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.
24 * 2003/06/25 - Shmulik Hen <shmulik.hen at intel dot com>
25 * - Fixed signed/unsigned calculation errors that caused load sharing
26 * to collapse to one slave under very heavy UDP Tx stress.
28 * 2003/08/06 - Amir Noam <amir.noam at intel dot com>
29 * - Add support for setting bond's MAC address with special
30 * handling required for ALB/TLB.
32 * 2003/12/01 - Shmulik Hen <shmulik.hen at intel dot com>
33 * - Code cleanup and style changes
35 * 2003/12/30 - Amir Noam <amir.noam at intel dot com>
36 * - Fixed: Cannot remove and re-enslave the original active slave.
38 * 2004/01/14 - Shmulik Hen <shmulik.hen at intel dot com>
39 * - Add capability to tag self generated packets in ALB/TLB modes.
42 //#define BONDING_DEBUG 1
44 #include <linux/skbuff.h>
45 #include <linux/netdevice.h>
46 #include <linux/etherdevice.h>
47 #include <linux/pkt_sched.h>
48 #include <linux/spinlock.h>
49 #include <linux/slab.h>
50 #include <linux/timer.h>
52 #include <linux/ipv6.h>
53 #include <linux/if_arp.h>
54 #include <linux/if_ether.h>
55 #include <linux/if_bonding.h>
56 #include <linux/if_vlan.h>
59 #include <asm/byteorder.h>
64 #define ALB_TIMER_TICKS_PER_SEC 10 /* should be a divisor of HZ */
65 #define BOND_TLB_REBALANCE_INTERVAL 10 /* In seconds, periodic re-balancing.
66 * Used for division - never set
69 #define BOND_ALB_LP_INTERVAL 1 /* In seconds, periodic send of
70 * learning packets to the switch
73 #define BOND_TLB_REBALANCE_TICKS (BOND_TLB_REBALANCE_INTERVAL \
74 * ALB_TIMER_TICKS_PER_SEC)
76 #define BOND_ALB_LP_TICKS (BOND_ALB_LP_INTERVAL \
77 * ALB_TIMER_TICKS_PER_SEC)
79 #define TLB_HASH_TABLE_SIZE 256 /* The size of the clients hash table.
80 * Note that this value MUST NOT be smaller
81 * because the key hash table is BYTE wide !
85 #define TLB_NULL_INDEX 0xffffffff
86 #define MAX_LP_BURST 3
89 #define RLB_HASH_TABLE_SIZE 256
90 #define RLB_NULL_INDEX 0xffffffff
91 #define RLB_UPDATE_DELAY 2*ALB_TIMER_TICKS_PER_SEC /* 2 seconds */
92 #define RLB_ARP_BURST_SIZE 2
93 #define RLB_UPDATE_RETRY 3 /* 3-ticks - must be smaller than the rlb
94 * rebalance interval (5 min).
96 /* RLB_PROMISC_TIMEOUT = 10 sec equals the time that the current slave is
97 * promiscuous after failover
99 #define RLB_PROMISC_TIMEOUT 10*ALB_TIMER_TICKS_PER_SEC
101 static const u8 mac_bcast[ETH_ALEN] = {0xff,0xff,0xff,0xff,0xff,0xff};
102 static const int alb_delta_in_ticks = HZ / ALB_TIMER_TICKS_PER_SEC;
105 struct learning_pkt {
106 u8 mac_dst[ETH_ALEN];
107 u8 mac_src[ETH_ALEN];
109 u8 padding[ETH_ZLEN - ETH_HLEN];
118 u8 mac_src[ETH_ALEN]; /* sender hardware address */
119 u32 ip_src; /* sender IP address */
120 u8 mac_dst[ETH_ALEN]; /* target hardware address */
121 u32 ip_dst; /* target IP address */
125 /* Forward declaration */
126 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[]);
128 static inline u8 _simple_hash(u8 *hash_start, int hash_size)
133 for (i = 0; i < hash_size; i++) {
134 hash ^= hash_start[i];
140 /*********************** tlb specific functions ***************************/
142 static inline void _lock_tx_hashtbl(struct bonding *bond)
144 spin_lock(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
147 static inline void _unlock_tx_hashtbl(struct bonding *bond)
149 spin_unlock(&(BOND_ALB_INFO(bond).tx_hashtbl_lock));
152 /* Caller must hold tx_hashtbl lock */
153 static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
156 entry->load_history = 1 + entry->tx_bytes /
157 BOND_TLB_REBALANCE_INTERVAL;
161 entry->tx_slave = NULL;
162 entry->next = TLB_NULL_INDEX;
163 entry->prev = TLB_NULL_INDEX;
166 static inline void tlb_init_slave(struct slave *slave)
168 SLAVE_TLB_INFO(slave).load = 0;
169 SLAVE_TLB_INFO(slave).head = TLB_NULL_INDEX;
172 /* Caller must hold bond lock for read */
173 static void tlb_clear_slave(struct bonding *bond, struct slave *slave, int save_load)
175 struct tlb_client_info *tx_hash_table;
178 _lock_tx_hashtbl(bond);
180 /* clear slave from tx_hashtbl */
181 tx_hash_table = BOND_ALB_INFO(bond).tx_hashtbl;
183 index = SLAVE_TLB_INFO(slave).head;
184 while (index != TLB_NULL_INDEX) {
185 u32 next_index = tx_hash_table[index].next;
186 tlb_init_table_entry(&tx_hash_table[index], save_load);
190 _unlock_tx_hashtbl(bond);
192 tlb_init_slave(slave);
195 /* Must be called before starting the monitor timer */
196 static int tlb_initialize(struct bonding *bond)
198 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
199 int size = TLB_HASH_TABLE_SIZE * sizeof(struct tlb_client_info);
202 spin_lock_init(&(bond_info->tx_hashtbl_lock));
204 _lock_tx_hashtbl(bond);
206 bond_info->tx_hashtbl = kmalloc(size, GFP_KERNEL);
207 if (!bond_info->tx_hashtbl) {
208 printk(KERN_ERR DRV_NAME
209 ": Error: %s: Failed to allocate TLB hash table\n",
211 _unlock_tx_hashtbl(bond);
215 memset(bond_info->tx_hashtbl, 0, size);
217 for (i = 0; i < TLB_HASH_TABLE_SIZE; i++) {
218 tlb_init_table_entry(&bond_info->tx_hashtbl[i], 1);
221 _unlock_tx_hashtbl(bond);
226 /* Must be called only after all slaves have been released */
227 static void tlb_deinitialize(struct bonding *bond)
229 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
231 _lock_tx_hashtbl(bond);
233 kfree(bond_info->tx_hashtbl);
234 bond_info->tx_hashtbl = NULL;
236 _unlock_tx_hashtbl(bond);
239 /* Caller must hold bond lock for read */
240 static struct slave *tlb_get_least_loaded_slave(struct bonding *bond)
242 struct slave *slave, *least_loaded;
246 /* Find the first enabled slave */
247 bond_for_each_slave(bond, slave, i) {
248 if (SLAVE_IS_OK(slave)) {
258 least_loaded = slave;
259 max_gap = (s64)(slave->speed << 20) - /* Convert to Megabit per sec */
260 (s64)(SLAVE_TLB_INFO(slave).load << 3); /* Bytes to bits */
262 /* Find the slave with the largest gap */
263 bond_for_each_slave_from(bond, slave, i, least_loaded) {
264 if (SLAVE_IS_OK(slave)) {
265 s64 gap = (s64)(slave->speed << 20) -
266 (s64)(SLAVE_TLB_INFO(slave).load << 3);
268 least_loaded = slave;
277 /* Caller must hold bond lock for read */
278 struct slave *tlb_choose_channel(struct bonding *bond, u32 hash_index, u32 skb_len)
280 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
281 struct tlb_client_info *hash_table;
282 struct slave *assigned_slave;
284 _lock_tx_hashtbl(bond);
286 hash_table = bond_info->tx_hashtbl;
287 assigned_slave = hash_table[hash_index].tx_slave;
288 if (!assigned_slave) {
289 assigned_slave = tlb_get_least_loaded_slave(bond);
291 if (assigned_slave) {
292 struct tlb_slave_info *slave_info =
293 &(SLAVE_TLB_INFO(assigned_slave));
294 u32 next_index = slave_info->head;
296 hash_table[hash_index].tx_slave = assigned_slave;
297 hash_table[hash_index].next = next_index;
298 hash_table[hash_index].prev = TLB_NULL_INDEX;
300 if (next_index != TLB_NULL_INDEX) {
301 hash_table[next_index].prev = hash_index;
304 slave_info->head = hash_index;
306 hash_table[hash_index].load_history;
310 if (assigned_slave) {
311 hash_table[hash_index].tx_bytes += skb_len;
314 _unlock_tx_hashtbl(bond);
316 return assigned_slave;
319 /*********************** rlb specific functions ***************************/
320 static inline void _lock_rx_hashtbl(struct bonding *bond)
322 spin_lock(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
325 static inline void _unlock_rx_hashtbl(struct bonding *bond)
327 spin_unlock(&(BOND_ALB_INFO(bond).rx_hashtbl_lock));
330 /* when an ARP REPLY is received from a client update its info
333 static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
335 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
336 struct rlb_client_info *client_info;
339 _lock_rx_hashtbl(bond);
341 hash_index = _simple_hash((u8*)&(arp->ip_src), sizeof(arp->ip_src));
342 client_info = &(bond_info->rx_hashtbl[hash_index]);
344 if ((client_info->assigned) &&
345 (client_info->ip_src == arp->ip_dst) &&
346 (client_info->ip_dst == arp->ip_src)) {
347 /* update the clients MAC address */
348 memcpy(client_info->mac_dst, arp->mac_src, ETH_ALEN);
349 client_info->ntt = 1;
350 bond_info->rx_ntt = 1;
353 _unlock_rx_hashtbl(bond);
356 static int rlb_arp_recv(struct sk_buff *skb, struct net_device *bond_dev, struct packet_type *ptype)
358 struct bonding *bond = bond_dev->priv;
359 struct arp_pkt *arp = (struct arp_pkt *)skb->data;
360 int res = NET_RX_DROP;
362 if (!(bond_dev->flags & IFF_MASTER)) {
367 dprintk("Packet has no ARP data\n");
371 if (skb->len < sizeof(struct arp_pkt)) {
372 dprintk("Packet is too small to be an ARP\n");
376 if (arp->op_code == htons(ARPOP_REPLY)) {
377 /* update rx hash table for this ARP */
378 rlb_update_entry_from_arp(bond, arp);
379 dprintk("Server received an ARP Reply from client\n");
382 res = NET_RX_SUCCESS;
390 /* Caller must hold bond lock for read */
391 static struct slave *rlb_next_rx_slave(struct bonding *bond)
393 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
394 struct slave *rx_slave, *slave, *start_at;
397 if (bond_info->next_rx_slave) {
398 start_at = bond_info->next_rx_slave;
400 start_at = bond->first_slave;
405 bond_for_each_slave_from(bond, slave, i, start_at) {
406 if (SLAVE_IS_OK(slave)) {
409 } else if (slave->speed > rx_slave->speed) {
416 bond_info->next_rx_slave = rx_slave->next;
422 /* teach the switch the mac of a disabled slave
423 * on the primary for fault tolerance
425 * Caller must hold bond->curr_slave_lock for write or bond lock for write
427 static void rlb_teach_disabled_mac_on_primary(struct bonding *bond, u8 addr[])
429 if (!bond->curr_active_slave) {
433 if (!bond->alb_info.primary_is_promisc) {
434 bond->alb_info.primary_is_promisc = 1;
435 dev_set_promiscuity(bond->curr_active_slave->dev, 1);
438 bond->alb_info.rlb_promisc_timeout_counter = 0;
440 alb_send_learning_packets(bond->curr_active_slave, addr);
443 /* slave being removed should not be active at this point
445 * Caller must hold bond lock for read
447 static void rlb_clear_slave(struct bonding *bond, struct slave *slave)
449 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
450 struct rlb_client_info *rx_hash_table;
451 u32 index, next_index;
453 /* clear slave from rx_hashtbl */
454 _lock_rx_hashtbl(bond);
456 rx_hash_table = bond_info->rx_hashtbl;
457 index = bond_info->rx_hashtbl_head;
458 for (; index != RLB_NULL_INDEX; index = next_index) {
459 next_index = rx_hash_table[index].next;
460 if (rx_hash_table[index].slave == slave) {
461 struct slave *assigned_slave = rlb_next_rx_slave(bond);
463 if (assigned_slave) {
464 rx_hash_table[index].slave = assigned_slave;
465 if (memcmp(rx_hash_table[index].mac_dst,
466 mac_bcast, ETH_ALEN)) {
467 bond_info->rx_hashtbl[index].ntt = 1;
468 bond_info->rx_ntt = 1;
469 /* A slave has been removed from the
470 * table because it is either disabled
471 * or being released. We must retry the
472 * update to avoid clients from not
473 * being updated & disconnecting when
476 bond_info->rlb_update_retry_counter =
479 } else { /* there is no active slave */
480 rx_hash_table[index].slave = NULL;
485 _unlock_rx_hashtbl(bond);
487 write_lock(&bond->curr_slave_lock);
489 if (slave != bond->curr_active_slave) {
490 rlb_teach_disabled_mac_on_primary(bond, slave->dev->dev_addr);
493 write_unlock(&bond->curr_slave_lock);
496 static void rlb_update_client(struct rlb_client_info *client_info)
500 if (!client_info->slave) {
504 for (i = 0; i < RLB_ARP_BURST_SIZE; i++) {
507 skb = arp_create(ARPOP_REPLY, ETH_P_ARP,
509 client_info->slave->dev,
511 client_info->mac_dst,
512 client_info->slave->dev->dev_addr,
513 client_info->mac_dst);
515 printk(KERN_ERR DRV_NAME
516 ": Error: failed to create an ARP packet\n");
520 skb->dev = client_info->slave->dev;
522 if (client_info->tag) {
523 skb = vlan_put_tag(skb, client_info->vlan_id);
525 printk(KERN_ERR DRV_NAME
526 ": Error: failed to insert VLAN tag\n");
535 /* sends ARP REPLIES that update the clients that need updating */
536 static void rlb_update_rx_clients(struct bonding *bond)
538 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
539 struct rlb_client_info *client_info;
542 _lock_rx_hashtbl(bond);
544 hash_index = bond_info->rx_hashtbl_head;
545 for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
546 client_info = &(bond_info->rx_hashtbl[hash_index]);
547 if (client_info->ntt) {
548 rlb_update_client(client_info);
549 if (bond_info->rlb_update_retry_counter == 0) {
550 client_info->ntt = 0;
555 /* do not update the entries again untill this counter is zero so that
556 * not to confuse the clients.
558 bond_info->rlb_update_delay_counter = RLB_UPDATE_DELAY;
560 _unlock_rx_hashtbl(bond);
563 /* The slave was assigned a new mac address - update the clients */
564 static void rlb_req_update_slave_clients(struct bonding *bond, struct slave *slave)
566 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
567 struct rlb_client_info *client_info;
571 _lock_rx_hashtbl(bond);
573 hash_index = bond_info->rx_hashtbl_head;
574 for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
575 client_info = &(bond_info->rx_hashtbl[hash_index]);
577 if ((client_info->slave == slave) &&
578 memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
579 client_info->ntt = 1;
584 // update the team's flag only after the whole iteration
586 bond_info->rx_ntt = 1;
588 bond_info->rlb_update_retry_counter = RLB_UPDATE_RETRY;
591 _unlock_rx_hashtbl(bond);
594 /* mark all clients using src_ip to be updated */
595 static void rlb_req_update_subnet_clients(struct bonding *bond, u32 src_ip)
597 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
598 struct rlb_client_info *client_info;
601 _lock_rx_hashtbl(bond);
603 hash_index = bond_info->rx_hashtbl_head;
604 for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
605 client_info = &(bond_info->rx_hashtbl[hash_index]);
607 if (!client_info->slave) {
608 printk(KERN_ERR DRV_NAME
609 ": Error: found a client with no channel in "
610 "the client's hash table\n");
613 /*update all clients using this src_ip, that are not assigned
614 * to the team's address (curr_active_slave) and have a known
615 * unicast mac address.
617 if ((client_info->ip_src == src_ip) &&
618 memcmp(client_info->slave->dev->dev_addr,
619 bond->dev->dev_addr, ETH_ALEN) &&
620 memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
621 client_info->ntt = 1;
622 bond_info->rx_ntt = 1;
626 _unlock_rx_hashtbl(bond);
629 /* Caller must hold both bond and ptr locks for read */
630 struct slave *rlb_choose_channel(struct sk_buff *skb, struct bonding *bond)
632 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
633 struct arp_pkt *arp = (struct arp_pkt *)skb->nh.raw;
634 struct slave *assigned_slave;
635 struct rlb_client_info *client_info;
638 _lock_rx_hashtbl(bond);
640 hash_index = _simple_hash((u8 *)&arp->ip_dst, sizeof(arp->ip_src));
641 client_info = &(bond_info->rx_hashtbl[hash_index]);
643 if (client_info->assigned) {
644 if ((client_info->ip_src == arp->ip_src) &&
645 (client_info->ip_dst == arp->ip_dst)) {
646 /* the entry is already assigned to this client */
647 if (memcmp(arp->mac_dst, mac_bcast, ETH_ALEN)) {
648 /* update mac address from arp */
649 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
652 assigned_slave = client_info->slave;
653 if (assigned_slave) {
654 _unlock_rx_hashtbl(bond);
655 return assigned_slave;
658 /* the entry is already assigned to some other client,
659 * move the old client to primary (curr_active_slave) so
660 * that the new client can be assigned to this entry.
662 if (bond->curr_active_slave &&
663 client_info->slave != bond->curr_active_slave) {
664 client_info->slave = bond->curr_active_slave;
665 rlb_update_client(client_info);
669 /* assign a new slave */
670 assigned_slave = rlb_next_rx_slave(bond);
672 if (assigned_slave) {
673 client_info->ip_src = arp->ip_src;
674 client_info->ip_dst = arp->ip_dst;
675 /* arp->mac_dst is broadcast for arp reqeusts.
676 * will be updated with clients actual unicast mac address
677 * upon receiving an arp reply.
679 memcpy(client_info->mac_dst, arp->mac_dst, ETH_ALEN);
680 client_info->slave = assigned_slave;
682 if (memcmp(client_info->mac_dst, mac_bcast, ETH_ALEN)) {
683 client_info->ntt = 1;
684 bond->alb_info.rx_ntt = 1;
686 client_info->ntt = 0;
689 if (!list_empty(&bond->vlan_list)) {
690 unsigned short vlan_id;
691 int res = vlan_get_tag(skb, &vlan_id);
693 client_info->tag = 1;
694 client_info->vlan_id = vlan_id;
698 if (!client_info->assigned) {
699 u32 prev_tbl_head = bond_info->rx_hashtbl_head;
700 bond_info->rx_hashtbl_head = hash_index;
701 client_info->next = prev_tbl_head;
702 if (prev_tbl_head != RLB_NULL_INDEX) {
703 bond_info->rx_hashtbl[prev_tbl_head].prev =
706 client_info->assigned = 1;
710 _unlock_rx_hashtbl(bond);
712 return assigned_slave;
715 /* chooses (and returns) transmit channel for arp reply
716 * does not choose channel for other arp types since they are
717 * sent on the curr_active_slave
719 static struct slave *rlb_arp_xmit(struct sk_buff *skb, struct bonding *bond)
721 struct arp_pkt *arp = (struct arp_pkt *)skb->nh.raw;
722 struct slave *tx_slave = NULL;
724 if (arp->op_code == __constant_htons(ARPOP_REPLY)) {
725 /* the arp must be sent on the selected
728 tx_slave = rlb_choose_channel(skb, bond);
730 memcpy(arp->mac_src,tx_slave->dev->dev_addr, ETH_ALEN);
732 dprintk("Server sent ARP Reply packet\n");
733 } else if (arp->op_code == __constant_htons(ARPOP_REQUEST)) {
734 /* Create an entry in the rx_hashtbl for this client as a
736 * When the arp reply is received the entry will be updated
737 * with the correct unicast address of the client.
739 rlb_choose_channel(skb, bond);
741 /* The ARP relpy packets must be delayed so that
742 * they can cancel out the influence of the ARP request.
744 bond->alb_info.rlb_update_delay_counter = RLB_UPDATE_DELAY;
746 /* arp requests are broadcast and are sent on the primary
747 * the arp request will collapse all clients on the subnet to
748 * the primary slave. We must register these clients to be
749 * updated with their assigned mac.
751 rlb_req_update_subnet_clients(bond, arp->ip_src);
752 dprintk("Server sent ARP Request packet\n");
758 /* Caller must hold bond lock for read */
759 static void rlb_rebalance(struct bonding *bond)
761 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
762 struct slave *assigned_slave;
763 struct rlb_client_info *client_info;
767 _lock_rx_hashtbl(bond);
770 hash_index = bond_info->rx_hashtbl_head;
771 for (; hash_index != RLB_NULL_INDEX; hash_index = client_info->next) {
772 client_info = &(bond_info->rx_hashtbl[hash_index]);
773 assigned_slave = rlb_next_rx_slave(bond);
774 if (assigned_slave && (client_info->slave != assigned_slave)) {
775 client_info->slave = assigned_slave;
776 client_info->ntt = 1;
781 /* update the team's flag only after the whole iteration */
783 bond_info->rx_ntt = 1;
785 _unlock_rx_hashtbl(bond);
788 /* Caller must hold rx_hashtbl lock */
789 static void rlb_init_table_entry(struct rlb_client_info *entry)
791 memset(entry, 0, sizeof(struct rlb_client_info));
792 entry->next = RLB_NULL_INDEX;
793 entry->prev = RLB_NULL_INDEX;
796 static int rlb_initialize(struct bonding *bond)
798 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
799 struct packet_type *pk_type = &(BOND_ALB_INFO(bond).rlb_pkt_type);
800 int size = RLB_HASH_TABLE_SIZE * sizeof(struct rlb_client_info);
803 spin_lock_init(&(bond_info->rx_hashtbl_lock));
805 _lock_rx_hashtbl(bond);
807 bond_info->rx_hashtbl = kmalloc(size, GFP_KERNEL);
808 if (!bond_info->rx_hashtbl) {
809 printk(KERN_ERR DRV_NAME
810 ": Error: %s: Failed to allocate RLB hash table\n",
812 _unlock_rx_hashtbl(bond);
816 bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
818 for (i = 0; i < RLB_HASH_TABLE_SIZE; i++) {
819 rlb_init_table_entry(bond_info->rx_hashtbl + i);
822 _unlock_rx_hashtbl(bond);
824 /*initialize packet type*/
825 pk_type->type = __constant_htons(ETH_P_ARP);
826 pk_type->dev = bond->dev;
827 pk_type->func = rlb_arp_recv;
829 /* register to receive ARPs */
830 dev_add_pack(pk_type);
835 static void rlb_deinitialize(struct bonding *bond)
837 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
839 dev_remove_pack(&(bond_info->rlb_pkt_type));
841 _lock_rx_hashtbl(bond);
843 kfree(bond_info->rx_hashtbl);
844 bond_info->rx_hashtbl = NULL;
845 bond_info->rx_hashtbl_head = RLB_NULL_INDEX;
847 _unlock_rx_hashtbl(bond);
850 static void rlb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
852 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
855 _lock_rx_hashtbl(bond);
857 curr_index = bond_info->rx_hashtbl_head;
858 while (curr_index != RLB_NULL_INDEX) {
859 struct rlb_client_info *curr = &(bond_info->rx_hashtbl[curr_index]);
860 u32 next_index = bond_info->rx_hashtbl[curr_index].next;
861 u32 prev_index = bond_info->rx_hashtbl[curr_index].prev;
863 if (curr->tag && (curr->vlan_id == vlan_id)) {
864 if (curr_index == bond_info->rx_hashtbl_head) {
865 bond_info->rx_hashtbl_head = next_index;
867 if (prev_index != RLB_NULL_INDEX) {
868 bond_info->rx_hashtbl[prev_index].next = next_index;
870 if (next_index != RLB_NULL_INDEX) {
871 bond_info->rx_hashtbl[next_index].prev = prev_index;
874 rlb_init_table_entry(curr);
877 curr_index = next_index;
880 _unlock_rx_hashtbl(bond);
883 /*********************** tlb/rlb shared functions *********************/
885 static void alb_send_learning_packets(struct slave *slave, u8 mac_addr[])
887 struct bonding *bond = bond_get_bond_by_slave(slave);
888 struct learning_pkt pkt;
889 int size = sizeof(struct learning_pkt);
892 memset(&pkt, 0, size);
893 memcpy(pkt.mac_dst, mac_addr, ETH_ALEN);
894 memcpy(pkt.mac_src, mac_addr, ETH_ALEN);
895 pkt.type = __constant_htons(ETH_P_LOOP);
897 for (i = 0; i < MAX_LP_BURST; i++) {
901 skb = dev_alloc_skb(size);
906 data = skb_put(skb, size);
907 memcpy(data, &pkt, size);
910 skb->nh.raw = data + ETH_HLEN;
911 skb->protocol = pkt.type;
912 skb->priority = TC_PRIO_CONTROL;
913 skb->dev = slave->dev;
915 if (!list_empty(&bond->vlan_list)) {
916 struct vlan_entry *vlan;
918 vlan = bond_next_vlan(bond,
919 bond->alb_info.current_alb_vlan);
921 bond->alb_info.current_alb_vlan = vlan;
927 skb = vlan_put_tag(skb, vlan->vlan_id);
929 printk(KERN_ERR DRV_NAME
930 ": Error: failed to insert VLAN tag\n");
939 /* hw is a boolean parameter that determines whether we should try and
940 * set the hw address of the device as well as the hw address of the
943 static int alb_set_slave_mac_addr(struct slave *slave, u8 addr[], int hw)
945 struct net_device *dev = slave->dev;
946 struct sockaddr s_addr;
949 memcpy(dev->dev_addr, addr, dev->addr_len);
953 /* for rlb each slave must have a unique hw mac addresses so that */
954 /* each slave will receive packets destined to a different mac */
955 memcpy(s_addr.sa_data, addr, dev->addr_len);
956 s_addr.sa_family = dev->type;
957 if (dev->set_mac_address(dev, &s_addr)) {
958 printk(KERN_ERR DRV_NAME
959 ": Error: dev->set_mac_address of dev %s failed! ALB "
960 "mode requires that the base driver support setting "
961 "the hw address also when the network device's "
962 "interface is open\n",
969 /* Caller must hold bond lock for write or curr_slave_lock for write*/
970 static void alb_swap_mac_addr(struct bonding *bond, struct slave *slave1, struct slave *slave2)
972 struct slave *disabled_slave = NULL;
973 u8 tmp_mac_addr[ETH_ALEN];
974 int slaves_state_differ;
976 slaves_state_differ = (SLAVE_IS_OK(slave1) != SLAVE_IS_OK(slave2));
978 memcpy(tmp_mac_addr, slave1->dev->dev_addr, ETH_ALEN);
979 alb_set_slave_mac_addr(slave1, slave2->dev->dev_addr, bond->alb_info.rlb_enabled);
980 alb_set_slave_mac_addr(slave2, tmp_mac_addr, bond->alb_info.rlb_enabled);
982 /* fasten the change in the switch */
983 if (SLAVE_IS_OK(slave1)) {
984 alb_send_learning_packets(slave1, slave1->dev->dev_addr);
985 if (bond->alb_info.rlb_enabled) {
986 /* inform the clients that the mac address
989 rlb_req_update_slave_clients(bond, slave1);
992 disabled_slave = slave1;
995 if (SLAVE_IS_OK(slave2)) {
996 alb_send_learning_packets(slave2, slave2->dev->dev_addr);
997 if (bond->alb_info.rlb_enabled) {
998 /* inform the clients that the mac address
1001 rlb_req_update_slave_clients(bond, slave2);
1004 disabled_slave = slave2;
1007 if (bond->alb_info.rlb_enabled && slaves_state_differ) {
1008 /* A disabled slave was assigned an active mac addr */
1009 rlb_teach_disabled_mac_on_primary(bond,
1010 disabled_slave->dev->dev_addr);
1015 * alb_change_hw_addr_on_detach
1016 * @bond: bonding we're working on
1017 * @slave: the slave that was just detached
1019 * We assume that @slave was already detached from the slave list.
1021 * If @slave's permanent hw address is different both from its current
1022 * address and from @bond's address, then somewhere in the bond there's
1023 * a slave that has @slave's permanet address as its current address.
1024 * We'll make sure that that slave no longer uses @slave's permanent address.
1026 * Caller must hold bond lock
1028 static void alb_change_hw_addr_on_detach(struct bonding *bond, struct slave *slave)
1033 perm_curr_diff = memcmp(slave->perm_hwaddr,
1034 slave->dev->dev_addr,
1036 perm_bond_diff = memcmp(slave->perm_hwaddr,
1037 bond->dev->dev_addr,
1040 if (perm_curr_diff && perm_bond_diff) {
1041 struct slave *tmp_slave;
1044 bond_for_each_slave(bond, tmp_slave, i) {
1045 if (!memcmp(slave->perm_hwaddr,
1046 tmp_slave->dev->dev_addr,
1054 alb_swap_mac_addr(bond, slave, tmp_slave);
1060 * alb_handle_addr_collision_on_attach
1061 * @bond: bonding we're working on
1062 * @slave: the slave that was just attached
1064 * checks uniqueness of slave's mac address and handles the case the
1065 * new slave uses the bonds mac address.
1067 * If the permanent hw address of @slave is @bond's hw address, we need to
1068 * find a different hw address to give @slave, that isn't in use by any other
1069 * slave in the bond. This address must be, of course, one of the premanent
1070 * addresses of the other slaves.
1072 * We go over the slave list, and for each slave there we compare its
1073 * permanent hw address with the current address of all the other slaves.
1074 * If no match was found, then we've found a slave with a permanent address
1075 * that isn't used by any other slave in the bond, so we can assign it to
1078 * assumption: this function is called before @slave is attached to the
1081 * caller must hold the bond lock for write since the mac addresses are compared
1082 * and may be swapped.
1084 static int alb_handle_addr_collision_on_attach(struct bonding *bond, struct slave *slave)
1086 struct slave *tmp_slave1, *tmp_slave2, *free_mac_slave;
1087 struct slave *has_bond_addr = bond->curr_active_slave;
1088 int i, j, found = 0;
1090 if (bond->slave_cnt == 0) {
1091 /* this is the first slave */
1095 /* if slave's mac address differs from bond's mac address
1096 * check uniqueness of slave's mac address against the other
1097 * slaves in the bond.
1099 if (memcmp(slave->perm_hwaddr, bond->dev->dev_addr, ETH_ALEN)) {
1100 bond_for_each_slave(bond, tmp_slave1, i) {
1101 if (!memcmp(tmp_slave1->dev->dev_addr, slave->dev->dev_addr,
1109 /* a slave was found that is using the mac address
1112 printk(KERN_ERR DRV_NAME
1113 ": Error: the hw address of slave %s is not "
1114 "unique - cannot enslave it!",
1122 /* The slave's address is equal to the address of the bond.
1123 * Search for a spare address in the bond for this slave.
1125 free_mac_slave = NULL;
1127 bond_for_each_slave(bond, tmp_slave1, i) {
1129 bond_for_each_slave(bond, tmp_slave2, j) {
1130 if (!memcmp(tmp_slave1->perm_hwaddr,
1131 tmp_slave2->dev->dev_addr,
1139 /* no slave has tmp_slave1's perm addr
1142 free_mac_slave = tmp_slave1;
1146 if (!has_bond_addr) {
1147 if (!memcmp(tmp_slave1->dev->dev_addr,
1148 bond->dev->dev_addr,
1151 has_bond_addr = tmp_slave1;
1156 if (free_mac_slave) {
1157 alb_set_slave_mac_addr(slave, free_mac_slave->perm_hwaddr,
1158 bond->alb_info.rlb_enabled);
1160 printk(KERN_WARNING DRV_NAME
1161 ": Warning: the hw address of slave %s is in use by "
1162 "the bond; giving it the hw address of %s\n",
1163 slave->dev->name, free_mac_slave->dev->name);
1165 } else if (has_bond_addr) {
1166 printk(KERN_ERR DRV_NAME
1167 ": Error: the hw address of slave %s is in use by the "
1168 "bond; couldn't find a slave with a free hw address to "
1169 "give it (this should not have happened)\n",
1178 * alb_set_mac_address
1182 * In TLB mode all slaves are configured to the bond's hw address, but set
1183 * their dev_addr field to different addresses (based on their permanent hw
1186 * For each slave, this function sets the interface to the new address and then
1187 * changes its dev_addr field to its previous value.
1189 * Unwinding assumes bond's mac address has not yet changed.
1191 static int alb_set_mac_address(struct bonding *bond, void *addr)
1194 struct slave *slave, *stop_at;
1195 char tmp_addr[ETH_ALEN];
1199 if (bond->alb_info.rlb_enabled) {
1203 bond_for_each_slave(bond, slave, i) {
1204 if (slave->dev->set_mac_address == NULL) {
1209 /* save net_device's current hw address */
1210 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1212 res = slave->dev->set_mac_address(slave->dev, addr);
1214 /* restore net_device's hw address */
1215 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1225 memcpy(sa.sa_data, bond->dev->dev_addr, bond->dev->addr_len);
1226 sa.sa_family = bond->dev->type;
1228 /* unwind from head to the slave that failed */
1230 bond_for_each_slave_from_to(bond, slave, i, bond->first_slave, stop_at) {
1231 memcpy(tmp_addr, slave->dev->dev_addr, ETH_ALEN);
1232 slave->dev->set_mac_address(slave->dev, &sa);
1233 memcpy(slave->dev->dev_addr, tmp_addr, ETH_ALEN);
1239 /************************ exported alb funcions ************************/
1241 int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
1245 res = tlb_initialize(bond);
1251 bond->alb_info.rlb_enabled = 1;
1252 /* initialize rlb */
1253 res = rlb_initialize(bond);
1255 tlb_deinitialize(bond);
1263 void bond_alb_deinitialize(struct bonding *bond)
1265 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1267 tlb_deinitialize(bond);
1269 if (bond_info->rlb_enabled) {
1270 rlb_deinitialize(bond);
1274 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
1276 struct bonding *bond = bond_dev->priv;
1277 struct ethhdr *eth_data;
1278 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1279 struct slave *tx_slave = NULL;
1280 static u32 ip_bcast = 0xffffffff;
1282 int do_tx_balance = 1;
1284 u8 *hash_start = NULL;
1287 skb->mac.raw = (unsigned char *)skb->data;
1288 eth_data = (struct ethhdr *)skb->data;
1290 /* make sure that the curr_active_slave and the slaves list do
1291 * not change during tx
1293 read_lock(&bond->lock);
1294 read_lock(&bond->curr_slave_lock);
1296 if (!BOND_IS_OK(bond)) {
1300 switch (ntohs(skb->protocol)) {
1302 if ((memcmp(eth_data->h_dest, mac_bcast, ETH_ALEN) == 0) ||
1303 (skb->nh.iph->daddr == ip_bcast)) {
1307 hash_start = (char*)&(skb->nh.iph->daddr);
1308 hash_size = sizeof(skb->nh.iph->daddr);
1311 if (memcmp(eth_data->h_dest, mac_bcast, ETH_ALEN) == 0) {
1316 hash_start = (char*)&(skb->nh.ipv6h->daddr);
1317 hash_size = sizeof(skb->nh.ipv6h->daddr);
1320 if (ipx_hdr(skb)->ipx_checksum !=
1321 __constant_htons(IPX_NO_CHECKSUM)) {
1322 /* something is wrong with this packet */
1327 if (ipx_hdr(skb)->ipx_type != IPX_TYPE_NCP) {
1328 /* The only protocol worth balancing in
1329 * this family since it has an "ARP" like
1336 hash_start = (char*)eth_data->h_dest;
1337 hash_size = ETH_ALEN;
1341 if (bond_info->rlb_enabled) {
1342 tx_slave = rlb_arp_xmit(skb, bond);
1350 if (do_tx_balance) {
1351 hash_index = _simple_hash(hash_start, hash_size);
1352 tx_slave = tlb_choose_channel(bond, hash_index, skb->len);
1356 /* unbalanced or unassigned, send through primary */
1357 tx_slave = bond->curr_active_slave;
1358 bond_info->unbalanced_load += skb->len;
1361 if (tx_slave && SLAVE_IS_OK(tx_slave)) {
1362 if (tx_slave != bond->curr_active_slave) {
1363 memcpy(eth_data->h_source,
1364 tx_slave->dev->dev_addr,
1368 res = bond_dev_queue_xmit(bond, skb, tx_slave->dev);
1371 tlb_clear_slave(bond, tx_slave, 0);
1377 /* no suitable interface, frame not sent */
1380 read_unlock(&bond->curr_slave_lock);
1381 read_unlock(&bond->lock);
1385 void bond_alb_monitor(struct bonding *bond)
1387 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1388 struct slave *slave;
1391 read_lock(&bond->lock);
1393 if (bond->kill_timers) {
1397 if (bond->slave_cnt == 0) {
1398 bond_info->tx_rebalance_counter = 0;
1399 bond_info->lp_counter = 0;
1403 bond_info->tx_rebalance_counter++;
1404 bond_info->lp_counter++;
1406 /* send learning packets */
1407 if (bond_info->lp_counter >= BOND_ALB_LP_TICKS) {
1408 /* change of curr_active_slave involves swapping of mac addresses.
1409 * in order to avoid this swapping from happening while
1410 * sending the learning packets, the curr_slave_lock must be held for
1413 read_lock(&bond->curr_slave_lock);
1415 bond_for_each_slave(bond, slave, i) {
1416 alb_send_learning_packets(slave,slave->dev->dev_addr);
1419 read_unlock(&bond->curr_slave_lock);
1421 bond_info->lp_counter = 0;
1424 /* rebalance tx traffic */
1425 if (bond_info->tx_rebalance_counter >= BOND_TLB_REBALANCE_TICKS) {
1427 read_lock(&bond->curr_slave_lock);
1429 bond_for_each_slave(bond, slave, i) {
1430 tlb_clear_slave(bond, slave, 1);
1431 if (slave == bond->curr_active_slave) {
1432 SLAVE_TLB_INFO(slave).load =
1433 bond_info->unbalanced_load /
1434 BOND_TLB_REBALANCE_INTERVAL;
1435 bond_info->unbalanced_load = 0;
1439 read_unlock(&bond->curr_slave_lock);
1441 bond_info->tx_rebalance_counter = 0;
1444 /* handle rlb stuff */
1445 if (bond_info->rlb_enabled) {
1446 /* the following code changes the promiscuity of the
1447 * the curr_active_slave. It needs to be locked with a
1448 * write lock to protect from other code that also
1449 * sets the promiscuity.
1451 write_lock(&bond->curr_slave_lock);
1453 if (bond_info->primary_is_promisc &&
1454 (++bond_info->rlb_promisc_timeout_counter >= RLB_PROMISC_TIMEOUT)) {
1456 bond_info->rlb_promisc_timeout_counter = 0;
1458 /* If the primary was set to promiscuous mode
1459 * because a slave was disabled then
1460 * it can now leave promiscuous mode.
1462 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1463 bond_info->primary_is_promisc = 0;
1466 write_unlock(&bond->curr_slave_lock);
1468 if (bond_info->rlb_rebalance) {
1469 bond_info->rlb_rebalance = 0;
1470 rlb_rebalance(bond);
1473 /* check if clients need updating */
1474 if (bond_info->rx_ntt) {
1475 if (bond_info->rlb_update_delay_counter) {
1476 --bond_info->rlb_update_delay_counter;
1478 rlb_update_rx_clients(bond);
1479 if (bond_info->rlb_update_retry_counter) {
1480 --bond_info->rlb_update_retry_counter;
1482 bond_info->rx_ntt = 0;
1489 mod_timer(&(bond_info->alb_timer), jiffies + alb_delta_in_ticks);
1491 read_unlock(&bond->lock);
1494 /* assumption: called before the slave is attached to the bond
1495 * and not locked by the bond lock
1497 int bond_alb_init_slave(struct bonding *bond, struct slave *slave)
1501 res = alb_set_slave_mac_addr(slave, slave->perm_hwaddr,
1502 bond->alb_info.rlb_enabled);
1507 /* caller must hold the bond lock for write since the mac addresses
1508 * are compared and may be swapped.
1510 write_lock_bh(&bond->lock);
1512 res = alb_handle_addr_collision_on_attach(bond, slave);
1514 write_unlock_bh(&bond->lock);
1520 tlb_init_slave(slave);
1522 /* order a rebalance ASAP */
1523 bond->alb_info.tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1525 if (bond->alb_info.rlb_enabled) {
1526 bond->alb_info.rlb_rebalance = 1;
1532 /* Caller must hold bond lock for write */
1533 void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave)
1535 if (bond->slave_cnt > 1) {
1536 alb_change_hw_addr_on_detach(bond, slave);
1539 tlb_clear_slave(bond, slave, 0);
1541 if (bond->alb_info.rlb_enabled) {
1542 bond->alb_info.next_rx_slave = NULL;
1543 rlb_clear_slave(bond, slave);
1547 /* Caller must hold bond lock for read */
1548 void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char link)
1550 struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
1552 if (link == BOND_LINK_DOWN) {
1553 tlb_clear_slave(bond, slave, 0);
1554 if (bond->alb_info.rlb_enabled) {
1555 rlb_clear_slave(bond, slave);
1557 } else if (link == BOND_LINK_UP) {
1558 /* order a rebalance ASAP */
1559 bond_info->tx_rebalance_counter = BOND_TLB_REBALANCE_TICKS;
1560 if (bond->alb_info.rlb_enabled) {
1561 bond->alb_info.rlb_rebalance = 1;
1562 /* If the updelay module parameter is smaller than the
1563 * forwarding delay of the switch the rebalance will
1564 * not work because the rebalance arp replies will
1565 * not be forwarded to the clients..
1572 * bond_alb_handle_active_change - assign new curr_active_slave
1573 * @bond: our bonding struct
1574 * @new_slave: new slave to assign
1576 * Set the bond->curr_active_slave to @new_slave and handle
1577 * mac address swapping and promiscuity changes as needed.
1579 * Caller must hold bond curr_slave_lock for write (or bond lock for write)
1581 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave)
1583 struct slave *swap_slave;
1586 if (bond->curr_active_slave == new_slave) {
1590 if (bond->curr_active_slave && bond->alb_info.primary_is_promisc) {
1591 dev_set_promiscuity(bond->curr_active_slave->dev, -1);
1592 bond->alb_info.primary_is_promisc = 0;
1593 bond->alb_info.rlb_promisc_timeout_counter = 0;
1596 swap_slave = bond->curr_active_slave;
1597 bond->curr_active_slave = new_slave;
1599 if (!new_slave || (bond->slave_cnt == 0)) {
1603 /* set the new curr_active_slave to the bonds mac address
1604 * i.e. swap mac addresses of old curr_active_slave and new curr_active_slave
1607 struct slave *tmp_slave;
1608 /* find slave that is holding the bond's mac address */
1609 bond_for_each_slave(bond, tmp_slave, i) {
1610 if (!memcmp(tmp_slave->dev->dev_addr,
1611 bond->dev->dev_addr, ETH_ALEN)) {
1612 swap_slave = tmp_slave;
1618 /* curr_active_slave must be set before calling alb_swap_mac_addr */
1620 /* swap mac address */
1621 alb_swap_mac_addr(bond, swap_slave, new_slave);
1623 /* set the new_slave to the bond mac address */
1624 alb_set_slave_mac_addr(new_slave, bond->dev->dev_addr,
1625 bond->alb_info.rlb_enabled);
1626 /* fasten bond mac on new current slave */
1627 alb_send_learning_packets(new_slave, bond->dev->dev_addr);
1631 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr)
1633 struct bonding *bond = bond_dev->priv;
1634 struct sockaddr *sa = addr;
1635 struct slave *slave, *swap_slave;
1639 if (!is_valid_ether_addr(sa->sa_data)) {
1640 return -EADDRNOTAVAIL;
1643 res = alb_set_mac_address(bond, addr);
1648 memcpy(bond_dev->dev_addr, sa->sa_data, bond_dev->addr_len);
1650 /* If there is no curr_active_slave there is nothing else to do.
1651 * Otherwise we'll need to pass the new address to it and handle
1654 if (!bond->curr_active_slave) {
1660 bond_for_each_slave(bond, slave, i) {
1661 if (!memcmp(slave->dev->dev_addr, bond_dev->dev_addr, ETH_ALEN)) {
1668 alb_swap_mac_addr(bond, swap_slave, bond->curr_active_slave);
1670 alb_set_slave_mac_addr(bond->curr_active_slave, bond_dev->dev_addr,
1671 bond->alb_info.rlb_enabled);
1673 alb_send_learning_packets(bond->curr_active_slave, bond_dev->dev_addr);
1674 if (bond->alb_info.rlb_enabled) {
1675 /* inform clients mac address has changed */
1676 rlb_req_update_slave_clients(bond, bond->curr_active_slave);
1683 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id)
1685 if (bond->alb_info.current_alb_vlan &&
1686 (bond->alb_info.current_alb_vlan->vlan_id == vlan_id)) {
1687 bond->alb_info.current_alb_vlan = NULL;
1690 if (bond->alb_info.rlb_enabled) {
1691 rlb_clear_vlan(bond, vlan_id);