[PATCH v6 30/43] arm64: rme: Prevent Device mappings for Realms

[Gavin Shan]

On 12/13/24 1:55 AM, Steven Price wrote:
> Physical device assignment is not yet supported by the RMM, so it
> doesn't make much sense to allow device mappings within the realm.
> Prevent them when the guest is a realm.
> 
> Signed-off-by: Steven Price <steven.price at arm.com>
> ---
> Changes from v5:
>   * Also prevent accesses in user_mem_abort()
> ---
>   arch/arm64/kvm/mmu.c | 12 ++++++++++++
>   1 file changed, 12 insertions(+)
> 
> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> index 9ede143ccef1..cef7c3dcbf99 100644
> --- a/arch/arm64/kvm/mmu.c
> +++ b/arch/arm64/kvm/mmu.c
> @@ -1149,6 +1149,10 @@ int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
>   	if (is_protected_kvm_enabled())
>   		return -EPERM;
>   
> +	/* We don't support mapping special pages into a Realm */
> +	if (kvm_is_realm(kvm))
> +		return -EINVAL;
> +

		return -EPERM;

>   	size += offset_in_page(guest_ipa);
>   	guest_ipa &= PAGE_MASK;
>   
> @@ -1725,6 +1729,14 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>   	if (exec_fault && device)
>   		return -ENOEXEC;
>   
> +	/*
> +	 * Don't allow device accesses to protected memory as we don't (yet)
> +	 * support protected devices.
> +	 */
> +	if (device && kvm_is_realm(kvm) &&
> +	    kvm_gpa_from_fault(kvm, fault_ipa) == fault_ipa)
> +		return -EINVAL;
> +

s/kvm_is_realm/vcpu_is_rec

I don't understand the check very well. What I understood is mem_abort() is called
only when kvm_gpa_from_fault(kvm, fault_ipa) != fault_ipa, meaning only the page
faults in the shared address space is handled by mem_abort(). So I guess we perhaps
need something like below.

	if (vcpu_is_rec(vcpu) && device)
		return -EPERM;

kvm_handle_guest_abort
   kvm_slot_can_be_private
     private_memslot_fault	// page fault in the private space is handled here
   io_mem_abort			// MMIO emulation is handled here
   user_mem_abort                // page fault in the shared space is handled here

>   	/*
>   	 * Potentially reduce shadow S2 permissions to match the guest's own
>   	 * S2. For exec faults, we'd only reach this point if the guest

Thanks,
Gavin