From: Jens Axboe Subject: Re: [PATCH] JBD: Fix DIO EIO error caused by race between free buffer and commit trasanction Date: Tue, 20 May 2008 11:30:21 +0200 Message-ID: <20080520093020.GL22369@kernel.dk> References: <20080514170856.GH24363@duck.suse.cz> <1210786872.3657.48.camel@localhost.localdomain> <20080514181444.GI24363@duck.suse.cz> <1210947250.3608.18.camel@localhost.localdomain> <1210957976.4231.31.camel@badari-desktop> <1210971693.3608.46.camel@localhost.localdomain> <20080518223739.GB11006@atrey.karlin.mff.cuni.cz> <1211227158.3663.25.camel@localhost.localdomain> <20080519132553.de9b78b0.akpm@linux-foundation.org> <1211234829.3663.39.camel@localhost.localdomain> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Andrew Morton , jack@suse.cz, pbadari@us.ibm.com, linux-ext4@vger.kernel.org, linux-kernel@vger.kernel.org To: Mingming Cao Return-path: Received: from brick.kernel.dk ([87.55.233.238]:5054 "EHLO kernel.dk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752746AbYETJaZ (ORCPT ); Tue, 20 May 2008 05:30:25 -0400 Content-Disposition: inline In-Reply-To: <1211234829.3663.39.camel@localhost.localdomain> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Mon, May 19 2008, Mingming Cao wrote: > On Mon, 2008-05-19 at 13:25 -0700, Andrew Morton wrote: > > On Mon, 19 May 2008 12:59:18 -0700 > > Mingming Cao wrote: > > > > > On Mon, 2008-05-19 at 00:37 +0200, Jan Kara wrote: > > > > Hi, > > > > > > > > > This patch fixed a few races between direct IO and kjournald commit > > > > > transaction. An unexpected EIO error gets returned to direct IO > > > > > caller when it failed to free those data buffers. This could be > > > > > reproduced easily with parallel direct write and buffered write to the > > > > > same file > > > > > > > > > > More specific, those races could cause journal_try_to_free_buffers() > > > > > fail to free the data buffers, when jbd is committing the transaction > > > > > that has those data buffers on its t_syncdata_list or t_locked_list. > > > > > journal_commit_transaction() still holds the reference to those > > > > > buffers before data reach to disk and buffers are removed from the > > > > > t_syncdata_list of t_locked_list. This prevent the concurrent > > > > > journal_try_to_free_buffers() to free those buffers at the same time, > > > > > but cause EIO error returns back to direct IO. > > > > > > > > > > With this patch, in case of direct IO and when try_to_free_buffers() failed, > > > > > let's waiting for journal_commit_transaction() to finish > > > > > flushing the current committing transaction's data buffers to disk, > > > > > then try to free those buffers again. > > > > If Andrew or Christoph wouldn't beat you for "inventive use" of > > > > gfp_mask, I'm fine with the patch as well ;). You can add > > > > Acked-by: Jan Kara > > > > > > > > > > This is less intrusive way to fix this problem. The gfp_mask was marked > > > as unused in try_to_free_page(). I looked at filesystems in the kernel, > > > there is only a few defined releasepage() callback, and only xfs checks > > > the flag(but not used). btrfs is actually using it though. I thought > > > about the way you have suggested, i.e.clean up this gfp_mask and and > > > replace with a flag. I am not entirely sure if it we need to change the > > > address_space_operations and fix all the filesystems for this matter. > > > > > > Andrew, what do you think? Is this approach acceptable? > > > > > > > > > > > Please ensure that the final patch is sufficiently well changelogged to > > permit me to remain asleep ;) > :-) > > The ->releasepage semantics are fairly ad-hoc and have grown over time. > > It'd be nice to prevent them from becoming vaguer than they are. > > > > It has been (approximately?) the case that code paths which really care > > about having the page released will set __GFP_WAIT (via GFP_KERNEL) > > whereas code paths which are happy with best-effort will clear > > __GFP_WAIT (with a "0'). And that's reasonsable - __GFP_WAIT here > > means "be synchronous" whereas !__GFP_WAIT means "be non-blocking". > > > > This make sense to me. > > > Is that old convention not sufficient here as well? Two problem areas > > I see are mm/vmscan.c and fs/splice.c (there may be others). > > > > > In mm/vmscan.c we probably don't want your new synchronous behaviour > > and it might well be deadlockable anyway. No probs, that's what > > __GFP_FS is for. > > > Sure. We could check __GFP_FS and __GFP_WAIT, and that make sense. > > > In fs/splice.c, reading the comment there I have a feeling that you've > > found another bug, and that splice _does_ want your new synchronous > > behaviour? > > Yes, it looks like page_cache_pipe_buf_steal() expects page is free > before removeing it by passing the GFP_KERNEL flag, but currently ext3 > could fails to releasepage when it called. In fact > try_to_release_page() return value is ignored in > page_cache_pipe_buf_steal(), should probably checked the failure case. > > > The other caller of try_to_release_page() in mm/splice.c is > fallback_migrate_page(), which does want the synchronous behaviour to > make sure buffers are dropped. So something like this, then? diff --git a/fs/splice.c b/fs/splice.c index 7815003..e08a2f5 100644 --- a/fs/splice.c +++ b/fs/splice.c @@ -58,8 +58,8 @@ static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe, */ wait_on_page_writeback(page); - if (PagePrivate(page)) - try_to_release_page(page, GFP_KERNEL); + if (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL)) + goto out_unlock; /* * If we succeeded in removing the mapping, set LRU flag @@ -75,6 +75,7 @@ static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe, * Raced with truncate or failed to remove page from current * address space, unlock and return failure. */ +out_unlock: unlock_page(page); return 1; } -- Jens Axboe