Message-ID: <CACT4Y+aDeSAARG0b9FjDFyWuhjb=YVxpGtsvBmoKnHo+0TF4gA@mail.gmail.com>
Date:   Thu, 9 Apr 2020 09:02:54 +0200
From:   Dmitry Vyukov <dvyukov@...gle.com>
To:     Stéphane Graber <stgraber@...ntu.com>
Cc:     Jann Horn <jannh@...gle.com>,
        Christian Brauner <christian.brauner@...ntu.com>,
        Jens Axboe <axboe@...nel.dk>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        kernel list <linux-kernel@...r.kernel.org>,
        linux-block <linux-block@...r.kernel.org>,
        Linux API <linux-api@...r.kernel.org>,
        Jonathan Corbet <corbet@....net>,
        Serge Hallyn <serge@...lyn.com>,
        "Rafael J. Wysocki" <rafael@...nel.org>, Tejun Heo <tj@...nel.org>,
        "David S. Miller" <davem@...emloft.net>,
        Saravana Kannan <saravanak@...gle.com>,
        Jan Kara <jack@...e.cz>, David Howells <dhowells@...hat.com>,
        Seth Forshee <seth.forshee@...onical.com>,
        David Rheinsberg <david.rheinsberg@...il.com>,
        Tom Gundersen <teg@...m.no>,
        Christian Kellner <ckellner@...hat.com>,
        "open list:DOCUMENTATION" <linux-doc@...r.kernel.org>,
        Network Development <netdev@...r.kernel.org>,
        Matthew Garrett <mjg59@...gle.com>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        syzkaller <syzkaller@...glegroups.com>
Subject: Re: [PATCH 0/8] loopfs

On Wed, Apr 8, 2020 at 6:41 PM Stéphane Graber <stgraber@...ntu.com> wrote:
>
> On Wed, Apr 8, 2020 at 12:24 PM Jann Horn <jannh@...gle.com> wrote:
> >
> > On Wed, Apr 8, 2020 at 5:23 PM Christian Brauner
> > <christian.brauner@...ntu.com> wrote:
> > > One of the use-cases for loopfs is to allow to dynamically allocate loop
> > > devices in sandboxed workloads without exposing /dev or
> > > /dev/loop-control to the workload in question and without having to
> > > implement a complex and also racy protocol to send around file
> > > descriptors for loop devices. With loopfs each mount is a new instance,
> > > i.e. loop devices created in one loopfs instance are independent of any
> > > loop devices created in another loopfs instance. This allows
> > > sufficiently privileged tools to have their own private stash of loop
> > > device instances. Dmitry has expressed his desire to use this for
> > > syzkaller in a private discussion. And various parties that want to use
> > > it are Cced here too.
> > >
> > > In addition, the loopfs filesystem can be mounted by user namespace root
> > > and is thus suitable for use in containers. Combined with syscall
> > > interception this makes it possible to securely delegate mounting of
> > > images on loop devices, i.e. when a user calls mount -o loop <image>
> > > <mountpoint> it will be possible to completely setup the loop device.
> > > The final mount syscall to actually perform the mount will be handled
> > > through syscall interception and be performed by a sufficiently
> > > privileged process. Syscall interception is already supported through a
> > > new seccomp feature we implemented in [1] and extended in [2] and is
> > > actively used in production workloads. The additional loopfs work will
> > > be used there and in various other workloads too. You'll find a short
> > > illustration how this works with syscall interception below in [4].
> >
> > Would that privileged process then allow you to mount your filesystem
> > images with things like ext4? As far as I know, the filesystem
> > maintainers don't generally consider "untrusted filesystem image" to
> > be a strongly enforced security boundary; and worse, if an attacker
> > has access to a loop device from which something like ext4 is mounted,
> > things like "struct ext4_dir_entry_2" will effectively be in shared
> > memory, and an attacker can trivially bypass e.g.
> > ext4_check_dir_entry(). At the moment, that's not a huge problem (for
> > anything other than kernel lockdown) because only root normally has
> > access to loop devices.
> >
> > Ubuntu carries an out-of-tree patch that afaik blocks the shared
> > memory thing: <https://git.launchpad.net/~ubuntu-kernel/ubuntu/+source/linux/+git/eoan/commit?id=4bc428fdf5500b7366313f166b7c9c50ee43f2c4>
> >
> > But even with that patch, I'm not super excited about exposing
> > filesystem image parsing attack surface to containers unless you run
> > the filesystem in a sandboxed environment (at which point you don't
> > need a loop device anymore either).
>
> So in general we certainly agree that you should never expose someone
> that you wouldn't trust with root on the host to syscall interception
> mounting of real kernel filesystems.
>
> But that's not all that our syscall interception logic can do. We have
> support for rewriting a normal filesystem mount attempt to instead use
> an available FUSE implementation. As far as the user is concerned,
> they ran "mount /dev/sdaX /mnt" and got that ext4 filesystem mounted
> on /mnt as requested, except that the container manager intercepted
> the mount attempt and instead spawned fuse2fs for that mount. This
> requires absolutely no change to the software the user is running.
>
> loopfs, with that interception mode, will let us also handle all cases
> where a loop would be used, similarly without needing any change to
> the software being run. If a piece of software calls the command
> "mount -o loop blah.img /mnt", the "mount" command will setup a loop
> device as it normally would (doing so through loopfs) and then will
> call the "mount" syscall, which will get intercepted and redirected to
> a FUSE implementation if so configured, resulting in the expected
> filesystem being mounted for the user.
>
> LXD with syscall interception offers both straight up privileged
> mounting using the kernel fs or using a FUSE based implementation.
> This is configurable on a per-filesystem and per-container basis.
>
> I hope that clarifies what we're doing here :)
>
> Stéphane


Hi Christian,

Our use case for loopfs in syzkaller would be isolation of several
test processes from each other.
Currently all loop devices and loop-control are global and cause test
processes to collide, which in turn causes non-reproducible coverage
and non-reproducible crashes. Ideally we give each test process its
own loopfs instance.

Powered by blists - more mailing lists