datapath/actions.c

   1 /*
   2  * Distributed under the terms of the GNU GPL version 2.
   3  * Copyright (c) 2007, 2008, 2009, 2010 Nicira Networks.
   4  *
   5  * Significant portions of this file may be copied from parts of the Linux
   6  * kernel, by Linus Torvalds and others.
   7  */
   8
   9 /* Functions for executing flow actions. */
  10
  11 #include <linux/skbuff.h>
  12 #include <linux/in.h>
  13 #include <linux/ip.h>
  14 #include <linux/tcp.h>
  15 #include <linux/udp.h>
  16 #include <linux/in6.h>
  17 #include <linux/if_vlan.h>
  18 #include <net/ip.h>
  19 #include <net/checksum.h>
  20 #include "datapath.h"
  21 #include "dp_dev.h"
  22 #include "actions.h"
  23 #include "openvswitch/datapath-protocol.h"
  24
  25 static struct sk_buff *
  26 make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
  27 {
  28         if (skb_shared(skb) || skb_cloned(skb)) {
  29                 struct sk_buff *nskb;
  30                 unsigned headroom = max(min_headroom, skb_headroom(skb));
  31
  32                 nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), gfp);
  33                 if (nskb) {
  34 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,24)
  35                         /* Before 2.6.24 these fields were not copied when
  36                          * doing an skb_copy_expand. */
  37                         nskb->ip_summed = skb->ip_summed;
  38                         nskb->csum = skb->csum;
  39 #endif
  40 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
  41                         /* These fields are copied in skb_clone but not in
  42                          * skb_copy or related functions.  We need to manually
  43                          * copy them over here. */
  44                         nskb->proto_data_valid = skb->proto_data_valid;
  45                         nskb->proto_csum_blank = skb->proto_csum_blank;
  46 #endif
  47                         kfree_skb(skb);
  48                         return nskb;
  49                 }
  50         } else {
  51                 unsigned int hdr_len = (skb_transport_offset(skb)
  52                                         + sizeof(struct tcphdr));
  53                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  54                         return skb;
  55         }
  56         kfree_skb(skb);
  57         return NULL;
  58 }
  59
  60
  61 static struct sk_buff *
  62 vlan_pull_tag(struct sk_buff *skb)
  63 {
  64         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  65         struct ethhdr *eh;
  66
  67
  68         /* Verify we were given a vlan packet */
  69         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
  70                 return skb;
  71
  72         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  73
  74         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  75
  76         skb->protocol = eh->h_proto;
  77         skb->mac_header += VLAN_HLEN;
  78
  79         return skb;
  80 }
  81
  82
  83 static struct sk_buff *
  84 modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  85                 struct odp_flow_key *key, const union odp_action *a,
  86                 int n_actions, gfp_t gfp)
  87 {
  88         u16 tci, mask;
  89
  90         if (a->type == ODPAT_SET_VLAN_VID) {
  91                 tci = ntohs(a->vlan_vid.vlan_vid);
  92                 mask = VLAN_VID_MASK;
  93                 key->dl_vlan = htons(tci & mask);
  94         } else {
  95                 tci = a->vlan_pcp.vlan_pcp << 13;
  96                 mask = VLAN_PCP_MASK;
  97         }
  98
  99         skb = make_writable(skb, VLAN_HLEN, gfp);
 100         if (!skb)
 101                 return ERR_PTR(-ENOMEM);
 102
 103         if (skb->protocol == htons(ETH_P_8021Q)) {
 104                 /* Modify vlan id, but maintain other TCI values */
 105                 struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
 106                 vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
 107         } else {
 108                 /* Add vlan header */
 109
 110                 /* Set up checksumming pointers for checksum-deferred packets
 111                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 112                  * when we send the packet out on the wire, and it will fail at
 113                  * that point because skb_checksum_setup() will not look inside
 114                  * an 802.1Q header. */
 115                 vswitch_skb_checksum_setup(skb);
 116
 117                 /* GSO is not implemented for packets with an 802.1Q header, so
 118                  * we have to do segmentation before we add that header.
 119                  *
 120                  * GSO does work with hardware-accelerated VLAN tagging, but we
 121                  * can't use hardware-accelerated VLAN tagging since it
 122                  * requires the device to have a VLAN group configured (with
 123                  * e.g. vconfig(8)) and we don't do that.
 124                  *
 125                  * Having to do this here may be a performance loss, since we
 126                  * can't take advantage of TSO hardware support, although it
 127                  * does not make a measurable network performance difference
 128                  * for 1G Ethernet.  Fixing that would require patching the
 129                  * kernel (either to add GSO support to the VLAN protocol or to
 130                  * support hardware-accelerated VLAN tagging without VLAN
 131                  * groups configured). */
 132                 if (skb_is_gso(skb)) {
 133                         struct sk_buff *segs;
 134
 135                         segs = skb_gso_segment(skb, 0);
 136                         kfree_skb(skb);
 137                         if (unlikely(IS_ERR(segs)))
 138                                 return ERR_CAST(segs);
 139
 140                         do {
 141                                 struct sk_buff *nskb = segs->next;
 142                                 int err;
 143
 144                                 segs->next = NULL;
 145
 146                                 segs = __vlan_put_tag(segs, tci);
 147                                 err = -ENOMEM;
 148                                 if (segs) {
 149                                         struct odp_flow_key segkey = *key;
 150                                         err = execute_actions(dp, segs,
 151                                                               &segkey, a + 1,
 152                                                               n_actions - 1,
 153                                                               gfp);
 154                                 }
 155
 156                                 if (unlikely(err)) {
 157                                         while ((segs = nskb)) {
 158                                                 nskb = segs->next;
 159                                                 segs->next = NULL;
 160                                                 kfree_skb(segs);
 161                                         }
 162                                         return ERR_PTR(err);
 163                                 }
 164
 165                                 segs = nskb;
 166                         } while (segs->next);
 167
 168                         skb = segs;
 169                 }
 170
 171                 /* The hardware-accelerated version of vlan_put_tag() works
 172                  * only for a device that has a VLAN group configured (with
 173                  * e.g. vconfig(8)), so call the software-only version
 174                  * __vlan_put_tag() directly instead.
 175                  */
 176                 skb = __vlan_put_tag(skb, tci);
 177                 if (!skb)
 178                         return ERR_PTR(-ENOMEM);
 179         }
 180
 181         return skb;
 182 }
 183
 184 static struct sk_buff *strip_vlan(struct sk_buff *skb,
 185                                   struct odp_flow_key *key, gfp_t gfp)
 186 {
 187         skb = make_writable(skb, 0, gfp);
 188         if (skb) {
 189                 vlan_pull_tag(skb);
 190                 key->dl_vlan = htons(ODP_VLAN_NONE);
 191         }
 192         return skb;
 193 }
 194
 195 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 196                                    const struct odp_action_dl_addr *a,
 197                                    gfp_t gfp)
 198 {
 199         skb = make_writable(skb, 0, gfp);
 200         if (skb) {
 201                 struct ethhdr *eh = eth_hdr(skb);
 202                 memcpy(a->type == ODPAT_SET_DL_SRC ? eh->h_source : eh->h_dest,
 203                        a->dl_addr, ETH_ALEN);
 204         }
 205         return skb;
 206 }
 207
 208 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 209  * covered by the sum has been changed from 'from' to 'to'.  If set,
 210  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 211  * Based on nf_proto_csum_replace4. */
 212 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 213                         __be32 from, __be32 to, int pseudohdr)
 214 {
 215         __be32 diff[] = { ~from, to };
 216
 217 /* On older kernels, CHECKSUM_PARTIAL and CHECKSUM_COMPLETE are both defined
 218  * as CHECKSUM_HW.  However, we can make some inferences so that we can update
 219  * the checksums appropriately. */
 220         enum {
 221                 CSUM_PARTIAL,   /* Partial checksum, skb->csum undefined. */
 222                 CSUM_PACKET,    /* In-packet checksum, skb->csum undefined. */
 223                 CSUM_COMPLETE,  /* In-packet checksum, skb->csum valid. */
 224         } csum_type;
 225
 226         csum_type = CSUM_PACKET;
 227 #ifndef CHECKSUM_HW
 228         /* Newer kernel, just map between kernel types and ours. */
 229         if (skb->ip_summed == CHECKSUM_PARTIAL)
 230                 csum_type = CSUM_PARTIAL;
 231         else if (skb->ip_summed == CHECKSUM_COMPLETE)
 232                 csum_type = CSUM_COMPLETE;
 233 #else
 234         /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
 235          * However, we should only get CHECKSUM_PARTIAL packets from Xen, which
 236          * uses some special fields to represent this (see below).  Since we
 237          * can only make one type work, pick the one that actually happens in
 238          * practice. */
 239         if (skb->ip_summed == CHECKSUM_HW)
 240                 csum_type = CSUM_COMPLETE;
 241 #endif
 242 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
 243         /* Xen has a special way of representing CHECKSUM_PARTIAL on older
 244          * kernels. */
 245         if (skb->proto_csum_blank)
 246                 csum_type = CSUM_PARTIAL;
 247 #endif
 248
 249         if (csum_type != CSUM_PARTIAL) {
 250                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 251                                 ~csum_unfold(*sum)));
 252                 if (csum_type == CSUM_COMPLETE && pseudohdr)
 253                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 254                                                 ~skb->csum);
 255         } else if (pseudohdr)
 256                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 257                                 csum_unfold(*sum)));
 258 }
 259
 260 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 261                                    struct odp_flow_key *key,
 262                                    const struct odp_action_nw_addr *a,
 263                                    gfp_t gfp)
 264 {
 265         if (key->dl_type != htons(ETH_P_IP))
 266                 return skb;
 267
 268         skb = make_writable(skb, 0, gfp);
 269         if (skb) {
 270                 struct iphdr *nh = ip_hdr(skb);
 271                 u32 *f = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 272                 u32 old = *f;
 273                 u32 new = a->nw_addr;
 274
 275                 if (key->nw_proto == IPPROTO_TCP) {
 276                         struct tcphdr *th = tcp_hdr(skb);
 277                         update_csum(&th->check, skb, old, new, 1);
 278                 } else if (key->nw_proto == IPPROTO_UDP) {
 279                         struct udphdr *th = udp_hdr(skb);
 280                         update_csum(&th->check, skb, old, new, 1);
 281                 }
 282                 update_csum(&nh->check, skb, old, new, 0);
 283                 *f = new;
 284         }
 285         return skb;
 286 }
 287
 288 static struct sk_buff *
 289 set_tp_port(struct sk_buff *skb, struct odp_flow_key *key,
 290             const struct odp_action_tp_port *a,
 291             gfp_t gfp)
 292 {
 293         int check_ofs;
 294
 295         if (key->dl_type != htons(ETH_P_IP))
 296                 return skb;
 297
 298         if (key->nw_proto == IPPROTO_TCP)
 299                 check_ofs = offsetof(struct tcphdr, check);
 300         else if (key->nw_proto == IPPROTO_UDP)
 301                 check_ofs = offsetof(struct udphdr, check);
 302         else
 303                 return skb;
 304
 305         skb = make_writable(skb, 0, gfp);
 306         if (skb) {
 307                 struct udphdr *th = udp_hdr(skb);
 308                 u16 *f = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 309                 u16 old = *f;
 310                 u16 new = a->tp_port;
 311                 update_csum((u16*)(skb_transport_header(skb) + check_ofs),
 312                                 skb, old, new, 0);
 313                 *f = new;
 314         }
 315         return skb;
 316 }
 317
 318 static inline unsigned packet_length(const struct sk_buff *skb)
 319 {
 320         unsigned length = skb->len - ETH_HLEN;
 321         if (skb->protocol == htons(ETH_P_8021Q))
 322                 length -= VLAN_HLEN;
 323         return length;
 324 }
 325
 326 int dp_xmit_skb(struct sk_buff *skb)
 327 {
 328         struct datapath *dp = skb->dev->br_port->dp;
 329         int len = skb->len;
 330
 331         if (packet_length(skb) > skb->dev->mtu && !skb_is_gso(skb)) {
 332                 printk(KERN_WARNING "%s: dropped over-mtu packet: %d > %d\n",
 333                        dp_name(dp), packet_length(skb), skb->dev->mtu);
 334                 kfree_skb(skb);
 335                 return -E2BIG;
 336         }
 337
 338         forward_ip_summed(skb);
 339         dev_queue_xmit(skb);
 340
 341         return len;
 342 }
 343
 344 static void
 345 do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 346 {
 347         struct net_bridge_port *p;
 348         struct net_device *dev;
 349
 350         if (!skb)
 351                 goto error;
 352
 353         p = dp->ports[out_port];
 354         if (!p)
 355                 goto error;
 356
 357         dev = skb->dev = p->dev;
 358         if (is_dp_dev(dev))
 359                 dp_dev_recv(dev, skb);
 360         else
 361                 dp_xmit_skb(skb);
 362         return;
 363
 364 error:
 365         kfree_skb(skb);
 366 }
 367
 368 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 369  * none.  */
 370 static int output_group(struct datapath *dp, __u16 group,
 371                         struct sk_buff *skb, gfp_t gfp)
 372 {
 373         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 374         int prev_port = -1;
 375         int i;
 376
 377         if (!g)
 378                 return -1;
 379         for (i = 0; i < g->n_ports; i++) {
 380                 struct net_bridge_port *p = dp->ports[g->ports[i]];
 381                 if (!p || skb->dev == p->dev)
 382                         continue;
 383                 if (prev_port != -1) {
 384                         struct sk_buff *clone = skb_clone(skb, gfp);
 385                         if (!clone)
 386                                 return -1;
 387                         do_output(dp, clone, prev_port);
 388                 }
 389                 prev_port = p->port_no;
 390         }
 391         return prev_port;
 392 }
 393
 394 static int
 395 output_control(struct datapath *dp, struct sk_buff *skb, u32 arg, gfp_t gfp)
 396 {
 397         skb = skb_clone(skb, gfp);
 398         if (!skb)
 399                 return -ENOMEM;
 400         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 401 }
 402
 403 /* Send a copy of this packet up to the sFlow agent, along with extra
 404  * information about what happened to it. */
 405 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 406                          const union odp_action *a, int n_actions,
 407                          gfp_t gfp, struct net_bridge_port *nbp)
 408 {
 409         struct odp_sflow_sample_header *hdr;
 410         unsigned int actlen = n_actions * sizeof(union odp_action);
 411         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 412         struct sk_buff *nskb;
 413
 414         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 415         if (!nskb)
 416                 return;
 417
 418         memcpy(__skb_push(nskb, actlen), a, actlen);
 419         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 420         hdr->n_actions = n_actions;
 421         hdr->sample_pool = atomic_read(&nbp->sflow_pool);
 422         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 423 }
 424
 425 /* Execute a list of actions against 'skb'. */
 426 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 427                     struct odp_flow_key *key,
 428                     const union odp_action *a, int n_actions,
 429                     gfp_t gfp)
 430 {
 431         /* Every output action needs a separate clone of 'skb', but the common
 432          * case is just a single output action, so that doing a clone and
 433          * then freeing the original skbuff is wasteful.  So the following code
 434          * is slightly obscure just to avoid that. */
 435         int prev_port = -1;
 436         int err;
 437
 438         if (dp->sflow_probability) {
 439                 struct net_bridge_port *p = skb->dev->br_port;
 440                 if (p) {
 441                         atomic_inc(&p->sflow_pool);
 442                         if (dp->sflow_probability == UINT_MAX ||
 443                             net_random() < dp->sflow_probability)
 444                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 445                 }
 446         }
 447
 448         for (; n_actions > 0; a++, n_actions--) {
 449                 WARN_ON_ONCE(skb_shared(skb));
 450                 if (prev_port != -1) {
 451                         do_output(dp, skb_clone(skb, gfp), prev_port);
 452                         prev_port = -1;
 453                 }
 454
 455                 switch (a->type) {
 456                 case ODPAT_OUTPUT:
 457                         prev_port = a->output.port;
 458                         break;
 459
 460                 case ODPAT_OUTPUT_GROUP:
 461                         prev_port = output_group(dp, a->output_group.group,
 462                                                  skb, gfp);
 463                         break;
 464
 465                 case ODPAT_CONTROLLER:
 466                         err = output_control(dp, skb, a->controller.arg, gfp);
 467                         if (err) {
 468                                 kfree_skb(skb);
 469                                 return err;
 470                         }
 471                         break;
 472
 473                 case ODPAT_SET_VLAN_VID:
 474                 case ODPAT_SET_VLAN_PCP:
 475                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 476                         if (IS_ERR(skb))
 477                                 return PTR_ERR(skb);
 478                         break;
 479
 480                 case ODPAT_STRIP_VLAN:
 481                         skb = strip_vlan(skb, key, gfp);
 482                         break;
 483
 484                 case ODPAT_SET_DL_SRC:
 485                 case ODPAT_SET_DL_DST:
 486                         skb = set_dl_addr(skb, &a->dl_addr, gfp);
 487                         break;
 488
 489                 case ODPAT_SET_NW_SRC:
 490                 case ODPAT_SET_NW_DST:
 491                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 492                         break;
 493
 494                 case ODPAT_SET_TP_SRC:
 495                 case ODPAT_SET_TP_DST:
 496                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 497                         break;
 498                 }
 499                 if (!skb)
 500                         return -ENOMEM;
 501         }
 502         if (prev_port != -1)
 503                 do_output(dp, skb, prev_port);
 504         else
 505                 kfree_skb(skb);
 506         return 0;
 507 }