[PATCH] mm, vmscan: Warn about possible deadlock at shirink_inactive_list

[Dave Chinner]

On Mon, Sep 21, 2015 at 08:09:54PM +0900, Tetsuo Handa wrote:
> This is a difficult-to-trigger silent hang up bug.
> 
> The kswapd is allowed to bypass too_many_isolated() check in
> shrink_inactive_list(). But the kswapd can be blocked by locks in
> shrink_page_list() in shrink_inactive_list(). If the task which is
> blocking the kswapd is trying to allocate memory with the locks held,
> it forms memory reclaim deadlock.

It's a known problem in XFS and I'm currently working on patches to
fix it by hoisting the memory allocations outside of the CIL context
lock.

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com