toggle quoted messageShow quoted text
From: gaoliming <email@example.com>
Sent: Monday, September 28, 2020 8:40 AM
To: firstname.lastname@example.org; Jeff Brasen <email@example.com>;
Subject: 回复: [edk2-discuss] GCD initialization and memory allocation HOBs
External email: Use caution opening links or attachments
-----邮件原件-----You mean the memory range specified by PHIT resource hob is selected by
<firstname.lastname@example.org> 代表 Jeff Brasen
发送时间: 2020年9月26日 1:08
收件人: email@example.com; firstname.lastname@example.org
主题: Re: [edk2-discuss] GCD initialization and memory allocation HOBs
#2 is the case we are in, we have data that is loaded by our pre-UEFI
bootloaders into memory that should behave as boot services memory (OS
can reclaim this) When we setup our HOB list we added a normal system
memory resource that covers both the HOB list and this other data. We
then add a memory allocation HOB to mark this memory as boot services
data. We are seeing cases where CoreInitializeMemoryServices Is
selecting the region that has this data and is both causing corruption
as well as preventing the memory allocation hob from being processed
correctly as the memory is not free when GCD parses it.
This memory range includes PEI memory and the allocated memory. So,
there is no free memory in this memory range. After
CoreInitializeMemoryServices(), do you meet with memory allocation
[JMB] So what we have is a system resource that includes both the PHIT memory area (From EfiMemoryBottom to EfiMemoryTop) as well as other data that is boot services memory that is marked by a memory allocation hob entry but is outside the PHIT region. In our case we see CoreInitializeMemoryServices selecting this region as it doesn't look for memory allocations in the case that it selections a region outside the PHIT region. We can split the system resource so that the memory isn't in question isn't in the same system resource which solves the issue for now (as we have enough memory in the EfiFreeMemoryBottom to EfiFreeMemoryTop region. However, if there isn't enough memory in this region (and assuming we don't have any memory before EfiMemoryBottom due to the same region) it will just go search other system resources which should have the same issue.
Note we don't build this with PEI as we don't use that phase as we just skip from SEC to DXE as our memory is already setup so there is less value to that phase.
From my reading of the PI spec there is no issue with have memory allocation entries that are outside phit areas so what we are setting up seems to be compliant with the spec.
I would suggest one of the two changes for this
1. Have CoreInitializeMemoryServices Handle memory allocation hobs
2. Update either the PI spec or have an EDK2 rule that makes sure that HOB lists are created in such a we aren't coding setup of the hob producer to the implementation of the dxe core.
Can you attach the failure boot log for further analysis?
[JMB] I don't think we actually see any logs related to this but it was causing our memory it be not marked as allocated which was causing some corruption, we can look at adding some prints to this case if this is useful.
We can easily work around this particular issue by splitting theimplementation of GCD services.
system memory resource so that GCD then goes to look for space in the
free region in the hob or at the beginning of the resource that
contains the HOB list. But if we don't have enough space in those
regions then it will look for other memory regions and it doesn't seem
like the best practice to design the HOB producer with the specific
It seems that if we have memory allocation data that is in outside the
HOB list itself we could hit this problem, and this seems valid per my
reading of the PI spec.
From: email@example.com <firstname.lastname@example.org> On Behalf Of
Sent: Friday, September 25, 2020 3:06 AM
To: email@example.com; Jeff Brasen <firstname.lastname@example.org>
Cc: Jan Bobek <email@example.com>; Ashish Singhal
Subject: Re: [edk2-discuss] GCD initialization and memory allocation
External email: Use caution opening links or attachments
On 09/24/20 20:40, Jeff Brasen wrote:
It looks like CoreInitializeMemoryServices
ec699b2c17/MdeModulePkg/Core/Dxe/Gcd/Gcd.c#L2116)Do you mean
does not honor any memory allocations outside of the HOB list itself.
Should we add code to this function to make sure the region this
function select does not end up over a reserved region that the HOB
producer phase marked as already allocated? Either that or codify
the specific requirements a bit more on to make sure cases like this
are invalid. (producer might have to split up system resource
(1) a physically reserved memory address range
(2) normal RAM (EFI_RESOURCE_SYSTEM_MEMORY) that is meant to be
"repurposed" (= pre-allocated) in PEI as a particular UEFI memory type?
EFI_HOB_RESOURCE_DESCRIPTOR "does not describe how memory is used
instead describes the attributes of the physical memory present".EFI_RESOURCE_MEMORY_RESERVED
CoreInitializeMemoryServices() honors (1).
CoreInitializeMemoryServices() does not honor (2); however, the later
function CoreInitializeGcdServices() does honor (2).
* Regarding (1):
My understanding is that the HOB producer phase should describe
"physically reserved" memory regions with
resource descriptor HOBs, and these should not overlap normal memorybranch
regions, which are described with EFI_RESOURCE_SYSTEM_MEMORY
Furthermore, whether (a part of) a memory resource is allocated or
*orthogonal* to whether the memory resource is "physically reserved
memory" or "normal RAM". In my opinion, the diagram at
Volume 2 (DXE Core Interface)
7.2.2 GCD Memory Resources
Figure 2. GCD Memory State Transitions
So the HOB producer phase should firstly describe the area in question
as EFI_RESOURCE_MEMORY_RESERVED (without overlapping any other
resource descriptor in memory address space).
Secondly, if the HOB producer phase wants to prevent DXE modules from
allocating reserved memory out of this resource, then the HOB producer
phase should *also* cover the entire range with a memalloc HOB that
The first step above (= describing the area with an
EFI_RESOURCE_MEMORY_RESERVED HOB) will activate the following
// Skip Resource Descriptor HOBs that do not describe tested
ResourceHob = Hob.ResourceDescriptor;
if (ResourceHob->ResourceType !=
* Regarding (2):
Normal (= not physically reserved) RAM that is allocated in PEI --
using memalloc HOBs with various UEFI memory types, such as BS Data or
-- is set aside in the CoreInitializeGcdServices() function.
Is that too late for your purpose? If so, why?
AIUI, CoreInitializeGcdServices() does make sure that DXE-phase calls
gDS->AllocateMemorySpace(), gBS->AllocatePool(), and friends, will not
trample over the original memalloc HOBs.