On Fri, Oct 21, 2022, Vitaly Kuznetsov wrote:
> Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> writes:
>
> > Sean Christopherson <seanjc@xxxxxxxxxx> writes:
> >> Mostly because it's somewhat arbitrary that earlier code ensures valid_bank_mask
> >> is set in the all_cpus=true case, e.g. arguably KVM doesn't need to do the var_cnt
> >> sanity check in the all_cpus case:
> >>
> >> all_cpus = send_ipi_ex.vp_set.format == HV_GENERIC_SET_ALL;
> >> if (all_cpus)
> >> goto check_and_send_ipi;
> >>
> >> valid_bank_mask = send_ipi_ex.vp_set.valid_bank_mask;
> >> if (hc->var_cnt != hweight64(valid_bank_mask))
> >> return HV_STATUS_INVALID_HYPERCALL_INPUT;
> >>
> >> if (!hc->var_cnt)
> >> goto ret_success;
> >>
> >
> > I think 'var_cnt' (== hweight64(valid_bank_mask)) has to be checked in
> > 'all_cpus' case, especially in kvm_hv_flush_tlb(): the code which reads
> > TLB flush entries will read them from the wrong offset (data_offset/
> > consumed_xmm_halves) otherwise. The problem is less severe in
> > kvm_hv_send_ipi() as there's no data after CPU banks.
> >
> > At the bare minimum, "KVM: x86: hyper-v: Handle
> > HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST{,EX} calls gently" patch from this
> > series will have to be adjusted. I *think* mandating var_cnt==0 in 'all_cpus'
> > is OK but I don't recall such requirement from TLFS, maybe it's safer to
> > just adjust 'data_offset'/'consumed_xmm_halves' even in 'all_cpus' case.
> >
> > Let me do some tests...
>
> "We can neither confirm nor deny the existence of the problem". Windows
> guests seem to be smart enough to avoid using *_EX hypercalls altogether
> for "all cpus" case (as non-ex versions are good enough). Let's keep
> allowing non-zero var_cnt for 'all cpus' case for now
Sounds good.
> and think about hardening it later...
Eh, no need to add more work for ourselves. I wasn't thinking about hardening so
much as slightly simplifying KVM code.
