From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from ipmail06.adl2.internode.on.net ([150.101.137.129]:23727 "EHLO ipmail06.adl2.internode.on.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1729585AbeLYXrh (ORCPT ); Tue, 25 Dec 2018 18:47:37 -0500 Date: Wed, 26 Dec 2018 10:47:32 +1100 From: Dave Chinner Subject: Re: Non-blocking socket stuck for multiple seconds on xfs_reclaim_inodes_ag() Message-ID: <20181225234732.GH4205@dastard> References: <20181129021800.GQ6311@dastard> <20181130021840.GV6311@dastard> <20181130064908.GX6311@dastard> <20181130074547.GY6311@dastard> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: Sender: linux-xfs-owner@vger.kernel.org List-ID: List-Id: xfs To: Kenton Varda Cc: Ivan Babrou , linux-xfs@vger.kernel.org, Shawn Bohrer Hi Kenton, On Thu, Dec 20, 2018 at 08:00:21PM -0800, Kenton Varda wrote: > When XFS blocks direct reclaim, our service pretty much grinds to a > halt on that machine, because everything is trying to allocate memory > all the time. For example, as alluded by the subject of this thread, > writing to a socket allocates memory, and thus will block waiting for > XFS to write back inodes. What we find really frustrating is that we > almost always have over 100GB of clean page cache that could be > reclaimed immediately, without blocking, yet we end up waiting for the > much-smaller inode cache to be written back to disk. Sure, it's frustrating. It frsutrates me that I've been trying for years to get memory reclaim behaviour changed so that we don't have to do this, but we are still stuck with it. But taking out your frustrations on the people who are trying to fix the problems you are seeing isn't productive. We are only a small team and we can't fix every problem that everyone reports immediately. Some things take time to fix. > We really can't accept random multi-second pauses. And that's the problem here. For the vast majority of XFS users, the alternative (i.e. not blocking reclaim) leads to substantially lower performance and the high risk of premature OOM kills. It basically moves the reclaim blocking problem to a different context (i.e. to journal space allocation) and that has even worse latency and has global filesystem scope rather than just the process doing reclaim. IOWs, there are relatively few applications that have such a significant dependency on memory reclaim having extremely low latency, but there are a lot that are dependent on memory reclaim being throttled harshly to keep the filesystem out of even worse breakdown conditions..... > Our current plan is > to roll out the patch Ivan linked to. Which is perfectly fine by me. I read the link, and it looks like it works just fine in your environment. In contrast, I ran the same patch on my performance bencmarks and saw between 15-30% degradation in performance on my inode cache heavy and mixed inode/page cache heavy memory pressure benchmarks. IOWs, that change still doesn't work for XFS in general. This is the beauty of open source software - you can make tweaks to it to refine it for your specific workloads easily when such changes aren't really suitable for the wider user base. I encourage people to make tweaks like this to optimise your systems for your workload. However, I also encourage people then to discuss the problems that lead to needing such tweaks with upstream so we are aware of the issues and can work towards either incorporating them or modifying infrastructure to avoid them altogether. Further, there's no need to be combative or get upset when upstream determines that the tweak isn't generally applicable or is hiding something deeper that needs fixing. All it means is that the upstream developers know there is a deeper underlying problem and want to fix the underlying problem rather than try to hide the problem for a specific workload. The fact we know there's a problem (and it's not just one workload situation that it affects) helps us prioritise what we need to fix. > But, if you have any other > suggestions, we'd love to hear them. It would be great if we could > agree on an upstream solution, and maybe solve Facebook's problem too. I've already mentioned the things we are working on to Shaun to keep the ball rolling on this. e.g. Darrick's pipelined background inode inactivation work: https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/commit/?h=djwong-wtf&id=85baea68d8e87803e6831bda7b5a3773cf0d8820 i.e. as I indicated above, inode reclaim can block on lots more things that just inode writeback. We can do transactions in inode reclaim, so we can block on log space and that might require *thousands* of IO to be completed before it can make progress. The async kswapd mod doesn't address this problem at all because inactivation occurs in the prune_icache() VFS shrinker context (i.e. ->destroy_inode()) rather than the XFS inode cache context that the FB and CF patches address. Also, the upcoming inode cache xarray conversion will provide us with many more per-inode state tags in the xarray that will allow us to track and execute multiple different post-VFS reclaim inode states directly. This will allow us to efficiently separate inodes that need transactions and/or IO to be reclaimed from inodes that can be reclaimed immediately, and it will allow efficient concurrent async processing of inodes that need IO to be reclaimed. IOWs, we're trying to solve *all* the blocking problems that we know that can occur in inode reclaim so that it all just works for everyone without tweaks being necessary. Yes, this takes longer than just addressing the specific symptom that is causing you problems, but the reality is while fixing things properly takes time to get right, everyone will benefit from it being fixed and not just one or two very specific, latency sensitive workloads. Cheers, Dave. -- Dave Chinner david@fromorbit.com