git.oblomov.eu Git - linux-2.6/blob - 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 <asm/byteorder.h>
  42 #include "bonding.h"
  43 #include "bond_alb.h"
  44
  45
  46 #define ALB_TIMER_TICKS_PER_SEC     10  /* should be a divisor of HZ */
  47 #define BOND_TLB_REBALANCE_INTERVAL 10  /* In seconds, periodic re-balancing.
  48                                          * Used for division - never set
  49                                          * to zero !!!
  50                                          */
  51 #define BOND_ALB_LP_INTERVAL        1   /* In seconds, periodic send of
  52                                          * learning packets to the switch
  53                                          */
  54
  55 #define BOND_TLB_REBALANCE_TICKS (BOND_TLB_REBALANCE_INTERVAL \
  56                                   * ALB_TIMER_TICKS_PER_SEC)
  57
  58 #define BOND_ALB_LP_TICKS (BOND_ALB_LP_INTERVAL \
  59                            * ALB_TIMER_TICKS_PER_SEC)
  60
  61 #define TLB_HASH_TABLE_SIZE 256 /* The size of the clients hash table.
  62                                  * Note that this value MUST NOT be smaller
  63                                  * because the key hash table is BYTE wide !
  64                                  */
  65
  66
  67 #define TLB_NULL_INDEX          0xffffffff
  68 #define MAX_LP_BURST            3
  69
  70 /* rlb defs */
  71 #define RLB_HASH_TABLE_SIZE     256
  72 #define RLB_NULL_INDEX          0xffffffff
  73 #define RLB_UPDATE_DELAY        2*ALB_TIMER_TICKS_PER_SEC /* 2 seconds */
  74 #define RLB_ARP_BURST_SIZE      2
  75 #define RLB_UPDATE_RETRY        3       /* 3-ticks - must be smaller than the rlb
  76                                          * rebalance interval (5 min).
  77                                          */
  78 /* RLB_PROMISC_TIMEOUT = 10 sec equals the time that the current slave is
  79  * promiscuous after failover
  80  */
  81 #define RLB_PROMISC_TIMEOUT     10*ALB_TIMER_TICKS_PER_SEC
  82
  83 static const u8 mac_bcast[ETH_ALEN] = {0xff,0xff,0xff,0xff,0xff,0xff};
  84 static const int alb_delta_in_ticks = HZ / ALB_TIMER_TICKS_PER_SEC;
  85
  86 #pragma pack(1)
  87 struct learning_pkt {
  88         u8 mac_dst[ETH_ALEN];
  89         u8 mac_src[ETH_ALEN];
  90         __be16 type;
  91         u8 padding[ETH_ZLEN - ETH_HLEN];
  92 };
  93
  94 struct arp_pkt {
  95         __be16  hw_addr_space;
  96         __be16  prot_addr_space;
  97         u8      hw_addr_len;
  98         u8      prot_addr_len;
  99         __be16  op_code;
 100         u8      mac_src[ETH_ALEN];      /* sender hardware address */
 101         __be32  ip_src;                 /* sender IP address */
 102         u8      mac_dst[ETH_ALEN];      /* target hardware address */
 103         __be32  ip_dst;                 /* target IP address */
 104 };
 105 #pragma pack()
 106
 107 static inline struct arp_pkt *arp_pkt(const struct sk_buff *skb)
 108 {
 109         return (struct arp_pkt *)skb_network_header(skb);
 110 }
 111
 112 /* Forward declaration */
 113 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[]);
 114
 115 static inline u8 _simple_hash(const u8 *hash_start, int hash_size)
 116 {
 117         int i;
 118         u8 hash = 0;
 119
 120         for (i = 0; i < hash_size; i++) {
 121                 hash ^= hash_start[i];
 122         }
 123
 124         return hash;
 125 }
 126
 127 /*********************** tlb specific functions ***************************/
 128
 129 static inline void _lock_tx_hashtbl(struct bonding *bond)
 130 {
 131         spin_lock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 132 }
 133
 134 static inline void _unlock_tx_hashtbl(struct bonding *bond)
 135 {
 136         spin_unlock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 137 }
 138
 139 /* Caller must hold tx_hashtbl lock */
 140 static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
 141 {
 142         if (save_load) {
 143                 entry->load_history = 1 + entry->tx_bytes /
 144                                       BOND_TLB_REBALANCE_INTERVAL;
 145                 entry->tx_bytes = 0;
 146         }
 147
 148         entry->tx_slave = NULL;
 149         entry->next = TLB_NULL_INDEX;
 150         entry->prev = TLB_NULL_INDEX;
 151 }
 152
 153 static inline void tlb_init_slave(struct slave *slave)
 154 {
 155         SLAVE_TLB_INFO(slave).load = 0;
 156         SLAVE_TLB_INFO(slave).head = TLB_NULL_INDEX;
 157 }
 158
 159 /* Caller must hold bond lock for read */
 160 static void tlb_clear_slave(struct bonding *bond, struct slave *slave, int save_load)
 161 {
 162         struct tlb_client_info *tx_hash_table;
 163         u32 index;
 164
 165         _lock_tx_hashtbl(bond);
 166
 167         /* clear slave from tx_hashtbl */
 168         tx_hash_table = BOND_ALB_INFO(bond).tx_hashtbl;
 169
 170         index = SLAVE_TLB_INFO(slave).head;
 171         while (index != TLB_NULL_INDEX) {
 172                 u32 next_index = tx_hash_table[index].next;
 173                 tlb_init_table_entry(&tx_hash_table[index], save_load);
 174                 index = next_index;
 175         }
 176
 177         tlb_init_slave(slave);
 178
 179         _unlock_tx_hashtbl(bond);
 180 }
 181
 182 /* Must be called before starting the monitor timer */
 183 static int tlb_initialize(struct bonding *bond)
 184 {
 185         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 186         int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
 187         struct tlb_client_info *new_hashtbl;
 188         int i;
 189
 190         spin_lock_init(&(bond_info->tx_hashtbl_lock));
 191
 192         new_hashtbl = kzalloc(size, GFP_KERNEL);
 193         if (!new_hashtbl) {
 194                 printk(KERN_ERR DRV_NAME
 195                        ": %s: Error: Failed to allocate TLB hash table\n",
 196                        bond->dev->name);
 197                 return -1;
 198         }
 199         _lock_tx_hashtbl(bond);
 200
 201         bond_info->tx_hashtbl = new_hashtbl;
 202
 203         for (i = 0; i < TLB_HASH_TABLE_SIZE; i++) {
 204                 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 1);
 205         }
 206
 207         _unlock_tx_hashtbl(bond);
 208
 209         return 0;
 210 }
 211
 212 /* Must be called only after all slaves have been released */
 213 static void tlb_deinitialize(struct bonding *bond)
 214 {
 215         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 216
 217         _lock_tx_hashtbl(bond);
 218
 219         kfree(bond_info->tx_hashtbl);
 220         bond_info->tx_hashtbl = NULL;
 221
 222         _unlock_tx_hashtbl(bond);
 223 }
 224
 225 /* Caller must hold bond lock for read */
 226 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
 227 {
 228         struct slave *slave, *least_loaded;
 229         s64 max_gap;
 230         int i, found = 0;
 231
 232         /* Find the first enabled slave */
 233         bond_for_each_slave(bond, slave, i) {
 234                 if (SLAVE_IS_OK(slave)) {
 235                         found = 1;
 236                         break;
 237                 }
 238         }
 239
 240         if (!found) {
 241                 return NULL;
 242         }
 243
 244         least_loaded = slave;
 245         max_gap = (s64)(slave->speed << 20) - /* Convert to Megabit per sec */
 246                         (s64)(SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
 247
 248         /* Find the slave with the largest gap */
 249         bond_for_each_slave_from(bond, slave, i, least_loaded) {
 250                 if (SLAVE_IS_OK(slave)) {
 251                         s64 gap = (s64)(slave->speed << 20) -
 252                                         (s64)(SLAVE_TLB_INFO(slave).load << 3);
 253                         if (max_gap < gap) {
 254                                 least_loaded = slave;
 255                                 max_gap = gap;
 256                         }
 257                 }
 258         }
 259
 260         return least_loaded;
 261 }
 262
 263 /* Caller must hold bond lock for read */
 264 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index, u32 skb_len)
 265 {
 266         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 267         struct tlb_client_info *hash_table;
 268         struct slave *assigned_slave;
 269
 270         _lock_tx_hashtbl(bond);
 271
 272         hash_table = bond_info->tx_hashtbl;
 273         assigned_slave = hash_table[hash_index].tx_slave;
 274         if (!assigned_slave) {
 275                 assigned_slave = tlb_get_least_loaded_slave(bond);
 276
 277                 if (assigned_slave) {
 278                         struct tlb_slave_info *slave_info =
 279                                 &(SLAVE_TLB_INFO(assigned_slave));
 280                         u32 next_index = slave_info->head;
 281
 282                         hash_table[hash_index].tx_slave = assigned_slave;
 283                         hash_table[hash_index].next = next_index;
 284                         hash_table[hash_index].prev = TLB_NULL_INDEX;
 285
 286                         if (next_index != TLB_NULL_INDEX) {
 287                                 hash_table[next_index].prev = hash_index;
 288                         }
 289
 290                         slave_info->head = hash_index;
 291                         slave_info->load +=
 292                                 hash_table[hash_index].load_history;
 293                 }
 294         }
 295
 296         if (assigned_slave) {
 297                 hash_table[hash_index].tx_bytes += skb_len;
 298         }
 299
 300         _unlock_tx_hashtbl(bond);
 301
 302         return assigned_slave;
 303 }
 304
 305 /*********************** rlb specific functions ***************************/
 306 static inline void _lock_rx_hashtbl(struct bonding *bond)
 307 {
 308         spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 309 }
 310
 311 static inline void _unlock_rx_hashtbl(struct bonding *bond)
 312 {
 313         spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 314 }
 315
 316 /* when an ARP REPLY is received from a client update its info
 317  * in the rx_hashtbl
 318  */
 319 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
 320 {
 321         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 322         struct rlb_client_info *client_info;
 323         u32 hash_index;
 324
 325         _lock_rx_hashtbl(bond);
 326
 327         hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
 328         client_info = &(bond_info->rx_hashtbl[hash_index]);
 329
 330         if ((client_info->assigned) &&
 331             (client_info->ip_src == arp->ip_dst) &&
 332             (client_info->ip_dst == arp->ip_src)) {
 333                 /* update the clients MAC address */
 334                 memcpy(client_info->mac_dst, arp->mac_src, ETH_ALEN);
 335                 client_info->ntt = 1;
 336                 bond_info->rx_ntt = 1;
 337         }
 338
 339         _unlock_rx_hashtbl(bond);
 340 }
 341
 342 static int rlb_arp_recv(struct sk_buff *skb, struct net_device *bond_dev, struct packet_type *ptype, struct net_device *orig_dev)
 343 {
 344         struct bonding *bond = bond_dev->priv;
 345         struct arp_pkt *arp = (struct arp_pkt *)skb->data;
 346         int res = NET_RX_DROP;
 347
 348         if (bond_dev->nd_net != &init_net)
 349                 goto out;
 350
 351         if (!(bond_dev->flags & IFF_MASTER))
 352                 goto out;
 353
 354         if (!arp) {
 355                 dprintk("Packet has no ARP data\n");
 356                 goto out;
 357         }
 358
 359         if (skb->len < sizeof(struct arp_pkt)) {
 360                 dprintk("Packet is too small to be an ARP\n");
 361                 goto out;
 362         }
 363
 364         if (arp->op_code == htons(ARPOP_REPLY)) {
 365                 /* update rx hash table for this ARP */
 366                 rlb_update_entry_from_arp(bond, arp);
 367                 dprintk("Server received an ARP Reply from client\n");
 368         }
 369
 370         res = NET_RX_SUCCESS;
 371
 372 out:
 373         dev_kfree_skb(skb);
 374
 375         return res;
 376 }
 377
 378 /* Caller must hold bond lock for read */
 379 static struct slave *rlb_next_rx_slave(struct bonding *bond)
 380 {
 381         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 382         struct slave *rx_slave, *slave, *start_at;
 383         int i = 0;
 384
 385         if (bond_info->next_rx_slave) {
 386                 start_at = bond_info->next_rx_slave;
 387         } else {
 388                 start_at = bond->first_slave;
 389         }
 390
 391         rx_slave = NULL;
 392
 393         bond_for_each_slave_from(bond, slave, i, start_at) {
 394                 if (SLAVE_IS_OK(slave)) {
 395                         if (!rx_slave) {
 396                                 rx_slave = slave;
 397                         } else if (slave->speed > rx_slave->speed) {
 398                                 rx_slave = slave;
 399                         }
 400                 }
 401         }
 402
 403         if (rx_slave) {
 404                 bond_info->next_rx_slave = rx_slave->next;
 405         }
 406
 407         return rx_slave;
 408 }
 409
 410 /* teach the switch the mac of a disabled slave
 411  * on the primary for fault tolerance
 412  *
 413  * Caller must hold bond->curr_slave_lock for write or bond lock for write
 414  */
 415 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
 416 {
 417         if (!bond->curr_active_slave) {
 418                 return;
 419         }
 420
 421         if (!bond->alb_info.primary_is_promisc) {
 422                 bond->alb_info.primary_is_promisc = 1;
 423                 dev_set_promiscuity(bond->curr_active_slave->dev, 1);
 424         }
 425
 426         bond->alb_info.rlb_promisc_timeout_counter = 0;
 427
 428         alb_send_learning_packets(bond->curr_active_slave, addr);
 429 }
 430
 431 /* slave being removed should not be active at this point
 432  *
 433  * Caller must hold bond lock for read
 434  */
 435 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
 436 {
 437         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 438         struct rlb_client_info *rx_hash_table;
 439         u32 index, next_index;
 440
 441         /* clear slave from rx_hashtbl */
 442         _lock_rx_hashtbl(bond);
 443
 444         rx_hash_table = bond_info->rx_hashtbl;
 445         index = bond_info->rx_hashtbl_head;
 446         for (; index != RLB_NULL_INDEX; index = next_index) {
 447                 next_index = rx_hash_table[index].next;
 448                 if (rx_hash_table[index].slave == slave) {
 449                         struct slave *assigned_slave = rlb_next_rx_slave(bond);
 450
 451                         if (assigned_slave) {
 452                                 rx_hash_table[index].slave = assigned_slave;
 453                                 if (memcmp(rx_hash_table[index].mac_dst,
 454                                            mac_bcast, ETH_ALEN)) {
 455                                         bond_info->rx_hashtbl[index].ntt = 1;
 456                                         bond_info->rx_ntt = 1;
 457                                         /* A slave has been removed from the
 458                                          * table because it is either disabled
 459                                          * or being released. We must retry the
 460                                          * update to avoid clients from not
 461                                          * being updated & disconnecting when
 462                                          * there is stress
 463                                          */
 464                                         bond_info->rlb_update_retry_counter =
 465                                                 RLB_UPDATE_RETRY;
 466                                 }
 467                         } else {  /* there is no active slave */
 468                                 rx_hash_table[index].slave = NULL;
 469                         }
 470                 }
 471         }
 472
 473         _unlock_rx_hashtbl(bond);
 474
 475         write_lock_bh(&bond->curr_slave_lock);
 476
 477         if (slave != bond->curr_active_slave) {
 478                 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
 479         }
 480
 481         write_unlock_bh(&bond->curr_slave_lock);
 482 }
 483
 484 static void rlb_update_client(struct rlb_client_info *client_info)
 485 {
 486         int i;
 487
 488         if (!client_info->slave) {
 489                 return;
 490         }
 491
 492         for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
 493                 struct sk_buff *skb;
 494
 495                 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
 496                                  client_info->ip_dst,
 497                                  client_info->slave->dev,
 498                                  client_info->ip_src,
 499                                  client_info->mac_dst,
 500                                  client_info->slave->dev->dev_addr,
 501                                  client_info->mac_dst);
 502                 if (!skb) {
 503                         printk(KERN_ERR DRV_NAME
 504                                ": %s: Error: failed to create an ARP packet\n",
 505                                client_info->slave->dev->master->name);
 506                         continue;
 507                 }
 508
 509                 skb->dev = client_info->slave->dev;
 510
 511                 if (client_info->tag) {
 512                         skb = vlan_put_tag(skb, client_info->vlan_id);
 513                         if (!skb) {
 514                                 printk(KERN_ERR DRV_NAME
 515                                        ": %s: Error: failed to insert VLAN tag\n",
 516                                        client_info->slave->dev->master->name);
 517                                 continue;
 518                         }
 519                 }
 520
 521                 arp_xmit(skb);
 522         }
 523 }
 524
 525 /* sends ARP REPLIES that update the clients that need updating */
 526 static void rlb_update_rx_clients(struct bonding *bond)
 527 {
 528         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 529         struct rlb_client_info *client_info;
 530         u32 hash_index;
 531
 532         _lock_rx_hashtbl(bond);
 533
 534         hash_index = bond_info->rx_hashtbl_head;
 535         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 536                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 537                 if (client_info->ntt) {
 538                         rlb_update_client(client_info);
 539                         if (bond_info->rlb_update_retry_counter == 0) {
 540                                 client_info->ntt = 0;
 541                         }
 542                 }
 543         }
 544
 545         /* do not update the entries again untill this counter is zero so that
 546          * not to confuse the clients.
 547          */
 548         bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
 549
 550         _unlock_rx_hashtbl(bond);
 551 }
 552
 553 /* The slave was assigned a new mac address - update the clients */
 554 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
 555 {
 556         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 557         struct rlb_client_info *client_info;
 558         int ntt = 0;
 559         u32 hash_index;
 560
 561         _lock_rx_hashtbl(bond);
 562
 563         hash_index = bond_info->rx_hashtbl_head;
 564         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 565                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 566
 567                 if ((client_info->slave == slave) &&
 568                     memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 569                         client_info->ntt = 1;
 570                         ntt = 1;
 571                 }
 572         }
 573
 574         // update the team's flag only after the whole iteration
 575         if (ntt) {
 576                 bond_info->rx_ntt = 1;
 577                 //fasten the change
 578                 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
 579         }
 580
 581         _unlock_rx_hashtbl(bond);
 582 }
 583
 584 /* mark all clients using src_ip to be updated */
 585 static void rlb_req_update_subnet_clients(struct bonding *bond, __be32 src_ip)
 586 {
 587         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 588         struct rlb_client_info *client_info;
 589         u32 hash_index;
 590
 591         _lock_rx_hashtbl(bond);
 592
 593         hash_index = bond_info->rx_hashtbl_head;
 594         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 595                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 596
 597                 if (!client_info->slave) {
 598                         printk(KERN_ERR DRV_NAME
 599                                ": %s: Error: found a client with no channel in "
 600                                "the client's hash table\n",
 601                                bond->dev->name);
 602                         continue;
 603                 }
 604                 /*update all clients using this src_ip, that are not assigned
 605                  * to the team's address (curr_active_slave) and have a known
 606                  * unicast mac address.
 607                  */
 608                 if ((client_info->ip_src == src_ip) &&
 609                     memcmp(client_info->slave->dev->dev_addr,
 610                            bond->dev->dev_addr, ETH_ALEN) &&
 611                     memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 612                         client_info->ntt = 1;
 613                         bond_info->rx_ntt = 1;
 614                 }
 615         }
 616
 617         _unlock_rx_hashtbl(bond);
 618 }
 619
 620 /* Caller must hold both bond and ptr locks for read */
 621 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
 622 {
 623         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 624         struct arp_pkt *arp = arp_pkt(skb);
 625         struct slave *assigned_slave;
 626         struct rlb_client_info *client_info;
 627         u32 hash_index = 0;
 628
 629         _lock_rx_hashtbl(bond);
 630
 631         hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_src));
 632         client_info = &(bond_info->rx_hashtbl[hash_index]);
 633
 634         if (client_info->assigned) {
 635                 if ((client_info->ip_src == arp->ip_src) &&
 636                     (client_info->ip_dst == arp->ip_dst)) {
 637                         /* the entry is already assigned to this client */
 638                         if (memcmp(arp->mac_dst, mac_bcast, ETH_ALEN)) {
 639                                 /* update mac address from arp */
 640                                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 641                         }
 642
 643                         assigned_slave = client_info->slave;
 644                         if (assigned_slave) {
 645                                 _unlock_rx_hashtbl(bond);
 646                                 return assigned_slave;
 647                         }
 648                 } else {
 649                         /* the entry is already assigned to some other client,
 650                          * move the old client to primary (curr_active_slave) so
 651                          * that the new client can be assigned to this entry.
 652                          */
 653                         if (bond->curr_active_slave &&
 654                             client_info->slave != bond->curr_active_slave) {
 655                                 client_info->slave = bond->curr_active_slave;
 656                                 rlb_update_client(client_info);
 657                         }
 658                 }
 659         }
 660         /* assign a new slave */
 661         assigned_slave = rlb_next_rx_slave(bond);
 662
 663         if (assigned_slave) {
 664                 client_info->ip_src = arp->ip_src;
 665                 client_info->ip_dst = arp->ip_dst;
 666                 /* arp->mac_dst is broadcast for arp reqeusts.
 667                  * will be updated with clients actual unicast mac address
 668                  * upon receiving an arp reply.
 669                  */
 670                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 671                 client_info->slave = assigned_slave;
 672
 673                 if (memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 674                         client_info->ntt = 1;
 675                         bond->alb_info.rx_ntt = 1;
 676                 } else {
 677                         client_info->ntt = 0;
 678                 }
 679
 680                 if (!list_empty(&bond->vlan_list)) {
 681                         if (!vlan_get_tag(skb, &client_info->vlan_id))
 682                                 client_info->tag = 1;
 683                 }
 684
 685                 if (!client_info->assigned) {
 686                         u32 prev_tbl_head = bond_info->rx_hashtbl_head;
 687                         bond_info->rx_hashtbl_head = hash_index;
 688                         client_info->next = prev_tbl_head;
 689                         if (prev_tbl_head != RLB_NULL_INDEX) {
 690                                 bond_info->rx_hashtbl[prev_tbl_head].prev =
 691                                         hash_index;
 692                         }
 693                         client_info->assigned = 1;
 694                 }
 695         }
 696
 697         _unlock_rx_hashtbl(bond);
 698
 699         return assigned_slave;
 700 }
 701
 702 /* chooses (and returns) transmit channel for arp reply
 703  * does not choose channel for other arp types since they are
 704  * sent on the curr_active_slave
 705  */
 706 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
 707 {
 708         struct arp_pkt *arp = arp_pkt(skb);
 709         struct slave *tx_slave = NULL;
 710
 711         if (arp->op_code == __constant_htons(ARPOP_REPLY)) {
 712                 /* the arp must be sent on the selected
 713                 * rx channel
 714                 */
 715                 tx_slave = rlb_choose_channel(skb, bond);
 716                 if (tx_slave) {
 717                         memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
 718                 }
 719                 dprintk("Server sent ARP Reply packet\n");
 720         } else if (arp->op_code == __constant_htons(ARPOP_REQUEST)) {
 721                 /* Create an entry in the rx_hashtbl for this client as a
 722                  * place holder.
 723                  * When the arp reply is received the entry will be updated
 724                  * with the correct unicast address of the client.
 725                  */
 726                 rlb_choose_channel(skb, bond);
 727
 728                 /* The ARP relpy packets must be delayed so that
 729                  * they can cancel out the influence of the ARP request.
 730                  */
 731                 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
 732
 733                 /* arp requests are broadcast and are sent on the primary
 734                  * the arp request will collapse all clients on the subnet to
 735                  * the primary slave. We must register these clients to be
 736                  * updated with their assigned mac.
 737                  */
 738                 rlb_req_update_subnet_clients(bond, arp->ip_src);
 739                 dprintk("Server sent ARP Request packet\n");
 740         }
 741
 742         return tx_slave;
 743 }
 744
 745 /* Caller must hold bond lock for read */
 746 static void rlb_rebalance(struct bonding *bond)
 747 {
 748         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 749         struct slave *assigned_slave;
 750         struct rlb_client_info *client_info;
 751         int ntt;
 752         u32 hash_index;
 753
 754         _lock_rx_hashtbl(bond);
 755
 756         ntt = 0;
 757         hash_index = bond_info->rx_hashtbl_head;
 758         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 759                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 760                 assigned_slave = rlb_next_rx_slave(bond);
 761                 if (assigned_slave && (client_info->slave != assigned_slave)) {
 762                         client_info->slave = assigned_slave;
 763                         client_info->ntt = 1;
 764                         ntt = 1;
 765                 }
 766         }
 767
 768         /* update the team's flag only after the whole iteration */
 769         if (ntt) {
 770                 bond_info->rx_ntt = 1;
 771         }
 772         _unlock_rx_hashtbl(bond);
 773 }
 774
 775 /* Caller must hold rx_hashtbl lock */
 776 static void rlb_init_table_entry(struct rlb_client_info *entry)
 777 {
 778         memset(entry, 0, sizeof(struct rlb_client_info));
 779         entry->next = RLB_NULL_INDEX;
 780         entry->prev = RLB_NULL_INDEX;
 781 }
 782
 783 static int rlb_initialize(struct bonding *bond)
 784 {
 785         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 786         struct packet_type *pk_type = &(BOND_ALB_INFO(bond).rlb_pkt_type);
 787         struct rlb_client_info  *new_hashtbl;
 788         int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
 789         int i;
 790
 791         spin_lock_init(&(bond_info->rx_hashtbl_lock));
 792
 793         new_hashtbl = kmalloc(size, GFP_KERNEL);
 794         if (!new_hashtbl) {
 795                 printk(KERN_ERR DRV_NAME
 796                        ": %s: Error: Failed to allocate RLB hash table\n",
 797                        bond->dev->name);
 798                 return -1;
 799         }
 800         _lock_rx_hashtbl(bond);
 801
 802         bond_info->rx_hashtbl = new_hashtbl;
 803
 804         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 805
 806         for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
 807                 rlb_init_table_entry(bond_info->rx_hashtbl + i);
 808         }
 809
 810         _unlock_rx_hashtbl(bond);
 811
 812         /*initialize packet type*/
 813         pk_type->type = __constant_htons(ETH_P_ARP);
 814         pk_type->dev = bond->dev;
 815         pk_type->func = rlb_arp_recv;
 816
 817         /* register to receive ARPs */
 818         dev_add_pack(pk_type);
 819
 820         return 0;
 821 }
 822
 823 static void rlb_deinitialize(struct bonding *bond)
 824 {
 825         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 826
 827         dev_remove_pack(&(bond_info->rlb_pkt_type));
 828
 829         _lock_rx_hashtbl(bond);
 830
 831         kfree(bond_info->rx_hashtbl);
 832         bond_info->rx_hashtbl = NULL;
 833         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 834
 835         _unlock_rx_hashtbl(bond);
 836 }
 837
 838 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
 839 {
 840         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 841         u32 curr_index;
 842
 843         _lock_rx_hashtbl(bond);
 844
 845         curr_index = bond_info->rx_hashtbl_head;
 846         while (curr_index != RLB_NULL_INDEX) {
 847                 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
 848                 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
 849                 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
 850
 851                 if (curr->tag && (curr->vlan_id == vlan_id)) {
 852                         if (curr_index == bond_info->rx_hashtbl_head) {
 853                                 bond_info->rx_hashtbl_head = next_index;
 854                         }
 855                         if (prev_index != RLB_NULL_INDEX) {
 856                                 bond_info->rx_hashtbl[prev_index].next = next_index;
 857                         }
 858                         if (next_index != RLB_NULL_INDEX) {
 859                                 bond_info->rx_hashtbl[next_index].prev = prev_index;
 860                         }
 861
 862                         rlb_init_table_entry(curr);
 863                 }
 864
 865                 curr_index = next_index;
 866         }
 867
 868         _unlock_rx_hashtbl(bond);
 869 }
 870
 871 /*********************** tlb/rlb shared functions *********************/
 872
 873 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
 874 {
 875         struct bonding *bond = bond_get_bond_by_slave(slave);
 876         struct learning_pkt pkt;
 877         int size = sizeof(struct learning_pkt);
 878         int i;
 879
 880         memset(&pkt, 0, size);
 881         memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
 882         memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
 883         pkt.type = __constant_htons(ETH_P_LOOP);
 884
 885         for (i = 0; i < MAX_LP_BURST; i++) {
 886                 struct sk_buff *skb;
 887                 char *data;
 888
 889                 skb = dev_alloc_skb(size);
 890                 if (!skb) {
 891                         return;
 892                 }
 893
 894                 data = skb_put(skb, size);
 895                 memcpy(data, &pkt, size);
 896
 897                 skb_reset_mac_header(skb);
 898                 skb->network_header = skb->mac_header + ETH_HLEN;
 899                 skb->protocol = pkt.type;
 900                 skb->priority = TC_PRIO_CONTROL;
 901                 skb->dev = slave->dev;
 902
 903                 if (!list_empty(&bond->vlan_list)) {
 904                         struct vlan_entry *vlan;
 905
 906                         vlan = bond_next_vlan(bond,
 907                                               bond->alb_info.current_alb_vlan);
 908
 909                         bond->alb_info.current_alb_vlan = vlan;
 910                         if (!vlan) {
 911                                 kfree_skb(skb);
 912                                 continue;
 913                         }
 914
 915                         skb = vlan_put_tag(skb, vlan->vlan_id);
 916                         if (!skb) {
 917                                 printk(KERN_ERR DRV_NAME
 918                                        ": %s: Error: failed to insert VLAN tag\n",
 919                                        bond->dev->name);
 920                                 continue;
 921                         }
 922                 }
 923
 924                 dev_queue_xmit(skb);
 925         }
 926 }
 927
 928 /* hw is a boolean parameter that determines whether we should try and
 929  * set the hw address of the device as well as the hw address of the
 930  * net_device
 931  */
 932 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[], int hw)
 933 {
 934         struct net_device *dev = slave->dev;
 935         struct sockaddr s_addr;
 936
 937         if (!hw) {
 938                 memcpy(dev->dev_addr, addr, dev->addr_len);
 939                 return 0;
 940         }
 941
 942         /* for rlb each slave must have a unique hw mac addresses so that */
 943         /* each slave will receive packets destined to a different mac */
 944         memcpy(s_addr.sa_data, addr, dev->addr_len);
 945         s_addr.sa_family = dev->type;
 946         if (dev_set_mac_address(dev, &s_addr)) {
 947                 printk(KERN_ERR DRV_NAME
 948                        ": %s: Error: dev_set_mac_address of dev %s failed! ALB "
 949                        "mode requires that the base driver support setting "
 950                        "the hw address also when the network device's "
 951                        "interface is open\n",
 952                        dev->master->name, dev->name);
 953                 return -EOPNOTSUPP;
 954         }
 955         return 0;
 956 }
 957
 958 /*
 959  * Swap MAC addresses between two slaves.
 960  *
 961  * Called with RTNL held, and no other locks.
 962  *
 963  */
 964
 965 static void alb_swap_mac_addr(struct bonding *bond, struct slave *slave1, struct slave *slave2)
 966 {
 967         u8 tmp_mac_addr[ETH_ALEN];
 968
 969         memcpy(tmp_mac_addr, slave1->dev->dev_addr, ETH_ALEN);
 970         alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr, bond->alb_info.rlb_enabled);
 971         alb_set_slave_mac_addr(slave2, tmp_mac_addr, bond->alb_info.rlb_enabled);
 972
 973 }
 974
 975 /*
 976  * Send learning packets after MAC address swap.
 977  *
 978  * Called with RTNL and no other locks
 979  */
 980 static void alb_fasten_mac_swap(struct bonding *bond, struct slave *slave1,
 981                                 struct slave *slave2)
 982 {
 983         int slaves_state_differ = (SLAVE_IS_OK(slave1) != SLAVE_IS_OK(slave2));
 984         struct slave *disabled_slave = NULL;
 985
 986         ASSERT_RTNL();
 987
 988         /* fasten the change in the switch */
 989         if (SLAVE_IS_OK(slave1)) {
 990                 alb_send_learning_packets(slave1, slave1->dev->dev_addr);
 991                 if (bond->alb_info.rlb_enabled) {
 992                         /* inform the clients that the mac address
 993                          * has changed
 994                          */
 995                         rlb_req_update_slave_clients(bond, slave1);
 996                 }
 997         } else {
 998                 disabled_slave = slave1;
 999         }
1000
1001         if (SLAVE_IS_OK(slave2)) {
1002                 alb_send_learning_packets(slave2, slave2->dev->dev_addr);
1003                 if (bond->alb_info.rlb_enabled) {
1004                         /* inform the clients that the mac address
1005                          * has changed
1006                          */
1007                         rlb_req_update_slave_clients(bond, slave2);
1008                 }
1009         } else {
1010                 disabled_slave = slave2;
1011         }
1012
1013         if (bond->alb_info.rlb_enabled && slaves_state_differ) {
1014                 /* A disabled slave was assigned an active mac addr */
1015                 rlb_teach_disabled_mac_on_primary(bond,
1016                                                   disabled_slave->dev->dev_addr);
1017         }
1018 }
1019
1020 /**
1021  * alb_change_hw_addr_on_detach
1022  * @bond: bonding we're working on
1023  * @slave: the slave that was just detached
1024  *
1025  * We assume that @slave was already detached from the slave list.
1026  *
1027  * If @slave's permanent hw address is different both from its current
1028  * address and from @bond's address, then somewhere in the bond there's
1029  * a slave that has @slave's permanet address as its current address.
1030  * We'll make sure that that slave no longer uses @slave's permanent address.
1031  *
1032  * Caller must hold RTNL and no other locks
1033  */
1034 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
1035 {
1036         int perm_curr_diff;
1037         int perm_bond_diff;
1038
1039         perm_curr_diff = memcmp(slave->perm_hwaddr,
1040                                 slave->dev->dev_addr,
1041                                 ETH_ALEN);
1042         perm_bond_diff = memcmp(slave->perm_hwaddr,
1043                                 bond->dev->dev_addr,
1044                                 ETH_ALEN);
1045
1046         if (perm_curr_diff && perm_bond_diff) {
1047                 struct slave *tmp_slave;
1048                 int i, found = 0;
1049
1050                 bond_for_each_slave(bond, tmp_slave, i) {
1051                         if (!memcmp(slave->perm_hwaddr,
1052                                     tmp_slave->dev->dev_addr,
1053                                     ETH_ALEN)) {
1054                                 found = 1;
1055                                 break;
1056                         }
1057                 }
1058
1059                 if (found) {
1060                         /* locking: needs RTNL and nothing else */
1061                         alb_swap_mac_addr(bond, slave, tmp_slave);
1062                         alb_fasten_mac_swap(bond, slave, tmp_slave);
1063                 }
1064         }
1065 }
1066
1067 /**
1068  * alb_handle_addr_collision_on_attach
1069  * @bond: bonding we're working on
1070  * @slave: the slave that was just attached
1071  *
1072  * checks uniqueness of slave's mac address and handles the case the
1073  * new slave uses the bonds mac address.
1074  *
1075  * If the permanent hw address of @slave is @bond's hw address, we need to
1076  * find a different hw address to give @slave, that isn't in use by any other
1077  * slave in the bond. This address must be, of course, one of the premanent
1078  * addresses of the other slaves.
1079  *
1080  * We go over the slave list, and for each slave there we compare its
1081  * permanent hw address with the current address of all the other slaves.
1082  * If no match was found, then we've found a slave with a permanent address
1083  * that isn't used by any other slave in the bond, so we can assign it to
1084  * @slave.
1085  *
1086  * assumption: this function is called before @slave is attached to the
1087  *             bond slave list.
1088  *
1089  * caller must hold the bond lock for write since the mac addresses are compared
1090  * and may be swapped.
1091  */
1092 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1093 {
1094         struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1095         struct slave *has_bond_addr = bond->curr_active_slave;
1096         int i, j, found = 0;
1097
1098         if (bond->slave_cnt == 0) {
1099                 /* this is the first slave */
1100                 return 0;
1101         }
1102
1103         /* if slave's mac address differs from bond's mac address
1104          * check uniqueness of slave's mac address against the other
1105          * slaves in the bond.
1106          */
1107         if (memcmp(slave->perm_hwaddr, bond->dev->dev_addr, ETH_ALEN)) {
1108                 bond_for_each_slave(bond, tmp_slave1, i) {
1109                         if (!memcmp(tmp_slave1->dev->dev_addr, slave->dev->dev_addr,
1110                                     ETH_ALEN)) {
1111                                 found = 1;
1112                                 break;
1113                         }
1114                 }
1115
1116                 if (!found)
1117                         return 0;
1118
1119                 /* Try setting slave mac to bond address and fall-through
1120                    to code handling that situation below... */
1121                 alb_set_slave_mac_addr(slave, bond->dev->dev_addr,
1122                                        bond->alb_info.rlb_enabled);
1123         }
1124
1125         /* The slave's address is equal to the address of the bond.
1126          * Search for a spare address in the bond for this slave.
1127          */
1128         free_mac_slave = NULL;
1129
1130         bond_for_each_slave(bond, tmp_slave1, i) {
1131                 found = 0;
1132                 bond_for_each_slave(bond, tmp_slave2, j) {
1133                         if (!memcmp(tmp_slave1->perm_hwaddr,
1134                                     tmp_slave2->dev->dev_addr,
1135                                     ETH_ALEN)) {
1136                                 found = 1;
1137                                 break;
1138                         }
1139                 }
1140
1141                 if (!found) {
1142                         /* no slave has tmp_slave1's perm addr
1143                          * as its curr addr
1144                          */
1145                         free_mac_slave = tmp_slave1;
1146                         break;
1147                 }
1148
1149                 if (!has_bond_addr) {
1150                         if (!memcmp(tmp_slave1->dev->dev_addr,
1151                                     bond->dev->dev_addr,
1152                                     ETH_ALEN)) {
1153
1154                                 has_bond_addr = tmp_slave1;
1155                         }
1156                 }
1157         }
1158
1159         if (free_mac_slave) {
1160                 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr,
1161                                        bond->alb_info.rlb_enabled);
1162
1163                 printk(KERN_WARNING DRV_NAME
1164                        ": %s: Warning: the hw address of slave %s is in use by "
1165                        "the bond; giving it the hw address of %s\n",
1166                        bond->dev->name, slave->dev->name, free_mac_slave->dev->name);
1167
1168         } else if (has_bond_addr) {
1169                 printk(KERN_ERR DRV_NAME
1170                        ": %s: Error: the hw address of slave %s is in use by the "
1171                        "bond; couldn't find a slave with a free hw address to "
1172                        "give it (this should not have happened)\n",
1173                        bond->dev->name, slave->dev->name);
1174                 return -EFAULT;
1175         }
1176
1177         return 0;
1178 }
1179
1180 /**
1181  * alb_set_mac_address
1182  * @bond:
1183  * @addr:
1184  *
1185  * In TLB mode all slaves are configured to the bond's hw address, but set
1186  * their dev_addr field to different addresses (based on their permanent hw
1187  * addresses).
1188  *
1189  * For each slave, this function sets the interface to the new address and then
1190  * changes its dev_addr field to its previous value.
1191  *
1192  * Unwinding assumes bond's mac address has not yet changed.
1193  */
1194 static int alb_set_mac_address(struct bonding *bond, void *addr)
1195 {
1196         struct sockaddr sa;
1197         struct slave *slave, *stop_at;
1198         char tmp_addr[ETH_ALEN];
1199         int res;
1200         int i;
1201
1202         if (bond->alb_info.rlb_enabled) {
1203                 return 0;
1204         }
1205
1206         bond_for_each_slave(bond, slave, i) {
1207                 if (slave->dev->set_mac_address == NULL) {
1208                         res = -EOPNOTSUPP;
1209                         goto unwind;
1210                 }
1211
1212                 /* save net_device's current hw address */
1213                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1214
1215                 res = dev_set_mac_address(slave->dev, addr);
1216
1217                 /* restore net_device's hw address */
1218                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1219
1220                 if (res) {
1221                         goto unwind;
1222                 }
1223         }
1224
1225         return 0;
1226
1227 unwind:
1228         memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1229         sa.sa_family = bond->dev->type;
1230
1231         /* unwind from head to the slave that failed */
1232         stop_at = slave;
1233         bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1234                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1235                 dev_set_mac_address(slave->dev, &sa);
1236                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1237         }
1238
1239         return res;
1240 }
1241
1242 /************************ exported alb funcions ************************/
1243
1244 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1245 {
1246         int res;
1247
1248         res = tlb_initialize(bond);
1249         if (res) {
1250                 return res;
1251         }
1252
1253         if (rlb_enabled) {
1254                 bond->alb_info.rlb_enabled = 1;
1255                 /* initialize rlb */
1256                 res = rlb_initialize(bond);
1257                 if (res) {
1258                         tlb_deinitialize(bond);
1259                         return res;
1260                 }
1261         } else {
1262                 bond->alb_info.rlb_enabled = 0;
1263         }
1264
1265         return 0;
1266 }
1267
1268 void bond_alb_deinitialize(struct bonding *bond)
1269 {
1270         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1271
1272         tlb_deinitialize(bond);
1273
1274         if (bond_info->rlb_enabled) {
1275                 rlb_deinitialize(bond);
1276         }
1277 }
1278
1279 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1280 {
1281         struct bonding *bond = bond_dev->priv;
1282         struct ethhdr *eth_data;
1283         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1284         struct slave *tx_slave = NULL;
1285         static const __be32 ip_bcast = htonl(0xffffffff);
1286         int hash_size = 0;
1287         int do_tx_balance = 1;
1288         u32 hash_index = 0;
1289         const u8 *hash_start = NULL;
1290         int res = 1;
1291
1292         skb_reset_mac_header(skb);
1293         eth_data = eth_hdr(skb);
1294
1295         /* make sure that the curr_active_slave and the slaves list do
1296          * not change during tx
1297          */
1298         read_lock(&bond->lock);
1299         read_lock(&bond->curr_slave_lock);
1300
1301         if (!BOND_IS_OK(bond)) {
1302                 goto out;
1303         }
1304
1305         switch (ntohs(skb->protocol)) {
1306         case ETH_P_IP: {
1307                 const struct iphdr *iph = ip_hdr(skb);
1308
1309                 if ((memcmp(eth_data->h_dest, mac_bcast, ETH_ALEN) == 0) ||
1310                     (iph->daddr == ip_bcast) ||
1311                     (iph->protocol == IPPROTO_IGMP)) {
1312                         do_tx_balance = 0;
1313                         break;
1314                 }
1315                 hash_start = (char *)&(iph->daddr);
1316                 hash_size = sizeof(iph->daddr);
1317         }
1318                 break;
1319         case ETH_P_IPV6:
1320                 if (memcmp(eth_data->h_dest, mac_bcast, ETH_ALEN) == 0) {
1321                         do_tx_balance = 0;
1322                         break;
1323                 }
1324
1325                 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1326                 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1327                 break;
1328         case ETH_P_IPX:
1329                 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1330                         /* something is wrong with this packet */
1331                         do_tx_balance = 0;
1332                         break;
1333                 }
1334
1335                 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1336                         /* The only protocol worth balancing in
1337                          * this family since it has an "ARP" like
1338                          * mechanism
1339                          */
1340                         do_tx_balance = 0;
1341                         break;
1342                 }
1343
1344                 hash_start = (char*)eth_data->h_dest;
1345                 hash_size = ETH_ALEN;
1346                 break;
1347         case ETH_P_ARP:
1348                 do_tx_balance = 0;
1349                 if (bond_info->rlb_enabled) {
1350                         tx_slave = rlb_arp_xmit(skb, bond);
1351                 }
1352                 break;
1353         default:
1354                 do_tx_balance = 0;
1355                 break;
1356         }
1357
1358         if (do_tx_balance) {
1359                 hash_index = _simple_hash(hash_start, hash_size);
1360                 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1361         }
1362
1363         if (!tx_slave) {
1364                 /* unbalanced or unassigned, send through primary */
1365                 tx_slave = bond->curr_active_slave;
1366                 bond_info->unbalanced_load += skb->len;
1367         }
1368
1369         if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1370                 if (tx_slave != bond->curr_active_slave) {
1371                         memcpy(eth_data->h_source,
1372                                tx_slave->dev->dev_addr,
1373                                ETH_ALEN);
1374                 }
1375
1376                 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1377         } else {
1378                 if (tx_slave) {
1379                         tlb_clear_slave(bond, tx_slave, 0);
1380                 }
1381         }
1382
1383 out:
1384         if (res) {
1385                 /* no suitable interface, frame not sent */
1386                 dev_kfree_skb(skb);
1387         }
1388         read_unlock(&bond->curr_slave_lock);
1389         read_unlock(&bond->lock);
1390         return 0;
1391 }
1392
1393 void bond_alb_monitor(struct work_struct *work)
1394 {
1395         struct bonding *bond = container_of(work, struct bonding,
1396                                             alb_work.work);
1397         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1398         struct slave *slave;
1399         int i;
1400
1401         read_lock(&bond->lock);
1402
1403         if (bond->kill_timers) {
1404                 goto out;
1405         }
1406
1407         if (bond->slave_cnt == 0) {
1408                 bond_info->tx_rebalance_counter = 0;
1409                 bond_info->lp_counter = 0;
1410                 goto re_arm;
1411         }
1412
1413         bond_info->tx_rebalance_counter++;
1414         bond_info->lp_counter++;
1415
1416         /* send learning packets */
1417         if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1418                 /* change of curr_active_slave involves swapping of mac addresses.
1419                  * in order to avoid this swapping from happening while
1420                  * sending the learning packets, the curr_slave_lock must be held for
1421                  * read.
1422                  */
1423                 read_lock(&bond->curr_slave_lock);
1424
1425                 bond_for_each_slave(bond, slave, i) {
1426                         alb_send_learning_packets(slave, slave->dev->dev_addr);
1427                 }
1428
1429                 read_unlock(&bond->curr_slave_lock);
1430
1431                 bond_info->lp_counter = 0;
1432         }
1433
1434         /* rebalance tx traffic */
1435         if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1436
1437                 read_lock(&bond->curr_slave_lock);
1438
1439                 bond_for_each_slave(bond, slave, i) {
1440                         tlb_clear_slave(bond, slave, 1);
1441                         if (slave == bond->curr_active_slave) {
1442                                 SLAVE_TLB_INFO(slave).load =
1443                                         bond_info->unbalanced_load /
1444                                                 BOND_TLB_REBALANCE_INTERVAL;
1445                                 bond_info->unbalanced_load = 0;
1446                         }
1447                 }
1448
1449                 read_unlock(&bond->curr_slave_lock);
1450
1451                 bond_info->tx_rebalance_counter = 0;
1452         }
1453
1454         /* handle rlb stuff */
1455         if (bond_info->rlb_enabled) {
1456                 if (bond_info->primary_is_promisc &&
1457                     (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1458
1459                         /*
1460                          * dev_set_promiscuity requires rtnl and
1461                          * nothing else.
1462                          */
1463                         read_unlock(&bond->lock);
1464                         rtnl_lock();
1465
1466                         bond_info->rlb_promisc_timeout_counter = 0;
1467
1468                         /* If the primary was set to promiscuous mode
1469                          * because a slave was disabled then
1470                          * it can now leave promiscuous mode.
1471                          */
1472                         dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1473                         bond_info->primary_is_promisc = 0;
1474
1475                         rtnl_unlock();
1476                         read_lock(&bond->lock);
1477                 }
1478
1479                 if (bond_info->rlb_rebalance) {
1480                         bond_info->rlb_rebalance = 0;
1481                         rlb_rebalance(bond);
1482                 }
1483
1484                 /* check if clients need updating */
1485                 if (bond_info->rx_ntt) {
1486                         if (bond_info->rlb_update_delay_counter) {
1487                                 --bond_info->rlb_update_delay_counter;
1488                         } else {
1489                                 rlb_update_rx_clients(bond);
1490                                 if (bond_info->rlb_update_retry_counter) {
1491                                         --bond_info->rlb_update_retry_counter;
1492                                 } else {
1493                                         bond_info->rx_ntt = 0;
1494                                 }
1495                         }
1496                 }
1497         }
1498
1499 re_arm:
1500         queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1501 out:
1502         read_unlock(&bond->lock);
1503 }
1504
1505 /* assumption: called before the slave is attached to the bond
1506  * and not locked by the bond lock
1507  */
1508 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1509 {
1510         int res;
1511
1512         res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr,
1513                                      bond->alb_info.rlb_enabled);
1514         if (res) {
1515                 return res;
1516         }
1517
1518         /* caller must hold the bond lock for write since the mac addresses
1519          * are compared and may be swapped.
1520          */
1521         read_lock(&bond->lock);
1522
1523         res = alb_handle_addr_collision_on_attach(bond, slave);
1524
1525         read_unlock(&bond->lock);
1526
1527         if (res) {
1528                 return res;
1529         }
1530
1531         tlb_init_slave(slave);
1532
1533         /* order a rebalance ASAP */
1534         bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1535
1536         if (bond->alb_info.rlb_enabled) {
1537                 bond->alb_info.rlb_rebalance = 1;
1538         }
1539
1540         return 0;
1541 }
1542
1543 /*
1544  * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1545  * if necessary.
1546  *
1547  * Caller must hold RTNL and no other locks
1548  */
1549 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1550 {
1551         if (bond->slave_cnt > 1) {
1552                 alb_change_hw_addr_on_detach(bond, slave);
1553         }
1554
1555         tlb_clear_slave(bond, slave, 0);
1556
1557         if (bond->alb_info.rlb_enabled) {
1558                 bond->alb_info.next_rx_slave = NULL;
1559                 rlb_clear_slave(bond, slave);
1560         }
1561 }
1562
1563 /* Caller must hold bond lock for read */
1564 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1565 {
1566         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1567
1568         if (link == BOND_LINK_DOWN) {
1569                 tlb_clear_slave(bond, slave, 0);
1570                 if (bond->alb_info.rlb_enabled) {
1571                         rlb_clear_slave(bond, slave);
1572                 }
1573         } else if (link == BOND_LINK_UP) {
1574                 /* order a rebalance ASAP */
1575                 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1576                 if (bond->alb_info.rlb_enabled) {
1577                         bond->alb_info.rlb_rebalance = 1;
1578                         /* If the updelay module parameter is smaller than the
1579                          * forwarding delay of the switch the rebalance will
1580                          * not work because the rebalance arp replies will
1581                          * not be forwarded to the clients..
1582                          */
1583                 }
1584         }
1585 }
1586
1587 /**
1588  * bond_alb_handle_active_change - assign new curr_active_slave
1589  * @bond: our bonding struct
1590  * @new_slave: new slave to assign
1591  *
1592  * Set the bond->curr_active_slave to @new_slave and handle
1593  * mac address swapping and promiscuity changes as needed.
1594  *
1595  * If new_slave is NULL, caller must hold curr_slave_lock or
1596  * bond->lock for write.
1597  *
1598  * If new_slave is not NULL, caller must hold RTNL, bond->lock for
1599  * read and curr_slave_lock for write.  Processing here may sleep, so
1600  * no other locks may be held.
1601  */
1602 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1603 {
1604         struct slave *swap_slave;
1605         int i;
1606
1607         if (bond->curr_active_slave == new_slave) {
1608                 return;
1609         }
1610
1611         if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1612                 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1613                 bond->alb_info.primary_is_promisc = 0;
1614                 bond->alb_info.rlb_promisc_timeout_counter = 0;
1615         }
1616
1617         swap_slave = bond->curr_active_slave;
1618         bond->curr_active_slave = new_slave;
1619
1620         if (!new_slave || (bond->slave_cnt == 0)) {
1621                 return;
1622         }
1623
1624         /* set the new curr_active_slave to the bonds mac address
1625          * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1626          */
1627         if (!swap_slave) {
1628                 struct slave *tmp_slave;
1629                 /* find slave that is holding the bond's mac address */
1630                 bond_for_each_slave(bond, tmp_slave, i) {
1631                         if (!memcmp(tmp_slave->dev->dev_addr,
1632                                     bond->dev->dev_addr, ETH_ALEN)) {
1633                                 swap_slave = tmp_slave;
1634                                 break;
1635                         }
1636                 }
1637         }
1638
1639         /*
1640          * Arrange for swap_slave and new_slave to temporarily be
1641          * ignored so we can mess with their MAC addresses without
1642          * fear of interference from transmit activity.
1643          */
1644         if (swap_slave) {
1645                 tlb_clear_slave(bond, swap_slave, 1);
1646         }
1647         tlb_clear_slave(bond, new_slave, 1);
1648
1649         write_unlock_bh(&bond->curr_slave_lock);
1650         read_unlock(&bond->lock);
1651
1652         ASSERT_RTNL();
1653
1654         /* curr_active_slave must be set before calling alb_swap_mac_addr */
1655         if (swap_slave) {
1656                 /* swap mac address */
1657                 alb_swap_mac_addr(bond, swap_slave, new_slave);
1658         } else {
1659                 /* set the new_slave to the bond mac address */
1660                 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr,
1661                                        bond->alb_info.rlb_enabled);
1662         }
1663
1664         if (swap_slave) {
1665                 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1666                 read_lock(&bond->lock);
1667         } else {
1668                 read_lock(&bond->lock);
1669                 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1670         }
1671
1672         write_lock_bh(&bond->curr_slave_lock);
1673 }
1674
1675 /*
1676  * Called with RTNL
1677  */
1678 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1679 {
1680         struct bonding *bond = bond_dev->priv;
1681         struct sockaddr *sa = addr;
1682         struct slave *slave, *swap_slave;
1683         int res;
1684         int i;
1685
1686         if (!is_valid_ether_addr(sa->sa_data)) {
1687                 return -EADDRNOTAVAIL;
1688         }
1689
1690         res = alb_set_mac_address(bond, addr);
1691         if (res) {
1692                 return res;
1693         }
1694
1695         memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1696
1697         /* If there is no curr_active_slave there is nothing else to do.
1698          * Otherwise we'll need to pass the new address to it and handle
1699          * duplications.
1700          */
1701         if (!bond->curr_active_slave) {
1702                 return 0;
1703         }
1704
1705         swap_slave = NULL;
1706
1707         bond_for_each_slave(bond, slave, i) {
1708                 if (!memcmp(slave->dev->dev_addr, bond_dev->dev_addr, ETH_ALEN)) {
1709                         swap_slave = slave;
1710                         break;
1711                 }
1712         }
1713
1714         write_unlock_bh(&bond->curr_slave_lock);
1715         read_unlock(&bond->lock);
1716
1717         if (swap_slave) {
1718                 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1719                 alb_fasten_mac_swap(bond, swap_slave, bond->curr_active_slave);
1720         } else {
1721                 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr,
1722                                        bond->alb_info.rlb_enabled);
1723
1724                 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1725                 if (bond->alb_info.rlb_enabled) {
1726                         /* inform clients mac address has changed */
1727                         rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1728                 }
1729         }
1730
1731         read_lock(&bond->lock);
1732         write_lock_bh(&bond->curr_slave_lock);
1733
1734         return 0;
1735 }
1736
1737 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1738 {
1739         if (bond->alb_info.current_alb_vlan &&
1740             (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1741                 bond->alb_info.current_alb_vlan = NULL;
1742         }
1743
1744         if (bond->alb_info.rlb_enabled) {
1745                 rlb_clear_vlan(bond, vlan_id);
1746         }
1747 }
1748