Re: [Qemu-devel] [PATCH v10 2/6] ppc: spapr: Introduce FWNMI capability

[David Gibson]

On Thu, Jul 04, 2019 at 10:33:11AM +0530, Aravinda Prasad wrote:
> 
> 
> On Thursday 04 July 2019 06:37 AM, David Gibson wrote:
> > On Wed, Jul 03, 2019 at 02:58:24PM +0530, Aravinda Prasad wrote:
> >>
> >>
> >> On Wednesday 03 July 2019 08:33 AM, David Gibson wrote:
> >>> On Tue, Jul 02, 2019 at 11:54:26AM +0530, Aravinda Prasad wrote:
> >>>>
> >>>>
> >>>> On Tuesday 02 July 2019 09:21 AM, David Gibson wrote:
> >>>>> On Wed, Jun 12, 2019 at 02:51:04PM +0530, Aravinda Prasad wrote:
> >>>>>> Introduce the KVM capability KVM_CAP_PPC_FWNMI so that
> >>>>>> the KVM causes guest exit with NMI as exit reason
> >>>>>> when it encounters a machine check exception on the
> >>>>>> address belonging to a guest. Without this capability
> >>>>>> enabled, KVM redirects machine check exceptions to
> >>>>>> guest's 0x200 vector.
> >>>>>>
> >>>>>> This patch also introduces fwnmi-mce capability to
> >>>>>> deal with the case when a guest with the
> >>>>>> KVM_CAP_PPC_FWNMI capability enabled is attempted
> >>>>>> to migrate to a host that does not support this
> >>>>>> capability.
> >>>>>>
> >>>>>> Signed-off-by: Aravinda Prasad <address@hidden>
> >>>>>> ---
> >>>>>>  hw/ppc/spapr.c         |    1 +
> >>>>>>  hw/ppc/spapr_caps.c    |   26 ++++++++++++++++++++++++++
> >>>>>>  include/hw/ppc/spapr.h |    4 +++-
> >>>>>>  target/ppc/kvm.c       |   19 +++++++++++++++++++
> >>>>>>  target/ppc/kvm_ppc.h   |   12 ++++++++++++
> >>>>>>  5 files changed, 61 insertions(+), 1 deletion(-)
> >>>>>>
> >>>>>> diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
> >>>>>> index 6dd8aaa..2ef86aa 100644
> >>>>>> --- a/hw/ppc/spapr.c
> >>>>>> +++ b/hw/ppc/spapr.c
> >>>>>> @@ -4360,6 +4360,7 @@ static void spapr_machine_class_init(ObjectClass 
> >>>>>> *oc, void *data)
> >>>>>>      smc->default_caps.caps[SPAPR_CAP_NESTED_KVM_HV] = SPAPR_CAP_OFF;
> >>>>>>      smc->default_caps.caps[SPAPR_CAP_LARGE_DECREMENTER] = 
> >>>>>> SPAPR_CAP_ON;
> >>>>>>      smc->default_caps.caps[SPAPR_CAP_CCF_ASSIST] = SPAPR_CAP_OFF;
> >>>>>> +    smc->default_caps.caps[SPAPR_CAP_FWNMI_MCE] = SPAPR_CAP_OFF;
> >>>>>>      spapr_caps_add_properties(smc, &error_abort);
> >>>>>>      smc->irq = &spapr_irq_dual;
> >>>>>>      smc->dr_phb_enabled = true;
> >>>>>> diff --git a/hw/ppc/spapr_caps.c b/hw/ppc/spapr_caps.c
> >>>>>> index 31b4661..2e92eb6 100644
> >>>>>> --- a/hw/ppc/spapr_caps.c
> >>>>>> +++ b/hw/ppc/spapr_caps.c
> >>>>>> @@ -479,6 +479,22 @@ static void 
> >>>>>> cap_ccf_assist_apply(SpaprMachineState *spapr, uint8_t val,
> >>>>>>      }
> >>>>>>  }
> >>>>>>  
> >>>>>> +static void cap_fwnmi_mce_apply(SpaprMachineState *spapr, uint8_t val,
> >>>>>> +                                Error **errp)
> >>>>>> +{
> >>>>>> +    if (!val) {
> >>>>>> +        return; /* Disabled by default */
> >>>>>> +    }
> >>>>>> +
> >>>>>> +    if (tcg_enabled()) {
> >>>>>> +        error_setg(errp,
> >>>>>> +"No Firmware Assisted Non-Maskable Interrupts support in TCG, try 
> >>>>>> cap-fwnmi-mce=off");
> >>>>>
> >>>>> Not allowing this for TCG creates an awkward incompatibility between
> >>>>> KVM and TCG guests.  I can't actually see any reason to ban it for TCG
> >>>>> - with the current code TCG won't ever generate NMIs, but I don't see
> >>>>> that anything will actually break.
> >>>>>
> >>>>> In fact, we do have an nmi monitor command, currently wired to the
> >>>>> spapr_nmi() function which resets each cpu, but it probably makes
> >>>>> sense to wire it up to the fwnmi stuff when present.
> >>>>
> >>>> Yes, but that nmi support is not enough to inject a synchronous error
> >>>> into the guest kernel. For example, we should provide the faulty address
> >>>> along with other information such as the type of error (slb multi-hit,
> >>>> memory error, TLB multi-hit) and when the error occurred (load/store)
> >>>> and whether the error was completely recovered or not. Without such
> >>>> information we cannot build the error log and pass it on to the guest
> >>>> kernel. Right now nmi monitor command takes cpu number as the only 
> >>>> argument.
> >>>
> >>> Obviously we can't inject an arbitrary MCE event with that monitor
> >>> command.  But isn't there some sort of catch-all / unknown type of MCE
> >>> event which we could inject?
> >>
> >> We have "unknown" type of error, but we should also pass an address in
> >> the MCE event log. Strictly speaking this address should be a valid
> >> address in the current CPU context as MCEs are synchronous errors
> >> triggered when we touch a bad address.
> > 
> > Well, some of them are.  At least historically both synchronous and
> > asnchronous MCEs were possible.  Are there really no versions where
> > you can report an MCE with unknown address?
> 
> I am not aware of any such versions. Will cross check.
> 
> > 
> >> We can pass a default address with every nmi, but I am not sure whether
> >> that will be practically helpful.
> >>
> >>> It seems very confusing to me to have 2 totally separate "nmi"
> >>> mechanisms.
> >>>
> >>>> So I think TCG support should be a separate patch by itself.
> >>>
> >>> Even if we don't wire up the monitor command, I still don't see
> >>> anything that this patch breaks - we can support the nmi-register and
> >>> nmi-interlock calls without ever actually creating MCE events.
> >>
> >> If we support nmi-register and nmi-interlock calls without the monitor
> >> command wire-up then we will be falsely claiming the nmi support to the
> >> guest while it is not actually supported.
> > 
> > How so?  AFAICT, from the point of view of the guest this is not
> > observably different from supporting the NMI mechanism but NMIs never
> > occurring.
> 
> A guest inserting a duplicate SLB will expect the machine check
> exception delivered to the handler registered via nmi,register.
> But we actually don't do that in TCG.

Ah, true, I was thinking of external hardware fault triggered MCEs
rather than software error ones like duplicate SLB.

That said, I strongly suspect TCG is buggy enough at present that
exact behaviour in rare error conditions like duplicate SLB is not
really a big problem in the scheme of things.

I really don't think we can enable this by default until we allow it
for TCG - we don't want starting a TCG guest to involve manually
switching other options.

We could consider allowing it for TCG but just printing a warning that
the behaviour may not be correct in some conditions - we do something
similar for some of the Spectre workarounds already.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature