Re: [PATCH] mm: swapfile: avoid split_swap_cluster() NULL pointer dereference

[Huang, Ying]

Rafael Aquini <aquini@xxxxxxxxxx> writes:

> On Wed, Sep 23, 2020 at 10:21:36AM +0800, Huang, Ying wrote:
>> Hi, Rafael,
>> 
>> Rafael Aquini <aquini@xxxxxxxxxx> writes:
>> 
>> > The swap area descriptor only gets struct swap_cluster_info *cluster_info
>> > allocated if the swapfile is backed by non-rotational storage.
>> > When the swap area is laid on top of ordinary disk spindles, lock_cluster()
>> > will naturally return NULL.
>> 
>> Thanks for reporting.  But the bug looks strange.  Because in a system
>> with only HDD swap devices, during THP swap out, the swap cluster
>> shouldn't be allocated, as in
>> 
>> shrink_page_list()
>>   add_to_swap()
>>     get_swap_page()
>>       get_swap_pages()
>>         swap_alloc_cluster()
>>
>
> The underlying problem is that swap_info_struct.cluster_info is always NULL 
> on the rotational storage case.

Yes.

> So, it's very easy to follow that constructions 
> like this one, in split_swap_cluster 
>
> ...
>         ci = lock_cluster(si, offset);
>         cluster_clear_huge(ci);
> ...
>
> will go for a NULL pointer dereference, in that case, given that lock_cluster 
> reads:
>
> ...
> 	struct swap_cluster_info *ci;
>         ci = si->cluster_info;
>         if (ci) {
>                 ci += offset / SWAPFILE_CLUSTER;
>                 spin_lock(&ci->lock);
>         }
>         return ci;
> ...

But on HDD, we shouldn't call split_swap_cluster() at all, because we
will not allocate swap cluster firstly.  So, if we run into this,
there should be some other bug, we need to figure it out.

Best Regards,
Huang, Ying