-current NFSv2 and NFSv3 issues

Hi all,

  We're seeing some data corruption and performance problems here with
NFSv3/TCP on a machine running -current NFS client against a 4.8 NFS
server.

  The problems can be reproduced by doing a 'make buildworld' over
NFS.  With NFSv2 & UDP we could buildworld with -j8 and the build
would usually - but not always - complete successfully. The times when
it did not appeared to possibly be a result of the mtime.tv_usec not
being checked for files over NFS (ie, one build step creates a .depend
file, and another step tries to use it before it "appears" over NFS).

  We decided to try NFSv3/TCP to see if we could get better
performance.  However with buildworld and -j8, we reliably see gcc or
some other toolchain component coredump during the build. With -j1
things complete succesfully, but buildworld -j1 of -current takes 5
hours over dedicated 100baseT network, and the ethernet never gets
even close to being saturated, so the underlying network transport
doesn't seem to be the bottleneck.

  We appear to have sufficient nfsd's & nfsiod's in that they don't
all seem to be incurring appreciable CPU time.

Any clues as to the data corruption issue? Should we expect NFSv3 over
TCP to outperform NFSv2 over UDP?

Thanks,
Ben