Re: [PATCH] x86/bugs: protect against userspace-userspace spectreRSB

From: Tim Chen <tim.c.chen@linux.intel.com>
To: Josh Poimboeuf <jpoimboe@redhat.com>, Jiri Kosina <jikos@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>,
	Ingo Molnar <mingo@redhat.com>,
	Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>,
	Borislav Petkov <bp@suse.de>, David Woodhouse <dwmw@amazon.co.uk>,
	Peter Zijlstra <peterz@infradead.org>,
	linux-kernel@vger.kernel.org, x86@kernel.org,
	Linus Torvalds <torvalds@linux-foundation.org>
Subject: Re: [PATCH] x86/bugs: protect against userspace-userspace spectreRSB
Date: Mon, 30 Jul 2018 10:59:50 -0700	[thread overview]
Message-ID: <d3ace1f0-15c0-e14e-c35b-24e880d7abc2@linux.intel.com> (raw)
In-Reply-To: <20180725171124.kklnori7tf56ngeu@treble>

On 07/25/2018 10:11 AM, Josh Poimboeuf wrote:
> On Wed, Jul 25, 2018 at 03:50:44PM +0200, Jiri Kosina wrote:
>> On Wed, 25 Jul 2018, Josh Poimboeuf wrote:
>>
>>>> The article "Spectre Returns! Speculation Attacks using the Return Stack
>>>> Buffer" [1] describes two new (sub-)variants of spectrev2-like attack,
>>>> making use solely of the RSB contents even on CPUs that don't fallback to
>>>> BTB on RSB underflow (Skylake+).
>>>>
>>>> Mitigate userspace-userspace attacks by always unconditionally filling RSB on
>>>> context switch when generic spectrev2 mitigation has been enabled.
>>>>
>>>> [1] https://arxiv.org/pdf/1807.07940.pdf
>>>>
>>>> Signed-off-by: Jiri Kosina <jkosina@suse.cz>
>>>
>>> While I generally agree with this patch, isn't it odd that we would do
>>> RSB filling on every context switch, but almost never do IBPB?
>>
>> Yeah, I have actually been wondering exactly the same, but that's what we 
>> have been doing so far on SKL+, so I didn't really want to mix this aspect 
>> in.
>>
>> I actually believe that in the name of consistency we should've been doing 
>> the RSB fills under the same conditions we're issuing IBPB even on SKL+; I 
>> can resend a patch that re-adjusts that, if that's the consensus.
> 
> True, in theory it might make more sense to only fill RSB when doing an
> IBPB.  But given the current state of almost never doing IBPB, that
> would be pointless.  RSB is cheap enough that we should just do it
> unconditionally on context switch.
> 
> BTW, I've heard that IBPB actually flushes RSB, though I haven't seen
> that officially documented anywhere.  Not that it matters given the
> current IBPB code.

I think that's correct.  IBPB does flushes the RSB.  But doing
RSB stuffing will be cheaper.

Tim

> 
> BTW^2, there was some discussion a few months back about offloading the
> "when to IBPB" decision to security modules, though I don't think I've
> ever seen official patches for that.
> 
> All that said, this patch is fine until if/when the IBPB strategy gets
> figured out.  RSB filling is cheap.
> 
> Reviewed-by: Josh Poimboeuf <jpoimboe@redhat.com>
> 