Re: [PATCH] slab: add a check for the first kmem_cache not to be destroyed

From: kwon
Date: Tue Jan 17 2017 - 21:36:43 EST



> On Jan 18, 2017, at 7:54 AM, David Rientjes <rientjes@xxxxxxxxxx> wrote:
>
> On Tue, 17 Jan 2017, kwon wrote:
>
>>>> diff --git a/mm/slab_common.c b/mm/slab_common.c
>>>> index 1dfc209..2d30ace 100644
>>>> --- a/mm/slab_common.c
>>>> +++ b/mm/slab_common.c
>>>> @@ -744,7 +744,7 @@ void kmem_cache_destroy(struct kmem_cache *s)
>>>> bool need_rcu_barrier = false;
>>>> int err;
>>>>
>>>> - if (unlikely(!s))
>>>> + if (unlikely(!s) || s->refcount == -1)
>>>> return;
>>>
>>> Hello, Kyunghwan.
>>>
>>> Few lines below, s->refcount is checked.
>>>
>>> if (s->refcount)
>>> goto unlock;
>>>
>>> Am I missing something?
>>>
>>> Thanks.
>>
>> Hello, Joonsoo.
>>
>> In case it is called the number of int size times. refcount would finally reach
>> to 0 since decreased every time the function called.
>>
>
> The only thing using create_boot_cache() should be the slab implementation
> itself, so I don't think we need to protect ourselves from doing something
> like kmem_cache_destroy(kmem_cache) or
> kmem_cache_destroy(kmem_cache_node) even a single time.

Agreed. I was aware of that though, I thought it would make its logic firm not
giving performance disadvantages. Sorry for distraction.