Implement QoS framework.
[sliver-openvswitch.git] / datapath / actions.c
index 4fc0a4a..ff67372 100644 (file)
 #include <linux/udp.h>
 #include <linux/in6.h>
 #include <linux/if_vlan.h>
+#include <net/inet_ecn.h>
 #include <net/ip.h>
 #include <net/checksum.h>
-#include "datapath.h"
-#include "dp_dev.h"
+
 #include "actions.h"
+#include "datapath.h"
 #include "openvswitch/datapath-protocol.h"
+#include "vport.h"
 
 static struct sk_buff *
 make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
@@ -31,19 +33,7 @@ make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
 
                nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), gfp);
                if (nskb) {
-#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,24)
-                       /* Before 2.6.24 these fields were not copied when
-                        * doing an skb_copy_expand. */
-                       nskb->ip_summed = skb->ip_summed;
-                       nskb->csum = skb->csum;
-#endif
-#if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
-                       /* These fields are copied in skb_clone but not in
-                        * skb_copy or related functions.  We need to manually
-                        * copy them over here. */
-                       nskb->proto_data_valid = skb->proto_data_valid;
-                       nskb->proto_csum_blank = skb->proto_csum_blank;
-#endif
+                       set_skb_csum_bits(skb, nskb);
                        kfree_skb(skb);
                        return nskb;
                }
@@ -57,6 +47,11 @@ make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
        return NULL;
 }
 
+static void set_tunnel(struct sk_buff *skb, struct odp_flow_key *key,
+                      __be32 tun_id)
+{
+       OVS_CB(skb)->tun_id = key->tun_id = tun_id;
+}
 
 static struct sk_buff *
 vlan_pull_tag(struct sk_buff *skb)
@@ -64,11 +59,14 @@ vlan_pull_tag(struct sk_buff *skb)
        struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
        struct ethhdr *eh;
 
-
        /* Verify we were given a vlan packet */
        if (vh->h_vlan_proto != htons(ETH_P_8021Q))
                return skb;
 
+       if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
+               skb->csum = csum_sub(skb->csum, csum_partial(skb->data
+                                       + ETH_HLEN, VLAN_HLEN, 0));
+
        memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
 
        eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
@@ -90,10 +88,11 @@ modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
        if (a->type == ODPAT_SET_VLAN_VID) {
                tci = ntohs(a->vlan_vid.vlan_vid);
                mask = VLAN_VID_MASK;
-               key->dl_vlan = htons(tci & mask);
+               key->dl_vlan = a->vlan_vid.vlan_vid;
        } else {
-               tci = a->vlan_pcp.vlan_pcp << 13;
+               tci = a->vlan_pcp.vlan_pcp << VLAN_PCP_SHIFT;
                mask = VLAN_PCP_MASK;
+               key->dl_vlan_pcp = a->vlan_pcp.vlan_pcp;
        }
 
        skb = make_writable(skb, VLAN_HLEN, gfp);
@@ -103,7 +102,16 @@ modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
        if (skb->protocol == htons(ETH_P_8021Q)) {
                /* Modify vlan id, but maintain other TCI values */
                struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
+               __be16 old_tci = vh->h_vlan_TCI;
+
                vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
+
+               if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE) {
+                       __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
+
+                       skb->csum = ~csum_partial((char *)diff, sizeof(diff),
+                                               ~skb->csum);
+               }
        } else {
                /* Add vlan header */
 
@@ -143,6 +151,9 @@ modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
 
                                segs->next = NULL;
 
+                               /* GSO can change the checksum type so update.*/
+                               compute_ip_summed(segs, true);
+
                                segs = __vlan_put_tag(segs, tci);
                                err = -ENOMEM;
                                if (segs) {
@@ -166,6 +177,7 @@ modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
                        } while (segs->next);
 
                        skb = segs;
+                       compute_ip_summed(skb, true);
                }
 
                /* The hardware-accelerated version of vlan_put_tag() works
@@ -176,6 +188,12 @@ modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
                skb = __vlan_put_tag(skb, tci);
                if (!skb)
                        return ERR_PTR(-ENOMEM);
+
+               /* GSO doesn't fix up the hardware computed checksum so this
+                * will only be hit in the non-GSO case. */
+               if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
+                       skb->csum = csum_add(skb->csum, csum_partial(skb->data
+                                               + ETH_HLEN, VLAN_HLEN, 0));
        }
 
        return skb;
@@ -193,14 +211,20 @@ static struct sk_buff *strip_vlan(struct sk_buff *skb,
 }
 
 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
+                                  struct odp_flow_key *key,
                                   const struct odp_action_dl_addr *a,
                                   gfp_t gfp)
 {
        skb = make_writable(skb, 0, gfp);
        if (skb) {
                struct ethhdr *eh = eth_hdr(skb);
-               memcpy(a->type == ODPAT_SET_DL_SRC ? eh->h_source : eh->h_dest,
-                      a->dl_addr, ETH_ALEN);
+               if (a->type == ODPAT_SET_DL_SRC) {
+                       memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
+                       memcpy(key->dl_src, a->dl_addr, ETH_ALEN);
+               } else {
+                       memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
+                       memcpy(key->dl_dst, a->dl_addr, ETH_ALEN);
+               }
        }
        return skb;
 }
@@ -214,42 +238,10 @@ static void update_csum(__sum16 *sum, struct sk_buff *skb,
 {
        __be32 diff[] = { ~from, to };
 
-/* On older kernels, CHECKSUM_PARTIAL and CHECKSUM_COMPLETE are both defined
- * as CHECKSUM_HW.  However, we can make some inferences so that we can update
- * the checksums appropriately. */
-       enum {
-               CSUM_PARTIAL,   /* Partial checksum, skb->csum undefined. */
-               CSUM_PACKET,    /* In-packet checksum, skb->csum undefined. */
-               CSUM_COMPLETE,  /* In-packet checksum, skb->csum valid. */
-       } csum_type;
-
-       csum_type = CSUM_PACKET;
-#ifndef CHECKSUM_HW
-       /* Newer kernel, just map between kernel types and ours. */
-       if (skb->ip_summed == CHECKSUM_PARTIAL)
-               csum_type = CSUM_PARTIAL;
-       else if (skb->ip_summed == CHECKSUM_COMPLETE)
-               csum_type = CSUM_COMPLETE;
-#else
-       /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
-        * However, we should only get CHECKSUM_PARTIAL packets from Xen, which
-        * uses some special fields to represent this (see below).  Since we
-        * can only make one type work, pick the one that actually happens in
-        * practice. */
-       if (skb->ip_summed == CHECKSUM_HW)
-               csum_type = CSUM_COMPLETE;
-#endif
-#if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
-       /* Xen has a special way of representing CHECKSUM_PARTIAL on older
-        * kernels. */
-       if (skb->proto_csum_blank)
-               csum_type = CSUM_PARTIAL;
-#endif
-
-       if (csum_type != CSUM_PARTIAL) {
+       if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
                *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
                                ~csum_unfold(*sum)));
-               if (csum_type == CSUM_COMPLETE && pseudohdr)
+               if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
                        skb->csum = ~csum_partial((char *)diff, sizeof(diff),
                                                ~skb->csum);
        } else if (pseudohdr)
@@ -281,6 +273,36 @@ static struct sk_buff *set_nw_addr(struct sk_buff *skb,
                }
                update_csum(&nh->check, skb, old, new, 0);
                *f = new;
+
+               if (a->type == ODPAT_SET_NW_SRC)
+                       key->nw_src = a->nw_addr;
+               else
+                       key->nw_dst = a->nw_addr;
+       }
+       return skb;
+}
+
+static struct sk_buff *set_nw_tos(struct sk_buff *skb,
+                                  struct odp_flow_key *key,
+                                  const struct odp_action_nw_tos *a,
+                                  gfp_t gfp)
+{
+       if (key->dl_type != htons(ETH_P_IP))
+               return skb;
+
+       skb = make_writable(skb, 0, gfp);
+       if (skb) {
+               struct iphdr *nh = ip_hdr(skb);
+               u8 *f = &nh->tos;
+               u8 old = *f;
+               u8 new;
+
+               /* Set the DSCP bits and preserve the ECN bits. */
+               new = a->nw_tos | (nh->tos & INET_ECN_MASK);
+               update_csum(&nh->check, skb, htons((uint16_t)old),
+                               htons((uint16_t)new), 0);
+               *f = new;
+               key->nw_tos = a->nw_tos;
        }
        return skb;
 }
@@ -311,6 +333,10 @@ set_tp_port(struct sk_buff *skb, struct odp_flow_key *key,
                update_csum((u16*)(skb_transport_header(skb) + check_ofs), 
                                skb, old, new, 0);
                *f = new;
+               if (a->type == ODPAT_SET_TP_SRC)
+                       key->tp_src = a->tp_port;
+               else
+                       key->tp_dst = a->tp_port;
        }
        return skb;
 }
@@ -323,42 +349,27 @@ static inline unsigned packet_length(const struct sk_buff *skb)
        return length;
 }
 
-int dp_xmit_skb(struct sk_buff *skb)
-{
-       struct datapath *dp = skb->dev->br_port->dp;
-       int len = skb->len;
-
-       if (packet_length(skb) > skb->dev->mtu && !skb_is_gso(skb)) {
-               printk(KERN_WARNING "%s: dropped over-mtu packet: %d > %d\n",
-                      dp_name(dp), packet_length(skb), skb->dev->mtu);
-               kfree_skb(skb);
-               return -E2BIG;
-       }
-
-       forward_ip_summed(skb);
-       dev_queue_xmit(skb);
-
-       return len;
-}
-
 static void
 do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 {
-       struct net_bridge_port *p;
-       struct net_device *dev;
+       struct dp_port *p;
+       int mtu;
 
        if (!skb)
                goto error;
 
-       p = dp->ports[out_port];
+       p = rcu_dereference(dp->ports[out_port]);
        if (!p)
                goto error;
 
-       dev = skb->dev = p->dev;
-       if (is_dp_dev(dev))
-               dp_dev_recv(dev, skb);
-       else
-               dp_xmit_skb(skb);
+       mtu = vport_get_mtu(p->vport);
+       if (packet_length(skb) > mtu && !skb_is_gso(skb)) {
+               printk(KERN_WARNING "%s: dropped over-mtu packet: %d > %d\n",
+                      dp_name(dp), packet_length(skb), mtu);
+               goto error;
+       }
+
+       vport_send(p->vport, skb);
        return;
 
 error:
@@ -377,8 +388,8 @@ static int output_group(struct datapath *dp, __u16 group,
        if (!g)
                return -1;
        for (i = 0; i < g->n_ports; i++) {
-               struct net_bridge_port *p = dp->ports[g->ports[i]];
-               if (!p || skb->dev == p->dev)
+               struct dp_port *p = rcu_dereference(dp->ports[g->ports[i]]);
+               if (!p || OVS_CB(skb)->dp_port == p)
                        continue;
                if (prev_port != -1) {
                        struct sk_buff *clone = skb_clone(skb, gfp);
@@ -404,7 +415,7 @@ output_control(struct datapath *dp, struct sk_buff *skb, u32 arg, gfp_t gfp)
  * information about what happened to it. */
 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
                         const union odp_action *a, int n_actions,
-                        gfp_t gfp, struct net_bridge_port *nbp)
+                        gfp_t gfp, struct dp_port *dp_port)
 {
        struct odp_sflow_sample_header *hdr;
        unsigned int actlen = n_actions * sizeof(union odp_action);
@@ -418,7 +429,7 @@ static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
        memcpy(__skb_push(nskb, actlen), a, actlen);
        hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
        hdr->n_actions = n_actions;
-       hdr->sample_pool = atomic_read(&nbp->sflow_pool);
+       hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
        dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 }
 
@@ -433,10 +444,11 @@ int execute_actions(struct datapath *dp, struct sk_buff *skb,
         * then freeing the original skbuff is wasteful.  So the following code
         * is slightly obscure just to avoid that. */
        int prev_port = -1;
+       u32 priority = skb->priority;
        int err;
 
        if (dp->sflow_probability) {
-               struct net_bridge_port *p = skb->dev->br_port;
+               struct dp_port *p = OVS_CB(skb)->dp_port;
                if (p) {
                        atomic_inc(&p->sflow_pool);
                        if (dp->sflow_probability == UINT_MAX ||
@@ -445,6 +457,8 @@ int execute_actions(struct datapath *dp, struct sk_buff *skb,
                }
        }
 
+       OVS_CB(skb)->tun_id = 0;
+
        for (; n_actions > 0; a++, n_actions--) {
                WARN_ON_ONCE(skb_shared(skb));
                if (prev_port != -1) {
@@ -470,6 +484,10 @@ int execute_actions(struct datapath *dp, struct sk_buff *skb,
                        }
                        break;
 
+               case ODPAT_SET_TUNNEL:
+                       set_tunnel(skb, key, a->tunnel.tun_id);
+                       break;
+
                case ODPAT_SET_VLAN_VID:
                case ODPAT_SET_VLAN_PCP:
                        skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
@@ -483,7 +501,7 @@ int execute_actions(struct datapath *dp, struct sk_buff *skb,
 
                case ODPAT_SET_DL_SRC:
                case ODPAT_SET_DL_DST:
-                       skb = set_dl_addr(skb, &a->dl_addr, gfp);
+                       skb = set_dl_addr(skb, key, &a->dl_addr, gfp);
                        break;
 
                case ODPAT_SET_NW_SRC:
@@ -491,10 +509,22 @@ int execute_actions(struct datapath *dp, struct sk_buff *skb,
                        skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
                        break;
 
+               case ODPAT_SET_NW_TOS:
+                       skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
+                       break;
+
                case ODPAT_SET_TP_SRC:
                case ODPAT_SET_TP_DST:
                        skb = set_tp_port(skb, key, &a->tp_port, gfp);
                        break;
+
+               case ODPAT_SET_PRIORITY:
+                       skb->priority = a->priority.priority;
+                       break;
+
+               case ODPAT_POP_PRIORITY:
+                       skb->priority = priority;
+                       break;
                }
                if (!skb)
                        return -ENOMEM;