[PATCH V3 5/8] net: mediatek: fix TX locking

[John Crispin]

Inside the TX path there is a lock inside the tx_map function. This is
however too late. The patch moves the lock to the start of the xmit
function right before the free count check of the DMA ring happens.
If we do not do this, the code becomes racy leading to TX stalls and
dropped packets. This happens as there are 2 netdevs running on the
same physical DMA ring.

Signed-off-by: John Crispin <blogic at openwrt.org>

---
Changes in V3
* fix a typo in the comment

 drivers/net/ethernet/mediatek/mtk_eth_soc.c |   20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/mediatek/mtk_eth_soc.c b/drivers/net/ethernet/mediatek/mtk_eth_soc.c
index 4ebc42e..7b76075 100644
--- a/drivers/net/ethernet/mediatek/mtk_eth_soc.c
+++ b/drivers/net/ethernet/mediatek/mtk_eth_soc.c
@@ -536,7 +536,6 @@ static int mtk_tx_map(struct sk_buff *skb, struct net_device *dev,
 	struct mtk_eth *eth = mac->hw;
 	struct mtk_tx_dma *itxd, *txd;
 	struct mtk_tx_buf *tx_buf;
-	unsigned long flags;
 	dma_addr_t mapped_addr;
 	unsigned int nr_frags;
 	int i, n_desc = 1;
@@ -568,11 +567,6 @@ static int mtk_tx_map(struct sk_buff *skb, struct net_device *dev,
 	if (unlikely(dma_mapping_error(&dev->dev, mapped_addr)))
 		return -ENOMEM;
 
-	/* normally we can rely on the stack not calling this more than once,
-	 * however we have 2 queues running ont he same ring so we need to lock
-	 * the ring access
-	 */
-	spin_lock_irqsave(&eth->page_lock, flags);
 	WRITE_ONCE(itxd->txd1, mapped_addr);
 	tx_buf->flags |= MTK_TX_FLAGS_SINGLE0;
 	dma_unmap_addr_set(tx_buf, dma_addr0, mapped_addr);
@@ -632,8 +626,6 @@ static int mtk_tx_map(struct sk_buff *skb, struct net_device *dev,
 	WRITE_ONCE(itxd->txd3, (TX_DMA_SWC | TX_DMA_PLEN0(skb_headlen(skb)) |
 				(!nr_frags * TX_DMA_LS0)));
 
-	spin_unlock_irqrestore(&eth->page_lock, flags);
-
 	netdev_sent_queue(dev, skb->len);
 	skb_tx_timestamp(skb);
 
@@ -661,8 +653,6 @@ err_dma:
 		itxd = mtk_qdma_phys_to_virt(ring, itxd->txd2);
 	} while (itxd != txd);
 
-	spin_unlock_irqrestore(&eth->page_lock, flags);
-
 	return -ENOMEM;
 }
 
@@ -712,14 +702,22 @@ static int mtk_start_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct mtk_eth *eth = mac->hw;
 	struct mtk_tx_ring *ring = &eth->tx_ring;
 	struct net_device_stats *stats = &dev->stats;
+	unsigned long flags;
 	bool gso = false;
 	int tx_num;
 
+	/* normally we can rely on the stack not calling this more than once,
+	 * however we have 2 queues running on the same ring so we need to lock
+	 * the ring access
+	 */
+	spin_lock_irqsave(&eth->page_lock, flags);
+
 	tx_num = mtk_cal_txd_req(skb);
 	if (unlikely(atomic_read(&ring->free_count) <= tx_num)) {
 		mtk_stop_queue(eth);
 		netif_err(eth, tx_queued, dev,
 			  "Tx Ring full when queue awake!\n");
+		spin_unlock_irqrestore(&eth->page_lock, flags);
 		return NETDEV_TX_BUSY;
 	}
 
@@ -747,10 +745,12 @@ static int mtk_start_xmit(struct sk_buff *skb, struct net_device *dev)
 			     ring->thresh))
 			mtk_wake_queue(eth);
 	}
+	spin_unlock_irqrestore(&eth->page_lock, flags);
 
 	return NETDEV_TX_OK;
 
 drop:
+	spin_unlock_irqrestore(&eth->page_lock, flags);
 	stats->tx_dropped++;
 	dev_kfree_skb(skb);
 	return NETDEV_TX_OK;
-- 
1.7.10.4