close() of an flock'd file is not atomic
John Baldwin
jhb at freebsd.org
Fri Mar 9 19:30:00 UTC 2012
On Thursday, March 08, 2012 5:39:19 pm Konstantin Belousov wrote:
> On Thu, Mar 08, 2012 at 03:39:07PM -0500, John Baldwin wrote:
> > On Wednesday, March 07, 2012 1:18:07 pm John Baldwin wrote:
> > > So I ran into this problem at work. Suppose you have a process that opens a
> > > read-write file descriptor with O_EXLOCK (so it has an flock()). It then
> > > writes out a binary into that file. Another process wants to execve() the
> > > file when it is ready, so it opens the file with O_EXLOCK (or O_SHLOCK), and
> > > will call execve() once it has locked the file. In theory, what should happen
> > > is that the second process should wait until the first process has finished
> > > and called close(). In practice what happens is that I occasionally see the
> > > second process fail with ETXTBUSY.
> > >
> > > The bug is that the vn_closefile() does the VOP_ADVLOCK() to unlock the file
> > > separately from the call to vn_close() which drops the writecount. Thus, the
> > > second process can do an open() and flock() of the file and subsequently call
> > > execve() after the first process has done the VOP_ADVLOCK(), but before it
> > > calls into vn_close(). In fact, since vn_close() requires a write lock on the
> > > vnode, this turns out to not be too hard to reproduce at all. Below is a
> > > simple test program that reproduces this constantly. To use, copy /bin/test
> > > to some other file (e.g. /tmp/foo) and make it writable (chmod a+w), then run
> > > ./flock_close_race /tmp/foo.
> > >
> > > The "fix" I came up with is to defer calling VOP_ADVLOCK() to release the lock
> > > until after vn_close() executes. However, even with that fix applied, my test
> > > case still fails. Now it is because open() with a given lock flag is
> > > non-atomic in that the open(O_RDWR) will call vn_open() and bump v_writecount
> > > before it blocks on the lock due to O_EXLOCK, so even though the 'exec_child'
> > > process has the fd locked, the writecount can still be bumped. One gross hack
> > > would be to defer the bump of the writecount to the caller of vn_open() if the
> > > caller passes in O_EXLOCK or O_SHLOCK, but that's a really gross kludge, plus
> > > it doesn't actually work. I ended up moving acquiring the lock into
> > > vn_open_cred(). The current patch I'm testing has both of these approaches,
> > > but the first one is #if 0'd out, and the second is #if 1'd.
> > >
> > > http://www.freebsd.org/~jhb/patches/flock_open_close.patch
> >
> > Based on some feedback from Konstantin, I've fixed some issues in the failure
> > path handling for VOP_ADVLOCK(). I've also removed the #if 0'd code mentioned
> > above, so the patch is now the actual change that I'm testing. So far it
> > handles both my workload at work and my test program without any issues.
>
> I think a comment is needed for a reason to call vn_writechk() second time.
Fixed.
> Could you, please, point me, where the FHASLOCK is set for O_EXLOCK | O_SHLOCK
> case in the patched kernel ?
It wasn't. :( I wonder how this was even working since close shouldn't have
been unlocking. I'll need to do some more testing. BTW, I ran into fhopen()
and found that I would need to put all this same logic into that, so I've split
the common code from fhopen() and vn_open_cred() into a new vn_open_vnode().
I think in general it improves both sets of code.
I'll upate the patch once I've done some more testing.
--
John Baldwin
More information about the freebsd-fs
mailing list