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