[PATCH] net: ethernet: ti: cpsw: fix net watchdog timeout

[Grygorii Strashko]

It was discovered that simple program which indefinitely sends 200b UDP
packets and runs on TI AM574x SoC (SMP) under RT Kernel triggers network
watchdog timeout in TI CPSW driver (<6 hours run). The network watchdog
timeout is triggered due to race between cpsw_ndo_start_xmit() and
cpsw_tx_handler() [NAPI]

cpsw_ndo_start_xmit()
	if (unlikely(!cpdma_check_free_tx_desc(txch))) {
		txq = netdev_get_tx_queue(ndev, q_idx);
		netif_tx_stop_queue(txq);

^^ as per [1] barier has to be used after set_bit() otherwise new value
might not be visible to other cpus
	}

cpsw_tx_handler()
	if (unlikely(netif_tx_queue_stopped(txq)))
		netif_tx_wake_queue(txq);

and when it happens ndev TX queue became disabled forever while driver's HW
TX queue is empty.

Fix this, by adding smp_mb__after_atomic() after netif_tx_stop_queue()
calls and double check for free TX descriptors after stopping ndev TX queue
- if there are free TX descriptors wake up ndev TX queue.

[1] https://www.kernel.org/doc/html/latest/core-api/atomic_ops.html
Signed-off-by: Grygorii Strashko <grygorii.strashko@xxxxxx>
---
 drivers/net/ethernet/ti/cpsw.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/ti/cpsw.c b/drivers/net/ethernet/ti/cpsw.c
index 10d7cbe..3805b13 100644
--- a/drivers/net/ethernet/ti/cpsw.c
+++ b/drivers/net/ethernet/ti/cpsw.c
@@ -1638,6 +1638,7 @@ static netdev_tx_t cpsw_ndo_start_xmit(struct sk_buff *skb,
 		q_idx = q_idx % cpsw->tx_ch_num;
 
 	txch = cpsw->txv[q_idx].ch;
+	txq = netdev_get_tx_queue(ndev, q_idx);
 	ret = cpsw_tx_packet_submit(priv, skb, txch);
 	if (unlikely(ret != 0)) {
 		cpsw_err(priv, tx_err, "desc submit failed\n");
@@ -1648,15 +1649,26 @@ static netdev_tx_t cpsw_ndo_start_xmit(struct sk_buff *skb,
 	 * tell the kernel to stop sending us tx frames.
 	 */
 	if (unlikely(!cpdma_check_free_tx_desc(txch))) {
-		txq = netdev_get_tx_queue(ndev, q_idx);
 		netif_tx_stop_queue(txq);
+
+		/* Barrier, so that stop_queue visible to other cpus */
+		smp_mb__after_atomic();
+
+		if (cpdma_check_free_tx_desc(txch))
+			netif_tx_wake_queue(txq);
 	}
 
 	return NETDEV_TX_OK;
 fail:
 	ndev->stats.tx_dropped++;
-	txq = netdev_get_tx_queue(ndev, skb_get_queue_mapping(skb));
 	netif_tx_stop_queue(txq);
+
+	/* Barrier, so that stop_queue visible to other cpus */
+	smp_mb__after_atomic();
+
+	if (cpdma_check_free_tx_desc(txch))
+		netif_tx_wake_queue(txq);
+
 	return NETDEV_TX_BUSY;
 }
 
-- 
2.10.5