On Wednesday 18 March 2009 18:13:13 Christoph Hellwig wrote:
> On Mon, Mar 16, 2009 at 09:19:21AM +1100, Dave Chinner wrote:
> > This is a VM problem where it calls fput() with the mmap_sem() held
> > in remove_vma(). It makes the incorrect assumption that filesystems
> > will never use the same lock in the IO path and the inode release path.
> > This can deadlock if you are really unlucky.
> I really wonder why other filesystems haven't hit this yet. Any chance
> we can get the fput moved out of mmap_sem to get rid of this class of
Yes I don't think there is any reason against holding the file refcount
a little longer, but it can be pretty nasty to do in practice because
of deep call chains and sometimes multiple fputs within a given lock
Possibly the easiest and quickest way will be to move aio's deferred
__fput into a usable API and try that. If there are any performance
problems, then we can try to move those fputs out from the mmap_sem
one at a time (eg. I don't expect fput for vma replacement/merging to
often cause the refcount to reach 0, and this is one of the harder
places; wheras fput for a simple munmap might be more often causing
refcount to reach 0 but it should be simpler to move out of mmap_sem
if it is a problem).