Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Content-Type: text/plain;
        charset=utf-8
Mime-Version: 1.0 (1.0)
Subject: Re: [PATCH] x86: uaccess: fix regression in unsafe_get_user
From:   Andy Lutomirski <luto@amacapital.net>
In-Reply-To: <alpine.DEB.2.21.1902161358160.1683@nanos.tec.linutronix.de>
Date:   Sat, 16 Feb 2019 14:50:15 -0800
Cc:     Jann Horn <jannh@google.com>, baloo@gandi.net,
        the arch/x86 maintainers <x86@kernel.org>,
        Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
        kernel list <linux-kernel@vger.kernel.org>,
        Pascal Bouchareine <pascal@gandi.net>
Content-Transfer-Encoding: quoted-printable
Message-Id: <4F2693EA-1553-4F09-9475-781305540DBC@amacapital.net>
References: <20190215235901.23541-1-baloo@gandi.net> <CAG48ez2tYizTKncevLF=AMQ2nm3D=SqGHH5bM5f-U0fhQ1nL9Q@mail.gmail.com> <alpine.DEB.2.21.1902161358160.1683@nanos.tec.linutronix.de>
To:     Thomas Gleixner <tglx@linutronix.de>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


> On Feb 16, 2019, at 12:18 PM, Thomas Gleixner <tglx@linutronix.de> wrote:
>=20
>> On Sat, 16 Feb 2019, Jann Horn wrote:
>>> On Sat, Feb 16, 2019 at 12:59 AM <baloo@gandi.net> wrote:
>>> When extracting an initramfs, a filename may be near an allocation bound=
ary.
>>> Should that happen, strncopy_from_user will invoke unsafe_get_user which=

>>> may cross the allocation boundary. Should that happen, unsafe_get_user w=
ill
>>> trigger a page fault, and strncopy_from_user would then bailout to
>>> byte_at_a_time behavior.
>>>=20
>>> unsafe_get_user is unsafe by nature, and rely on pagefault to detect bou=
ndaries.
>>> After 9da3f2b74054 ("x86/fault: BUG() when uaccess helpers fault on kern=
el addresses")
>>> it may no longer rely on pagefault as the new page fault handler would
>>> trigger a BUG().
>>>=20
>>> This commit allows unsafe_get_user to explicitly trigger pagefaults and
>>> handle them directly with the error target label.
>>=20
>> Oof. So basically the init code is full of things that just call
>> syscalls instead of using VFS functions (which don't actually exist
>> for everything), and the VFS syscalls use getname_flags(), which uses
>> strncpy_from_user(), which can access out-of-bounds pages on
>> architectures that set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, and
>> that in summary means that all the init code is potentially prone to
>> tripping over this?
>=20
> Not all init code. It should be only the initramfs decompression.
>=20
>> I don't particularly like this approach to fixing it, but I also don't
>> have any better ideas, so I guess unless someone else has a bright
>> idea, this patch might have to go in.
>=20
> So we know that this happens in the context of decompress() which calls
> flush_buffer() for every chunk. flush_buffer() gets the start_address and
> the length. We also know that the fault can only happen within:
>=20
>    start_address <=3D fault_address < start_address + length + 8;
>=20
> So something like the untested workaround below should cover the initramfs=

> oddity and avoid to weaken the protection for all other cases.

What is the actual problem?  We=E2=80=99re not actually demand-faulting this=
 data, are we?  Are we just overrunning the buffer because the from_user hel=
pers are too clever?  Can we fix it for real by having the fancy helpers do *=
aligned* loads so that they don=E2=80=99t overrun the buffer?  Heck, this mi=
ght be faster, too.