diff mbox

[v2,06/11] net: calxedaxgmac: fix race with tx queue stop/wake

Message ID 1377832428-18117-6-git-send-email-robherring2@gmail.com
State Changes Requested, archived
Delegated to: David Miller
Headers show

Commit Message

Rob Herring Aug. 30, 2013, 3:13 a.m. UTC 
From: Rob Herring <rob.herring@calxeda.com>

Since the xgmac transmit start and completion work locklessly, it is
possible for xgmac_xmit to stop the tx queue after the xgmac_tx_complete
has run resulting in the tx queue never being woken up. Fix this by
ensuring that ring buffer index updates are visible and recheck the ring
space after stopping the queue.

The implementation used here was copied from
drivers/net/ethernet/broadcom/tg3.c.

Signed-off-by: Rob Herring <rob.herring@calxeda.com>
---
v2:
- drop netif_tx_lock
- use netif_start_queue instead of netif_wake_queue

 drivers/net/ethernet/calxeda/xgmac.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

Comments

Ben Hutchings Aug. 30, 2013, 5:12 p.m. UTC | #1 
On Thu, 2013-08-29 at 22:13 -0500, Rob Herring wrote:
> From: Rob Herring <rob.herring@calxeda.com>
> 
> Since the xgmac transmit start and completion work locklessly, it is
> possible for xgmac_xmit to stop the tx queue after the xgmac_tx_complete
> has run resulting in the tx queue never being woken up. Fix this by
> ensuring that ring buffer index updates are visible and recheck the ring
> space after stopping the queue.
> 
> The implementation used here was copied from
> drivers/net/ethernet/broadcom/tg3.c.
> 
> Signed-off-by: Rob Herring <rob.herring@calxeda.com>
> ---
> v2:
> - drop netif_tx_lock
> - use netif_start_queue instead of netif_wake_queue
[...]
> @@ -1125,10 +1130,15 @@ static netdev_tx_t xgmac_xmit(struct sk_buff *skb, struct net_device *dev)
>  
>  	priv->tx_head = dma_ring_incr(entry, DMA_TX_RING_SZ);
>  
> -	if (dma_ring_space(priv->tx_head, priv->tx_tail, DMA_TX_RING_SZ) <
> -	    MAX_SKB_FRAGS)
> +	/* Ensure tx_head update is visible to tx completion */
> +	smp_mb();
> +	if (unlikely(tx_dma_ring_space(priv) < MAX_SKB_FRAGS)) {
[...]

On re-reading, I see there's an off-by-one bug lingering here.  You use
up to 1 + MAX_SKB_FRAGS descriptors per skb.  So you must stop the queue
when the ring space is <= MAX_SKB_FRAGS.

Ben.
diff mbox

Patch

diff --git a/drivers/net/ethernet/calxeda/xgmac.c b/drivers/net/ethernet/calxeda/xgmac.c
index f630855..54be55b 100644
--- a/drivers/net/ethernet/calxeda/xgmac.c
+++ b/drivers/net/ethernet/calxeda/xgmac.c
@@ -410,6 +410,9 @@  struct xgmac_priv {
 #define dma_ring_space(h, t, s)	CIRC_SPACE(h, t, s)
 #define dma_ring_cnt(h, t, s)	CIRC_CNT(h, t, s)
 
+#define tx_dma_ring_space(p) \
+	dma_ring_space((p)->tx_head, (p)->tx_tail, DMA_TX_RING_SZ)
+
 /* XGMAC Descriptor Access Helpers */
 static inline void desc_set_buf_len(struct xgmac_dma_desc *p, u32 buf_sz)
 {
@@ -886,8 +889,10 @@  static void xgmac_tx_complete(struct xgmac_priv *priv)
 		priv->tx_tail = dma_ring_incr(entry, DMA_TX_RING_SZ);
 	}
 
-	if (dma_ring_space(priv->tx_head, priv->tx_tail, DMA_TX_RING_SZ) >
-	    MAX_SKB_FRAGS)
+	/* Ensure tx_tail is visible to xgmac_xmit */
+	smp_mb();
+	if (unlikely(netif_queue_stopped(priv->dev) &&
+	    (tx_dma_ring_space(priv) > MAX_SKB_FRAGS)))
 		netif_wake_queue(priv->dev);
 }
 
@@ -1125,10 +1130,15 @@  static netdev_tx_t xgmac_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	priv->tx_head = dma_ring_incr(entry, DMA_TX_RING_SZ);
 
-	if (dma_ring_space(priv->tx_head, priv->tx_tail, DMA_TX_RING_SZ) <
-	    MAX_SKB_FRAGS)
+	/* Ensure tx_head update is visible to tx completion */
+	smp_mb();
+	if (unlikely(tx_dma_ring_space(priv) < MAX_SKB_FRAGS)) {
 		netif_stop_queue(dev);
-
+		/* Ensure netif_stop_queue is visible to tx completion */
+		smp_mb();
+		if (tx_dma_ring_space(priv) > MAX_SKB_FRAGS)
+			netif_start_queue(dev);
+	}
 	return NETDEV_TX_OK;
 }