On 8 September 2016 at 08:57, Ard Biesheuvel <[email protected]> wrote: > On 8 September 2016 at 08:54, Laszlo Ersek <[email protected]> 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 <[email protected]> >>> --- >>> 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 <[email protected]> >> >> 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 ...
Pushed, thanks. _______________________________________________ edk2-devel mailing list [email protected] https://lists.01.org/mailman/listinfo/edk2-devel
