Re: Race in shrink_cache

[Daniel Phillips <phillips@arcor.de>]

On Thursday 05 September 2002 09:07, Andrew Morton wrote:
> Daniel Phillips wrote:
> > 
> > ...
> > /*
> >  * We must not allow an anon page
> >  * with no buffers to be visible on
> >  * the LRU, so we unlock the page after
> >  * taking the lru lock
> >  */
> > 
> > That is, what's scary about an anon page without buffers?
> 
> ooop.  That's an akpm comment.  umm, err..
> 
> It solves this BUG:
> 
> http://www.cs.helsinki.fi/linux/linux-kernel/2001-37/0594.html
> 
> Around the 2.4.10 timeframe, Andrea started putting anon pages
> on the LRU.  Then he backed that out, then put it in again.  I
> think this comment dates from the time when anon pages were
> not on the LRU.  So there's a little window there where the
> page is unlocked, we've just dropped its swapdev buffers, the page is
> on the LRU and pagemap_lru_lock is not held.
> 
> So another CPU came in, found the page on the LRU, saw that it had
> no ->mapping and no ->buffers and went BUG.
> 
> The fix was to take pagemap_lru_lock before unlocking the page.
> 
> The comment is stale.

With the atomic_dec_and_lock strategy, the page would be freed immediately on 
the buffers being released, and with the lru=1 strategy it doesn't matter 
in terms of correctness whether the page ends up on the lru or not, so I was 
inclined not to worry about this anyway, still, when you a dire-looking 
comment like that...

You said something about your lru locking strategy in 2.5.33-mm2.  I have not
reverse engineered it yet, would you care to wax poetic?

-- 
Daniel