Re: [PATCH 1/2] KVM: arm64: nv: Allocate the shadow S2 MMUs individually

[Karl Mehltretter <kmehltretter@gmail.com>]

On Tue, Aug 04, 2026 at 03:31:13PM +0100, Marc Zyngier wrote:
> See the hack below that seems to work OK.
> 

Hi Marc,

Since error handling has come up again, I should clarify why
v2 did not use this sketch verbatim.

> -	kvm_init_nested(kvm);
> +	ret = kvm_init_nested(kvm);
> +	if (ret)
> +		return ret;
>  
>  	ret = kvm_share_hyp(kvm, kvm + 1);
>  	if (ret)

Once kvm_init_nested() has allocated the pointer table, failure in
kvm_share_hyp() or any subsequent VM initialisation step leaks that
table.

v2 therefore calls kvm_init_nested() after the earlier initialisation
steps have succeeded and frees the table if a later step in
kvm_arch_init_vm() fails.

> +		for (int i = 0; !ret && i < S2_MMU_PER_VCPU; i++)
> +			ret = init_nested_s2_mmu(kvm, &tmp[i]);

> +		if (ret) {
> +			for (int i = 0; i < S2_MMU_PER_VCPU; i++)
> +				kvm_free_stage2_pgd(&tmp[i]);

This cleanup includes the entry whose initialisation failed and any
entries whose initialisation was never attempted. An uninitialised
entry has no valid mmu->arch, but kvm_free_stage2_pgd() immediately
derives kvm from mmu->arch before checking mmu->pgt. It can therefore
dereference an invalid pointer.

This is why v2 tracks successfully initialised MMUs and frees only
those entries.

Thanks,
Karl