On Thu, Jun 20, 2024, David Hildenbrand wrote:
On 20.06.24 16:29, Jason Gunthorpe wrote:
On Thu, Jun 20, 2024 at 04:01:08PM +0200, David Hildenbrand wrote:
On 20.06.24 15:55, Jason Gunthorpe wrote:
On Thu, Jun 20, 2024 at 09:32:11AM +0100, Fuad Tabba wrote:
Regarding huge pages: assume the huge page (e.g., 1 GiB hugetlb) is shared, now the VM requests to make one subpage private.
I think the general CC model has the shared/private setup earlier on the VM lifecycle with large runs of contiguous pages. It would only become a problem if you intend to to high rate fine granual shared/private switching. Which is why I am asking what the actual "why" is here.
I am not an expert on that, but I remember that the way memory shared<->private conversion happens can heavily depend on the VM use case,
Yeah, I forget the details, but there are scenarios where the guest will share (and unshare) memory at 4KiB (give or take) granularity, at runtime. There's an RFC[*] for making SWIOTLB operate at 2MiB is driven by the same underlying problems.
But even if Linux-as-a-guest were better behaved, we (the host) can't prevent the guest from doing suboptimal conversions. In practice, killing the guest or refusing to convert memory isn't an option, i.e. we can't completely push the problem into the guest
https://lore.kernel.org/all/20240112055251.36101-1-vannapurve@google.com
and that under pKVM we might see more frequent conversion, without even going to user space.
How to handle that without eventually running into a double memory-allocation? (in the worst case, allocating a 1GiB huge page for shared and for private memory).
I expect you'd take the linear range of 1G of PFNs and fragment it into three ranges private/shared/private that span the same 1G.
When you construct a page table (ie a S2) that holds these three ranges and has permission to access all the memory you want the page table to automatically join them back together into 1GB entry.
When you construct a page table that has only access to the shared, then you'd only install the shared hole at its natural best size.
So, I think there are two challenges - how to build an allocator and uAPI to manage this sort of stuff so you can keep track of any fractured pfns and ensure things remain in physical order.
Then how to re-consolidate this for the KVM side of the world.
Exactly!
guest_memfd, or something like it, is just really a good answer. You have it obtain the huge folio, and keep track on its own which sub pages can be mapped to a VMA because they are shared. KVM will obtain the PFNs directly from the fd and KVM will not see the shared holes. This means your S2's can be trivially constructed correctly.
No need to double allocate..
Yes, that's why my thinking so far was:
Let guest_memfd (or something like that) consume huge pages (somehow, let it access the hugetlb reserves). Preallocate that memory once, as the VM starts up: just like we do with hugetlb in VMs.
Let KVM track which parts are shared/private, and if required, let it map only the shared parts to user space. KVM has all information to make these decisions.
If we could disallow pinning any shared pages, that would make life a lot easier, but I think there were reasons for why we might require it. To convert shared->private, simply unmap that folio (only the shared parts could possibly be mapped) from all user page tables.
Of course, there might be alternatives, and I'll be happy to learn about them. The allcoator part would be fairly easy, and the uAPI part would similarly be comparably easy. So far the theory :)
I'm kind of surprised the CC folks don't want the same thing for exactly the same reason. It is much easier to recover the huge mappings for the S2 in the presence of shared holes if you track it this way. Even CC will have this problem, to some degree, too.
Precisely! RH (and therefore, me) is primarily interested in existing guest_memfd users at this point ("CC"), and I don't see an easy way to get that running with huge pages in the existing model reasonably well ...
This is the general direction guest_memfd is headed, but getting there is easier said than done. E.g. as alluded to above, "simply unmap that folio" is quite difficult, bordering on infeasible if the kernel is allowed to gup() shared guest_memfd memory.