From mboxrd@z Thu Jan 1 00:00:00 1970 From: =?ISO-8859-1?Q?Fr=E9d=E9ric_Weisbecker?= Subject: Re: [PATCH 1/1] vfs: umount_begin BKL pushdown v2 Date: Fri, 24 Apr 2009 11:16:18 +0200 Message-ID: References: <20090423191934.GW1926@parisc-linux.org> <1240556813-8739-1-git-send-email-abogani@texware.it> <1240556813-8739-2-git-send-email-abogani@texware.it> <20090424071312.GE8633@ZenIV.linux.org.uk> <20090424071853.GG8633@ZenIV.linux.org.uk> <20090424080634.GG24912@elte.hu> <20090424085017.GB28592@infradead.org> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Cc: Ingo Molnar , Al Viro , Alessio Igor Bogani , Jonathan Corbet , Peter Zijlstra , LKML , LFSDEV , Linus Torvalds , Matthew Wilcox To: Christoph Hellwig Return-path: In-Reply-To: <20090424085017.GB28592@infradead.org> Sender: linux-kernel-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org 2009/4/24 Christoph Hellwig : >> Furthermore, by doing this you are also hindering the >> tip:kill-the-BKL effort (which has been ongoing for a year chipping >> away at various BKL details) which facilitated these changes. >> Alessio did these fixes to fix bugs he can trigger in that tree. >> >> You've also not explained why you have done it this way. It would >> cost you almost nothing to apply these bits into a separate branch >> and merge that branch into your main tree. Lots of other maintainer >> are doing that. > > Having a separate kill the BKL tree is a stupid idea. Locking changes > need deep subsystem knowledge and should always go through the subsystem > trees. I disagree with you. The kill-the-BKL tree does not only aggregate patches to turn the BKL into more traditional locks. The Bkl has been converted to a common mutex in this tree, making it losing its common horrid properties: - release/reacquire on schedule - not preemptable - can be reacquired recursively by a same task Such a basis is very useful because we can easily find these places which won't support a usual lock conversion without reworking the locking scheme. This is a necessary preliminary for the Bkl removal. All the places which have been designed very tightly with Bkl properties are rapidly detected with lockdep in this tree and reworked, still using lockdep, code reviewing and the help of this Bkl-to-mutex conversion. The work done with this tree can be merged inside and also on the matching subsytem tree for each patchset. That's a very sane workflow IMHO. Thanks, Frederic.