datapath/actions.c

   1 /*
   2  * Distributed under the terms of the GNU GPL version 2.
   3  * Copyright (c) 2007, 2008, 2009 Nicira Networks.
   4  */
   5
   6 /* Functions for executing flow actions. */
   7
   8 #include <linux/skbuff.h>
   9 #include <linux/in.h>
  10 #include <linux/ip.h>
  11 #include <linux/tcp.h>
  12 #include <linux/udp.h>
  13 #include <linux/in6.h>
  14 #include <linux/if_vlan.h>
  15 #include <net/ip.h>
  16 #include <net/checksum.h>
  17 #include "datapath.h"
  18 #include "dp_dev.h"
  19 #include "actions.h"
  20 #include "openvswitch/datapath-protocol.h"
  21
  22 struct sk_buff *
  23 make_writable(struct sk_buff *skb, gfp_t gfp)
  24 {
  25         if (skb_shared(skb) || skb_cloned(skb)) {
  26                 struct sk_buff *nskb = skb_copy(skb, gfp);
  27                 if (nskb) {
  28                         kfree_skb(skb);
  29                         return nskb;
  30                 }
  31         } else {
  32                 unsigned int hdr_len = (skb_transport_offset(skb)
  33                                         + sizeof(struct tcphdr));
  34                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  35                         return skb;
  36         }
  37         kfree_skb(skb);
  38         return NULL;
  39 }
  40
  41
  42 static struct sk_buff *
  43 vlan_pull_tag(struct sk_buff *skb)
  44 {
  45         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  46         struct ethhdr *eh;
  47
  48
  49         /* Verify we were given a vlan packet */
  50         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
  51                 return skb;
  52
  53         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  54
  55         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  56
  57         skb->protocol = eh->h_proto;
  58         skb->mac_header += VLAN_HLEN;
  59
  60         return skb;
  61 }
  62
  63
  64 static struct sk_buff *
  65 modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  66                 struct odp_flow_key *key, const union odp_action *a,
  67                 int n_actions, gfp_t gfp)
  68 {
  69         u16 tci, mask;
  70
  71         if (a->type == ODPAT_SET_VLAN_VID) {
  72                 tci = ntohs(a->vlan_vid.vlan_vid);
  73                 mask = VLAN_VID_MASK;
  74                 key->dl_vlan = htons(tci & mask);
  75         } else {
  76                 tci = a->vlan_pcp.vlan_pcp << 13;
  77                 mask = VLAN_PCP_MASK;
  78         }
  79
  80         skb = make_writable(skb, gfp);
  81         if (!skb)
  82                 return ERR_PTR(-ENOMEM);
  83
  84         if (skb->protocol == htons(ETH_P_8021Q)) {
  85                 /* Modify vlan id, but maintain other TCI values */
  86                 struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  87                 vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
  88         } else {
  89                 /* Add vlan header */
  90
  91                 /* Set up checksumming pointers for checksum-deferred packets
  92                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
  93                  * when we send the packet out on the wire, and it will fail at
  94                  * that point because skb_checksum_setup() will not look inside
  95                  * an 802.1Q header. */
  96                 skb_checksum_setup(skb);
  97
  98                 /* GSO is not implemented for packets with an 802.1Q header, so
  99                  * we have to do segmentation before we add that header.
 100                  *
 101                  * GSO does work with hardware-accelerated VLAN tagging, but we
 102                  * can't use hardware-accelerated VLAN tagging since it
 103                  * requires the device to have a VLAN group configured (with
 104                  * e.g. vconfig(8)) and we don't do that.
 105                  *
 106                  * Having to do this here may be a performance loss, since we
 107                  * can't take advantage of TSO hardware support, although it
 108                  * does not make a measurable network performance difference
 109                  * for 1G Ethernet.  Fixing that would require patching the
 110                  * kernel (either to add GSO support to the VLAN protocol or to
 111                  * support hardware-accelerated VLAN tagging without VLAN
 112                  * groups configured). */
 113                 if (skb_is_gso(skb)) {
 114                         struct sk_buff *segs;
 115
 116                         segs = skb_gso_segment(skb, 0);
 117                         kfree_skb(skb);
 118                         if (unlikely(IS_ERR(segs)))
 119                                 return ERR_CAST(segs);
 120
 121                         do {
 122                                 struct sk_buff *nskb = segs->next;
 123                                 int err;
 124
 125                                 segs->next = NULL;
 126
 127                                 segs = __vlan_put_tag(segs, tci);
 128                                 err = -ENOMEM;
 129                                 if (segs) {
 130                                         struct odp_flow_key segkey = *key;
 131                                         err = execute_actions(dp, segs,
 132                                                               &segkey, a + 1,
 133                                                               n_actions - 1,
 134                                                               gfp);
 135                                 }
 136
 137                                 if (unlikely(err)) {
 138                                         while ((segs = nskb)) {
 139                                                 nskb = segs->next;
 140                                                 segs->next = NULL;
 141                                                 kfree_skb(segs);
 142                                         }
 143                                         return ERR_PTR(err);
 144                                 }
 145
 146                                 segs = nskb;
 147                         } while (segs->next);
 148
 149                         skb = segs;
 150                 }
 151
 152                 /* The hardware-accelerated version of vlan_put_tag() works
 153                  * only for a device that has a VLAN group configured (with
 154                  * e.g. vconfig(8)), so call the software-only version
 155                  * __vlan_put_tag() directly instead.
 156                  */
 157                 skb = __vlan_put_tag(skb, tci);
 158                 if (!skb)
 159                         return ERR_PTR(-ENOMEM);
 160         }
 161
 162         return skb;
 163 }
 164
 165 static struct sk_buff *strip_vlan(struct sk_buff *skb,
 166                                   struct odp_flow_key *key, gfp_t gfp)
 167 {
 168         skb = make_writable(skb, gfp);
 169         if (skb) {
 170                 vlan_pull_tag(skb);
 171                 key->dl_vlan = htons(ODP_VLAN_NONE);
 172         }
 173         return skb;
 174 }
 175
 176 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 177                                    const struct odp_action_dl_addr *a,
 178                                    gfp_t gfp)
 179 {
 180         skb = make_writable(skb, gfp);
 181         if (skb) {
 182                 struct ethhdr *eh = eth_hdr(skb);
 183                 memcpy(a->type == ODPAT_SET_DL_SRC ? eh->h_source : eh->h_dest,
 184                        a->dl_addr, ETH_ALEN);
 185         }
 186         return skb;
 187 }
 188
 189 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 190  * covered by the sum has been changed from 'from' to 'to'.  If set,
 191  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 192  * Based on nf_proto_csum_replace4. */
 193 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 194                         __be32 from, __be32 to, int pseudohdr)
 195 {
 196         __be32 diff[] = { ~from, to };
 197         if (skb->ip_summed != CHECKSUM_PARTIAL) {
 198                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 199                                 ~csum_unfold(*sum)));
 200                 if (skb->ip_summed == CHECKSUM_COMPLETE && pseudohdr)
 201                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 202                                                 ~skb->csum);
 203         } else if (pseudohdr)
 204                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 205                                 csum_unfold(*sum)));
 206 }
 207
 208 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 209                                    struct odp_flow_key *key,
 210                                    const struct odp_action_nw_addr *a,
 211                                    gfp_t gfp)
 212 {
 213         if (key->dl_type != htons(ETH_P_IP))
 214                 return skb;
 215
 216         skb = make_writable(skb, gfp);
 217         if (skb) {
 218                 struct iphdr *nh = ip_hdr(skb);
 219                 u32 *f = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 220                 u32 old = *f;
 221                 u32 new = a->nw_addr;
 222
 223                 if (key->nw_proto == IPPROTO_TCP) {
 224                         struct tcphdr *th = tcp_hdr(skb);
 225                         update_csum(&th->check, skb, old, new, 1);
 226                 } else if (key->nw_proto == IPPROTO_UDP) {
 227                         struct udphdr *th = udp_hdr(skb);
 228                         update_csum(&th->check, skb, old, new, 1);
 229                 }
 230                 update_csum(&nh->check, skb, old, new, 0);
 231                 *f = new;
 232         }
 233         return skb;
 234 }
 235
 236 static struct sk_buff *
 237 set_tp_port(struct sk_buff *skb, struct odp_flow_key *key,
 238             const struct odp_action_tp_port *a,
 239             gfp_t gfp)
 240 {
 241         int check_ofs;
 242
 243         if (key->dl_type != htons(ETH_P_IP))
 244                 return skb;
 245
 246         if (key->nw_proto == IPPROTO_TCP)
 247                 check_ofs = offsetof(struct tcphdr, check);
 248         else if (key->nw_proto == IPPROTO_UDP)
 249                 check_ofs = offsetof(struct udphdr, check);
 250         else
 251                 return skb;
 252
 253         skb = make_writable(skb, gfp);
 254         if (skb) {
 255                 struct udphdr *th = udp_hdr(skb);
 256                 u16 *f = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 257                 u16 old = *f;
 258                 u16 new = a->tp_port;
 259                 update_csum((u16*)((u8*)skb->data + check_ofs),
 260                             skb, old, new, 1);
 261                 *f = new;
 262         }
 263         return skb;
 264 }
 265
 266 static inline unsigned packet_length(const struct sk_buff *skb)
 267 {
 268         unsigned length = skb->len - ETH_HLEN;
 269         if (skb->protocol == htons(ETH_P_8021Q))
 270                 length -= VLAN_HLEN;
 271         return length;
 272 }
 273
 274 int dp_xmit_skb(struct sk_buff *skb)
 275 {
 276         struct datapath *dp = skb->dev->br_port->dp;
 277         int len = skb->len;
 278
 279         if (packet_length(skb) > skb->dev->mtu && !skb_is_gso(skb)) {
 280                 printk(KERN_WARNING "%s: dropped over-mtu packet: %d > %d\n",
 281                        dp_name(dp), packet_length(skb), skb->dev->mtu);
 282                 kfree_skb(skb);
 283                 return -E2BIG;
 284         }
 285
 286         dev_queue_xmit(skb);
 287
 288         return len;
 289 }
 290
 291 static void
 292 do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 293 {
 294         struct net_bridge_port *p;
 295         struct net_device *dev;
 296
 297         if (!skb)
 298                 goto error;
 299
 300         p = dp->ports[out_port];
 301         if (!p)
 302                 goto error;
 303
 304         dev = skb->dev = p->dev;
 305         if (is_dp_dev(dev))
 306                 dp_dev_recv(dev, skb);
 307         else
 308                 dp_xmit_skb(skb);
 309         return;
 310
 311 error:
 312         kfree_skb(skb);
 313 }
 314
 315 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 316  * none.  */
 317 static int output_group(struct datapath *dp, __u16 group,
 318                         struct sk_buff *skb, gfp_t gfp)
 319 {
 320         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 321         int prev_port = -1;
 322         int i;
 323
 324         if (!g)
 325                 return -1;
 326         for (i = 0; i < g->n_ports; i++) {
 327                 struct net_bridge_port *p = dp->ports[g->ports[i]];
 328                 if (!p || skb->dev == p->dev)
 329                         continue;
 330                 if (prev_port != -1) {
 331                         struct sk_buff *clone = skb_clone(skb, gfp);
 332                         if (!clone)
 333                                 return -1;
 334                         do_output(dp, clone, prev_port);
 335                 }
 336                 prev_port = p->port_no;
 337         }
 338         return prev_port;
 339 }
 340
 341 static int
 342 output_control(struct datapath *dp, struct sk_buff *skb, u32 arg, gfp_t gfp)
 343 {
 344         skb = skb_clone(skb, gfp);
 345         if (!skb)
 346                 return -ENOMEM;
 347         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 348 }
 349
 350 /* Execute a list of actions against 'skb'. */
 351 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 352                     struct odp_flow_key *key,
 353                     const union odp_action *a, int n_actions,
 354                     gfp_t gfp)
 355 {
 356         /* Every output action needs a separate clone of 'skb', but the common
 357          * case is just a single output action, so that doing a clone and
 358          * then freeing the original skbuff is wasteful.  So the following code
 359          * is slightly obscure just to avoid that. */
 360         int prev_port = -1;
 361         int err = 0;
 362         for (; n_actions > 0; a++, n_actions--) {
 363                 WARN_ON_ONCE(skb_shared(skb));
 364                 if (prev_port != -1) {
 365                         do_output(dp, skb_clone(skb, gfp), prev_port);
 366                         prev_port = -1;
 367                 }
 368
 369                 switch (a->type) {
 370                 case ODPAT_OUTPUT:
 371                         prev_port = a->output.port;
 372                         break;
 373
 374                 case ODPAT_OUTPUT_GROUP:
 375                         prev_port = output_group(dp, a->output_group.group,
 376                                                  skb, gfp);
 377                         break;
 378
 379                 case ODPAT_CONTROLLER:
 380                         err = output_control(dp, skb, a->controller.arg, gfp);
 381                         if (err) {
 382                                 kfree_skb(skb);
 383                                 return err;
 384                         }
 385                         break;
 386
 387                 case ODPAT_SET_VLAN_VID:
 388                 case ODPAT_SET_VLAN_PCP:
 389                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 390                         if (IS_ERR(skb))
 391                                 return PTR_ERR(skb);
 392                         break;
 393
 394                 case ODPAT_STRIP_VLAN:
 395                         skb = strip_vlan(skb, key, gfp);
 396                         break;
 397
 398                 case ODPAT_SET_DL_SRC:
 399                 case ODPAT_SET_DL_DST:
 400                         skb = set_dl_addr(skb, &a->dl_addr, gfp);
 401                         break;
 402
 403                 case ODPAT_SET_NW_SRC:
 404                 case ODPAT_SET_NW_DST:
 405                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 406                         break;
 407
 408                 case ODPAT_SET_TP_SRC:
 409                 case ODPAT_SET_TP_DST:
 410                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 411                         break;
 412                 }
 413                 if (!skb)
 414                         return -ENOMEM;
 415         }
 416         if (prev_port != -1)
 417                 do_output(dp, skb, prev_port);
 418         else
 419                 kfree_skb(skb);
 420         return err;
 421 }