Skip to contentSkip to conversation

net: systemport and bcmgenet OOM fixes

PATCH In review Florian Fainelli 2 participants 4 messages Sep 8, 2014

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

cover net: systemport and bcmgenet OOM fixes

Hi David,

These two patches fix similar Out of Memory code paths in the SYSTEMPORT and GENET drivers. Under high memory pressure, we could produce an OOPS by passing a NULL pointer to dma_unmap_single().

Thanks!

Florian Fainelli (2):
net: systemport: check harder for out of memory conditions
net: bcmgenet: check harder for out of memory conditions

 drivers/net/ethernet/broadcom/bcmsysport.c     | 31 ++++++++++++++----------
 drivers/net/ethernet/broadcom/genet/bcmgenet.c | 33 +++++++++++++++-----------
 2 files changed, 38 insertions(+), 26 deletions(-)
Signature · 1 line
1.9.1
#1 lore raw

PATCH 1/2 net: systemport: check harder for out of memory conditions

There is a potential case where we might be failing to refill a control block, leaving it with both a NULL skb pointer *and* a NULL dma_unmap_addr.

The way we process incoming packets, by first calling dma_unmap_single(), and then only checking for a potential NULL skb can lead to situations where do pass a NULL dma_unmap_addr() to dma_unmap_single(), resulting in an oops.

Fix this my moving the NULL skb check earlier, since no backing skb also means no corresponding DMA mapping for this packet.

---
 drivers/net/ethernet/broadcom/bcmsysport.c | 31 ++++++++++++++++++------------
 1 file changed, 19 insertions(+), 12 deletions(-)
drivers/net/ethernet/broadcom/bcmsysport.c +19 -12
index 6f4e18644bd4..d9b9170ed2fc 100644
@@ -534,6 +534,25 @@ static unsigned int bcm_sysport_desc_rx(struct bcm_sysport_priv *priv,
 	while ((processed < to_process) && (processed < budget)) {
 		cb = &priv->rx_cbs[priv->rx_read_ptr];
 		skb = cb->skb;
+
+		processed++;
+		priv->rx_read_ptr++;
+
+		if (priv->rx_read_ptr == priv->num_rx_bds)
+			priv->rx_read_ptr = 0;
+
+		/* We do not have a backing SKB, so we do not a corresponding
+		 * DMA mapping for this incoming packet since
+		 * bcm_sysport_rx_refill always either has both skb and mapping
+		 * or none.
+		 */
+		if (unlikely(!skb)) {
+			netif_err(priv, rx_err, ndev, "out of memory!\n");
+			ndev->stats.rx_dropped++;
+			ndev->stats.rx_errors++;
+			goto refill;
+		}
+
 		dma_unmap_single(kdev, dma_unmap_addr(cb, dma_addr),
 				 RX_BUF_LENGTH, DMA_FROM_DEVICE);
 
@@ -543,23 +562,11 @@ static unsigned int bcm_sysport_desc_rx(struct bcm_sysport_priv *priv,
 		status = (rsb->rx_status_len >> DESC_STATUS_SHIFT) &
 			  DESC_STATUS_MASK;
 
-		processed++;
-		priv->rx_read_ptr++;
-		if (priv->rx_read_ptr == priv->num_rx_bds)
-			priv->rx_read_ptr = 0;
-
 		netif_dbg(priv, rx_status, ndev,
 			  "p=%d, c=%d, rd_ptr=%d, len=%d, flag=0x%04x\n",
 			  p_index, priv->rx_c_index, priv->rx_read_ptr,
 			  len, status);
 
-		if (unlikely(!skb)) {
-			netif_err(priv, rx_err, ndev, "out of memory!\n");
-			ndev->stats.rx_dropped++;
-			ndev->stats.rx_errors++;
-			goto refill;
-		}
-
 		if (unlikely(!(status & DESC_EOP) || !(status & DESC_SOP))) {
 			netif_err(priv, rx_status, ndev, "fragmented packet!\n");
 			ndev->stats.rx_dropped++;
Signature · 1 line
1.9.1
#2 lore raw

PATCH 2/2 net: bcmgenet: check harder for out of memory conditions

In reply to Florian Fainelli

There is a potential case where we might be failing to refill a control block, leaving it with both a NULL skb pointer *and* a NULL dma_unmap_addr.

The way we process incoming packets, by first calling dma_unmap_single(), and then only checking for a potential NULL skb can lead to situations where do pass a NULL dma_unmap_addr() to dma_unmap_single(), resulting in an oops.

Fix this my moving the NULL skb check earlier, since no backing skb also means no corresponding DMA mapping for this packet.

---
 drivers/net/ethernet/broadcom/genet/bcmgenet.c | 33 +++++++++++++++-----------
 1 file changed, 19 insertions(+), 14 deletions(-)
drivers/net/ethernet/broadcom/genet/bcmgenet.c +19 -14
index 3f9d4de8173c..cdef86a03862 100644
@@ -1274,12 +1274,29 @@ static unsigned int bcmgenet_desc_rx(struct bcmgenet_priv *priv,
 
 	while ((rxpktprocessed < rxpkttoprocess) &&
 	       (rxpktprocessed < budget)) {
+		cb = &priv->rx_cbs[priv->rx_read_ptr];
+		skb = cb->skb;
+
+		rxpktprocessed++;
+
+		priv->rx_read_ptr++;
+		priv->rx_read_ptr &= (priv->num_rx_bds - 1);
+
+		/* We do not have a backing SKB, so we do not have a
+		 * corresponding DMA mapping for this incoming packet since
+		 * bcmgenet_rx_refill always either has both skb and mapping or
+		 * none.
+		 */
+		if (unlikely(!skb)) {
+			dev->stats.rx_dropped++;
+			dev->stats.rx_errors++;
+			goto refill;
+		}
+
 		/* Unmap the packet contents such that we can use the
 		 * RSV from the 64 bytes descriptor when enabled and save
 		 * a 32-bits register read
 		 */
-		cb = &priv->rx_cbs[priv->rx_read_ptr];
-		skb = cb->skb;
 		dma_unmap_single(&dev->dev, dma_unmap_addr(cb, dma_addr),
 				 priv->rx_buf_len, DMA_FROM_DEVICE);
 
@@ -1307,18 +1324,6 @@ static unsigned int bcmgenet_desc_rx(struct bcmgenet_priv *priv,
 			  __func__, p_index, priv->rx_c_index,
 			  priv->rx_read_ptr, dma_length_status);
 
-		rxpktprocessed++;
-
-		priv->rx_read_ptr++;
-		priv->rx_read_ptr &= (priv->num_rx_bds - 1);
-
-		/* out of memory, just drop packets at the hardware level */
-		if (unlikely(!skb)) {
-			dev->stats.rx_dropped++;
-			dev->stats.rx_errors++;
-			goto refill;
-		}
-
 		if (unlikely(!(dma_flag & DMA_EOP) || !(dma_flag & DMA_SOP))) {
 			netif_err(priv, rx_status, dev,
 				  "dropping fragmented packet!\n");
Signature · 1 line
1.9.1
maintainer
#3 lore raw
In reply to Florian Fainelli

From: Florian Fainelli <[email protected]>
Date: Mon, 8 Sep 2014 11:37:50 -0700

Quoted 3 lines

These two patches fix similar Out of Memory code paths in the SYSTEMPORT and GENET drivers. Under high memory pressure, we could produce an OOPS by passing a NULL pointer to dma_unmap_single().

Series applied, thanks Florian.