Re: [PATCH] IOMMU: make domctl handler tolerate NULL domain

[Jan Beulich <jbeulich@xxxxxxxx>]

On 19.04.2022 17:39, Andrew Cooper wrote:
> On 19/04/2022 11:59, Jan Beulich wrote:
>> On 19.04.2022 12:49, Andrew Cooper wrote:
>>> On 19/04/2022 10:39, Jan Beulich wrote:
>>> Furthermore, under what circumstances is test_assign_device legitimate
>>> when passing DOMID_INVALID ?  This has been broken for 3 years now
>>> without report, so it's clearly an unused codepath under both xl's and
>>> xapi's idea of passthrough.
>> I guess xend had a way to drive the domctl this way.
> 
> Looking at the xend code, it always called with domid 0.
> 
> I can't see any evidence that there has actually been a caller passing
> DOMID_INVALID, but this is an utter mess.
> 
>>  Iirc this was
>> to find out whether a device is assignable at all, without needing
>> to know of any particular valid domain.
> 
> In a correctly architected IOMMU subsystem (which Xen most definitely
> does not have at this juncture), that question is "Does the device have
> an upstream IOMMU".
> 
> Xen doesn't currently have a working idea of an x86 system with IOMMUs
> but not covering all devices.  While such a system is unlikely to exist
> in reality, there are cases where quirks create asymmetric coverage. 
> Either way, this is fully subsumed by the future work to assign IOMMU
> groups.
> 
> Also at the moment, because Xen doesn't support per-device IOMMU
> contexts, another check not performed is whether this devices identity
> maps are compatible with the identity maps in the target domain.  Extra
> complexity here is that it will not occur on a single system
> (Conflicting RMRRs/IVHDs on a single system is definitely a firmware
> bug, not an accurate description of the system); only with migration
> between systems where equivalent logical devices have differing identity
> requirements on different systems.
> 
> 
> I don't see anything useful the "with a domid" version gets you.

Hence I guess someone thought allowing DOMID_INVALID there would be a
good idea, irrespective of whether xend actually did things this way.

Jan