Re: [PATCH] locking/mutex: remove redundant argument from __mutex_lock_common()

[Waiman Long <longman@redhat.com>]

On 9/2/23 15:40, Waiman Long wrote:
> On 9/2/23 13:06, Michał Mirosław wrote:
>> On Wed, Aug 30, 2023 at 07:54:13PM -0400, Waiman Long wrote:
>>> On 8/30/23 18:12, Michał Mirosław wrote:
>>>> use_ww_ctx is equivalent to ww_ctx != NULL. The one case where
>>>> use_ww_ctx was true but ww_ctx == NULL leads to the same
>>>> __mutex_add_waiter() call via __ww_mutex_add_waiter().
>>> I think ww_mutex_lock() can be called with a NULL ctx. Your patch will
>>> effectively change those ww_mutex_lock() to be equivalent to 
>>> mutex_lock().
>>> So it is a behavioral change.
>> Isn't ww_mutex_lock() with ctx = NULL expected to behave like 
>> mutex_lock()?
>>
>>>> Since now __ww_mutex_add_waiter() is called only with ww_mutex != 
>>>> NULL,
>>>> remove the branch there.
>> [...]
>>>> @@ -627,12 +624,11 @@ __mutex_lock_common(struct mutex *lock, 
>>>> unsigned int state, unsigned int subclas
>>>>        debug_mutex_lock_common(lock, &waiter);
>>>>        waiter.task = current;
>>>> -    if (use_ww_ctx)
>>>> -        waiter.ww_ctx = ww_ctx;
>>>> +    waiter.ww_ctx = ww_ctx;
>>> This one is fine.
>>>> lock_contended(&lock->dep_map, ip);
>>>> -    if (!use_ww_ctx) {
>>>> +    if (!ww_ctx) {
>>> That change will break ww_mutex.
>> I see that there is the rt_mutex version that stubs out
>> __ww_mutex_add_waiter(), but its ww_mutex_lock() doesn't use
>> __mutex_lock_common() at all. With the RT version out of the picture, we
>> can see that __ww_mutex_add_waiter(), when passed ww_ctx == NULL, just
>> forwards the work to __ww_waiter_add() with the same arguments
>> and returns 0 -- making the path exactly as the !use_ww_ctx branch.
>>
>> Note: There is a lot of templating-via-preprocessor code here and I
>> might have missed something. I'll appreciate hints here as maybe it
>> could be made simpler or better understood.
>
> Yes, I have misread the code thinking that __ww_waiter_add() with a 
> NULL third argument is different from __mutex_add_waiter(). They the 
> same in this case for the non-PREEMPT_RT kernel. For the PREEMPT_RT 
> kernel, however, they are still different.

OTOH, the rtmutex code will not call __ww_mutex_add_waiter() with NULL 
ww_ctx. So in that sense, the patch is probably OK. You will need to 
expand the patch description to also describe the case for PREEMPT_RT 
kernel.

Cheers,
Longman