Re: [PATCH v3 00/14] x86/hvm: {svm,vmx} {c,h} cleanup

[Jan Beulich <jbeulich@xxxxxxxx>]

On 27.02.2023 13:06, Andrew Cooper wrote:
> On 27/02/2023 11:33 am, Jan Beulich wrote:
>> On 27.02.2023 12:15, Andrew Cooper wrote:
>>> On 27/02/2023 10:46 am, Jan Beulich wrote:
>>>> On 24.02.2023 22:33, Andrew Cooper wrote:
>>>>> But I think we want to change tact slightly at this point, so I'm not
>>>>> going to do any further tweaking on commit.
>>>>>
>>>>> Next, I think we want to rename asm/hvm/svm/svm.h to asm/hvm/svm.h,
>>>>> updating the non-SVM include paths as we go.  Probably best to
>>>>> chain-include the other svm/hvm/svm/*.h headers temporarily, so we've
>>>>> only got one tree-wide cleanup of the external include paths.
>>>>>
>>>>> Quick tangent - I will be making all of that cpu_has_svm_*
>>>>> infrastructure disappear by moving it into the normal CPUID handling,
>>>>> but I've not had sufficient time to finish that yet.
>>>>>
>>>>> Next, hvm/svm/nestedsvm.h can merge straight into hvm/svm.h, and
>>>>> disappear (after my decoupling patch has gone in).
>>>> Why would you want to fold hvm/svm/nestedsvm.h into hvm/svm/svm.h?
>>>> The latter doesn't use anything from the former, does it?
>>> It's about what else uses them.
>>>
>>> hvm_vcpu needs both svm_vcpu and nestedsvm, so both headers are always
>>> included in tandem.
>> Well, yes, that's how things are today. But can you explain to me why
>> hvm_vcpu has to know nestedsvm's layout?
> 
> Because it's part of the same single memory allocation.

Which keeps growing, and sooner or later we'll need to find something
again to split off, so we won't exceed a page in size. The nested
structures would, to me, look to be prime candidates for such.

>> If the field was a pointer,
>> a forward decl of that struct would suffice, and any entity in the
>> rest of Xen not caring about struct nestedsvm would no longer see it
>> (and hence also no longer be re-built if a change is made there).
> 
> Yes, you could hide it as a pointer.  The cost of doing so is an
> unnecessary extra memory allocation, and extra pointer handling on
> create/destroy, not to mention extra pointer chasing in memory during use.
> 
>>> nestedsvm is literally just one struct now, and no subsystem ought to
>>> have multiple headers when one will do.
>> When one will do, yes. Removing build dependencies is a good reason
>> to have separate headers, though.
> 
> Its not the only only option, and an #ifdef CONFIG_NESTED_VIRT inside
> the struct would be an equally acceptable way of doing this which
> wouldn't involve making an extra memory allocation.

That would make it a build-time decision, but then on NESTED_VIRT=y
hypervisors there might still be guests not meaning to use that
functionality, and for quite some time that may actually be a majority.

> Everything you've posed here is way out of scope for Xenia's series. 

There was never an intention to extend the scope of the work she's doing.
Instead I was trying to limit the scope by suggesting to avoid a piece
of rework which later may want undoing.

Jan