diff --git a/drivers/net/ethernet/amd/xgbe/xgbe-desc.c b/drivers/net/ethernet/amd/xgbe/xgbe-desc.c
index 62cb4d3388df..51b68d1299fe 100644
--- a/drivers/net/ethernet/amd/xgbe/xgbe-desc.c
+++ b/drivers/net/ethernet/amd/xgbe/xgbe-desc.c
@@ -378,7 +378,7 @@ static void xgbe_wrapper_tx_descriptor_init(struct xgbe_prv_data *pdata)
 
 		ring->cur = 0;
 		ring->dirty = 0;
-		ring->tx.queue_stopped = 0;
+		memset(&ring->tx, 0, sizeof(ring->tx));
 
 		hw_if->tx_desc_init(channel);
 	}
@@ -422,8 +422,7 @@ static void xgbe_wrapper_rx_descriptor_init(struct xgbe_prv_data *pdata)
 
 		ring->cur = 0;
 		ring->dirty = 0;
-		ring->rx.realloc_index = 0;
-		ring->rx.realloc_threshold = 0;
+		memset(&ring->rx, 0, sizeof(ring->rx));
 
 		hw_if->rx_desc_init(channel);
 	}
diff --git a/drivers/net/ethernet/amd/xgbe/xgbe-dev.c b/drivers/net/ethernet/amd/xgbe/xgbe-dev.c
index b3d2433b52bb..53f5f66ec2ee 100644
--- a/drivers/net/ethernet/amd/xgbe/xgbe-dev.c
+++ b/drivers/net/ethernet/amd/xgbe/xgbe-dev.c
@@ -1325,6 +1325,29 @@ static void xgbe_config_dcb_pfc(struct xgbe_prv_data *pdata)
 	xgbe_config_flow_control(pdata);
 }
 
+static void xgbe_tx_start_xmit(struct xgbe_channel *channel,
+			       struct xgbe_ring *ring)
+{
+	struct xgbe_prv_data *pdata = channel->pdata;
+	struct xgbe_ring_data *rdata;
+
+	/* Issue a poll command to Tx DMA by writing address
+	 * of next immediate free descriptor */
+	rdata = XGBE_GET_DESC_DATA(ring, ring->cur);
+	XGMAC_DMA_IOWRITE(channel, DMA_CH_TDTR_LO,
+			  lower_32_bits(rdata->rdesc_dma));
+
+	/* Start the Tx coalescing timer */
+	if (pdata->tx_usecs && !channel->tx_timer_active) {
+		channel->tx_timer_active = 1;
+		hrtimer_start(&channel->tx_timer,
+			      ktime_set(0, pdata->tx_usecs * NSEC_PER_USEC),
+			      HRTIMER_MODE_REL);
+	}
+
+	ring->tx.xmit_more = 0;
+}
+
 static void xgbe_dev_xmit(struct xgbe_channel *channel)
 {
 	struct xgbe_prv_data *pdata = channel->pdata;
@@ -1528,20 +1551,13 @@ static void xgbe_dev_xmit(struct xgbe_channel *channel)
 	/* Make sure ownership is written to the descriptor */
 	wmb();
 
-	/* Issue a poll command to Tx DMA by writing address
-	 * of next immediate free descriptor */
 	ring->cur++;
-	rdata = XGBE_GET_DESC_DATA(ring, ring->cur);
-	XGMAC_DMA_IOWRITE(channel, DMA_CH_TDTR_LO,
-			  lower_32_bits(rdata->rdesc_dma));
-
-	/* Start the Tx coalescing timer */
-	if (pdata->tx_usecs && !channel->tx_timer_active) {
-		channel->tx_timer_active = 1;
-		hrtimer_start(&channel->tx_timer,
-			      ktime_set(0, pdata->tx_usecs * NSEC_PER_USEC),
-			      HRTIMER_MODE_REL);
-	}
+	if (!packet->skb->xmit_more ||
+	    netif_xmit_stopped(netdev_get_tx_queue(pdata->netdev,
+						   channel->queue_index)))
+		xgbe_tx_start_xmit(channel, ring);
+	else
+		ring->tx.xmit_more = 1;
 
 	DBGPR("  %s: descriptors %u to %u written\n",
 	      channel->name, start_index & (ring->rdesc_count - 1),
@@ -2802,6 +2818,7 @@ void xgbe_init_function_ptrs_dev(struct xgbe_hw_if *hw_if)
 	hw_if->rx_desc_reset = xgbe_rx_desc_reset;
 	hw_if->is_last_desc = xgbe_is_last_desc;
 	hw_if->is_context_desc = xgbe_is_context_desc;
+	hw_if->tx_start_xmit = xgbe_tx_start_xmit;
 
 	/* For FLOW ctrl */
 	hw_if->config_tx_flow_control = xgbe_config_tx_flow_control;
diff --git a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
index f9635281c7df..02c104dc2aa4 100644
--- a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
+++ b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
@@ -225,6 +225,28 @@ static inline unsigned int xgbe_tx_avail_desc(struct xgbe_ring *ring)
 	return (ring->rdesc_count - (ring->cur - ring->dirty));
 }
 
+static int xgbe_maybe_stop_tx_queue(struct xgbe_channel *channel,
+				    struct xgbe_ring *ring, unsigned int count)
+{
+	struct xgbe_prv_data *pdata = channel->pdata;
+
+	if (count > xgbe_tx_avail_desc(ring)) {
+		DBGPR("  Tx queue stopped, not enough descriptors available\n");
+		netif_stop_subqueue(pdata->netdev, channel->queue_index);
+		ring->tx.queue_stopped = 1;
+
+		/* If we haven't notified the hardware because of xmit_more
+		 * support, tell it now
+		 */
+		if (ring->tx.xmit_more)
+			pdata->hw_if.tx_start_xmit(channel, ring);
+
+		return NETDEV_TX_BUSY;
+	}
+
+	return 0;
+}
+
 static int xgbe_calc_rx_buf_size(struct net_device *netdev, unsigned int mtu)
 {
 	unsigned int rx_buf_size;
@@ -1199,6 +1221,8 @@ static void xgbe_packet_info(struct xgbe_prv_data *pdata,
 	unsigned int len;
 	unsigned int i;
 
+	packet->skb = skb;
+
 	context_desc = 0;
 	packet->rdesc_count = 0;
 
@@ -1447,13 +1471,9 @@ static int xgbe_xmit(struct sk_buff *skb, struct net_device *netdev)
 	xgbe_packet_info(pdata, ring, skb, packet);
 
 	/* Check that there are enough descriptors available */
-	if (packet->rdesc_count > xgbe_tx_avail_desc(ring)) {
-		DBGPR("  Tx queue stopped, not enough descriptors available\n");
-		netif_stop_subqueue(netdev, channel->queue_index);
-		ring->tx.queue_stopped = 1;
-		ret = NETDEV_TX_BUSY;
+	ret = xgbe_maybe_stop_tx_queue(channel, ring, packet->rdesc_count);
+	if (ret)
 		goto tx_netdev_return;
-	}
 
 	ret = xgbe_prep_tso(skb, packet);
 	if (ret) {
@@ -1480,6 +1500,11 @@ static int xgbe_xmit(struct sk_buff *skb, struct net_device *netdev)
 	xgbe_print_pkt(netdev, skb, true);
 #endif
 
+	/* Stop the queue in advance if there may not be enough descriptors */
+	xgbe_maybe_stop_tx_queue(channel, ring, XGBE_TX_MAX_DESCS);
+
+	ret = NETDEV_TX_OK;
+
 tx_netdev_return:
 	spin_unlock_irqrestore(&ring->lock, flags);
 
diff --git a/drivers/net/ethernet/amd/xgbe/xgbe.h b/drivers/net/ethernet/amd/xgbe/xgbe.h
index ec0d37a917ae..eb3387398c6f 100644
--- a/drivers/net/ethernet/amd/xgbe/xgbe.h
+++ b/drivers/net/ethernet/amd/xgbe/xgbe.h
@@ -140,6 +140,17 @@
 
 #define XGBE_TX_MAX_BUF_SIZE	(0x3fff & ~(64 - 1))
 
+/* Descriptors required for maximum contigous TSO/GSO packet */
+#define XGBE_TX_MAX_SPLIT	((GSO_MAX_SIZE / XGBE_TX_MAX_BUF_SIZE) + 1)
+
+/* Maximum possible descriptors needed for an SKB:
+ * - Maximum number of SKB frags
+ * - Maximum descriptors for contiguous TSO/GSO packet
+ * - Possible context descriptor
+ * - Possible TSO header descriptor
+ */
+#define XGBE_TX_MAX_DESCS	(MAX_SKB_FRAGS + XGBE_TX_MAX_SPLIT + 2)
+
 #define XGBE_RX_MIN_BUF_SIZE	(ETH_FRAME_LEN + ETH_FCS_LEN + VLAN_HLEN)
 #define XGBE_RX_BUF_ALIGN	64
 #define XGBE_SKB_ALLOC_SIZE	256
@@ -225,6 +236,8 @@
 struct xgbe_prv_data;
 
 struct xgbe_packet_data {
+	struct sk_buff *skb;
+
 	unsigned int attributes;
 
 	unsigned int errors;
@@ -360,6 +373,7 @@ struct xgbe_ring {
 	union {
 		struct {
 			unsigned int queue_stopped;
+			unsigned int xmit_more;
 			unsigned short cur_mss;
 			unsigned short cur_vlan_ctag;
 		} tx;
@@ -523,6 +537,7 @@ struct xgbe_hw_if {
 	void (*tx_desc_reset)(struct xgbe_ring_data *);
 	int (*is_last_desc)(struct xgbe_ring_desc *);
 	int (*is_context_desc)(struct xgbe_ring_desc *);
+	void (*tx_start_xmit)(struct xgbe_channel *, struct xgbe_ring *);
 
 	/* For FLOW ctrl */
 	int (*config_tx_flow_control)(struct xgbe_prv_data *);