Re: Linux 5.13+ as Xen dom0 crashes on Ryzen CPU (ucode loading related?)

From: Juergen Gross
Date: Tue Sep 14 2021 - 04:42:05 EST


On 14.09.21 10:33, Mike Rapoport wrote:
On Tue, Sep 14, 2021 at 09:14:38AM +0200, Juergen Gross wrote:
On 13.09.21 14:50, Marek Marczykowski-Górecki wrote:
Hi,

Since 5.13, the Xen (PV) dom0 crashes on boot, before even printing the
kernel version.
Test environment:
- Xen 4.14.2
- AMD Ryzen 5 4500U (reported also on AMD Ryzen 7 4750U)
- Linux 5.13.13, confirmed also on 5.14

The crash happens only if the initramfs has earlycpio with microcode.
I don't have a serial console, but I've got a photo with crash message
(from Xen, Linux doesn't managed to print anything):
https://user-images.githubusercontent.com/726704/133084966-5038f37e-001b-4688-9f90-83d09be3dc2d.jpg

Transcription of some of it:

mapping kernel into physical memory
about to get started
(XEN) Pagetable walk from ffffffff82810888:
(XEN) L4[0x1ff] = 0000000332815067 0000000000002815
(XEN) L3[0x1fe] = 0000000332816067 0000000000002816
(XEN) L2[0x014] = 0000000334018067 0000000000004018
(XEN) L1[0x010] = 0000000332810067 0000000000002810
(XEN) domain_crash_sync called from entry.S: fault at ffff82d04033e790 x86_64/entry.S#domain_crash_page_fault
(XEN) Domain 0 (vcpu#0) crashed on cpu#0:
(XEN) ----[ Xen-4.14.2 x86_64 debug=n Not tainted ]----
(XEN) CPU: 0
(XEN) RIP: e033:[<0000000000000000>]

The domain's run state seems to be completely clobbered.

Did you try to boot the kernel with "earlyprintk=xen" to get some idea
how far it progressed?

I could imagine that doing the early reservations after the call of
e820__memory_setup() is problematic, as Xen PV guests have a hook in
this function performing some rather extended actions.

Right, among them it may relocate initrd:

https://elixir.bootlin.com/linux/latest/source/arch/x86/xen/setup.c#L872
and this may cause the reported crash.

I'm not sure the call of early_reserve_memory() can be moved just before
the e820__memory_setup() call. If this is possibel it should be done
IMO, if not then the reservations which have been at the start of
setup_arch() might need to go there again.

early_reserve_memory() can be moved to the beginning of setup_arch().

IMO this should be the preferred fix. I will write a patch to do that.

Anther possibility is to move initrd relocation out of xen_setup_memory()
and maybe even integrate it somehow in reserve_initrd().

This would be rather complicated as xen_setup_memory() is changing the
memory map from one large memory chunk to match the memory map of the
host in case the system is running as dom0 (the need to do so has
historical reasons, changing that is no option). The initrd needs to be
moved in case it is using memory which is conflicting with the new
layout.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature
Description: OpenPGP digital signature