Re: [nfsv4] review of "draft-ietf-nfsv4-layrec-00"

[Thomas Haynes <[e-mail]>]

> On Mar 1, 2024, at 3:09 AM, Jeff Layton <[e-mail]> wrote:
> 
> On Thu, 2024-02-29 at 20:45 -0800, Rick Macklem wrote:
>> On Thu, Feb 29, 2024 at 8:09 AM Jeff Layton <[e-mail]> wrote:
>>> 
>>> Most of this document looks fine to me, but it has this statement in
>>> section 2:
>>> 
>>> "The metadata server MUST NOT have been resilvering the file such
>>> that it has a different layout (set of mirror instances) than the
>>> client before the restart of the metadata server."
>>> 
>>> The rationale for this seems clear, but it seems like resilvering the
>>> file while the clients hold any sort of RW layout would be a bad idea,
>>> no?
>> I think this is a good point. It seems to me that, although the MDS can
>> use the LayoutReturn with stateid of all zeros to decide to resilver a file,
>> the actual resilvering cannot start until the MDS's grace period has ended.
>> 
>> (Until the grace period ends, the rebooted MDS cannot know that no client
>> is still using a layout, I think? If the MDS has a record of all clients that
>> held a layout for the file before the reboot, then it needs to wait until those
>> clients complete reclaim, but it may be simpler to just delay resilvering
>> until grace is over.)
>> 
>> Lets see what Tom thinks?
>> 
> 
> Yep. In fact, it says this in the doc too:
> 
> "Further, the metadata server MUST NOT start a new resilvering of the
> file during the grace period."


Thanks for pointing this one out - caused me to delve deeper into my code than I wanted. :-)


> 
> It might be simpler to lay out the conditions under which you _can_
> start resilvering, but maybe that's too restrictive.


Yes, I’ll revisit this whole issue and resolve it.


> -- 
> Jeff Layton <[e-mail] <mailto:[e-mail]>>