Re: [PATCH 3/3] KVM SVM: Add Bus Lock Detect support

[Sean Christopherson <seanjc@google.com>]

On Tue, Jul 09, 2024, Jim Mattson wrote:
> On Tue, Jul 9, 2024 at 7:25 PM Ravi Bangoria <ravi.bangoria@amd.com> wrote:
> >
> > Sean,
> >
> > Apologies for the delay. I was waiting for Bus Lock Threshold patches to be
> > posted upstream:
> >
> >   https://lore.kernel.org/r/20240709175145.9986-1-manali.shukla@amd.com
> >
> > On 12-Jun-24 7:12 AM, Sean Christopherson wrote:
> > > On Wed, Jun 05, 2024, Ravi Bangoria wrote:
> > >> On 6/5/2024 8:38 PM, Sean Christopherson wrote:
> > >>> Some of the problems on Intel were due to the awful FMS-based feature detection,
> > >>> but those weren't the only hiccups.  E.g. IIRC, we never sorted out what should
> > >>> happen if both the host and guest want bus-lock #DBs.
> > >>
> > >> I've to check about vcpu->guest_debug part, but keeping that aside, host and
> > >> guest can use Bus Lock Detect in parallel because, DEBUG_CTL MSR and DR6
> > >> register are save/restored in VMCB, hardware cause a VMEXIT_EXCEPTION_1 for
> > >> guest #DB(when intercepted) and hardware raises #DB on host when it's for the
> > >> host.
> > >
> > > I'm talking about the case where the host wants to do something in response to
> > > bus locks that occurred in the guest.  E.g. if the host is taking punitive action,
> > > say by stalling the vCPU, then the guest kernel could bypass that behavior by
> > > enabling bus lock detect itself.
> > >
> > > Maybe it's moot point in practice, since it sounds like Bus Lock Threshold will
> > > be available at the same time.
> > >
> > > Ugh, and if we wanted to let the host handle guest-induced #DBs, we'd need code
> > > to keep Bus Lock Detect enabled in the guest since it resides in DEBUG_CTL.  Bah.
> > >
> > > So I guess if the vcpu->guest_debug part is fairly straightforward, it probably
> > > makes to virtualize Bus Lock Detect because the only reason not to virtualize it
> > > would actually require more work/code in KVM.
> >
> > KVM forwards #DB to Qemu when vcpu->guest_debug is set and it's Qemu's
> > responsibility to re-inject exception when Bus Lock Trap is enabled
> > inside the guest. I realized that it is broken so I've prepared a
> > Qemu patch, embedding it at the end.
> >
> > > I'd still love to see Bus Lock Threshold support sooner than later though :-)
> >
> > With Bus Lock Threshold enabled, I assume the changes introduced by this
> > patch plus Qemu fix are sufficient to support Bus Lock Trap inside the
> > guest?
> 
> In any case, it seems that commit 76ea438b4afc ("KVM: X86: Expose bus
> lock debug exception to guest") prematurely advertised the presence of
> X86_FEATURE_BUS_LOCK to userspace on non-Intel platforms. We should
> probably either accept these changes or fix up that commit. Either
> way, something should be done for all active branches back to v5.15.

Drat.  Yeah, we need a patch to clear BUS_LOCK_DETECT in svm_set_cpu_caps(), marked
for stable@.  Then this series can remove that clearing.

At least I caught it for CET[*]!  It'd be nice to not have to rely on humans to
detect potential issues like this, but I can't think of a way to programmatically
handle this situation without incurring an annoying amount of overhead and/or
duplicate code between VMX and SVM.

[*] https://lore.kernel.org/all/ZjLRnisdUgeYgg8i@google.com