datapath/datapath.c

   1 /*
   2  * Distributed under the terms of the GNU GPL version 2.
   3  * Copyright (c) 2007, 2008 The Board of Trustees of The Leland
   4  * Stanford Junior University
   5  */
   6
   7 /* Functions for managing the dp interface/device. */
   8
   9 #include <linux/init.h>
  10 #include <linux/module.h>
  11 #include <linux/if_arp.h>
  12 #include <linux/if_bridge.h>
  13 #include <linux/if_vlan.h>
  14 #include <linux/in.h>
  15 #include <net/genetlink.h>
  16 #include <linux/ip.h>
  17 #include <linux/delay.h>
  18 #include <linux/etherdevice.h>
  19 #include <linux/kernel.h>
  20 #include <linux/kthread.h>
  21 #include <linux/mutex.h>
  22 #include <linux/rtnetlink.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/version.h>
  25 #include <linux/ethtool.h>
  26 #include <linux/random.h>
  27 #include <asm/system.h>
  28 #include <linux/netfilter_bridge.h>
  29 #include <linux/inetdevice.h>
  30 #include <linux/list.h>
  31
  32 #include "openflow-netlink.h"
  33 #include "datapath.h"
  34 #include "table.h"
  35 #include "chain.h"
  36 #include "forward.h"
  37 #include "flow.h"
  38 #include "datapath_t.h"
  39
  40 #include "compat.h"
  41
  42
  43 /* Number of milliseconds between runs of the maintenance thread. */
  44 #define MAINT_SLEEP_MSECS 1000
  45
  46 #define BRIDGE_PORT_NO_FLOOD    0x00000001
  47
  48 #define UINT32_MAX                        4294967295U
  49 #define MAX(X, Y) ((X) > (Y) ? (X) : (Y))
  50
  51 struct net_bridge_port {
  52         u16     port_no;
  53         u32 flags;
  54         struct datapath *dp;
  55         struct net_device *dev;
  56         struct list_head node; /* Element in datapath.ports. */
  57 };
  58
  59 static struct genl_family dp_genl_family;
  60 static struct genl_multicast_group mc_group;
  61
  62 int dp_dev_setup(struct net_device *dev);
  63
  64 /* It's hard to imagine wanting more than one datapath, but... */
  65 #define DP_MAX 32
  66
  67 /* datapaths.  Protected on the read side by rcu_read_lock, on the write side
  68  * by dp_mutex.
  69  *
  70  * It is safe to access the datapath and net_bridge_port structures with just
  71  * the dp_mutex, but to access the chain you need to take the rcu_read_lock
  72  * also (because dp_mutex doesn't prevent flows from being destroyed).
  73  */
  74 static struct datapath *dps[DP_MAX];
  75 static DEFINE_MUTEX(dp_mutex);
  76
  77 static int dp_maint_func(void *data);
  78 static int send_port_status(struct net_bridge_port *p, uint8_t status);
  79
  80
  81 /* nla_unreserve - reduce amount of space reserved by nla_reserve
  82  * @skb: socket buffer from which to recover room
  83  * @nla: netlink attribute to adjust
  84  * @len: amount by which to reduce attribute payload
  85  *
  86  * Reduces amount of space reserved by a call to nla_reserve.
  87  *
  88  * No other attributes may be added between calling nla_reserve and this
  89  * function, since it will create a hole in the message.
  90  */
  91 void nla_unreserve(struct sk_buff *skb, struct nlattr *nla, int len)
  92 {
  93         skb->tail -= len;
  94         skb->len  -= len;
  95
  96         nla->nla_len -= len;
  97 }
  98
  99 static void *
 100 alloc_openflow_skb(struct datapath *dp, size_t openflow_len, uint8_t type,
 101                    const struct sender *sender, struct sk_buff **pskb)
 102 {
 103         size_t genl_len;
 104         struct sk_buff *skb;
 105         struct nlattr *attr;
 106         struct ofp_header *oh;
 107
 108         genl_len = nlmsg_total_size(GENL_HDRLEN + dp_genl_family.hdrsize);
 109         genl_len += nla_total_size(sizeof(uint32_t)); /* DP_GENL_A_DP_IDX */
 110         genl_len += nla_total_size(openflow_len);    /* DP_GENL_A_OPENFLOW */
 111         skb = *pskb = genlmsg_new(genl_len, GFP_ATOMIC);
 112         if (!skb) {
 113                 if (net_ratelimit())
 114                         printk("alloc_openflow_skb: genlmsg_new failed\n");
 115                 return NULL;
 116         }
 117
 118         /* Assemble the Generic Netlink wrapper. */
 119         if (!genlmsg_put(skb,
 120                          sender ? sender->pid : 0,
 121                          sender ? sender->seq : 0,
 122                          &dp_genl_family, 0, DP_GENL_C_OPENFLOW))
 123                 BUG();
 124         if (nla_put_u32(skb, DP_GENL_A_DP_IDX, dp->dp_idx) < 0)
 125                 BUG();
 126         attr = nla_reserve(skb, DP_GENL_A_OPENFLOW, openflow_len);
 127         BUG_ON(!attr);
 128         nlmsg_end(skb, (struct nlmsghdr *) skb->data);
 129
 130         /* Fill in the header. */
 131         oh = nla_data(attr);
 132         oh->version = OFP_VERSION;
 133         oh->type = type;
 134         oh->length = htons(openflow_len);
 135         oh->xid = sender ? sender->xid : 0;
 136
 137         return oh;
 138 }
 139
 140 static void
 141 resize_openflow_skb(struct sk_buff *skb,
 142                     struct ofp_header *oh, size_t new_length)
 143 {
 144         struct nlattr *attr;
 145
 146         BUG_ON(new_length > ntohs(oh->length));
 147         attr = ((void *) oh) - NLA_HDRLEN;
 148         nla_unreserve(skb, attr, ntohs(oh->length) - new_length);
 149         oh->length = htons(new_length);
 150         nlmsg_end(skb, (struct nlmsghdr *) skb->data);
 151 }
 152
 153 static int
 154 send_openflow_skb(struct sk_buff *skb, const struct sender *sender)
 155 {
 156         int err = (sender
 157                    ? genlmsg_unicast(skb, sender->pid)
 158                    : genlmsg_multicast(skb, 0, mc_group.id, GFP_ATOMIC));
 159         if (err && net_ratelimit())
 160                 printk(KERN_WARNING "send_openflow_skb: send failed: %d\n",
 161                        err);
 162         return err;
 163 }
 164
 165 /* Generates a unique datapath id.  It incorporates the datapath index
 166  * and a hardware address, if available.  If not, it generates a random
 167  * one.
 168  */
 169 static
 170 uint64_t gen_datapath_id(uint16_t dp_idx)
 171 {
 172         uint64_t id;
 173         int i;
 174         struct net_device *dev;
 175
 176         /* The top 16 bits are used to identify the datapath.  The lower 48 bits
 177          * use an interface address.  */
 178         id = (uint64_t)dp_idx << 48;
 179         if ((dev = dev_get_by_name(&init_net, "ctl0"))
 180                         || (dev = dev_get_by_name(&init_net, "eth0"))) {
 181                 for (i=0; i<ETH_ALEN; i++) {
 182                         id |= (uint64_t)dev->dev_addr[i] << (8*(ETH_ALEN-1 - i));
 183                 }
 184                 dev_put(dev);
 185         } else {
 186                 /* Randomly choose the lower 48 bits if we cannot find an
 187                  * address and mark the most significant bit to indicate that
 188                  * this was randomly generated. */
 189                 uint8_t rand[ETH_ALEN];
 190                 get_random_bytes(rand, ETH_ALEN);
 191                 id |= (uint64_t)1 << 63;
 192                 for (i=0; i<ETH_ALEN; i++) {
 193                         id |= (uint64_t)rand[i] << (8*(ETH_ALEN-1 - i));
 194                 }
 195         }
 196
 197         return id;
 198 }
 199
 200 /* Creates a new datapath numbered 'dp_idx'.  Returns 0 for success or a
 201  * negative error code.
 202  *
 203  * Not called with any locks. */
 204 static int new_dp(int dp_idx)
 205 {
 206         struct datapath *dp;
 207         int err;
 208
 209         if (dp_idx < 0 || dp_idx >= DP_MAX)
 210                 return -EINVAL;
 211
 212         if (!try_module_get(THIS_MODULE))
 213                 return -ENODEV;
 214
 215         mutex_lock(&dp_mutex);
 216         dp = rcu_dereference(dps[dp_idx]);
 217         if (dp != NULL) {
 218                 err = -EEXIST;
 219                 goto err_unlock;
 220         }
 221
 222         err = -ENOMEM;
 223         dp = kzalloc(sizeof *dp, GFP_KERNEL);
 224         if (dp == NULL)
 225                 goto err_unlock;
 226
 227         dp->dp_idx = dp_idx;
 228         dp->id = gen_datapath_id(dp_idx);
 229         dp->chain = chain_create(dp);
 230         if (dp->chain == NULL)
 231                 goto err_free_dp;
 232         INIT_LIST_HEAD(&dp->port_list);
 233
 234 #if 0
 235         /* Setup our "of" device */
 236         dp->dev.priv = dp;
 237         rtnl_lock();
 238         err = dp_dev_setup(&dp->dev);
 239         rtnl_unlock();
 240         if (err != 0)
 241                 printk("datapath: problem setting up 'of' device\n");
 242 #endif
 243
 244         dp->flags = 0;
 245         dp->miss_send_len = OFP_DEFAULT_MISS_SEND_LEN;
 246
 247         dp->dp_task = kthread_run(dp_maint_func, dp, "dp%d", dp_idx);
 248         if (IS_ERR(dp->dp_task))
 249                 goto err_free_dp;
 250
 251         rcu_assign_pointer(dps[dp_idx], dp);
 252         mutex_unlock(&dp_mutex);
 253
 254         return 0;
 255
 256 err_free_dp:
 257         kfree(dp);
 258 err_unlock:
 259         mutex_unlock(&dp_mutex);
 260         module_put(THIS_MODULE);
 261                 return err;
 262 }
 263
 264 /* Find and return a free port number under 'dp'.  Called under dp_mutex. */
 265 static int find_portno(struct datapath *dp)
 266 {
 267         int i;
 268         for (i = 0; i < OFPP_MAX; i++)
 269                 if (dp->ports[i] == NULL)
 270                         return i;
 271         return -EXFULL;
 272 }
 273
 274 static struct net_bridge_port *new_nbp(struct datapath *dp,
 275                                                                            struct net_device *dev)
 276 {
 277         struct net_bridge_port *p;
 278         int port_no;
 279
 280         port_no = find_portno(dp);
 281         if (port_no < 0)
 282                 return ERR_PTR(port_no);
 283
 284         p = kzalloc(sizeof(*p), GFP_KERNEL);
 285         if (p == NULL)
 286                 return ERR_PTR(-ENOMEM);
 287
 288         p->dp = dp;
 289         dev_hold(dev);
 290         p->dev = dev;
 291         p->port_no = port_no;
 292
 293         return p;
 294 }
 295
 296 /* Called with dp_mutex. */
 297 int add_switch_port(struct datapath *dp, struct net_device *dev)
 298 {
 299         struct net_bridge_port *p;
 300
 301         if (dev->flags & IFF_LOOPBACK || dev->type != ARPHRD_ETHER)
 302                 return -EINVAL;
 303
 304         if (dev->br_port != NULL)
 305                 return -EBUSY;
 306
 307         p = new_nbp(dp, dev);
 308         if (IS_ERR(p))
 309                 return PTR_ERR(p);
 310
 311         dev_hold(dev);
 312         rcu_assign_pointer(dev->br_port, p);
 313         rtnl_lock();
 314         dev_set_promiscuity(dev, 1);
 315         rtnl_unlock();
 316
 317         rcu_assign_pointer(dp->ports[p->port_no], p);
 318         list_add_rcu(&p->node, &dp->port_list);
 319
 320         /* Notify the ctlpath that this port has been added */
 321         send_port_status(p, OFPPR_ADD);
 322
 323         return 0;
 324 }
 325
 326 /* Delete 'p' from switch.
 327  * Called with dp_mutex. */
 328 static int del_switch_port(struct net_bridge_port *p)
 329 {
 330         /* First drop references to device. */
 331         rtnl_lock();
 332         dev_set_promiscuity(p->dev, -1);
 333         rtnl_unlock();
 334         list_del_rcu(&p->node);
 335         rcu_assign_pointer(p->dp->ports[p->port_no], NULL);
 336         rcu_assign_pointer(p->dev->br_port, NULL);
 337
 338         /* Then wait until no one is still using it, and destroy it. */
 339         synchronize_rcu();
 340
 341         /* Notify the ctlpath that this port no longer exists */
 342         send_port_status(p, OFPPR_DELETE);
 343
 344         dev_put(p->dev);
 345         kfree(p);
 346
 347         return 0;
 348 }
 349
 350 /* Called with dp_mutex. */
 351 static void del_dp(struct datapath *dp)
 352 {
 353         struct net_bridge_port *p, *n;
 354
 355 #if 0
 356         /* Unregister the "of" device of this dp */
 357         rtnl_lock();
 358         unregister_netdevice(&dp->dev);
 359         rtnl_unlock();
 360 #endif
 361
 362         kthread_stop(dp->dp_task);
 363
 364         /* Drop references to DP. */
 365         list_for_each_entry_safe (p, n, &dp->port_list, node)
 366                 del_switch_port(p);
 367         rcu_assign_pointer(dps[dp->dp_idx], NULL);
 368
 369         /* Wait until no longer in use, then destroy it. */
 370         synchronize_rcu();
 371         chain_destroy(dp->chain);
 372         kfree(dp);
 373         module_put(THIS_MODULE);
 374 }
 375
 376 static int dp_maint_func(void *data)
 377 {
 378         struct datapath *dp = (struct datapath *) data;
 379
 380         while (!kthread_should_stop()) {
 381 #if 1
 382                 chain_timeout(dp->chain);
 383 #else
 384                 int count = chain_timeout(dp->chain);
 385                 chain_print_stats(dp->chain);
 386                 if (count)
 387                         printk("%d flows timed out\n", count);
 388 #endif
 389                 msleep_interruptible(MAINT_SLEEP_MSECS);
 390         }
 391
 392         return 0;
 393 }
 394
 395 /*
 396  * Used as br_handle_frame_hook.  (Cannot run bridge at the same time, even on
 397  * different set of devices!)  Returns 0 if *pskb should be processed further,
 398  * 1 if *pskb is handled. */
 399 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22)
 400 /* Called with rcu_read_lock. */
 401 static struct sk_buff *dp_frame_hook(struct net_bridge_port *p,
 402                                          struct sk_buff *skb)
 403 {
 404         struct ethhdr *eh = eth_hdr(skb);
 405         struct sk_buff *skb_local = NULL;
 406
 407
 408         if (compare_ether_addr(eh->h_dest, skb->dev->dev_addr) == 0)
 409                 return skb;
 410
 411         if (is_broadcast_ether_addr(eh->h_dest)
 412                                 || is_multicast_ether_addr(eh->h_dest)
 413                                 || is_local_ether_addr(eh->h_dest))
 414                 skb_local = skb_clone(skb, GFP_ATOMIC);
 415
 416         /* Push the Ethernet header back on. */
 417         if (skb->protocol == htons(ETH_P_8021Q))
 418                 skb_push(skb, VLAN_ETH_HLEN);
 419         else
 420                 skb_push(skb, ETH_HLEN);
 421
 422         fwd_port_input(p->dp->chain, skb, p->port_no);
 423
 424         return skb_local;
 425 }
 426 #elif LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
 427 static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb)
 428 {
 429         /* Push the Ethernet header back on. */
 430         if ((*pskb)->protocol == htons(ETH_P_8021Q))
 431                 skb_push(*pskb, VLAN_ETH_HLEN);
 432         else
 433                 skb_push(*pskb, ETH_HLEN);
 434
 435         fwd_port_input(p->dp->chain, *pskb, p->port_no);
 436         return 1;
 437 }
 438 #else
 439 /* NB: This has only been tested on 2.4.35 */
 440
 441 /* Called without any locks (?) */
 442 static void dp_frame_hook(struct sk_buff *skb)
 443 {
 444         struct net_bridge_port *p = skb->dev->br_port;
 445
 446         /* Push the Ethernet header back on. */
 447         if (skb->protocol == htons(ETH_P_8021Q))
 448                 skb_push(skb, VLAN_ETH_HLEN);
 449         else
 450                 skb_push(skb, ETH_HLEN);
 451
 452         if (p) {
 453                 rcu_read_lock();
 454                 fwd_port_input(p->dp->chain, skb, p->port_no);
 455                 rcu_read_unlock();
 456         } else
 457                 kfree_skb(skb);
 458 }
 459 #endif
 460
 461 /* Forwarding output path.
 462  * Based on net/bridge/br_forward.c. */
 463
 464 /* Don't forward packets to originating port or with flooding disabled */
 465 static inline int should_deliver(const struct net_bridge_port *p,
 466                         const struct sk_buff *skb)
 467 {
 468         if ((skb->dev == p->dev) || (p->flags & BRIDGE_PORT_NO_FLOOD)) {
 469                 return 0;
 470         }
 471
 472         return 1;
 473 }
 474
 475 static inline unsigned packet_length(const struct sk_buff *skb)
 476 {
 477         int length = skb->len - ETH_HLEN;
 478         if (skb->protocol == htons(ETH_P_8021Q))
 479                 length -= VLAN_HLEN;
 480         return length;
 481 }
 482
 483 static int
 484 flood(struct datapath *dp, struct sk_buff *skb)
 485 {
 486         struct net_bridge_port *p;
 487         int prev_port;
 488
 489         prev_port = -1;
 490         list_for_each_entry_rcu (p, &dp->port_list, node) {
 491                 if (!should_deliver(p, skb))
 492                         continue;
 493                 if (prev_port != -1) {
 494                         struct sk_buff *clone = skb_clone(skb, GFP_ATOMIC);
 495                         if (!clone) {
 496                                 kfree_skb(skb);
 497                                 return -ENOMEM;
 498                         }
 499                         dp_output_port(dp, clone, prev_port);
 500                 }
 501                 prev_port = p->port_no;
 502         }
 503         if (prev_port != -1)
 504                 dp_output_port(dp, skb, prev_port);
 505         else
 506                 kfree_skb(skb);
 507
 508         return 0;
 509 }
 510
 511 /* Marks 'skb' as having originated from 'in_port' in 'dp'.
 512    FIXME: how are devices reference counted? */
 513 int dp_set_origin(struct datapath *dp, uint16_t in_port,
 514                            struct sk_buff *skb)
 515 {
 516         if (in_port < OFPP_MAX && dp->ports[in_port]) {
 517                 skb->dev = dp->ports[in_port]->dev;
 518                 return 0;
 519         }
 520         return -ENOENT;
 521 }
 522
 523 /* Takes ownership of 'skb' and transmits it to 'out_port' on 'dp'.
 524  */
 525 int dp_output_port(struct datapath *dp, struct sk_buff *skb, int out_port)
 526 {
 527         struct net_bridge_port *p;
 528         int len = skb->len;
 529
 530         BUG_ON(!skb);
 531         if (out_port == OFPP_FLOOD)
 532                 return flood(dp, skb);
 533         else if (out_port == OFPP_CONTROLLER)
 534                 return dp_output_control(dp, skb, fwd_save_skb(skb), 0,
 535                                                   OFPR_ACTION);
 536         else if (out_port == OFPP_TABLE) {
 537                 struct sw_flow_key key;
 538                 struct sw_flow *flow;
 539
 540                 flow_extract(skb, skb->dev->br_port->port_no, &key);
 541                 flow = chain_lookup(dp->chain, &key);
 542                 if (likely(flow != NULL)) {
 543                         flow_used(flow, skb);
 544                         execute_actions(dp, skb, &key, flow->actions, flow->n_actions);
 545                         return 0;
 546                 }
 547                 return -ESRCH;
 548         } else if (out_port >= OFPP_MAX)
 549                 goto bad_port;
 550
 551         p = dp->ports[out_port];
 552         if (p == NULL)
 553                 goto bad_port;
 554
 555         skb->dev = p->dev;
 556         if (packet_length(skb) > skb->dev->mtu) {
 557                 printk("dropped over-mtu packet: %d > %d\n",
 558                                         packet_length(skb), skb->dev->mtu);
 559                 kfree_skb(skb);
 560                 return -E2BIG;
 561         }
 562
 563         dev_queue_xmit(skb);
 564
 565         return len;
 566
 567 bad_port:
 568         kfree_skb(skb);
 569         if (net_ratelimit())
 570                 printk("can't forward to bad port %d\n", out_port);
 571         return -ENOENT;
 572 }
 573
 574 /* Takes ownership of 'skb' and transmits it to 'dp''s control path.  If
 575  * 'buffer_id' != -1, then only the first 64 bytes of 'skb' are sent;
 576  * otherwise, all of 'skb' is sent.  'reason' indicates why 'skb' is being
 577  * sent. 'max_len' sets the maximum number of bytes that the caller
 578  * wants to be sent; a value of 0 indicates the entire packet should be
 579  * sent. */
 580 int
 581 dp_output_control(struct datapath *dp, struct sk_buff *skb,
 582                            uint32_t buffer_id, size_t max_len, int reason)
 583 {
 584         /* FIXME?  Can we avoid creating a new skbuff in the case where we
 585          * forward the whole packet? */
 586         struct sk_buff *f_skb;
 587         struct ofp_packet_in *opi;
 588         size_t fwd_len, opi_len;
 589         int err;
 590
 591         fwd_len = skb->len;
 592         if ((buffer_id != (uint32_t) -1) && max_len)
 593                 fwd_len = min(fwd_len, max_len);
 594
 595         opi_len = offsetof(struct ofp_packet_in, data) + fwd_len;
 596         opi = alloc_openflow_skb(dp, opi_len, OFPT_PACKET_IN, NULL, &f_skb);
 597         opi->buffer_id      = htonl(buffer_id);
 598         opi->total_len      = htons(skb->len);
 599         opi->in_port        = htons(skb->dev->br_port->port_no);
 600         opi->reason         = reason;
 601         opi->pad            = 0;
 602         memcpy(opi->data, skb_mac_header(skb), fwd_len);
 603         err = send_openflow_skb(f_skb, NULL);
 604
 605         kfree_skb(skb);
 606
 607         return err;
 608 }
 609
 610 static void fill_port_desc(struct net_bridge_port *p, struct ofp_phy_port *desc)
 611 {
 612         desc->port_no = htons(p->port_no);
 613         strncpy(desc->name, p->dev->name, OFP_MAX_PORT_NAME_LEN);
 614         desc->name[OFP_MAX_PORT_NAME_LEN-1] = '\0';
 615         memcpy(desc->hw_addr, p->dev->dev_addr, ETH_ALEN);
 616         desc->flags = htonl(p->flags);
 617         desc->features = 0;
 618         desc->speed = 0;
 619
 620 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,4,24)
 621         if (p->dev->ethtool_ops && p->dev->ethtool_ops->get_settings) {
 622                 struct ethtool_cmd ecmd = { .cmd = ETHTOOL_GSET };
 623
 624                 if (!p->dev->ethtool_ops->get_settings(p->dev, &ecmd)) {
 625                         if (ecmd.supported & SUPPORTED_10baseT_Half)
 626                                 desc->features |= OFPPF_10MB_HD;
 627                         if (ecmd.supported & SUPPORTED_10baseT_Full)
 628                                 desc->features |= OFPPF_10MB_FD;
 629                         if (ecmd.supported & SUPPORTED_100baseT_Half)
 630                                 desc->features |= OFPPF_100MB_HD;
 631                         if (ecmd.supported & SUPPORTED_100baseT_Full)
 632                                 desc->features |= OFPPF_100MB_FD;
 633                         if (ecmd.supported & SUPPORTED_1000baseT_Half)
 634                                 desc->features |= OFPPF_1GB_HD;
 635                         if (ecmd.supported & SUPPORTED_1000baseT_Full)
 636                                 desc->features |= OFPPF_1GB_FD;
 637                         /* 10Gbps half-duplex doesn't exist... */
 638                         if (ecmd.supported & SUPPORTED_10000baseT_Full)
 639                                 desc->features |= OFPPF_10GB_FD;
 640
 641                         desc->features = htonl(desc->features);
 642                         desc->speed = htonl(ecmd.speed);
 643                 }
 644         }
 645 #endif
 646 }
 647
 648 static int
 649 fill_features_reply(struct datapath *dp, struct ofp_switch_features *ofr)
 650 {
 651         struct net_bridge_port *p;
 652         int port_count = 0;
 653
 654         ofr->datapath_id    = cpu_to_be64(dp->id);
 655
 656         ofr->n_exact        = htonl(2 * TABLE_HASH_MAX_FLOWS);
 657         ofr->n_compression  = 0;                                           /* Not supported */
 658         ofr->n_general      = htonl(TABLE_LINEAR_MAX_FLOWS);
 659         ofr->buffer_mb      = htonl(UINT32_MAX);
 660         ofr->n_buffers      = htonl(N_PKT_BUFFERS);
 661         ofr->capabilities   = htonl(OFP_SUPPORTED_CAPABILITIES);
 662         ofr->actions        = htonl(OFP_SUPPORTED_ACTIONS);
 663
 664         list_for_each_entry_rcu (p, &dp->port_list, node) {
 665                 fill_port_desc(p, &ofr->ports[port_count]);
 666                 port_count++;
 667         }
 668
 669         return port_count;
 670 }
 671
 672 int
 673 dp_send_features_reply(struct datapath *dp, const struct sender *sender)
 674 {
 675         struct sk_buff *skb;
 676         struct ofp_switch_features *ofr;
 677         size_t ofr_len, port_max_len;
 678         int port_count;
 679
 680         /* Overallocate. */
 681         port_max_len = sizeof(struct ofp_phy_port) * OFPP_MAX;
 682         ofr = alloc_openflow_skb(dp, sizeof(*ofr) + port_max_len,
 683                                  OFPT_FEATURES_REPLY, sender, &skb);
 684         if (!ofr)
 685                 return -ENOMEM;
 686
 687         /* Fill. */
 688         port_count = fill_features_reply(dp, ofr);
 689
 690         /* Shrink to fit. */
 691         ofr_len = sizeof(*ofr) + (sizeof(struct ofp_phy_port) * port_count);
 692         resize_openflow_skb(skb, &ofr->header, ofr_len);
 693         return send_openflow_skb(skb, sender);
 694 }
 695
 696 int
 697 dp_send_config_reply(struct datapath *dp, const struct sender *sender)
 698 {
 699         struct sk_buff *skb;
 700         struct ofp_switch_config *osc;
 701
 702         osc = alloc_openflow_skb(dp, sizeof *osc, OFPT_GET_CONFIG_REPLY, sender,
 703                                  &skb);
 704         if (!osc)
 705                 return -ENOMEM;
 706
 707         osc->flags = htons(dp->flags);
 708         osc->miss_send_len = htons(dp->miss_send_len);
 709
 710         return send_openflow_skb(skb, sender);
 711 }
 712
 713 int
 714 dp_update_port_flags(struct datapath *dp, const struct ofp_phy_port *opp)
 715 {
 716         struct net_bridge_port *p;
 717
 718         p = dp->ports[htons(opp->port_no)];
 719
 720         /* Make sure the port id hasn't changed since this was sent */
 721         if (!p || memcmp(opp->hw_addr, p->dev->dev_addr, ETH_ALEN) != 0)
 722                 return -1;
 723
 724         p->flags = htonl(opp->flags);
 725
 726         return 0;
 727 }
 728
 729
 730 static int
 731 send_port_status(struct net_bridge_port *p, uint8_t status)
 732 {
 733         struct sk_buff *skb;
 734         struct ofp_port_status *ops;
 735
 736         ops = alloc_openflow_skb(p->dp, sizeof *ops, OFPT_PORT_STATUS, NULL,
 737                                  &skb);
 738         if (!ops)
 739                 return -ENOMEM;
 740         ops->reason = status;
 741         fill_port_desc(p, &ops->desc);
 742
 743         return send_openflow_skb(skb, NULL);
 744 }
 745
 746 int
 747 dp_send_flow_expired(struct datapath *dp, struct sw_flow *flow)
 748 {
 749         struct sk_buff *skb;
 750         struct ofp_flow_expired *ofe;
 751         unsigned long duration_j;
 752
 753         ofe = alloc_openflow_skb(dp, sizeof *ofe, OFPT_FLOW_EXPIRED, 0, &skb);
 754         if (!ofe)
 755                 return -ENOMEM;
 756
 757         flow_fill_match(&ofe->match, &flow->key);
 758         duration_j = (flow->timeout - HZ * flow->max_idle) - flow->init_time;
 759         ofe->duration   = htonl(duration_j / HZ);
 760         ofe->packet_count   = cpu_to_be64(flow->packet_count);
 761         ofe->byte_count     = cpu_to_be64(flow->byte_count);
 762         return send_openflow_skb(skb, NULL);
 763 }
 764
 765 static void
 766 fill_flow_stats(struct ofp_flow_stats *ofs, struct sw_flow *flow,
 767                 int table_idx)
 768 {
 769         ofs->match.wildcards = htons(flow->key.wildcards);
 770         ofs->match.in_port   = flow->key.in_port;
 771         memcpy(ofs->match.dl_src, flow->key.dl_src, ETH_ALEN);
 772         memcpy(ofs->match.dl_dst, flow->key.dl_dst, ETH_ALEN);
 773         ofs->match.dl_vlan   = flow->key.dl_vlan;
 774         ofs->match.dl_type   = flow->key.dl_type;
 775         ofs->match.nw_src    = flow->key.nw_src;
 776         ofs->match.nw_dst    = flow->key.nw_dst;
 777         ofs->match.nw_proto  = flow->key.nw_proto;
 778         memset(ofs->match.pad, 0, sizeof ofs->match.pad);
 779         ofs->match.tp_src    = flow->key.tp_src;
 780         ofs->match.tp_dst    = flow->key.tp_dst;
 781         ofs->duration        = htonl((jiffies - flow->init_time) / HZ);
 782         ofs->table_id        = table_idx;
 783         ofs->packet_count    = cpu_to_be64(flow->packet_count);
 784         ofs->byte_count      = cpu_to_be64(flow->byte_count);
 785 }
 786
 787 int
 788 dp_send_flow_stats(struct datapath *dp, const struct sender *sender,
 789                    const struct ofp_match *match)
 790 {
 791         struct sk_buff *skb;
 792         struct ofp_flow_stats_reply *fsr;
 793         size_t header_size, fudge, flow_size;
 794         struct sw_flow_key match_key;
 795         int table_idx, n_flows, max_flows;
 796
 797         header_size = offsetof(struct ofp_flow_stats_reply, flows);
 798         fudge = 128;
 799         flow_size = sizeof fsr->flows[0];
 800         max_flows = (NLMSG_GOODSIZE - header_size - fudge) / flow_size;
 801         fsr = alloc_openflow_skb(dp, header_size + max_flows * flow_size,
 802                                  OFPT_FLOW_STATS_REPLY, sender, &skb);
 803         if (!fsr)
 804                 return -ENOMEM;
 805
 806         n_flows = 0;
 807         flow_extract_match(&match_key, match);
 808         for (table_idx = 0; table_idx < dp->chain->n_tables; table_idx++) {
 809                 struct sw_table *table = dp->chain->tables[table_idx];
 810                 struct swt_iterator iter;
 811
 812                 if (n_flows >= max_flows) {
 813                         break;
 814                 }
 815
 816                 if (!table->iterator(table, &iter)) {
 817                         if (net_ratelimit())
 818                                 printk("iterator failed for table %d\n",
 819                                        table_idx);
 820                         continue;
 821                 }
 822
 823                 for (; iter.flow; table->iterator_next(&iter)) {
 824                         if (flow_matches(&match_key, &iter.flow->key)) {
 825                                 fill_flow_stats(&fsr->flows[n_flows],
 826                                                 iter.flow, table_idx);
 827                                 if (++n_flows >= max_flows) {
 828                                         break;
 829                                 }
 830                         }
 831                 }
 832                 table->iterator_destroy(&iter);
 833         }
 834         resize_openflow_skb(skb, &fsr->header,
 835                             header_size + flow_size * n_flows);
 836         return send_openflow_skb(skb, sender);
 837 }
 838
 839 static int
 840 fill_port_stats_reply(struct datapath *dp, struct ofp_port_stats_reply *psr)
 841 {
 842         struct net_bridge_port *p;
 843         int port_count = 0;
 844
 845         list_for_each_entry_rcu (p, &dp->port_list, node) {
 846                 struct ofp_port_stats *ps = &psr->ports[port_count++];
 847                 struct net_device_stats *stats = p->dev->get_stats(p->dev);
 848                 ps->port_no = htons(p->port_no);
 849                 memset(ps->pad, 0, sizeof ps->pad);
 850                 ps->rx_count = cpu_to_be64(stats->rx_packets);
 851                 ps->tx_count = cpu_to_be64(stats->tx_packets);
 852                 ps->drop_count = cpu_to_be64(stats->rx_dropped
 853                                              + stats->tx_dropped);
 854         }
 855
 856         return port_count;
 857 }
 858
 859 int
 860 dp_send_port_stats(struct datapath *dp, const struct sender *sender)
 861 {
 862         struct sk_buff *skb;
 863         struct ofp_port_stats_reply *psr;
 864         size_t psr_len, port_max_len;
 865         int port_count;
 866
 867         /* Overallocate. */
 868         port_max_len = sizeof(struct ofp_port_stats) * OFPP_MAX;
 869         psr = alloc_openflow_skb(dp, sizeof *psr + port_max_len,
 870                                  OFPT_PORT_STATS_REPLY, sender, &skb);
 871         if (!psr)
 872                 return -ENOMEM;
 873
 874         /* Fill. */
 875         port_count = fill_port_stats_reply(dp, psr);
 876
 877         /* Shrink to fit. */
 878         psr_len = sizeof *psr + sizeof(struct ofp_port_stats) * port_count;
 879         resize_openflow_skb(skb, &psr->header, psr_len);
 880         return send_openflow_skb(skb, sender);
 881 }
 882
 883 int
 884 dp_send_table_stats(struct datapath *dp, const struct sender *sender)
 885 {
 886         struct sk_buff *skb;
 887         struct ofp_table_stats_reply *tsr;
 888         int i, n_tables;
 889
 890         n_tables = dp->chain->n_tables;
 891         tsr = alloc_openflow_skb(dp, (offsetof(struct ofp_table_stats_reply,
 892                                                tables)
 893                                       + sizeof tsr->tables[0] * n_tables),
 894                                  OFPT_TABLE_STATS_REPLY, sender, &skb);
 895         if (!tsr)
 896                 return -ENOMEM;
 897         for (i = 0; i < n_tables; i++) {
 898                 struct ofp_table_stats *ots = &tsr->tables[i];
 899                 struct sw_table_stats stats;
 900                 dp->chain->tables[i]->stats(dp->chain->tables[i], &stats);
 901                 strncpy(ots->name, stats.name, sizeof ots->name);
 902                 ots->table_id = i;
 903                 ots->pad[0] = ots->pad[1] = 0;
 904                 ots->max_entries = htonl(stats.max_flows);
 905                 ots->active_count = htonl(stats.n_flows);
 906                 ots->matched_count = cpu_to_be64(0); /* FIXME */
 907         }
 908         return send_openflow_skb(skb, sender);
 909 }
 910
 911 /* Generic Netlink interface.
 912  *
 913  * See netlink(7) for an introduction to netlink.  See
 914  * http://linux-net.osdl.org/index.php/Netlink for more information and
 915  * pointers on how to work with netlink and Generic Netlink in the kernel and
 916  * in userspace. */
 917
 918 static struct genl_family dp_genl_family = {
 919         .id = GENL_ID_GENERATE,
 920         .hdrsize = 0,
 921         .name = DP_GENL_FAMILY_NAME,
 922         .version = 1,
 923         .maxattr = DP_GENL_A_MAX,
 924 };
 925
 926 /* Attribute policy: what each attribute may contain.  */
 927 static struct nla_policy dp_genl_policy[DP_GENL_A_MAX + 1] = {
 928         [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
 929         [DP_GENL_A_MC_GROUP] = { .type = NLA_U32 },
 930         [DP_GENL_A_PORTNAME] = { .type = NLA_STRING }
 931 };
 932
 933 static int dp_genl_add(struct sk_buff *skb, struct genl_info *info)
 934 {
 935         if (!info->attrs[DP_GENL_A_DP_IDX])
 936                 return -EINVAL;
 937
 938         return new_dp(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
 939 }
 940
 941 static struct genl_ops dp_genl_ops_add_dp = {
 942         .cmd = DP_GENL_C_ADD_DP,
 943         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
 944         .policy = dp_genl_policy,
 945         .doit = dp_genl_add,
 946         .dumpit = NULL,
 947 };
 948
 949 struct datapath *dp_get(int dp_idx)
 950 {
 951         if (dp_idx < 0 || dp_idx > DP_MAX)
 952                 return NULL;
 953         return rcu_dereference(dps[dp_idx]);
 954 }
 955
 956 static int dp_genl_del(struct sk_buff *skb, struct genl_info *info)
 957 {
 958         struct datapath *dp;
 959         int err;
 960
 961         if (!info->attrs[DP_GENL_A_DP_IDX])
 962                 return -EINVAL;
 963
 964         mutex_lock(&dp_mutex);
 965         dp = dp_get(nla_get_u32((info->attrs[DP_GENL_A_DP_IDX])));
 966         if (!dp)
 967                 err = -ENOENT;
 968         else {
 969                 del_dp(dp);
 970                 err = 0;
 971         }
 972         mutex_unlock(&dp_mutex);
 973         return err;
 974 }
 975
 976 static struct genl_ops dp_genl_ops_del_dp = {
 977         .cmd = DP_GENL_C_DEL_DP,
 978         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
 979         .policy = dp_genl_policy,
 980         .doit = dp_genl_del,
 981         .dumpit = NULL,
 982 };
 983
 984 /* Queries a datapath for related information.  Currently the only relevant
 985  * information is the datapath's multicast group ID.  Really we want one
 986  * multicast group per datapath, but because of locking issues[*] we can't
 987  * easily get one.  Thus, every datapath will currently return the same
 988  * global multicast group ID, but in the future it would be nice to fix that.
 989  *
 990  * [*] dp_genl_add, to add a new datapath, is called under the genl_lock
 991  *       mutex, and genl_register_mc_group, called to acquire a new multicast
 992  *       group ID, also acquires genl_lock, thus deadlock.
 993  */
 994 static int dp_genl_query(struct sk_buff *skb, struct genl_info *info)
 995 {
 996         struct datapath *dp;
 997         struct sk_buff *ans_skb = NULL;
 998         int dp_idx;
 999         int err = -ENOMEM;
1000
1001         if (!info->attrs[DP_GENL_A_DP_IDX])
1002                 return -EINVAL;
1003
1004         rcu_read_lock();
1005         dp_idx = nla_get_u32((info->attrs[DP_GENL_A_DP_IDX]));
1006         dp = dp_get(dp_idx);
1007         if (!dp)
1008                 err = -ENOENT;
1009         else {
1010                 void *data;
1011                 ans_skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1012                 if (!ans_skb) {
1013                         err = -ENOMEM;
1014                         goto err;
1015                 }
1016                 data = genlmsg_put_reply(ans_skb, info, &dp_genl_family,
1017                                          0, DP_GENL_C_QUERY_DP);
1018                 if (data == NULL) {
1019                         err = -ENOMEM;
1020                         goto err;
1021                 }
1022                 NLA_PUT_U32(ans_skb, DP_GENL_A_DP_IDX, dp_idx);
1023                 NLA_PUT_U32(ans_skb, DP_GENL_A_MC_GROUP, mc_group.id);
1024
1025                 genlmsg_end(ans_skb, data);
1026                 err = genlmsg_reply(ans_skb, info);
1027                 if (!err)
1028                         ans_skb = NULL;
1029         }
1030 err:
1031 nla_put_failure:
1032         if (ans_skb)
1033                 kfree_skb(ans_skb);
1034         rcu_read_unlock();
1035         return err;
1036 }
1037
1038 static struct genl_ops dp_genl_ops_query_dp = {
1039         .cmd = DP_GENL_C_QUERY_DP,
1040         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1041         .policy = dp_genl_policy,
1042         .doit = dp_genl_query,
1043         .dumpit = NULL,
1044 };
1045
1046 static int dp_genl_add_del_port(struct sk_buff *skb, struct genl_info *info)
1047 {
1048         struct datapath *dp;
1049         struct net_device *port;
1050         int err;
1051
1052         if (!info->attrs[DP_GENL_A_DP_IDX] || !info->attrs[DP_GENL_A_PORTNAME])
1053                 return -EINVAL;
1054
1055         /* Get datapath. */
1056         mutex_lock(&dp_mutex);
1057         dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
1058         if (!dp) {
1059                 err = -ENOENT;
1060                 goto out;
1061         }
1062
1063         /* Get interface to add/remove. */
1064         port = dev_get_by_name(&init_net,
1065                         nla_data(info->attrs[DP_GENL_A_PORTNAME]));
1066         if (!port) {
1067                 err = -ENOENT;
1068                 goto out;
1069         }
1070
1071         /* Execute operation. */
1072         if (info->genlhdr->cmd == DP_GENL_C_ADD_PORT)
1073                 err = add_switch_port(dp, port);
1074         else {
1075                 if (port->br_port == NULL || port->br_port->dp != dp) {
1076                         err = -ENOENT;
1077                         goto out_put;
1078                 }
1079                 err = del_switch_port(port->br_port);
1080         }
1081
1082 out_put:
1083         dev_put(port);
1084 out:
1085         mutex_unlock(&dp_mutex);
1086         return err;
1087 }
1088
1089 static struct genl_ops dp_genl_ops_add_port = {
1090         .cmd = DP_GENL_C_ADD_PORT,
1091         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1092         .policy = dp_genl_policy,
1093         .doit = dp_genl_add_del_port,
1094         .dumpit = NULL,
1095 };
1096
1097 static struct genl_ops dp_genl_ops_del_port = {
1098         .cmd = DP_GENL_C_DEL_PORT,
1099         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1100         .policy = dp_genl_policy,
1101         .doit = dp_genl_add_del_port,
1102         .dumpit = NULL,
1103 };
1104
1105 static int dp_genl_openflow(struct sk_buff *skb, struct genl_info *info)
1106 {
1107         struct nlattr *va = info->attrs[DP_GENL_A_OPENFLOW];
1108         struct datapath *dp;
1109         struct ofp_header *oh;
1110         struct sender sender;
1111         int err;
1112
1113         if (!info->attrs[DP_GENL_A_DP_IDX] || !va)
1114                 return -EINVAL;
1115
1116         rcu_read_lock();
1117         dp = dp_get(nla_get_u32(info->attrs[DP_GENL_A_DP_IDX]));
1118         if (!dp) {
1119                 err = -ENOENT;
1120                 goto out;
1121         }
1122
1123         if (nla_len(va) < sizeof(struct ofp_header)) {
1124                 err = -EINVAL;
1125                 goto out;
1126         }
1127         oh = nla_data(va);
1128
1129         sender.xid = oh->xid;
1130         sender.pid = info->snd_pid;
1131         sender.seq = info->snd_seq;
1132         err = fwd_control_input(dp->chain, &sender, nla_data(va), nla_len(va));
1133
1134 out:
1135         rcu_read_unlock();
1136         return err;
1137 }
1138
1139 static struct nla_policy dp_genl_openflow_policy[DP_GENL_A_MAX + 1] = {
1140         [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1141 };
1142
1143 static struct genl_ops dp_genl_ops_openflow = {
1144         .cmd = DP_GENL_C_OPENFLOW,
1145         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1146         .policy = dp_genl_openflow_policy,
1147         .doit = dp_genl_openflow,
1148         .dumpit = NULL,
1149 };
1150
1151 static struct nla_policy dp_genl_benchmark_policy[DP_GENL_A_MAX + 1] = {
1152         [DP_GENL_A_DP_IDX] = { .type = NLA_U32 },
1153         [DP_GENL_A_NPACKETS] = { .type = NLA_U32 },
1154         [DP_GENL_A_PSIZE] = { .type = NLA_U32 },
1155 };
1156
1157 static struct genl_ops dp_genl_ops_benchmark_nl = {
1158         .cmd = DP_GENL_C_BENCHMARK_NL,
1159         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privilege. */
1160         .policy = dp_genl_benchmark_policy,
1161         .doit = dp_genl_benchmark_nl,
1162         .dumpit = NULL,
1163 };
1164
1165 static struct genl_ops *dp_genl_all_ops[] = {
1166         /* Keep this operation first.  Generic Netlink dispatching
1167          * looks up operations with linear search, so we want it at the
1168          * front. */
1169         &dp_genl_ops_openflow,
1170
1171         &dp_genl_ops_add_dp,
1172         &dp_genl_ops_del_dp,
1173         &dp_genl_ops_query_dp,
1174         &dp_genl_ops_add_port,
1175         &dp_genl_ops_del_port,
1176         &dp_genl_ops_benchmark_nl,
1177 };
1178
1179 static int dp_init_netlink(void)
1180 {
1181         int err;
1182         int i;
1183
1184         err = genl_register_family(&dp_genl_family);
1185         if (err)
1186                 return err;
1187
1188         for (i = 0; i < ARRAY_SIZE(dp_genl_all_ops); i++) {
1189                 err = genl_register_ops(&dp_genl_family, dp_genl_all_ops[i]);
1190                 if (err)
1191                         goto err_unregister;
1192         }
1193
1194         strcpy(mc_group.name, "openflow");
1195         err = genl_register_mc_group(&dp_genl_family, &mc_group);
1196         if (err < 0)
1197                 goto err_unregister;
1198
1199         return 0;
1200
1201 err_unregister:
1202         genl_unregister_family(&dp_genl_family);
1203                 return err;
1204 }
1205
1206 static void dp_uninit_netlink(void)
1207 {
1208         genl_unregister_family(&dp_genl_family);
1209 }
1210
1211 #define DRV_NAME                "openflow"
1212 #define DRV_VERSION      VERSION
1213 #define DRV_DESCRIPTION "OpenFlow switching datapath implementation"
1214 #define DRV_COPYRIGHT   "Copyright (c) 2007, 2008 The Board of Trustees of The Leland Stanford Junior University"
1215
1216
1217 static int __init dp_init(void)
1218 {
1219         int err;
1220
1221         printk(KERN_INFO DRV_NAME ": " DRV_DESCRIPTION "\n");
1222         printk(KERN_INFO DRV_NAME ": " VERSION" built on "__DATE__" "__TIME__"\n");
1223         printk(KERN_INFO DRV_NAME ": " DRV_COPYRIGHT "\n");
1224
1225         err = flow_init();
1226         if (err)
1227                 goto error;
1228
1229         err = dp_init_netlink();
1230         if (err)
1231                 goto error_flow_exit;
1232
1233         /* Hook into callback used by the bridge to intercept packets.
1234          * Parasites we are. */
1235         if (br_handle_frame_hook)
1236                 printk("openflow: hijacking bridge hook\n");
1237         br_handle_frame_hook = dp_frame_hook;
1238
1239         return 0;
1240
1241 error_flow_exit:
1242         flow_exit();
1243 error:
1244         printk(KERN_EMERG "openflow: failed to install!");
1245         return err;
1246 }
1247
1248 static void dp_cleanup(void)
1249 {
1250         fwd_exit();
1251         dp_uninit_netlink();
1252         flow_exit();
1253         br_handle_frame_hook = NULL;
1254 }
1255
1256 module_init(dp_init);
1257 module_exit(dp_cleanup);
1258
1259 MODULE_DESCRIPTION(DRV_DESCRIPTION);
1260 MODULE_AUTHOR(DRV_COPYRIGHT);
1261 MODULE_LICENSE("GPL");