Re: [PATCH v10 4/7] KVM: arm64: Enable ring-based dirty memory tracking

[Gavin Shan <gshan@redhat.com>]

Hi Zenghui,

On 1/15/23 10:20 PM, Zenghui Yu wrote:
> On 2022/11/10 18:49, Gavin Shan wrote:
>> Enable ring-based dirty memory tracking on ARM64:
>>
>>   - Enable CONFIG_HAVE_KVM_DIRTY_RING_ACQ_REL.
>>
>>   - Enable CONFIG_NEED_KVM_DIRTY_RING_WITH_BITMAP.
>>
>>   - Set KVM_DIRTY_LOG_PAGE_OFFSET for the ring buffer's physical page
>>     offset.
>>
>>   - Add ARM64 specific kvm_arch_allow_write_without_running_vcpu() to
>>     keep the site of saving vgic/its tables out of the no-running-vcpu
>>     radar.
> 
> And we have KVM_DEV_ARM_VGIC_SAVE_PENDING_TABLES.. On receiving it, the
> emulated VGIC will write all pending bits (if any) into pending tables
> (which reside in guest memory) and doesn't require a running vcpu
> context.
> 
> The no-running-vcpu WARN can be triggered with the
> kvm-unit-tests/its-pending-migration case. I run it using QEMU, which
> has nothing to do with the dirty ring atm.
> 
> Or are there already discussions about it that I haven't noticed?
> 
> |void mark_page_dirty_in_slot(struct kvm *kvm,
> |                 const struct kvm_memory_slot *memslot,
> |                 gfn_t gfn)
> |{
> |    WARN_ON_ONCE(!vcpu && !kvm_arch_allow_write_without_running_vcpu(kvm));
> 

It's a new case we never noticed. Could you please share the QEMU command lines
to start the guest? I need to reproduce the issue on my side firstly.

The fix would be simply to extending kvm->arch.vgic.save_its_tables_in_progress
from 'bool' to a bit map (e.g. kvm->arch.vgic.dirty_guest_memory_flags) and introduce
two separate flags for ITS table and VGIC3 pending bits separately. Alternatively,
we can also introduce another 'bool kvm->arch.vgic.save_vgic_v3_tables_in_progress'
to cover the new case.


Thanks,
Gavin