Re: [PATCH v3 21/35] x86/bugs: Determine relevant vulnerabilities based on attack vector controls.

[Josh Poimboeuf]

On Tue, Feb 11, 2025 at 07:04:44PM +0000, Kaplan, David wrote:
> To explain my thinking a bit more, mitigate_cross_thread is intended
> to enable cross-thread mitigations for any vulnerabilities the
> hardware may have.  That does not necessarily require disabling SMT.
> The required cross-thread mitigation is defined by each vulnerability.
> 
> For many vulnerabilities (like MDS), mitigation requires disabling
> SMT.  mds_apply_mitigation() queries the status of the cross-thread
> attack vector and will disable SMT if needed.
> 
> For GDS, mitigating cross-thread attacks does not require disabling
> SMT, just enabling the mitigation in the MSR.
> 
> To be fair, it doesn't make much sense to disable all the attack
> vectors except mitigate_cross_thread, but for correctness it seemed
> like enabling the mitigation in this case was the right thing.
> 
> I don't really want to tie mitigate_cross_thread to SMT disable
> because of cases like this where there is a cross-thread attack
> mitigation that is different from disabling SMT.  You could also
> imagine bugs that might be even more limited, where perhaps they're
> only relevant for say user->kernel but also have a cross-thread
> component.

But that "cross-thread" thing doesn't even make sense as a vector.

Think about it this way.  For cross-thread attacks:

  - CPU thread A is the attacker.  It's running in either user or guest.

  - CPU thread B is the victim.  It's running in either kernel, user, or
    host.

So ALL cross-thread attacks have to include one of the following:

  - user->kernel
  - user->user
  - guest->host
  - guest->guest

So by definition, a cross-thread attack must also involve at least one
of those four main vectors.

So cross-thread can't be a standalone vector.  Rather, it's a dependent
vector or "sub-vector".

If a user wants to be protected from user->user, of course that includes
wanting to be protected from *cross-thread* user->user.

And if they *don't* care about user->user, why would they care about
*cross-thread* user->user?

What users *really* care about (and why there exists such a distinction
in the first place) is the functional/performance impact of disabling
SMT.

So a flag to allow the vectors to disable SMT makes more sense, e.g.,

  mitigate_disable_smt=on

And maybe also an additional flag which says "I've enabled core
scheduling or some other isolation scheme, don't worry about any of the
SMT-specific mitigations like STIBP":

  mitigate_smt_safe=on

But the standalone "cross-thread" vector doesn't fit at all.

-- 
Josh