On Mon, Dec 19, 2022 at 04:05:08PM -0800, Darrick J. Wong wrote: > From: Darrick J. Wong <djwong@xxxxxxxxxx> > > The online fsck stress tests deadlocked a test VM the other night. The > deadlock happened because: > > 1. kswapd tried to prune the sb inode list, xfs found that it needed to > inactivate an inode and that the queue was long enough that it should > wait for the worker. It was holding shrinker_rwsem. > > 2. The inactivation worker allocated a transaction and then stalled > trying to obtain the AGI buffer lock. > > 3. An online repair function called unregister_shrinker while in > transaction context and holding the AGI lock. It also tried to grab > shrinker_rwsem. > > #3 shouldn't happen and was easily fixed, but seeing as we designed > background inodegc to avoid stalling reclaim, I feel that #1 shouldn't > be happening either. Fix xfs_inodegc_want_flush_work to avoid stalling > background reclaim on inode inactivation. Yup, I see what you are saying, but I specifically left GFP_KERNEL reclaim to throttle on inodegc if necessary. kswapd runs in GFP_KERNEL context - it's the only memory reclaim context where it is actually safe to block in this manner in the filesystem. The question becomes one of whether we allow kswapd to ignore inodegc queue limits. Before the background inodegc, we used to do all the xfs_inactive() work in ->destroy_inode, directly from the VFS inode cache shrinker context run by GFP_KERNEL memory reclaim contexts such as kswapd(). IOWs, causing kswapd to wait while we run inactivation in the VFS inode cache shrinker context is something we've done for a long time. kswapd did: vfs inode shrinker destroy inode xfs_inactive() mark inode inactivated update radix tree tags And then the XFS inode shrinker walks INACTIVATED inodes and frees them. Now the code does: vfs inode shrinker NEED_INACTIVE -> inodegc queue if (throttle needed) flush_work(gc->work); return Then some time later in a different context inodegc then runs and we do: xfs_inactive INACTIVATED And then sometime later in a different context the XFS inode shrinker walks INACTIVATED inodes and frees them. Essentially, without throttling the bulk of the memory freeing work is shifted from the immediate GFP_KERNEL reclaim context to some later GFP_KERNEL reclaim context. For small memory reclaim scans (i.e. low priority shrinker invocations) this doesn't matter - we aren't at risk of OOM in these situations, and these are the typical situations in which memory reclaim is needed. Generally speaking the the background work is fast enough that memory is freed before shortages escalate. However, when we have large inode caches and a significant memory demand event, we get into the situation where reclaim might be scanning a hundreds of thousands of cached VFS inodes at a time. We kinda need to stall kswapd in these cases - it may be the only reclaim context that can free inodes - and we need to avoid kswapd dumping all the freeable inodes on the node to a queue that isn't being processed, decided it has no more memory that can be freed because all the VFS inodes have gone and there's nothing reclaimable on the XFS inode lists, so it declares OOM when we have hundreds of thousands of freeable inodes sitting there waiting for inodegc to run. I note that inodes queued for inodegc are not counted as reclaimable XFS inodes, so for the purposes of shrinkers those inodes queued on inodegc disappear from the shrinker accounting compeltely. Maybe we could consider a separate bug hidden by the inodegc throttling - maybe that's the bug that makes throttling kswapd on inodegc necessary. But this re-inforces my point - I don't think we can just make kswapd skip inodegc throttling without further investion into the effect on performance under memory pressure as well as OOM killer resistance. It's likely a solvable issue, but given how complex and subtle the inode shrinker interactions can be, I'd like to make sure we don't wake a sleeping dragon here.... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx