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 #include <linux/skbuff.h>
  24 #include <linux/netdevice.h>
  25 #include <linux/etherdevice.h>
  26 #include <linux/pkt_sched.h>
  27 #include <linux/spinlock.h>
  28 #include <linux/slab.h>
  29 #include <linux/timer.h>
  30 #include <linux/ip.h>
  31 #include <linux/ipv6.h>
  32 #include <linux/if_arp.h>
  33 #include <linux/if_ether.h>
  34 #include <linux/if_bonding.h>
  35 #include <linux/if_vlan.h>
  36 #include <linux/in.h>
  37 #include <net/ipx.h>
  38 #include <net/arp.h>
  39 #include <net/ipv6.h>
  40 #include <asm/byteorder.h>
  41 #include "bonding.h"
  42 #include "bond_alb.h"
  43
  44
  45 #define ALB_TIMER_TICKS_PER_SEC     10  /* should be a divisor of HZ */
  46 #define BOND_TLB_REBALANCE_INTERVAL 10  /* In seconds, periodic re-balancing.
  47                                          * Used for division - never set
  48                                          * to zero !!!
  49                                          */
  50 #define BOND_ALB_LP_INTERVAL        1   /* In seconds, periodic send of
  51                                          * learning packets to the switch
  52                                          */
  53
  54 #define BOND_TLB_REBALANCE_TICKS (BOND_TLB_REBALANCE_INTERVAL \
  55                                   * ALB_TIMER_TICKS_PER_SEC)
  56
  57 #define BOND_ALB_LP_TICKS (BOND_ALB_LP_INTERVAL \
  58                            * ALB_TIMER_TICKS_PER_SEC)
  59
  60 #define TLB_HASH_TABLE_SIZE 256 /* The size of the clients hash table.
  61                                  * Note that this value MUST NOT be smaller
  62                                  * because the key hash table is BYTE wide !
  63                                  */
  64
  65
  66 #define TLB_NULL_INDEX          0xffffffff
  67 #define MAX_LP_BURST            3
  68
  69 /* rlb defs */
  70 #define RLB_HASH_TABLE_SIZE     256
  71 #define RLB_NULL_INDEX          0xffffffff
  72 #define RLB_UPDATE_DELAY        2*ALB_TIMER_TICKS_PER_SEC /* 2 seconds */
  73 #define RLB_ARP_BURST_SIZE      2
  74 #define RLB_UPDATE_RETRY        3       /* 3-ticks - must be smaller than the rlb
  75                                          * rebalance interval (5 min).
  76                                          */
  77 /* RLB_PROMISC_TIMEOUT = 10 sec equals the time that the current slave is
  78  * promiscuous after failover
  79  */
  80 #define RLB_PROMISC_TIMEOUT     10*ALB_TIMER_TICKS_PER_SEC
  81
  82 static const u8 mac_bcast[ETH_ALEN] = {0xff,0xff,0xff,0xff,0xff,0xff};
  83 static const u8 mac_v6_allmcast[ETH_ALEN] = {0x33,0x33,0x00,0x00,0x00,0x01};
  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         /* skip this if we've already freed the tx hash table */
 171         if (tx_hash_table) {
 172                 index = SLAVE_TLB_INFO(slave).head;
 173                 while (index != TLB_NULL_INDEX) {
 174                         u32 next_index = tx_hash_table[index].next;
 175                         tlb_init_table_entry(&tx_hash_table[index], save_load);
 176                         index = next_index;
 177                 }
 178         }
 179
 180         tlb_init_slave(slave);
 181
 182         _unlock_tx_hashtbl(bond);
 183 }
 184
 185 /* Must be called before starting the monitor timer */
 186 static int tlb_initialize(struct bonding *bond)
 187 {
 188         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 189         int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
 190         struct tlb_client_info *new_hashtbl;
 191         int i;
 192
 193         spin_lock_init(&(bond_info->tx_hashtbl_lock));
 194
 195         new_hashtbl = kzalloc(size, GFP_KERNEL);
 196         if (!new_hashtbl) {
 197                 printk(KERN_ERR DRV_NAME
 198                        ": %s: Error: Failed to allocate TLB hash table\n",
 199                        bond->dev->name);
 200                 return -1;
 201         }
 202         _lock_tx_hashtbl(bond);
 203
 204         bond_info->tx_hashtbl = new_hashtbl;
 205
 206         for (i = 0; i < TLB_HASH_TABLE_SIZE; i++) {
 207                 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 1);
 208         }
 209
 210         _unlock_tx_hashtbl(bond);
 211
 212         return 0;
 213 }
 214
 215 /* Must be called only after all slaves have been released */
 216 static void tlb_deinitialize(struct bonding *bond)
 217 {
 218         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 219
 220         _lock_tx_hashtbl(bond);
 221
 222         kfree(bond_info->tx_hashtbl);
 223         bond_info->tx_hashtbl = NULL;
 224
 225         _unlock_tx_hashtbl(bond);
 226 }
 227
 228 /* Caller must hold bond lock for read */
 229 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
 230 {
 231         struct slave *slave, *least_loaded;
 232         s64 max_gap;
 233         int i, found = 0;
 234
 235         /* Find the first enabled slave */
 236         bond_for_each_slave(bond, slave, i) {
 237                 if (SLAVE_IS_OK(slave)) {
 238                         found = 1;
 239                         break;
 240                 }
 241         }
 242
 243         if (!found) {
 244                 return NULL;
 245         }
 246
 247         least_loaded = slave;
 248         max_gap = (s64)(slave->speed << 20) - /* Convert to Megabit per sec */
 249                         (s64)(SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
 250
 251         /* Find the slave with the largest gap */
 252         bond_for_each_slave_from(bond, slave, i, least_loaded) {
 253                 if (SLAVE_IS_OK(slave)) {
 254                         s64 gap = (s64)(slave->speed << 20) -
 255                                         (s64)(SLAVE_TLB_INFO(slave).load << 3);
 256                         if (max_gap < gap) {
 257                                 least_loaded = slave;
 258                                 max_gap = gap;
 259                         }
 260                 }
 261         }
 262
 263         return least_loaded;
 264 }
 265
 266 /* Caller must hold bond lock for read */
 267 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index, u32 skb_len)
 268 {
 269         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 270         struct tlb_client_info *hash_table;
 271         struct slave *assigned_slave;
 272
 273         _lock_tx_hashtbl(bond);
 274
 275         hash_table = bond_info->tx_hashtbl;
 276         assigned_slave = hash_table[hash_index].tx_slave;
 277         if (!assigned_slave) {
 278                 assigned_slave = tlb_get_least_loaded_slave(bond);
 279
 280                 if (assigned_slave) {
 281                         struct tlb_slave_info *slave_info =
 282                                 &(SLAVE_TLB_INFO(assigned_slave));
 283                         u32 next_index = slave_info->head;
 284
 285                         hash_table[hash_index].tx_slave = assigned_slave;
 286                         hash_table[hash_index].next = next_index;
 287                         hash_table[hash_index].prev = TLB_NULL_INDEX;
 288
 289                         if (next_index != TLB_NULL_INDEX) {
 290                                 hash_table[next_index].prev = hash_index;
 291                         }
 292
 293                         slave_info->head = hash_index;
 294                         slave_info->load +=
 295                                 hash_table[hash_index].load_history;
 296                 }
 297         }
 298
 299         if (assigned_slave) {
 300                 hash_table[hash_index].tx_bytes += skb_len;
 301         }
 302
 303         _unlock_tx_hashtbl(bond);
 304
 305         return assigned_slave;
 306 }
 307
 308 /*********************** rlb specific functions ***************************/
 309 static inline void _lock_rx_hashtbl(struct bonding *bond)
 310 {
 311         spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 312 }
 313
 314 static inline void _unlock_rx_hashtbl(struct bonding *bond)
 315 {
 316         spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 317 }
 318
 319 /* when an ARP REPLY is received from a client update its info
 320  * in the rx_hashtbl
 321  */
 322 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
 323 {
 324         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 325         struct rlb_client_info *client_info;
 326         u32 hash_index;
 327
 328         _lock_rx_hashtbl(bond);
 329
 330         hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
 331         client_info = &(bond_info->rx_hashtbl[hash_index]);
 332
 333         if ((client_info->assigned) &&
 334             (client_info->ip_src == arp->ip_dst) &&
 335             (client_info->ip_dst == arp->ip_src)) {
 336                 /* update the clients MAC address */
 337                 memcpy(client_info->mac_dst, arp->mac_src, ETH_ALEN);
 338                 client_info->ntt = 1;
 339                 bond_info->rx_ntt = 1;
 340         }
 341
 342         _unlock_rx_hashtbl(bond);
 343 }
 344
 345 static int rlb_arp_recv(struct sk_buff *skb, struct net_device *bond_dev, struct packet_type *ptype, struct net_device *orig_dev)
 346 {
 347         struct bonding *bond;
 348         struct arp_pkt *arp = (struct arp_pkt *)skb->data;
 349         int res = NET_RX_DROP;
 350
 351         if (dev_net(bond_dev) != &init_net)
 352                 goto out;
 353
 354         while (bond_dev->priv_flags & IFF_802_1Q_VLAN)
 355                 bond_dev = vlan_dev_real_dev(bond_dev);
 356
 357         if (!(bond_dev->priv_flags & IFF_BONDING) ||
 358             !(bond_dev->flags & IFF_MASTER))
 359                 goto out;
 360
 361         if (!arp) {
 362                 pr_debug("Packet has no ARP data\n");
 363                 goto out;
 364         }
 365
 366         if (skb->len < sizeof(struct arp_pkt)) {
 367                 pr_debug("Packet is too small to be an ARP\n");
 368                 goto out;
 369         }
 370
 371         if (arp->op_code == htons(ARPOP_REPLY)) {
 372                 /* update rx hash table for this ARP */
 373                 printk("rar: update orig %s bond_dev %s\n", orig_dev->name,
 374                        bond_dev->name);
 375                 bond = netdev_priv(bond_dev);
 376                 rlb_update_entry_from_arp(bond, arp);
 377                 pr_debug("Server received an ARP Reply from client\n");
 378         }
 379
 380         res = NET_RX_SUCCESS;
 381
 382 out:
 383         dev_kfree_skb(skb);
 384
 385         return res;
 386 }
 387
 388 /* Caller must hold bond lock for read */
 389 static struct slave *rlb_next_rx_slave(struct bonding *bond)
 390 {
 391         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 392         struct slave *rx_slave, *slave, *start_at;
 393         int i = 0;
 394
 395         if (bond_info->next_rx_slave) {
 396                 start_at = bond_info->next_rx_slave;
 397         } else {
 398                 start_at = bond->first_slave;
 399         }
 400
 401         rx_slave = NULL;
 402
 403         bond_for_each_slave_from(bond, slave, i, start_at) {
 404                 if (SLAVE_IS_OK(slave)) {
 405                         if (!rx_slave) {
 406                                 rx_slave = slave;
 407                         } else if (slave->speed > rx_slave->speed) {
 408                                 rx_slave = slave;
 409                         }
 410                 }
 411         }
 412
 413         if (rx_slave) {
 414                 bond_info->next_rx_slave = rx_slave->next;
 415         }
 416
 417         return rx_slave;
 418 }
 419
 420 /* teach the switch the mac of a disabled slave
 421  * on the primary for fault tolerance
 422  *
 423  * Caller must hold bond->curr_slave_lock for write or bond lock for write
 424  */
 425 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
 426 {
 427         if (!bond->curr_active_slave) {
 428                 return;
 429         }
 430
 431         if (!bond->alb_info.primary_is_promisc) {
 432                 if (!dev_set_promiscuity(bond->curr_active_slave->dev, 1))
 433                         bond->alb_info.primary_is_promisc = 1;
 434                 else
 435                         bond->alb_info.primary_is_promisc = 0;
 436         }
 437
 438         bond->alb_info.rlb_promisc_timeout_counter = 0;
 439
 440         alb_send_learning_packets(bond->curr_active_slave, addr);
 441 }
 442
 443 /* slave being removed should not be active at this point
 444  *
 445  * Caller must hold bond lock for read
 446  */
 447 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
 448 {
 449         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 450         struct rlb_client_info *rx_hash_table;
 451         u32 index, next_index;
 452
 453         /* clear slave from rx_hashtbl */
 454         _lock_rx_hashtbl(bond);
 455
 456         rx_hash_table = bond_info->rx_hashtbl;
 457         index = bond_info->rx_hashtbl_head;
 458         for (; index != RLB_NULL_INDEX; index = next_index) {
 459                 next_index = rx_hash_table[index].next;
 460                 if (rx_hash_table[index].slave == slave) {
 461                         struct slave *assigned_slave = rlb_next_rx_slave(bond);
 462
 463                         if (assigned_slave) {
 464                                 rx_hash_table[index].slave = assigned_slave;
 465                                 if (memcmp(rx_hash_table[index].mac_dst,
 466                                            mac_bcast, ETH_ALEN)) {
 467                                         bond_info->rx_hashtbl[index].ntt = 1;
 468                                         bond_info->rx_ntt = 1;
 469                                         /* A slave has been removed from the
 470                                          * table because it is either disabled
 471                                          * or being released. We must retry the
 472                                          * update to avoid clients from not
 473                                          * being updated & disconnecting when
 474                                          * there is stress
 475                                          */
 476                                         bond_info->rlb_update_retry_counter =
 477                                                 RLB_UPDATE_RETRY;
 478                                 }
 479                         } else {  /* there is no active slave */
 480                                 rx_hash_table[index].slave = NULL;
 481                         }
 482                 }
 483         }
 484
 485         _unlock_rx_hashtbl(bond);
 486
 487         write_lock_bh(&bond->curr_slave_lock);
 488
 489         if (slave != bond->curr_active_slave) {
 490                 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
 491         }
 492
 493         write_unlock_bh(&bond->curr_slave_lock);
 494 }
 495
 496 static void rlb_update_client(struct rlb_client_info *client_info)
 497 {
 498         int i;
 499
 500         if (!client_info->slave) {
 501                 return;
 502         }
 503
 504         for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
 505                 struct sk_buff *skb;
 506
 507                 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
 508                                  client_info->ip_dst,
 509                                  client_info->slave->dev,
 510                                  client_info->ip_src,
 511                                  client_info->mac_dst,
 512                                  client_info->slave->dev->dev_addr,
 513                                  client_info->mac_dst);
 514                 if (!skb) {
 515                         printk(KERN_ERR DRV_NAME
 516                                ": %s: Error: failed to create an ARP packet\n",
 517                                client_info->slave->dev->master->name);
 518                         continue;
 519                 }
 520
 521                 skb->dev = client_info->slave->dev;
 522
 523                 if (client_info->tag) {
 524                         skb = vlan_put_tag(skb, client_info->vlan_id);
 525                         if (!skb) {
 526                                 printk(KERN_ERR DRV_NAME
 527                                        ": %s: Error: failed to insert VLAN tag\n",
 528                                        client_info->slave->dev->master->name);
 529                                 continue;
 530                         }
 531                 }
 532
 533                 arp_xmit(skb);
 534         }
 535 }
 536
 537 /* sends ARP REPLIES that update the clients that need updating */
 538 static void rlb_update_rx_clients(struct bonding *bond)
 539 {
 540         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 541         struct rlb_client_info *client_info;
 542         u32 hash_index;
 543
 544         _lock_rx_hashtbl(bond);
 545
 546         hash_index = bond_info->rx_hashtbl_head;
 547         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 548                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 549                 if (client_info->ntt) {
 550                         rlb_update_client(client_info);
 551                         if (bond_info->rlb_update_retry_counter == 0) {
 552                                 client_info->ntt = 0;
 553                         }
 554                 }
 555         }
 556
 557         /* do not update the entries again untill this counter is zero so that
 558          * not to confuse the clients.
 559          */
 560         bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
 561
 562         _unlock_rx_hashtbl(bond);
 563 }
 564
 565 /* The slave was assigned a new mac address - update the clients */
 566 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
 567 {
 568         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 569         struct rlb_client_info *client_info;
 570         int ntt = 0;
 571         u32 hash_index;
 572
 573         _lock_rx_hashtbl(bond);
 574
 575         hash_index = bond_info->rx_hashtbl_head;
 576         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 577                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 578
 579                 if ((client_info->slave == slave) &&
 580                     memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 581                         client_info->ntt = 1;
 582                         ntt = 1;
 583                 }
 584         }
 585
 586         // update the team's flag only after the whole iteration
 587         if (ntt) {
 588                 bond_info->rx_ntt = 1;
 589                 //fasten the change
 590                 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
 591         }
 592
 593         _unlock_rx_hashtbl(bond);
 594 }
 595
 596 /* mark all clients using src_ip to be updated */
 597 static void rlb_req_update_subnet_clients(struct bonding *bond, __be32 src_ip)
 598 {
 599         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 600         struct rlb_client_info *client_info;
 601         u32 hash_index;
 602
 603         _lock_rx_hashtbl(bond);
 604
 605         hash_index = bond_info->rx_hashtbl_head;
 606         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 607                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 608
 609                 if (!client_info->slave) {
 610                         printk(KERN_ERR DRV_NAME
 611                                ": %s: Error: found a client with no channel in "
 612                                "the client's hash table\n",
 613                                bond->dev->name);
 614                         continue;
 615                 }
 616                 /*update all clients using this src_ip, that are not assigned
 617                  * to the team's address (curr_active_slave) and have a known
 618                  * unicast mac address.
 619                  */
 620                 if ((client_info->ip_src == src_ip) &&
 621                     memcmp(client_info->slave->dev->dev_addr,
 622                            bond->dev->dev_addr, ETH_ALEN) &&
 623                     memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 624                         client_info->ntt = 1;
 625                         bond_info->rx_ntt = 1;
 626                 }
 627         }
 628
 629         _unlock_rx_hashtbl(bond);
 630 }
 631
 632 /* Caller must hold both bond and ptr locks for read */
 633 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
 634 {
 635         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 636         struct arp_pkt *arp = arp_pkt(skb);
 637         struct slave *assigned_slave;
 638         struct rlb_client_info *client_info;
 639         u32 hash_index = 0;
 640
 641         _lock_rx_hashtbl(bond);
 642
 643         hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_src));
 644         client_info = &(bond_info->rx_hashtbl[hash_index]);
 645
 646         if (client_info->assigned) {
 647                 if ((client_info->ip_src == arp->ip_src) &&
 648                     (client_info->ip_dst == arp->ip_dst)) {
 649                         /* the entry is already assigned to this client */
 650                         if (memcmp(arp->mac_dst, mac_bcast, ETH_ALEN)) {
 651                                 /* update mac address from arp */
 652                                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 653                         }
 654
 655                         assigned_slave = client_info->slave;
 656                         if (assigned_slave) {
 657                                 _unlock_rx_hashtbl(bond);
 658                                 return assigned_slave;
 659                         }
 660                 } else {
 661                         /* the entry is already assigned to some other client,
 662                          * move the old client to primary (curr_active_slave) so
 663                          * that the new client can be assigned to this entry.
 664                          */
 665                         if (bond->curr_active_slave &&
 666                             client_info->slave != bond->curr_active_slave) {
 667                                 client_info->slave = bond->curr_active_slave;
 668                                 rlb_update_client(client_info);
 669                         }
 670                 }
 671         }
 672         /* assign a new slave */
 673         assigned_slave = rlb_next_rx_slave(bond);
 674
 675         if (assigned_slave) {
 676                 client_info->ip_src = arp->ip_src;
 677                 client_info->ip_dst = arp->ip_dst;
 678                 /* arp->mac_dst is broadcast for arp reqeusts.
 679                  * will be updated with clients actual unicast mac address
 680                  * upon receiving an arp reply.
 681                  */
 682                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 683                 client_info->slave = assigned_slave;
 684
 685                 if (memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
 686                         client_info->ntt = 1;
 687                         bond->alb_info.rx_ntt = 1;
 688                 } else {
 689                         client_info->ntt = 0;
 690                 }
 691
 692                 if (!list_empty(&bond->vlan_list)) {
 693                         if (!vlan_get_tag(skb, &client_info->vlan_id))
 694                                 client_info->tag = 1;
 695                 }
 696
 697                 if (!client_info->assigned) {
 698                         u32 prev_tbl_head = bond_info->rx_hashtbl_head;
 699                         bond_info->rx_hashtbl_head = hash_index;
 700                         client_info->next = prev_tbl_head;
 701                         if (prev_tbl_head != RLB_NULL_INDEX) {
 702                                 bond_info->rx_hashtbl[prev_tbl_head].prev =
 703                                         hash_index;
 704                         }
 705                         client_info->assigned = 1;
 706                 }
 707         }
 708
 709         _unlock_rx_hashtbl(bond);
 710
 711         return assigned_slave;
 712 }
 713
 714 /* chooses (and returns) transmit channel for arp reply
 715  * does not choose channel for other arp types since they are
 716  * sent on the curr_active_slave
 717  */
 718 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
 719 {
 720         struct arp_pkt *arp = arp_pkt(skb);
 721         struct slave *tx_slave = NULL;
 722
 723         if (arp->op_code == htons(ARPOP_REPLY)) {
 724                 /* the arp must be sent on the selected
 725                 * rx channel
 726                 */
 727                 tx_slave = rlb_choose_channel(skb, bond);
 728                 if (tx_slave) {
 729                         memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
 730                 }
 731                 pr_debug("Server sent ARP Reply packet\n");
 732         } else if (arp->op_code == htons(ARPOP_REQUEST)) {
 733                 /* Create an entry in the rx_hashtbl for this client as a
 734                  * place holder.
 735                  * When the arp reply is received the entry will be updated
 736                  * with the correct unicast address of the client.
 737                  */
 738                 rlb_choose_channel(skb, bond);
 739
 740                 /* The ARP relpy packets must be delayed so that
 741                  * they can cancel out the influence of the ARP request.
 742                  */
 743                 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
 744
 745                 /* arp requests are broadcast and are sent on the primary
 746                  * the arp request will collapse all clients on the subnet to
 747                  * the primary slave. We must register these clients to be
 748                  * updated with their assigned mac.
 749                  */
 750                 rlb_req_update_subnet_clients(bond, arp->ip_src);
 751                 pr_debug("Server sent ARP Request packet\n");
 752         }
 753
 754         return tx_slave;
 755 }
 756
 757 /* Caller must hold bond lock for read */
 758 static void rlb_rebalance(struct bonding *bond)
 759 {
 760         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 761         struct slave *assigned_slave;
 762         struct rlb_client_info *client_info;
 763         int ntt;
 764         u32 hash_index;
 765
 766         _lock_rx_hashtbl(bond);
 767
 768         ntt = 0;
 769         hash_index = bond_info->rx_hashtbl_head;
 770         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 771                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 772                 assigned_slave = rlb_next_rx_slave(bond);
 773                 if (assigned_slave && (client_info->slave != assigned_slave)) {
 774                         client_info->slave = assigned_slave;
 775                         client_info->ntt = 1;
 776                         ntt = 1;
 777                 }
 778         }
 779
 780         /* update the team's flag only after the whole iteration */
 781         if (ntt) {
 782                 bond_info->rx_ntt = 1;
 783         }
 784         _unlock_rx_hashtbl(bond);
 785 }
 786
 787 /* Caller must hold rx_hashtbl lock */
 788 static void rlb_init_table_entry(struct rlb_client_info *entry)
 789 {
 790         memset(entry, 0, sizeof(struct rlb_client_info));
 791         entry->next = RLB_NULL_INDEX;
 792         entry->prev = RLB_NULL_INDEX;
 793 }
 794
 795 static int rlb_initialize(struct bonding *bond)
 796 {
 797         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 798         struct packet_type *pk_type = &(BOND_ALB_INFO(bond).rlb_pkt_type);
 799         struct rlb_client_info  *new_hashtbl;
 800         int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
 801         int i;
 802
 803         spin_lock_init(&(bond_info->rx_hashtbl_lock));
 804
 805         new_hashtbl = kmalloc(size, GFP_KERNEL);
 806         if (!new_hashtbl) {
 807                 printk(KERN_ERR DRV_NAME
 808                        ": %s: Error: Failed to allocate RLB hash table\n",
 809                        bond->dev->name);
 810                 return -1;
 811         }
 812         _lock_rx_hashtbl(bond);
 813
 814         bond_info->rx_hashtbl = new_hashtbl;
 815
 816         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 817
 818         for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
 819                 rlb_init_table_entry(bond_info->rx_hashtbl + i);
 820         }
 821
 822         _unlock_rx_hashtbl(bond);
 823
 824         /*initialize packet type*/
 825         pk_type->type = cpu_to_be16(ETH_P_ARP);
 826         pk_type->dev = NULL;
 827         pk_type->func = rlb_arp_recv;
 828
 829         /* register to receive ARPs */
 830         dev_add_pack(pk_type);
 831
 832         return 0;
 833 }
 834
 835 static void rlb_deinitialize(struct bonding *bond)
 836 {
 837         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 838
 839         dev_remove_pack(&(bond_info->rlb_pkt_type));
 840
 841         _lock_rx_hashtbl(bond);
 842
 843         kfree(bond_info->rx_hashtbl);
 844         bond_info->rx_hashtbl = NULL;
 845         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 846
 847         _unlock_rx_hashtbl(bond);
 848 }
 849
 850 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
 851 {
 852         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 853         u32 curr_index;
 854
 855         _lock_rx_hashtbl(bond);
 856
 857         curr_index = bond_info->rx_hashtbl_head;
 858         while (curr_index != RLB_NULL_INDEX) {
 859                 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
 860                 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
 861                 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
 862
 863                 if (curr->tag && (curr->vlan_id == vlan_id)) {
 864                         if (curr_index == bond_info->rx_hashtbl_head) {
 865                                 bond_info->rx_hashtbl_head = next_index;
 866                         }
 867                         if (prev_index != RLB_NULL_INDEX) {
 868                                 bond_info->rx_hashtbl[prev_index].next = next_index;
 869                         }
 870                         if (next_index != RLB_NULL_INDEX) {
 871                                 bond_info->rx_hashtbl[next_index].prev = prev_index;
 872                         }
 873
 874                         rlb_init_table_entry(curr);
 875                 }
 876
 877                 curr_index = next_index;
 878         }
 879
 880         _unlock_rx_hashtbl(bond);
 881 }
 882
 883 /*********************** tlb/rlb shared functions *********************/
 884
 885 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
 886 {
 887         struct bonding *bond = bond_get_bond_by_slave(slave);
 888         struct learning_pkt pkt;
 889         int size = sizeof(struct learning_pkt);
 890         int i;
 891
 892         memset(&pkt, 0, size);
 893         memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
 894         memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
 895         pkt.type = cpu_to_be16(ETH_P_LOOP);
 896
 897         for (i = 0; i < MAX_LP_BURST; i++) {
 898                 struct sk_buff *skb;
 899                 char *data;
 900
 901                 skb = dev_alloc_skb(size);
 902                 if (!skb) {
 903                         return;
 904                 }
 905
 906                 data = skb_put(skb, size);
 907                 memcpy(data, &pkt, size);
 908
 909                 skb_reset_mac_header(skb);
 910                 skb->network_header = skb->mac_header + ETH_HLEN;
 911                 skb->protocol = pkt.type;
 912                 skb->priority = TC_PRIO_CONTROL;
 913                 skb->dev = slave->dev;
 914
 915                 if (!list_empty(&bond->vlan_list)) {
 916                         struct vlan_entry *vlan;
 917
 918                         vlan = bond_next_vlan(bond,
 919                                               bond->alb_info.current_alb_vlan);
 920
 921                         bond->alb_info.current_alb_vlan = vlan;
 922                         if (!vlan) {
 923                                 kfree_skb(skb);
 924                                 continue;
 925                         }
 926
 927                         skb = vlan_put_tag(skb, vlan->vlan_id);
 928                         if (!skb) {
 929                                 printk(KERN_ERR DRV_NAME
 930                                        ": %s: Error: failed to insert VLAN tag\n",
 931                                        bond->dev->name);
 932                                 continue;
 933                         }
 934                 }
 935
 936                 dev_queue_xmit(skb);
 937         }
 938 }
 939
 940 /* hw is a boolean parameter that determines whether we should try and
 941  * set the hw address of the device as well as the hw address of the
 942  * net_device
 943  */
 944 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[], int hw)
 945 {
 946         struct net_device *dev = slave->dev;
 947         struct sockaddr s_addr;
 948
 949         if (!hw) {
 950                 memcpy(dev->dev_addr, addr, dev->addr_len);
 951                 return 0;
 952         }
 953
 954         /* for rlb each slave must have a unique hw mac addresses so that */
 955         /* each slave will receive packets destined to a different mac */
 956         memcpy(s_addr.sa_data, addr, dev->addr_len);
 957         s_addr.sa_family = dev->type;
 958         if (dev_set_mac_address(dev, &s_addr)) {
 959                 printk(KERN_ERR DRV_NAME
 960                        ": %s: Error: dev_set_mac_address of dev %s failed! ALB "
 961                        "mode requires that the base driver support setting "
 962                        "the hw address also when the network device's "
 963                        "interface is open\n",
 964                        dev->master->name, dev->name);
 965                 return -EOPNOTSUPP;
 966         }
 967         return 0;
 968 }
 969
 970 /*
 971  * Swap MAC addresses between two slaves.
 972  *
 973  * Called with RTNL held, and no other locks.
 974  *
 975  */
 976
 977 static void alb_swap_mac_addr(struct bonding *bond, struct slave *slave1, struct slave *slave2)
 978 {
 979         u8 tmp_mac_addr[ETH_ALEN];
 980
 981         memcpy(tmp_mac_addr, slave1->dev->dev_addr, ETH_ALEN);
 982         alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr, bond->alb_info.rlb_enabled);
 983         alb_set_slave_mac_addr(slave2, tmp_mac_addr, bond->alb_info.rlb_enabled);
 984
 985 }
 986
 987 /*
 988  * Send learning packets after MAC address swap.
 989  *
 990  * Called with RTNL and no other locks
 991  */
 992 static void alb_fasten_mac_swap(struct bonding *bond, struct slave *slave1,
 993                                 struct slave *slave2)
 994 {
 995         int slaves_state_differ = (SLAVE_IS_OK(slave1) != SLAVE_IS_OK(slave2));
 996         struct slave *disabled_slave = NULL;
 997
 998         ASSERT_RTNL();
 999
1000         /* fasten the change in the switch */
1001         if (SLAVE_IS_OK(slave1)) {
1002                 alb_send_learning_packets(slave1, slave1->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, slave1);
1008                 }
1009         } else {
1010                 disabled_slave = slave1;
1011         }
1012
1013         if (SLAVE_IS_OK(slave2)) {
1014                 alb_send_learning_packets(slave2, slave2->dev->dev_addr);
1015                 if (bond->alb_info.rlb_enabled) {
1016                         /* inform the clients that the mac address
1017                          * has changed
1018                          */
1019                         rlb_req_update_slave_clients(bond, slave2);
1020                 }
1021         } else {
1022                 disabled_slave = slave2;
1023         }
1024
1025         if (bond->alb_info.rlb_enabled && slaves_state_differ) {
1026                 /* A disabled slave was assigned an active mac addr */
1027                 rlb_teach_disabled_mac_on_primary(bond,
1028                                                   disabled_slave->dev->dev_addr);
1029         }
1030 }
1031
1032 /**
1033  * alb_change_hw_addr_on_detach
1034  * @bond: bonding we're working on
1035  * @slave: the slave that was just detached
1036  *
1037  * We assume that @slave was already detached from the slave list.
1038  *
1039  * If @slave's permanent hw address is different both from its current
1040  * address and from @bond's address, then somewhere in the bond there's
1041  * a slave that has @slave's permanet address as its current address.
1042  * We'll make sure that that slave no longer uses @slave's permanent address.
1043  *
1044  * Caller must hold RTNL and no other locks
1045  */
1046 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
1047 {
1048         int perm_curr_diff;
1049         int perm_bond_diff;
1050
1051         perm_curr_diff = memcmp(slave->perm_hwaddr,
1052                                 slave->dev->dev_addr,
1053                                 ETH_ALEN);
1054         perm_bond_diff = memcmp(slave->perm_hwaddr,
1055                                 bond->dev->dev_addr,
1056                                 ETH_ALEN);
1057
1058         if (perm_curr_diff && perm_bond_diff) {
1059                 struct slave *tmp_slave;
1060                 int i, found = 0;
1061
1062                 bond_for_each_slave(bond, tmp_slave, i) {
1063                         if (!memcmp(slave->perm_hwaddr,
1064                                     tmp_slave->dev->dev_addr,
1065                                     ETH_ALEN)) {
1066                                 found = 1;
1067                                 break;
1068                         }
1069                 }
1070
1071                 if (found) {
1072                         /* locking: needs RTNL and nothing else */
1073                         alb_swap_mac_addr(bond, slave, tmp_slave);
1074                         alb_fasten_mac_swap(bond, slave, tmp_slave);
1075                 }
1076         }
1077 }
1078
1079 /**
1080  * alb_handle_addr_collision_on_attach
1081  * @bond: bonding we're working on
1082  * @slave: the slave that was just attached
1083  *
1084  * checks uniqueness of slave's mac address and handles the case the
1085  * new slave uses the bonds mac address.
1086  *
1087  * If the permanent hw address of @slave is @bond's hw address, we need to
1088  * find a different hw address to give @slave, that isn't in use by any other
1089  * slave in the bond. This address must be, of course, one of the premanent
1090  * addresses of the other slaves.
1091  *
1092  * We go over the slave list, and for each slave there we compare its
1093  * permanent hw address with the current address of all the other slaves.
1094  * If no match was found, then we've found a slave with a permanent address
1095  * that isn't used by any other slave in the bond, so we can assign it to
1096  * @slave.
1097  *
1098  * assumption: this function is called before @slave is attached to the
1099  *             bond slave list.
1100  *
1101  * caller must hold the bond lock for write since the mac addresses are compared
1102  * and may be swapped.
1103  */
1104 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1105 {
1106         struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1107         struct slave *has_bond_addr = bond->curr_active_slave;
1108         int i, j, found = 0;
1109
1110         if (bond->slave_cnt == 0) {
1111                 /* this is the first slave */
1112                 return 0;
1113         }
1114
1115         /* if slave's mac address differs from bond's mac address
1116          * check uniqueness of slave's mac address against the other
1117          * slaves in the bond.
1118          */
1119         if (memcmp(slave->perm_hwaddr, bond->dev->dev_addr, ETH_ALEN)) {
1120                 bond_for_each_slave(bond, tmp_slave1, i) {
1121                         if (!memcmp(tmp_slave1->dev->dev_addr, slave->dev->dev_addr,
1122                                     ETH_ALEN)) {
1123                                 found = 1;
1124                                 break;
1125                         }
1126                 }
1127
1128                 if (!found)
1129                         return 0;
1130
1131                 /* Try setting slave mac to bond address and fall-through
1132                    to code handling that situation below... */
1133                 alb_set_slave_mac_addr(slave, bond->dev->dev_addr,
1134                                        bond->alb_info.rlb_enabled);
1135         }
1136
1137         /* The slave's address is equal to the address of the bond.
1138          * Search for a spare address in the bond for this slave.
1139          */
1140         free_mac_slave = NULL;
1141
1142         bond_for_each_slave(bond, tmp_slave1, i) {
1143                 found = 0;
1144                 bond_for_each_slave(bond, tmp_slave2, j) {
1145                         if (!memcmp(tmp_slave1->perm_hwaddr,
1146                                     tmp_slave2->dev->dev_addr,
1147                                     ETH_ALEN)) {
1148                                 found = 1;
1149                                 break;
1150                         }
1151                 }
1152
1153                 if (!found) {
1154                         /* no slave has tmp_slave1's perm addr
1155                          * as its curr addr
1156                          */
1157                         free_mac_slave = tmp_slave1;
1158                         break;
1159                 }
1160
1161                 if (!has_bond_addr) {
1162                         if (!memcmp(tmp_slave1->dev->dev_addr,
1163                                     bond->dev->dev_addr,
1164                                     ETH_ALEN)) {
1165
1166                                 has_bond_addr = tmp_slave1;
1167                         }
1168                 }
1169         }
1170
1171         if (free_mac_slave) {
1172                 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr,
1173                                        bond->alb_info.rlb_enabled);
1174
1175                 printk(KERN_WARNING DRV_NAME
1176                        ": %s: Warning: the hw address of slave %s is in use by "
1177                        "the bond; giving it the hw address of %s\n",
1178                        bond->dev->name, slave->dev->name, free_mac_slave->dev->name);
1179
1180         } else if (has_bond_addr) {
1181                 printk(KERN_ERR DRV_NAME
1182                        ": %s: Error: the hw address of slave %s is in use by the "
1183                        "bond; couldn't find a slave with a free hw address to "
1184                        "give it (this should not have happened)\n",
1185                        bond->dev->name, slave->dev->name);
1186                 return -EFAULT;
1187         }
1188
1189         return 0;
1190 }
1191
1192 /**
1193  * alb_set_mac_address
1194  * @bond:
1195  * @addr:
1196  *
1197  * In TLB mode all slaves are configured to the bond's hw address, but set
1198  * their dev_addr field to different addresses (based on their permanent hw
1199  * addresses).
1200  *
1201  * For each slave, this function sets the interface to the new address and then
1202  * changes its dev_addr field to its previous value.
1203  *
1204  * Unwinding assumes bond's mac address has not yet changed.
1205  */
1206 static int alb_set_mac_address(struct bonding *bond, void *addr)
1207 {
1208         struct sockaddr sa;
1209         struct slave *slave, *stop_at;
1210         char tmp_addr[ETH_ALEN];
1211         int res;
1212         int i;
1213
1214         if (bond->alb_info.rlb_enabled) {
1215                 return 0;
1216         }
1217
1218         bond_for_each_slave(bond, slave, i) {
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         return 0;
1232
1233 unwind:
1234         memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1235         sa.sa_family = bond->dev->type;
1236
1237         /* unwind from head to the slave that failed */
1238         stop_at = slave;
1239         bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1240                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1241                 dev_set_mac_address(slave->dev, &sa);
1242                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1243         }
1244
1245         return res;
1246 }
1247
1248 /************************ exported alb funcions ************************/
1249
1250 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1251 {
1252         int res;
1253
1254         res = tlb_initialize(bond);
1255         if (res) {
1256                 return res;
1257         }
1258
1259         if (rlb_enabled) {
1260                 bond->alb_info.rlb_enabled = 1;
1261                 /* initialize rlb */
1262                 res = rlb_initialize(bond);
1263                 if (res) {
1264                         tlb_deinitialize(bond);
1265                         return res;
1266                 }
1267         } else {
1268                 bond->alb_info.rlb_enabled = 0;
1269         }
1270
1271         return 0;
1272 }
1273
1274 void bond_alb_deinitialize(struct bonding *bond)
1275 {
1276         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1277
1278         tlb_deinitialize(bond);
1279
1280         if (bond_info->rlb_enabled) {
1281                 rlb_deinitialize(bond);
1282         }
1283 }
1284
1285 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1286 {
1287         struct bonding *bond = netdev_priv(bond_dev);
1288         struct ethhdr *eth_data;
1289         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1290         struct slave *tx_slave = NULL;
1291         static const __be32 ip_bcast = htonl(0xffffffff);
1292         int hash_size = 0;
1293         int do_tx_balance = 1;
1294         u32 hash_index = 0;
1295         const u8 *hash_start = NULL;
1296         int res = 1;
1297         struct ipv6hdr *ip6hdr;
1298
1299         skb_reset_mac_header(skb);
1300         eth_data = eth_hdr(skb);
1301
1302         /* make sure that the curr_active_slave and the slaves list do
1303          * not change during tx
1304          */
1305         read_lock(&bond->lock);
1306         read_lock(&bond->curr_slave_lock);
1307
1308         if (!BOND_IS_OK(bond)) {
1309                 goto out;
1310         }
1311
1312         switch (ntohs(skb->protocol)) {
1313         case ETH_P_IP: {
1314                 const struct iphdr *iph = ip_hdr(skb);
1315
1316                 if ((memcmp(eth_data->h_dest, mac_bcast, ETH_ALEN) == 0) ||
1317                     (iph->daddr == ip_bcast) ||
1318                     (iph->protocol == IPPROTO_IGMP)) {
1319                         do_tx_balance = 0;
1320                         break;
1321                 }
1322                 hash_start = (char *)&(iph->daddr);
1323                 hash_size = sizeof(iph->daddr);
1324         }
1325                 break;
1326         case ETH_P_IPV6:
1327                 /* IPv6 doesn't really use broadcast mac address, but leave
1328                  * that here just in case.
1329                  */
1330                 if (memcmp(eth_data->h_dest, mac_bcast, ETH_ALEN) == 0) {
1331                         do_tx_balance = 0;
1332                         break;
1333                 }
1334
1335                 /* IPv6 uses all-nodes multicast as an equivalent to
1336                  * broadcasts in IPv4.
1337                  */
1338                 if (memcmp(eth_data->h_dest, mac_v6_allmcast, ETH_ALEN) == 0) {
1339                         do_tx_balance = 0;
1340                         break;
1341                 }
1342
1343                 /* Additianally, DAD probes should not be tx-balanced as that
1344                  * will lead to false positives for duplicate addresses and
1345                  * prevent address configuration from working.
1346                  */
1347                 ip6hdr = ipv6_hdr(skb);
1348                 if (ipv6_addr_any(&ip6hdr->saddr)) {
1349                         do_tx_balance = 0;
1350                         break;
1351                 }
1352
1353                 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1354                 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1355                 break;
1356         case ETH_P_IPX:
1357                 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1358                         /* something is wrong with this packet */
1359                         do_tx_balance = 0;
1360                         break;
1361                 }
1362
1363                 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1364                         /* The only protocol worth balancing in
1365                          * this family since it has an "ARP" like
1366                          * mechanism
1367                          */
1368                         do_tx_balance = 0;
1369                         break;
1370                 }
1371
1372                 hash_start = (char*)eth_data->h_dest;
1373                 hash_size = ETH_ALEN;
1374                 break;
1375         case ETH_P_ARP:
1376                 do_tx_balance = 0;
1377                 if (bond_info->rlb_enabled) {
1378                         tx_slave = rlb_arp_xmit(skb, bond);
1379                 }
1380                 break;
1381         default:
1382                 do_tx_balance = 0;
1383                 break;
1384         }
1385
1386         if (do_tx_balance) {
1387                 hash_index = _simple_hash(hash_start, hash_size);
1388                 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1389         }
1390
1391         if (!tx_slave) {
1392                 /* unbalanced or unassigned, send through primary */
1393                 tx_slave = bond->curr_active_slave;
1394                 bond_info->unbalanced_load += skb->len;
1395         }
1396
1397         if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1398                 if (tx_slave != bond->curr_active_slave) {
1399                         memcpy(eth_data->h_source,
1400                                tx_slave->dev->dev_addr,
1401                                ETH_ALEN);
1402                 }
1403
1404                 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1405         } else {
1406                 if (tx_slave) {
1407                         tlb_clear_slave(bond, tx_slave, 0);
1408                 }
1409         }
1410
1411 out:
1412         if (res) {
1413                 /* no suitable interface, frame not sent */
1414                 dev_kfree_skb(skb);
1415         }
1416         read_unlock(&bond->curr_slave_lock);
1417         read_unlock(&bond->lock);
1418         return 0;
1419 }
1420
1421 void bond_alb_monitor(struct work_struct *work)
1422 {
1423         struct bonding *bond = container_of(work, struct bonding,
1424                                             alb_work.work);
1425         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1426         struct slave *slave;
1427         int i;
1428
1429         read_lock(&bond->lock);
1430
1431         if (bond->kill_timers) {
1432                 goto out;
1433         }
1434
1435         if (bond->slave_cnt == 0) {
1436                 bond_info->tx_rebalance_counter = 0;
1437                 bond_info->lp_counter = 0;
1438                 goto re_arm;
1439         }
1440
1441         bond_info->tx_rebalance_counter++;
1442         bond_info->lp_counter++;
1443
1444         /* send learning packets */
1445         if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1446                 /* change of curr_active_slave involves swapping of mac addresses.
1447                  * in order to avoid this swapping from happening while
1448                  * sending the learning packets, the curr_slave_lock must be held for
1449                  * read.
1450                  */
1451                 read_lock(&bond->curr_slave_lock);
1452
1453                 bond_for_each_slave(bond, slave, i) {
1454                         alb_send_learning_packets(slave, slave->dev->dev_addr);
1455                 }
1456
1457                 read_unlock(&bond->curr_slave_lock);
1458
1459                 bond_info->lp_counter = 0;
1460         }
1461
1462         /* rebalance tx traffic */
1463         if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1464
1465                 read_lock(&bond->curr_slave_lock);
1466
1467                 bond_for_each_slave(bond, slave, i) {
1468                         tlb_clear_slave(bond, slave, 1);
1469                         if (slave == bond->curr_active_slave) {
1470                                 SLAVE_TLB_INFO(slave).load =
1471                                         bond_info->unbalanced_load /
1472                                                 BOND_TLB_REBALANCE_INTERVAL;
1473                                 bond_info->unbalanced_load = 0;
1474                         }
1475                 }
1476
1477                 read_unlock(&bond->curr_slave_lock);
1478
1479                 bond_info->tx_rebalance_counter = 0;
1480         }
1481
1482         /* handle rlb stuff */
1483         if (bond_info->rlb_enabled) {
1484                 if (bond_info->primary_is_promisc &&
1485                     (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1486
1487                         /*
1488                          * dev_set_promiscuity requires rtnl and
1489                          * nothing else.
1490                          */
1491                         read_unlock(&bond->lock);
1492                         rtnl_lock();
1493
1494                         bond_info->rlb_promisc_timeout_counter = 0;
1495
1496                         /* If the primary was set to promiscuous mode
1497                          * because a slave was disabled then
1498                          * it can now leave promiscuous mode.
1499                          */
1500                         dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1501                         bond_info->primary_is_promisc = 0;
1502
1503                         rtnl_unlock();
1504                         read_lock(&bond->lock);
1505                 }
1506
1507                 if (bond_info->rlb_rebalance) {
1508                         bond_info->rlb_rebalance = 0;
1509                         rlb_rebalance(bond);
1510                 }
1511
1512                 /* check if clients need updating */
1513                 if (bond_info->rx_ntt) {
1514                         if (bond_info->rlb_update_delay_counter) {
1515                                 --bond_info->rlb_update_delay_counter;
1516                         } else {
1517                                 rlb_update_rx_clients(bond);
1518                                 if (bond_info->rlb_update_retry_counter) {
1519                                         --bond_info->rlb_update_retry_counter;
1520                                 } else {
1521                                         bond_info->rx_ntt = 0;
1522                                 }
1523                         }
1524                 }
1525         }
1526
1527 re_arm:
1528         queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1529 out:
1530         read_unlock(&bond->lock);
1531 }
1532
1533 /* assumption: called before the slave is attached to the bond
1534  * and not locked by the bond lock
1535  */
1536 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1537 {
1538         int res;
1539
1540         res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr,
1541                                      bond->alb_info.rlb_enabled);
1542         if (res) {
1543                 return res;
1544         }
1545
1546         /* caller must hold the bond lock for write since the mac addresses
1547          * are compared and may be swapped.
1548          */
1549         read_lock(&bond->lock);
1550
1551         res = alb_handle_addr_collision_on_attach(bond, slave);
1552
1553         read_unlock(&bond->lock);
1554
1555         if (res) {
1556                 return res;
1557         }
1558
1559         tlb_init_slave(slave);
1560
1561         /* order a rebalance ASAP */
1562         bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1563
1564         if (bond->alb_info.rlb_enabled) {
1565                 bond->alb_info.rlb_rebalance = 1;
1566         }
1567
1568         return 0;
1569 }
1570
1571 /*
1572  * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1573  * if necessary.
1574  *
1575  * Caller must hold RTNL and no other locks
1576  */
1577 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1578 {
1579         if (bond->slave_cnt > 1) {
1580                 alb_change_hw_addr_on_detach(bond, slave);
1581         }
1582
1583         tlb_clear_slave(bond, slave, 0);
1584
1585         if (bond->alb_info.rlb_enabled) {
1586                 bond->alb_info.next_rx_slave = NULL;
1587                 rlb_clear_slave(bond, slave);
1588         }
1589 }
1590
1591 /* Caller must hold bond lock for read */
1592 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1593 {
1594         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1595
1596         if (link == BOND_LINK_DOWN) {
1597                 tlb_clear_slave(bond, slave, 0);
1598                 if (bond->alb_info.rlb_enabled) {
1599                         rlb_clear_slave(bond, slave);
1600                 }
1601         } else if (link == BOND_LINK_UP) {
1602                 /* order a rebalance ASAP */
1603                 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1604                 if (bond->alb_info.rlb_enabled) {
1605                         bond->alb_info.rlb_rebalance = 1;
1606                         /* If the updelay module parameter is smaller than the
1607                          * forwarding delay of the switch the rebalance will
1608                          * not work because the rebalance arp replies will
1609                          * not be forwarded to the clients..
1610                          */
1611                 }
1612         }
1613 }
1614
1615 /**
1616  * bond_alb_handle_active_change - assign new curr_active_slave
1617  * @bond: our bonding struct
1618  * @new_slave: new slave to assign
1619  *
1620  * Set the bond->curr_active_slave to @new_slave and handle
1621  * mac address swapping and promiscuity changes as needed.
1622  *
1623  * If new_slave is NULL, caller must hold curr_slave_lock or
1624  * bond->lock for write.
1625  *
1626  * If new_slave is not NULL, caller must hold RTNL, bond->lock for
1627  * read and curr_slave_lock for write.  Processing here may sleep, so
1628  * no other locks may be held.
1629  */
1630 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1631 {
1632         struct slave *swap_slave;
1633         int i;
1634
1635         if (bond->curr_active_slave == new_slave) {
1636                 return;
1637         }
1638
1639         if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1640                 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1641                 bond->alb_info.primary_is_promisc = 0;
1642                 bond->alb_info.rlb_promisc_timeout_counter = 0;
1643         }
1644
1645         swap_slave = bond->curr_active_slave;
1646         bond->curr_active_slave = new_slave;
1647
1648         if (!new_slave || (bond->slave_cnt == 0)) {
1649                 return;
1650         }
1651
1652         /* set the new curr_active_slave to the bonds mac address
1653          * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1654          */
1655         if (!swap_slave) {
1656                 struct slave *tmp_slave;
1657                 /* find slave that is holding the bond's mac address */
1658                 bond_for_each_slave(bond, tmp_slave, i) {
1659                         if (!memcmp(tmp_slave->dev->dev_addr,
1660                                     bond->dev->dev_addr, ETH_ALEN)) {
1661                                 swap_slave = tmp_slave;
1662                                 break;
1663                         }
1664                 }
1665         }
1666
1667         /*
1668          * Arrange for swap_slave and new_slave to temporarily be
1669          * ignored so we can mess with their MAC addresses without
1670          * fear of interference from transmit activity.
1671          */
1672         if (swap_slave) {
1673                 tlb_clear_slave(bond, swap_slave, 1);
1674         }
1675         tlb_clear_slave(bond, new_slave, 1);
1676
1677         write_unlock_bh(&bond->curr_slave_lock);
1678         read_unlock(&bond->lock);
1679
1680         ASSERT_RTNL();
1681
1682         /* curr_active_slave must be set before calling alb_swap_mac_addr */
1683         if (swap_slave) {
1684                 /* swap mac address */
1685                 alb_swap_mac_addr(bond, swap_slave, new_slave);
1686         } else {
1687                 /* set the new_slave to the bond mac address */
1688                 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr,
1689                                        bond->alb_info.rlb_enabled);
1690         }
1691
1692         if (swap_slave) {
1693                 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1694                 read_lock(&bond->lock);
1695         } else {
1696                 read_lock(&bond->lock);
1697                 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1698         }
1699
1700         write_lock_bh(&bond->curr_slave_lock);
1701 }
1702
1703 /*
1704  * Called with RTNL
1705  */
1706 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1707 {
1708         struct bonding *bond = netdev_priv(bond_dev);
1709         struct sockaddr *sa = addr;
1710         struct slave *slave, *swap_slave;
1711         int res;
1712         int i;
1713
1714         if (!is_valid_ether_addr(sa->sa_data)) {
1715                 return -EADDRNOTAVAIL;
1716         }
1717
1718         res = alb_set_mac_address(bond, addr);
1719         if (res) {
1720                 return res;
1721         }
1722
1723         memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1724
1725         /* If there is no curr_active_slave there is nothing else to do.
1726          * Otherwise we'll need to pass the new address to it and handle
1727          * duplications.
1728          */
1729         if (!bond->curr_active_slave) {
1730                 return 0;
1731         }
1732
1733         swap_slave = NULL;
1734
1735         bond_for_each_slave(bond, slave, i) {
1736                 if (!memcmp(slave->dev->dev_addr, bond_dev->dev_addr, ETH_ALEN)) {
1737                         swap_slave = slave;
1738                         break;
1739                 }
1740         }
1741
1742         write_unlock_bh(&bond->curr_slave_lock);
1743         read_unlock(&bond->lock);
1744
1745         if (swap_slave) {
1746                 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1747                 alb_fasten_mac_swap(bond, swap_slave, bond->curr_active_slave);
1748         } else {
1749                 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr,
1750                                        bond->alb_info.rlb_enabled);
1751
1752                 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1753                 if (bond->alb_info.rlb_enabled) {
1754                         /* inform clients mac address has changed */
1755                         rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1756                 }
1757         }
1758
1759         read_lock(&bond->lock);
1760         write_lock_bh(&bond->curr_slave_lock);
1761
1762         return 0;
1763 }
1764
1765 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1766 {
1767         if (bond->alb_info.current_alb_vlan &&
1768             (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1769                 bond->alb_info.current_alb_vlan = NULL;
1770         }
1771
1772         if (bond->alb_info.rlb_enabled) {
1773                 rlb_clear_vlan(bond, vlan_id);
1774         }
1775 }
1776