Skip to contentSkip to conversation

lan78xx: fix ip header misalignment

PATCH In review Yuriy M. Kaminskiy 2 participants 2 messages Jan 21, 2019 – Jan 27, 2019

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 lan78xx: fix ip header misalignment

lan78xx.c:rx_submit() allocates space for frame-to-be-received with netdev_alloc_skb_ip_align(), which misalign start of buffer by 2 bytes in expectation that frame will start from 14-byte ethernet header, then ip header; if start of buffer misaligned by 2 bytes, ip header will be 16-byte aligned.

Unfortunately, usb frame that is sent by lan78xx starts with another 10-byte header (lan78xx_rx(): rx_cmd_a/rx_cmd_b/rx_cmd_c), *then* follows ethernet header, and *then* ip header (which ends up being misaligned).

This issue was observed on arm platform (where misaligned 32-bit word access triggers exception and leaves traces in /proc/cpu/alignment, see
https://github.com/raspberrypi/linux/issues/2599 ; for me, about any ipv6 traffic that hits machine - ping -I eth0 ip6-allnodes, tcp/udp packets, etc triggered increase in this counter, with ip6_datagram_recv_common_ctl, icmpv6_echo_reply, etc as culprit).

If we just allocate skb data without any misalignment tricks, ip header will end up and at offset 24 (8-byte aligned).

Patch attached; runtime-tested with raspbian fork of stable/4.14.y [4.14.92] on Raspberry pi 3B+ (it is slightly different from mainline, but patch should not have any conflicts, all affected code is pretty same).

P.S. I'm not subscribed, please CC me on reply.

From 2bd6b0a11e222be2df97da948924c71bf13d7192 Mon Sep 17 00:00:00 2001
From: "Yuriy M. Kaminskiy" <[email protected]>
Date: Mon, 21 Jan 2019 02:51:24 +0300
Subject: [PATCH] lan78xx: fix ip header misalignment

As lan78xx prepends 10-byte header before ether_hdr, skb->data misalignment trick by netdev_alloc_skb_ip_align() made things worse (ip_hdr becomes always misaligned).

See https://github.com/raspberrypi/linux/issues/2599

---
 drivers/net/usb/lan78xx.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/net/usb/lan78xx.c +1 -1
index 9b782cdf8..d64b0d3b8 100644
@@ -3143,7 +3143,7 @@ static int rx_submit(struct lan78xx_net *dev, struct urb *urb, gfp_t flags)
 	size_t size = dev->rx_urb_size;
 	int ret = 0;
 
-	skb = netdev_alloc_skb_ip_align(dev->net, size);
+	skb = netdev_alloc_skb(dev->net, size);
 	if (!skb) {
 		usb_free_urb(urb);
 		return -ENOMEM;
Signature · 1 line
2.11.0
#1 lore raw

Hi Yuiry,
Thanks for the patch.
Changes look good to me.

Thanks,
-Raghu

Quoted 33 lines

-----Original Message-----
From: Yuriy M. Kaminskiy <[email protected]>
Sent: Monday, January 21, 2019 6:18 PM
To: [email protected]
Cc: Woojung Huh - C21699 <[email protected]>; UNGLinuxDriver <[email protected]>
Subject: [PATCH] lan78xx: fix ip header misalignment

lan78xx.c:rx_submit() allocates space for frame-to-be-received with netdev_alloc_skb_ip_align(), which misalign start of buffer by 2 bytes in expectation that frame will start from 14-byte ethernet header, then ip header; if start of buffer misaligned by 2 bytes, ip header will be 16-byte aligned.

Unfortunately, usb frame that is sent by lan78xx starts with another 10-byte header (lan78xx_rx(): rx_cmd_a/rx_cmd_b/rx_cmd_c), *then* follows ethernet header, and *then* ip header (which ends up being misaligned).

This issue was observed on arm platform (where misaligned 32-bit word access triggers exception and leaves traces in /proc/cpu/alignment, see
https://github.com/raspberrypi/linux/issues/2599 ; for me, about any ipv6 traffic that hits machine - ping -I eth0 ip6-allnodes, tcp/udp packets, etc triggered increase in this counter, with ip6_datagram_recv_common_ctl, icmpv6_echo_reply, etc as culprit).

If we just allocate skb data without any misalignment tricks, ip header will end up and at offset 24 (8-byte aligned).

Patch attached; runtime-tested with raspbian fork of stable/4.14.y [4.14.92] on Raspberry pi 3B+ (it is slightly different from mainline, but patch should not have any conflicts, all affected code is pretty same).

P.S. I'm not subscribed, please CC me on reply.