Re: [PATCH v2 3/3] NFS: Fix a memory leak in nfs_readdir

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 1 Dec 2010 17:42:08 -0800 Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Wed, Dec 1, 2010 at 5:22 PM, Andrew Morton <akpm@linux-foundation.org> wrote:
> >
> > What we're talking about is races against memory reclaim, unmount, etc.
> 
> Ahh. Those I can believe in. Although I think they'd almost
> incidentally be fixed by making inode freeing (which is where the
> 'struct address_space' is embedded) RCU-safe, which we're going to do
> anyway in 38. Then we could make the vmscan code just be a rcu-read
> section.

I didn't know that aspect of it.  It will be nice to plug this race -
it's been there for so long because nobody was able to think of an
acceptable way of fixing it by direct means (synchronous locking,
refcounting, etc).  Taking a ref on the inode doesn't work, because we
can't run iput_final() in direct-reclaim contexts (lock ordering snafus).

vmscan is the problematic path - I _think_ all other code paths which
remove pagecache have an inode ref.  But this assumes that
inode->i_mapping points at inode->i_data!  Need to think about the
situation where it points at a different inode's i_data - in that case
these callers may have a ref on the wrong inode.

> Of course,  I do think the race is basically impossible to hit in
> practice regardless.

Actually I was able to hit the race back in late 2.5 or thereabouts. 
Really massive memory pressure caused vmscan->icache_shrinker to free
the inode/address_space while another CPU in vmscan was playing with the
address_space.  That was quite a debugging session ;)