From mboxrd@z Thu Jan 1 00:00:00 1970 From: Nick Piggin Subject: Patch "fs: use RCU read side protection in d_validate" broken Date: Mon, 15 Nov 2010 16:11:20 +1100 Message-ID: <20101115051120.GA4092@amd> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Christoph Hellwig , Al Viro , linux-fsdevel@vger.kernel.org To: Linus Torvalds Return-path: Received: from ipmail05.adl6.internode.on.net ([150.101.137.143]:58641 "EHLO ipmail05.adl6.internode.on.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750709Ab0KOFLZ (ORCPT ); Mon, 15 Nov 2010 00:11:25 -0500 Content-Disposition: inline Sender: linux-fsdevel-owner@vger.kernel.org List-ID: This patch is totally broken. You can't just dget() a dentry with nothing but RCU critical section open. The patch in my tree this is claimed to be split out of, at least attemptet to do some locking, I don't know why that was stripped out. But I didn't get that quite right myself at which point I decided to just forget about it entirely. Christoph, why did you think such a patch is worth getting merged, btw? I saw no hint of a justification in your changelog. I mean, in my tree at least there was a _rationale_ that dcache_lock is going away and this marginally made the locking simpler. But it doesn't make sense in the current tree, even if the merged patch was _not_ buggy -- what were you trying to do, make ncpfs's readdir go really fast? Breaking things out when they make sense and stand as patches on their own. That is very important. When they do _not_ make sense on their own, then that is when they do not make sense to be broken out and merged ahead of their series. Linus, I cooked up a fix for it, but decided it's stupid to bother with any complexity here. Please revert 3825bdb7ed920845961f32f364454bee5f469abb If I end up wanting it again for dcache series, I can do it with proper locking and proper justification. Thanks, Nick