses.org wrote:
>>>>> On Thu, Apr 25, 2013 at 01:30:58PM +0000, Myklebust, Trond wrote:
>>>>>> On Thu, 2013-04-25 at 09:29 -0400, bfields@fieldses.org wrote:
>>>>>> 
>>>>>>> My position is that we simply have no idea what order of magnitude even
>>>>>>> delay should be.  And that in such a situation exponential backoff such
>>>>>>> as implemented in the synchronous case seems the reasonable default as
>>>>>>> it guarantees at worst doubling the delay while still bounding the
>>>>>>> long-term average frequency of retries.
>>>>>> 
>>>>>> So we start with a 15 second delay, and then go to 60 seconds?
>>>>> 
>>>>> I agree that a server should normally be doing the wait on its own if
>>>>> the wait would be on the order of an rpc round trip.
>>>>> 
>>>>> So I'd be inclined to start with a delay that was an order of magnitude
>>>>> or two more than a round trip.
>>>>> 
>>>>> And I'd expect NFS isn't common on networks with 1-second latencies.
>>>>> 
>>>>> So the 1/10 second we're using in the synchronous case sounds closer to
>>>>> the right ballpark to me.
>>>> 
>>>> OK, then. Now all I need is actual motivation for changing the existing
>>>> code other than handwaving arguments about "polling is better than flat
>>>> waits".
>>>> What actual use cases are impacting us now, other than the AIX design
>>>> decision to force CLOSE to retry at least once before succeeding?
>>> 
>>> Nah, I've got nothing, and I agree that the AIX problem is there bug.
>>> 
>>> Just for fun I looked at re-checked the Linux server cases.  As far as I
>>> can tell they are:
>>> 
>>> 	- delegations: returned immediately on detection of any
>>> 	  conflict.  The current behavior in the sync case looks
>>> 	  reasonable to me.
>>> 	- allocation failures: not really sure it's the best error, but
>>> 	  it seems to be all the protocol offers.  We probably don't
>>> 	  care much what the client does in this case.
>>> 	- some rare cases that would probably indicate bugs (e.g.,
>>> 	  attempting to destroy a client while other rpc's from that
>>> 	  client are running.)  Again we don't care what the client does
>>> 	  here.
>>> 	- the 4.1 slot-inuse case.
>>> 
>>> We also by default map four errors (ETIMEDOUT, EAGAIN, EWOULDBLOCK,
>>> ENOMEM) to delay.  I thought I remembered one of those being used by
>>> some HFS system, but can't actually find an example now.  A quick grep
>>> doesn't show anything interesting.
>> 
>> It's worth mentioning that servers that have frozen state (say, in preparation for Transparent State Migration) may use NFS4ERR_DELAY to prevent clients from modifying open or lock state until that state has transitioned to a destination server.
> 
> I thought they'd decided they'll be forced to find a different way to do
> that?
> 
> (The issue being that it only works if you're using 4.1, and if the
> session state itself isn't part of the state to be transferred.
> Otherwise you're forced to modify the state anyway since NFS4ERR_DELAY
> is seqid-modifying.)

Either way, migration is not a performance-critical path that needs 1second or less response times on those NFS4ERR_DELAY replies.

Trond