Re: [PATCH RESEND V12 2/8] fuse: 32-bit user space ioctl compat for fuse device

From: Alessio Balsini <balsini@android.com>
To: Arnd Bergmann <arnd@kernel.org>
Cc: Alessio Balsini <balsini@android.com>,
	Miklos Szeredi <miklos@szeredi.hu>,
	Akilesh Kailash <akailash@google.com>,
	Amir Goldstein <amir73il@gmail.com>,
	Antonio SJ Musumeci <trapexit@spawn.link>,
	David Anderson <dvander@google.com>,
	Giuseppe Scrivano <gscrivan@redhat.com>,
	Jann Horn <jannh@google.com>, Jens Axboe <axboe@kernel.dk>,
	Martijn Coenen <maco@android.com>,
	Palmer Dabbelt <palmer@dabbelt.com>,
	Paul Lawrence <paullawrence@google.com>,
	Peng Tao <bergwolf@gmail.com>,
	Stefano Duo <duostefano93@gmail.com>,
	Zimuzo Ezeozue <zezeozue@google.com>, wuyan <wu-yan@tcl.com>,
	fuse-devel@lists.sourceforge.net,
	Android Kernel Team <kernel-team@android.com>,
	Linux FS-devel Mailing List <linux-fsdevel@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH RESEND V12 2/8] fuse: 32-bit user space ioctl compat for fuse device
Date: Fri, 19 Mar 2021 15:21:32 +0000	[thread overview]
Message-ID: <YFTBfKslrZsHYQPi@google.com> (raw)
In-Reply-To: <CAK8P3a36ToSbvW1F_0w0gCiWGCoZgFwoLHmQ7Tz2jtwV++VrWA@mail.gmail.com>

On Thu, Mar 18, 2021 at 10:15:43PM +0100, Arnd Bergmann wrote:
> On Thu, Mar 18, 2021 at 5:13 PM Alessio Balsini <balsini@android.com> wrote:
> > On Tue, Mar 16, 2021 at 07:53:06PM +0100, Arnd Bergmann wrote:
> > > On Mon, Jan 25, 2021 at 4:48 PM Alessio Balsini <balsini@android.com> wrote:
> > > >
> >
> > Thanks for spotting this possible criticality.
> >
> > I noticed that 32-bit users pace was unable to use the
> > FUSE_DEV_IOC_CLONE ioctl on 64-bit kernels, so this change avoid this
> > issue by forcing the kernel to interpret 32 and 64 bit
> > FUSE_DEV_IOC_CLONE command as if they were the same.
> 
> As far as I can tell from the kernel headers, the command code should
> be the same for both 32-bit and 64-bit tasks: 0x8004e500.
> Can you find out what exact value you see in the user space that was
> causing problems, and how it ended up with a different value than
> the 64-bit version?
> 
> If there are two possible command codes, I'd suggest you just change
> the driver to handle both variants explicitly, but not any other one.
> 
> > This is the simplest solution I could find as the UAPI is not changed
> > as, as you mentioned, the argument doesn't require any conversion.
> >
> > I understand that this might limit possible future extensions of the
> > FUSE_DEV_IOC_XXX ioctls if their in/out argument changed depending on
> > the architecture, but only at that point we can switch to using the
> > compat layer, right?
> >
> > What I'm worried about is the direction, do you think this would be an
> > issue?
> >
> > I can start working on a compat layer fix meanwhile.
> 
> For a proper well-designed ioctl interface, compat support should not
> need anything beyond the '.compat_ioctl = compat_ptr_ioctl'
> assignment.
> 
>        Arnd

You are right and now I can see what happened here.

When I introduce the PASSTHROUGH ioctl, because of the 'void *', the
command mismatches on _IOC_SIZE(nr). I solved this by only testing
_IOC_NUMBER and _IOC_TYPE, implicitely (mistakenly) removing the
_IOC_SIZE check.  So, the fuse_dev_ioctl was correctly rejecting the
ioctl request from 32-bit userspace because of the wrong size and I was
just forcing it to digest the wrong data regardless.
Ouch!

The commit message for this patch would still be incorrect, but I posted
a fix here to bring the ioctl checking back to the original behavior:

  https://lore.kernel.org/lkml/20210319150514.1315985-1-balsini@android.com/

Thanks for spotting this!
Alessio