Re: [RFC PATCH 9/9] kvm_main.c: handle atomic memslot update

[Emanuele Giuseppe Esposito]

Am 27/09/2022 um 11:22 schrieb David Hildenbrand:
> On 27.09.22 10:35, Emanuele Giuseppe Esposito wrote:
>>
>>
>> Am 27/09/2022 um 09:46 schrieb David Hildenbrand:
>>> On 09.09.22 12:45, Emanuele Giuseppe Esposito wrote:
>>>> When kvm_vm_ioctl_set_memory_region_list() is invoked, we need
>>>> to make sure that all memslots are updated in the inactive list
>>>> and then swap (preferreably only once) the lists, so that all
>>>> changes are visible immediately.
>>>>
>>>> The only issue is that DELETE and MOVE need to perform 2 swaps:
>>>> firstly replace old memslot with invalid, and then remove invalid.
>>>>
>>>
>>> I'm curious, how would a resize (grow/shrink) or a split be handled?
>>>
>>
>> There are only 4 operations possible in KVM: KVM_MR_{DELETE, MOVE,
>> CREATE, FLAGS_ONLY}.
>>
>> A resize should be implemented in QEMU as DELETE+CREATE.
>>
>> Therefore a resize on memslot X will be implemented as:
>> First pass on the userspace operations:
>>     invalidate memslot X;
>>     swap_memslot_list(); // NOW it is visible to the guest
>>
>> What guest sees: memslot X is invalid, so MMU keeps retrying the page
>> fault
>>
>> Second pass:
>>     create new memslot X
>>     delete old memslot X
> 
> Thanks a lot for the very nice explanation!

Anytime :)

> Does the invalidation already free up memslot metadata (especially the
> rmaps) or will we end up temporarily allocating twice the memslot metadata?
> 

Invalidation creates a new temporary identical memslot, I am not sure
about the rmaps. It is anyways the same code as it was done before and
if I understand correctly, a new slot is required to keep the old
intact, in case something goes wrong and we need to revert.

Thanks,
Emanuele