bonding/alb: properly access headers in bond_alb_xmit()

syzbot managed to send an IPX packet through bond_alb_xmit()
and af_packet and triggered a use-after-free.

First, bond_alb_xmit() was using ipx_hdr() helper to reach
the IPX header, but ipx_hdr() was using the transport offset
instead of the network offset. In the particular syzbot
report transport offset was 0xFFFF

This patch removes ipx_hdr() since it was only (mis)used from bonding.

Then we need to make sure IPv4/IPv6/IPX headers are pulled
in skb->head before dereferencing anything.

BUG: KASAN: use-after-free in bond_alb_xmit+0x153a/0x1590 drivers/net/bonding/bond_alb.c:1452
Read of size 2 at addr ffff8801ce56dfff by task syz-executor.2/18108
 (if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) ...)

Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Call Trace:
 [<ffffffff8441fc42>] __dump_stack lib/dump_stack.c:17 [inline]
 [<ffffffff8441fc42>] dump_stack+0x14d/0x20b lib/dump_stack.c:53
 [<ffffffff81a7dec4>] print_address_description+0x6f/0x20b mm/kasan/report.c:282
 [<ffffffff81a7e0ec>] kasan_report_error mm/kasan/report.c:380 [inline]
 [<ffffffff81a7e0ec>] kasan_report mm/kasan/report.c:438 [inline]
 [<ffffffff81a7e0ec>] kasan_report.cold+0x8c/0x2a0 mm/kasan/report.c:422
 [<ffffffff81a7dc4f>] __asan_report_load_n_noabort+0xf/0x20 mm/kasan/report.c:469
 [<ffffffff82c8c00a>] bond_alb_xmit+0x153a/0x1590 drivers/net/bonding/bond_alb.c:1452
 [<ffffffff82c60c74>] __bond_start_xmit drivers/net/bonding/bond_main.c:4199 [inline]
 [<ffffffff82c60c74>] bond_start_xmit+0x4f4/0x1570 drivers/net/bonding/bond_main.c:4224
 [<ffffffff83baa558>] __netdev_start_xmit include/linux/netdevice.h:4525 [inline]
 [<ffffffff83baa558>] netdev_start_xmit include/linux/netdevice.h:4539 [inline]
 [<ffffffff83baa558>] xmit_one net/core/dev.c:3611 [inline]
 [<ffffffff83baa558>] dev_hard_start_xmit+0x168/0x910 net/core/dev.c:3627
 [<ffffffff83bacf35>] __dev_queue_xmit+0x1f55/0x33b0 net/core/dev.c:4238
 [<ffffffff83bae3a8>] dev_queue_xmit+0x18/0x20 net/core/dev.c:4278
 [<ffffffff84339189>] packet_snd net/packet/af_packet.c:3226 [inline]
 [<ffffffff84339189>] packet_sendmsg+0x4919/0x70b0 net/packet/af_packet.c:3252
 [<ffffffff83b1ac0c>] sock_sendmsg_nosec net/socket.c:673 [inline]
 [<ffffffff83b1ac0c>] sock_sendmsg+0x12c/0x160 net/socket.c:684
 [<ffffffff83b1f5a2>] __sys_sendto+0x262/0x380 net/socket.c:1996
 [<ffffffff83b1f700>] SYSC_sendto net/socket.c:2008 [inline]
 [<ffffffff83b1f700>] SyS_sendto+0x40/0x60 net/socket.c:2004

Fixes: 1da177e4 ("Linux-2.6.12-rc2")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Cc: Jay Vosburgh <j.vosburgh@gmail.com>
Cc: Veaceslav Falico <vfalico@gmail.com>
Cc: Andy Gospodarek <andy@greyhouse.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/bonding/bond_alb.c

@@ -1383,26 +1383,31 @@ netdev_tx_t bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 	bool do_tx_balance = true;
 	u32 hash_index = 0;
 	const u8 *hash_start = NULL;
-	struct ipv6hdr *ip6hdr;
 
 	skb_reset_mac_header(skb);
 	eth_data = eth_hdr(skb);
 
 	switch (ntohs(skb->protocol)) {
 	case ETH_P_IP: {
-		const struct iphdr *iph = ip_hdr(skb);
+		const struct iphdr *iph;
 
 		if (is_broadcast_ether_addr(eth_data->h_dest) ||
-		    iph->daddr == ip_bcast ||
-		    iph->protocol == IPPROTO_IGMP) {
+		    !pskb_network_may_pull(skb, sizeof(*iph))) {
+			do_tx_balance = false;
+			break;
+		}
+		iph = ip_hdr(skb);
+		if (iph->daddr == ip_bcast || iph->protocol == IPPROTO_IGMP) {
 			do_tx_balance = false;
 			break;
 		}
 		hash_start = (char *)&(iph->daddr);
 		hash_size = sizeof(iph->daddr);
-	}
 		break;
-	case ETH_P_IPV6:
+	}
+	case ETH_P_IPV6: {
+		const struct ipv6hdr *ip6hdr;
+
 		/* IPv6 doesn't really use broadcast mac address, but leave
 		 * that here just in case.
 		 */
@@ -1419,7 +1424,11 @@ netdev_tx_t bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 			break;
 		}
 
-		/* Additianally, DAD probes should not be tx-balanced as that
+		if (!pskb_network_may_pull(skb, sizeof(*ip6hdr))) {
+			do_tx_balance = false;
+			break;
+		}
+		/* Additionally, DAD probes should not be tx-balanced as that
 		 * will lead to false positives for duplicate addresses and
 		 * prevent address configuration from working.
 		 */
@@ -1429,17 +1438,26 @@ netdev_tx_t bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 			break;
 		}
 
-		hash_start = (char *)&(ipv6_hdr(skb)->daddr);
-		hash_size = sizeof(ipv6_hdr(skb)->daddr);
+		hash_start = (char *)&ip6hdr->daddr;
+		hash_size = sizeof(ip6hdr->daddr);
 		break;
-	case ETH_P_IPX:
-		if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
+	}
+	case ETH_P_IPX: {
+		const struct ipxhdr *ipxhdr;
+
+		if (pskb_network_may_pull(skb, sizeof(*ipxhdr))) {
+			do_tx_balance = false;
+			break;
+		}
+		ipxhdr = (struct ipxhdr *)skb_network_header(skb);
+
+		if (ipxhdr->ipx_checksum != IPX_NO_CHECKSUM) {
 			/* something is wrong with this packet */
 			do_tx_balance = false;
 			break;
 		}
 
-		if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
+		if (ipxhdr->ipx_type != IPX_TYPE_NCP) {
 			/* The only protocol worth balancing in
 			 * this family since it has an "ARP" like
 			 * mechanism
@@ -1448,9 +1466,11 @@ netdev_tx_t bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 			break;
 		}
 
+		eth_data = eth_hdr(skb);
 		hash_start = (char *)eth_data->h_dest;
 		hash_size = ETH_ALEN;
 		break;
+	}
 	case ETH_P_ARP:
 		do_tx_balance = false;
 		if (bond_info->rlb_enabled)

include/net/ipx.h

@@ -47,11 +47,6 @@ struct ipxhdr {
 /* From af_ipx.c */
 extern int sysctl_ipx_pprop_broadcasting;
 
-static __inline__ struct ipxhdr *ipx_hdr(struct sk_buff *skb)
-{
-	return (struct ipxhdr *)skb_transport_header(skb);
-}
-
 struct ipx_interface {
 	/* IPX address */
 	__be32			if_netnum;