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;
Date:   Sat, 20 Jan 2018 08:56:34 -0800
From:   Alexei Starovoitov <alexei.starovoitov@gmail.com>
To:     Dan Williams <dan.j.williams@intel.com>
Cc:     Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Mark Rutland <mark.rutland@arm.com>,
        Kernel Hardening <kernel-hardening@lists.openwall.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Catalin Marinas <catalin.marinas@arm.com>,
        Will Deacon <will.deacon@arm.com>,
        "H. Peter Anvin" <hpa@zytor.com>,
        Elena Reshetova <elena.reshetova@intel.com>,
        linux-arch <linux-arch@vger.kernel.org>,
        Andi Kleen <ak@linux.intel.com>,
        Jonathan Corbet <corbet@lwn.net>, X86 ML <x86@kernel.org>,
        Russell King <linux@armlinux.org.uk>,
        Ingo Molnar <mingo@redhat.com>,
        Andrew Honig <ahonig@google.com>,
        Alan Cox <alan@linux.intel.com>,
        Tom Lendacky <thomas.lendacky@amd.com>,
        Kees Cook <keescook@chromium.org>,
        Al Viro <viro@zeniv.linux.org.uk>,
        Andy Lutomirski <luto@kernel.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Andrew Morton <akpm@linux-foundation.org>,
        Jim Mattson <jmattson@google.com>,
        Christian Lamparter <chunkeey@gmail.com>,
        Greg KH <gregkh@linuxfoundation.org>,
        Linux Wireless List <linux-wireless@vger.kernel.org>,
        stable@vger.kernel.org, Paolo Bonzini <pbonzini@redhat.com>,
        Johannes Berg <johannes@sipsolutions.net>,
        Linus Torvalds <torvalds@linux-foundation.org>,
        "David S. Miller" <davem@davemloft.net>
Subject: Re: [PATCH v4 00/10] prevent bounds-check bypass via speculative
 execution
Message-ID: <20180120165631.e7c3kipmhb5sckor@ast-mbp>
References: <151632009605.21271.11304291057104672116.stgit@dwillia2-desk3.amr.corp.intel.com>
 <CAPcyv4iBHTahi66deUT1iSPpA2W2wjYZSaiUYUjpdK10d5Rt7g@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAPcyv4iBHTahi66deUT1iSPpA2W2wjYZSaiUYUjpdK10d5Rt7g@mail.gmail.com>
User-Agent: NeoMutt/20170421 (1.8.2)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Fri, Jan 19, 2018 at 10:58:44PM -0800, Dan Williams wrote:
> On Thu, Jan 18, 2018 at 4:01 PM, Dan Williams <dan.j.williams@intel.com> wrote:
> > Changes since v3 [1]
> > * Drop 'ifence_array_ptr' and associated compile-time + run-time
> >   switching and just use the masking approach all the time.
> >
> > * Convert 'get_user' to use pointer sanitization via masking rather than
> >   lfence. '__get_user' and associated paths still rely on
> >   lfence. (Linus)
> >
> >       "Basically, the rule is trivial: find all 'stac' users, and use
> >        address masking if those users already integrate the limit
> >        check, and lfence they don't."
> >
> > * At syscall entry sanitize the syscall number under speculation
> >   to remove a user controlled pointer de-reference in kernel
> >   space.  (Linus)
> >
> > * Fix a raw lfence in the kvm code (added for v4.15-rc8) to use
> >   'array_ptr'.
> >
> > * Propose 'array_idx' as a way to sanitize user input that is
> >   later used as an array index, but where the validation is
> >   happening in a different code block than the array reference.
> >   (Christian).
> >
> > * Fix grammar in speculation.txt (Kees)
> >
> > ---
> >
> > Quoting Mark's original RFC:
> >
> > "Recently, Google Project Zero discovered several classes of attack
> > against speculative execution. One of these, known as variant-1, allows
> > explicit bounds checks to be bypassed under speculation, providing an
> > arbitrary read gadget. Further details can be found on the GPZ blog [2]
> > and the Documentation patch in this series."
> >
> > A precondition of using this attack on the kernel is to get a user
> > controlled pointer de-referenced (under speculation) in privileged code.
> > The primary source of user controlled pointers in the kernel is the
> > arguments passed to 'get_user' and '__get_user'. An example of other
> > user controlled pointers are user-controlled array / pointer offsets.
> >
> > Better tooling is needed to find more arrays / pointers with user
> > controlled indices / offsets that can be converted to use 'array_ptr' or
> > 'array_idx'. A few are included in this set, and these are not expected
> > to be complete. That said, the 'get_user' protections raise the bar on
> > finding a vulnerable gadget in the kernel.
> >
> > These patches are also available via the 'nospec-v4' git branch here:
> >
> >     git://git.kernel.org/pub/scm/linux/kernel/git/djbw/linux nospec-v4
> 
> I've pushed out a nospec-v4.1 with the below minor cleanup, a fixup of
> the changelog for "kvm, x86: fix spectre-v1 mitigation", and added
> Paolo's ack.
> 
>      git://git.kernel.org/pub/scm/linux/kernel/git/djbw/linux nospec-v4.1
> 
> diff --git a/include/linux/nospec.h b/include/linux/nospec.h
> index 8af35be1869e..b8a9222e34d1 100644
> --- a/include/linux/nospec.h
> +++ b/include/linux/nospec.h
> @@ -37,7 +37,7 @@ static inline unsigned long array_ptr_mask(unsigned
> long idx, unsigned long sz)
>         unsigned long _i = (idx);                                       \
>         unsigned long _mask = array_ptr_mask(_i, (sz));                 \
>                                                                         \
> -       __u._ptr = _arr + (_i & _mask);                                 \
> +       __u._ptr = _arr + _i;                                           \
>         __u._bit &= _mask;                                              \
>         __u._ptr;                                                       \

hmm. I'm not sure it's the right thing to do, since the macro
is forcing cpu to speculate subsequent load from null instead
of valid pointer.
As Linus said: "
 So that __u._bit masking wasn't masking
 the pointer, it was masking the value that was *read* from the
 pointer, so that you could know that an invalid access returned
 0/NULL, not just the first value in the array.
"
imo just
  return _arr + (_i & _mask);
is enough. No need for union games.
The cpu will speculate the load from _arr[0] if _i is out of bounds
which is the same as if user passed _i == 0 which would have passed
bounds check anyway, so I don't see any data leak from populating
cache with _arr[0] data. In-bounds access can do that just as well
without any speculation.