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 BONDING_DEBUG 1
  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 static const u8 mac_bcast[ETH_ALEN] = {0xff,0xff,0xff,0xff,0xff,0xff};
  85 static const u8 mac_v6_allmcast[ETH_ALEN] = {0x33,0x33,0x00,0x00,0x00,0x01};
  86 static const int alb_delta_in_ticks = HZ / ALB_TIMER_TICKS_PER_SEC;
  87
  88 #pragma pack(1)
  89 struct learning_pkt {
  90         u8 mac_dst[ETH_ALEN];
  91         u8 mac_src[ETH_ALEN];
  92         __be16 type;
  93         u8 padding[ETH_ZLEN - ETH_HLEN];
  94 };
  95
  96 struct arp_pkt {
  97         __be16  hw_addr_space;
  98         __be16  prot_addr_space;
  99         u8      hw_addr_len;
 100         u8      prot_addr_len;
 101         __be16  op_code;
 102         u8      mac_src[ETH_ALEN];      /* sender hardware address */
 103         __be32  ip_src;                 /* sender IP address */
 104         u8      mac_dst[ETH_ALEN];      /* target hardware address */
 105         __be32  ip_dst;                 /* target IP address */
 106 };
 107 #pragma pack()
 108
 109 static inline struct arp_pkt *arp_pkt(const struct sk_buff *skb)
 110 {
 111         return (struct arp_pkt *)skb_network_header(skb);
 112 }
 113
 114 /* Forward declaration */
 115 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[]);
 116
 117 static inline u8 _simple_hash(const u8 *hash_start, int hash_size)
 118 {
 119         int i;
 120         u8 hash = 0;
 121
 122         for (i = 0; i < hash_size; i++) {
 123                 hash ^= hash_start[i];
 124         }
 125
 126         return hash;
 127 }
 128
 129 /*********************** tlb specific functions ***************************/
 130
 131 static inline void _lock_tx_hashtbl(struct bonding *bond)
 132 {
 133         spin_lock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 134 }
 135
 136 static inline void _unlock_tx_hashtbl(struct bonding *bond)
 137 {
 138         spin_unlock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 139 }
 140
 141 /* Caller must hold tx_hashtbl lock */
 142 static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
 143 {
 144         if (save_load) {
 145                 entry->load_history = 1 + entry->tx_bytes /
 146                                       BOND_TLB_REBALANCE_INTERVAL;
 147                 entry->tx_bytes = 0;
 148         }
 149
 150         entry->tx_slave = NULL;
 151         entry->next = TLB_NULL_INDEX;
 152         entry->prev = TLB_NULL_INDEX;
 153 }
 154
 155 static inline void tlb_init_slave(struct slave *slave)
 156 {
 157         SLAVE_TLB_INFO(slave).load = 0;
 158         SLAVE_TLB_INFO(slave).head = TLB_NULL_INDEX;
 159 }
 160
 161 /* Caller must hold bond lock for read */
 162 static void tlb_clear_slave(struct bonding *bond, struct slave *slave, int save_load)
 163 {
 164         struct tlb_client_info *tx_hash_table;
 165         u32 index;
 166
 167         _lock_tx_hashtbl(bond);
 168
 169         /* clear slave from tx_hashtbl */
 170         tx_hash_table = BOND_ALB_INFO(bond).tx_hashtbl;
 171
 172         /* skip this if we've already freed the tx hash table */
 173         if (tx_hash_table) {
 174                 index = SLAVE_TLB_INFO(slave).head;
 175                 while (index != TLB_NULL_INDEX) {
 176                         u32 next_index = tx_hash_table[index].next;
 177                         tlb_init_table_entry(&tx_hash_table[index], save_load);
 178                         index = next_index;
 179                 }
 180         }
 181
 182         tlb_init_slave(slave);
 183
 184         _unlock_tx_hashtbl(bond);
 185 }
 186
 187 /* Must be called before starting the monitor timer */
 188 static int tlb_initialize(struct bonding *bond)
 189 {
 190         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 191         int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
 192         struct tlb_client_info *new_hashtbl;
 193         int i;
 194
 195         spin_lock_init(&(bond_info->tx_hashtbl_lock));
 196
 197         new_hashtbl = kzalloc(size, GFP_KERNEL);
 198         if (!new_hashtbl) {
 199                 printk(KERN_ERR DRV_NAME
 200                        ": %s: Error: Failed to allocate TLB hash table\n",
 201                        bond->dev->name);
 202                 return -1;
 203         }
 204         _lock_tx_hashtbl(bond);
 205
 206         bond_info->tx_hashtbl = new_hashtbl;
 207
 208         for (i = 0; i < TLB_HASH_TABLE_SIZE; i++) {
 209                 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 1);
 210         }
 211
 212         _unlock_tx_hashtbl(bond);
 213
 214         return 0;
 215 }
 216
 217 /* Must be called only after all slaves have been released */
 218 static void tlb_deinitialize(struct bonding *bond)
 219 {
 220         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 221
 222         _lock_tx_hashtbl(bond);
 223
 224         kfree(bond_info->tx_hashtbl);
 225         bond_info->tx_hashtbl = NULL;
 226
 227         _unlock_tx_hashtbl(bond);
 228 }
 229
 230 /* Caller must hold bond lock for read */
 231 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
 232 {
 233         struct slave *slave, *least_loaded;
 234         s64 max_gap;
 235         int i, found = 0;
 236
 237         /* Find the first enabled slave */
 238         bond_for_each_slave(bond, slave, i) {
 239                 if (SLAVE_IS_OK(slave)) {
 240                         found = 1;
 241                         break;
 242                 }
 243         }
 244
 245         if (!found) {
 246                 return NULL;
 247         }
 248
 249         least_loaded = slave;
 250         max_gap = (s64)(slave->speed << 20) - /* Convert to Megabit per sec */
 251                         (s64)(SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
 252
 253         /* Find the slave with the largest gap */
 254         bond_for_each_slave_from(bond, slave, i, least_loaded) {
 255                 if (SLAVE_IS_OK(slave)) {
 256                         s64 gap = (s64)(slave->speed << 20) -
 257                                         (s64)(SLAVE_TLB_INFO(slave).load << 3);
 258                         if (max_gap < gap) {
 259                                 least_loaded = slave;
 260                                 max_gap = gap;
 261                         }
 262                 }
 263         }
 264
 265         return least_loaded;
 266 }
 267
 268 /* Caller must hold bond lock for read */
 269 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index, u32 skb_len)
 270 {
 271         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 272         struct tlb_client_info *hash_table;
 273         struct slave *assigned_slave;
 274
 275         _lock_tx_hashtbl(bond);
 276
 277         hash_table = bond_info->tx_hashtbl;
 278         assigned_slave = hash_table[hash_index].tx_slave;
 279         if (!assigned_slave) {
 280                 assigned_slave = tlb_get_least_loaded_slave(bond);
 281
 282                 if (assigned_slave) {
 283                         struct tlb_slave_info *slave_info =
 284                                 &(SLAVE_TLB_INFO(assigned_slave));
 285                         u32 next_index = slave_info->head;
 286
 287                         hash_table[hash_index].tx_slave = assigned_slave;
 288                         hash_table[hash_index].next = next_index;
 289                         hash_table[hash_index].prev = TLB_NULL_INDEX;
 290
 291                         if (next_index != TLB_NULL_INDEX) {
 292                                 hash_table[next_index].prev = hash_index;
 293                         }
 294
 295                         slave_info->head = hash_index;
 296                         slave_info->load +=
 297                                 hash_table[hash_index].load_history;
 298                 }
 299         }
 300
 301         if (assigned_slave) {
 302                 hash_table[hash_index].tx_bytes += skb_len;
 303         }
 304
 305         _unlock_tx_hashtbl(bond);
 306
 307         return assigned_slave;
 308 }
 309
 310 /*********************** rlb specific functions ***************************/
 311 static inline void _lock_rx_hashtbl(struct bonding *bond)
 312 {
 313         spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 314 }
 315
 316 static inline void _unlock_rx_hashtbl(struct bonding *bond)
 317 {
 318         spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 319 }
 320
 321 /* when an ARP REPLY is received from a client update its info
 322  * in the rx_hashtbl
 323  */
 324 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
 325 {
 326         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 327         struct rlb_client_info *client_info;
 328         u32 hash_index;
 329
 330         _lock_rx_hashtbl(bond);
 331
 332         hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
 333         client_info = &(bond_info->rx_hashtbl[hash_index]);
 334
 335         if ((client_info->assigned) &&
 336             (client_info->ip_src == arp->ip_dst) &&
 337             (client_info->ip_dst == arp->ip_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 = bond_dev->priv;
 350         struct arp_pkt *arp = (struct arp_pkt *)skb->data;
 351         int res = NET_RX_DROP;
 352
 353         if (dev_net(bond_dev) != &init_net)
 354                 goto out;
 355
 356         if (!(bond_dev->flags & IFF_MASTER))
 357                 goto out;
 358
 359         if (!arp) {
 360                 dprintk("Packet has no ARP data\n");
 361                 goto out;
 362         }
 363
 364         if (skb->len < sizeof(struct arp_pkt)) {
 365                 dprintk("Packet is too small to be an ARP\n");
 366                 goto out;
 367         }
 368
 369         if (arp->op_code == htons(ARPOP_REPLY)) {
 370                 /* update rx hash table for this ARP */
 371                 rlb_update_entry_from_arp(bond, arp);
 372                 dprintk("Server received an ARP Reply from client\n");
 373         }
 374
 375         res = NET_RX_SUCCESS;
 376
 377 out:
 378         dev_kfree_skb(skb);
 379
 380         return res;
 381 }
 382
 383 /* Caller must hold bond lock for read */
 384 static struct slave *rlb_next_rx_slave(struct bonding *bond)
 385 {
 386         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 387         struct slave *rx_slave, *slave, *start_at;
 388         int i = 0;
 389
 390         if (bond_info->next_rx_slave) {
 391                 start_at = bond_info->next_rx_slave;
 392         } else {
 393                 start_at = bond->first_slave;
 394         }
 395
 396         rx_slave = NULL;
 397
 398         bond_for_each_slave_from(bond, slave, i, start_at) {
 399                 if (SLAVE_IS_OK(slave)) {
 400                         if (!rx_slave) {
 401                                 rx_slave = slave;
 402                         } else if (slave->speed > rx_slave->speed) {
 403                                 rx_slave = slave;
 404                         }
 405                 }
 406         }
 407
 408         if (rx_slave) {
 409                 bond_info->next_rx_slave = rx_slave->next;
 410         }
 411
 412         return rx_slave;
 413 }
 414
 415 /* teach the switch the mac of a disabled slave
 416  * on the primary for fault tolerance
 417  *
 418  * Caller must hold bond->curr_slave_lock for write or bond lock for write
 419  */
 420 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
 421 {
 422         if (!bond->curr_active_slave) {
 423                 return;
 424         }
 425
 426         if (!bond->alb_info.primary_is_promisc) {
 427                 if (!dev_set_promiscuity(bond->curr_active_slave->dev, 1))
 428                         bond->alb_info.primary_is_promisc = 1;
 429                 else
 430                         bond->alb_info.primary_is_promisc = 0;
 431         }
 432
 433         bond->alb_info.rlb_promisc_timeout_counter = 0;
 434
 435         alb_send_learning_packets(bond->curr_active_slave, addr);
 436 }
 437
 438 /* slave being removed should not be active at this point
 439  *
 440  * Caller must hold bond lock for read
 441  */
 442 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
 443 {
 444         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 445         struct rlb_client_info *rx_hash_table;
 446         u32 index, next_index;
 447
 448         /* clear slave from rx_hashtbl */
 449         _lock_rx_hashtbl(bond);
 450
 451         rx_hash_table = bond_info->rx_hashtbl;
 452         index = bond_info->rx_hashtbl_head;
 453         for (; index != RLB_NULL_INDEX; index = next_index) {
 454                 next_index = rx_hash_table[index].next;
 455                 if (rx_hash_table[index].slave == slave) {
 456                         struct slave *assigned_slave = rlb_next_rx_slave(bond);
 457
 458                         if (assigned_slave) {
 459                                 rx_hash_table[index].slave = assigned_slave;
 460                                 if (memcmp(rx_hash_table[index].mac_dst,
 461                                            mac_bcast, ETH_ALEN)) {
 462                                         bond_info->rx_hashtbl[index].ntt = 1;
 463                                         bond_info->rx_ntt = 1;
 464                                         /* A slave has been removed from the
 465                                          * table because it is either disabled
 466                                          * or being released. We must retry the
 467                                          * update to avoid clients from not
 468                                          * being updated & disconnecting when
 469                                          * there is stress
 470                                          */
 471                                         bond_info->rlb_update_retry_counter =
 472                                                 RLB_UPDATE_RETRY;
 473                                 }
 474                         } else {  /* there is no active slave */
 475                                 rx_hash_table[index].slave = NULL;
 476                         }
 477                 }
 478         }
 479
 480         _unlock_rx_hashtbl(bond);
 481
 482         write_lock_bh(&bond->curr_slave_lock);
 483
 484         if (slave != bond->curr_active_slave) {
 485                 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
 486         }
 487
 488         write_unlock_bh(&bond->curr_slave_lock);
 489 }
 490
 491 static void rlb_update_client(struct rlb_client_info *client_info)
 492 {
 493         int i;
 494
 495         if (!client_info->slave) {
 496                 return;
 497         }
 498
 499         for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
 500                 struct sk_buff *skb;
 501
 502                 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
 503                                  client_info->ip_dst,
 504                                  client_info->slave->dev,
 505                                  client_info->ip_src,
 506                                  client_info->mac_dst,
 507                                  client_info->slave->dev->dev_addr,
 508                                  client_info->mac_dst);
 509                 if (!skb) {
 510                         printk(KERN_ERR DRV_NAME
 511                                ": %s: Error: failed to create an ARP packet\n",
 512                                client_info->slave->dev->master->name);
 513                         continue;
 514                 }
 515
 516                 skb->dev = client_info->slave->dev;
 517
 518                 if (client_info->tag) {
 519                         skb = vlan_put_tag(skb, client_info->vlan_id);
 520                         if (!skb) {
 521                                 printk(KERN_ERR DRV_NAME
 522                                        ": %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 untill 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                     memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 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                         printk(KERN_ERR DRV_NAME
 606                                ": %s: Error: found a client with no channel in "
 607                                "the client's hash table\n",
 608                                bond->dev->name);
 609                         continue;
 610                 }
 611                 /*update all clients using this src_ip, that are not assigned
 612                  * to the team's address (curr_active_slave) and have a known
 613                  * unicast mac address.
 614                  */
 615                 if ((client_info->ip_src == src_ip) &&
 616                     memcmp(client_info->slave->dev->dev_addr,
 617                            bond->dev->dev_addr, ETH_ALEN) &&
 618                     memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 619                         client_info->ntt = 1;
 620                         bond_info->rx_ntt = 1;
 621                 }
 622         }
 623
 624         _unlock_rx_hashtbl(bond);
 625 }
 626
 627 /* Caller must hold both bond and ptr locks for read */
 628 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
 629 {
 630         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 631         struct arp_pkt *arp = arp_pkt(skb);
 632         struct slave *assigned_slave;
 633         struct rlb_client_info *client_info;
 634         u32 hash_index = 0;
 635
 636         _lock_rx_hashtbl(bond);
 637
 638         hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_src));
 639         client_info = &(bond_info->rx_hashtbl[hash_index]);
 640
 641         if (client_info->assigned) {
 642                 if ((client_info->ip_src == arp->ip_src) &&
 643                     (client_info->ip_dst == arp->ip_dst)) {
 644                         /* the entry is already assigned to this client */
 645                         if (memcmp(arp->mac_dst, mac_bcast, ETH_ALEN)) {
 646                                 /* update mac address from arp */
 647                                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 648                         }
 649
 650                         assigned_slave = client_info->slave;
 651                         if (assigned_slave) {
 652                                 _unlock_rx_hashtbl(bond);
 653                                 return assigned_slave;
 654                         }
 655                 } else {
 656                         /* the entry is already assigned to some other client,
 657                          * move the old client to primary (curr_active_slave) so
 658                          * that the new client can be assigned to this entry.
 659                          */
 660                         if (bond->curr_active_slave &&
 661                             client_info->slave != bond->curr_active_slave) {
 662                                 client_info->slave = bond->curr_active_slave;
 663                                 rlb_update_client(client_info);
 664                         }
 665                 }
 666         }
 667         /* assign a new slave */
 668         assigned_slave = rlb_next_rx_slave(bond);
 669
 670         if (assigned_slave) {
 671                 client_info->ip_src = arp->ip_src;
 672                 client_info->ip_dst = arp->ip_dst;
 673                 /* arp->mac_dst is broadcast for arp reqeusts.
 674                  * will be updated with clients actual unicast mac address
 675                  * upon receiving an arp reply.
 676                  */
 677                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 678                 client_info->slave = assigned_slave;
 679
 680                 if (memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 681                         client_info->ntt = 1;
 682                         bond->alb_info.rx_ntt = 1;
 683                 } else {
 684                         client_info->ntt = 0;
 685                 }
 686
 687                 if (!list_empty(&bond->vlan_list)) {
 688                         if (!vlan_get_tag(skb, &client_info->vlan_id))
 689                                 client_info->tag = 1;
 690                 }
 691
 692                 if (!client_info->assigned) {
 693                         u32 prev_tbl_head = bond_info->rx_hashtbl_head;
 694                         bond_info->rx_hashtbl_head = hash_index;
 695                         client_info->next = prev_tbl_head;
 696                         if (prev_tbl_head != RLB_NULL_INDEX) {
 697                                 bond_info->rx_hashtbl[prev_tbl_head].prev =
 698                                         hash_index;
 699                         }
 700                         client_info->assigned = 1;
 701                 }
 702         }
 703
 704         _unlock_rx_hashtbl(bond);
 705
 706         return assigned_slave;
 707 }
 708
 709 /* chooses (and returns) transmit channel for arp reply
 710  * does not choose channel for other arp types since they are
 711  * sent on the curr_active_slave
 712  */
 713 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
 714 {
 715         struct arp_pkt *arp = arp_pkt(skb);
 716         struct slave *tx_slave = NULL;
 717
 718         if (arp->op_code == htons(ARPOP_REPLY)) {
 719                 /* the arp must be sent on the selected
 720                 * rx channel
 721                 */
 722                 tx_slave = rlb_choose_channel(skb, bond);
 723                 if (tx_slave) {
 724                         memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
 725                 }
 726                 dprintk("Server sent ARP Reply packet\n");
 727         } else if (arp->op_code == htons(ARPOP_REQUEST)) {
 728                 /* Create an entry in the rx_hashtbl for this client as a
 729                  * place holder.
 730                  * When the arp reply is received the entry will be updated
 731                  * with the correct unicast address of the client.
 732                  */
 733                 rlb_choose_channel(skb, bond);
 734
 735                 /* The ARP relpy packets must be delayed so that
 736                  * they can cancel out the influence of the ARP request.
 737                  */
 738                 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
 739
 740                 /* arp requests are broadcast and are sent on the primary
 741                  * the arp request will collapse all clients on the subnet to
 742                  * the primary slave. We must register these clients to be
 743                  * updated with their assigned mac.
 744                  */
 745                 rlb_req_update_subnet_clients(bond, arp->ip_src);
 746                 dprintk("Server sent ARP Request packet\n");
 747         }
 748
 749         return tx_slave;
 750 }
 751
 752 /* Caller must hold bond lock for read */
 753 static void rlb_rebalance(struct bonding *bond)
 754 {
 755         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 756         struct slave *assigned_slave;
 757         struct rlb_client_info *client_info;
 758         int ntt;
 759         u32 hash_index;
 760
 761         _lock_rx_hashtbl(bond);
 762
 763         ntt = 0;
 764         hash_index = bond_info->rx_hashtbl_head;
 765         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 766                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 767                 assigned_slave = rlb_next_rx_slave(bond);
 768                 if (assigned_slave && (client_info->slave != assigned_slave)) {
 769                         client_info->slave = assigned_slave;
 770                         client_info->ntt = 1;
 771                         ntt = 1;
 772                 }
 773         }
 774
 775         /* update the team's flag only after the whole iteration */
 776         if (ntt) {
 777                 bond_info->rx_ntt = 1;
 778         }
 779         _unlock_rx_hashtbl(bond);
 780 }
 781
 782 /* Caller must hold rx_hashtbl lock */
 783 static void rlb_init_table_entry(struct rlb_client_info *entry)
 784 {
 785         memset(entry, 0, sizeof(struct rlb_client_info));
 786         entry->next = RLB_NULL_INDEX;
 787         entry->prev = RLB_NULL_INDEX;
 788 }
 789
 790 static int rlb_initialize(struct bonding *bond)
 791 {
 792         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 793         struct packet_type *pk_type = &(BOND_ALB_INFO(bond).rlb_pkt_type);
 794         struct rlb_client_info  *new_hashtbl;
 795         int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
 796         int i;
 797
 798         spin_lock_init(&(bond_info->rx_hashtbl_lock));
 799
 800         new_hashtbl = kmalloc(size, GFP_KERNEL);
 801         if (!new_hashtbl) {
 802                 printk(KERN_ERR DRV_NAME
 803                        ": %s: Error: Failed to allocate RLB hash table\n",
 804                        bond->dev->name);
 805                 return -1;
 806         }
 807         _lock_rx_hashtbl(bond);
 808
 809         bond_info->rx_hashtbl = new_hashtbl;
 810
 811         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 812
 813         for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
 814                 rlb_init_table_entry(bond_info->rx_hashtbl + i);
 815         }
 816
 817         _unlock_rx_hashtbl(bond);
 818
 819         /*initialize packet type*/
 820         pk_type->type = __constant_htons(ETH_P_ARP);
 821         pk_type->dev = bond->dev;
 822         pk_type->func = rlb_arp_recv;
 823
 824         /* register to receive ARPs */
 825         dev_add_pack(pk_type);
 826
 827         return 0;
 828 }
 829
 830 static void rlb_deinitialize(struct bonding *bond)
 831 {
 832         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 833
 834         dev_remove_pack(&(bond_info->rlb_pkt_type));
 835
 836         _lock_rx_hashtbl(bond);
 837
 838         kfree(bond_info->rx_hashtbl);
 839         bond_info->rx_hashtbl = NULL;
 840         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 841
 842         _unlock_rx_hashtbl(bond);
 843 }
 844
 845 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
 846 {
 847         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 848         u32 curr_index;
 849
 850         _lock_rx_hashtbl(bond);
 851
 852         curr_index = bond_info->rx_hashtbl_head;
 853         while (curr_index != RLB_NULL_INDEX) {
 854                 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
 855                 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
 856                 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
 857
 858                 if (curr->tag && (curr->vlan_id == vlan_id)) {
 859                         if (curr_index == bond_info->rx_hashtbl_head) {
 860                                 bond_info->rx_hashtbl_head = next_index;
 861                         }
 862                         if (prev_index != RLB_NULL_INDEX) {
 863                                 bond_info->rx_hashtbl[prev_index].next = next_index;
 864                         }
 865                         if (next_index != RLB_NULL_INDEX) {
 866                                 bond_info->rx_hashtbl[next_index].prev = prev_index;
 867                         }
 868
 869                         rlb_init_table_entry(curr);
 870                 }
 871
 872                 curr_index = next_index;
 873         }
 874
 875         _unlock_rx_hashtbl(bond);
 876 }
 877
 878 /*********************** tlb/rlb shared functions *********************/
 879
 880 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
 881 {
 882         struct bonding *bond = bond_get_bond_by_slave(slave);
 883         struct learning_pkt pkt;
 884         int size = sizeof(struct learning_pkt);
 885         int i;
 886
 887         memset(&pkt, 0, size);
 888         memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
 889         memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
 890         pkt.type = __constant_htons(ETH_P_LOOP);
 891
 892         for (i = 0; i < MAX_LP_BURST; i++) {
 893                 struct sk_buff *skb;
 894                 char *data;
 895
 896                 skb = dev_alloc_skb(size);
 897                 if (!skb) {
 898                         return;
 899                 }
 900
 901                 data = skb_put(skb, size);
 902                 memcpy(data, &pkt, size);
 903
 904                 skb_reset_mac_header(skb);
 905                 skb->network_header = skb->mac_header + ETH_HLEN;
 906                 skb->protocol = pkt.type;
 907                 skb->priority = TC_PRIO_CONTROL;
 908                 skb->dev = slave->dev;
 909
 910                 if (!list_empty(&bond->vlan_list)) {
 911                         struct vlan_entry *vlan;
 912
 913                         vlan = bond_next_vlan(bond,
 914                                               bond->alb_info.current_alb_vlan);
 915
 916                         bond->alb_info.current_alb_vlan = vlan;
 917                         if (!vlan) {
 918                                 kfree_skb(skb);
 919                                 continue;
 920                         }
 921
 922                         skb = vlan_put_tag(skb, vlan->vlan_id);
 923                         if (!skb) {
 924                                 printk(KERN_ERR DRV_NAME
 925                                        ": %s: Error: failed to insert VLAN tag\n",
 926                                        bond->dev->name);
 927                                 continue;
 928                         }
 929                 }
 930
 931                 dev_queue_xmit(skb);
 932         }
 933 }
 934
 935 /* hw is a boolean parameter that determines whether we should try and
 936  * set the hw address of the device as well as the hw address of the
 937  * net_device
 938  */
 939 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[], int hw)
 940 {
 941         struct net_device *dev = slave->dev;
 942         struct sockaddr s_addr;
 943
 944         if (!hw) {
 945                 memcpy(dev->dev_addr, addr, dev->addr_len);
 946                 return 0;
 947         }
 948
 949         /* for rlb each slave must have a unique hw mac addresses so that */
 950         /* each slave will receive packets destined to a different mac */
 951         memcpy(s_addr.sa_data, addr, dev->addr_len);
 952         s_addr.sa_family = dev->type;
 953         if (dev_set_mac_address(dev, &s_addr)) {
 954                 printk(KERN_ERR DRV_NAME
 955                        ": %s: Error: dev_set_mac_address of dev %s failed! ALB "
 956                        "mode requires that the base driver support setting "
 957                        "the hw address also when the network device's "
 958                        "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 = memcmp(slave->perm_hwaddr,
1047                                 slave->dev->dev_addr,
1048                                 ETH_ALEN);
1049         perm_bond_diff = memcmp(slave->perm_hwaddr,
1050                                 bond->dev->dev_addr,
1051                                 ETH_ALEN);
1052
1053         if (perm_curr_diff && perm_bond_diff) {
1054                 struct slave *tmp_slave;
1055                 int i, found = 0;
1056
1057                 bond_for_each_slave(bond, tmp_slave, i) {
1058                         if (!memcmp(slave->perm_hwaddr,
1059                                     tmp_slave->dev->dev_addr,
1060                                     ETH_ALEN)) {
1061                                 found = 1;
1062                                 break;
1063                         }
1064                 }
1065
1066                 if (found) {
1067                         /* locking: needs RTNL and nothing else */
1068                         alb_swap_mac_addr(bond, slave, tmp_slave);
1069                         alb_fasten_mac_swap(bond, slave, tmp_slave);
1070                 }
1071         }
1072 }
1073
1074 /**
1075  * alb_handle_addr_collision_on_attach
1076  * @bond: bonding we're working on
1077  * @slave: the slave that was just attached
1078  *
1079  * checks uniqueness of slave's mac address and handles the case the
1080  * new slave uses the bonds mac address.
1081  *
1082  * If the permanent hw address of @slave is @bond's hw address, we need to
1083  * find a different hw address to give @slave, that isn't in use by any other
1084  * slave in the bond. This address must be, of course, one of the premanent
1085  * addresses of the other slaves.
1086  *
1087  * We go over the slave list, and for each slave there we compare its
1088  * permanent hw address with the current address of all the other slaves.
1089  * If no match was found, then we've found a slave with a permanent address
1090  * that isn't used by any other slave in the bond, so we can assign it to
1091  * @slave.
1092  *
1093  * assumption: this function is called before @slave is attached to the
1094  *             bond slave list.
1095  *
1096  * caller must hold the bond lock for write since the mac addresses are compared
1097  * and may be swapped.
1098  */
1099 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1100 {
1101         struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1102         struct slave *has_bond_addr = bond->curr_active_slave;
1103         int i, j, found = 0;
1104
1105         if (bond->slave_cnt == 0) {
1106                 /* this is the first slave */
1107                 return 0;
1108         }
1109
1110         /* if slave's mac address differs from bond's mac address
1111          * check uniqueness of slave's mac address against the other
1112          * slaves in the bond.
1113          */
1114         if (memcmp(slave->perm_hwaddr, bond->dev->dev_addr, ETH_ALEN)) {
1115                 bond_for_each_slave(bond, tmp_slave1, i) {
1116                         if (!memcmp(tmp_slave1->dev->dev_addr, slave->dev->dev_addr,
1117                                     ETH_ALEN)) {
1118                                 found = 1;
1119                                 break;
1120                         }
1121                 }
1122
1123                 if (!found)
1124                         return 0;
1125
1126                 /* Try setting slave mac to bond address and fall-through
1127                    to code handling that situation below... */
1128                 alb_set_slave_mac_addr(slave, bond->dev->dev_addr,
1129                                        bond->alb_info.rlb_enabled);
1130         }
1131
1132         /* The slave's address is equal to the address of the bond.
1133          * Search for a spare address in the bond for this slave.
1134          */
1135         free_mac_slave = NULL;
1136
1137         bond_for_each_slave(bond, tmp_slave1, i) {
1138                 found = 0;
1139                 bond_for_each_slave(bond, tmp_slave2, j) {
1140                         if (!memcmp(tmp_slave1->perm_hwaddr,
1141                                     tmp_slave2->dev->dev_addr,
1142                                     ETH_ALEN)) {
1143                                 found = 1;
1144                                 break;
1145                         }
1146                 }
1147
1148                 if (!found) {
1149                         /* no slave has tmp_slave1's perm addr
1150                          * as its curr addr
1151                          */
1152                         free_mac_slave = tmp_slave1;
1153                         break;
1154                 }
1155
1156                 if (!has_bond_addr) {
1157                         if (!memcmp(tmp_slave1->dev->dev_addr,
1158                                     bond->dev->dev_addr,
1159                                     ETH_ALEN)) {
1160
1161                                 has_bond_addr = tmp_slave1;
1162                         }
1163                 }
1164         }
1165
1166         if (free_mac_slave) {
1167                 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr,
1168                                        bond->alb_info.rlb_enabled);
1169
1170                 printk(KERN_WARNING DRV_NAME
1171                        ": %s: Warning: the hw address of slave %s is in use by "
1172                        "the bond; giving it the hw address of %s\n",
1173                        bond->dev->name, slave->dev->name, free_mac_slave->dev->name);
1174
1175         } else if (has_bond_addr) {
1176                 printk(KERN_ERR DRV_NAME
1177                        ": %s: Error: the hw address of slave %s is in use by the "
1178                        "bond; couldn't find a slave with a free hw address to "
1179                        "give it (this should not have happened)\n",
1180                        bond->dev->name, slave->dev->name);
1181                 return -EFAULT;
1182         }
1183
1184         return 0;
1185 }
1186
1187 /**
1188  * alb_set_mac_address
1189  * @bond:
1190  * @addr:
1191  *
1192  * In TLB mode all slaves are configured to the bond's hw address, but set
1193  * their dev_addr field to different addresses (based on their permanent hw
1194  * addresses).
1195  *
1196  * For each slave, this function sets the interface to the new address and then
1197  * changes its dev_addr field to its previous value.
1198  *
1199  * Unwinding assumes bond's mac address has not yet changed.
1200  */
1201 static int alb_set_mac_address(struct bonding *bond, void *addr)
1202 {
1203         struct sockaddr sa;
1204         struct slave *slave, *stop_at;
1205         char tmp_addr[ETH_ALEN];
1206         int res;
1207         int i;
1208
1209         if (bond->alb_info.rlb_enabled) {
1210                 return 0;
1211         }
1212
1213         bond_for_each_slave(bond, slave, i) {
1214                 if (slave->dev->set_mac_address == NULL) {
1215                         res = -EOPNOTSUPP;
1216                         goto unwind;
1217                 }
1218
1219                 /* save net_device's current hw address */
1220                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1221
1222                 res = dev_set_mac_address(slave->dev, addr);
1223
1224                 /* restore net_device's hw address */
1225                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1226
1227                 if (res) {
1228                         goto unwind;
1229                 }
1230         }
1231
1232         return 0;
1233
1234 unwind:
1235         memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1236         sa.sa_family = bond->dev->type;
1237
1238         /* unwind from head to the slave that failed */
1239         stop_at = slave;
1240         bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1241                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1242                 dev_set_mac_address(slave->dev, &sa);
1243                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1244         }
1245
1246         return res;
1247 }
1248
1249 /************************ exported alb funcions ************************/
1250
1251 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1252 {
1253         int res;
1254
1255         res = tlb_initialize(bond);
1256         if (res) {
1257                 return res;
1258         }
1259
1260         if (rlb_enabled) {
1261                 bond->alb_info.rlb_enabled = 1;
1262                 /* initialize rlb */
1263                 res = rlb_initialize(bond);
1264                 if (res) {
1265                         tlb_deinitialize(bond);
1266                         return res;
1267                 }
1268         } else {
1269                 bond->alb_info.rlb_enabled = 0;
1270         }
1271
1272         return 0;
1273 }
1274
1275 void bond_alb_deinitialize(struct bonding *bond)
1276 {
1277         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1278
1279         tlb_deinitialize(bond);
1280
1281         if (bond_info->rlb_enabled) {
1282                 rlb_deinitialize(bond);
1283         }
1284 }
1285
1286 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1287 {
1288         struct bonding *bond = bond_dev->priv;
1289         struct ethhdr *eth_data;
1290         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1291         struct slave *tx_slave = NULL;
1292         static const __be32 ip_bcast = htonl(0xffffffff);
1293         int hash_size = 0;
1294         int do_tx_balance = 1;
1295         u32 hash_index = 0;
1296         const u8 *hash_start = NULL;
1297         int res = 1;
1298         struct ipv6hdr *ip6hdr;
1299
1300         skb_reset_mac_header(skb);
1301         eth_data = eth_hdr(skb);
1302
1303         /* make sure that the curr_active_slave and the slaves list do
1304          * not change during tx
1305          */
1306         read_lock(&bond->lock);
1307         read_lock(&bond->curr_slave_lock);
1308
1309         if (!BOND_IS_OK(bond)) {
1310                 goto out;
1311         }
1312
1313         switch (ntohs(skb->protocol)) {
1314         case ETH_P_IP: {
1315                 const struct iphdr *iph = ip_hdr(skb);
1316
1317                 if ((memcmp(eth_data->h_dest, mac_bcast, ETH_ALEN) == 0) ||
1318                     (iph->daddr == ip_bcast) ||
1319                     (iph->protocol == IPPROTO_IGMP)) {
1320                         do_tx_balance = 0;
1321                         break;
1322                 }
1323                 hash_start = (char *)&(iph->daddr);
1324                 hash_size = sizeof(iph->daddr);
1325         }
1326                 break;
1327         case ETH_P_IPV6:
1328                 /* IPv6 doesn't really use broadcast mac address, but leave
1329                  * that here just in case.
1330                  */
1331                 if (memcmp(eth_data->h_dest, mac_bcast, ETH_ALEN) == 0) {
1332                         do_tx_balance = 0;
1333                         break;
1334                 }
1335
1336                 /* IPv6 uses all-nodes multicast as an equivalent to
1337                  * broadcasts in IPv4.
1338                  */
1339                 if (memcmp(eth_data->h_dest, mac_v6_allmcast, ETH_ALEN) == 0) {
1340                         do_tx_balance = 0;
1341                         break;
1342                 }
1343
1344                 /* Additianally, DAD probes should not be tx-balanced as that
1345                  * will lead to false positives for duplicate addresses and
1346                  * prevent address configuration from working.
1347                  */
1348                 ip6hdr = ipv6_hdr(skb);
1349                 if (ipv6_addr_any(&ip6hdr->saddr)) {
1350                         do_tx_balance = 0;
1351                         break;
1352                 }
1353
1354                 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1355                 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1356                 break;
1357         case ETH_P_IPX:
1358                 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1359                         /* something is wrong with this packet */
1360                         do_tx_balance = 0;
1361                         break;
1362                 }
1363
1364                 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1365                         /* The only protocol worth balancing in
1366                          * this family since it has an "ARP" like
1367                          * mechanism
1368                          */
1369                         do_tx_balance = 0;
1370                         break;
1371                 }
1372
1373                 hash_start = (char*)eth_data->h_dest;
1374                 hash_size = ETH_ALEN;
1375                 break;
1376         case ETH_P_ARP:
1377                 do_tx_balance = 0;
1378                 if (bond_info->rlb_enabled) {
1379                         tx_slave = rlb_arp_xmit(skb, bond);
1380                 }
1381                 break;
1382         default:
1383                 do_tx_balance = 0;
1384                 break;
1385         }
1386
1387         if (do_tx_balance) {
1388                 hash_index = _simple_hash(hash_start, hash_size);
1389                 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1390         }
1391
1392         if (!tx_slave) {
1393                 /* unbalanced or unassigned, send through primary */
1394                 tx_slave = bond->curr_active_slave;
1395                 bond_info->unbalanced_load += skb->len;
1396         }
1397
1398         if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1399                 if (tx_slave != bond->curr_active_slave) {
1400                         memcpy(eth_data->h_source,
1401                                tx_slave->dev->dev_addr,
1402                                ETH_ALEN);
1403                 }
1404
1405                 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1406         } else {
1407                 if (tx_slave) {
1408                         tlb_clear_slave(bond, tx_slave, 0);
1409                 }
1410         }
1411
1412 out:
1413         if (res) {
1414                 /* no suitable interface, frame not sent */
1415                 dev_kfree_skb(skb);
1416         }
1417         read_unlock(&bond->curr_slave_lock);
1418         read_unlock(&bond->lock);
1419         return 0;
1420 }
1421
1422 void bond_alb_monitor(struct work_struct *work)
1423 {
1424         struct bonding *bond = container_of(work, struct bonding,
1425                                             alb_work.work);
1426         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1427         struct slave *slave;
1428         int i;
1429
1430         read_lock(&bond->lock);
1431
1432         if (bond->kill_timers) {
1433                 goto out;
1434         }
1435
1436         if (bond->slave_cnt == 0) {
1437                 bond_info->tx_rebalance_counter = 0;
1438                 bond_info->lp_counter = 0;
1439                 goto re_arm;
1440         }
1441
1442         bond_info->tx_rebalance_counter++;
1443         bond_info->lp_counter++;
1444
1445         /* send learning packets */
1446         if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1447                 /* change of curr_active_slave involves swapping of mac addresses.
1448                  * in order to avoid this swapping from happening while
1449                  * sending the learning packets, the curr_slave_lock must be held for
1450                  * read.
1451                  */
1452                 read_lock(&bond->curr_slave_lock);
1453
1454                 bond_for_each_slave(bond, slave, i) {
1455                         alb_send_learning_packets(slave, slave->dev->dev_addr);
1456                 }
1457
1458                 read_unlock(&bond->curr_slave_lock);
1459
1460                 bond_info->lp_counter = 0;
1461         }
1462
1463         /* rebalance tx traffic */
1464         if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1465
1466                 read_lock(&bond->curr_slave_lock);
1467
1468                 bond_for_each_slave(bond, slave, i) {
1469                         tlb_clear_slave(bond, slave, 1);
1470                         if (slave == bond->curr_active_slave) {
1471                                 SLAVE_TLB_INFO(slave).load =
1472                                         bond_info->unbalanced_load /
1473                                                 BOND_TLB_REBALANCE_INTERVAL;
1474                                 bond_info->unbalanced_load = 0;
1475                         }
1476                 }
1477
1478                 read_unlock(&bond->curr_slave_lock);
1479
1480                 bond_info->tx_rebalance_counter = 0;
1481         }
1482
1483         /* handle rlb stuff */
1484         if (bond_info->rlb_enabled) {
1485                 if (bond_info->primary_is_promisc &&
1486                     (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1487
1488                         /*
1489                          * dev_set_promiscuity requires rtnl and
1490                          * nothing else.
1491                          */
1492                         read_unlock(&bond->lock);
1493                         rtnl_lock();
1494
1495                         bond_info->rlb_promisc_timeout_counter = 0;
1496
1497                         /* If the primary was set to promiscuous mode
1498                          * because a slave was disabled then
1499                          * it can now leave promiscuous mode.
1500                          */
1501                         dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1502                         bond_info->primary_is_promisc = 0;
1503
1504                         rtnl_unlock();
1505                         read_lock(&bond->lock);
1506                 }
1507
1508                 if (bond_info->rlb_rebalance) {
1509                         bond_info->rlb_rebalance = 0;
1510                         rlb_rebalance(bond);
1511                 }
1512
1513                 /* check if clients need updating */
1514                 if (bond_info->rx_ntt) {
1515                         if (bond_info->rlb_update_delay_counter) {
1516                                 --bond_info->rlb_update_delay_counter;
1517                         } else {
1518                                 rlb_update_rx_clients(bond);
1519                                 if (bond_info->rlb_update_retry_counter) {
1520                                         --bond_info->rlb_update_retry_counter;
1521                                 } else {
1522                                         bond_info->rx_ntt = 0;
1523                                 }
1524                         }
1525                 }
1526         }
1527
1528 re_arm:
1529         queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1530 out:
1531         read_unlock(&bond->lock);
1532 }
1533
1534 /* assumption: called before the slave is attached to the bond
1535  * and not locked by the bond lock
1536  */
1537 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1538 {
1539         int res;
1540
1541         res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr,
1542                                      bond->alb_info.rlb_enabled);
1543         if (res) {
1544                 return res;
1545         }
1546
1547         /* caller must hold the bond lock for write since the mac addresses
1548          * are compared and may be swapped.
1549          */
1550         read_lock(&bond->lock);
1551
1552         res = alb_handle_addr_collision_on_attach(bond, slave);
1553
1554         read_unlock(&bond->lock);
1555
1556         if (res) {
1557                 return res;
1558         }
1559
1560         tlb_init_slave(slave);
1561
1562         /* order a rebalance ASAP */
1563         bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1564
1565         if (bond->alb_info.rlb_enabled) {
1566                 bond->alb_info.rlb_rebalance = 1;
1567         }
1568
1569         return 0;
1570 }
1571
1572 /*
1573  * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1574  * if necessary.
1575  *
1576  * Caller must hold RTNL and no other locks
1577  */
1578 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1579 {
1580         if (bond->slave_cnt > 1) {
1581                 alb_change_hw_addr_on_detach(bond, slave);
1582         }
1583
1584         tlb_clear_slave(bond, slave, 0);
1585
1586         if (bond->alb_info.rlb_enabled) {
1587                 bond->alb_info.next_rx_slave = NULL;
1588                 rlb_clear_slave(bond, slave);
1589         }
1590 }
1591
1592 /* Caller must hold bond lock for read */
1593 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1594 {
1595         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1596
1597         if (link == BOND_LINK_DOWN) {
1598                 tlb_clear_slave(bond, slave, 0);
1599                 if (bond->alb_info.rlb_enabled) {
1600                         rlb_clear_slave(bond, slave);
1601                 }
1602         } else if (link == BOND_LINK_UP) {
1603                 /* order a rebalance ASAP */
1604                 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1605                 if (bond->alb_info.rlb_enabled) {
1606                         bond->alb_info.rlb_rebalance = 1;
1607                         /* If the updelay module parameter is smaller than the
1608                          * forwarding delay of the switch the rebalance will
1609                          * not work because the rebalance arp replies will
1610                          * not be forwarded to the clients..
1611                          */
1612                 }
1613         }
1614 }
1615
1616 /**
1617  * bond_alb_handle_active_change - assign new curr_active_slave
1618  * @bond: our bonding struct
1619  * @new_slave: new slave to assign
1620  *
1621  * Set the bond->curr_active_slave to @new_slave and handle
1622  * mac address swapping and promiscuity changes as needed.
1623  *
1624  * If new_slave is NULL, caller must hold curr_slave_lock or
1625  * bond->lock for write.
1626  *
1627  * If new_slave is not NULL, caller must hold RTNL, bond->lock for
1628  * read and curr_slave_lock for write.  Processing here may sleep, so
1629  * no other locks may be held.
1630  */
1631 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1632 {
1633         struct slave *swap_slave;
1634         int i;
1635
1636         if (bond->curr_active_slave == new_slave) {
1637                 return;
1638         }
1639
1640         if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1641                 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1642                 bond->alb_info.primary_is_promisc = 0;
1643                 bond->alb_info.rlb_promisc_timeout_counter = 0;
1644         }
1645
1646         swap_slave = bond->curr_active_slave;
1647         bond->curr_active_slave = new_slave;
1648
1649         if (!new_slave || (bond->slave_cnt == 0)) {
1650                 return;
1651         }
1652
1653         /* set the new curr_active_slave to the bonds mac address
1654          * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1655          */
1656         if (!swap_slave) {
1657                 struct slave *tmp_slave;
1658                 /* find slave that is holding the bond's mac address */
1659                 bond_for_each_slave(bond, tmp_slave, i) {
1660                         if (!memcmp(tmp_slave->dev->dev_addr,
1661                                     bond->dev->dev_addr, ETH_ALEN)) {
1662                                 swap_slave = tmp_slave;
1663                                 break;
1664                         }
1665                 }
1666         }
1667
1668         /*
1669          * Arrange for swap_slave and new_slave to temporarily be
1670          * ignored so we can mess with their MAC addresses without
1671          * fear of interference from transmit activity.
1672          */
1673         if (swap_slave) {
1674                 tlb_clear_slave(bond, swap_slave, 1);
1675         }
1676         tlb_clear_slave(bond, new_slave, 1);
1677
1678         write_unlock_bh(&bond->curr_slave_lock);
1679         read_unlock(&bond->lock);
1680
1681         ASSERT_RTNL();
1682
1683         /* curr_active_slave must be set before calling alb_swap_mac_addr */
1684         if (swap_slave) {
1685                 /* swap mac address */
1686                 alb_swap_mac_addr(bond, swap_slave, new_slave);
1687         } else {
1688                 /* set the new_slave to the bond mac address */
1689                 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr,
1690                                        bond->alb_info.rlb_enabled);
1691         }
1692
1693         if (swap_slave) {
1694                 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1695                 read_lock(&bond->lock);
1696         } else {
1697                 read_lock(&bond->lock);
1698                 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1699         }
1700
1701         write_lock_bh(&bond->curr_slave_lock);
1702 }
1703
1704 /*
1705  * Called with RTNL
1706  */
1707 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1708 {
1709         struct bonding *bond = bond_dev->priv;
1710         struct sockaddr *sa = addr;
1711         struct slave *slave, *swap_slave;
1712         int res;
1713         int i;
1714
1715         if (!is_valid_ether_addr(sa->sa_data)) {
1716                 return -EADDRNOTAVAIL;
1717         }
1718
1719         res = alb_set_mac_address(bond, addr);
1720         if (res) {
1721                 return res;
1722         }
1723
1724         memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1725
1726         /* If there is no curr_active_slave there is nothing else to do.
1727          * Otherwise we'll need to pass the new address to it and handle
1728          * duplications.
1729          */
1730         if (!bond->curr_active_slave) {
1731                 return 0;
1732         }
1733
1734         swap_slave = NULL;
1735
1736         bond_for_each_slave(bond, slave, i) {
1737                 if (!memcmp(slave->dev->dev_addr, bond_dev->dev_addr, ETH_ALEN)) {
1738                         swap_slave = slave;
1739                         break;
1740                 }
1741         }
1742
1743         write_unlock_bh(&bond->curr_slave_lock);
1744         read_unlock(&bond->lock);
1745
1746         if (swap_slave) {
1747                 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1748                 alb_fasten_mac_swap(bond, swap_slave, bond->curr_active_slave);
1749         } else {
1750                 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr,
1751                                        bond->alb_info.rlb_enabled);
1752
1753                 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1754                 if (bond->alb_info.rlb_enabled) {
1755                         /* inform clients mac address has changed */
1756                         rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1757                 }
1758         }
1759
1760         read_lock(&bond->lock);
1761         write_lock_bh(&bond->curr_slave_lock);
1762
1763         return 0;
1764 }
1765
1766 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1767 {
1768         if (bond->alb_info.current_alb_vlan &&
1769             (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1770                 bond->alb_info.current_alb_vlan = NULL;
1771         }
1772
1773         if (bond->alb_info.rlb_enabled) {
1774                 rlb_clear_vlan(bond, vlan_id);
1775         }
1776 }
1777