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. */
97 /* SLB specific bonding info. */
98 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
99 int rebalance_interval; /* Interval between rebalances, in ms. */
100 long long int next_rebalance; /* Next rebalancing time. */
101 bool send_learning_packets;
103 /* BM_STABLE specific bonding info. */
104 tag_type stb_tag; /* Tag associated with this bond. */
107 enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
108 struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
109 long long int miimon_interval; /* Miimon status refresh interval. */
110 long long int miimon_next_update; /* Time of next miimon update. */
112 /* Legacy compatibility. */
113 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
115 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
116 * where we can't otherwise provide revalidation feedback to the client.
117 * That's only unixctl commands now; I hope no other cases will arise. */
118 struct tag_set unixctl_tags;
121 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
123 static void bond_entry_reset(struct bond *);
124 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
125 static bool bond_is_link_up(struct bond *, struct netdev *);
126 static void bond_enable_slave(struct bond_slave *, bool enable,
128 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
129 static void bond_choose_active_slave(struct bond *, struct tag_set *);
130 static bool bond_is_tcp_hash(const struct bond *);
131 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
133 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan);
134 static struct bond_entry *lookup_bond_entry(const struct bond *,
137 static tag_type bond_get_active_slave_tag(const struct bond *);
138 static struct bond_slave *choose_output_slave(const struct bond *,
141 static void bond_update_fake_slave_stats(struct bond *);
143 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
144 * stores the mode in '*balance' and returns true. Otherwise returns false
145 * without modifying '*balance'. */
147 bond_mode_from_string(enum bond_mode *balance, const char *s)
149 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
151 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
153 } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
154 *balance = BM_STABLE;
155 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
163 /* Returns a string representing 'balance'. */
165 bond_mode_to_string(enum bond_mode balance) {
168 return "balance-tcp";
170 return "balance-slb";
174 return "active-backup";
179 /* Attempts to parse 's' as the name of a bond link status detection mode. If
180 * successful, stores the mode in '*detect' and returns true. Otherwise
181 * returns false without modifying '*detect'. */
183 bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
185 if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
186 *detect = BLSM_CARRIER;
187 } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
188 *detect = BLSM_MIIMON;
195 /* Returns a string representing 'detect'. */
197 bond_detect_mode_to_string(enum bond_detect_mode detect)
208 /* Creates and returns a new bond whose configuration is initially taken from
211 * The caller should register each slave on the new bond by calling
212 * bond_slave_register(). */
214 bond_create(const struct bond_settings *s)
218 bond = xzalloc(sizeof *bond);
219 hmap_init(&bond->slaves);
220 bond->no_slaves_tag = tag_create_random();
221 bond->stb_tag = tag_create_random();
222 bond->miimon_next_update = LLONG_MAX;
223 bond->next_fake_iface_update = LLONG_MAX;
225 bond_reconfigure(bond, s);
227 tag_set_init(&bond->unixctl_tags);
234 bond_destroy(struct bond *bond)
236 struct bond_slave *slave, *next_slave;
242 hmap_remove(&all_bonds, &bond->hmap_node);
244 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
245 hmap_remove(&bond->slaves, &slave->hmap_node);
246 /* Client owns 'slave->netdev'. */
250 hmap_destroy(&bond->slaves);
254 netdev_monitor_destroy(bond->monitor);
260 /* Updates 'bond''s overall configuration to 's'.
262 * The caller should register each slave on 'bond' by calling
263 * bond_slave_register(). This is optional if none of the slaves'
264 * configuration has changed. In any case it can't hurt.
266 * Returns true if the configuration has changed in such a way that requires
270 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
272 bool revalidate = false;
274 if (!bond->name || strcmp(bond->name, s->name)) {
276 hmap_remove(&all_bonds, &bond->hmap_node);
279 bond->name = xstrdup(s->name);
280 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
283 bond->detect = s->detect;
284 bond->miimon_interval = s->miimon_interval;
285 bond->updelay = s->up_delay;
286 bond->downdelay = s->down_delay;
287 bond->rebalance_interval = s->rebalance_interval;
289 if (bond->balance != s->balance) {
290 bond->balance = s->balance;
294 if (bond->detect == BLSM_CARRIER) {
295 struct bond_slave *slave;
297 if (!bond->monitor) {
298 bond->monitor = netdev_monitor_create();
301 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
302 netdev_monitor_add(bond->monitor, slave->netdev);
305 netdev_monitor_destroy(bond->monitor);
306 bond->monitor = NULL;
308 if (bond->miimon_next_update == LLONG_MAX) {
309 bond->miimon_next_update = time_msec() + bond->miimon_interval;
314 if (bond->next_fake_iface_update == LLONG_MAX) {
315 bond->next_fake_iface_update = time_msec();
318 bond->next_fake_iface_update = LLONG_MAX;
321 if (bond->bond_revalidate) {
323 bond->bond_revalidate = false;
326 if (bond->balance == BM_AB || !bond->hash || revalidate) {
327 bond_entry_reset(bond);
333 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
334 * arbitrary client-provided pointer that uniquely identifies a slave within a
335 * bond. If 'slave_' already exists within 'bond' then this function
336 * reconfigures the existing slave.
338 * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
339 * across restarts and distributed vswitch instances. It should be unique per
340 * slave, and preferably consistent across restarts and reconfigurations.
342 * 'netdev' must be the network device that 'slave_' represents. It is owned
343 * by the client, so the client must not close it before either unregistering
344 * 'slave_' or destroying 'bond'.
347 bond_slave_register(struct bond *bond, void *slave_, uint16_t stb_id,
348 struct netdev *netdev)
350 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
353 slave = xzalloc(sizeof *slave);
355 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
358 slave->delay_expires = LLONG_MAX;
359 slave->up = bond_is_link_up(bond, netdev);
360 slave->name = xstrdup(netdev_get_name(netdev));
361 bond->bond_revalidate = true;
363 slave->enabled = false;
364 bond_enable_slave(slave, slave->up, NULL);
367 if (slave->stb_id != stb_id) {
368 slave->stb_id = stb_id;
369 bond->bond_revalidate = true;
372 if (slave->netdev != netdev) {
375 netdev_monitor_remove(bond->monitor, slave->netdev);
377 netdev_monitor_add(bond->monitor, netdev);
379 slave->netdev = netdev;
383 slave->name = xstrdup(netdev_get_name(netdev));
386 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
387 * then this function has no effect.
389 * Unregistering a slave invalidates all flows. */
391 bond_slave_unregister(struct bond *bond, const void *slave_)
393 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
401 netdev_monitor_remove(bond->monitor, slave->netdev);
404 bond_enable_slave(slave, false, NULL);
406 del_active = bond->active_slave == slave;
408 struct bond_entry *e;
409 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
410 if (e->slave == slave) {
418 hmap_remove(&bond->slaves, &slave->hmap_node);
419 /* Client owns 'slave->netdev'. */
426 bond_choose_active_slave(bond, &tags);
427 bond->send_learning_packets = true;
431 /* Should be called on each slave in 'bond' before bond_run() to indicate the
432 * results of lacp_slave_may_enable() on 'slave_'. */
434 bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
437 bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
440 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
441 * allow tagged flows to be invalidated.
443 * The caller should check bond_should_send_learning_packets() afterward. */
445 bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
447 struct bond_slave *slave;
448 bool is_tcp_hash = bond_is_tcp_hash(bond);
450 bond->lacp_negotiated = lacp_negotiated;
452 /* Update link status. */
453 if (bond->detect == BLSM_CARRIER
454 || time_msec() >= bond->miimon_next_update)
456 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
457 slave->up = bond_is_link_up(bond, slave->netdev);
459 bond->miimon_next_update = time_msec() + bond->miimon_interval;
462 /* Enable slaves based on link status and LACP feedback. */
463 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
464 bond_link_status_update(slave, tags);
466 if (!bond->active_slave || !bond->active_slave->enabled) {
467 bond_choose_active_slave(bond, tags);
470 /* Update fake bond interface stats. */
471 if (time_msec() >= bond->next_fake_iface_update) {
472 bond_update_fake_slave_stats(bond);
473 bond->next_fake_iface_update = time_msec() + 1000;
476 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
477 bond->bond_revalidate = true;
480 if (bond->bond_revalidate) {
481 bond->bond_revalidate = false;
483 bond_entry_reset(bond);
484 if (bond->balance != BM_STABLE) {
485 struct bond_slave *slave;
487 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
488 tag_set_add(tags, slave->tag);
491 tag_set_add(tags, bond->stb_tag);
493 tag_set_add(tags, bond->no_slaves_tag);
496 /* Invalidate any tags required by */
497 tag_set_union(tags, &bond->unixctl_tags);
498 tag_set_init(&bond->unixctl_tags);
501 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
503 bond_wait(struct bond *bond)
505 struct bond_slave *slave;
507 if (bond->detect == BLSM_CARRIER) {
508 netdev_monitor_poll_wait(bond->monitor);
510 poll_timer_wait_until(bond->miimon_next_update);
513 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
514 if (slave->delay_expires != LLONG_MAX) {
515 poll_timer_wait_until(slave->delay_expires);
519 if (bond->next_fake_iface_update != LLONG_MAX) {
520 poll_timer_wait_until(bond->next_fake_iface_update);
523 /* Ensure that any saved tags get revalidated right away. */
524 if (!tag_set_is_empty(&bond->unixctl_tags)) {
525 poll_immediate_wake();
528 /* We don't wait for bond->next_rebalance because rebalancing can only run
529 * at a flow account checkpoint. ofproto does checkpointing on its own
530 * schedule and bond_rebalance() gets called afterward, so we'd just be
531 * waking up for no purpose. */
534 /* MAC learning table interaction. */
537 may_send_learning_packets(const struct bond *bond)
539 return !bond->lacp_negotiated && bond->balance != BM_AB;
542 /* Returns true if 'bond' needs the client to send out packets to assist with
543 * MAC learning on 'bond'. If this function returns true, then the client
544 * should iterate through its MAC learning table for the bridge on which 'bond'
545 * is located. For each MAC that has been learned on a port other than 'bond',
546 * it should call bond_send_learning_packet().
548 * This function will only return true if 'bond' is in SLB mode and LACP is not
549 * negotiated. Otherwise sending learning packets isn't necessary.
551 * Calling this function resets the state that it checks. */
553 bond_should_send_learning_packets(struct bond *bond)
555 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
556 bond->send_learning_packets = false;
560 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
562 * See bond_should_send_learning_packets() for description of usage. */
564 bond_send_learning_packet(struct bond *bond,
565 const uint8_t eth_src[ETH_ADDR_LEN],
568 struct bond_slave *slave;
569 struct ofpbuf packet;
573 assert(may_send_learning_packets(bond));
574 if (!bond->active_slave) {
575 /* Nowhere to send the learning packet. */
579 memset(&flow, 0, sizeof flow);
580 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
581 slave = choose_output_slave(bond, &flow, vlan);
583 ofpbuf_init(&packet, 0);
584 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
587 eth_set_vlan_tci(&packet, htons(vlan));
589 error = netdev_send(slave->netdev, &packet);
590 ofpbuf_uninit(&packet);
595 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
596 * Ethernet destination address of 'eth_dst', should be admitted.
598 * The return value is one of the following:
600 * - BV_ACCEPT: Admit the packet.
602 * - BV_DROP: Drop the packet.
604 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
605 * Ethernet source address and VLAN. If there is none, or if the packet
606 * is on the learned port, then admit the packet. If a different port has
607 * been learned, however, drop the packet (and do not use it for MAC
611 bond_check_admissibility(struct bond *bond, const void *slave_,
612 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
614 /* Admit all packets if LACP has been negotiated, because that means that
615 * the remote switch is aware of the bond and will "do the right thing". */
616 if (bond->lacp_negotiated) {
620 /* Drop all multicast packets on inactive slaves. */
621 if (eth_addr_is_multicast(eth_dst)) {
622 *tags |= bond_get_active_slave_tag(bond);
623 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
628 /* Drop all packets for which we have learned a different input port,
629 * because we probably sent the packet on one slave and got it back on the
630 * other. Gratuitous ARP packets are an exception to this rule: the host
631 * has moved to another switch. The exception to the exception is if we
632 * locked the learning table to avoid reflections on bond slaves. */
633 return BV_DROP_IF_MOVED;
636 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
637 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
638 * NULL if the packet should be dropped because no slaves are enabled.
640 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
641 * should be a VID only (i.e. excluding the PCP bits). Second,
642 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
643 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
644 * packet belongs to (so for an access port it will be the access port's VLAN).
646 * Adds a tag to '*tags' that associates the flow with the returned slave.
649 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
650 uint16_t vlan, tag_type *tags)
652 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
654 *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
657 *tags |= bond->no_slaves_tag;
665 bond_is_balanced(const struct bond *bond)
667 return bond->balance == BM_SLB || bond->balance == BM_TCP;
670 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
672 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
676 if (bond_is_balanced(bond)) {
677 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
681 static struct bond_slave *
682 bond_slave_from_bal_node(struct list *bal)
684 return CONTAINER_OF(bal, struct bond_slave, bal_node);
688 log_bals(struct bond *bond, const struct list *bals)
690 if (VLOG_IS_DBG_ENABLED()) {
691 struct ds ds = DS_EMPTY_INITIALIZER;
692 const struct bond_slave *slave;
694 LIST_FOR_EACH (slave, bal_node, bals) {
696 ds_put_char(&ds, ',');
698 ds_put_format(&ds, " %s %"PRIu64"kB",
699 slave->name, slave->tx_bytes / 1024);
701 if (!slave->enabled) {
702 ds_put_cstr(&ds, " (disabled)");
704 if (!list_is_empty(&slave->entries)) {
705 struct bond_entry *e;
707 ds_put_cstr(&ds, " (");
708 LIST_FOR_EACH (e, list_node, &slave->entries) {
709 if (&e->list_node != list_front(&slave->entries)) {
710 ds_put_cstr(&ds, " + ");
712 ds_put_format(&ds, "h%td: %"PRIu64"kB",
713 e - bond->hash, e->tx_bytes / 1024);
715 ds_put_cstr(&ds, ")");
718 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
723 /* Shifts 'hash' from its current slave to 'to'. */
725 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
728 struct bond_slave *from = hash->slave;
729 struct bond *bond = from->bond;
730 uint64_t delta = hash->tx_bytes;
732 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
733 "from %s to %s (now carrying %"PRIu64"kB and "
734 "%"PRIu64"kB load, respectively)",
735 bond->name, delta / 1024, hash - bond->hash,
736 from->name, to->name,
737 (from->tx_bytes - delta) / 1024,
738 (to->tx_bytes + delta) / 1024);
740 /* Shift load away from 'from' to 'to'. */
741 from->tx_bytes -= delta;
742 to->tx_bytes += delta;
744 /* Arrange for flows to be revalidated. */
745 tag_set_add(set, hash->tag);
747 hash->tag = tag_create_random();
750 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
751 * given that doing so must decrease the ratio of the load on the two slaves by
752 * at least 0.1. Returns NULL if there is no appropriate entry.
754 * The list of entries isn't sorted. I don't know of a reason to prefer to
755 * shift away small hashes or large hashes. */
756 static struct bond_entry *
757 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
759 struct bond_entry *e;
761 if (list_is_short(&from->entries)) {
762 /* 'from' carries no more than one MAC hash, so shifting load away from
763 * it would be pointless. */
767 LIST_FOR_EACH (e, list_node, &from->entries) {
768 double old_ratio, new_ratio;
771 if (to_tx_bytes == 0) {
772 /* Nothing on the new slave, move it. */
777 old_ratio = (double)from->tx_bytes / to_tx_bytes;
778 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
779 if (old_ratio - new_ratio > 0.1) {
780 /* Would decrease the ratio, move it. */
788 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
791 insert_bal(struct list *bals, struct bond_slave *slave)
793 struct bond_slave *pos;
795 LIST_FOR_EACH (pos, bal_node, bals) {
796 if (slave->tx_bytes > pos->tx_bytes) {
800 list_insert(&pos->bal_node, &slave->bal_node);
803 /* Removes 'slave' from its current list and then inserts it into 'bals' so
804 * that descending order of 'tx_bytes' is maintained. */
806 reinsert_bal(struct list *bals, struct bond_slave *slave)
808 list_remove(&slave->bal_node);
809 insert_bal(bals, slave);
812 /* If 'bond' needs rebalancing, does so.
814 * The caller should have called bond_account() for each active flow, to ensure
815 * that flow data is consistently accounted at this point. */
817 bond_rebalance(struct bond *bond, struct tag_set *tags)
819 struct bond_slave *slave;
820 struct bond_entry *e;
823 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
826 bond->next_rebalance = time_msec() + bond->rebalance_interval;
828 /* Add each bond_entry to its slave's 'entries' list.
829 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
830 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
832 list_init(&slave->entries);
834 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
835 if (e->slave && e->tx_bytes) {
836 e->slave->tx_bytes += e->tx_bytes;
837 list_push_back(&e->slave->entries, &e->list_node);
841 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
843 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
844 * with a proper list sort algorithm. */
846 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
847 if (slave->enabled) {
848 insert_bal(&bals, slave);
851 log_bals(bond, &bals);
853 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
854 while (!list_is_short(&bals)) {
855 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
856 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
859 overload = from->tx_bytes - to->tx_bytes;
860 if (overload < to->tx_bytes >> 5 || overload < 100000) {
861 /* The extra load on 'from' (and all less-loaded slaves), compared
862 * to that of 'to' (the least-loaded slave), is less than ~3%, or
863 * it is less than ~1Mbps. No point in rebalancing. */
867 /* 'from' is carrying significantly more load than 'to', and that load
868 * is split across at least two different hashes. */
869 e = choose_entry_to_migrate(from, to->tx_bytes);
871 bond_shift_load(e, to, tags);
873 /* Delete element from from->entries.
875 * We don't add the element to to->hashes. That would only allow
876 * 'e' to be migrated to another slave in this rebalancing run, and
877 * there is no point in doing that. */
878 list_remove(&e->list_node);
880 /* Re-sort 'bals'. */
881 reinsert_bal(&bals, from);
882 reinsert_bal(&bals, to);
884 /* Can't usefully migrate anything away from 'from'.
885 * Don't reconsider it. */
886 list_remove(&from->bal_node);
890 /* Implement exponentially weighted moving average. A weight of 1/2 causes
891 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
892 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
893 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
901 /* Bonding unixctl user interface functions. */
904 bond_find(const char *name)
908 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
910 if (!strcmp(bond->name, name)) {
917 static struct bond_slave *
918 bond_lookup_slave(struct bond *bond, const char *slave_name)
920 struct bond_slave *slave;
922 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
923 if (!strcmp(slave->name, slave_name)) {
931 bond_unixctl_list(struct unixctl_conn *conn,
932 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
934 struct ds ds = DS_EMPTY_INITIALIZER;
935 const struct bond *bond;
937 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
939 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
940 const struct bond_slave *slave;
943 ds_put_format(&ds, "%s\t%s\t",
944 bond->name, bond_mode_to_string(bond->balance));
947 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
949 ds_put_cstr(&ds, ", ");
951 ds_put_cstr(&ds, slave->name);
953 ds_put_char(&ds, '\n');
955 unixctl_command_reply(conn, 200, ds_cstr(&ds));
960 bond_unixctl_show(struct unixctl_conn *conn,
961 const char *args, void *aux OVS_UNUSED)
963 struct ds ds = DS_EMPTY_INITIALIZER;
964 const struct bond_slave *slave;
965 const struct bond *bond;
967 bond = bond_find(args);
969 unixctl_command_reply(conn, 501, "no such bond");
973 ds_put_format(&ds, "bond_mode: %s\n",
974 bond_mode_to_string(bond->balance));
976 if (bond->balance != BM_AB) {
977 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
978 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
981 ds_put_format(&ds, "bond-detect-mode: %s\n",
982 bond->monitor ? "carrier" : "miimon");
984 if (!bond->monitor) {
985 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
986 bond->miimon_interval);
989 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
990 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
992 if (bond_is_balanced(bond)) {
993 ds_put_format(&ds, "next rebalance: %lld ms\n",
994 bond->next_rebalance - time_msec());
997 ds_put_format(&ds, "lacp_negotiated: %s\n",
998 bond->lacp_negotiated ? "true" : "false");
1000 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1001 struct bond_entry *be;
1005 ds_put_format(&ds, "\nslave %s: %s\n",
1006 slave->name, slave->enabled ? "enabled" : "disabled");
1007 if (slave == bond->active_slave) {
1008 ds_put_cstr(&ds, "\tactive slave\n");
1010 if (slave->delay_expires != LLONG_MAX) {
1011 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1012 slave->enabled ? "downdelay" : "updelay",
1013 slave->delay_expires - time_msec());
1016 ds_put_format(&ds, "\tlacp_may_enable: %s\n",
1017 slave->lacp_may_enable ? "true" : "false");
1019 if (!bond_is_balanced(bond)) {
1024 memset(&flow, 0, sizeof flow);
1025 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1026 int hash = be - bond->hash;
1028 if (be->slave != slave) {
1032 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1033 hash, be->tx_bytes / 1024);
1035 if (bond->balance != BM_SLB) {
1039 /* XXX How can we list the MACs assigned to hashes? */
1042 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1047 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1048 void *aux OVS_UNUSED)
1050 char *args = (char *) args_;
1051 char *save_ptr = NULL;
1052 char *bond_s, *hash_s, *slave_s;
1054 struct bond_slave *slave;
1055 struct bond_entry *entry;
1058 bond_s = strtok_r(args, " ", &save_ptr);
1059 hash_s = strtok_r(NULL, " ", &save_ptr);
1060 slave_s = strtok_r(NULL, " ", &save_ptr);
1062 unixctl_command_reply(conn, 501,
1063 "usage: bond/migrate BOND HASH SLAVE");
1067 bond = bond_find(bond_s);
1069 unixctl_command_reply(conn, 501, "no such bond");
1073 if (bond->balance != BM_SLB) {
1074 unixctl_command_reply(conn, 501, "not an SLB bond");
1078 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1079 hash = atoi(hash_s) & BOND_MASK;
1081 unixctl_command_reply(conn, 501, "bad hash");
1085 slave = bond_lookup_slave(bond, slave_s);
1087 unixctl_command_reply(conn, 501, "no such slave");
1091 if (!slave->enabled) {
1092 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1096 entry = &bond->hash[hash];
1097 tag_set_add(&bond->unixctl_tags, entry->tag);
1098 entry->slave = slave;
1099 entry->tag = tag_create_random();
1100 unixctl_command_reply(conn, 200, "migrated");
1104 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1105 void *aux OVS_UNUSED)
1107 char *args = (char *) args_;
1108 char *save_ptr = NULL;
1109 char *bond_s, *slave_s;
1111 struct bond_slave *slave;
1113 bond_s = strtok_r(args, " ", &save_ptr);
1114 slave_s = strtok_r(NULL, " ", &save_ptr);
1116 unixctl_command_reply(conn, 501,
1117 "usage: bond/set-active-slave BOND SLAVE");
1121 bond = bond_find(bond_s);
1123 unixctl_command_reply(conn, 501, "no such bond");
1127 slave = bond_lookup_slave(bond, slave_s);
1129 unixctl_command_reply(conn, 501, "no such slave");
1133 if (!slave->enabled) {
1134 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1138 if (bond->active_slave != slave) {
1139 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1140 bond->active_slave = slave;
1141 bond->active_slave->tag = tag_create_random();
1142 VLOG_INFO("bond %s: active interface is now %s",
1143 bond->name, slave->name);
1144 bond->send_learning_packets = true;
1145 unixctl_command_reply(conn, 200, "done");
1147 unixctl_command_reply(conn, 200, "no change");
1152 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1154 char *args = (char *) args_;
1155 char *save_ptr = NULL;
1156 char *bond_s, *slave_s;
1158 struct bond_slave *slave;
1160 bond_s = strtok_r(args, " ", &save_ptr);
1161 slave_s = strtok_r(NULL, " ", &save_ptr);
1163 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1164 enable ? "enable" : "disable");
1165 unixctl_command_reply(conn, 501, usage);
1170 bond = bond_find(bond_s);
1172 unixctl_command_reply(conn, 501, "no such bond");
1176 slave = bond_lookup_slave(bond, slave_s);
1178 unixctl_command_reply(conn, 501, "no such slave");
1182 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1183 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1187 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1188 void *aux OVS_UNUSED)
1190 enable_slave(conn, args, true);
1194 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1195 void *aux OVS_UNUSED)
1197 enable_slave(conn, args, false);
1201 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1202 void *aux OVS_UNUSED)
1204 char *args = (char *) args_;
1205 uint8_t mac[ETH_ADDR_LEN];
1209 char *mac_s, *vlan_s;
1210 char *save_ptr = NULL;
1212 mac_s = strtok_r(args, " ", &save_ptr);
1213 vlan_s = strtok_r(NULL, " ", &save_ptr);
1216 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1217 unixctl_command_reply(conn, 501, "invalid vlan");
1221 vlan = OFP_VLAN_NONE;
1224 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1225 == ETH_ADDR_SCAN_COUNT) {
1226 hash = bond_hash_src(mac, vlan) & BOND_MASK;
1228 hash_cstr = xasprintf("%u", hash);
1229 unixctl_command_reply(conn, 200, hash_cstr);
1232 unixctl_command_reply(conn, 501, "invalid mac");
1239 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1240 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1241 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1242 unixctl_command_register("bond/set-active-slave",
1243 bond_unixctl_set_active_slave, NULL);
1244 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1246 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1248 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1252 bond_entry_reset(struct bond *bond)
1254 if (bond->balance != BM_AB) {
1255 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1258 bond->hash = xmalloc(hash_len);
1260 memset(bond->hash, 0, hash_len);
1262 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1269 static struct bond_slave *
1270 bond_slave_lookup(struct bond *bond, const void *slave_)
1272 struct bond_slave *slave;
1274 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1276 if (slave->aux == slave_) {
1285 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1287 return (bond->detect == BLSM_CARRIER
1288 ? netdev_get_carrier(netdev)
1289 : netdev_get_miimon(netdev));
1293 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1295 struct bond *bond = slave->bond;
1296 slave->delay_expires = LLONG_MAX;
1297 if (enable != slave->enabled) {
1298 slave->enabled = enable;
1299 if (!slave->enabled) {
1300 VLOG_WARN("interface %s: disabled", slave->name);
1302 tag_set_add(tags, slave->tag);
1305 VLOG_WARN("interface %s: enabled", slave->name);
1306 slave->tag = tag_create_random();
1309 if (bond->balance == BM_STABLE) {
1310 bond->bond_revalidate = true;
1316 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1318 struct bond *bond = slave->bond;
1321 up = slave->up && slave->lacp_may_enable;
1322 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1323 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1324 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1325 slave->name, up ? "up" : "down");
1326 if (up == slave->enabled) {
1327 slave->delay_expires = LLONG_MAX;
1328 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1329 slave->name, up ? "disabled" : "enabled");
1331 int delay = (bond->lacp_negotiated ? 0
1332 : up ? bond->updelay : bond->downdelay);
1333 slave->delay_expires = time_msec() + delay;
1335 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1338 up ? "enabled" : "disabled",
1345 if (time_msec() >= slave->delay_expires) {
1346 bond_enable_slave(slave, up, tags);
1351 bond_is_tcp_hash(const struct bond *bond)
1353 return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
1354 && bond->lacp_negotiated;
1358 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan)
1360 return hash_bytes(mac, ETH_ADDR_LEN, vlan);
1364 bond_hash_tcp(const struct flow *flow, uint16_t vlan)
1366 struct flow hash_flow = *flow;
1367 hash_flow.vlan_tci = vlan;
1369 /* The symmetric quality of this hash function is not required, but
1370 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1371 * purposes, so we use it out of convenience. */
1372 return flow_hash_symmetric_l4(&hash_flow, 0);
1376 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1378 assert(bond->balance != BM_AB);
1380 return (bond_is_tcp_hash(bond)
1381 ? bond_hash_tcp(flow, vlan)
1382 : bond_hash_src(flow->dl_src, vlan));
1385 static struct bond_entry *
1386 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1389 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1392 /* This function uses Highest Random Weight hashing to choose an output slave.
1393 * This approach only reassigns a minimal number of flows when slaves are
1394 * enabled or disabled. Unfortunately, it has O(n) performance against the
1395 * number of slaves. There exist algorithms which are O(1), but have slightly
1396 * more complex implementations and require the use of memory. This may need
1397 * to be reimplemented if it becomes a performance bottleneck. */
1398 static struct bond_slave *
1399 choose_stb_slave(const struct bond *bond, const struct flow *flow,
1402 struct bond_slave *best, *slave;
1403 uint32_t best_hash, flow_hash;
1407 flow_hash = bond_hash(bond, flow, vlan);
1408 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1409 if (slave->enabled) {
1412 hash = hash_2words(flow_hash, slave->stb_id);
1413 if (!best || hash > best_hash) {
1423 static struct bond_slave *
1424 choose_output_slave(const struct bond *bond, const struct flow *flow,
1427 struct bond_entry *e;
1429 switch (bond->balance) {
1431 return bond->active_slave;
1434 return choose_stb_slave(bond, flow, vlan);
1437 e = lookup_bond_entry(bond, flow, vlan);
1438 if (!e->slave || !e->slave->enabled) {
1439 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1440 struct bond_slave, hmap_node);
1441 if (!e->slave->enabled) {
1442 e->slave = bond->active_slave;
1444 e->tag = tag_create_random();
1453 static struct bond_slave *
1454 bond_choose_slave(const struct bond *bond)
1456 struct bond_slave *slave, *best;
1458 /* Find an enabled slave. */
1459 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1460 if (slave->enabled) {
1465 /* All interfaces are disabled. Find an interface that will be enabled
1466 * after its updelay expires. */
1468 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1469 if (slave->delay_expires != LLONG_MAX
1470 && slave->lacp_may_enable
1471 && (!best || slave->delay_expires < best->delay_expires)) {
1479 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1481 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1482 struct bond_slave *old_active_slave = bond->active_slave;
1484 bond->active_slave = bond_choose_slave(bond);
1485 if (bond->active_slave) {
1486 if (bond->active_slave->enabled) {
1487 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1488 bond->name, bond->active_slave->name);
1490 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1491 "remaining %lld ms updelay (since no interface was "
1492 "enabled)", bond->name, bond->active_slave->name,
1493 bond->active_slave->delay_expires - time_msec());
1494 bond_enable_slave(bond->active_slave, true, tags);
1497 if (!old_active_slave) {
1498 tag_set_add(tags, bond->no_slaves_tag);
1501 bond->send_learning_packets = true;
1502 } else if (old_active_slave) {
1503 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1507 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1508 * there is no active slave. */
1510 bond_get_active_slave_tag(const struct bond *bond)
1512 return (bond->active_slave
1513 ? bond->active_slave->tag
1514 : bond->no_slaves_tag);
1517 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1518 * bond interface. */
1520 bond_update_fake_slave_stats(struct bond *bond)
1522 struct netdev_stats bond_stats;
1523 struct bond_slave *slave;
1524 struct netdev *bond_dev;
1526 memset(&bond_stats, 0, sizeof bond_stats);
1528 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1529 struct netdev_stats slave_stats;
1531 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1532 /* XXX: We swap the stats here because they are swapped back when
1533 * reported by the internal device. The reason for this is
1534 * internal devices normally represent packets going into the
1535 * system but when used as fake bond device they represent packets
1536 * leaving the system. We really should do this in the internal
1537 * device itself because changing it here reverses the counts from
1538 * the perspective of the switch. However, the internal device
1539 * doesn't know what type of device it represents so we have to do
1540 * it here for now. */
1541 bond_stats.tx_packets += slave_stats.rx_packets;
1542 bond_stats.tx_bytes += slave_stats.rx_bytes;
1543 bond_stats.rx_packets += slave_stats.tx_packets;
1544 bond_stats.rx_bytes += slave_stats.tx_bytes;
1548 if (!netdev_open_default(bond->name, &bond_dev)) {
1549 netdev_set_stats(bond_dev, &bond_stats);
1550 netdev_close(bond_dev);