Re: [PATCH] mm/mempolicy: fix potential mpol_new leak in shared_policy_replace

[Miaohe Lin <linmiaohe@huawei.com>]

On 2022/3/15 23:27, Michal Hocko wrote:
> On Tue 15-03-22 21:42:29, Miaohe Lin wrote:
>> On 2022/3/15 0:44, Michal Hocko wrote:
>>> On Fri 11-03-22 17:36:24, Miaohe Lin wrote:
>>>> If mpol_new is allocated but not used in restart loop, mpol_new will be
>>>> freed via mpol_put before returning to the caller. But refcnt is not
>>>> initialized yet, so mpol_put could not do the right things and might
>>>> leak the unused mpol_new.
>>>
>>> The code is really hideous but is there really any bug there? AFAICS the
>>> new policy is only allocated in if (n->end > end) branch and that one
>>> will set the reference count on the retry. Or am I missing something?
>>>
>>
>> Many thanks for your comment.
>> IIUC, new policy is allocated via the below code:
>>
>> shared_policy_replace:
>> 	alloc_new:
>> 		write_unlock(&sp->lock);
>> 		ret = -ENOMEM;
>> 		n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
>> 		if (!n_new)
>> 			goto err_out;
>> 		mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
>> 		if (!mpol_new)
>> 			goto err_out;
>> 		goto restart;
>>
>> And mpol_new' reference count will be set before used in n->end > end case. But
>> if that is "not" the case, i.e. mpol_new is not inserted into the rb_tree, mpol_new
>> will be freed via mpol_put before return:
> 
> One thing I have missed previously is that the lock is dropped during
> the allocation so I guess the memory policy could have been changed
> during that time. Is this possible? Have you explored this possibility?
> Is this a theoretical problem or it can be triggered intentionally.
> 

This is found via code investigation. I think this could be triggered if there
are many concurrent mpol_set_shared_policy in place. But the user-visible effect
might be obscure as only sizeof(struct mempolicy) bytes leaks possiblely every time.

> These details would be really interesting for the changelog so that we
> can judge how important this would be.

This might not be that important as this issue should have been well-concealed for
almost ten years (since commit 42288fe366c4 ("mm: mempolicy: Convert shared_policy mutex to spinlock")).

> 

Thanks.