netfilter: ipset: Fix suspicious rcu_dereference_protected()

When destroying all sets, we are either in pernet exit phase or
are executing a "destroy all sets command" from userspace. The latter
was taken into account in ip_set_dereference() (nfnetlink mutex is held),
but the former was not. The patch adds the required check to
rcu_dereference_protected() in ip_set_dereference().

Fixes: 4e7aaa6 ("netfilter: ipset: Fix race between namespace cleanup and gc in the list:set type")
Reported-by: syzbot+b62c37cdd58103293a5a@syzkaller.appspotmail.com
Reported-by: syzbot+cfbe1da5fdfc39efc293@syzkaller.appspotmail.com
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202406141556.e0b6f17e-lkp@intel.com
Signed-off-by: Jozsef Kadlecsik <kadlec@netfilter.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>

net/netfilter/ipset/ip_set_core.c

@@ -53,12 +53,13 @@ MODULE_DESCRIPTION("core IP set support");
 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_IPSET);
 
 /* When the nfnl mutex or ip_set_ref_lock is held: */
-#define ip_set_dereference(p)		\
-	rcu_dereference_protected(p,	\
+#define ip_set_dereference(inst)	\
+	rcu_dereference_protected((inst)->ip_set_list,	\
 		lockdep_nfnl_is_held(NFNL_SUBSYS_IPSET) || \
-		lockdep_is_held(&ip_set_ref_lock))
+		lockdep_is_held(&ip_set_ref_lock) || \
+		(inst)->is_deleted)
 #define ip_set(inst, id)		\
-	ip_set_dereference((inst)->ip_set_list)[id]
+	ip_set_dereference(inst)[id]
 #define ip_set_ref_netlink(inst,id)	\
 	rcu_dereference_raw((inst)->ip_set_list)[id]
 #define ip_set_dereference_nfnl(p)	\
@@ -1133,7 +1134,7 @@ static int ip_set_create(struct sk_buff *skb, const struct nfnl_info *info,
 		if (!list)
 			goto cleanup;
 		/* nfnl mutex is held, both lists are valid */
-		tmp = ip_set_dereference(inst->ip_set_list);
+		tmp = ip_set_dereference(inst);
 		memcpy(list, tmp, sizeof(struct ip_set *) * inst->ip_set_max);
 		rcu_assign_pointer(inst->ip_set_list, list);
 		/* Make sure all current packets have passed through */