Re: [PATCH 6.4 000/292] 6.4.5-rc1 review

[Muchun Song <muchun.song@linux.dev>]

> On Jul 27, 2023, at 00:52, Alexander Potapenko <glider@google.com> wrote:
> 
> On Tue, Jul 25, 2023 at 6:21 PM Alexander Potapenko <glider@google.com> wrote:
>> 
>> On Tue, Jul 25, 2023 at 3:39 PM Naresh Kamboju
>> <naresh.kamboju@linaro.org> wrote:
>>> 
>>> On Tue, 25 Jul 2023 at 17:22, Alexander Potapenko <glider@google.com> wrote:
>>>> 
>>>> On Tue, Jul 25, 2023 at 11:59 AM Alexander Potapenko <glider@google.com> wrote:
>>>>> 
>>>>> On Mon, Jul 24, 2023 at 2:10 PM Naresh Kamboju
>>>>> <naresh.kamboju@linaro.org> wrote:
>>>>>> 
>>>>>> On Mon, 24 Jul 2023 at 15:50, Alexander Potapenko <glider@google.com> wrote:
>>>>>>> 
>>>>>>> On Sat, Jul 22, 2023 at 6:37 PM Linus Torvalds
>>>>>>> <torvalds@linux-foundation.org> wrote:
>>>>>>>> 
>>>>>>>> [ Removed the stable reviewers, bringing in the kfence people ]
>>>>>>>> 
>>>>>>>> See
>>>>>>>> 
>>>>>>>>  https://lore.kernel.org/lkml/CA+G9fYvgy22wiY=c3wLOrCM6o33636abhtEynXhJkqxJh4ca0A@mail.gmail.com/
>>>>>>>> 
>>>>>>>> for the original report. The warning was introduced in 8f0b36497303
>>>>>>>> ("mm: kfence: fix objcgs vector allocation"), and Google doesn't find
>>>>>>>> any other cases of this.
>>>>>>>> 
>>>>>>>> Anybody?
>>>>>>>> 
>>>>>>>>                    Linus
>>>>>>>> 
>>>>>>> 
>> 
>> Muchun, any chance you know under what circumstances a KFENCE object
>> has its meta->objcg set to a non-NULL value?
>> It seems to be a quite rare case, and I've only seen it in live
>> radix_tree_node objects.
>> Since the check here:
>> https://elixir.bootlin.com/linux/latest/source/mm/kfence/core.c#L1097
>> ensures that this value is NULL when the object is freed, where is the
>> code that is supposed to zero it?
>> Could there be a race somewhere?
> 
> 
> I am still puzzled about what is going on.
> 
> As far as I can see, when KFENCE pool is initialized, for ith object
> page in the pool its page_slab()->memcg_data is set to a value derived
> from kfence_metadata[i].objcg
> Because KFENCE objects always occupy one page, no two objects are
> expected to share memcg_data at any time.
> 
> When slab_alloc_node() is called, it first invokes
> slab_pre_alloc_hook(), figures out the obj_cgroup and charges it for
> the allocated memory. The obj_cgroup is returned to slab_alloc_node()
> and after KFENCE allocation succeeds is passed to
> slab_post_alloc_hook(), which then writes obj_cgroup to
> *(page_slab(object)->memcg_data).
> 
> When an object is deallocated, slab_free() calls
> memcg_slab_free_hook(), which zeroes *(page_slab(object)->memcg_data)
> and passes the object to kfence_free().
> At this point the object's meta->objcg must be NULL, so the warning
> should not be firing.

At least, totally agree. This call  stack comes from slab_free() which
makes sure memcg_slab_free_hook() is called before kfence_free(), so
meta->objcg must be NULL. Otherwise, seems something is corrupted. So
I really want to know what's the value of "meta->objcg" when the warning
is firing (e.g. whether it is a valid pointer or does the last bit is
set with MEMCG_DATA_OBJCGS). Maybe we could improve the warning message,
e.g. print the current value of "meta->objcg".

Thanks.