Re: [PATCH v7 34/45] kvm: rme: Hide KVM_CAP_READONLY_MEM for realm guests

[Gavin Shan <gshan@redhat.com>]

On 4/8/25 2:34 AM, Steven Price wrote:
> On 04/03/2025 11:51, Gavin Shan wrote:
>> On 2/14/25 2:14 AM, Steven Price wrote:
>>> For protected memory read only isn't supported. While it may be possible
>>> to support read only for unprotected memory, this isn't supported at the
>>> present time.
>>>
>>> Signed-off-by: Steven Price <steven.price@arm.com>
>>> ---
>>>    arch/arm64/kvm/arm.c | 2 +-
>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>
>> It's worthy to explain why KVM_CAP_READONLY_MEM isn't supported and its
>> negative impact. It's something to be done in the future if I'm correct.
> 
> I'll add to the commit message:
> 
>      Note that this does mean that e.g. ROM (or flash) data cannot be
>      emulated correctly by the VMM.
> 

Please also to mention this if you agree: At present, there is no exposed
APIs from RMM allowing to specifying stage-2 page-table entry's permission.
read-only regions for ROM and flash have to be backed up by read-write stage-2
page-table entries. It's going to rely on the stage-1 page-table to have the
proper permissions for those read-only regions.

>>  From QEMU's perspective, all ROM data, which is populated by it, can
>> be written. It conflicts to the natural limit: all ROM data should be
>> read-only.
> 
> Yes this is my understanding of the main impact. I'm not sure how useful
> (shared) ROM/flash emulation is. It can certainly be added in the future
> if needed. Protected read-only memory I don't believe is useful - the
> only sane response I can see from a write fault in that case is killing
> the guest.
> 

Yes, VMM is still able to write to those regions even they're read-only
since they're emulated. For misbehaving guest where those regions are also
mapped as read-write, the data resident in those regions can be corrupted
by guest. It's not the expected output.

Since RMM doesn't have exposed APIs allowing to specify page-table entry's
permissions, meaning all entries have read-write permissions, we have to
give read-write permission to those read-only regions for now. In long run,
it's something to be fixed, starting from RMM.

Thanks,
Gavin

> Thanks,
> Steve
> 
>> QEMU
>> ====
>> rom_add_blob
>>    rom_set_mr
>>      memory_region_set_readonly
>>        memory_region_transaction_commit
>>          kvm_region_commit
>>            kvm_set_phys_mem
>>              kvm_mem_flags                                    // flag
>> KVM_MEM_READONLY is missed
>>              kvm_set_user_memory_region
>>                kvm_vm_ioctl(KVM_SET_USER_MEMORY_REGION2)
>>
>> non-secure host
>> ===============
>> rec_exit_sync_dabt
>>    kvm_handle_guest_abort
>>      user_mem_abort
>>        __kvm_faultin_pfn                       // writable == true
>>          realm_map_ipa
>>            WARN_ON(!(prot & KVM_PGTABLE_PROT_W)
>>
>> non-secure host
>> ===============
>> kvm_realm_enable_cap(KVM_CAP_ARM_RME_POPULATE_REALM)
>>    kvm_populate_realm
>>      __kvm_faultin_pfn                      // writable == true
>>        realm_create_protected_data_page
>>
>>> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
>>> index 1f3674e95f03..0f1d65f87e2b 100644
>>> --- a/arch/arm64/kvm/arm.c
>>> +++ b/arch/arm64/kvm/arm.c
>>> @@ -348,7 +348,6 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm,
>>> long ext)
>>>        case KVM_CAP_ONE_REG:
>>>        case KVM_CAP_ARM_PSCI:
>>>        case KVM_CAP_ARM_PSCI_0_2:
>>> -    case KVM_CAP_READONLY_MEM:
>>>        case KVM_CAP_MP_STATE:
>>>        case KVM_CAP_IMMEDIATE_EXIT:
>>>        case KVM_CAP_VCPU_EVENTS:
>>> @@ -362,6 +361,7 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm,
>>> long ext)
>>>        case KVM_CAP_COUNTER_OFFSET:
>>>            r = 1;
>>>            break;
>>> +    case KVM_CAP_READONLY_MEM:
>>>        case KVM_CAP_SET_GUEST_DEBUG:
>>>            r = !kvm_is_realm(kvm);
>>>            break;
>>
>> Thanks,
>> Gavin
>>
>