Re: [Bug Report] Linux v6.15-rc7 boot failure on Xen-4.17

[Juergen Gross <jgross@suse.com>]

[-- Attachment #1.1.1: Type: text/plain, Size: 3013 bytes --]

On 28.05.25 10:57, Jürgen Groß wrote:
> On 28.05.25 10:26, Xin Li wrote:
>> On 5/28/2025 12:27 AM, Xin Li wrote:
>>> On 5/27/2025 11:49 PM, Juergen Gross wrote:
>>>> On 28.05.25 07:11, Jürgen Groß wrote:
>>>>> On 27.05.25 21:29, Andrew Cooper wrote:
>>>>>> On 27/05/2025 8:21 pm, Xin Li wrote:
>>>>>>>> On May 27, 2025, at 11:36 AM, Jürgen Groß <jgross@suse.com> wrote:
>>>>>>>>
>>>>>>>> On 27.05.25 19:54, Xin Li wrote:
>>>>>>>>> On 5/27/2025 10:46 AM, Pawan Gupta wrote:
>>>>>>>>>>> Attached is the serial console log and my kernel config.
>>>>>>>>>> Serial logs aren't telling much. I do not have a Xen setup to test, 
>>>>>>>>>> without
>>>>>>>>>> Xen the config that you provided is booting a KVM guest just fine.
>>>>>>>>> Yeah, as I replied to Juergen, the same kernel binary boots fine as
>>>>>>>>> "native".
>>>>>>>>> Unfortunately when booting as dom0 on Xen, it keeps rebooting w/o
>>>>>>>>> helpful log.
>>>>>>>> What about booting Xen on bare metal, i.e. no KVM being involved?
>>>>>>> The same exact problem happens on Intel Simics.  And I got to see it’s a 
>>>>>>> NX page fault in dom0 kernel during apply alternatives.
>>>>>>
>>>>>> In which case it's likely that there's an opencoded PTE update, rather
>>>>>> than using the hooks (which are suitably paravirt'd).
>>>>>
>>>>> I'd suspect a bug when NOT using 2M pages for execmem.
>>>>>
>>>>> I'll have a look.
>>>>
>>>> Could you have a try using "nohugevmalloc" dom0 kernel boot parameter?
>>>>
>>>
>>> Tried in a KVM guest, still the same problem, and nothing new in the
>>> serial log.
>>
>> Attached is a dom0 log with stack traces.
>>
>> But I really did NOT change anything to make it happen...
> 
> Thanks.
> 
> I think this might be related to Xen not advertising X86_FEATURE_PSE.
> 
> This will use PAGE_KERNEL page protection for execmem_alloc() page protection,
> while with X86_FEATURE_PSE PAGE_KERNEL_ROX is being used.
> 
> For the kernel (so not in a module) there is no execmem_restore_rox() call
> involved, so the NX bit will be kept for kernel side ITS thunks.
> 
> Peter, can you confirm my suspicion?

I just made a small test on my (rather old) system:

I verified that kernel 6.15 is booting fine as Xen dom0 (ITS mitigation
not needed due to old cpu). Then I modified alternative.c to apply the
ITS mitigations nevertheless, which made the kernel crash as Xen dom0.

With the following additional modification boot was working again:

diff --git a/arch/x86/mm/init.c b/arch/x86/mm/init.c
index bfa444a7dbb0..fac4f9d26132 100644
--- a/arch/x86/mm/init.c
+++ b/arch/x86/mm/init.c
@@ -1090,7 +1090,7 @@ struct execmem_info __init *execmem_arch_setup(void)
                 pgprot = PAGE_KERNEL_ROX;
                 flags = EXECMEM_KASAN_SHADOW | EXECMEM_ROX_CACHE;
         } else {
-               pgprot = PAGE_KERNEL;
+               pgprot = PAGE_KERNEL_EXEC;
                 flags = EXECMEM_KASAN_SHADOW;
         }


Juergen

[-- Attachment #1.1.2: OpenPGP public key --]
[-- Type: application/pgp-keys, Size: 3743 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]