> On 7 Nov 2019, at 16:00, Christophe de Dinechin <christophe.de.dinechin@xxxxxxxxx> wrote: > > > >> On 6 Nov 2019, at 00:25, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote: >> >> On Tue, Nov 05, 2019 at 11:37:50AM -0800, Sean Christopherson wrote: >>> On Tue, Nov 05, 2019 at 05:17:37PM +0100, Vitaly Kuznetsov wrote: >>>> Virtualized guests may pick a different strategy to mitigate hardware >>>> vulnerabilities when it comes to hyper-threading: disable SMT completely, >>>> use core scheduling, or, for example, opt in for STIBP. Making the >>>> decision, however, requires an extra bit of information which is currently >>>> missing: does the topology the guest see match hardware or if it is 'fake' >>>> and two vCPUs which look like different cores from guest's perspective can >>>> actually be scheduled on the same physical core. Disabling SMT or doing >>>> core scheduling only makes sense when the topology is trustworthy. >>>> >>>> Add two feature bits to KVM: KVM_FEATURE_TRUSTWORTHY_SMT with the meaning >>>> that KVM_HINTS_TRUSTWORTHY_SMT bit answers the question if the exposed SMT >>>> topology is actually trustworthy. It would, of course, be possible to get >>>> away with a single bit (e.g. 'KVM_FEATURE_FAKE_SMT') and not lose backwards >>>> compatibility but the current approach looks more straightforward. >>> >>> I'd stay away from "trustworthy", especially if this is controlled by >>> userspace. Whether or not the hint is trustworthy is purely up to the >>> guest. Right now it doesn't really matter, but that will change as we >>> start moving pieces of the host out of the guest's TCB. >>> >>> It may make sense to split the two (or even three?) cases, e.g. >>> KVM_FEATURE_NO_SMT and KVM_FEATURE_ACCURATE_TOPOLOGY. KVM can easily >>> enforce NO_SMT _today_, i.e. allow it to be set if and only if SMT is >>> truly disabled. Verifying that the topology exposed to the guest is legit >>> is a completely different beast. >> >> Scratch the ACCURATE_TOPOLOGY idea, I doubt there's a real use case for >> setting ACCURATE_TOPOLOGY and not KVM_HINTS_REALTIME. A feature flag to >> state that SMT is disabled seems simple and useful. A bit such as NoNonArchitecturalCoreSharing can be set even when host SMT is enabled. For example, when host use core-scheduling to group together vCPUs that run as sibling hyperthreads. Therefore, I wouldn’t want to tie the feature-flag semantics to host SMT being enabled/disabled. It’s just true that this bit can be set when host SMT is disabled. > > I share that concern about the naming, although I do see some > value in exposing the cpu_smt_possible() result. I think it’s easier > to state that something does not work than to state something does > work. > > Also, with respect to mitigation, we may want to split the two cases > that Paolo outlined, i.e. have KVM_HINTS_REALTIME, > KVM_HINTS_CORES_CROSSTALK and > KVM_HINTS_CORES_LEAKING, > where CORES_CROSSTALKS indicates there may be some > cross-talk between what the guest thinks are isolated cores, > and CORES_LEAKING indicates that cores may leak data > to some other guest. > > The problem with my approach is that it is shouting “don’t trust me” > a bit too loudly. I don’t see a value in exposing CORES_LEAKING to guest. As guest have nothing to do with it. -Liran
