Re: [PATCH] block: replace __getblk_slow misfix by grow_dev_page fix

From: Jens Axboe <jaxboe@fusionio.com>
To: Hugh Dickins <hughd@google.com>
Cc: "Richard W.M. Jones" <rjones@redhat.com>,
	Jeff Moyer <jmoyer@redhat.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Torsten Hilbrich <torsten.hilbrich@secunet.com>,
	Josh Boyer <jwboyer@redhat.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] block: replace __getblk_slow misfix by grow_dev_page fix
Date: Thu, 23 Aug 2012 12:20:14 +0200	[thread overview]
Message-ID: <503603DE.5060905@fusionio.com> (raw)
In-Reply-To: <alpine.LSU.2.00.1208222143440.1917@eggly.anvils>

On 08/23/2012 06:56 AM, Hugh Dickins wrote:
> [PATCH] block: replace __getblk_slow misfix by grow_dev_page fix
> 
> Commit 91f68c89d8f3 ("block: fix infinite loop in __getblk_slow")
> is not good: a successful call to grow_buffers() cannot guarantee
> that the page won't be reclaimed before the immediate next call to
> __find_get_block(), which is why there was always a loop there.
> 
> Yesterday I got "EXT4-fs error (device loop0): __ext4_get_inode_loc:3595:
> inode #19278: block 664: comm cc1: unable to read itable block" on console,
> which pointed to this commit.
> 
> I've been trying to bisect for weeks, why kbuild-on-ext4-on-loop-on-tmpfs
> sometimes fails from a missing header file, under memory pressure on
> ppc G5.  I've never seen this on x86, and I've never seen it on 3.5-rc7
> itself, despite that commit being in there: bisection pointed to an
> irrelevant pinctrl merge, but hard to tell when failure takes between
> 18 minutes and 38 hours (but so far it's happened quicker on 3.6-rc2).
> 
> (I've since found such __ext4_get_inode_loc errors in /var/log/messages
> from previous weeks: why the message never appeared on console until
> yesterday morning is a mystery for another day.)
> 
> Revert 91f68c89d8f3, restoring __getblk_slow() to how it was (plus
> a checkpatch nitfix).  Simplify the interface between grow_buffers()
> and grow_dev_page(), and avoid the infinite loop beyond end of device
> by instead checking init_page_buffers()'s end_block there (I presume
> that's more efficient than a repeated call to blkdev_max_block()),
> returning -ENXIO to __getblk_slow() in that case.
> 
> And remove akpm's ten-year-old "__getblk() cannot fail ... weird"
> comment, but that is worrying: are all users of __getblk() really
> now prepared for a NULL bh beyond end of device, or will some oops??

Hugh, I tentatively applied this one, awaiting some test feedback before
pushing it upstream this cycle.

-- 
Jens Axboe