]> git.karo-electronics.de Git - mv-sheeva.git/commitdiff
net/mpc5200: Fix locking on fec_mpc52xx driver
authorAsier Llano <a.llano@ziv.es>
Wed, 9 Dec 2009 04:29:10 +0000 (20:29 -0800)
committerDavid S. Miller <davem@davemloft.net>
Wed, 9 Dec 2009 04:29:10 +0000 (20:29 -0800)
Fix the locking scheme on the fec_mpc52xx driver.  This device can
receive IRQs from three sources; the FEC itself, the tx DMA, and the
rx DMA.  Mutual exclusion was handled by taking a spin_lock() in the
critical regions, but because the handlers are run with IRQs enabled,
spin_lock() is insufficient and the driver can end up interrupting
a critical region anyway from another IRQ.

Asier Llano discovered that this occurs when an error IRQ is raised
in the middle of handling rx irqs which resulted in an sk_buff memory
leak.

In addition, locking is spotty at best in the driver and inspection
revealed quite a few places with insufficient locking.

This patch is based on Asier's initial work, but reworks a number of
things so that locks are held for as short a time as possible, so
that spin_lock_irqsave() is used everywhere, and so the locks are
dropped when calling into the network stack (because the lock only
protects the hardware interface; not the network stack).

Boot tested on a lite5200 with an NFS root.  Has not been performance
tested.

Signed-off-by: Asier Llano <a.llano@ziv.es>
Signed-off-by: Grant Likely <grant.likely@secretlab.ca>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/fec_mpc52xx.c

index 6407672b28e96aac3491835c85eaa88684773474..848e8407ea8f5407a53eb5464a2f487e131e2ee9 100644 (file)
@@ -85,11 +85,15 @@ MODULE_PARM_DESC(debug, "debugging messages level");
 
 static void mpc52xx_fec_tx_timeout(struct net_device *dev)
 {
+       struct mpc52xx_fec_priv *priv = netdev_priv(dev);
+       unsigned long flags;
+
        dev_warn(&dev->dev, "transmit timed out\n");
 
+       spin_lock_irqsave(&priv->lock, flags);
        mpc52xx_fec_reset(dev);
-
        dev->stats.tx_errors++;
+       spin_unlock_irqrestore(&priv->lock, flags);
 
        netif_wake_queue(dev);
 }
@@ -135,28 +139,32 @@ static void mpc52xx_fec_free_rx_buffers(struct net_device *dev, struct bcom_task
        }
 }
 
+static void
+mpc52xx_fec_rx_submit(struct net_device *dev, struct sk_buff *rskb)
+{
+       struct mpc52xx_fec_priv *priv = netdev_priv(dev);
+       struct bcom_fec_bd *bd;
+
+       bd = (struct bcom_fec_bd *) bcom_prepare_next_buffer(priv->rx_dmatsk);
+       bd->status = FEC_RX_BUFFER_SIZE;
+       bd->skb_pa = dma_map_single(dev->dev.parent, rskb->data,
+                                   FEC_RX_BUFFER_SIZE, DMA_FROM_DEVICE);
+       bcom_submit_next_buffer(priv->rx_dmatsk, rskb);
+}
+
 static int mpc52xx_fec_alloc_rx_buffers(struct net_device *dev, struct bcom_task *rxtsk)
 {
-       while (!bcom_queue_full(rxtsk)) {
-               struct sk_buff *skb;
-               struct bcom_fec_bd *bd;
+       struct sk_buff *skb;
 
+       while (!bcom_queue_full(rxtsk)) {
                skb = dev_alloc_skb(FEC_RX_BUFFER_SIZE);
-               if (skb == NULL)
+               if (!skb)
                        return -EAGAIN;
 
                /* zero out the initial receive buffers to aid debugging */
                memset(skb->data, 0, FEC_RX_BUFFER_SIZE);
-
-               bd = (struct bcom_fec_bd *)bcom_prepare_next_buffer(rxtsk);
-
-               bd->status = FEC_RX_BUFFER_SIZE;
-               bd->skb_pa = dma_map_single(dev->dev.parent, skb->data,
-                               FEC_RX_BUFFER_SIZE, DMA_FROM_DEVICE);
-
-               bcom_submit_next_buffer(rxtsk, skb);
+               mpc52xx_fec_rx_submit(dev, skb);
        }
-
        return 0;
 }
 
@@ -328,13 +336,12 @@ static int mpc52xx_fec_start_xmit(struct sk_buff *skb, struct net_device *dev)
                                    DMA_TO_DEVICE);
 
        bcom_submit_next_buffer(priv->tx_dmatsk, skb);
+       spin_unlock_irqrestore(&priv->lock, flags);
 
        if (bcom_queue_full(priv->tx_dmatsk)) {
                netif_stop_queue(dev);
        }
 
-       spin_unlock_irqrestore(&priv->lock, flags);
-
        return NETDEV_TX_OK;
 }
 
@@ -359,9 +366,9 @@ static irqreturn_t mpc52xx_fec_tx_interrupt(int irq, void *dev_id)
 {
        struct net_device *dev = dev_id;
        struct mpc52xx_fec_priv *priv = netdev_priv(dev);
+       unsigned long flags;
 
-       spin_lock(&priv->lock);
-
+       spin_lock_irqsave(&priv->lock, flags);
        while (bcom_buffer_done(priv->tx_dmatsk)) {
                struct sk_buff *skb;
                struct bcom_fec_bd *bd;
@@ -372,11 +379,10 @@ static irqreturn_t mpc52xx_fec_tx_interrupt(int irq, void *dev_id)
 
                dev_kfree_skb_irq(skb);
        }
+       spin_unlock_irqrestore(&priv->lock, flags);
 
        netif_wake_queue(dev);
 
-       spin_unlock(&priv->lock);
-
        return IRQ_HANDLED;
 }
 
@@ -384,67 +390,60 @@ static irqreturn_t mpc52xx_fec_rx_interrupt(int irq, void *dev_id)
 {
        struct net_device *dev = dev_id;
        struct mpc52xx_fec_priv *priv = netdev_priv(dev);
+       struct sk_buff *rskb; /* received sk_buff */
+       struct sk_buff *skb;  /* new sk_buff to enqueue in its place */
+       struct bcom_fec_bd *bd;
+       u32 status, physaddr;
+       int length;
+       unsigned long flags;
+
+       spin_lock_irqsave(&priv->lock, flags);
 
        while (bcom_buffer_done(priv->rx_dmatsk)) {
-               struct sk_buff *skb;
-               struct sk_buff *rskb;
-               struct bcom_fec_bd *bd;
-               u32 status;
 
                rskb = bcom_retrieve_buffer(priv->rx_dmatsk, &status,
-                               (struct bcom_bd **)&bd);
-               dma_unmap_single(dev->dev.parent, bd->skb_pa, rskb->len,
-                                DMA_FROM_DEVICE);
+                                           (struct bcom_bd **)&bd);
+               physaddr = bd->skb_pa;
 
                /* Test for errors in received frame */
                if (status & BCOM_FEC_RX_BD_ERRORS) {
                        /* Drop packet and reuse the buffer */
-                       bd = (struct bcom_fec_bd *)
-                               bcom_prepare_next_buffer(priv->rx_dmatsk);
-
-                       bd->status = FEC_RX_BUFFER_SIZE;
-                       bd->skb_pa = dma_map_single(dev->dev.parent,
-                                       rskb->data,
-                                       FEC_RX_BUFFER_SIZE, DMA_FROM_DEVICE);
-
-                       bcom_submit_next_buffer(priv->rx_dmatsk, rskb);
-
+                       mpc52xx_fec_rx_submit(dev, rskb);
                        dev->stats.rx_dropped++;
-
                        continue;
                }
 
                /* skbs are allocated on open, so now we allocate a new one,
                 * and remove the old (with the packet) */
                skb = dev_alloc_skb(FEC_RX_BUFFER_SIZE);
-               if (skb) {
-                       /* Process the received skb */
-                       int length = status & BCOM_FEC_RX_BD_LEN_MASK;
-
-                       skb_put(rskb, length - 4);      /* length without CRC32 */
-
-                       rskb->dev = dev;
-                       rskb->protocol = eth_type_trans(rskb, dev);
-
-                       netif_rx(rskb);
-               } else {
+               if (!skb) {
                        /* Can't get a new one : reuse the same & drop pkt */
-                       dev_notice(&dev->dev, "Memory squeeze, dropping packet.\n");
+                       dev_notice(&dev->dev, "Low memory - dropped packet.\n");
+                       mpc52xx_fec_rx_submit(dev, rskb);
                        dev->stats.rx_dropped++;
-
-                       skb = rskb;
+                       continue;
                }
 
-               bd = (struct bcom_fec_bd *)
-                       bcom_prepare_next_buffer(priv->rx_dmatsk);
+               /* Enqueue the new sk_buff back on the hardware */
+               mpc52xx_fec_rx_submit(dev, skb);
 
-               bd->status = FEC_RX_BUFFER_SIZE;
-               bd->skb_pa = dma_map_single(dev->dev.parent, skb->data,
-                               FEC_RX_BUFFER_SIZE, DMA_FROM_DEVICE);
+               /* Process the received skb - Drop the spin lock while
+                * calling into the network stack */
+               spin_unlock_irqrestore(&priv->lock, flags);
 
-               bcom_submit_next_buffer(priv->rx_dmatsk, skb);
+               dma_unmap_single(dev->dev.parent, physaddr, rskb->len,
+                                DMA_FROM_DEVICE);
+               length = status & BCOM_FEC_RX_BD_LEN_MASK;
+               skb_put(rskb, length - 4);      /* length without CRC32 */
+               rskb->dev = dev;
+               rskb->protocol = eth_type_trans(rskb, dev);
+               netif_rx(rskb);
+
+               spin_lock_irqsave(&priv->lock, flags);
        }
 
+       spin_unlock_irqrestore(&priv->lock, flags);
+
        return IRQ_HANDLED;
 }
 
@@ -454,6 +453,7 @@ static irqreturn_t mpc52xx_fec_interrupt(int irq, void *dev_id)
        struct mpc52xx_fec_priv *priv = netdev_priv(dev);
        struct mpc52xx_fec __iomem *fec = priv->fec;
        u32 ievent;
+       unsigned long flags;
 
        ievent = in_be32(&fec->ievent);
 
@@ -471,9 +471,10 @@ static irqreturn_t mpc52xx_fec_interrupt(int irq, void *dev_id)
                if (net_ratelimit() && (ievent & FEC_IEVENT_XFIFO_ERROR))
                        dev_warn(&dev->dev, "FEC_IEVENT_XFIFO_ERROR\n");
 
+               spin_lock_irqsave(&priv->lock, flags);
                mpc52xx_fec_reset(dev);
+               spin_unlock_irqrestore(&priv->lock, flags);
 
-               netif_wake_queue(dev);
                return IRQ_HANDLED;
        }
 
@@ -768,6 +769,8 @@ static void mpc52xx_fec_reset(struct net_device *dev)
        bcom_enable(priv->tx_dmatsk);
 
        mpc52xx_fec_start(dev);
+
+       netif_wake_queue(dev);
 }