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