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 #define ALB_TIMER_TICKS_PER_SEC     10  /* should be a divisor of HZ */
  48 #define BOND_TLB_REBALANCE_INTERVAL 10  /* In seconds, periodic re-balancing.
  49                                          * Used for division - never set
  50                                          * to zero !!!
  51                                          */
  52 #define BOND_ALB_LP_INTERVAL        1   /* In seconds, periodic send of
  53                                          * learning packets to the switch
  54                                          */
  55
  56 #define BOND_TLB_REBALANCE_TICKS (BOND_TLB_REBALANCE_INTERVAL \
  57                                   * ALB_TIMER_TICKS_PER_SEC)
  58
  59 #define BOND_ALB_LP_TICKS (BOND_ALB_LP_INTERVAL \
  60                            * ALB_TIMER_TICKS_PER_SEC)
  61
  62 #define TLB_HASH_TABLE_SIZE 256 /* The size of the clients hash table.
  63                                  * Note that this value MUST NOT be smaller
  64                                  * because the key hash table is BYTE wide !
  65                                  */
  66
  67
  68 #define TLB_NULL_INDEX          0xffffffff
  69 #define MAX_LP_BURST            3
  70
  71 /* rlb defs */
  72 #define RLB_HASH_TABLE_SIZE     256
  73 #define RLB_NULL_INDEX          0xffffffff
  74 #define RLB_UPDATE_DELAY        2*ALB_TIMER_TICKS_PER_SEC /* 2 seconds */
  75 #define RLB_ARP_BURST_SIZE      2
  76 #define RLB_UPDATE_RETRY        3       /* 3-ticks - must be smaller than the rlb
  77                                          * rebalance interval (5 min).
  78                                          */
  79 /* RLB_PROMISC_TIMEOUT = 10 sec equals the time that the current slave is
  80  * promiscuous after failover
  81  */
  82 #define RLB_PROMISC_TIMEOUT     10*ALB_TIMER_TICKS_PER_SEC
  83
  84 #ifndef __long_aligned
  85 #define __long_aligned __attribute__((aligned((sizeof(long)))))
  86 #endif
  87 static const u8 mac_bcast[ETH_ALEN] __long_aligned = {
  88         0xff, 0xff, 0xff, 0xff, 0xff, 0xff
  89 };
  90 static const u8 mac_v6_allmcast[ETH_ALEN] __long_aligned = {
  91         0x33, 0x33, 0x00, 0x00, 0x00, 0x01
  92 };
  93 static const int alb_delta_in_ticks = HZ / ALB_TIMER_TICKS_PER_SEC;
  94
  95 #pragma pack(1)
  96 struct learning_pkt {
  97         u8 mac_dst[ETH_ALEN];
  98         u8 mac_src[ETH_ALEN];
  99         __be16 type;
 100         u8 padding[ETH_ZLEN - ETH_HLEN];
 101 };
 102
 103 struct arp_pkt {
 104         __be16  hw_addr_space;
 105         __be16  prot_addr_space;
 106         u8      hw_addr_len;
 107         u8      prot_addr_len;
 108         __be16  op_code;
 109         u8      mac_src[ETH_ALEN];      /* sender hardware address */
 110         __be32  ip_src;                 /* sender IP address */
 111         u8      mac_dst[ETH_ALEN];      /* target hardware address */
 112         __be32  ip_dst;                 /* target IP address */
 113 };
 114 #pragma pack()
 115
 116 static inline struct arp_pkt *arp_pkt(const struct sk_buff *skb)
 117 {
 118         return (struct arp_pkt *)skb_network_header(skb);
 119 }
 120
 121 /* Forward declaration */
 122 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[]);
 123
 124 static inline u8 _simple_hash(const u8 *hash_start, int hash_size)
 125 {
 126         int i;
 127         u8 hash = 0;
 128
 129         for (i = 0; i < hash_size; i++) {
 130                 hash ^= hash_start[i];
 131         }
 132
 133         return hash;
 134 }
 135
 136 /*********************** tlb specific functions ***************************/
 137
 138 static inline void _lock_tx_hashtbl(struct bonding *bond)
 139 {
 140         spin_lock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 141 }
 142
 143 static inline void _unlock_tx_hashtbl(struct bonding *bond)
 144 {
 145         spin_unlock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
 146 }
 147
 148 /* Caller must hold tx_hashtbl lock */
 149 static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
 150 {
 151         if (save_load) {
 152                 entry->load_history = 1 + entry->tx_bytes /
 153                                       BOND_TLB_REBALANCE_INTERVAL;
 154                 entry->tx_bytes = 0;
 155         }
 156
 157         entry->tx_slave = NULL;
 158         entry->next = TLB_NULL_INDEX;
 159         entry->prev = TLB_NULL_INDEX;
 160 }
 161
 162 static inline void tlb_init_slave(struct slave *slave)
 163 {
 164         SLAVE_TLB_INFO(slave).load = 0;
 165         SLAVE_TLB_INFO(slave).head = TLB_NULL_INDEX;
 166 }
 167
 168 /* Caller must hold bond lock for read */
 169 static void tlb_clear_slave(struct bonding *bond, struct slave *slave, int save_load)
 170 {
 171         struct tlb_client_info *tx_hash_table;
 172         u32 index;
 173
 174         _lock_tx_hashtbl(bond);
 175
 176         /* clear slave from tx_hashtbl */
 177         tx_hash_table = BOND_ALB_INFO(bond).tx_hashtbl;
 178
 179         /* skip this if we've already freed the tx hash table */
 180         if (tx_hash_table) {
 181                 index = SLAVE_TLB_INFO(slave).head;
 182                 while (index != TLB_NULL_INDEX) {
 183                         u32 next_index = tx_hash_table[index].next;
 184                         tlb_init_table_entry(&tx_hash_table[index], save_load);
 185                         index = next_index;
 186                 }
 187         }
 188
 189         tlb_init_slave(slave);
 190
 191         _unlock_tx_hashtbl(bond);
 192 }
 193
 194 /* Must be called before starting the monitor timer */
 195 static int tlb_initialize(struct bonding *bond)
 196 {
 197         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 198         int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
 199         struct tlb_client_info *new_hashtbl;
 200         int i;
 201
 202         spin_lock_init(&(bond_info->tx_hashtbl_lock));
 203
 204         new_hashtbl = kzalloc(size, GFP_KERNEL);
 205         if (!new_hashtbl) {
 206                 pr_err("%s: Error: Failed to allocate TLB hash table\n",
 207                        bond->dev->name);
 208                 return -1;
 209         }
 210         _lock_tx_hashtbl(bond);
 211
 212         bond_info->tx_hashtbl = new_hashtbl;
 213
 214         for (i = 0; i < TLB_HASH_TABLE_SIZE; i++) {
 215                 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 1);
 216         }
 217
 218         _unlock_tx_hashtbl(bond);
 219
 220         return 0;
 221 }
 222
 223 /* Must be called only after all slaves have been released */
 224 static void tlb_deinitialize(struct bonding *bond)
 225 {
 226         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 227
 228         _lock_tx_hashtbl(bond);
 229
 230         kfree(bond_info->tx_hashtbl);
 231         bond_info->tx_hashtbl = NULL;
 232
 233         _unlock_tx_hashtbl(bond);
 234 }
 235
 236 static long long compute_gap(struct slave *slave)
 237 {
 238         return (s64) (slave->speed << 20) - /* Convert to Megabit per sec */
 239                (s64) (SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
 240 }
 241
 242 /* Caller must hold bond lock for read */
 243 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
 244 {
 245         struct slave *slave, *least_loaded;
 246         long long max_gap;
 247         int i;
 248
 249         least_loaded = NULL;
 250         max_gap = LLONG_MIN;
 251
 252         /* Find the slave with the largest gap */
 253         bond_for_each_slave(bond, slave, i) {
 254                 if (SLAVE_IS_OK(slave)) {
 255                         long long gap = compute_gap(slave);
 256
 257                         if (max_gap < gap) {
 258                                 least_loaded = slave;
 259                                 max_gap = gap;
 260                         }
 261                 }
 262         }
 263
 264         return least_loaded;
 265 }
 266
 267 /* Caller must hold bond lock for read */
 268 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index, u32 skb_len)
 269 {
 270         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 271         struct tlb_client_info *hash_table;
 272         struct slave *assigned_slave;
 273
 274         _lock_tx_hashtbl(bond);
 275
 276         hash_table = bond_info->tx_hashtbl;
 277         assigned_slave = hash_table[hash_index].tx_slave;
 278         if (!assigned_slave) {
 279                 assigned_slave = tlb_get_least_loaded_slave(bond);
 280
 281                 if (assigned_slave) {
 282                         struct tlb_slave_info *slave_info =
 283                                 &(SLAVE_TLB_INFO(assigned_slave));
 284                         u32 next_index = slave_info->head;
 285
 286                         hash_table[hash_index].tx_slave = assigned_slave;
 287                         hash_table[hash_index].next = next_index;
 288                         hash_table[hash_index].prev = TLB_NULL_INDEX;
 289
 290                         if (next_index != TLB_NULL_INDEX) {
 291                                 hash_table[next_index].prev = hash_index;
 292                         }
 293
 294                         slave_info->head = hash_index;
 295                         slave_info->load +=
 296                                 hash_table[hash_index].load_history;
 297                 }
 298         }
 299
 300         if (assigned_slave) {
 301                 hash_table[hash_index].tx_bytes += skb_len;
 302         }
 303
 304         _unlock_tx_hashtbl(bond);
 305
 306         return assigned_slave;
 307 }
 308
 309 /*********************** rlb specific functions ***************************/
 310 static inline void _lock_rx_hashtbl(struct bonding *bond)
 311 {
 312         spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 313 }
 314
 315 static inline void _unlock_rx_hashtbl(struct bonding *bond)
 316 {
 317         spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
 318 }
 319
 320 /* when an ARP REPLY is received from a client update its info
 321  * in the rx_hashtbl
 322  */
 323 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
 324 {
 325         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 326         struct rlb_client_info *client_info;
 327         u32 hash_index;
 328
 329         _lock_rx_hashtbl(bond);
 330
 331         hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
 332         client_info = &(bond_info->rx_hashtbl[hash_index]);
 333
 334         if ((client_info->assigned) &&
 335             (client_info->ip_src == arp->ip_dst) &&
 336             (client_info->ip_dst == arp->ip_src) &&
 337             (compare_ether_addr_64bits(client_info->mac_dst, arp->mac_src))) {
 338                 /* update the clients MAC address */
 339                 memcpy(client_info->mac_dst, arp->mac_src, ETH_ALEN);
 340                 client_info->ntt = 1;
 341                 bond_info->rx_ntt = 1;
 342         }
 343
 344         _unlock_rx_hashtbl(bond);
 345 }
 346
 347 static int rlb_arp_recv(struct sk_buff *skb, struct net_device *bond_dev, struct packet_type *ptype, struct net_device *orig_dev)
 348 {
 349         struct bonding *bond;
 350         struct arp_pkt *arp = (struct arp_pkt *)skb->data;
 351         int res = NET_RX_DROP;
 352
 353         while (bond_dev->priv_flags & IFF_802_1Q_VLAN)
 354                 bond_dev = vlan_dev_real_dev(bond_dev);
 355
 356         if (!(bond_dev->priv_flags & IFF_BONDING) ||
 357             !(bond_dev->flags & IFF_MASTER))
 358                 goto out;
 359
 360         if (!arp) {
 361                 pr_debug("Packet has no ARP data\n");
 362                 goto out;
 363         }
 364
 365         if (!pskb_may_pull(skb, arp_hdr_len(bond_dev)))
 366                 goto out;
 367
 368         if (skb->len < sizeof(struct arp_pkt)) {
 369                 pr_debug("Packet is too small to be an ARP\n");
 370                 goto out;
 371         }
 372
 373         if (arp->op_code == htons(ARPOP_REPLY)) {
 374                 /* update rx hash table for this ARP */
 375                 bond = netdev_priv(bond_dev);
 376                 rlb_update_entry_from_arp(bond, arp);
 377                 pr_debug("Server received an ARP Reply from client\n");
 378         }
 379
 380         res = NET_RX_SUCCESS;
 381
 382 out:
 383         dev_kfree_skb(skb);
 384
 385         return res;
 386 }
 387
 388 /* Caller must hold bond lock for read */
 389 static struct slave *rlb_next_rx_slave(struct bonding *bond)
 390 {
 391         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 392         struct slave *rx_slave, *slave, *start_at;
 393         int i = 0;
 394
 395         if (bond_info->next_rx_slave) {
 396                 start_at = bond_info->next_rx_slave;
 397         } else {
 398                 start_at = bond->first_slave;
 399         }
 400
 401         rx_slave = NULL;
 402
 403         bond_for_each_slave_from(bond, slave, i, start_at) {
 404                 if (SLAVE_IS_OK(slave)) {
 405                         if (!rx_slave) {
 406                                 rx_slave = slave;
 407                         } else if (slave->speed > rx_slave->speed) {
 408                                 rx_slave = slave;
 409                         }
 410                 }
 411         }
 412
 413         if (rx_slave) {
 414                 bond_info->next_rx_slave = rx_slave->next;
 415         }
 416
 417         return rx_slave;
 418 }
 419
 420 /* teach the switch the mac of a disabled slave
 421  * on the primary for fault tolerance
 422  *
 423  * Caller must hold bond->curr_slave_lock for write or bond lock for write
 424  */
 425 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
 426 {
 427         if (!bond->curr_active_slave) {
 428                 return;
 429         }
 430
 431         if (!bond->alb_info.primary_is_promisc) {
 432                 if (!dev_set_promiscuity(bond->curr_active_slave->dev, 1))
 433                         bond->alb_info.primary_is_promisc = 1;
 434                 else
 435                         bond->alb_info.primary_is_promisc = 0;
 436         }
 437
 438         bond->alb_info.rlb_promisc_timeout_counter = 0;
 439
 440         alb_send_learning_packets(bond->curr_active_slave, addr);
 441 }
 442
 443 /* slave being removed should not be active at this point
 444  *
 445  * Caller must hold bond lock for read
 446  */
 447 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
 448 {
 449         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 450         struct rlb_client_info *rx_hash_table;
 451         u32 index, next_index;
 452
 453         /* clear slave from rx_hashtbl */
 454         _lock_rx_hashtbl(bond);
 455
 456         rx_hash_table = bond_info->rx_hashtbl;
 457         index = bond_info->rx_hashtbl_head;
 458         for (; index != RLB_NULL_INDEX; index = next_index) {
 459                 next_index = rx_hash_table[index].next;
 460                 if (rx_hash_table[index].slave == slave) {
 461                         struct slave *assigned_slave = rlb_next_rx_slave(bond);
 462
 463                         if (assigned_slave) {
 464                                 rx_hash_table[index].slave = assigned_slave;
 465                                 if (compare_ether_addr_64bits(rx_hash_table[index].mac_dst,
 466                                                               mac_bcast)) {
 467                                         bond_info->rx_hashtbl[index].ntt = 1;
 468                                         bond_info->rx_ntt = 1;
 469                                         /* A slave has been removed from the
 470                                          * table because it is either disabled
 471                                          * or being released. We must retry the
 472                                          * update to avoid clients from not
 473                                          * being updated & disconnecting when
 474                                          * there is stress
 475                                          */
 476                                         bond_info->rlb_update_retry_counter =
 477                                                 RLB_UPDATE_RETRY;
 478                                 }
 479                         } else {  /* there is no active slave */
 480                                 rx_hash_table[index].slave = NULL;
 481                         }
 482                 }
 483         }
 484
 485         _unlock_rx_hashtbl(bond);
 486
 487         write_lock_bh(&bond->curr_slave_lock);
 488
 489         if (slave != bond->curr_active_slave) {
 490                 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
 491         }
 492
 493         write_unlock_bh(&bond->curr_slave_lock);
 494 }
 495
 496 static void rlb_update_client(struct rlb_client_info *client_info)
 497 {
 498         int i;
 499
 500         if (!client_info->slave) {
 501                 return;
 502         }
 503
 504         for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
 505                 struct sk_buff *skb;
 506
 507                 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
 508                                  client_info->ip_dst,
 509                                  client_info->slave->dev,
 510                                  client_info->ip_src,
 511                                  client_info->mac_dst,
 512                                  client_info->slave->dev->dev_addr,
 513                                  client_info->mac_dst);
 514                 if (!skb) {
 515                         pr_err("%s: Error: failed to create an ARP packet\n",
 516                                client_info->slave->dev->master->name);
 517                         continue;
 518                 }
 519
 520                 skb->dev = client_info->slave->dev;
 521
 522                 if (client_info->tag) {
 523                         skb = vlan_put_tag(skb, client_info->vlan_id);
 524                         if (!skb) {
 525                                 pr_err("%s: Error: failed to insert VLAN tag\n",
 526                                        client_info->slave->dev->master->name);
 527                                 continue;
 528                         }
 529                 }
 530
 531                 arp_xmit(skb);
 532         }
 533 }
 534
 535 /* sends ARP REPLIES that update the clients that need updating */
 536 static void rlb_update_rx_clients(struct bonding *bond)
 537 {
 538         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 539         struct rlb_client_info *client_info;
 540         u32 hash_index;
 541
 542         _lock_rx_hashtbl(bond);
 543
 544         hash_index = bond_info->rx_hashtbl_head;
 545         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 546                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 547                 if (client_info->ntt) {
 548                         rlb_update_client(client_info);
 549                         if (bond_info->rlb_update_retry_counter == 0) {
 550                                 client_info->ntt = 0;
 551                         }
 552                 }
 553         }
 554
 555         /* do not update the entries again until this counter is zero so that
 556          * not to confuse the clients.
 557          */
 558         bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
 559
 560         _unlock_rx_hashtbl(bond);
 561 }
 562
 563 /* The slave was assigned a new mac address - update the clients */
 564 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
 565 {
 566         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 567         struct rlb_client_info *client_info;
 568         int ntt = 0;
 569         u32 hash_index;
 570
 571         _lock_rx_hashtbl(bond);
 572
 573         hash_index = bond_info->rx_hashtbl_head;
 574         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 575                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 576
 577                 if ((client_info->slave == slave) &&
 578                     compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 579                         client_info->ntt = 1;
 580                         ntt = 1;
 581                 }
 582         }
 583
 584         // update the team's flag only after the whole iteration
 585         if (ntt) {
 586                 bond_info->rx_ntt = 1;
 587                 //fasten the change
 588                 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
 589         }
 590
 591         _unlock_rx_hashtbl(bond);
 592 }
 593
 594 /* mark all clients using src_ip to be updated */
 595 static void rlb_req_update_subnet_clients(struct bonding *bond, __be32 src_ip)
 596 {
 597         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 598         struct rlb_client_info *client_info;
 599         u32 hash_index;
 600
 601         _lock_rx_hashtbl(bond);
 602
 603         hash_index = bond_info->rx_hashtbl_head;
 604         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 605                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 606
 607                 if (!client_info->slave) {
 608                         pr_err("%s: Error: found a client with no channel in the client's hash table\n",
 609                                bond->dev->name);
 610                         continue;
 611                 }
 612                 /*update all clients using this src_ip, that are not assigned
 613                  * to the team's address (curr_active_slave) and have a known
 614                  * unicast mac address.
 615                  */
 616                 if ((client_info->ip_src == src_ip) &&
 617                     compare_ether_addr_64bits(client_info->slave->dev->dev_addr,
 618                            bond->dev->dev_addr) &&
 619                     compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 620                         client_info->ntt = 1;
 621                         bond_info->rx_ntt = 1;
 622                 }
 623         }
 624
 625         _unlock_rx_hashtbl(bond);
 626 }
 627
 628 /* Caller must hold both bond and ptr locks for read */
 629 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
 630 {
 631         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 632         struct arp_pkt *arp = arp_pkt(skb);
 633         struct slave *assigned_slave;
 634         struct rlb_client_info *client_info;
 635         u32 hash_index = 0;
 636
 637         _lock_rx_hashtbl(bond);
 638
 639         hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_src));
 640         client_info = &(bond_info->rx_hashtbl[hash_index]);
 641
 642         if (client_info->assigned) {
 643                 if ((client_info->ip_src == arp->ip_src) &&
 644                     (client_info->ip_dst == arp->ip_dst)) {
 645                         /* the entry is already assigned to this client */
 646                         if (compare_ether_addr_64bits(arp->mac_dst, mac_bcast)) {
 647                                 /* update mac address from arp */
 648                                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 649                         }
 650
 651                         assigned_slave = client_info->slave;
 652                         if (assigned_slave) {
 653                                 _unlock_rx_hashtbl(bond);
 654                                 return assigned_slave;
 655                         }
 656                 } else {
 657                         /* the entry is already assigned to some other client,
 658                          * move the old client to primary (curr_active_slave) so
 659                          * that the new client can be assigned to this entry.
 660                          */
 661                         if (bond->curr_active_slave &&
 662                             client_info->slave != bond->curr_active_slave) {
 663                                 client_info->slave = bond->curr_active_slave;
 664                                 rlb_update_client(client_info);
 665                         }
 666                 }
 667         }
 668         /* assign a new slave */
 669         assigned_slave = rlb_next_rx_slave(bond);
 670
 671         if (assigned_slave) {
 672                 client_info->ip_src = arp->ip_src;
 673                 client_info->ip_dst = arp->ip_dst;
 674                 /* arp->mac_dst is broadcast for arp reqeusts.
 675                  * will be updated with clients actual unicast mac address
 676                  * upon receiving an arp reply.
 677                  */
 678                 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
 679                 client_info->slave = assigned_slave;
 680
 681                 if (compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
 682                         client_info->ntt = 1;
 683                         bond->alb_info.rx_ntt = 1;
 684                 } else {
 685                         client_info->ntt = 0;
 686                 }
 687
 688                 if (bond->vlgrp) {
 689                         if (!vlan_get_tag(skb, &client_info->vlan_id))
 690                                 client_info->tag = 1;
 691                 }
 692
 693                 if (!client_info->assigned) {
 694                         u32 prev_tbl_head = bond_info->rx_hashtbl_head;
 695                         bond_info->rx_hashtbl_head = hash_index;
 696                         client_info->next = prev_tbl_head;
 697                         if (prev_tbl_head != RLB_NULL_INDEX) {
 698                                 bond_info->rx_hashtbl[prev_tbl_head].prev =
 699                                         hash_index;
 700                         }
 701                         client_info->assigned = 1;
 702                 }
 703         }
 704
 705         _unlock_rx_hashtbl(bond);
 706
 707         return assigned_slave;
 708 }
 709
 710 /* chooses (and returns) transmit channel for arp reply
 711  * does not choose channel for other arp types since they are
 712  * sent on the curr_active_slave
 713  */
 714 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
 715 {
 716         struct arp_pkt *arp = arp_pkt(skb);
 717         struct slave *tx_slave = NULL;
 718
 719         if (arp->op_code == htons(ARPOP_REPLY)) {
 720                 /* the arp must be sent on the selected
 721                 * rx channel
 722                 */
 723                 tx_slave = rlb_choose_channel(skb, bond);
 724                 if (tx_slave) {
 725                         memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
 726                 }
 727                 pr_debug("Server sent ARP Reply packet\n");
 728         } else if (arp->op_code == htons(ARPOP_REQUEST)) {
 729                 /* Create an entry in the rx_hashtbl for this client as a
 730                  * place holder.
 731                  * When the arp reply is received the entry will be updated
 732                  * with the correct unicast address of the client.
 733                  */
 734                 rlb_choose_channel(skb, bond);
 735
 736                 /* The ARP relpy packets must be delayed so that
 737                  * they can cancel out the influence of the ARP request.
 738                  */
 739                 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
 740
 741                 /* arp requests are broadcast and are sent on the primary
 742                  * the arp request will collapse all clients on the subnet to
 743                  * the primary slave. We must register these clients to be
 744                  * updated with their assigned mac.
 745                  */
 746                 rlb_req_update_subnet_clients(bond, arp->ip_src);
 747                 pr_debug("Server sent ARP Request packet\n");
 748         }
 749
 750         return tx_slave;
 751 }
 752
 753 /* Caller must hold bond lock for read */
 754 static void rlb_rebalance(struct bonding *bond)
 755 {
 756         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 757         struct slave *assigned_slave;
 758         struct rlb_client_info *client_info;
 759         int ntt;
 760         u32 hash_index;
 761
 762         _lock_rx_hashtbl(bond);
 763
 764         ntt = 0;
 765         hash_index = bond_info->rx_hashtbl_head;
 766         for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
 767                 client_info = &(bond_info->rx_hashtbl[hash_index]);
 768                 assigned_slave = rlb_next_rx_slave(bond);
 769                 if (assigned_slave && (client_info->slave != assigned_slave)) {
 770                         client_info->slave = assigned_slave;
 771                         client_info->ntt = 1;
 772                         ntt = 1;
 773                 }
 774         }
 775
 776         /* update the team's flag only after the whole iteration */
 777         if (ntt) {
 778                 bond_info->rx_ntt = 1;
 779         }
 780         _unlock_rx_hashtbl(bond);
 781 }
 782
 783 /* Caller must hold rx_hashtbl lock */
 784 static void rlb_init_table_entry(struct rlb_client_info *entry)
 785 {
 786         memset(entry, 0, sizeof(struct rlb_client_info));
 787         entry->next = RLB_NULL_INDEX;
 788         entry->prev = RLB_NULL_INDEX;
 789 }
 790
 791 static int rlb_initialize(struct bonding *bond)
 792 {
 793         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 794         struct packet_type *pk_type = &(BOND_ALB_INFO(bond).rlb_pkt_type);
 795         struct rlb_client_info  *new_hashtbl;
 796         int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
 797         int i;
 798
 799         spin_lock_init(&(bond_info->rx_hashtbl_lock));
 800
 801         new_hashtbl = kmalloc(size, GFP_KERNEL);
 802         if (!new_hashtbl) {
 803                 pr_err("%s: Error: Failed to allocate RLB hash table\n",
 804                        bond->dev->name);
 805                 return -1;
 806         }
 807         _lock_rx_hashtbl(bond);
 808
 809         bond_info->rx_hashtbl = new_hashtbl;
 810
 811         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 812
 813         for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
 814                 rlb_init_table_entry(bond_info->rx_hashtbl + i);
 815         }
 816
 817         _unlock_rx_hashtbl(bond);
 818
 819         /*initialize packet type*/
 820         pk_type->type = cpu_to_be16(ETH_P_ARP);
 821         pk_type->dev = bond->dev;
 822         pk_type->func = rlb_arp_recv;
 823
 824         /* register to receive ARPs */
 825         dev_add_pack(pk_type);
 826
 827         return 0;
 828 }
 829
 830 static void rlb_deinitialize(struct bonding *bond)
 831 {
 832         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 833
 834         dev_remove_pack(&(bond_info->rlb_pkt_type));
 835
 836         _lock_rx_hashtbl(bond);
 837
 838         kfree(bond_info->rx_hashtbl);
 839         bond_info->rx_hashtbl = NULL;
 840         bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
 841
 842         _unlock_rx_hashtbl(bond);
 843 }
 844
 845 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
 846 {
 847         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
 848         u32 curr_index;
 849
 850         _lock_rx_hashtbl(bond);
 851
 852         curr_index = bond_info->rx_hashtbl_head;
 853         while (curr_index != RLB_NULL_INDEX) {
 854                 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
 855                 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
 856                 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
 857
 858                 if (curr->tag && (curr->vlan_id == vlan_id)) {
 859                         if (curr_index == bond_info->rx_hashtbl_head) {
 860                                 bond_info->rx_hashtbl_head = next_index;
 861                         }
 862                         if (prev_index != RLB_NULL_INDEX) {
 863                                 bond_info->rx_hashtbl[prev_index].next = next_index;
 864                         }
 865                         if (next_index != RLB_NULL_INDEX) {
 866                                 bond_info->rx_hashtbl[next_index].prev = prev_index;
 867                         }
 868
 869                         rlb_init_table_entry(curr);
 870                 }
 871
 872                 curr_index = next_index;
 873         }
 874
 875         _unlock_rx_hashtbl(bond);
 876 }
 877
 878 /*********************** tlb/rlb shared functions *********************/
 879
 880 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
 881 {
 882         struct bonding *bond = bond_get_bond_by_slave(slave);
 883         struct learning_pkt pkt;
 884         int size = sizeof(struct learning_pkt);
 885         int i;
 886
 887         memset(&pkt, 0, size);
 888         memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
 889         memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
 890         pkt.type = cpu_to_be16(ETH_P_LOOP);
 891
 892         for (i = 0; i < MAX_LP_BURST; i++) {
 893                 struct sk_buff *skb;
 894                 char *data;
 895
 896                 skb = dev_alloc_skb(size);
 897                 if (!skb) {
 898                         return;
 899                 }
 900
 901                 data = skb_put(skb, size);
 902                 memcpy(data, &pkt, size);
 903
 904                 skb_reset_mac_header(skb);
 905                 skb->network_header = skb->mac_header + ETH_HLEN;
 906                 skb->protocol = pkt.type;
 907                 skb->priority = TC_PRIO_CONTROL;
 908                 skb->dev = slave->dev;
 909
 910                 if (bond->vlgrp) {
 911                         struct vlan_entry *vlan;
 912
 913                         vlan = bond_next_vlan(bond,
 914                                               bond->alb_info.current_alb_vlan);
 915
 916                         bond->alb_info.current_alb_vlan = vlan;
 917                         if (!vlan) {
 918                                 kfree_skb(skb);
 919                                 continue;
 920                         }
 921
 922                         skb = vlan_put_tag(skb, vlan->vlan_id);
 923                         if (!skb) {
 924                                 pr_err("%s: Error: failed to insert VLAN tag\n",
 925                                        bond->dev->name);
 926                                 continue;
 927                         }
 928                 }
 929
 930                 dev_queue_xmit(skb);
 931         }
 932 }
 933
 934 /* hw is a boolean parameter that determines whether we should try and
 935  * set the hw address of the device as well as the hw address of the
 936  * net_device
 937  */
 938 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[], int hw)
 939 {
 940         struct net_device *dev = slave->dev;
 941         struct sockaddr s_addr;
 942
 943         if (!hw) {
 944                 memcpy(dev->dev_addr, addr, dev->addr_len);
 945                 return 0;
 946         }
 947
 948         /* for rlb each slave must have a unique hw mac addresses so that */
 949         /* each slave will receive packets destined to a different mac */
 950         memcpy(s_addr.sa_data, addr, dev->addr_len);
 951         s_addr.sa_family = dev->type;
 952         if (dev_set_mac_address(dev, &s_addr)) {
 953                 pr_err("%s: Error: dev_set_mac_address of dev %s failed!\n"
 954                        "ALB mode requires that the base driver support setting the hw address also when the network device's interface is open\n",
 955                        dev->master->name, dev->name);
 956                 return -EOPNOTSUPP;
 957         }
 958         return 0;
 959 }
 960
 961 /*
 962  * Swap MAC addresses between two slaves.
 963  *
 964  * Called with RTNL held, and no other locks.
 965  *
 966  */
 967
 968 static void alb_swap_mac_addr(struct bonding *bond, struct slave *slave1, struct slave *slave2)
 969 {
 970         u8 tmp_mac_addr[ETH_ALEN];
 971
 972         memcpy(tmp_mac_addr, slave1->dev->dev_addr, ETH_ALEN);
 973         alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr, bond->alb_info.rlb_enabled);
 974         alb_set_slave_mac_addr(slave2, tmp_mac_addr, bond->alb_info.rlb_enabled);
 975
 976 }
 977
 978 /*
 979  * Send learning packets after MAC address swap.
 980  *
 981  * Called with RTNL and no other locks
 982  */
 983 static void alb_fasten_mac_swap(struct bonding *bond, struct slave *slave1,
 984                                 struct slave *slave2)
 985 {
 986         int slaves_state_differ = (SLAVE_IS_OK(slave1) != SLAVE_IS_OK(slave2));
 987         struct slave *disabled_slave = NULL;
 988
 989         ASSERT_RTNL();
 990
 991         /* fasten the change in the switch */
 992         if (SLAVE_IS_OK(slave1)) {
 993                 alb_send_learning_packets(slave1, slave1->dev->dev_addr);
 994                 if (bond->alb_info.rlb_enabled) {
 995                         /* inform the clients that the mac address
 996                          * has changed
 997                          */
 998                         rlb_req_update_slave_clients(bond, slave1);
 999                 }
1000         } else {
1001                 disabled_slave = slave1;
1002         }
1003
1004         if (SLAVE_IS_OK(slave2)) {
1005                 alb_send_learning_packets(slave2, slave2->dev->dev_addr);
1006                 if (bond->alb_info.rlb_enabled) {
1007                         /* inform the clients that the mac address
1008                          * has changed
1009                          */
1010                         rlb_req_update_slave_clients(bond, slave2);
1011                 }
1012         } else {
1013                 disabled_slave = slave2;
1014         }
1015
1016         if (bond->alb_info.rlb_enabled && slaves_state_differ) {
1017                 /* A disabled slave was assigned an active mac addr */
1018                 rlb_teach_disabled_mac_on_primary(bond,
1019                                                   disabled_slave->dev->dev_addr);
1020         }
1021 }
1022
1023 /**
1024  * alb_change_hw_addr_on_detach
1025  * @bond: bonding we're working on
1026  * @slave: the slave that was just detached
1027  *
1028  * We assume that @slave was already detached from the slave list.
1029  *
1030  * If @slave's permanent hw address is different both from its current
1031  * address and from @bond's address, then somewhere in the bond there's
1032  * a slave that has @slave's permanet address as its current address.
1033  * We'll make sure that that slave no longer uses @slave's permanent address.
1034  *
1035  * Caller must hold RTNL and no other locks
1036  */
1037 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
1038 {
1039         int perm_curr_diff;
1040         int perm_bond_diff;
1041
1042         perm_curr_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
1043                                                    slave->dev->dev_addr);
1044         perm_bond_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
1045                                                    bond->dev->dev_addr);
1046
1047         if (perm_curr_diff && perm_bond_diff) {
1048                 struct slave *tmp_slave;
1049                 int i, found = 0;
1050
1051                 bond_for_each_slave(bond, tmp_slave, i) {
1052                         if (!compare_ether_addr_64bits(slave->perm_hwaddr,
1053                                                        tmp_slave->dev->dev_addr)) {
1054                                 found = 1;
1055                                 break;
1056                         }
1057                 }
1058
1059                 if (found) {
1060                         /* locking: needs RTNL and nothing else */
1061                         alb_swap_mac_addr(bond, slave, tmp_slave);
1062                         alb_fasten_mac_swap(bond, slave, tmp_slave);
1063                 }
1064         }
1065 }
1066
1067 /**
1068  * alb_handle_addr_collision_on_attach
1069  * @bond: bonding we're working on
1070  * @slave: the slave that was just attached
1071  *
1072  * checks uniqueness of slave's mac address and handles the case the
1073  * new slave uses the bonds mac address.
1074  *
1075  * If the permanent hw address of @slave is @bond's hw address, we need to
1076  * find a different hw address to give @slave, that isn't in use by any other
1077  * slave in the bond. This address must be, of course, one of the premanent
1078  * addresses of the other slaves.
1079  *
1080  * We go over the slave list, and for each slave there we compare its
1081  * permanent hw address with the current address of all the other slaves.
1082  * If no match was found, then we've found a slave with a permanent address
1083  * that isn't used by any other slave in the bond, so we can assign it to
1084  * @slave.
1085  *
1086  * assumption: this function is called before @slave is attached to the
1087  *             bond slave list.
1088  *
1089  * caller must hold the bond lock for write since the mac addresses are compared
1090  * and may be swapped.
1091  */
1092 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1093 {
1094         struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1095         struct slave *has_bond_addr = bond->curr_active_slave;
1096         int i, j, found = 0;
1097
1098         if (bond->slave_cnt == 0) {
1099                 /* this is the first slave */
1100                 return 0;
1101         }
1102
1103         /* if slave's mac address differs from bond's mac address
1104          * check uniqueness of slave's mac address against the other
1105          * slaves in the bond.
1106          */
1107         if (compare_ether_addr_64bits(slave->perm_hwaddr, bond->dev->dev_addr)) {
1108                 bond_for_each_slave(bond, tmp_slave1, i) {
1109                         if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1110                                                        slave->dev->dev_addr)) {
1111                                 found = 1;
1112                                 break;
1113                         }
1114                 }
1115
1116                 if (!found)
1117                         return 0;
1118
1119                 /* Try setting slave mac to bond address and fall-through
1120                    to code handling that situation below... */
1121                 alb_set_slave_mac_addr(slave, bond->dev->dev_addr,
1122                                        bond->alb_info.rlb_enabled);
1123         }
1124
1125         /* The slave's address is equal to the address of the bond.
1126          * Search for a spare address in the bond for this slave.
1127          */
1128         free_mac_slave = NULL;
1129
1130         bond_for_each_slave(bond, tmp_slave1, i) {
1131                 found = 0;
1132                 bond_for_each_slave(bond, tmp_slave2, j) {
1133                         if (!compare_ether_addr_64bits(tmp_slave1->perm_hwaddr,
1134                                                        tmp_slave2->dev->dev_addr)) {
1135                                 found = 1;
1136                                 break;
1137                         }
1138                 }
1139
1140                 if (!found) {
1141                         /* no slave has tmp_slave1's perm addr
1142                          * as its curr addr
1143                          */
1144                         free_mac_slave = tmp_slave1;
1145                         break;
1146                 }
1147
1148                 if (!has_bond_addr) {
1149                         if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1150                                                        bond->dev->dev_addr)) {
1151
1152                                 has_bond_addr = tmp_slave1;
1153                         }
1154                 }
1155         }
1156
1157         if (free_mac_slave) {
1158                 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr,
1159                                        bond->alb_info.rlb_enabled);
1160
1161                 pr_warning("%s: Warning: the hw address of slave %s is in use by the bond; giving it the hw address of %s\n",
1162                            bond->dev->name, slave->dev->name,
1163                            free_mac_slave->dev->name);
1164
1165         } else if (has_bond_addr) {
1166                 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",
1167                        bond->dev->name, slave->dev->name);
1168                 return -EFAULT;
1169         }
1170
1171         return 0;
1172 }
1173
1174 /**
1175  * alb_set_mac_address
1176  * @bond:
1177  * @addr:
1178  *
1179  * In TLB mode all slaves are configured to the bond's hw address, but set
1180  * their dev_addr field to different addresses (based on their permanent hw
1181  * addresses).
1182  *
1183  * For each slave, this function sets the interface to the new address and then
1184  * changes its dev_addr field to its previous value.
1185  *
1186  * Unwinding assumes bond's mac address has not yet changed.
1187  */
1188 static int alb_set_mac_address(struct bonding *bond, void *addr)
1189 {
1190         struct sockaddr sa;
1191         struct slave *slave, *stop_at;
1192         char tmp_addr[ETH_ALEN];
1193         int res;
1194         int i;
1195
1196         if (bond->alb_info.rlb_enabled) {
1197                 return 0;
1198         }
1199
1200         bond_for_each_slave(bond, slave, i) {
1201                 /* save net_device's current hw address */
1202                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1203
1204                 res = dev_set_mac_address(slave->dev, addr);
1205
1206                 /* restore net_device's hw address */
1207                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1208
1209                 if (res)
1210                         goto unwind;
1211         }
1212
1213         return 0;
1214
1215 unwind:
1216         memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1217         sa.sa_family = bond->dev->type;
1218
1219         /* unwind from head to the slave that failed */
1220         stop_at = slave;
1221         bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1222                 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1223                 dev_set_mac_address(slave->dev, &sa);
1224                 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1225         }
1226
1227         return res;
1228 }
1229
1230 /************************ exported alb funcions ************************/
1231
1232 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1233 {
1234         int res;
1235
1236         res = tlb_initialize(bond);
1237         if (res) {
1238                 return res;
1239         }
1240
1241         if (rlb_enabled) {
1242                 bond->alb_info.rlb_enabled = 1;
1243                 /* initialize rlb */
1244                 res = rlb_initialize(bond);
1245                 if (res) {
1246                         tlb_deinitialize(bond);
1247                         return res;
1248                 }
1249         } else {
1250                 bond->alb_info.rlb_enabled = 0;
1251         }
1252
1253         return 0;
1254 }
1255
1256 void bond_alb_deinitialize(struct bonding *bond)
1257 {
1258         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1259
1260         tlb_deinitialize(bond);
1261
1262         if (bond_info->rlb_enabled) {
1263                 rlb_deinitialize(bond);
1264         }
1265 }
1266
1267 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1268 {
1269         struct bonding *bond = netdev_priv(bond_dev);
1270         struct ethhdr *eth_data;
1271         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1272         struct slave *tx_slave = NULL;
1273         static const __be32 ip_bcast = htonl(0xffffffff);
1274         int hash_size = 0;
1275         int do_tx_balance = 1;
1276         u32 hash_index = 0;
1277         const u8 *hash_start = NULL;
1278         int res = 1;
1279         struct ipv6hdr *ip6hdr;
1280
1281         skb_reset_mac_header(skb);
1282         eth_data = eth_hdr(skb);
1283
1284         /* make sure that the curr_active_slave and the slaves list do
1285          * not change during tx
1286          */
1287         read_lock(&bond->lock);
1288         read_lock(&bond->curr_slave_lock);
1289
1290         if (!BOND_IS_OK(bond)) {
1291                 goto out;
1292         }
1293
1294         switch (ntohs(skb->protocol)) {
1295         case ETH_P_IP: {
1296                 const struct iphdr *iph = ip_hdr(skb);
1297
1298                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast) ||
1299                     (iph->daddr == ip_bcast) ||
1300                     (iph->protocol == IPPROTO_IGMP)) {
1301                         do_tx_balance = 0;
1302                         break;
1303                 }
1304                 hash_start = (char *)&(iph->daddr);
1305                 hash_size = sizeof(iph->daddr);
1306         }
1307                 break;
1308         case ETH_P_IPV6:
1309                 /* IPv6 doesn't really use broadcast mac address, but leave
1310                  * that here just in case.
1311                  */
1312                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast)) {
1313                         do_tx_balance = 0;
1314                         break;
1315                 }
1316
1317                 /* IPv6 uses all-nodes multicast as an equivalent to
1318                  * broadcasts in IPv4.
1319                  */
1320                 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_v6_allmcast)) {
1321                         do_tx_balance = 0;
1322                         break;
1323                 }
1324
1325                 /* Additianally, DAD probes should not be tx-balanced as that
1326                  * will lead to false positives for duplicate addresses and
1327                  * prevent address configuration from working.
1328                  */
1329                 ip6hdr = ipv6_hdr(skb);
1330                 if (ipv6_addr_any(&ip6hdr->saddr)) {
1331                         do_tx_balance = 0;
1332                         break;
1333                 }
1334
1335                 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1336                 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1337                 break;
1338         case ETH_P_IPX:
1339                 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1340                         /* something is wrong with this packet */
1341                         do_tx_balance = 0;
1342                         break;
1343                 }
1344
1345                 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1346                         /* The only protocol worth balancing in
1347                          * this family since it has an "ARP" like
1348                          * mechanism
1349                          */
1350                         do_tx_balance = 0;
1351                         break;
1352                 }
1353
1354                 hash_start = (char*)eth_data->h_dest;
1355                 hash_size = ETH_ALEN;
1356                 break;
1357         case ETH_P_ARP:
1358                 do_tx_balance = 0;
1359                 if (bond_info->rlb_enabled) {
1360                         tx_slave = rlb_arp_xmit(skb, bond);
1361                 }
1362                 break;
1363         default:
1364                 do_tx_balance = 0;
1365                 break;
1366         }
1367
1368         if (do_tx_balance) {
1369                 hash_index = _simple_hash(hash_start, hash_size);
1370                 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1371         }
1372
1373         if (!tx_slave) {
1374                 /* unbalanced or unassigned, send through primary */
1375                 tx_slave = bond->curr_active_slave;
1376                 bond_info->unbalanced_load += skb->len;
1377         }
1378
1379         if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1380                 if (tx_slave != bond->curr_active_slave) {
1381                         memcpy(eth_data->h_source,
1382                                tx_slave->dev->dev_addr,
1383                                ETH_ALEN);
1384                 }
1385
1386                 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1387         } else {
1388                 if (tx_slave) {
1389                         tlb_clear_slave(bond, tx_slave, 0);
1390                 }
1391         }
1392
1393 out:
1394         if (res) {
1395                 /* no suitable interface, frame not sent */
1396                 dev_kfree_skb(skb);
1397         }
1398         read_unlock(&bond->curr_slave_lock);
1399         read_unlock(&bond->lock);
1400         return NETDEV_TX_OK;
1401 }
1402
1403 void bond_alb_monitor(struct work_struct *work)
1404 {
1405         struct bonding *bond = container_of(work, struct bonding,
1406                                             alb_work.work);
1407         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1408         struct slave *slave;
1409         int i;
1410
1411         read_lock(&bond->lock);
1412
1413         if (bond->kill_timers) {
1414                 goto out;
1415         }
1416
1417         if (bond->slave_cnt == 0) {
1418                 bond_info->tx_rebalance_counter = 0;
1419                 bond_info->lp_counter = 0;
1420                 goto re_arm;
1421         }
1422
1423         bond_info->tx_rebalance_counter++;
1424         bond_info->lp_counter++;
1425
1426         /* send learning packets */
1427         if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1428                 /* change of curr_active_slave involves swapping of mac addresses.
1429                  * in order to avoid this swapping from happening while
1430                  * sending the learning packets, the curr_slave_lock must be held for
1431                  * read.
1432                  */
1433                 read_lock(&bond->curr_slave_lock);
1434
1435                 bond_for_each_slave(bond, slave, i) {
1436                         alb_send_learning_packets(slave, slave->dev->dev_addr);
1437                 }
1438
1439                 read_unlock(&bond->curr_slave_lock);
1440
1441                 bond_info->lp_counter = 0;
1442         }
1443
1444         /* rebalance tx traffic */
1445         if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1446
1447                 read_lock(&bond->curr_slave_lock);
1448
1449                 bond_for_each_slave(bond, slave, i) {
1450                         tlb_clear_slave(bond, slave, 1);
1451                         if (slave == bond->curr_active_slave) {
1452                                 SLAVE_TLB_INFO(slave).load =
1453                                         bond_info->unbalanced_load /
1454                                                 BOND_TLB_REBALANCE_INTERVAL;
1455                                 bond_info->unbalanced_load = 0;
1456                         }
1457                 }
1458
1459                 read_unlock(&bond->curr_slave_lock);
1460
1461                 bond_info->tx_rebalance_counter = 0;
1462         }
1463
1464         /* handle rlb stuff */
1465         if (bond_info->rlb_enabled) {
1466                 if (bond_info->primary_is_promisc &&
1467                     (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1468
1469                         /*
1470                          * dev_set_promiscuity requires rtnl and
1471                          * nothing else.
1472                          */
1473                         read_unlock(&bond->lock);
1474                         rtnl_lock();
1475
1476                         bond_info->rlb_promisc_timeout_counter = 0;
1477
1478                         /* If the primary was set to promiscuous mode
1479                          * because a slave was disabled then
1480                          * it can now leave promiscuous mode.
1481                          */
1482                         dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1483                         bond_info->primary_is_promisc = 0;
1484
1485                         rtnl_unlock();
1486                         read_lock(&bond->lock);
1487                 }
1488
1489                 if (bond_info->rlb_rebalance) {
1490                         bond_info->rlb_rebalance = 0;
1491                         rlb_rebalance(bond);
1492                 }
1493
1494                 /* check if clients need updating */
1495                 if (bond_info->rx_ntt) {
1496                         if (bond_info->rlb_update_delay_counter) {
1497                                 --bond_info->rlb_update_delay_counter;
1498                         } else {
1499                                 rlb_update_rx_clients(bond);
1500                                 if (bond_info->rlb_update_retry_counter) {
1501                                         --bond_info->rlb_update_retry_counter;
1502                                 } else {
1503                                         bond_info->rx_ntt = 0;
1504                                 }
1505                         }
1506                 }
1507         }
1508
1509 re_arm:
1510         queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1511 out:
1512         read_unlock(&bond->lock);
1513 }
1514
1515 /* assumption: called before the slave is attached to the bond
1516  * and not locked by the bond lock
1517  */
1518 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1519 {
1520         int res;
1521
1522         res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr,
1523                                      bond->alb_info.rlb_enabled);
1524         if (res) {
1525                 return res;
1526         }
1527
1528         /* caller must hold the bond lock for write since the mac addresses
1529          * are compared and may be swapped.
1530          */
1531         read_lock(&bond->lock);
1532
1533         res = alb_handle_addr_collision_on_attach(bond, slave);
1534
1535         read_unlock(&bond->lock);
1536
1537         if (res) {
1538                 return res;
1539         }
1540
1541         tlb_init_slave(slave);
1542
1543         /* order a rebalance ASAP */
1544         bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1545
1546         if (bond->alb_info.rlb_enabled) {
1547                 bond->alb_info.rlb_rebalance = 1;
1548         }
1549
1550         return 0;
1551 }
1552
1553 /*
1554  * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1555  * if necessary.
1556  *
1557  * Caller must hold RTNL and no other locks
1558  */
1559 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1560 {
1561         if (bond->slave_cnt > 1) {
1562                 alb_change_hw_addr_on_detach(bond, slave);
1563         }
1564
1565         tlb_clear_slave(bond, slave, 0);
1566
1567         if (bond->alb_info.rlb_enabled) {
1568                 bond->alb_info.next_rx_slave = NULL;
1569                 rlb_clear_slave(bond, slave);
1570         }
1571 }
1572
1573 /* Caller must hold bond lock for read */
1574 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1575 {
1576         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1577
1578         if (link == BOND_LINK_DOWN) {
1579                 tlb_clear_slave(bond, slave, 0);
1580                 if (bond->alb_info.rlb_enabled) {
1581                         rlb_clear_slave(bond, slave);
1582                 }
1583         } else if (link == BOND_LINK_UP) {
1584                 /* order a rebalance ASAP */
1585                 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1586                 if (bond->alb_info.rlb_enabled) {
1587                         bond->alb_info.rlb_rebalance = 1;
1588                         /* If the updelay module parameter is smaller than the
1589                          * forwarding delay of the switch the rebalance will
1590                          * not work because the rebalance arp replies will
1591                          * not be forwarded to the clients..
1592                          */
1593                 }
1594         }
1595 }
1596
1597 /**
1598  * bond_alb_handle_active_change - assign new curr_active_slave
1599  * @bond: our bonding struct
1600  * @new_slave: new slave to assign
1601  *
1602  * Set the bond->curr_active_slave to @new_slave and handle
1603  * mac address swapping and promiscuity changes as needed.
1604  *
1605  * If new_slave is NULL, caller must hold curr_slave_lock or
1606  * bond->lock for write.
1607  *
1608  * If new_slave is not NULL, caller must hold RTNL, bond->lock for
1609  * read and curr_slave_lock for write.  Processing here may sleep, so
1610  * no other locks may be held.
1611  */
1612 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1613         __releases(&bond->curr_slave_lock)
1614         __releases(&bond->lock)
1615         __acquires(&bond->lock)
1616         __acquires(&bond->curr_slave_lock)
1617 {
1618         struct slave *swap_slave;
1619         int i;
1620
1621         if (bond->curr_active_slave == new_slave) {
1622                 return;
1623         }
1624
1625         if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1626                 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1627                 bond->alb_info.primary_is_promisc = 0;
1628                 bond->alb_info.rlb_promisc_timeout_counter = 0;
1629         }
1630
1631         swap_slave = bond->curr_active_slave;
1632         bond->curr_active_slave = new_slave;
1633
1634         if (!new_slave || (bond->slave_cnt == 0)) {
1635                 return;
1636         }
1637
1638         /* set the new curr_active_slave to the bonds mac address
1639          * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1640          */
1641         if (!swap_slave) {
1642                 struct slave *tmp_slave;
1643                 /* find slave that is holding the bond's mac address */
1644                 bond_for_each_slave(bond, tmp_slave, i) {
1645                         if (!compare_ether_addr_64bits(tmp_slave->dev->dev_addr,
1646                                                        bond->dev->dev_addr)) {
1647                                 swap_slave = tmp_slave;
1648                                 break;
1649                         }
1650                 }
1651         }
1652
1653         /*
1654          * Arrange for swap_slave and new_slave to temporarily be
1655          * ignored so we can mess with their MAC addresses without
1656          * fear of interference from transmit activity.
1657          */
1658         if (swap_slave) {
1659                 tlb_clear_slave(bond, swap_slave, 1);
1660         }
1661         tlb_clear_slave(bond, new_slave, 1);
1662
1663         write_unlock_bh(&bond->curr_slave_lock);
1664         read_unlock(&bond->lock);
1665
1666         ASSERT_RTNL();
1667
1668         /* curr_active_slave must be set before calling alb_swap_mac_addr */
1669         if (swap_slave) {
1670                 /* swap mac address */
1671                 alb_swap_mac_addr(bond, swap_slave, new_slave);
1672         } else {
1673                 /* set the new_slave to the bond mac address */
1674                 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr,
1675                                        bond->alb_info.rlb_enabled);
1676         }
1677
1678         if (swap_slave) {
1679                 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1680                 read_lock(&bond->lock);
1681         } else {
1682                 read_lock(&bond->lock);
1683                 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1684         }
1685
1686         write_lock_bh(&bond->curr_slave_lock);
1687 }
1688
1689 /*
1690  * Called with RTNL
1691  */
1692 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1693         __acquires(&bond->lock)
1694         __releases(&bond->lock)
1695 {
1696         struct bonding *bond = netdev_priv(bond_dev);
1697         struct sockaddr *sa = addr;
1698         struct slave *slave, *swap_slave;
1699         int res;
1700         int i;
1701
1702         if (!is_valid_ether_addr(sa->sa_data)) {
1703                 return -EADDRNOTAVAIL;
1704         }
1705
1706         res = alb_set_mac_address(bond, addr);
1707         if (res) {
1708                 return res;
1709         }
1710
1711         memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1712
1713         /* If there is no curr_active_slave there is nothing else to do.
1714          * Otherwise we'll need to pass the new address to it and handle
1715          * duplications.
1716          */
1717         if (!bond->curr_active_slave) {
1718                 return 0;
1719         }
1720
1721         swap_slave = NULL;
1722
1723         bond_for_each_slave(bond, slave, i) {
1724                 if (!compare_ether_addr_64bits(slave->dev->dev_addr,
1725                                                bond_dev->dev_addr)) {
1726                         swap_slave = slave;
1727                         break;
1728                 }
1729         }
1730
1731         if (swap_slave) {
1732                 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1733                 alb_fasten_mac_swap(bond, swap_slave, bond->curr_active_slave);
1734         } else {
1735                 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr,
1736                                        bond->alb_info.rlb_enabled);
1737
1738                 read_lock(&bond->lock);
1739                 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1740                 if (bond->alb_info.rlb_enabled) {
1741                         /* inform clients mac address has changed */
1742                         rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1743                 }
1744                 read_unlock(&bond->lock);
1745         }
1746
1747         return 0;
1748 }
1749
1750 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1751 {
1752         if (bond->alb_info.current_alb_vlan &&
1753             (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1754                 bond->alb_info.current_alb_vlan = NULL;
1755         }
1756
1757         if (bond->alb_info.rlb_enabled) {
1758                 rlb_clear_vlan(bond, vlan_id);
1759         }
1760 }
1761