Re: [PATCH] LoongArch: KVM: Preserve memslot arch flags on KVM_MR_FLAGS_ONLY

[Tao Cui <cui.tao@linux.dev>]

在 2026/8/28 17:48, Zeng Chi 写道:
> From: Zeng Chi <zengchi@kylinos.cn>
> 
> kvm_arch_prepare_memory_region() computes new->arch.flags, i.e. whether
> a memslot is KVM_MEM_HUGEPAGE_CAPABLE or KVM_MEM_HUGEPAGE_INCAPABLE, only
> for KVM_MR_CREATE and KVM_MR_MOVE, and returns early for every other
> change.  But the generic code allocates a zeroed memslot for every
> change and never copies old->arch, so after a KVM_MR_FLAGS_ONLY update,
> e.g. toggling KVM_MEM_LOG_DIRTY_PAGES for live migration, the active
> memslot has arch.flags == 0.
> 
> With both flags clear, fault_supports_huge_mapping() falls through to
> the alignment check on the HVA range alone, which no longer verifies
> that the GPA and HVA have the same offset within a PMD.  A memslot that
> was marked KVM_MEM_HUGEPAGE_INCAPABLE because of a GPA/HVA offset
> mismatch can then be mapped with PMD entries on read faults, and since
> kvm_map_page() aligns the gfn and the pfn independently, the guest ends
> up accessing the wrong host pages, exactly the "d -> f, e -> g" case
> described in the comment above the check.
> 
> Carry the arch flags over from the old memslot for KVM_MR_FLAGS_ONLY,
> as the GPA, HVA and size are guaranteed to be unchanged for that case.
> 
> Fixes: 7ab6fb505b2a ("LoongArch: KVM: Optimization for memslot hugepage checking")
> Cc: stable@vger.kernel.org
> Signed-off-by: Zeng Chi <zengchi@kylinos.cn>
> ---
>  arch/loongarch/kvm/mmu.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> diff --git a/arch/loongarch/kvm/mmu.c b/arch/loongarch/kvm/mmu.c
> index e104897aa532..4ca7c6ea8f08 100644
> --- a/arch/loongarch/kvm/mmu.c
> +++ b/arch/loongarch/kvm/mmu.c
> @@ -383,6 +383,16 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm, const struct kvm_memory_slot
>  	hva_t hva_start;
>  	size_t size, gpa_offset, hva_offset;
>  
> +	/*
> +	 * The generic code allocates a fresh, zeroed memslot for every change,
> +	 * so the arch flags computed below must be carried over when only the
> +	 * userspace flags change, e.g. when dirty logging is toggled.
> +	 */
> +	if (change == KVM_MR_FLAGS_ONLY) {
> +		new->arch = old->arch;
> +		return 0;
> +	}
> +
>  	if ((change != KVM_MR_MOVE) && (change != KVM_MR_CREATE))
>  		return 0;
>  	/*

I verified this on a Loongson-3A6000 host with a KVM selftest that uses
a GPA/HVA offset-mismatched memslot and toggles dirty logging: the
unpatched kernel maps the wrong host pages into the guest after the
toggle, and with this patch the mapping stays correct.

Tested-by: Tao Cui <cuitao@kylinos.cn>
Reviewed-by: Tao Cui <cuitao@kylinos.cn>