Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 31 Mar 2023 23:56:10 +0000
In-Reply-To: <20230322111951.vfrm2xf4o5kmtte6@wittgenstein> (message from
 Christian Brauner on Wed, 22 Mar 2023 12:19:51 +0100)
Mime-Version: 1.0
Message-ID: <diqzmt3sqxut.fsf@ackerleytng-cloudtop.c.googlers.com>
Subject: Re: [RFC PATCH v2 1/2] mm: restrictedmem: Allow userspace to specify
 mount for memfd_restricted
From:   Ackerley Tng <ackerleytng@google.com>
To:     Christian Brauner <brauner@kernel.org>
Cc:     kvm@vger.kernel.org, linux-api@vger.kernel.org,
        linux-arch@vger.kernel.org, linux-doc@vger.kernel.org,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org, qemu-devel@nongnu.org, aarcange@redhat.com,
        ak@linux.intel.com, akpm@linux-foundation.org, arnd@arndb.de,
        bfields@fieldses.org, bp@alien8.de, chao.p.peng@linux.intel.com,
        corbet@lwn.net, dave.hansen@intel.com, david@redhat.com,
        ddutile@redhat.com, dhildenb@redhat.com, hpa@zytor.com,
        hughd@google.com, jlayton@kernel.org, jmattson@google.com,
        joro@8bytes.org, jun.nakajima@intel.com,
        kirill.shutemov@linux.intel.com, linmiaohe@huawei.com,
        luto@kernel.org, mail@maciej.szmigiero.name, mhocko@suse.com,
        michael.roth@amd.com, mingo@redhat.com, naoya.horiguchi@nec.com,
        pbonzini@redhat.com, qperret@google.com, rppt@kernel.org,
        seanjc@google.com, shuah@kernel.org, steven.price@arm.com,
        tabba@google.com, tglx@linutronix.de, vannapurve@google.com,
        vbabka@suse.cz, vkuznets@redhat.com, wanpengli@tencent.com,
        wei.w.wang@intel.com, x86@kernel.org, yu.c.zhang@linux.intel.com
Content-Type: text/plain; charset="UTF-8"; format=flowed; delsp=yes
Precedence: bulk

Christian Brauner <brauner@kernel.org> writes:

> On Tue, Mar 21, 2023 at 08:15:32PM +0000, Ackerley Tng wrote:
>> By default, the backing shmem file for a restrictedmem fd is created
>> on shmem's kernel space mount.

>> ...

Thanks for reviewing this patch!


> This looks like you can just pass in some tmpfs fd and you just use it
> to identify the mnt and then you create a restricted memfd area in that
> instance. So if I did:

> mount -t tmpfs tmpfs /mnt
> mknod /mnt/bla c 0 0
> fd = open("/mnt/bla")
> memfd_restricted(fd)

> then it would create a memfd restricted entry in the tmpfs instance
> using the arbitrary dummy device node to infer the tmpfs instance.

> Looking at the older thread briefly and the cover letter. Afaict, the
> new mount api shouldn't figure into the design of this. fsopen() returns
> fds referencing a VFS-internal fs_context object. They can't be used to
> create or lookup files or identify mounts. The mount doesn't exist at
> that time. Not even a superblock might exist at the time before
> fsconfig(FSCONFIG_CMD_CREATE).

> When fsmount() is called after superblock setup then it's similar to any
> other fd from open() or open_tree() or whatever (glossing over some
> details that are irrelevant here). Difference is that open_tree() and
> fsmount() would refer to the root of a mount.

This is correct, memfd_restricted() needs an fd returned from fsmount()
and not fsopen(). Usage examples of this new parameter in
memfd_restricted() are available in selftests.


> At first I wondered why this doesn't just use standard *at() semantics
> but I guess the restricted memfd is unlinked and doesn't show up in the
> tmpfs instance.

> So if you go down that route then I would suggest to enforce that the
> provided fd refer to the root of a tmpfs mount. IOW, it can't just be an
> arbitrary file descriptor in a tmpfs instance. That seems cleaner to me:

> sb = f_path->mnt->mnt_sb;
> sb->s_magic == TMPFS_MAGIC && f_path->mnt->mnt_root == sb->s_root

> and has much tigher semantics than just allowing any kind of fd.

Thanks for your suggestion, I've tightened the semantics as you
suggested. memfd_restricted() now only accepts fds representing the root
of the mount.


> Another wrinkly I find odd but that's for you to judge is that this
> bypasses the permission model of the tmpfs instance. IOW, as long as you
> have a handle to the root of a tmpfs mount you can just create
> restricted memfds in there. So if I provided a completely sandboxed
> service - running in a user namespace or whatever - with an fd to the
> host's tmpfs instance they can just create restricted memfds in there no
> questions asked.

> Maybe that's fine but it's certainly something to spell out and think
> about the implications.

Thanks for pointing this out! I added a permissions check in RFC v3, and
clarified the permissions model (please see patch 1 of 2):
https://lore.kernel.org/lkml/cover.1680306489.git.ackerleytng@google.com/