Dave Chinner wrote:
> On Thu, Jun 05, 2014 at 09:45:26PM +0900, Tetsuo Handa wrote:
> > This means that, under rare circumstances, it is possible that all processes
> > other than kswapd are trapped into too_many_isolated()/congestion_wait()
> > loop
> > while kswapd is sleeping because this loop assumes that somebody else shall
> > wake up kswapd and kswapd shall perform operations for making
> > too_many_isolated() to return 0. However, we cannot guarantee that kswapd is
> > waken by somebody nor kswapd is not blocked by blocking operations inside
> > shrinker functions (e.g. mutex_lock()).
> So what you are saying is that kswapd is having problems with
> getting blocked on locks held by processes in direct reclaim?
> What are the stack traces that demonstrate such a dependency loop?
If a normal task's GFP_KERNEL memory allocation called a shrinker function and
the shrinker function does GFP_WAIT-able allocation with a mutex held, there
is a possibility that kswapd is waken up due to GFP_WAIT-able allocation and
kswapd calls the shrinker function, and the kswapd is blocked at trying to
hold the same mutex inside the shrinker function, isn't it?
Since ttm_dma_pool_shrink_count()/ttm_dma_pool_shrink_scan() holds a mutex
and ttm_dma_pool_shrink_scan() does GFP_WAIT-able allocation, I think such
a dependency loop is possible.