[PATCH net v2] netfilter: ipset: add synchronize_rcu() in destroy to close use-after-free race

[PATCH net v2] netfilter: ipset: add synchronize_rcu() in destroy to close use-after-free race

[Cen Zhang (Microsoft)]

The child set refcount decrement in list_set_del() was moved from an RCU
callback to the synchronous path, so that userspace sees accurate
reference counts immediately. However, this broke an implicit invariant:
previously ref could only reach zero after an RCU grace period,
guaranteeing all RCU readers had finished before destroy could proceed.
Now ref can hit zero while readers still hold a stale index, and
ip_set_destroy() NULLs the slot out from under them:

  CPU 0 (softirq)                CPU 1 (control path)
  ---                            ---
  rcu_read_lock()
  index = e->id
                                 list_set_del():
                                   list_del_rcu(e)
                                   ip_set_put_byindex(index) // ref->0
                                 ip_set_destroy():
                                   ip_set_list[index] = NULL
  ip_set_rcu_get(index) -> NULL
  BUG_ON(!set)                   // crash

  kernel BUG at net/netfilter/ipset/ip_set_core.c:754!
    ip_set_test  <- list_set_kadt <- ip_set_test <- set_match_v1

Insert synchronize_rcu() in ip_set_destroy() after confirming ref == 0
but before NULLing the slot, so it only pays the RCU wait cost when
actually destroying. Because synchronize_rcu() sleeps, ip_set_ref_lock
must be dropped first, which splits the critical section in two. A
recheck of ref/ref_netlink is therefore needed in the second section,
since a concurrent netlink dump continuation (which does not hold
nfnl_lock) may have incremented ref_netlink in the interim.

The bulk _destroy_all_sets() path does not need this recheck because
its is_destroyed flag prevents dump from taking new references.

Fixes: 439cd39ea136 ("netfilter: ipset: list:set: Decrease refcount synchronously on deletion and replace")
Reported-by: AutonomousCodeSecurity@microsoft.com
Reported-by: Xiang Mei (Microsoft) <xmei5@asu.edu>
Reported-by: Cen Zhang (Microsoft) <blbllhy@gmail.com>
Closes: https://lore.kernel.org/all/20260820010617.46851-1-blbllhy@gmail.com/
Signed-off-by: Cen Zhang (Microsoft) <blbllhy@gmail.com>
---
 net/netfilter/ipset/ip_set_core.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/net/netfilter/ipset/ip_set_core.c b/net/netfilter/ipset/ip_set_core.c
index 0a86a170ba90..1295bea7944a 100644
--- a/net/netfilter/ipset/ip_set_core.c
+++ b/net/netfilter/ipset/ip_set_core.c
@@ -1225,6 +1225,8 @@ _destroy_all_sets(struct ip_set_net *inst)
 	/* Must wait for flush to be really finished  */
 	if (need_wait)
 		rcu_barrier();
+	/* Wait for RCU readers before NULLing slots */
+	synchronize_rcu();
 	for (i = 0; i < inst->ip_set_max; i++) {
 		set = ip_set(inst, i);
 		if (set) {
@@ -1286,6 +1288,17 @@ static int ip_set_destroy(struct sk_buff *skb, const struct nfnl_info *info,
 			ret = -IPSET_ERR_BUSY;
 			goto out;
 		}
+		read_unlock_bh(&ip_set_ref_lock);
+
+		/* Wait for RCU readers before NULLing slot */
+		synchronize_rcu();
+
+		read_lock_bh(&ip_set_ref_lock);
+		/* Dump may have taken a ref while lock was dropped */
+		if (s->ref || s->ref_netlink) {
+			ret = -IPSET_ERR_BUSY;
+			goto out;
+		}
 		features = s->type->features;
 		ip_set(inst, i) = NULL;
 		read_unlock_bh(&ip_set_ref_lock);
-- 
2.55.0