qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

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


From: Aravinda Prasad
Subject: Re: [Qemu-devel] [Qemu-ppc] [PATCH v10 2/6] ppc: spapr: Introduce FWNMI capability
Date: Mon, 8 Jul 2019 15:56:20 +0530
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.6.0


On Friday 05 July 2019 06:53 PM, Greg Kurz wrote:
> On Fri, 5 Jul 2019 16:49:17 +0530
> Aravinda Prasad <address@hidden> wrote:
> 
>>
>>
>> On Friday 05 July 2019 06:37 AM, David Gibson wrote:
>>> 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.
>>
>> I think we better not enable this by default until we enhance TCG to
>> support fwnmi.
>>
> 
> If we ever enhance TCG... until this get done, I concur with David's
> idea of just printing a warning. System emulation+TCG is more a CI
> or developer thing: we just want FWNMI not to break anything, even
> if it doesn't work. KVM is the real life scenario we want to support.
> If the feature is valuable, and I think it is, it should be the
> default otherwise fewer people will have a chance to take benefit
> from it.

ok.

> 
>>>
>>
> 

-- 
Regards,
Aravinda




reply via email to

[Prev in Thread] Current Thread [Next in Thread]