Re: [PATCH] ArmVirtPkg: restrict mapping attributes of normal memory to EFI_MEMORY_WB

[Ard Biesheuvel <ard.biesheuvel@linaro.org>]

On 8 September 2016 at 08:54, Laszlo Ersek <lersek@redhat.com> wrote:
> On 09/08/16 09:50, Ard Biesheuvel wrote:
>> In general, on an ARM system, mapping normal memory as device memory may
>> have unintended side effects, given that unaligned accesses or loads and
>> stores with special semantics (e.g., load/store exclusive) may fault or
>> may not work as expected.
>>
>> Under KVM, the situation is even worse, since the host may not expect the
>> guest to perform uncached accesses, and so writes to such an uncached
>> region may get lost completely.
>>
>> Since the only safe mapping type under KVM is EFI_MEMORY_WB, remove all
>> other memory type attributes.
>>
>> Contributed-under: TianoCore Contribution Agreement 1.0
>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
>> ---
>>  ArmVirtPkg/HighMemDxe/HighMemDxe.c                                   | 3 +--
>>  ArmVirtPkg/Library/ArmVirtMemoryInitPeiLib/ArmVirtMemoryInitPeiLib.c | 3 ---
>>  2 files changed, 1 insertion(+), 5 deletions(-)
>>
>> diff --git a/ArmVirtPkg/HighMemDxe/HighMemDxe.c b/ArmVirtPkg/HighMemDxe/HighMemDxe.c
>> index 7fd7e8e9a539..4d56e6236b54 100644
>> --- a/ArmVirtPkg/HighMemDxe/HighMemDxe.c
>> +++ b/ArmVirtPkg/HighMemDxe/HighMemDxe.c
>> @@ -78,8 +78,7 @@ InitializeHighMemDxe (
>>            Status = gDS->AddMemorySpace (
>>                            EfiGcdMemoryTypeSystemMemory,
>>                            CurBase, CurSize,
>> -                          EFI_MEMORY_WB | EFI_MEMORY_WC |
>> -                          EFI_MEMORY_WT | EFI_MEMORY_UC);
>> +                          EFI_MEMORY_WB);
>>
>>            if (EFI_ERROR (Status)) {
>>              DEBUG ((EFI_D_ERROR,
>> diff --git a/ArmVirtPkg/Library/ArmVirtMemoryInitPeiLib/ArmVirtMemoryInitPeiLib.c b/ArmVirtPkg/Library/ArmVirtMemoryInitPeiLib/ArmVirtMemoryInitPeiLib.c
>> index 251e5314e61d..6f3e54b7afcb 100644
>> --- a/ArmVirtPkg/Library/ArmVirtMemoryInitPeiLib/ArmVirtMemoryInitPeiLib.c
>> +++ b/ArmVirtPkg/Library/ArmVirtMemoryInitPeiLib/ArmVirtMemoryInitPeiLib.c
>> @@ -68,9 +68,6 @@ MemoryPeim (
>>    ResourceAttributes = (
>>        EFI_RESOURCE_ATTRIBUTE_PRESENT |
>>        EFI_RESOURCE_ATTRIBUTE_INITIALIZED |
>> -      EFI_RESOURCE_ATTRIBUTE_UNCACHEABLE |
>> -      EFI_RESOURCE_ATTRIBUTE_WRITE_COMBINEABLE |
>> -      EFI_RESOURCE_ATTRIBUTE_WRITE_THROUGH_CACHEABLE |
>>        EFI_RESOURCE_ATTRIBUTE_WRITE_BACK_CACHEABLE |
>>        EFI_RESOURCE_ATTRIBUTE_TESTED
>>    );
>>
>
> Reviewed-by: Laszlo Ersek <lersek@redhat.com>
>
> Did you encounter any specific problem with these?
>

No, but I am looking into using the new optimized BaseMemoryLibOptDxe,
and the AARCH64 version uses DC ZVA instructions for ZeroMem() [i.e.,
zero a cacheline, which is much faster than writing zeroes, obviously]
That does not work on uncached memory, and so while I was removing
/that/ attribute from ArmVirtPkg, I realized that WC and WT are not
safe either.

For bare metal platforms, we will keep WC and WT, but for ArmVirtQemu,
they don't make sense. Not sure about Xen though ...