Re: [PATCH] Revert "floppy: reintroduce O_NDELAY fix"

From: Denis Efremov <efremov@linux.com>
To: Jiri Kosina <jikos@kernel.org>
Cc: linux-block@vger.kernel.org, linux-kernel@vger.kernel.org,
	regressions@lists.linux.dev, Mark Hounschell <markh@compro.net>,
	Wim Osterholt <wim@djo.tudelft.nl>,
	Kurt Garloff <kurt@garloff.de>,
	stable@vger.kernel.org
Subject: Re: [PATCH] Revert "floppy: reintroduce O_NDELAY fix"
Date: Wed, 18 Aug 2021 18:53:58 +0300	[thread overview]
Message-ID: <5ca95b90-c6e0-99b3-b129-75dc05cfb1d4@linux.com> (raw)
In-Reply-To: <nycvar.YFH.7.76.2108160914070.8253@cbobk.fhfr.pm>

Hi,

On 8/16/21 10:17 AM, Jiri Kosina wrote:
> On Sun, 8 Aug 2021, Denis Efremov wrote:
> 
>> The patch breaks userspace implementations (e.g. fdutils) and introduces
>> regressions in behaviour. Previously, it was possible to O_NDELAY open a
>> floppy device with no media inserted or with write protected media without
>> an error. Some userspace tools use this particular behavior for probing.
>>
>> It's not the first time when we revert this patch. Previous revert is in
>> commit f2791e7eadf4 (Revert "floppy: refactor open() flags handling").
>>
>> This reverts commit 8a0c014cd20516ade9654fc13b51345ec58e7be8.
> 
> By reverting it you bring back the bugs that were fixed by it

I agree with you, that O_NDELAY is broken for floppies (and always been).
However, just by removing O_NDELAY we break many existing tools that use
it for probing and ioctl-only opens. With the patch tools fail to open the
device without a diskette and try to read a diskette if there is one (this is
not as fast on a real hardware as in QEMU).
I think that there should be a better fix that doesn't break existing tools.
It appears that people still use software that depends on O_NDELAY in floppies.
Same patch was already reverted in 2016 (presumably) by the same reason.

>  -- e.g. the 
> possibility to livelock mmap() on the returned fd to keep waiting on the 
> page unlock bit forever

As far as I understand this is a problem only for syzkaller.
And this is not a security issue nowadays since most distributions
(I don't know exceptions) require at least "disk" group to access floppies.
Do you know a link for the syzkaller reproducer?

> or the functionality bug reported at [1], and 
> likely others.
> 
> [1] https://bugzilla.suse.com/show_bug.cgi?id=1181018

The patch starts to return -ENXIO for O_NDELAY|O_RDONLY opens and devices
without a diskette. I don't think this is an expected behavior during 
libblkid probing.

Probably there is a better fix for [1], maybe even an additional workaround
for floppies in libblkid. They already have workarounds for cdroms
https://github.com/karelzak/util-linux/commit/dc30fd4383e57a0440cdb0e16ba5c4336a43b290

I started to add simple tests https://lkml.org/lkml/2021/8/18/845
However, I failed to reproduce mount bug [1], probably
because I don't know how to configure cloudinit properly. I tried to reproduce
a mount fail bug with open("/dev/fd0", O_NDELAY|O_RDONLY) and mount("/dev/fd0", ...)
but it works. Looks like there should be something else in between...

Regards,
Denis