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:   Mon, 10 Feb 2020 11:51:17 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Kees Cook <keescook@chromium.org>
Cc:     Andy Lutomirski <luto@amacapital.net>,
        Kristen Carlson Accardi <kristen@linux.intel.com>,
        tglx@linutronix.de, mingo@redhat.com, bp@alien8.de, hpa@zytor.com,
        arjan@linux.intel.com, rick.p.edgecombe@intel.com, x86@kernel.org,
        linux-kernel@vger.kernel.org, kernel-hardening@lists.openwall.com
Subject: Re: [RFC PATCH 06/11] x86: make sure _etext includes function
 sections
Message-ID: <20200210105117.GE14879@hirez.programming.kicks-ass.net>
References: <75f0bd0365857ba4442ee69016b63764a8d2ad68.camel@linux.intel.com>
 <B413445A-F1F0-4FB7-AA9F-C5FF4CEFF5F5@amacapital.net>
 <20200207092423.GC14914@hirez.programming.kicks-ass.net>
 <202002091742.7B1E6BF19@keescook>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <202002091742.7B1E6BF19@keescook>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Sun, Feb 09, 2020 at 05:43:40PM -0800, Kees Cook wrote:
> On Fri, Feb 07, 2020 at 10:24:23AM +0100, Peter Zijlstra wrote:
> > On Thu, Feb 06, 2020 at 12:02:36PM -0800, Andy Lutomirski wrote:
> > > Also, in the shiny new era of
> > > Intel-CPUs-can’t-handle-Jcc-spanning-a-cacheline, function alignment
> > > may actually matter.
> > 
> > *groan*, indeed. I just went and looked that up. I missed this one in
> > all the other fuss :/
> > 
> > So per:
> > 
> >   https://www.intel.com/content/dam/support/us/en/documents/processors/mitigations-jump-conditional-code-erratum.pdf
> > 
> > the toolchain mitigations only work if the offset in the ifetch window
> > (32 bytes) is preserved. Which seems to suggest we ought to align all
> > functions to 32byte before randomizing it, otherwise we're almost
> > guaranteed to change this offset by the act of randomizing.
> 
> Wheee! This sounds like in needs to be fixed generally, yes? (And I see
> "FUNCTION_ALIGN" macro is currently 16 bytes...

It depends a bit on how it all works I suppose (I'm not too clear on the
details).

Suppose the linker appends translation units at (at least) 32 bytes
alignment, but the function alignment inside the translation unit is
smaller, then it could still work, because the assembler (which is going
to insert NOPs to avoid instructions being in the 'wrong' place) can
still know the offset.

If the linker is going to be fancy (say LTO) and move code around inside
sections/translation units, then this goes out the window obviously.

The same with this fine-grained-randomization, if the section alignment
is smaller than 32 bytes, the offset is going to change and the
mitigation will be nullified.

I'll leave it to others to figure out the exact details. But afaict it
should be possible to have fine-grained-randomization and preserve the
workaround in the end.