drivers/net/bonding/bond_alb.c

   1 /*
   2  * Copyright(c) 1999 - 2004 Intel Corporation. All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License as published by the
   6  * Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful, but
  10  * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  11  * or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along
  15  * with this program; if not, write to the Free Software Foundation, Inc.,
  16  * 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  17  *
  18  * The full GNU General Public License is included in this distribution in the
  19  * file called LICENSE.
  20  *
  21  */
  22
  23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  24
  25 #include <linux/skbuff.h>
  26 #include <linux/netdevice.h>
  27 #include <linux/etherdevice.h>
  28 #include <linux/pkt_sched.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/slab.h>
  31 #include <linux/timer.h>
  32 #include <linux/ip.h>
  33 #include <linux/ipv6.h>
  34 #include <linux/if_arp.h>
  35 #include <linux/if_ether.h>
  36 #include <linux/if_bonding.h>
  37 #include <linux/if_vlan.h>
  38 #include <linux/in.h>
  39 #include <net/ipx.h>
  40 #include <net/arp.h>
  41 #include <net/ipv6.h>
  42 #include <asm/byteorder.h>
  43 #include "bonding.h"
  44 #include "bond_alb.h"
  45
  46
  47
  48 #ifndef __long_aligned
  49 #define __long_aligned __attribute__((aligned((sizeof(long)))))
  50 #endif
  51 static const u8 mac_bcast[ETH_ALEN] __long_aligned = {
  52         0xff, 0xff, 0xff, 0xff, 0xff, 0xff
  53 };
  54 static const u8 mac_v6_allmcast[ETH_ALEN] __long_aligned = {
  55         0x33, 0x33, 0x00, 0x00, 0x00, 0x01
  56 };
  57 static const int alb_delta_in_ticks = HZ / ALB_TIMER_TICKS_PER_SEC;
  58
  59 #pragma pack(1)
  60 struct learning_pkt {
  61         u8 mac_dst[ETH_ALEN];
  62         u8 mac_src[ETH_ALEN];
  63         __be16 type;
  64         u8 padding[ETH_ZLEN - ETH_HLEN];
  65 };
  66
  67 struct arp_pkt {
  68         __be16  hw_addr_space;
  69         __be16  prot_addr_space;
  70         u8      hw_addr_len;
  71         u8      prot_addr_len;
  72         __be16  op_code;
  73         u8      mac_src[ETH_ALEN];      /* sender hardware address */
  74         __be32  ip_src;                 /* sender IP address */
  75         u8      mac_dst[ETH_ALEN];      /* target hardware address */
  76         __be32  ip_dst;                 /* target IP address */
  77 };
  78 #pragma pack()
  79
  80 static inline struct arp_pkt *arp_pkt(const struct sk_buff *skb)
  81 {
  82         return (struct arp_pkt *)skb_network_header(skb);
  83 }
  84
  85 /* Forward declaration */
  86 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[]);
  87
  88 static inline u8 _simple_hash(const u8 *hash_start, int hash_size)
  89 {
  90         int i;
  91         u8 hash = 0;
  92
  93         for (i = 0; i < hash_size; i++) {
  94                 hash ^= hash_start[i];
  95         }
  96
  97         return hash;
  98 }
  99
 100 /*********************** tlb specific functions ***************************/
 101
 102 static inline void _lock_tx_hashtbl_bh(struct bonding *bond)
 103 {
 104         spin_lock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 105 }
 106
 107 static inline void _unlock_tx_hashtbl_bh(struct bonding *bond)
 108 {
 109         spin_unlock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 110 }
 111
 112 static inline void _lock_tx_hashtbl(struct bonding *bond)
 113 {
 114         spin_lock(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 115 }
 116
 117 static inline void _unlock_tx_hashtbl(struct bonding *bond)
 118 {
 119         spin_unlock(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 120 }
 121
 122 /* Caller must hold tx_hashtbl lock */
 123 static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
 124 {
 125         if (save_load) {
 126                 entry->load_history = 1 + entry->tx_bytes /
 127                                       BOND_TLB_REBALANCE_INTERVAL;
 128                 entry->tx_bytes = 0;
 129         }
 130
 131         entry->tx_slave = NULL;
 132         entry->next = TLB_NULL_INDEX;
 133         entry->prev = TLB_NULL_INDEX;
 134 }
 135
 136 static inline void tlb_init_slave(struct slave *slave)
 137 {
 138         SLAVE_TLB_INFO(slave).load = 0;
 139         SLAVE_TLB_INFO(slave).head = TLB_NULL_INDEX;
 140 }
 141
 142 /* Caller must hold bond lock for read, BH disabled */
 143 static void __tlb_clear_slave(struct bonding *bond, struct slave *slave,
 144                          int save_load)
 145 {
 146         struct tlb_client_info *tx_hash_table;
 147         u32 index;
 148
 149         /* clear slave from tx_hashtbl */
 150         tx_hash_table = BOND_ALB_INFO(bond).tx_hashtbl;
 151
 152         /* skip this if we've already freed the tx hash table */
 153         if (tx_hash_table) {
 154                 index = SLAVE_TLB_INFO(slave).head;
 155                 while (index != TLB_NULL_INDEX) {
 156                         u32 next_index = tx_hash_table[index].next;
 157                         tlb_init_table_entry(&tx_hash_table[index], save_load);
 158                         index = next_index;
 159                 }
 160         }
 161
 162         tlb_init_slave(slave);
 163 }
 164
 165 /* Caller must hold bond lock for read */
 166 static void tlb_clear_slave(struct bonding *bond, struct slave *slave,
 167                          int save_load)
 168 {
 169         _lock_tx_hashtbl_bh(bond);
 170         __tlb_clear_slave(bond, slave, save_load);
 171         _unlock_tx_hashtbl_bh(bond);
 172 }
 173
 174 /* Must be called before starting the monitor timer */
 175 static int tlb_initialize(struct bonding *bond)
 176 {
 177         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 178         int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
 179         struct tlb_client_info *new_hashtbl;
 180         int i;
 181
 182         new_hashtbl = kzalloc(size, GFP_KERNEL);
 183         if (!new_hashtbl)
 184                 return -1;
 185
 186         _lock_tx_hashtbl_bh(bond);
 187
 188         bond_info->tx_hashtbl = new_hashtbl;
 189
 190         for (i = 0; i < TLB_HASH_TABLE_SIZE; i++) {
 191                 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 0);
 192         }
 193
 194         _unlock_tx_hashtbl_bh(bond);
 195
 196         return 0;
 197 }
 198
 199 /* Must be called only after all slaves have been released */
 200 static void tlb_deinitialize(struct bonding *bond)
 201 {
 202         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 203
 204         _lock_tx_hashtbl_bh(bond);
 205
 206         kfree(bond_info->tx_hashtbl);
 207         bond_info->tx_hashtbl = NULL;
 208
 209         _unlock_tx_hashtbl_bh(bond);
 210 }
 211
 212 static long long compute_gap(struct slave *slave)
 213 {
 214         return (s64) (slave->speed << 20) - /* Convert to Megabit per sec */
 215                (s64) (SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
 216 }
 217
 218 /* Caller must hold bond lock for read */
 219 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
 220 {
 221         struct slave *slave, *least_loaded;
 222         long long max_gap;
 223         int i;
 224
 225         least_loaded = NULL;
 226         max_gap = LLONG_MIN;
 227
 228         /* Find the slave with the largest gap */
 229         bond_for_each_slave(bond, slave, i) {
 230                 if (SLAVE_IS_OK(slave)) {
 231                         long long gap = compute_gap(slave);
 232
 233                         if (max_gap < gap) {
 234                                 least_loaded = slave;
 235                                 max_gap = gap;
 236                         }
 237                 }
 238         }
 239
 240         return least_loaded;
 241 }
 242
 243 static struct slave *__tlb_choose_channel(struct bonding *bond, u32 hash_index,
 244                                                 u32 skb_len)
 245 {
 246         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 247         struct tlb_client_info *hash_table;
 248         struct slave *assigned_slave;
 249
 250         hash_table = bond_info->tx_hashtbl;
 251         assigned_slave = hash_table[hash_index].tx_slave;
 252         if (!assigned_slave) {
 253                 assigned_slave = tlb_get_least_loaded_slave(bond);
 254
 255                 if (assigned_slave) {
 256                         struct tlb_slave_info *slave_info =
 257                                 &(SLAVE_TLB_INFO(assigned_slave));
 258                         u32 next_index = slave_info->head;
 259
 260                         hash_table[hash_index].tx_slave = assigned_slave;
 261                         hash_table[hash_index].next = next_index;
 262                         hash_table[hash_index].prev = TLB_NULL_INDEX;
 263
 264                         if (next_index != TLB_NULL_INDEX) {
 265                                 hash_table[next_index].prev = hash_index;
 266                         }
 267
 268                         slave_info->head = hash_index;
 269                         slave_info->load +=
 270                                 hash_table[hash_index].load_history;
 271                 }
 272         }
 273
 274         if (assigned_slave) {
 275                 hash_table[hash_index].tx_bytes += skb_len;
 276         }
 277
 278         return assigned_slave;
 279 }
 280
 281 /* Caller must hold bond lock for read */
 282 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index,
 283                                         u32 skb_len)
 284 {
 285         struct slave *tx_slave;
 286         /*
 287          * We don't need to disable softirq here, becase
 288          * tlb_choose_channel() is only called by bond_alb_xmit()
 289          * which already has softirq disabled.
 290          */
 291         _lock_tx_hashtbl(bond);
 292         tx_slave = __tlb_choose_channel(bond, hash_index, skb_len);
 293         _unlock_tx_hashtbl(bond);
 294         return tx_slave;
 295 }
 296
 297 /*********************** rlb specific functions ***************************/
 298 static inline void _lock_rx_hashtbl_bh(struct bonding *bond)
 299 {
 300         spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 301 }
 302
 303 static inline void _unlock_rx_hashtbl_bh(struct bonding *bond)
 304 {
 305         spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 306 }
 307
 308 static inline void _lock_rx_hashtbl(struct bonding *bond)
 309 {
 310         spin_lock(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 311 }
 312
 313 static inline void _unlock_rx_hashtbl(struct bonding *bond)
 314 {
 315         spin_unlock(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 316 }
 317
 318 /* when an ARP REPLY is received from a client update its info
 319  * in the rx_hashtbl
 320  */
 321 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
 322 {
 323         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 324         struct rlb_client_info *client_info;
 325         u32 hash_index;
 326
 327         _lock_rx_hashtbl_bh(bond);
 328
 329         hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
 330         client_info = &(bond_info->rx_hashtbl[hash_index]);
 331
 332         if ((client_info->assigned) &&
 333             (client_info->ip_src == arp->ip_dst) &&
 334             (client_info->ip_dst == arp->ip_src) &&
 335             (compare_ether_addr_64bits(client_info->mac_dst, arp->mac_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_bh(bond);
 343 }
 344
 345 static int rlb_arp_recv(struct sk_buff *skb, struct bonding *bond,
 346                          struct slave *slave)
 347 {
 348         struct arp_pkt *arp;
 349
 350         if (skb->protocol != cpu_to_be16(ETH_P_ARP))
 351                 goto out;
 352
 353         arp = (struct arp_pkt *) skb->data;
 354         if (!arp) {
 355                 pr_debug("Packet has no ARP data\n");
 356                 goto out;
 357         }
 358
 359         if (!pskb_may_pull(skb, arp_hdr_len(bond->dev)))
 360                 goto out;
 361
 362         if (skb->len < sizeof(struct arp_pkt)) {
 363                 pr_debug("Packet is too small to be an ARP\n");
 364                 goto out;
 365         }
 366
 367         if (arp->op_code == htons(ARPOP_REPLY)) {
 368                 /* update rx hash table for this ARP */
 369                 rlb_update_entry_from_arp(bond, arp);
 370                 pr_debug("Server received an ARP Reply from client\n");
 371         }
 372 out:
 373         return RX_HANDLER_ANOTHER;
 374 }
 375
 376 /* Caller must hold bond lock for read */
 377 static struct slave *rlb_next_rx_slave(struct bonding *bond)
 378 {
 379         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 380         struct slave *rx_slave, *slave, *start_at;
 381         int i = 0;
 382
 383         if (bond_info->next_rx_slave) {
 384                 start_at = bond_info->next_rx_slave;
 385         } else {
 386                 start_at = bond->first_slave;
 387         }
 388
 389         rx_slave = NULL;
 390
 391         bond_for_each_slave_from(bond, slave, i, start_at) {
 392                 if (SLAVE_IS_OK(slave)) {
 393                         if (!rx_slave) {
 394                                 rx_slave = slave;
 395                         } else if (slave->speed > rx_slave->speed) {
 396                                 rx_slave = slave;
 397                         }
 398                 }
 399         }
 400
 401         if (rx_slave) {
 402                 bond_info->next_rx_slave = rx_slave->next;
 403         }
 404
 405         return rx_slave;
 406 }
 407
 408 /* teach the switch the mac of a disabled slave
 409  * on the primary for fault tolerance
 410  *
 411  * Caller must hold bond->curr_slave_lock for write or bond lock for write
 412  */
 413 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
 414 {
 415         if (!bond->curr_active_slave) {
 416                 return;
 417         }
 418
 419         if (!bond->alb_info.primary_is_promisc) {
 420                 if (!dev_set_promiscuity(bond->curr_active_slave->dev, 1))
 421                         bond->alb_info.primary_is_promisc = 1;
 422                 else
 423                         bond->alb_info.primary_is_promisc = 0;
 424         }
 425
 426         bond->alb_info.rlb_promisc_timeout_counter = 0;
 427
 428         alb_send_learning_packets(bond->curr_active_slave, addr);
 429 }
 430
 431 /* slave being removed should not be active at this point
 432  *
 433  * Caller must hold bond lock for read
 434  */
 435 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
 436 {
 437         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 438         struct rlb_client_info *rx_hash_table;
 439         u32 index, next_index;
 440
 441         /* clear slave from rx_hashtbl */
 442         _lock_rx_hashtbl_bh(bond);
 443
 444         rx_hash_table = bond_info->rx_hashtbl;
 445         index = bond_info->rx_hashtbl_head;
 446         for (; index != RLB_NULL_INDEX; index = next_index) {
 447                 next_index = rx_hash_table[index].next;
 448                 if (rx_hash_table[index].slave == slave) {
 449                         struct slave *assigned_slave = rlb_next_rx_slave(bond);
 450
 451                         if (assigned_slave) {
 452                                 rx_hash_table[index].slave = assigned_slave;
 453                                 if (compare_ether_addr_64bits(rx_hash_table[index].mac_dst,
 454                                                               mac_bcast)) {
 455                                         bond_info->rx_hashtbl[index].ntt = 1;
 456                                         bond_info->rx_ntt = 1;
 457                                         /* A slave has been removed from the
 458                                          * table because it is either disabled
 459                                          * or being released. We must retry the
 460                                          * update to avoid clients from not
 461                                          * being updated & disconnecting when
 462                                          * there is stress
 463                                          */
 464                                         bond_info->rlb_update_retry_counter =
 465                                                 RLB_UPDATE_RETRY;
 466                                 }
 467                         } else {  /* there is no active slave */
 468                                 rx_hash_table[index].slave = NULL;
 469                         }
 470                 }
 471         }
 472
 473         _unlock_rx_hashtbl_bh(bond);
 474
 475         write_lock_bh(&bond->curr_slave_lock);
 476
 477         if (slave != bond->curr_active_slave) {
 478                 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
 479         }
 480
 481         write_unlock_bh(&bond->curr_slave_lock);
 482 }
 483
 484 static void rlb_update_client(struct rlb_client_info *client_info)
 485 {
 486         int i;
 487
 488         if (!client_info->slave) {
 489                 return;
 490         }
 491
 492         for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
 493                 struct sk_buff *skb;
 494
 495                 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
 496                                  client_info->ip_dst,
 497                                  client_info->slave->dev,
 498                                  client_info->ip_src,
 499                                  client_info->mac_dst,
 500                                  client_info->slave->dev->dev_addr,
 501                                  client_info->mac_dst);
 502                 if (!skb) {
 503                         pr_err("%s: Error: failed to create an ARP packet\n",
 504                                client_info->slave->dev->master->name);
 505                         continue;
 506                 }
 507
 508                 skb->dev = client_info->slave->dev;
 509
 510                 if (client_info->tag) {
 511                         skb = vlan_put_tag(skb, client_info->vlan_id);
 512                         if (!skb) {
 513                                 pr_err("%s: Error: failed to insert VLAN tag\n",
 514                                        client_info->slave->dev->master->name);
 515                                 continue;
 516                         }
 517                 }
 518
 519                 arp_xmit(skb);
 520         }
 521 }
 522
 523 /* sends ARP REPLIES that update the clients that need updating */
 524 static void rlb_update_rx_clients(struct bonding *bond)
 525 {
 526         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 527         struct rlb_client_info *client_info;
 528         u32 hash_index;
 529
 530         _lock_rx_hashtbl_bh(bond);
 531
 532         hash_index = bond_info->rx_hashtbl_head;
 533         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 534                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 535                 if (client_info->ntt) {
 536                         rlb_update_client(client_info);
 537                         if (bond_info->rlb_update_retry_counter == 0) {
 538                                 client_info->ntt = 0;
 539                         }
 540                 }
 541         }
 542
 543         /* do not update the entries again until this counter is zero so that
 544          * not to confuse the clients.
 545          */
 546         bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
 547
 548         _unlock_rx_hashtbl_bh(bond);
 549 }
 550
 551 /* The slave was assigned a new mac address - update the clients */
 552 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
 553 {
 554         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 555         struct rlb_client_info *client_info;
 556         int ntt = 0;
 557         u32 hash_index;
 558
 559         _lock_rx_hashtbl_bh(bond);
 560
 561         hash_index = bond_info->rx_hashtbl_head;
 562         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 563                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 564
 565                 if ((client_info->slave == slave) &&
 566                     compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 567                         client_info->ntt = 1;
 568                         ntt = 1;
 569                 }
 570         }
 571
 572         // update the team's flag only after the whole iteration
 573         if (ntt) {
 574                 bond_info->rx_ntt = 1;
 575                 //fasten the change
 576                 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
 577         }
 578
 579         _unlock_rx_hashtbl_bh(bond);
 580 }
 581
 582 /* mark all clients using src_ip to be updated */
 583 static void rlb_req_update_subnet_clients(struct bonding *bond, __be32 src_ip)
 584 {
 585         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 586         struct rlb_client_info *client_info;
 587         u32 hash_index;
 588
 589         _lock_rx_hashtbl(bond);
 590
 591         hash_index = bond_info->rx_hashtbl_head;
 592         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 593                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 594
 595                 if (!client_info->slave) {
 596                         pr_err("%s: Error: found a client with no channel in the client's hash table\n",
 597                                bond->dev->name);
 598                         continue;
 599                 }
 600                 /*update all clients using this src_ip, that are not assigned
 601                  * to the team's address (curr_active_slave) and have a known
 602                  * unicast mac address.
 603                  */
 604                 if ((client_info->ip_src == src_ip) &&
 605                     compare_ether_addr_64bits(client_info->slave->dev->dev_addr,
 606                            bond->dev->dev_addr) &&
 607                     compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 608                         client_info->ntt = 1;
 609                         bond_info->rx_ntt = 1;
 610                 }
 611         }
 612
 613         _unlock_rx_hashtbl(bond);
 614 }
 615
 616 /* Caller must hold both bond and ptr locks for read */
 617 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
 618 {
 619         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 620         struct arp_pkt *arp = arp_pkt(skb);
 621         struct slave *assigned_slave;
 622         struct rlb_client_info *client_info;
 623         u32 hash_index = 0;
 624
 625         _lock_rx_hashtbl(bond);
 626
 627         hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_dst));
 628         client_info = &(bond_info->rx_hashtbl[hash_index]);
 629
 630         if (client_info->assigned) {
 631                 if ((client_info->ip_src == arp->ip_src) &&
 632                     (client_info->ip_dst == arp->ip_dst)) {
 633                         /* the entry is already assigned to this client */
 634                         if (compare_ether_addr_64bits(arp->mac_dst, mac_bcast)) {
 635                                 /* update mac address from arp */
 636                                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 637                         }
 638
 639                         assigned_slave = client_info->slave;
 640                         if (assigned_slave) {
 641                                 _unlock_rx_hashtbl(bond);
 642                                 return assigned_slave;
 643                         }
 644                 } else {
 645                         /* the entry is already assigned to some other client,
 646                          * move the old client to primary (curr_active_slave) so
 647                          * that the new client can be assigned to this entry.
 648                          */
 649                         if (bond->curr_active_slave &&
 650                             client_info->slave != bond->curr_active_slave) {
 651                                 client_info->slave = bond->curr_active_slave;
 652                                 rlb_update_client(client_info);
 653                         }
 654                 }
 655         }
 656         /* assign a new slave */
 657         assigned_slave = rlb_next_rx_slave(bond);
 658
 659         if (assigned_slave) {
 660                 client_info->ip_src = arp->ip_src;
 661                 client_info->ip_dst = arp->ip_dst;
 662                 /* arp->mac_dst is broadcast for arp reqeusts.
 663                  * will be updated with clients actual unicast mac address
 664                  * upon receiving an arp reply.
 665                  */
 666                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 667                 client_info->slave = assigned_slave;
 668
 669                 if (compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 670                         client_info->ntt = 1;
 671                         bond->alb_info.rx_ntt = 1;
 672                 } else {
 673                         client_info->ntt = 0;
 674                 }
 675
 676                 if (bond_vlan_used(bond)) {
 677                         if (!vlan_get_tag(skb, &client_info->vlan_id))
 678                                 client_info->tag = 1;
 679                 }
 680
 681                 if (!client_info->assigned) {
 682                         u32 prev_tbl_head = bond_info->rx_hashtbl_head;
 683                         bond_info->rx_hashtbl_head = hash_index;
 684                         client_info->next = prev_tbl_head;
 685                         if (prev_tbl_head != RLB_NULL_INDEX) {
 686                                 bond_info->rx_hashtbl[prev_tbl_head].prev =
 687                                         hash_index;
 688                         }
 689                         client_info->assigned = 1;
 690                 }
 691         }
 692
 693         _unlock_rx_hashtbl(bond);
 694
 695         return assigned_slave;
 696 }
 697
 698 /* chooses (and returns) transmit channel for arp reply
 699  * does not choose channel for other arp types since they are
 700  * sent on the curr_active_slave
 701  */
 702 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
 703 {
 704         struct arp_pkt *arp = arp_pkt(skb);
 705         struct slave *tx_slave = NULL;
 706
 707         if (arp->op_code == htons(ARPOP_REPLY)) {
 708                 /* the arp must be sent on the selected
 709                 * rx channel
 710                 */
 711                 tx_slave = rlb_choose_channel(skb, bond);
 712                 if (tx_slave) {
 713                         memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
 714                 }
 715                 pr_debug("Server sent ARP Reply packet\n");
 716         } else if (arp->op_code == htons(ARPOP_REQUEST)) {
 717                 /* Create an entry in the rx_hashtbl for this client as a
 718                  * place holder.
 719                  * When the arp reply is received the entry will be updated
 720                  * with the correct unicast address of the client.
 721                  */
 722                 rlb_choose_channel(skb, bond);
 723
 724                 /* The ARP reply packets must be delayed so that
 725                  * they can cancel out the influence of the ARP request.
 726                  */
 727                 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
 728
 729                 /* arp requests are broadcast and are sent on the primary
 730                  * the arp request will collapse all clients on the subnet to
 731                  * the primary slave. We must register these clients to be
 732                  * updated with their assigned mac.
 733                  */
 734                 rlb_req_update_subnet_clients(bond, arp->ip_src);
 735                 pr_debug("Server sent ARP Request packet\n");
 736         }
 737
 738         return tx_slave;
 739 }
 740
 741 /* Caller must hold bond lock for read */
 742 static void rlb_rebalance(struct bonding *bond)
 743 {
 744         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 745         struct slave *assigned_slave;
 746         struct rlb_client_info *client_info;
 747         int ntt;
 748         u32 hash_index;
 749
 750         _lock_rx_hashtbl_bh(bond);
 751
 752         ntt = 0;
 753         hash_index = bond_info->rx_hashtbl_head;
 754         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 755                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 756                 assigned_slave = rlb_next_rx_slave(bond);
 757                 if (assigned_slave && (client_info->slave != assigned_slave)) {
 758                         client_info->slave = assigned_slave;
 759                         client_info->ntt = 1;
 760                         ntt = 1;
 761                 }
 762         }
 763
 764         /* update the team's flag only after the whole iteration */
 765         if (ntt) {
 766                 bond_info->rx_ntt = 1;
 767         }
 768         _unlock_rx_hashtbl_bh(bond);
 769 }
 770
 771 /* Caller must hold rx_hashtbl lock */
 772 static void rlb_init_table_entry(struct rlb_client_info *entry)
 773 {
 774         memset(entry, 0, sizeof(struct rlb_client_info));
 775         entry->next = RLB_NULL_INDEX;
 776         entry->prev = RLB_NULL_INDEX;
 777 }
 778
 779 static int rlb_initialize(struct bonding *bond)
 780 {
 781         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 782         struct rlb_client_info  *new_hashtbl;
 783         int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
 784         int i;
 785
 786         new_hashtbl = kmalloc(size, GFP_KERNEL);
 787         if (!new_hashtbl)
 788                 return -1;
 789
 790         _lock_rx_hashtbl_bh(bond);
 791
 792         bond_info->rx_hashtbl = new_hashtbl;
 793
 794         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 795
 796         for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
 797                 rlb_init_table_entry(bond_info->rx_hashtbl + i);
 798         }
 799
 800         _unlock_rx_hashtbl_bh(bond);
 801
 802         /* register to receive ARPs */
 803         bond->recv_probe = rlb_arp_recv;
 804
 805         return 0;
 806 }
 807
 808 static void rlb_deinitialize(struct bonding *bond)
 809 {
 810         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 811
 812         _lock_rx_hashtbl_bh(bond);
 813
 814         kfree(bond_info->rx_hashtbl);
 815         bond_info->rx_hashtbl = NULL;
 816         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 817
 818         _unlock_rx_hashtbl_bh(bond);
 819 }
 820
 821 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
 822 {
 823         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 824         u32 curr_index;
 825
 826         _lock_rx_hashtbl_bh(bond);
 827
 828         curr_index = bond_info->rx_hashtbl_head;
 829         while (curr_index != RLB_NULL_INDEX) {
 830                 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
 831                 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
 832                 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
 833
 834                 if (curr->tag && (curr->vlan_id == vlan_id)) {
 835                         if (curr_index == bond_info->rx_hashtbl_head) {
 836                                 bond_info->rx_hashtbl_head = next_index;
 837                         }
 838                         if (prev_index != RLB_NULL_INDEX) {
 839                                 bond_info->rx_hashtbl[prev_index].next = next_index;
 840                         }
 841                         if (next_index != RLB_NULL_INDEX) {
 842                                 bond_info->rx_hashtbl[next_index].prev = prev_index;
 843                         }
 844
 845                         rlb_init_table_entry(curr);
 846                 }
 847
 848                 curr_index = next_index;
 849         }
 850
 851         _unlock_rx_hashtbl_bh(bond);
 852 }
 853
 854 /*********************** tlb/rlb shared functions *********************/
 855
 856 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
 857 {
 858         struct bonding *bond = bond_get_bond_by_slave(slave);
 859         struct learning_pkt pkt;
 860         int size = sizeof(struct learning_pkt);
 861         int i;
 862
 863         memset(&pkt, 0, size);
 864         memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
 865         memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
 866         pkt.type = cpu_to_be16(ETH_P_LOOP);
 867
 868         for (i = 0; i < MAX_LP_BURST; i++) {
 869                 struct sk_buff *skb;
 870                 char *data;
 871
 872                 skb = dev_alloc_skb(size);
 873                 if (!skb) {
 874                         return;
 875                 }
 876
 877                 data = skb_put(skb, size);
 878                 memcpy(data, &pkt, size);
 879
 880                 skb_reset_mac_header(skb);
 881                 skb->network_header = skb->mac_header + ETH_HLEN;
 882                 skb->protocol = pkt.type;
 883                 skb->priority = TC_PRIO_CONTROL;
 884                 skb->dev = slave->dev;
 885
 886                 if (bond_vlan_used(bond)) {
 887                         struct vlan_entry *vlan;
 888
 889                         vlan = bond_next_vlan(bond,
 890                                               bond->alb_info.current_alb_vlan);
 891
 892                         bond->alb_info.current_alb_vlan = vlan;
 893                         if (!vlan) {
 894                                 kfree_skb(skb);
 895                                 continue;
 896                         }
 897
 898                         skb = vlan_put_tag(skb, vlan->vlan_id);
 899                         if (!skb) {
 900                                 pr_err("%s: Error: failed to insert VLAN tag\n",
 901                                        bond->dev->name);
 902                                 continue;
 903                         }
 904                 }
 905
 906                 dev_queue_xmit(skb);
 907         }
 908 }
 909
 910 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[])
 911 {
 912         struct net_device *dev = slave->dev;
 913         struct sockaddr s_addr;
 914
 915         if (slave->bond->params.mode == BOND_MODE_TLB) {
 916                 memcpy(dev->dev_addr, addr, dev->addr_len);
 917                 return 0;
 918         }
 919
 920         /* for rlb each slave must have a unique hw mac addresses so that */
 921         /* each slave will receive packets destined to a different mac */
 922         memcpy(s_addr.sa_data, addr, dev->addr_len);
 923         s_addr.sa_family = dev->type;
 924         if (dev_set_mac_address(dev, &s_addr)) {
 925                 pr_err("%s: Error: dev_set_mac_address of dev %s failed!\n"
 926                        "ALB mode requires that the base driver support setting the hw address also when the network device's interface is open\n",
 927                        dev->master->name, dev->name);
 928                 return -EOPNOTSUPP;
 929         }
 930         return 0;
 931 }
 932
 933 /*
 934  * Swap MAC addresses between two slaves.
 935  *
 936  * Called with RTNL held, and no other locks.
 937  *
 938  */
 939
 940 static void alb_swap_mac_addr(struct bonding *bond, struct slave *slave1, struct slave *slave2)
 941 {
 942         u8 tmp_mac_addr[ETH_ALEN];
 943
 944         memcpy(tmp_mac_addr, slave1->dev->dev_addr, ETH_ALEN);
 945         alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr);
 946         alb_set_slave_mac_addr(slave2, tmp_mac_addr);
 947
 948 }
 949
 950 /*
 951  * Send learning packets after MAC address swap.
 952  *
 953  * Called with RTNL and no other locks
 954  */
 955 static void alb_fasten_mac_swap(struct bonding *bond, struct slave *slave1,
 956                                 struct slave *slave2)
 957 {
 958         int slaves_state_differ = (SLAVE_IS_OK(slave1) != SLAVE_IS_OK(slave2));
 959         struct slave *disabled_slave = NULL;
 960
 961         ASSERT_RTNL();
 962
 963         /* fasten the change in the switch */
 964         if (SLAVE_IS_OK(slave1)) {
 965                 alb_send_learning_packets(slave1, slave1->dev->dev_addr);
 966                 if (bond->alb_info.rlb_enabled) {
 967                         /* inform the clients that the mac address
 968                          * has changed
 969                          */
 970                         rlb_req_update_slave_clients(bond, slave1);
 971                 }
 972         } else {
 973                 disabled_slave = slave1;
 974         }
 975
 976         if (SLAVE_IS_OK(slave2)) {
 977                 alb_send_learning_packets(slave2, slave2->dev->dev_addr);
 978                 if (bond->alb_info.rlb_enabled) {
 979                         /* inform the clients that the mac address
 980                          * has changed
 981                          */
 982                         rlb_req_update_slave_clients(bond, slave2);
 983                 }
 984         } else {
 985                 disabled_slave = slave2;
 986         }
 987
 988         if (bond->alb_info.rlb_enabled && slaves_state_differ) {
 989                 /* A disabled slave was assigned an active mac addr */
 990                 rlb_teach_disabled_mac_on_primary(bond,
 991                                                   disabled_slave->dev->dev_addr);
 992         }
 993 }
 994
 995 /**
 996  * alb_change_hw_addr_on_detach
 997  * @bond: bonding we're working on
 998  * @slave: the slave that was just detached
 999  *
1000  * We assume that @slave was already detached from the slave list.
1001  *
1002  * If @slave's permanent hw address is different both from its current
1003  * address and from @bond's address, then somewhere in the bond there's
1004  * a slave that has @slave's permanet address as its current address.
1005  * We'll make sure that that slave no longer uses @slave's permanent address.
1006  *
1007  * Caller must hold RTNL and no other locks
1008  */
1009 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
1010 {
1011         int perm_curr_diff;
1012         int perm_bond_diff;
1013
1014         perm_curr_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
1015                                                    slave->dev->dev_addr);
1016         perm_bond_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
1017                                                    bond->dev->dev_addr);
1018
1019         if (perm_curr_diff && perm_bond_diff) {
1020                 struct slave *tmp_slave;
1021                 int i, found = 0;
1022
1023                 bond_for_each_slave(bond, tmp_slave, i) {
1024                         if (!compare_ether_addr_64bits(slave->perm_hwaddr,
1025                                                        tmp_slave->dev->dev_addr)) {
1026                                 found = 1;
1027                                 break;
1028                         }
1029                 }
1030
1031                 if (found) {
1032                         /* locking: needs RTNL and nothing else */
1033                         alb_swap_mac_addr(bond, slave, tmp_slave);
1034                         alb_fasten_mac_swap(bond, slave, tmp_slave);
1035                 }
1036         }
1037 }
1038
1039 /**
1040  * alb_handle_addr_collision_on_attach
1041  * @bond: bonding we're working on
1042  * @slave: the slave that was just attached
1043  *
1044  * checks uniqueness of slave's mac address and handles the case the
1045  * new slave uses the bonds mac address.
1046  *
1047  * If the permanent hw address of @slave is @bond's hw address, we need to
1048  * find a different hw address to give @slave, that isn't in use by any other
1049  * slave in the bond. This address must be, of course, one of the permanent
1050  * addresses of the other slaves.
1051  *
1052  * We go over the slave list, and for each slave there we compare its
1053  * permanent hw address with the current address of all the other slaves.
1054  * If no match was found, then we've found a slave with a permanent address
1055  * that isn't used by any other slave in the bond, so we can assign it to
1056  * @slave.
1057  *
1058  * assumption: this function is called before @slave is attached to the
1059  *             bond slave list.
1060  *
1061  * caller must hold the bond lock for write since the mac addresses are compared
1062  * and may be swapped.
1063  */
1064 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1065 {
1066         struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1067         struct slave *has_bond_addr = bond->curr_active_slave;
1068         int i, j, found = 0;
1069
1070         if (bond->slave_cnt == 0) {
1071                 /* this is the first slave */
1072                 return 0;
1073         }
1074
1075         /* if slave's mac address differs from bond's mac address
1076          * check uniqueness of slave's mac address against the other
1077          * slaves in the bond.
1078          */
1079         if (compare_ether_addr_64bits(slave->perm_hwaddr, bond->dev->dev_addr)) {
1080                 bond_for_each_slave(bond, tmp_slave1, i) {
1081                         if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1082                                                        slave->dev->dev_addr)) {
1083                                 found = 1;
1084                                 break;
1085                         }
1086                 }
1087
1088                 if (!found)
1089                         return 0;
1090
1091                 /* Try setting slave mac to bond address and fall-through
1092                    to code handling that situation below... */
1093                 alb_set_slave_mac_addr(slave, bond->dev->dev_addr);
1094         }
1095
1096         /* The slave's address is equal to the address of the bond.
1097          * Search for a spare address in the bond for this slave.
1098          */
1099         free_mac_slave = NULL;
1100
1101         bond_for_each_slave(bond, tmp_slave1, i) {
1102                 found = 0;
1103                 bond_for_each_slave(bond, tmp_slave2, j) {
1104                         if (!compare_ether_addr_64bits(tmp_slave1->perm_hwaddr,
1105                                                        tmp_slave2->dev->dev_addr)) {
1106                                 found = 1;
1107                                 break;
1108                         }
1109                 }
1110
1111                 if (!found) {
1112                         /* no slave has tmp_slave1's perm addr
1113                          * as its curr addr
1114                          */
1115                         free_mac_slave = tmp_slave1;
1116                         break;
1117                 }
1118
1119                 if (!has_bond_addr) {
1120                         if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1121                                                        bond->dev->dev_addr)) {
1122
1123                                 has_bond_addr = tmp_slave1;
1124                         }
1125                 }
1126         }
1127
1128         if (free_mac_slave) {
1129                 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr);
1130
1131                 pr_warning("%s: Warning: the hw address of slave %s is in use by the bond; giving it the hw address of %s\n",
1132                            bond->dev->name, slave->dev->name,
1133                            free_mac_slave->dev->name);
1134
1135         } else if (has_bond_addr) {
1136                 pr_err("%s: Error: the hw address of slave %s is in use by the bond; couldn't find a slave with a free hw address to give it (this should not have happened)\n",
1137                        bond->dev->name, slave->dev->name);
1138                 return -EFAULT;
1139         }
1140
1141         return 0;
1142 }
1143
1144 /**
1145  * alb_set_mac_address
1146  * @bond:
1147  * @addr:
1148  *
1149  * In TLB mode all slaves are configured to the bond's hw address, but set
1150  * their dev_addr field to different addresses (based on their permanent hw
1151  * addresses).
1152  *
1153  * For each slave, this function sets the interface to the new address and then
1154  * changes its dev_addr field to its previous value.
1155  *
1156  * Unwinding assumes bond's mac address has not yet changed.
1157  */
1158 static int alb_set_mac_address(struct bonding *bond, void *addr)
1159 {
1160         struct sockaddr sa;
1161         struct slave *slave, *stop_at;
1162         char tmp_addr[ETH_ALEN];
1163         int res;
1164         int i;
1165
1166         if (bond->alb_info.rlb_enabled) {
1167                 return 0;
1168         }
1169
1170         bond_for_each_slave(bond, slave, i) {
1171                 /* save net_device's current hw address */
1172                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1173
1174                 res = dev_set_mac_address(slave->dev, addr);
1175
1176                 /* restore net_device's hw address */
1177                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1178
1179                 if (res)
1180                         goto unwind;
1181         }
1182
1183         return 0;
1184
1185 unwind:
1186         memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1187         sa.sa_family = bond->dev->type;
1188
1189         /* unwind from head to the slave that failed */
1190         stop_at = slave;
1191         bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1192                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1193                 dev_set_mac_address(slave->dev, &sa);
1194                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1195         }
1196
1197         return res;
1198 }
1199
1200 /************************ exported alb funcions ************************/
1201
1202 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1203 {
1204         int res;
1205
1206         res = tlb_initialize(bond);
1207         if (res) {
1208                 return res;
1209         }
1210
1211         if (rlb_enabled) {
1212                 bond->alb_info.rlb_enabled = 1;
1213                 /* initialize rlb */
1214                 res = rlb_initialize(bond);
1215                 if (res) {
1216                         tlb_deinitialize(bond);
1217                         return res;
1218                 }
1219         } else {
1220                 bond->alb_info.rlb_enabled = 0;
1221         }
1222
1223         return 0;
1224 }
1225
1226 void bond_alb_deinitialize(struct bonding *bond)
1227 {
1228         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1229
1230         tlb_deinitialize(bond);
1231
1232         if (bond_info->rlb_enabled) {
1233                 rlb_deinitialize(bond);
1234         }
1235 }
1236
1237 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1238 {
1239         struct bonding *bond = netdev_priv(bond_dev);
1240         struct ethhdr *eth_data;
1241         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1242         struct slave *tx_slave = NULL;
1243         static const __be32 ip_bcast = htonl(0xffffffff);
1244         int hash_size = 0;
1245         int do_tx_balance = 1;
1246         u32 hash_index = 0;
1247         const u8 *hash_start = NULL;
1248         int res = 1;
1249         struct ipv6hdr *ip6hdr;
1250
1251         skb_reset_mac_header(skb);
1252         eth_data = eth_hdr(skb);
1253
1254         /* make sure that the curr_active_slave do not change during tx
1255          */
1256         read_lock(&bond->curr_slave_lock);
1257
1258         switch (ntohs(skb->protocol)) {
1259         case ETH_P_IP: {
1260                 const struct iphdr *iph = ip_hdr(skb);
1261
1262                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast) ||
1263                     (iph->daddr == ip_bcast) ||
1264                     (iph->protocol == IPPROTO_IGMP)) {
1265                         do_tx_balance = 0;
1266                         break;
1267                 }
1268                 hash_start = (char *)&(iph->daddr);
1269                 hash_size = sizeof(iph->daddr);
1270         }
1271                 break;
1272         case ETH_P_IPV6:
1273                 /* IPv6 doesn't really use broadcast mac address, but leave
1274                  * that here just in case.
1275                  */
1276                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast)) {
1277                         do_tx_balance = 0;
1278                         break;
1279                 }
1280
1281                 /* IPv6 uses all-nodes multicast as an equivalent to
1282                  * broadcasts in IPv4.
1283                  */
1284                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_v6_allmcast)) {
1285                         do_tx_balance = 0;
1286                         break;
1287                 }
1288
1289                 /* Additianally, DAD probes should not be tx-balanced as that
1290                  * will lead to false positives for duplicate addresses and
1291                  * prevent address configuration from working.
1292                  */
1293                 ip6hdr = ipv6_hdr(skb);
1294                 if (ipv6_addr_any(&ip6hdr->saddr)) {
1295                         do_tx_balance = 0;
1296                         break;
1297                 }
1298
1299                 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1300                 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1301                 break;
1302         case ETH_P_IPX:
1303                 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1304                         /* something is wrong with this packet */
1305                         do_tx_balance = 0;
1306                         break;
1307                 }
1308
1309                 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1310                         /* The only protocol worth balancing in
1311                          * this family since it has an "ARP" like
1312                          * mechanism
1313                          */
1314                         do_tx_balance = 0;
1315                         break;
1316                 }
1317
1318                 hash_start = (char*)eth_data->h_dest;
1319                 hash_size = ETH_ALEN;
1320                 break;
1321         case ETH_P_ARP:
1322                 do_tx_balance = 0;
1323                 if (bond_info->rlb_enabled) {
1324                         tx_slave = rlb_arp_xmit(skb, bond);
1325                 }
1326                 break;
1327         default:
1328                 do_tx_balance = 0;
1329                 break;
1330         }
1331
1332         if (do_tx_balance) {
1333                 hash_index = _simple_hash(hash_start, hash_size);
1334                 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1335         }
1336
1337         if (!tx_slave) {
1338                 /* unbalanced or unassigned, send through primary */
1339                 tx_slave = bond->curr_active_slave;
1340                 bond_info->unbalanced_load += skb->len;
1341         }
1342
1343         if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1344                 if (tx_slave != bond->curr_active_slave) {
1345                         memcpy(eth_data->h_source,
1346                                tx_slave->dev->dev_addr,
1347                                ETH_ALEN);
1348                 }
1349
1350                 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1351         } else {
1352                 if (tx_slave) {
1353                         _lock_tx_hashtbl(bond);
1354                         __tlb_clear_slave(bond, tx_slave, 0);
1355                         _unlock_tx_hashtbl(bond);
1356                 }
1357         }
1358
1359         if (res) {
1360                 /* no suitable interface, frame not sent */
1361                 dev_kfree_skb(skb);
1362         }
1363         read_unlock(&bond->curr_slave_lock);
1364
1365         return NETDEV_TX_OK;
1366 }
1367
1368 void bond_alb_monitor(struct work_struct *work)
1369 {
1370         struct bonding *bond = container_of(work, struct bonding,
1371                                             alb_work.work);
1372         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1373         struct slave *slave;
1374         int i;
1375
1376         read_lock(&bond->lock);
1377
1378         if (bond->slave_cnt == 0) {
1379                 bond_info->tx_rebalance_counter = 0;
1380                 bond_info->lp_counter = 0;
1381                 goto re_arm;
1382         }
1383
1384         bond_info->tx_rebalance_counter++;
1385         bond_info->lp_counter++;
1386
1387         /* send learning packets */
1388         if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1389                 /* change of curr_active_slave involves swapping of mac addresses.
1390                  * in order to avoid this swapping from happening while
1391                  * sending the learning packets, the curr_slave_lock must be held for
1392                  * read.
1393                  */
1394                 read_lock(&bond->curr_slave_lock);
1395
1396                 bond_for_each_slave(bond, slave, i) {
1397                         alb_send_learning_packets(slave, slave->dev->dev_addr);
1398                 }
1399
1400                 read_unlock(&bond->curr_slave_lock);
1401
1402                 bond_info->lp_counter = 0;
1403         }
1404
1405         /* rebalance tx traffic */
1406         if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1407
1408                 read_lock(&bond->curr_slave_lock);
1409
1410                 bond_for_each_slave(bond, slave, i) {
1411                         tlb_clear_slave(bond, slave, 1);
1412                         if (slave == bond->curr_active_slave) {
1413                                 SLAVE_TLB_INFO(slave).load =
1414                                         bond_info->unbalanced_load /
1415                                                 BOND_TLB_REBALANCE_INTERVAL;
1416                                 bond_info->unbalanced_load = 0;
1417                         }
1418                 }
1419
1420                 read_unlock(&bond->curr_slave_lock);
1421
1422                 bond_info->tx_rebalance_counter = 0;
1423         }
1424
1425         /* handle rlb stuff */
1426         if (bond_info->rlb_enabled) {
1427                 if (bond_info->primary_is_promisc &&
1428                     (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1429
1430                         /*
1431                          * dev_set_promiscuity requires rtnl and
1432                          * nothing else.  Avoid race with bond_close.
1433                          */
1434                         read_unlock(&bond->lock);
1435                         if (!rtnl_trylock()) {
1436                                 read_lock(&bond->lock);
1437                                 goto re_arm;
1438                         }
1439
1440                         bond_info->rlb_promisc_timeout_counter = 0;
1441
1442                         /* If the primary was set to promiscuous mode
1443                          * because a slave was disabled then
1444                          * it can now leave promiscuous mode.
1445                          */
1446                         dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1447                         bond_info->primary_is_promisc = 0;
1448
1449                         rtnl_unlock();
1450                         read_lock(&bond->lock);
1451                 }
1452
1453                 if (bond_info->rlb_rebalance) {
1454                         bond_info->rlb_rebalance = 0;
1455                         rlb_rebalance(bond);
1456                 }
1457
1458                 /* check if clients need updating */
1459                 if (bond_info->rx_ntt) {
1460                         if (bond_info->rlb_update_delay_counter) {
1461                                 --bond_info->rlb_update_delay_counter;
1462                         } else {
1463                                 rlb_update_rx_clients(bond);
1464                                 if (bond_info->rlb_update_retry_counter) {
1465                                         --bond_info->rlb_update_retry_counter;
1466                                 } else {
1467                                         bond_info->rx_ntt = 0;
1468                                 }
1469                         }
1470                 }
1471         }
1472
1473 re_arm:
1474         queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1475
1476         read_unlock(&bond->lock);
1477 }
1478
1479 /* assumption: called before the slave is attached to the bond
1480  * and not locked by the bond lock
1481  */
1482 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1483 {
1484         int res;
1485
1486         res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr);
1487         if (res) {
1488                 return res;
1489         }
1490
1491         /* caller must hold the bond lock for write since the mac addresses
1492          * are compared and may be swapped.
1493          */
1494         read_lock(&bond->lock);
1495
1496         res = alb_handle_addr_collision_on_attach(bond, slave);
1497
1498         read_unlock(&bond->lock);
1499
1500         if (res) {
1501                 return res;
1502         }
1503
1504         tlb_init_slave(slave);
1505
1506         /* order a rebalance ASAP */
1507         bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1508
1509         if (bond->alb_info.rlb_enabled) {
1510                 bond->alb_info.rlb_rebalance = 1;
1511         }
1512
1513         return 0;
1514 }
1515
1516 /*
1517  * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1518  * if necessary.
1519  *
1520  * Caller must hold RTNL and no other locks
1521  */
1522 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1523 {
1524         if (bond->slave_cnt > 1) {
1525                 alb_change_hw_addr_on_detach(bond, slave);
1526         }
1527
1528         tlb_clear_slave(bond, slave, 0);
1529
1530         if (bond->alb_info.rlb_enabled) {
1531                 bond->alb_info.next_rx_slave = NULL;
1532                 rlb_clear_slave(bond, slave);
1533         }
1534 }
1535
1536 /* Caller must hold bond lock for read */
1537 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1538 {
1539         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1540
1541         if (link == BOND_LINK_DOWN) {
1542                 tlb_clear_slave(bond, slave, 0);
1543                 if (bond->alb_info.rlb_enabled) {
1544                         rlb_clear_slave(bond, slave);
1545                 }
1546         } else if (link == BOND_LINK_UP) {
1547                 /* order a rebalance ASAP */
1548                 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1549                 if (bond->alb_info.rlb_enabled) {
1550                         bond->alb_info.rlb_rebalance = 1;
1551                         /* If the updelay module parameter is smaller than the
1552                          * forwarding delay of the switch the rebalance will
1553                          * not work because the rebalance arp replies will
1554                          * not be forwarded to the clients..
1555                          */
1556                 }
1557         }
1558 }
1559
1560 /**
1561  * bond_alb_handle_active_change - assign new curr_active_slave
1562  * @bond: our bonding struct
1563  * @new_slave: new slave to assign
1564  *
1565  * Set the bond->curr_active_slave to @new_slave and handle
1566  * mac address swapping and promiscuity changes as needed.
1567  *
1568  * If new_slave is NULL, caller must hold curr_slave_lock or
1569  * bond->lock for write.
1570  *
1571  * If new_slave is not NULL, caller must hold RTNL, bond->lock for
1572  * read and curr_slave_lock for write.  Processing here may sleep, so
1573  * no other locks may be held.
1574  */
1575 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1576         __releases(&bond->curr_slave_lock)
1577         __releases(&bond->lock)
1578         __acquires(&bond->lock)
1579         __acquires(&bond->curr_slave_lock)
1580 {
1581         struct slave *swap_slave;
1582         int i;
1583
1584         if (bond->curr_active_slave == new_slave) {
1585                 return;
1586         }
1587
1588         if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1589                 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1590                 bond->alb_info.primary_is_promisc = 0;
1591                 bond->alb_info.rlb_promisc_timeout_counter = 0;
1592         }
1593
1594         swap_slave = bond->curr_active_slave;
1595         bond->curr_active_slave = new_slave;
1596
1597         if (!new_slave || (bond->slave_cnt == 0)) {
1598                 return;
1599         }
1600
1601         /* set the new curr_active_slave to the bonds mac address
1602          * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1603          */
1604         if (!swap_slave) {
1605                 struct slave *tmp_slave;
1606                 /* find slave that is holding the bond's mac address */
1607                 bond_for_each_slave(bond, tmp_slave, i) {
1608                         if (!compare_ether_addr_64bits(tmp_slave->dev->dev_addr,
1609                                                        bond->dev->dev_addr)) {
1610                                 swap_slave = tmp_slave;
1611                                 break;
1612                         }
1613                 }
1614         }
1615
1616         /*
1617          * Arrange for swap_slave and new_slave to temporarily be
1618          * ignored so we can mess with their MAC addresses without
1619          * fear of interference from transmit activity.
1620          */
1621         if (swap_slave) {
1622                 tlb_clear_slave(bond, swap_slave, 1);
1623         }
1624         tlb_clear_slave(bond, new_slave, 1);
1625
1626         write_unlock_bh(&bond->curr_slave_lock);
1627         read_unlock(&bond->lock);
1628
1629         ASSERT_RTNL();
1630
1631         /* curr_active_slave must be set before calling alb_swap_mac_addr */
1632         if (swap_slave) {
1633                 /* swap mac address */
1634                 alb_swap_mac_addr(bond, swap_slave, new_slave);
1635         } else {
1636                 /* set the new_slave to the bond mac address */
1637                 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr);
1638         }
1639
1640         if (swap_slave) {
1641                 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1642                 read_lock(&bond->lock);
1643         } else {
1644                 read_lock(&bond->lock);
1645                 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1646         }
1647
1648         write_lock_bh(&bond->curr_slave_lock);
1649 }
1650
1651 /*
1652  * Called with RTNL
1653  */
1654 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1655         __acquires(&bond->lock)
1656         __releases(&bond->lock)
1657 {
1658         struct bonding *bond = netdev_priv(bond_dev);
1659         struct sockaddr *sa = addr;
1660         struct slave *slave, *swap_slave;
1661         int res;
1662         int i;
1663
1664         if (!is_valid_ether_addr(sa->sa_data)) {
1665                 return -EADDRNOTAVAIL;
1666         }
1667
1668         res = alb_set_mac_address(bond, addr);
1669         if (res) {
1670                 return res;
1671         }
1672
1673         memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1674
1675         /* If there is no curr_active_slave there is nothing else to do.
1676          * Otherwise we'll need to pass the new address to it and handle
1677          * duplications.
1678          */
1679         if (!bond->curr_active_slave) {
1680                 return 0;
1681         }
1682
1683         swap_slave = NULL;
1684
1685         bond_for_each_slave(bond, slave, i) {
1686                 if (!compare_ether_addr_64bits(slave->dev->dev_addr,
1687                                                bond_dev->dev_addr)) {
1688                         swap_slave = slave;
1689                         break;
1690                 }
1691         }
1692
1693         if (swap_slave) {
1694                 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1695                 alb_fasten_mac_swap(bond, swap_slave, bond->curr_active_slave);
1696         } else {
1697                 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr);
1698
1699                 read_lock(&bond->lock);
1700                 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1701                 if (bond->alb_info.rlb_enabled) {
1702                         /* inform clients mac address has changed */
1703                         rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1704                 }
1705                 read_unlock(&bond->lock);
1706         }
1707
1708         return 0;
1709 }
1710
1711 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1712 {
1713         if (bond->alb_info.current_alb_vlan &&
1714             (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1715                 bond->alb_info.current_alb_vlan = NULL;
1716         }
1717
1718         if (bond->alb_info.rlb_enabled) {
1719                 rlb_clear_vlan(bond, vlan_id);
1720         }
1721 }
1722