Re: [RFC PATCH 39/39] KVM: guest_memfd: Dynamically split/reconstruct HugeTLB page

[Yan Zhao <yan.y.zhao@intel.com>]

On Fri, Apr 25, 2025 at 03:45:20PM -0700, Ackerley Tng wrote:
> Yan Zhao <yan.y.zhao@intel.com> writes:
> 
> > On Thu, Apr 24, 2025 at 11:15:11AM -0700, Ackerley Tng wrote:
> >> Vishal Annapurve <vannapurve@google.com> writes:
> >> 
> >> > On Thu, Apr 24, 2025 at 1:15 AM Yan Zhao <yan.y.zhao@intel.com> wrote:
> >> >>
> >> >> On Thu, Apr 24, 2025 at 01:55:51PM +0800, Chenyi Qiang wrote:
> >> >> >
> >> >> >
> >> >> > On 4/24/2025 12:25 PM, Yan Zhao wrote:
> >> >> > > On Thu, Apr 24, 2025 at 09:09:22AM +0800, Yan Zhao wrote:
> >> >> > >> On Wed, Apr 23, 2025 at 03:02:02PM -0700, Ackerley Tng wrote:
> >> >> > >>> Yan Zhao <yan.y.zhao@intel.com> writes:
> >> >> > >>>
> >> >> > >>>> On Tue, Sep 10, 2024 at 11:44:10PM +0000, Ackerley Tng wrote:
> >> >> > >>>>> +/*
> >> >> > >>>>> + * Allocates and then caches a folio in the filemap. Returns a folio with
> >> >> > >>>>> + * refcount of 2: 1 after allocation, and 1 taken by the filemap.
> >> >> > >>>>> + */
> >> >> > >>>>> +static struct folio *kvm_gmem_hugetlb_alloc_and_cache_folio(struct inode *inode,
> >> >> > >>>>> +                                                           pgoff_t index)
> >> >> > >>>>> +{
> >> >> > >>>>> +       struct kvm_gmem_hugetlb *hgmem;
> >> >> > >>>>> +       pgoff_t aligned_index;
> >> >> > >>>>> +       struct folio *folio;
> >> >> > >>>>> +       int nr_pages;
> >> >> > >>>>> +       int ret;
> >> >> > >>>>> +
> >> >> > >>>>> +       hgmem = kvm_gmem_hgmem(inode);
> >> >> > >>>>> +       folio = kvm_gmem_hugetlb_alloc_folio(hgmem->h, hgmem->spool);
> >> >> > >>>>> +       if (IS_ERR(folio))
> >> >> > >>>>> +               return folio;
> >> >> > >>>>> +
> >> >> > >>>>> +       nr_pages = 1UL << huge_page_order(hgmem->h);
> >> >> > >>>>> +       aligned_index = round_down(index, nr_pages);
> >> >> > >>>> Maybe a gap here.
> >> >> > >>>>
> >> >> > >>>> When a guest_memfd is bound to a slot where slot->base_gfn is not aligned to
> >> >> > >>>> 2M/1G and slot->gmem.pgoff is 0, even if an index is 2M/1G aligned, the
> >> >> > >>>> corresponding GFN is not 2M/1G aligned.
> >> >> > >>>
> >> >> > >>> Thanks for looking into this.
> >> >> > >>>
> >> >> > >>> In 1G page support for guest_memfd, the offset and size are always
> >> >> > >>> hugepage aligned to the hugepage size requested at guest_memfd creation
> >> >> > >>> time, and it is true that when binding to a memslot, slot->base_gfn and
> >> >> > >>> slot->npages may not be hugepage aligned.
> >> >> > >>>
> >> >> > >>>>
> >> >> > >>>> However, TDX requires that private huge pages be 2M aligned in GFN.
> >> >> > >>>>
> >> >> > >>>
> >> >> > >>> IIUC other factors also contribute to determining the mapping level in
> >> >> > >>> the guest page tables, like lpage_info and .private_max_mapping_level()
> >> >> > >>> in kvm_x86_ops.
> >> >> > >>>
> >> >> > >>> If slot->base_gfn and slot->npages are not hugepage aligned, lpage_info
> >> >> > >>> will track that and not allow faulting into guest page tables at higher
> >> >> > >>> granularity.
> >> >> > >>
> >> >> > >> lpage_info only checks the alignments of slot->base_gfn and
> >> >> > >> slot->base_gfn + npages. e.g.,
> >> >> > >>
> >> >> > >> if slot->base_gfn is 8K, npages is 8M, then for this slot,
> >> >> > >> lpage_info[2M][0].disallow_lpage = 1, which is for GFN [4K, 2M+8K);
> >> >> > >> lpage_info[2M][1].disallow_lpage = 0, which is for GFN [2M+8K, 4M+8K);
> >> >> > >> lpage_info[2M][2].disallow_lpage = 0, which is for GFN [4M+8K, 6M+8K);
> >> >> > >> lpage_info[2M][3].disallow_lpage = 1, which is for GFN [6M+8K, 8M+8K);
> >> >> >
> >> >> > Should it be?
> >> >> > lpage_info[2M][0].disallow_lpage = 1, which is for GFN [8K, 2M);
> >> >> > lpage_info[2M][1].disallow_lpage = 0, which is for GFN [2M, 4M);
> >> >> > lpage_info[2M][2].disallow_lpage = 0, which is for GFN [4M, 6M);
> >> >> > lpage_info[2M][3].disallow_lpage = 0, which is for GFN [6M, 8M);
> >> >> > lpage_info[2M][4].disallow_lpage = 1, which is for GFN [8M, 8M+8K);
> >> >> Right. Good catch. Thanks!
> >> >>
> >> >> Let me update the example as below:
> >> >> slot->base_gfn is 2 (for GPA 8KB), npages 2000 (for a 8MB range)
> >> >>
> >> >> lpage_info[2M][0].disallow_lpage = 1, which is for GPA [8KB, 2MB);
> >> >> lpage_info[2M][1].disallow_lpage = 0, which is for GPA [2MB, 4MB);
> >> >> lpage_info[2M][2].disallow_lpage = 0, which is for GPA [4MB, 6MB);
> >> >> lpage_info[2M][3].disallow_lpage = 0, which is for GPA [6MB, 8MB);
> >> >> lpage_info[2M][4].disallow_lpage = 1, which is for GPA [8MB, 8MB+8KB);
> >> >>
> >> >> lpage_info indicates that a 2MB mapping is alllowed to cover GPA 4MB and GPA
> >> >> 4MB+16KB. However, their aligned_index values lead guest_memfd to allocate two
> >> >> 2MB folios, whose physical addresses may not be contiguous.
> >> >>
> >> >> Additionally, if the guest accesses two GPAs, e.g., GPA 2MB+8KB and GPA 4MB,
> >> >> KVM could create two 2MB mappings to cover GPA ranges [2MB, 4MB), [4MB, 6MB).
> >> >> However, guest_memfd just allocates the same 2MB folio for both faults.
> >> >>
> >> >>
> >> >> >
> >> >> > >>
> >> >> > >>   ---------------------------------------------------------
> >> >> > >>   |          |  |          |  |          |  |          |  |
> >> >> > >>   8K        2M 2M+8K      4M  4M+8K     6M  6M+8K     8M  8M+8K
> >> >> > >>
> >> >> > >> For GFN 6M and GFN 6M+4K, as they both belong to lpage_info[2M][2], huge
> >> >> > >> page is allowed. Also, they have the same aligned_index 2 in guest_memfd.
> >> >> > >> So, guest_memfd allocates the same huge folio of 2M order for them.
> >> >> > > Sorry, sent too fast this morning. The example is not right. The correct
> >> >> > > one is:
> >> >> > >
> >> >> > > For GFN 4M and GFN 4M+16K, lpage_info indicates that 2M is allowed. So,
> >> >> > > KVM will create a 2M mapping for them.
> >> >> > >
> >> >> > > However, in guest_memfd, GFN 4M and GFN 4M+16K do not correspond to the
> >> >> > > same 2M folio and physical addresses may not be contiguous.
> >> >
> >> > Then during binding, guest memfd offset misalignment with hugepage
> >> > should be same as gfn misalignment. i.e.
> >> >
> >> > (offset & ~huge_page_mask(h)) == ((slot->base_gfn << PAGE_SHIFT) &
> >> > ~huge_page_mask(h));
> >> >
> >> > For non guest_memfd backed scenarios, KVM allows slot gfn ranges that
> >> > are not hugepage aligned, so guest_memfd should also be able to
> >> > support non-hugepage aligned memslots.
> >> >
> >> 
> >> I drew up a picture [1] which hopefully clarifies this.
> >> 
> >> Thanks for pointing this out, I understand better now and we will add an
> >> extra constraint during memslot binding of guest_memfd to check that gfn
> >> offsets within a hugepage must be guest_memfd offsets.
> > I'm a bit confused.
> >
> > As "index = gfn - slot->base_gfn + slot->gmem.pgoff", do you mean you are going
> > to force "slot->base_gfn == slot->gmem.pgoff" ?
> >
> > For some memory region, e.g., "pc.ram", it's divided into 2 parts:
> > - one with offset 0, size 0x80000000(2G),
> >   positioned at GPA 0, which is below GPA 4G;
> > - one with offset 0x80000000(2G), size 0x80000000(2G),
> >   positioned at GPA 0x100000000(4G), which is above GPA 4G.
> >
> > For the second part, its slot->base_gfn is 0x100000000, while slot->gmem.pgoff
> > is 0x80000000.
> >
> 
> Nope I don't mean to enforce that they are equal, we just need the
> offsets within the page to be equal.
> 
> I edited Vishal's code snippet, perhaps it would help explain better:
> 
> page_size is the size of the hugepage, so in our example,
> 
>   page_size = SZ_2M;
>   page_mask = ~(page_size - 1);
page_mask = page_size - 1  ?

>   offset_within_page = slot->gmem.pgoff & page_mask;
>   gfn_within_page = (slot->base_gfn << PAGE_SHIFT) & page_mask;
> 
> We will enforce that
> 
>   offset_within_page == gfn_within_page;
For "pc.ram", if it has 2.5G below 4G, it would be configured as follows
- slot 1: slot->gmem.pgoff=0, base GPA 0, size=2.5G
- slot 2: slot->gmem.pgoff=2.5G, base GPA 4G, size=1.5G

When binding these two slots to the same guest_memfd created with flag
KVM_GUEST_MEMFD_HUGE_1GB: 
- binding the 1st slot will succeed;
- binding the 2nd slot will fail.

What options does userspace have in this scenario?
It can't reduce the flag to KVM_GUEST_MEMFD_HUGE_2MB. Adjusting the gmem.pgoff
isn't ideal either.

What about something similar as below?

diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
index d2feacd14786..87c33704a748 100644
--- a/virt/kvm/guest_memfd.c
+++ b/virt/kvm/guest_memfd.c
@@ -1842,8 +1842,16 @@ __kvm_gmem_get_pfn(struct file *file, struct kvm_memory_slot *slot,
        }

        *pfn = folio_file_pfn(folio, index);
-       if (max_order)
-               *max_order = folio_order(folio);
+       if (max_order) {
+               int order;
+
+               order = folio_order(folio);
+
+               while (order > 0 && ((slot->base_gfn ^ slot->gmem.pgoff) & ((1 << order) - 1)))
+                       order--;
+
+               *max_order = order;
+       }

        *is_prepared = folio_test_uptodate(folio);
        return folio;


> >> Adding checks at binding time will allow hugepage-unaligned offsets (to
> >> be at parity with non-guest_memfd backing memory) but still fix this
> >> issue.
> >> 
> >> lpage_info will make sure that ranges near the bounds will be
> >> fragmented, but the hugepages in the middle will still be mappable as
> >> hugepages.
> >> 
> >> [1] https://lpc.events/event/18/contributions/1764/attachments/1409/3706/binding-must-have-same-alignment.svg