From: Jan Kara Subject: Re: fsync on ext[34] working only by an accident Date: Thu, 10 Sep 2009 10:50:56 +0200 Message-ID: <20090910085056.GA607@duck.suse.cz> References: <20090908132601.GA17778@duck.suse.cz> <20090910064605.GA8690@skywalker.linux.vnet.ibm.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Jan Kara , linux-ext4@vger.kernel.org To: "Aneesh Kumar K.V" Return-path: Received: from cantor2.suse.de ([195.135.220.15]:46883 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752508AbZIJIuy (ORCPT ); Thu, 10 Sep 2009 04:50:54 -0400 Content-Disposition: inline In-Reply-To: <20090910064605.GA8690@skywalker.linux.vnet.ibm.com> Sender: linux-ext4-owner@vger.kernel.org List-ID: Hi, On Thu 10-09-09 12:16:05, Aneesh Kumar K.V wrote: > On Tue, Sep 08, 2009 at 03:26:01PM +0200, Jan Kara wrote: > > When looking at how ext3/4 handles fsync, I've realized I don't > > understand how writing out inode on fsync can work. The problem is that > > ext3/4 mostly calls ext?_mark_inode_dirty() which actually does *not* dirty > > the inode. It just copies the in-memory inode content to disk buffer. > > So in particular the inode looks clean to VFS and our check in > > ext?_sync_file() shouldn't trigger. > > The only obvious case when we call mark_inode_dirty() is from write_end > > functions when we update i_size but that's clearly not enough. Now I did > > some research why things seem to be actually working. The trick is that > > when allocating block, we call vfs_dq_alloc_block() which calls > > mark_inode_dirty(). But that's all what's keeping our fsync / writeout > > logic from breaking! > > ext4_handle_dirty_metadata should do mark_inode_dirty right ? > __ext4_handle_dirty_metadata -> mark_buffer_dirty ->__set_page_dirty > -> __mark_inode_dirty -> list_move(&inode->i_list, &sb->s_dirty); ext4_handle_dirty_metadata() marks the buffer dirty only when we do not have a journal (BTW, the inode that gets dirtied in the nojournal case is the block-device one, not the one whose metadata we mark as dirty, so it won't work there either - but Google guys are working on this I think). With a journal the function just calls jbd2_journal_dirty_metadata which does nothing with the inode. > > There are even some cases when the logic actually is broken (I've tested > > it and it really does not work) - for example when you create an empty > > file, the inode won't get written when you fsync it. > > So what we should IMHO do is to convert all ext?_mark_inode_dirty() > > calls to simple mark_inode_dirty() (or even maybe introduce and use > > mark_inode_dirty_datasync() where appropriate). It will cost us some more > > CPU and stack space but if we optimize ext3_dirty_inode() for the case > > where handle is already started, it shouldn't be too bad. Honza -- Jan Kara SUSE Labs, CR