Re: [NFS] rpc.mountd crashes when extensively using netgroups

[Jeff Layton <jlayton@redhat.com>]

On Thu, 2 Aug 2007 12:05:36 -0400
"J. Bruce Fields" <bfields@fieldses.org> wrote:

> On Thu, Aug 02, 2007 at 11:32:55AM -0400, Jeff Layton wrote:
> > On Tue, 31 Jul 2007 10:48:24 -0400
> > "J. Bruce Fields" <bfields@fieldses.org> wrote:
> > > That's an interesting problem.  Thanks for the report!
> > > 
> > > I don't believe that long comma-delimited string actually has any
> > > meaning to the kernel--as far as the kernel is concerned, it's just an
> > > opaque object that will be passed back to mountd later (along with a
> > > path name) to get export options.
> > > 
> > > So I suppose that string could be replaced by a hash, or maybe even just
> > > by the ip address of the particular host--the disadvantage to the latter
> > > being that it would require the kernel to keep a separate export for
> > > each client address.
> > 
> > I started having a look at this today. The original patches that I
> > proposed to clean up the rmtab a few months ago also eliminated this
> > comma-delimited string. Neil had valid objections to it at the time,
> > but if we switched to using the IP address as a cache key like Bruce
> > describes then doing that becomes more reasonable.
> > 
> > The only downside I see is the one Bruce points out -- the size of
> > the kernel export cache would increase. I don't have a feel for whether
> > this is a show stopper, however.
> 
> Yeah, there might be some risk of solving that problem at the expense of
> people with tons of clients all matching *.example.com.  The actual
> export objects are actually pretty small--68 bytes, last I checked?
> You'd also need two upcalls to mountd per client (for ip address and
> export, as opposed to just ip address).  I don't know what other costs
> there might be.
> 

Yeah, that could make using the ipaddr too costly...

> What about the idea of hashing the comma-delimited list and passing
> that?  You'd want hash large enough to be collision-free--might as well
> use some cryptographic hash, I guess, though it may be mild overkill.
> Is there any reason why that wouldn't work?
> 

Hashing would certainly work. The only issue though is that this
doesn't take care of the second problem, which is that NFSCLNT_IDMAX is
too small for some environments. So this change would need to be
coupled with an increase in this hardcoded value, change m_hostname to
be dynamically allocated, or change the comma-delimited string to some
other scheme entirely (maybe a linked list of some sort?).

As far as the hashing scheme, I guess we could use the glibc crypt()
routine that does md5 hashing. I'd like to avoid adding dependencies
on new libs if possible.

> Could you remind me what problems you were trying to fix with your rmtab
> cleanup?
>

Those patches were really intended to get showmount -a to look correct.
My first pass at it changed the m_hostname field to hold an actual
hostname, got rid of the comma-separated strings and just had mountd do
repeated calls to client_check to check the host's validity.

This was problematic though with a multi-homed client though since the
export table key then became a hostname rather than the comma-delimited
string and more than 1 ip addr could resolve to it. If we use the ipaddr
as a key, then this problem goes away, but then we have the other issues
you've detailed above.

-- 
Jeff Layton <jlayton@redhat.com>