From: Jan Kara Subject: Re: Ordered mode rewrite patch Date: Mon, 14 Apr 2008 19:47:22 +0200 Message-ID: <20080414174722.GG15824@duck.suse.cz> References: <20080408121149.GB6901@duck.suse.cz> <20080411184258.GA2757@unused.rdu.redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Nathan Grennan , Marcelo Tosatti , linux-ext4@vger.kernel.org To: Josef Bacik Return-path: Received: from styx.suse.cz ([82.119.242.94]:33959 "EHLO mail.suse.cz" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1762866AbYDNRrY (ORCPT ); Mon, 14 Apr 2008 13:47:24 -0400 Content-Disposition: inline In-Reply-To: <20080411184258.GA2757@unused.rdu.redhat.com> Sender: linux-ext4-owner@vger.kernel.org List-ID: Hi Josef, On Fri 11-04-08 14:42:58, Josef Bacik wrote: > On Tue, Apr 08, 2008 at 02:11:49PM +0200, Jan Kara wrote: > > Hello, > > > > attached is a jumbo patch that reverses locking order of transaction > > start and page lock in ext3 and rewrites handling of ordered data mode in > > JBD and ext3. Note that the patch will break compilation of ext4 and OCFS2. > > The patch survives LTP run on my test machine so it shouldn't eat your data > > immediately but bugs are of course possible... > > I'm very interested in any results (both positive and negative) you could > > get with it :). Thanks for testing it. > > > > Honza > I just hit a problem with your patch. In journal_destroy() we do a > iput(journal->j_inode) and then kfree the journal, so when the iput comes back > into journal_release_jbd_inode we are doing a use after free which in my case > resulted in a panic. I was going to fix it but I figure since this is still in > transit you would have to just rewrite it so I'm not going to attach one, just > giving you a heads up. Let me know if my explanation wasn't clear enough. Thanks for letting me know. But I don't quite get it: journal_destroy() does: if (journal->j_inode) iput(journal->j_inode); if (journal->j_revoke) journal_destroy_revoke(journal); kfree(journal->j_wbuf); kfree(journal); } iput() will eventually call generic_forget_inode() which calls clear_inode() in which we do journal_release_jbd_inode(). But at that moment journal still exists to I don't see any use after free... But possibly there could be some other inode in the inode list but which one is it? Honza -- Jan Kara SUSE Labs, CR