Re: [PATCH v1 2/4] xen-version: Add third parameter (len) to the do_version hypercall.

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 09/10/15 13:46, Ian Campbell wrote:
> On Fri, 2015-10-09 at 13:29 +0100, Andrew Cooper wrote:
>> On 09/10/15 09:25, Jan Beulich wrote:
>>>>>> On 09.10.15 at 04:56, <konrad.wilk@oracle.com> wrote:
>>>> All existing commands ignore the parameter so this does
>>>> not break the ABI.
>>> Does it not? What about the debug mode clobbering of hypercall
>>> argument registers?
>> That is an implementation detail of the hypervisor.  It is irrelevant to
>> guests whether Xen chooses to clobber the spare registers or not.
> Or in other words the effect here is to clobber one _less_ register, and
> the guest cannot have been relying on a register getting so clobbered (if
> nothing else it doesn't happen in debug=n builds).

Correct - that is definitely a better phrasing.

> The flip side is that we are now no longer clobbering that register even
> for existing sub-ops which do not use it (since the clobbering doesn't go
> down to the subop level). So there is a risk that a guest may come to
> depend on that register not being clobbered and then fail older debug=y
> hypervisors.
>
> This second scenario doesn't seem especially likely to me.
>
> Do we not already have one or two hypercalls where subops consume different
> numbers of parameters anyway? HYPERVISOR_sched_op I think has this property
> and we've not been too concerned.

SCHEDOP_shutdown(suspend) effectively has a third parameter for PV
guests, but that is done entirely by the toolstack using the VMs
register context.

Xen isn't aware of it, and the duration of do_sched_op() does have the
register clobbered.

~Andrew