[PATCH] mv643xx_eth: Fix handling of small, unaligned fragments
authorPaul Janzen <pcj@linux.sez.to>
Mon, 16 Jan 2006 23:52:13 +0000 (16:52 -0700)
committerJeff Garzik <jgarzik@pobox.com>
Tue, 17 Jan 2006 12:23:37 +0000 (07:23 -0500)
Fix handling of small, unaligned fragments.
It also solves a potential deadlock if skb_linearize() returns -ENOMEM.

Signed-off-by: Paul Janzen <pcj@linux.sez.to>
Signed-off-by: Dale Farnsworth <dale@farnsworth.org>
 mv643xx_eth.c |   54 +++++++++++++++++++++++++++++++-----------------------
 1 file changed, 31 insertions(+), 23 deletions(-)
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
drivers/net/mv643xx_eth.c

index 4afb954092a6a35bdf5c0a2c75900597e3e92f55..e01b03c7615eb9e2ad76c6268381cd7db1cb94b0 100644 (file)
@@ -1093,6 +1093,25 @@ static int mv643xx_poll(struct net_device *dev, int *budget)
 }
 #endif
 
+/* Hardware can't handle unaligned fragments smaller than 9 bytes.
+ * This helper function detects that case.
+ */
+
+static inline unsigned int has_tiny_unaligned_frags(struct sk_buff *skb)
+{
+        unsigned int frag;
+        skb_frag_t *fragp;
+
+        for (frag = 0; frag < skb_shinfo(skb)->nr_frags; frag++) {
+                fragp = &skb_shinfo(skb)->frags[frag];
+                if (fragp->size <= 8 && fragp->page_offset & 0x7)
+                        return 1;
+
+        }
+        return 0;
+}
+
+
 /*
  * mv643xx_eth_start_xmit
  *
@@ -1136,12 +1155,19 @@ static int mv643xx_eth_start_xmit(struct sk_buff *skb, struct net_device *dev)
                return 1;
        }
 
+#ifdef MV643XX_CHECKSUM_OFFLOAD_TX
+       if (has_tiny_unaligned_frags(skb)) {
+               if ((skb_linearize(skb, GFP_ATOMIC) != 0)) {
+                       stats->tx_dropped++;
+                       printk(KERN_DEBUG "%s: failed to linearize tiny "
+                                       "unaligned fragment\n", dev->name);
+                       return 1;
+               }
+       }
+
        spin_lock_irqsave(&mp->lock, flags);
 
-       /* Update packet info data structure -- DMA owned, first last */
-#ifdef MV643XX_CHECKSUM_OFFLOAD_TX
        if (!skb_shinfo(skb)->nr_frags) {
-linear:
                if (skb->ip_summed != CHECKSUM_HW) {
                        /* Errata BTS #50, IHL must be 5 if no HW checksum */
                        pkt_info.cmd_sts = ETH_TX_ENABLE_INTERRUPT |
@@ -1183,26 +1209,6 @@ linear:
        } else {
                unsigned int frag;
 
-               /* Since hardware can't handle unaligned fragments smaller
-                * than 9 bytes, if we find any, we linearize the skb
-                * and start again.  When I've seen it, it's always been
-                * the first frag (probably near the end of the page),
-                * but we check all frags to be safe.
-                */
-               for (frag = 0; frag < skb_shinfo(skb)->nr_frags; frag++) {
-                       skb_frag_t *fragp;
-
-                       fragp = &skb_shinfo(skb)->frags[frag];
-                       if (fragp->size <= 8 && fragp->page_offset & 0x7) {
-                               skb_linearize(skb, GFP_ATOMIC);
-                               printk(KERN_DEBUG "%s: unaligned tiny fragment"
-                                               "%d of %d, fixed\n",
-                                               dev->name, frag,
-                                               skb_shinfo(skb)->nr_frags);
-                               goto linear;
-                       }
-               }
-
                /* first frag which is skb header */
                pkt_info.byte_cnt = skb_headlen(skb);
                pkt_info.buf_ptr = dma_map_single(NULL, skb->data,
@@ -1288,6 +1294,8 @@ linear:
                }
        }
 #else
+       spin_lock_irqsave(&mp->lock, flags);
+
        pkt_info.cmd_sts = ETH_TX_ENABLE_INTERRUPT | ETH_TX_FIRST_DESC |
                                                        ETH_TX_LAST_DESC;
        pkt_info.l4i_chk = 0;