Re: PATCH [v3 0/2] NFSD: use CB_GETATTR to handle GETATTR conflict with write delegation

[Chuck Lever <chuck.lever@oracle.com>]

On Sat, Feb 17, 2024 at 12:34:45AM +0100, Dan Shelton wrote:
> On Thu, 15 Feb 2024 at 23:11, Dai Ngo <dai.ngo@oracle.com> wrote:
> >
> > Currently GETATTR conflict with a write delegation is handled by
> > recalling the delegation before replying to the GETATTR.
> >
> > This patch series add supports for CB_GETATTR callback to get the latest
> > change_info and size information of the file from the client that holds
> > the delegation to reply to the GETATTR from the second client.
> >
> > NOTE: this patch series is mostly the same as the previous patches which
> > were backed out when un unrelated problem of NFSD server hang on reboot
> > was reported.
> >
> > The only difference is the wait_on_bit() in nfsd4_deleg_getattr_conflict was
> > replaced with wait_on_bit_timeout() with 30ms timeout to avoid a potential
> > DOS attack by exhausting NFSD kernel threads with GETATTR conflicts.
> 
> I have a concern about this static and very tiny timeout.
> What will happen if the ICMPv6 latency is well over 30ms, like 660ms
> (average 250mbit/s satellite latency)?

CB_GETATTR is an optimization for write delegation. Without
CB_GETATTR, or if the client does not respond within 30ms, the
server recalls the delegation. We expect no impact on clients
that connect on a high bandwidth-latency product link.

To lengthen that timeout would require the implementation of a
mechanism for NFSD to defer requests without tying up an NFSD
thread. So for the moment, the proposed CB_GETATTR implementation
will help fast local clients but should not negatively impact
remote clients, and we cannot in good faith provide a tunable
to extend that timeout beyond a few dozen milliseconds.


> Would that not ruin delegations?

As stated above, it should not impact write delegation, and Dai can
correct me if I'm wrong, but I believe CB_GETATTR is not used if the
server has offered a read delegation. NFSD implemented only read
delegation until very recently.

IIRC, there is instrumentation in v6.6 or v6.7's NFSD to measure
how often a CB_GETATTR might have been beneficial. I can provide
more detail when I'm in front of my desktop computer.


-- 
Chuck Lever