Re: [RFC PATCH v1 3/4] KVM: arm64: GICv4.1: Restore VLPI's pending state to physical side

[Shenming Lu <lushenming@huawei.com>]

On 2020/11/24 16:44, Marc Zyngier wrote:
> On 2020-11-24 08:10, Shenming Lu wrote:
>> On 2020/11/23 17:27, Marc Zyngier wrote:
>>> On 2020-11-23 06:54, Shenming Lu wrote:
>>>> From: Zenghui Yu <yuzenghui@huawei.com>
>>>>
>>>> When setting the forwarding path of a VLPI, it is more consistent to
>>>
>>> I'm not sure it is more consistent. It is a *new* behaviour, because it only
>>> matters for migration, which has been so far unsupported.
>>
>> Alright, consistent may not be accurate...
>> But I have doubt that whether there is really no need to transfer the
>> pending states
>> from kvm'vgic to VPT in set_forwarding regardless of migration, and the similar
>> for unset_forwarding.
> 
> If you have to transfer that state outside of the a save/restore, it means that
> you have missed the programming of the PCI endpoint. This is an established
> restriction that the MSI programming must occur *after* the translation has
> been established using MAPI/MAPTI (see the large comment at the beginning of
> vgic-v4.c).
> 
> If you want to revisit this, fair enough. But you will need a lot more than
> just opportunistically transfer the pending state.

Thanks, I will look at what you mentioned.

> 
>>
>>>
>>>> also transfer the pending state from irq->pending_latch to VPT (especially
>>>> in migration, the pending states of VLPIs are restored into kvm’s vgic
>>>> first). And we currently send "INT+VSYNC" to trigger a VLPI to pending.
>>>>
>>>> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
>>>> Signed-off-by: Shenming Lu <lushenming@huawei.com>
>>>> ---
>>>>  arch/arm64/kvm/vgic/vgic-v4.c | 12 ++++++++++++
>>>>  1 file changed, 12 insertions(+)
>>>>
>>>> diff --git a/arch/arm64/kvm/vgic/vgic-v4.c b/arch/arm64/kvm/vgic/vgic-v4.c
>>>> index b5fa73c9fd35..cc3ab9cea182 100644
>>>> --- a/arch/arm64/kvm/vgic/vgic-v4.c
>>>> +++ b/arch/arm64/kvm/vgic/vgic-v4.c
>>>> @@ -418,6 +418,18 @@ int kvm_vgic_v4_set_forwarding(struct kvm *kvm, int virq,
>>>>      irq->host_irq    = virq;
>>>>      atomic_inc(&map.vpe->vlpi_count);
>>>>
>>>> +    /* Transfer pending state */
>>>> +    ret = irq_set_irqchip_state(irq->host_irq,
>>>> +                    IRQCHIP_STATE_PENDING,
>>>> +                    irq->pending_latch);
>>>> +    WARN_RATELIMIT(ret, "IRQ %d", irq->host_irq);
>>>> +
>>>> +    /*
>>>> +     * Let it be pruned from ap_list later and don't bother
>>>> +     * the List Register.
>>>> +     */
>>>> +    irq->pending_latch = false;
>>>
>>> It occurs to me that calling into irq_set_irqchip_state() for a large
>>> number of interrupts can take a significant amount of time. It is also
>>> odd that you dump the VPT with the VPE unmapped, but rely on the VPE
>>> being mapped for the opposite operation.
>>>
>>> Shouldn't these be symmetric, all performed while the VPE is unmapped?
>>> It would also save a lot of ITS traffic.
>>>
>>
>> My thought was to use the existing interface directly without unmapping...
>>
>> If you want to unmap the vPE and poke the VPT here, as I said in the cover
>> letter, set/unset_forwarding might also be called when all devices are running
>> at normal run time, in which case the unmapping of the vPE is not allowed...
> 
> No, I'm suggesting that you don't do anything here, but instead as a by-product
> of restoring the ITS tables. What goes wrong if you use the
> KVM_DEV_ARM_ITS_RESTORE_TABLE backend instead?

There is an issue if we do it in the restoring of the ITS tables: the transferring
of the pending state needs the irq to be marked as hw before, which is done by the
pass-through device, but the configuring of the forwarding path of the VLPI depends
on the restoring of the vgic first... It is a circular dependency.

> 
>> Another possible solution is to add a new dedicated interface to QEMU
>> to transfer
>> these pending states to HW in GIC VM state change handler corresponding to
>> save_pending_tables?
> 
> Userspace has no way to know we use GICv4, and I intend to keep it
> completely out of the loop. The API is already pretty tortuous, and
> I really don't want to add any extra complexity to it.
> 
> Thanks,
> 
>         M.