drivers/net/bonding/bond_alb.c

   1 /*
   2  * Copyright(c) 1999 - 2004 Intel Corporation. All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License as published by the
   6  * Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful, but
  10  * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  11  * or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along
  15  * with this program; if not, write to the Free Software Foundation, Inc.,
  16  * 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  17  *
  18  * The full GNU General Public License is included in this distribution in the
  19  * file called LICENSE.
  20  *
  21  */
  22
  23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  24
  25 #include <linux/skbuff.h>
  26 #include <linux/netdevice.h>
  27 #include <linux/etherdevice.h>
  28 #include <linux/pkt_sched.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/slab.h>
  31 #include <linux/timer.h>
  32 #include <linux/ip.h>
  33 #include <linux/ipv6.h>
  34 #include <linux/if_arp.h>
  35 #include <linux/if_ether.h>
  36 #include <linux/if_bonding.h>
  37 #include <linux/if_vlan.h>
  38 #include <linux/in.h>
  39 #include <net/ipx.h>
  40 #include <net/arp.h>
  41 #include <net/ipv6.h>
  42 #include <asm/byteorder.h>
  43 #include "bonding.h"
  44 #include "bond_alb.h"
  45
  46
  47 #define ALB_TIMER_TICKS_PER_SEC     10  /* should be a divisor of HZ */
  48 #define BOND_TLB_REBALANCE_INTERVAL 10  /* In seconds, periodic re-balancing.
  49                                          * Used for division - never set
  50                                          * to zero !!!
  51                                          */
  52 #define BOND_ALB_LP_INTERVAL        1   /* In seconds, periodic send of
  53                                          * learning packets to the switch
  54                                          */
  55
  56 #define BOND_TLB_REBALANCE_TICKS (BOND_TLB_REBALANCE_INTERVAL \
  57                                   * ALB_TIMER_TICKS_PER_SEC)
  58
  59 #define BOND_ALB_LP_TICKS (BOND_ALB_LP_INTERVAL \
  60                            * ALB_TIMER_TICKS_PER_SEC)
  61
  62 #define TLB_HASH_TABLE_SIZE 256 /* The size of the clients hash table.
  63                                  * Note that this value MUST NOT be smaller
  64                                  * because the key hash table is BYTE wide !
  65                                  */
  66
  67
  68 #define TLB_NULL_INDEX          0xffffffff
  69 #define MAX_LP_BURST            3
  70
  71 /* rlb defs */
  72 #define RLB_HASH_TABLE_SIZE     256
  73 #define RLB_NULL_INDEX          0xffffffff
  74 #define RLB_UPDATE_DELAY        2*ALB_TIMER_TICKS_PER_SEC /* 2 seconds */
  75 #define RLB_ARP_BURST_SIZE      2
  76 #define RLB_UPDATE_RETRY        3       /* 3-ticks - must be smaller than the rlb
  77                                          * rebalance interval (5 min).
  78                                          */
  79 /* RLB_PROMISC_TIMEOUT = 10 sec equals the time that the current slave is
  80  * promiscuous after failover
  81  */
  82 #define RLB_PROMISC_TIMEOUT     10*ALB_TIMER_TICKS_PER_SEC
  83
  84 #ifndef __long_aligned
  85 #define __long_aligned __attribute__((aligned((sizeof(long)))))
  86 #endif
  87 static const u8 mac_bcast[ETH_ALEN] __long_aligned = {
  88         0xff, 0xff, 0xff, 0xff, 0xff, 0xff
  89 };
  90 static const u8 mac_v6_allmcast[ETH_ALEN] __long_aligned = {
  91         0x33, 0x33, 0x00, 0x00, 0x00, 0x01
  92 };
  93 static const int alb_delta_in_ticks = HZ / ALB_TIMER_TICKS_PER_SEC;
  94
  95 #pragma pack(1)
  96 struct learning_pkt {
  97         u8 mac_dst[ETH_ALEN];
  98         u8 mac_src[ETH_ALEN];
  99         __be16 type;
 100         u8 padding[ETH_ZLEN - ETH_HLEN];
 101 };
 102
 103 struct arp_pkt {
 104         __be16  hw_addr_space;
 105         __be16  prot_addr_space;
 106         u8      hw_addr_len;
 107         u8      prot_addr_len;
 108         __be16  op_code;
 109         u8      mac_src[ETH_ALEN];      /* sender hardware address */
 110         __be32  ip_src;                 /* sender IP address */
 111         u8      mac_dst[ETH_ALEN];      /* target hardware address */
 112         __be32  ip_dst;                 /* target IP address */
 113 };
 114 #pragma pack()
 115
 116 static inline struct arp_pkt *arp_pkt(const struct sk_buff *skb)
 117 {
 118         return (struct arp_pkt *)skb_network_header(skb);
 119 }
 120
 121 /* Forward declaration */
 122 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[]);
 123
 124 static inline u8 _simple_hash(const u8 *hash_start, int hash_size)
 125 {
 126         int i;
 127         u8 hash = 0;
 128
 129         for (i = 0; i < hash_size; i++) {
 130                 hash ^= hash_start[i];
 131         }
 132
 133         return hash;
 134 }
 135
 136 /*********************** tlb specific functions ***************************/
 137
 138 static inline void _lock_tx_hashtbl(struct bonding *bond)
 139 {
 140         spin_lock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 141 }
 142
 143 static inline void _unlock_tx_hashtbl(struct bonding *bond)
 144 {
 145         spin_unlock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 146 }
 147
 148 /* Caller must hold tx_hashtbl lock */
 149 static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
 150 {
 151         if (save_load) {
 152                 entry->load_history = 1 + entry->tx_bytes /
 153                                       BOND_TLB_REBALANCE_INTERVAL;
 154                 entry->tx_bytes = 0;
 155         }
 156
 157         entry->tx_slave = NULL;
 158         entry->next = TLB_NULL_INDEX;
 159         entry->prev = TLB_NULL_INDEX;
 160 }
 161
 162 static inline void tlb_init_slave(struct slave *slave)
 163 {
 164         SLAVE_TLB_INFO(slave).load = 0;
 165         SLAVE_TLB_INFO(slave).head = TLB_NULL_INDEX;
 166 }
 167
 168 /* Caller must hold bond lock for read */
 169 static void tlb_clear_slave(struct bonding *bond, struct slave *slave, int save_load)
 170 {
 171         struct tlb_client_info *tx_hash_table;
 172         u32 index;
 173
 174         _lock_tx_hashtbl(bond);
 175
 176         /* clear slave from tx_hashtbl */
 177         tx_hash_table = BOND_ALB_INFO(bond).tx_hashtbl;
 178
 179         /* skip this if we've already freed the tx hash table */
 180         if (tx_hash_table) {
 181                 index = SLAVE_TLB_INFO(slave).head;
 182                 while (index != TLB_NULL_INDEX) {
 183                         u32 next_index = tx_hash_table[index].next;
 184                         tlb_init_table_entry(&tx_hash_table[index], save_load);
 185                         index = next_index;
 186                 }
 187         }
 188
 189         tlb_init_slave(slave);
 190
 191         _unlock_tx_hashtbl(bond);
 192 }
 193
 194 /* Must be called before starting the monitor timer */
 195 static int tlb_initialize(struct bonding *bond)
 196 {
 197         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 198         int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
 199         struct tlb_client_info *new_hashtbl;
 200         int i;
 201
 202         spin_lock_init(&(bond_info->tx_hashtbl_lock));
 203
 204         new_hashtbl = kzalloc(size, GFP_KERNEL);
 205         if (!new_hashtbl) {
 206                 pr_err("%s: Error: Failed to allocate TLB hash table\n",
 207                        bond->dev->name);
 208                 return -1;
 209         }
 210         _lock_tx_hashtbl(bond);
 211
 212         bond_info->tx_hashtbl = new_hashtbl;
 213
 214         for (i = 0; i < TLB_HASH_TABLE_SIZE; i++) {
 215                 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 1);
 216         }
 217
 218         _unlock_tx_hashtbl(bond);
 219
 220         return 0;
 221 }
 222
 223 /* Must be called only after all slaves have been released */
 224 static void tlb_deinitialize(struct bonding *bond)
 225 {
 226         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 227
 228         _lock_tx_hashtbl(bond);
 229
 230         kfree(bond_info->tx_hashtbl);
 231         bond_info->tx_hashtbl = NULL;
 232
 233         _unlock_tx_hashtbl(bond);
 234 }
 235
 236 static long long compute_gap(struct slave *slave)
 237 {
 238         return (s64) (slave->speed << 20) - /* Convert to Megabit per sec */
 239                (s64) (SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
 240 }
 241
 242 /* Caller must hold bond lock for read */
 243 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
 244 {
 245         struct slave *slave, *least_loaded;
 246         long long max_gap;
 247         int i;
 248
 249         least_loaded = NULL;
 250         max_gap = LLONG_MIN;
 251
 252         /* Find the slave with the largest gap */
 253         bond_for_each_slave(bond, slave, i) {
 254                 if (SLAVE_IS_OK(slave)) {
 255                         long long gap = compute_gap(slave);
 256
 257                         if (max_gap < gap) {
 258                                 least_loaded = slave;
 259                                 max_gap = gap;
 260                         }
 261                 }
 262         }
 263
 264         return least_loaded;
 265 }
 266
 267 /* Caller must hold bond lock for read */
 268 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index, u32 skb_len)
 269 {
 270         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 271         struct tlb_client_info *hash_table;
 272         struct slave *assigned_slave;
 273
 274         _lock_tx_hashtbl(bond);
 275
 276         hash_table = bond_info->tx_hashtbl;
 277         assigned_slave = hash_table[hash_index].tx_slave;
 278         if (!assigned_slave) {
 279                 assigned_slave = tlb_get_least_loaded_slave(bond);
 280
 281                 if (assigned_slave) {
 282                         struct tlb_slave_info *slave_info =
 283                                 &(SLAVE_TLB_INFO(assigned_slave));
 284                         u32 next_index = slave_info->head;
 285
 286                         hash_table[hash_index].tx_slave = assigned_slave;
 287                         hash_table[hash_index].next = next_index;
 288                         hash_table[hash_index].prev = TLB_NULL_INDEX;
 289
 290                         if (next_index != TLB_NULL_INDEX) {
 291                                 hash_table[next_index].prev = hash_index;
 292                         }
 293
 294                         slave_info->head = hash_index;
 295                         slave_info->load +=
 296                                 hash_table[hash_index].load_history;
 297                 }
 298         }
 299
 300         if (assigned_slave) {
 301                 hash_table[hash_index].tx_bytes += skb_len;
 302         }
 303
 304         _unlock_tx_hashtbl(bond);
 305
 306         return assigned_slave;
 307 }
 308
 309 /*********************** rlb specific functions ***************************/
 310 static inline void _lock_rx_hashtbl(struct bonding *bond)
 311 {
 312         spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 313 }
 314
 315 static inline void _unlock_rx_hashtbl(struct bonding *bond)
 316 {
 317         spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 318 }
 319
 320 /* when an ARP REPLY is received from a client update its info
 321  * in the rx_hashtbl
 322  */
 323 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
 324 {
 325         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 326         struct rlb_client_info *client_info;
 327         u32 hash_index;
 328
 329         _lock_rx_hashtbl(bond);
 330
 331         hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
 332         client_info = &(bond_info->rx_hashtbl[hash_index]);
 333
 334         if ((client_info->assigned) &&
 335             (client_info->ip_src == arp->ip_dst) &&
 336             (client_info->ip_dst == arp->ip_src) &&
 337             (compare_ether_addr_64bits(client_info->mac_dst, arp->mac_src))) {
 338                 /* update the clients MAC address */
 339                 memcpy(client_info->mac_dst, arp->mac_src, ETH_ALEN);
 340                 client_info->ntt = 1;
 341                 bond_info->rx_ntt = 1;
 342         }
 343
 344         _unlock_rx_hashtbl(bond);
 345 }
 346
 347 static int rlb_arp_recv(struct sk_buff *skb, struct net_device *bond_dev, struct packet_type *ptype, struct net_device *orig_dev)
 348 {
 349         struct bonding *bond;
 350         struct arp_pkt *arp = (struct arp_pkt *)skb->data;
 351         int res = NET_RX_DROP;
 352
 353         while (bond_dev->priv_flags & IFF_802_1Q_VLAN)
 354                 bond_dev = vlan_dev_real_dev(bond_dev);
 355
 356         if (!(bond_dev->priv_flags & IFF_BONDING) ||
 357             !(bond_dev->flags & IFF_MASTER))
 358                 goto out;
 359
 360         if (!arp) {
 361                 pr_debug("Packet has no ARP data\n");
 362                 goto out;
 363         }
 364
 365         if (skb->len < sizeof(struct arp_pkt)) {
 366                 pr_debug("Packet is too small to be an ARP\n");
 367                 goto out;
 368         }
 369
 370         if (arp->op_code == htons(ARPOP_REPLY)) {
 371                 /* update rx hash table for this ARP */
 372                 bond = netdev_priv(bond_dev);
 373                 rlb_update_entry_from_arp(bond, arp);
 374                 pr_debug("Server received an ARP Reply from client\n");
 375         }
 376
 377         res = NET_RX_SUCCESS;
 378
 379 out:
 380         dev_kfree_skb(skb);
 381
 382         return res;
 383 }
 384
 385 /* Caller must hold bond lock for read */
 386 static struct slave *rlb_next_rx_slave(struct bonding *bond)
 387 {
 388         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 389         struct slave *rx_slave, *slave, *start_at;
 390         int i = 0;
 391
 392         if (bond_info->next_rx_slave) {
 393                 start_at = bond_info->next_rx_slave;
 394         } else {
 395                 start_at = bond->first_slave;
 396         }
 397
 398         rx_slave = NULL;
 399
 400         bond_for_each_slave_from(bond, slave, i, start_at) {
 401                 if (SLAVE_IS_OK(slave)) {
 402                         if (!rx_slave) {
 403                                 rx_slave = slave;
 404                         } else if (slave->speed > rx_slave->speed) {
 405                                 rx_slave = slave;
 406                         }
 407                 }
 408         }
 409
 410         if (rx_slave) {
 411                 bond_info->next_rx_slave = rx_slave->next;
 412         }
 413
 414         return rx_slave;
 415 }
 416
 417 /* teach the switch the mac of a disabled slave
 418  * on the primary for fault tolerance
 419  *
 420  * Caller must hold bond->curr_slave_lock for write or bond lock for write
 421  */
 422 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
 423 {
 424         if (!bond->curr_active_slave) {
 425                 return;
 426         }
 427
 428         if (!bond->alb_info.primary_is_promisc) {
 429                 if (!dev_set_promiscuity(bond->curr_active_slave->dev, 1))
 430                         bond->alb_info.primary_is_promisc = 1;
 431                 else
 432                         bond->alb_info.primary_is_promisc = 0;
 433         }
 434
 435         bond->alb_info.rlb_promisc_timeout_counter = 0;
 436
 437         alb_send_learning_packets(bond->curr_active_slave, addr);
 438 }
 439
 440 /* slave being removed should not be active at this point
 441  *
 442  * Caller must hold bond lock for read
 443  */
 444 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
 445 {
 446         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 447         struct rlb_client_info *rx_hash_table;
 448         u32 index, next_index;
 449
 450         /* clear slave from rx_hashtbl */
 451         _lock_rx_hashtbl(bond);
 452
 453         rx_hash_table = bond_info->rx_hashtbl;
 454         index = bond_info->rx_hashtbl_head;
 455         for (; index != RLB_NULL_INDEX; index = next_index) {
 456                 next_index = rx_hash_table[index].next;
 457                 if (rx_hash_table[index].slave == slave) {
 458                         struct slave *assigned_slave = rlb_next_rx_slave(bond);
 459
 460                         if (assigned_slave) {
 461                                 rx_hash_table[index].slave = assigned_slave;
 462                                 if (compare_ether_addr_64bits(rx_hash_table[index].mac_dst,
 463                                                               mac_bcast)) {
 464                                         bond_info->rx_hashtbl[index].ntt = 1;
 465                                         bond_info->rx_ntt = 1;
 466                                         /* A slave has been removed from the
 467                                          * table because it is either disabled
 468                                          * or being released. We must retry the
 469                                          * update to avoid clients from not
 470                                          * being updated & disconnecting when
 471                                          * there is stress
 472                                          */
 473                                         bond_info->rlb_update_retry_counter =
 474                                                 RLB_UPDATE_RETRY;
 475                                 }
 476                         } else {  /* there is no active slave */
 477                                 rx_hash_table[index].slave = NULL;
 478                         }
 479                 }
 480         }
 481
 482         _unlock_rx_hashtbl(bond);
 483
 484         write_lock_bh(&bond->curr_slave_lock);
 485
 486         if (slave != bond->curr_active_slave) {
 487                 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
 488         }
 489
 490         write_unlock_bh(&bond->curr_slave_lock);
 491 }
 492
 493 static void rlb_update_client(struct rlb_client_info *client_info)
 494 {
 495         int i;
 496
 497         if (!client_info->slave) {
 498                 return;
 499         }
 500
 501         for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
 502                 struct sk_buff *skb;
 503
 504                 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
 505                                  client_info->ip_dst,
 506                                  client_info->slave->dev,
 507                                  client_info->ip_src,
 508                                  client_info->mac_dst,
 509                                  client_info->slave->dev->dev_addr,
 510                                  client_info->mac_dst);
 511                 if (!skb) {
 512                         pr_err("%s: Error: failed to create an ARP packet\n",
 513                                client_info->slave->dev->master->name);
 514                         continue;
 515                 }
 516
 517                 skb->dev = client_info->slave->dev;
 518
 519                 if (client_info->tag) {
 520                         skb = vlan_put_tag(skb, client_info->vlan_id);
 521                         if (!skb) {
 522                                 pr_err("%s: Error: failed to insert VLAN tag\n",
 523                                        client_info->slave->dev->master->name);
 524                                 continue;
 525                         }
 526                 }
 527
 528                 arp_xmit(skb);
 529         }
 530 }
 531
 532 /* sends ARP REPLIES that update the clients that need updating */
 533 static void rlb_update_rx_clients(struct bonding *bond)
 534 {
 535         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 536         struct rlb_client_info *client_info;
 537         u32 hash_index;
 538
 539         _lock_rx_hashtbl(bond);
 540
 541         hash_index = bond_info->rx_hashtbl_head;
 542         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 543                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 544                 if (client_info->ntt) {
 545                         rlb_update_client(client_info);
 546                         if (bond_info->rlb_update_retry_counter == 0) {
 547                                 client_info->ntt = 0;
 548                         }
 549                 }
 550         }
 551
 552         /* do not update the entries again until this counter is zero so that
 553          * not to confuse the clients.
 554          */
 555         bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
 556
 557         _unlock_rx_hashtbl(bond);
 558 }
 559
 560 /* The slave was assigned a new mac address - update the clients */
 561 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
 562 {
 563         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 564         struct rlb_client_info *client_info;
 565         int ntt = 0;
 566         u32 hash_index;
 567
 568         _lock_rx_hashtbl(bond);
 569
 570         hash_index = bond_info->rx_hashtbl_head;
 571         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 572                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 573
 574                 if ((client_info->slave == slave) &&
 575                     compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 576                         client_info->ntt = 1;
 577                         ntt = 1;
 578                 }
 579         }
 580
 581         // update the team's flag only after the whole iteration
 582         if (ntt) {
 583                 bond_info->rx_ntt = 1;
 584                 //fasten the change
 585                 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
 586         }
 587
 588         _unlock_rx_hashtbl(bond);
 589 }
 590
 591 /* mark all clients using src_ip to be updated */
 592 static void rlb_req_update_subnet_clients(struct bonding *bond, __be32 src_ip)
 593 {
 594         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 595         struct rlb_client_info *client_info;
 596         u32 hash_index;
 597
 598         _lock_rx_hashtbl(bond);
 599
 600         hash_index = bond_info->rx_hashtbl_head;
 601         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 602                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 603
 604                 if (!client_info->slave) {
 605                         pr_err("%s: Error: found a client with no channel in the client's hash table\n",
 606                                bond->dev->name);
 607                         continue;
 608                 }
 609                 /*update all clients using this src_ip, that are not assigned
 610                  * to the team's address (curr_active_slave) and have a known
 611                  * unicast mac address.
 612                  */
 613                 if ((client_info->ip_src == src_ip) &&
 614                     compare_ether_addr_64bits(client_info->slave->dev->dev_addr,
 615                            bond->dev->dev_addr) &&
 616                     compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 617                         client_info->ntt = 1;
 618                         bond_info->rx_ntt = 1;
 619                 }
 620         }
 621
 622         _unlock_rx_hashtbl(bond);
 623 }
 624
 625 /* Caller must hold both bond and ptr locks for read */
 626 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
 627 {
 628         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 629         struct arp_pkt *arp = arp_pkt(skb);
 630         struct slave *assigned_slave;
 631         struct rlb_client_info *client_info;
 632         u32 hash_index = 0;
 633
 634         _lock_rx_hashtbl(bond);
 635
 636         hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_src));
 637         client_info = &(bond_info->rx_hashtbl[hash_index]);
 638
 639         if (client_info->assigned) {
 640                 if ((client_info->ip_src == arp->ip_src) &&
 641                     (client_info->ip_dst == arp->ip_dst)) {
 642                         /* the entry is already assigned to this client */
 643                         if (compare_ether_addr_64bits(arp->mac_dst, mac_bcast)) {
 644                                 /* update mac address from arp */
 645                                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 646                         }
 647
 648                         assigned_slave = client_info->slave;
 649                         if (assigned_slave) {
 650                                 _unlock_rx_hashtbl(bond);
 651                                 return assigned_slave;
 652                         }
 653                 } else {
 654                         /* the entry is already assigned to some other client,
 655                          * move the old client to primary (curr_active_slave) so
 656                          * that the new client can be assigned to this entry.
 657                          */
 658                         if (bond->curr_active_slave &&
 659                             client_info->slave != bond->curr_active_slave) {
 660                                 client_info->slave = bond->curr_active_slave;
 661                                 rlb_update_client(client_info);
 662                         }
 663                 }
 664         }
 665         /* assign a new slave */
 666         assigned_slave = rlb_next_rx_slave(bond);
 667
 668         if (assigned_slave) {
 669                 client_info->ip_src = arp->ip_src;
 670                 client_info->ip_dst = arp->ip_dst;
 671                 /* arp->mac_dst is broadcast for arp reqeusts.
 672                  * will be updated with clients actual unicast mac address
 673                  * upon receiving an arp reply.
 674                  */
 675                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 676                 client_info->slave = assigned_slave;
 677
 678                 if (compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 679                         client_info->ntt = 1;
 680                         bond->alb_info.rx_ntt = 1;
 681                 } else {
 682                         client_info->ntt = 0;
 683                 }
 684
 685                 if (bond->vlgrp) {
 686                         if (!vlan_get_tag(skb, &client_info->vlan_id))
 687                                 client_info->tag = 1;
 688                 }
 689
 690                 if (!client_info->assigned) {
 691                         u32 prev_tbl_head = bond_info->rx_hashtbl_head;
 692                         bond_info->rx_hashtbl_head = hash_index;
 693                         client_info->next = prev_tbl_head;
 694                         if (prev_tbl_head != RLB_NULL_INDEX) {
 695                                 bond_info->rx_hashtbl[prev_tbl_head].prev =
 696                                         hash_index;
 697                         }
 698                         client_info->assigned = 1;
 699                 }
 700         }
 701
 702         _unlock_rx_hashtbl(bond);
 703
 704         return assigned_slave;
 705 }
 706
 707 /* chooses (and returns) transmit channel for arp reply
 708  * does not choose channel for other arp types since they are
 709  * sent on the curr_active_slave
 710  */
 711 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
 712 {
 713         struct arp_pkt *arp = arp_pkt(skb);
 714         struct slave *tx_slave = NULL;
 715
 716         if (arp->op_code == htons(ARPOP_REPLY)) {
 717                 /* the arp must be sent on the selected
 718                 * rx channel
 719                 */
 720                 tx_slave = rlb_choose_channel(skb, bond);
 721                 if (tx_slave) {
 722                         memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
 723                 }
 724                 pr_debug("Server sent ARP Reply packet\n");
 725         } else if (arp->op_code == htons(ARPOP_REQUEST)) {
 726                 /* Create an entry in the rx_hashtbl for this client as a
 727                  * place holder.
 728                  * When the arp reply is received the entry will be updated
 729                  * with the correct unicast address of the client.
 730                  */
 731                 rlb_choose_channel(skb, bond);
 732
 733                 /* The ARP relpy packets must be delayed so that
 734                  * they can cancel out the influence of the ARP request.
 735                  */
 736                 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
 737
 738                 /* arp requests are broadcast and are sent on the primary
 739                  * the arp request will collapse all clients on the subnet to
 740                  * the primary slave. We must register these clients to be
 741                  * updated with their assigned mac.
 742                  */
 743                 rlb_req_update_subnet_clients(bond, arp->ip_src);
 744                 pr_debug("Server sent ARP Request packet\n");
 745         }
 746
 747         return tx_slave;
 748 }
 749
 750 /* Caller must hold bond lock for read */
 751 static void rlb_rebalance(struct bonding *bond)
 752 {
 753         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 754         struct slave *assigned_slave;
 755         struct rlb_client_info *client_info;
 756         int ntt;
 757         u32 hash_index;
 758
 759         _lock_rx_hashtbl(bond);
 760
 761         ntt = 0;
 762         hash_index = bond_info->rx_hashtbl_head;
 763         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 764                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 765                 assigned_slave = rlb_next_rx_slave(bond);
 766                 if (assigned_slave && (client_info->slave != assigned_slave)) {
 767                         client_info->slave = assigned_slave;
 768                         client_info->ntt = 1;
 769                         ntt = 1;
 770                 }
 771         }
 772
 773         /* update the team's flag only after the whole iteration */
 774         if (ntt) {
 775                 bond_info->rx_ntt = 1;
 776         }
 777         _unlock_rx_hashtbl(bond);
 778 }
 779
 780 /* Caller must hold rx_hashtbl lock */
 781 static void rlb_init_table_entry(struct rlb_client_info *entry)
 782 {
 783         memset(entry, 0, sizeof(struct rlb_client_info));
 784         entry->next = RLB_NULL_INDEX;
 785         entry->prev = RLB_NULL_INDEX;
 786 }
 787
 788 static int rlb_initialize(struct bonding *bond)
 789 {
 790         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 791         struct packet_type *pk_type = &(BOND_ALB_INFO(bond).rlb_pkt_type);
 792         struct rlb_client_info  *new_hashtbl;
 793         int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
 794         int i;
 795
 796         spin_lock_init(&(bond_info->rx_hashtbl_lock));
 797
 798         new_hashtbl = kmalloc(size, GFP_KERNEL);
 799         if (!new_hashtbl) {
 800                 pr_err("%s: Error: Failed to allocate RLB hash table\n",
 801                        bond->dev->name);
 802                 return -1;
 803         }
 804         _lock_rx_hashtbl(bond);
 805
 806         bond_info->rx_hashtbl = new_hashtbl;
 807
 808         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 809
 810         for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
 811                 rlb_init_table_entry(bond_info->rx_hashtbl + i);
 812         }
 813
 814         _unlock_rx_hashtbl(bond);
 815
 816         /*initialize packet type*/
 817         pk_type->type = cpu_to_be16(ETH_P_ARP);
 818         pk_type->dev = bond->dev;
 819         pk_type->func = rlb_arp_recv;
 820
 821         /* register to receive ARPs */
 822         dev_add_pack(pk_type);
 823
 824         return 0;
 825 }
 826
 827 static void rlb_deinitialize(struct bonding *bond)
 828 {
 829         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 830
 831         dev_remove_pack(&(bond_info->rlb_pkt_type));
 832
 833         _lock_rx_hashtbl(bond);
 834
 835         kfree(bond_info->rx_hashtbl);
 836         bond_info->rx_hashtbl = NULL;
 837         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 838
 839         _unlock_rx_hashtbl(bond);
 840 }
 841
 842 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
 843 {
 844         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 845         u32 curr_index;
 846
 847         _lock_rx_hashtbl(bond);
 848
 849         curr_index = bond_info->rx_hashtbl_head;
 850         while (curr_index != RLB_NULL_INDEX) {
 851                 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
 852                 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
 853                 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
 854
 855                 if (curr->tag && (curr->vlan_id == vlan_id)) {
 856                         if (curr_index == bond_info->rx_hashtbl_head) {
 857                                 bond_info->rx_hashtbl_head = next_index;
 858                         }
 859                         if (prev_index != RLB_NULL_INDEX) {
 860                                 bond_info->rx_hashtbl[prev_index].next = next_index;
 861                         }
 862                         if (next_index != RLB_NULL_INDEX) {
 863                                 bond_info->rx_hashtbl[next_index].prev = prev_index;
 864                         }
 865
 866                         rlb_init_table_entry(curr);
 867                 }
 868
 869                 curr_index = next_index;
 870         }
 871
 872         _unlock_rx_hashtbl(bond);
 873 }
 874
 875 /*********************** tlb/rlb shared functions *********************/
 876
 877 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
 878 {
 879         struct bonding *bond = bond_get_bond_by_slave(slave);
 880         struct learning_pkt pkt;
 881         int size = sizeof(struct learning_pkt);
 882         int i;
 883
 884         memset(&pkt, 0, size);
 885         memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
 886         memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
 887         pkt.type = cpu_to_be16(ETH_P_LOOP);
 888
 889         for (i = 0; i < MAX_LP_BURST; i++) {
 890                 struct sk_buff *skb;
 891                 char *data;
 892
 893                 skb = dev_alloc_skb(size);
 894                 if (!skb) {
 895                         return;
 896                 }
 897
 898                 data = skb_put(skb, size);
 899                 memcpy(data, &pkt, size);
 900
 901                 skb_reset_mac_header(skb);
 902                 skb->network_header = skb->mac_header + ETH_HLEN;
 903                 skb->protocol = pkt.type;
 904                 skb->priority = TC_PRIO_CONTROL;
 905                 skb->dev = slave->dev;
 906
 907                 if (bond->vlgrp) {
 908                         struct vlan_entry *vlan;
 909
 910                         vlan = bond_next_vlan(bond,
 911                                               bond->alb_info.current_alb_vlan);
 912
 913                         bond->alb_info.current_alb_vlan = vlan;
 914                         if (!vlan) {
 915                                 kfree_skb(skb);
 916                                 continue;
 917                         }
 918
 919                         skb = vlan_put_tag(skb, vlan->vlan_id);
 920                         if (!skb) {
 921                                 pr_err("%s: Error: failed to insert VLAN tag\n",
 922                                        bond->dev->name);
 923                                 continue;
 924                         }
 925                 }
 926
 927                 dev_queue_xmit(skb);
 928         }
 929 }
 930
 931 /* hw is a boolean parameter that determines whether we should try and
 932  * set the hw address of the device as well as the hw address of the
 933  * net_device
 934  */
 935 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[], int hw)
 936 {
 937         struct net_device *dev = slave->dev;
 938         struct sockaddr s_addr;
 939
 940         if (!hw) {
 941                 memcpy(dev->dev_addr, addr, dev->addr_len);
 942                 return 0;
 943         }
 944
 945         /* for rlb each slave must have a unique hw mac addresses so that */
 946         /* each slave will receive packets destined to a different mac */
 947         memcpy(s_addr.sa_data, addr, dev->addr_len);
 948         s_addr.sa_family = dev->type;
 949         if (dev_set_mac_address(dev, &s_addr)) {
 950                 pr_err("%s: Error: dev_set_mac_address of dev %s failed!\n"
 951                        "ALB mode requires that the base driver support setting the hw address also when the network device's interface is open\n",
 952                        dev->master->name, dev->name);
 953                 return -EOPNOTSUPP;
 954         }
 955         return 0;
 956 }
 957
 958 /*
 959  * Swap MAC addresses between two slaves.
 960  *
 961  * Called with RTNL held, and no other locks.
 962  *
 963  */
 964
 965 static void alb_swap_mac_addr(struct bonding *bond, struct slave *slave1, struct slave *slave2)
 966 {
 967         u8 tmp_mac_addr[ETH_ALEN];
 968
 969         memcpy(tmp_mac_addr, slave1->dev->dev_addr, ETH_ALEN);
 970         alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr, bond->alb_info.rlb_enabled);
 971         alb_set_slave_mac_addr(slave2, tmp_mac_addr, bond->alb_info.rlb_enabled);
 972
 973 }
 974
 975 /*
 976  * Send learning packets after MAC address swap.
 977  *
 978  * Called with RTNL and no other locks
 979  */
 980 static void alb_fasten_mac_swap(struct bonding *bond, struct slave *slave1,
 981                                 struct slave *slave2)
 982 {
 983         int slaves_state_differ = (SLAVE_IS_OK(slave1) != SLAVE_IS_OK(slave2));
 984         struct slave *disabled_slave = NULL;
 985
 986         ASSERT_RTNL();
 987
 988         /* fasten the change in the switch */
 989         if (SLAVE_IS_OK(slave1)) {
 990                 alb_send_learning_packets(slave1, slave1->dev->dev_addr);
 991                 if (bond->alb_info.rlb_enabled) {
 992                         /* inform the clients that the mac address
 993                          * has changed
 994                          */
 995                         rlb_req_update_slave_clients(bond, slave1);
 996                 }
 997         } else {
 998                 disabled_slave = slave1;
 999         }
1000
1001         if (SLAVE_IS_OK(slave2)) {
1002                 alb_send_learning_packets(slave2, slave2->dev->dev_addr);
1003                 if (bond->alb_info.rlb_enabled) {
1004                         /* inform the clients that the mac address
1005                          * has changed
1006                          */
1007                         rlb_req_update_slave_clients(bond, slave2);
1008                 }
1009         } else {
1010                 disabled_slave = slave2;
1011         }
1012
1013         if (bond->alb_info.rlb_enabled && slaves_state_differ) {
1014                 /* A disabled slave was assigned an active mac addr */
1015                 rlb_teach_disabled_mac_on_primary(bond,
1016                                                   disabled_slave->dev->dev_addr);
1017         }
1018 }
1019
1020 /**
1021  * alb_change_hw_addr_on_detach
1022  * @bond: bonding we're working on
1023  * @slave: the slave that was just detached
1024  *
1025  * We assume that @slave was already detached from the slave list.
1026  *
1027  * If @slave's permanent hw address is different both from its current
1028  * address and from @bond's address, then somewhere in the bond there's
1029  * a slave that has @slave's permanet address as its current address.
1030  * We'll make sure that that slave no longer uses @slave's permanent address.
1031  *
1032  * Caller must hold RTNL and no other locks
1033  */
1034 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
1035 {
1036         int perm_curr_diff;
1037         int perm_bond_diff;
1038
1039         perm_curr_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
1040                                                    slave->dev->dev_addr);
1041         perm_bond_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
1042                                                    bond->dev->dev_addr);
1043
1044         if (perm_curr_diff && perm_bond_diff) {
1045                 struct slave *tmp_slave;
1046                 int i, found = 0;
1047
1048                 bond_for_each_slave(bond, tmp_slave, i) {
1049                         if (!compare_ether_addr_64bits(slave->perm_hwaddr,
1050                                                        tmp_slave->dev->dev_addr)) {
1051                                 found = 1;
1052                                 break;
1053                         }
1054                 }
1055
1056                 if (found) {
1057                         /* locking: needs RTNL and nothing else */
1058                         alb_swap_mac_addr(bond, slave, tmp_slave);
1059                         alb_fasten_mac_swap(bond, slave, tmp_slave);
1060                 }
1061         }
1062 }
1063
1064 /**
1065  * alb_handle_addr_collision_on_attach
1066  * @bond: bonding we're working on
1067  * @slave: the slave that was just attached
1068  *
1069  * checks uniqueness of slave's mac address and handles the case the
1070  * new slave uses the bonds mac address.
1071  *
1072  * If the permanent hw address of @slave is @bond's hw address, we need to
1073  * find a different hw address to give @slave, that isn't in use by any other
1074  * slave in the bond. This address must be, of course, one of the premanent
1075  * addresses of the other slaves.
1076  *
1077  * We go over the slave list, and for each slave there we compare its
1078  * permanent hw address with the current address of all the other slaves.
1079  * If no match was found, then we've found a slave with a permanent address
1080  * that isn't used by any other slave in the bond, so we can assign it to
1081  * @slave.
1082  *
1083  * assumption: this function is called before @slave is attached to the
1084  *             bond slave list.
1085  *
1086  * caller must hold the bond lock for write since the mac addresses are compared
1087  * and may be swapped.
1088  */
1089 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1090 {
1091         struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1092         struct slave *has_bond_addr = bond->curr_active_slave;
1093         int i, j, found = 0;
1094
1095         if (bond->slave_cnt == 0) {
1096                 /* this is the first slave */
1097                 return 0;
1098         }
1099
1100         /* if slave's mac address differs from bond's mac address
1101          * check uniqueness of slave's mac address against the other
1102          * slaves in the bond.
1103          */
1104         if (compare_ether_addr_64bits(slave->perm_hwaddr, bond->dev->dev_addr)) {
1105                 bond_for_each_slave(bond, tmp_slave1, i) {
1106                         if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1107                                                        slave->dev->dev_addr)) {
1108                                 found = 1;
1109                                 break;
1110                         }
1111                 }
1112
1113                 if (!found)
1114                         return 0;
1115
1116                 /* Try setting slave mac to bond address and fall-through
1117                    to code handling that situation below... */
1118                 alb_set_slave_mac_addr(slave, bond->dev->dev_addr,
1119                                        bond->alb_info.rlb_enabled);
1120         }
1121
1122         /* The slave's address is equal to the address of the bond.
1123          * Search for a spare address in the bond for this slave.
1124          */
1125         free_mac_slave = NULL;
1126
1127         bond_for_each_slave(bond, tmp_slave1, i) {
1128                 found = 0;
1129                 bond_for_each_slave(bond, tmp_slave2, j) {
1130                         if (!compare_ether_addr_64bits(tmp_slave1->perm_hwaddr,
1131                                                        tmp_slave2->dev->dev_addr)) {
1132                                 found = 1;
1133                                 break;
1134                         }
1135                 }
1136
1137                 if (!found) {
1138                         /* no slave has tmp_slave1's perm addr
1139                          * as its curr addr
1140                          */
1141                         free_mac_slave = tmp_slave1;
1142                         break;
1143                 }
1144
1145                 if (!has_bond_addr) {
1146                         if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1147                                                        bond->dev->dev_addr)) {
1148
1149                                 has_bond_addr = tmp_slave1;
1150                         }
1151                 }
1152         }
1153
1154         if (free_mac_slave) {
1155                 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr,
1156                                        bond->alb_info.rlb_enabled);
1157
1158                 pr_warning("%s: Warning: the hw address of slave %s is in use by the bond; giving it the hw address of %s\n",
1159                            bond->dev->name, slave->dev->name,
1160                            free_mac_slave->dev->name);
1161
1162         } else if (has_bond_addr) {
1163                 pr_err("%s: Error: the hw address of slave %s is in use by the bond; couldn't find a slave with a free hw address to give it (this should not have happened)\n",
1164                        bond->dev->name, slave->dev->name);
1165                 return -EFAULT;
1166         }
1167
1168         return 0;
1169 }
1170
1171 /**
1172  * alb_set_mac_address
1173  * @bond:
1174  * @addr:
1175  *
1176  * In TLB mode all slaves are configured to the bond's hw address, but set
1177  * their dev_addr field to different addresses (based on their permanent hw
1178  * addresses).
1179  *
1180  * For each slave, this function sets the interface to the new address and then
1181  * changes its dev_addr field to its previous value.
1182  *
1183  * Unwinding assumes bond's mac address has not yet changed.
1184  */
1185 static int alb_set_mac_address(struct bonding *bond, void *addr)
1186 {
1187         struct sockaddr sa;
1188         struct slave *slave, *stop_at;
1189         char tmp_addr[ETH_ALEN];
1190         int res;
1191         int i;
1192
1193         if (bond->alb_info.rlb_enabled) {
1194                 return 0;
1195         }
1196
1197         bond_for_each_slave(bond, slave, i) {
1198                 /* save net_device's current hw address */
1199                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1200
1201                 res = dev_set_mac_address(slave->dev, addr);
1202
1203                 /* restore net_device's hw address */
1204                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1205
1206                 if (res)
1207                         goto unwind;
1208         }
1209
1210         return 0;
1211
1212 unwind:
1213         memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1214         sa.sa_family = bond->dev->type;
1215
1216         /* unwind from head to the slave that failed */
1217         stop_at = slave;
1218         bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1219                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1220                 dev_set_mac_address(slave->dev, &sa);
1221                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1222         }
1223
1224         return res;
1225 }
1226
1227 /************************ exported alb funcions ************************/
1228
1229 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1230 {
1231         int res;
1232
1233         res = tlb_initialize(bond);
1234         if (res) {
1235                 return res;
1236         }
1237
1238         if (rlb_enabled) {
1239                 bond->alb_info.rlb_enabled = 1;
1240                 /* initialize rlb */
1241                 res = rlb_initialize(bond);
1242                 if (res) {
1243                         tlb_deinitialize(bond);
1244                         return res;
1245                 }
1246         } else {
1247                 bond->alb_info.rlb_enabled = 0;
1248         }
1249
1250         return 0;
1251 }
1252
1253 void bond_alb_deinitialize(struct bonding *bond)
1254 {
1255         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1256
1257         tlb_deinitialize(bond);
1258
1259         if (bond_info->rlb_enabled) {
1260                 rlb_deinitialize(bond);
1261         }
1262 }
1263
1264 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1265 {
1266         struct bonding *bond = netdev_priv(bond_dev);
1267         struct ethhdr *eth_data;
1268         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1269         struct slave *tx_slave = NULL;
1270         static const __be32 ip_bcast = htonl(0xffffffff);
1271         int hash_size = 0;
1272         int do_tx_balance = 1;
1273         u32 hash_index = 0;
1274         const u8 *hash_start = NULL;
1275         int res = 1;
1276         struct ipv6hdr *ip6hdr;
1277
1278         skb_reset_mac_header(skb);
1279         eth_data = eth_hdr(skb);
1280
1281         /* make sure that the curr_active_slave and the slaves list do
1282          * not change during tx
1283          */
1284         read_lock(&bond->lock);
1285         read_lock(&bond->curr_slave_lock);
1286
1287         if (!BOND_IS_OK(bond)) {
1288                 goto out;
1289         }
1290
1291         switch (ntohs(skb->protocol)) {
1292         case ETH_P_IP: {
1293                 const struct iphdr *iph = ip_hdr(skb);
1294
1295                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast) ||
1296                     (iph->daddr == ip_bcast) ||
1297                     (iph->protocol == IPPROTO_IGMP)) {
1298                         do_tx_balance = 0;
1299                         break;
1300                 }
1301                 hash_start = (char *)&(iph->daddr);
1302                 hash_size = sizeof(iph->daddr);
1303         }
1304                 break;
1305         case ETH_P_IPV6:
1306                 /* IPv6 doesn't really use broadcast mac address, but leave
1307                  * that here just in case.
1308                  */
1309                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast)) {
1310                         do_tx_balance = 0;
1311                         break;
1312                 }
1313
1314                 /* IPv6 uses all-nodes multicast as an equivalent to
1315                  * broadcasts in IPv4.
1316                  */
1317                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_v6_allmcast)) {
1318                         do_tx_balance = 0;
1319                         break;
1320                 }
1321
1322                 /* Additianally, DAD probes should not be tx-balanced as that
1323                  * will lead to false positives for duplicate addresses and
1324                  * prevent address configuration from working.
1325                  */
1326                 ip6hdr = ipv6_hdr(skb);
1327                 if (ipv6_addr_any(&ip6hdr->saddr)) {
1328                         do_tx_balance = 0;
1329                         break;
1330                 }
1331
1332                 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1333                 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1334                 break;
1335         case ETH_P_IPX:
1336                 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1337                         /* something is wrong with this packet */
1338                         do_tx_balance = 0;
1339                         break;
1340                 }
1341
1342                 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1343                         /* The only protocol worth balancing in
1344                          * this family since it has an "ARP" like
1345                          * mechanism
1346                          */
1347                         do_tx_balance = 0;
1348                         break;
1349                 }
1350
1351                 hash_start = (char*)eth_data->h_dest;
1352                 hash_size = ETH_ALEN;
1353                 break;
1354         case ETH_P_ARP:
1355                 do_tx_balance = 0;
1356                 if (bond_info->rlb_enabled) {
1357                         tx_slave = rlb_arp_xmit(skb, bond);
1358                 }
1359                 break;
1360         default:
1361                 do_tx_balance = 0;
1362                 break;
1363         }
1364
1365         if (do_tx_balance) {
1366                 hash_index = _simple_hash(hash_start, hash_size);
1367                 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1368         }
1369
1370         if (!tx_slave) {
1371                 /* unbalanced or unassigned, send through primary */
1372                 tx_slave = bond->curr_active_slave;
1373                 bond_info->unbalanced_load += skb->len;
1374         }
1375
1376         if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1377                 if (tx_slave != bond->curr_active_slave) {
1378                         memcpy(eth_data->h_source,
1379                                tx_slave->dev->dev_addr,
1380                                ETH_ALEN);
1381                 }
1382
1383                 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1384         } else {
1385                 if (tx_slave) {
1386                         tlb_clear_slave(bond, tx_slave, 0);
1387                 }
1388         }
1389
1390 out:
1391         if (res) {
1392                 /* no suitable interface, frame not sent */
1393                 dev_kfree_skb(skb);
1394         }
1395         read_unlock(&bond->curr_slave_lock);
1396         read_unlock(&bond->lock);
1397         return NETDEV_TX_OK;
1398 }
1399
1400 void bond_alb_monitor(struct work_struct *work)
1401 {
1402         struct bonding *bond = container_of(work, struct bonding,
1403                                             alb_work.work);
1404         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1405         struct slave *slave;
1406         int i;
1407
1408         read_lock(&bond->lock);
1409
1410         if (bond->kill_timers) {
1411                 goto out;
1412         }
1413
1414         if (bond->slave_cnt == 0) {
1415                 bond_info->tx_rebalance_counter = 0;
1416                 bond_info->lp_counter = 0;
1417                 goto re_arm;
1418         }
1419
1420         bond_info->tx_rebalance_counter++;
1421         bond_info->lp_counter++;
1422
1423         /* send learning packets */
1424         if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1425                 /* change of curr_active_slave involves swapping of mac addresses.
1426                  * in order to avoid this swapping from happening while
1427                  * sending the learning packets, the curr_slave_lock must be held for
1428                  * read.
1429                  */
1430                 read_lock(&bond->curr_slave_lock);
1431
1432                 bond_for_each_slave(bond, slave, i) {
1433                         alb_send_learning_packets(slave, slave->dev->dev_addr);
1434                 }
1435
1436                 read_unlock(&bond->curr_slave_lock);
1437
1438                 bond_info->lp_counter = 0;
1439         }
1440
1441         /* rebalance tx traffic */
1442         if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1443
1444                 read_lock(&bond->curr_slave_lock);
1445
1446                 bond_for_each_slave(bond, slave, i) {
1447                         tlb_clear_slave(bond, slave, 1);
1448                         if (slave == bond->curr_active_slave) {
1449                                 SLAVE_TLB_INFO(slave).load =
1450                                         bond_info->unbalanced_load /
1451                                                 BOND_TLB_REBALANCE_INTERVAL;
1452                                 bond_info->unbalanced_load = 0;
1453                         }
1454                 }
1455
1456                 read_unlock(&bond->curr_slave_lock);
1457
1458                 bond_info->tx_rebalance_counter = 0;
1459         }
1460
1461         /* handle rlb stuff */
1462         if (bond_info->rlb_enabled) {
1463                 if (bond_info->primary_is_promisc &&
1464                     (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1465
1466                         /*
1467                          * dev_set_promiscuity requires rtnl and
1468                          * nothing else.
1469                          */
1470                         read_unlock(&bond->lock);
1471                         rtnl_lock();
1472
1473                         bond_info->rlb_promisc_timeout_counter = 0;
1474
1475                         /* If the primary was set to promiscuous mode
1476                          * because a slave was disabled then
1477                          * it can now leave promiscuous mode.
1478                          */
1479                         dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1480                         bond_info->primary_is_promisc = 0;
1481
1482                         rtnl_unlock();
1483                         read_lock(&bond->lock);
1484                 }
1485
1486                 if (bond_info->rlb_rebalance) {
1487                         bond_info->rlb_rebalance = 0;
1488                         rlb_rebalance(bond);
1489                 }
1490
1491                 /* check if clients need updating */
1492                 if (bond_info->rx_ntt) {
1493                         if (bond_info->rlb_update_delay_counter) {
1494                                 --bond_info->rlb_update_delay_counter;
1495                         } else {
1496                                 rlb_update_rx_clients(bond);
1497                                 if (bond_info->rlb_update_retry_counter) {
1498                                         --bond_info->rlb_update_retry_counter;
1499                                 } else {
1500                                         bond_info->rx_ntt = 0;
1501                                 }
1502                         }
1503                 }
1504         }
1505
1506 re_arm:
1507         queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1508 out:
1509         read_unlock(&bond->lock);
1510 }
1511
1512 /* assumption: called before the slave is attached to the bond
1513  * and not locked by the bond lock
1514  */
1515 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1516 {
1517         int res;
1518
1519         res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr,
1520                                      bond->alb_info.rlb_enabled);
1521         if (res) {
1522                 return res;
1523         }
1524
1525         /* caller must hold the bond lock for write since the mac addresses
1526          * are compared and may be swapped.
1527          */
1528         read_lock(&bond->lock);
1529
1530         res = alb_handle_addr_collision_on_attach(bond, slave);
1531
1532         read_unlock(&bond->lock);
1533
1534         if (res) {
1535                 return res;
1536         }
1537
1538         tlb_init_slave(slave);
1539
1540         /* order a rebalance ASAP */
1541         bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1542
1543         if (bond->alb_info.rlb_enabled) {
1544                 bond->alb_info.rlb_rebalance = 1;
1545         }
1546
1547         return 0;
1548 }
1549
1550 /*
1551  * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1552  * if necessary.
1553  *
1554  * Caller must hold RTNL and no other locks
1555  */
1556 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1557 {
1558         if (bond->slave_cnt > 1) {
1559                 alb_change_hw_addr_on_detach(bond, slave);
1560         }
1561
1562         tlb_clear_slave(bond, slave, 0);
1563
1564         if (bond->alb_info.rlb_enabled) {
1565                 bond->alb_info.next_rx_slave = NULL;
1566                 rlb_clear_slave(bond, slave);
1567         }
1568 }
1569
1570 /* Caller must hold bond lock for read */
1571 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1572 {
1573         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1574
1575         if (link == BOND_LINK_DOWN) {
1576                 tlb_clear_slave(bond, slave, 0);
1577                 if (bond->alb_info.rlb_enabled) {
1578                         rlb_clear_slave(bond, slave);
1579                 }
1580         } else if (link == BOND_LINK_UP) {
1581                 /* order a rebalance ASAP */
1582                 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1583                 if (bond->alb_info.rlb_enabled) {
1584                         bond->alb_info.rlb_rebalance = 1;
1585                         /* If the updelay module parameter is smaller than the
1586                          * forwarding delay of the switch the rebalance will
1587                          * not work because the rebalance arp replies will
1588                          * not be forwarded to the clients..
1589                          */
1590                 }
1591         }
1592 }
1593
1594 /**
1595  * bond_alb_handle_active_change - assign new curr_active_slave
1596  * @bond: our bonding struct
1597  * @new_slave: new slave to assign
1598  *
1599  * Set the bond->curr_active_slave to @new_slave and handle
1600  * mac address swapping and promiscuity changes as needed.
1601  *
1602  * If new_slave is NULL, caller must hold curr_slave_lock or
1603  * bond->lock for write.
1604  *
1605  * If new_slave is not NULL, caller must hold RTNL, bond->lock for
1606  * read and curr_slave_lock for write.  Processing here may sleep, so
1607  * no other locks may be held.
1608  */
1609 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1610         __releases(&bond->curr_slave_lock)
1611         __releases(&bond->lock)
1612         __acquires(&bond->lock)
1613         __acquires(&bond->curr_slave_lock)
1614 {
1615         struct slave *swap_slave;
1616         int i;
1617
1618         if (bond->curr_active_slave == new_slave) {
1619                 return;
1620         }
1621
1622         if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1623                 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1624                 bond->alb_info.primary_is_promisc = 0;
1625                 bond->alb_info.rlb_promisc_timeout_counter = 0;
1626         }
1627
1628         swap_slave = bond->curr_active_slave;
1629         bond->curr_active_slave = new_slave;
1630
1631         if (!new_slave || (bond->slave_cnt == 0)) {
1632                 return;
1633         }
1634
1635         /* set the new curr_active_slave to the bonds mac address
1636          * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1637          */
1638         if (!swap_slave) {
1639                 struct slave *tmp_slave;
1640                 /* find slave that is holding the bond's mac address */
1641                 bond_for_each_slave(bond, tmp_slave, i) {
1642                         if (!compare_ether_addr_64bits(tmp_slave->dev->dev_addr,
1643                                                        bond->dev->dev_addr)) {
1644                                 swap_slave = tmp_slave;
1645                                 break;
1646                         }
1647                 }
1648         }
1649
1650         /*
1651          * Arrange for swap_slave and new_slave to temporarily be
1652          * ignored so we can mess with their MAC addresses without
1653          * fear of interference from transmit activity.
1654          */
1655         if (swap_slave) {
1656                 tlb_clear_slave(bond, swap_slave, 1);
1657         }
1658         tlb_clear_slave(bond, new_slave, 1);
1659
1660         write_unlock_bh(&bond->curr_slave_lock);
1661         read_unlock(&bond->lock);
1662
1663         ASSERT_RTNL();
1664
1665         /* curr_active_slave must be set before calling alb_swap_mac_addr */
1666         if (swap_slave) {
1667                 /* swap mac address */
1668                 alb_swap_mac_addr(bond, swap_slave, new_slave);
1669         } else {
1670                 /* set the new_slave to the bond mac address */
1671                 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr,
1672                                        bond->alb_info.rlb_enabled);
1673         }
1674
1675         if (swap_slave) {
1676                 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1677                 read_lock(&bond->lock);
1678         } else {
1679                 read_lock(&bond->lock);
1680                 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1681         }
1682
1683         write_lock_bh(&bond->curr_slave_lock);
1684 }
1685
1686 /*
1687  * Called with RTNL
1688  */
1689 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1690         __acquires(&bond->lock)
1691         __releases(&bond->lock)
1692 {
1693         struct bonding *bond = netdev_priv(bond_dev);
1694         struct sockaddr *sa = addr;
1695         struct slave *slave, *swap_slave;
1696         int res;
1697         int i;
1698
1699         if (!is_valid_ether_addr(sa->sa_data)) {
1700                 return -EADDRNOTAVAIL;
1701         }
1702
1703         res = alb_set_mac_address(bond, addr);
1704         if (res) {
1705                 return res;
1706         }
1707
1708         memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1709
1710         /* If there is no curr_active_slave there is nothing else to do.
1711          * Otherwise we'll need to pass the new address to it and handle
1712          * duplications.
1713          */
1714         if (!bond->curr_active_slave) {
1715                 return 0;
1716         }
1717
1718         swap_slave = NULL;
1719
1720         bond_for_each_slave(bond, slave, i) {
1721                 if (!compare_ether_addr_64bits(slave->dev->dev_addr,
1722                                                bond_dev->dev_addr)) {
1723                         swap_slave = slave;
1724                         break;
1725                 }
1726         }
1727
1728         if (swap_slave) {
1729                 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1730                 alb_fasten_mac_swap(bond, swap_slave, bond->curr_active_slave);
1731         } else {
1732                 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr,
1733                                        bond->alb_info.rlb_enabled);
1734
1735                 read_lock(&bond->lock);
1736                 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1737                 if (bond->alb_info.rlb_enabled) {
1738                         /* inform clients mac address has changed */
1739                         rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1740                 }
1741                 read_unlock(&bond->lock);
1742         }
1743
1744         return 0;
1745 }
1746
1747 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1748 {
1749         if (bond->alb_info.current_alb_vlan &&
1750             (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1751                 bond->alb_info.current_alb_vlan = NULL;
1752         }
1753
1754         if (bond->alb_info.rlb_enabled) {
1755                 rlb_clear_vlan(bond, vlan_id);
1756         }
1757 }
1758