CVE-2026-99585: KVM: arm64: Use kvm_s2_fault_vma_info in gmem_abort()
In the Linux kernel, the following vulnerability has been resolved:
KVM: arm64: Use kvms2faultvmainfo in gmemabort()
gmemabort() maps at s2fd->faultipa, which HPFAREL2 holds at 4K granularity whatever the page size. The generic page-table code aligns it, but pkvmpgtablestage2map() looks up existing mappings over [addr, addr + size), so on a pKVM host with pages larger than 4K a guestmemfd-backed guest that faults past the first 4K of a page can find its neighbour's mapping, get -EAGAIN and take the same fault forever.
The memory fault exit reports s2fd->faultipa too, which for a nested guest's fault is the nested IPA rather than the canonical one kvmgmemgetpfn() failed on.
Take the addresses from kvms2faultvmainfo instead, as usermemabort() does, and report the canonical gfn in the memory fault exit. kvms2faultgetvmainfo() itself isn't called: a guestmemfd memslot's userspaceaddr doesn't need to be backed by a VMA.
Frequently Asked Questions
Which systems are affected in practice?
The issue affects pKVM hosts using page sizes larger than 4K with guest_memfd-backed guests. The problematic case occurs when a guest faults beyond the first 4K portion of a page.
What is required to trigger the fault loop?
A guest_memfd-backed guest must fault at an address past the first 4K of a host page on a pKVM host whose pages are larger than 4K. The lookup can then encounter a neighboring mapping, return -EAGAIN, and repeatedly fault on the same access.
How can an existing affected system be recognized?
Affected systems may show a guest repeatedly taking the same memory fault without progress because gmem_abort() receives -EAGAIN after finding a neighboring mapping. For nested guests, memory fault exits can also report the nested IPA rather than the canonical GFN associated with the failed guest_memfd lookup.