Re: [edk2-devel] [PATCH v2] ArmVirt: Allow memory attributes protocol to be disabled on first boot

["Laszlo Ersek" <lersek@redhat.com>]

On 12/7/23 11:06, Ard Biesheuvel wrote:
> From: Ard Biesheuvel <ardb@kernel.org>
> 
> Shim's PE loader uses the EFI memory attributes protocol in a way that
> results in an immediate crash when invoking the loaded image, unless the
> base and size of its executable segment are both aligned to 4k.
> 
> If this is not the case, it will strip the memory allocation of its
> executable permissions, but fail to add them back for the executable
> region, resulting in non-executable code. Unfortunately, the PE loader
> does not even bother invoking the protocol in this case (as it notices
> the misalignment), making it very hard for system firmware to work
> around this by attempting to infer the intent of the caller.
> 
> So let's introduce a QEMU command line option to indicate that the
> protocol should not be exposed at all on the first boot, which is when
> the issue is triggered. (fbaa64.efi is broken but grubaa64.efi boots
> fine)
> 
>   -fw_cfg opt/org.tianocore/UninstallMemAttrProtocolOnFirstBoot,string=y
> 
> Also introduce a fixed boolean PCD that sets the default.
> 
> Cc: Laszlo Ersek <lersek@redhat.com>
> Cc: Gerd Hoffmann <kraxel@redhat.com>
> Cc: Oliver Steffen <osteffen@redhat.com>
> Cc: Alexander Graf <graf@amazon.com>
> Cc: Oliver Smith-Denny <osde@linux.microsoft.com>
> Cc: Taylor Beebe <taylor.d.beebe@gmail.com>
> Cc: Peter Jones <pjones@redhat.com>
> Cc: Leif Lindholm <quic_llindhol@quicinc.com>
> Link: https://gitlab.com/qemu-project/qemu/-/issues/1990
> Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
> ---
>  ArmVirtPkg/ArmVirtPkg.dec                                            |  6 ++
>  ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBootManagerLib.inf |  7 ++
>  ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBm.c               | 85 ++++++++++++++++++++
>  3 files changed, 98 insertions(+)
> 
> diff --git a/ArmVirtPkg/ArmVirtPkg.dec b/ArmVirtPkg/ArmVirtPkg.dec
> index 0f2d7873279f..c55978f75c19 100644
> --- a/ArmVirtPkg/ArmVirtPkg.dec
> +++ b/ArmVirtPkg/ArmVirtPkg.dec
> @@ -68,3 +68,9 @@ [PcdsFixedAtBuild, PcdsPatchableInModule]
>    # Cloud Hypervisor has no other way to pass Rsdp address to the guest except use a PCD.
> 
>    #
> 
>    gArmVirtTokenSpaceGuid.PcdCloudHvAcpiRsdpBaseAddress|0x0|UINT64|0x00000005
> 
> +
> 
> +  ##
> 
> +  # Whether the EFI memory attribus protocol should be uninstalled before
> 
> +  # invoking the OS loader on the first boot. This may be needed to work around
> 
> +  # problematic builds of shim that use the protocol incorrectly.
> 
> +  gArmVirtTokenSpaceGuid.PcdUninstallMemAttrProtocolOnFirstBoot|FALSE|BOOLEAN|0x00000006
> 

(1) could be a feature PCD (although it couldn't be patchable-in-module
then, and perhaps we don't consider this a "feature")

(2) typo: "attribus"

(3) for some reason, I see double line breaks.

> diff --git a/ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBootManagerLib.inf b/ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBootManagerLib.inf
> index 997eb1a4429f..5d119af6a3b3 100644
> --- a/ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBootManagerLib.inf
> +++ b/ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBootManagerLib.inf
> @@ -16,6 +16,7 @@ [Defines]
>    MODULE_TYPE                    = DXE_DRIVER
> 
>    VERSION_STRING                 = 1.0
> 
>    LIBRARY_CLASS                  = PlatformBootManagerLib|DXE_DRIVER
> 
> +  CONSTRUCTOR                    = PlatformBootManagerLibConstructor
> 
>  
> 
>  #
> 
>  # The following information is for reference only and not required by the build tools.
> 
> @@ -46,6 +47,7 @@ [LibraryClasses]
>    PcdLib
> 
>    PlatformBmPrintScLib
> 
>    QemuBootOrderLib
> 
> +  QemuFwCfgSimpleParserLib
> 
>    QemuLoadImageLib
> 
>    ReportStatusCodeLib
> 
>    TpmPlatformHierarchyLib
> 
> @@ -55,6 +57,7 @@ [LibraryClasses]
>    UefiRuntimeServicesTableLib
> 
>  
> 
>  [FixedPcd]
> 
> +  gArmVirtTokenSpaceGuid.PcdUninstallMemAttrProtocolOnFirstBoot
> 
>    gEfiMdePkgTokenSpaceGuid.PcdUartDefaultBaudRate
> 
>    gEfiMdePkgTokenSpaceGuid.PcdUartDefaultDataBits
> 
>    gEfiMdePkgTokenSpaceGuid.PcdUartDefaultParity
> 
> @@ -73,5 +76,9 @@ [Guids]
>  [Protocols]
> 
>    gEfiFirmwareVolume2ProtocolGuid
> 
>    gEfiGraphicsOutputProtocolGuid
> 
> +  gEfiMemoryAttributeProtocolGuid
> 
>    gEfiPciRootBridgeIoProtocolGuid
> 
>    gVirtioDeviceProtocolGuid
> 
> +
> 
> +[Depex]
> 
> +  gEfiVariableArchProtocolGuid
> 

I've made an effort to read through the v1 discussion (exhausting). Some
quetions remain:

(4) Why the change from an explicit call from AfterConsole to a
constructor? Was AfterConsole too late somehow?

I think constructors should be the last resort.

(5) Is the depex really necessary? BDS is supposed to start when all
drivers have been dispatched, and so by that time, all of the UEFI
architectural protocols should be available. (BDS will launch UEFI
drivers, and all the UEFI drivers have an implicit depex on all the
architectural protocols.)

> diff --git a/ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBm.c b/ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBm.c
> index 85c01351b09d..5306d9ea0a05 100644
> --- a/ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBm.c
> +++ b/ArmVirtPkg/Library/PlatformBootManagerLib/PlatformBm.c
> @@ -16,6 +16,7 @@
>  #include <Library/PcdLib.h>
> 
>  #include <Library/PlatformBmPrintScLib.h>
> 
>  #include <Library/QemuBootOrderLib.h>
> 
> +#include <Library/QemuFwCfgSimpleParserLib.h>
> 
>  #include <Library/TpmPlatformHierarchyLib.h>
> 
>  #include <Library/UefiBootManagerLib.h>
> 
>  #include <Protocol/DevicePath.h>
> 
> @@ -1274,3 +1275,87 @@ PlatformBootManagerUnableToBoot (
>      EfiBootManagerBoot (&BootManagerMenu);
> 
>    }
> 
>  }
> 
> +
> 
> +/**
> 
> +  Uninstall the EFI memory attribute protocol if it exists.
> 
> +**/
> 
> +STATIC
> 
> +VOID
> 
> +UninstallEfiMemoryAttributesProtocol (
> 
> +  VOID
> 
> +  )
> 
> +{
> 
> +  EFI_STATUS  Status;
> 
> +  EFI_HANDLE  Handle;
> 
> +  UINTN       Size;
> 
> +  VOID        *MemoryAttributeProtocol;
> 
> +
> 
> +  Size   = sizeof (Handle);
> 
> +  Status = gBS->LocateHandle (
> 
> +                  ByProtocol,
> 
> +                  &gEfiMemoryAttributeProtocolGuid,
> 
> +                  NULL,
> 
> +                  &Size,
> 
> +                  &Handle
> 
> +                  );
> 
> +
> 
> +  if (EFI_ERROR (Status)) {
> 
> +    ASSERT (Status == EFI_NOT_FOUND);
> 
> +    return;
> 
> +  }
> 
> +
> 
> +  Status = gBS->HandleProtocol (
> 
> +                  Handle,
> 
> +                  &gEfiMemoryAttributeProtocolGuid,
> 
> +                  &MemoryAttributeProtocol
> 
> +                  );
> 
> +  ASSERT_EFI_ERROR (Status);
> 
> +
> 
> +  Status = gBS->UninstallProtocolInterface (
> 
> +                  Handle,
> 
> +                  &gEfiMemoryAttributeProtocolGuid,
> 
> +                  MemoryAttributeProtocol
> 
> +                  );
> 
> +  ASSERT_EFI_ERROR (Status);
> 
> +}

Looks OK to me.

> 
> +
> 
> +EFI_STATUS
> 
> +EFIAPI
> 
> +PlatformBootManagerLibConstructor (
> 
> +  IN EFI_HANDLE        ImageHandle,
> 
> +  IN EFI_SYSTEM_TABLE  *SystemTable
> 
> +  )
> 
> +{
> 
> +  BOOLEAN  Uninstall;
> 
> +  UINTN    VarSize;
> 
> +  UINT32   Attr;
> 
> +
> 
> +  //
> 
> +  // Work around shim's terminally broken use of the EFI memory attributes
> 
> +  // protocol, by uninstalling it if requested on the QEMU command line.
> 
> +  //
> 
> +  // E.g.,
> 
> +  //       -fw_cfg opt/org.tianocore/UninstallMemAttrProtocolOnFirstBoot,string=y
> 
> +  //
> 
> +  // This is only needed on the first boot, when fbaa64.efi is being invoked to
> 
> +  // set the boot order variables. Subsequent boots involving GRUB are not
> 
> +  // affected.
> 
> +  //
> 
> +  VarSize = 0;
> 
> +  if (gRT->GetVariable (
> 
> +             L"BootOrder",
> 
> +             &gEfiGlobalVariableGuid,
> 
> +             &Attr,

(6) "Attr" is optional; we could / should pass NULL here.

> 
> +             &VarSize,
> 
> +             NULL
> 
> +             ) == EFI_NOT_FOUND)
> 
> +  {
> 
> +    Uninstall = FixedPcdGetBool (PcdUninstallMemAttrProtocolOnFirstBoot);
> 
> +    QemuFwCfgParseBool ("opt/org.tianocore/UninstallMemAttrProtocolOnFirstBoot", &Uninstall);
> 
> +    if (Uninstall) {
> 
> +      UninstallEfiMemoryAttributesProtocol ();
> 
> +    }
> 
> +  }
> 
> +
> 
> +  return EFI_SUCCESS;
> 
> +}
> 

(7) Tying back to my point (4) -- I understand this is a hack anyway,
but I'm still uncomfortable with platform BDS uninstalling a protocol
that is owned by / provided by the CPU driver. Feels like a significant
layering violation.

Can we modify the CPU driver instead, to listen to a new event group,
upon which being signaled, the CPU driver would uninstall the protocol
(and close the listening event)?

This PlatformBootManagerLib instance would act more or less the same
(I'd suggest signaling the event group from within AfterConsole, in case
the PCD default and/or the fw_cfg knob dictated that), but the protocol
uninstallation would occur in "ArmPkg/Drivers/CpuDxe".

In more technical terms, the layering violation IMO is that we mess with
CpuDxe's "mCpuHandle" and "mMemoryAttribute" static variables from
within BDS. Adding the new event group requires more boiler-plate code
for sure, but there's a small code-size benefit as well: we'd not have
to look up either the handle (with LocateHandle) or the protocol
interface (with HandleProtocol), as CpuDxe inherently knows those
(mCpuHandle, mMemoryAttribute).

Thanks for considering (and sorry for butting in this late...)
Laszlo



-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#112233): https://edk2.groups.io/g/devel/message/112233
Mute This Topic: https://groups.io/mt/103031504/7686176
Group Owner: devel+owner@edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/leave/12367111/7686176/1913456212/xyzzy [rebecca@openfw.io]
-=-=-=-=-=-=-=-=-=-=-=-