2 * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
26 #include "dynamic-string.h"
34 #include "poll-loop.h"
40 VLOG_DEFINE_THIS_MODULE(bond);
42 /* Bit-mask for hashing a flow down to a bucket.
43 * There are (BOND_MASK + 1) buckets. */
44 #define BOND_MASK 0xff
46 /* A hash bucket for mapping a flow to a slave.
47 * "struct bond" has an array of (BOND_MASK + 1) of these. */
49 struct bond_slave *slave; /* Assigned slave, NULL if unassigned. */
50 uint64_t tx_bytes; /* Count of bytes recently transmitted. */
51 tag_type tag; /* Tag for entry<->slave association. */
52 struct list list_node; /* In bond_slave's 'entries' list. */
55 /* A bond slave, that is, one of the links comprising a bond. */
57 struct hmap_node hmap_node; /* In struct bond's slaves hmap. */
58 struct bond *bond; /* The bond that contains this slave. */
59 void *aux; /* Client-provided handle for this slave. */
61 struct netdev *netdev; /* Network device, owned by the client. */
62 unsigned int change_seq; /* Tracks changes in 'netdev'. */
63 char *name; /* Name (a copy of netdev_get_name(netdev)). */
66 long long delay_expires; /* Time after which 'enabled' may change. */
67 bool enabled; /* May be chosen for flows? */
68 bool may_enable; /* Client considers this slave bondable. */
69 tag_type tag; /* Tag associated with this slave. */
71 /* Rebalancing info. Used only by bond_rebalance(). */
72 struct list bal_node; /* In bond_rebalance()'s 'bals' list. */
73 struct list entries; /* 'struct bond_entry's assigned here. */
74 uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
76 /* BM_STABLE specific bonding info. */
77 uint32_t stb_id; /* ID used for 'stb_slaves' ordering. */
80 /* A bond, that is, a set of network devices grouped to improve performance or
83 struct hmap_node hmap_node; /* In 'all_bonds' hmap. */
84 char *name; /* Name provided by client. */
90 enum bond_mode balance; /* Balancing mode, one of BM_*. */
91 struct bond_slave *active_slave;
92 tag_type no_slaves_tag; /* Tag for flows when all slaves disabled. */
93 int updelay, downdelay; /* Delay before slave goes up/down, in ms. */
94 bool lacp_negotiated; /* LACP negotiations were successful. */
95 bool bond_revalidate; /* True if flows need revalidation. */
96 uint32_t basis; /* Basis for flow hash function. */
98 /* SLB specific bonding info. */
99 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
100 int rebalance_interval; /* Interval between rebalances, in ms. */
101 long long int next_rebalance; /* Next rebalancing time. */
102 bool send_learning_packets;
104 /* BM_STABLE specific bonding info. */
105 tag_type stb_tag; /* Tag associated with this bond. */
107 /* Legacy compatibility. */
108 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
110 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
111 * where we can't otherwise provide revalidation feedback to the client.
112 * That's only unixctl commands now; I hope no other cases will arise. */
113 struct tag_set unixctl_tags;
116 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
118 static void bond_entry_reset(struct bond *);
119 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
120 static void bond_enable_slave(struct bond_slave *, bool enable,
122 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
123 static void bond_choose_active_slave(struct bond *, struct tag_set *);
124 static bool bond_is_tcp_hash(const struct bond *);
125 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
126 uint16_t vlan, uint32_t basis);
127 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan,
129 static struct bond_entry *lookup_bond_entry(const struct bond *,
132 static tag_type bond_get_active_slave_tag(const struct bond *);
133 static struct bond_slave *choose_output_slave(const struct bond *,
136 static void bond_update_fake_slave_stats(struct bond *);
138 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
139 * stores the mode in '*balance' and returns true. Otherwise returns false
140 * without modifying '*balance'. */
142 bond_mode_from_string(enum bond_mode *balance, const char *s)
144 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
146 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
148 } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
149 *balance = BM_STABLE;
150 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
158 /* Returns a string representing 'balance'. */
160 bond_mode_to_string(enum bond_mode balance) {
163 return "balance-tcp";
165 return "balance-slb";
169 return "active-backup";
175 /* Creates and returns a new bond whose configuration is initially taken from
178 * The caller should register each slave on the new bond by calling
179 * bond_slave_register(). */
181 bond_create(const struct bond_settings *s)
185 bond = xzalloc(sizeof *bond);
186 hmap_init(&bond->slaves);
187 bond->no_slaves_tag = tag_create_random();
188 bond->stb_tag = tag_create_random();
189 bond->next_fake_iface_update = LLONG_MAX;
191 bond_reconfigure(bond, s);
193 tag_set_init(&bond->unixctl_tags);
200 bond_destroy(struct bond *bond)
202 struct bond_slave *slave, *next_slave;
208 hmap_remove(&all_bonds, &bond->hmap_node);
210 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
211 hmap_remove(&bond->slaves, &slave->hmap_node);
212 /* Client owns 'slave->netdev'. */
216 hmap_destroy(&bond->slaves);
223 /* Updates 'bond''s overall configuration to 's'.
225 * The caller should register each slave on 'bond' by calling
226 * bond_slave_register(). This is optional if none of the slaves'
227 * configuration has changed. In any case it can't hurt.
229 * Returns true if the configuration has changed in such a way that requires
233 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
235 bool revalidate = false;
237 if (!bond->name || strcmp(bond->name, s->name)) {
239 hmap_remove(&all_bonds, &bond->hmap_node);
242 bond->name = xstrdup(s->name);
243 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
246 bond->updelay = s->up_delay;
247 bond->downdelay = s->down_delay;
248 bond->rebalance_interval = s->rebalance_interval;
250 if (bond->balance != s->balance) {
251 bond->balance = s->balance;
255 if (bond->basis != s->basis) {
256 bond->basis = s->basis;
261 if (bond->next_fake_iface_update == LLONG_MAX) {
262 bond->next_fake_iface_update = time_msec();
265 bond->next_fake_iface_update = LLONG_MAX;
268 if (bond->bond_revalidate) {
270 bond->bond_revalidate = false;
273 if (bond->balance == BM_AB || !bond->hash || revalidate) {
274 bond_entry_reset(bond);
281 bond_slave_set_netdev__(struct bond_slave *slave, struct netdev *netdev)
283 if (slave->netdev != netdev) {
284 slave->netdev = netdev;
285 slave->change_seq = 0;
289 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
290 * arbitrary client-provided pointer that uniquely identifies a slave within a
291 * bond. If 'slave_' already exists within 'bond' then this function
292 * reconfigures the existing slave.
294 * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
295 * across restarts and distributed vswitch instances. It should be unique per
296 * slave, and preferably consistent across restarts and reconfigurations.
298 * 'netdev' must be the network device that 'slave_' represents. It is owned
299 * by the client, so the client must not close it before either unregistering
300 * 'slave_' or destroying 'bond'.
303 bond_slave_register(struct bond *bond, void *slave_, uint32_t stb_id,
304 struct netdev *netdev)
306 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
309 slave = xzalloc(sizeof *slave);
311 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
314 slave->delay_expires = LLONG_MAX;
315 slave->name = xstrdup(netdev_get_name(netdev));
316 bond->bond_revalidate = true;
318 slave->enabled = false;
319 bond_enable_slave(slave, netdev_get_carrier(netdev), NULL);
322 if (slave->stb_id != stb_id) {
323 slave->stb_id = stb_id;
324 bond->bond_revalidate = true;
327 bond_slave_set_netdev__(slave, netdev);
330 slave->name = xstrdup(netdev_get_name(netdev));
333 /* Updates the network device to be used with 'slave_' to 'netdev'.
335 * This is useful if the caller closes and re-opens the network device
336 * registered with bond_slave_register() but doesn't need to change anything
339 bond_slave_set_netdev(struct bond *bond, void *slave_, struct netdev *netdev)
341 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
343 bond_slave_set_netdev__(slave, netdev);
347 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
348 * then this function has no effect.
350 * Unregistering a slave invalidates all flows. */
352 bond_slave_unregister(struct bond *bond, const void *slave_)
354 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
361 bond_enable_slave(slave, false, NULL);
363 del_active = bond->active_slave == slave;
365 struct bond_entry *e;
366 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
367 if (e->slave == slave) {
375 hmap_remove(&bond->slaves, &slave->hmap_node);
376 /* Client owns 'slave->netdev'. */
383 bond_choose_active_slave(bond, &tags);
384 bond->send_learning_packets = true;
388 /* Should be called on each slave in 'bond' before bond_run() to indicate
389 * whether or not 'slave_' may be enabled. This function is intended to allow
390 * other protocols to have some impact on bonding decisions. For example LACP
391 * or high level link monitoring protocols may decide that a given slave should
392 * not be able to send traffic. */
394 bond_slave_set_may_enable(struct bond *bond, void *slave_, bool may_enable)
396 bond_slave_lookup(bond, slave_)->may_enable = may_enable;
399 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
400 * allow tagged flows to be invalidated.
402 * The caller should check bond_should_send_learning_packets() afterward. */
404 bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
406 struct bond_slave *slave;
407 bool is_tcp_hash = bond_is_tcp_hash(bond);
409 bond->lacp_negotiated = lacp_negotiated;
411 /* Enable slaves based on link status and LACP feedback. */
412 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
413 bond_link_status_update(slave, tags);
414 slave->change_seq = netdev_change_seq(slave->netdev);
416 if (!bond->active_slave || !bond->active_slave->enabled) {
417 bond_choose_active_slave(bond, tags);
420 /* Update fake bond interface stats. */
421 if (time_msec() >= bond->next_fake_iface_update) {
422 bond_update_fake_slave_stats(bond);
423 bond->next_fake_iface_update = time_msec() + 1000;
426 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
427 bond->bond_revalidate = true;
430 if (bond->bond_revalidate) {
431 bond->bond_revalidate = false;
433 bond_entry_reset(bond);
434 if (bond->balance != BM_STABLE) {
435 struct bond_slave *slave;
437 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
438 tag_set_add(tags, slave->tag);
441 tag_set_add(tags, bond->stb_tag);
443 tag_set_add(tags, bond->no_slaves_tag);
446 /* Invalidate any tags required by */
447 tag_set_union(tags, &bond->unixctl_tags);
448 tag_set_init(&bond->unixctl_tags);
451 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
453 bond_wait(struct bond *bond)
455 struct bond_slave *slave;
457 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
458 if (slave->delay_expires != LLONG_MAX) {
459 poll_timer_wait_until(slave->delay_expires);
462 if (slave->change_seq != netdev_change_seq(slave->netdev)) {
463 poll_immediate_wake();
467 if (bond->next_fake_iface_update != LLONG_MAX) {
468 poll_timer_wait_until(bond->next_fake_iface_update);
471 /* Ensure that any saved tags get revalidated right away. */
472 if (!tag_set_is_empty(&bond->unixctl_tags)) {
473 poll_immediate_wake();
476 /* We don't wait for bond->next_rebalance because rebalancing can only run
477 * at a flow account checkpoint. ofproto does checkpointing on its own
478 * schedule and bond_rebalance() gets called afterward, so we'd just be
479 * waking up for no purpose. */
482 /* MAC learning table interaction. */
485 may_send_learning_packets(const struct bond *bond)
487 return !bond->lacp_negotiated && bond->balance != BM_AB;
490 /* Returns true if 'bond' needs the client to send out packets to assist with
491 * MAC learning on 'bond'. If this function returns true, then the client
492 * should iterate through its MAC learning table for the bridge on which 'bond'
493 * is located. For each MAC that has been learned on a port other than 'bond',
494 * it should call bond_send_learning_packet().
496 * This function will only return true if 'bond' is in SLB mode and LACP is not
497 * negotiated. Otherwise sending learning packets isn't necessary.
499 * Calling this function resets the state that it checks. */
501 bond_should_send_learning_packets(struct bond *bond)
503 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
504 bond->send_learning_packets = false;
508 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
510 * See bond_should_send_learning_packets() for description of usage. */
512 bond_send_learning_packet(struct bond *bond,
513 const uint8_t eth_src[ETH_ADDR_LEN],
516 struct bond_slave *slave;
517 struct ofpbuf packet;
521 assert(may_send_learning_packets(bond));
522 if (!bond->active_slave) {
523 /* Nowhere to send the learning packet. */
527 memset(&flow, 0, sizeof flow);
528 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
529 slave = choose_output_slave(bond, &flow, vlan);
531 ofpbuf_init(&packet, 0);
532 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
535 eth_push_vlan(&packet, htons(vlan));
537 error = netdev_send(slave->netdev, &packet);
538 ofpbuf_uninit(&packet);
543 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
544 * Ethernet destination address of 'eth_dst', should be admitted.
546 * The return value is one of the following:
548 * - BV_ACCEPT: Admit the packet.
550 * - BV_DROP: Drop the packet.
552 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
553 * Ethernet source address and VLAN. If there is none, or if the packet
554 * is on the learned port, then admit the packet. If a different port has
555 * been learned, however, drop the packet (and do not use it for MAC
559 bond_check_admissibility(struct bond *bond, const void *slave_,
560 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
562 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
564 /* LACP bonds have very loose admissibility restrictions because we can
565 * assume the remote switch is aware of the bond and will "do the right
566 * thing". However, as a precaution we drop packets on disabled slaves
567 * because no correctly implemented partner switch should be sending
568 * packets to them. */
569 if (bond->lacp_negotiated) {
570 return slave->enabled ? BV_ACCEPT : BV_DROP;
573 /* Drop all multicast packets on inactive slaves. */
574 if (eth_addr_is_multicast(eth_dst)) {
575 *tags |= bond_get_active_slave_tag(bond);
576 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
581 /* Drop all packets which arrive on backup slaves. This is similar to how
582 * Linux bonding handles active-backup bonds. */
583 if (bond->balance == BM_AB) {
584 *tags |= bond_get_active_slave_tag(bond);
585 if (bond->active_slave != slave) {
586 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
588 VLOG_WARN_RL(&rl, "active-backup bond received packet on backup"
589 " slave (%s) destined for " ETH_ADDR_FMT,
590 slave->name, ETH_ADDR_ARGS(eth_dst));
595 /* Drop all packets for which we have learned a different input port,
596 * because we probably sent the packet on one slave and got it back on the
597 * other. Gratuitous ARP packets are an exception to this rule: the host
598 * has moved to another switch. The exception to the exception is if we
599 * locked the learning table to avoid reflections on bond slaves. */
600 return BV_DROP_IF_MOVED;
603 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
604 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
605 * NULL if the packet should be dropped because no slaves are enabled.
607 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
608 * should be a VID only (i.e. excluding the PCP bits). Second,
609 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
610 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
611 * packet belongs to (so for an access port it will be the access port's VLAN).
613 * Adds a tag to '*tags' that associates the flow with the returned slave.
616 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
617 uint16_t vlan, tag_type *tags)
619 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
621 *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
624 *tags |= bond->no_slaves_tag;
632 bond_is_balanced(const struct bond *bond)
634 return bond->balance == BM_SLB || bond->balance == BM_TCP;
637 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
639 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
642 if (bond_is_balanced(bond)) {
643 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
647 static struct bond_slave *
648 bond_slave_from_bal_node(struct list *bal)
650 return CONTAINER_OF(bal, struct bond_slave, bal_node);
654 log_bals(struct bond *bond, const struct list *bals)
656 if (VLOG_IS_DBG_ENABLED()) {
657 struct ds ds = DS_EMPTY_INITIALIZER;
658 const struct bond_slave *slave;
660 LIST_FOR_EACH (slave, bal_node, bals) {
662 ds_put_char(&ds, ',');
664 ds_put_format(&ds, " %s %"PRIu64"kB",
665 slave->name, slave->tx_bytes / 1024);
667 if (!slave->enabled) {
668 ds_put_cstr(&ds, " (disabled)");
670 if (!list_is_empty(&slave->entries)) {
671 struct bond_entry *e;
673 ds_put_cstr(&ds, " (");
674 LIST_FOR_EACH (e, list_node, &slave->entries) {
675 if (&e->list_node != list_front(&slave->entries)) {
676 ds_put_cstr(&ds, " + ");
678 ds_put_format(&ds, "h%td: %"PRIu64"kB",
679 e - bond->hash, e->tx_bytes / 1024);
681 ds_put_cstr(&ds, ")");
684 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
689 /* Shifts 'hash' from its current slave to 'to'. */
691 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
694 struct bond_slave *from = hash->slave;
695 struct bond *bond = from->bond;
696 uint64_t delta = hash->tx_bytes;
698 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
699 "from %s to %s (now carrying %"PRIu64"kB and "
700 "%"PRIu64"kB load, respectively)",
701 bond->name, delta / 1024, hash - bond->hash,
702 from->name, to->name,
703 (from->tx_bytes - delta) / 1024,
704 (to->tx_bytes + delta) / 1024);
706 /* Shift load away from 'from' to 'to'. */
707 from->tx_bytes -= delta;
708 to->tx_bytes += delta;
710 /* Arrange for flows to be revalidated. */
711 tag_set_add(set, hash->tag);
713 hash->tag = tag_create_random();
716 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
717 * given that doing so must decrease the ratio of the load on the two slaves by
718 * at least 0.1. Returns NULL if there is no appropriate entry.
720 * The list of entries isn't sorted. I don't know of a reason to prefer to
721 * shift away small hashes or large hashes. */
722 static struct bond_entry *
723 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
725 struct bond_entry *e;
727 if (list_is_short(&from->entries)) {
728 /* 'from' carries no more than one MAC hash, so shifting load away from
729 * it would be pointless. */
733 LIST_FOR_EACH (e, list_node, &from->entries) {
734 double old_ratio, new_ratio;
737 if (to_tx_bytes == 0) {
738 /* Nothing on the new slave, move it. */
743 old_ratio = (double)from->tx_bytes / to_tx_bytes;
744 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
745 if (old_ratio - new_ratio > 0.1) {
746 /* Would decrease the ratio, move it. */
754 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
757 insert_bal(struct list *bals, struct bond_slave *slave)
759 struct bond_slave *pos;
761 LIST_FOR_EACH (pos, bal_node, bals) {
762 if (slave->tx_bytes > pos->tx_bytes) {
766 list_insert(&pos->bal_node, &slave->bal_node);
769 /* Removes 'slave' from its current list and then inserts it into 'bals' so
770 * that descending order of 'tx_bytes' is maintained. */
772 reinsert_bal(struct list *bals, struct bond_slave *slave)
774 list_remove(&slave->bal_node);
775 insert_bal(bals, slave);
778 /* If 'bond' needs rebalancing, does so.
780 * The caller should have called bond_account() for each active flow, to ensure
781 * that flow data is consistently accounted at this point. */
783 bond_rebalance(struct bond *bond, struct tag_set *tags)
785 struct bond_slave *slave;
786 struct bond_entry *e;
789 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
792 bond->next_rebalance = time_msec() + bond->rebalance_interval;
794 /* Add each bond_entry to its slave's 'entries' list.
795 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
796 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
798 list_init(&slave->entries);
800 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
801 if (e->slave && e->tx_bytes) {
802 e->slave->tx_bytes += e->tx_bytes;
803 list_push_back(&e->slave->entries, &e->list_node);
807 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
809 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
810 * with a proper list sort algorithm. */
812 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
813 if (slave->enabled) {
814 insert_bal(&bals, slave);
817 log_bals(bond, &bals);
819 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
820 while (!list_is_short(&bals)) {
821 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
822 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
825 overload = from->tx_bytes - to->tx_bytes;
826 if (overload < to->tx_bytes >> 5 || overload < 100000) {
827 /* The extra load on 'from' (and all less-loaded slaves), compared
828 * to that of 'to' (the least-loaded slave), is less than ~3%, or
829 * it is less than ~1Mbps. No point in rebalancing. */
833 /* 'from' is carrying significantly more load than 'to', and that load
834 * is split across at least two different hashes. */
835 e = choose_entry_to_migrate(from, to->tx_bytes);
837 bond_shift_load(e, to, tags);
839 /* Delete element from from->entries.
841 * We don't add the element to to->hashes. That would only allow
842 * 'e' to be migrated to another slave in this rebalancing run, and
843 * there is no point in doing that. */
844 list_remove(&e->list_node);
846 /* Re-sort 'bals'. */
847 reinsert_bal(&bals, from);
848 reinsert_bal(&bals, to);
850 /* Can't usefully migrate anything away from 'from'.
851 * Don't reconsider it. */
852 list_remove(&from->bal_node);
856 /* Implement exponentially weighted moving average. A weight of 1/2 causes
857 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
858 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
859 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
867 /* Bonding unixctl user interface functions. */
870 bond_find(const char *name)
874 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
876 if (!strcmp(bond->name, name)) {
883 static struct bond_slave *
884 bond_lookup_slave(struct bond *bond, const char *slave_name)
886 struct bond_slave *slave;
888 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
889 if (!strcmp(slave->name, slave_name)) {
897 bond_unixctl_list(struct unixctl_conn *conn,
898 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
900 struct ds ds = DS_EMPTY_INITIALIZER;
901 const struct bond *bond;
903 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
905 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
906 const struct bond_slave *slave;
909 ds_put_format(&ds, "%s\t%s\t",
910 bond->name, bond_mode_to_string(bond->balance));
913 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
915 ds_put_cstr(&ds, ", ");
917 ds_put_cstr(&ds, slave->name);
919 ds_put_char(&ds, '\n');
921 unixctl_command_reply(conn, 200, ds_cstr(&ds));
926 bond_unixctl_show(struct unixctl_conn *conn,
927 const char *args, void *aux OVS_UNUSED)
929 struct ds ds = DS_EMPTY_INITIALIZER;
930 const struct bond_slave *slave;
931 const struct bond *bond;
933 bond = bond_find(args);
935 unixctl_command_reply(conn, 501, "no such bond");
939 ds_put_format(&ds, "bond_mode: %s\n",
940 bond_mode_to_string(bond->balance));
942 if (bond->balance != BM_AB) {
943 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
944 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
947 ds_put_format(&ds, "bond-hash-basis: %"PRIu32"\n", bond->basis);
949 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
950 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
952 if (bond_is_balanced(bond)) {
953 ds_put_format(&ds, "next rebalance: %lld ms\n",
954 bond->next_rebalance - time_msec());
957 ds_put_format(&ds, "lacp_negotiated: %s\n",
958 bond->lacp_negotiated ? "true" : "false");
960 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
961 struct bond_entry *be;
964 ds_put_format(&ds, "\nslave %s: %s\n",
965 slave->name, slave->enabled ? "enabled" : "disabled");
966 if (slave == bond->active_slave) {
967 ds_put_cstr(&ds, "\tactive slave\n");
969 if (slave->delay_expires != LLONG_MAX) {
970 ds_put_format(&ds, "\t%s expires in %lld ms\n",
971 slave->enabled ? "downdelay" : "updelay",
972 slave->delay_expires - time_msec());
975 ds_put_format(&ds, "\tmay_enable: %s\n",
976 slave->may_enable ? "true" : "false");
978 if (!bond_is_balanced(bond)) {
983 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
984 int hash = be - bond->hash;
986 if (be->slave != slave) {
990 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
991 hash, be->tx_bytes / 1024);
993 if (bond->balance != BM_SLB) {
997 /* XXX How can we list the MACs assigned to hashes? */
1000 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1005 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1006 void *aux OVS_UNUSED)
1008 char *args = (char *) args_;
1009 char *save_ptr = NULL;
1010 char *bond_s, *hash_s, *slave_s;
1012 struct bond_slave *slave;
1013 struct bond_entry *entry;
1016 bond_s = strtok_r(args, " ", &save_ptr);
1017 hash_s = strtok_r(NULL, " ", &save_ptr);
1018 slave_s = strtok_r(NULL, " ", &save_ptr);
1020 unixctl_command_reply(conn, 501,
1021 "usage: bond/migrate BOND HASH SLAVE");
1025 bond = bond_find(bond_s);
1027 unixctl_command_reply(conn, 501, "no such bond");
1031 if (bond->balance != BM_SLB) {
1032 unixctl_command_reply(conn, 501, "not an SLB bond");
1036 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1037 hash = atoi(hash_s) & BOND_MASK;
1039 unixctl_command_reply(conn, 501, "bad hash");
1043 slave = bond_lookup_slave(bond, slave_s);
1045 unixctl_command_reply(conn, 501, "no such slave");
1049 if (!slave->enabled) {
1050 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1054 entry = &bond->hash[hash];
1055 tag_set_add(&bond->unixctl_tags, entry->tag);
1056 entry->slave = slave;
1057 entry->tag = tag_create_random();
1058 unixctl_command_reply(conn, 200, "migrated");
1062 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1063 void *aux OVS_UNUSED)
1065 char *args = (char *) args_;
1066 char *save_ptr = NULL;
1067 char *bond_s, *slave_s;
1069 struct bond_slave *slave;
1071 bond_s = strtok_r(args, " ", &save_ptr);
1072 slave_s = strtok_r(NULL, " ", &save_ptr);
1074 unixctl_command_reply(conn, 501,
1075 "usage: bond/set-active-slave BOND SLAVE");
1079 bond = bond_find(bond_s);
1081 unixctl_command_reply(conn, 501, "no such bond");
1085 slave = bond_lookup_slave(bond, slave_s);
1087 unixctl_command_reply(conn, 501, "no such slave");
1091 if (!slave->enabled) {
1092 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1096 if (bond->active_slave != slave) {
1097 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1098 bond->active_slave = slave;
1099 bond->active_slave->tag = tag_create_random();
1100 VLOG_INFO("bond %s: active interface is now %s",
1101 bond->name, slave->name);
1102 bond->send_learning_packets = true;
1103 unixctl_command_reply(conn, 200, "done");
1105 unixctl_command_reply(conn, 200, "no change");
1110 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1112 char *args = (char *) args_;
1113 char *save_ptr = NULL;
1114 char *bond_s, *slave_s;
1116 struct bond_slave *slave;
1118 bond_s = strtok_r(args, " ", &save_ptr);
1119 slave_s = strtok_r(NULL, " ", &save_ptr);
1121 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1122 enable ? "enable" : "disable");
1123 unixctl_command_reply(conn, 501, usage);
1128 bond = bond_find(bond_s);
1130 unixctl_command_reply(conn, 501, "no such bond");
1134 slave = bond_lookup_slave(bond, slave_s);
1136 unixctl_command_reply(conn, 501, "no such slave");
1140 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1141 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1145 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1146 void *aux OVS_UNUSED)
1148 enable_slave(conn, args, true);
1152 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1153 void *aux OVS_UNUSED)
1155 enable_slave(conn, args, false);
1159 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1160 void *aux OVS_UNUSED)
1162 char *args = (char *) args_;
1163 uint8_t mac[ETH_ADDR_LEN];
1168 char *mac_s, *vlan_s, *basis_s;
1169 char *save_ptr = NULL;
1171 mac_s = strtok_r(args, " ", &save_ptr);
1172 vlan_s = strtok_r(NULL, " ", &save_ptr);
1173 basis_s = strtok_r(NULL, " ", &save_ptr);
1176 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1177 unixctl_command_reply(conn, 501, "invalid vlan");
1181 vlan = OFP_VLAN_NONE;
1185 if (sscanf(basis_s, "%"PRIu32, &basis) != 1) {
1186 unixctl_command_reply(conn, 501, "invalid basis");
1193 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1194 == ETH_ADDR_SCAN_COUNT) {
1195 hash = bond_hash_src(mac, vlan, basis) & BOND_MASK;
1197 hash_cstr = xasprintf("%u", hash);
1198 unixctl_command_reply(conn, 200, hash_cstr);
1201 unixctl_command_reply(conn, 501, "invalid mac");
1208 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1209 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1210 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1211 unixctl_command_register("bond/set-active-slave",
1212 bond_unixctl_set_active_slave, NULL);
1213 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1215 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1217 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1221 bond_entry_reset(struct bond *bond)
1223 if (bond->balance != BM_AB) {
1224 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1227 bond->hash = xmalloc(hash_len);
1229 memset(bond->hash, 0, hash_len);
1231 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1238 static struct bond_slave *
1239 bond_slave_lookup(struct bond *bond, const void *slave_)
1241 struct bond_slave *slave;
1243 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1245 if (slave->aux == slave_) {
1254 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1256 struct bond *bond = slave->bond;
1257 slave->delay_expires = LLONG_MAX;
1258 if (enable != slave->enabled) {
1259 slave->enabled = enable;
1260 if (!slave->enabled) {
1261 VLOG_WARN("interface %s: disabled", slave->name);
1263 tag_set_add(tags, slave->tag);
1266 VLOG_WARN("interface %s: enabled", slave->name);
1267 slave->tag = tag_create_random();
1270 if (bond->balance == BM_STABLE) {
1271 bond->bond_revalidate = true;
1277 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1279 struct bond *bond = slave->bond;
1282 up = netdev_get_carrier(slave->netdev) && slave->may_enable;
1283 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1284 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1285 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1286 slave->name, up ? "up" : "down");
1287 if (up == slave->enabled) {
1288 slave->delay_expires = LLONG_MAX;
1289 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1290 slave->name, up ? "disabled" : "enabled");
1292 int delay = (bond->lacp_negotiated ? 0
1293 : up ? bond->updelay : bond->downdelay);
1294 slave->delay_expires = time_msec() + delay;
1296 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1299 up ? "enabled" : "disabled",
1306 if (time_msec() >= slave->delay_expires) {
1307 bond_enable_slave(slave, up, tags);
1312 bond_is_tcp_hash(const struct bond *bond)
1314 return (bond->balance == BM_TCP && bond->lacp_negotiated)
1315 || bond->balance == BM_STABLE;
1319 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan, uint32_t basis)
1321 return hash_3words(hash_bytes(mac, ETH_ADDR_LEN, 0), vlan, basis);
1325 bond_hash_tcp(const struct flow *flow, uint16_t vlan, uint32_t basis)
1327 struct flow hash_flow = *flow;
1328 hash_flow.vlan_tci = htons(vlan);
1330 /* The symmetric quality of this hash function is not required, but
1331 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1332 * purposes, so we use it out of convenience. */
1333 return flow_hash_symmetric_l4(&hash_flow, basis);
1337 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1339 assert(bond->balance != BM_AB);
1341 return (bond_is_tcp_hash(bond)
1342 ? bond_hash_tcp(flow, vlan, bond->basis)
1343 : bond_hash_src(flow->dl_src, vlan, bond->basis));
1346 static struct bond_entry *
1347 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1350 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1353 /* This function uses Highest Random Weight hashing to choose an output slave.
1354 * This approach only reassigns a minimal number of flows when slaves are
1355 * enabled or disabled. Unfortunately, it has O(n) performance against the
1356 * number of slaves. There exist algorithms which are O(1), but have slightly
1357 * more complex implementations and require the use of memory. This may need
1358 * to be reimplemented if it becomes a performance bottleneck. */
1359 static struct bond_slave *
1360 choose_stb_slave(const struct bond *bond, const struct flow *flow,
1363 struct bond_slave *best, *slave;
1364 uint32_t best_hash, flow_hash;
1368 flow_hash = bond_hash(bond, flow, vlan);
1369 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1370 if (slave->enabled) {
1373 hash = hash_2words(flow_hash, slave->stb_id);
1374 if (!best || hash > best_hash) {
1384 static struct bond_slave *
1385 choose_output_slave(const struct bond *bond, const struct flow *flow,
1388 struct bond_entry *e;
1390 switch (bond->balance) {
1392 return bond->active_slave;
1395 return choose_stb_slave(bond, flow, vlan);
1398 e = lookup_bond_entry(bond, flow, vlan);
1399 if (!e->slave || !e->slave->enabled) {
1400 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1401 struct bond_slave, hmap_node);
1402 if (!e->slave->enabled) {
1403 e->slave = bond->active_slave;
1405 e->tag = tag_create_random();
1414 static struct bond_slave *
1415 bond_choose_slave(const struct bond *bond)
1417 struct bond_slave *slave, *best;
1419 /* Find an enabled slave. */
1420 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1421 if (slave->enabled) {
1426 /* All interfaces are disabled. Find an interface that will be enabled
1427 * after its updelay expires. */
1429 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1430 if (slave->delay_expires != LLONG_MAX
1431 && slave->may_enable
1432 && (!best || slave->delay_expires < best->delay_expires)) {
1440 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1442 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1443 struct bond_slave *old_active_slave = bond->active_slave;
1445 bond->active_slave = bond_choose_slave(bond);
1446 if (bond->active_slave) {
1447 if (bond->active_slave->enabled) {
1448 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1449 bond->name, bond->active_slave->name);
1451 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1452 "remaining %lld ms updelay (since no interface was "
1453 "enabled)", bond->name, bond->active_slave->name,
1454 bond->active_slave->delay_expires - time_msec());
1455 bond_enable_slave(bond->active_slave, true, tags);
1458 if (!old_active_slave) {
1459 tag_set_add(tags, bond->no_slaves_tag);
1462 bond->send_learning_packets = true;
1463 } else if (old_active_slave) {
1464 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1468 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1469 * there is no active slave. */
1471 bond_get_active_slave_tag(const struct bond *bond)
1473 return (bond->active_slave
1474 ? bond->active_slave->tag
1475 : bond->no_slaves_tag);
1478 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1479 * bond interface. */
1481 bond_update_fake_slave_stats(struct bond *bond)
1483 struct netdev_stats bond_stats;
1484 struct bond_slave *slave;
1485 struct netdev *bond_dev;
1487 memset(&bond_stats, 0, sizeof bond_stats);
1489 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1490 struct netdev_stats slave_stats;
1492 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1493 /* XXX: We swap the stats here because they are swapped back when
1494 * reported by the internal device. The reason for this is
1495 * internal devices normally represent packets going into the
1496 * system but when used as fake bond device they represent packets
1497 * leaving the system. We really should do this in the internal
1498 * device itself because changing it here reverses the counts from
1499 * the perspective of the switch. However, the internal device
1500 * doesn't know what type of device it represents so we have to do
1501 * it here for now. */
1502 bond_stats.tx_packets += slave_stats.rx_packets;
1503 bond_stats.tx_bytes += slave_stats.rx_bytes;
1504 bond_stats.rx_packets += slave_stats.tx_packets;
1505 bond_stats.rx_bytes += slave_stats.tx_bytes;
1509 if (!netdev_open(bond->name, "system", &bond_dev)) {
1510 netdev_set_stats(bond_dev, &bond_stats);
1511 netdev_close(bond_dev);