tmpfs race conditions

[Joerg Sonnenberger <joerg%britannica.bec.de@localhost>]

Hi all,
I have a build machine building pkgsrc where I see some very strange
issues. The individual chroots share a read-only null-mounted base
directory with a number of tmpfs instances on top for most of the
writeable parts. Symptoms of the some data corruption or race
conditions:

(1) checksum failures for the distfiles after fetching them from the
local mirror via HTTP. An example from the currently running build:

  => Checksum SHA1 OK for wxPython-src-2.8.12.0.tar.bz2
  => Checksum RMD160 OK for wxPython-src-2.8.12.0.tar.bz2
  => Checksum SHA1 OK for wxPython-src-2.8.12.0.tar.bz2
  checksum: Checksum RMD160 mismatch for wxPython-src-2.8.12.0.tar.bz2

(2) libreoffice failing with a syntax error in one of the message files.
This appears to be a pretty regular error case in one of the three
libreoffice builds and it tends to stick around for a while, if it
happened.

(3) Concurrent build fails with errors noone else is seeing and that
often don't make sense when looking at the dependencies. Don't have a
specific sample at hand.

(4) automoc crashes at exit time, hitting a libc assertion about
pthread_specific with unset destructor. The code has been patched so
that this should simply not be possible. I haven't seen this one lately,
but that doesn't mean anything. 

Any idea what to look for? I've been running memtest86+ for a while to
rule out hardware issue, nothing found in two days at least.

Joerg