xen-netback: add support for IPv6 checksum offload from guest
authorPaul Durrant <Paul.Durrant@citrix.com>
Wed, 16 Oct 2013 16:50:29 +0000 (17:50 +0100)
committerDavid S. Miller <davem@davemloft.net>
Thu, 17 Oct 2013 19:35:16 +0000 (15:35 -0400)
For performance of VM to VM traffic on a single host it is better to avoid
calculation of TCP/UDP checksum in the sending frontend. To allow this this
patch adds the code necessary to set up partial checksum for IPv6 packets
and xenstore flag feature-ipv6-csum-offload to advertise that fact to
frontends.

Signed-off-by: Paul Durrant <paul.durrant@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/xen-netback/netback.c
drivers/net/xen-netback/xenbus.c

index f3e591c611ded744ec0b6ea2cefe9b7dd3d568c7..4271f8d1da7ad3cb6df996c41da50ba0deeb98d0 100644 (file)
@@ -109,15 +109,12 @@ static inline unsigned long idx_to_kaddr(struct xenvif *vif,
        return (unsigned long)pfn_to_kaddr(idx_to_pfn(vif, idx));
 }
 
-/*
- * This is the amount of packet we copy rather than map, so that the
- * guest can't fiddle with the contents of the headers while we do
- * packet processing on them (netfilter, routing, etc).
+/* This is a miniumum size for the linear area to avoid lots of
+ * calls to __pskb_pull_tail() as we set up checksum offsets. The
+ * value 128 was chosen as it covers all IPv4 and most likely
+ * IPv6 headers.
  */
-#define PKT_PROT_LEN    (ETH_HLEN + \
-                        VLAN_HLEN + \
-                        sizeof(struct iphdr) + MAX_IPOPTLEN + \
-                        sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
+#define PKT_PROT_LEN 128
 
 static u16 frag_get_pending_idx(skb_frag_t *frag)
 {
@@ -1118,61 +1115,74 @@ static int xenvif_set_skb_gso(struct xenvif *vif,
        return 0;
 }
 
-static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
+static inline void maybe_pull_tail(struct sk_buff *skb, unsigned int len)
+{
+       if (skb_is_nonlinear(skb) && skb_headlen(skb) < len) {
+               /* If we need to pullup then pullup to the max, so we
+                * won't need to do it again.
+                */
+               int target = min_t(int, skb->len, MAX_TCP_HEADER);
+               __pskb_pull_tail(skb, target - skb_headlen(skb));
+       }
+}
+
+static int checksum_setup_ip(struct xenvif *vif, struct sk_buff *skb,
+                            int recalculate_partial_csum)
 {
-       struct iphdr *iph;
+       struct iphdr *iph = (void *)skb->data;
+       unsigned int header_size;
+       unsigned int off;
        int err = -EPROTO;
-       int recalculate_partial_csum = 0;
 
-       /*
-        * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
-        * peers can fail to set NETRXF_csum_blank when sending a GSO
-        * frame. In this case force the SKB to CHECKSUM_PARTIAL and
-        * recalculate the partial checksum.
-        */
-       if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
-               vif->rx_gso_checksum_fixup++;
-               skb->ip_summed = CHECKSUM_PARTIAL;
-               recalculate_partial_csum = 1;
-       }
+       off = sizeof(struct iphdr);
 
-       /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
-       if (skb->ip_summed != CHECKSUM_PARTIAL)
-               return 0;
+       header_size = skb->network_header + off + MAX_IPOPTLEN;
+       maybe_pull_tail(skb, header_size);
 
-       if (skb->protocol != htons(ETH_P_IP))
-               goto out;
+       off = iph->ihl * 4;
 
-       iph = (void *)skb->data;
        switch (iph->protocol) {
        case IPPROTO_TCP:
-               if (!skb_partial_csum_set(skb, 4 * iph->ihl,
+               if (!skb_partial_csum_set(skb, off,
                                          offsetof(struct tcphdr, check)))
                        goto out;
 
                if (recalculate_partial_csum) {
                        struct tcphdr *tcph = tcp_hdr(skb);
+
+                       header_size = skb->network_header +
+                               off +
+                               sizeof(struct tcphdr);
+                       maybe_pull_tail(skb, header_size);
+
                        tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
-                                                        skb->len - iph->ihl*4,
+                                                        skb->len - off,
                                                         IPPROTO_TCP, 0);
                }
                break;
        case IPPROTO_UDP:
-               if (!skb_partial_csum_set(skb, 4 * iph->ihl,
+               if (!skb_partial_csum_set(skb, off,
                                          offsetof(struct udphdr, check)))
                        goto out;
 
                if (recalculate_partial_csum) {
                        struct udphdr *udph = udp_hdr(skb);
+
+                       header_size = skb->network_header +
+                               off +
+                               sizeof(struct udphdr);
+                       maybe_pull_tail(skb, header_size);
+
                        udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
-                                                        skb->len - iph->ihl*4,
+                                                        skb->len - off,
                                                         IPPROTO_UDP, 0);
                }
                break;
        default:
                if (net_ratelimit())
                        netdev_err(vif->dev,
-                                  "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
+                                  "Attempting to checksum a non-TCP/UDP packet, "
+                                  "dropping a protocol %d packet\n",
                                   iph->protocol);
                goto out;
        }
@@ -1183,6 +1193,158 @@ out:
        return err;
 }
 
+static int checksum_setup_ipv6(struct xenvif *vif, struct sk_buff *skb,
+                              int recalculate_partial_csum)
+{
+       int err = -EPROTO;
+       struct ipv6hdr *ipv6h = (void *)skb->data;
+       u8 nexthdr;
+       unsigned int header_size;
+       unsigned int off;
+       bool fragment;
+       bool done;
+
+       done = false;
+
+       off = sizeof(struct ipv6hdr);
+
+       header_size = skb->network_header + off;
+       maybe_pull_tail(skb, header_size);
+
+       nexthdr = ipv6h->nexthdr;
+
+       while ((off <= sizeof(struct ipv6hdr) + ntohs(ipv6h->payload_len)) &&
+              !done) {
+               switch (nexthdr) {
+               case IPPROTO_DSTOPTS:
+               case IPPROTO_HOPOPTS:
+               case IPPROTO_ROUTING: {
+                       struct ipv6_opt_hdr *hp = (void *)(skb->data + off);
+
+                       header_size = skb->network_header +
+                               off +
+                               sizeof(struct ipv6_opt_hdr);
+                       maybe_pull_tail(skb, header_size);
+
+                       nexthdr = hp->nexthdr;
+                       off += ipv6_optlen(hp);
+                       break;
+               }
+               case IPPROTO_AH: {
+                       struct ip_auth_hdr *hp = (void *)(skb->data + off);
+
+                       header_size = skb->network_header +
+                               off +
+                               sizeof(struct ip_auth_hdr);
+                       maybe_pull_tail(skb, header_size);
+
+                       nexthdr = hp->nexthdr;
+                       off += (hp->hdrlen+2)<<2;
+                       break;
+               }
+               case IPPROTO_FRAGMENT:
+                       fragment = true;
+                       /* fall through */
+               default:
+                       done = true;
+                       break;
+               }
+       }
+
+       if (!done) {
+               if (net_ratelimit())
+                       netdev_err(vif->dev, "Failed to parse packet header\n");
+               goto out;
+       }
+
+       if (fragment) {
+               if (net_ratelimit())
+                       netdev_err(vif->dev, "Packet is a fragment!\n");
+               goto out;
+       }
+
+       switch (nexthdr) {
+       case IPPROTO_TCP:
+               if (!skb_partial_csum_set(skb, off,
+                                         offsetof(struct tcphdr, check)))
+                       goto out;
+
+               if (recalculate_partial_csum) {
+                       struct tcphdr *tcph = tcp_hdr(skb);
+
+                       header_size = skb->network_header +
+                               off +
+                               sizeof(struct tcphdr);
+                       maybe_pull_tail(skb, header_size);
+
+                       tcph->check = ~csum_ipv6_magic(&ipv6h->saddr,
+                                                      &ipv6h->daddr,
+                                                      skb->len - off,
+                                                      IPPROTO_TCP, 0);
+               }
+               break;
+       case IPPROTO_UDP:
+               if (!skb_partial_csum_set(skb, off,
+                                         offsetof(struct udphdr, check)))
+                       goto out;
+
+               if (recalculate_partial_csum) {
+                       struct udphdr *udph = udp_hdr(skb);
+
+                       header_size = skb->network_header +
+                               off +
+                               sizeof(struct udphdr);
+                       maybe_pull_tail(skb, header_size);
+
+                       udph->check = ~csum_ipv6_magic(&ipv6h->saddr,
+                                                      &ipv6h->daddr,
+                                                      skb->len - off,
+                                                      IPPROTO_UDP, 0);
+               }
+               break;
+       default:
+               if (net_ratelimit())
+                       netdev_err(vif->dev,
+                                  "Attempting to checksum a non-TCP/UDP packet, "
+                                  "dropping a protocol %d packet\n",
+                                  nexthdr);
+               goto out;
+       }
+
+       err = 0;
+
+out:
+       return err;
+}
+
+static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
+{
+       int err = -EPROTO;
+       int recalculate_partial_csum = 0;
+
+       /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
+        * peers can fail to set NETRXF_csum_blank when sending a GSO
+        * frame. In this case force the SKB to CHECKSUM_PARTIAL and
+        * recalculate the partial checksum.
+        */
+       if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
+               vif->rx_gso_checksum_fixup++;
+               skb->ip_summed = CHECKSUM_PARTIAL;
+               recalculate_partial_csum = 1;
+       }
+
+       /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
+       if (skb->ip_summed != CHECKSUM_PARTIAL)
+               return 0;
+
+       if (skb->protocol == htons(ETH_P_IP))
+               err = checksum_setup_ip(vif, skb, recalculate_partial_csum);
+       else if (skb->protocol == htons(ETH_P_IPV6))
+               err = checksum_setup_ipv6(vif, skb, recalculate_partial_csum);
+
+       return err;
+}
+
 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
 {
        unsigned long now = jiffies;
@@ -1428,12 +1590,7 @@ static int xenvif_tx_submit(struct xenvif *vif, int budget)
 
                xenvif_fill_frags(vif, skb);
 
-               /*
-                * If the initial fragment was < PKT_PROT_LEN then
-                * pull through some bytes from the other fragments to
-                * increase the linear region to PKT_PROT_LEN bytes.
-                */
-               if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
+               if (skb_is_nonlinear(skb) && skb_headlen(skb) < PKT_PROT_LEN) {
                        int target = min_t(int, skb->len, PKT_PROT_LEN);
                        __pskb_pull_tail(skb, target - skb_headlen(skb));
                }
index ad27b15242cdae90785bf7530fd7ac5ce1c89729..108e7523017aa0d81d0f9bcdd5c7a416637b4e36 100644 (file)
@@ -105,6 +105,15 @@ static int netback_probe(struct xenbus_device *dev,
                        goto abort_transaction;
                }
 
+               /* We support partial checksum setup for IPv6 packets */
+               err = xenbus_printf(xbt, dev->nodename,
+                                   "feature-ipv6-csum-offload",
+                                   "%d", 1);
+               if (err) {
+                       message = "writing feature-ipv6-csum-offload";
+                       goto abort_transaction;
+               }
+
                /* We support rx-copy path. */
                err = xenbus_printf(xbt, dev->nodename,
                                    "feature-rx-copy", "%d", 1);