ArmVirtPkg/ArmVirtQemu: always build XIP code with strict alignment
PATCH 1/1 ArmVirtPkg/ArmVirtQemu: always build XIP code with strict alignment
From: Ard Biesheuvel <[email protected]>
The optimization that enabled entry with MMU and caches enabled at EL1 removed the strict alignment requirement for XIP code (roughly, any code that might execute with the MMU and caches off, which means SEC and PEI phase modules but also *all* BASE libraries), on the basis that QEMU can only run guest payloads at EL2 in TCG emulation, which used to ignore alignment violations, and execution at EL1 would always occur with the MMU enabled.
This assumption no longer holds: not only does QEMU now enforce strict alignment for memory accesses with device semantics, there are also cases where this code might execute at EL2 under virtualization (i.e., under NV2 nested virtualization) where the strict alignment is required too.
The latter case could be optimized too, by enabling VHE and pretending execution is occurring at EL1, which would allow the existing logic for entry with the MMU enabled to be reused. However, this would leave non-VHE CPUs behind.
So in summary, strict alignment needs to be enforced for any code that may execute with the MMU off, so drop the override that sets the XIP flags to the empty string.
--- ArmVirtPkg/ArmVirtQemu.dsc | 2 -- 1 file changed, 2 deletions(-)
ArmVirtPkg/ArmVirtQemu.dsc +0 -1
@@ -124,8 +124,6 @@ [LibraryClasses.common.UEFI_DRIVER] [BuildOptions] !if $(CAVIUM_ERRATUM_27456) == TRUE GCC:*_*_AARCH64_PP_FLAGS = -DCAVIUM_ERRATUM_27456 -!else
- GCC:*_*_AARCH64_CC_XIPFLAGS ==
!endif
!include NetworkPkg/NetworkBuildOptions.dsc.inc
--
2.44.0.769.g3c40516874-goog
Quoted 25 lines · On Fri, 19 Apr 2024 19:50:13 +0200 Ard Biesheuvel <[email protected]> wrote:
From: Ard Biesheuvel <[email protected]>
The optimization that enabled entry with MMU and caches enabled at EL1 removed the strict alignment requirement for XIP code (roughly, any code that might execute with the MMU and caches off, which means SEC and PEI phase modules but also *all* BASE libraries), on the basis that QEMU can only run guest payloads at EL2 in TCG emulation, which used to ignore alignment violations, and execution at EL1 would always occur with the MMU enabled.
This assumption no longer holds: not only does QEMU now enforce strict alignment for memory accesses with device semantics, there are also cases where this code might execute at EL2 under virtualization (i.e., under NV2 nested virtualization) where the strict alignment is required too.
The latter case could be optimized too, by enabling VHE and pretending execution is occurring at EL1, which would allow the existing logic for entry with the MMU enabled to be reused. However, this would leave non-VHE CPUs behind.
So in summary, strict alignment needs to be enforced for any code that may execute with the MMU off, so drop the override that sets the XIP flags to the empty string.
Quoted 6 lines
Cc Jonathan Cameron <[email protected]>Cc Richard Henderson <[email protected]>Cc Philippe Mathieu-Daud� <[email protected]>Cc Idan Horowitz <[email protected]>Signed-off-by Ard Biesheuvel <[email protected]>
This matches what I was testing with locally. Thanks sorting this out.
Quoted 28 lines
--- ArmVirtPkg/ArmVirtQemu.dsc | 2 -- 1 file changed, 2 deletions(-)ArmVirtPkg/ArmVirtQemu.dsc +0
-3@@ -124,8 +124,6 @@ [LibraryClasses.common.UEFI_DRIVER] [BuildOptions] !if $(CAVIUM_ERRATUM_27456) == TRUE GCC:*_*_AARCH64_PP_FLAGS = -DCAVIUM_ERRATUM_27456 -!else - GCC:*_*_AARCH64_CC_XIPFLAGS == !endif !include NetworkPkg/NetworkBuildOptions.dsc.inc --2.44.0.769.g3c40516874-goog