zero bytes in files written to NFS

[Thomas Klausner <wiz%NetBSD.org@localhost>]

Hi!

For a few months now (at least since November) I see broken files when
writing from NetBSD-current/amd64 to a Synology (Linux) file server
via NFS.  I don't have a completely reliable way to reproduce this,
but it happens quite a lot when I run pkgsrc/archivers/torrentzip on
large files (1-50GB) on file systems that are mounted via NFS. (The
torrentzip tool reads the complete zip archive and replaces it with
particular zlib settings and zip metadata.)

I thought it could be a problem with the NFS server, but I've replaced
the server hardware, and it still happens.

I thought it could be a bug in the torrentzip tool, but it also
happens (very rarely) when downloading files using filezilla (sftp)
from a remote server directly to the NFS mounted device.

The breakage always looks the same way - a block of zero (NUL) bytes
in the middle of the file. The amounts differ, but usually it's a
couple of kB of zeroes.

So far I have not noticed this issue on local file systems.

So I assume it's a bug in NFS and/or UVM.

Has anyone else seen something similar?
 Thomas