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"
35 #include "poll-loop.h"
41 VLOG_DEFINE_THIS_MODULE(bond);
43 COVERAGE_DEFINE(bond_process_lacp);
45 /* Bit-mask for hashing a flow down to a bucket.
46 * There are (BOND_MASK + 1) buckets. */
47 #define BOND_MASK 0xff
49 /* A hash bucket for mapping a flow to a slave.
50 * "struct bond" has an array of (BOND_MASK + 1) of these. */
52 struct bond_slave *slave; /* Assigned slave, NULL if unassigned. */
53 uint64_t tx_bytes; /* Count of bytes recently transmitted. */
54 tag_type tag; /* Tag for entry<->slave association. */
55 struct list list_node; /* In bond_slave's 'entries' list. */
58 /* A bond slave, that is, one of the links comprising a bond. */
60 struct hmap_node hmap_node; /* In struct bond's slaves hmap. */
61 struct bond *bond; /* The bond that contains this slave. */
62 void *aux; /* Client-provided handle for this slave. */
64 struct netdev *netdev; /* Network device, owned by the client. */
65 char *name; /* Name (a copy of netdev_get_name(netdev)). */
68 long long delay_expires; /* Time after which 'enabled' may change. */
69 bool up; /* Last link status read from netdev. */
70 bool enabled; /* May be chosen for flows? */
71 tag_type tag; /* Tag associated with this slave. */
73 /* Rebalancing info. Used only by bond_rebalance(). */
74 struct list bal_node; /* In bond_rebalance()'s 'bals' list. */
75 struct list entries; /* 'struct bond_entry's assigned here. */
76 uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
79 /* A bond, that is, a set of network devices grouped to improve performance or
82 struct hmap_node hmap_node; /* In 'all_bonds' hmap. */
83 char *name; /* Name provided by client. */
89 enum bond_mode balance; /* Balancing mode, one of BM_*. */
90 struct bond_slave *active_slave;
91 tag_type no_slaves_tag; /* Tag for flows when all slaves disabled. */
92 int updelay, downdelay; /* Delay before slave goes up/down, in ms. */
94 /* SLB specific bonding info. */
95 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
96 int rebalance_interval; /* Interval between rebalances, in ms. */
97 long long int next_rebalance; /* Next rebalancing time. */
98 bool send_learning_packets;
101 struct lacp *lacp; /* LACP object. NULL if LACP is disabled. */
104 enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
105 struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
106 long long int miimon_interval; /* Miimon status refresh interval. */
107 long long int miimon_next_update; /* Time of next miimon update. */
109 /* Legacy compatibility. */
110 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
112 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
113 * where we can't otherwise provide revalidation feedback to the client.
114 * That's only unixctl commands now; I hope no other cases will arise. */
115 struct tag_set unixctl_tags;
118 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
120 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
121 static bool bond_is_link_up(struct bond *, struct netdev *);
122 static void bond_enable_slave(struct bond_slave *, bool enable,
124 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
125 static void bond_choose_active_slave(struct bond *, struct tag_set *);
126 static bool bond_is_tcp_hash(const struct bond *);
127 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
129 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan);
130 static struct bond_entry *lookup_bond_entry(const struct bond *,
133 static tag_type bond_get_active_slave_tag(const struct bond *);
134 static struct bond_slave *choose_output_slave(const struct bond *,
137 static void bond_update_fake_slave_stats(struct bond *);
139 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
140 * stores the mode in '*balance' and returns true. Otherwise returns false
141 * without modifying '*balance'. */
143 bond_mode_from_string(enum bond_mode *balance, const char *s)
145 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
147 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
149 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
157 /* Returns a string representing 'balance'. */
159 bond_mode_to_string(enum bond_mode balance) {
162 return "balance-tcp";
164 return "balance-slb";
166 return "active-backup";
171 /* Attempts to parse 's' as the name of a bond link status detection mode. If
172 * successful, stores the mode in '*detect' and returns true. Otherwise
173 * returns false without modifying '*detect'. */
175 bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
177 if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
178 *detect = BLSM_CARRIER;
179 } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
180 *detect = BLSM_MIIMON;
187 /* Returns a string representing 'detect'. */
189 bond_detect_mode_to_string(enum bond_detect_mode detect)
200 /* Creates and returns a new bond whose configuration is initially taken from
203 * The caller should register each slave on the new bond by calling
204 * bond_slave_register(). */
206 bond_create(const struct bond_settings *s)
210 bond = xzalloc(sizeof *bond);
211 hmap_init(&bond->slaves);
212 bond->no_slaves_tag = tag_create_random();
213 bond->miimon_next_update = LLONG_MAX;
214 bond->next_fake_iface_update = LLONG_MAX;
216 bond_reconfigure(bond, s);
218 tag_set_init(&bond->unixctl_tags);
225 bond_destroy(struct bond *bond)
227 struct bond_slave *slave, *next_slave;
233 hmap_remove(&all_bonds, &bond->hmap_node);
235 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
236 hmap_remove(&bond->slaves, &slave->hmap_node);
237 /* Client owns 'slave->netdev'. */
241 hmap_destroy(&bond->slaves);
245 lacp_destroy(bond->lacp);
247 netdev_monitor_destroy(bond->monitor);
253 /* Updates 'bond''s overall configuration to 's'.
255 * The caller should register each slave on 'bond' by calling
256 * bond_slave_register(). This is optional if none of the slaves'
257 * configuration has changed, except that it is mandatory if 's' enables LACP
258 * and 'bond' previously didn't have LACP enabled. In any case it can't
261 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
263 if (!bond->name || strcmp(bond->name, s->name)) {
265 hmap_remove(&all_bonds, &bond->hmap_node);
268 bond->name = xstrdup(s->name);
269 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
272 bond->balance = s->balance;
273 bond->detect = s->detect;
274 bond->miimon_interval = s->miimon_interval;
275 bond->updelay = s->up_delay;
276 bond->downdelay = s->down_delay;
277 bond->rebalance_interval = s->rebalance_interval;
279 if (bond->balance != BM_AB) {
281 bond->hash = xcalloc(BOND_MASK + 1, sizeof *bond->hash);
282 bond->next_rebalance = time_msec() + bond->rebalance_interval;
291 if (bond->detect == BLSM_CARRIER) {
292 struct bond_slave *slave;
294 if (!bond->monitor) {
295 bond->monitor = netdev_monitor_create();
298 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
299 netdev_monitor_add(bond->monitor, slave->netdev);
302 netdev_monitor_destroy(bond->monitor);
303 bond->monitor = NULL;
305 if (bond->miimon_next_update == LLONG_MAX) {
306 bond->miimon_next_update = time_msec() + bond->miimon_interval;
312 bond->lacp = lacp_create();
314 lacp_configure(bond->lacp, s->lacp);
316 lacp_destroy(bond->lacp);
321 if (bond->next_fake_iface_update == LLONG_MAX) {
322 bond->next_fake_iface_update = time_msec();
325 bond->next_fake_iface_update = LLONG_MAX;
329 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
330 * arbitrary client-provided pointer that uniquely identifies a slave within a
331 * bond. If 'slave_' already exists within 'bond' then this function
332 * reconfigures the existing slave.
334 * 'netdev' must be the network device that 'slave_' represents. It is owned
335 * by the client, so the client must not close it before either unregistering
336 * 'slave_' or destroying 'bond'.
338 * If 'bond' has a LACP configuration then 'lacp_settings' must point to LACP
339 * settings for 'slave_'; otherwise 'lacp_settings' is ignored.
342 bond_slave_register(struct bond *bond, void *slave_, struct netdev *netdev,
343 const struct lacp_slave_settings *lacp_settings)
345 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
348 slave = xzalloc(sizeof *slave);
350 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
353 slave->delay_expires = LLONG_MAX;
354 slave->up = bond_is_link_up(bond, netdev);
355 slave->enabled = slave->up;
358 slave->netdev = netdev;
360 slave->name = xstrdup(netdev_get_name(netdev));
363 assert(lacp_settings != NULL);
364 lacp_slave_register(bond->lacp, slave, lacp_settings);
368 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
369 * then this function has no effect.
371 * Unregistering a slave invalidates all flows. */
373 bond_slave_unregister(struct bond *bond, const void *slave_)
375 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
382 del_active = bond->active_slave == slave;
384 struct bond_entry *e;
385 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
386 if (e->slave == slave) {
394 hmap_remove(&bond->slaves, &slave->hmap_node);
395 /* Client owns 'slave->netdev'. */
402 bond_choose_active_slave(bond, &tags);
403 bond->send_learning_packets = true;
407 /* Callback for lacp_run(). */
409 bond_send_pdu_cb(void *slave_, const struct lacp_pdu *pdu)
411 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 10);
412 struct bond_slave *slave = slave_;
413 uint8_t ea[ETH_ADDR_LEN];
416 error = netdev_get_etheraddr(slave->netdev, ea);
418 struct lacp_pdu *packet_pdu;
419 struct ofpbuf packet;
421 ofpbuf_init(&packet, 0);
422 packet_pdu = eth_compose(&packet, eth_addr_lacp, ea, ETH_TYPE_LACP,
425 error = netdev_send(slave->netdev, &packet);
427 VLOG_WARN_RL(&rl, "bond %s: sending LACP PDU on slave %s failed "
429 slave->bond->name, slave->name, strerror(error));
431 ofpbuf_uninit(&packet);
433 VLOG_ERR_RL(&rl, "bond %s: cannot obtain Ethernet address of slave "
435 slave->bond->name, slave->name, strerror(error));
439 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
440 * allow tagged flows to be invalidated.
442 * The caller should check bond_should_send_learning_packets() afterward. */
444 bond_run(struct bond *bond, struct tag_set *tags)
446 struct bond_slave *slave;
448 /* Update link status. */
449 if (bond->detect == BLSM_CARRIER
450 || time_msec() >= bond->miimon_next_update)
452 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
453 slave->up = bond_is_link_up(bond, slave->netdev);
455 bond->miimon_next_update = time_msec() + bond->miimon_interval;
460 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
461 lacp_slave_enable(bond->lacp, slave, slave->enabled);
464 lacp_run(bond->lacp, bond_send_pdu_cb);
467 /* Enable slaves based on link status and LACP feedback. */
468 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
469 bond_link_status_update(slave, tags);
471 if (!bond->active_slave || !bond->active_slave->enabled) {
472 bond_choose_active_slave(bond, tags);
475 /* Update fake bond interface stats. */
476 if (time_msec() >= bond->next_fake_iface_update) {
477 bond_update_fake_slave_stats(bond);
478 bond->next_fake_iface_update = time_msec() + 1000;
481 /* Invalidate any tags required by */
482 tag_set_union(tags, &bond->unixctl_tags);
483 tag_set_init(&bond->unixctl_tags);
486 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
488 bond_wait(struct bond *bond)
490 struct bond_slave *slave;
492 if (bond->detect == BLSM_CARRIER) {
493 netdev_monitor_poll_wait(bond->monitor);
495 poll_timer_wait_until(bond->miimon_next_update);
498 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
499 if (slave->delay_expires != LLONG_MAX) {
500 poll_timer_wait_until(slave->delay_expires);
504 if (bond->next_fake_iface_update != LLONG_MAX) {
505 poll_timer_wait_until(bond->next_fake_iface_update);
508 /* Ensure that any saved tags get revalidated right away. */
509 if (!tag_set_is_empty(&bond->unixctl_tags)) {
510 poll_immediate_wake();
513 /* We don't wait for bond->next_rebalance because rebalancing can only run
514 * at a flow account checkpoint. ofproto does checkpointing on its own
515 * schedule and bond_rebalance() gets called afterward, so we'd just be
516 * waking up for no purpose. */
519 /* MAC learning table interaction. */
522 may_send_learning_packets(const struct bond *bond)
524 return !lacp_negotiated(bond->lacp) && bond->balance != BM_AB;
527 /* Returns true if 'bond' needs the client to send out packets to assist with
528 * MAC learning on 'bond'. If this function returns true, then the client
529 * should iterate through its MAC learning table for the bridge on which 'bond'
530 * is located. For each MAC that has been learned on a port other than 'bond',
531 * it should call bond_send_learning_packet().
533 * This function will only return true if 'bond' is in SLB mode and LACP is not
534 * negotiated. Otherwise sending learning packets isn't necessary.
536 * Calling this function resets the state that it checks. */
538 bond_should_send_learning_packets(struct bond *bond)
540 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
541 bond->send_learning_packets = false;
545 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
547 * See bond_should_send_learning_packets() for description of usage. */
549 bond_send_learning_packet(struct bond *bond,
550 const uint8_t eth_src[ETH_ADDR_LEN],
553 struct bond_slave *slave;
554 struct ofpbuf packet;
558 assert(may_send_learning_packets(bond));
559 if (!bond->active_slave) {
560 /* Nowhere to send the learning packet. */
564 memset(&flow, 0, sizeof flow);
565 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
566 slave = choose_output_slave(bond, &flow, vlan);
568 ofpbuf_init(&packet, 0);
569 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
572 eth_set_vlan_tci(&packet, htons(vlan));
574 error = netdev_send(slave->netdev, &packet);
575 ofpbuf_uninit(&packet);
580 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
581 * Ethernet destination address of 'eth_dst', should be admitted.
583 * The return value is one of the following:
585 * - BV_ACCEPT: Admit the packet.
587 * - BV_DROP: Drop the packet.
589 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
590 * Ethernet source address and VLAN. If there is none, or if the packet
591 * is on the learned port, then admit the packet. If a different port has
592 * been learned, however, drop the packet (and do not use it for MAC
596 bond_check_admissibility(struct bond *bond, const void *slave_,
597 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
599 /* Admit all packets if LACP has been negotiated, because that means that
600 * the remote switch is aware of the bond and will "do the right thing". */
601 if (lacp_negotiated(bond->lacp)) {
605 /* Drop all multicast packets on inactive slaves. */
606 if (eth_addr_is_multicast(eth_dst)) {
607 *tags |= bond_get_active_slave_tag(bond);
608 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
613 /* Drop all packets for which we have learned a different input port,
614 * because we probably sent the packet on one slave and got it back on the
615 * other. Gratuitous ARP packets are an exception to this rule: the host
616 * has moved to another switch. The exception to the exception is if we
617 * locked the learning table to avoid reflections on bond slaves. */
618 return BV_DROP_IF_MOVED;
621 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
622 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
623 * NULL if the packet should be dropped because no slaves are enabled.
625 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
626 * should be a VID only (i.e. excluding the PCP bits). Second,
627 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
628 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
629 * packet belongs to (so for an access port it will be the access port's VLAN).
631 * Adds a tag to '*tags' that associates the flow with the returned slave.
634 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
635 uint16_t vlan, tag_type *tags)
637 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
642 *tags |= bond->no_slaves_tag;
647 /* Processes LACP packet 'packet', which was received on 'slave_' within
650 * The client should use this function to pass along LACP messages received on
651 * any of 'bond''s slaves. */
653 bond_process_lacp(struct bond *bond, void *slave_, const struct ofpbuf *packet)
656 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
657 const struct lacp_pdu *pdu = parse_lacp_packet(packet);
659 COVERAGE_INC(bond_process_lacp);
660 lacp_process_pdu(bond->lacp, slave, pdu);
667 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
669 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
672 switch (bond->balance) {
679 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
687 static struct bond_slave *
688 bond_slave_from_bal_node(struct list *bal)
690 return CONTAINER_OF(bal, struct bond_slave, bal_node);
694 log_bals(struct bond *bond, const struct list *bals)
696 if (VLOG_IS_DBG_ENABLED()) {
697 struct ds ds = DS_EMPTY_INITIALIZER;
698 const struct bond_slave *slave;
700 LIST_FOR_EACH (slave, bal_node, bals) {
702 ds_put_char(&ds, ',');
704 ds_put_format(&ds, " %s %"PRIu64"kB",
705 slave->name, slave->tx_bytes / 1024);
707 if (!slave->enabled) {
708 ds_put_cstr(&ds, " (disabled)");
710 if (!list_is_empty(&slave->entries)) {
711 struct bond_entry *e;
713 ds_put_cstr(&ds, " (");
714 LIST_FOR_EACH (e, list_node, &slave->entries) {
715 if (&e->list_node != list_front(&slave->entries)) {
716 ds_put_cstr(&ds, " + ");
718 ds_put_format(&ds, "h%td: %"PRIu64"kB",
719 e - bond->hash, e->tx_bytes / 1024);
721 ds_put_cstr(&ds, ")");
724 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
729 /* Shifts 'hash' from its current slave to 'to'. */
731 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
734 struct bond_slave *from = hash->slave;
735 struct bond *bond = from->bond;
736 uint64_t delta = hash->tx_bytes;
738 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
739 "from %s to %s (now carrying %"PRIu64"kB and "
740 "%"PRIu64"kB load, respectively)",
741 bond->name, delta / 1024, hash - bond->hash,
742 from->name, to->name,
743 (from->tx_bytes - delta) / 1024,
744 (to->tx_bytes + delta) / 1024);
746 /* Shift load away from 'from' to 'to'. */
747 from->tx_bytes -= delta;
748 to->tx_bytes += delta;
750 /* Arrange for flows to be revalidated. */
751 tag_set_add(set, hash->tag);
753 hash->tag = tag_create_random();
756 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
757 * given that doing so must decrease the ratio of the load on the two slaves by
758 * at least 0.1. Returns NULL if there is no appropriate entry.
760 * The list of entries isn't sorted. I don't know of a reason to prefer to
761 * shift away small hashes or large hashes. */
762 static struct bond_entry *
763 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
765 struct bond_entry *e;
767 if (list_is_short(&from->entries)) {
768 /* 'from' carries no more than one MAC hash, so shifting load away from
769 * it would be pointless. */
773 LIST_FOR_EACH (e, list_node, &from->entries) {
774 double old_ratio, new_ratio;
777 if (to_tx_bytes == 0) {
778 /* Nothing on the new slave, move it. */
783 old_ratio = (double)from->tx_bytes / to_tx_bytes;
784 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
785 if (old_ratio - new_ratio > 0.1) {
786 /* Would decrease the ratio, move it. */
794 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
797 insert_bal(struct list *bals, struct bond_slave *slave)
799 struct bond_slave *pos;
801 LIST_FOR_EACH (pos, bal_node, bals) {
802 if (slave->tx_bytes > pos->tx_bytes) {
806 list_insert(&pos->bal_node, &slave->bal_node);
809 /* Removes 'slave' from its current list and then inserts it into 'bals' so
810 * that descending order of 'tx_bytes' is maintained. */
812 reinsert_bal(struct list *bals, struct bond_slave *slave)
814 list_remove(&slave->bal_node);
815 insert_bal(bals, slave);
818 /* If 'bond' needs rebalancing, does so.
820 * The caller should have called bond_account() for each active flow, to ensure
821 * that flow data is consistently accounted at this point. */
823 bond_rebalance(struct bond *bond, struct tag_set *tags)
825 struct bond_slave *slave;
826 struct bond_entry *e;
829 if (bond->balance == BM_AB || time_msec() < bond->next_rebalance) {
832 bond->next_rebalance = time_msec() + bond->rebalance_interval;
834 /* Add each bond_entry to its slave's 'entries' list.
835 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
836 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
838 list_init(&slave->entries);
840 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
841 if (e->slave && e->tx_bytes) {
842 e->slave->tx_bytes += e->tx_bytes;
843 list_push_back(&e->slave->entries, &e->list_node);
847 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
849 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
850 * with a proper list sort algorithm. */
852 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
853 if (slave->enabled) {
854 insert_bal(&bals, slave);
857 log_bals(bond, &bals);
859 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
860 while (!list_is_short(&bals)) {
861 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
862 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
865 overload = from->tx_bytes - to->tx_bytes;
866 if (overload < to->tx_bytes >> 5 || overload < 100000) {
867 /* The extra load on 'from' (and all less-loaded slaves), compared
868 * to that of 'to' (the least-loaded slave), is less than ~3%, or
869 * it is less than ~1Mbps. No point in rebalancing. */
873 /* 'from' is carrying significantly more load than 'to', and that load
874 * is split across at least two different hashes. */
875 e = choose_entry_to_migrate(from, to->tx_bytes);
877 bond_shift_load(e, to, tags);
879 /* Delete element from from->entries.
881 * We don't add the element to to->hashes. That would only allow
882 * 'e' to be migrated to another slave in this rebalancing run, and
883 * there is no point in doing that. */
884 list_remove(&e->list_node);
886 /* Re-sort 'bals'. */
887 reinsert_bal(&bals, from);
888 reinsert_bal(&bals, to);
890 /* Can't usefully migrate anything away from 'from'.
891 * Don't reconsider it. */
892 list_remove(&from->bal_node);
896 /* Implement exponentially weighted moving average. A weight of 1/2 causes
897 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
898 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
899 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
907 /* Bonding unixctl user interface functions. */
910 bond_find(const char *name)
914 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
916 if (!strcmp(bond->name, name)) {
923 static struct bond_slave *
924 bond_lookup_slave(struct bond *bond, const char *slave_name)
926 struct bond_slave *slave;
928 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
929 if (!strcmp(slave->name, slave_name)) {
937 bond_unixctl_list(struct unixctl_conn *conn,
938 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
940 struct ds ds = DS_EMPTY_INITIALIZER;
941 const struct bond *bond;
943 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
945 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
946 const struct bond_slave *slave;
949 ds_put_format(&ds, "%s\t%s\t",
950 bond->name, bond_mode_to_string(bond->balance));
953 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
955 ds_put_cstr(&ds, ", ");
957 ds_put_cstr(&ds, slave->name);
959 ds_put_char(&ds, '\n');
961 unixctl_command_reply(conn, 200, ds_cstr(&ds));
966 bond_unixctl_show(struct unixctl_conn *conn,
967 const char *args, void *aux OVS_UNUSED)
969 struct ds ds = DS_EMPTY_INITIALIZER;
970 const struct bond_slave *slave;
971 const struct bond *bond;
973 bond = bond_find(args);
975 unixctl_command_reply(conn, 501, "no such bond");
979 ds_put_format(&ds, "bond_mode: %s\n",
980 bond_mode_to_string(bond->balance));
983 ds_put_format(&ds, "lacp: %s\n",
984 lacp_is_active(bond->lacp) ? "active" : "passive");
986 ds_put_cstr(&ds, "lacp: off\n");
989 if (bond->balance != BM_AB) {
990 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
991 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
994 ds_put_format(&ds, "bond-detect-mode: %s\n",
995 bond->monitor ? "carrier" : "miimon");
997 if (!bond->monitor) {
998 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
999 bond->miimon_interval);
1002 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
1003 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
1005 if (bond->balance != BM_AB) {
1006 ds_put_format(&ds, "next rebalance: %lld ms\n",
1007 bond->next_rebalance - time_msec());
1010 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1011 struct bond_entry *be;
1015 ds_put_format(&ds, "\nslave %s: %s\n",
1016 slave->name, slave->enabled ? "enabled" : "disabled");
1017 if (slave == bond->active_slave) {
1018 ds_put_cstr(&ds, "\tactive slave\n");
1020 if (slave->delay_expires != LLONG_MAX) {
1021 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1022 slave->enabled ? "downdelay" : "updelay",
1023 slave->delay_expires - time_msec());
1026 if (bond->balance == BM_AB) {
1031 memset(&flow, 0, sizeof flow);
1032 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1033 int hash = be - bond->hash;
1035 if (be->slave != slave) {
1039 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1040 hash, be->tx_bytes / 1024);
1042 if (bond->balance != BM_SLB) {
1046 /* XXX How can we list the MACs assigned to hashes? */
1049 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1054 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1055 void *aux OVS_UNUSED)
1057 char *args = (char *) args_;
1058 char *save_ptr = NULL;
1059 char *bond_s, *hash_s, *slave_s;
1061 struct bond_slave *slave;
1062 struct bond_entry *entry;
1065 bond_s = strtok_r(args, " ", &save_ptr);
1066 hash_s = strtok_r(NULL, " ", &save_ptr);
1067 slave_s = strtok_r(NULL, " ", &save_ptr);
1069 unixctl_command_reply(conn, 501,
1070 "usage: bond/migrate BOND HASH SLAVE");
1074 bond = bond_find(bond_s);
1076 unixctl_command_reply(conn, 501, "no such bond");
1080 if (bond->balance != BM_SLB) {
1081 unixctl_command_reply(conn, 501, "not an SLB bond");
1085 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1086 hash = atoi(hash_s) & BOND_MASK;
1088 unixctl_command_reply(conn, 501, "bad hash");
1092 slave = bond_lookup_slave(bond, slave_s);
1094 unixctl_command_reply(conn, 501, "no such slave");
1098 if (!slave->enabled) {
1099 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1103 entry = &bond->hash[hash];
1104 tag_set_add(&bond->unixctl_tags, entry->tag);
1105 entry->slave = slave;
1106 entry->tag = tag_create_random();
1107 unixctl_command_reply(conn, 200, "migrated");
1111 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1112 void *aux OVS_UNUSED)
1114 char *args = (char *) args_;
1115 char *save_ptr = NULL;
1116 char *bond_s, *slave_s;
1118 struct bond_slave *slave;
1120 bond_s = strtok_r(args, " ", &save_ptr);
1121 slave_s = strtok_r(NULL, " ", &save_ptr);
1123 unixctl_command_reply(conn, 501,
1124 "usage: bond/set-active-slave BOND SLAVE");
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 if (!slave->enabled) {
1141 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1145 if (bond->active_slave != slave) {
1146 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1147 bond->active_slave = slave;
1148 bond->active_slave->tag = tag_create_random();
1149 VLOG_INFO("bond %s: active interface is now %s",
1150 bond->name, slave->name);
1151 bond->send_learning_packets = true;
1152 unixctl_command_reply(conn, 200, "done");
1154 unixctl_command_reply(conn, 200, "no change");
1159 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1161 char *args = (char *) args_;
1162 char *save_ptr = NULL;
1163 char *bond_s, *slave_s;
1165 struct bond_slave *slave;
1167 bond_s = strtok_r(args, " ", &save_ptr);
1168 slave_s = strtok_r(NULL, " ", &save_ptr);
1170 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1171 enable ? "enable" : "disable");
1172 unixctl_command_reply(conn, 501, usage);
1177 bond = bond_find(bond_s);
1179 unixctl_command_reply(conn, 501, "no such bond");
1183 slave = bond_lookup_slave(bond, slave_s);
1185 unixctl_command_reply(conn, 501, "no such slave");
1189 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1190 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1194 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1195 void *aux OVS_UNUSED)
1197 enable_slave(conn, args, true);
1201 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1202 void *aux OVS_UNUSED)
1204 enable_slave(conn, args, false);
1208 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1209 void *aux OVS_UNUSED)
1211 char *args = (char *) args_;
1212 uint8_t mac[ETH_ADDR_LEN];
1216 char *mac_s, *vlan_s;
1217 char *save_ptr = NULL;
1219 mac_s = strtok_r(args, " ", &save_ptr);
1220 vlan_s = strtok_r(NULL, " ", &save_ptr);
1223 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1224 unixctl_command_reply(conn, 501, "invalid vlan");
1228 vlan = OFP_VLAN_NONE;
1231 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1232 == ETH_ADDR_SCAN_COUNT) {
1233 hash = bond_hash_src(mac, vlan) & BOND_MASK;
1235 hash_cstr = xasprintf("%u", hash);
1236 unixctl_command_reply(conn, 200, hash_cstr);
1239 unixctl_command_reply(conn, 501, "invalid mac");
1248 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1249 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1250 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1251 unixctl_command_register("bond/set-active-slave",
1252 bond_unixctl_set_active_slave, NULL);
1253 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1255 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1257 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1260 static struct bond_slave *
1261 bond_slave_lookup(struct bond *bond, const void *slave_)
1263 struct bond_slave *slave;
1265 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1267 if (slave->aux == slave_) {
1276 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1278 return (bond->detect == BLSM_CARRIER
1279 ? netdev_get_carrier(netdev)
1280 : netdev_get_miimon(netdev));
1284 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1286 slave->delay_expires = LLONG_MAX;
1287 if (enable != slave->enabled) {
1288 slave->enabled = enable;
1289 if (!slave->enabled) {
1290 VLOG_WARN("interface %s: disabled", slave->name);
1291 tag_set_add(tags, slave->tag);
1293 VLOG_WARN("interface %s: enabled", slave->name);
1294 slave->tag = tag_create_random();
1300 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1302 struct bond *bond = slave->bond;
1305 up = slave->up && lacp_slave_may_enable(bond->lacp, slave);
1306 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1307 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1308 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1309 slave->name, up ? "up" : "down");
1310 if (up == slave->enabled) {
1311 slave->delay_expires = LLONG_MAX;
1312 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1313 slave->name, up ? "disabled" : "enabled");
1315 int delay = (lacp_negotiated(bond->lacp) ? 0
1316 : up ? bond->updelay : bond->downdelay);
1317 slave->delay_expires = time_msec() + delay;
1319 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1322 up ? "enabled" : "disabled",
1329 if (time_msec() >= slave->delay_expires) {
1330 bond_enable_slave(slave, up, tags);
1335 bond_is_tcp_hash(const struct bond *bond)
1337 return bond->balance == BM_TCP && lacp_negotiated(bond->lacp);
1341 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan)
1343 return hash_bytes(mac, ETH_ADDR_LEN, vlan);
1347 bond_hash_tcp(const struct flow *flow, uint16_t vlan)
1349 struct flow hash_flow = *flow;
1350 hash_flow.vlan_tci = vlan;
1352 /* The symmetric quality of this hash function is not required, but
1353 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1354 * purposes, so we use it out of convenience. */
1355 return flow_hash_symmetric_l4(&hash_flow, 0);
1358 static struct bond_entry *
1359 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1362 assert(bond->balance != BM_AB);
1363 return &bond->hash[(bond_is_tcp_hash(bond)
1364 ? bond_hash_tcp(flow, vlan)
1365 : bond_hash_src(flow->dl_src, vlan)) & BOND_MASK];
1368 static struct bond_slave *
1369 choose_output_slave(const struct bond *bond, const struct flow *flow,
1372 struct bond_entry *e;
1374 switch (bond->balance) {
1376 return bond->active_slave;
1380 e = lookup_bond_entry(bond, flow, vlan);
1381 if (!e->slave || !e->slave->enabled) {
1382 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1383 struct bond_slave, hmap_node);
1384 if (!e->slave->enabled) {
1385 e->slave = bond->active_slave;
1387 e->tag = tag_create_random();
1396 static struct bond_slave *
1397 bond_choose_slave(const struct bond *bond)
1399 struct bond_slave *slave, *best;
1401 /* Find an enabled slave. */
1402 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1403 if (slave->enabled) {
1408 /* All interfaces are disabled. Find an interface that will be enabled
1409 * after its updelay expires. */
1411 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1412 if (slave->delay_expires != LLONG_MAX
1413 && lacp_slave_may_enable(bond->lacp, slave)
1414 && (!best || slave->delay_expires < best->delay_expires)) {
1422 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1424 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1425 struct bond_slave *old_active_slave = bond->active_slave;
1427 bond->active_slave = bond_choose_slave(bond);
1428 if (bond->active_slave) {
1429 if (bond->active_slave->enabled) {
1430 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1431 bond->name, bond->active_slave->name);
1433 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1434 "remaining %lld ms updelay (since no interface was "
1435 "enabled)", bond->name, bond->active_slave->name,
1436 bond->active_slave->delay_expires - time_msec());
1437 bond_enable_slave(bond->active_slave, true, tags);
1440 if (!old_active_slave) {
1441 tag_set_add(tags, bond->no_slaves_tag);
1444 bond->send_learning_packets = true;
1445 } else if (old_active_slave) {
1446 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1450 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1451 * there is no active slave. */
1453 bond_get_active_slave_tag(const struct bond *bond)
1455 return (bond->active_slave
1456 ? bond->active_slave->tag
1457 : bond->no_slaves_tag);
1460 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1461 * bond interface. */
1463 bond_update_fake_slave_stats(struct bond *bond)
1465 struct netdev_stats bond_stats;
1466 struct bond_slave *slave;
1467 struct netdev *bond_dev;
1469 memset(&bond_stats, 0, sizeof bond_stats);
1471 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1472 struct netdev_stats slave_stats;
1474 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1475 /* XXX: We swap the stats here because they are swapped back when
1476 * reported by the internal device. The reason for this is
1477 * internal devices normally represent packets going into the
1478 * system but when used as fake bond device they represent packets
1479 * leaving the system. We really should do this in the internal
1480 * device itself because changing it here reverses the counts from
1481 * the perspective of the switch. However, the internal device
1482 * doesn't know what type of device it represents so we have to do
1483 * it here for now. */
1484 bond_stats.tx_packets += slave_stats.rx_packets;
1485 bond_stats.tx_bytes += slave_stats.rx_bytes;
1486 bond_stats.rx_packets += slave_stats.tx_packets;
1487 bond_stats.rx_bytes += slave_stats.tx_bytes;
1491 if (!netdev_open_default(bond->name, &bond_dev)) {
1492 netdev_set_stats(bond_dev, &bond_stats);
1493 netdev_close(bond_dev);