On 2024/4/29 22:32, Nicolas Bouchinet wrote:
>
> On 4/29/24 15:35, Chengming Zhou wrote:
>> On 2024/4/29 20:59, Nicolas Bouchinet wrote:
>>> On 4/29/24 11:09, Nicolas Bouchinet wrote:
>>>> Hi Vlastimil,
>>>>
>>>> thanks for your review and your proposal.
>>>>
>>>> On 4/29/24 10:52, Vlastimil Babka wrote:
>>>>> On 4/25/24 5:14 PM, Chengming Zhou wrote:
>>>>>> On 2024/4/25 23:02, Nicolas Bouchinet wrote:
>>>>> Thanks for finding the bug and the fix!
>>>>>
>>>>>>> Hy,
>>>>>>>
>>>>>>> First of all, thanks a lot for your time.
>>>>>>>
>>>>>>> On 4/25/24 10:36, Chengming Zhou wrote:
>>>>>>>> On 2024/4/24 20:47, Nicolas Bouchinet wrote:
>>>>>>>>> From: Nicolas Bouchinet<nicolas.bouchinet@xxxxxxxxxxx>
>>>>>>>>>
>>>>>>>>> Commit 284f17ac13fe ("mm/slub: handle bulk and single object freeing
>>>>>>>>> separately") splits single and bulk object freeing in two functions
>>>>>>>>> slab_free() and slab_free_bulk() which leads slab_free() to call
>>>>>>>>> slab_free_hook() directly instead of slab_free_freelist_hook().
>>>>>>>> Right.
>>>>>>>> y not suitable for a stable-candidate fix we need
>>>>>>>>> If `init_on_free` is set, slab_free_hook() zeroes the object.
>>>>>>>>> Afterward, if `slub_debug=F` and `CONFIG_SLAB_FREELIST_HARDENED` are
>>>>>>>>> set, the do_slab_free() slowpath executes freelist consistency
>>>>>>>>> checks and try to decode a zeroed freepointer which leads to a
>>>>>>>>> "Freepointer corrupt" detection in check_object().
>>>>>>>> IIUC, the "freepointer" can be checked on the free path only when
>>>>>>>> it's outside the object memory. Here slab_free_hook() zeroed the
>>>>>>>> freepointer and caused the problem.
>>>>>>>>
>>>>>>>> But why we should zero the memory outside the object_size? It seems
>>>>>>>> more reasonable to only zero the object_size when init_on_free is set?
>>>>>>> The original purpose was to avoid leaking information through the object and its metadata / tracking information as described in init_on_free initial Commit 6471384af2a6 ("mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options").
>>>>>>>
>>>>>>> I have to admit I didn't read the entire lore about the original patchset yet, though it could be interesting to know a bit more the threat models, specifically regarding the object metadata init.
>>>>>> Thank you for the reference! I also don't get why it needs to zero
>>>>>> the metadata and tracking information.
>>>>> Hmm taking a step back, it seems really suboptimal to initialize the
>>>>> outside-object freepointer as part of init_on_free:
>>>>>
>>>>> - the freeing itself will always set it one way or another, in this case
>>>>> free_to_partial_list() will do set_freepointer() after free_debug_processing()
>>>>>
>>>>> - we lose the ability to detect if the allocated slab object's user wrote to
>>>>> it, which is a buffer overflow
>> Ah, right, this ability seems important for debugging overflow problem.
>>
>>>>> So the best option to me would be to adjust the init in slab_free_hook() to
>>>>> avoid the outside-object freepointer similarly to how it avoids the red zone.
>> Agree.
>>
>>>>> We'll still not have the buffer overflow detection ability for bulk free
>>>>> where slab_free_freelist_hook() will set the free pointer before we reach
>>>>> the checks, but changing that is most likely not worth the trouble, and
>>>>> especially not suitable for a stable-candidate fix we need here.
>>>> It seems like a good alternative to me, I'll push a V2 patch with those changes.
>>>>
>>>> I help maintaining the Linux-Hardened patchset in which we have a slab object canary feature that helps detecting overflows. It is located just after the object freepointer.
>>>
>>> I've tried a patch where the freepointer is avoided but it results in the same bug. It seems that the commit 0f181f9fbea8bc7ea ("mm/slub.c: init_on_free=1 should wipe freelist ptr for bulk allocations") inits the freepointer on allocation if init_on_free is set in order to return a clean initialized object to the caller.
>>>
>> Good catch! You may need to change maybe_wipe_obj_freeptr() too,
>> I haven't tested this, not sure whether it works for you. :)
>>
>> diff --git a/mm/slub.c b/mm/slub.c
>> index 3e33ff900d35..3f250a167cb5 100644
>> --- a/mm/slub.c
>> +++ b/mm/slub.c
>> @@ -3796,7 +3796,8 @@ static void *__slab_alloc_node(struct kmem_cache *s,
>> static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
>> void *obj)
>> {
>> - if (unlikely(slab_want_init_on_free(s)) && obj)
>> + if (unlikely(slab_want_init_on_free(s)) && obj &&
>> + !freeptr_outside_object(s))
>> memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
>> 0, sizeof(void *));
>> }
>>
>> Thanks!
>
> Indeed since check_object() avoids objects for which freepointer is in the object and since val is equal to SLUB_RED_ACTIVE in our specific case it should work. Do you want me to add you as Co-authored ?
>
Ok, it's great. Thanks!
