PCI passthrough corrupting data?

[Thor Lancelot Simon <tls%rek.tjls.com@localhost>]

I have a NetBSD/amd64 (5.99.11 as of a few days ago) dom0 with
NetBSD/i386 domUs, each of which has a passed-through Intel gigabit
ethernet controller on wm0.  That physical interface has two vlans
stacked on it, vlan5 and vlan3.

If configure wm0 ip4csum tcp4csum udp4csum tso4 and vlan5 (by which
the system reaches its default gateway) the same way, I see data
corruption on TCP streams.  For example, when I try to pkg_add from
ftp.netbsd.org, for large .tgz files I often get MD5 checksum mismatches
and pkg_add aborts (after a few tries it will succeed).

If I turn off tso4 and tcp4csum on wm0 and the vlans stacked on top,
things work better but TCP reports receive checksum errors.

Is anyone else using PCI passthrough in production?  Are you seeing
any signs of data corruption?  This system is a Dell server with Xeon
CPUs with cache ECC and ECC memory.

-- 
Thor Lancelot Simon                                        
tls%rek.tjls.com@localhost
    "Even experienced UNIX users occasionally enter rm *.* at the UNIX
     prompt only to realize too late that they have removed the wrong
     segment of the directory structure." - Microsoft WSS whitepaper