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 void bond_entry_reset(struct bond *);
121 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
122 static bool bond_is_link_up(struct bond *, struct netdev *);
123 static void bond_enable_slave(struct bond_slave *, bool enable,
125 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
126 static void bond_choose_active_slave(struct bond *, struct tag_set *);
127 static bool bond_is_tcp_hash(const struct bond *);
128 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
130 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan);
131 static struct bond_entry *lookup_bond_entry(const struct bond *,
134 static tag_type bond_get_active_slave_tag(const struct bond *);
135 static struct bond_slave *choose_output_slave(const struct bond *,
138 static void bond_update_fake_slave_stats(struct bond *);
140 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
141 * stores the mode in '*balance' and returns true. Otherwise returns false
142 * without modifying '*balance'. */
144 bond_mode_from_string(enum bond_mode *balance, const char *s)
146 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
148 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
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";
167 return "active-backup";
172 /* Attempts to parse 's' as the name of a bond link status detection mode. If
173 * successful, stores the mode in '*detect' and returns true. Otherwise
174 * returns false without modifying '*detect'. */
176 bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
178 if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
179 *detect = BLSM_CARRIER;
180 } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
181 *detect = BLSM_MIIMON;
188 /* Returns a string representing 'detect'. */
190 bond_detect_mode_to_string(enum bond_detect_mode detect)
201 /* Creates and returns a new bond whose configuration is initially taken from
204 * The caller should register each slave on the new bond by calling
205 * bond_slave_register(). */
207 bond_create(const struct bond_settings *s)
211 bond = xzalloc(sizeof *bond);
212 hmap_init(&bond->slaves);
213 bond->no_slaves_tag = tag_create_random();
214 bond->miimon_next_update = LLONG_MAX;
215 bond->next_fake_iface_update = LLONG_MAX;
217 bond_reconfigure(bond, s);
219 tag_set_init(&bond->unixctl_tags);
226 bond_destroy(struct bond *bond)
228 struct bond_slave *slave, *next_slave;
234 hmap_remove(&all_bonds, &bond->hmap_node);
236 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
237 hmap_remove(&bond->slaves, &slave->hmap_node);
238 /* Client owns 'slave->netdev'. */
242 hmap_destroy(&bond->slaves);
246 lacp_destroy(bond->lacp);
248 netdev_monitor_destroy(bond->monitor);
254 /* Updates 'bond''s overall configuration to 's'.
256 * The caller should register each slave on 'bond' by calling
257 * bond_slave_register(). This is optional if none of the slaves'
258 * configuration has changed, except that it is mandatory if 's' enables LACP
259 * and 'bond' previously didn't have LACP enabled. In any case it can't
262 * Returns true if the configuration has changed in such a way that requires
266 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
268 bool revalidate = false;
270 if (!bond->name || strcmp(bond->name, s->name)) {
272 hmap_remove(&all_bonds, &bond->hmap_node);
275 bond->name = xstrdup(s->name);
276 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
279 bond->detect = s->detect;
280 bond->miimon_interval = s->miimon_interval;
281 bond->updelay = s->up_delay;
282 bond->downdelay = s->down_delay;
283 bond->rebalance_interval = s->rebalance_interval;
285 if (bond->balance != s->balance) {
286 bond->balance = s->balance;
290 if (bond->detect == BLSM_CARRIER) {
291 struct bond_slave *slave;
293 if (!bond->monitor) {
294 bond->monitor = netdev_monitor_create();
297 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
298 netdev_monitor_add(bond->monitor, slave->netdev);
301 netdev_monitor_destroy(bond->monitor);
302 bond->monitor = NULL;
304 if (bond->miimon_next_update == LLONG_MAX) {
305 bond->miimon_next_update = time_msec() + bond->miimon_interval;
311 bond->lacp = lacp_create();
313 lacp_configure(bond->lacp, s->lacp);
315 lacp_destroy(bond->lacp);
320 if (bond->next_fake_iface_update == LLONG_MAX) {
321 bond->next_fake_iface_update = time_msec();
324 bond->next_fake_iface_update = LLONG_MAX;
327 if (bond->balance == BM_AB || !bond->hash || revalidate) {
328 bond_entry_reset(bond);
334 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
335 * arbitrary client-provided pointer that uniquely identifies a slave within a
336 * bond. If 'slave_' already exists within 'bond' then this function
337 * reconfigures the existing slave.
339 * 'netdev' must be the network device that 'slave_' represents. It is owned
340 * by the client, so the client must not close it before either unregistering
341 * 'slave_' or destroying 'bond'.
343 * If 'bond' has a LACP configuration then 'lacp_settings' must point to LACP
344 * settings for 'slave_'; otherwise 'lacp_settings' is ignored.
347 bond_slave_register(struct bond *bond, void *slave_, struct netdev *netdev,
348 const struct lacp_slave_settings *lacp_settings)
350 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
353 slave = xzalloc(sizeof *slave);
355 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
358 slave->delay_expires = LLONG_MAX;
359 slave->up = bond_is_link_up(bond, netdev);
360 slave->enabled = false;
361 bond_enable_slave(slave, slave->up, NULL);
364 slave->netdev = netdev;
366 slave->name = xstrdup(netdev_get_name(netdev));
369 assert(lacp_settings != NULL);
370 lacp_slave_register(bond->lacp, slave, lacp_settings);
374 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
375 * then this function has no effect.
377 * Unregistering a slave invalidates all flows. */
379 bond_slave_unregister(struct bond *bond, const void *slave_)
381 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
388 bond_enable_slave(slave, false, NULL);
390 del_active = bond->active_slave == slave;
392 struct bond_entry *e;
393 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
394 if (e->slave == slave) {
402 hmap_remove(&bond->slaves, &slave->hmap_node);
403 /* Client owns 'slave->netdev'. */
410 bond_choose_active_slave(bond, &tags);
411 bond->send_learning_packets = true;
415 /* Callback for lacp_run(). */
417 bond_send_pdu_cb(void *slave_, const struct lacp_pdu *pdu)
419 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 10);
420 struct bond_slave *slave = slave_;
421 uint8_t ea[ETH_ADDR_LEN];
424 error = netdev_get_etheraddr(slave->netdev, ea);
426 struct lacp_pdu *packet_pdu;
427 struct ofpbuf packet;
429 ofpbuf_init(&packet, 0);
430 packet_pdu = eth_compose(&packet, eth_addr_lacp, ea, ETH_TYPE_LACP,
433 error = netdev_send(slave->netdev, &packet);
435 VLOG_WARN_RL(&rl, "bond %s: sending LACP PDU on slave %s failed "
437 slave->bond->name, slave->name, strerror(error));
439 ofpbuf_uninit(&packet);
441 VLOG_ERR_RL(&rl, "bond %s: cannot obtain Ethernet address of slave "
443 slave->bond->name, slave->name, strerror(error));
447 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
448 * allow tagged flows to be invalidated.
450 * The caller should check bond_should_send_learning_packets() afterward. */
452 bond_run(struct bond *bond, struct tag_set *tags)
454 struct bond_slave *slave;
455 bool is_tcp_hash = bond_is_tcp_hash(bond);
457 /* Update link status. */
458 if (bond->detect == BLSM_CARRIER
459 || time_msec() >= bond->miimon_next_update)
461 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
462 slave->up = bond_is_link_up(bond, slave->netdev);
464 bond->miimon_next_update = time_msec() + bond->miimon_interval;
469 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
470 lacp_slave_enable(bond->lacp, slave, slave->enabled);
473 lacp_run(bond->lacp, bond_send_pdu_cb);
476 /* Enable slaves based on link status and LACP feedback. */
477 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
478 bond_link_status_update(slave, tags);
480 if (!bond->active_slave || !bond->active_slave->enabled) {
481 bond_choose_active_slave(bond, tags);
484 /* Update fake bond interface stats. */
485 if (time_msec() >= bond->next_fake_iface_update) {
486 bond_update_fake_slave_stats(bond);
487 bond->next_fake_iface_update = time_msec() + 1000;
490 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
491 struct bond_slave *slave;
493 bond_entry_reset(bond);
494 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
495 tag_set_add(tags, slave->tag);
499 /* Invalidate any tags required by */
500 tag_set_union(tags, &bond->unixctl_tags);
501 tag_set_init(&bond->unixctl_tags);
504 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
506 bond_wait(struct bond *bond)
508 struct bond_slave *slave;
510 if (bond->detect == BLSM_CARRIER) {
511 netdev_monitor_poll_wait(bond->monitor);
513 poll_timer_wait_until(bond->miimon_next_update);
516 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
517 if (slave->delay_expires != LLONG_MAX) {
518 poll_timer_wait_until(slave->delay_expires);
522 if (bond->next_fake_iface_update != LLONG_MAX) {
523 poll_timer_wait_until(bond->next_fake_iface_update);
526 /* Ensure that any saved tags get revalidated right away. */
527 if (!tag_set_is_empty(&bond->unixctl_tags)) {
528 poll_immediate_wake();
531 /* We don't wait for bond->next_rebalance because rebalancing can only run
532 * at a flow account checkpoint. ofproto does checkpointing on its own
533 * schedule and bond_rebalance() gets called afterward, so we'd just be
534 * waking up for no purpose. */
537 /* MAC learning table interaction. */
540 may_send_learning_packets(const struct bond *bond)
542 return !lacp_negotiated(bond->lacp) && bond->balance != BM_AB;
545 /* Returns true if 'bond' needs the client to send out packets to assist with
546 * MAC learning on 'bond'. If this function returns true, then the client
547 * should iterate through its MAC learning table for the bridge on which 'bond'
548 * is located. For each MAC that has been learned on a port other than 'bond',
549 * it should call bond_send_learning_packet().
551 * This function will only return true if 'bond' is in SLB mode and LACP is not
552 * negotiated. Otherwise sending learning packets isn't necessary.
554 * Calling this function resets the state that it checks. */
556 bond_should_send_learning_packets(struct bond *bond)
558 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
559 bond->send_learning_packets = false;
563 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
565 * See bond_should_send_learning_packets() for description of usage. */
567 bond_send_learning_packet(struct bond *bond,
568 const uint8_t eth_src[ETH_ADDR_LEN],
571 struct bond_slave *slave;
572 struct ofpbuf packet;
576 assert(may_send_learning_packets(bond));
577 if (!bond->active_slave) {
578 /* Nowhere to send the learning packet. */
582 memset(&flow, 0, sizeof flow);
583 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
584 slave = choose_output_slave(bond, &flow, vlan);
586 ofpbuf_init(&packet, 0);
587 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
590 eth_set_vlan_tci(&packet, htons(vlan));
592 error = netdev_send(slave->netdev, &packet);
593 ofpbuf_uninit(&packet);
598 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
599 * Ethernet destination address of 'eth_dst', should be admitted.
601 * The return value is one of the following:
603 * - BV_ACCEPT: Admit the packet.
605 * - BV_DROP: Drop the packet.
607 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
608 * Ethernet source address and VLAN. If there is none, or if the packet
609 * is on the learned port, then admit the packet. If a different port has
610 * been learned, however, drop the packet (and do not use it for MAC
614 bond_check_admissibility(struct bond *bond, const void *slave_,
615 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
617 /* Admit all packets if LACP has been negotiated, because that means that
618 * the remote switch is aware of the bond and will "do the right thing". */
619 if (lacp_negotiated(bond->lacp)) {
623 /* Drop all multicast packets on inactive slaves. */
624 if (eth_addr_is_multicast(eth_dst)) {
625 *tags |= bond_get_active_slave_tag(bond);
626 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
631 /* Drop all packets for which we have learned a different input port,
632 * because we probably sent the packet on one slave and got it back on the
633 * other. Gratuitous ARP packets are an exception to this rule: the host
634 * has moved to another switch. The exception to the exception is if we
635 * locked the learning table to avoid reflections on bond slaves. */
636 return BV_DROP_IF_MOVED;
639 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
640 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
641 * NULL if the packet should be dropped because no slaves are enabled.
643 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
644 * should be a VID only (i.e. excluding the PCP bits). Second,
645 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
646 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
647 * packet belongs to (so for an access port it will be the access port's VLAN).
649 * Adds a tag to '*tags' that associates the flow with the returned slave.
652 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
653 uint16_t vlan, tag_type *tags)
655 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
660 *tags |= bond->no_slaves_tag;
665 /* Processes LACP packet 'packet', which was received on 'slave_' within
668 * The client should use this function to pass along LACP messages received on
669 * any of 'bond''s slaves. */
671 bond_process_lacp(struct bond *bond, void *slave_, const struct ofpbuf *packet)
674 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
675 const struct lacp_pdu *pdu = parse_lacp_packet(packet);
677 COVERAGE_INC(bond_process_lacp);
678 lacp_process_pdu(bond->lacp, slave, pdu);
686 bond_is_balanced(const struct bond *bond)
688 return bond->balance == BM_SLB || bond->balance == BM_TCP;
691 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
693 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
697 if (bond_is_balanced(bond)) {
698 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
702 static struct bond_slave *
703 bond_slave_from_bal_node(struct list *bal)
705 return CONTAINER_OF(bal, struct bond_slave, bal_node);
709 log_bals(struct bond *bond, const struct list *bals)
711 if (VLOG_IS_DBG_ENABLED()) {
712 struct ds ds = DS_EMPTY_INITIALIZER;
713 const struct bond_slave *slave;
715 LIST_FOR_EACH (slave, bal_node, bals) {
717 ds_put_char(&ds, ',');
719 ds_put_format(&ds, " %s %"PRIu64"kB",
720 slave->name, slave->tx_bytes / 1024);
722 if (!slave->enabled) {
723 ds_put_cstr(&ds, " (disabled)");
725 if (!list_is_empty(&slave->entries)) {
726 struct bond_entry *e;
728 ds_put_cstr(&ds, " (");
729 LIST_FOR_EACH (e, list_node, &slave->entries) {
730 if (&e->list_node != list_front(&slave->entries)) {
731 ds_put_cstr(&ds, " + ");
733 ds_put_format(&ds, "h%td: %"PRIu64"kB",
734 e - bond->hash, e->tx_bytes / 1024);
736 ds_put_cstr(&ds, ")");
739 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
744 /* Shifts 'hash' from its current slave to 'to'. */
746 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
749 struct bond_slave *from = hash->slave;
750 struct bond *bond = from->bond;
751 uint64_t delta = hash->tx_bytes;
753 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
754 "from %s to %s (now carrying %"PRIu64"kB and "
755 "%"PRIu64"kB load, respectively)",
756 bond->name, delta / 1024, hash - bond->hash,
757 from->name, to->name,
758 (from->tx_bytes - delta) / 1024,
759 (to->tx_bytes + delta) / 1024);
761 /* Shift load away from 'from' to 'to'. */
762 from->tx_bytes -= delta;
763 to->tx_bytes += delta;
765 /* Arrange for flows to be revalidated. */
766 tag_set_add(set, hash->tag);
768 hash->tag = tag_create_random();
771 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
772 * given that doing so must decrease the ratio of the load on the two slaves by
773 * at least 0.1. Returns NULL if there is no appropriate entry.
775 * The list of entries isn't sorted. I don't know of a reason to prefer to
776 * shift away small hashes or large hashes. */
777 static struct bond_entry *
778 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
780 struct bond_entry *e;
782 if (list_is_short(&from->entries)) {
783 /* 'from' carries no more than one MAC hash, so shifting load away from
784 * it would be pointless. */
788 LIST_FOR_EACH (e, list_node, &from->entries) {
789 double old_ratio, new_ratio;
792 if (to_tx_bytes == 0) {
793 /* Nothing on the new slave, move it. */
798 old_ratio = (double)from->tx_bytes / to_tx_bytes;
799 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
800 if (old_ratio - new_ratio > 0.1) {
801 /* Would decrease the ratio, move it. */
809 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
812 insert_bal(struct list *bals, struct bond_slave *slave)
814 struct bond_slave *pos;
816 LIST_FOR_EACH (pos, bal_node, bals) {
817 if (slave->tx_bytes > pos->tx_bytes) {
821 list_insert(&pos->bal_node, &slave->bal_node);
824 /* Removes 'slave' from its current list and then inserts it into 'bals' so
825 * that descending order of 'tx_bytes' is maintained. */
827 reinsert_bal(struct list *bals, struct bond_slave *slave)
829 list_remove(&slave->bal_node);
830 insert_bal(bals, slave);
833 /* If 'bond' needs rebalancing, does so.
835 * The caller should have called bond_account() for each active flow, to ensure
836 * that flow data is consistently accounted at this point. */
838 bond_rebalance(struct bond *bond, struct tag_set *tags)
840 struct bond_slave *slave;
841 struct bond_entry *e;
844 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
847 bond->next_rebalance = time_msec() + bond->rebalance_interval;
849 /* Add each bond_entry to its slave's 'entries' list.
850 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
851 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
853 list_init(&slave->entries);
855 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
856 if (e->slave && e->tx_bytes) {
857 e->slave->tx_bytes += e->tx_bytes;
858 list_push_back(&e->slave->entries, &e->list_node);
862 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
864 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
865 * with a proper list sort algorithm. */
867 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
868 if (slave->enabled) {
869 insert_bal(&bals, slave);
872 log_bals(bond, &bals);
874 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
875 while (!list_is_short(&bals)) {
876 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
877 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
880 overload = from->tx_bytes - to->tx_bytes;
881 if (overload < to->tx_bytes >> 5 || overload < 100000) {
882 /* The extra load on 'from' (and all less-loaded slaves), compared
883 * to that of 'to' (the least-loaded slave), is less than ~3%, or
884 * it is less than ~1Mbps. No point in rebalancing. */
888 /* 'from' is carrying significantly more load than 'to', and that load
889 * is split across at least two different hashes. */
890 e = choose_entry_to_migrate(from, to->tx_bytes);
892 bond_shift_load(e, to, tags);
894 /* Delete element from from->entries.
896 * We don't add the element to to->hashes. That would only allow
897 * 'e' to be migrated to another slave in this rebalancing run, and
898 * there is no point in doing that. */
899 list_remove(&e->list_node);
901 /* Re-sort 'bals'. */
902 reinsert_bal(&bals, from);
903 reinsert_bal(&bals, to);
905 /* Can't usefully migrate anything away from 'from'.
906 * Don't reconsider it. */
907 list_remove(&from->bal_node);
911 /* Implement exponentially weighted moving average. A weight of 1/2 causes
912 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
913 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
914 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
922 /* Bonding unixctl user interface functions. */
925 bond_find(const char *name)
929 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
931 if (!strcmp(bond->name, name)) {
938 static struct bond_slave *
939 bond_lookup_slave(struct bond *bond, const char *slave_name)
941 struct bond_slave *slave;
943 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
944 if (!strcmp(slave->name, slave_name)) {
952 bond_unixctl_list(struct unixctl_conn *conn,
953 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
955 struct ds ds = DS_EMPTY_INITIALIZER;
956 const struct bond *bond;
958 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
960 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
961 const struct bond_slave *slave;
964 ds_put_format(&ds, "%s\t%s\t",
965 bond->name, bond_mode_to_string(bond->balance));
968 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
970 ds_put_cstr(&ds, ", ");
972 ds_put_cstr(&ds, slave->name);
974 ds_put_char(&ds, '\n');
976 unixctl_command_reply(conn, 200, ds_cstr(&ds));
981 bond_unixctl_show(struct unixctl_conn *conn,
982 const char *args, void *aux OVS_UNUSED)
984 struct ds ds = DS_EMPTY_INITIALIZER;
985 const struct bond_slave *slave;
986 const struct bond *bond;
988 bond = bond_find(args);
990 unixctl_command_reply(conn, 501, "no such bond");
994 ds_put_format(&ds, "bond_mode: %s\n",
995 bond_mode_to_string(bond->balance));
998 ds_put_format(&ds, "lacp: %s\n",
999 lacp_is_active(bond->lacp) ? "active" : "passive");
1001 ds_put_cstr(&ds, "lacp: off\n");
1004 if (bond->balance != BM_AB) {
1005 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
1006 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
1009 ds_put_format(&ds, "bond-detect-mode: %s\n",
1010 bond->monitor ? "carrier" : "miimon");
1012 if (!bond->monitor) {
1013 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
1014 bond->miimon_interval);
1017 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
1018 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
1020 if (bond_is_balanced(bond)) {
1021 ds_put_format(&ds, "next rebalance: %lld ms\n",
1022 bond->next_rebalance - time_msec());
1025 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1026 struct bond_entry *be;
1030 ds_put_format(&ds, "\nslave %s: %s\n",
1031 slave->name, slave->enabled ? "enabled" : "disabled");
1032 if (slave == bond->active_slave) {
1033 ds_put_cstr(&ds, "\tactive slave\n");
1035 if (slave->delay_expires != LLONG_MAX) {
1036 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1037 slave->enabled ? "downdelay" : "updelay",
1038 slave->delay_expires - time_msec());
1041 if (!bond_is_balanced(bond)) {
1046 memset(&flow, 0, sizeof flow);
1047 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1048 int hash = be - bond->hash;
1050 if (be->slave != slave) {
1054 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1055 hash, be->tx_bytes / 1024);
1057 if (bond->balance != BM_SLB) {
1061 /* XXX How can we list the MACs assigned to hashes? */
1064 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1069 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1070 void *aux OVS_UNUSED)
1072 char *args = (char *) args_;
1073 char *save_ptr = NULL;
1074 char *bond_s, *hash_s, *slave_s;
1076 struct bond_slave *slave;
1077 struct bond_entry *entry;
1080 bond_s = strtok_r(args, " ", &save_ptr);
1081 hash_s = strtok_r(NULL, " ", &save_ptr);
1082 slave_s = strtok_r(NULL, " ", &save_ptr);
1084 unixctl_command_reply(conn, 501,
1085 "usage: bond/migrate BOND HASH SLAVE");
1089 bond = bond_find(bond_s);
1091 unixctl_command_reply(conn, 501, "no such bond");
1095 if (bond->balance != BM_SLB) {
1096 unixctl_command_reply(conn, 501, "not an SLB bond");
1100 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1101 hash = atoi(hash_s) & BOND_MASK;
1103 unixctl_command_reply(conn, 501, "bad hash");
1107 slave = bond_lookup_slave(bond, slave_s);
1109 unixctl_command_reply(conn, 501, "no such slave");
1113 if (!slave->enabled) {
1114 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1118 entry = &bond->hash[hash];
1119 tag_set_add(&bond->unixctl_tags, entry->tag);
1120 entry->slave = slave;
1121 entry->tag = tag_create_random();
1122 unixctl_command_reply(conn, 200, "migrated");
1126 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1127 void *aux OVS_UNUSED)
1129 char *args = (char *) args_;
1130 char *save_ptr = NULL;
1131 char *bond_s, *slave_s;
1133 struct bond_slave *slave;
1135 bond_s = strtok_r(args, " ", &save_ptr);
1136 slave_s = strtok_r(NULL, " ", &save_ptr);
1138 unixctl_command_reply(conn, 501,
1139 "usage: bond/set-active-slave BOND SLAVE");
1143 bond = bond_find(bond_s);
1145 unixctl_command_reply(conn, 501, "no such bond");
1149 slave = bond_lookup_slave(bond, slave_s);
1151 unixctl_command_reply(conn, 501, "no such slave");
1155 if (!slave->enabled) {
1156 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1160 if (bond->active_slave != slave) {
1161 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1162 bond->active_slave = slave;
1163 bond->active_slave->tag = tag_create_random();
1164 VLOG_INFO("bond %s: active interface is now %s",
1165 bond->name, slave->name);
1166 bond->send_learning_packets = true;
1167 unixctl_command_reply(conn, 200, "done");
1169 unixctl_command_reply(conn, 200, "no change");
1174 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1176 char *args = (char *) args_;
1177 char *save_ptr = NULL;
1178 char *bond_s, *slave_s;
1180 struct bond_slave *slave;
1182 bond_s = strtok_r(args, " ", &save_ptr);
1183 slave_s = strtok_r(NULL, " ", &save_ptr);
1185 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1186 enable ? "enable" : "disable");
1187 unixctl_command_reply(conn, 501, usage);
1192 bond = bond_find(bond_s);
1194 unixctl_command_reply(conn, 501, "no such bond");
1198 slave = bond_lookup_slave(bond, slave_s);
1200 unixctl_command_reply(conn, 501, "no such slave");
1204 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1205 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1209 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1210 void *aux OVS_UNUSED)
1212 enable_slave(conn, args, true);
1216 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1217 void *aux OVS_UNUSED)
1219 enable_slave(conn, args, false);
1223 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1224 void *aux OVS_UNUSED)
1226 char *args = (char *) args_;
1227 uint8_t mac[ETH_ADDR_LEN];
1231 char *mac_s, *vlan_s;
1232 char *save_ptr = NULL;
1234 mac_s = strtok_r(args, " ", &save_ptr);
1235 vlan_s = strtok_r(NULL, " ", &save_ptr);
1238 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1239 unixctl_command_reply(conn, 501, "invalid vlan");
1243 vlan = OFP_VLAN_NONE;
1246 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1247 == ETH_ADDR_SCAN_COUNT) {
1248 hash = bond_hash_src(mac, vlan) & BOND_MASK;
1250 hash_cstr = xasprintf("%u", hash);
1251 unixctl_command_reply(conn, 200, hash_cstr);
1254 unixctl_command_reply(conn, 501, "invalid mac");
1263 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1264 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1265 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1266 unixctl_command_register("bond/set-active-slave",
1267 bond_unixctl_set_active_slave, NULL);
1268 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1270 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1272 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1276 bond_entry_reset(struct bond *bond)
1278 if (bond->balance != BM_AB) {
1279 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1282 bond->hash = xmalloc(hash_len);
1284 memset(bond->hash, 0, hash_len);
1286 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1293 static struct bond_slave *
1294 bond_slave_lookup(struct bond *bond, const void *slave_)
1296 struct bond_slave *slave;
1298 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1300 if (slave->aux == slave_) {
1309 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1311 return (bond->detect == BLSM_CARRIER
1312 ? netdev_get_carrier(netdev)
1313 : netdev_get_miimon(netdev));
1317 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1319 slave->delay_expires = LLONG_MAX;
1320 if (enable != slave->enabled) {
1321 slave->enabled = enable;
1322 if (!slave->enabled) {
1323 VLOG_WARN("interface %s: disabled", slave->name);
1325 tag_set_add(tags, slave->tag);
1328 VLOG_WARN("interface %s: enabled", slave->name);
1329 slave->tag = tag_create_random();
1335 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1337 struct bond *bond = slave->bond;
1340 up = slave->up && lacp_slave_may_enable(bond->lacp, slave);
1341 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1342 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1343 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1344 slave->name, up ? "up" : "down");
1345 if (up == slave->enabled) {
1346 slave->delay_expires = LLONG_MAX;
1347 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1348 slave->name, up ? "disabled" : "enabled");
1350 int delay = (lacp_negotiated(bond->lacp) ? 0
1351 : up ? bond->updelay : bond->downdelay);
1352 slave->delay_expires = time_msec() + delay;
1354 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1357 up ? "enabled" : "disabled",
1364 if (time_msec() >= slave->delay_expires) {
1365 bond_enable_slave(slave, up, tags);
1370 bond_is_tcp_hash(const struct bond *bond)
1372 return bond->balance == BM_TCP && lacp_negotiated(bond->lacp);
1376 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan)
1378 return hash_bytes(mac, ETH_ADDR_LEN, vlan);
1382 bond_hash_tcp(const struct flow *flow, uint16_t vlan)
1384 struct flow hash_flow = *flow;
1385 hash_flow.vlan_tci = vlan;
1387 /* The symmetric quality of this hash function is not required, but
1388 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1389 * purposes, so we use it out of convenience. */
1390 return flow_hash_symmetric_l4(&hash_flow, 0);
1393 static struct bond_entry *
1394 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1397 assert(bond->balance != BM_AB);
1398 return &bond->hash[(bond_is_tcp_hash(bond)
1399 ? bond_hash_tcp(flow, vlan)
1400 : bond_hash_src(flow->dl_src, vlan)) & BOND_MASK];
1403 static struct bond_slave *
1404 choose_output_slave(const struct bond *bond, const struct flow *flow,
1407 struct bond_entry *e;
1409 switch (bond->balance) {
1411 return bond->active_slave;
1415 e = lookup_bond_entry(bond, flow, vlan);
1416 if (!e->slave || !e->slave->enabled) {
1417 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1418 struct bond_slave, hmap_node);
1419 if (!e->slave->enabled) {
1420 e->slave = bond->active_slave;
1422 e->tag = tag_create_random();
1431 static struct bond_slave *
1432 bond_choose_slave(const struct bond *bond)
1434 struct bond_slave *slave, *best;
1436 /* Find an enabled slave. */
1437 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1438 if (slave->enabled) {
1443 /* All interfaces are disabled. Find an interface that will be enabled
1444 * after its updelay expires. */
1446 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1447 if (slave->delay_expires != LLONG_MAX
1448 && lacp_slave_may_enable(bond->lacp, slave)
1449 && (!best || slave->delay_expires < best->delay_expires)) {
1457 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1459 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1460 struct bond_slave *old_active_slave = bond->active_slave;
1462 bond->active_slave = bond_choose_slave(bond);
1463 if (bond->active_slave) {
1464 if (bond->active_slave->enabled) {
1465 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1466 bond->name, bond->active_slave->name);
1468 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1469 "remaining %lld ms updelay (since no interface was "
1470 "enabled)", bond->name, bond->active_slave->name,
1471 bond->active_slave->delay_expires - time_msec());
1472 bond_enable_slave(bond->active_slave, true, tags);
1475 if (!old_active_slave) {
1476 tag_set_add(tags, bond->no_slaves_tag);
1479 bond->send_learning_packets = true;
1480 } else if (old_active_slave) {
1481 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1485 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1486 * there is no active slave. */
1488 bond_get_active_slave_tag(const struct bond *bond)
1490 return (bond->active_slave
1491 ? bond->active_slave->tag
1492 : bond->no_slaves_tag);
1495 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1496 * bond interface. */
1498 bond_update_fake_slave_stats(struct bond *bond)
1500 struct netdev_stats bond_stats;
1501 struct bond_slave *slave;
1502 struct netdev *bond_dev;
1504 memset(&bond_stats, 0, sizeof bond_stats);
1506 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1507 struct netdev_stats slave_stats;
1509 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1510 /* XXX: We swap the stats here because they are swapped back when
1511 * reported by the internal device. The reason for this is
1512 * internal devices normally represent packets going into the
1513 * system but when used as fake bond device they represent packets
1514 * leaving the system. We really should do this in the internal
1515 * device itself because changing it here reverses the counts from
1516 * the perspective of the switch. However, the internal device
1517 * doesn't know what type of device it represents so we have to do
1518 * it here for now. */
1519 bond_stats.tx_packets += slave_stats.rx_packets;
1520 bond_stats.tx_bytes += slave_stats.rx_bytes;
1521 bond_stats.rx_packets += slave_stats.tx_packets;
1522 bond_stats.rx_bytes += slave_stats.tx_bytes;
1526 if (!netdev_open_default(bond->name, &bond_dev)) {
1527 netdev_set_stats(bond_dev, &bond_stats);
1528 netdev_close(bond_dev);