On Fri, May 28, 2021 at 02:33:42PM +0800, Ian Kent wrote: > There have been a few instances of contention on the kernfs_mutex during > path walks, a case on very large IBM systems seen by myself, a report by > Brice Goglin and followed up by Fox Chen, and I've since seen a couple > of other reports by CoreOS users. > > The common thread is a large number of kernfs path walks leading to > slowness of path walks due to kernfs_mutex contention. > > The problem being that changes to the VFS over some time have increased > it's concurrency capabilities to an extent that kernfs's use of a mutex > is no longer appropriate. There's also an issue of walks for non-existent > paths causing contention if there are quite a few of them which is a less > common problem. > > This patch series is relatively straight forward. > > All it does is add the ability to take advantage of VFS negative dentry > caching to avoid needless dentry alloc/free cycles for lookups of paths > that don't exit and change the kernfs_mutex to a read/write semaphore. > > The patch that tried to stay in VFS rcu-walk mode during path walks has > been dropped for two reasons. First, it doesn't actually give very much > improvement and, second, if there's a place where mistakes could go > unnoticed it would be in that path. This makes the patch series simpler > to review and reduces the likelihood of problems going unnoticed and > popping up later. > > The patch to use a revision to identify if a directory has changed has > also been dropped. If the directory has changed the dentry revision > needs to be updated to avoid subsequent rb tree searches and after > changing to use a read/write semaphore the update also requires a lock. > But the d_lock is the only lock available at this point which might > itself be contended. Fox, can you take some time and test these to verify it all still works properly with your benchmarks? thanks, greg k-h
