Re: [PATCH v6 14/15] vfio: Document dual stage control

[Eric Auger]

Hi Yi,

On 8/17/20 9:00 AM, Liu, Yi L wrote:
> Hi Eric,
> 
>> From: Auger Eric <eric.auger@xxxxxxxxxx>
>> Sent: Sunday, August 16, 2020 7:52 PM
>>
>> Hi Yi,
>>
>> On 7/28/20 8:27 AM, Liu Yi L wrote:
>>> From: Eric Auger <eric.auger@xxxxxxxxxx>
>>>
>>> The VFIO API was enhanced to support nested stage control: a bunch of> new
>> ioctls and usage guideline.
>>>
>>> Let's document the process to follow to set up nested mode.
>>>
>>> Cc: Kevin Tian <kevin.tian@xxxxxxxxx>
>>> CC: Jacob Pan <jacob.jun.pan@xxxxxxxxxxxxxxx>
>>> Cc: Alex Williamson <alex.williamson@xxxxxxxxxx>
>>> Cc: Eric Auger <eric.auger@xxxxxxxxxx>
>>> Cc: Jean-Philippe Brucker <jean-philippe@xxxxxxxxxx>
>>> Cc: Joerg Roedel <joro@xxxxxxxxxx>
>>> Cc: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx>
>>> Reviewed-by: Stefan Hajnoczi <stefanha@xxxxxxxxxx>
>>> Signed-off-by: Eric Auger <eric.auger@xxxxxxxxxx>
>>> Signed-off-by: Liu Yi L <yi.l.liu@xxxxxxxxx>
>>> ---
>>> v5 -> v6:
>>> *) tweak per Eric's comments.
>>>
>>> v3 -> v4:
>>> *) add review-by from Stefan Hajnoczi
>>>
>>> v2 -> v3:
>>> *) address comments from Stefan Hajnoczi
>>>
>>> v1 -> v2:
>>> *) new in v2, compared with Eric's original version, pasid table bind
>>>    and fault reporting is removed as this series doesn't cover them.
>>>    Original version from Eric.
>>>    https://lkml.org/lkml/2020/3/20/700
>>> ---
>>>  Documentation/driver-api/vfio.rst | 75
>> +++++++++++++++++++++++++++++++++++++++
>>>  1 file changed, 75 insertions(+)
>>>
>>> diff --git a/Documentation/driver-api/vfio.rst b/Documentation/driver-api/vfio.rst
>>> index f1a4d3c..c0d43f0 100644
>>> --- a/Documentation/driver-api/vfio.rst
>>> +++ b/Documentation/driver-api/vfio.rst
>>> @@ -239,6 +239,81 @@ group and can access them as follows::
>>>  	/* Gratuitous device reset and go... */
>>>  	ioctl(device, VFIO_DEVICE_RESET);
>>>
>>> +IOMMU Dual Stage Control
>>> +------------------------
>>> +
>>> +Some IOMMUs support 2 stages/levels of translation. Stage corresponds
>>> +to the ARM terminology while level corresponds to Intel's terminology.
>>> +In the following text we use either without distinction.
>>> +
>>> +This is useful when the guest is exposed with a virtual IOMMU and some
>>> +devices are assigned to the guest through VFIO. Then the guest OS can
>>> +use stage-1 (GIOVA -> GPA or GVA->GPA), while the hypervisor uses stage
>>> +2 for VM isolation (GPA -> HPA).
>>> +
>>> +Under dual stage translation, the guest gets ownership of the stage-1 page
>>> +tables and also owns stage-1 configuration structures. The hypervisor owns
>>> +the root configuration structure (for security reason), including stage-2
>>> +configuration.
>> This is only true for vtd. On ARM the stage2 cfg is the Context
>> Descriptor table (aka PASID table). root cfg only store the GPA of the
>> CD table.
> 
> I've a check with you on the meaning of "configuration structures".
> For Vt-d, does it mean the root table/context table/pasid table? if
> I'm correct, then how about below description?
Yes I agree
> 
> "Under dual stage translation, the guest gets ownership of the stage-1
> configuration structures or page tables.
Actually on ARM the guest both owns the S1 configuration (CD table) and
S1 page tables ;-)

on Intel I understand the guest only owns the S1 page tables.

If confirmed, you may use such kind of explicit statement.

Thanks

Eric

 This depends on vendor. The
> hypervisor owns the root configuration structure (for security reason),
> including stage-2 configuration."
> 
>>  This works as long as configuration structures and page table
>>> +formats are compatible between the virtual IOMMU and the physical IOMMU.
>>> +
>>> +Assuming the HW supports it, this nested mode is selected by choosing the
>>> +VFIO_TYPE1_NESTING_IOMMU type through:
>>> +
>>> +    ioctl(container, VFIO_SET_IOMMU, VFIO_TYPE1_NESTING_IOMMU);
>>> +
>>> +This forces the hypervisor to use the stage-2, leaving stage-1 available
>>> +for guest usage. The stage-1 format and binding method are vendor specific
>> . There are reported in the nesting capability ...
> 
> got it.
> 
> "The stage-1 format and binding method are reported in nesting capability."
> 
>>> +and reported in nesting cap (VFIO_IOMMU_TYPE1_INFO_CAP_NESTING) through
>>> +VFIO_IOMMU_GET_INFO:
>>> +
>>> +    ioctl(container->fd, VFIO_IOMMU_GET_INFO, &nesting_info);
>>> +
>>> +The nesting cap info is available only after NESTING_IOMMU is selected.
>>> +If underlying IOMMU doesn't support nesting, VFIO_SET_IOMMU fails and
>> If the underlying
> 
> got it.
> 
>>> +userspace should try other IOMMU types. Details of the nesting cap info
>>> +can be found in Documentation/userspace-api/iommu.rst.
>>> +
>>> +The stage-1 page table can be bound to the IOMMU in two methods: directly>
>> +or indirectly. Direct binding requires userspace to notify VFIO of every
>> Not sure we shall use this direct/indirect terminology. I don't think
>> this is part of either ARM or Intel SPEC.
>>
>> Suggestion: On Intel, the stage1 page table info are mediated by the
>> userspace for each PASID. On ARM, the userspace directly passes the GPA
>> of the whole PASID table. Currently only Intel's binding is supported.
> 
> got it. this is what we want to say by ditect/indirect terminology.
> 
> Regards,
> Yi Liu
> 
>>> +guest stage-1 page table binding, while indirect binding allows userspace
>>> +to bind once with an intermediate structure (e.g. PASID table) which
>>> +indirectly links to guest stage-1 page tables. The actual binding method
>>> +depends on IOMMU vendor. Currently only the direct binding capability (
>>> +IOMMU_NESTING_FEAT_BIND_PGTBL) is supported:
>>> +
>>> +    nesting_op->flags = VFIO_IOMMU_NESTING_OP_BIND_PGTBL;
>>> +    memcpy(&nesting_op->data, &bind_data, sizeof(bind_data));
>>> +    ioctl(container->fd, VFIO_IOMMU_NESTING_OP, nesting_op);
>>> +
>>> +When multiple stage-1 page tables are supported on a device, each page
>>> +table is associated with a PASID (Process Address Space ID) to differentiate
>>> +with each other. In such case, userspace should include PASID in the
>>> +bind_data when issuing direct binding request.
>>> +
>>> +PASID could be managed per-device or system-wide which, again, depends on
>>> +IOMMU vendor and is reported in nesting cap info. When system-wide policy
>>> +is reported (IOMMU_NESTING_FEAT_SYSWIDE_PASID), e.g. as by Intel platforms,
>>> +userspace *must* allocate PASID from VFIO before attempting binding of
>>> +stage-1 page table:
>>> +
>>> +    req.flags = VFIO_IOMMU_ALLOC_PASID;
>>> +    ioctl(container, VFIO_IOMMU_PASID_REQUEST, &req);
>>> +
>>> +Once the stage-1 page table is bound to the IOMMU, the guest is allowed to
>>> +fully manage its mapping at its disposal. The IOMMU walks nested stage-1
>>> +and stage-2 page tables when serving DMA requests from assigned device, and
>>> +may cache the stage-1 mapping in the IOTLB. When required (IOMMU_NESTING_
>>> +FEAT_CACHE_INVLD), userspace *must* forward guest stage-1 invalidation to
>>> +the host, so the IOTLB is invalidated:
>>> +
>>> +    nesting_op->flags = VFIO_IOMMU_NESTING_OP_CACHE_INVLD;
>>> +    memcpy(&nesting_op->data, &cache_inv_data, sizeof(cache_inv_data));
>>> +    ioctl(container->fd, VFIO_IOMMU_NESTING_OP, nesting_op);
>>> +
>>> +Forwarded invalidations can happen at various granularity levels (page
>>> +level, context level, etc.)
>>> +
>>>  VFIO User API
>>>  -------------------------------------------------------------------------------
>>>
>>>
>> Thanks
>>
>> Eric
>