Re: [Xen-devel] [PATCH v3 59/62] xen/arm: Add a hypercall for device mmio mapping

[Shannon Zhao <zhaoshenglong@xxxxxxxxxx>]

On 2016/1/8 5:40, Daniel De Graaf wrote:
> On 01/07/2016 05:50 AM, Jan Beulich wrote:
>>>>> On 07.01.16 at 10:11, <zhaoshenglong@xxxxxxxxxx> wrote:
>>> Hi Jan,
>>>
>>> On 2016/1/7 15:45, Jan Beulich wrote:
>>>>>>> On 07.01.16 at 07:58, <zhaoshenglong@xxxxxxxxxx> wrote:
>>>>>> On 2015/11/17 19:04, Jan Beulich wrote:
>>>>>>>>>>>>>> On 17.11.15 at 10:40, <shannon.zhao@xxxxxxxxxx> wrote:
>>>>>>>>>>>> --- a/xen/arch/arm/mm.c
>>>>>>>>>>>> +++ b/xen/arch/arm/mm.c
>>>>>>>>>>>> @@ -1138,6 +1138,10 @@ int xenmem_add_to_physmap_one(
>>>>>>>>>>>>           rcu_unlock_domain(od);
>>>>>>>>>>>>           break;
>>>>>>>>>>>>       }
>>>>>>>>>>>> +    case XENMAPSPACE_dev_mmio:
>>>>>>>>>>>> +        rc = map_dev_mmio_region(d, gpfn, 1, idx);
>>>>>>>>>>>> +        return rc;
>>>>>>>>>>>> +        break;
>>>>>>>> Blindly for any kind of domain? The XSM check in the
>>>>>>>> XENMEM_add_to_physmap_batch handler (in common code) doesn't
>>>>>>>> even know which map space is to be used...
>>>>>>
>>>>>> Sorry, I know little about XSM. Could you suggest me how to add the
>>>>>> check for this new type here?
>>>> I'm sorry to push back here, but did you at least try to derive
>>>> what is wanted from the multitude of other XSM checks present
>>>> throughout the tree?
>>>
>>> IIUC, you mean that it doean't need to change the XSM check itself, but
>>> we should check if the current->domain is hardware domain and it maps
>>> the space to itself before the XSM check, right?
>>
>> No, I actually think that you need to add a new, secondary XSM
>> check. But you may want to consult with Daniel (who so far wasn't
>> even Cc-ed).
> 
> Looking at the original patch, I am not sure if I understand the
> checks: it seems like the iomem_access_permitted check is being done
> on the guest's page range instead of the actual IO memory, which
> ends up allowing the guest to map anything as long as it maps it in
> the right guest area. 
Yeah, since it's hard to know the MMIO range from the DSDT in XEN, we
permit full mmio capabilities for Dom0 and deny mmio access for some
devices e.g. uart. Then when Dom0 add those devices, call
XENMEM_add_to_physmap to map their MMIO ranges. This looks similar with
what x86 does.

    /* The hardware domain is initially permitted full I/O capabilities. */
    rc |= ioports_permit_access(d, 0, 0xFFFF);
    rc |= iomem_permit_access(d, 0UL, ~0UL);
    rc |= irqs_permit_access(d, 1, nr_irqs_gsi - 1);

> The iomem_permit_access call there also seems
> to be redundant because it is the same range that was just checked.
>
Ah, I'll drop this at next version.

> If the [start_gfn, start_gfn + nr) memory range actually describes
> the physical addresses, then this operation is taking advantage of
> the existing XSM checks on XEN_DOMCTL_iomem_permission, and the
> only XSM check that is needed would be that current->domain has
> permission to modify (d)'s mappings - and this is done by the
> xsm_add_to_physmap check in XENMEM_add_to_physmap.

Thanks,
-- 
Shannon


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel