Before releasing the global mutex, we only unlink the hashtable
from the hash list, its proc file is still not unregistered at
this point. So syzbot could trigger a race condition where a
parallel htable_create() could register the same file immediately
after the mutex is released.
Move htable_remove_proc_entry() back to mutex protection to
fix this. And, fold htable_destroy() into htable_put() to make
the code slightly easier to understand.
Reported-and-tested-by: syzbot+d195fd3b9a364ddd6731@syzkaller.appspotmail.com
Fixes: c4a3922d2d20 ("netfilter: xt_hashlimit: reduce hashlimit_mutex scope for htable_put()")
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
                remove_proc_entry(hinfo->name, parent);
 }
 
-static void htable_destroy(struct xt_hashlimit_htable *hinfo)
-{
-       cancel_delayed_work_sync(&hinfo->gc_work);
-       htable_remove_proc_entry(hinfo);
-       htable_selective_cleanup(hinfo, true);
-       kfree(hinfo->name);
-       vfree(hinfo);
-}
-
 static struct xt_hashlimit_htable *htable_find_get(struct net *net,
                                                   const char *name,
                                                   u_int8_t family)
 {
        if (refcount_dec_and_mutex_lock(&hinfo->use, &hashlimit_mutex)) {
                hlist_del(&hinfo->node);
+               htable_remove_proc_entry(hinfo);
                mutex_unlock(&hashlimit_mutex);
-               htable_destroy(hinfo);
+
+               cancel_delayed_work_sync(&hinfo->gc_work);
+               htable_selective_cleanup(hinfo, true);
+               kfree(hinfo->name);
+               vfree(hinfo);
        }
 }