Re: [RFC] Patch to allow distributed flock

[Trond Myklebust <trond.myklebust@fys.uio.no>]

På to , 24/06/2004 klokka 19:10, skreiv Ken Preslan:
> Hi,
> 
> I'd like to start a discussion about changing the VFS so it allows
> flocks to be enforced between machines in a cluster filesystem (such as
> GFS).  The purpose of GFS it so allow local filesystem semantics
> to a filesystem shared between the nodes of a cluster of tightly-coupled
> machines.  As such, flock is probably expected to work across the cluster.
> 
> What are everyone's thoughts on a patch such as this?

If you defer updating the VFS until after the ->lock() call returns,
then it makes it difficult to protect yourself against races (as I
argued about the POSIX lock interface on the list yesterday).

If you have the underlying filesystem call flock_lock_file() itself,
then that gives it the freedom to implement its own locking scheme
around that call.
For instance NFS has a thread that is supposed to reclaim locks if the
server reboots. We take a non-exclusive lock on an rwsem to ensure that
we block it while there are outstanding locking RPC calls, however that
rwsem has to be released before we return from the ->lock() call, and so
there exists a race after the rwsem was released until the
inode->i_flock list is updated.

Cheers,
  Trond