On Fri, Aug 09, 2019 at 07:30:00AM -0400, Mikulas Patocka wrote: > > > On Fri, 9 Aug 2019, Dave Chinner wrote: > > > And, FWIW, there's an argument to be made here that the underlying > > bug is dm_bufio_shrink_scan() blocking kswapd by waiting on IO > > completions while holding a mutex that other IO-level reclaim > > contexts require to make progress. > > > > Cheers, > > > > Dave. > > The IO-level reclaim contexts should use GFP_NOIO. If the dm-bufio > shrinker is called with GFP_NOIO, it cannot be blocked by kswapd, because: No, you misunderstand. I'm talking about blocking kswapd being wrong. i.e. Blocking kswapd in shrinkers causes problems because th ememory reclaim code does not expect kswapd to be arbitrarily delayed by waiting on IO. We've had this problem with the XFS inode cache shrinker for years, and there are many reports of extremely long reclaim latencies for both direct and kswapd reclaim that result from kswapd not making progress while waiting in shrinkers for IO to complete. The work I'm currently doing to fix this XFS problem can be found here: https://lore.kernel.org/linux-fsdevel/20190801021752.4986-1-david@xxxxxxxxxxxxx/ i.e. the point I'm making is that waiting for IO in kswapd reclaim context is considered harmful - kswapd context shrinker reclaim should be as non-blocking as possible, and any back-off to wait for IO to complete should be done by the high level reclaim core once it's completed an entire reclaim scan cycle of everything.... What follows from that, and is pertinent for in this situation, is that if you don't block kswapd, then other reclaim contexts are not going to get stuck waiting for it regardless of the reclaim context they use. Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
