[brandz-discuss] Jumbo frames

[Edward Pilatowicz]

the first thing that comes to mind is that linux default to nfsv3 where
as the gobal zone will default to nfsv4.  so the first thing that i
would try is when you mount the nfs filesystem in the global zone be
sure to specify "-o vers=3" to the mount command.

another issue to be aware of is that linux actually has different default
mount parameters (that are passed to the mount system call) than solaris
does.  you could compare the nfs mount options by running the following
command in the global zone:
	echo "::nfs_vfs -v" | mdb -k | less

finally, comparing some snoop traces of the nfs traffic between
the global zone and the nfs server and the linux zone and the nfs
server might reveal something.

there are a few more comments below...

On Sun, Nov 26, 2006 at 10:55:19AM -0800, Scott L. Burson wrote:
> Hi,
>
> I am observing a strange little problem with my Linux zone.  (I am
> running build 49.)
>
> The Linux zone NFS-mounts a server over the LAN.  I just installed a
> Gigabit Ethernet switch (was previously using 100Mb).  If I set the
> server to use jumbo frames (MTU = 9000), I get NFS timeouts in the
> Linux zone -- but not in the global zone, which has the same server
> mounted!  When I set the MTU on the server back to 1500, the symptom
> disappears.
>
> I know -- I wouldn't believe it either if I weren't seeing it :-)  It
> wouldn't be quite so weird (and wouldn't be a BrandZ question) if the
> global zone had the same problem, but it absolutely doesn't -- the
> symptom only occurs in the Linux zone.
>
> Notes:
>
> (1) When I say "NFS timeouts" I mean that accesses seem to hang, or
> are very slow, and I get the "NFS server not responding still trying"
> message.  And part of what clued me in to the MTU thing was that some
> accesses succeed immediately -- for instance, I can stat a single
> file just fine -- but it seems that anything that might cause the
> server to respond with a jumbo frame (e.g. running an executable off
> the server) has the problem.
>

hm.  my guess would be that operations that require larger packets
are the ones that are failing.

by any chance, are you doing operations that require file locking?

> (2) If it matters, the server is an old Linux 2.4 box.
>
> (3) Is there some way for the Linux zone to just inherit the NFS
> mount in question from the global zone, rather than having to mount
> it again?
>

no.  nfs mounts can not be shared between zones.
(the same goes for native zones as well.)

> (4) I note with some surprise that attempting to set the MTU on the
> local machine (in the global zone) fails with "invalid argument".
> Doesn't the bge driver support this??  (Not a BrandZ question,
> obviously)
>

hm.  i've never actually tried to do this but it seems like you should
be able to.  my first guess here would be that you're falling victim to
the super user friendly command line syntax for ifconfig.  i know that
i get tripped up by it all the time.  ;)

ed