Re: [PATCH 2/5] blktrace: fix debugfs use after free

From: Luis Chamberlain <mcgrof@kernel.org>
To: Eric Sandeen <sandeen@sandeen.net>
Cc: axboe@kernel.dk, viro@zeniv.linux.org.uk, bvanassche@acm.org,
	gregkh@linuxfoundation.org, rostedt@goodmis.org,
	mingo@redhat.com, jack@suse.cz, ming.lei@redhat.com,
	nstange@suse.de, akpm@linux-foundation.org, mhocko@suse.com,
	yukuai3@huawei.com, linux-block@vger.kernel.org,
	linux-fsdevel@vger.kernel.org, linux-mm@kvack.org,
	linux-kernel@vger.kernel.org, Omar Sandoval <osandov@fb.com>,
	Hannes Reinecke <hare@suse.com>, Michal Hocko <mhocko@kernel.org>,
	syzbot+603294af2d01acfdd6da@syzkaller.appspotmail.com
Subject: Re: [PATCH 2/5] blktrace: fix debugfs use after free
Date: Thu, 16 Apr 2020 01:20:42 +0000	[thread overview]
Message-ID: <20200416012042.GD11244@42.do-not-panic.com> (raw)
In-Reply-To: <924950e6-e016-25b2-4ee1-b5ea9f752c12@sandeen.net>

On Wed, Apr 15, 2020 at 08:02:04PM -0500, Eric Sandeen wrote:
> 
> 
> On 4/15/20 7:56 PM, Luis Chamberlain wrote:
> > On Wed, Apr 15, 2020 at 12:38:26PM -0500, Eric Sandeen wrote:
> >> On 4/13/20 11:18 PM, Luis Chamberlain wrote:
> >>> On commit 6ac93117ab00 ("blktrace: use existing disk debugfs directory")
> >>> merged on v4.12 Omar fixed the original blktrace code for request-based
> >>> drivers (multiqueue). This however left in place a possible crash, if you
> >>> happen to abuse blktrace in a way it was not intended.
> >>>
> >>> Namely, if you loop adding a device, setup the blktrace with BLKTRACESETUP,
> >>> forget to BLKTRACETEARDOWN, and then just remove the device you end up
> >>> with a panic:
> >>
> >> I think this patch makes this all cleaner anyway, but - without the apparent
> >> loop bug mentioned by Bart which allows removal of the loop device while blktrace
> >> is active (if I read that right), can this still happen?
> > 
> > I have not tested that, but some modifications of the break-blktrace
> > program could enable us to test that
> 
> FWIW, I modified it to modprobe & rmmod scsi_debug instead of the loop ioctls,
> and the module can't be unloaded after the blktrace is started since it's busy.
> 
> Not sure that's equivalent tho.

Given what Bart mentioned about sd_open, that might be the saving grace
for blocking async behaviour on scsi. If it only refcounted the
request_queue as the loop driver it would have been exposed to the
same bug.

> > however I don't think the race
> > would be possible after patch 3/5 "blktrace: refcount the request_queue
> > during ioctl" is merged, as removal then a pending blktrace would
> > refcount the request_queue and the removal would have to wait until
> > the refcount is decremeneted, until after the blktrace ioctl.
> 
> I'm out of my depth in the block layer, not sure who's supposed to take
> refs on what. ;)

I'm new to all this code, only been a few weeks looking into it, but am
trying to do my best ot make sense of it. So the above is what I can
tell would be needed.

  Luis