Re: Xen-Error: Disabling IOMMU on Stepping C2 5520 Host-Bridge

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 25/07/2021 14:55, Marek Marczykowski-Górecki wrote:
> On Sun, Jul 25, 2021 at 02:31:17PM +0200, luja wrote:
>> This Z600 is equipped with 0B54h mainboard as can be seen with dmi-decode.
>>
>> The manual states that 0B54h mainboard has the "newer C2 stepping",
>> so it is *not* affected by Intel "spec update" (nota bene: Intel updates the
>> spec, others report erratas) bugs  
> The code above checks for rev 0x13, and the spec (page 11) clearly says that 
> rev
> 0x13 is stepping B-3. Stepping C-2 is rev 0x22. So, if this check
> triggers for you, I'm afraid you have the affected chipset.

The ID in hardware is the authoritative information.  Sounds like the
Z600 manual is wrong.

>> So the way Xen detects the "bug" (pci rev 13) is not sufficient, as my Z600
>> shows pci rev13 with lspci but 0xB54h (board rev only on Z600) with dmidecode
>> I would suggest first to have an override xen kernel boot option to disable 
>> the disablement in this code section. Or just patch this part out of the Xen 
>> code and rebuild xen. If this stuff really crashes, one will see it.
> Patching it out is out of the question, this check if there for a
> reason.

Using interrupt remapping on these systems does cause it to cease
functioning.

>> So please build a new xen without this stupid disablement or please add an 
>> override boot command for it.
>>
>> Please see the attached upgrade manual of Z600 and the errata "spec update" 
>> by Intel.
>> You see that the C2 stepping is not affected by the bugs refered to in the 
>> xen code,
>> so removing that section or adding better detection of the mask revision (B3 
>> vs. C2)  of 5520 host bridge would allow  many users to operate Qubes4.
> Maybe someone else has an alternative idea?

The logic in Xen is broken.  I've tried fixing it before for XenServer,
but was objected to, and the patch is still in the patchqueue.

The errata is with the Queued Invalidation, which (in Xen) is tied to
interrupt remapping.  The rest of the IOMMU works fine.

The current status quo is that if Xen boots with an Intel gen1 IOMMU, it
will be happy with DMA remapping but no IRQ remapping.  If Xen boots on
this specific buggy system, it will turn the entire IOMMU off in
protest, which leaves the system less secure than booting on the
previous generation of hardware.

The correct behaviour is to just disable interrupt remapping in this
case, which brings Xen's behaviour in line with adjacent generations of
hardware.

~Andrew