Re: [PATCH RFC 1/1] nfsd: Initial implementation of NFSv4 Courteous Server

[Bruce Fields <bfields@fieldses.org>]

On Wed, Jun 16, 2021 at 07:29:37PM +0000, Chuck Lever III wrote:
> 
> 
> > On Jun 16, 2021, at 3:25 PM, Dai Ngo <dai.ngo@oracle.com> wrote:
> > 
> > On 6/16/21 9:32 AM, Chuck Lever III wrote:
> >>> On Jun 16, 2021, at 12:02 PM, J. Bruce Fields <bfields@fieldses.org> wrote:
> >>> 
> >>> On Thu, Jun 03, 2021 at 02:14:38PM -0400, Dai Ngo wrote:
> >>>> . instead of destroy the client anf all its state on conflict, only destroy
> >>>> the state that is conflicted with the current request.
> >>> The other todos I think have to be done before we merge, but this one I
> >>> think can wait.
> >> I agree on both points: this one can wait, but the others
> >> should be done before merge.
> > 
> > yes, will do.
> > 
> >> 
> >> 
> >>>> . destroy the COURTESY_CLIENT either after a fixed period of time to release
> >>>> resources or as reacting to memory pressure.
> >>> I think we need something here, but it can be pretty simple.
> >> We should work out a policy now.
> >> 
> >> A lower bound is good to have. Keep courtesy clients at least
> >> this long. Average network partition length times two as a shot
> >> in the dark. Or it could be N times the lease expiry time.
> >> 
> >> An upper bound is harder to guess at. Obviously these things
> >> will go away when the server reboots. The laundromat could
> >> handle this sooner. However using a shrinker might be nicer and
> >> more Linux-y, keeping the clients as long as practical, without
> >> the need for adding another administrative setting.
> > 
> > Can we start out with a simple 12 or 24 hours to accommodate long
> > network outages for this phase?
> 
> Sure. Let's go with 24 hours.
> 
> Bill suggested adding a "clear_locks" like mechanism that could be
> used to throw out all courteous clients at once. Maybe another
> phase 2 project!

For what it's worth, you can forcibly expire a client by writing
"expire" to /proc/fs/nfsd/client/xxx/ctl.  So it shouldn't be hard to
script this, if we add some kind of "courtesy" flag to
client_info_show() and/or a number of seconds since the most recent
renew.

Maybe adding a command like "expire_if_courtesy" would also simplify
that and avoid a race where the renew comes in simultaneously with the
expire command.

Or we could just add a single call to clear all courtesy clients.  But
the per-client approach would allow more flexibility if you wanted (e.g.
to throw out only clients over a certain age).

--b.