Issue With real-time patches on 5.15.y

Issue With real-time patches on 5.15.y

[Joseph Salisbury]

Hello,

I ran into a build failure after applying the rt40 real-time patches[0] 
to an Ubuntu 22.04 kernel. The Ubuntu kernel is based on upstream stable 
5.15.y(5.15.34 currently).  I believe this patch is the specific one 
that introduces the build issue: 
0086-locking-Allow-to-include-asm-spinlock_types.h-from-l.patch.

After a bisect, the following commit was identified as the cause for the 
build failure:
b50854eca0e0 ("x86/pkru: Remove useless include")

This commit was added to mainline as of 5.16-rc1 and not Cc'd to stable, 
so it should not exist in 5.15.y.  However, it was purposely added to 
the Ubuntu kernel to enable x86 AMX support[1].

The real-time kernel build failure can be resolved by reverting commit 
b50854eca0e0.  The failure seems to be due to the removal of an include 
of xstate.h from pkru.h and caused spinlock_t to not be defined.  The 
commit would only be reverted for the real-time kernel and not any other 
kernels.  I wanted to see if reverting the commit is the proper 
approach, or if cherry-picking additional commits might be a better 
solution in preparation for additional changes that might be coming in 
the future?

Any suggestions would be greatly appreciated.

Thanks,

Joe


[0] 
https://cdn.kernel.org/pub/linux/kernel/projects/rt/5.15/older/patches-5.15.34-rt40.tar.gz
[1] https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1967750

Re: Issue With real-time patches on 5.15.y

[Joseph Salisbury]

On 5/5/22 19:18, Joseph Salisbury wrote:
> Hello,
>
> I ran into a build failure after applying the rt40 real-time 
> patches[0] to an Ubuntu 22.04 kernel. The Ubuntu kernel is based on 
> upstream stable 5.15.y(5.15.34 currently).  I believe this patch is 
> the specific one that introduces the build issue: 
> 0086-locking-Allow-to-include-asm-spinlock_types.h-from-l.patch.
>
> After a bisect, the following commit was identified as the cause for 
> the build failure:
> b50854eca0e0 ("x86/pkru: Remove useless include")
>
> This commit was added to mainline as of 5.16-rc1 and not Cc'd to 
> stable, so it should not exist in 5.15.y.  However, it was purposely 
> added to the Ubuntu kernel to enable x86 AMX support[1].
>
> The real-time kernel build failure can be resolved by reverting commit 
> b50854eca0e0.  The failure seems to be due to the removal of an 
> include of xstate.h from pkru.h and caused spinlock_t to not be 
> defined.  The commit would only be reverted for the real-time kernel 
> and not any other kernels.  I wanted to see if reverting the commit is 
> the proper approach, or if cherry-picking additional commits might be 
> a better solution in preparation for additional changes that might be 
> coming in the future?
>
> Any suggestions would be greatly appreciated.
>
> Thanks,
>
> Joe
>
>
> [0] 
> https://cdn.kernel.org/pub/linux/kernel/projects/rt/5.15/older/patches-5.15.34-rt40.tar.gz
> [1] https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1967750
Also, the following is the build failure that happens:
arch/x86/include/asm/pgtable.h:55:8: error: unknown type name 'spinlock_t'
    55 | extern spinlock_t pgd_lock;
       |        ^~~~~~~~~~

Re: Issue With real-time patches on 5.15.y

[Dave Hansen]

On 5/5/22 16:18, Joseph Salisbury wrote:
> The real-time kernel build failure can be resolved by reverting commit
> b50854eca0e0.  The failure seems to be due to the removal of an include
> of xstate.h from pkru.h and caused spinlock_t to not be defined.  The
> commit would only be reverted for the real-time kernel and not any other
> kernels.  I wanted to see if reverting the commit is the proper
> approach, or if cherry-picking additional commits might be a better
> solution in preparation for additional changes that might be coming in
> the future?
> 
> Any suggestions would be greatly appreciated.

I thought you distro folks were the franenkernel experts. :)

But, seriously...  This isn't rocket science.  Just look at the pkru.h
in your tree and figure out what includes it needs.  If it needs FPU
gunk, include the FPU header.  If you get a compile error for spinlock_t
...  well ...  find the code using spinlock_t and make sure it includes
spinlock.h.

Re: Issue With real-time patches on 5.15.y

[Joseph Salisbury]

On 5/5/22 19:35, Dave Hansen wrote:
> On 5/5/22 16:18, Joseph Salisbury wrote:
>> The real-time kernel build failure can be resolved by reverting commit
>> b50854eca0e0.  The failure seems to be due to the removal of an include
>> of xstate.h from pkru.h and caused spinlock_t to not be defined.  The
>> commit would only be reverted for the real-time kernel and not any other
>> kernels.  I wanted to see if reverting the commit is the proper
>> approach, or if cherry-picking additional commits might be a better
>> solution in preparation for additional changes that might be coming in
>> the future?
>>
>> Any suggestions would be greatly appreciated.
> I thought you distro folks were the franenkernel experts. :)
>
> But, seriously...  This isn't rocket science.  Just look at the pkru.h
> in your tree and figure out what includes it needs.  If it needs FPU
> gunk, include the FPU header.  If you get a compile error for spinlock_t
> ...  well ...  find the code using spinlock_t and make sure it includes
> spinlock.h.
>
Thanks for the advice, Dave.

Your suggestion will probably be the easiest approach.  I was just 
trying to limit the amount of frankenkernel changes, since there are 
already so many  :-)  It's always best to stay as close to mainline as 
possible.

Re: Issue With real-time patches on 5.15.y

[Mike Galbraith]

On Thu, 2022-05-05 at 19:20 -0400, Joseph Salisbury wrote:
> ...
> Also, the following is the build failure that happens:
> arch/x86/include/asm/pgtable.h:55:8: error: unknown type name 'spinlock_t'
>     55 | extern spinlock_t pgd_lock;
>        |        ^~~~~~~~~~

35fa745286ac4 was merged into mainline 5.17-rc1 to squash that buglet.

	-Mike

Re: Issue With real-time patches on 5.15.y

[Joseph Salisbury]

On 5/9/22 06:14, Mike Galbraith wrote:
> On Thu, 2022-05-05 at 19:20 -0400, Joseph Salisbury wrote:
>> ...
>> Also, the following is the build failure that happens:
>> arch/x86/include/asm/pgtable.h:55:8: error: unknown type name 'spinlock_t'
>>      55 | extern spinlock_t pgd_lock;
>>         |        ^~~~~~~~~~
> 35fa745286ac4 was merged into mainline 5.17-rc1 to squash that buglet.
>
> 	-Mike
Thank you for the pointer, Mike!  This commit fixes the build issue I 
was seeing.

Thanks,

Joe