Received-SPF: softfail (google.com: domain of transitioning linux-kernel-owner@vger.kernel.org does not designate 23.128.96.19 as permitted sender) client-ip=23.128.96.19;
MIME-Version: 1.0
References: <20220420004241.2093-1-joao@overdrivepizza.com>
 <20220420074044.GC2731@worktop.programming.kicks-ass.net> <20220420151714.fderdz4dzea75rvg@treble>
In-Reply-To: <20220420151714.fderdz4dzea75rvg@treble>
From:   Nick Desaulniers <ndesaulniers@google.com>
Date:   Wed, 20 Apr 2022 10:12:59 -0700
Message-ID: <CAKwvOdm6J9TFUWJA2Q7VRjx1CpBme9z_Dp99+7YTjN5Jqp5=Aw@mail.gmail.com>
Subject: Re: [RFC PATCH 00/11] Kernel FineIBT Support
To:     Josh Poimboeuf <jpoimboe@redhat.com>, joao@overdrivepizza.com,
        hjl.tools@gmail.com, Fangrui Song <maskray@google.com>
Cc:     Peter Zijlstra <peterz@infradead.org>,
        linux-kernel@vger.kernel.org, linux-hardening@vger.kernel.org,
        andrew.cooper3@citrix.com, keescook@chromium.org,
        samitolvanen@google.com, mark.rutland@arm.com,
        alyssa.milburn@linux.intel.com, gabriel.gomes@linux.intel.com,
        rick.p.edgecombe@intel.com
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Apr 20, 2022 at 8:17 AM Josh Poimboeuf <jpoimboe@redhat.com> wrote:
>
> On Wed, Apr 20, 2022 at 09:40:44AM +0200, Peter Zijlstra wrote:
> > On Tue, Apr 19, 2022 at 05:42:30PM -0700, joao@overdrivepizza.com wrote:
> > > @PeterZ @JoshP
> > >
> > > I'm a bit unaware of the details on why the objtool approach to bypass ENDBRs
> > > was removed from the IBT series. Is this approach now sensible considering that
> > > it is a requirement for a new/enhanced feature? If not, how extending the Linker
> > > to emit already fixed offsets sounds like?
> >
> > Josh hates objtool modifying actualy code. He much prefers objtool only
> > emits out of band data.
> >
> > Now, I did sneak in that jump_label nop'ing, and necessity (broken
> > compilers) had us do the KCOV nop'ing in noinstr, but if you look at the
> > recent objtool series here:
> >
> >   https://lkml.kernel.org/r/cover.1650300597.git.jpoimboe@redhat.com
> >
> > you'll see his thoughs on that :-)
> >
> > Now, I obviously don't mind, it's easy enough to figure out what objtool
> > actually does with something like:
> >
> >   $ OBJTOOL_ARGS="--backup" make O=ibt-build/ -j$lots vmlinux
> >   $ objdiff.sh ibt-build/vmlinux.o
> >
> > Where objdiff.sh is the below crummy script.
> >
> > Now, one compromise that I did get out of Josh was that he objected less
> > to rewriting relocations than to rewriting the immediates. From my
> > testing the relocations got us the vast majority of direct call sites,
> > very few are immediates.
> >
> > Josh, any way you might reconsider all that? :-)
>
> If I remember correctly, the goal of --ibt-fix-direct was to avoid
> hitting unnecessary ENDBRs, which basically decode to NOPs, so the
> ghastly hack wasn't worth it.
>
> If FineIBT needs it, I could reconsider.  But I think there's a strong
> case to be made that the linker should be doing that instead.

That sounds reasonable to me (and reminds me of linker relaxation).
Joao, can you please work with Fangrui (LLD) and HJ (GNU binutils) to
determine how feasible this would be? I assume code outside the kernel
might enjoy such an optimization, too.  When that's the case, then it
probably makes more sense to "upstream" such "optimizations" from the
kernel-specific objtool into the toolchains.
-- 
Thanks,
~Nick Desaulniers