Merge git://www.linux-watchdog.org/linux-watchdog
[linux-2.6-block.git] / 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, see <http://www.gnu.org/licenses/>.
16  *
17  * The full GNU General Public License is included in this distribution in the
18  * file called LICENSE.
19  *
20  */
21
22 #include <linux/skbuff.h>
23 #include <linux/netdevice.h>
24 #include <linux/etherdevice.h>
25 #include <linux/pkt_sched.h>
26 #include <linux/spinlock.h>
27 #include <linux/slab.h>
28 #include <linux/timer.h>
29 #include <linux/ip.h>
30 #include <linux/ipv6.h>
31 #include <linux/if_arp.h>
32 #include <linux/if_ether.h>
33 #include <linux/if_bonding.h>
34 #include <linux/if_vlan.h>
35 #include <linux/in.h>
36 #include <net/ipx.h>
37 #include <net/arp.h>
38 #include <net/ipv6.h>
39 #include <asm/byteorder.h>
40 #include "bonding.h"
41 #include "bond_alb.h"
42
43
44
45 #ifndef __long_aligned
46 #define __long_aligned __attribute__((aligned((sizeof(long)))))
47 #endif
48 static const u8 mac_bcast[ETH_ALEN] __long_aligned = {
49         0xff, 0xff, 0xff, 0xff, 0xff, 0xff
50 };
51 static const u8 mac_v6_allmcast[ETH_ALEN] __long_aligned = {
52         0x33, 0x33, 0x00, 0x00, 0x00, 0x01
53 };
54 static const int alb_delta_in_ticks = HZ / ALB_TIMER_TICKS_PER_SEC;
55
56 #pragma pack(1)
57 struct learning_pkt {
58         u8 mac_dst[ETH_ALEN];
59         u8 mac_src[ETH_ALEN];
60         __be16 type;
61         u8 padding[ETH_ZLEN - ETH_HLEN];
62 };
63
64 struct arp_pkt {
65         __be16  hw_addr_space;
66         __be16  prot_addr_space;
67         u8      hw_addr_len;
68         u8      prot_addr_len;
69         __be16  op_code;
70         u8      mac_src[ETH_ALEN];      /* sender hardware address */
71         __be32  ip_src;                 /* sender IP address */
72         u8      mac_dst[ETH_ALEN];      /* target hardware address */
73         __be32  ip_dst;                 /* target IP address */
74 };
75 #pragma pack()
76
77 static inline struct arp_pkt *arp_pkt(const struct sk_buff *skb)
78 {
79         return (struct arp_pkt *)skb_network_header(skb);
80 }
81
82 /* Forward declaration */
83 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[],
84                                       bool strict_match);
85 static void rlb_purge_src_ip(struct bonding *bond, struct arp_pkt *arp);
86 static void rlb_src_unlink(struct bonding *bond, u32 index);
87 static void rlb_src_link(struct bonding *bond, u32 ip_src_hash,
88                          u32 ip_dst_hash);
89
90 static inline u8 _simple_hash(const u8 *hash_start, int hash_size)
91 {
92         int i;
93         u8 hash = 0;
94
95         for (i = 0; i < hash_size; i++)
96                 hash ^= hash_start[i];
97
98         return hash;
99 }
100
101 /*********************** tlb specific functions ***************************/
102
103 static inline void _lock_tx_hashtbl_bh(struct bonding *bond)
104 {
105         spin_lock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
106 }
107
108 static inline void _unlock_tx_hashtbl_bh(struct bonding *bond)
109 {
110         spin_unlock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
111 }
112
113 static inline void _lock_tx_hashtbl(struct bonding *bond)
114 {
115         spin_lock(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
116 }
117
118 static inline void _unlock_tx_hashtbl(struct bonding *bond)
119 {
120         spin_unlock(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
121 }
122
123 /* Caller must hold tx_hashtbl lock */
124 static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
125 {
126         if (save_load) {
127                 entry->load_history = 1 + entry->tx_bytes /
128                                       BOND_TLB_REBALANCE_INTERVAL;
129                 entry->tx_bytes = 0;
130         }
131
132         entry->tx_slave = NULL;
133         entry->next = TLB_NULL_INDEX;
134         entry->prev = TLB_NULL_INDEX;
135 }
136
137 static inline void tlb_init_slave(struct slave *slave)
138 {
139         SLAVE_TLB_INFO(slave).load = 0;
140         SLAVE_TLB_INFO(slave).head = TLB_NULL_INDEX;
141 }
142
143 /* Caller must hold bond lock for read, BH disabled */
144 static void __tlb_clear_slave(struct bonding *bond, struct slave *slave,
145                          int save_load)
146 {
147         struct tlb_client_info *tx_hash_table;
148         u32 index;
149
150         /* clear slave from tx_hashtbl */
151         tx_hash_table = BOND_ALB_INFO(bond).tx_hashtbl;
152
153         /* skip this if we've already freed the tx hash table */
154         if (tx_hash_table) {
155                 index = SLAVE_TLB_INFO(slave).head;
156                 while (index != TLB_NULL_INDEX) {
157                         u32 next_index = tx_hash_table[index].next;
158                         tlb_init_table_entry(&tx_hash_table[index], save_load);
159                         index = next_index;
160                 }
161         }
162
163         tlb_init_slave(slave);
164 }
165
166 /* Caller must hold bond lock for read */
167 static void tlb_clear_slave(struct bonding *bond, struct slave *slave,
168                          int save_load)
169 {
170         _lock_tx_hashtbl_bh(bond);
171         __tlb_clear_slave(bond, slave, save_load);
172         _unlock_tx_hashtbl_bh(bond);
173 }
174
175 /* Must be called before starting the monitor timer */
176 static int tlb_initialize(struct bonding *bond)
177 {
178         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
179         int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
180         struct tlb_client_info *new_hashtbl;
181         int i;
182
183         new_hashtbl = kzalloc(size, GFP_KERNEL);
184         if (!new_hashtbl)
185                 return -1;
186
187         _lock_tx_hashtbl_bh(bond);
188
189         bond_info->tx_hashtbl = new_hashtbl;
190
191         for (i = 0; i < TLB_HASH_TABLE_SIZE; i++)
192                 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 0);
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         struct tlb_up_slave *arr;
204
205         _lock_tx_hashtbl_bh(bond);
206
207         kfree(bond_info->tx_hashtbl);
208         bond_info->tx_hashtbl = NULL;
209
210         _unlock_tx_hashtbl_bh(bond);
211
212         arr = rtnl_dereference(bond_info->slave_arr);
213         if (arr)
214                 kfree_rcu(arr, rcu);
215 }
216
217 static long long compute_gap(struct slave *slave)
218 {
219         return (s64) (slave->speed << 20) - /* Convert to Megabit per sec */
220                (s64) (SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
221 }
222
223 /* Caller must hold bond lock for read */
224 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
225 {
226         struct slave *slave, *least_loaded;
227         struct list_head *iter;
228         long long max_gap;
229
230         least_loaded = NULL;
231         max_gap = LLONG_MIN;
232
233         /* Find the slave with the largest gap */
234         bond_for_each_slave_rcu(bond, slave, iter) {
235                 if (bond_slave_can_tx(slave)) {
236                         long long gap = compute_gap(slave);
237
238                         if (max_gap < gap) {
239                                 least_loaded = slave;
240                                 max_gap = gap;
241                         }
242                 }
243         }
244
245         return least_loaded;
246 }
247
248 static struct slave *__tlb_choose_channel(struct bonding *bond, u32 hash_index,
249                                                 u32 skb_len)
250 {
251         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
252         struct tlb_client_info *hash_table;
253         struct slave *assigned_slave;
254
255         hash_table = bond_info->tx_hashtbl;
256         assigned_slave = hash_table[hash_index].tx_slave;
257         if (!assigned_slave) {
258                 assigned_slave = tlb_get_least_loaded_slave(bond);
259
260                 if (assigned_slave) {
261                         struct tlb_slave_info *slave_info =
262                                 &(SLAVE_TLB_INFO(assigned_slave));
263                         u32 next_index = slave_info->head;
264
265                         hash_table[hash_index].tx_slave = assigned_slave;
266                         hash_table[hash_index].next = next_index;
267                         hash_table[hash_index].prev = TLB_NULL_INDEX;
268
269                         if (next_index != TLB_NULL_INDEX)
270                                 hash_table[next_index].prev = hash_index;
271
272                         slave_info->head = hash_index;
273                         slave_info->load +=
274                                 hash_table[hash_index].load_history;
275                 }
276         }
277
278         if (assigned_slave)
279                 hash_table[hash_index].tx_bytes += skb_len;
280
281         return assigned_slave;
282 }
283
284 /* Caller must hold bond lock for read */
285 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index,
286                                         u32 skb_len)
287 {
288         struct slave *tx_slave;
289         /*
290          * We don't need to disable softirq here, becase
291          * tlb_choose_channel() is only called by bond_alb_xmit()
292          * which already has softirq disabled.
293          */
294         _lock_tx_hashtbl(bond);
295         tx_slave = __tlb_choose_channel(bond, hash_index, skb_len);
296         _unlock_tx_hashtbl(bond);
297         return tx_slave;
298 }
299
300 /*********************** rlb specific functions ***************************/
301 static inline void _lock_rx_hashtbl_bh(struct bonding *bond)
302 {
303         spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
304 }
305
306 static inline void _unlock_rx_hashtbl_bh(struct bonding *bond)
307 {
308         spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
309 }
310
311 static inline void _lock_rx_hashtbl(struct bonding *bond)
312 {
313         spin_lock(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
314 }
315
316 static inline void _unlock_rx_hashtbl(struct bonding *bond)
317 {
318         spin_unlock(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
319 }
320
321 /* when an ARP REPLY is received from a client update its info
322  * in the rx_hashtbl
323  */
324 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
325 {
326         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
327         struct rlb_client_info *client_info;
328         u32 hash_index;
329
330         _lock_rx_hashtbl_bh(bond);
331
332         hash_index = _simple_hash((u8 *)&(arp->ip_src), sizeof(arp->ip_src));
333         client_info = &(bond_info->rx_hashtbl[hash_index]);
334
335         if ((client_info->assigned) &&
336             (client_info->ip_src == arp->ip_dst) &&
337             (client_info->ip_dst == arp->ip_src) &&
338             (!ether_addr_equal_64bits(client_info->mac_dst, arp->mac_src))) {
339                 /* update the clients MAC address */
340                 ether_addr_copy(client_info->mac_dst, arp->mac_src);
341                 client_info->ntt = 1;
342                 bond_info->rx_ntt = 1;
343         }
344
345         _unlock_rx_hashtbl_bh(bond);
346 }
347
348 static int rlb_arp_recv(const struct sk_buff *skb, struct bonding *bond,
349                         struct slave *slave)
350 {
351         struct arp_pkt *arp, _arp;
352
353         if (skb->protocol != cpu_to_be16(ETH_P_ARP))
354                 goto out;
355
356         arp = skb_header_pointer(skb, 0, sizeof(_arp), &_arp);
357         if (!arp)
358                 goto out;
359
360         /* We received an ARP from arp->ip_src.
361          * We might have used this IP address previously (on the bonding host
362          * itself or on a system that is bridged together with the bond).
363          * However, if arp->mac_src is different than what is stored in
364          * rx_hashtbl, some other host is now using the IP and we must prevent
365          * sending out client updates with this IP address and the old MAC
366          * address.
367          * Clean up all hash table entries that have this address as ip_src but
368          * have a different mac_src.
369          */
370         rlb_purge_src_ip(bond, arp);
371
372         if (arp->op_code == htons(ARPOP_REPLY)) {
373                 /* update rx hash table for this ARP */
374                 rlb_update_entry_from_arp(bond, arp);
375                 netdev_dbg(bond->dev, "Server received an ARP Reply from client\n");
376         }
377 out:
378         return RX_HANDLER_ANOTHER;
379 }
380
381 /* Caller must hold bond lock for read */
382 static struct slave *rlb_next_rx_slave(struct bonding *bond)
383 {
384         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
385         struct slave *before = NULL, *rx_slave = NULL, *slave;
386         struct list_head *iter;
387         bool found = false;
388
389         bond_for_each_slave(bond, slave, iter) {
390                 if (!bond_slave_can_tx(slave))
391                         continue;
392                 if (!found) {
393                         if (!before || before->speed < slave->speed)
394                                 before = slave;
395                 } else {
396                         if (!rx_slave || rx_slave->speed < slave->speed)
397                                 rx_slave = slave;
398                 }
399                 if (slave == bond_info->rx_slave)
400                         found = true;
401         }
402         /* we didn't find anything after the current or we have something
403          * better before and up to the current slave
404          */
405         if (!rx_slave || (before && rx_slave->speed < before->speed))
406                 rx_slave = before;
407
408         if (rx_slave)
409                 bond_info->rx_slave = rx_slave;
410
411         return rx_slave;
412 }
413
414 /* Caller must hold rcu_read_lock() for read */
415 static struct slave *__rlb_next_rx_slave(struct bonding *bond)
416 {
417         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
418         struct slave *before = NULL, *rx_slave = NULL, *slave;
419         struct list_head *iter;
420         bool found = false;
421
422         bond_for_each_slave_rcu(bond, slave, iter) {
423                 if (!bond_slave_can_tx(slave))
424                         continue;
425                 if (!found) {
426                         if (!before || before->speed < slave->speed)
427                                 before = slave;
428                 } else {
429                         if (!rx_slave || rx_slave->speed < slave->speed)
430                                 rx_slave = slave;
431                 }
432                 if (slave == bond_info->rx_slave)
433                         found = true;
434         }
435         /* we didn't find anything after the current or we have something
436          * better before and up to the current slave
437          */
438         if (!rx_slave || (before && rx_slave->speed < before->speed))
439                 rx_slave = before;
440
441         if (rx_slave)
442                 bond_info->rx_slave = rx_slave;
443
444         return rx_slave;
445 }
446
447 /* teach the switch the mac of a disabled slave
448  * on the primary for fault tolerance
449  *
450  * Caller must hold bond->curr_slave_lock for write or bond lock for write
451  */
452 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
453 {
454         struct slave *curr_active = bond_deref_active_protected(bond);
455
456         if (!curr_active)
457                 return;
458
459         if (!bond->alb_info.primary_is_promisc) {
460                 if (!dev_set_promiscuity(curr_active->dev, 1))
461                         bond->alb_info.primary_is_promisc = 1;
462                 else
463                         bond->alb_info.primary_is_promisc = 0;
464         }
465
466         bond->alb_info.rlb_promisc_timeout_counter = 0;
467
468         alb_send_learning_packets(curr_active, addr, true);
469 }
470
471 /* slave being removed should not be active at this point
472  *
473  * Caller must hold rtnl.
474  */
475 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
476 {
477         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
478         struct rlb_client_info *rx_hash_table;
479         u32 index, next_index;
480
481         /* clear slave from rx_hashtbl */
482         _lock_rx_hashtbl_bh(bond);
483
484         rx_hash_table = bond_info->rx_hashtbl;
485         index = bond_info->rx_hashtbl_used_head;
486         for (; index != RLB_NULL_INDEX; index = next_index) {
487                 next_index = rx_hash_table[index].used_next;
488                 if (rx_hash_table[index].slave == slave) {
489                         struct slave *assigned_slave = rlb_next_rx_slave(bond);
490
491                         if (assigned_slave) {
492                                 rx_hash_table[index].slave = assigned_slave;
493                                 if (!ether_addr_equal_64bits(rx_hash_table[index].mac_dst,
494                                                              mac_bcast)) {
495                                         bond_info->rx_hashtbl[index].ntt = 1;
496                                         bond_info->rx_ntt = 1;
497                                         /* A slave has been removed from the
498                                          * table because it is either disabled
499                                          * or being released. We must retry the
500                                          * update to avoid clients from not
501                                          * being updated & disconnecting when
502                                          * there is stress
503                                          */
504                                         bond_info->rlb_update_retry_counter =
505                                                 RLB_UPDATE_RETRY;
506                                 }
507                         } else {  /* there is no active slave */
508                                 rx_hash_table[index].slave = NULL;
509                         }
510                 }
511         }
512
513         _unlock_rx_hashtbl_bh(bond);
514
515         write_lock_bh(&bond->curr_slave_lock);
516
517         if (slave != bond_deref_active_protected(bond))
518                 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
519
520         write_unlock_bh(&bond->curr_slave_lock);
521 }
522
523 static void rlb_update_client(struct rlb_client_info *client_info)
524 {
525         int i;
526
527         if (!client_info->slave)
528                 return;
529
530         for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
531                 struct sk_buff *skb;
532
533                 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
534                                  client_info->ip_dst,
535                                  client_info->slave->dev,
536                                  client_info->ip_src,
537                                  client_info->mac_dst,
538                                  client_info->slave->dev->dev_addr,
539                                  client_info->mac_dst);
540                 if (!skb) {
541                         netdev_err(client_info->slave->bond->dev,
542                                    "failed to create an ARP packet\n");
543                         continue;
544                 }
545
546                 skb->dev = client_info->slave->dev;
547
548                 if (client_info->vlan_id) {
549                         skb = vlan_put_tag(skb, htons(ETH_P_8021Q), client_info->vlan_id);
550                         if (!skb) {
551                                 netdev_err(client_info->slave->bond->dev,
552                                            "failed to insert VLAN tag\n");
553                                 continue;
554                         }
555                 }
556
557                 arp_xmit(skb);
558         }
559 }
560
561 /* sends ARP REPLIES that update the clients that need updating */
562 static void rlb_update_rx_clients(struct bonding *bond)
563 {
564         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
565         struct rlb_client_info *client_info;
566         u32 hash_index;
567
568         _lock_rx_hashtbl_bh(bond);
569
570         hash_index = bond_info->rx_hashtbl_used_head;
571         for (; hash_index != RLB_NULL_INDEX;
572              hash_index = client_info->used_next) {
573                 client_info = &(bond_info->rx_hashtbl[hash_index]);
574                 if (client_info->ntt) {
575                         rlb_update_client(client_info);
576                         if (bond_info->rlb_update_retry_counter == 0)
577                                 client_info->ntt = 0;
578                 }
579         }
580
581         /* do not update the entries again until this counter is zero so that
582          * not to confuse the clients.
583          */
584         bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
585
586         _unlock_rx_hashtbl_bh(bond);
587 }
588
589 /* The slave was assigned a new mac address - update the clients */
590 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
591 {
592         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
593         struct rlb_client_info *client_info;
594         int ntt = 0;
595         u32 hash_index;
596
597         _lock_rx_hashtbl_bh(bond);
598
599         hash_index = bond_info->rx_hashtbl_used_head;
600         for (; hash_index != RLB_NULL_INDEX;
601              hash_index = client_info->used_next) {
602                 client_info = &(bond_info->rx_hashtbl[hash_index]);
603
604                 if ((client_info->slave == slave) &&
605                     !ether_addr_equal_64bits(client_info->mac_dst, mac_bcast)) {
606                         client_info->ntt = 1;
607                         ntt = 1;
608                 }
609         }
610
611         /* update the team's flag only after the whole iteration */
612         if (ntt) {
613                 bond_info->rx_ntt = 1;
614                 /* fasten the change */
615                 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
616         }
617
618         _unlock_rx_hashtbl_bh(bond);
619 }
620
621 /* mark all clients using src_ip to be updated */
622 static void rlb_req_update_subnet_clients(struct bonding *bond, __be32 src_ip)
623 {
624         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
625         struct rlb_client_info *client_info;
626         u32 hash_index;
627
628         _lock_rx_hashtbl(bond);
629
630         hash_index = bond_info->rx_hashtbl_used_head;
631         for (; hash_index != RLB_NULL_INDEX;
632              hash_index = client_info->used_next) {
633                 client_info = &(bond_info->rx_hashtbl[hash_index]);
634
635                 if (!client_info->slave) {
636                         netdev_err(bond->dev, "found a client with no channel in the client's hash table\n");
637                         continue;
638                 }
639                 /*update all clients using this src_ip, that are not assigned
640                  * to the team's address (curr_active_slave) and have a known
641                  * unicast mac address.
642                  */
643                 if ((client_info->ip_src == src_ip) &&
644                     !ether_addr_equal_64bits(client_info->slave->dev->dev_addr,
645                                              bond->dev->dev_addr) &&
646                     !ether_addr_equal_64bits(client_info->mac_dst, mac_bcast)) {
647                         client_info->ntt = 1;
648                         bond_info->rx_ntt = 1;
649                 }
650         }
651
652         _unlock_rx_hashtbl(bond);
653 }
654
655 /* Caller must hold both bond and ptr locks for read */
656 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
657 {
658         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
659         struct arp_pkt *arp = arp_pkt(skb);
660         struct slave *assigned_slave, *curr_active_slave;
661         struct rlb_client_info *client_info;
662         u32 hash_index = 0;
663
664         _lock_rx_hashtbl(bond);
665
666         curr_active_slave = rcu_dereference(bond->curr_active_slave);
667
668         hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_dst));
669         client_info = &(bond_info->rx_hashtbl[hash_index]);
670
671         if (client_info->assigned) {
672                 if ((client_info->ip_src == arp->ip_src) &&
673                     (client_info->ip_dst == arp->ip_dst)) {
674                         /* the entry is already assigned to this client */
675                         if (!ether_addr_equal_64bits(arp->mac_dst, mac_bcast)) {
676                                 /* update mac address from arp */
677                                 ether_addr_copy(client_info->mac_dst, arp->mac_dst);
678                         }
679                         ether_addr_copy(client_info->mac_src, arp->mac_src);
680
681                         assigned_slave = client_info->slave;
682                         if (assigned_slave) {
683                                 _unlock_rx_hashtbl(bond);
684                                 return assigned_slave;
685                         }
686                 } else {
687                         /* the entry is already assigned to some other client,
688                          * move the old client to primary (curr_active_slave) so
689                          * that the new client can be assigned to this entry.
690                          */
691                         if (curr_active_slave &&
692                             client_info->slave != curr_active_slave) {
693                                 client_info->slave = curr_active_slave;
694                                 rlb_update_client(client_info);
695                         }
696                 }
697         }
698         /* assign a new slave */
699         assigned_slave = __rlb_next_rx_slave(bond);
700
701         if (assigned_slave) {
702                 if (!(client_info->assigned &&
703                       client_info->ip_src == arp->ip_src)) {
704                         /* ip_src is going to be updated,
705                          * fix the src hash list
706                          */
707                         u32 hash_src = _simple_hash((u8 *)&arp->ip_src,
708                                                     sizeof(arp->ip_src));
709                         rlb_src_unlink(bond, hash_index);
710                         rlb_src_link(bond, hash_src, hash_index);
711                 }
712
713                 client_info->ip_src = arp->ip_src;
714                 client_info->ip_dst = arp->ip_dst;
715                 /* arp->mac_dst is broadcast for arp reqeusts.
716                  * will be updated with clients actual unicast mac address
717                  * upon receiving an arp reply.
718                  */
719                 ether_addr_copy(client_info->mac_dst, arp->mac_dst);
720                 ether_addr_copy(client_info->mac_src, arp->mac_src);
721                 client_info->slave = assigned_slave;
722
723                 if (!ether_addr_equal_64bits(client_info->mac_dst, mac_bcast)) {
724                         client_info->ntt = 1;
725                         bond->alb_info.rx_ntt = 1;
726                 } else {
727                         client_info->ntt = 0;
728                 }
729
730                 if (vlan_get_tag(skb, &client_info->vlan_id))
731                         client_info->vlan_id = 0;
732
733                 if (!client_info->assigned) {
734                         u32 prev_tbl_head = bond_info->rx_hashtbl_used_head;
735                         bond_info->rx_hashtbl_used_head = hash_index;
736                         client_info->used_next = prev_tbl_head;
737                         if (prev_tbl_head != RLB_NULL_INDEX) {
738                                 bond_info->rx_hashtbl[prev_tbl_head].used_prev =
739                                         hash_index;
740                         }
741                         client_info->assigned = 1;
742                 }
743         }
744
745         _unlock_rx_hashtbl(bond);
746
747         return assigned_slave;
748 }
749
750 /* chooses (and returns) transmit channel for arp reply
751  * does not choose channel for other arp types since they are
752  * sent on the curr_active_slave
753  */
754 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
755 {
756         struct arp_pkt *arp = arp_pkt(skb);
757         struct slave *tx_slave = NULL;
758
759         /* Don't modify or load balance ARPs that do not originate locally
760          * (e.g.,arrive via a bridge).
761          */
762         if (!bond_slave_has_mac_rx(bond, arp->mac_src))
763                 return NULL;
764
765         if (arp->op_code == htons(ARPOP_REPLY)) {
766                 /* the arp must be sent on the selected
767                 * rx channel
768                 */
769                 tx_slave = rlb_choose_channel(skb, bond);
770                 if (tx_slave)
771                         ether_addr_copy(arp->mac_src, tx_slave->dev->dev_addr);
772                 netdev_dbg(bond->dev, "Server sent ARP Reply packet\n");
773         } else if (arp->op_code == htons(ARPOP_REQUEST)) {
774                 /* Create an entry in the rx_hashtbl for this client as a
775                  * place holder.
776                  * When the arp reply is received the entry will be updated
777                  * with the correct unicast address of the client.
778                  */
779                 rlb_choose_channel(skb, bond);
780
781                 /* The ARP reply packets must be delayed so that
782                  * they can cancel out the influence of the ARP request.
783                  */
784                 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
785
786                 /* arp requests are broadcast and are sent on the primary
787                  * the arp request will collapse all clients on the subnet to
788                  * the primary slave. We must register these clients to be
789                  * updated with their assigned mac.
790                  */
791                 rlb_req_update_subnet_clients(bond, arp->ip_src);
792                 netdev_dbg(bond->dev, "Server sent ARP Request packet\n");
793         }
794
795         return tx_slave;
796 }
797
798 /* Caller must hold bond lock for read */
799 static void rlb_rebalance(struct bonding *bond)
800 {
801         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
802         struct slave *assigned_slave;
803         struct rlb_client_info *client_info;
804         int ntt;
805         u32 hash_index;
806
807         _lock_rx_hashtbl_bh(bond);
808
809         ntt = 0;
810         hash_index = bond_info->rx_hashtbl_used_head;
811         for (; hash_index != RLB_NULL_INDEX;
812              hash_index = client_info->used_next) {
813                 client_info = &(bond_info->rx_hashtbl[hash_index]);
814                 assigned_slave = __rlb_next_rx_slave(bond);
815                 if (assigned_slave && (client_info->slave != assigned_slave)) {
816                         client_info->slave = assigned_slave;
817                         client_info->ntt = 1;
818                         ntt = 1;
819                 }
820         }
821
822         /* update the team's flag only after the whole iteration */
823         if (ntt)
824                 bond_info->rx_ntt = 1;
825         _unlock_rx_hashtbl_bh(bond);
826 }
827
828 /* Caller must hold rx_hashtbl lock */
829 static void rlb_init_table_entry_dst(struct rlb_client_info *entry)
830 {
831         entry->used_next = RLB_NULL_INDEX;
832         entry->used_prev = RLB_NULL_INDEX;
833         entry->assigned = 0;
834         entry->slave = NULL;
835         entry->vlan_id = 0;
836 }
837 static void rlb_init_table_entry_src(struct rlb_client_info *entry)
838 {
839         entry->src_first = RLB_NULL_INDEX;
840         entry->src_prev = RLB_NULL_INDEX;
841         entry->src_next = RLB_NULL_INDEX;
842 }
843
844 static void rlb_init_table_entry(struct rlb_client_info *entry)
845 {
846         memset(entry, 0, sizeof(struct rlb_client_info));
847         rlb_init_table_entry_dst(entry);
848         rlb_init_table_entry_src(entry);
849 }
850
851 static void rlb_delete_table_entry_dst(struct bonding *bond, u32 index)
852 {
853         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
854         u32 next_index = bond_info->rx_hashtbl[index].used_next;
855         u32 prev_index = bond_info->rx_hashtbl[index].used_prev;
856
857         if (index == bond_info->rx_hashtbl_used_head)
858                 bond_info->rx_hashtbl_used_head = next_index;
859         if (prev_index != RLB_NULL_INDEX)
860                 bond_info->rx_hashtbl[prev_index].used_next = next_index;
861         if (next_index != RLB_NULL_INDEX)
862                 bond_info->rx_hashtbl[next_index].used_prev = prev_index;
863 }
864
865 /* unlink a rlb hash table entry from the src list */
866 static void rlb_src_unlink(struct bonding *bond, u32 index)
867 {
868         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
869         u32 next_index = bond_info->rx_hashtbl[index].src_next;
870         u32 prev_index = bond_info->rx_hashtbl[index].src_prev;
871
872         bond_info->rx_hashtbl[index].src_next = RLB_NULL_INDEX;
873         bond_info->rx_hashtbl[index].src_prev = RLB_NULL_INDEX;
874
875         if (next_index != RLB_NULL_INDEX)
876                 bond_info->rx_hashtbl[next_index].src_prev = prev_index;
877
878         if (prev_index == RLB_NULL_INDEX)
879                 return;
880
881         /* is prev_index pointing to the head of this list? */
882         if (bond_info->rx_hashtbl[prev_index].src_first == index)
883                 bond_info->rx_hashtbl[prev_index].src_first = next_index;
884         else
885                 bond_info->rx_hashtbl[prev_index].src_next = next_index;
886
887 }
888
889 static void rlb_delete_table_entry(struct bonding *bond, u32 index)
890 {
891         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
892         struct rlb_client_info *entry = &(bond_info->rx_hashtbl[index]);
893
894         rlb_delete_table_entry_dst(bond, index);
895         rlb_init_table_entry_dst(entry);
896
897         rlb_src_unlink(bond, index);
898 }
899
900 /* add the rx_hashtbl[ip_dst_hash] entry to the list
901  * of entries with identical ip_src_hash
902  */
903 static void rlb_src_link(struct bonding *bond, u32 ip_src_hash, u32 ip_dst_hash)
904 {
905         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
906         u32 next;
907
908         bond_info->rx_hashtbl[ip_dst_hash].src_prev = ip_src_hash;
909         next = bond_info->rx_hashtbl[ip_src_hash].src_first;
910         bond_info->rx_hashtbl[ip_dst_hash].src_next = next;
911         if (next != RLB_NULL_INDEX)
912                 bond_info->rx_hashtbl[next].src_prev = ip_dst_hash;
913         bond_info->rx_hashtbl[ip_src_hash].src_first = ip_dst_hash;
914 }
915
916 /* deletes all rx_hashtbl entries with  arp->ip_src if their mac_src does
917  * not match arp->mac_src */
918 static void rlb_purge_src_ip(struct bonding *bond, struct arp_pkt *arp)
919 {
920         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
921         u32 ip_src_hash = _simple_hash((u8 *)&(arp->ip_src), sizeof(arp->ip_src));
922         u32 index;
923
924         _lock_rx_hashtbl_bh(bond);
925
926         index = bond_info->rx_hashtbl[ip_src_hash].src_first;
927         while (index != RLB_NULL_INDEX) {
928                 struct rlb_client_info *entry = &(bond_info->rx_hashtbl[index]);
929                 u32 next_index = entry->src_next;
930                 if (entry->ip_src == arp->ip_src &&
931                     !ether_addr_equal_64bits(arp->mac_src, entry->mac_src))
932                                 rlb_delete_table_entry(bond, index);
933                 index = next_index;
934         }
935         _unlock_rx_hashtbl_bh(bond);
936 }
937
938 static int rlb_initialize(struct bonding *bond)
939 {
940         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
941         struct rlb_client_info  *new_hashtbl;
942         int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
943         int i;
944
945         new_hashtbl = kmalloc(size, GFP_KERNEL);
946         if (!new_hashtbl)
947                 return -1;
948
949         _lock_rx_hashtbl_bh(bond);
950
951         bond_info->rx_hashtbl = new_hashtbl;
952
953         bond_info->rx_hashtbl_used_head = RLB_NULL_INDEX;
954
955         for (i = 0; i < RLB_HASH_TABLE_SIZE; i++)
956                 rlb_init_table_entry(bond_info->rx_hashtbl + i);
957
958         _unlock_rx_hashtbl_bh(bond);
959
960         /* register to receive ARPs */
961         bond->recv_probe = rlb_arp_recv;
962
963         return 0;
964 }
965
966 static void rlb_deinitialize(struct bonding *bond)
967 {
968         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
969
970         _lock_rx_hashtbl_bh(bond);
971
972         kfree(bond_info->rx_hashtbl);
973         bond_info->rx_hashtbl = NULL;
974         bond_info->rx_hashtbl_used_head = RLB_NULL_INDEX;
975
976         _unlock_rx_hashtbl_bh(bond);
977 }
978
979 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
980 {
981         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
982         u32 curr_index;
983
984         _lock_rx_hashtbl_bh(bond);
985
986         curr_index = bond_info->rx_hashtbl_used_head;
987         while (curr_index != RLB_NULL_INDEX) {
988                 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
989                 u32 next_index = bond_info->rx_hashtbl[curr_index].used_next;
990
991                 if (curr->vlan_id == vlan_id)
992                         rlb_delete_table_entry(bond, curr_index);
993
994                 curr_index = next_index;
995         }
996
997         _unlock_rx_hashtbl_bh(bond);
998 }
999
1000 /*********************** tlb/rlb shared functions *********************/
1001
1002 static void alb_send_lp_vid(struct slave *slave, u8 mac_addr[],
1003                             __be16 vlan_proto, u16 vid)
1004 {
1005         struct learning_pkt pkt;
1006         struct sk_buff *skb;
1007         int size = sizeof(struct learning_pkt);
1008         char *data;
1009
1010         memset(&pkt, 0, size);
1011         ether_addr_copy(pkt.mac_dst, mac_addr);
1012         ether_addr_copy(pkt.mac_src, mac_addr);
1013         pkt.type = cpu_to_be16(ETH_P_LOOPBACK);
1014
1015         skb = dev_alloc_skb(size);
1016         if (!skb)
1017                 return;
1018
1019         data = skb_put(skb, size);
1020         memcpy(data, &pkt, size);
1021
1022         skb_reset_mac_header(skb);
1023         skb->network_header = skb->mac_header + ETH_HLEN;
1024         skb->protocol = pkt.type;
1025         skb->priority = TC_PRIO_CONTROL;
1026         skb->dev = slave->dev;
1027
1028         if (vid) {
1029                 skb = vlan_put_tag(skb, vlan_proto, vid);
1030                 if (!skb) {
1031                         netdev_err(slave->bond->dev, "failed to insert VLAN tag\n");
1032                         return;
1033                 }
1034         }
1035
1036         dev_queue_xmit(skb);
1037 }
1038
1039 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[],
1040                                       bool strict_match)
1041 {
1042         struct bonding *bond = bond_get_bond_by_slave(slave);
1043         struct net_device *upper;
1044         struct list_head *iter;
1045         struct bond_vlan_tag *tags;
1046
1047         /* send untagged */
1048         alb_send_lp_vid(slave, mac_addr, 0, 0);
1049
1050         /* loop through all devices and see if we need to send a packet
1051          * for that device.
1052          */
1053         rcu_read_lock();
1054         netdev_for_each_all_upper_dev_rcu(bond->dev, upper, iter) {
1055                 if (is_vlan_dev(upper) && vlan_get_encap_level(upper) == 0) {
1056                         if (strict_match &&
1057                             ether_addr_equal_64bits(mac_addr,
1058                                                     upper->dev_addr)) {
1059                                 alb_send_lp_vid(slave, mac_addr,
1060                                                 vlan_dev_vlan_proto(upper),
1061                                                 vlan_dev_vlan_id(upper));
1062                         } else if (!strict_match) {
1063                                 alb_send_lp_vid(slave, upper->dev_addr,
1064                                                 vlan_dev_vlan_proto(upper),
1065                                                 vlan_dev_vlan_id(upper));
1066                         }
1067                 }
1068
1069                 /* If this is a macvlan device, then only send updates
1070                  * when strict_match is turned off.
1071                  */
1072                 if (netif_is_macvlan(upper) && !strict_match) {
1073                         tags = bond_verify_device_path(bond->dev, upper, 0);
1074                         if (IS_ERR_OR_NULL(tags))
1075                                 BUG();
1076                         alb_send_lp_vid(slave, upper->dev_addr,
1077                                         tags[0].vlan_proto, tags[0].vlan_id);
1078                         kfree(tags);
1079                 }
1080         }
1081         rcu_read_unlock();
1082 }
1083
1084 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[])
1085 {
1086         struct net_device *dev = slave->dev;
1087         struct sockaddr s_addr;
1088
1089         if (BOND_MODE(slave->bond) == BOND_MODE_TLB) {
1090                 memcpy(dev->dev_addr, addr, dev->addr_len);
1091                 return 0;
1092         }
1093
1094         /* for rlb each slave must have a unique hw mac addresses so that */
1095         /* each slave will receive packets destined to a different mac */
1096         memcpy(s_addr.sa_data, addr, dev->addr_len);
1097         s_addr.sa_family = dev->type;
1098         if (dev_set_mac_address(dev, &s_addr)) {
1099                 netdev_err(slave->bond->dev, "dev_set_mac_address of dev %s failed! ALB mode requires that the base driver support setting the hw address also when the network device's interface is open\n",
1100                            dev->name);
1101                 return -EOPNOTSUPP;
1102         }
1103         return 0;
1104 }
1105
1106 /*
1107  * Swap MAC addresses between two slaves.
1108  *
1109  * Called with RTNL held, and no other locks.
1110  *
1111  */
1112
1113 static void alb_swap_mac_addr(struct slave *slave1, struct slave *slave2)
1114 {
1115         u8 tmp_mac_addr[ETH_ALEN];
1116
1117         ether_addr_copy(tmp_mac_addr, slave1->dev->dev_addr);
1118         alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr);
1119         alb_set_slave_mac_addr(slave2, tmp_mac_addr);
1120
1121 }
1122
1123 /*
1124  * Send learning packets after MAC address swap.
1125  *
1126  * Called with RTNL and no other locks
1127  */
1128 static void alb_fasten_mac_swap(struct bonding *bond, struct slave *slave1,
1129                                 struct slave *slave2)
1130 {
1131         int slaves_state_differ = (bond_slave_can_tx(slave1) != bond_slave_can_tx(slave2));
1132         struct slave *disabled_slave = NULL;
1133
1134         ASSERT_RTNL();
1135
1136         /* fasten the change in the switch */
1137         if (bond_slave_can_tx(slave1)) {
1138                 alb_send_learning_packets(slave1, slave1->dev->dev_addr, false);
1139                 if (bond->alb_info.rlb_enabled) {
1140                         /* inform the clients that the mac address
1141                          * has changed
1142                          */
1143                         rlb_req_update_slave_clients(bond, slave1);
1144                 }
1145         } else {
1146                 disabled_slave = slave1;
1147         }
1148
1149         if (bond_slave_can_tx(slave2)) {
1150                 alb_send_learning_packets(slave2, slave2->dev->dev_addr, false);
1151                 if (bond->alb_info.rlb_enabled) {
1152                         /* inform the clients that the mac address
1153                          * has changed
1154                          */
1155                         rlb_req_update_slave_clients(bond, slave2);
1156                 }
1157         } else {
1158                 disabled_slave = slave2;
1159         }
1160
1161         if (bond->alb_info.rlb_enabled && slaves_state_differ) {
1162                 /* A disabled slave was assigned an active mac addr */
1163                 rlb_teach_disabled_mac_on_primary(bond,
1164                                                   disabled_slave->dev->dev_addr);
1165         }
1166 }
1167
1168 /**
1169  * alb_change_hw_addr_on_detach
1170  * @bond: bonding we're working on
1171  * @slave: the slave that was just detached
1172  *
1173  * We assume that @slave was already detached from the slave list.
1174  *
1175  * If @slave's permanent hw address is different both from its current
1176  * address and from @bond's address, then somewhere in the bond there's
1177  * a slave that has @slave's permanet address as its current address.
1178  * We'll make sure that that slave no longer uses @slave's permanent address.
1179  *
1180  * Caller must hold RTNL and no other locks
1181  */
1182 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
1183 {
1184         int perm_curr_diff;
1185         int perm_bond_diff;
1186         struct slave *found_slave;
1187
1188         perm_curr_diff = !ether_addr_equal_64bits(slave->perm_hwaddr,
1189                                                   slave->dev->dev_addr);
1190         perm_bond_diff = !ether_addr_equal_64bits(slave->perm_hwaddr,
1191                                                   bond->dev->dev_addr);
1192
1193         if (perm_curr_diff && perm_bond_diff) {
1194                 found_slave = bond_slave_has_mac(bond, slave->perm_hwaddr);
1195
1196                 if (found_slave) {
1197                         /* locking: needs RTNL and nothing else */
1198                         alb_swap_mac_addr(slave, found_slave);
1199                         alb_fasten_mac_swap(bond, slave, found_slave);
1200                 }
1201         }
1202 }
1203
1204 /**
1205  * alb_handle_addr_collision_on_attach
1206  * @bond: bonding we're working on
1207  * @slave: the slave that was just attached
1208  *
1209  * checks uniqueness of slave's mac address and handles the case the
1210  * new slave uses the bonds mac address.
1211  *
1212  * If the permanent hw address of @slave is @bond's hw address, we need to
1213  * find a different hw address to give @slave, that isn't in use by any other
1214  * slave in the bond. This address must be, of course, one of the permanent
1215  * addresses of the other slaves.
1216  *
1217  * We go over the slave list, and for each slave there we compare its
1218  * permanent hw address with the current address of all the other slaves.
1219  * If no match was found, then we've found a slave with a permanent address
1220  * that isn't used by any other slave in the bond, so we can assign it to
1221  * @slave.
1222  *
1223  * assumption: this function is called before @slave is attached to the
1224  *             bond slave list.
1225  */
1226 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1227 {
1228         struct slave *has_bond_addr = rcu_access_pointer(bond->curr_active_slave);
1229         struct slave *tmp_slave1, *free_mac_slave = NULL;
1230         struct list_head *iter;
1231
1232         if (!bond_has_slaves(bond)) {
1233                 /* this is the first slave */
1234                 return 0;
1235         }
1236
1237         /* if slave's mac address differs from bond's mac address
1238          * check uniqueness of slave's mac address against the other
1239          * slaves in the bond.
1240          */
1241         if (!ether_addr_equal_64bits(slave->perm_hwaddr, bond->dev->dev_addr)) {
1242                 if (!bond_slave_has_mac(bond, slave->dev->dev_addr))
1243                         return 0;
1244
1245                 /* Try setting slave mac to bond address and fall-through
1246                    to code handling that situation below... */
1247                 alb_set_slave_mac_addr(slave, bond->dev->dev_addr);
1248         }
1249
1250         /* The slave's address is equal to the address of the bond.
1251          * Search for a spare address in the bond for this slave.
1252          */
1253         bond_for_each_slave(bond, tmp_slave1, iter) {
1254                 if (!bond_slave_has_mac(bond, tmp_slave1->perm_hwaddr)) {
1255                         /* no slave has tmp_slave1's perm addr
1256                          * as its curr addr
1257                          */
1258                         free_mac_slave = tmp_slave1;
1259                         break;
1260                 }
1261
1262                 if (!has_bond_addr) {
1263                         if (ether_addr_equal_64bits(tmp_slave1->dev->dev_addr,
1264                                                     bond->dev->dev_addr)) {
1265
1266                                 has_bond_addr = tmp_slave1;
1267                         }
1268                 }
1269         }
1270
1271         if (free_mac_slave) {
1272                 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr);
1273
1274                 netdev_warn(bond->dev, "the hw address of slave %s is in use by the bond; giving it the hw address of %s\n",
1275                             slave->dev->name, free_mac_slave->dev->name);
1276
1277         } else if (has_bond_addr) {
1278                 netdev_err(bond->dev, "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",
1279                            slave->dev->name);
1280                 return -EFAULT;
1281         }
1282
1283         return 0;
1284 }
1285
1286 /**
1287  * alb_set_mac_address
1288  * @bond:
1289  * @addr:
1290  *
1291  * In TLB mode all slaves are configured to the bond's hw address, but set
1292  * their dev_addr field to different addresses (based on their permanent hw
1293  * addresses).
1294  *
1295  * For each slave, this function sets the interface to the new address and then
1296  * changes its dev_addr field to its previous value.
1297  *
1298  * Unwinding assumes bond's mac address has not yet changed.
1299  */
1300 static int alb_set_mac_address(struct bonding *bond, void *addr)
1301 {
1302         struct slave *slave, *rollback_slave;
1303         struct list_head *iter;
1304         struct sockaddr sa;
1305         char tmp_addr[ETH_ALEN];
1306         int res;
1307
1308         if (bond->alb_info.rlb_enabled)
1309                 return 0;
1310
1311         bond_for_each_slave(bond, slave, iter) {
1312                 /* save net_device's current hw address */
1313                 ether_addr_copy(tmp_addr, slave->dev->dev_addr);
1314
1315                 res = dev_set_mac_address(slave->dev, addr);
1316
1317                 /* restore net_device's hw address */
1318                 ether_addr_copy(slave->dev->dev_addr, tmp_addr);
1319
1320                 if (res)
1321                         goto unwind;
1322         }
1323
1324         return 0;
1325
1326 unwind:
1327         memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1328         sa.sa_family = bond->dev->type;
1329
1330         /* unwind from head to the slave that failed */
1331         bond_for_each_slave(bond, rollback_slave, iter) {
1332                 if (rollback_slave == slave)
1333                         break;
1334                 ether_addr_copy(tmp_addr, rollback_slave->dev->dev_addr);
1335                 dev_set_mac_address(rollback_slave->dev, &sa);
1336                 ether_addr_copy(rollback_slave->dev->dev_addr, tmp_addr);
1337         }
1338
1339         return res;
1340 }
1341
1342 /************************ exported alb funcions ************************/
1343
1344 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1345 {
1346         int res;
1347
1348         res = tlb_initialize(bond);
1349         if (res)
1350                 return res;
1351
1352         if (rlb_enabled) {
1353                 bond->alb_info.rlb_enabled = 1;
1354                 /* initialize rlb */
1355                 res = rlb_initialize(bond);
1356                 if (res) {
1357                         tlb_deinitialize(bond);
1358                         return res;
1359                 }
1360         } else {
1361                 bond->alb_info.rlb_enabled = 0;
1362         }
1363
1364         return 0;
1365 }
1366
1367 void bond_alb_deinitialize(struct bonding *bond)
1368 {
1369         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1370
1371         tlb_deinitialize(bond);
1372
1373         if (bond_info->rlb_enabled)
1374                 rlb_deinitialize(bond);
1375 }
1376
1377 static int bond_do_alb_xmit(struct sk_buff *skb, struct bonding *bond,
1378                 struct slave *tx_slave)
1379 {
1380         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1381         struct ethhdr *eth_data = eth_hdr(skb);
1382
1383         if (!tx_slave) {
1384                 /* unbalanced or unassigned, send through primary */
1385                 tx_slave = rcu_dereference(bond->curr_active_slave);
1386                 if (bond->params.tlb_dynamic_lb)
1387                         bond_info->unbalanced_load += skb->len;
1388         }
1389
1390         if (tx_slave && bond_slave_can_tx(tx_slave)) {
1391                 if (tx_slave != rcu_dereference(bond->curr_active_slave)) {
1392                         ether_addr_copy(eth_data->h_source,
1393                                         tx_slave->dev->dev_addr);
1394                 }
1395
1396                 bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1397                 goto out;
1398         }
1399
1400         if (tx_slave && bond->params.tlb_dynamic_lb) {
1401                 _lock_tx_hashtbl(bond);
1402                 __tlb_clear_slave(bond, tx_slave, 0);
1403                 _unlock_tx_hashtbl(bond);
1404         }
1405
1406         /* no suitable interface, frame not sent */
1407         dev_kfree_skb_any(skb);
1408 out:
1409         return NETDEV_TX_OK;
1410 }
1411
1412 static int bond_tlb_update_slave_arr(struct bonding *bond,
1413                                      struct slave *skipslave)
1414 {
1415         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1416         struct slave *tx_slave;
1417         struct list_head *iter;
1418         struct tlb_up_slave *new_arr, *old_arr;
1419
1420         new_arr = kzalloc(offsetof(struct tlb_up_slave, arr[bond->slave_cnt]),
1421                           GFP_ATOMIC);
1422         if (!new_arr)
1423                 return -ENOMEM;
1424
1425         bond_for_each_slave(bond, tx_slave, iter) {
1426                 if (!bond_slave_can_tx(tx_slave))
1427                         continue;
1428                 if (skipslave == tx_slave)
1429                         continue;
1430                 new_arr->arr[new_arr->count++] = tx_slave;
1431         }
1432
1433         old_arr = rtnl_dereference(bond_info->slave_arr);
1434         rcu_assign_pointer(bond_info->slave_arr, new_arr);
1435         if (old_arr)
1436                 kfree_rcu(old_arr, rcu);
1437
1438         return 0;
1439 }
1440
1441 int bond_tlb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1442 {
1443         struct bonding *bond = netdev_priv(bond_dev);
1444         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1445         struct ethhdr *eth_data;
1446         struct slave *tx_slave = NULL;
1447         u32 hash_index;
1448
1449         skb_reset_mac_header(skb);
1450         eth_data = eth_hdr(skb);
1451
1452         /* Do not TX balance any multicast or broadcast */
1453         if (!is_multicast_ether_addr(eth_data->h_dest)) {
1454                 switch (skb->protocol) {
1455                 case htons(ETH_P_IP):
1456                 case htons(ETH_P_IPX):
1457                     /* In case of IPX, it will falback to L2 hash */
1458                 case htons(ETH_P_IPV6):
1459                         hash_index = bond_xmit_hash(bond, skb);
1460                         if (bond->params.tlb_dynamic_lb) {
1461                                 tx_slave = tlb_choose_channel(bond,
1462                                                               hash_index & 0xFF,
1463                                                               skb->len);
1464                         } else {
1465                                 struct tlb_up_slave *slaves;
1466
1467                                 slaves = rcu_dereference(bond_info->slave_arr);
1468                                 if (slaves && slaves->count)
1469                                         tx_slave = slaves->arr[hash_index %
1470                                                                slaves->count];
1471                         }
1472                         break;
1473                 }
1474         }
1475         return bond_do_alb_xmit(skb, bond, tx_slave);
1476 }
1477
1478 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1479 {
1480         struct bonding *bond = netdev_priv(bond_dev);
1481         struct ethhdr *eth_data;
1482         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1483         struct slave *tx_slave = NULL;
1484         static const __be32 ip_bcast = htonl(0xffffffff);
1485         int hash_size = 0;
1486         bool do_tx_balance = true;
1487         u32 hash_index = 0;
1488         const u8 *hash_start = NULL;
1489         struct ipv6hdr *ip6hdr;
1490
1491         skb_reset_mac_header(skb);
1492         eth_data = eth_hdr(skb);
1493
1494         switch (ntohs(skb->protocol)) {
1495         case ETH_P_IP: {
1496                 const struct iphdr *iph = ip_hdr(skb);
1497
1498                 if (ether_addr_equal_64bits(eth_data->h_dest, mac_bcast) ||
1499                     (iph->daddr == ip_bcast) ||
1500                     (iph->protocol == IPPROTO_IGMP)) {
1501                         do_tx_balance = false;
1502                         break;
1503                 }
1504                 hash_start = (char *)&(iph->daddr);
1505                 hash_size = sizeof(iph->daddr);
1506         }
1507                 break;
1508         case ETH_P_IPV6:
1509                 /* IPv6 doesn't really use broadcast mac address, but leave
1510                  * that here just in case.
1511                  */
1512                 if (ether_addr_equal_64bits(eth_data->h_dest, mac_bcast)) {
1513                         do_tx_balance = false;
1514                         break;
1515                 }
1516
1517                 /* IPv6 uses all-nodes multicast as an equivalent to
1518                  * broadcasts in IPv4.
1519                  */
1520                 if (ether_addr_equal_64bits(eth_data->h_dest, mac_v6_allmcast)) {
1521                         do_tx_balance = false;
1522                         break;
1523                 }
1524
1525                 /* Additianally, DAD probes should not be tx-balanced as that
1526                  * will lead to false positives for duplicate addresses and
1527                  * prevent address configuration from working.
1528                  */
1529                 ip6hdr = ipv6_hdr(skb);
1530                 if (ipv6_addr_any(&ip6hdr->saddr)) {
1531                         do_tx_balance = false;
1532                         break;
1533                 }
1534
1535                 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1536                 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1537                 break;
1538         case ETH_P_IPX:
1539                 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1540                         /* something is wrong with this packet */
1541                         do_tx_balance = false;
1542                         break;
1543                 }
1544
1545                 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1546                         /* The only protocol worth balancing in
1547                          * this family since it has an "ARP" like
1548                          * mechanism
1549                          */
1550                         do_tx_balance = false;
1551                         break;
1552                 }
1553
1554                 hash_start = (char *)eth_data->h_dest;
1555                 hash_size = ETH_ALEN;
1556                 break;
1557         case ETH_P_ARP:
1558                 do_tx_balance = false;
1559                 if (bond_info->rlb_enabled)
1560                         tx_slave = rlb_arp_xmit(skb, bond);
1561                 break;
1562         default:
1563                 do_tx_balance = false;
1564                 break;
1565         }
1566
1567         if (do_tx_balance) {
1568                 hash_index = _simple_hash(hash_start, hash_size);
1569                 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1570         }
1571
1572         return bond_do_alb_xmit(skb, bond, tx_slave);
1573 }
1574
1575 void bond_alb_monitor(struct work_struct *work)
1576 {
1577         struct bonding *bond = container_of(work, struct bonding,
1578                                             alb_work.work);
1579         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1580         struct list_head *iter;
1581         struct slave *slave;
1582
1583         if (!bond_has_slaves(bond)) {
1584                 bond_info->tx_rebalance_counter = 0;
1585                 bond_info->lp_counter = 0;
1586                 goto re_arm;
1587         }
1588
1589         rcu_read_lock();
1590
1591         bond_info->tx_rebalance_counter++;
1592         bond_info->lp_counter++;
1593
1594         /* send learning packets */
1595         if (bond_info->lp_counter >= BOND_ALB_LP_TICKS(bond)) {
1596                 bool strict_match;
1597
1598                 /* change of curr_active_slave involves swapping of mac addresses.
1599                  * in order to avoid this swapping from happening while
1600                  * sending the learning packets, the curr_slave_lock must be held for
1601                  * read.
1602                  */
1603                 read_lock(&bond->curr_slave_lock);
1604
1605                 bond_for_each_slave_rcu(bond, slave, iter) {
1606                         /* If updating current_active, use all currently
1607                          * user mac addreses (!strict_match).  Otherwise, only
1608                          * use mac of the slave device.
1609                          * In RLB mode, we always use strict matches.
1610                          */
1611                         strict_match = (slave != rcu_access_pointer(bond->curr_active_slave) ||
1612                                         bond_info->rlb_enabled);
1613                         alb_send_learning_packets(slave, slave->dev->dev_addr,
1614                                                   strict_match);
1615                 }
1616
1617                 read_unlock(&bond->curr_slave_lock);
1618
1619                 bond_info->lp_counter = 0;
1620         }
1621
1622         /* rebalance tx traffic */
1623         if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1624
1625                 read_lock(&bond->curr_slave_lock);
1626
1627                 bond_for_each_slave_rcu(bond, slave, iter) {
1628                         tlb_clear_slave(bond, slave, 1);
1629                         if (slave == rcu_access_pointer(bond->curr_active_slave)) {
1630                                 SLAVE_TLB_INFO(slave).load =
1631                                         bond_info->unbalanced_load /
1632                                                 BOND_TLB_REBALANCE_INTERVAL;
1633                                 bond_info->unbalanced_load = 0;
1634                         }
1635                 }
1636
1637                 read_unlock(&bond->curr_slave_lock);
1638
1639                 bond_info->tx_rebalance_counter = 0;
1640         }
1641
1642         /* handle rlb stuff */
1643         if (bond_info->rlb_enabled) {
1644                 if (bond_info->primary_is_promisc &&
1645                     (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1646
1647                         /*
1648                          * dev_set_promiscuity requires rtnl and
1649                          * nothing else.  Avoid race with bond_close.
1650                          */
1651                         rcu_read_unlock();
1652                         if (!rtnl_trylock())
1653                                 goto re_arm;
1654
1655                         bond_info->rlb_promisc_timeout_counter = 0;
1656
1657                         /* If the primary was set to promiscuous mode
1658                          * because a slave was disabled then
1659                          * it can now leave promiscuous mode.
1660                          */
1661                         dev_set_promiscuity(rtnl_dereference(bond->curr_active_slave)->dev,
1662                                             -1);
1663                         bond_info->primary_is_promisc = 0;
1664
1665                         rtnl_unlock();
1666                         rcu_read_lock();
1667                 }
1668
1669                 if (bond_info->rlb_rebalance) {
1670                         bond_info->rlb_rebalance = 0;
1671                         rlb_rebalance(bond);
1672                 }
1673
1674                 /* check if clients need updating */
1675                 if (bond_info->rx_ntt) {
1676                         if (bond_info->rlb_update_delay_counter) {
1677                                 --bond_info->rlb_update_delay_counter;
1678                         } else {
1679                                 rlb_update_rx_clients(bond);
1680                                 if (bond_info->rlb_update_retry_counter)
1681                                         --bond_info->rlb_update_retry_counter;
1682                                 else
1683                                         bond_info->rx_ntt = 0;
1684                         }
1685                 }
1686         }
1687         rcu_read_unlock();
1688 re_arm:
1689         queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1690 }
1691
1692 /* assumption: called before the slave is attached to the bond
1693  * and not locked by the bond lock
1694  */
1695 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1696 {
1697         int res;
1698
1699         res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr);
1700         if (res)
1701                 return res;
1702
1703         res = alb_handle_addr_collision_on_attach(bond, slave);
1704         if (res)
1705                 return res;
1706
1707         tlb_init_slave(slave);
1708
1709         /* order a rebalance ASAP */
1710         bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1711
1712         if (bond->alb_info.rlb_enabled)
1713                 bond->alb_info.rlb_rebalance = 1;
1714
1715         return 0;
1716 }
1717
1718 /*
1719  * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1720  * if necessary.
1721  *
1722  * Caller must hold RTNL and no other locks
1723  */
1724 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1725 {
1726         if (bond_has_slaves(bond))
1727                 alb_change_hw_addr_on_detach(bond, slave);
1728
1729         tlb_clear_slave(bond, slave, 0);
1730
1731         if (bond->alb_info.rlb_enabled) {
1732                 bond->alb_info.rx_slave = NULL;
1733                 rlb_clear_slave(bond, slave);
1734         }
1735
1736         if (bond_is_nondyn_tlb(bond))
1737                 if (bond_tlb_update_slave_arr(bond, slave))
1738                         pr_err("Failed to build slave-array for TLB mode.\n");
1739
1740 }
1741
1742 /* Caller must hold bond lock for read */
1743 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1744 {
1745         struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1746
1747         if (link == BOND_LINK_DOWN) {
1748                 tlb_clear_slave(bond, slave, 0);
1749                 if (bond->alb_info.rlb_enabled)
1750                         rlb_clear_slave(bond, slave);
1751         } else if (link == BOND_LINK_UP) {
1752                 /* order a rebalance ASAP */
1753                 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1754                 if (bond->alb_info.rlb_enabled) {
1755                         bond->alb_info.rlb_rebalance = 1;
1756                         /* If the updelay module parameter is smaller than the
1757                          * forwarding delay of the switch the rebalance will
1758                          * not work because the rebalance arp replies will
1759                          * not be forwarded to the clients..
1760                          */
1761                 }
1762         }
1763
1764         if (bond_is_nondyn_tlb(bond)) {
1765                 if (bond_tlb_update_slave_arr(bond, NULL))
1766                         pr_err("Failed to build slave-array for TLB mode.\n");
1767         }
1768 }
1769
1770 /**
1771  * bond_alb_handle_active_change - assign new curr_active_slave
1772  * @bond: our bonding struct
1773  * @new_slave: new slave to assign
1774  *
1775  * Set the bond->curr_active_slave to @new_slave and handle
1776  * mac address swapping and promiscuity changes as needed.
1777  *
1778  * If new_slave is NULL, caller must hold curr_slave_lock or
1779  * bond->lock for write.
1780  *
1781  * If new_slave is not NULL, caller must hold RTNL, curr_slave_lock
1782  * for write.  Processing here may sleep, so no other locks may be held.
1783  */
1784 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1785         __releases(&bond->curr_slave_lock)
1786         __acquires(&bond->curr_slave_lock)
1787 {
1788         struct slave *swap_slave;
1789         struct slave *curr_active;
1790
1791         curr_active = rcu_dereference_protected(bond->curr_active_slave,
1792                                                 !new_slave ||
1793                                                 lockdep_is_held(&bond->curr_slave_lock));
1794         if (curr_active == new_slave)
1795                 return;
1796
1797         if (curr_active && bond->alb_info.primary_is_promisc) {
1798                 dev_set_promiscuity(curr_active->dev, -1);
1799                 bond->alb_info.primary_is_promisc = 0;
1800                 bond->alb_info.rlb_promisc_timeout_counter = 0;
1801         }
1802
1803         swap_slave = curr_active;
1804         rcu_assign_pointer(bond->curr_active_slave, new_slave);
1805
1806         if (!new_slave || !bond_has_slaves(bond))
1807                 return;
1808
1809         /* set the new curr_active_slave to the bonds mac address
1810          * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1811          */
1812         if (!swap_slave)
1813                 swap_slave = bond_slave_has_mac(bond, bond->dev->dev_addr);
1814
1815         /*
1816          * Arrange for swap_slave and new_slave to temporarily be
1817          * ignored so we can mess with their MAC addresses without
1818          * fear of interference from transmit activity.
1819          */
1820         if (swap_slave)
1821                 tlb_clear_slave(bond, swap_slave, 1);
1822         tlb_clear_slave(bond, new_slave, 1);
1823
1824         write_unlock_bh(&bond->curr_slave_lock);
1825
1826         ASSERT_RTNL();
1827
1828         /* in TLB mode, the slave might flip down/up with the old dev_addr,
1829          * and thus filter bond->dev_addr's packets, so force bond's mac
1830          */
1831         if (BOND_MODE(bond) == BOND_MODE_TLB) {
1832                 struct sockaddr sa;
1833                 u8 tmp_addr[ETH_ALEN];
1834
1835                 ether_addr_copy(tmp_addr, new_slave->dev->dev_addr);
1836
1837                 memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1838                 sa.sa_family = bond->dev->type;
1839                 /* we don't care if it can't change its mac, best effort */
1840                 dev_set_mac_address(new_slave->dev, &sa);
1841
1842                 ether_addr_copy(new_slave->dev->dev_addr, tmp_addr);
1843         }
1844
1845         /* curr_active_slave must be set before calling alb_swap_mac_addr */
1846         if (swap_slave) {
1847                 /* swap mac address */
1848                 alb_swap_mac_addr(swap_slave, new_slave);
1849                 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1850         } else {
1851                 /* set the new_slave to the bond mac address */
1852                 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr);
1853                 alb_send_learning_packets(new_slave, bond->dev->dev_addr,
1854                                           false);
1855         }
1856
1857         write_lock_bh(&bond->curr_slave_lock);
1858 }
1859
1860 /*
1861  * Called with RTNL
1862  */
1863 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1864         __acquires(&bond->lock)
1865         __releases(&bond->lock)
1866 {
1867         struct bonding *bond = netdev_priv(bond_dev);
1868         struct sockaddr *sa = addr;
1869         struct slave *curr_active;
1870         struct slave *swap_slave;
1871         int res;
1872
1873         if (!is_valid_ether_addr(sa->sa_data))
1874                 return -EADDRNOTAVAIL;
1875
1876         res = alb_set_mac_address(bond, addr);
1877         if (res)
1878                 return res;
1879
1880         memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1881
1882         /* If there is no curr_active_slave there is nothing else to do.
1883          * Otherwise we'll need to pass the new address to it and handle
1884          * duplications.
1885          */
1886         curr_active = rtnl_dereference(bond->curr_active_slave);
1887         if (!curr_active)
1888                 return 0;
1889
1890         swap_slave = bond_slave_has_mac(bond, bond_dev->dev_addr);
1891
1892         if (swap_slave) {
1893                 alb_swap_mac_addr(swap_slave, curr_active);
1894                 alb_fasten_mac_swap(bond, swap_slave, curr_active);
1895         } else {
1896                 alb_set_slave_mac_addr(curr_active, bond_dev->dev_addr);
1897
1898                 read_lock(&bond->lock);
1899                 alb_send_learning_packets(curr_active,
1900                                           bond_dev->dev_addr, false);
1901                 if (bond->alb_info.rlb_enabled) {
1902                         /* inform clients mac address has changed */
1903                         rlb_req_update_slave_clients(bond, curr_active);
1904                 }
1905                 read_unlock(&bond->lock);
1906         }
1907
1908         return 0;
1909 }
1910
1911 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1912 {
1913         if (bond->alb_info.rlb_enabled)
1914                 rlb_clear_vlan(bond, vlan_id);
1915 }
1916