bug in ixgbe_atr

-----Original Message-----
From: Sowmini Varadhan [mailto:[email protected]]
Sent: Thursday, October 13, 2016 6:44 PM
To: Duyck, Alexander H <[email protected]>; [email protected]
Subject: bug in ixgbe_atr

When I was playing around with TPACKET_V2, I think I ran into a bug in ixgbe_atr(): if I get here with an sk_buff that has, e.g., just 14 bytes in the header, then the skb_network_header is invalid. I found my kernel sometimes wandering off into ipv6_find_hdr() in an attempt to get the l4_proto, and then complaining that "IPv6 header not found\n"
(this was an ipv4 packet).

I think we want to use skb_header_pointer in ixgbe_atr to get the network header itself.. I tried the patch below, and it works for simple (non-vlan, basic) ethernet header, but probably needs more refinement to work for more complex encapsulations?

And other drivers may need a similar fix too, I've not checked yet.

--Sowmini

drivers/net/ethernet/intel/ixgbe/ixgbe_main.c +11 -5

index a244d9a..be453c6 100644
@@ -7627,6 +7627,10 @@ static void ixgbe_atr(struct ixgbe_ring *ring,
 		struct iphdr *ipv4;
 		struct ipv6hdr *ipv6;
 	} hdr;
+	union {
+		struct iphdr ipv4;
+		struct ipv6hdr ipv6;
+	} ip_hdr;
 	struct tcphdr *th;
 	unsigned int hlen;
 	struct sk_buff *skb;

@@ -7667,13 +7671,15 @@ static void ixgbe_atr(struct ixgbe_ring *ring,
 	}
 
 	/* Currently only IPv4/IPv6 with TCP is supported */
-	switch (hdr.ipv4->version) {
-	case IPVERSION:
+	switch (ntohs(first->protocol)) {
+	case ETH_P_IP:
+		skb_header_pointer(skb, ETH_HLEN, sizeof (struct iphdr),
+				   &ip_hdr);
 		/* access ihl as u8 to avoid unaligned access on ia64 */
-		hlen = (hdr.network[0] & 0x0F) << 2;
-		l4_proto = hdr.ipv4->protocol;
+		hlen = ip_hdr.ipv4.ihl << 2;
+		l4_proto = ip_hdr.ipv4.protocol;
 		break;
-	case 6:
+	case ETH_P_IPV6:
 		hlen = hdr.network - skb->data;
 		l4_proto = ipv6_find_hdr(skb, &hlen, IPPROTO_TCP, NULL,
 NULL);

hlen -= hdr.network - skb->data;

drivers/net/ethernet/intel/ixgbe/ixgbe_main.c static void ixgbe_atr(struct ixgbe_ring *ring, · line 7681

Quoted 9 lines

drivers/net/ethernet/intel/ixgbe/ixgbe_main.c +5 -2

+	case ETH_P_IP:
+		skb_header_pointer(skb, ETH_HLEN, sizeof (struct iphdr),
+				   &ip_hdr);
 		/* access ihl as u8 to avoid unaligned access on ia64 */
-		hlen = (hdr.network[0] & 0x0F) << 2;
-		l4_proto = hdr.ipv4->protocol;
+		hlen = ip_hdr.ipv4.ihl << 2;
+		l4_proto = ip_hdr.ipv4.protocol;
 		break;

The problem is this will break other stuff, for example I have seen the ihl access actually cause problems with unaligned accesses as some architectures decide to pull it as a u32 and then mask it.

My advice would be to keep this simple. Add a check to make sure we have room for at least skb_headlen(skb) - 40 >= hrd.raw - skb->data.

Messing with the protocol bits will break stuff since there is support for tunneling also floating around in here now.

I believe we are planning on dropping this code in favor of ndo_rx_flow_steer in the future. If we do that then the whole problem becomes moot.

-----Original Message-----
From: Sowmini Varadhan [mailto:[email protected]]
Sent: Thursday, October 13, 2016 8:49 PM
To: Duyck, Alexander H <[email protected]>
Cc: [email protected]
Subject: Re: bug in ixgbe_atr

Quoted 9 lines · Alexander H

drivers/net/ethernet/intel/ixgbe/ixgbe_main.c static void ixgbe_atr(struct ixgbe_ring *ring, · line 7681

Quoted 9 lines

drivers/net/ethernet/intel/ixgbe/ixgbe_main.c +5 -2

+	case ETH_P_IP:
+		skb_header_pointer(skb, ETH_HLEN, sizeof (struct iphdr),
+				   &ip_hdr);
 		/* access ihl as u8 to avoid unaligned access on ia64 */
-		hlen = (hdr.network[0] & 0x0F) << 2;
-		l4_proto = hdr.ipv4->protocol;
+		hlen = ip_hdr.ipv4.ihl << 2;
+		l4_proto = ip_hdr.ipv4.protocol;
 		break;

:

Quoted 3 lines

The problem is this will break other stuff, for example I have seen the ihl access actually cause problems with unaligned accesses as some architectures decide to pull it as a u32 and then mask it.

Yes, I noticed that u8 comment for ia64.. if that's the only issue here, we could just reset hdr.network to &ip_hdr..

However, I suspect the above patch is probably not going to work for the vlan case (it was just a first-pass hack)

Quoted 2 lines

My advice would be to keep this simple. Add a check to make sure we have room for at least skb_headlen(skb) - 40 >= hrd.raw - skb->data.

I don't parse that- the hdr union in ixgbe_atr doesnt have a ->raw field. Can you explain?

Quoted 6 lines

Messing with the protocol bits will break stuff since there is support for tunneling also floating around in here now.

I believe we are planning on dropping this code in favor of ndo_rx_flow_steer in the future. If we do that then the whole problem becomes moot.

Dropping it is fine with me I guess - maybe just return, if the skb_headlen() doesnt have enough bytes for a network header, i.e., skb_headlen is at least ETH_HLEN + sizeof (struct iphdr) for ETH_P_IP, or ETH_HLEN + sizeof (struct ipv6hdr) for ETH_P_IPV6?

--Sowmini

Sorry I was thinking of a different piece of code. In the case of the atr code it would be hdr.network, not hdr.raw. Basically the thought was to validate that there is enough data in skb_headlen() that we can verify that from where the network header should be we have at least 40 bytes of data as that would be the minimum needed for a TCP header and an IPv4 header, or just an IPv6 header. We would probably need a separate follow-up for the TCP header after we validate network header.

Quoted 5 lines

Dropping it is fine with me I guess - maybe just return, if the skb_headlen() doesnt have enough bytes for a network header, i.e., skb_headlen
is at least ETH_HLEN + sizeof (struct iphdr) for ETH_P_IP, or ETH_HLEN + sizeof (struct ipv6hdr) for ETH_P_IPV6?

Right that is kind of what I was thinking. If we validate that we have at least 40 before inspecting the network header, and at least 20 before we validate the TCP header that would work for me.