Re: [RFC Part2 PATCH 07/30] mm: add support to split the large THP based on RMP violation

From: Dave Hansen
Date: Thu Mar 25 2021 - 12:00:54 EST


On 3/25/21 8:24 AM, Brijesh Singh wrote:
> On 3/25/21 9:48 AM, Dave Hansen wrote:
>> On 3/24/21 10:04 AM, Brijesh Singh wrote:
>>> When SEV-SNP is enabled globally in the system, a write from the hypervisor
>>> can raise an RMP violation. We can resolve the RMP violation by splitting
>>> the virtual address to a lower page level.
>>>
>>> e.g
>>> - guest made a page shared in the RMP entry so that the hypervisor
>>> can write to it.
>>> - the hypervisor has mapped the pfn as a large page. A write access
>>> will cause an RMP violation if one of the pages within the 2MB region
>>> is a guest private page.
>>>
>>> The above RMP violation can be resolved by simply splitting the large
>>> page.
>> What if the large page is provided by hugetlbfs?
> I was not able to find a method to split the large pages in the
> hugetlbfs. Unfortunately, at this time a VMM cannot use the backing
> memory from the hugetlbfs pool. An SEV-SNP aware VMM can use either
> transparent hugepage or small pages.

That's really, really nasty. Especially since it might not be evident
until long after boot and the guest is killed.

It's even nastier because hugetlbfs is actually a great fit for SEV-SNP
memory. It's physically contiguous, so it would keep you from having to
fracture the direct map all the way down to 4k, it also can't be
reclaimed (just like all SEV memory).

I think the minimal thing you can do here is to fail to add memory to
the RMP in the first place if you can't split it. That way, users will
at least fail to _start_ their VM versus dying randomly for no good reason.

Even better would be to come up with a stronger contract between host
and guest. I really don't think the host should be exposed to random
RMP faults on the direct map. If the guest wants to share memory, then
it needs to tell the host and give the host an opportunity to move the
guest physical memory. It might, for instance, sequester all the shared
pages in a single spot to minimize direct map fragmentation.

I'll let the other x86 folks chime in on this, but I really think this
needs a different approach than what's being proposed.