Re: [PATCH] ram-block-attributes: Avoid the overkill of shared memory with hugetlbfs backend

[Chenyi Qiang <chenyi.qiang@intel.com>]

On 10/17/2025 11:13 PM, David Hildenbrand wrote:
> On 17.10.25 10:14, Chenyi Qiang wrote:
>> Currently, private memory and shared memory have different backend in
>> CoCo VMs. It is possible for users to specify the shared memory with
>> hugetlbfs backend while private memory with guest_memfd backend only
>> supports 4K page size. In this case, ram_block->page_size is different
>> from the host page size which will trigger the assertion when getting
>> block size. Relax the restriction to allow shared memory to use
>> hugetlbfs backend.
>>
>> Fixes: 5d6483edaa92 ("ram-block-attributes: Introduce RamBlockAttributes to manage RAMBlock with guest_memfd")
>> Signed-off-by: Chenyi Qiang <chenyi.qiang@intel.com>
>> ---
>>   system/ram-block-attributes.c | 7 ++++---
>>   1 file changed, 4 insertions(+), 3 deletions(-)
>>
>> diff --git a/system/ram-block-attributes.c b/system/ram-block-attributes.c
>> index 68e8a027032..0f39ccf9090 100644
>> --- a/system/ram-block-attributes.c
>> +++ b/system/ram-block-attributes.c
>> @@ -28,10 +28,11 @@ ram_block_attributes_get_block_size(const RamBlockAttributes *attr)
>>        * Because page conversion could be manipulated in the size of at least 4K
>>        * or 4K aligned, Use the host page size as the granularity to track the
>>        * memory attribute.
>> +     * When hugetlbfs is used as backend of shared memory, ram_block->page_size
>> +     * is different from host page size. So it is not appropriate to use
>> +     * ram_block->page_size here.
> 
> But are we sure everything else is working as expected and that this is not a check that prevents other code from doing the wrong thing?

I think so. The block size must be 4K due to the page conversion could be in the size of 4K and we use "bitmap" to track the status.
I originally missed the case of hugetlb so added an assert() here. But it is allowed to use hugetlb as shared memory backend
before shared device assignment patches were introduced.

> 
> I recall that punching holes was problematic as the VM shares/unshared 4k chunks.

I can see the kvm_convert_memory() will skip ram_block_discard_range() if using hugetlb backend.
It will cause the double-memory consumption (*). Any other problem?

[*] https://lore.kernel.org/qemu-devel/5f65ec5d-7079-4b32-8e40-b9ece3bdd47c@redhat.com/

>