tg3: Enforce DMA mapping / skb assignment ordering

Michael Chan noted that there is nothing in the code that would prevent
the compiler from delaying the access of the "mapping" member of the
newly arrived packet until much later.  If this happened after the
skb = NULL assignment, it is possible for the driver to pass a bad
dma_addr value to pci_unmap_single().  To enforce this ordering, we need
a write memory barrier.  The pairing read memory barrier already exists
in tg3_rx_prodring_xfer() under the comments starting with
"Ensure that updates to the...".

Signed-off-by: Matt Carlson <mcarlson@broadcom.com>
Signed-off-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/drivers/net/tg3.c b/drivers/net/tg3.c
index 385434ff3960bcf32ac0fb24a9232060972f6f63..344490e872b4d34071b5ce9f40358070911a220c 100644 (file)
--- a/drivers/net/tg3.c
+++ b/drivers/net/tg3.c
@@ -4659,11 +4659,16 @@ static int tg3_rx(struct tg3_napi *tnapi, int budget)
                        if (skb_size < 0)
                                goto drop_it;
 
-                       ri->skb = NULL;
-
                        pci_unmap_single(tp->pdev, dma_addr, skb_size,
                                         PCI_DMA_FROMDEVICE);
 
+                       /* Ensure that the update to the skb happens
+                        * after the usage of the old DMA mapping.
+                        */
+                       smp_wmb();
+
+                       ri->skb = NULL;
+
                        skb_put(skb, len);
                } else {
                        struct sk_buff *copy_skb;