Merge branch 'net-ll_temac-Bugfixes'

Esben Haabendal says:

====================
net: ll_temac: Bugfixes

Fix a number of bugs which have been present since the first commit.

The bugs fixed in patch 1,2 and 4 have all been observed in real systems, and
was relatively easy to reproduce given an appropriate stress setup.

Changes since v1:

- Changed error handling of of dma_map_single() in temac_start_xmit() to drop
  packet instead of returning NETDEV_TX_BUSY.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/xilinx/ll_temac.h

@@ -375,10 +375,14 @@ struct temac_local {
 	int tx_bd_next;
 	int tx_bd_tail;
 	int rx_bd_ci;
+	int rx_bd_tail;
 
 	/* DMA channel control setup */
 	u32 tx_chnl_ctrl;
 	u32 rx_chnl_ctrl;
+	u8 coalesce_count_rx;
+
+	struct delayed_work restart_work;
 };
 
 /* Wrappers for temac_ior()/temac_iow() function pointers above */

drivers/net/ethernet/xilinx/ll_temac_main.c

@@ -51,6 +51,7 @@
 #include <linux/ip.h>
 #include <linux/slab.h>
 #include <linux/interrupt.h>
+#include <linux/workqueue.h>
 #include <linux/dma-mapping.h>
 #include <linux/processor.h>
 #include <linux/platform_data/xilinx-ll-temac.h>
@@ -367,6 +368,8 @@ static int temac_dma_bd_init(struct net_device *ndev)
 		skb_dma_addr = dma_map_single(ndev->dev.parent, skb->data,
 					      XTE_MAX_JUMBO_FRAME_SIZE,
 					      DMA_FROM_DEVICE);
+		if (dma_mapping_error(ndev->dev.parent, skb_dma_addr))
+			goto out;
 		lp->rx_bd_v[i].phys = cpu_to_be32(skb_dma_addr);
 		lp->rx_bd_v[i].len = cpu_to_be32(XTE_MAX_JUMBO_FRAME_SIZE);
 		lp->rx_bd_v[i].app0 = cpu_to_be32(STS_CTRL_APP0_IRQONEND);
@@ -387,12 +390,13 @@ static int temac_dma_bd_init(struct net_device *ndev)
 	lp->tx_bd_next = 0;
 	lp->tx_bd_tail = 0;
 	lp->rx_bd_ci = 0;
+	lp->rx_bd_tail = RX_BD_NUM - 1;
 
 	/* Enable RX DMA transfers */
 	wmb();
 	lp->dma_out(lp, RX_CURDESC_PTR,  lp->rx_bd_p);
 	lp->dma_out(lp, RX_TAILDESC_PTR,
-		       lp->rx_bd_p + (sizeof(*lp->rx_bd_v) * (RX_BD_NUM - 1)));
+		       lp->rx_bd_p + (sizeof(*lp->rx_bd_v) * lp->rx_bd_tail));
 
 	/* Prepare for TX DMA transfer */
 	lp->dma_out(lp, TX_CURDESC_PTR, lp->tx_bd_p);
@@ -788,6 +792,9 @@ static void temac_start_xmit_done(struct net_device *ndev)
 		stat = be32_to_cpu(cur_p->app0);
 	}
 
+	/* Matches barrier in temac_start_xmit */
+	smp_mb();
+
 	netif_wake_queue(ndev);
 }
 
@@ -830,9 +837,19 @@ temac_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 	cur_p = &lp->tx_bd_v[lp->tx_bd_tail];
 
 	if (temac_check_tx_bd_space(lp, num_frag + 1)) {
-		if (!netif_queue_stopped(ndev))
+		if (netif_queue_stopped(ndev))
+			return NETDEV_TX_BUSY;
+
 		netif_stop_queue(ndev);
+
+		/* Matches barrier in temac_start_xmit_done */
+		smp_mb();
+
+		/* Space might have just been freed - check again */
+		if (temac_check_tx_bd_space(lp, num_frag))
 			return NETDEV_TX_BUSY;
+
+		netif_wake_queue(ndev);
 	}
 
 	cur_p->app0 = 0;
@@ -850,12 +867,16 @@ temac_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 	skb_dma_addr = dma_map_single(ndev->dev.parent, skb->data,
 				      skb_headlen(skb), DMA_TO_DEVICE);
 	cur_p->len = cpu_to_be32(skb_headlen(skb));
+	if (WARN_ON_ONCE(dma_mapping_error(ndev->dev.parent, skb_dma_addr))) {
+		dev_kfree_skb_any(skb);
+		ndev->stats.tx_dropped++;
+		return NETDEV_TX_OK;
+	}
 	cur_p->phys = cpu_to_be32(skb_dma_addr);
 	ptr_to_txbd((void *)skb, cur_p);
 
 	for (ii = 0; ii < num_frag; ii++) {
-		lp->tx_bd_tail++;
-		if (lp->tx_bd_tail >= TX_BD_NUM)
+		if (++lp->tx_bd_tail >= TX_BD_NUM)
 			lp->tx_bd_tail = 0;
 
 		cur_p = &lp->tx_bd_v[lp->tx_bd_tail];
@@ -863,6 +884,27 @@ temac_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 					      skb_frag_address(frag),
 					      skb_frag_size(frag),
 					      DMA_TO_DEVICE);
+		if (dma_mapping_error(ndev->dev.parent, skb_dma_addr)) {
+			if (--lp->tx_bd_tail < 0)
+				lp->tx_bd_tail = TX_BD_NUM - 1;
+			cur_p = &lp->tx_bd_v[lp->tx_bd_tail];
+			while (--ii >= 0) {
+				--frag;
+				dma_unmap_single(ndev->dev.parent,
+						 be32_to_cpu(cur_p->phys),
+						 skb_frag_size(frag),
+						 DMA_TO_DEVICE);
+				if (--lp->tx_bd_tail < 0)
+					lp->tx_bd_tail = TX_BD_NUM - 1;
+				cur_p = &lp->tx_bd_v[lp->tx_bd_tail];
+			}
+			dma_unmap_single(ndev->dev.parent,
+					 be32_to_cpu(cur_p->phys),
+					 skb_headlen(skb), DMA_TO_DEVICE);
+			dev_kfree_skb_any(skb);
+			ndev->stats.tx_dropped++;
+			return NETDEV_TX_OK;
+		}
 		cur_p->phys = cpu_to_be32(skb_dma_addr);
 		cur_p->len = cpu_to_be32(skb_frag_size(frag));
 		cur_p->app0 = 0;
@@ -884,31 +926,56 @@ temac_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 	return NETDEV_TX_OK;
 }
 
+static int ll_temac_recv_buffers_available(struct temac_local *lp)
+{
+	int available;
+
+	if (!lp->rx_skb[lp->rx_bd_ci])
+		return 0;
+	available = 1 + lp->rx_bd_tail - lp->rx_bd_ci;
+	if (available <= 0)
+		available += RX_BD_NUM;
+	return available;
+}
 
 static void ll_temac_recv(struct net_device *ndev)
 {
 	struct temac_local *lp = netdev_priv(ndev);
-	struct sk_buff *skb, *new_skb;
-	unsigned int bdstat;
-	struct cdmac_bd *cur_p;
-	dma_addr_t tail_p, skb_dma_addr;
-	int length;
 	unsigned long flags;
+	int rx_bd;
+	bool update_tail = false;
 
 	spin_lock_irqsave(&lp->rx_lock, flags);
 
-	tail_p = lp->rx_bd_p + sizeof(*lp->rx_bd_v) * lp->rx_bd_ci;
-	cur_p = &lp->rx_bd_v[lp->rx_bd_ci];
+	/* Process all received buffers, passing them on network
+	 * stack.  After this, the buffer descriptors will be in an
+	 * un-allocated stage, where no skb is allocated for it, and
+	 * they are therefore not available for TEMAC/DMA.
+	 */
+	do {
+		struct cdmac_bd *bd = &lp->rx_bd_v[lp->rx_bd_ci];
+		struct sk_buff *skb = lp->rx_skb[lp->rx_bd_ci];
+		unsigned int bdstat = be32_to_cpu(bd->app0);
+		int length;
 
-	bdstat = be32_to_cpu(cur_p->app0);
-	while ((bdstat & STS_CTRL_APP0_CMPLT)) {
+		/* While this should not normally happen, we can end
+		 * here when GFP_ATOMIC allocations fail, and we
+		 * therefore have un-allocated buffers.
+		 */
+		if (!skb)
+			break;
 
-		skb = lp->rx_skb[lp->rx_bd_ci];
-		length = be32_to_cpu(cur_p->app4) & 0x3FFF;
+		/* Loop over all completed buffer descriptors */
+		if (!(bdstat & STS_CTRL_APP0_CMPLT))
+			break;
 
-		dma_unmap_single(ndev->dev.parent, be32_to_cpu(cur_p->phys),
+		dma_unmap_single(ndev->dev.parent, be32_to_cpu(bd->phys),
 				 XTE_MAX_JUMBO_FRAME_SIZE, DMA_FROM_DEVICE);
+		/* The buffer is not valid for DMA anymore */
+		bd->phys = 0;
+		bd->len = 0;
 
+		length = be32_to_cpu(bd->app4) & 0x3FFF;
 		skb_put(skb, length);
 		skb->protocol = eth_type_trans(skb, ndev);
 		skb_checksum_none_assert(skb);
@@ -923,43 +990,102 @@ static void ll_temac_recv(struct net_device *ndev)
 			 * (back) for proper IP checksum byte order
 			 * (be16).
 			 */
-			skb->csum = htons(be32_to_cpu(cur_p->app3) & 0xFFFF);
+			skb->csum = htons(be32_to_cpu(bd->app3) & 0xFFFF);
 			skb->ip_summed = CHECKSUM_COMPLETE;
 		}
 
 		if (!skb_defer_rx_timestamp(skb))
 			netif_rx(skb);
+		/* The skb buffer is now owned by network stack above */
+		lp->rx_skb[lp->rx_bd_ci] = NULL;
 
 		ndev->stats.rx_packets++;
 		ndev->stats.rx_bytes += length;
 
-		new_skb = netdev_alloc_skb_ip_align(ndev,
-						XTE_MAX_JUMBO_FRAME_SIZE);
-		if (!new_skb) {
-			spin_unlock_irqrestore(&lp->rx_lock, flags);
-			return;
+		rx_bd = lp->rx_bd_ci;
+		if (++lp->rx_bd_ci >= RX_BD_NUM)
+			lp->rx_bd_ci = 0;
+	} while (rx_bd != lp->rx_bd_tail);
+
+	/* DMA operations will halt when the last buffer descriptor is
+	 * processed (ie. the one pointed to by RX_TAILDESC_PTR).
+	 * When that happens, no more interrupt events will be
+	 * generated.  No IRQ_COAL or IRQ_DLY, and not even an
+	 * IRQ_ERR.  To avoid stalling, we schedule a delayed work
+	 * when there is a potential risk of that happening.  The work
+	 * will call this function, and thus re-schedule itself until
+	 * enough buffers are available again.
+	 */
+	if (ll_temac_recv_buffers_available(lp) < lp->coalesce_count_rx)
+		schedule_delayed_work(&lp->restart_work, HZ / 1000);
+
+	/* Allocate new buffers for those buffer descriptors that were
+	 * passed to network stack.  Note that GFP_ATOMIC allocations
+	 * can fail (e.g. when a larger burst of GFP_ATOMIC
+	 * allocations occurs), so while we try to allocate all
+	 * buffers in the same interrupt where they were processed, we
+	 * continue with what we could get in case of allocation
+	 * failure.  Allocation of remaining buffers will be retried
+	 * in following calls.
+	 */
+	while (1) {
+		struct sk_buff *skb;
+		struct cdmac_bd *bd;
+		dma_addr_t skb_dma_addr;
+
+		rx_bd = lp->rx_bd_tail + 1;
+		if (rx_bd >= RX_BD_NUM)
+			rx_bd = 0;
+		bd = &lp->rx_bd_v[rx_bd];
+
+		if (bd->phys)
+			break;	/* All skb's allocated */
+
+		skb = netdev_alloc_skb_ip_align(ndev, XTE_MAX_JUMBO_FRAME_SIZE);
+		if (!skb) {
+			dev_warn(&ndev->dev, "skb alloc failed\n");
+			break;
 		}
 
-		cur_p->app0 = cpu_to_be32(STS_CTRL_APP0_IRQONEND);
-		skb_dma_addr = dma_map_single(ndev->dev.parent, new_skb->data,
+		skb_dma_addr = dma_map_single(ndev->dev.parent, skb->data,
 					      XTE_MAX_JUMBO_FRAME_SIZE,
 					      DMA_FROM_DEVICE);
-		cur_p->phys = cpu_to_be32(skb_dma_addr);
-		cur_p->len = cpu_to_be32(XTE_MAX_JUMBO_FRAME_SIZE);
-		lp->rx_skb[lp->rx_bd_ci] = new_skb;
+		if (WARN_ON_ONCE(dma_mapping_error(ndev->dev.parent,
+						   skb_dma_addr))) {
+			dev_kfree_skb_any(skb);
+			break;
+		}
 
-		lp->rx_bd_ci++;
-		if (lp->rx_bd_ci >= RX_BD_NUM)
-			lp->rx_bd_ci = 0;
+		bd->phys = cpu_to_be32(skb_dma_addr);
+		bd->len = cpu_to_be32(XTE_MAX_JUMBO_FRAME_SIZE);
+		bd->app0 = cpu_to_be32(STS_CTRL_APP0_IRQONEND);
+		lp->rx_skb[rx_bd] = skb;
+
+		lp->rx_bd_tail = rx_bd;
+		update_tail = true;
+	}
 
-		cur_p = &lp->rx_bd_v[lp->rx_bd_ci];
-		bdstat = be32_to_cpu(cur_p->app0);
+	/* Move tail pointer when buffers have been allocated */
+	if (update_tail) {
+		lp->dma_out(lp, RX_TAILDESC_PTR,
+			lp->rx_bd_p + sizeof(*lp->rx_bd_v) * lp->rx_bd_tail);
 	}
-	lp->dma_out(lp, RX_TAILDESC_PTR, tail_p);
 
 	spin_unlock_irqrestore(&lp->rx_lock, flags);
 }
 
+/* Function scheduled to ensure a restart in case of DMA halt
+ * condition caused by running out of buffer descriptors.
+ */
+static void ll_temac_restart_work_func(struct work_struct *work)
+{
+	struct temac_local *lp = container_of(work, struct temac_local,
+					      restart_work.work);
+	struct net_device *ndev = lp->ndev;
+
+	ll_temac_recv(ndev);
+}
+
 static irqreturn_t ll_temac_tx_irq(int irq, void *_ndev)
 {
 	struct net_device *ndev = _ndev;
@@ -1052,6 +1178,8 @@ static int temac_stop(struct net_device *ndev)
 
 	dev_dbg(&ndev->dev, "temac_close()\n");
 
+	cancel_delayed_work_sync(&lp->restart_work);
+
 	free_irq(lp->tx_irq, ndev);
 	free_irq(lp->rx_irq, ndev);
 
@@ -1173,6 +1301,7 @@ static int temac_probe(struct platform_device *pdev)
 	lp->dev = &pdev->dev;
 	lp->options = XTE_OPTION_DEFAULTS;
 	spin_lock_init(&lp->rx_lock);
+	INIT_DELAYED_WORK(&lp->restart_work, ll_temac_restart_work_func);
 
 	/* Setup mutex for synchronization of indirect register access */
 	if (pdata) {
@@ -1279,6 +1408,7 @@ static int temac_probe(struct platform_device *pdev)
 		 */
 		lp->tx_chnl_ctrl = 0x10220000;
 		lp->rx_chnl_ctrl = 0xff070000;
+		lp->coalesce_count_rx = 0x07;
 
 		/* Finished with the DMA node; drop the reference */
 		of_node_put(dma_np);
@@ -1310,11 +1440,14 @@ static int temac_probe(struct platform_device *pdev)
 				(pdata->tx_irq_count << 16);
 		else
 			lp->tx_chnl_ctrl = 0x10220000;
-		if (pdata->rx_irq_timeout || pdata->rx_irq_count)
+		if (pdata->rx_irq_timeout || pdata->rx_irq_count) {
 			lp->rx_chnl_ctrl = (pdata->rx_irq_timeout << 24) |
 				(pdata->rx_irq_count << 16);
-		else
+			lp->coalesce_count_rx = pdata->rx_irq_count;
+		} else {
 			lp->rx_chnl_ctrl = 0xff070000;
+			lp->coalesce_count_rx = 0x07;
+		}
 	}
 
 	/* Error handle returned DMA RX and TX interrupts */