From 9ed7de1afad3ebc08a3ce379bee0bbbbd9702414 Mon Sep 17 00:00:00 2001
From: Robert Lubos <robert.lubos@nordicsemi.no>
Date: Fri, 10 Nov 2023 10:23:34 +0100
Subject: [PATCH] [nrf fromlist] net: iface: Introduce TX mutex locking

A recent iface lock removal in ed17320c3d2e43c76b09dc359d3b371e9d23732d
exposed issues with concurrent access on TX to drivers that are not
re-entrant.

Reverting that commit does not really solve the problem, as it would
still exist if multiple Traffic Class queues are in use.

Therefore, introduce a separate mutex for TX data path, protecting the
L2/driver from concurrent transfers from several threads.

Upstream PR: https://github.com/zephyrproject-rtos/zephyr/pull/65049

Signed-off-by: Robert Lubos <robert.lubos@nordicsemi.no>
---
 include/zephyr/net/net_if.h  | 29 +++++++++++++++++++++++++++++
 subsys/net/ip/net_if.c       |  3 +++
 subsys/net/l2/ethernet/arp.c |  2 ++
 3 files changed, 34 insertions(+)

diff --git a/include/zephyr/net/net_if.h b/include/zephyr/net/net_if.h
index b06bf74de03..92a59ad3b73 100644
--- a/include/zephyr/net/net_if.h
+++ b/include/zephyr/net/net_if.h
@@ -212,6 +212,9 @@ enum net_if_flag {
 	/** IPv6 Multicast Listener Discovery disabled. */
 	NET_IF_IPV6_NO_MLD,
 
+	/** Mutex locking on TX data path disabled on the interface. */
+	NET_IF_NO_TX_LOCK,
+
 /** @cond INTERNAL_HIDDEN */
 	/* Total number of flags - must be at the end of the enum */
 	NET_IF_NUM_FLAGS
@@ -613,6 +616,7 @@ struct net_if {
 #endif
 
 	struct k_mutex lock;
+	struct k_mutex tx_lock;
 };
 
 static inline void net_if_lock(struct net_if *iface)
@@ -629,6 +633,31 @@ static inline void net_if_unlock(struct net_if *iface)
 	k_mutex_unlock(&iface->lock);
 }
 
+static inline bool net_if_flag_is_set(struct net_if *iface,
+				      enum net_if_flag value);
+
+static inline void net_if_tx_lock(struct net_if *iface)
+{
+	NET_ASSERT(iface);
+
+	if (net_if_flag_is_set(iface, NET_IF_NO_TX_LOCK)) {
+		return;
+	}
+
+	(void)k_mutex_lock(&iface->tx_lock, K_FOREVER);
+}
+
+static inline void net_if_tx_unlock(struct net_if *iface)
+{
+	NET_ASSERT(iface);
+
+	if (net_if_flag_is_set(iface, NET_IF_NO_TX_LOCK)) {
+		return;
+	}
+
+	k_mutex_unlock(&iface->tx_lock);
+}
+
 /**
  * @brief Set a value in network interface flags
  *
diff --git a/subsys/net/ip/net_if.c b/subsys/net/ip/net_if.c
index 035e18da74f..026eaa88b48 100644
--- a/subsys/net/ip/net_if.c
+++ b/subsys/net/ip/net_if.c
@@ -265,7 +265,9 @@ static bool net_if_tx(struct net_if *iface, struct net_pkt *pkt)
 			}
 		}
 
+		net_if_tx_lock(iface);
 		status = net_if_l2(iface)->send(iface, pkt);
+		net_if_tx_unlock(iface);
 
 		if (IS_ENABLED(CONFIG_NET_PKT_TXTIME_STATS)) {
 			uint32_t end_tick = k_cycle_get_32();
@@ -437,6 +439,7 @@ static inline void init_iface(struct net_if *iface)
 #endif
 
 	k_mutex_init(&iface->lock);
+	k_mutex_init(&iface->tx_lock);
 
 	api->init(iface);
 }
diff --git a/subsys/net/l2/ethernet/arp.c b/subsys/net/l2/ethernet/arp.c
index 35a2ec63a64..39a06ed64a0 100644
--- a/subsys/net/l2/ethernet/arp.c
+++ b/subsys/net/l2/ethernet/arp.c
@@ -534,7 +534,9 @@ static void arp_update(struct net_if *iface,
 		 * the pkt are not counted twice and the packet filter
 		 * callbacks are only called once.
 		 */
+		net_if_tx_lock(iface);
 		ret = net_if_l2(iface)->send(iface, pkt);
+		net_if_tx_unlock(iface);
 		if (ret < 0) {
 			net_pkt_unref(pkt);
 		}