On Tue, 2021-11-30 at 11:05 -0500, Bruce Fields wrote:

On Tue, Nov 30, 2021 at 03:36:43PM +0000, Chuck Lever III wrote:

quoted

I am a little concerned that we are trying to optimize a case
that won't happen during practice. pynfs does not reflect any
kind of realistic or reasonable client behavior -- it's designed
to test very specific server operations.

I wonder how hard this problem would be to hit in normal use.  I
mean, a
few hundred or a thousand clients doesn't sound that crazy.  This
case
depends on an open deny, but you could hit the same problem with file
locks.  Would it be that weird to have a client trying to get a write
lock on a file read-locked by a bunch of other clients?

That's a scenario that is subject to starvation problems anyway.
Particularly so on NFSv4.0, which lacks CB_NOTIFY_LOCK.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com