On 5/26/2025 8:10 PM, David Hildenbrand wrote:
> On 26.05.25 12:19, Chenyi Qiang wrote:
>>
>>
>> On 5/26/2025 5:17 PM, David Hildenbrand wrote:
>>> On 20.05.25 12:28, Chenyi Qiang wrote:
>>>> The current error handling is simple with the following assumption:
>>>> - QEMU will quit instead of resuming the guest if kvm_convert_memory()
>>>> fails, thus no need to do rollback.
>>>> - The convert range is required to be in the desired state. It is not
>>>> allowed to handle the mixture case.
>>>> - The conversion from shared to private is a non-failure operation.
>>>>
>>>> This is sufficient for now as complext error handling is not required.
>>>> For future extension, add some potential error handling.
>>>> - For private to shared conversion, do the rollback operation if
>>>> ram_block_attribute_notify_to_populated() fails.
>>>> - For shared to private conversion, still assert it as a non-failure
>>>> operation for now. It could be an easy fail path with in-place
>>>> conversion, which will likely have to retry the conversion until it
>>>> works in the future.
>>>> - For mixture case, process individual blocks for ease of rollback.
>>>>
>>>> Signed-off-by: Chenyi Qiang <chenyi.qi...@intel.com>
>>>> ---
>>>> system/ram-block-attribute.c | 116 ++++++++++++++++++++++++++
>>>> +--------
>>>> 1 file changed, 90 insertions(+), 26 deletions(-)
>>>>
>>>> diff --git a/system/ram-block-attribute.c b/system/ram-block-
>>>> attribute.c
>>>> index 387501b569..0af3396aa4 100644
>>>> --- a/system/ram-block-attribute.c
>>>> +++ b/system/ram-block-attribute.c
>>>> @@ -289,7 +289,12 @@ static int
>>>> ram_block_attribute_notify_to_discard(RamBlockAttribute *attr,
>>>> }
>>>> ret = rdl->notify_discard(rdl, &tmp);
>>>> if (ret) {
>>>> - break;
>>>> + /*
>>>> + * The current to_private listeners (VFIO dma_unmap and
>>>> + * KVM set_attribute_private) are non-failing operations.
>>>> + * TODO: add rollback operations if it is allowed to fail.
>>>> + */
>>>> + g_assert(ret);
>>>> }
>>>> }
>>>>
>>>
>>> If it's not allowed to fail for now, then patch #8 does not make sense
>>> and should be dropped :)
>>
>> It was intended for future extension as in-place conversion to_private
>> allows it to fail. So I add the patch #8.
>>
>> But as you mentioned, since the conversion path is changing, and maybe
>> it is easier to handle from KVM code directly. Let me drop patch #8 and
>> wait for the in-place conversion to mature.
>
> Makes sense. I'm afraid it might all be a bit complicated to handle:
> vfio can fail private -> shared conversion and KVM the shared -> private
> conversion.
>
> So recovering ... will not be straight forward once multiple pages are
> converted.
>
>>
>>>
>>> The implementations (vfio) should likely exit() instead on unexpected
>>> errors when discarding.
>>
>> After drop patch #8, maybe keep vfio discard handling as it was. Adding
>> an additional exit() is also OK to me since it's non-fail case.
>>
>>>
>>>
>>>
>>> Why not squash all the below into the corresponding patch? Looks mostly
>>> like handling partial conversions correctly (as discussed previously)?
>>
>> I extract these two handling 1) mixture conversion; 2) operation
>> rollback into this individual patch because they are not the practical
>> cases and are untested.
>>
>> For 1), I still don't see any real case which will convert a range with
>> mixture attributes.
>
> Okay. I thought we were not sure if the guest could trigger that?
Yes. At least I didn't see any statement in TDX GHCI spec to mention
such mixture case. And in TDX KVM code, it will check the start and end
gpa to have the same attribute before exiting to userspace: (maybe with
the assumption that the whole range shares the same attribute?)
vt_is_tdx_private_gpa(vcpu->kvm, gpa) !=
vt_is_tdx_private_gpa(vcpu->kvm, gpa + size - 1)
>
> I think it would be better to just include the "mixture" handling in the
> original patch.
[...]
>
>>
>> For 2), current failure of memory conversion (as seen in kvm_cpu_exec()
>> ->kvm_convert_memory()) will cause the QEMU to quit instead of resuming
>> guest. Doing the rollback seems useless at present.
>
> Makes sense.
OK, then I will move the mixture handling in the original patch and
still keep the rollback operation separately.
>
