2 * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
26 #include "dynamic-string.h"
34 #include "poll-loop.h"
40 VLOG_DEFINE_THIS_MODULE(bond);
42 /* Bit-mask for hashing a flow down to a bucket.
43 * There are (BOND_MASK + 1) buckets. */
44 #define BOND_MASK 0xff
46 /* A hash bucket for mapping a flow to a slave.
47 * "struct bond" has an array of (BOND_MASK + 1) of these. */
49 struct bond_slave *slave; /* Assigned slave, NULL if unassigned. */
50 uint64_t tx_bytes; /* Count of bytes recently transmitted. */
51 tag_type tag; /* Tag for entry<->slave association. */
52 struct list list_node; /* In bond_slave's 'entries' list. */
55 /* A bond slave, that is, one of the links comprising a bond. */
57 struct hmap_node hmap_node; /* In struct bond's slaves hmap. */
58 struct bond *bond; /* The bond that contains this slave. */
59 void *aux; /* Client-provided handle for this slave. */
61 struct netdev *netdev; /* Network device, owned by the client. */
62 char *name; /* Name (a copy of netdev_get_name(netdev)). */
65 long long delay_expires; /* Time after which 'enabled' may change. */
66 bool up; /* Last link status read from netdev. */
67 bool enabled; /* May be chosen for flows? */
68 bool lacp_may_enable; /* LACP considers this interface bondable. */
69 tag_type tag; /* Tag associated with this slave. */
71 /* Rebalancing info. Used only by bond_rebalance(). */
72 struct list bal_node; /* In bond_rebalance()'s 'bals' list. */
73 struct list entries; /* 'struct bond_entry's assigned here. */
74 uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
76 /* BM_STABLE specific bonding info. */
77 uint16_t stb_id; /* ID used for 'stb_slaves' ordering. */
78 size_t stb_idx; /* Index in 'bond''s 'stb_slaves' array.
79 Undefined value if participating in a
80 BTM_STABLE bond or not enabled. */
83 /* A bond, that is, a set of network devices grouped to improve performance or
86 struct hmap_node hmap_node; /* In 'all_bonds' hmap. */
87 char *name; /* Name provided by client. */
93 enum bond_mode balance; /* Balancing mode, one of BM_*. */
94 struct bond_slave *active_slave;
95 tag_type no_slaves_tag; /* Tag for flows when all slaves disabled. */
96 int updelay, downdelay; /* Delay before slave goes up/down, in ms. */
97 bool lacp_negotiated; /* LACP negotiations were successful. */
99 /* SLB specific bonding info. */
100 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
101 int rebalance_interval; /* Interval between rebalances, in ms. */
102 long long int next_rebalance; /* Next rebalancing time. */
103 bool send_learning_packets;
105 /* BM_STABLE specific bonding info. */
106 struct bond_slave **stb_slaves; /* Ordered list of enabled slaves. */
107 size_t n_stb_slaves; /* Number of slaves in 'stb_slaves'. */
108 size_t len_stb_slaves; /* Slaves allocated in 'stb_slaves'. */
109 bool stb_need_sort; /* True if stb_slaves is not sorted. */
110 tag_type stb_tag; /* Tag associated with this bond. */
113 enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
114 struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
115 long long int miimon_interval; /* Miimon status refresh interval. */
116 long long int miimon_next_update; /* Time of next miimon update. */
118 /* Legacy compatibility. */
119 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
121 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
122 * where we can't otherwise provide revalidation feedback to the client.
123 * That's only unixctl commands now; I hope no other cases will arise. */
124 struct tag_set unixctl_tags;
127 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
129 static void bond_entry_reset(struct bond *);
130 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
131 static bool bond_is_link_up(struct bond *, struct netdev *);
132 static void bond_enable_slave(struct bond_slave *, bool enable,
134 static bool bond_stb_sort(struct bond *);
135 static void bond_stb_enable_slave(struct bond_slave *);
136 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
137 static void bond_choose_active_slave(struct bond *, struct tag_set *);
138 static bool bond_is_tcp_hash(const struct bond *);
139 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
141 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan);
142 static struct bond_entry *lookup_bond_entry(const struct bond *,
145 static tag_type bond_get_active_slave_tag(const struct bond *);
146 static struct bond_slave *choose_output_slave(const struct bond *,
149 static void bond_update_fake_slave_stats(struct bond *);
151 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
152 * stores the mode in '*balance' and returns true. Otherwise returns false
153 * without modifying '*balance'. */
155 bond_mode_from_string(enum bond_mode *balance, const char *s)
157 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
159 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
161 } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
162 *balance = BM_STABLE;
163 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
171 /* Returns a string representing 'balance'. */
173 bond_mode_to_string(enum bond_mode balance) {
176 return "balance-tcp";
178 return "balance-slb";
182 return "active-backup";
187 /* Attempts to parse 's' as the name of a bond link status detection mode. If
188 * successful, stores the mode in '*detect' and returns true. Otherwise
189 * returns false without modifying '*detect'. */
191 bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
193 if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
194 *detect = BLSM_CARRIER;
195 } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
196 *detect = BLSM_MIIMON;
203 /* Returns a string representing 'detect'. */
205 bond_detect_mode_to_string(enum bond_detect_mode detect)
216 /* Creates and returns a new bond whose configuration is initially taken from
219 * The caller should register each slave on the new bond by calling
220 * bond_slave_register(). */
222 bond_create(const struct bond_settings *s)
226 bond = xzalloc(sizeof *bond);
227 hmap_init(&bond->slaves);
228 bond->no_slaves_tag = tag_create_random();
229 bond->miimon_next_update = LLONG_MAX;
230 bond->next_fake_iface_update = LLONG_MAX;
232 bond_reconfigure(bond, s);
234 tag_set_init(&bond->unixctl_tags);
241 bond_destroy(struct bond *bond)
243 struct bond_slave *slave, *next_slave;
249 hmap_remove(&all_bonds, &bond->hmap_node);
251 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
252 hmap_remove(&bond->slaves, &slave->hmap_node);
253 /* Client owns 'slave->netdev'. */
257 hmap_destroy(&bond->slaves);
261 netdev_monitor_destroy(bond->monitor);
267 /* Updates 'bond''s overall configuration to 's'.
269 * The caller should register each slave on 'bond' by calling
270 * bond_slave_register(). This is optional if none of the slaves'
271 * configuration has changed. In any case it can't hurt.
273 * Returns true if the configuration has changed in such a way that requires
277 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
279 bool revalidate = false;
281 if (!bond->name || strcmp(bond->name, s->name)) {
283 hmap_remove(&all_bonds, &bond->hmap_node);
286 bond->name = xstrdup(s->name);
287 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
290 bond->detect = s->detect;
291 bond->miimon_interval = s->miimon_interval;
292 bond->updelay = s->up_delay;
293 bond->downdelay = s->down_delay;
294 bond->rebalance_interval = s->rebalance_interval;
296 if (bond->balance != s->balance) {
297 bond->balance = s->balance;
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;
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;
328 if (bond->balance != BM_STABLE) {
329 free(bond->stb_slaves);
330 bond->stb_slaves = NULL;
332 } else if (!bond->stb_tag) {
333 struct bond_slave *slave;
335 bond->stb_tag = tag_create_random();
337 assert(!bond->stb_slaves);
338 bond->n_stb_slaves = 0;
339 bond->len_stb_slaves = 0;
340 bond->stb_slaves = NULL;
342 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
343 if (slave->enabled) {
344 bond_stb_enable_slave(slave);
349 if (bond->balance == BM_AB || !bond->hash || revalidate) {
350 bond_entry_reset(bond);
356 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
357 * arbitrary client-provided pointer that uniquely identifies a slave within a
358 * bond. If 'slave_' already exists within 'bond' then this function
359 * reconfigures the existing slave.
361 * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
362 * across restarts and distributed vswitch instances. It should be unique per
363 * slave, and preferably consistent across restarts and reconfigurations.
365 * 'netdev' must be the network device that 'slave_' represents. It is owned
366 * by the client, so the client must not close it before either unregistering
367 * 'slave_' or destroying 'bond'.
370 bond_slave_register(struct bond *bond, void *slave_, uint16_t stb_id,
371 struct netdev *netdev)
373 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
376 slave = xzalloc(sizeof *slave);
378 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
381 slave->delay_expires = LLONG_MAX;
382 slave->up = bond_is_link_up(bond, netdev);
383 slave->name = xstrdup(netdev_get_name(netdev));
385 slave->enabled = false;
386 bond_enable_slave(slave, slave->up, NULL);
389 if (slave->stb_id != stb_id) {
390 bond->stb_need_sort = true;
391 slave->stb_id = stb_id;
394 slave->netdev = netdev;
396 slave->name = xstrdup(netdev_get_name(netdev));
399 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
400 * then this function has no effect.
402 * Unregistering a slave invalidates all flows. */
404 bond_slave_unregister(struct bond *bond, const void *slave_)
406 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
413 bond_enable_slave(slave, false, NULL);
415 del_active = bond->active_slave == slave;
417 struct bond_entry *e;
418 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
419 if (e->slave == slave) {
427 hmap_remove(&bond->slaves, &slave->hmap_node);
428 /* Client owns 'slave->netdev'. */
435 bond_choose_active_slave(bond, &tags);
436 bond->send_learning_packets = true;
440 /* Should be called on each slave in 'bond' before bond_run() to indicate the
441 * results of lacp_slave_may_enable() on 'slave_'. */
443 bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
446 bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
449 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
450 * allow tagged flows to be invalidated.
452 * The caller should check bond_should_send_learning_packets() afterward. */
454 bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
456 struct bond_slave *slave;
457 bool is_tcp_hash = bond_is_tcp_hash(bond);
459 bond->lacp_negotiated = lacp_negotiated;
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;
471 /* Enable slaves based on link status and LACP feedback. */
472 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
473 bond_link_status_update(slave, tags);
475 if (!bond->active_slave || !bond->active_slave->enabled) {
476 bond_choose_active_slave(bond, tags);
479 /* Update fake bond interface stats. */
480 if (time_msec() >= bond->next_fake_iface_update) {
481 bond_update_fake_slave_stats(bond);
482 bond->next_fake_iface_update = time_msec() + 1000;
485 if (bond_stb_sort(bond) || is_tcp_hash != bond_is_tcp_hash(bond)) {
487 bond_entry_reset(bond);
488 if (bond->balance != BM_STABLE) {
489 struct bond_slave *slave;
491 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
492 tag_set_add(tags, slave->tag);
495 tag_set_add(tags, bond->stb_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 !bond->lacp_negotiated && 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 (bond->lacp_negotiated) {
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);
657 *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
660 *tags |= bond->no_slaves_tag;
668 bond_is_balanced(const struct bond *bond)
670 return bond->balance == BM_SLB || bond->balance == BM_TCP;
673 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
675 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
679 if (bond_is_balanced(bond)) {
680 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
684 static struct bond_slave *
685 bond_slave_from_bal_node(struct list *bal)
687 return CONTAINER_OF(bal, struct bond_slave, bal_node);
691 log_bals(struct bond *bond, const struct list *bals)
693 if (VLOG_IS_DBG_ENABLED()) {
694 struct ds ds = DS_EMPTY_INITIALIZER;
695 const struct bond_slave *slave;
697 LIST_FOR_EACH (slave, bal_node, bals) {
699 ds_put_char(&ds, ',');
701 ds_put_format(&ds, " %s %"PRIu64"kB",
702 slave->name, slave->tx_bytes / 1024);
704 if (!slave->enabled) {
705 ds_put_cstr(&ds, " (disabled)");
707 if (!list_is_empty(&slave->entries)) {
708 struct bond_entry *e;
710 ds_put_cstr(&ds, " (");
711 LIST_FOR_EACH (e, list_node, &slave->entries) {
712 if (&e->list_node != list_front(&slave->entries)) {
713 ds_put_cstr(&ds, " + ");
715 ds_put_format(&ds, "h%td: %"PRIu64"kB",
716 e - bond->hash, e->tx_bytes / 1024);
718 ds_put_cstr(&ds, ")");
721 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
726 /* Shifts 'hash' from its current slave to 'to'. */
728 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
731 struct bond_slave *from = hash->slave;
732 struct bond *bond = from->bond;
733 uint64_t delta = hash->tx_bytes;
735 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
736 "from %s to %s (now carrying %"PRIu64"kB and "
737 "%"PRIu64"kB load, respectively)",
738 bond->name, delta / 1024, hash - bond->hash,
739 from->name, to->name,
740 (from->tx_bytes - delta) / 1024,
741 (to->tx_bytes + delta) / 1024);
743 /* Shift load away from 'from' to 'to'. */
744 from->tx_bytes -= delta;
745 to->tx_bytes += delta;
747 /* Arrange for flows to be revalidated. */
748 tag_set_add(set, hash->tag);
750 hash->tag = tag_create_random();
753 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
754 * given that doing so must decrease the ratio of the load on the two slaves by
755 * at least 0.1. Returns NULL if there is no appropriate entry.
757 * The list of entries isn't sorted. I don't know of a reason to prefer to
758 * shift away small hashes or large hashes. */
759 static struct bond_entry *
760 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
762 struct bond_entry *e;
764 if (list_is_short(&from->entries)) {
765 /* 'from' carries no more than one MAC hash, so shifting load away from
766 * it would be pointless. */
770 LIST_FOR_EACH (e, list_node, &from->entries) {
771 double old_ratio, new_ratio;
774 if (to_tx_bytes == 0) {
775 /* Nothing on the new slave, move it. */
780 old_ratio = (double)from->tx_bytes / to_tx_bytes;
781 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
782 if (old_ratio - new_ratio > 0.1) {
783 /* Would decrease the ratio, move it. */
791 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
794 insert_bal(struct list *bals, struct bond_slave *slave)
796 struct bond_slave *pos;
798 LIST_FOR_EACH (pos, bal_node, bals) {
799 if (slave->tx_bytes > pos->tx_bytes) {
803 list_insert(&pos->bal_node, &slave->bal_node);
806 /* Removes 'slave' from its current list and then inserts it into 'bals' so
807 * that descending order of 'tx_bytes' is maintained. */
809 reinsert_bal(struct list *bals, struct bond_slave *slave)
811 list_remove(&slave->bal_node);
812 insert_bal(bals, slave);
815 /* If 'bond' needs rebalancing, does so.
817 * The caller should have called bond_account() for each active flow, to ensure
818 * that flow data is consistently accounted at this point. */
820 bond_rebalance(struct bond *bond, struct tag_set *tags)
822 struct bond_slave *slave;
823 struct bond_entry *e;
826 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
829 bond->next_rebalance = time_msec() + bond->rebalance_interval;
831 /* Add each bond_entry to its slave's 'entries' list.
832 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
833 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
835 list_init(&slave->entries);
837 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
838 if (e->slave && e->tx_bytes) {
839 e->slave->tx_bytes += e->tx_bytes;
840 list_push_back(&e->slave->entries, &e->list_node);
844 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
846 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
847 * with a proper list sort algorithm. */
849 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
850 if (slave->enabled) {
851 insert_bal(&bals, slave);
854 log_bals(bond, &bals);
856 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
857 while (!list_is_short(&bals)) {
858 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
859 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
862 overload = from->tx_bytes - to->tx_bytes;
863 if (overload < to->tx_bytes >> 5 || overload < 100000) {
864 /* The extra load on 'from' (and all less-loaded slaves), compared
865 * to that of 'to' (the least-loaded slave), is less than ~3%, or
866 * it is less than ~1Mbps. No point in rebalancing. */
870 /* 'from' is carrying significantly more load than 'to', and that load
871 * is split across at least two different hashes. */
872 e = choose_entry_to_migrate(from, to->tx_bytes);
874 bond_shift_load(e, to, tags);
876 /* Delete element from from->entries.
878 * We don't add the element to to->hashes. That would only allow
879 * 'e' to be migrated to another slave in this rebalancing run, and
880 * there is no point in doing that. */
881 list_remove(&e->list_node);
883 /* Re-sort 'bals'. */
884 reinsert_bal(&bals, from);
885 reinsert_bal(&bals, to);
887 /* Can't usefully migrate anything away from 'from'.
888 * Don't reconsider it. */
889 list_remove(&from->bal_node);
893 /* Implement exponentially weighted moving average. A weight of 1/2 causes
894 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
895 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
896 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
904 /* Bonding unixctl user interface functions. */
907 bond_find(const char *name)
911 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
913 if (!strcmp(bond->name, name)) {
920 static struct bond_slave *
921 bond_lookup_slave(struct bond *bond, const char *slave_name)
923 struct bond_slave *slave;
925 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
926 if (!strcmp(slave->name, slave_name)) {
934 bond_unixctl_list(struct unixctl_conn *conn,
935 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
937 struct ds ds = DS_EMPTY_INITIALIZER;
938 const struct bond *bond;
940 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
942 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
943 const struct bond_slave *slave;
946 ds_put_format(&ds, "%s\t%s\t",
947 bond->name, bond_mode_to_string(bond->balance));
950 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
952 ds_put_cstr(&ds, ", ");
954 ds_put_cstr(&ds, slave->name);
956 ds_put_char(&ds, '\n');
958 unixctl_command_reply(conn, 200, ds_cstr(&ds));
963 bond_unixctl_show(struct unixctl_conn *conn,
964 const char *args, void *aux OVS_UNUSED)
966 struct ds ds = DS_EMPTY_INITIALIZER;
967 const struct bond_slave *slave;
968 const struct bond *bond;
970 bond = bond_find(args);
972 unixctl_command_reply(conn, 501, "no such bond");
976 ds_put_format(&ds, "bond_mode: %s\n",
977 bond_mode_to_string(bond->balance));
979 if (bond->balance != BM_AB) {
980 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
981 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
984 ds_put_format(&ds, "bond-detect-mode: %s\n",
985 bond->monitor ? "carrier" : "miimon");
987 if (!bond->monitor) {
988 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
989 bond->miimon_interval);
992 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
993 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
995 if (bond_is_balanced(bond)) {
996 ds_put_format(&ds, "next rebalance: %lld ms\n",
997 bond->next_rebalance - time_msec());
1000 ds_put_format(&ds, "lacp_negotiated: %s\n",
1001 bond->lacp_negotiated ? "true" : "false");
1003 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1004 struct bond_entry *be;
1008 ds_put_format(&ds, "\nslave %s: %s\n",
1009 slave->name, slave->enabled ? "enabled" : "disabled");
1010 if (slave == bond->active_slave) {
1011 ds_put_cstr(&ds, "\tactive slave\n");
1013 if (slave->delay_expires != LLONG_MAX) {
1014 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1015 slave->enabled ? "downdelay" : "updelay",
1016 slave->delay_expires - time_msec());
1019 ds_put_format(&ds, "\tlacp_may_enable: %s\n",
1020 slave->lacp_may_enable ? "true" : "false");
1022 if (!bond_is_balanced(bond)) {
1027 memset(&flow, 0, sizeof flow);
1028 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1029 int hash = be - bond->hash;
1031 if (be->slave != slave) {
1035 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1036 hash, be->tx_bytes / 1024);
1038 if (bond->balance != BM_SLB) {
1042 /* XXX How can we list the MACs assigned to hashes? */
1045 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1050 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1051 void *aux OVS_UNUSED)
1053 char *args = (char *) args_;
1054 char *save_ptr = NULL;
1055 char *bond_s, *hash_s, *slave_s;
1057 struct bond_slave *slave;
1058 struct bond_entry *entry;
1061 bond_s = strtok_r(args, " ", &save_ptr);
1062 hash_s = strtok_r(NULL, " ", &save_ptr);
1063 slave_s = strtok_r(NULL, " ", &save_ptr);
1065 unixctl_command_reply(conn, 501,
1066 "usage: bond/migrate BOND HASH SLAVE");
1070 bond = bond_find(bond_s);
1072 unixctl_command_reply(conn, 501, "no such bond");
1076 if (bond->balance != BM_SLB) {
1077 unixctl_command_reply(conn, 501, "not an SLB bond");
1081 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1082 hash = atoi(hash_s) & BOND_MASK;
1084 unixctl_command_reply(conn, 501, "bad hash");
1088 slave = bond_lookup_slave(bond, slave_s);
1090 unixctl_command_reply(conn, 501, "no such slave");
1094 if (!slave->enabled) {
1095 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1099 entry = &bond->hash[hash];
1100 tag_set_add(&bond->unixctl_tags, entry->tag);
1101 entry->slave = slave;
1102 entry->tag = tag_create_random();
1103 unixctl_command_reply(conn, 200, "migrated");
1107 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1108 void *aux OVS_UNUSED)
1110 char *args = (char *) args_;
1111 char *save_ptr = NULL;
1112 char *bond_s, *slave_s;
1114 struct bond_slave *slave;
1116 bond_s = strtok_r(args, " ", &save_ptr);
1117 slave_s = strtok_r(NULL, " ", &save_ptr);
1119 unixctl_command_reply(conn, 501,
1120 "usage: bond/set-active-slave BOND SLAVE");
1124 bond = bond_find(bond_s);
1126 unixctl_command_reply(conn, 501, "no such bond");
1130 slave = bond_lookup_slave(bond, slave_s);
1132 unixctl_command_reply(conn, 501, "no such slave");
1136 if (!slave->enabled) {
1137 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1141 if (bond->active_slave != slave) {
1142 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1143 bond->active_slave = slave;
1144 bond->active_slave->tag = tag_create_random();
1145 VLOG_INFO("bond %s: active interface is now %s",
1146 bond->name, slave->name);
1147 bond->send_learning_packets = true;
1148 unixctl_command_reply(conn, 200, "done");
1150 unixctl_command_reply(conn, 200, "no change");
1155 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1157 char *args = (char *) args_;
1158 char *save_ptr = NULL;
1159 char *bond_s, *slave_s;
1161 struct bond_slave *slave;
1163 bond_s = strtok_r(args, " ", &save_ptr);
1164 slave_s = strtok_r(NULL, " ", &save_ptr);
1166 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1167 enable ? "enable" : "disable");
1168 unixctl_command_reply(conn, 501, usage);
1173 bond = bond_find(bond_s);
1175 unixctl_command_reply(conn, 501, "no such bond");
1179 slave = bond_lookup_slave(bond, slave_s);
1181 unixctl_command_reply(conn, 501, "no such slave");
1185 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1186 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1190 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1191 void *aux OVS_UNUSED)
1193 enable_slave(conn, args, true);
1197 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1198 void *aux OVS_UNUSED)
1200 enable_slave(conn, args, false);
1204 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1205 void *aux OVS_UNUSED)
1207 char *args = (char *) args_;
1208 uint8_t mac[ETH_ADDR_LEN];
1212 char *mac_s, *vlan_s;
1213 char *save_ptr = NULL;
1215 mac_s = strtok_r(args, " ", &save_ptr);
1216 vlan_s = strtok_r(NULL, " ", &save_ptr);
1219 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1220 unixctl_command_reply(conn, 501, "invalid vlan");
1224 vlan = OFP_VLAN_NONE;
1227 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1228 == ETH_ADDR_SCAN_COUNT) {
1229 hash = bond_hash_src(mac, vlan) & BOND_MASK;
1231 hash_cstr = xasprintf("%u", hash);
1232 unixctl_command_reply(conn, 200, hash_cstr);
1235 unixctl_command_reply(conn, 501, "invalid mac");
1242 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1243 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1244 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1245 unixctl_command_register("bond/set-active-slave",
1246 bond_unixctl_set_active_slave, NULL);
1247 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1249 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1251 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1255 bond_entry_reset(struct bond *bond)
1257 if (bond->balance != BM_AB) {
1258 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1261 bond->hash = xmalloc(hash_len);
1263 memset(bond->hash, 0, hash_len);
1265 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1272 static struct bond_slave *
1273 bond_slave_lookup(struct bond *bond, const void *slave_)
1275 struct bond_slave *slave;
1277 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1279 if (slave->aux == slave_) {
1288 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1290 return (bond->detect == BLSM_CARRIER
1291 ? netdev_get_carrier(netdev)
1292 : netdev_get_miimon(netdev));
1296 bond_stb_sort_cmp__(const void *a_, const void *b_)
1298 const struct bond_slave *const *ap = a_;
1299 const struct bond_slave *const *bp = b_;
1300 const struct bond_slave *a = *ap;
1301 const struct bond_slave *b = *bp;
1302 uint16_t aid = a->stb_id;
1303 uint16_t bid = b->stb_id;
1305 return aid < bid ? -1 : aid > bid;
1309 bond_stb_sort(struct bond *bond)
1313 if (!bond->stb_slaves || !bond->stb_need_sort) {
1316 bond->stb_need_sort = false;
1318 qsort(bond->stb_slaves, bond->n_stb_slaves, sizeof *bond->stb_slaves,
1319 bond_stb_sort_cmp__);
1321 for (i = 0; i < bond->n_stb_slaves; i++) {
1322 bond->stb_slaves[i]->stb_idx = i;
1329 bond_stb_enable_slave(struct bond_slave *slave)
1331 struct bond *bond = slave->bond;
1333 if (bond->balance != BM_STABLE) {
1337 bond->stb_need_sort = true;
1339 if (slave->enabled) {
1340 if (bond->len_stb_slaves <= bond->n_stb_slaves) {
1341 bond->stb_slaves = x2nrealloc(bond->stb_slaves,
1342 &bond->len_stb_slaves,
1343 sizeof *bond->stb_slaves);
1346 slave->stb_idx = bond->n_stb_slaves++;
1347 bond->stb_slaves[slave->stb_idx] = slave;
1349 size_t index = slave->stb_idx;
1350 bond->stb_slaves[index] = bond->stb_slaves[--bond->n_stb_slaves];
1351 bond->stb_slaves[index]->stb_idx = index;
1356 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1358 slave->delay_expires = LLONG_MAX;
1359 if (enable != slave->enabled) {
1360 slave->enabled = enable;
1361 if (!slave->enabled) {
1362 VLOG_WARN("interface %s: disabled", slave->name);
1364 tag_set_add(tags, slave->tag);
1367 VLOG_WARN("interface %s: enabled", slave->name);
1368 slave->tag = tag_create_random();
1370 bond_stb_enable_slave(slave);
1375 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1377 struct bond *bond = slave->bond;
1380 up = slave->up && slave->lacp_may_enable;
1381 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1382 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1383 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1384 slave->name, up ? "up" : "down");
1385 if (up == slave->enabled) {
1386 slave->delay_expires = LLONG_MAX;
1387 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1388 slave->name, up ? "disabled" : "enabled");
1390 int delay = (bond->lacp_negotiated ? 0
1391 : up ? bond->updelay : bond->downdelay);
1392 slave->delay_expires = time_msec() + delay;
1394 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1397 up ? "enabled" : "disabled",
1404 if (time_msec() >= slave->delay_expires) {
1405 bond_enable_slave(slave, up, tags);
1410 bond_is_tcp_hash(const struct bond *bond)
1412 return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
1413 && bond->lacp_negotiated;
1417 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan)
1419 return hash_bytes(mac, ETH_ADDR_LEN, vlan);
1423 bond_hash_tcp(const struct flow *flow, uint16_t vlan)
1425 struct flow hash_flow = *flow;
1426 hash_flow.vlan_tci = vlan;
1428 /* The symmetric quality of this hash function is not required, but
1429 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1430 * purposes, so we use it out of convenience. */
1431 return flow_hash_symmetric_l4(&hash_flow, 0);
1435 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1437 assert(bond->balance != BM_AB);
1439 return (bond_is_tcp_hash(bond)
1440 ? bond_hash_tcp(flow, vlan)
1441 : bond_hash_src(flow->dl_src, vlan));
1444 static struct bond_entry *
1445 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1448 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1451 static struct bond_slave *
1452 choose_output_slave(const struct bond *bond, const struct flow *flow,
1455 struct bond_entry *e;
1457 switch (bond->balance) {
1459 return bond->active_slave;
1462 if (bond->n_stb_slaves) {
1463 return bond->stb_slaves[bond_hash(bond, flow, vlan)
1464 % bond->n_stb_slaves];
1466 return bond->active_slave;
1471 e = lookup_bond_entry(bond, flow, vlan);
1472 if (!e->slave || !e->slave->enabled) {
1473 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1474 struct bond_slave, hmap_node);
1475 if (!e->slave->enabled) {
1476 e->slave = bond->active_slave;
1478 e->tag = tag_create_random();
1487 static struct bond_slave *
1488 bond_choose_slave(const struct bond *bond)
1490 struct bond_slave *slave, *best;
1492 /* Find an enabled slave. */
1493 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1494 if (slave->enabled) {
1499 /* All interfaces are disabled. Find an interface that will be enabled
1500 * after its updelay expires. */
1502 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1503 if (slave->delay_expires != LLONG_MAX
1504 && slave->lacp_may_enable
1505 && (!best || slave->delay_expires < best->delay_expires)) {
1513 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1515 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1516 struct bond_slave *old_active_slave = bond->active_slave;
1518 bond->active_slave = bond_choose_slave(bond);
1519 if (bond->active_slave) {
1520 if (bond->active_slave->enabled) {
1521 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1522 bond->name, bond->active_slave->name);
1524 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1525 "remaining %lld ms updelay (since no interface was "
1526 "enabled)", bond->name, bond->active_slave->name,
1527 bond->active_slave->delay_expires - time_msec());
1528 bond_enable_slave(bond->active_slave, true, tags);
1531 if (!old_active_slave) {
1532 tag_set_add(tags, bond->no_slaves_tag);
1535 bond->send_learning_packets = true;
1536 } else if (old_active_slave) {
1537 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1541 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1542 * there is no active slave. */
1544 bond_get_active_slave_tag(const struct bond *bond)
1546 return (bond->active_slave
1547 ? bond->active_slave->tag
1548 : bond->no_slaves_tag);
1551 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1552 * bond interface. */
1554 bond_update_fake_slave_stats(struct bond *bond)
1556 struct netdev_stats bond_stats;
1557 struct bond_slave *slave;
1558 struct netdev *bond_dev;
1560 memset(&bond_stats, 0, sizeof bond_stats);
1562 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1563 struct netdev_stats slave_stats;
1565 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1566 /* XXX: We swap the stats here because they are swapped back when
1567 * reported by the internal device. The reason for this is
1568 * internal devices normally represent packets going into the
1569 * system but when used as fake bond device they represent packets
1570 * leaving the system. We really should do this in the internal
1571 * device itself because changing it here reverses the counts from
1572 * the perspective of the switch. However, the internal device
1573 * doesn't know what type of device it represents so we have to do
1574 * it here for now. */
1575 bond_stats.tx_packets += slave_stats.rx_packets;
1576 bond_stats.tx_bytes += slave_stats.rx_bytes;
1577 bond_stats.rx_packets += slave_stats.tx_packets;
1578 bond_stats.rx_bytes += slave_stats.tx_bytes;
1582 if (!netdev_open_default(bond->name, &bond_dev)) {
1583 netdev_set_stats(bond_dev, &bond_stats);
1584 netdev_close(bond_dev);