On Fri, May 13, 2022 at 8:21 AM Jon Kohler <jon@xxxxxxxxxxx> wrote: > > > > > On May 12, 2022, at 11:50 PM, Jim Mattson <jmattson@xxxxxxxxxx> wrote: > > > > On Thu, May 12, 2022 at 8:19 PM Jon Kohler <jon@xxxxxxxxxxx> wrote: > >> > >> > >> > >>> On May 12, 2022, at 11:06 PM, Jim Mattson <jmattson@xxxxxxxxxx> wrote: > >>> > >>> On Thu, May 12, 2022 at 5:50 PM Jon Kohler <jon@xxxxxxxxxxx> wrote: > >>> > >>>> You mentioned if someone was concerned about performance, are you > >>>> saying they also critically care about performance, such that they are > >>>> willing to *not* use IBPB at all, and instead just use taskset and hope > >>>> nothing ever gets scheduled on there, and then hope that the hypervisor > >>>> does the job for them? > >>> > >>> I am saying that IBPB is not the only viable mitigation for > >>> cross-process indirect branch steering. Proper scheduling can also > >>> solve the problem, without the overhead of IBPB. Say that you have two > >>> security domains: trusted and untrusted. If you have a two-socket > >>> system, and you always run trusted workloads on socket#0 and untrusted > >>> workloads on socket#1, IBPB is completely superfluous. However, if the > >>> hypervisor chooses to schedule a vCPU thread from virtual socket#0 > >>> after a vCPU thread from virtual socket#1 on the same logical > >>> processor, then it *must* execute an IBPB between those two vCPU > >>> threads. Otherwise, it has introduced a non-architectural > >>> vulnerability that the guest can't possibly be aware of. > >>> > >>> If you can't trust your OS to schedule tasks where you tell it to > >>> schedule them, can you really trust it to provide you with any kind of > >>> inter-process security? > >> > >> Fair enough, so going forward: > >> Should this be mandatory in all cases? How this whole effort came > >> was that a user could configure their KVM host with conditional > >> IBPB, but this particular mitigation is now always on no matter what. > >> > >> In our previous patch review threads, Sean and I mostly settled on making > >> this particular avenue active only when a user configures always_ibpb, such > >> that for cases like the one you describe (and others like it that come up in > >> the future) can be covered easily, but for cond_ibpb, we can document > >> that it doesn’t cover this case. > >> > >> Would that be acceptable here? > > > > That would make me unhappy. We use cond_ibpb, and I don't want to > > switch to always_ibpb, yet I do want this barrier. > > Ok gotcha, which I think is a good point for cloud providers, since the > workload(s) are especially opaque. > > How about this: I could work up a v5 patch here where this was at minimum > a system level knob (similar to other mitigation knobs) and documented > In more detail. That way folks who might want more control here have the > basic ability to do that without recompiling the kernel. Such a “knob” would > be on by default, such that there is no functional regression here. > > Would that be ok with you as a middle ground? That would be great. Module parameter or sysctl is fine with me. Thanks! > Thanks again, > Jon > > > > >>> > >>>> Would this be the expectation of just KVM? Or all hypervisors on the > >>>> market? > >>> > >>> Any hypervisor that doesn't do this is broken, but that won't keep it > >>> off the market. :-) > >> > >> Very true :) > >> >