Re: page allocator regression on nommu

[Paul Mundt <lethal@linux-sh.org>]

On Tue, Sep 01, 2009 at 02:46:45PM +0100, David Howells wrote:
> Paul Mundt <lethal@linux-sh.org> wrote:
> 
> > Yeah, that looks a bit suspect. __put_nommu_region() is safe to be called
> > without a call to add_nommu_region(), but we happen to trip over the
> > BUG_ON() in this case because we've never made a single addition to the
> > region tree.
> > 
> > We probably ought to just up_write() and return if nommu_region_tree ==
> > RB_ROOT, which is what I'll do unless David objects.
> 
> I think that's the wrong thing to do.  I think we're better moving the call to
> add_nommu_region() to above the "/* set up the mapping */" comment.  We hold
> the region semaphore at this point, so the fact that it winds up in the tree
> briefly won't cause a race, and it means __put_nommu_region() can be used with
> impunity to correctly clean up.
> 
[snip]

> From: David Howells <dhowells@redhat.com>
> Subject: [PATCH] NOMMU: Fix error handling in do_mmap_pgoff()
> 
> Fix the error handling in do_mmap_pgoff().  If do_mmap_shared_file() or
> do_mmap_private() fail, we jump to the error_put_region label at which point we
> cann __put_nommu_region() on the region - but we haven't yet added the region
> to the tree, and so __put_nommu_region() may BUG because the region tree is
> empty or it may corrupt the region tree.
> 
> To get around this, we can afford to add the region to the region tree before
> calling do_mmap_shared_file() or do_mmap_private() as we keep nommu_region_sem
> write-locked, so no-one can race with us by seeing a transient region.
> 
> Signed-off-by: David Howells <dhowells@redhat.com>

Agreed, that does look cleaner. After playing around with it a bit, I concede
that the BUG_ON() is definitely worth preserving. :-)

Acked-by: Paul Mundt <lethal@linux-sh.org>