KVM: arm64: Allow module-owned pages in host stage-2 range adjustment

[PATCH] KVM: arm64: Allow module-owned pages in host stage-2 range adjustment

Posted by kyle-jk.liao@mediatek.com 9 hours ago

From: "kyle-jk.liao" <kyle-jk.liao@mediatek.com>

The host_stage2_adjust_range() function triggers a WARN_ON when
encountering pages marked as both PKVM_NOPAGE and PKVM_MODULE_OWNED_PAGE.
This occurs when load_unaligned_zeropad() accesses memory that crosses
page boundaries, where one of the pages is module-owned.

Extend the check to accept pages with the PKVM_MODULE_OWNED_PAGE flag
in addition to PKVM_NOPAGE, preventing spurious warnings during normal
module memory access patterns.

Signed-off-by: kyle-jk.liao <kyle-jk.liao@mediatek.com>
---
 arch/arm64/kvm/hyp/nvhe/mem_protect.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/kvm/hyp/nvhe/mem_protect.c b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
index 49db32f3ddf7..baaabf4b5112 100644
--- a/arch/arm64/kvm/hyp/nvhe/mem_protect.c
+++ b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
@@ -495,6 +495,7 @@ static int host_stage2_adjust_range(u64 addr, struct kvm_mem_range *range)
 	u64 granule;
 	s8 level;
 	int ret;
+	enum pkvm_page_state state;
 
 	hyp_assert_lock_held(&host_mmu.lock);
 	ret = kvm_pgtable_get_leaf(&host_mmu.pgt, addr, &pte, &level);
@@ -505,8 +506,9 @@ static int host_stage2_adjust_range(u64 addr, struct kvm_mem_range *range)
 		return -EAGAIN;
 
 	if (pte) {
+		state = get_host_state(hyp_phys_to_page(addr));
 		WARN_ON(addr_is_memory(addr) &&
-			get_host_state(hyp_phys_to_page(addr)) != PKVM_NOPAGE);
+			(state != PKVM_NOPAGE && state != (PKVM_NOPAGE | PKVM_MODULE_OWNED_PAGE)));
 		return -EPERM;
 	}
 
-- 
2.45.2

Re: [PATCH] KVM: arm64: Allow module-owned pages in host stage-2 range adjustment

Posted by Marc Zyngier 3 hours ago

On Mon, 09 Feb 2026 03:24:13 +0000,
<kyle-jk.liao@mediatek.com> wrote:
> 
> From: "kyle-jk.liao" <kyle-jk.liao@mediatek.com>
> 
> The host_stage2_adjust_range() function triggers a WARN_ON when
> encountering pages marked as both PKVM_NOPAGE and PKVM_MODULE_OWNED_PAGE.
> This occurs when load_unaligned_zeropad() accesses memory that crosses
> page boundaries, where one of the pages is module-owned.
> 
> Extend the check to accept pages with the PKVM_MODULE_OWNED_PAGE flag
> in addition to PKVM_NOPAGE, preventing spurious warnings during normal
> module memory access patterns.
> 
> Signed-off-by: kyle-jk.liao <kyle-jk.liao@mediatek.com>
> ---
>  arch/arm64/kvm/hyp/nvhe/mem_protect.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/kvm/hyp/nvhe/mem_protect.c b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
> index 49db32f3ddf7..baaabf4b5112 100644
> --- a/arch/arm64/kvm/hyp/nvhe/mem_protect.c
> +++ b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
> @@ -495,6 +495,7 @@ static int host_stage2_adjust_range(u64 addr, struct kvm_mem_range *range)
>  	u64 granule;
>  	s8 level;
>  	int ret;
> +	enum pkvm_page_state state;
>  
>  	hyp_assert_lock_held(&host_mmu.lock);
>  	ret = kvm_pgtable_get_leaf(&host_mmu.pgt, addr, &pte, &level);
> @@ -505,8 +506,9 @@ static int host_stage2_adjust_range(u64 addr, struct kvm_mem_range *range)
>  		return -EAGAIN;
>  
>  	if (pte) {
> +		state = get_host_state(hyp_phys_to_page(addr));
>  		WARN_ON(addr_is_memory(addr) &&
> -			get_host_state(hyp_phys_to_page(addr)) != PKVM_NOPAGE);
> +			(state != PKVM_NOPAGE && state != (PKVM_NOPAGE | PKVM_MODULE_OWNED_PAGE)));
>  		return -EPERM;
>  	}
>  

None of that exists upstream.

	M.

-- 
Without deviation from the norm, progress is not possible.