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 * Returns true if the configuration has changed in such a way that requires
265 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
267 bool revalidate = false;
269 if (!bond->name || strcmp(bond->name, s->name)) {
271 hmap_remove(&all_bonds, &bond->hmap_node);
274 bond->name = xstrdup(s->name);
275 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
278 bond->detect = s->detect;
279 bond->miimon_interval = s->miimon_interval;
280 bond->updelay = s->up_delay;
281 bond->downdelay = s->down_delay;
282 bond->rebalance_interval = s->rebalance_interval;
284 if (bond->balance != s->balance) {
285 bond->balance = s->balance;
289 if (bond->balance != BM_AB) {
291 bond->hash = xcalloc(BOND_MASK + 1, sizeof *bond->hash);
292 bond->next_rebalance = time_msec() + bond->rebalance_interval;
301 if (bond->detect == BLSM_CARRIER) {
302 struct bond_slave *slave;
304 if (!bond->monitor) {
305 bond->monitor = netdev_monitor_create();
308 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
309 netdev_monitor_add(bond->monitor, slave->netdev);
312 netdev_monitor_destroy(bond->monitor);
313 bond->monitor = NULL;
315 if (bond->miimon_next_update == LLONG_MAX) {
316 bond->miimon_next_update = time_msec() + bond->miimon_interval;
322 bond->lacp = lacp_create();
324 lacp_configure(bond->lacp, s->lacp);
326 lacp_destroy(bond->lacp);
331 if (bond->next_fake_iface_update == LLONG_MAX) {
332 bond->next_fake_iface_update = time_msec();
335 bond->next_fake_iface_update = LLONG_MAX;
341 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
342 * arbitrary client-provided pointer that uniquely identifies a slave within a
343 * bond. If 'slave_' already exists within 'bond' then this function
344 * reconfigures the existing slave.
346 * 'netdev' must be the network device that 'slave_' represents. It is owned
347 * by the client, so the client must not close it before either unregistering
348 * 'slave_' or destroying 'bond'.
350 * If 'bond' has a LACP configuration then 'lacp_settings' must point to LACP
351 * settings for 'slave_'; otherwise 'lacp_settings' is ignored.
354 bond_slave_register(struct bond *bond, void *slave_, struct netdev *netdev,
355 const struct lacp_slave_settings *lacp_settings)
357 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
360 slave = xzalloc(sizeof *slave);
362 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
365 slave->delay_expires = LLONG_MAX;
366 slave->up = bond_is_link_up(bond, netdev);
367 slave->enabled = slave->up;
370 slave->netdev = netdev;
372 slave->name = xstrdup(netdev_get_name(netdev));
375 assert(lacp_settings != NULL);
376 lacp_slave_register(bond->lacp, slave, lacp_settings);
380 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
381 * then this function has no effect.
383 * Unregistering a slave invalidates all flows. */
385 bond_slave_unregister(struct bond *bond, const void *slave_)
387 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
394 del_active = bond->active_slave == slave;
396 struct bond_entry *e;
397 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
398 if (e->slave == slave) {
406 hmap_remove(&bond->slaves, &slave->hmap_node);
407 /* Client owns 'slave->netdev'. */
414 bond_choose_active_slave(bond, &tags);
415 bond->send_learning_packets = true;
419 /* Callback for lacp_run(). */
421 bond_send_pdu_cb(void *slave_, const struct lacp_pdu *pdu)
423 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 10);
424 struct bond_slave *slave = slave_;
425 uint8_t ea[ETH_ADDR_LEN];
428 error = netdev_get_etheraddr(slave->netdev, ea);
430 struct lacp_pdu *packet_pdu;
431 struct ofpbuf packet;
433 ofpbuf_init(&packet, 0);
434 packet_pdu = eth_compose(&packet, eth_addr_lacp, ea, ETH_TYPE_LACP,
437 error = netdev_send(slave->netdev, &packet);
439 VLOG_WARN_RL(&rl, "bond %s: sending LACP PDU on slave %s failed "
441 slave->bond->name, slave->name, strerror(error));
443 ofpbuf_uninit(&packet);
445 VLOG_ERR_RL(&rl, "bond %s: cannot obtain Ethernet address of slave "
447 slave->bond->name, slave->name, strerror(error));
451 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
452 * allow tagged flows to be invalidated.
454 * The caller should check bond_should_send_learning_packets() afterward. */
456 bond_run(struct bond *bond, struct tag_set *tags)
458 struct bond_slave *slave;
459 bool is_tcp_hash = bond_is_tcp_hash(bond);
461 /* Update link status. */
462 if (bond->detect == BLSM_CARRIER
463 || time_msec() >= bond->miimon_next_update)
465 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
466 slave->up = bond_is_link_up(bond, slave->netdev);
468 bond->miimon_next_update = time_msec() + bond->miimon_interval;
473 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
474 lacp_slave_enable(bond->lacp, slave, slave->enabled);
477 lacp_run(bond->lacp, bond_send_pdu_cb);
480 /* Enable slaves based on link status and LACP feedback. */
481 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
482 bond_link_status_update(slave, tags);
484 if (!bond->active_slave || !bond->active_slave->enabled) {
485 bond_choose_active_slave(bond, tags);
488 /* Update fake bond interface stats. */
489 if (time_msec() >= bond->next_fake_iface_update) {
490 bond_update_fake_slave_stats(bond);
491 bond->next_fake_iface_update = time_msec() + 1000;
494 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
495 struct bond_slave *slave;
497 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
498 tag_set_add(tags, slave->tag);
502 /* Invalidate any tags required by */
503 tag_set_union(tags, &bond->unixctl_tags);
504 tag_set_init(&bond->unixctl_tags);
507 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
509 bond_wait(struct bond *bond)
511 struct bond_slave *slave;
513 if (bond->detect == BLSM_CARRIER) {
514 netdev_monitor_poll_wait(bond->monitor);
516 poll_timer_wait_until(bond->miimon_next_update);
519 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
520 if (slave->delay_expires != LLONG_MAX) {
521 poll_timer_wait_until(slave->delay_expires);
525 if (bond->next_fake_iface_update != LLONG_MAX) {
526 poll_timer_wait_until(bond->next_fake_iface_update);
529 /* Ensure that any saved tags get revalidated right away. */
530 if (!tag_set_is_empty(&bond->unixctl_tags)) {
531 poll_immediate_wake();
534 /* We don't wait for bond->next_rebalance because rebalancing can only run
535 * at a flow account checkpoint. ofproto does checkpointing on its own
536 * schedule and bond_rebalance() gets called afterward, so we'd just be
537 * waking up for no purpose. */
540 /* MAC learning table interaction. */
543 may_send_learning_packets(const struct bond *bond)
545 return !lacp_negotiated(bond->lacp) && bond->balance != BM_AB;
548 /* Returns true if 'bond' needs the client to send out packets to assist with
549 * MAC learning on 'bond'. If this function returns true, then the client
550 * should iterate through its MAC learning table for the bridge on which 'bond'
551 * is located. For each MAC that has been learned on a port other than 'bond',
552 * it should call bond_send_learning_packet().
554 * This function will only return true if 'bond' is in SLB mode and LACP is not
555 * negotiated. Otherwise sending learning packets isn't necessary.
557 * Calling this function resets the state that it checks. */
559 bond_should_send_learning_packets(struct bond *bond)
561 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
562 bond->send_learning_packets = false;
566 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
568 * See bond_should_send_learning_packets() for description of usage. */
570 bond_send_learning_packet(struct bond *bond,
571 const uint8_t eth_src[ETH_ADDR_LEN],
574 struct bond_slave *slave;
575 struct ofpbuf packet;
579 assert(may_send_learning_packets(bond));
580 if (!bond->active_slave) {
581 /* Nowhere to send the learning packet. */
585 memset(&flow, 0, sizeof flow);
586 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
587 slave = choose_output_slave(bond, &flow, vlan);
589 ofpbuf_init(&packet, 0);
590 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
593 eth_set_vlan_tci(&packet, htons(vlan));
595 error = netdev_send(slave->netdev, &packet);
596 ofpbuf_uninit(&packet);
601 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
602 * Ethernet destination address of 'eth_dst', should be admitted.
604 * The return value is one of the following:
606 * - BV_ACCEPT: Admit the packet.
608 * - BV_DROP: Drop the packet.
610 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
611 * Ethernet source address and VLAN. If there is none, or if the packet
612 * is on the learned port, then admit the packet. If a different port has
613 * been learned, however, drop the packet (and do not use it for MAC
617 bond_check_admissibility(struct bond *bond, const void *slave_,
618 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
620 /* Admit all packets if LACP has been negotiated, because that means that
621 * the remote switch is aware of the bond and will "do the right thing". */
622 if (lacp_negotiated(bond->lacp)) {
626 /* Drop all multicast packets on inactive slaves. */
627 if (eth_addr_is_multicast(eth_dst)) {
628 *tags |= bond_get_active_slave_tag(bond);
629 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
634 /* Drop all packets for which we have learned a different input port,
635 * because we probably sent the packet on one slave and got it back on the
636 * other. Gratuitous ARP packets are an exception to this rule: the host
637 * has moved to another switch. The exception to the exception is if we
638 * locked the learning table to avoid reflections on bond slaves. */
639 return BV_DROP_IF_MOVED;
642 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
643 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
644 * NULL if the packet should be dropped because no slaves are enabled.
646 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
647 * should be a VID only (i.e. excluding the PCP bits). Second,
648 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
649 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
650 * packet belongs to (so for an access port it will be the access port's VLAN).
652 * Adds a tag to '*tags' that associates the flow with the returned slave.
655 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
656 uint16_t vlan, tag_type *tags)
658 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
663 *tags |= bond->no_slaves_tag;
668 /* Processes LACP packet 'packet', which was received on 'slave_' within
671 * The client should use this function to pass along LACP messages received on
672 * any of 'bond''s slaves. */
674 bond_process_lacp(struct bond *bond, void *slave_, const struct ofpbuf *packet)
677 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
678 const struct lacp_pdu *pdu = parse_lacp_packet(packet);
680 COVERAGE_INC(bond_process_lacp);
681 lacp_process_pdu(bond->lacp, slave, pdu);
688 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
690 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
693 switch (bond->balance) {
700 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
708 static struct bond_slave *
709 bond_slave_from_bal_node(struct list *bal)
711 return CONTAINER_OF(bal, struct bond_slave, bal_node);
715 log_bals(struct bond *bond, const struct list *bals)
717 if (VLOG_IS_DBG_ENABLED()) {
718 struct ds ds = DS_EMPTY_INITIALIZER;
719 const struct bond_slave *slave;
721 LIST_FOR_EACH (slave, bal_node, bals) {
723 ds_put_char(&ds, ',');
725 ds_put_format(&ds, " %s %"PRIu64"kB",
726 slave->name, slave->tx_bytes / 1024);
728 if (!slave->enabled) {
729 ds_put_cstr(&ds, " (disabled)");
731 if (!list_is_empty(&slave->entries)) {
732 struct bond_entry *e;
734 ds_put_cstr(&ds, " (");
735 LIST_FOR_EACH (e, list_node, &slave->entries) {
736 if (&e->list_node != list_front(&slave->entries)) {
737 ds_put_cstr(&ds, " + ");
739 ds_put_format(&ds, "h%td: %"PRIu64"kB",
740 e - bond->hash, e->tx_bytes / 1024);
742 ds_put_cstr(&ds, ")");
745 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
750 /* Shifts 'hash' from its current slave to 'to'. */
752 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
755 struct bond_slave *from = hash->slave;
756 struct bond *bond = from->bond;
757 uint64_t delta = hash->tx_bytes;
759 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
760 "from %s to %s (now carrying %"PRIu64"kB and "
761 "%"PRIu64"kB load, respectively)",
762 bond->name, delta / 1024, hash - bond->hash,
763 from->name, to->name,
764 (from->tx_bytes - delta) / 1024,
765 (to->tx_bytes + delta) / 1024);
767 /* Shift load away from 'from' to 'to'. */
768 from->tx_bytes -= delta;
769 to->tx_bytes += delta;
771 /* Arrange for flows to be revalidated. */
772 tag_set_add(set, hash->tag);
774 hash->tag = tag_create_random();
777 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
778 * given that doing so must decrease the ratio of the load on the two slaves by
779 * at least 0.1. Returns NULL if there is no appropriate entry.
781 * The list of entries isn't sorted. I don't know of a reason to prefer to
782 * shift away small hashes or large hashes. */
783 static struct bond_entry *
784 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
786 struct bond_entry *e;
788 if (list_is_short(&from->entries)) {
789 /* 'from' carries no more than one MAC hash, so shifting load away from
790 * it would be pointless. */
794 LIST_FOR_EACH (e, list_node, &from->entries) {
795 double old_ratio, new_ratio;
798 if (to_tx_bytes == 0) {
799 /* Nothing on the new slave, move it. */
804 old_ratio = (double)from->tx_bytes / to_tx_bytes;
805 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
806 if (old_ratio - new_ratio > 0.1) {
807 /* Would decrease the ratio, move it. */
815 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
818 insert_bal(struct list *bals, struct bond_slave *slave)
820 struct bond_slave *pos;
822 LIST_FOR_EACH (pos, bal_node, bals) {
823 if (slave->tx_bytes > pos->tx_bytes) {
827 list_insert(&pos->bal_node, &slave->bal_node);
830 /* Removes 'slave' from its current list and then inserts it into 'bals' so
831 * that descending order of 'tx_bytes' is maintained. */
833 reinsert_bal(struct list *bals, struct bond_slave *slave)
835 list_remove(&slave->bal_node);
836 insert_bal(bals, slave);
839 /* If 'bond' needs rebalancing, does so.
841 * The caller should have called bond_account() for each active flow, to ensure
842 * that flow data is consistently accounted at this point. */
844 bond_rebalance(struct bond *bond, struct tag_set *tags)
846 struct bond_slave *slave;
847 struct bond_entry *e;
850 if (bond->balance == BM_AB || time_msec() < bond->next_rebalance) {
853 bond->next_rebalance = time_msec() + bond->rebalance_interval;
855 /* Add each bond_entry to its slave's 'entries' list.
856 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
857 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
859 list_init(&slave->entries);
861 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
862 if (e->slave && e->tx_bytes) {
863 e->slave->tx_bytes += e->tx_bytes;
864 list_push_back(&e->slave->entries, &e->list_node);
868 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
870 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
871 * with a proper list sort algorithm. */
873 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
874 if (slave->enabled) {
875 insert_bal(&bals, slave);
878 log_bals(bond, &bals);
880 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
881 while (!list_is_short(&bals)) {
882 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
883 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
886 overload = from->tx_bytes - to->tx_bytes;
887 if (overload < to->tx_bytes >> 5 || overload < 100000) {
888 /* The extra load on 'from' (and all less-loaded slaves), compared
889 * to that of 'to' (the least-loaded slave), is less than ~3%, or
890 * it is less than ~1Mbps. No point in rebalancing. */
894 /* 'from' is carrying significantly more load than 'to', and that load
895 * is split across at least two different hashes. */
896 e = choose_entry_to_migrate(from, to->tx_bytes);
898 bond_shift_load(e, to, tags);
900 /* Delete element from from->entries.
902 * We don't add the element to to->hashes. That would only allow
903 * 'e' to be migrated to another slave in this rebalancing run, and
904 * there is no point in doing that. */
905 list_remove(&e->list_node);
907 /* Re-sort 'bals'. */
908 reinsert_bal(&bals, from);
909 reinsert_bal(&bals, to);
911 /* Can't usefully migrate anything away from 'from'.
912 * Don't reconsider it. */
913 list_remove(&from->bal_node);
917 /* Implement exponentially weighted moving average. A weight of 1/2 causes
918 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
919 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
920 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
928 /* Bonding unixctl user interface functions. */
931 bond_find(const char *name)
935 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
937 if (!strcmp(bond->name, name)) {
944 static struct bond_slave *
945 bond_lookup_slave(struct bond *bond, const char *slave_name)
947 struct bond_slave *slave;
949 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
950 if (!strcmp(slave->name, slave_name)) {
958 bond_unixctl_list(struct unixctl_conn *conn,
959 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
961 struct ds ds = DS_EMPTY_INITIALIZER;
962 const struct bond *bond;
964 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
966 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
967 const struct bond_slave *slave;
970 ds_put_format(&ds, "%s\t%s\t",
971 bond->name, bond_mode_to_string(bond->balance));
974 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
976 ds_put_cstr(&ds, ", ");
978 ds_put_cstr(&ds, slave->name);
980 ds_put_char(&ds, '\n');
982 unixctl_command_reply(conn, 200, ds_cstr(&ds));
987 bond_unixctl_show(struct unixctl_conn *conn,
988 const char *args, void *aux OVS_UNUSED)
990 struct ds ds = DS_EMPTY_INITIALIZER;
991 const struct bond_slave *slave;
992 const struct bond *bond;
994 bond = bond_find(args);
996 unixctl_command_reply(conn, 501, "no such bond");
1000 ds_put_format(&ds, "bond_mode: %s\n",
1001 bond_mode_to_string(bond->balance));
1004 ds_put_format(&ds, "lacp: %s\n",
1005 lacp_is_active(bond->lacp) ? "active" : "passive");
1007 ds_put_cstr(&ds, "lacp: off\n");
1010 if (bond->balance != BM_AB) {
1011 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
1012 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
1015 ds_put_format(&ds, "bond-detect-mode: %s\n",
1016 bond->monitor ? "carrier" : "miimon");
1018 if (!bond->monitor) {
1019 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
1020 bond->miimon_interval);
1023 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
1024 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
1026 if (bond->balance != BM_AB) {
1027 ds_put_format(&ds, "next rebalance: %lld ms\n",
1028 bond->next_rebalance - time_msec());
1031 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1032 struct bond_entry *be;
1036 ds_put_format(&ds, "\nslave %s: %s\n",
1037 slave->name, slave->enabled ? "enabled" : "disabled");
1038 if (slave == bond->active_slave) {
1039 ds_put_cstr(&ds, "\tactive slave\n");
1041 if (slave->delay_expires != LLONG_MAX) {
1042 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1043 slave->enabled ? "downdelay" : "updelay",
1044 slave->delay_expires - time_msec());
1047 if (bond->balance == BM_AB) {
1052 memset(&flow, 0, sizeof flow);
1053 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1054 int hash = be - bond->hash;
1056 if (be->slave != slave) {
1060 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1061 hash, be->tx_bytes / 1024);
1063 if (bond->balance != BM_SLB) {
1067 /* XXX How can we list the MACs assigned to hashes? */
1070 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1075 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1076 void *aux OVS_UNUSED)
1078 char *args = (char *) args_;
1079 char *save_ptr = NULL;
1080 char *bond_s, *hash_s, *slave_s;
1082 struct bond_slave *slave;
1083 struct bond_entry *entry;
1086 bond_s = strtok_r(args, " ", &save_ptr);
1087 hash_s = strtok_r(NULL, " ", &save_ptr);
1088 slave_s = strtok_r(NULL, " ", &save_ptr);
1090 unixctl_command_reply(conn, 501,
1091 "usage: bond/migrate BOND HASH SLAVE");
1095 bond = bond_find(bond_s);
1097 unixctl_command_reply(conn, 501, "no such bond");
1101 if (bond->balance != BM_SLB) {
1102 unixctl_command_reply(conn, 501, "not an SLB bond");
1106 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1107 hash = atoi(hash_s) & BOND_MASK;
1109 unixctl_command_reply(conn, 501, "bad hash");
1113 slave = bond_lookup_slave(bond, slave_s);
1115 unixctl_command_reply(conn, 501, "no such slave");
1119 if (!slave->enabled) {
1120 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1124 entry = &bond->hash[hash];
1125 tag_set_add(&bond->unixctl_tags, entry->tag);
1126 entry->slave = slave;
1127 entry->tag = tag_create_random();
1128 unixctl_command_reply(conn, 200, "migrated");
1132 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1133 void *aux OVS_UNUSED)
1135 char *args = (char *) args_;
1136 char *save_ptr = NULL;
1137 char *bond_s, *slave_s;
1139 struct bond_slave *slave;
1141 bond_s = strtok_r(args, " ", &save_ptr);
1142 slave_s = strtok_r(NULL, " ", &save_ptr);
1144 unixctl_command_reply(conn, 501,
1145 "usage: bond/set-active-slave BOND SLAVE");
1149 bond = bond_find(bond_s);
1151 unixctl_command_reply(conn, 501, "no such bond");
1155 slave = bond_lookup_slave(bond, slave_s);
1157 unixctl_command_reply(conn, 501, "no such slave");
1161 if (!slave->enabled) {
1162 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1166 if (bond->active_slave != slave) {
1167 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1168 bond->active_slave = slave;
1169 bond->active_slave->tag = tag_create_random();
1170 VLOG_INFO("bond %s: active interface is now %s",
1171 bond->name, slave->name);
1172 bond->send_learning_packets = true;
1173 unixctl_command_reply(conn, 200, "done");
1175 unixctl_command_reply(conn, 200, "no change");
1180 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1182 char *args = (char *) args_;
1183 char *save_ptr = NULL;
1184 char *bond_s, *slave_s;
1186 struct bond_slave *slave;
1188 bond_s = strtok_r(args, " ", &save_ptr);
1189 slave_s = strtok_r(NULL, " ", &save_ptr);
1191 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1192 enable ? "enable" : "disable");
1193 unixctl_command_reply(conn, 501, usage);
1198 bond = bond_find(bond_s);
1200 unixctl_command_reply(conn, 501, "no such bond");
1204 slave = bond_lookup_slave(bond, slave_s);
1206 unixctl_command_reply(conn, 501, "no such slave");
1210 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1211 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1215 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1216 void *aux OVS_UNUSED)
1218 enable_slave(conn, args, true);
1222 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1223 void *aux OVS_UNUSED)
1225 enable_slave(conn, args, false);
1229 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1230 void *aux OVS_UNUSED)
1232 char *args = (char *) args_;
1233 uint8_t mac[ETH_ADDR_LEN];
1237 char *mac_s, *vlan_s;
1238 char *save_ptr = NULL;
1240 mac_s = strtok_r(args, " ", &save_ptr);
1241 vlan_s = strtok_r(NULL, " ", &save_ptr);
1244 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1245 unixctl_command_reply(conn, 501, "invalid vlan");
1249 vlan = OFP_VLAN_NONE;
1252 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1253 == ETH_ADDR_SCAN_COUNT) {
1254 hash = bond_hash_src(mac, vlan) & BOND_MASK;
1256 hash_cstr = xasprintf("%u", hash);
1257 unixctl_command_reply(conn, 200, hash_cstr);
1260 unixctl_command_reply(conn, 501, "invalid mac");
1269 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1270 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1271 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1272 unixctl_command_register("bond/set-active-slave",
1273 bond_unixctl_set_active_slave, NULL);
1274 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1276 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1278 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1281 static struct bond_slave *
1282 bond_slave_lookup(struct bond *bond, const void *slave_)
1284 struct bond_slave *slave;
1286 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1288 if (slave->aux == slave_) {
1297 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1299 return (bond->detect == BLSM_CARRIER
1300 ? netdev_get_carrier(netdev)
1301 : netdev_get_miimon(netdev));
1305 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1307 slave->delay_expires = LLONG_MAX;
1308 if (enable != slave->enabled) {
1309 slave->enabled = enable;
1310 if (!slave->enabled) {
1311 VLOG_WARN("interface %s: disabled", slave->name);
1312 tag_set_add(tags, slave->tag);
1314 VLOG_WARN("interface %s: enabled", slave->name);
1315 slave->tag = tag_create_random();
1321 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1323 struct bond *bond = slave->bond;
1326 up = slave->up && lacp_slave_may_enable(bond->lacp, slave);
1327 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1328 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1329 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1330 slave->name, up ? "up" : "down");
1331 if (up == slave->enabled) {
1332 slave->delay_expires = LLONG_MAX;
1333 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1334 slave->name, up ? "disabled" : "enabled");
1336 int delay = (lacp_negotiated(bond->lacp) ? 0
1337 : up ? bond->updelay : bond->downdelay);
1338 slave->delay_expires = time_msec() + delay;
1340 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1343 up ? "enabled" : "disabled",
1350 if (time_msec() >= slave->delay_expires) {
1351 bond_enable_slave(slave, up, tags);
1356 bond_is_tcp_hash(const struct bond *bond)
1358 return bond->balance == BM_TCP && lacp_negotiated(bond->lacp);
1362 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan)
1364 return hash_bytes(mac, ETH_ADDR_LEN, vlan);
1368 bond_hash_tcp(const struct flow *flow, uint16_t vlan)
1370 struct flow hash_flow = *flow;
1371 hash_flow.vlan_tci = vlan;
1373 /* The symmetric quality of this hash function is not required, but
1374 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1375 * purposes, so we use it out of convenience. */
1376 return flow_hash_symmetric_l4(&hash_flow, 0);
1379 static struct bond_entry *
1380 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1383 assert(bond->balance != BM_AB);
1384 return &bond->hash[(bond_is_tcp_hash(bond)
1385 ? bond_hash_tcp(flow, vlan)
1386 : bond_hash_src(flow->dl_src, vlan)) & BOND_MASK];
1389 static struct bond_slave *
1390 choose_output_slave(const struct bond *bond, const struct flow *flow,
1393 struct bond_entry *e;
1395 switch (bond->balance) {
1397 return bond->active_slave;
1401 e = lookup_bond_entry(bond, flow, vlan);
1402 if (!e->slave || !e->slave->enabled) {
1403 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1404 struct bond_slave, hmap_node);
1405 if (!e->slave->enabled) {
1406 e->slave = bond->active_slave;
1408 e->tag = tag_create_random();
1417 static struct bond_slave *
1418 bond_choose_slave(const struct bond *bond)
1420 struct bond_slave *slave, *best;
1422 /* Find an enabled slave. */
1423 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1424 if (slave->enabled) {
1429 /* All interfaces are disabled. Find an interface that will be enabled
1430 * after its updelay expires. */
1432 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1433 if (slave->delay_expires != LLONG_MAX
1434 && lacp_slave_may_enable(bond->lacp, slave)
1435 && (!best || slave->delay_expires < best->delay_expires)) {
1443 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1445 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1446 struct bond_slave *old_active_slave = bond->active_slave;
1448 bond->active_slave = bond_choose_slave(bond);
1449 if (bond->active_slave) {
1450 if (bond->active_slave->enabled) {
1451 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1452 bond->name, bond->active_slave->name);
1454 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1455 "remaining %lld ms updelay (since no interface was "
1456 "enabled)", bond->name, bond->active_slave->name,
1457 bond->active_slave->delay_expires - time_msec());
1458 bond_enable_slave(bond->active_slave, true, tags);
1461 if (!old_active_slave) {
1462 tag_set_add(tags, bond->no_slaves_tag);
1465 bond->send_learning_packets = true;
1466 } else if (old_active_slave) {
1467 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1471 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1472 * there is no active slave. */
1474 bond_get_active_slave_tag(const struct bond *bond)
1476 return (bond->active_slave
1477 ? bond->active_slave->tag
1478 : bond->no_slaves_tag);
1481 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1482 * bond interface. */
1484 bond_update_fake_slave_stats(struct bond *bond)
1486 struct netdev_stats bond_stats;
1487 struct bond_slave *slave;
1488 struct netdev *bond_dev;
1490 memset(&bond_stats, 0, sizeof bond_stats);
1492 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1493 struct netdev_stats slave_stats;
1495 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1496 /* XXX: We swap the stats here because they are swapped back when
1497 * reported by the internal device. The reason for this is
1498 * internal devices normally represent packets going into the
1499 * system but when used as fake bond device they represent packets
1500 * leaving the system. We really should do this in the internal
1501 * device itself because changing it here reverses the counts from
1502 * the perspective of the switch. However, the internal device
1503 * doesn't know what type of device it represents so we have to do
1504 * it here for now. */
1505 bond_stats.tx_packets += slave_stats.rx_packets;
1506 bond_stats.tx_bytes += slave_stats.rx_bytes;
1507 bond_stats.rx_packets += slave_stats.tx_packets;
1508 bond_stats.rx_bytes += slave_stats.tx_bytes;
1512 if (!netdev_open_default(bond->name, &bond_dev)) {
1513 netdev_set_stats(bond_dev, &bond_stats);
1514 netdev_close(bond_dev);