Question on mountd

[Ian Lepore]

On Thu, 2017-08-10 at 19:31 +0200, Olavi Kumpulainen wrote:
> Thanks Ian,
> 
> Just a small follow up question - The reason to why -S works is that
> mountd cannot interrupt the execution thread of a NFS request from a
> client?
> Is this true in a multicore system as well?
> 

With -S, mountd sends a NFSSVC_SUSPENDNFSD command to nfsd, which I
presume makes nfsd finish any client requests in progress and stop
handling new client requests, then mountd manipulates the in-kernel
data, then it sends NFSSVC_RESUMENFSD.  I think the net effect is that
no client requests which rely on the in-kernel mount-export data can be
processed while mountd is manipulating the data.  You would think this
would be the default behavior, but maybe the delays are so bad for some
use cases that it was made optional with -S.

I should note I'm not an expert on this stuff, I'm making a lot of
assumptions from a little bit of looking at code.

-- Ian

> /Olavi
> 
> 
> > 
> > On 10 Aug 2017, at 17:51 , Ian Lepore <ian at FreeBSD.org> wrote:
> > 
> > On Thu, 2017-08-10 at 16:56 +0200, Olavi Kumpulainen wrote:
> > > 
> > > Hi guys,
> > > 
> > > I notice that mount SIGHUP’s mountd every time mount succeeds.
> > > The
> > > SIGHUP causes mountd to remount all exported directories.If this
> > > happens while a NFS-client is accessing a share, an access error
> > > may
> > > occur.
> > > For this purpose, there is an option to mount, -S, which locks
> > > nfsd
> > > while the remount is executing.
> > > 
> > > Can anyone of you share why mount needs to SIGHUP mountd in the
> > > first
> > > place? It would make sense if mountd is restarted whenever
> > > /etc/exports is modified, but always seems like overkill.
> > > 
> > Based on looking through the mountd code a bit... When a new
> > filesystem
> > is mounted, it may be mounted on one of the mount points listed in
> > /etc/exports.  If there was no fs mounted there previously, then
> > mountd
> > might have failed to set the in-kernel export attributes the last
> > time
> > it processed the exports file, so it has to do the processing again
> > after the mount to update the in-kernel export data.  It would be
> > really complex for mountd to try to figure out the minimal set of
> > "what
> > changed" after a mount succeeds, so it just completely reprocesses
> > the
> > exports file, first removing all export data from the kernel, then
> > re-
> > applying it all.
> > 
> > So, all in all, I think the right fix for this is to add
> > "mountd_flags="-r -S" to your rc.conf file (-r is a default from
> > /etc/defaults/rc.conf).
> > 
> > -- Ian
>