[PATCH net] wireguard: fix race in wg_index_hashtable_replace()

[Jason A. Donenfeld]

Hey Eric,

On Tue, Sep 8, 2020 at 4:59 PM Eric Dumazet <edumazet at google.com> wrote:
>
> syzbot got a NULL dereference in wg_index_hashtable_replace() [1]
>
> Issue here is that right after checking hlist_unhashed(&old->index_hash)
> another cpu might have removed @old already from the hash.
>
> Since we are dealing with a very unlikely case, we can simply
> acquire the table lock earlier.

That's a nice bug. It looks like this is triggered by a teardown race,
when wg_index_hashtable_replace races with wg_index_hashtable_remove.

Since all the other hashtable mutator functions are protected by that
spinlock, it doesn't seem harmful to fix this by doing the same, even
if formally that spinlock is supposed to protect hash bucket heads
rather than entry pointers.

I'm playing with your patch and a variant of it, which I'll have
queued up in my tree in the next hour or so.

Thanks a lot for triaging this.

Jason