wireguard: allowedips: add WGALLOWEDIP_F_REMOVE_ME flag

The current netlink API for WireGuard does not directly support removal
of allowed ips from a peer. A user can remove an allowed ip from a peer
in one of two ways:

1. By using the WGPEER_F_REPLACE_ALLOWEDIPS flag and providing a new
   list of allowed ips which omits the allowed ip that is to be removed.
2. By reassigning an allowed ip to a "dummy" peer then removing that
   peer with WGPEER_F_REMOVE_ME.

With the first approach, the driver completely rebuilds the allowed ip
list for a peer. If my current configuration is such that a peer has
allowed ips 192.168.0.2 and 192.168.0.3 and I want to remove 192.168.0.2
the actual transition looks like this.

[192.168.0.2, 192.168.0.3] <-- Initial state
[]                         <-- Step 1: Allowed ips removed for peer
[192.168.0.3]              <-- Step 2: Allowed ips added back for peer

This is true even if the allowed ip list is small and the update does
not need to be batched into multiple WG_CMD_SET_DEVICE requests, as the
removal and subsequent addition of ips is non-atomic within a single
request. Consequently, wg_allowedips_lookup_dst and
wg_allowedips_lookup_src may return NULL while reconfiguring a peer even
for packets bound for ips a user did not intend to remove leading to
unintended interruptions in connectivity. This presents in userspace as
failed calls to sendto and sendmsg for UDP sockets. In my case, I ran
netperf while repeatedly reconfiguring the allowed ips for a peer with
wg.

/usr/local/bin/netperf -H 10.102.73.72 -l 10m -t UDP_STREAM -- -R 1 -m 1024
send_data: data send error: No route to host (errno 113)
netperf: send_omni: send_data failed: No route to host

While this may not be of particular concern for environments where peers
and allowed ips are mostly static, systems like Cilium manage peers and
allowed ips in a dynamic environment where peers (i.e. Kubernetes nodes)
and allowed ips (i.e. pods running on those nodes) can frequently
change making WGPEER_F_REPLACE_ALLOWEDIPS problematic.

The second approach avoids any possible connectivity interruptions
but is hacky and less direct, requiring the creation of a temporary
peer just to dispose of an allowed ip.

Introduce a new flag called WGALLOWEDIP_F_REMOVE_ME which in the same
way that WGPEER_F_REMOVE_ME allows a user to remove a single peer from
a WireGuard device's configuration allows a user to remove an ip from a
peer's set of allowed ips. This enables incremental updates to a
device's configuration without any connectivity blips or messy
workarounds.

A corresponding patch for wg extends the existing `wg set` interface to
leverage this feature.

$ wg set wg0 peer <PUBKEY> allowed-ips +192.168.88.0/24,-192.168.0.1/32

When '+' or '-' is prepended to any ip in the list, wg clears
WGPEER_F_REPLACE_ALLOWEDIPS and sets the WGALLOWEDIP_F_REMOVE_ME flag on
any ip prefixed with '-'.

Signed-off-by: Jordan Rife <jordan@jrife.io>
[Jason: minor style nits, fixes to selftest, bump of wireguard-tools version]
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Link: https://patch.msgid.link/20250521212707.1767879-5-Jason@zx2c4.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>

1 files changed, 71 insertions, 31 deletions

diff --git a/drivers/net/wireguard/allowedips.c b/drivers/net/wireguard/allowedips.c
index 4b8528206cc8..09f7fcd7da78 100644
--- a/drivers/net/wireguard/allowedips.c
+++ b/drivers/net/wireguard/allowedips.c
@@ -249,6 +249,52 @@ static int add(struct allowedips_node __rcu **trie, u8 bits, const u8 *key,
 	return 0;
 }
 
+static void remove_node(struct allowedips_node *node, struct mutex *lock)
+{
+	struct allowedips_node *child, **parent_bit, *parent;
+	bool free_parent;
+
+	list_del_init(&node->peer_list);
+	RCU_INIT_POINTER(node->peer, NULL);
+	if (node->bit[0] && node->bit[1])
+		return;
+	child = rcu_dereference_protected(node->bit[!rcu_access_pointer(node->bit[0])],
+					  lockdep_is_held(lock));
+	if (child)
+		child->parent_bit_packed = node->parent_bit_packed;
+	parent_bit = (struct allowedips_node **)(node->parent_bit_packed & ~3UL);
+	*parent_bit = child;
+	parent = (void *)parent_bit -
+			offsetof(struct allowedips_node, bit[node->parent_bit_packed & 1]);
+	free_parent = !rcu_access_pointer(node->bit[0]) && !rcu_access_pointer(node->bit[1]) &&
+			(node->parent_bit_packed & 3) <= 1 && !rcu_access_pointer(parent->peer);
+	if (free_parent)
+		child = rcu_dereference_protected(parent->bit[!(node->parent_bit_packed & 1)],
+						  lockdep_is_held(lock));
+	call_rcu(&node->rcu, node_free_rcu);
+	if (!free_parent)
+		return;
+	if (child)
+		child->parent_bit_packed = parent->parent_bit_packed;
+	*(struct allowedips_node **)(parent->parent_bit_packed & ~3UL) = child;
+	call_rcu(&parent->rcu, node_free_rcu);
+}
+
+static int remove(struct allowedips_node __rcu **trie, u8 bits, const u8 *key,
+		  u8 cidr, struct wg_peer *peer, struct mutex *lock)
+{
+	struct allowedips_node *node;
+
+	if (unlikely(cidr > bits))
+		return -EINVAL;
+	if (!rcu_access_pointer(*trie) || !node_placement(*trie, key, cidr, bits, &node, lock) ||
+	    peer != rcu_access_pointer(node->peer))
+		return 0;
+
+	remove_node(node, lock);
+	return 0;
+}
+
 void wg_allowedips_init(struct allowedips *table)
 {
 	table->root4 = table->root6 = NULL;
@@ -300,44 +346,38 @@ int wg_allowedips_insert_v6(struct allowedips *table, const struct in6_addr *ip,
 	return add(&table->root6, 128, key, cidr, peer, lock);
 }
 
+int wg_allowedips_remove_v4(struct allowedips *table, const struct in_addr *ip,
+			    u8 cidr, struct wg_peer *peer, struct mutex *lock)
+{
+	/* Aligned so it can be passed to fls */
+	u8 key[4] __aligned(__alignof(u32));
+
+	++table->seq;
+	swap_endian(key, (const u8 *)ip, 32);
+	return remove(&table->root4, 32, key, cidr, peer, lock);
+}
+
+int wg_allowedips_remove_v6(struct allowedips *table, const struct in6_addr *ip,
+			    u8 cidr, struct wg_peer *peer, struct mutex *lock)
+{
+	/* Aligned so it can be passed to fls64 */
+	u8 key[16] __aligned(__alignof(u64));
+
+	++table->seq;
+	swap_endian(key, (const u8 *)ip, 128);
+	return remove(&table->root6, 128, key, cidr, peer, lock);
+}
+
 void wg_allowedips_remove_by_peer(struct allowedips *table,
 				  struct wg_peer *peer, struct mutex *lock)
 {
-	struct allowedips_node *node, *child, **parent_bit, *parent, *tmp;
-	bool free_parent;
+	struct allowedips_node *node, *tmp;
 
 	if (list_empty(&peer->allowedips_list))
 		return;
 	++table->seq;
-	list_for_each_entry_safe(node, tmp, &peer->allowedips_list, peer_list) {
-		list_del_init(&node->peer_list);
-		RCU_INIT_POINTER(node->peer, NULL);
-		if (node->bit[0] && node->bit[1])
-			continue;
-		child = rcu_dereference_protected(node->bit[!rcu_access_pointer(node->bit[0])],
-						  lockdep_is_held(lock));
-		if (child)
-			child->parent_bit_packed = node->parent_bit_packed;
-		parent_bit = (struct allowedips_node **)(node->parent_bit_packed & ~3UL);
-		*parent_bit = child;
-		parent = (void *)parent_bit -
-			 offsetof(struct allowedips_node, bit[node->parent_bit_packed & 1]);
-		free_parent = !rcu_access_pointer(node->bit[0]) &&
-			      !rcu_access_pointer(node->bit[1]) &&
-			      (node->parent_bit_packed & 3) <= 1 &&
-			      !rcu_access_pointer(parent->peer);
-		if (free_parent)
-			child = rcu_dereference_protected(
-					parent->bit[!(node->parent_bit_packed & 1)],
-					lockdep_is_held(lock));
-		call_rcu(&node->rcu, node_free_rcu);
-		if (!free_parent)
-			continue;
-		if (child)
-			child->parent_bit_packed = parent->parent_bit_packed;
-		*(struct allowedips_node **)(parent->parent_bit_packed & ~3UL) = child;
-		call_rcu(&parent->rcu, node_free_rcu);
-	}
+	list_for_each_entry_safe(node, tmp, &peer->allowedips_list, peer_list)
+		remove_node(node, lock);
 }
 
 int wg_allowedips_read_node(struct allowedips_node *node, u8 ip[16], u8 *cidr)