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 slave->netdev = netdev;
374 slave->name = xstrdup(netdev_get_name(netdev));
377 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
378 * then this function has no effect.
380 * Unregistering a slave invalidates all flows. */
382 bond_slave_unregister(struct bond *bond, const void *slave_)
384 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
391 bond_enable_slave(slave, false, NULL);
393 del_active = bond->active_slave == slave;
395 struct bond_entry *e;
396 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
397 if (e->slave == slave) {
405 hmap_remove(&bond->slaves, &slave->hmap_node);
406 /* Client owns 'slave->netdev'. */
413 bond_choose_active_slave(bond, &tags);
414 bond->send_learning_packets = true;
418 /* Should be called on each slave in 'bond' before bond_run() to indicate the
419 * results of lacp_slave_may_enable() on 'slave_'. */
421 bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
424 bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
427 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
428 * allow tagged flows to be invalidated.
430 * The caller should check bond_should_send_learning_packets() afterward. */
432 bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
434 struct bond_slave *slave;
435 bool is_tcp_hash = bond_is_tcp_hash(bond);
437 bond->lacp_negotiated = lacp_negotiated;
439 /* Update link status. */
440 if (bond->detect == BLSM_CARRIER
441 || time_msec() >= bond->miimon_next_update)
443 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
444 slave->up = bond_is_link_up(bond, slave->netdev);
446 bond->miimon_next_update = time_msec() + bond->miimon_interval;
449 /* Enable slaves based on link status and LACP feedback. */
450 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
451 bond_link_status_update(slave, tags);
453 if (!bond->active_slave || !bond->active_slave->enabled) {
454 bond_choose_active_slave(bond, tags);
457 /* Update fake bond interface stats. */
458 if (time_msec() >= bond->next_fake_iface_update) {
459 bond_update_fake_slave_stats(bond);
460 bond->next_fake_iface_update = time_msec() + 1000;
463 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
464 bond->bond_revalidate = true;
467 if (bond->bond_revalidate) {
468 bond->bond_revalidate = false;
470 bond_entry_reset(bond);
471 if (bond->balance != BM_STABLE) {
472 struct bond_slave *slave;
474 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
475 tag_set_add(tags, slave->tag);
478 tag_set_add(tags, bond->stb_tag);
480 tag_set_add(tags, bond->no_slaves_tag);
483 /* Invalidate any tags required by */
484 tag_set_union(tags, &bond->unixctl_tags);
485 tag_set_init(&bond->unixctl_tags);
488 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
490 bond_wait(struct bond *bond)
492 struct bond_slave *slave;
494 if (bond->detect == BLSM_CARRIER) {
495 netdev_monitor_poll_wait(bond->monitor);
497 poll_timer_wait_until(bond->miimon_next_update);
500 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
501 if (slave->delay_expires != LLONG_MAX) {
502 poll_timer_wait_until(slave->delay_expires);
506 if (bond->next_fake_iface_update != LLONG_MAX) {
507 poll_timer_wait_until(bond->next_fake_iface_update);
510 /* Ensure that any saved tags get revalidated right away. */
511 if (!tag_set_is_empty(&bond->unixctl_tags)) {
512 poll_immediate_wake();
515 /* We don't wait for bond->next_rebalance because rebalancing can only run
516 * at a flow account checkpoint. ofproto does checkpointing on its own
517 * schedule and bond_rebalance() gets called afterward, so we'd just be
518 * waking up for no purpose. */
521 /* MAC learning table interaction. */
524 may_send_learning_packets(const struct bond *bond)
526 return !bond->lacp_negotiated && bond->balance != BM_AB;
529 /* Returns true if 'bond' needs the client to send out packets to assist with
530 * MAC learning on 'bond'. If this function returns true, then the client
531 * should iterate through its MAC learning table for the bridge on which 'bond'
532 * is located. For each MAC that has been learned on a port other than 'bond',
533 * it should call bond_send_learning_packet().
535 * This function will only return true if 'bond' is in SLB mode and LACP is not
536 * negotiated. Otherwise sending learning packets isn't necessary.
538 * Calling this function resets the state that it checks. */
540 bond_should_send_learning_packets(struct bond *bond)
542 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
543 bond->send_learning_packets = false;
547 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
549 * See bond_should_send_learning_packets() for description of usage. */
551 bond_send_learning_packet(struct bond *bond,
552 const uint8_t eth_src[ETH_ADDR_LEN],
555 struct bond_slave *slave;
556 struct ofpbuf packet;
560 assert(may_send_learning_packets(bond));
561 if (!bond->active_slave) {
562 /* Nowhere to send the learning packet. */
566 memset(&flow, 0, sizeof flow);
567 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
568 slave = choose_output_slave(bond, &flow, vlan);
570 ofpbuf_init(&packet, 0);
571 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
574 eth_set_vlan_tci(&packet, htons(vlan));
576 error = netdev_send(slave->netdev, &packet);
577 ofpbuf_uninit(&packet);
582 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
583 * Ethernet destination address of 'eth_dst', should be admitted.
585 * The return value is one of the following:
587 * - BV_ACCEPT: Admit the packet.
589 * - BV_DROP: Drop the packet.
591 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
592 * Ethernet source address and VLAN. If there is none, or if the packet
593 * is on the learned port, then admit the packet. If a different port has
594 * been learned, however, drop the packet (and do not use it for MAC
598 bond_check_admissibility(struct bond *bond, const void *slave_,
599 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
601 /* Admit all packets if LACP has been negotiated, because that means that
602 * the remote switch is aware of the bond and will "do the right thing". */
603 if (bond->lacp_negotiated) {
607 /* Drop all multicast packets on inactive slaves. */
608 if (eth_addr_is_multicast(eth_dst)) {
609 *tags |= bond_get_active_slave_tag(bond);
610 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
615 /* Drop all packets for which we have learned a different input port,
616 * because we probably sent the packet on one slave and got it back on the
617 * other. Gratuitous ARP packets are an exception to this rule: the host
618 * has moved to another switch. The exception to the exception is if we
619 * locked the learning table to avoid reflections on bond slaves. */
620 return BV_DROP_IF_MOVED;
623 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
624 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
625 * NULL if the packet should be dropped because no slaves are enabled.
627 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
628 * should be a VID only (i.e. excluding the PCP bits). Second,
629 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
630 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
631 * packet belongs to (so for an access port it will be the access port's VLAN).
633 * Adds a tag to '*tags' that associates the flow with the returned slave.
636 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
637 uint16_t vlan, tag_type *tags)
639 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
641 *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
644 *tags |= bond->no_slaves_tag;
652 bond_is_balanced(const struct bond *bond)
654 return bond->balance == BM_SLB || bond->balance == BM_TCP;
657 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
659 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
663 if (bond_is_balanced(bond)) {
664 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
668 static struct bond_slave *
669 bond_slave_from_bal_node(struct list *bal)
671 return CONTAINER_OF(bal, struct bond_slave, bal_node);
675 log_bals(struct bond *bond, const struct list *bals)
677 if (VLOG_IS_DBG_ENABLED()) {
678 struct ds ds = DS_EMPTY_INITIALIZER;
679 const struct bond_slave *slave;
681 LIST_FOR_EACH (slave, bal_node, bals) {
683 ds_put_char(&ds, ',');
685 ds_put_format(&ds, " %s %"PRIu64"kB",
686 slave->name, slave->tx_bytes / 1024);
688 if (!slave->enabled) {
689 ds_put_cstr(&ds, " (disabled)");
691 if (!list_is_empty(&slave->entries)) {
692 struct bond_entry *e;
694 ds_put_cstr(&ds, " (");
695 LIST_FOR_EACH (e, list_node, &slave->entries) {
696 if (&e->list_node != list_front(&slave->entries)) {
697 ds_put_cstr(&ds, " + ");
699 ds_put_format(&ds, "h%td: %"PRIu64"kB",
700 e - bond->hash, e->tx_bytes / 1024);
702 ds_put_cstr(&ds, ")");
705 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
710 /* Shifts 'hash' from its current slave to 'to'. */
712 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
715 struct bond_slave *from = hash->slave;
716 struct bond *bond = from->bond;
717 uint64_t delta = hash->tx_bytes;
719 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
720 "from %s to %s (now carrying %"PRIu64"kB and "
721 "%"PRIu64"kB load, respectively)",
722 bond->name, delta / 1024, hash - bond->hash,
723 from->name, to->name,
724 (from->tx_bytes - delta) / 1024,
725 (to->tx_bytes + delta) / 1024);
727 /* Shift load away from 'from' to 'to'. */
728 from->tx_bytes -= delta;
729 to->tx_bytes += delta;
731 /* Arrange for flows to be revalidated. */
732 tag_set_add(set, hash->tag);
734 hash->tag = tag_create_random();
737 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
738 * given that doing so must decrease the ratio of the load on the two slaves by
739 * at least 0.1. Returns NULL if there is no appropriate entry.
741 * The list of entries isn't sorted. I don't know of a reason to prefer to
742 * shift away small hashes or large hashes. */
743 static struct bond_entry *
744 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
746 struct bond_entry *e;
748 if (list_is_short(&from->entries)) {
749 /* 'from' carries no more than one MAC hash, so shifting load away from
750 * it would be pointless. */
754 LIST_FOR_EACH (e, list_node, &from->entries) {
755 double old_ratio, new_ratio;
758 if (to_tx_bytes == 0) {
759 /* Nothing on the new slave, move it. */
764 old_ratio = (double)from->tx_bytes / to_tx_bytes;
765 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
766 if (old_ratio - new_ratio > 0.1) {
767 /* Would decrease the ratio, move it. */
775 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
778 insert_bal(struct list *bals, struct bond_slave *slave)
780 struct bond_slave *pos;
782 LIST_FOR_EACH (pos, bal_node, bals) {
783 if (slave->tx_bytes > pos->tx_bytes) {
787 list_insert(&pos->bal_node, &slave->bal_node);
790 /* Removes 'slave' from its current list and then inserts it into 'bals' so
791 * that descending order of 'tx_bytes' is maintained. */
793 reinsert_bal(struct list *bals, struct bond_slave *slave)
795 list_remove(&slave->bal_node);
796 insert_bal(bals, slave);
799 /* If 'bond' needs rebalancing, does so.
801 * The caller should have called bond_account() for each active flow, to ensure
802 * that flow data is consistently accounted at this point. */
804 bond_rebalance(struct bond *bond, struct tag_set *tags)
806 struct bond_slave *slave;
807 struct bond_entry *e;
810 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
813 bond->next_rebalance = time_msec() + bond->rebalance_interval;
815 /* Add each bond_entry to its slave's 'entries' list.
816 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
817 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
819 list_init(&slave->entries);
821 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
822 if (e->slave && e->tx_bytes) {
823 e->slave->tx_bytes += e->tx_bytes;
824 list_push_back(&e->slave->entries, &e->list_node);
828 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
830 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
831 * with a proper list sort algorithm. */
833 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
834 if (slave->enabled) {
835 insert_bal(&bals, slave);
838 log_bals(bond, &bals);
840 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
841 while (!list_is_short(&bals)) {
842 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
843 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
846 overload = from->tx_bytes - to->tx_bytes;
847 if (overload < to->tx_bytes >> 5 || overload < 100000) {
848 /* The extra load on 'from' (and all less-loaded slaves), compared
849 * to that of 'to' (the least-loaded slave), is less than ~3%, or
850 * it is less than ~1Mbps. No point in rebalancing. */
854 /* 'from' is carrying significantly more load than 'to', and that load
855 * is split across at least two different hashes. */
856 e = choose_entry_to_migrate(from, to->tx_bytes);
858 bond_shift_load(e, to, tags);
860 /* Delete element from from->entries.
862 * We don't add the element to to->hashes. That would only allow
863 * 'e' to be migrated to another slave in this rebalancing run, and
864 * there is no point in doing that. */
865 list_remove(&e->list_node);
867 /* Re-sort 'bals'. */
868 reinsert_bal(&bals, from);
869 reinsert_bal(&bals, to);
871 /* Can't usefully migrate anything away from 'from'.
872 * Don't reconsider it. */
873 list_remove(&from->bal_node);
877 /* Implement exponentially weighted moving average. A weight of 1/2 causes
878 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
879 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
880 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
888 /* Bonding unixctl user interface functions. */
891 bond_find(const char *name)
895 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
897 if (!strcmp(bond->name, name)) {
904 static struct bond_slave *
905 bond_lookup_slave(struct bond *bond, const char *slave_name)
907 struct bond_slave *slave;
909 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
910 if (!strcmp(slave->name, slave_name)) {
918 bond_unixctl_list(struct unixctl_conn *conn,
919 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
921 struct ds ds = DS_EMPTY_INITIALIZER;
922 const struct bond *bond;
924 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
926 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
927 const struct bond_slave *slave;
930 ds_put_format(&ds, "%s\t%s\t",
931 bond->name, bond_mode_to_string(bond->balance));
934 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
936 ds_put_cstr(&ds, ", ");
938 ds_put_cstr(&ds, slave->name);
940 ds_put_char(&ds, '\n');
942 unixctl_command_reply(conn, 200, ds_cstr(&ds));
947 bond_unixctl_show(struct unixctl_conn *conn,
948 const char *args, void *aux OVS_UNUSED)
950 struct ds ds = DS_EMPTY_INITIALIZER;
951 const struct bond_slave *slave;
952 const struct bond *bond;
954 bond = bond_find(args);
956 unixctl_command_reply(conn, 501, "no such bond");
960 ds_put_format(&ds, "bond_mode: %s\n",
961 bond_mode_to_string(bond->balance));
963 if (bond->balance != BM_AB) {
964 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
965 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
968 ds_put_format(&ds, "bond-detect-mode: %s\n",
969 bond->monitor ? "carrier" : "miimon");
971 if (!bond->monitor) {
972 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
973 bond->miimon_interval);
976 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
977 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
979 if (bond_is_balanced(bond)) {
980 ds_put_format(&ds, "next rebalance: %lld ms\n",
981 bond->next_rebalance - time_msec());
984 ds_put_format(&ds, "lacp_negotiated: %s\n",
985 bond->lacp_negotiated ? "true" : "false");
987 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
988 struct bond_entry *be;
992 ds_put_format(&ds, "\nslave %s: %s\n",
993 slave->name, slave->enabled ? "enabled" : "disabled");
994 if (slave == bond->active_slave) {
995 ds_put_cstr(&ds, "\tactive slave\n");
997 if (slave->delay_expires != LLONG_MAX) {
998 ds_put_format(&ds, "\t%s expires in %lld ms\n",
999 slave->enabled ? "downdelay" : "updelay",
1000 slave->delay_expires - time_msec());
1003 ds_put_format(&ds, "\tlacp_may_enable: %s\n",
1004 slave->lacp_may_enable ? "true" : "false");
1006 if (!bond_is_balanced(bond)) {
1011 memset(&flow, 0, sizeof flow);
1012 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1013 int hash = be - bond->hash;
1015 if (be->slave != slave) {
1019 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1020 hash, be->tx_bytes / 1024);
1022 if (bond->balance != BM_SLB) {
1026 /* XXX How can we list the MACs assigned to hashes? */
1029 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1034 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1035 void *aux OVS_UNUSED)
1037 char *args = (char *) args_;
1038 char *save_ptr = NULL;
1039 char *bond_s, *hash_s, *slave_s;
1041 struct bond_slave *slave;
1042 struct bond_entry *entry;
1045 bond_s = strtok_r(args, " ", &save_ptr);
1046 hash_s = strtok_r(NULL, " ", &save_ptr);
1047 slave_s = strtok_r(NULL, " ", &save_ptr);
1049 unixctl_command_reply(conn, 501,
1050 "usage: bond/migrate BOND HASH SLAVE");
1054 bond = bond_find(bond_s);
1056 unixctl_command_reply(conn, 501, "no such bond");
1060 if (bond->balance != BM_SLB) {
1061 unixctl_command_reply(conn, 501, "not an SLB bond");
1065 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1066 hash = atoi(hash_s) & BOND_MASK;
1068 unixctl_command_reply(conn, 501, "bad hash");
1072 slave = bond_lookup_slave(bond, slave_s);
1074 unixctl_command_reply(conn, 501, "no such slave");
1078 if (!slave->enabled) {
1079 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1083 entry = &bond->hash[hash];
1084 tag_set_add(&bond->unixctl_tags, entry->tag);
1085 entry->slave = slave;
1086 entry->tag = tag_create_random();
1087 unixctl_command_reply(conn, 200, "migrated");
1091 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1092 void *aux OVS_UNUSED)
1094 char *args = (char *) args_;
1095 char *save_ptr = NULL;
1096 char *bond_s, *slave_s;
1098 struct bond_slave *slave;
1100 bond_s = strtok_r(args, " ", &save_ptr);
1101 slave_s = strtok_r(NULL, " ", &save_ptr);
1103 unixctl_command_reply(conn, 501,
1104 "usage: bond/set-active-slave BOND SLAVE");
1108 bond = bond_find(bond_s);
1110 unixctl_command_reply(conn, 501, "no such bond");
1114 slave = bond_lookup_slave(bond, slave_s);
1116 unixctl_command_reply(conn, 501, "no such slave");
1120 if (!slave->enabled) {
1121 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1125 if (bond->active_slave != slave) {
1126 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1127 bond->active_slave = slave;
1128 bond->active_slave->tag = tag_create_random();
1129 VLOG_INFO("bond %s: active interface is now %s",
1130 bond->name, slave->name);
1131 bond->send_learning_packets = true;
1132 unixctl_command_reply(conn, 200, "done");
1134 unixctl_command_reply(conn, 200, "no change");
1139 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1141 char *args = (char *) args_;
1142 char *save_ptr = NULL;
1143 char *bond_s, *slave_s;
1145 struct bond_slave *slave;
1147 bond_s = strtok_r(args, " ", &save_ptr);
1148 slave_s = strtok_r(NULL, " ", &save_ptr);
1150 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1151 enable ? "enable" : "disable");
1152 unixctl_command_reply(conn, 501, usage);
1157 bond = bond_find(bond_s);
1159 unixctl_command_reply(conn, 501, "no such bond");
1163 slave = bond_lookup_slave(bond, slave_s);
1165 unixctl_command_reply(conn, 501, "no such slave");
1169 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1170 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1174 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1175 void *aux OVS_UNUSED)
1177 enable_slave(conn, args, true);
1181 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1182 void *aux OVS_UNUSED)
1184 enable_slave(conn, args, false);
1188 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1189 void *aux OVS_UNUSED)
1191 char *args = (char *) args_;
1192 uint8_t mac[ETH_ADDR_LEN];
1196 char *mac_s, *vlan_s;
1197 char *save_ptr = NULL;
1199 mac_s = strtok_r(args, " ", &save_ptr);
1200 vlan_s = strtok_r(NULL, " ", &save_ptr);
1203 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1204 unixctl_command_reply(conn, 501, "invalid vlan");
1208 vlan = OFP_VLAN_NONE;
1211 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1212 == ETH_ADDR_SCAN_COUNT) {
1213 hash = bond_hash_src(mac, vlan) & BOND_MASK;
1215 hash_cstr = xasprintf("%u", hash);
1216 unixctl_command_reply(conn, 200, hash_cstr);
1219 unixctl_command_reply(conn, 501, "invalid mac");
1226 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1227 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1228 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1229 unixctl_command_register("bond/set-active-slave",
1230 bond_unixctl_set_active_slave, NULL);
1231 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1233 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1235 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1239 bond_entry_reset(struct bond *bond)
1241 if (bond->balance != BM_AB) {
1242 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1245 bond->hash = xmalloc(hash_len);
1247 memset(bond->hash, 0, hash_len);
1249 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1256 static struct bond_slave *
1257 bond_slave_lookup(struct bond *bond, const void *slave_)
1259 struct bond_slave *slave;
1261 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1263 if (slave->aux == slave_) {
1272 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1274 return (bond->detect == BLSM_CARRIER
1275 ? netdev_get_carrier(netdev)
1276 : netdev_get_miimon(netdev));
1280 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1282 struct bond *bond = slave->bond;
1283 slave->delay_expires = LLONG_MAX;
1284 if (enable != slave->enabled) {
1285 slave->enabled = enable;
1286 if (!slave->enabled) {
1287 VLOG_WARN("interface %s: disabled", slave->name);
1289 tag_set_add(tags, slave->tag);
1292 VLOG_WARN("interface %s: enabled", slave->name);
1293 slave->tag = tag_create_random();
1296 if (bond->balance == BM_STABLE) {
1297 bond->bond_revalidate = true;
1303 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1305 struct bond *bond = slave->bond;
1308 up = slave->up && slave->lacp_may_enable;
1309 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1310 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1311 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1312 slave->name, up ? "up" : "down");
1313 if (up == slave->enabled) {
1314 slave->delay_expires = LLONG_MAX;
1315 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1316 slave->name, up ? "disabled" : "enabled");
1318 int delay = (bond->lacp_negotiated ? 0
1319 : up ? bond->updelay : bond->downdelay);
1320 slave->delay_expires = time_msec() + delay;
1322 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1325 up ? "enabled" : "disabled",
1332 if (time_msec() >= slave->delay_expires) {
1333 bond_enable_slave(slave, up, tags);
1338 bond_is_tcp_hash(const struct bond *bond)
1340 return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
1341 && bond->lacp_negotiated;
1345 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan)
1347 return hash_bytes(mac, ETH_ADDR_LEN, vlan);
1351 bond_hash_tcp(const struct flow *flow, uint16_t vlan)
1353 struct flow hash_flow = *flow;
1354 hash_flow.vlan_tci = vlan;
1356 /* The symmetric quality of this hash function is not required, but
1357 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1358 * purposes, so we use it out of convenience. */
1359 return flow_hash_symmetric_l4(&hash_flow, 0);
1363 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1365 assert(bond->balance != BM_AB);
1367 return (bond_is_tcp_hash(bond)
1368 ? bond_hash_tcp(flow, vlan)
1369 : bond_hash_src(flow->dl_src, vlan));
1372 static struct bond_entry *
1373 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1376 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1379 /* This function uses Highest Random Weight hashing to choose an output slave.
1380 * This approach only reassigns a minimal number of flows when slaves are
1381 * enabled or disabled. Unfortunately, it has O(n) performance against the
1382 * number of slaves. There exist algorithms which are O(1), but have slightly
1383 * more complex implementations and require the use of memory. This may need
1384 * to be reimplemented if it becomes a performance bottleneck. */
1385 static struct bond_slave *
1386 choose_stb_slave(const struct bond *bond, const struct flow *flow,
1389 struct bond_slave *best, *slave;
1390 uint32_t best_hash, flow_hash;
1394 flow_hash = bond_hash(bond, flow, vlan);
1395 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1396 if (slave->enabled) {
1399 hash = hash_2words(flow_hash, slave->stb_id);
1400 if (!best || hash > best_hash) {
1410 static struct bond_slave *
1411 choose_output_slave(const struct bond *bond, const struct flow *flow,
1414 struct bond_entry *e;
1416 switch (bond->balance) {
1418 return bond->active_slave;
1421 return choose_stb_slave(bond, flow, vlan);
1424 e = lookup_bond_entry(bond, flow, vlan);
1425 if (!e->slave || !e->slave->enabled) {
1426 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1427 struct bond_slave, hmap_node);
1428 if (!e->slave->enabled) {
1429 e->slave = bond->active_slave;
1431 e->tag = tag_create_random();
1440 static struct bond_slave *
1441 bond_choose_slave(const struct bond *bond)
1443 struct bond_slave *slave, *best;
1445 /* Find an enabled slave. */
1446 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1447 if (slave->enabled) {
1452 /* All interfaces are disabled. Find an interface that will be enabled
1453 * after its updelay expires. */
1455 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1456 if (slave->delay_expires != LLONG_MAX
1457 && slave->lacp_may_enable
1458 && (!best || slave->delay_expires < best->delay_expires)) {
1466 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1468 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1469 struct bond_slave *old_active_slave = bond->active_slave;
1471 bond->active_slave = bond_choose_slave(bond);
1472 if (bond->active_slave) {
1473 if (bond->active_slave->enabled) {
1474 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1475 bond->name, bond->active_slave->name);
1477 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1478 "remaining %lld ms updelay (since no interface was "
1479 "enabled)", bond->name, bond->active_slave->name,
1480 bond->active_slave->delay_expires - time_msec());
1481 bond_enable_slave(bond->active_slave, true, tags);
1484 if (!old_active_slave) {
1485 tag_set_add(tags, bond->no_slaves_tag);
1488 bond->send_learning_packets = true;
1489 } else if (old_active_slave) {
1490 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1494 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1495 * there is no active slave. */
1497 bond_get_active_slave_tag(const struct bond *bond)
1499 return (bond->active_slave
1500 ? bond->active_slave->tag
1501 : bond->no_slaves_tag);
1504 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1505 * bond interface. */
1507 bond_update_fake_slave_stats(struct bond *bond)
1509 struct netdev_stats bond_stats;
1510 struct bond_slave *slave;
1511 struct netdev *bond_dev;
1513 memset(&bond_stats, 0, sizeof bond_stats);
1515 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1516 struct netdev_stats slave_stats;
1518 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1519 /* XXX: We swap the stats here because they are swapped back when
1520 * reported by the internal device. The reason for this is
1521 * internal devices normally represent packets going into the
1522 * system but when used as fake bond device they represent packets
1523 * leaving the system. We really should do this in the internal
1524 * device itself because changing it here reverses the counts from
1525 * the perspective of the switch. However, the internal device
1526 * doesn't know what type of device it represents so we have to do
1527 * it here for now. */
1528 bond_stats.tx_packets += slave_stats.rx_packets;
1529 bond_stats.tx_bytes += slave_stats.rx_bytes;
1530 bond_stats.rx_packets += slave_stats.tx_packets;
1531 bond_stats.rx_bytes += slave_stats.tx_bytes;
1535 if (!netdev_open_default(bond->name, &bond_dev)) {
1536 netdev_set_stats(bond_dev, &bond_stats);
1537 netdev_close(bond_dev);