Re: [Xen-devel] [PATCH] evtchn: make support for different ABIs tunable

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 07/08/2019 16:08, Jan Beulich wrote:
> On 07.08.2019 17:00, Andrew Cooper wrote:
>> On 07/08/2019 15:30, Jan Beulich wrote:
>>> On 07.08.2019 15:41, Andrew Cooper wrote:
>>>> Furthermore, if there is this problem for event channels, then
>>>> there is
>>>> almost certainly a related problem for grant tables.
>>>>
>>>> The control in Xen should be expressed in a positive form, or the
>>>> logic
>>>> will become a tangle.  It should be a bit permitting the use of the
>>>> FIFO
>>>> ABI, rather than a bit saying "oh actually, you can't use that".
>>>>
>>>> That said, it might be easier to declare FIFO to be "event channel
>>>> v2",
>>>> and specify max_{grant,evntchn}_ver instead.
>>>
>>> I'm not sure assuming linear (or actually any) ordering between
>>> variants is a good thing. Yes, right now we only have gnttab
>>> v1 and v2 and evtchn 2l and fifo, which could be considered v1
>>> and v2 as you suggest. However, assuming a 3rd variant surfaces,
>>> why would it be that one has to expose v2 just to make v3
>>> usable? In particular gnttab v2 has various issues (which is why
>>> you introduced a way to disable its use in the first place), yet
>>> I'd hope we'd end up with a less quirky v3 if one ever becomes
>>> necessary. And in turn I'd hope we could hide v2 from any v3
>>> users.
>>>
>>> IOW I think a bitmap to permit use of "advanced" versions is
>>> more future proof. (As a side note, I don't think we want to
>>> introduce a disable for the respective v1 interfaces.)
>>
>> We absolutely do want a way to turn everything off.
>>
>> The inability to turn the Xen extensions off for HVM guests (even if
>> only for debugging purposes) is a severely short sighted decision.
>
> For HVM perhaps, but not for PV.

Right...

I'm confused as to what in my sentence is in any way unclear.

>
>> It is also a feature which has been requested repeatedly by users in the
>> past, and I am very deliberately building a way to do this into the
>> CPUID work.
>>
>> However, it is an unreasonable request to bundle into this bugfix, hence
>> why I didn't suggest it.
>
> There's no bug fix here, as there's no bug (in Xen).

?  I didn't say it was a bug in Xen, but the change is specifically to
fix a bug.

>
>> Now I think about it, things like available ABIs really should be in the
>> Xen hypervisor CPUID leaves, but again, that ship sailed a decade ago.
>
> For comparison, do you know of any CPU architecture making _all_
> of its basic insns and other functionality available conditionally
> only?

No, but that's also not what I'm suggesting.

For an HVM guest, literally everything Xen-specific, from hypercalls to
CPUID leaves to xenstored and PV drivers, is an extension on a base
IBM-compatible PC.

There should always have been a way of of running an HVM guest without
any of this, and we (the Xen community) really do have users which want
to be able to do this.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel