Date: Fri, 5 Jan 2018 02:40:55 -0800
From: Paul Turner <pjt@google.com>
To: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Alexei Starovoitov <alexei.starovoitov@gmail.com>,
        David Woodhouse <dwmw@amazon.co.uk>,
        Andi Kleen <ak@linux.intel.com>,
        LKML <linux-kernel@vger.kernel.org>,
        Greg Kroah-Hartman <gregkh@linux-foundation.org>,
        Tim Chen <tim.c.chen@linux.intel.com>,
        Dave Hansen <dave.hansen@intel.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Kees Cook <keescook@google.com>,
        Rik van Riel <riel@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Andy Lutomirski <luto@amacapital.net>,
        Jiri Kosina <jikos@kernel.org>,
        One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk>
Subject: Re: [PATCH v3 01/13] x86/retpoline: Add initial retpoline support
Message-ID: <20180105104055.GA253582@google.com>
References: <1515058213.12987.89.camel@amazon.co.uk>
 <20180104143710.8961-1-dwmw@amazon.co.uk>
 <20180104181744.komdplek7nfdvlsw@ast-mbp>
 <CA+55aFxRwAEe1VH-E_HMPf+jpxAUWaUkwLbJLXG2U44qHga+Dg@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CA+55aFxRwAEe1VH-E_HMPf+jpxAUWaUkwLbJLXG2U44qHga+Dg@mail.gmail.com>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org

On Thu, Jan 04, 2018 at 10:25:35AM -0800, Linus Torvalds wrote:
> On Thu, Jan 4, 2018 at 10:17 AM, Alexei Starovoitov
> <alexei.starovoitov@gmail.com> wrote:
> >
> > Clearly Paul's approach to retpoline without lfence is faster.

Using pause rather than lfence does not represent a fundamental difference here.

A protected indirect branch is always adding ~25-30 cycles of overhead.

That this can be avoided in practice is a function of two key factors:
(1) Kernel code uses fewer indirect branches.
(2) The overhead can be avoided for hot indirect branches via devirtualization.
  e.g. the semantic equivalent of,
    if (ptr == foo)
      foo();
    else
      (*ptr)();
  Allowing foo() to be called directly, even though it was provided as an
  indirect.

> > I'm guessing it wasn't shared with amazon/intel until now and
> > this set of patches going to adopt it, right?
> >
> > Paul, could you share a link to a set of alternative gcc patches
> > that do retpoline similar to llvm diff ?
> 
> What is the alternative approach? Is it literally just doing a
> 
>       call 1f
> 1:    mov real_target,(%rsp)
>        ret
> 
> on the assumption that the "ret" will always just predict to that "1"
> due to the call stack?
> 
>                 Linus