Re: [PATCH 1/3] mm/slub: directly load freelist from cpu partial slab in the likely case

[Chengming Zhou <zhouchengming@bytedance.com>]

On 2024/1/23 16:24, Vlastimil Babka wrote:
> On 1/23/24 03:51, Chengming Zhou wrote:
>> On 2024/1/23 01:13, Vlastimil Babka wrote:
>>> On 1/19/24 04:53, Chengming Zhou wrote:
>>>> On 2024/1/19 06:14, Christoph Lameter (Ampere) wrote:
>>>>> On Thu, 18 Jan 2024, Chengming Zhou wrote:
>>>>>
>>>>>> So get_freelist() has two cases to handle: cpu slab and cpu partial list slab.
>>>>>> The latter is NOT frozen, so need to remove "VM_BUG_ON(!new.frozen)" from it.
>>>>>
>>>>> Right so keep the check if it is the former?
>>>>>
>>>>
>>>> Ok, I get it. Maybe like this:
>>>
>>> I think that's just too ugly for a VM_BUG_ON(). I'd just remove the check
>>> and be done with that.
>>
>> Ok with me.
>>
>>>
>>> I have a somewhat different point. You reused get_freelist() but in fact
>>> it's more like freeze_slab(), but that one uses slab_update_freelist() and
>>> we are under the local_lock so we want the cheaper __slab_update_freelist(),
>>> which get_freelist() has and I guess that's why you reused that one.
>>
>> Right, we already have the lock_lock, so reuse get_freelist().
>>
>>>
>>> However get_freelist() also assumes it can return NULL if the freelist is
>>> empty. If that's possible to happen on the percpu partial list, we should
>>> not "goto load_freelist;" but rather create a new label above that, above
>>> the "if (!freelist) {" block that handles the case.
>>>
>>> If that's not possible to happen (needs careful audit) and we have guarantee
>>
>> Yes, it's not possible for now.
>>
>>> that slabs on percpu partial list must have non-empty freelist, then we
>>> probably instead want a new __freeze_slab() variant that is like
>>> freeze_slab(), but uses __slab_update_freelist() and probably also has
>>> VM_BUG_ON(!freelist) before returning it?
>>>
>>
>> Instead of introducing another new function, how about still reusing get_freelist()
>> and VM_BUG_ON(!freelist) after calling it? I feel this is simpler.
> 
> Could you measure if introducing new function that sets new.frozen = 1; has
> any performance benefit? If not, we can reuse get_freelist() as you say.
> Thanks!
> 

I just tested using the new function: __freeze_slab() that uses __slab_update_freelist()
and sets new.frozen = 1, but found the performance is a little worse than reusing
get_freelist().

The reason I think maybe more code memory footprint? I don't look deep into that.

Anyway it looks better to reuse get_freelist(), I will update a version later.

Thanks!