Content-Type: text/plain; charset=utf-8
Mime-Version: 1.0 (Mac OS X Mail 10.3 \(3273\))
Subject: Re: [RFC] x86: Avoid CR3 load on compatibility mode with PTI
From: Nadav Amit <nadav.amit@gmail.com>
In-Reply-To: <20180116004128.us5uprkzrr5gf4li@gmail.com>
Date: Mon, 15 Jan 2018 19:49:16 -0800
Cc: Dave Hansen <dave.hansen@linux.intel.com>,
        LKML <linux-kernel@vger.kernel.org>,
        Andy Lutomirski <luto@kernel.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>,
        "H. Peter Anvin" <hpa@zytor.com>,
        the arch/x86 maintainers <x86@kernel.org>, w@1wt.eu,
        Linus Torvalds <torvalds@linux-foundation.org>
Content-Transfer-Encoding: 8BIT
Message-Id: <5D6CD440-B20F-4ABF-8B02-EE87205B661D@gmail.com>
References: <20180114201306.3554-1-namit@vmware.com>
 <57a8fa6b-a1d1-d440-ce13-b1d06d265584@linux.intel.com>
 <3D823F02-89EF-48D9-913D-5E65391F6F9D@gmail.com>
 <20180116004128.us5uprkzrr5gf4li@gmail.com>
To: Ingo Molnar <mingo@kernel.org>
Sender: linux-kernel-owner@vger.kernel.org

Ingo Molnar <mingo@kernel.org> wrote:

> 
> * Nadav Amit <nadav.amit@gmail.com> wrote:
> 
>>> Also, what's the end goal here?  Run old 32-bit binaries better?  You
>>> want to weaken the security of the whole implementation to do that?
>>> Sounds like a bad tradeoff to me.
>> 
>> As Willy noted in this thread, I think that some users may be interested in 
>> running 32-bit Apache/Nginx/Redis to get the performance back without 
>> sacrificing security.
> 
> Note that it is a flawed assumption to think that this is possible, as they might 
> in many cases not be getting their performance back: 32-bit binaries for the same 
> general CPU bound computation can easily be 5% slower than 64-bit binaries (as 
> long as the larger cache footprint of 64-bit data doesn't fall out of key caches), 
> but can be up to 30% slower for certain computations.
> 
> In fact, depending on how kernel heavy the web workload is (for example how much 
> CGI processing versus IO it does, etc.), a 32-bit binary could be distinctly 
> _slower_ than even a PTI-enabled 64-bit binary.

Obviously you are right - I didn’t argue otherwise - and I think it is also
reflected in the results (Redis LRANGE results). Yet, arguably the workloads
that are affected the most by PTI are those with a high number of syscalls
and interrupts, in which user computation time is relatively small.

> So we are trading a 5-15% slowdown (PTI) for another 5-15% slowdown, plus we are
> losing the soft-SMEP feature on older CPUs that PTI enables, which is a pretty 
> powerful mitigation technique.

This soft-SMEP can be kept by keeping PTI if SMEP is unsupported. Although
we trade slowdowns, they are different ones, which allows the user to make
his best decision.

> Yes, I suspect in some (maybe many) cases it would be a speedup, but I really 
> don't like the underlying assumptions and tradeoffs here. (Not that I like any of 
> this whole Meltdown debacle TBH.)

To make sure that I understand correctly - the assumptions are that
disabling PTI on compatibility mode would: (1) Benefit some workloads; (2)
Be useful, even if we only consider CPUs with SMEP; and (3) Secure.

Under these assumptions, the tradeoff is slightly greater code complexity
for considerably better performance of 32-bit code; in some common cases
this makes 32-bit code to perform significantly better than 64-bit code.

Am I missing something? My main concern was initially security, but so far
from your aggregated feedback I did not see something concrete which cannot
relatively easily be addressed.