Race in NFS in 6.0-RC1?

[Danny Braniss]

> I've started using NFS in 6.0 a little more heavily lately, as since the
> em(4) wedge has been fixed I can actually use it reliably.
> Unfortunately there appears to be a problem.  Twice, now, in less than
> 24 hours the client has paniced under load.  Both times it was building
> OpenOffice in an NFS-mounted /usr/ports.  In case it matters, it's a
> soft mount from another 6.0 box over an em(4) interface with an MTU of
> 9000.
> 
> Both times it was a panic from vnlru while trying to flush a vnode and
> both times it was a null-pointer dereference in nfs_putpages() at
> nfs_bio.c:301.  In both cases vp->v_data was null.  The vnode itself
> looks fine to my eyes, although there may well be FreeBSD-specific
> subtleties that I'm missing.  I've just entered a PR for this problem,
> kern/87967.  I'll keep the cores around; if anyone wants more
> information from them, let me know.  As may be apparent, I can reproduce
> this fairly easily, although it takes a few minutes for it to trigger.
> 
> The worrying thing about this is, in fact, its reproducibility.

This looks very similar to a problem we have with a 5.4 box running samba,
it has an em(4), no jumbo packets, but is heavely doing nfs - the files
are on a netapp filer. the problem is not easely reprodusable, but it happens.

danny