close() of an flock'd file is not atomic

[Peter Holm]

On Mon, Jul 09, 2012 at 04:48:32PM -0400, John Baldwin wrote:
> On Monday, July 09, 2012 4:40:07 pm Konstantin Belousov wrote:
> > On Mon, Jul 09, 2012 at 11:38:15AM -0400, John Baldwin wrote:
> > > Here now is the tested version of the actual fix after the vn_open_vnode()
> > > changes were committed.  This is hopefully easier to parse now.
> > > 
> > > http://www.FreeBSD.org/~jhb/patches/flock_open_close4.patch
> > 
> > Do you need atomic op to set FHASLOCK in vn_open_cred ? I do not think
> > *fp can be shared with other thread there.
> 
> Oh, that's true.  I had just preserved it from the original code.
> 
> > I thought that vrele() call in vn_closefile() would need a
> > vn_start_write() or vn_start_secondary_write() dance around it, but
> > now I believe it is not needed, since ufs_inactive() handles start of
> > secondary writes on its own. Still, it would be good if Peter could test
> > the patch with snapshotting load just be to safe there.
> 
> Ok.  I'm happy to have pho@ test it, but the test will have to use file 
> locking along with snapshots to exercise this case.
> 

Verified your scenario on a pristine head and it fails like this:

$ uname -a
FreeBSD x4.osted.lan 10.0-CURRENT FreeBSD 10.0-CURRENT #0 r234951
$ /usr/bin/time -h ./flock_open_close.sh 
flock_open_close: execv(/mnt/test): Text file busy
FAIL
        3,79s real              0,24s user              0,78s sys
$ 

Not a problem with your patch. The patch has further been stress tested
for 24 hours without any problems showing up.

- Peter