Re: [PATCH v4 15/32] vfio: introduce KVM-owned IOMMU type

[Matthew Rosato <mjrosato@linux.ibm.com>]

On 3/14/22 5:38 PM, Jason Gunthorpe wrote:
> On Mon, Mar 14, 2022 at 03:44:34PM -0400, Matthew Rosato wrote:
> 
>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
>> index 9394aa9444c1..0bec97077d61 100644
>> +++ b/drivers/vfio/vfio_iommu_type1.c
>> @@ -77,6 +77,7 @@ struct vfio_iommu {
>>   	bool			nesting;
>>   	bool			dirty_page_tracking;
>>   	bool			container_open;
>> +	bool			kvm;
>>   	struct list_head	emulated_iommu_groups;
>>   };
>>   
>> @@ -2203,7 +2204,12 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
>>   		goto out_free_group;
>>   
>>   	ret = -EIO;
>> -	domain->domain = iommu_domain_alloc(bus);
>> +
>> +	if (iommu->kvm)
>> +		domain->domain = iommu_domain_alloc_type(bus, IOMMU_DOMAIN_KVM);
>> +	else
>> +		domain->domain = iommu_domain_alloc(bus);
>> +
>>   	if (!domain->domain)
>>   		goto out_free_domain;
>>   
>> @@ -2552,6 +2558,9 @@ static void *vfio_iommu_type1_open(unsigned long arg)
>>   	case VFIO_TYPE1v2_IOMMU:
>>   		iommu->v2 = true;
>>   		break;
>> +	case VFIO_KVM_IOMMU:
>> +		iommu->kvm = true;
>> +		break;
> 
> Same remark for this - but more - this is called KVM but it doesn't
> accept a kvm FD or any thing else to link the domain to the KVM
> in-use.

Right...  The name is poor, but with the current design the KVM 
association comes shortly after.  To summarize, with this series, the 
following relevant steps occur:

1) VFIO_SET_IOMMU: Indicate we wish to use the alternate IOMMU domain
	-> At this point, the IOMMU will reject any maps (no KVM, no guest 
table anchor)
2) KVM ioctl "start":
	-> Register the KVM with the IOMMU domain
	-> At this point, IOMMU will still reject any maps (no guest table anchor)
3) KVM ioctl "register ioat"
	-> Register the guest DMA table head with the IOMMU domain
	-> now IOMMU maps are allowed

The rationale for splitting steps 1 and 2 are that VFIO_SET_IOMMU 
doesn't have a mechanism for specifying more than the type as an arg, 
no?  Otherwise yes, you could specify a kvm fd at this point and it 
would have some other advantages (e.g. skip notifier).  But we still 
can't use the IOMMU for mapping until step 3.

The rationale for splitting steps 2 and 3 are twofold:  1) during init, 
we simply don't know where the guest anchor will be when we allocate the 
domain and 2) because the guest can technically clear and re-initialize 
their DMA space during the life of the guest, moving the location of the 
table anchor.  We would receive another ioctl operation to unregister 
the guest table anchor and again reject any map operation until a new 
table location is provided.