[OpenAFS] zero length core files with OpenAFS 1.6.5

Marc Dionne marc.c.dionne@gmail.com
Mon, 5 May 2014 18:45:34 -0400

On Mon, May 5, 2014 at 6:31 PM, Renata Maria Dart
<renata@slac.stanford.edu> wrote:
> Hi, we had a user report that he was only getting 0-length core files
> from his segfaulting executable when writing the core into AFS from a
> rhel6 64-bit linux system.  We have confirmed that under OpenAFS 1.6.5
> this appears to be so.  The executable can write a normal core file on
> that same linux system if writing to nfs or /tmp.  Is this a known
> issue that might be fixed in a later version of OpenAFS?
> A bit more information...we had a rhel6 64-bit linux system running
> OpenAFS 1.6.1 and a back-level linux kernel 2.6.32-358.23.2 that could
> successfully write a core file into AFS.  We upgraded the version of
> OpenAFS to 1.6.5 leaving the kernel at the same older level and now
> only get 0-length core files when writing to AFS.  The same failure
> occurs with OpenAFS 1.6.5 and newer linux kernels like
> 2.6.32-431.11.2.  Core files written to OpenAFS 1.6.2 are also fine.
> Thanks,
> Renata

There was a bug introduced between 1.6.2 and 1.6.3 that could cause a
problem with core dumps on certain kernels, this was fixed in 1.6.6.

Another potential issue is that the kernel will check that the owner
of the file matches the current uid - in case of a mismatch between
the AFS id and the local uid, the file will get created but no content
will be written, resulting in a 0-size core file.