Re: [PATCH] sparc/pci: Make pci_poke_lock a raw_spinlock_t.

[Waiman Long <llong@redhat.com>]

On 11/27/24 12:44 PM, Guenter Roeck wrote:
> On 11/27/24 08:53, Sebastian Andrzej Siewior wrote:
>> On 2024-11-27 08:02:50 [-0800], Guenter Roeck wrote:
>>> On 11/27/24 07:39, Andreas Larsson wrote:
>>>> Even though this is for sparc64, there is work being done looking into
>>>> enabling RT for sparc32. If the amount of fixes needed to keep
>>>> PROVE_RAW_LOCK_NESTING enabled is quite small at the moment I'd rather
>>>> see it enabled for sparc rather than risking it becoming worse in the
>>>> future.
>>
>> Okay. So you seem to be in favour of fixing the sparc64 splats Guenter
>> reported?
>>
>>>> I don't know what the situation is for other architectures that 
>>>> does not
>>>> support RT.
>>>>
>>>
>>> For my part I still don't understand why PROVE_RAW_LOCK_NESTING is 
>>> no longer
>>> a configurable option, or in other words why it is mandated even for 
>>> architectures
>>> not supporting RT. To me this means that I'll either have to disable 
>>> PROVE_LOCKING
>>> for sparc or live with endless warning backtraces. The latter 
>>> obscures real
>>> problems, so it is a no-go.
>>
>> It is documented in Documentation/locking/locktypes.rst how the locks
>> should nest. It is just nobody enabled it on sparc64 and tested. The
>> option was meant temporary until the big read blocks are cleared.
>>
>
> That doesn't explain why PROVE_RAW_LOCK_NESTING is now mandatory if
> PROVE_LOCKING is enabled, even on architectures where is was not tested.
> I am all for testing, but that doesn't include making it mandatory
> even where it is known to fail. Enabling it by default, sure, no problem.
> Dropping the option entirely after it is proven to no longer needed,
> also no problem. But force-enabling it even where untested or, worse,
> known to fail, is two steps too far.

The main reason for enforcing PROVE_RAW_LOCK_NESTING with PROVE_LOCKING 
is due to the fact that PREEMPT_RT kernel is much less tested than the 
non-RT kernel. I do agree that we shouldn't force this on arches that 
don't support PREEMPT_RT. However, once an arch decides to support 
PREEMPT_RT, they have to fix all these raw_spinlock nesting problems.

Cheers,
Longman