Re: [PATCH 0/3] mountd: disabling turning off TCP listeners (v2)

[Jeff Layton <jlayton@redhat.com>]

On Fri, 25 Oct 2013 08:23:29 -0400
Steve Dickson <SteveD@redhat.com> wrote:

> 
> 
> On 25/10/13 07:43, Jeff Layton wrote:
> > On Thu, 24 Oct 2013 15:45:03 -0400
> > Steve Dickson <SteveD@redhat.com> wrote:
> > 
> >>
> >>
> >> On 24/10/13 14:45, Jeff Layton wrote:
> >>> On Thu, 24 Oct 2013 14:17:10 -0400
> >>> Steve Dickson <steved@redhat.com> wrote:
> >>>
> >>>> [ Here is the second try for these patches incorporating the code review..]
> >>>>
> >>>> Recently it was pointed out to me that the [-n | --no-tcp] flags 
> >>>> were broken in mountd. Sure enough they are and they broke 
> >>>> when nfs-utils moved to using libtirpc, which was years ago.
> >>>>
> >>>> Obviously nobody is using these flags since has not been 
> >>>> notice until now, but it seemed to me it no longer makes 
> >>>> any sense to have flags. We really want people to use TCP 
> >>>> so why should there be a way to turn it off? It should be
> >>>> the opposite... They should be able to turn off UDP listeners
> >>>> not TCP... 
> >>>>
> >>>>
> >>>> Steve Dickson (3):
> >>>>   mountd: Use protocol bit fields to turn protocols off.
> >>>>   mountd: Deprecate the ability to disable TCP listeners.
> >>>>   mountd: Add the ability to disable UDP listeners.
> >>>>
> >>>>  support/include/rpcmisc.h |  2 +-
> >>>>  support/nfs/rpcmisc.c     | 19 ++++++++++++++-----
> >>>>  support/nfs/svc_create.c  |  5 +++++
> >>>>  utils/mountd/mountd.c     | 17 ++++++++++++-----
> >>>>  utils/mountd/mountd.man   |  6 +++---
> >>>>  5 files changed, 35 insertions(+), 14 deletions(-)
> >>>>
> >>>
> >>> Sorry I'm coming in late on this...
> >> np... I was expecting more push back! ;-) 
> >>
> >>>
> >>> I don't think we want to remove the ability to disable TCP listeners.
> >>>
> >>> Why, you ask? We've been on a multi-year effort to move people to
> >>> NFSv4, and with that, there's no reason to have mountd listen on the
> >>> network at all.
> >> True...
> >>
> >>>
> >>> So personally, I think it would make sense to:
> >>>
> >>> a) allow people to disable listening on UDP in addition to TCP
> >> I see no reason whatsoever to turn off TCP listeners especially
> >> since that is the protocol of choice... something we have 
> >> be spouting about for years...  
> >>  
> > 
> > There are reasons to be able to turn off TCP listeners:
> > 
> > If you're running a NFSv4-only server, there's no reason to allow it to
> > listen on TCP _or_ UDP sockets. I think that sort of environment is
> > going to become more prevalent in the future, not less.
> I ideally it would be best not to have mountd at all on NFSv4-only server.
> Basically, have the kernel get its exports like it gets it ID mappings.
> Until that day comes, which I hope fill be soon, the TCP listener 
> only effects v3 mounts and we definitely want people to use TCP
> with v3. 
>  

We want people to use TCP for NFS protocol with v3. In general however,
we do *not* want them using TCP for the MNT protocol. We've had many
problems in the past with clients hitting reserved port exhaustion due
to using TCP for short-lived sockets to carry MNT traffic. This is the
reason that the client defaults to UDP for MNT traffic.

>  
> > 
> >>>
> >>> ...or...
> >>>
> >>> b) add an option that prevents it from listening on any sockets for a
> >>>    v4-only configuration
> >> In this case it would optimal to not even start mountd, unfortunately
> >> due to exports reasons, it not possible... but it should be!! :-) 
> >>   
> > 
> > Right, mountd has 2 jobs:
> > 
> > 1) respond to MNT protocol requests from clients
> > 
> > ...and...
> > 
> > 2) feed exports info to the kernel
> > 
> > For v4, you obviously don't need the first role, so being able to
> > disable network listeners is a good thing in such a configuration.
> Again, I would rather build an v4 only environment where mountd
> does not even run... 
> 

I'd be fine with that. But until that materializes we're stuck with
running mountd in some form on the server.

>  
> > 
> >>>
> >>> In addition, we generally do want people to use UDP for the MNT
> >>> protocol because it's less apt to cause issues with reserved port
> >>> exhaustion. Given that it'll continue to listen on a UDP socket by
> >>> default, that last point is less of an issue, but that might be a good
> >>> reason to rethink this whole plan.
> >>>
> >> I did think of this.... UDP is on by default... Is up the admin... 
> >>
> > 
> > That's good. I have no objection to adding an option to disable UDP
> > listeners if the admin chooses. I just think it would be best to fix
> > the ability to disable TCP listeners as well instead of removing it.
> > 


-- 
Jeff Layton <jlayton@redhat.com>