Re: [PATCH 00/14] SUNRPC: clean up server thread management.

["J. Bruce Fields" <bfields@fieldses.org>]

On Sun, Nov 21, 2021 at 10:38:40PM -0500, J. Bruce Fields wrote:
> On Sun, Nov 21, 2021 at 09:37:53PM -0500, J. Bruce Fields wrote:
> > On Mon, Nov 22, 2021 at 12:13:08PM +1100, NeilBrown wrote:
> > > On Mon, 22 Nov 2021, J. Bruce Fields wrote:
> > > > On Sun, Nov 21, 2021 at 07:56:39PM -0500, J. Bruce Fields wrote:
> > > > > On Mon, Nov 22, 2021 at 10:50:34AM +1100, NeilBrown wrote:
> > > > > > On Thu, 18 Nov 2021, J. Bruce Fields wrote:
> > > > > > > On Wed, Nov 17, 2021 at 11:46:49AM +1100, NeilBrown wrote:
> > > > > > > > I have a dream of making nfsd threads start and stop dynamically.
> > > > > > > 
> > > > > > > It's a good dream!
> > > > > > > 
> > > > > > > I haven't had a chance to look at these at all yet, I just kicked off
> > > > > > > tests to run overnight, and woke up to the below.
> > > > > > > 
> > > > > > > This happened on the client, probably the first time it attempted to do
> > > > > > > an nfsv4 mount, so something went wrong with setup of the callback
> > > > > > > server.
> > > > > > 
> > > > > > I cannot reproduce this and cannot see any way it could possible happen.
> > > > > 
> > > > > Huh.  Well, it's possible I mixed up the results somehow.  I'll see if I
> > > > > can reproduce tonight or tomorrow.
> > > > > 
> > > > > > Could you please confirm the patches were applied on a vanilla 5.1.6-rc1
> > > > > > kernel, and that you don't have the "pool_mode" module parameter set.
> > > > > 
> > > > > /sys/module/sunrpc/parameters/pool_mode is "global", the default.
> > > > 
> > > > Oh, and yes, this is what I was testing, should just be 5.16-rc1 plus
> > > > your 14 patches:
> > > > 
> > > > 	http://git.linux-nfs.org/?p=bfields/linux-topics.git;a=shortlog;h=659e13af1f8702776704676937932f332265d85e
> > 
> > OK, tried again and it did indeed reproduce in the same spot.
> > 
> > > I did find a possible problem.  Very first patch.
> > > in fs/nfsd/nfsctl.c, in _write_ports_addfd()
> > >   if (!err && !nn->nfsd_serv->sv_nrthreads && !xchg(&nn->keep_active, 1))
> > > 
> > > should be "err >= 0" rather than "!err".  That could result in a
> > > use-after free, which can make anything explode.
> > > If not too much trouble, could you just tweek that line and see what
> > > happens?
> > 
> > Like the following?  Same divide-by-zero, I'm afraid.
> 
> Hm, playing with reproducer; it takes more than one mount.  My simplest
> reproducer is:
> 
> 	mount -overs=3 server:/path /mnt/
> 	umount /mnt/
> 	mount -overs=4.0 server:/path /mnt/
> 
> ... and the client crashes here.

Also: the problem starts with the last patch ("lockd: use
svc_set_num_threads() for thread start and stop").

--b.