Re: [PATCH v3 5/7] x86: guard against straight-line speculation past RET

[Jan Beulich <jbeulich@xxxxxxxx>]

On 10.11.2020 15:08, Roger Pau Monné wrote:
> On Tue, Nov 10, 2020 at 02:19:40PM +0100, Jan Beulich wrote:
>> On 10.11.2020 12:16, Roger Pau Monné wrote:
>>> On Tue, Nov 10, 2020 at 11:06:46AM +0100, Jan Beulich wrote:
>>>> On 10.11.2020 10:31, Roger Pau Monné wrote:
>>>>> On Fri, Oct 23, 2020 at 10:38:04AM +0200, Jan Beulich wrote:
>>>>>> Under certain conditions CPUs can speculate into the instruction stream
>>>>>> past a RET instruction. Guard against this just like 3b7dab93f240
>>>>>> ("x86/spec-ctrl: Protect against CALL/JMP straight-line speculation")
>>>>>> did - by inserting an "INT $3" insn. It's merely the mechanics of how to
>>>>>> achieve this that differ: A set of macros gets introduced to post-
>>>>>> process RET insns issued by the compiler (or living in assembly files).
>>>>>>
>>>>>> Unfortunately for clang this requires further features their built-in
>>>>>> assembler doesn't support: We need to be able to override insn mnemonics
>>>>>> produced by the compiler (which may be impossible, if internally
>>>>>> assembly mnemonics never get generated), and we want to use \(text)
>>>>>> escaping / quoting in the auxiliary macro.
>>>>>
>>>>> Could this have an option to enable/disable at build time?
>>>>
>>>> Well, a subsequent patch adds a config option for this, which in
>>>> the worst case could be turned off. I'm afraid though I'm not
>>>> clear about the question, because ...
>>>>
>>>>> FreeBSD will drop GNU as quite soon from base, and albeit it can be
>>>>> installed as a package I would like to be able to build Xen using a
>>>>> toolchain based on LLVM exclusively.
>>>>
>>>> ... it's not clear to me what the implications here are: Are you
>>>> saying -no-integrated-as is not going to function anymore, unless
>>>> people explicitly install gas? If that's not what you meant to
>>>> indicate, then I don't see how building would become impossible.
>>>
>>> I'm still inquiring about this, but I would say that when gas is
>>> removed from FreeBSD then the 'as' command would be mapped to llvm-as,
>>> and thus -no-integrated-as would hit the same issues as the integrated
>>> as. So far in Xen we have assumed that -no-integrated-as would
>>> fallback to an as capable of doing what the integrated clang as
>>> doesn't support, but that might not be the case.
>>
>> At which point Xen couldn't be built anyway, I expect. If llvm-as
>> isn't sufficiently gas-compatible, we've lost (right now at least).
>>
>>> Ideally we would have to re-run the tests with -no-integrated-as, in
>>> order to assert that the external as is really capable of what the
>>> internal one is not.
>>
>> And if it doesn't, what would we do other than failing the build
>> (which it would also if we didn't do the 2nd round of checks)?
> 
> I would always prefer a clear message (ie: your toolstack is not
> capable of building Xen) rather than a weird build time failure.

Fair point in general.

> Also we could maybe disable certain options by default if the
> toolstack doesn't have the required support to build them?

We could, but I'm afraid this will go down the route of embedding
tool chain capabilities in xen/.config, which I continue to not
consider a good idea (and the thread got stalled, as expected).

In fact (also to Andrew and Anthony), recently I've become aware
of another shortcoming of this model: Our kernel packages contain
.config files for the various architectures and specific per-
architecture flavors. It used to be easy to update them on any
system, until the tool chain capability checks got introduced.
Now, in order to update them, one has to use the precise versions
of the various tool chain parts that will be used on the build
hosts, or else an error may result (for unexpected changes to
the file), or one may unknowingly turn off options that are
expected to be enabled.

Put more generally - if I handed someone a specific .config, I'd
expect their resulting binary to contain what I did set up. Or
for them to report back that they can't build the thing. But it
should not be the case that the .config got silently changed and
certain functionality disabled just because they use a different
tool chain.

> Has anyone reported this shortcoming to upstream llvm, so they are
> aware and can work on this or maybe provide an alternative way to
> achieve the same result?

I didn't and I'm unaware of anyone else possibly having done so.
That said, I consider it sort of obvious though that the goal of
replacing the GNU tool chain implies being fully compatible (and
presumably better in certain areas).

Jan