Hi, Stefan: Thanks for your time investigating this bug. Op Sat, Aug 16, 2014 at 09:55:11PM +0200 schreef Stefan Fritsch: > On Friday 27 June 2014 11:37:18, Joost van Baal-Ilić wrote: > > While libapr1 defaults to fcntl() locking it also supports flock(), > > which does not have the problems outlined above. A patch is > > attached which makes libapr1 use flock() even if fcntl() locking is > > available. > > flock does not support locking on NFS (at least according to its man > page), while fcntl does. I am undecided which is worse. > > Since there does not seem to be a good solution, I am downgrading the > severity of this bug. The default behavior for Linux NFS clients is to convert flock() calls into fcntl(). See nfs(5), in particular the local_lock option (the default for that option is ‘none’). To summarize: - fcntl() locking is broken for everyone (even if you're using a non-threaded mpm you will still run into problems if you open the locked file twice); - flock() is broken only for people using NFS (rare) AND a non-Linux kernel (very rare, in Debian at least). Therefore the proposed solution is good, imho. Leaving severity to your judgement, nevertheless. Bye, Joost
Attachment:
signature.asc
Description: Digital signature