2 * Copyright(c) 1999 - 2004 Intel Corporation. All rights reserved.
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.
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
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.
18 * The full GNU General Public License is included in this distribution in the
19 * file called LICENSE.
23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
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>
42 #include <asm/byteorder.h>
48 #ifndef __long_aligned
49 #define __long_aligned __attribute__((aligned((sizeof(long)))))
51 static const u8 mac_bcast[ETH_ALEN] __long_aligned = {
52 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
54 static const u8 mac_v6_allmcast[ETH_ALEN] __long_aligned = {
55 0x33, 0x33, 0x00, 0x00, 0x00, 0x01
57 static const int alb_delta_in_ticks = HZ / ALB_TIMER_TICKS_PER_SEC;
64 u8 padding[ETH_ZLEN - ETH_HLEN];
69 __be16 prot_addr_space;
73 u8 mac_src[ETH_ALEN]; /* sender hardware address */
74 __be32 ip_src; /* sender IP address */
75 u8 mac_dst[ETH_ALEN]; /* target hardware address */
76 __be32 ip_dst; /* target IP address */
80 static inline struct arp_pkt *arp_pkt(const struct sk_buff *skb)
82 return (struct arp_pkt *)skb_network_header(skb);
85 /* Forward declaration */
86 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[]);
88 static inline u8 _simple_hash(const u8 *hash_start, int hash_size)
93 for (i = 0; i < hash_size; i++) {
94 hash ^= hash_start[i];
100 /*********************** tlb specific functions ***************************/
102 static inline void _lock_tx_hashtbl(struct bonding *bond)
104 spin_lock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
107 static inline void _unlock_tx_hashtbl(struct bonding *bond)
109 spin_unlock_bh(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
112 /* Caller must hold tx_hashtbl lock */
113 static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
116 entry->load_history = 1 + entry->tx_bytes /
117 BOND_TLB_REBALANCE_INTERVAL;
121 entry->tx_slave = NULL;
122 entry->next = TLB_NULL_INDEX;
123 entry->prev = TLB_NULL_INDEX;
126 static inline void tlb_init_slave(struct slave *slave)
128 SLAVE_TLB_INFO(slave).load = 0;
129 SLAVE_TLB_INFO(slave).head = TLB_NULL_INDEX;
132 /* Caller must hold bond lock for read */
133 static void tlb_clear_slave(struct bonding *bond, struct slave *slave, int save_load)
135 struct tlb_client_info *tx_hash_table;
138 _lock_tx_hashtbl(bond);
140 /* clear slave from tx_hashtbl */
141 tx_hash_table = BOND_ALB_INFO(bond).tx_hashtbl;
143 /* skip this if we've already freed the tx hash table */
145 index = SLAVE_TLB_INFO(slave).head;
146 while (index != TLB_NULL_INDEX) {
147 u32 next_index = tx_hash_table[index].next;
148 tlb_init_table_entry(&tx_hash_table[index], save_load);
153 tlb_init_slave(slave);
155 _unlock_tx_hashtbl(bond);
158 /* Must be called before starting the monitor timer */
159 static int tlb_initialize(struct bonding *bond)
161 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
162 int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
163 struct tlb_client_info *new_hashtbl;
166 new_hashtbl = kzalloc(size, GFP_KERNEL);
168 pr_err("%s: Error: Failed to allocate TLB hash table\n",
172 _lock_tx_hashtbl(bond);
174 bond_info->tx_hashtbl = new_hashtbl;
176 for (i = 0; i < TLB_HASH_TABLE_SIZE; i++) {
177 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 0);
180 _unlock_tx_hashtbl(bond);
185 /* Must be called only after all slaves have been released */
186 static void tlb_deinitialize(struct bonding *bond)
188 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
190 _lock_tx_hashtbl(bond);
192 kfree(bond_info->tx_hashtbl);
193 bond_info->tx_hashtbl = NULL;
195 _unlock_tx_hashtbl(bond);
198 static long long compute_gap(struct slave *slave)
200 return (s64) (slave->speed << 20) - /* Convert to Megabit per sec */
201 (s64) (SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
204 /* Caller must hold bond lock for read */
205 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
207 struct slave *slave, *least_loaded;
214 /* Find the slave with the largest gap */
215 bond_for_each_slave(bond, slave, i) {
216 if (SLAVE_IS_OK(slave)) {
217 long long gap = compute_gap(slave);
220 least_loaded = slave;
229 /* Caller must hold bond lock for read */
230 static struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index, u32 skb_len)
232 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
233 struct tlb_client_info *hash_table;
234 struct slave *assigned_slave;
236 _lock_tx_hashtbl(bond);
238 hash_table = bond_info->tx_hashtbl;
239 assigned_slave = hash_table[hash_index].tx_slave;
240 if (!assigned_slave) {
241 assigned_slave = tlb_get_least_loaded_slave(bond);
243 if (assigned_slave) {
244 struct tlb_slave_info *slave_info =
245 &(SLAVE_TLB_INFO(assigned_slave));
246 u32 next_index = slave_info->head;
248 hash_table[hash_index].tx_slave = assigned_slave;
249 hash_table[hash_index].next = next_index;
250 hash_table[hash_index].prev = TLB_NULL_INDEX;
252 if (next_index != TLB_NULL_INDEX) {
253 hash_table[next_index].prev = hash_index;
256 slave_info->head = hash_index;
258 hash_table[hash_index].load_history;
262 if (assigned_slave) {
263 hash_table[hash_index].tx_bytes += skb_len;
266 _unlock_tx_hashtbl(bond);
268 return assigned_slave;
271 /*********************** rlb specific functions ***************************/
272 static inline void _lock_rx_hashtbl(struct bonding *bond)
274 spin_lock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
277 static inline void _unlock_rx_hashtbl(struct bonding *bond)
279 spin_unlock_bh(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
282 /* when an ARP REPLY is received from a client update its info
285 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
287 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
288 struct rlb_client_info *client_info;
291 _lock_rx_hashtbl(bond);
293 hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
294 client_info = &(bond_info->rx_hashtbl[hash_index]);
296 if ((client_info->assigned) &&
297 (client_info->ip_src == arp->ip_dst) &&
298 (client_info->ip_dst == arp->ip_src) &&
299 (compare_ether_addr_64bits(client_info->mac_dst, arp->mac_src))) {
300 /* update the clients MAC address */
301 memcpy(client_info->mac_dst, arp->mac_src, ETH_ALEN);
302 client_info->ntt = 1;
303 bond_info->rx_ntt = 1;
306 _unlock_rx_hashtbl(bond);
309 static void rlb_arp_recv(struct sk_buff *skb, struct bonding *bond,
314 if (skb->protocol != cpu_to_be16(ETH_P_ARP))
317 arp = (struct arp_pkt *) skb->data;
319 pr_debug("Packet has no ARP data\n");
323 if (!pskb_may_pull(skb, arp_hdr_len(bond->dev)))
326 if (skb->len < sizeof(struct arp_pkt)) {
327 pr_debug("Packet is too small to be an ARP\n");
331 if (arp->op_code == htons(ARPOP_REPLY)) {
332 /* update rx hash table for this ARP */
333 rlb_update_entry_from_arp(bond, arp);
334 pr_debug("Server received an ARP Reply from client\n");
338 /* Caller must hold bond lock for read */
339 static struct slave *rlb_next_rx_slave(struct bonding *bond)
341 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
342 struct slave *rx_slave, *slave, *start_at;
345 if (bond_info->next_rx_slave) {
346 start_at = bond_info->next_rx_slave;
348 start_at = bond->first_slave;
353 bond_for_each_slave_from(bond, slave, i, start_at) {
354 if (SLAVE_IS_OK(slave)) {
357 } else if (slave->speed > rx_slave->speed) {
364 bond_info->next_rx_slave = rx_slave->next;
370 /* teach the switch the mac of a disabled slave
371 * on the primary for fault tolerance
373 * Caller must hold bond->curr_slave_lock for write or bond lock for write
375 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
377 if (!bond->curr_active_slave) {
381 if (!bond->alb_info.primary_is_promisc) {
382 if (!dev_set_promiscuity(bond->curr_active_slave->dev, 1))
383 bond->alb_info.primary_is_promisc = 1;
385 bond->alb_info.primary_is_promisc = 0;
388 bond->alb_info.rlb_promisc_timeout_counter = 0;
390 alb_send_learning_packets(bond->curr_active_slave, addr);
393 /* slave being removed should not be active at this point
395 * Caller must hold bond lock for read
397 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
399 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
400 struct rlb_client_info *rx_hash_table;
401 u32 index, next_index;
403 /* clear slave from rx_hashtbl */
404 _lock_rx_hashtbl(bond);
406 rx_hash_table = bond_info->rx_hashtbl;
407 index = bond_info->rx_hashtbl_head;
408 for (; index != RLB_NULL_INDEX; index = next_index) {
409 next_index = rx_hash_table[index].next;
410 if (rx_hash_table[index].slave == slave) {
411 struct slave *assigned_slave = rlb_next_rx_slave(bond);
413 if (assigned_slave) {
414 rx_hash_table[index].slave = assigned_slave;
415 if (compare_ether_addr_64bits(rx_hash_table[index].mac_dst,
417 bond_info->rx_hashtbl[index].ntt = 1;
418 bond_info->rx_ntt = 1;
419 /* A slave has been removed from the
420 * table because it is either disabled
421 * or being released. We must retry the
422 * update to avoid clients from not
423 * being updated & disconnecting when
426 bond_info->rlb_update_retry_counter =
429 } else { /* there is no active slave */
430 rx_hash_table[index].slave = NULL;
435 _unlock_rx_hashtbl(bond);
437 write_lock_bh(&bond->curr_slave_lock);
439 if (slave != bond->curr_active_slave) {
440 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
443 write_unlock_bh(&bond->curr_slave_lock);
446 static void rlb_update_client(struct rlb_client_info *client_info)
450 if (!client_info->slave) {
454 for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
457 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
459 client_info->slave->dev,
461 client_info->mac_dst,
462 client_info->slave->dev->dev_addr,
463 client_info->mac_dst);
465 pr_err("%s: Error: failed to create an ARP packet\n",
466 client_info->slave->dev->master->name);
470 skb->dev = client_info->slave->dev;
472 if (client_info->tag) {
473 skb = vlan_put_tag(skb, client_info->vlan_id);
475 pr_err("%s: Error: failed to insert VLAN tag\n",
476 client_info->slave->dev->master->name);
485 /* sends ARP REPLIES that update the clients that need updating */
486 static void rlb_update_rx_clients(struct bonding *bond)
488 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
489 struct rlb_client_info *client_info;
492 _lock_rx_hashtbl(bond);
494 hash_index = bond_info->rx_hashtbl_head;
495 for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
496 client_info = &(bond_info->rx_hashtbl[hash_index]);
497 if (client_info->ntt) {
498 rlb_update_client(client_info);
499 if (bond_info->rlb_update_retry_counter == 0) {
500 client_info->ntt = 0;
505 /* do not update the entries again until this counter is zero so that
506 * not to confuse the clients.
508 bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
510 _unlock_rx_hashtbl(bond);
513 /* The slave was assigned a new mac address - update the clients */
514 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
516 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
517 struct rlb_client_info *client_info;
521 _lock_rx_hashtbl(bond);
523 hash_index = bond_info->rx_hashtbl_head;
524 for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
525 client_info = &(bond_info->rx_hashtbl[hash_index]);
527 if ((client_info->slave == slave) &&
528 compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
529 client_info->ntt = 1;
534 // update the team's flag only after the whole iteration
536 bond_info->rx_ntt = 1;
538 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
541 _unlock_rx_hashtbl(bond);
544 /* mark all clients using src_ip to be updated */
545 static void rlb_req_update_subnet_clients(struct bonding *bond, __be32 src_ip)
547 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
548 struct rlb_client_info *client_info;
551 _lock_rx_hashtbl(bond);
553 hash_index = bond_info->rx_hashtbl_head;
554 for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
555 client_info = &(bond_info->rx_hashtbl[hash_index]);
557 if (!client_info->slave) {
558 pr_err("%s: Error: found a client with no channel in the client's hash table\n",
562 /*update all clients using this src_ip, that are not assigned
563 * to the team's address (curr_active_slave) and have a known
564 * unicast mac address.
566 if ((client_info->ip_src == src_ip) &&
567 compare_ether_addr_64bits(client_info->slave->dev->dev_addr,
568 bond->dev->dev_addr) &&
569 compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
570 client_info->ntt = 1;
571 bond_info->rx_ntt = 1;
575 _unlock_rx_hashtbl(bond);
578 /* Caller must hold both bond and ptr locks for read */
579 static struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
581 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
582 struct arp_pkt *arp = arp_pkt(skb);
583 struct slave *assigned_slave;
584 struct rlb_client_info *client_info;
587 _lock_rx_hashtbl(bond);
589 hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_dst));
590 client_info = &(bond_info->rx_hashtbl[hash_index]);
592 if (client_info->assigned) {
593 if ((client_info->ip_src == arp->ip_src) &&
594 (client_info->ip_dst == arp->ip_dst)) {
595 /* the entry is already assigned to this client */
596 if (compare_ether_addr_64bits(arp->mac_dst, mac_bcast)) {
597 /* update mac address from arp */
598 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
601 assigned_slave = client_info->slave;
602 if (assigned_slave) {
603 _unlock_rx_hashtbl(bond);
604 return assigned_slave;
607 /* the entry is already assigned to some other client,
608 * move the old client to primary (curr_active_slave) so
609 * that the new client can be assigned to this entry.
611 if (bond->curr_active_slave &&
612 client_info->slave != bond->curr_active_slave) {
613 client_info->slave = bond->curr_active_slave;
614 rlb_update_client(client_info);
618 /* assign a new slave */
619 assigned_slave = rlb_next_rx_slave(bond);
621 if (assigned_slave) {
622 client_info->ip_src = arp->ip_src;
623 client_info->ip_dst = arp->ip_dst;
624 /* arp->mac_dst is broadcast for arp reqeusts.
625 * will be updated with clients actual unicast mac address
626 * upon receiving an arp reply.
628 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
629 client_info->slave = assigned_slave;
631 if (compare_ether_addr_64bits(client_info->mac_dst, mac_bcast)) {
632 client_info->ntt = 1;
633 bond->alb_info.rx_ntt = 1;
635 client_info->ntt = 0;
638 if (bond_vlan_used(bond)) {
639 if (!vlan_get_tag(skb, &client_info->vlan_id))
640 client_info->tag = 1;
643 if (!client_info->assigned) {
644 u32 prev_tbl_head = bond_info->rx_hashtbl_head;
645 bond_info->rx_hashtbl_head = hash_index;
646 client_info->next = prev_tbl_head;
647 if (prev_tbl_head != RLB_NULL_INDEX) {
648 bond_info->rx_hashtbl[prev_tbl_head].prev =
651 client_info->assigned = 1;
655 _unlock_rx_hashtbl(bond);
657 return assigned_slave;
660 /* chooses (and returns) transmit channel for arp reply
661 * does not choose channel for other arp types since they are
662 * sent on the curr_active_slave
664 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
666 struct arp_pkt *arp = arp_pkt(skb);
667 struct slave *tx_slave = NULL;
669 if (arp->op_code == htons(ARPOP_REPLY)) {
670 /* the arp must be sent on the selected
673 tx_slave = rlb_choose_channel(skb, bond);
675 memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
677 pr_debug("Server sent ARP Reply packet\n");
678 } else if (arp->op_code == htons(ARPOP_REQUEST)) {
679 /* Create an entry in the rx_hashtbl for this client as a
681 * When the arp reply is received the entry will be updated
682 * with the correct unicast address of the client.
684 rlb_choose_channel(skb, bond);
686 /* The ARP reply packets must be delayed so that
687 * they can cancel out the influence of the ARP request.
689 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
691 /* arp requests are broadcast and are sent on the primary
692 * the arp request will collapse all clients on the subnet to
693 * the primary slave. We must register these clients to be
694 * updated with their assigned mac.
696 rlb_req_update_subnet_clients(bond, arp->ip_src);
697 pr_debug("Server sent ARP Request packet\n");
703 /* Caller must hold bond lock for read */
704 static void rlb_rebalance(struct bonding *bond)
706 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
707 struct slave *assigned_slave;
708 struct rlb_client_info *client_info;
712 _lock_rx_hashtbl(bond);
715 hash_index = bond_info->rx_hashtbl_head;
716 for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
717 client_info = &(bond_info->rx_hashtbl[hash_index]);
718 assigned_slave = rlb_next_rx_slave(bond);
719 if (assigned_slave && (client_info->slave != assigned_slave)) {
720 client_info->slave = assigned_slave;
721 client_info->ntt = 1;
726 /* update the team's flag only after the whole iteration */
728 bond_info->rx_ntt = 1;
730 _unlock_rx_hashtbl(bond);
733 /* Caller must hold rx_hashtbl lock */
734 static void rlb_init_table_entry(struct rlb_client_info *entry)
736 memset(entry, 0, sizeof(struct rlb_client_info));
737 entry->next = RLB_NULL_INDEX;
738 entry->prev = RLB_NULL_INDEX;
741 static int rlb_initialize(struct bonding *bond)
743 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
744 struct rlb_client_info *new_hashtbl;
745 int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
748 new_hashtbl = kmalloc(size, GFP_KERNEL);
750 pr_err("%s: Error: Failed to allocate RLB hash table\n",
754 _lock_rx_hashtbl(bond);
756 bond_info->rx_hashtbl = new_hashtbl;
758 bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
760 for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
761 rlb_init_table_entry(bond_info->rx_hashtbl + i);
764 _unlock_rx_hashtbl(bond);
766 /* register to receive ARPs */
767 bond->recv_probe = rlb_arp_recv;
772 static void rlb_deinitialize(struct bonding *bond)
774 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
776 _lock_rx_hashtbl(bond);
778 kfree(bond_info->rx_hashtbl);
779 bond_info->rx_hashtbl = NULL;
780 bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
782 _unlock_rx_hashtbl(bond);
785 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
787 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
790 _lock_rx_hashtbl(bond);
792 curr_index = bond_info->rx_hashtbl_head;
793 while (curr_index != RLB_NULL_INDEX) {
794 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
795 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
796 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
798 if (curr->tag && (curr->vlan_id == vlan_id)) {
799 if (curr_index == bond_info->rx_hashtbl_head) {
800 bond_info->rx_hashtbl_head = next_index;
802 if (prev_index != RLB_NULL_INDEX) {
803 bond_info->rx_hashtbl[prev_index].next = next_index;
805 if (next_index != RLB_NULL_INDEX) {
806 bond_info->rx_hashtbl[next_index].prev = prev_index;
809 rlb_init_table_entry(curr);
812 curr_index = next_index;
815 _unlock_rx_hashtbl(bond);
818 /*********************** tlb/rlb shared functions *********************/
820 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
822 struct bonding *bond = bond_get_bond_by_slave(slave);
823 struct learning_pkt pkt;
824 int size = sizeof(struct learning_pkt);
827 memset(&pkt, 0, size);
828 memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
829 memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
830 pkt.type = cpu_to_be16(ETH_P_LOOP);
832 for (i = 0; i < MAX_LP_BURST; i++) {
836 skb = dev_alloc_skb(size);
841 data = skb_put(skb, size);
842 memcpy(data, &pkt, size);
844 skb_reset_mac_header(skb);
845 skb->network_header = skb->mac_header + ETH_HLEN;
846 skb->protocol = pkt.type;
847 skb->priority = TC_PRIO_CONTROL;
848 skb->dev = slave->dev;
850 if (bond_vlan_used(bond)) {
851 struct vlan_entry *vlan;
853 vlan = bond_next_vlan(bond,
854 bond->alb_info.current_alb_vlan);
856 bond->alb_info.current_alb_vlan = vlan;
862 skb = vlan_put_tag(skb, vlan->vlan_id);
864 pr_err("%s: Error: failed to insert VLAN tag\n",
874 /* hw is a boolean parameter that determines whether we should try and
875 * set the hw address of the device as well as the hw address of the
878 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[], int hw)
880 struct net_device *dev = slave->dev;
881 struct sockaddr s_addr;
884 memcpy(dev->dev_addr, addr, dev->addr_len);
888 /* for rlb each slave must have a unique hw mac addresses so that */
889 /* each slave will receive packets destined to a different mac */
890 memcpy(s_addr.sa_data, addr, dev->addr_len);
891 s_addr.sa_family = dev->type;
892 if (dev_set_mac_address(dev, &s_addr)) {
893 pr_err("%s: Error: dev_set_mac_address of dev %s failed!\n"
894 "ALB mode requires that the base driver support setting the hw address also when the network device's interface is open\n",
895 dev->master->name, dev->name);
902 * Swap MAC addresses between two slaves.
904 * Called with RTNL held, and no other locks.
908 static void alb_swap_mac_addr(struct bonding *bond, struct slave *slave1, struct slave *slave2)
910 u8 tmp_mac_addr[ETH_ALEN];
912 memcpy(tmp_mac_addr, slave1->dev->dev_addr, ETH_ALEN);
913 alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr, bond->alb_info.rlb_enabled);
914 alb_set_slave_mac_addr(slave2, tmp_mac_addr, bond->alb_info.rlb_enabled);
919 * Send learning packets after MAC address swap.
921 * Called with RTNL and no other locks
923 static void alb_fasten_mac_swap(struct bonding *bond, struct slave *slave1,
924 struct slave *slave2)
926 int slaves_state_differ = (SLAVE_IS_OK(slave1) != SLAVE_IS_OK(slave2));
927 struct slave *disabled_slave = NULL;
931 /* fasten the change in the switch */
932 if (SLAVE_IS_OK(slave1)) {
933 alb_send_learning_packets(slave1, slave1->dev->dev_addr);
934 if (bond->alb_info.rlb_enabled) {
935 /* inform the clients that the mac address
938 rlb_req_update_slave_clients(bond, slave1);
941 disabled_slave = slave1;
944 if (SLAVE_IS_OK(slave2)) {
945 alb_send_learning_packets(slave2, slave2->dev->dev_addr);
946 if (bond->alb_info.rlb_enabled) {
947 /* inform the clients that the mac address
950 rlb_req_update_slave_clients(bond, slave2);
953 disabled_slave = slave2;
956 if (bond->alb_info.rlb_enabled && slaves_state_differ) {
957 /* A disabled slave was assigned an active mac addr */
958 rlb_teach_disabled_mac_on_primary(bond,
959 disabled_slave->dev->dev_addr);
964 * alb_change_hw_addr_on_detach
965 * @bond: bonding we're working on
966 * @slave: the slave that was just detached
968 * We assume that @slave was already detached from the slave list.
970 * If @slave's permanent hw address is different both from its current
971 * address and from @bond's address, then somewhere in the bond there's
972 * a slave that has @slave's permanet address as its current address.
973 * We'll make sure that that slave no longer uses @slave's permanent address.
975 * Caller must hold RTNL and no other locks
977 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
982 perm_curr_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
983 slave->dev->dev_addr);
984 perm_bond_diff = compare_ether_addr_64bits(slave->perm_hwaddr,
985 bond->dev->dev_addr);
987 if (perm_curr_diff && perm_bond_diff) {
988 struct slave *tmp_slave;
991 bond_for_each_slave(bond, tmp_slave, i) {
992 if (!compare_ether_addr_64bits(slave->perm_hwaddr,
993 tmp_slave->dev->dev_addr)) {
1000 /* locking: needs RTNL and nothing else */
1001 alb_swap_mac_addr(bond, slave, tmp_slave);
1002 alb_fasten_mac_swap(bond, slave, tmp_slave);
1008 * alb_handle_addr_collision_on_attach
1009 * @bond: bonding we're working on
1010 * @slave: the slave that was just attached
1012 * checks uniqueness of slave's mac address and handles the case the
1013 * new slave uses the bonds mac address.
1015 * If the permanent hw address of @slave is @bond's hw address, we need to
1016 * find a different hw address to give @slave, that isn't in use by any other
1017 * slave in the bond. This address must be, of course, one of the permanent
1018 * addresses of the other slaves.
1020 * We go over the slave list, and for each slave there we compare its
1021 * permanent hw address with the current address of all the other slaves.
1022 * If no match was found, then we've found a slave with a permanent address
1023 * that isn't used by any other slave in the bond, so we can assign it to
1026 * assumption: this function is called before @slave is attached to the
1029 * caller must hold the bond lock for write since the mac addresses are compared
1030 * and may be swapped.
1032 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1034 struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1035 struct slave *has_bond_addr = bond->curr_active_slave;
1036 int i, j, found = 0;
1038 if (bond->slave_cnt == 0) {
1039 /* this is the first slave */
1043 /* if slave's mac address differs from bond's mac address
1044 * check uniqueness of slave's mac address against the other
1045 * slaves in the bond.
1047 if (compare_ether_addr_64bits(slave->perm_hwaddr, bond->dev->dev_addr)) {
1048 bond_for_each_slave(bond, tmp_slave1, i) {
1049 if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1050 slave->dev->dev_addr)) {
1059 /* Try setting slave mac to bond address and fall-through
1060 to code handling that situation below... */
1061 alb_set_slave_mac_addr(slave, bond->dev->dev_addr,
1062 bond->alb_info.rlb_enabled);
1065 /* The slave's address is equal to the address of the bond.
1066 * Search for a spare address in the bond for this slave.
1068 free_mac_slave = NULL;
1070 bond_for_each_slave(bond, tmp_slave1, i) {
1072 bond_for_each_slave(bond, tmp_slave2, j) {
1073 if (!compare_ether_addr_64bits(tmp_slave1->perm_hwaddr,
1074 tmp_slave2->dev->dev_addr)) {
1081 /* no slave has tmp_slave1's perm addr
1084 free_mac_slave = tmp_slave1;
1088 if (!has_bond_addr) {
1089 if (!compare_ether_addr_64bits(tmp_slave1->dev->dev_addr,
1090 bond->dev->dev_addr)) {
1092 has_bond_addr = tmp_slave1;
1097 if (free_mac_slave) {
1098 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr,
1099 bond->alb_info.rlb_enabled);
1101 pr_warning("%s: Warning: the hw address of slave %s is in use by the bond; giving it the hw address of %s\n",
1102 bond->dev->name, slave->dev->name,
1103 free_mac_slave->dev->name);
1105 } else if (has_bond_addr) {
1106 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",
1107 bond->dev->name, slave->dev->name);
1115 * alb_set_mac_address
1119 * In TLB mode all slaves are configured to the bond's hw address, but set
1120 * their dev_addr field to different addresses (based on their permanent hw
1123 * For each slave, this function sets the interface to the new address and then
1124 * changes its dev_addr field to its previous value.
1126 * Unwinding assumes bond's mac address has not yet changed.
1128 static int alb_set_mac_address(struct bonding *bond, void *addr)
1131 struct slave *slave, *stop_at;
1132 char tmp_addr[ETH_ALEN];
1136 if (bond->alb_info.rlb_enabled) {
1140 bond_for_each_slave(bond, slave, i) {
1141 /* save net_device's current hw address */
1142 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1144 res = dev_set_mac_address(slave->dev, addr);
1146 /* restore net_device's hw address */
1147 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1156 memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1157 sa.sa_family = bond->dev->type;
1159 /* unwind from head to the slave that failed */
1161 bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1162 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1163 dev_set_mac_address(slave->dev, &sa);
1164 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1170 /************************ exported alb funcions ************************/
1172 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1176 res = tlb_initialize(bond);
1182 bond->alb_info.rlb_enabled = 1;
1183 /* initialize rlb */
1184 res = rlb_initialize(bond);
1186 tlb_deinitialize(bond);
1190 bond->alb_info.rlb_enabled = 0;
1196 void bond_alb_deinitialize(struct bonding *bond)
1198 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1200 tlb_deinitialize(bond);
1202 if (bond_info->rlb_enabled) {
1203 rlb_deinitialize(bond);
1207 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1209 struct bonding *bond = netdev_priv(bond_dev);
1210 struct ethhdr *eth_data;
1211 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1212 struct slave *tx_slave = NULL;
1213 static const __be32 ip_bcast = htonl(0xffffffff);
1215 int do_tx_balance = 1;
1217 const u8 *hash_start = NULL;
1219 struct ipv6hdr *ip6hdr;
1221 skb_reset_mac_header(skb);
1222 eth_data = eth_hdr(skb);
1224 /* make sure that the curr_active_slave do not change during tx
1226 read_lock(&bond->curr_slave_lock);
1228 switch (ntohs(skb->protocol)) {
1230 const struct iphdr *iph = ip_hdr(skb);
1232 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast) ||
1233 (iph->daddr == ip_bcast) ||
1234 (iph->protocol == IPPROTO_IGMP)) {
1238 hash_start = (char *)&(iph->daddr);
1239 hash_size = sizeof(iph->daddr);
1243 /* IPv6 doesn't really use broadcast mac address, but leave
1244 * that here just in case.
1246 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_bcast)) {
1251 /* IPv6 uses all-nodes multicast as an equivalent to
1252 * broadcasts in IPv4.
1254 if (!compare_ether_addr_64bits(eth_data->h_dest, mac_v6_allmcast)) {
1259 /* Additianally, DAD probes should not be tx-balanced as that
1260 * will lead to false positives for duplicate addresses and
1261 * prevent address configuration from working.
1263 ip6hdr = ipv6_hdr(skb);
1264 if (ipv6_addr_any(&ip6hdr->saddr)) {
1269 hash_start = (char *)&(ipv6_hdr(skb)->daddr);
1270 hash_size = sizeof(ipv6_hdr(skb)->daddr);
1273 if (ipx_hdr(skb)->ipx_checksum != IPX_NO_CHECKSUM) {
1274 /* something is wrong with this packet */
1279 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1280 /* The only protocol worth balancing in
1281 * this family since it has an "ARP" like
1288 hash_start = (char*)eth_data->h_dest;
1289 hash_size = ETH_ALEN;
1293 if (bond_info->rlb_enabled) {
1294 tx_slave = rlb_arp_xmit(skb, bond);
1302 if (do_tx_balance) {
1303 hash_index = _simple_hash(hash_start, hash_size);
1304 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1308 /* unbalanced or unassigned, send through primary */
1309 tx_slave = bond->curr_active_slave;
1310 bond_info->unbalanced_load += skb->len;
1313 if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1314 if (tx_slave != bond->curr_active_slave) {
1315 memcpy(eth_data->h_source,
1316 tx_slave->dev->dev_addr,
1320 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1323 tlb_clear_slave(bond, tx_slave, 0);
1328 /* no suitable interface, frame not sent */
1331 read_unlock(&bond->curr_slave_lock);
1333 return NETDEV_TX_OK;
1336 void bond_alb_monitor(struct work_struct *work)
1338 struct bonding *bond = container_of(work, struct bonding,
1340 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1341 struct slave *slave;
1344 read_lock(&bond->lock);
1346 if (bond->slave_cnt == 0) {
1347 bond_info->tx_rebalance_counter = 0;
1348 bond_info->lp_counter = 0;
1352 bond_info->tx_rebalance_counter++;
1353 bond_info->lp_counter++;
1355 /* send learning packets */
1356 if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1357 /* change of curr_active_slave involves swapping of mac addresses.
1358 * in order to avoid this swapping from happening while
1359 * sending the learning packets, the curr_slave_lock must be held for
1362 read_lock(&bond->curr_slave_lock);
1364 bond_for_each_slave(bond, slave, i) {
1365 alb_send_learning_packets(slave, slave->dev->dev_addr);
1368 read_unlock(&bond->curr_slave_lock);
1370 bond_info->lp_counter = 0;
1373 /* rebalance tx traffic */
1374 if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1376 read_lock(&bond->curr_slave_lock);
1378 bond_for_each_slave(bond, slave, i) {
1379 tlb_clear_slave(bond, slave, 1);
1380 if (slave == bond->curr_active_slave) {
1381 SLAVE_TLB_INFO(slave).load =
1382 bond_info->unbalanced_load /
1383 BOND_TLB_REBALANCE_INTERVAL;
1384 bond_info->unbalanced_load = 0;
1388 read_unlock(&bond->curr_slave_lock);
1390 bond_info->tx_rebalance_counter = 0;
1393 /* handle rlb stuff */
1394 if (bond_info->rlb_enabled) {
1395 if (bond_info->primary_is_promisc &&
1396 (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1399 * dev_set_promiscuity requires rtnl and
1400 * nothing else. Avoid race with bond_close.
1402 read_unlock(&bond->lock);
1403 if (!rtnl_trylock()) {
1404 read_lock(&bond->lock);
1408 bond_info->rlb_promisc_timeout_counter = 0;
1410 /* If the primary was set to promiscuous mode
1411 * because a slave was disabled then
1412 * it can now leave promiscuous mode.
1414 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1415 bond_info->primary_is_promisc = 0;
1418 read_lock(&bond->lock);
1421 if (bond_info->rlb_rebalance) {
1422 bond_info->rlb_rebalance = 0;
1423 rlb_rebalance(bond);
1426 /* check if clients need updating */
1427 if (bond_info->rx_ntt) {
1428 if (bond_info->rlb_update_delay_counter) {
1429 --bond_info->rlb_update_delay_counter;
1431 rlb_update_rx_clients(bond);
1432 if (bond_info->rlb_update_retry_counter) {
1433 --bond_info->rlb_update_retry_counter;
1435 bond_info->rx_ntt = 0;
1442 queue_delayed_work(bond->wq, &bond->alb_work, alb_delta_in_ticks);
1444 read_unlock(&bond->lock);
1447 /* assumption: called before the slave is attached to the bond
1448 * and not locked by the bond lock
1450 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1454 res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr,
1455 bond->alb_info.rlb_enabled);
1460 /* caller must hold the bond lock for write since the mac addresses
1461 * are compared and may be swapped.
1463 read_lock(&bond->lock);
1465 res = alb_handle_addr_collision_on_attach(bond, slave);
1467 read_unlock(&bond->lock);
1473 tlb_init_slave(slave);
1475 /* order a rebalance ASAP */
1476 bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1478 if (bond->alb_info.rlb_enabled) {
1479 bond->alb_info.rlb_rebalance = 1;
1486 * Remove slave from tlb and rlb hash tables, and fix up MAC addresses
1489 * Caller must hold RTNL and no other locks
1491 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1493 if (bond->slave_cnt > 1) {
1494 alb_change_hw_addr_on_detach(bond, slave);
1497 tlb_clear_slave(bond, slave, 0);
1499 if (bond->alb_info.rlb_enabled) {
1500 bond->alb_info.next_rx_slave = NULL;
1501 rlb_clear_slave(bond, slave);
1505 /* Caller must hold bond lock for read */
1506 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1508 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1510 if (link == BOND_LINK_DOWN) {
1511 tlb_clear_slave(bond, slave, 0);
1512 if (bond->alb_info.rlb_enabled) {
1513 rlb_clear_slave(bond, slave);
1515 } else if (link == BOND_LINK_UP) {
1516 /* order a rebalance ASAP */
1517 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1518 if (bond->alb_info.rlb_enabled) {
1519 bond->alb_info.rlb_rebalance = 1;
1520 /* If the updelay module parameter is smaller than the
1521 * forwarding delay of the switch the rebalance will
1522 * not work because the rebalance arp replies will
1523 * not be forwarded to the clients..
1530 * bond_alb_handle_active_change - assign new curr_active_slave
1531 * @bond: our bonding struct
1532 * @new_slave: new slave to assign
1534 * Set the bond->curr_active_slave to @new_slave and handle
1535 * mac address swapping and promiscuity changes as needed.
1537 * If new_slave is NULL, caller must hold curr_slave_lock or
1538 * bond->lock for write.
1540 * If new_slave is not NULL, caller must hold RTNL, bond->lock for
1541 * read and curr_slave_lock for write. Processing here may sleep, so
1542 * no other locks may be held.
1544 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1545 __releases(&bond->curr_slave_lock)
1546 __releases(&bond->lock)
1547 __acquires(&bond->lock)
1548 __acquires(&bond->curr_slave_lock)
1550 struct slave *swap_slave;
1553 if (bond->curr_active_slave == new_slave) {
1557 if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1558 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1559 bond->alb_info.primary_is_promisc = 0;
1560 bond->alb_info.rlb_promisc_timeout_counter = 0;
1563 swap_slave = bond->curr_active_slave;
1564 bond->curr_active_slave = new_slave;
1566 if (!new_slave || (bond->slave_cnt == 0)) {
1570 /* set the new curr_active_slave to the bonds mac address
1571 * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1574 struct slave *tmp_slave;
1575 /* find slave that is holding the bond's mac address */
1576 bond_for_each_slave(bond, tmp_slave, i) {
1577 if (!compare_ether_addr_64bits(tmp_slave->dev->dev_addr,
1578 bond->dev->dev_addr)) {
1579 swap_slave = tmp_slave;
1586 * Arrange for swap_slave and new_slave to temporarily be
1587 * ignored so we can mess with their MAC addresses without
1588 * fear of interference from transmit activity.
1591 tlb_clear_slave(bond, swap_slave, 1);
1593 tlb_clear_slave(bond, new_slave, 1);
1595 write_unlock_bh(&bond->curr_slave_lock);
1596 read_unlock(&bond->lock);
1600 /* curr_active_slave must be set before calling alb_swap_mac_addr */
1602 /* swap mac address */
1603 alb_swap_mac_addr(bond, swap_slave, new_slave);
1605 /* set the new_slave to the bond mac address */
1606 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr,
1607 bond->alb_info.rlb_enabled);
1611 alb_fasten_mac_swap(bond, swap_slave, new_slave);
1612 read_lock(&bond->lock);
1614 read_lock(&bond->lock);
1615 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1618 write_lock_bh(&bond->curr_slave_lock);
1624 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1625 __acquires(&bond->lock)
1626 __releases(&bond->lock)
1628 struct bonding *bond = netdev_priv(bond_dev);
1629 struct sockaddr *sa = addr;
1630 struct slave *slave, *swap_slave;
1634 if (!is_valid_ether_addr(sa->sa_data)) {
1635 return -EADDRNOTAVAIL;
1638 res = alb_set_mac_address(bond, addr);
1643 memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1645 /* If there is no curr_active_slave there is nothing else to do.
1646 * Otherwise we'll need to pass the new address to it and handle
1649 if (!bond->curr_active_slave) {
1655 bond_for_each_slave(bond, slave, i) {
1656 if (!compare_ether_addr_64bits(slave->dev->dev_addr,
1657 bond_dev->dev_addr)) {
1664 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1665 alb_fasten_mac_swap(bond, swap_slave, bond->curr_active_slave);
1667 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr,
1668 bond->alb_info.rlb_enabled);
1670 read_lock(&bond->lock);
1671 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1672 if (bond->alb_info.rlb_enabled) {
1673 /* inform clients mac address has changed */
1674 rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1676 read_unlock(&bond->lock);
1682 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1684 if (bond->alb_info.current_alb_vlan &&
1685 (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1686 bond->alb_info.current_alb_vlan = NULL;
1689 if (bond->alb_info.rlb_enabled) {
1690 rlb_clear_vlan(bond, vlan_id);