2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 Nicira, Inc.
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.
27 #include "dynamic-string.h"
36 #include "poll-loop.h"
43 VLOG_DEFINE_THIS_MODULE(bond);
45 /* Bit-mask for hashing a flow down to a bucket.
46 * There are (BOND_MASK + 1) buckets. */
47 #define BOND_MASK 0xff
49 /* A hash bucket for mapping a flow to a slave.
50 * "struct bond" has an array of (BOND_MASK + 1) of these. */
52 struct bond_slave *slave; /* Assigned slave, NULL if unassigned. */
53 uint64_t tx_bytes; /* Count of bytes recently transmitted. */
54 tag_type tag; /* Tag for entry<->facet association. */
55 struct list list_node; /* In bond_slave's 'entries' list. */
58 /* A bond slave, that is, one of the links comprising a bond. */
60 struct hmap_node hmap_node; /* In struct bond's slaves hmap. */
61 struct bond *bond; /* The bond that contains this slave. */
62 void *aux; /* Client-provided handle for this slave. */
64 struct netdev *netdev; /* Network device, owned by the client. */
65 unsigned int change_seq; /* Tracks changes in 'netdev'. */
66 char *name; /* Name (a copy of netdev_get_name(netdev)). */
69 long long delay_expires; /* Time after which 'enabled' may change. */
70 bool enabled; /* May be chosen for flows? */
71 bool may_enable; /* Client considers this slave bondable. */
72 tag_type tag; /* Tag associated with this slave. */
74 /* Rebalancing info. Used only by bond_rebalance(). */
75 struct list bal_node; /* In bond_rebalance()'s 'bals' list. */
76 struct list entries; /* 'struct bond_entry's assigned here. */
77 uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
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 enum lacp_status lacp_status; /* Status of LACP negotiations. */
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 /* Legacy compatibility. */
105 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
107 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
108 * where we can't otherwise provide revalidation feedback to the client.
109 * That's only unixctl commands now; I hope no other cases will arise. */
110 struct tag_set unixctl_tags;
115 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
117 static void bond_entry_reset(struct bond *);
118 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
119 static void bond_enable_slave(struct bond_slave *, bool enable,
121 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
122 static void bond_choose_active_slave(struct bond *, struct tag_set *);
123 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
124 uint16_t vlan, uint32_t basis);
125 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan,
127 static struct bond_entry *lookup_bond_entry(const struct bond *,
130 static tag_type bond_get_active_slave_tag(const struct bond *);
131 static struct bond_slave *choose_output_slave(const struct bond *,
133 struct flow_wildcards *,
134 uint16_t vlan, tag_type *tags);
135 static void bond_update_fake_slave_stats(struct bond *);
137 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
138 * stores the mode in '*balance' and returns true. Otherwise returns false
139 * without modifying '*balance'. */
141 bond_mode_from_string(enum bond_mode *balance, const char *s)
143 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
145 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
147 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
155 /* Returns a string representing 'balance'. */
157 bond_mode_to_string(enum bond_mode balance) {
160 return "balance-tcp";
162 return "balance-slb";
164 return "active-backup";
170 /* Creates and returns a new bond whose configuration is initially taken from
173 * The caller should register each slave on the new bond by calling
174 * bond_slave_register(). */
176 bond_create(const struct bond_settings *s)
180 bond = xzalloc(sizeof *bond);
181 hmap_init(&bond->slaves);
182 bond->no_slaves_tag = tag_create_random();
183 bond->next_fake_iface_update = LLONG_MAX;
186 bond_reconfigure(bond, s);
188 tag_set_init(&bond->unixctl_tags);
194 bond_ref(const struct bond *bond_)
196 struct bond *bond = CONST_CAST(struct bond *, bond_);
198 ovs_assert(bond->ref_cnt > 0);
205 bond_unref(struct bond *bond)
207 struct bond_slave *slave, *next_slave;
213 ovs_assert(bond->ref_cnt > 0);
214 if (--bond->ref_cnt) {
218 hmap_remove(&all_bonds, &bond->hmap_node);
220 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
221 hmap_remove(&bond->slaves, &slave->hmap_node);
222 /* Client owns 'slave->netdev'. */
226 hmap_destroy(&bond->slaves);
233 /* Updates 'bond''s overall configuration to 's'.
235 * The caller should register each slave on 'bond' by calling
236 * bond_slave_register(). This is optional if none of the slaves'
237 * configuration has changed. In any case it can't hurt.
239 * Returns true if the configuration has changed in such a way that requires
243 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
245 bool revalidate = false;
247 if (!bond->name || strcmp(bond->name, s->name)) {
249 hmap_remove(&all_bonds, &bond->hmap_node);
252 bond->name = xstrdup(s->name);
253 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
256 bond->updelay = s->up_delay;
257 bond->downdelay = s->down_delay;
259 if (bond->rebalance_interval != s->rebalance_interval) {
260 bond->rebalance_interval = s->rebalance_interval;
264 if (bond->balance != s->balance) {
265 bond->balance = s->balance;
269 if (bond->basis != s->basis) {
270 bond->basis = s->basis;
275 if (bond->next_fake_iface_update == LLONG_MAX) {
276 bond->next_fake_iface_update = time_msec();
279 bond->next_fake_iface_update = LLONG_MAX;
282 if (bond->bond_revalidate) {
284 bond->bond_revalidate = false;
287 if (bond->balance == BM_AB || !bond->hash || revalidate) {
288 bond_entry_reset(bond);
295 bond_slave_set_netdev__(struct bond_slave *slave, struct netdev *netdev)
297 if (slave->netdev != netdev) {
298 slave->netdev = netdev;
299 slave->change_seq = 0;
303 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
304 * arbitrary client-provided pointer that uniquely identifies a slave within a
305 * bond. If 'slave_' already exists within 'bond' then this function
306 * reconfigures the existing slave.
308 * 'netdev' must be the network device that 'slave_' represents. It is owned
309 * by the client, so the client must not close it before either unregistering
310 * 'slave_' or destroying 'bond'.
313 bond_slave_register(struct bond *bond, void *slave_, struct netdev *netdev)
315 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
318 slave = xzalloc(sizeof *slave);
320 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
323 slave->delay_expires = LLONG_MAX;
324 slave->name = xstrdup(netdev_get_name(netdev));
325 bond->bond_revalidate = true;
327 slave->enabled = false;
328 bond_enable_slave(slave, netdev_get_carrier(netdev), NULL);
331 bond_slave_set_netdev__(slave, netdev);
334 slave->name = xstrdup(netdev_get_name(netdev));
337 /* Updates the network device to be used with 'slave_' to 'netdev'.
339 * This is useful if the caller closes and re-opens the network device
340 * registered with bond_slave_register() but doesn't need to change anything
343 bond_slave_set_netdev(struct bond *bond, void *slave_, struct netdev *netdev)
345 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
347 bond_slave_set_netdev__(slave, netdev);
351 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
352 * then this function has no effect.
354 * Unregistering a slave invalidates all flows. */
356 bond_slave_unregister(struct bond *bond, const void *slave_)
358 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
365 bond_enable_slave(slave, false, NULL);
367 del_active = bond->active_slave == slave;
369 struct bond_entry *e;
370 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
371 if (e->slave == slave) {
379 hmap_remove(&bond->slaves, &slave->hmap_node);
380 /* Client owns 'slave->netdev'. */
387 bond_choose_active_slave(bond, &tags);
388 bond->send_learning_packets = true;
392 /* Should be called on each slave in 'bond' before bond_run() to indicate
393 * whether or not 'slave_' may be enabled. This function is intended to allow
394 * other protocols to have some impact on bonding decisions. For example LACP
395 * or high level link monitoring protocols may decide that a given slave should
396 * not be able to send traffic. */
398 bond_slave_set_may_enable(struct bond *bond, void *slave_, bool may_enable)
400 bond_slave_lookup(bond, slave_)->may_enable = may_enable;
403 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
404 * allow tagged flows to be invalidated.
406 * The caller should check bond_should_send_learning_packets() afterward. */
408 bond_run(struct bond *bond, struct tag_set *tags, enum lacp_status lacp_status)
410 struct bond_slave *slave;
412 if (bond->lacp_status != lacp_status) {
413 bond->lacp_status = lacp_status;
414 bond->bond_revalidate = true;
417 /* Enable slaves based on link status and LACP feedback. */
418 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
419 bond_link_status_update(slave, tags);
420 slave->change_seq = netdev_change_seq(slave->netdev);
422 if (!bond->active_slave || !bond->active_slave->enabled) {
423 bond_choose_active_slave(bond, tags);
426 /* Update fake bond interface stats. */
427 if (time_msec() >= bond->next_fake_iface_update) {
428 bond_update_fake_slave_stats(bond);
429 bond->next_fake_iface_update = time_msec() + 1000;
432 if (bond->bond_revalidate) {
433 struct bond_slave *slave;
435 bond->bond_revalidate = false;
436 bond_entry_reset(bond);
437 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
438 tag_set_add(tags, slave->tag);
440 tag_set_add(tags, bond->no_slaves_tag);
443 /* Invalidate any tags required by */
444 tag_set_union(tags, &bond->unixctl_tags);
445 tag_set_init(&bond->unixctl_tags);
448 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
450 bond_wait(struct bond *bond)
452 struct bond_slave *slave;
454 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
455 if (slave->delay_expires != LLONG_MAX) {
456 poll_timer_wait_until(slave->delay_expires);
459 if (slave->change_seq != netdev_change_seq(slave->netdev)) {
460 poll_immediate_wake();
464 if (bond->next_fake_iface_update != LLONG_MAX) {
465 poll_timer_wait_until(bond->next_fake_iface_update);
468 /* Ensure that any saved tags get revalidated right away. */
469 if (!tag_set_is_empty(&bond->unixctl_tags)) {
470 poll_immediate_wake();
473 /* We don't wait for bond->next_rebalance because rebalancing can only run
474 * at a flow account checkpoint. ofproto does checkpointing on its own
475 * schedule and bond_rebalance() gets called afterward, so we'd just be
476 * waking up for no purpose. */
479 /* MAC learning table interaction. */
482 may_send_learning_packets(const struct bond *bond)
484 return bond->lacp_status == LACP_DISABLED
485 && (bond->balance == BM_SLB || bond->balance == BM_AB)
486 && bond->active_slave;
489 /* Returns true if 'bond' needs the client to send out packets to assist with
490 * MAC learning on 'bond'. If this function returns true, then the client
491 * should iterate through its MAC learning table for the bridge on which 'bond'
492 * is located. For each MAC that has been learned on a port other than 'bond',
493 * it should call bond_compose_learning_packet().
495 * This function will only return true if 'bond' is in SLB or active-backup
496 * mode and LACP is not negotiated. Otherwise sending learning packets isn't
499 * Calling this function resets the state that it checks. */
501 bond_should_send_learning_packets(struct bond *bond)
503 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
504 bond->send_learning_packets = false;
508 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
510 * See bond_should_send_learning_packets() for description of usage. The
511 * caller should send the composed packet on the port associated with
512 * port_aux and takes ownership of the returned ofpbuf. */
514 bond_compose_learning_packet(struct bond *bond,
515 const uint8_t eth_src[ETH_ADDR_LEN],
516 uint16_t vlan, void **port_aux)
518 struct bond_slave *slave;
519 struct ofpbuf *packet;
523 ovs_assert(may_send_learning_packets(bond));
525 memset(&flow, 0, sizeof flow);
526 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
527 slave = choose_output_slave(bond, &flow, NULL, vlan, &tags);
529 packet = ofpbuf_new(0);
530 compose_rarp(packet, eth_src);
532 eth_push_vlan(packet, htons(vlan));
535 *port_aux = slave->aux;
539 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
540 * Ethernet destination address of 'eth_dst', should be admitted.
542 * The return value is one of the following:
544 * - BV_ACCEPT: Admit the packet.
546 * - BV_DROP: Drop the packet.
548 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
549 * Ethernet source address and VLAN. If there is none, or if the packet
550 * is on the learned port, then admit the packet. If a different port has
551 * been learned, however, drop the packet (and do not use it for MAC
555 bond_check_admissibility(struct bond *bond, const void *slave_,
556 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
558 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
564 /* LACP bonds have very loose admissibility restrictions because we can
565 * assume the remote switch is aware of the bond and will "do the right
566 * thing". However, as a precaution we drop packets on disabled slaves
567 * because no correctly implemented partner switch should be sending
570 * If LACP is configured, but LACP negotiations have been unsuccessful, we
571 * drop all incoming traffic. */
572 switch (bond->lacp_status) {
573 case LACP_NEGOTIATED: return slave->enabled ? BV_ACCEPT : BV_DROP;
574 case LACP_CONFIGURED: return BV_DROP;
575 case LACP_DISABLED: break;
578 /* Drop all multicast packets on inactive slaves. */
579 if (eth_addr_is_multicast(eth_dst)) {
580 *tags |= bond_get_active_slave_tag(bond);
581 if (bond->active_slave != slave) {
586 switch (bond->balance) {
588 /* Drop all packets which arrive on backup slaves. This is similar to
589 * how Linux bonding handles active-backup bonds. */
590 *tags |= bond_get_active_slave_tag(bond);
591 if (bond->active_slave != slave) {
592 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
594 VLOG_DBG_RL(&rl, "active-backup bond received packet on backup"
595 " slave (%s) destined for " ETH_ADDR_FMT,
596 slave->name, ETH_ADDR_ARGS(eth_dst));
602 /* TCP balanced bonds require successful LACP negotiated. Based on the
603 * above check, LACP is off on this bond. Therfore, we drop all
604 * incoming traffic. */
608 /* Drop all packets for which we have learned a different input port,
609 * because we probably sent the packet on one slave and got it back on
610 * the other. Gratuitous ARP packets are an exception to this rule:
611 * the host has moved to another switch. The exception to the
612 * exception is if we locked the learning table to avoid reflections on
614 return BV_DROP_IF_MOVED;
620 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
621 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
622 * NULL if the packet should be dropped because no slaves are enabled.
624 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
625 * should be a VID only (i.e. excluding the PCP bits). Second,
626 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
627 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
628 * packet belongs to (so for an access port it will be the access port's VLAN).
630 * Adds a tag to '*tags' that associates the flow with the returned slave.
632 * If 'wc' is non-NULL, bitwise-OR's 'wc' with the set of bits that were
633 * significant in the selection. At some point earlier, 'wc' should
634 * have been initialized (e.g., by flow_wildcards_init_catchall()).
637 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
638 struct flow_wildcards *wc, uint16_t vlan,
641 struct bond_slave *slave = choose_output_slave(bond, flow, wc, vlan, tags);
646 *tags |= bond->no_slaves_tag;
654 bond_is_balanced(const struct bond *bond)
656 return bond->rebalance_interval
657 && (bond->balance == BM_SLB || bond->balance == BM_TCP);
660 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
662 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
665 if (bond_is_balanced(bond)) {
666 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
670 static struct bond_slave *
671 bond_slave_from_bal_node(struct list *bal)
673 return CONTAINER_OF(bal, struct bond_slave, bal_node);
677 log_bals(struct bond *bond, const struct list *bals)
679 if (VLOG_IS_DBG_ENABLED()) {
680 struct ds ds = DS_EMPTY_INITIALIZER;
681 const struct bond_slave *slave;
683 LIST_FOR_EACH (slave, bal_node, bals) {
685 ds_put_char(&ds, ',');
687 ds_put_format(&ds, " %s %"PRIu64"kB",
688 slave->name, slave->tx_bytes / 1024);
690 if (!slave->enabled) {
691 ds_put_cstr(&ds, " (disabled)");
693 if (!list_is_empty(&slave->entries)) {
694 struct bond_entry *e;
696 ds_put_cstr(&ds, " (");
697 LIST_FOR_EACH (e, list_node, &slave->entries) {
698 if (&e->list_node != list_front(&slave->entries)) {
699 ds_put_cstr(&ds, " + ");
701 ds_put_format(&ds, "h%td: %"PRIu64"kB",
702 e - bond->hash, e->tx_bytes / 1024);
704 ds_put_cstr(&ds, ")");
707 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
712 /* Shifts 'hash' from its current slave to 'to'. */
714 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
717 struct bond_slave *from = hash->slave;
718 struct bond *bond = from->bond;
719 uint64_t delta = hash->tx_bytes;
721 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
722 "from %s to %s (now carrying %"PRIu64"kB and "
723 "%"PRIu64"kB load, respectively)",
724 bond->name, delta / 1024, hash - bond->hash,
725 from->name, to->name,
726 (from->tx_bytes - delta) / 1024,
727 (to->tx_bytes + delta) / 1024);
729 /* Shift load away from 'from' to 'to'. */
730 from->tx_bytes -= delta;
731 to->tx_bytes += delta;
733 /* Arrange for flows to be revalidated. */
734 tag_set_add(set, hash->tag);
736 hash->tag = tag_create_random();
739 /* Picks and returns a bond_entry to migrate from 'from' (the most heavily
740 * loaded bond slave) to a bond slave that has 'to_tx_bytes' bytes of load,
741 * given that doing so must decrease the ratio of the load on the two slaves by
742 * at least 0.1. Returns NULL if there is no appropriate entry.
744 * The list of entries isn't sorted. I don't know of a reason to prefer to
745 * shift away small hashes or large hashes. */
746 static struct bond_entry *
747 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
749 struct bond_entry *e;
751 if (list_is_short(&from->entries)) {
752 /* 'from' carries no more than one MAC hash, so shifting load away from
753 * it would be pointless. */
757 LIST_FOR_EACH (e, list_node, &from->entries) {
758 double old_ratio, new_ratio;
761 if (to_tx_bytes == 0) {
762 /* Nothing on the new slave, move it. */
767 old_ratio = (double)from->tx_bytes / to_tx_bytes;
768 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
769 if (old_ratio - new_ratio > 0.1
770 && fabs(new_ratio - 1.0) < fabs(old_ratio - 1.0)) {
771 /* We're aiming for an ideal ratio of 1, meaning both the 'from'
772 and 'to' slave have the same load. Therefore, we only move an
773 entry if it decreases the load on 'from', and brings us closer
774 to equal traffic load. */
782 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
785 insert_bal(struct list *bals, struct bond_slave *slave)
787 struct bond_slave *pos;
789 LIST_FOR_EACH (pos, bal_node, bals) {
790 if (slave->tx_bytes > pos->tx_bytes) {
794 list_insert(&pos->bal_node, &slave->bal_node);
797 /* Removes 'slave' from its current list and then inserts it into 'bals' so
798 * that descending order of 'tx_bytes' is maintained. */
800 reinsert_bal(struct list *bals, struct bond_slave *slave)
802 list_remove(&slave->bal_node);
803 insert_bal(bals, slave);
806 /* If 'bond' needs rebalancing, does so.
808 * The caller should have called bond_account() for each active flow, to ensure
809 * that flow data is consistently accounted at this point. */
811 bond_rebalance(struct bond *bond, struct tag_set *tags)
813 struct bond_slave *slave;
814 struct bond_entry *e;
817 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
820 bond->next_rebalance = time_msec() + bond->rebalance_interval;
822 /* Add each bond_entry to its slave's 'entries' list.
823 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
824 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
826 list_init(&slave->entries);
828 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
829 if (e->slave && e->tx_bytes) {
830 e->slave->tx_bytes += e->tx_bytes;
831 list_push_back(&e->slave->entries, &e->list_node);
835 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
837 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
838 * with a proper list sort algorithm. */
840 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
841 if (slave->enabled) {
842 insert_bal(&bals, slave);
845 log_bals(bond, &bals);
847 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
848 while (!list_is_short(&bals)) {
849 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
850 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
853 overload = from->tx_bytes - to->tx_bytes;
854 if (overload < to->tx_bytes >> 5 || overload < 100000) {
855 /* The extra load on 'from' (and all less-loaded slaves), compared
856 * to that of 'to' (the least-loaded slave), is less than ~3%, or
857 * it is less than ~1Mbps. No point in rebalancing. */
861 /* 'from' is carrying significantly more load than 'to'. Pick a hash
862 * to move from 'from' to 'to'. */
863 e = choose_entry_to_migrate(from, to->tx_bytes);
865 bond_shift_load(e, to, tags);
867 /* Delete element from from->entries.
869 * We don't add the element to to->hashes. That would only allow
870 * 'e' to be migrated to another slave in this rebalancing run, and
871 * there is no point in doing that. */
872 list_remove(&e->list_node);
874 /* Re-sort 'bals'. */
875 reinsert_bal(&bals, from);
876 reinsert_bal(&bals, to);
878 /* Can't usefully migrate anything away from 'from'.
879 * Don't reconsider it. */
880 list_remove(&from->bal_node);
884 /* Implement exponentially weighted moving average. A weight of 1/2 causes
885 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
886 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
887 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
895 /* Bonding unixctl user interface functions. */
898 bond_find(const char *name)
902 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
904 if (!strcmp(bond->name, name)) {
911 static struct bond_slave *
912 bond_lookup_slave(struct bond *bond, const char *slave_name)
914 struct bond_slave *slave;
916 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
917 if (!strcmp(slave->name, slave_name)) {
925 bond_unixctl_list(struct unixctl_conn *conn,
926 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
927 void *aux OVS_UNUSED)
929 struct ds ds = DS_EMPTY_INITIALIZER;
930 const struct bond *bond;
932 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
934 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
935 const struct bond_slave *slave;
938 ds_put_format(&ds, "%s\t%s\t",
939 bond->name, bond_mode_to_string(bond->balance));
942 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
944 ds_put_cstr(&ds, ", ");
946 ds_put_cstr(&ds, slave->name);
948 ds_put_char(&ds, '\n');
950 unixctl_command_reply(conn, ds_cstr(&ds));
955 bond_print_details(struct ds *ds, const struct bond *bond)
957 struct shash slave_shash = SHASH_INITIALIZER(&slave_shash);
958 const struct shash_node **sorted_slaves = NULL;
959 const struct bond_slave *slave;
962 ds_put_format(ds, "---- %s ----\n", bond->name);
963 ds_put_format(ds, "bond_mode: %s\n",
964 bond_mode_to_string(bond->balance));
966 ds_put_format(ds, "bond-hash-basis: %"PRIu32"\n", bond->basis);
968 ds_put_format(ds, "updelay: %d ms\n", bond->updelay);
969 ds_put_format(ds, "downdelay: %d ms\n", bond->downdelay);
971 if (bond_is_balanced(bond)) {
972 ds_put_format(ds, "next rebalance: %lld ms\n",
973 bond->next_rebalance - time_msec());
976 ds_put_cstr(ds, "lacp_status: ");
977 switch (bond->lacp_status) {
978 case LACP_NEGOTIATED:
979 ds_put_cstr(ds, "negotiated\n");
981 case LACP_CONFIGURED:
982 ds_put_cstr(ds, "configured\n");
985 ds_put_cstr(ds, "off\n");
988 ds_put_cstr(ds, "<unknown>\n");
992 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
993 shash_add(&slave_shash, slave->name, slave);
995 sorted_slaves = shash_sort(&slave_shash);
997 for (i = 0; i < shash_count(&slave_shash); i++) {
998 struct bond_entry *be;
1000 slave = sorted_slaves[i]->data;
1003 ds_put_format(ds, "\nslave %s: %s\n",
1004 slave->name, slave->enabled ? "enabled" : "disabled");
1005 if (slave == bond->active_slave) {
1006 ds_put_cstr(ds, "\tactive slave\n");
1008 if (slave->delay_expires != LLONG_MAX) {
1009 ds_put_format(ds, "\t%s expires in %lld ms\n",
1010 slave->enabled ? "downdelay" : "updelay",
1011 slave->delay_expires - time_msec());
1014 ds_put_format(ds, "\tmay_enable: %s\n",
1015 slave->may_enable ? "true" : "false");
1017 if (!bond_is_balanced(bond)) {
1022 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1023 int hash = be - bond->hash;
1025 if (be->slave != slave) {
1029 ds_put_format(ds, "\thash %d: %"PRIu64" kB load\n",
1030 hash, be->tx_bytes / 1024);
1032 /* XXX How can we list the MACs assigned to hashes of SLB bonds? */
1035 shash_destroy(&slave_shash);
1036 free(sorted_slaves);
1037 ds_put_cstr(ds, "\n");
1041 bond_unixctl_show(struct unixctl_conn *conn,
1042 int argc, const char *argv[],
1043 void *aux OVS_UNUSED)
1045 struct ds ds = DS_EMPTY_INITIALIZER;
1048 const struct bond *bond = bond_find(argv[1]);
1051 unixctl_command_reply_error(conn, "no such bond");
1054 bond_print_details(&ds, bond);
1056 const struct bond *bond;
1058 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
1059 bond_print_details(&ds, bond);
1063 unixctl_command_reply(conn, ds_cstr(&ds));
1068 bond_unixctl_migrate(struct unixctl_conn *conn,
1069 int argc OVS_UNUSED, const char *argv[],
1070 void *aux OVS_UNUSED)
1072 const char *bond_s = argv[1];
1073 const char *hash_s = argv[2];
1074 const char *slave_s = argv[3];
1076 struct bond_slave *slave;
1077 struct bond_entry *entry;
1080 bond = bond_find(bond_s);
1082 unixctl_command_reply_error(conn, "no such bond");
1086 if (bond->balance != BM_SLB) {
1087 unixctl_command_reply_error(conn, "not an SLB bond");
1091 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1092 hash = atoi(hash_s) & BOND_MASK;
1094 unixctl_command_reply_error(conn, "bad hash");
1098 slave = bond_lookup_slave(bond, slave_s);
1100 unixctl_command_reply_error(conn, "no such slave");
1104 if (!slave->enabled) {
1105 unixctl_command_reply_error(conn, "cannot migrate to disabled slave");
1109 entry = &bond->hash[hash];
1110 tag_set_add(&bond->unixctl_tags, entry->tag);
1111 entry->slave = slave;
1112 entry->tag = tag_create_random();
1113 unixctl_command_reply(conn, "migrated");
1117 bond_unixctl_set_active_slave(struct unixctl_conn *conn,
1118 int argc OVS_UNUSED, const char *argv[],
1119 void *aux OVS_UNUSED)
1121 const char *bond_s = argv[1];
1122 const char *slave_s = argv[2];
1124 struct bond_slave *slave;
1126 bond = bond_find(bond_s);
1128 unixctl_command_reply_error(conn, "no such bond");
1132 slave = bond_lookup_slave(bond, slave_s);
1134 unixctl_command_reply_error(conn, "no such slave");
1138 if (!slave->enabled) {
1139 unixctl_command_reply_error(conn, "cannot make disabled slave active");
1143 if (bond->active_slave != slave) {
1144 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1145 bond->active_slave = slave;
1146 bond->active_slave->tag = tag_create_random();
1147 VLOG_INFO("bond %s: active interface is now %s",
1148 bond->name, slave->name);
1149 bond->send_learning_packets = true;
1150 unixctl_command_reply(conn, "done");
1152 unixctl_command_reply(conn, "no change");
1157 enable_slave(struct unixctl_conn *conn, const char *argv[], bool enable)
1159 const char *bond_s = argv[1];
1160 const char *slave_s = argv[2];
1162 struct bond_slave *slave;
1164 bond = bond_find(bond_s);
1166 unixctl_command_reply_error(conn, "no such bond");
1170 slave = bond_lookup_slave(bond, slave_s);
1172 unixctl_command_reply_error(conn, "no such slave");
1176 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1177 unixctl_command_reply(conn, enable ? "enabled" : "disabled");
1181 bond_unixctl_enable_slave(struct unixctl_conn *conn,
1182 int argc OVS_UNUSED, const char *argv[],
1183 void *aux OVS_UNUSED)
1185 enable_slave(conn, argv, true);
1189 bond_unixctl_disable_slave(struct unixctl_conn *conn,
1190 int argc OVS_UNUSED, const char *argv[],
1191 void *aux OVS_UNUSED)
1193 enable_slave(conn, argv, false);
1197 bond_unixctl_hash(struct unixctl_conn *conn, int argc, const char *argv[],
1198 void *aux OVS_UNUSED)
1200 const char *mac_s = argv[1];
1201 const char *vlan_s = argc > 2 ? argv[2] : NULL;
1202 const char *basis_s = argc > 3 ? argv[3] : NULL;
1203 uint8_t mac[ETH_ADDR_LEN];
1210 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1211 unixctl_command_reply_error(conn, "invalid vlan");
1219 if (sscanf(basis_s, "%"PRIu32, &basis) != 1) {
1220 unixctl_command_reply_error(conn, "invalid basis");
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, basis) & BOND_MASK;
1231 hash_cstr = xasprintf("%u", hash);
1232 unixctl_command_reply(conn, hash_cstr);
1235 unixctl_command_reply_error(conn, "invalid mac");
1242 unixctl_command_register("bond/list", "", 0, 0, bond_unixctl_list, NULL);
1243 unixctl_command_register("bond/show", "[port]", 0, 1, bond_unixctl_show,
1245 unixctl_command_register("bond/migrate", "port hash slave", 3, 3,
1246 bond_unixctl_migrate, NULL);
1247 unixctl_command_register("bond/set-active-slave", "port slave", 2, 2,
1248 bond_unixctl_set_active_slave, NULL);
1249 unixctl_command_register("bond/enable-slave", "port slave", 2, 2,
1250 bond_unixctl_enable_slave, NULL);
1251 unixctl_command_register("bond/disable-slave", "port slave", 2, 2,
1252 bond_unixctl_disable_slave, NULL);
1253 unixctl_command_register("bond/hash", "mac [vlan] [basis]", 1, 3,
1254 bond_unixctl_hash, NULL);
1258 bond_entry_reset(struct bond *bond)
1260 if (bond->balance != BM_AB) {
1261 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1264 bond->hash = xmalloc(hash_len);
1266 memset(bond->hash, 0, hash_len);
1268 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1275 static struct bond_slave *
1276 bond_slave_lookup(struct bond *bond, const void *slave_)
1278 struct bond_slave *slave;
1280 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1282 if (slave->aux == slave_) {
1291 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1293 slave->delay_expires = LLONG_MAX;
1294 if (enable != slave->enabled) {
1295 slave->enabled = enable;
1296 if (!slave->enabled) {
1297 VLOG_INFO("interface %s: disabled", slave->name);
1299 tag_set_add(tags, slave->tag);
1302 VLOG_INFO("interface %s: enabled", slave->name);
1303 slave->tag = tag_create_random();
1309 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1311 struct bond *bond = slave->bond;
1314 up = netdev_get_carrier(slave->netdev) && slave->may_enable;
1315 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1316 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1317 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1318 slave->name, up ? "up" : "down");
1319 if (up == slave->enabled) {
1320 slave->delay_expires = LLONG_MAX;
1321 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1322 slave->name, up ? "disabled" : "enabled");
1324 int delay = (bond->lacp_status != LACP_DISABLED ? 0
1325 : up ? bond->updelay : bond->downdelay);
1326 slave->delay_expires = time_msec() + delay;
1328 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1331 up ? "enabled" : "disabled",
1338 if (time_msec() >= slave->delay_expires) {
1339 bond_enable_slave(slave, up, tags);
1344 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan, uint32_t basis)
1346 return hash_3words(hash_bytes(mac, ETH_ADDR_LEN, 0), vlan, basis);
1350 bond_hash_tcp(const struct flow *flow, uint16_t vlan, uint32_t basis)
1352 struct flow hash_flow = *flow;
1353 hash_flow.vlan_tci = htons(vlan);
1355 /* The symmetric quality of this hash function is not required, but
1356 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1357 * purposes, so we use it out of convenience. */
1358 return flow_hash_symmetric_l4(&hash_flow, basis);
1362 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1364 ovs_assert(bond->balance == BM_TCP || bond->balance == BM_SLB);
1366 return (bond->balance == BM_TCP
1367 ? bond_hash_tcp(flow, vlan, bond->basis)
1368 : bond_hash_src(flow->dl_src, vlan, bond->basis));
1371 static struct bond_entry *
1372 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1375 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1378 static struct bond_slave *
1379 choose_output_slave(const struct bond *bond, const struct flow *flow,
1380 struct flow_wildcards *wc, uint16_t vlan, tag_type *tags)
1382 struct bond_entry *e;
1384 if (bond->lacp_status == LACP_CONFIGURED) {
1385 /* LACP has been configured on this bond but negotiations were
1386 * unsuccussful. Drop all traffic. */
1390 switch (bond->balance) {
1392 return bond->active_slave;
1395 if (bond->lacp_status != LACP_NEGOTIATED) {
1396 /* Must have LACP negotiations for TCP balanced bonds. */
1400 flow_mask_hash_fields(flow, wc, NX_HASH_FIELDS_SYMMETRIC_L4);
1405 flow_mask_hash_fields(flow, wc, NX_HASH_FIELDS_ETH_SRC);
1407 e = lookup_bond_entry(bond, flow, vlan);
1408 if (!e->slave || !e->slave->enabled) {
1409 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1410 struct bond_slave, hmap_node);
1411 if (!e->slave->enabled) {
1412 e->slave = bond->active_slave;
1414 e->tag = tag_create_random();
1424 static struct bond_slave *
1425 bond_choose_slave(const struct bond *bond)
1427 struct bond_slave *slave, *best;
1429 /* Find an enabled slave. */
1430 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1431 if (slave->enabled) {
1436 /* All interfaces are disabled. Find an interface that will be enabled
1437 * after its updelay expires. */
1439 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1440 if (slave->delay_expires != LLONG_MAX
1441 && slave->may_enable
1442 && (!best || slave->delay_expires < best->delay_expires)) {
1450 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1452 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1453 struct bond_slave *old_active_slave = bond->active_slave;
1455 bond->active_slave = bond_choose_slave(bond);
1456 if (bond->active_slave) {
1457 if (bond->active_slave->enabled) {
1458 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1459 bond->name, bond->active_slave->name);
1461 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1462 "remaining %lld ms updelay (since no interface was "
1463 "enabled)", bond->name, bond->active_slave->name,
1464 bond->active_slave->delay_expires - time_msec());
1465 bond_enable_slave(bond->active_slave, true, tags);
1468 if (!old_active_slave) {
1469 tag_set_add(tags, bond->no_slaves_tag);
1472 bond->send_learning_packets = true;
1473 } else if (old_active_slave) {
1474 VLOG_INFO_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1478 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1479 * there is no active slave. */
1481 bond_get_active_slave_tag(const struct bond *bond)
1483 return (bond->active_slave
1484 ? bond->active_slave->tag
1485 : bond->no_slaves_tag);
1488 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1489 * bond interface. */
1491 bond_update_fake_slave_stats(struct bond *bond)
1493 struct netdev_stats bond_stats;
1494 struct bond_slave *slave;
1495 struct netdev *bond_dev;
1497 memset(&bond_stats, 0, sizeof bond_stats);
1499 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1500 struct netdev_stats slave_stats;
1502 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1503 /* XXX: We swap the stats here because they are swapped back when
1504 * reported by the internal device. The reason for this is
1505 * internal devices normally represent packets going into the
1506 * system but when used as fake bond device they represent packets
1507 * leaving the system. We really should do this in the internal
1508 * device itself because changing it here reverses the counts from
1509 * the perspective of the switch. However, the internal device
1510 * doesn't know what type of device it represents so we have to do
1511 * it here for now. */
1512 bond_stats.tx_packets += slave_stats.rx_packets;
1513 bond_stats.tx_bytes += slave_stats.rx_bytes;
1514 bond_stats.rx_packets += slave_stats.tx_packets;
1515 bond_stats.rx_bytes += slave_stats.tx_bytes;
1519 if (!netdev_open(bond->name, "system", &bond_dev)) {
1520 netdev_set_stats(bond_dev, &bond_stats);
1521 netdev_close(bond_dev);