Re: OVMF.fd and placement of EfiBootServicesData

[Laszlo Ersek <lersek@redhat.com>]

On 10/04/16 04:45, spam collector wrote:
> 
> 
> ----- Original Message -----
>> From: "Laszlo Ersek" <lersek@redhat.com>
>> To: "spam collector" <spamcollector@cableone.net>
>> Cc: edk2-devel@ml01.01.org
>> Sent: Monday, October 3, 2016 2:42:08 AM
>> Subject: Re: [edk2] OVMF.fd and placement of EfiBootServicesData
>>
>> On 10/03/16 05:16, spam collector wrote:
>>> Hello,
>>> Forgive me for not searching for this first, but gmane.org has
>>> been down for a little while. Therefore, forgive me if this
>>> question has been asked and answered.
>>> I am using OVMF.fd and QEMU for Windows and tried to load
>>> a large file to physical address 0x00800000. When an error
>>> was returned, I found that OVMF has reserved an amount of
>>> EfiBootServicesData within/around that location.
>>> Here is a memory dump using the BootServices MemoryMap Service:
>>> Start: 0x00000000->0x0009FFFF, Pgs: 160, EfiConventionalMem
>>> Start: 0x00100000->0x003FFFFF, Pgs: 768, EfiConventionalMem
>>> Start: 0x00400000->0x0040BFFF, Pgs: 12, EfiBootServicesCode
>>> Start: 0x0040C000->0x0081FFFF, Pgs: 1044, EfiConventionalMem
>>> Start: 0x00820000->0x00FFFFFF, Pgs: 2016, EfiBootServicesData
>>> Start: 0x01000000->0x0BFFFFFF, Pgs: 1044, EfiConventionalMem
>>> ... and so on
>>> Without the idea of "you should make your code relocatable,
>>> i.e.: not care where it is in memory", or if I say that I *must* have
>>> the memory from 0x00800000 to 0x00FFFFFFavailable,
>>> without re-building OVMF.fd, is there a way to tell the system
>>> to use a different address for that portion of EfiBootServiceCode?
>>> I tried manipulating NvVars with a few entries to no avail.
>>> I added the "LoadFixedAddressConfigurationTable" entry hoping
>>> that I could set a minimum location, too without success.
>>> So, without rebuilding OVMF.fd, which would require everyone
>>> whom wished to use my code to do so, or at least download it
>>> from me, and I don't like modified packages of other's work
>>> floating around, is there a why to tell OVMF.fd/the EFI system
>>> to not use memory below a certain address?
>>> Thank you in advance for your suggestions,
>>
>> Sorry, I don't think I can help.
>>
>> You can read about the OVMF memory map in the (now somewhat outdated)
>> OVMF whitepaper at
>>
>>   http://www.linux-kvm.org/downloads/lersek/ovmf-whitepaper-c770f8c.txt
>>
>> section
>>
>>   A comprehensive memory map of OVMF
>>
>> You'd like to place your stuff at 8192 KB .. 16384 KB, but as you can
>> see from the map, that exact range is heavily used by OVMF itself.
>>
>> (As I said, some of those exact values are no longer current, due to the
>> following commits since:
>>
>>   08df58ec3043 OvmfPkg: raise DXEFV size to 9 MB
>>   2f7b34b20842 OvmfPkg: raise DXEFV size to 10 MB
>>   45d870815156 OvmfPkg/PlatformPei: rebase and resize the permanent PEI
>>                memory for S3
>>   6b04cca4d697 OvmfPkg: remove PcdS3AcpiReservedMemoryBase,
>>                PcdS3AcpiReservedMemorySize
>>
>> but the main layout remains the same.)
>>
>> Please consider making your code relocatable.
> 
> Without sounding rude, and please do not take it as so, making my
> code relocatable due to the fact that the other code is not relocatable,
> sounds a little hypocritical. :-)

Not taking it as rude, but you are wrong about this. OVMF is a platform
firmware, including SEC and PEI phase modules. Such modules are fully
expected to operate with a fixed address map.

In comparison your module is (I assume) a UEFI_DRIVER /
UEFI_APPLICATION, or maybe a DXE_DRIVER.
- If it's a UEFI_DRIVER or UEFI_APPLICATION, then it is expected to work
without hardcoded addresses.
- If it is a DXE_DRIVER, then it might be okay for it to operate with
fixed addresses. However, a DXE_DRIVER module is meant to be shipped by
the platform vendor. For that, your module would have to be integrated
with OVMF -- it should either be part of the public edk2 tree (or your
edk2 tree), or you should be providing a prebuilt OVMF binary that
contained it.

> Anyway, please don't take offense to that, it is just what first
> came to mind when I read it.  Eventually I will make my code relocatable
> but at the moment that would be quite a re-write of many other parts
> of this project, and I am not ready or willing to do that yet. Yet...
> 
> May I please suggest to the maintainers of this project to add a function
> that will allow a NvVars entry to indicate where the base of this 
> "comprehensive memory map of OVMF" starts.  Allow an entry within the 
> NvVars file to "suggest" a base address and let the code accommodate 
> as close as it can.

(1) You mention the NvVars file as storage for non-volatile variables.
If you are using OVMF with QEMU, I strongly suggest to abandon this
setup, and adopt the pflash based variable store instead. The behavior
of the NvVars file is not fully compatible with the UEFI spec, and it
had been invented as a stop-gap solution until real variable storage
would be implemented.

(2) The addresses that you seem interested in are build-time constants,
and even reset vector and SEC phase code depend on them. In those
phases, variable services don't exist.

If you want to move OVMF's fixed memory map to a different spot, for
your own builds, please grep OvmfPkg for the PCDs that are listed to the
right of the table that is in section "A comprehensive memory map of
OVMF" of the whitepaper, and customize them as you see fit. (And then
rebuild OVMF, of course.)

We have sought to make this as painless as possible -- for our own
sanity's sake too --; that is, most of those PCDs are already being
computed from other (more basic) PCDs in the same set, either at build
time or at runtime. It's just that the "root" PCDs have to be build-time
constants. For example, please refer to the MEMFD_BASE_ADDRESS macro in
"OvmfPkg.fdf.inc" and elsewhere.

> Or even just a flag that states to use higher
> area memory when set, or lower area memory when clear.  Some OS writers
> like to use high memory and leave low memory available while others
> like to use low memory and leave the remaining higher memory available.

If you want to move OVMF's fixed areas to different places, specific
addresses are necessary. Please refer to the PCDs mentioned above.

The current map works with at least Linux guests, Windows guests, OSX
guests (I've heard from users), and FreeBSD guests (based on their
Wiki). I'm quite sure the address constraints currently encoded in your
module (a boot loader perhaps?) could break on suitable physical
firmware as well; OVMF does nothing invalid in this regard.

Laszlo