ceph vfs

[Jeremy Allison]

On Tue, Apr 16, 2013 at 11:50:53AM -0700, Jeremy Allison wrote:
> On Tue, Apr 16, 2013 at 10:13:46AM -0700, Jeremy Allison wrote:
> > On Tue, Apr 16, 2013 at 10:51:13AM -0500, Sam Lang wrote:
> > > I still have one failing test from smbtorture on the ceph vfs module.
> > > base.open.ctemp test fails with EPERM.  It looks like reply_ctemp() in
> > > source3/smbd/reply.c is doing a mkstemp with the path returned by
> > > realpath, so in the ceph case, its doing:  mkstemp("//TMXXXXXX")
> > > 
> > > Unsurprisingly, that fails.  On a local filesystem (such as a
> > > directory like /tmp/foo), realpath resolves the path to
> > > /tmp/foo/TMXXXXXX.  But with ceph, no separate mountpoint exists
> > > locally.  How is ctemp meant to be handled by backend vfs modules that
> > > don't have a separate mount running on the smbd server?
> > 
> > Oh, you just found a bug in the VFS :-). Thanks !
> > 
> > reply_ctemp() is a *really* old DOS-style call
> > that no modern clients do.
> 
> Just to be clear, there are vendors shipping
> cloud gateway products that also have no local
> mounts that have never run into this problem,
> so it's not an issue in actual customer deployments.

FYI. I have a patch that fixes this issue
by removing the mkstemp call and using
only Samba VFS calls.

Jeremy.