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. */
80 /* A bond, that is, a set of network devices grouped to improve performance or
83 struct hmap_node hmap_node; /* In 'all_bonds' hmap. */
84 char *name; /* Name provided by client. */
90 enum bond_mode balance; /* Balancing mode, one of BM_*. */
91 struct bond_slave *active_slave;
92 tag_type no_slaves_tag; /* Tag for flows when all slaves disabled. */
93 int updelay, downdelay; /* Delay before slave goes up/down, in ms. */
94 bool lacp_negotiated; /* LACP negotiations were successful. */
95 bool bond_revalidate; /* True if flows need revalidation. */
96 uint32_t basis; /* Basis for flow hash function. */
98 /* SLB specific bonding info. */
99 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
100 int rebalance_interval; /* Interval between rebalances, in ms. */
101 long long int next_rebalance; /* Next rebalancing time. */
102 bool send_learning_packets;
104 /* BM_STABLE specific bonding info. */
105 tag_type stb_tag; /* Tag associated with this bond. */
108 enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
109 struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
110 long long int miimon_interval; /* Miimon status refresh interval. */
111 long long int miimon_next_update; /* Time of next miimon update. */
113 /* Legacy compatibility. */
114 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
116 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
117 * where we can't otherwise provide revalidation feedback to the client.
118 * That's only unixctl commands now; I hope no other cases will arise. */
119 struct tag_set unixctl_tags;
122 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
124 static void bond_entry_reset(struct bond *);
125 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
126 static bool bond_is_link_up(struct bond *, struct netdev *);
127 static void bond_enable_slave(struct bond_slave *, bool enable,
129 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
130 static void bond_choose_active_slave(struct bond *, struct tag_set *);
131 static bool bond_is_tcp_hash(const struct bond *);
132 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
133 uint16_t vlan, uint32_t basis);
134 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan,
136 static struct bond_entry *lookup_bond_entry(const struct bond *,
139 static tag_type bond_get_active_slave_tag(const struct bond *);
140 static struct bond_slave *choose_output_slave(const struct bond *,
143 static void bond_update_fake_slave_stats(struct bond *);
145 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
146 * stores the mode in '*balance' and returns true. Otherwise returns false
147 * without modifying '*balance'. */
149 bond_mode_from_string(enum bond_mode *balance, const char *s)
151 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
153 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
155 } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
156 *balance = BM_STABLE;
157 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
165 /* Returns a string representing 'balance'. */
167 bond_mode_to_string(enum bond_mode balance) {
170 return "balance-tcp";
172 return "balance-slb";
176 return "active-backup";
181 /* Attempts to parse 's' as the name of a bond link status detection mode. If
182 * successful, stores the mode in '*detect' and returns true. Otherwise
183 * returns false without modifying '*detect'. */
185 bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
187 if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
188 *detect = BLSM_CARRIER;
189 } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
190 *detect = BLSM_MIIMON;
197 /* Returns a string representing 'detect'. */
199 bond_detect_mode_to_string(enum bond_detect_mode detect)
210 /* Creates and returns a new bond whose configuration is initially taken from
213 * The caller should register each slave on the new bond by calling
214 * bond_slave_register(). */
216 bond_create(const struct bond_settings *s)
220 bond = xzalloc(sizeof *bond);
221 hmap_init(&bond->slaves);
222 bond->no_slaves_tag = tag_create_random();
223 bond->stb_tag = tag_create_random();
224 bond->miimon_next_update = LLONG_MAX;
225 bond->next_fake_iface_update = LLONG_MAX;
227 bond_reconfigure(bond, s);
229 tag_set_init(&bond->unixctl_tags);
236 bond_destroy(struct bond *bond)
238 struct bond_slave *slave, *next_slave;
244 hmap_remove(&all_bonds, &bond->hmap_node);
246 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
247 hmap_remove(&bond->slaves, &slave->hmap_node);
248 /* Client owns 'slave->netdev'. */
252 hmap_destroy(&bond->slaves);
256 netdev_monitor_destroy(bond->monitor);
262 /* Updates 'bond''s overall configuration to 's'.
264 * The caller should register each slave on 'bond' by calling
265 * bond_slave_register(). This is optional if none of the slaves'
266 * configuration has changed. In any case it can't hurt.
268 * Returns true if the configuration has changed in such a way that requires
272 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
274 bool revalidate = false;
276 if (!bond->name || strcmp(bond->name, s->name)) {
278 hmap_remove(&all_bonds, &bond->hmap_node);
281 bond->name = xstrdup(s->name);
282 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
285 bond->detect = s->detect;
286 bond->miimon_interval = s->miimon_interval;
287 bond->updelay = s->up_delay;
288 bond->downdelay = s->down_delay;
289 bond->rebalance_interval = s->rebalance_interval;
291 if (bond->balance != s->balance) {
292 bond->balance = s->balance;
296 if (bond->basis != s->basis) {
297 bond->basis = s->basis;
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->bond_revalidate) {
330 bond->bond_revalidate = false;
333 if (bond->balance == BM_AB || !bond->hash || revalidate) {
334 bond_entry_reset(bond);
340 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
341 * arbitrary client-provided pointer that uniquely identifies a slave within a
342 * bond. If 'slave_' already exists within 'bond' then this function
343 * reconfigures the existing slave.
345 * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
346 * across restarts and distributed vswitch instances. It should be unique per
347 * slave, and preferably consistent across restarts and reconfigurations.
349 * 'netdev' must be the network device that 'slave_' represents. It is owned
350 * by the client, so the client must not close it before either unregistering
351 * 'slave_' or destroying 'bond'.
354 bond_slave_register(struct bond *bond, void *slave_, uint16_t stb_id,
355 struct netdev *netdev)
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->name = xstrdup(netdev_get_name(netdev));
368 bond->bond_revalidate = true;
370 slave->enabled = false;
371 bond_enable_slave(slave, slave->up, NULL);
374 if (slave->stb_id != stb_id) {
375 slave->stb_id = stb_id;
376 bond->bond_revalidate = true;
379 if (slave->netdev != netdev) {
382 netdev_monitor_remove(bond->monitor, slave->netdev);
384 netdev_monitor_add(bond->monitor, netdev);
386 slave->netdev = netdev;
390 slave->name = xstrdup(netdev_get_name(netdev));
393 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
394 * then this function has no effect.
396 * Unregistering a slave invalidates all flows. */
398 bond_slave_unregister(struct bond *bond, const void *slave_)
400 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
408 netdev_monitor_remove(bond->monitor, slave->netdev);
411 bond_enable_slave(slave, false, NULL);
413 del_active = bond->active_slave == slave;
415 struct bond_entry *e;
416 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
417 if (e->slave == slave) {
425 hmap_remove(&bond->slaves, &slave->hmap_node);
426 /* Client owns 'slave->netdev'. */
433 bond_choose_active_slave(bond, &tags);
434 bond->send_learning_packets = true;
438 /* Should be called on each slave in 'bond' before bond_run() to indicate the
439 * results of lacp_slave_may_enable() on 'slave_'. */
441 bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
444 bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
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, bool lacp_negotiated)
454 struct bond_slave *slave;
455 bool is_tcp_hash = bond_is_tcp_hash(bond);
457 bond->lacp_negotiated = lacp_negotiated;
459 /* Update link status. */
460 if (bond->detect == BLSM_CARRIER
461 || time_msec() >= bond->miimon_next_update)
463 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
464 slave->up = bond_is_link_up(bond, slave->netdev);
466 bond->miimon_next_update = time_msec() + bond->miimon_interval;
469 /* Enable slaves based on link status and LACP feedback. */
470 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
471 bond_link_status_update(slave, tags);
473 if (!bond->active_slave || !bond->active_slave->enabled) {
474 bond_choose_active_slave(bond, tags);
477 /* Update fake bond interface stats. */
478 if (time_msec() >= bond->next_fake_iface_update) {
479 bond_update_fake_slave_stats(bond);
480 bond->next_fake_iface_update = time_msec() + 1000;
483 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
484 bond->bond_revalidate = true;
487 if (bond->bond_revalidate) {
488 bond->bond_revalidate = false;
490 bond_entry_reset(bond);
491 if (bond->balance != BM_STABLE) {
492 struct bond_slave *slave;
494 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
495 tag_set_add(tags, slave->tag);
498 tag_set_add(tags, bond->stb_tag);
500 tag_set_add(tags, bond->no_slaves_tag);
503 /* Invalidate any tags required by */
504 tag_set_union(tags, &bond->unixctl_tags);
505 tag_set_init(&bond->unixctl_tags);
508 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
510 bond_wait(struct bond *bond)
512 struct bond_slave *slave;
514 if (bond->detect == BLSM_CARRIER) {
515 netdev_monitor_poll_wait(bond->monitor);
517 poll_timer_wait_until(bond->miimon_next_update);
520 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
521 if (slave->delay_expires != LLONG_MAX) {
522 poll_timer_wait_until(slave->delay_expires);
526 if (bond->next_fake_iface_update != LLONG_MAX) {
527 poll_timer_wait_until(bond->next_fake_iface_update);
530 /* Ensure that any saved tags get revalidated right away. */
531 if (!tag_set_is_empty(&bond->unixctl_tags)) {
532 poll_immediate_wake();
535 /* We don't wait for bond->next_rebalance because rebalancing can only run
536 * at a flow account checkpoint. ofproto does checkpointing on its own
537 * schedule and bond_rebalance() gets called afterward, so we'd just be
538 * waking up for no purpose. */
541 /* MAC learning table interaction. */
544 may_send_learning_packets(const struct bond *bond)
546 return !bond->lacp_negotiated && bond->balance != BM_AB;
549 /* Returns true if 'bond' needs the client to send out packets to assist with
550 * MAC learning on 'bond'. If this function returns true, then the client
551 * should iterate through its MAC learning table for the bridge on which 'bond'
552 * is located. For each MAC that has been learned on a port other than 'bond',
553 * it should call bond_send_learning_packet().
555 * This function will only return true if 'bond' is in SLB mode and LACP is not
556 * negotiated. Otherwise sending learning packets isn't necessary.
558 * Calling this function resets the state that it checks. */
560 bond_should_send_learning_packets(struct bond *bond)
562 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
563 bond->send_learning_packets = false;
567 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
569 * See bond_should_send_learning_packets() for description of usage. */
571 bond_send_learning_packet(struct bond *bond,
572 const uint8_t eth_src[ETH_ADDR_LEN],
575 struct bond_slave *slave;
576 struct ofpbuf packet;
580 assert(may_send_learning_packets(bond));
581 if (!bond->active_slave) {
582 /* Nowhere to send the learning packet. */
586 memset(&flow, 0, sizeof flow);
587 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
588 slave = choose_output_slave(bond, &flow, vlan);
590 ofpbuf_init(&packet, 0);
591 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
594 eth_set_vlan_tci(&packet, htons(vlan));
596 error = netdev_send(slave->netdev, &packet);
597 ofpbuf_uninit(&packet);
602 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
603 * Ethernet destination address of 'eth_dst', should be admitted.
605 * The return value is one of the following:
607 * - BV_ACCEPT: Admit the packet.
609 * - BV_DROP: Drop the packet.
611 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
612 * Ethernet source address and VLAN. If there is none, or if the packet
613 * is on the learned port, then admit the packet. If a different port has
614 * been learned, however, drop the packet (and do not use it for MAC
618 bond_check_admissibility(struct bond *bond, const void *slave_,
619 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
621 /* Admit all packets if LACP has been negotiated, because that means that
622 * the remote switch is aware of the bond and will "do the right thing". */
623 if (bond->lacp_negotiated) {
627 /* Drop all multicast packets on inactive slaves. */
628 if (eth_addr_is_multicast(eth_dst)) {
629 *tags |= bond_get_active_slave_tag(bond);
630 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
635 /* Drop all packets for which we have learned a different input port,
636 * because we probably sent the packet on one slave and got it back on the
637 * other. Gratuitous ARP packets are an exception to this rule: the host
638 * has moved to another switch. The exception to the exception is if we
639 * locked the learning table to avoid reflections on bond slaves. */
640 return BV_DROP_IF_MOVED;
643 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
644 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
645 * NULL if the packet should be dropped because no slaves are enabled.
647 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
648 * should be a VID only (i.e. excluding the PCP bits). Second,
649 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
650 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
651 * packet belongs to (so for an access port it will be the access port's VLAN).
653 * Adds a tag to '*tags' that associates the flow with the returned slave.
656 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
657 uint16_t vlan, tag_type *tags)
659 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
661 *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
664 *tags |= bond->no_slaves_tag;
672 bond_is_balanced(const struct bond *bond)
674 return bond->balance == BM_SLB || bond->balance == BM_TCP;
677 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
679 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
683 if (bond_is_balanced(bond)) {
684 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
688 static struct bond_slave *
689 bond_slave_from_bal_node(struct list *bal)
691 return CONTAINER_OF(bal, struct bond_slave, bal_node);
695 log_bals(struct bond *bond, const struct list *bals)
697 if (VLOG_IS_DBG_ENABLED()) {
698 struct ds ds = DS_EMPTY_INITIALIZER;
699 const struct bond_slave *slave;
701 LIST_FOR_EACH (slave, bal_node, bals) {
703 ds_put_char(&ds, ',');
705 ds_put_format(&ds, " %s %"PRIu64"kB",
706 slave->name, slave->tx_bytes / 1024);
708 if (!slave->enabled) {
709 ds_put_cstr(&ds, " (disabled)");
711 if (!list_is_empty(&slave->entries)) {
712 struct bond_entry *e;
714 ds_put_cstr(&ds, " (");
715 LIST_FOR_EACH (e, list_node, &slave->entries) {
716 if (&e->list_node != list_front(&slave->entries)) {
717 ds_put_cstr(&ds, " + ");
719 ds_put_format(&ds, "h%td: %"PRIu64"kB",
720 e - bond->hash, e->tx_bytes / 1024);
722 ds_put_cstr(&ds, ")");
725 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
730 /* Shifts 'hash' from its current slave to 'to'. */
732 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
735 struct bond_slave *from = hash->slave;
736 struct bond *bond = from->bond;
737 uint64_t delta = hash->tx_bytes;
739 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
740 "from %s to %s (now carrying %"PRIu64"kB and "
741 "%"PRIu64"kB load, respectively)",
742 bond->name, delta / 1024, hash - bond->hash,
743 from->name, to->name,
744 (from->tx_bytes - delta) / 1024,
745 (to->tx_bytes + delta) / 1024);
747 /* Shift load away from 'from' to 'to'. */
748 from->tx_bytes -= delta;
749 to->tx_bytes += delta;
751 /* Arrange for flows to be revalidated. */
752 tag_set_add(set, hash->tag);
754 hash->tag = tag_create_random();
757 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
758 * given that doing so must decrease the ratio of the load on the two slaves by
759 * at least 0.1. Returns NULL if there is no appropriate entry.
761 * The list of entries isn't sorted. I don't know of a reason to prefer to
762 * shift away small hashes or large hashes. */
763 static struct bond_entry *
764 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
766 struct bond_entry *e;
768 if (list_is_short(&from->entries)) {
769 /* 'from' carries no more than one MAC hash, so shifting load away from
770 * it would be pointless. */
774 LIST_FOR_EACH (e, list_node, &from->entries) {
775 double old_ratio, new_ratio;
778 if (to_tx_bytes == 0) {
779 /* Nothing on the new slave, move it. */
784 old_ratio = (double)from->tx_bytes / to_tx_bytes;
785 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
786 if (old_ratio - new_ratio > 0.1) {
787 /* Would decrease the ratio, move it. */
795 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
798 insert_bal(struct list *bals, struct bond_slave *slave)
800 struct bond_slave *pos;
802 LIST_FOR_EACH (pos, bal_node, bals) {
803 if (slave->tx_bytes > pos->tx_bytes) {
807 list_insert(&pos->bal_node, &slave->bal_node);
810 /* Removes 'slave' from its current list and then inserts it into 'bals' so
811 * that descending order of 'tx_bytes' is maintained. */
813 reinsert_bal(struct list *bals, struct bond_slave *slave)
815 list_remove(&slave->bal_node);
816 insert_bal(bals, slave);
819 /* If 'bond' needs rebalancing, does so.
821 * The caller should have called bond_account() for each active flow, to ensure
822 * that flow data is consistently accounted at this point. */
824 bond_rebalance(struct bond *bond, struct tag_set *tags)
826 struct bond_slave *slave;
827 struct bond_entry *e;
830 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
833 bond->next_rebalance = time_msec() + bond->rebalance_interval;
835 /* Add each bond_entry to its slave's 'entries' list.
836 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
837 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
839 list_init(&slave->entries);
841 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
842 if (e->slave && e->tx_bytes) {
843 e->slave->tx_bytes += e->tx_bytes;
844 list_push_back(&e->slave->entries, &e->list_node);
848 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
850 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
851 * with a proper list sort algorithm. */
853 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
854 if (slave->enabled) {
855 insert_bal(&bals, slave);
858 log_bals(bond, &bals);
860 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
861 while (!list_is_short(&bals)) {
862 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
863 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
866 overload = from->tx_bytes - to->tx_bytes;
867 if (overload < to->tx_bytes >> 5 || overload < 100000) {
868 /* The extra load on 'from' (and all less-loaded slaves), compared
869 * to that of 'to' (the least-loaded slave), is less than ~3%, or
870 * it is less than ~1Mbps. No point in rebalancing. */
874 /* 'from' is carrying significantly more load than 'to', and that load
875 * is split across at least two different hashes. */
876 e = choose_entry_to_migrate(from, to->tx_bytes);
878 bond_shift_load(e, to, tags);
880 /* Delete element from from->entries.
882 * We don't add the element to to->hashes. That would only allow
883 * 'e' to be migrated to another slave in this rebalancing run, and
884 * there is no point in doing that. */
885 list_remove(&e->list_node);
887 /* Re-sort 'bals'. */
888 reinsert_bal(&bals, from);
889 reinsert_bal(&bals, to);
891 /* Can't usefully migrate anything away from 'from'.
892 * Don't reconsider it. */
893 list_remove(&from->bal_node);
897 /* Implement exponentially weighted moving average. A weight of 1/2 causes
898 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
899 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
900 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
908 /* Bonding unixctl user interface functions. */
911 bond_find(const char *name)
915 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
917 if (!strcmp(bond->name, name)) {
924 static struct bond_slave *
925 bond_lookup_slave(struct bond *bond, const char *slave_name)
927 struct bond_slave *slave;
929 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
930 if (!strcmp(slave->name, slave_name)) {
938 bond_unixctl_list(struct unixctl_conn *conn,
939 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
941 struct ds ds = DS_EMPTY_INITIALIZER;
942 const struct bond *bond;
944 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
946 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
947 const struct bond_slave *slave;
950 ds_put_format(&ds, "%s\t%s\t",
951 bond->name, bond_mode_to_string(bond->balance));
954 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
956 ds_put_cstr(&ds, ", ");
958 ds_put_cstr(&ds, slave->name);
960 ds_put_char(&ds, '\n');
962 unixctl_command_reply(conn, 200, ds_cstr(&ds));
967 bond_unixctl_show(struct unixctl_conn *conn,
968 const char *args, void *aux OVS_UNUSED)
970 struct ds ds = DS_EMPTY_INITIALIZER;
971 const struct bond_slave *slave;
972 const struct bond *bond;
974 bond = bond_find(args);
976 unixctl_command_reply(conn, 501, "no such bond");
980 ds_put_format(&ds, "bond_mode: %s\n",
981 bond_mode_to_string(bond->balance));
983 if (bond->balance != BM_AB) {
984 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
985 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
988 ds_put_format(&ds, "bond-hash-basis: %"PRIu32"\n", bond->basis);
990 ds_put_format(&ds, "bond-detect-mode: %s\n",
991 bond->monitor ? "carrier" : "miimon");
993 if (!bond->monitor) {
994 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
995 bond->miimon_interval);
998 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
999 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
1001 if (bond_is_balanced(bond)) {
1002 ds_put_format(&ds, "next rebalance: %lld ms\n",
1003 bond->next_rebalance - time_msec());
1006 ds_put_format(&ds, "lacp_negotiated: %s\n",
1007 bond->lacp_negotiated ? "true" : "false");
1009 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1010 struct bond_entry *be;
1014 ds_put_format(&ds, "\nslave %s: %s\n",
1015 slave->name, slave->enabled ? "enabled" : "disabled");
1016 if (slave == bond->active_slave) {
1017 ds_put_cstr(&ds, "\tactive slave\n");
1019 if (slave->delay_expires != LLONG_MAX) {
1020 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1021 slave->enabled ? "downdelay" : "updelay",
1022 slave->delay_expires - time_msec());
1025 ds_put_format(&ds, "\tlacp_may_enable: %s\n",
1026 slave->lacp_may_enable ? "true" : "false");
1028 if (!bond_is_balanced(bond)) {
1033 memset(&flow, 0, sizeof flow);
1034 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1035 int hash = be - bond->hash;
1037 if (be->slave != slave) {
1041 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1042 hash, be->tx_bytes / 1024);
1044 if (bond->balance != BM_SLB) {
1048 /* XXX How can we list the MACs assigned to hashes? */
1051 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1056 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1057 void *aux OVS_UNUSED)
1059 char *args = (char *) args_;
1060 char *save_ptr = NULL;
1061 char *bond_s, *hash_s, *slave_s;
1063 struct bond_slave *slave;
1064 struct bond_entry *entry;
1067 bond_s = strtok_r(args, " ", &save_ptr);
1068 hash_s = strtok_r(NULL, " ", &save_ptr);
1069 slave_s = strtok_r(NULL, " ", &save_ptr);
1071 unixctl_command_reply(conn, 501,
1072 "usage: bond/migrate BOND HASH SLAVE");
1076 bond = bond_find(bond_s);
1078 unixctl_command_reply(conn, 501, "no such bond");
1082 if (bond->balance != BM_SLB) {
1083 unixctl_command_reply(conn, 501, "not an SLB bond");
1087 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1088 hash = atoi(hash_s) & BOND_MASK;
1090 unixctl_command_reply(conn, 501, "bad hash");
1094 slave = bond_lookup_slave(bond, slave_s);
1096 unixctl_command_reply(conn, 501, "no such slave");
1100 if (!slave->enabled) {
1101 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1105 entry = &bond->hash[hash];
1106 tag_set_add(&bond->unixctl_tags, entry->tag);
1107 entry->slave = slave;
1108 entry->tag = tag_create_random();
1109 unixctl_command_reply(conn, 200, "migrated");
1113 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1114 void *aux OVS_UNUSED)
1116 char *args = (char *) args_;
1117 char *save_ptr = NULL;
1118 char *bond_s, *slave_s;
1120 struct bond_slave *slave;
1122 bond_s = strtok_r(args, " ", &save_ptr);
1123 slave_s = strtok_r(NULL, " ", &save_ptr);
1125 unixctl_command_reply(conn, 501,
1126 "usage: bond/set-active-slave BOND SLAVE");
1130 bond = bond_find(bond_s);
1132 unixctl_command_reply(conn, 501, "no such bond");
1136 slave = bond_lookup_slave(bond, slave_s);
1138 unixctl_command_reply(conn, 501, "no such slave");
1142 if (!slave->enabled) {
1143 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1147 if (bond->active_slave != slave) {
1148 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1149 bond->active_slave = slave;
1150 bond->active_slave->tag = tag_create_random();
1151 VLOG_INFO("bond %s: active interface is now %s",
1152 bond->name, slave->name);
1153 bond->send_learning_packets = true;
1154 unixctl_command_reply(conn, 200, "done");
1156 unixctl_command_reply(conn, 200, "no change");
1161 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1163 char *args = (char *) args_;
1164 char *save_ptr = NULL;
1165 char *bond_s, *slave_s;
1167 struct bond_slave *slave;
1169 bond_s = strtok_r(args, " ", &save_ptr);
1170 slave_s = strtok_r(NULL, " ", &save_ptr);
1172 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1173 enable ? "enable" : "disable");
1174 unixctl_command_reply(conn, 501, usage);
1179 bond = bond_find(bond_s);
1181 unixctl_command_reply(conn, 501, "no such bond");
1185 slave = bond_lookup_slave(bond, slave_s);
1187 unixctl_command_reply(conn, 501, "no such slave");
1191 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1192 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1196 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1197 void *aux OVS_UNUSED)
1199 enable_slave(conn, args, true);
1203 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1204 void *aux OVS_UNUSED)
1206 enable_slave(conn, args, false);
1210 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1211 void *aux OVS_UNUSED)
1213 char *args = (char *) args_;
1214 uint8_t mac[ETH_ADDR_LEN];
1219 char *mac_s, *vlan_s, *basis_s;
1220 char *save_ptr = NULL;
1222 mac_s = strtok_r(args, " ", &save_ptr);
1223 vlan_s = strtok_r(NULL, " ", &save_ptr);
1224 basis_s = strtok_r(NULL, " ", &save_ptr);
1227 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1228 unixctl_command_reply(conn, 501, "invalid vlan");
1232 vlan = OFP_VLAN_NONE;
1236 if (sscanf(basis_s, "%"PRIu32, &basis) != 1) {
1237 unixctl_command_reply(conn, 501, "invalid basis");
1244 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1245 == ETH_ADDR_SCAN_COUNT) {
1246 hash = bond_hash_src(mac, vlan, basis) & BOND_MASK;
1248 hash_cstr = xasprintf("%u", hash);
1249 unixctl_command_reply(conn, 200, hash_cstr);
1252 unixctl_command_reply(conn, 501, "invalid mac");
1259 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1260 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1261 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1262 unixctl_command_register("bond/set-active-slave",
1263 bond_unixctl_set_active_slave, NULL);
1264 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1266 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1268 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1272 bond_entry_reset(struct bond *bond)
1274 if (bond->balance != BM_AB) {
1275 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1278 bond->hash = xmalloc(hash_len);
1280 memset(bond->hash, 0, hash_len);
1282 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1289 static struct bond_slave *
1290 bond_slave_lookup(struct bond *bond, const void *slave_)
1292 struct bond_slave *slave;
1294 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1296 if (slave->aux == slave_) {
1305 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1307 return (bond->detect == BLSM_CARRIER
1308 ? netdev_get_carrier(netdev)
1309 : netdev_get_miimon(netdev));
1313 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1315 struct bond *bond = slave->bond;
1316 slave->delay_expires = LLONG_MAX;
1317 if (enable != slave->enabled) {
1318 slave->enabled = enable;
1319 if (!slave->enabled) {
1320 VLOG_WARN("interface %s: disabled", slave->name);
1322 tag_set_add(tags, slave->tag);
1325 VLOG_WARN("interface %s: enabled", slave->name);
1326 slave->tag = tag_create_random();
1329 if (bond->balance == BM_STABLE) {
1330 bond->bond_revalidate = true;
1336 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1338 struct bond *bond = slave->bond;
1341 up = slave->up && slave->lacp_may_enable;
1342 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1343 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1344 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1345 slave->name, up ? "up" : "down");
1346 if (up == slave->enabled) {
1347 slave->delay_expires = LLONG_MAX;
1348 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1349 slave->name, up ? "disabled" : "enabled");
1351 int delay = (bond->lacp_negotiated ? 0
1352 : up ? bond->updelay : bond->downdelay);
1353 slave->delay_expires = time_msec() + delay;
1355 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1358 up ? "enabled" : "disabled",
1365 if (time_msec() >= slave->delay_expires) {
1366 bond_enable_slave(slave, up, tags);
1371 bond_is_tcp_hash(const struct bond *bond)
1373 return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
1374 && bond->lacp_negotiated;
1378 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan, uint32_t basis)
1380 return hash_3words(hash_bytes(mac, ETH_ADDR_LEN, 0), vlan, basis);
1384 bond_hash_tcp(const struct flow *flow, uint16_t vlan, uint32_t basis)
1386 struct flow hash_flow = *flow;
1387 hash_flow.vlan_tci = vlan;
1389 /* The symmetric quality of this hash function is not required, but
1390 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1391 * purposes, so we use it out of convenience. */
1392 return flow_hash_symmetric_l4(&hash_flow, basis);
1396 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1398 assert(bond->balance != BM_AB);
1400 return (bond_is_tcp_hash(bond)
1401 ? bond_hash_tcp(flow, vlan, bond->basis)
1402 : bond_hash_src(flow->dl_src, vlan, bond->basis));
1405 static struct bond_entry *
1406 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1409 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1412 /* This function uses Highest Random Weight hashing to choose an output slave.
1413 * This approach only reassigns a minimal number of flows when slaves are
1414 * enabled or disabled. Unfortunately, it has O(n) performance against the
1415 * number of slaves. There exist algorithms which are O(1), but have slightly
1416 * more complex implementations and require the use of memory. This may need
1417 * to be reimplemented if it becomes a performance bottleneck. */
1418 static struct bond_slave *
1419 choose_stb_slave(const struct bond *bond, const struct flow *flow,
1422 struct bond_slave *best, *slave;
1423 uint32_t best_hash, flow_hash;
1427 flow_hash = bond_hash(bond, flow, vlan);
1428 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1429 if (slave->enabled) {
1432 hash = hash_2words(flow_hash, slave->stb_id);
1433 if (!best || hash > best_hash) {
1443 static struct bond_slave *
1444 choose_output_slave(const struct bond *bond, const struct flow *flow,
1447 struct bond_entry *e;
1449 switch (bond->balance) {
1451 return bond->active_slave;
1454 return choose_stb_slave(bond, flow, vlan);
1457 e = lookup_bond_entry(bond, flow, vlan);
1458 if (!e->slave || !e->slave->enabled) {
1459 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1460 struct bond_slave, hmap_node);
1461 if (!e->slave->enabled) {
1462 e->slave = bond->active_slave;
1464 e->tag = tag_create_random();
1473 static struct bond_slave *
1474 bond_choose_slave(const struct bond *bond)
1476 struct bond_slave *slave, *best;
1478 /* Find an enabled slave. */
1479 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1480 if (slave->enabled) {
1485 /* All interfaces are disabled. Find an interface that will be enabled
1486 * after its updelay expires. */
1488 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1489 if (slave->delay_expires != LLONG_MAX
1490 && slave->lacp_may_enable
1491 && (!best || slave->delay_expires < best->delay_expires)) {
1499 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1501 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1502 struct bond_slave *old_active_slave = bond->active_slave;
1504 bond->active_slave = bond_choose_slave(bond);
1505 if (bond->active_slave) {
1506 if (bond->active_slave->enabled) {
1507 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1508 bond->name, bond->active_slave->name);
1510 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1511 "remaining %lld ms updelay (since no interface was "
1512 "enabled)", bond->name, bond->active_slave->name,
1513 bond->active_slave->delay_expires - time_msec());
1514 bond_enable_slave(bond->active_slave, true, tags);
1517 if (!old_active_slave) {
1518 tag_set_add(tags, bond->no_slaves_tag);
1521 bond->send_learning_packets = true;
1522 } else if (old_active_slave) {
1523 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1527 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1528 * there is no active slave. */
1530 bond_get_active_slave_tag(const struct bond *bond)
1532 return (bond->active_slave
1533 ? bond->active_slave->tag
1534 : bond->no_slaves_tag);
1537 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1538 * bond interface. */
1540 bond_update_fake_slave_stats(struct bond *bond)
1542 struct netdev_stats bond_stats;
1543 struct bond_slave *slave;
1544 struct netdev *bond_dev;
1546 memset(&bond_stats, 0, sizeof bond_stats);
1548 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1549 struct netdev_stats slave_stats;
1551 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1552 /* XXX: We swap the stats here because they are swapped back when
1553 * reported by the internal device. The reason for this is
1554 * internal devices normally represent packets going into the
1555 * system but when used as fake bond device they represent packets
1556 * leaving the system. We really should do this in the internal
1557 * device itself because changing it here reverses the counts from
1558 * the perspective of the switch. However, the internal device
1559 * doesn't know what type of device it represents so we have to do
1560 * it here for now. */
1561 bond_stats.tx_packets += slave_stats.rx_packets;
1562 bond_stats.tx_bytes += slave_stats.rx_bytes;
1563 bond_stats.rx_packets += slave_stats.tx_packets;
1564 bond_stats.rx_bytes += slave_stats.tx_bytes;
1568 if (!netdev_open_default(bond->name, &bond_dev)) {
1569 netdev_set_stats(bond_dev, &bond_stats);
1570 netdev_close(bond_dev);