net: lan78xx: Fix race in tx pending skb size calculation

The skb size calculation in lan78xx_tx_bh is in race with the start_xmit,
which could lead to rare kernel oopses. So protect the whole skb walk with
a spin lock. As a benefit we can unlink the skb directly.

This patch was tested on Raspberry Pi 3B+

Link: raspberrypi/linux#2608
Fixes: 55d7de9 ("Microchip's LAN7800 family USB 2/3 to 10/100/1000 Ethernet")
Cc: stable <[email protected]>
Signed-off-by: Floris Bos <[email protected]>
Signed-off-by: Stefan Wahren <[email protected]>
Signed-off-by: David S. Miller <[email protected]>

drivers/net/usb/lan78xx.c

@@ -3344,6 +3344,7 @@ static void lan78xx_tx_bh(struct lan78xx_net *dev)
 	pkt_cnt = 0;
 	count = 0;
 	length = 0;
+	spin_lock_irqsave(&tqp->lock, flags);
 	for (skb = tqp->next; pkt_cnt < tqp->qlen; skb = skb->next) {
 		if (skb_is_gso(skb)) {
 			if (pkt_cnt) {
@@ -3352,7 +3353,8 @@ static void lan78xx_tx_bh(struct lan78xx_net *dev)
 			}
 			count = 1;
 			length = skb->len - TX_OVERHEAD;
-			skb2 = skb_dequeue(tqp);
+			__skb_unlink(skb, tqp);
+			spin_unlock_irqrestore(&tqp->lock, flags);
 			goto gso_skb;
 		}
 
@@ -3361,6 +3363,7 @@ static void lan78xx_tx_bh(struct lan78xx_net *dev)
 		skb_totallen = skb->len + roundup(skb_totallen, sizeof(u32));
 		pkt_cnt++;
 	}
+	spin_unlock_irqrestore(&tqp->lock, flags);
 
 	/* copy to a single skb */
 	skb = alloc_skb(skb_totallen, GFP_ATOMIC);