On Tue, Sep 24, 2024 at 06:45:55PM GMT, Linus Torvalds wrote: > On Tue, 24 Sept 2024 at 17:17, Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > > > FWIW, I think all this "how do we cache inodes better" discussion is > > somehwat glossing over a more important question we need to think > > about first: do we even need a fully fledged inode cache anymore? > > I'd be more than happy to try. I have to admit that it's largely my > mental picture (ie "inode caches don't really matter"), and why I was > surprised that the inode locks even show up on benchmarks. > > If we just always drop inodes when the refcount goes to zero and never > have any cached inodes outside of other references, I think most of > the reasons for the superblock inode list just disappear. Most of the > heavy lifting has long since been moved to the dentry shrinking. > > I'd worry that we have a lot of tuning that depends on the whole inode > LRU thing (that iirc takes the mapping size etc into account). > > But maybe it would be worth trying to make I_DONTCACHE the default, > with the aim of removing the independent inode lifetimes entirely... I would be rather surprised if we didn't see regressions for some workloads but who knows. If we wanted to test this then one easy way would be to add a FS_INODE_NOCACHE fs_flag and have inode_init_always() raise I_DONTCACHE based on that. That way we can experiment with this on a per-fs basis.
