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 /* Caller must hold bond lock for read */
 237 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
 238 {
 239         struct slave *slave, *least_loaded;
 240         s64 max_gap;
 241         int i, found = 0;
 242
 243         /* Find the first enabled slave */
 244         bond_for_each_slave(bond, slave, i) {
 245                 if (SLAVE_IS_OK(slave)) {
 246                         found = 1;
 247                         break;
 248                 }
 249         }
 250
 251         if (!found) {
 252                 return NULL;
 253         }
 254
 255         least_loaded = slave;
 256         max_gap = (s64)(slave->speed << 20) - /* Convert to Megabit per sec */
 257                         (s64)(SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
 258
 259         /* Find the slave with the largest gap */
 260         bond_for_each_slave_from(bond, slave, i, least_loaded) {
 261                 if (SLAVE_IS_OK(slave)) {
 262                         s64 gap = (s64)(slave->speed << 20) -
 263                                         (s64)(SLAVE_TLB_INFO(slave).load << 3);
 264                         if (max_gap < gap) {
 265                                 least_loaded = slave;
 266                                 max_gap = gap;
 267                         }
 268                 }
 269         }
 270
 271         return least_loaded;
 272 }
 273
 274 /* Caller must hold bond lock for read */
 275 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index, u32 skb_len)
 276 {
 277         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 278         struct tlb_client_info *hash_table;
 279         struct slave *assigned_slave;
 280
 281         _lock_tx_hashtbl(bond);
 282
 283         hash_table = bond_info->tx_hashtbl;
 284         assigned_slave = hash_table[hash_index].tx_slave;
 285         if (!assigned_slave) {
 286                 assigned_slave = tlb_get_least_loaded_slave(bond);
 287
 288                 if (assigned_slave) {
 289                         struct tlb_slave_info *slave_info =
 290                                 &(SLAVE_TLB_INFO(assigned_slave));
 291                         u32 next_index = slave_info->head;
 292
 293                         hash_table[hash_index].tx_slave = assigned_slave;
 294                         hash_table[hash_index].next = next_index;
 295                         hash_table[hash_index].prev = TLB_NULL_INDEX;
 296
 297                         if (next_index != TLB_NULL_INDEX) {
 298                                 hash_table[next_index].prev = hash_index;
 299                         }
 300
 301                         slave_info->head = hash_index;
 302                         slave_info->load +=
 303                                 hash_table[hash_index].load_history;
 304                 }
 305         }
 306
 307         if (assigned_slave) {
 308                 hash_table[hash_index].tx_bytes += skb_len;
 309         }
 310
 311         _unlock_tx_hashtbl(bond);
 312
 313         return assigned_slave;
 314 }
 315
 316 /*********************** rlb specific functions ***************************/
 317 static inline void _lock_rx_hashtbl(struct bonding *bond)
 318 {
 319         spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 320 }
 321
 322 static inline void _unlock_rx_hashtbl(struct bonding *bond)
 323 {
 324         spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 325 }
 326
 327 /* when an ARP REPLY is received from a client update its info
 328  * in the rx_hashtbl
 329  */
 330 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
 331 {
 332         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 333         struct rlb_client_info *client_info;
 334         u32 hash_index;
 335
 336         _lock_rx_hashtbl(bond);
 337
 338         hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
 339         client_info = &(bond_info->rx_hashtbl[hash_index]);
 340
 341         if ((client_info->assigned) &&
 342             (client_info->ip_src == arp->ip_dst) &&
 343             (client_info->ip_dst == arp->ip_src) &&
 344             (compare_ether_addr_64bits(client_info->mac_dst, arp->mac_src))) {
 345                 /* update the clients MAC address */
 346                 memcpy(client_info->mac_dst, arp->mac_src, ETH_ALEN);
 347                 client_info->ntt = 1;
 348                 bond_info->rx_ntt = 1;
 349         }
 350
 351         _unlock_rx_hashtbl(bond);
 352 }
 353
 354 static int rlb_arp_recv(struct sk_buff *skb, struct net_device *bond_dev, struct packet_type *ptype, struct net_device *orig_dev)
 355 {
 356         struct bonding *bond;
 357         struct arp_pkt *arp = (struct arp_pkt *)skb->data;
 358         int res = NET_RX_DROP;
 359
 360         while (bond_dev->priv_flags & IFF_802_1Q_VLAN)
 361                 bond_dev = vlan_dev_real_dev(bond_dev);
 362
 363         if (!(bond_dev->priv_flags & IFF_BONDING) ||
 364             !(bond_dev->flags & IFF_MASTER))
 365                 goto out;
 366
 367         if (!arp) {
 368                 pr_debug("Packet has no ARP data\n");
 369                 goto out;
 370         }
 371
 372         if (skb->len < sizeof(struct arp_pkt)) {
 373                 pr_debug("Packet is too small to be an ARP\n");
 374                 goto out;
 375         }
 376
 377         if (arp->op_code == htons(ARPOP_REPLY)) {
 378                 /* update rx hash table for this ARP */
 379                 bond = netdev_priv(bond_dev);
 380                 rlb_update_entry_from_arp(bond, arp);
 381                 pr_debug("Server received an ARP Reply from client\n");
 382         }
 383
 384         res = NET_RX_SUCCESS;
 385
 386 out:
 387         dev_kfree_skb(skb);
 388
 389         return res;
 390 }
 391
 392 /* Caller must hold bond lock for read */
 393 static struct slave *rlb_next_rx_slave(struct bonding *bond)
 394 {
 395         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 396         struct slave *rx_slave, *slave, *start_at;
 397         int i = 0;
 398
 399         if (bond_info->next_rx_slave) {
 400                 start_at = bond_info->next_rx_slave;
 401         } else {
 402                 start_at = bond->first_slave;
 403         }
 404
 405         rx_slave = NULL;
 406
 407         bond_for_each_slave_from(bond, slave, i, start_at) {
 408                 if (SLAVE_IS_OK(slave)) {
 409                         if (!rx_slave) {
 410                                 rx_slave = slave;
 411                         } else if (slave->speed > rx_slave->speed) {
 412                                 rx_slave = slave;
 413                         }
 414                 }
 415         }
 416
 417         if (rx_slave) {
 418                 bond_info->next_rx_slave = rx_slave->next;
 419         }
 420
 421         return rx_slave;
 422 }
 423
 424 /* teach the switch the mac of a disabled slave
 425  * on the primary for fault tolerance
 426  *
 427  * Caller must hold bond->curr_slave_lock for write or bond lock for write
 428  */
 429 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
 430 {
 431         if (!bond->curr_active_slave) {
 432                 return;
 433         }
 434
 435         if (!bond->alb_info.primary_is_promisc) {
 436                 if (!dev_set_promiscuity(bond->curr_active_slave->dev, 1))
 437                         bond->alb_info.primary_is_promisc = 1;
 438                 else
 439                         bond->alb_info.primary_is_promisc = 0;
 440         }
 441
 442         bond->alb_info.rlb_promisc_timeout_counter = 0;
 443
 444         alb_send_learning_packets(bond->curr_active_slave, addr);
 445 }
 446
 447 /* slave being removed should not be active at this point
 448  *
 449  * Caller must hold bond lock for read
 450  */
 451 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
 452 {
 453         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 454         struct rlb_client_info *rx_hash_table;
 455         u32 index, next_index;
 456
 457         /* clear slave from rx_hashtbl */
 458         _lock_rx_hashtbl(bond);
 459
 460         rx_hash_table = bond_info->rx_hashtbl;
 461         index = bond_info->rx_hashtbl_head;
 462         for (; index != RLB_NULL_INDEX; index = next_index) {
 463                 next_index = rx_hash_table[index].next;
 464                 if (rx_hash_table[index].slave == slave) {
 465                         struct slave *assigned_slave = rlb_next_rx_slave(bond);
 466
 467                         if (assigned_slave) {
 468                                 rx_hash_table[index].slave = assigned_slave;
 469                                 if (compare_ether_addr_64bits(rx_hash_table[index].mac_dst,
 470                                                               mac_bcast)) {
 471                                         bond_info->rx_hashtbl[index].ntt = 1;
 472                                         bond_info->rx_ntt = 1;
 473                                         /* A slave has been removed from the
 474                                          * table because it is either disabled
 475                                          * or being released. We must retry the
 476                                          * update to avoid clients from not
 477                                          * being updated & disconnecting when
 478                                          * there is stress
 479                                          */
 480                                         bond_info->rlb_update_retry_counter =
 481                                                 RLB_UPDATE_RETRY;
 482                                 }
 483                         } else {  /* there is no active slave */
 484                                 rx_hash_table[index].slave = NULL;
 485                         }
 486                 }
 487         }
 488
 489         _unlock_rx_hashtbl(bond);
 490
 491         write_lock_bh(&bond->curr_slave_lock);
 492
 493         if (slave != bond->curr_active_slave) {
 494                 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
 495         }
 496
 497         write_unlock_bh(&bond->curr_slave_lock);
 498 }
 499
 500 static void rlb_update_client(struct rlb_client_info *client_info)
 501 {
 502         int i;
 503
 504         if (!client_info->slave) {
 505                 return;
 506         }
 507
 508         for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
 509                 struct sk_buff *skb;
 510
 511                 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
 512                                  client_info->ip_dst,
 513                                  client_info->slave->dev,
 514                                  client_info->ip_src,
 515                                  client_info->mac_dst,
 516                                  client_info->slave->dev->dev_addr,
 517                                  client_info->mac_dst);
 518                 if (!skb) {
 519                         pr_err("%s: Error: failed to create an ARP packet\n",
 520                                client_info->slave->dev->master->name);
 521                         continue;
 522                 }
 523
 524                 skb->dev = client_info->slave->dev;
 525
 526                 if (client_info->tag) {
 527                         skb = vlan_put_tag(skb, client_info->vlan_id);
 528                         if (!skb) {
 529                                 pr_err("%s: Error: failed to insert VLAN tag\n",
 530                                        client_info->slave->dev->master->name);
 531                                 continue;
 532                         }
 533                 }
 534
 535                 arp_xmit(skb);
 536         }
 537 }
 538
 539 /* sends ARP REPLIES that update the clients that need updating */
 540 static void rlb_update_rx_clients(struct bonding *bond)
 541 {
 542         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 543         struct rlb_client_info *client_info;
 544         u32 hash_index;
 545
 546         _lock_rx_hashtbl(bond);
 547
 548         hash_index = bond_info->rx_hashtbl_head;
 549         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 550                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 551                 if (client_info->ntt) {
 552                         rlb_update_client(client_info);
 553                         if (bond_info->rlb_update_retry_counter == 0) {
 554                                 client_info->ntt = 0;
 555                         }
 556                 }
 557         }
 558
 559         /* do not update the entries again until this counter is zero so that
 560          * not to confuse the clients.
 561          */
 562         bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
 563
 564         _unlock_rx_hashtbl(bond);
 565 }
 566
 567 /* The slave was assigned a new mac address - update the clients */
 568 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
 569 {
 570         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 571         struct rlb_client_info *client_info;
 572         int ntt = 0;
 573         u32 hash_index;
 574
 575         _lock_rx_hashtbl(bond);
 576
 577         hash_index = bond_info->rx_hashtbl_head;
 578         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 579                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 580
 581                 if ((client_info->slave == slave) &&
 582                     compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 583                         client_info->ntt = 1;
 584                         ntt = 1;
 585                 }
 586         }
 587
 588         // update the team's flag only after the whole iteration
 589         if (ntt) {
 590                 bond_info->rx_ntt = 1;
 591                 //fasten the change
 592                 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
 593         }
 594
 595         _unlock_rx_hashtbl(bond);
 596 }
 597
 598 /* mark all clients using src_ip to be updated */
 599 static void rlb_req_update_subnet_clients(struct bonding *bond, __be32 src_ip)
 600 {
 601         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 602         struct rlb_client_info *client_info;
 603         u32 hash_index;
 604
 605         _lock_rx_hashtbl(bond);
 606
 607         hash_index = bond_info->rx_hashtbl_head;
 608         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 609                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 610
 611                 if (!client_info->slave) {
 612                         pr_err("%s: Error: found a client with no channel in the client's hash table\n",
 613                                bond->dev->name);
 614                         continue;
 615                 }
 616                 /*update all clients using this src_ip, that are not assigned
 617                  * to the team's address (curr_active_slave) and have a known
 618                  * unicast mac address.
 619                  */
 620                 if ((client_info->ip_src == src_ip) &&
 621                     compare_ether_addr_64bits(client_info->slave->dev->dev_addr,
 622                            bond->dev->dev_addr) &&
 623                     compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 624                         client_info->ntt = 1;
 625                         bond_info->rx_ntt = 1;
 626                 }
 627         }
 628
 629         _unlock_rx_hashtbl(bond);
 630 }
 631
 632 /* Caller must hold both bond and ptr locks for read */
 633 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
 634 {
 635         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 636         struct arp_pkt *arp = arp_pkt(skb);
 637         struct slave *assigned_slave;
 638         struct rlb_client_info *client_info;
 639         u32 hash_index = 0;
 640
 641         _lock_rx_hashtbl(bond);
 642
 643         hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_src));
 644         client_info = &(bond_info->rx_hashtbl[hash_index]);
 645
 646         if (client_info->assigned) {
 647                 if ((client_info->ip_src == arp->ip_src) &&
 648                     (client_info->ip_dst == arp->ip_dst)) {
 649                         /* the entry is already assigned to this client */
 650                         if (compare_ether_addr_64bits(arp->mac_dst, mac_bcast)) {
 651                                 /* update mac address from arp */
 652                                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 653                         }
 654
 655                         assigned_slave = client_info->slave;
 656                         if (assigned_slave) {
 657                                 _unlock_rx_hashtbl(bond);
 658                                 return assigned_slave;
 659                         }
 660                 } else {
 661                         /* the entry is already assigned to some other client,
 662                          * move the old client to primary (curr_active_slave) so
 663                          * that the new client can be assigned to this entry.
 664                          */
 665                         if (bond->curr_active_slave &&
 666                             client_info->slave != bond->curr_active_slave) {
 667                                 client_info->slave = bond->curr_active_slave;
 668                                 rlb_update_client(client_info);
 669                         }
 670                 }
 671         }
 672         /* assign a new slave */
 673         assigned_slave = rlb_next_rx_slave(bond);
 674
 675         if (assigned_slave) {
 676                 client_info->ip_src = arp->ip_src;
 677                 client_info->ip_dst = arp->ip_dst;
 678                 /* arp->mac_dst is broadcast for arp reqeusts.
 679                  * will be updated with clients actual unicast mac address
 680                  * upon receiving an arp reply.
 681                  */
 682                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 683                 client_info->slave = assigned_slave;
 684
 685                 if (compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 686                         client_info->ntt = 1;
 687                         bond->alb_info.rx_ntt = 1;
 688                 } else {
 689                         client_info->ntt = 0;
 690                 }
 691
 692                 if (!list_empty(&bond->vlan_list)) {
 693                         if (!vlan_get_tag(skb, &client_info->vlan_id))
 694                                 client_info->tag = 1;
 695                 }
 696
 697                 if (!client_info->assigned) {
 698                         u32 prev_tbl_head = bond_info->rx_hashtbl_head;
 699                         bond_info->rx_hashtbl_head = hash_index;
 700                         client_info->next = prev_tbl_head;
 701                         if (prev_tbl_head != RLB_NULL_INDEX) {
 702                                 bond_info->rx_hashtbl[prev_tbl_head].prev =
 703                                         hash_index;
 704                         }
 705                         client_info->assigned = 1;
 706                 }
 707         }
 708
 709         _unlock_rx_hashtbl(bond);
 710
 711         return assigned_slave;
 712 }
 713
 714 /* chooses (and returns) transmit channel for arp reply
 715  * does not choose channel for other arp types since they are
 716  * sent on the curr_active_slave
 717  */
 718 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
 719 {
 720         struct arp_pkt *arp = arp_pkt(skb);
 721         struct slave *tx_slave = NULL;
 722
 723         if (arp->op_code == htons(ARPOP_REPLY)) {
 724                 /* the arp must be sent on the selected
 725                 * rx channel
 726                 */
 727                 tx_slave = rlb_choose_channel(skb, bond);
 728                 if (tx_slave) {
 729                         memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
 730                 }
 731                 pr_debug("Server sent ARP Reply packet\n");
 732         } else if (arp->op_code == htons(ARPOP_REQUEST)) {
 733                 /* Create an entry in the rx_hashtbl for this client as a
 734                  * place holder.
 735                  * When the arp reply is received the entry will be updated
 736                  * with the correct unicast address of the client.
 737                  */
 738                 rlb_choose_channel(skb, bond);
 739
 740                 /* The ARP relpy packets must be delayed so that
 741                  * they can cancel out the influence of the ARP request.
 742                  */
 743                 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
 744
 745                 /* arp requests are broadcast and are sent on the primary
 746                  * the arp request will collapse all clients on the subnet to
 747                  * the primary slave. We must register these clients to be
 748                  * updated with their assigned mac.
 749                  */
 750                 rlb_req_update_subnet_clients(bond, arp->ip_src);
 751                 pr_debug("Server sent ARP Request packet\n");
 752         }
 753
 754         return tx_slave;
 755 }
 756
 757 /* Caller must hold bond lock for read */
 758 static void rlb_rebalance(struct bonding *bond)
 759 {
 760         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 761         struct slave *assigned_slave;
 762         struct rlb_client_info *client_info;
 763         int ntt;
 764         u32 hash_index;
 765
 766         _lock_rx_hashtbl(bond);
 767
 768         ntt = 0;
 769         hash_index = bond_info->rx_hashtbl_head;
 770         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 771                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 772                 assigned_slave = rlb_next_rx_slave(bond);
 773                 if (assigned_slave && (client_info->slave != assigned_slave)) {
 774                         client_info->slave = assigned_slave;
 775                         client_info->ntt = 1;
 776                         ntt = 1;
 777                 }
 778         }
 779
 780         /* update the team's flag only after the whole iteration */
 781         if (ntt) {
 782                 bond_info->rx_ntt = 1;
 783         }
 784         _unlock_rx_hashtbl(bond);
 785 }
 786
 787 /* Caller must hold rx_hashtbl lock */
 788 static void rlb_init_table_entry(struct rlb_client_info *entry)
 789 {
 790         memset(entry, 0, sizeof(struct rlb_client_info));
 791         entry->next = RLB_NULL_INDEX;
 792         entry->prev = RLB_NULL_INDEX;
 793 }
 794
 795 static int rlb_initialize(struct bonding *bond)
 796 {
 797         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 798         struct packet_type *pk_type = &(BOND_ALB_INFO(bond).rlb_pkt_type);
 799         struct rlb_client_info  *new_hashtbl;
 800         int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
 801         int i;
 802
 803         spin_lock_init(&(bond_info->rx_hashtbl_lock));
 804
 805         new_hashtbl = kmalloc(size, GFP_KERNEL);
 806         if (!new_hashtbl) {
 807                 pr_err("%s: Error: Failed to allocate RLB hash table\n",
 808                        bond->dev->name);
 809                 return -1;
 810         }
 811         _lock_rx_hashtbl(bond);
 812
 813         bond_info->rx_hashtbl = new_hashtbl;
 814
 815         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 816
 817         for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
 818                 rlb_init_table_entry(bond_info->rx_hashtbl + i);
 819         }
 820
 821         _unlock_rx_hashtbl(bond);
 822
 823         /*initialize packet type*/
 824         pk_type->type = cpu_to_be16(ETH_P_ARP);
 825         pk_type->dev = NULL;
 826         pk_type->func = rlb_arp_recv;
 827
 828         /* register to receive ARPs */
 829         dev_add_pack(pk_type);
 830
 831         return 0;
 832 }
 833
 834 static void rlb_deinitialize(struct bonding *bond)
 835 {
 836         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 837
 838         dev_remove_pack(&(bond_info->rlb_pkt_type));
 839
 840         _lock_rx_hashtbl(bond);
 841
 842         kfree(bond_info->rx_hashtbl);
 843         bond_info->rx_hashtbl = NULL;
 844         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 845
 846         _unlock_rx_hashtbl(bond);
 847 }
 848
 849 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
 850 {
 851         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 852         u32 curr_index;
 853
 854         _lock_rx_hashtbl(bond);
 855
 856         curr_index = bond_info->rx_hashtbl_head;
 857         while (curr_index != RLB_NULL_INDEX) {
 858                 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
 859                 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
 860                 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
 861
 862                 if (curr->tag && (curr->vlan_id == vlan_id)) {
 863                         if (curr_index == bond_info->rx_hashtbl_head) {
 864                                 bond_info->rx_hashtbl_head = next_index;
 865                         }
 866                         if (prev_index != RLB_NULL_INDEX) {
 867                                 bond_info->rx_hashtbl[prev_index].next = next_index;
 868                         }
 869                         if (next_index != RLB_NULL_INDEX) {
 870                                 bond_info->rx_hashtbl[next_index].prev = prev_index;
 871                         }
 872
 873                         rlb_init_table_entry(curr);
 874                 }
 875
 876                 curr_index = next_index;
 877         }
 878
 879         _unlock_rx_hashtbl(bond);
 880 }
 881
 882 /*********************** tlb/rlb shared functions *********************/
 883
 884 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
 885 {
 886         struct bonding *bond = bond_get_bond_by_slave(slave);
 887         struct learning_pkt pkt;
 888         int size = sizeof(struct learning_pkt);
 889         int i;
 890
 891         memset(&pkt, 0, size);
 892         memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
 893         memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
 894         pkt.type = cpu_to_be16(ETH_P_LOOP);
 895
 896         for (i = 0; i < MAX_LP_BURST; i++) {
 897                 struct sk_buff *skb;
 898                 char *data;
 899
 900                 skb = dev_alloc_skb(size);
 901                 if (!skb) {
 902                         return;
 903                 }
 904
 905                 data = skb_put(skb, size);
 906                 memcpy(data, &pkt, size);
 907
 908                 skb_reset_mac_header(skb);
 909                 skb->network_header = skb->mac_header + ETH_HLEN;
 910                 skb->protocol = pkt.type;
 911                 skb->priority = TC_PRIO_CONTROL;
 912                 skb->dev = slave->dev;
 913
 914                 if (!list_empty(&bond->vlan_list)) {
 915                         struct vlan_entry *vlan;
 916
 917                         vlan = bond_next_vlan(bond,
 918                                               bond->alb_info.current_alb_vlan);
 919
 920                         bond->alb_info.current_alb_vlan = vlan;
 921                         if (!vlan) {
 922                                 kfree_skb(skb);
 923                                 continue;
 924                         }
 925
 926                         skb = vlan_put_tag(skb, vlan->vlan_id);
 927                         if (!skb) {
 928                                 pr_err("%s: Error: failed to insert VLAN tag\n",
 929                                        bond->dev->name);
 930                                 continue;
 931                         }
 932                 }
 933
 934                 dev_queue_xmit(skb);
 935         }
 936 }
 937
 938 /* hw is a boolean parameter that determines whether we should try and
 939  * set the hw address of the device as well as the hw address of the
 940  * net_device
 941  */
 942 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[], int hw)
 943 {
 944         struct net_device *dev = slave->dev;
 945         struct sockaddr s_addr;
 946
 947         if (!hw) {
 948                 memcpy(dev->dev_addr, addr, dev->addr_len);
 949                 return 0;
 950         }
 951
 952         /* for rlb each slave must have a unique hw mac addresses so that */
 953         /* each slave will receive packets destined to a different mac */
 954         memcpy(s_addr.sa_data, addr, dev->addr_len);
 955         s_addr.sa_family = dev->type;
 956         if (dev_set_mac_address(dev, &s_addr)) {
 957                 pr_err("%s: Error: dev_set_mac_address of dev %s failed!\n"
 958                        "ALB mode requires that the base driver support setting the hw address also when the network device's interface is open\n",
 959                        dev->master->name, dev->name);
 960                 return -EOPNOTSUPP;
 961         }
 962         return 0;
 963 }
 964
 965 /*
 966  * Swap MAC addresses between two slaves.
 967  *
 968  * Called with RTNL held, and no other locks.
 969  *
 970  */
 971
 972 static void alb_swap_mac_addr(struct bonding *bond, struct slave *slave1, struct slave *slave2)
 973 {
 974         u8 tmp_mac_addr[ETH_ALEN];
 975
 976         memcpy(tmp_mac_addr, slave1->dev->dev_addr, ETH_ALEN);
 977         alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr, bond->alb_info.rlb_enabled);
 978         alb_set_slave_mac_addr(slave2, tmp_mac_addr, bond->alb_info.rlb_enabled);
 979
 980 }
 981
 982 /*
 983  * Send learning packets after MAC address swap.
 984  *
 985  * Called with RTNL and no other locks
 986  */
 987 static void alb_fasten_mac_swap(struct bonding *bond, struct slave *slave1,
 988                                 struct slave *slave2)
 989 {
 990         int slaves_state_differ = (SLAVE_IS_OK(slave1) != SLAVE_IS_OK(slave2));
 991         struct slave *disabled_slave = NULL;
 992
 993         ASSERT_RTNL();
 994
 995         /* fasten the change in the switch */
 996         if (SLAVE_IS_OK(slave1)) {
 997                 alb_send_learning_packets(slave1, slave1->dev->dev_addr);
 998                 if (bond->alb_info.rlb_enabled) {
 999                         /* inform the clients that the mac address
1000                          * has changed
1001                          */
1002                         rlb_req_update_slave_clients(bond, slave1);
1003                 }
1004         } else {
1005                 disabled_slave = slave1;
1006         }
1007
1008         if (SLAVE_IS_OK(slave2)) {
1009                 alb_send_learning_packets(slave2, slave2->dev->dev_addr);
1010                 if (bond->alb_info.rlb_enabled) {
1011                         /* inform the clients that the mac address
1012                          * has changed
1013                          */
1014                         rlb_req_update_slave_clients(bond, slave2);
1015                 }
1016         } else {
1017                 disabled_slave = slave2;
1018         }
1019
1020         if (bond->alb_info.rlb_enabled && slaves_state_differ) {
1021                 /* A disabled slave was assigned an active mac addr */
1022                 rlb_teach_disabled_mac_on_primary(bond,
1023                                                   disabled_slave->dev->dev_addr);
1024         }
1025 }
1026
1027 /**
1028  * alb_change_hw_addr_on_detach
1029  * @bond: bonding we're working on
1030  * @slave: the slave that was just detached
1031  *
1032  * We assume that @slave was already detached from the slave list.
1033  *
1034  * If @slave's permanent hw address is different both from its current
1035  * address and from @bond's address, then somewhere in the bond there's
1036  * a slave that has @slave's permanet address as its current address.
1037  * We'll make sure that that slave no longer uses @slave's permanent address.
1038  *
1039  * Caller must hold RTNL and no other locks
1040  */
1041 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
1042 {
1043         int perm_curr_diff;
1044         int perm_bond_diff;
1045
1046         perm_curr_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
1047                                                    slave->dev->dev_addr);
1048         perm_bond_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
1049                                                    bond->dev->dev_addr);
1050
1051         if (perm_curr_diff && perm_bond_diff) {
1052                 struct slave *tmp_slave;
1053                 int i, found = 0;
1054
1055                 bond_for_each_slave(bond, tmp_slave, i) {
1056                         if (!compare_ether_addr_64bits(slave->perm_hwaddr,
1057                                                        tmp_slave->dev->dev_addr)) {
1058                                 found = 1;
1059                                 break;
1060                         }
1061                 }
1062
1063                 if (found) {
1064                         /* locking: needs RTNL and nothing else */
1065                         alb_swap_mac_addr(bond, slave, tmp_slave);
1066                         alb_fasten_mac_swap(bond, slave, tmp_slave);
1067                 }
1068         }
1069 }
1070
1071 /**
1072  * alb_handle_addr_collision_on_attach
1073  * @bond: bonding we're working on
1074  * @slave: the slave that was just attached
1075  *
1076  * checks uniqueness of slave's mac address and handles the case the
1077  * new slave uses the bonds mac address.
1078  *
1079  * If the permanent hw address of @slave is @bond's hw address, we need to
1080  * find a different hw address to give @slave, that isn't in use by any other
1081  * slave in the bond. This address must be, of course, one of the premanent
1082  * addresses of the other slaves.
1083  *
1084  * We go over the slave list, and for each slave there we compare its
1085  * permanent hw address with the current address of all the other slaves.
1086  * If no match was found, then we've found a slave with a permanent address
1087  * that isn't used by any other slave in the bond, so we can assign it to
1088  * @slave.
1089  *
1090  * assumption: this function is called before @slave is attached to the
1091  *             bond slave list.
1092  *
1093  * caller must hold the bond lock for write since the mac addresses are compared
1094  * and may be swapped.
1095  */
1096 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1097 {
1098         struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1099         struct slave *has_bond_addr = bond->curr_active_slave;
1100         int i, j, found = 0;
1101
1102         if (bond->slave_cnt == 0) {
1103                 /* this is the first slave */
1104                 return 0;
1105         }
1106
1107         /* if slave's mac address differs from bond's mac address
1108          * check uniqueness of slave's mac address against the other
1109          * slaves in the bond.
1110          */
1111         if (compare_ether_addr_64bits(slave->perm_hwaddr, bond->dev->dev_addr)) {
1112                 bond_for_each_slave(bond, tmp_slave1, i) {
1113                         if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1114                                                        slave->dev->dev_addr)) {
1115                                 found = 1;
1116                                 break;
1117                         }
1118                 }
1119
1120                 if (!found)
1121                         return 0;
1122
1123                 /* Try setting slave mac to bond address and fall-through
1124                    to code handling that situation below... */
1125                 alb_set_slave_mac_addr(slave, bond->dev->dev_addr,
1126                                        bond->alb_info.rlb_enabled);
1127         }
1128
1129         /* The slave's address is equal to the address of the bond.
1130          * Search for a spare address in the bond for this slave.
1131          */
1132         free_mac_slave = NULL;
1133
1134         bond_for_each_slave(bond, tmp_slave1, i) {
1135                 found = 0;
1136                 bond_for_each_slave(bond, tmp_slave2, j) {
1137                         if (!compare_ether_addr_64bits(tmp_slave1->perm_hwaddr,
1138                                                        tmp_slave2->dev->dev_addr)) {
1139                                 found = 1;
1140                                 break;
1141                         }
1142                 }
1143
1144                 if (!found) {
1145                         /* no slave has tmp_slave1's perm addr
1146                          * as its curr addr
1147                          */
1148                         free_mac_slave = tmp_slave1;
1149                         break;
1150                 }
1151
1152                 if (!has_bond_addr) {
1153                         if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1154                                                        bond->dev->dev_addr)) {
1155
1156                                 has_bond_addr = tmp_slave1;
1157                         }
1158                 }
1159         }
1160
1161         if (free_mac_slave) {
1162                 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr,
1163                                        bond->alb_info.rlb_enabled);
1164
1165                 pr_warning("%s: Warning: the hw address of slave %s is in use by the bond; giving it the hw address of %s\n",
1166                            bond->dev->name, slave->dev->name,
1167                            free_mac_slave->dev->name);
1168
1169         } else if (has_bond_addr) {
1170                 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",
1171                        bond->dev->name, slave->dev->name);
1172                 return -EFAULT;
1173         }
1174
1175         return 0;
1176 }
1177
1178 /**
1179  * alb_set_mac_address
1180  * @bond:
1181  * @addr:
1182  *
1183  * In TLB mode all slaves are configured to the bond's hw address, but set
1184  * their dev_addr field to different addresses (based on their permanent hw
1185  * addresses).
1186  *
1187  * For each slave, this function sets the interface to the new address and then
1188  * changes its dev_addr field to its previous value.
1189  *
1190  * Unwinding assumes bond's mac address has not yet changed.
1191  */
1192 static int alb_set_mac_address(struct bonding *bond, void *addr)
1193 {
1194         struct sockaddr sa;
1195         struct slave *slave, *stop_at;
1196         char tmp_addr[ETH_ALEN];
1197         int res;
1198         int i;
1199
1200         if (bond->alb_info.rlb_enabled) {
1201                 return 0;
1202         }
1203
1204         bond_for_each_slave(bond, slave, i) {
1205                 /* save net_device's current hw address */
1206                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1207
1208                 res = dev_set_mac_address(slave->dev, addr);
1209
1210                 /* restore net_device's hw address */
1211                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1212
1213                 if (res)
1214                         goto unwind;
1215         }
1216
1217         return 0;
1218
1219 unwind:
1220         memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1221         sa.sa_family = bond->dev->type;
1222
1223         /* unwind from head to the slave that failed */
1224         stop_at = slave;
1225         bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1226                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1227                 dev_set_mac_address(slave->dev, &sa);
1228                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1229         }
1230
1231         return res;
1232 }
1233
1234 /************************ exported alb funcions ************************/
1235
1236 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1237 {
1238         int res;
1239
1240         res = tlb_initialize(bond);
1241         if (res) {
1242                 return res;
1243         }
1244
1245         if (rlb_enabled) {
1246                 bond->alb_info.rlb_enabled = 1;
1247                 /* initialize rlb */
1248                 res = rlb_initialize(bond);
1249                 if (res) {
1250                         tlb_deinitialize(bond);
1251                         return res;
1252                 }
1253         } else {
1254                 bond->alb_info.rlb_enabled = 0;
1255         }
1256
1257         return 0;
1258 }
1259
1260 void bond_alb_deinitialize(struct bonding *bond)
1261 {
1262         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1263
1264         tlb_deinitialize(bond);
1265
1266         if (bond_info->rlb_enabled) {
1267                 rlb_deinitialize(bond);
1268         }
1269 }
1270
1271 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1272 {
1273         struct bonding *bond = netdev_priv(bond_dev);
1274         struct ethhdr *eth_data;
1275         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1276         struct slave *tx_slave = NULL;
1277         static const __be32 ip_bcast = htonl(0xffffffff);
1278         int hash_size = 0;
1279         int do_tx_balance = 1;
1280         u32 hash_index = 0;
1281         const u8 *hash_start = NULL;
1282         int res = 1;
1283         struct ipv6hdr *ip6hdr;
1284
1285         skb_reset_mac_header(skb);
1286         eth_data = eth_hdr(skb);
1287
1288         /* make sure that the curr_active_slave and the slaves list do
1289          * not change during tx
1290          */
1291         read_lock(&bond->lock);
1292         read_lock(&bond->curr_slave_lock);
1293
1294         if (!BOND_IS_OK(bond)) {
1295                 goto out;
1296         }
1297
1298         switch (ntohs(skb->protocol)) {
1299         case ETH_P_IP: {
1300                 const struct iphdr *iph = ip_hdr(skb);
1301
1302                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast) ||
1303                     (iph->daddr == ip_bcast) ||
1304                     (iph->protocol == IPPROTO_IGMP)) {
1305                         do_tx_balance = 0;
1306                         break;
1307                 }
1308                 hash_start = (char *)&(iph->daddr);
1309                 hash_size = sizeof(iph->daddr);
1310         }
1311                 break;
1312         case ETH_P_IPV6:
1313                 /* IPv6 doesn't really use broadcast mac address, but leave
1314                  * that here just in case.
1315                  */
1316                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast)) {
1317                         do_tx_balance = 0;
1318                         break;
1319                 }
1320
1321                 /* IPv6 uses all-nodes multicast as an equivalent to
1322                  * broadcasts in IPv4.
1323                  */
1324                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_v6_allmcast)) {
1325                         do_tx_balance = 0;
1326                         break;
1327                 }
1328
1329                 /* Additianally, DAD probes should not be tx-balanced as that
1330                  * will lead to false positives for duplicate addresses and
1331                  * prevent address configuration from working.
1332                  */
1333                 ip6hdr = ipv6_hdr(skb);
1334                 if (ipv6_addr_any(&ip6hdr->saddr)) {
1335                         do_tx_balance = 0;
1336                         break;
1337                 }
1338
1339                 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1340                 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1341                 break;
1342         case ETH_P_IPX:
1343                 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1344                         /* something is wrong with this packet */
1345                         do_tx_balance = 0;
1346                         break;
1347                 }
1348
1349                 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1350                         /* The only protocol worth balancing in
1351                          * this family since it has an "ARP" like
1352                          * mechanism
1353                          */
1354                         do_tx_balance = 0;
1355                         break;
1356                 }
1357
1358                 hash_start = (char*)eth_data->h_dest;
1359                 hash_size = ETH_ALEN;
1360                 break;
1361         case ETH_P_ARP:
1362                 do_tx_balance = 0;
1363                 if (bond_info->rlb_enabled) {
1364                         tx_slave = rlb_arp_xmit(skb, bond);
1365                 }
1366                 break;
1367         default:
1368                 do_tx_balance = 0;
1369                 break;
1370         }
1371
1372         if (do_tx_balance) {
1373                 hash_index = _simple_hash(hash_start, hash_size);
1374                 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1375         }
1376
1377         if (!tx_slave) {
1378                 /* unbalanced or unassigned, send through primary */
1379                 tx_slave = bond->curr_active_slave;
1380                 bond_info->unbalanced_load += skb->len;
1381         }
1382
1383         if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1384                 if (tx_slave != bond->curr_active_slave) {
1385                         memcpy(eth_data->h_source,
1386                                tx_slave->dev->dev_addr,
1387                                ETH_ALEN);
1388                 }
1389
1390                 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1391         } else {
1392                 if (tx_slave) {
1393                         tlb_clear_slave(bond, tx_slave, 0);
1394                 }
1395         }
1396
1397 out:
1398         if (res) {
1399                 /* no suitable interface, frame not sent */
1400                 dev_kfree_skb(skb);
1401         }
1402         read_unlock(&bond->curr_slave_lock);
1403         read_unlock(&bond->lock);
1404         return NETDEV_TX_OK;
1405 }
1406
1407 void bond_alb_monitor(struct work_struct *work)
1408 {
1409         struct bonding *bond = container_of(work, struct bonding,
1410                                             alb_work.work);
1411         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1412         struct slave *slave;
1413         int i;
1414
1415         read_lock(&bond->lock);
1416
1417         if (bond->kill_timers) {
1418                 goto out;
1419         }
1420
1421         if (bond->slave_cnt == 0) {
1422                 bond_info->tx_rebalance_counter = 0;
1423                 bond_info->lp_counter = 0;
1424                 goto re_arm;
1425         }
1426
1427         bond_info->tx_rebalance_counter++;
1428         bond_info->lp_counter++;
1429
1430         /* send learning packets */
1431         if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1432                 /* change of curr_active_slave involves swapping of mac addresses.
1433                  * in order to avoid this swapping from happening while
1434                  * sending the learning packets, the curr_slave_lock must be held for
1435                  * read.
1436                  */
1437                 read_lock(&bond->curr_slave_lock);
1438
1439                 bond_for_each_slave(bond, slave, i) {
1440                         alb_send_learning_packets(slave, slave->dev->dev_addr);
1441                 }
1442
1443                 read_unlock(&bond->curr_slave_lock);
1444
1445                 bond_info->lp_counter = 0;
1446         }
1447
1448         /* rebalance tx traffic */
1449         if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1450
1451                 read_lock(&bond->curr_slave_lock);
1452
1453                 bond_for_each_slave(bond, slave, i) {
1454                         tlb_clear_slave(bond, slave, 1);
1455                         if (slave == bond->curr_active_slave) {
1456                                 SLAVE_TLB_INFO(slave).load =
1457                                         bond_info->unbalanced_load /
1458                                                 BOND_TLB_REBALANCE_INTERVAL;
1459                                 bond_info->unbalanced_load = 0;
1460                         }
1461                 }
1462
1463                 read_unlock(&bond->curr_slave_lock);
1464
1465                 bond_info->tx_rebalance_counter = 0;
1466         }
1467
1468         /* handle rlb stuff */
1469         if (bond_info->rlb_enabled) {
1470                 if (bond_info->primary_is_promisc &&
1471                     (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1472
1473                         /*
1474                          * dev_set_promiscuity requires rtnl and
1475                          * nothing else.
1476                          */
1477                         read_unlock(&bond->lock);
1478                         rtnl_lock();
1479
1480                         bond_info->rlb_promisc_timeout_counter = 0;
1481
1482                         /* If the primary was set to promiscuous mode
1483                          * because a slave was disabled then
1484                          * it can now leave promiscuous mode.
1485                          */
1486                         dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1487                         bond_info->primary_is_promisc = 0;
1488
1489                         rtnl_unlock();
1490                         read_lock(&bond->lock);
1491                 }
1492
1493                 if (bond_info->rlb_rebalance) {
1494                         bond_info->rlb_rebalance = 0;
1495                         rlb_rebalance(bond);
1496                 }
1497
1498                 /* check if clients need updating */
1499                 if (bond_info->rx_ntt) {
1500                         if (bond_info->rlb_update_delay_counter) {
1501                                 --bond_info->rlb_update_delay_counter;
1502                         } else {
1503                                 rlb_update_rx_clients(bond);
1504                                 if (bond_info->rlb_update_retry_counter) {
1505                                         --bond_info->rlb_update_retry_counter;
1506                                 } else {
1507                                         bond_info->rx_ntt = 0;
1508                                 }
1509                         }
1510                 }
1511         }
1512
1513 re_arm:
1514         queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1515 out:
1516         read_unlock(&bond->lock);
1517 }
1518
1519 /* assumption: called before the slave is attached to the bond
1520  * and not locked by the bond lock
1521  */
1522 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1523 {
1524         int res;
1525
1526         res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr,
1527                                      bond->alb_info.rlb_enabled);
1528         if (res) {
1529                 return res;
1530         }
1531
1532         /* caller must hold the bond lock for write since the mac addresses
1533          * are compared and may be swapped.
1534          */
1535         read_lock(&bond->lock);
1536
1537         res = alb_handle_addr_collision_on_attach(bond, slave);
1538
1539         read_unlock(&bond->lock);
1540
1541         if (res) {
1542                 return res;
1543         }
1544
1545         tlb_init_slave(slave);
1546
1547         /* order a rebalance ASAP */
1548         bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1549
1550         if (bond->alb_info.rlb_enabled) {
1551                 bond->alb_info.rlb_rebalance = 1;
1552         }
1553
1554         return 0;
1555 }
1556
1557 /*
1558  * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1559  * if necessary.
1560  *
1561  * Caller must hold RTNL and no other locks
1562  */
1563 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1564 {
1565         if (bond->slave_cnt > 1) {
1566                 alb_change_hw_addr_on_detach(bond, slave);
1567         }
1568
1569         tlb_clear_slave(bond, slave, 0);
1570
1571         if (bond->alb_info.rlb_enabled) {
1572                 bond->alb_info.next_rx_slave = NULL;
1573                 rlb_clear_slave(bond, slave);
1574         }
1575 }
1576
1577 /* Caller must hold bond lock for read */
1578 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1579 {
1580         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1581
1582         if (link == BOND_LINK_DOWN) {
1583                 tlb_clear_slave(bond, slave, 0);
1584                 if (bond->alb_info.rlb_enabled) {
1585                         rlb_clear_slave(bond, slave);
1586                 }
1587         } else if (link == BOND_LINK_UP) {
1588                 /* order a rebalance ASAP */
1589                 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1590                 if (bond->alb_info.rlb_enabled) {
1591                         bond->alb_info.rlb_rebalance = 1;
1592                         /* If the updelay module parameter is smaller than the
1593                          * forwarding delay of the switch the rebalance will
1594                          * not work because the rebalance arp replies will
1595                          * not be forwarded to the clients..
1596                          */
1597                 }
1598         }
1599 }
1600
1601 /**
1602  * bond_alb_handle_active_change - assign new curr_active_slave
1603  * @bond: our bonding struct
1604  * @new_slave: new slave to assign
1605  *
1606  * Set the bond->curr_active_slave to @new_slave and handle
1607  * mac address swapping and promiscuity changes as needed.
1608  *
1609  * If new_slave is NULL, caller must hold curr_slave_lock or
1610  * bond->lock for write.
1611  *
1612  * If new_slave is not NULL, caller must hold RTNL, bond->lock for
1613  * read and curr_slave_lock for write.  Processing here may sleep, so
1614  * no other locks may be held.
1615  */
1616 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1617         __releases(&bond->curr_slave_lock)
1618         __releases(&bond->lock)
1619         __acquires(&bond->lock)
1620         __acquires(&bond->curr_slave_lock)
1621 {
1622         struct slave *swap_slave;
1623         int i;
1624
1625         if (bond->curr_active_slave == new_slave) {
1626                 return;
1627         }
1628
1629         if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1630                 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1631                 bond->alb_info.primary_is_promisc = 0;
1632                 bond->alb_info.rlb_promisc_timeout_counter = 0;
1633         }
1634
1635         swap_slave = bond->curr_active_slave;
1636         bond->curr_active_slave = new_slave;
1637
1638         if (!new_slave || (bond->slave_cnt == 0)) {
1639                 return;
1640         }
1641
1642         /* set the new curr_active_slave to the bonds mac address
1643          * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1644          */
1645         if (!swap_slave) {
1646                 struct slave *tmp_slave;
1647                 /* find slave that is holding the bond's mac address */
1648                 bond_for_each_slave(bond, tmp_slave, i) {
1649                         if (!compare_ether_addr_64bits(tmp_slave->dev->dev_addr,
1650                                                        bond->dev->dev_addr)) {
1651                                 swap_slave = tmp_slave;
1652                                 break;
1653                         }
1654                 }
1655         }
1656
1657         /*
1658          * Arrange for swap_slave and new_slave to temporarily be
1659          * ignored so we can mess with their MAC addresses without
1660          * fear of interference from transmit activity.
1661          */
1662         if (swap_slave) {
1663                 tlb_clear_slave(bond, swap_slave, 1);
1664         }
1665         tlb_clear_slave(bond, new_slave, 1);
1666
1667         write_unlock_bh(&bond->curr_slave_lock);
1668         read_unlock(&bond->lock);
1669
1670         ASSERT_RTNL();
1671
1672         /* curr_active_slave must be set before calling alb_swap_mac_addr */
1673         if (swap_slave) {
1674                 /* swap mac address */
1675                 alb_swap_mac_addr(bond, swap_slave, new_slave);
1676         } else {
1677                 /* set the new_slave to the bond mac address */
1678                 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr,
1679                                        bond->alb_info.rlb_enabled);
1680         }
1681
1682         if (swap_slave) {
1683                 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1684                 read_lock(&bond->lock);
1685         } else {
1686                 read_lock(&bond->lock);
1687                 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1688         }
1689
1690         write_lock_bh(&bond->curr_slave_lock);
1691 }
1692
1693 /*
1694  * Called with RTNL
1695  */
1696 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1697         __acquires(&bond->lock)
1698         __releases(&bond->lock)
1699 {
1700         struct bonding *bond = netdev_priv(bond_dev);
1701         struct sockaddr *sa = addr;
1702         struct slave *slave, *swap_slave;
1703         int res;
1704         int i;
1705
1706         if (!is_valid_ether_addr(sa->sa_data)) {
1707                 return -EADDRNOTAVAIL;
1708         }
1709
1710         res = alb_set_mac_address(bond, addr);
1711         if (res) {
1712                 return res;
1713         }
1714
1715         memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1716
1717         /* If there is no curr_active_slave there is nothing else to do.
1718          * Otherwise we'll need to pass the new address to it and handle
1719          * duplications.
1720          */
1721         if (!bond->curr_active_slave) {
1722                 return 0;
1723         }
1724
1725         swap_slave = NULL;
1726
1727         bond_for_each_slave(bond, slave, i) {
1728                 if (!compare_ether_addr_64bits(slave->dev->dev_addr,
1729                                                bond_dev->dev_addr)) {
1730                         swap_slave = slave;
1731                         break;
1732                 }
1733         }
1734
1735         if (swap_slave) {
1736                 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1737                 alb_fasten_mac_swap(bond, swap_slave, bond->curr_active_slave);
1738         } else {
1739                 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr,
1740                                        bond->alb_info.rlb_enabled);
1741
1742                 read_lock(&bond->lock);
1743                 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1744                 if (bond->alb_info.rlb_enabled) {
1745                         /* inform clients mac address has changed */
1746                         rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1747                 }
1748                 read_unlock(&bond->lock);
1749         }
1750
1751         return 0;
1752 }
1753
1754 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1755 {
1756         if (bond->alb_info.current_alb_vlan &&
1757             (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1758                 bond->alb_info.current_alb_vlan = NULL;
1759         }
1760
1761         if (bond->alb_info.rlb_enabled) {
1762                 rlb_clear_vlan(bond, vlan_id);
1763         }
1764 }
1765