From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1032033Ab2CORjq (ORCPT ); Thu, 15 Mar 2012 13:39:46 -0400 Received: from zeniv.linux.org.uk ([195.92.253.2]:49915 "EHLO ZenIV.linux.org.uk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1030464Ab2CORjn (ORCPT ); Thu, 15 Mar 2012 13:39:43 -0400 Date: Thu, 15 Mar 2012 17:39:37 +0000 From: Al Viro To: Linus Torvalds Cc: Thomas Gleixner , LKML , Ingo Molnar , Peter Zijlstra , Nick Piggin Subject: Re: [patch 1/5] seqlock: Remove unused functions Message-ID: <20120315173937.GC8943@ZenIV.linux.org.uk> References: <20120314170736.617746873@linutronix.de> <20120314170918.404528547@linutronix.de> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Thu, Mar 15, 2012 at 09:29:50AM -0700, Linus Torvalds wrote: > So I have to say, I hate this entire series. > > Seriously, what the heck is the point of this churn? It's all entirely > pointless searc-and-replace as far as I can tell, with absolutely zero > upside. > > It makes the low-level filesystems have to be aware of things that > they don't want to know and *shouldn't* know. Why should a filesystem > care that d_lock is a seqlock, and have to use a locking function that > they've never seen before and is very specialized? > > The "seq" part of the dentry is something only the lookup code and the > internal dentry code should care about. NOBODY ELSE should ever care. *nod* There's another issue I have with that on API level, leaving aside any questions of that being a good fit for dcache. It's simply a bad interface: we have variants that lock and play with d_seq, variants that play with d_seq alone and, most commonly used, variant that locks but does not touch d_seq at all. IOW, we have traded "writes to d_seq must be under d_lock" with "update-seq-without-locking primitive must be used after we'd used lock-without-touching-seq one". Which is not an improvement at all. Sure, you can make a direct product out of anything; that doesn't make the result a natural object. The _only_ relationship between d_seq and d_lock is that the latter happens to be serializing updates of the former. For RT there's another one - ->d_lock taken to protect ->d_seq modifications really should not be preempted in favour of anything that might do read_seqcount_begin on ->d_seq. The biggest such section is in __d_move(), AFAICS, and it's not _that_ big; can't RT simply have them protected by whatever it has that really prevents preempt? IOW, instead of all that stuff, how about about_to_modify_seq_holding_lock(&dentry->d_seq, &dentry->d_lock); done_modifying_seq(&dentry->d_seq, &dentry->d_lock); around those 3 or 4 areas in fs/dcache.c, to give RT the missing information?