From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jan Kara Subject: Re: [PATCH] ext4/jbd2: don't wait (forever) for stale tid caused by wraparound Date: Thu, 21 Mar 2013 23:41:43 +0100 Message-ID: <20130321224143.GA5066@quack.suse.cz> References: <1363412062.3937.196.camel@deadeye.wl.decadent.org.uk> <20130318025401.GA12611@thunk.org> <20130321204638.GA6116@quack.suse.cz> <20130321210940.GD21877@thunk.org> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Jan Kara , Ben Hutchings , George Barnett , linux-ext4@vger.kernel.org, Debian kernel maintainers To: Theodore Ts'o Return-path: Received: from cantor2.suse.de ([195.135.220.15]:54631 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753680Ab3CUWlq (ORCPT ); Thu, 21 Mar 2013 18:41:46 -0400 Content-Disposition: inline In-Reply-To: <20130321210940.GD21877@thunk.org> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Thu 21-03-13 17:09:40, Ted Tso wrote: > On Thu, Mar 21, 2013 at 09:46:38PM +0100, Jan Kara wrote: > > Good catch! But shouldn't we rather fix jbd2_log_wait_commit() instead of > > inventing new function? > > In most of the places where we call jbd2_log_start_commit(), we're > actually starting the current running transaction. So the fact that > we pass in a tid, and we're having to validate that the tid is > actually a valid one, is a bit of a waste. So in the long run I think > it's worth rethinking whether or not jbd2_log_{start,wait}_commit() > should exist in their current form, or whether we should reorganize > their functionality (i.e., by having a jbd2_start_running_commit(), > for example.). Piling on fixes to jbd2_log_wait_commit() would make > it get even more complicated, and I think if we separate out the > various ways in which we use these functions, we can make the code > simpler and easier to read. I don't find jbd2_log_wait_commit() that complex that it couldn't bear another if :) But given there are really two waiting operations that make sense: a) request commit of running transaction and wait for it b) wait for committing transaction then I agree there may be a better interface. OTOH I'm somewhat curious about the new interface because the only race-free way of identifying a transaction you want to wait for is using its tid. > In fact, I had started making this rather large set of changes when I > decided it would be better to save that kind of wholesale refactoring > for the next merge window. So the reason why I ended up fixing the > patch the way I did was to keep things simple. > > Also as I mentioned in the commit description, by using a single > function I was also able to optimize the locking the locking somewhat. Yeah. I'm not as much opposed to the new function doing start commit & wait but what I dislike is the fact that we have still exposed the function jbd2_log_wait_commit() which can possibly lockup if tid overflows. I agree there aren't currently any other callers where this could happen but in a few years who knows... Honza -- Jan Kara SUSE Labs, CR