Re: [PATCH net-next] ibmveth: v1 calculate correct gso_size and set gso_type

On Thu, Nov 3, 2016 at 8:40 AM, Brian King [off-list ref] wrote:

On 10/27/2016 10:26 AM, Eric Dumazet wrote:

quoted

On Wed, 2016-10-26 at 11:09 +1100, Jon Maxwell wrote:

quoted

We recently encountered a bug where a few customers using ibmveth on the
same LPAR hit an issue where a TCP session hung when large receive was
enabled. Closer analysis revealed that the session was stuck because the
one side was advertising a zero window repeatedly.

We narrowed this down to the fact the ibmveth driver did not set gso_size
which is translated by TCP into the MSS later up the stack. The MSS is
used to calculate the TCP window size and as that was abnormally large,
it was calculating a zero window, even although the sockets receive buffer
was completely empty.

We were able to reproduce this and worked with IBM to fix this. Thanks Tom
and Marcelo for all your help and review on this.

The patch fixes both our internal reproduction tests and our customers tests.

Signed-off-by: Jon Maxwell <redacted>
---
 drivers/net/ethernet/ibm/ibmveth.c | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/drivers/net/ethernet/ibm/ibmveth.c b/drivers/net/ethernet/ibm/ibmveth.c
index 29c05d0..c51717e 100644
--- a/drivers/net/ethernet/ibm/ibmveth.c
+++ b/drivers/net/ethernet/ibm/ibmveth.c

@@ -1182,6 +1182,8 @@ static int ibmveth_poll(struct napi_struct *napi, int budget)
     int frames_processed = 0;
     unsigned long lpar_rc;
     struct iphdr *iph;
+    bool large_packet = 0;
+    u16 hdr_len = ETH_HLEN + sizeof(struct tcphdr);

 restart_poll:
     while (frames_processed < budget) {

@@ -1236,10 +1238,28 @@ static int ibmveth_poll(struct napi_struct *napi, int budget)
                                             iph->check = 0;
                                             iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
                                             adapter->rx_large_packets++;
+                                            large_packet = 1;
                                     }
                             }
                     }

+                    if (skb->len > netdev->mtu) {
+                            iph = (struct iphdr *)skb->data;
+                            if (be16_to_cpu(skb->protocol) == ETH_P_IP &&
+                                iph->protocol == IPPROTO_TCP) {
+                                    hdr_len += sizeof(struct iphdr);
+                                    skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
+                                    skb_shinfo(skb)->gso_size = netdev->mtu - hdr_len;
+                            } else if (be16_to_cpu(skb->protocol) == ETH_P_IPV6 &&
+                                       iph->protocol == IPPROTO_TCP) {
+                                    hdr_len += sizeof(struct ipv6hdr);
+                                    skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
+                                    skb_shinfo(skb)->gso_size = netdev->mtu - hdr_len;
+                            }
+                            if (!large_packet)
+                                    adapter->rx_large_packets++;
+                    }
+

This might break forwarding and PMTU discovery.

You force gso_size to device mtu, regardless of real MSS used by the TCP
sender.

Don't you have the MSS provided in RX descriptor, instead of guessing
the value ?

We've had some further discussions on this with the Virtual I/O Server (VIOS)
development team. The large receive aggregation in the VIOS (AIX based) is actually
being done by software in the VIOS. What they may be able to do is when performing
this aggregation, they could look at the packet lengths of all the packets being
aggregated and take the largest packet size within the aggregation unit, minus the
header length and return that to the virtual ethernet client which we could then stuff
into gso_size. They are currently assessing how feasible this would be to do and whether
it would impact other bits of the code. However, assuming this does end up being an option,
would this address the concerns here or is that going to break something else I'm
not thinking of?

I was discussing this with a colleague and although this is better than
what we have so far. We wonder if there could be a corner case where
it ends up with a smaller value than the current MSS. For example if
the application sent a burst of small TCP packets with the PUSH
bit set. In that case they may not be coalesced by GRO. The VIOS could
probably be coded to detect that condition and use the previous MSS.
But that may not necessarily be the current MSS.

The ibmveth driver passes the MSS via gso_size to the VIOS. Either as the
3rd argument of ibmveth_send() or via tcp_hdr(skb)->check which is
presumably over-written when the VIOS does the TSO. Would it be possible
to keep a copy of this value on the TX side on the VIOS before it over-written
and then some how pass that up to the RX side along with frame and set
gso_size to that which should be the current MSS?

Regards

Jon

Unfortunately, I don't think we'd have a good way to get gso_segs set correctly as I don't
see how that would get passed back up the interface.

Thanks,

Brian


--
Brian King
Power Linux I/O
IBM Linux Technology Center