Hi Steve,

On 04/11/18 12:30, Steve Capper wrote:
> Hello,
> I am trying to debug a boot problem for a recent build of EDK2 in an
> Arm FVP model.
> I built EDK2 from the following branches:
> Edk2: a146c532c754106431b063fec9985a838afd82be
> Edk2-platforms: e74f53df8b18e4aed0c9df0942ce0c30f78a68e2
> Toolchain: Debian Stretch latest (though I have also tried Linaros
> GCC 5 latest).
> Build command:
> GCC5_AARCH64_PREFIX=aarch64-linux-gnu- build -a AARCH64 -t GCC5 -b
> DEBUG -p Platform/ARM/VExpressPkg/ArmVExpress-FVP-AArch64.dsc
> The boot log can be found in the attached file. I *think* this is due
> to the Cpu Arch protocol not actually being ready when
> CoreConvertSpace is called by the SetMemoryAttributes from the
> NorFlash code.
> I have found that this code path fails:
> https://github.com/tianocore/edk2/blob/a146c532c754106431b063fec9985a838afd82be/MdeModulePkg/Core/Dxe/Gcd/Gcd.c#L893
> The only modification I have made is to edk2-platforms to get more
> debug output (PcdDebugPrintErrorLevel).
> Unfortunately, I'm not sure how to fix this?

This is a bit of a problem.

The gDS->SetMemorySpaceAttributes() DXE service is allowed to fail if
the CPU architectural protocol has not been installed yet. This is
documented in the Platform Init spec (although riddled by a few typos --
EFI_NOT_AVAILABLE_YET is documented under the preceding
GetMemorySpaceDescriptor() section, and not in the correct
SetMemorySpaceAttributes() section).

The usual solution for DXE drivers is to add an explicit dependency on
gEfiCpuArchProtocolGuid to the DEPEX section in the INF file. You can
see an example in commit f9a8be423cdd
("ArmVirtualizationPkg/PciHostBridgeDxe: MMIO aperture must not be
uncached", 2015-02-23).

And that's where the issue is. If you check the [Depex] section in
"ArmPlatformPkg/Drivers/NorFlashDxe/NorFlashDxe.inf", you find

> [Depex]
>   #
>   # NorFlashDxe must be loaded before VariableRuntimeDxe in case empty flash 
> needs populating with default values
>   #
>   BEFORE gVariableRuntimeDxeFileGuid

Why is BEFORE technically an issue here? Because, it cannot be combined
with other depex opcodes. From the PI spec:

> If this opcode is present in a dependency expression, it must be the
> first and only opcode in the expression. If is appears in any other
> location in the dependency expression, then the dependency expression
> is evaluated to FALSE.

So we can't simply combine the existent BEFORE depex with "AND
gEfiCpuArchProtocolGuid", like you see in the commit f9a8be423cdd
example above.

Now, more to the point, a "BEFORE" depex is *always* a hack. From the PI

> This opcode tells the DXE Dispatcher that the DXE driver that is
> associated with this dependency expression must be dispatched just
> before the DXE driver with the file name specified by GUID. This means
> that as soon as the dependency expression for the DXE driver specified
> by GUID evaluates to TRUE, then this DXE driver must be placed in the
> dispatch queue just before the DXE driver with the file name specified
> by GUID.

Sometimes there's really no way around such a hack, to express an
inter-module dependency, but even then, there is a better way to
implement it. Namely,
- a custom protocol GUID can be invented,
- the pre-requisite module can install a NULL protocol interface with
  that GUID,
- and
  - either the protocol GUID can be added directly to the [depex]
    section of the dependent module (if the latter module is under the
    platform's control),
  - or the platform DSC file can hook a NULL-class library instance into
    the dependent module, such that the lib instance makes the dependent
    module *inherit* a [depex] on the protocol GUID.

In other words, if we want to order the dispatch of a
platform-independent module (that lives in another Package), due to
platform-specific circumstances, behind another module that our platform
owns, then we can change our platform DSC to "imbue" the module with a
dependency on a custom protocol GUID. We can then install that NULL
protocol in our own module. Under this scheme the usual dependency
resolution will order things the right way, and all the depex-es remain
composable with other GUIDs.

OK -- so *should* we rewrite this BEFORE depex, like described above? I
don't think so. In my opinion, the *idea* of this dependency is wrong.
The BEFORE depex was introduced in commit 6acb379fbcf8
("ArmPlatformPkg/CTA9x4: Remove Variable Storage FD file from FDF",
2011-03-31), and the idea there was to replace the build-time
pre-formatting of the flash variable store with runtime formatting. You
might ask why this replacement needed an explicit ordering hint between
drivers (i.e. why any DEPEXes had to be touched at all). Here's why:

- For *writing* non-volatile UEFI variables, a platform-independent
  protocol dependency chain already exists. At the least abstract level,
  we have the FVB (firmware volume block) protocol service, e.g.
  NorFlashDxe. At the next level, we have the platform-independent FTW
  (fault tolerant write) driver. Finally, at the top (the most abstract
  level), we have the non-volatile variable write service / driver.

- However, for *reading* non-volatile UEFI variables, no such dependency
  chain exists. The variable driver (the top level from the previous
  bullet) will just go ahead and read various headers from the flash
  chip. If those headers are not correctly formatted, the variable
  driver blows up and the platform will not boot. To solve this issue,
  platform FDF files usually pre-format a variable store template at
  build-time, which will at once satisfy the "read-side startup" of the
  variable driver.

Unfortunately, this requires FDF files to contain various ugly hexdumps,
and people are tempted to replace those with runtime formatting of the
varstore headers. In turn though, they have to employ hacks for ordering
the variable driver behind their runtime formatting. Hence the BEFORE
depex. :(

So, what can we do? Given that NorFlashDxe is used by several platforms,
the ugly BEFORE depex should be removed. Once BEFORE is removed, you can
simply add "gEfiCpuArchProtocolGuid", and then the current issue will be
solved. But, what shall take the place of BEFORE?

Using ArmVirtQemu as an example platform, the answer is "nothing",
because ArmVirtQemu already pre-formats the varstore template -- please
see "ArmVirtPkg/VarStore.fdf.inc". For all other platforms that don't do
this -- and I don't see
"Platform/ARM/VExpressPkg/ArmVExpress-FVP-AArch64.fdf" doing it --,
there are two choices:

- build a varstore template in *all* of the platform FDF files where the
  platform DSC file includes the NorFlashDxe driver -- in this case, the
  BEFORE needs no replacement, it can simply be dropped,

- or implement the custom NULL protocol trick that I described above.
  Platforms that don't actually *depend* on this trick, such as
  ArmVirtQemu, will simply ignore the new protocol instance in the
  protocol database.

For the second option, there are several examples under OvmfPkg and
ArmVirtPkg. One is:

- "MdeModulePkg/Include/Guid/PlatformHasAcpi.h" defines the custom

- "ArmVirtPkg/ArmVirtQemu.dsc" includes the
  "ArmVirtPkg/PlatformHasAcpiDtDxe" platform driver, which installs a
  NULL protocol interface with the above custom GUID under certain

- "EmbeddedPkg/Library/PlatformHasAcpiLib" is a library instance that
  does nothing; it just has an empty constructor function, and a
  dependency on the GUID.

- "ArmVirtPkg/ArmVirt.dsc.inc" hooks the library into
  "MdeModulePkg/Universal/Acpi/AcpiTableDxe", imbuing the latter with a
  dependency on the protocol GUID.

As a result, "MdeModulePkg/Universal/Acpi/AcpiTableDxe" will not be
dispatched until (and, unless!) "ArmVirtPkg/PlatformHasAcpiDtDxe"
installs the GUID. If the GUID is never installed, then
"MdeModulePkg/Universal/Acpi/AcpiTableDxe" will never be dispatched.

(1) introduce a custom GUID for "NorFlashDxe has formatted the variable
    store headers for the variable driver to read";
(2) install the GUID in NorFlashDxe once it's done verifying and/or
    formatting the headers;
(3) introduce a custom library instance with an empty constructor
    function, and a DEPEX on the GUID;
(4) hook the library instance into
    in all the platform DSC files where the platform build does not
    pre-format a varstore template in the FDF file;
(5) replace the BEFORE depex with gEfiCpuArchProtocolGuid.

edk2-devel mailing list

Reply via email to