From mboxrd@z Thu Jan 1 00:00:00 1970 From: Dave Chinner Subject: Re: Are there u32 atomic bitops? (or dealing w/ i_flags) Date: Wed, 19 Dec 2012 08:30:43 +1100 Message-ID: <20121218213043.GF15182@dastard> References: <20121218015705.GH4939@ZenIV.linux.org.uk> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Al Viro , linux-kernel@vger.kernel.org, Linux FS Devel To: Andy Lutomirski Return-path: Received: from ipmail06.adl6.internode.on.net ([150.101.137.145]:10739 "EHLO ipmail06.adl6.internode.on.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754203Ab2LRVat (ORCPT ); Tue, 18 Dec 2012 16:30:49 -0500 Content-Disposition: inline In-Reply-To: Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Mon, Dec 17, 2012 at 06:42:44PM -0800, Andy Lutomirski wrote: > On Mon, Dec 17, 2012 at 5:57 PM, Al Viro wrote: > > On Mon, Dec 17, 2012 at 05:10:21PM -0800, Andy Lutomirski wrote: > >> I want to change inode->i_flags access to be atomic -- there are some > >> locking oddities right now, I think, and I want to use a new inode > >> flag to signal mtime updates from page_mkwrite. The problem is that > >> i_flags is an unsigned int, and making it an unsigned long seems like > >> a waste, but there aren't any u32 atomic bitops. > > > > ... and atomic accesses cost more. A lot more on some architectures. > > FWIW, atomic_t *is* 32bit on 32bit architectures, which still doesn't > > make it a good idea. > > Are atomic_set_mask and atomic_clear_mask as fast as set_bit and > friends on all archs? > > In any case, i_flags looks like it's rarely written, so I find it a > bit hard to believe that making it atomic would hurt. Isn't > atomic_read equivalent to non-atomic reads everywhere? > > I want page_mkwrite to set a flag (without taking i_mutex) but *not* > call file_update_time and then to have the writeback paths update the > inode time. Deadlocks ahoy! We don't currently take the i_mutex anywhere in the writeback path as the writeback path is nested inside the i_mutex. Hence this seems like an extremely dangerous thing to do... > (This, along with stable pages, is the major cause of > long sleeps in my application.) OTOH, maybe I should just use i_state > and i_lock for this. Or, perhaps, export O_CMTIME to fcntl and/or open? Cheers, Dave. -- Dave Chinner david@fromorbit.com