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

[Rafael Aquini <aquini@redhat.com>]

On Wed, Sep 23, 2020 at 10:21:36AM +0800, Huang, Ying wrote:
> Hi, Rafael,
> 
> Rafael Aquini <aquini@redhat.com> 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. 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;
...