--

Hi all

I did fast experiment already and it looks like this Itanium data structure removal solves my issue. Later this week I should have access to machine with 200 physical CPU so I will be able to run test against 1600 vCPU using Qemu.

To reproduce test 3 things are needed:
1) Change in the edk2: Removal of Itanium data structure in MdePkg/Include/Ppi/SecPlatformInformation.

2) Change in Qemu: Default limit is 288 / 255 in
hw/i386/pc_q35.c and hw/i386/pc.c

3) Change in KVM: Default limit is 1024 in
arch/x86/include/asm/kvm_host.h

At the end I am using Qemu to run modified firmware:

$QEMU \
-accel kvm \
-m 4G -M q35,kernel-irqchip=on,smm=on \
-smp cpus=1024,maxcpus=1024 \
-global mch.extended-tseg-mbytes=128 \
\
-drive if=pflash,format=raw,file=${CODE},readonly=on \
-drive if=pflash,format=raw,file=${VARS} \
\
-chardev stdio,id=fwlog \
-device isa-debugcon,iobase=0x402,chardev=fwlog \
\
"$@"

W dniu 15.11.2022 o 01:29, Kinney, Michael D pisze:

Hi Pedro,
After Pawel runs an experiment, we can think about how to address.
1. Code First process for spec change to remove ItaniumHealthFlags field
2. Code first process for spec change for a new GUID HIB value that
does not have ItaniumHealthFlags field
3. Code First process to allow multiple instances of the GUIDed HOB.

After checking the code it looks like this ITANIUM_HANDOFF_STATUS is only referenced in this header file (MdePkg/Include/Ppi/SecPlatformInformation.h) and nowhere else.

I would like to prepare pull request for the mailing list with this data structure removed. From what I saw the whole Itanium support has been already removed back in 2019.

You mentioned code first approach, so can I just remove unused part and submit a patch? Or should I introduce intermediate data structure with new GUID to not break any compatibility, and then at some point in the future remove old one?

Please advise

Best,
Pawel

I prefer option (1) or (2) because is reduces the temp RAM usage in HOBs as the number of IA32/X64 CPUs increase.
(3) may be required to scale above 64KB HOB size limit even with the reduced per CPU size.
Mike
*From:*devel@edk2.groups.io <devel@edk2.groups.io> *On Behalf Of *Pedro Falcato
*Sent:* Monday, November 14, 2022 3:30 PM
*To:* devel@edk2.groups.io; Kinney, Michael D <michael.d.kinney@...>
*Cc:* Pawel Polawski <ppolawsk@...>; Dong, Eric <eric.dong@...>; Laszlo Ersek <lersek@...>; Ni, Ray <ray.ni@...>; Kumar, Rahul R <rahul.r.kumar@...>
*Subject:* Re: [edk2-devel] 1024 VCPU limitation
On Mon, Nov 7, 2022 at 5:28 PM Michael D Kinney <michael.d.kinney@... <mailto:michael.d.kinney@...>> wrote:
Hi Pawel,
I see the following union involved in the size of this structure.
typedefunion{
  IA32_HANDOFF_STATUS       IA32HealthFlags;
  X64_HANDOFF_STATUS        x64HealthFlags;
  ITANIUM_HANDOFF_STATUS ItaniumHealthFlags;
} EFI_SEC_PLATFORM_INFORMATION_RECORD;
IA32 is 4 bytes per CPU
X64 is 4 bytes per CPU
Itanium is 56 bytes per CPU
We have removed the Itanium content from edk2 repo and it look like
we missed this
union.
Hi Mike,
I just want to note that I don't think you can remove ITANIUM_HANDOFF_STATUS (in an upstreamable way at least) since it's specified in the EFI PI spec, and it would also break any sort of ABI.
Maybe once you update the spec? Or maybe we could find a way to pass these handoff statuses in multiple HOBs, or have v2 HOBs with UINT32 lengths (with an appropriate spec update).
If you comment out the following line from the union does it resolve
the issue?
https://github.com/tianocore/edk2/blob/7c0ad2c33810ead45b7919f8f8d0e282dae52e71/MdePkg/Include/Ppi/SecPlatformInformation.h#L137 <https://github.com/tianocore/edk2/blob/7c0ad2c33810ead45b7919f8f8d0e282dae52e71/MdePkg/Include/Ppi/SecPlatformInformation.h#L137>
I know this only increases the total number of CPUs that can be
handled by a single 64kb HOB, so we would run into
it again at a higher number of CPUs.  However, I think this gets the
overhead per CPU down to 8 bytes, which should
scale to about 8091 CPUs.
Thanks,
Mike
*From:* Pawel Polawski <ppolawsk@...
<mailto:ppolawsk@...>>
*Sent:* Monday, November 7, 2022 3:52 AM
*To:* devel@edk2.groups.io <mailto:devel@edk2.groups.io>
*Cc:* Dong, Eric <eric.dong@... <mailto:eric.dong@...>>;
Laszlo Ersek <lersek@... <mailto:lersek@...>>; Ni, Ray
<ray.ni@... <mailto:ray.ni@...>>; Kumar, Rahul R
<rahul.r.kumar@... <mailto:rahul.r.kumar@...>>; Kinney,
Michael D <michael.d.kinney@...
<mailto:michael.d.kinney@...>>
*Subject:* [edk2-devel] 1024 VCPU limitation
Hi All,
I am trying to run edk2 with more than 1024 VCPU. It looks like it
is not possible
at the moment and results in an ASSERT trigger.
In the past the topic has been analyzed by Laszlo Ersek [1]. It
turns out that the limit
is result of HOB default allocation being limited to ~64KB, quoting
original email thread:
"""
If "NumberOfProcessors" is large enough, such as ~1024, then
"BistInformationSize" will exceed ~64KB, and PeiServicesAllocatePool()
will fail with EFI_OUT_OF_RESOURCES. The reason is that pool allocations
in PEI are implemented with memory alloaction HOBs, and HOBs can't be
larger than ~64KB. (See PeiAllocatePool() in
"MdeModulePkg/Core/Pei/Memory/MemoryServices.c".)
"""
Even with HOB allocation being changed, I am afraid it may break some
compatibility on the DXE level. This is the reason I am looking for
a more universal solution.
I believe the same limitation exists for the physical x86 platforms
with more than 1024 CPU.
If someone has encountered the same issue or has knowledge that
workaround / solution for
this already exists or is being developed?
[1]
https://listman.redhat.com/archives/edk2-devel-archive/2021-June/msg01493 <https://listman.redhat.com/archives/edk2-devel-archive/2021-June/msg01493>
Best regards,
Pawel
--
*Paweł Poławski*
Red Hat <https://www.redhat.com/> Virtualization
ppolawsk@... <mailto:ppolawsk@...>
@RedHat <https://twitter.com/redhat> Red Hat
<https://www.linkedin.com/company/red-hat> Red Hat
<https://www.facebook.com/RedHatInc>
<https://red.ht/sig>
--
Pedro Falcato