Re: [PATCH 1/5] debugobjects: Fix the misuse of global variables in fill_pool()

[Thomas Gleixner <tglx@linutronix.de>]

On Tue, Sep 03 2024 at 19:14, Leizhen wrote:
> On 2024/9/3 17:37, Thomas Gleixner wrote:
>> On Tue, Sep 03 2024 at 15:00, Leizhen wrote:
>>>>> @@ -84,10 +85,7 @@ static int __data_racy                       debug_objects_fixups __read_mostly;
>>>>>  static int __data_racy                 debug_objects_warnings __read_mostly;
>>>>>  static int __data_racy                 debug_objects_enabled __read_mostly
>>>>>                                         = CONFIG_DEBUG_OBJECTS_ENABLE_DEFAULT;
>>>>> -static int __data_racy                 debug_objects_pool_size __read_mostly
>>>>> -                                       = ODEBUG_POOL_SIZE;
>>>>> -static int __data_racy                 debug_objects_pool_min_level __read_mostly
>>>>> -                                       = ODEBUG_POOL_MIN_LEVEL;
>>>>> +static int __data_racy                 obj_pool_min_free = ODEBUG_POOL_SIZE;
>>>
>>> Sorry, I rechecked it again. After this patch, obj_pool_min_free is referenced in the
>>> same way as obj_pool_max_used. The only race point is located in debug_stats_show().
>>> However, this reference point does not need to be included in the race analysis. So
>>> there is no need to add __data_racy for obj_pool_min_free.
>> 
>> The read races against the write, so KCSAN can detect it and complain, no?
>
> Oh, I just saw that there were a lot of other global variables in that function
> that didn't mask KCSAN's detection. So I'll recheck each global variable.
> However, for obj_pool_min_free, it seems that it would be better to just add
> READ_ONCE() in debug_stats_show(). This does not prevent the compiler from
> optimizing variable references in the lock.
>
> # define __data_racy volatile

This is only when KCSAN is enabled. Otherwise it's empty.

And if you do a READ_ONCE() then you need a corresponding WRITE_ONCE()
to make sense. __data_racy is much simpler for that.

Thanks,

        tglx