From mboxrd@z Thu Jan 1 00:00:00 1970 From: Nick Piggin Subject: [announce] vfs-scale git tree update Date: Wed, 5 Jan 2011 21:25:34 +1100 Message-ID: <20110105102534.GA6679@amd> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii To: Linus Torvalds , Al Viro , linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org Return-path: Content-Disposition: inline Sender: linux-kernel-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org Git tree is here: git://git.kernel.org/pub/scm/linux/kernel/git/npiggin/linux-npiggin.git Branch is: vfs-scale-working Changes since last posting: * Updated to 2.6.37 (Documentation/filesystems/Locking clash) * Switch names of vfsmount scalable counter helpers suggested by Andreas * Most significant are changes in Documentation/filesystems/path-lookup.txt attempt to make it more readable, consistent and informative. Add some interesting rcu-walk path lookup success and behaviour statistics. Status: * Linus is planning to merge. It's never too late for review, though. * linux-next has been uneventful, but I don't think it nearly covers all interesting and fiddly use cases. * Still has the barrier-less __seqcount optimisation that Linus didn't like; I like the idea of a seqcount-switch API, but it just didn't seem to fit well here. Let's leave that on the todo list? Future dcache / name lookup work: * Per-zone LRUs. Patch is simple and ready, but performance bisecting might be a bit easier if we hold off. Also inode LRUs should be done at the same time. * Filesystems will need to start implementing rcu-walk aware dentry and permission ops. They've got simple examples to follow. * Rename scaling. The rename seqlock can explode on large systems, getting into strange conditions where lookup performance crashes. It is also a global lock for renames. Quite simple to break it up and fix lookup performance and provide linear vfs scalability for parallel same-directory renames (if they are in different directories). Doesn't need to be merged yet, though. * Further optimise name string copying and comparison (may be as much as 10-20% in that). * rcu-walk for symlinks. A bit tricky, not impossible.