2 * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
26 #include "dynamic-string.h"
34 #include "poll-loop.h"
40 VLOG_DEFINE_THIS_MODULE(bond);
42 /* Bit-mask for hashing a flow down to a bucket.
43 * There are (BOND_MASK + 1) buckets. */
44 #define BOND_MASK 0xff
46 /* A hash bucket for mapping a flow to a slave.
47 * "struct bond" has an array of (BOND_MASK + 1) of these. */
49 struct bond_slave *slave; /* Assigned slave, NULL if unassigned. */
50 uint64_t tx_bytes; /* Count of bytes recently transmitted. */
51 tag_type tag; /* Tag for entry<->slave association. */
52 struct list list_node; /* In bond_slave's 'entries' list. */
55 /* A bond slave, that is, one of the links comprising a bond. */
57 struct hmap_node hmap_node; /* In struct bond's slaves hmap. */
58 struct bond *bond; /* The bond that contains this slave. */
59 void *aux; /* Client-provided handle for this slave. */
61 struct netdev *netdev; /* Network device, owned by the client. */
62 char *name; /* Name (a copy of netdev_get_name(netdev)). */
65 long long delay_expires; /* Time after which 'enabled' may change. */
66 bool up; /* Last link status read from netdev. */
67 bool enabled; /* May be chosen for flows? */
68 bool lacp_may_enable; /* LACP considers this interface bondable. */
69 tag_type tag; /* Tag associated with this slave. */
71 /* Rebalancing info. Used only by bond_rebalance(). */
72 struct list bal_node; /* In bond_rebalance()'s 'bals' list. */
73 struct list entries; /* 'struct bond_entry's assigned here. */
74 uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
76 /* BM_STABLE specific bonding info. */
77 uint16_t stb_id; /* ID used for 'stb_slaves' ordering. */
78 size_t stb_idx; /* Index in 'bond''s 'stb_slaves' array.
79 Undefined value if participating in a
80 BTM_STABLE bond or not enabled. */
83 /* A bond, that is, a set of network devices grouped to improve performance or
86 struct hmap_node hmap_node; /* In 'all_bonds' hmap. */
87 char *name; /* Name provided by client. */
93 enum bond_mode balance; /* Balancing mode, one of BM_*. */
94 struct bond_slave *active_slave;
95 tag_type no_slaves_tag; /* Tag for flows when all slaves disabled. */
96 int updelay, downdelay; /* Delay before slave goes up/down, in ms. */
97 bool lacp_negotiated; /* LACP negotiations were successful. */
99 /* SLB specific bonding info. */
100 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
101 int rebalance_interval; /* Interval between rebalances, in ms. */
102 long long int next_rebalance; /* Next rebalancing time. */
103 bool send_learning_packets;
105 /* BM_STABLE specific bonding info. */
106 struct bond_slave **stb_slaves; /* Ordered list of enabled slaves. */
107 size_t n_stb_slaves; /* Number of slaves in 'stb_slaves'. */
108 size_t len_stb_slaves; /* Slaves allocated in 'stb_slaves'. */
109 bool stb_need_sort; /* True if stb_slaves is not sorted. */
113 enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
114 struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
115 long long int miimon_interval; /* Miimon status refresh interval. */
116 long long int miimon_next_update; /* Time of next miimon update. */
118 /* Legacy compatibility. */
119 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
121 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
122 * where we can't otherwise provide revalidation feedback to the client.
123 * That's only unixctl commands now; I hope no other cases will arise. */
124 struct tag_set unixctl_tags;
127 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
129 static void bond_entry_reset(struct bond *);
130 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
131 static bool bond_is_link_up(struct bond *, struct netdev *);
132 static void bond_enable_slave(struct bond_slave *, bool enable,
134 static bool bond_stb_sort(struct bond *);
135 static void bond_stb_enable_slave(struct bond_slave *);
136 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
137 static void bond_choose_active_slave(struct bond *, struct tag_set *);
138 static bool bond_is_tcp_hash(const struct bond *);
139 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
141 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan);
142 static struct bond_entry *lookup_bond_entry(const struct bond *,
145 static tag_type bond_get_active_slave_tag(const struct bond *);
146 static struct bond_slave *choose_output_slave(const struct bond *,
149 static void bond_update_fake_slave_stats(struct bond *);
151 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
152 * stores the mode in '*balance' and returns true. Otherwise returns false
153 * without modifying '*balance'. */
155 bond_mode_from_string(enum bond_mode *balance, const char *s)
157 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
159 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
161 } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
162 *balance = BM_STABLE;
163 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
171 /* Returns a string representing 'balance'. */
173 bond_mode_to_string(enum bond_mode balance) {
176 return "balance-tcp";
178 return "balance-slb";
182 return "active-backup";
187 /* Attempts to parse 's' as the name of a bond link status detection mode. If
188 * successful, stores the mode in '*detect' and returns true. Otherwise
189 * returns false without modifying '*detect'. */
191 bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
193 if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
194 *detect = BLSM_CARRIER;
195 } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
196 *detect = BLSM_MIIMON;
203 /* Returns a string representing 'detect'. */
205 bond_detect_mode_to_string(enum bond_detect_mode detect)
216 /* Creates and returns a new bond whose configuration is initially taken from
219 * The caller should register each slave on the new bond by calling
220 * bond_slave_register(). */
222 bond_create(const struct bond_settings *s)
226 bond = xzalloc(sizeof *bond);
227 hmap_init(&bond->slaves);
228 bond->no_slaves_tag = tag_create_random();
229 bond->miimon_next_update = LLONG_MAX;
230 bond->next_fake_iface_update = LLONG_MAX;
232 bond_reconfigure(bond, s);
234 tag_set_init(&bond->unixctl_tags);
241 bond_destroy(struct bond *bond)
243 struct bond_slave *slave, *next_slave;
249 hmap_remove(&all_bonds, &bond->hmap_node);
251 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
252 hmap_remove(&bond->slaves, &slave->hmap_node);
253 /* Client owns 'slave->netdev'. */
257 hmap_destroy(&bond->slaves);
261 netdev_monitor_destroy(bond->monitor);
267 /* Updates 'bond''s overall configuration to 's'.
269 * The caller should register each slave on 'bond' by calling
270 * bond_slave_register(). This is optional if none of the slaves'
271 * configuration has changed. In any case it can't hurt.
273 * Returns true if the configuration has changed in such a way that requires
277 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
279 bool revalidate = false;
281 if (!bond->name || strcmp(bond->name, s->name)) {
283 hmap_remove(&all_bonds, &bond->hmap_node);
286 bond->name = xstrdup(s->name);
287 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
290 bond->detect = s->detect;
291 bond->miimon_interval = s->miimon_interval;
292 bond->updelay = s->up_delay;
293 bond->downdelay = s->down_delay;
294 bond->rebalance_interval = s->rebalance_interval;
296 if (bond->balance != s->balance) {
297 bond->balance = s->balance;
301 if (bond->detect == BLSM_CARRIER) {
302 struct bond_slave *slave;
304 if (!bond->monitor) {
305 bond->monitor = netdev_monitor_create();
308 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
309 netdev_monitor_add(bond->monitor, slave->netdev);
312 netdev_monitor_destroy(bond->monitor);
313 bond->monitor = NULL;
315 if (bond->miimon_next_update == LLONG_MAX) {
316 bond->miimon_next_update = time_msec() + bond->miimon_interval;
321 if (bond->next_fake_iface_update == LLONG_MAX) {
322 bond->next_fake_iface_update = time_msec();
325 bond->next_fake_iface_update = LLONG_MAX;
328 if (bond->balance != BM_STABLE) {
329 free(bond->stb_slaves);
330 bond->stb_slaves = NULL;
331 } else if (!bond->stb_slaves) {
332 struct bond_slave *slave;
334 bond->n_stb_slaves = 0;
335 bond->len_stb_slaves = 0;
336 bond->stb_slaves = NULL;
338 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
339 if (slave->enabled) {
340 bond_stb_enable_slave(slave);
345 if (bond->balance == BM_AB || !bond->hash || revalidate) {
346 bond_entry_reset(bond);
352 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
353 * arbitrary client-provided pointer that uniquely identifies a slave within a
354 * bond. If 'slave_' already exists within 'bond' then this function
355 * reconfigures the existing slave.
357 * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
358 * across restarts and distributed vswitch instances. It should be unique per
359 * slave, and preferably consistent across restarts and reconfigurations.
361 * 'netdev' must be the network device that 'slave_' represents. It is owned
362 * by the client, so the client must not close it before either unregistering
363 * 'slave_' or destroying 'bond'.
366 bond_slave_register(struct bond *bond, void *slave_, uint16_t stb_id,
367 struct netdev *netdev)
369 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
372 slave = xzalloc(sizeof *slave);
374 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
377 slave->delay_expires = LLONG_MAX;
378 slave->up = bond_is_link_up(bond, netdev);
379 slave->enabled = false;
380 bond_enable_slave(slave, slave->up, NULL);
383 if (slave->stb_id != stb_id) {
384 bond->stb_need_sort = true;
385 slave->stb_id = stb_id;
388 slave->netdev = netdev;
390 slave->name = xstrdup(netdev_get_name(netdev));
393 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
394 * then this function has no effect.
396 * Unregistering a slave invalidates all flows. */
398 bond_slave_unregister(struct bond *bond, const void *slave_)
400 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
407 bond_enable_slave(slave, false, NULL);
409 del_active = bond->active_slave == slave;
411 struct bond_entry *e;
412 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
413 if (e->slave == slave) {
421 hmap_remove(&bond->slaves, &slave->hmap_node);
422 /* Client owns 'slave->netdev'. */
429 bond_choose_active_slave(bond, &tags);
430 bond->send_learning_packets = true;
434 /* Should be called on each slave in 'bond' before bond_run() to indicate the
435 * results of lacp_slave_may_enable() on 'slave_'. */
437 bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
440 bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
443 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
444 * allow tagged flows to be invalidated.
446 * The caller should check bond_should_send_learning_packets() afterward. */
448 bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
450 struct bond_slave *slave;
451 bool is_tcp_hash = bond_is_tcp_hash(bond);
453 bond->lacp_negotiated = lacp_negotiated;
455 /* Update link status. */
456 if (bond->detect == BLSM_CARRIER
457 || time_msec() >= bond->miimon_next_update)
459 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
460 slave->up = bond_is_link_up(bond, slave->netdev);
462 bond->miimon_next_update = time_msec() + bond->miimon_interval;
465 /* Enable slaves based on link status and LACP feedback. */
466 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
467 bond_link_status_update(slave, tags);
469 if (!bond->active_slave || !bond->active_slave->enabled) {
470 bond_choose_active_slave(bond, tags);
473 /* Update fake bond interface stats. */
474 if (time_msec() >= bond->next_fake_iface_update) {
475 bond_update_fake_slave_stats(bond);
476 bond->next_fake_iface_update = time_msec() + 1000;
479 if (bond_stb_sort(bond) || is_tcp_hash != bond_is_tcp_hash(bond)) {
480 struct bond_slave *slave;
482 bond_entry_reset(bond);
483 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
484 tag_set_add(tags, slave->tag);
488 /* Invalidate any tags required by */
489 tag_set_union(tags, &bond->unixctl_tags);
490 tag_set_init(&bond->unixctl_tags);
493 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
495 bond_wait(struct bond *bond)
497 struct bond_slave *slave;
499 if (bond->detect == BLSM_CARRIER) {
500 netdev_monitor_poll_wait(bond->monitor);
502 poll_timer_wait_until(bond->miimon_next_update);
505 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
506 if (slave->delay_expires != LLONG_MAX) {
507 poll_timer_wait_until(slave->delay_expires);
511 if (bond->next_fake_iface_update != LLONG_MAX) {
512 poll_timer_wait_until(bond->next_fake_iface_update);
515 /* Ensure that any saved tags get revalidated right away. */
516 if (!tag_set_is_empty(&bond->unixctl_tags)) {
517 poll_immediate_wake();
520 /* We don't wait for bond->next_rebalance because rebalancing can only run
521 * at a flow account checkpoint. ofproto does checkpointing on its own
522 * schedule and bond_rebalance() gets called afterward, so we'd just be
523 * waking up for no purpose. */
526 /* MAC learning table interaction. */
529 may_send_learning_packets(const struct bond *bond)
531 return !bond->lacp_negotiated && bond->balance != BM_AB;
534 /* Returns true if 'bond' needs the client to send out packets to assist with
535 * MAC learning on 'bond'. If this function returns true, then the client
536 * should iterate through its MAC learning table for the bridge on which 'bond'
537 * is located. For each MAC that has been learned on a port other than 'bond',
538 * it should call bond_send_learning_packet().
540 * This function will only return true if 'bond' is in SLB mode and LACP is not
541 * negotiated. Otherwise sending learning packets isn't necessary.
543 * Calling this function resets the state that it checks. */
545 bond_should_send_learning_packets(struct bond *bond)
547 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
548 bond->send_learning_packets = false;
552 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
554 * See bond_should_send_learning_packets() for description of usage. */
556 bond_send_learning_packet(struct bond *bond,
557 const uint8_t eth_src[ETH_ADDR_LEN],
560 struct bond_slave *slave;
561 struct ofpbuf packet;
565 assert(may_send_learning_packets(bond));
566 if (!bond->active_slave) {
567 /* Nowhere to send the learning packet. */
571 memset(&flow, 0, sizeof flow);
572 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
573 slave = choose_output_slave(bond, &flow, vlan);
575 ofpbuf_init(&packet, 0);
576 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
579 eth_set_vlan_tci(&packet, htons(vlan));
581 error = netdev_send(slave->netdev, &packet);
582 ofpbuf_uninit(&packet);
587 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
588 * Ethernet destination address of 'eth_dst', should be admitted.
590 * The return value is one of the following:
592 * - BV_ACCEPT: Admit the packet.
594 * - BV_DROP: Drop the packet.
596 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
597 * Ethernet source address and VLAN. If there is none, or if the packet
598 * is on the learned port, then admit the packet. If a different port has
599 * been learned, however, drop the packet (and do not use it for MAC
603 bond_check_admissibility(struct bond *bond, const void *slave_,
604 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
606 /* Admit all packets if LACP has been negotiated, because that means that
607 * the remote switch is aware of the bond and will "do the right thing". */
608 if (bond->lacp_negotiated) {
612 /* Drop all multicast packets on inactive slaves. */
613 if (eth_addr_is_multicast(eth_dst)) {
614 *tags |= bond_get_active_slave_tag(bond);
615 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
620 /* Drop all packets for which we have learned a different input port,
621 * because we probably sent the packet on one slave and got it back on the
622 * other. Gratuitous ARP packets are an exception to this rule: the host
623 * has moved to another switch. The exception to the exception is if we
624 * locked the learning table to avoid reflections on bond slaves. */
625 return BV_DROP_IF_MOVED;
628 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
629 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
630 * NULL if the packet should be dropped because no slaves are enabled.
632 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
633 * should be a VID only (i.e. excluding the PCP bits). Second,
634 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
635 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
636 * packet belongs to (so for an access port it will be the access port's VLAN).
638 * Adds a tag to '*tags' that associates the flow with the returned slave.
641 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
642 uint16_t vlan, tag_type *tags)
644 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
649 *tags |= bond->no_slaves_tag;
657 bond_is_balanced(const struct bond *bond)
659 return bond->balance == BM_SLB || bond->balance == BM_TCP;
662 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
664 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
668 if (bond_is_balanced(bond)) {
669 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
673 static struct bond_slave *
674 bond_slave_from_bal_node(struct list *bal)
676 return CONTAINER_OF(bal, struct bond_slave, bal_node);
680 log_bals(struct bond *bond, const struct list *bals)
682 if (VLOG_IS_DBG_ENABLED()) {
683 struct ds ds = DS_EMPTY_INITIALIZER;
684 const struct bond_slave *slave;
686 LIST_FOR_EACH (slave, bal_node, bals) {
688 ds_put_char(&ds, ',');
690 ds_put_format(&ds, " %s %"PRIu64"kB",
691 slave->name, slave->tx_bytes / 1024);
693 if (!slave->enabled) {
694 ds_put_cstr(&ds, " (disabled)");
696 if (!list_is_empty(&slave->entries)) {
697 struct bond_entry *e;
699 ds_put_cstr(&ds, " (");
700 LIST_FOR_EACH (e, list_node, &slave->entries) {
701 if (&e->list_node != list_front(&slave->entries)) {
702 ds_put_cstr(&ds, " + ");
704 ds_put_format(&ds, "h%td: %"PRIu64"kB",
705 e - bond->hash, e->tx_bytes / 1024);
707 ds_put_cstr(&ds, ")");
710 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
715 /* Shifts 'hash' from its current slave to 'to'. */
717 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
720 struct bond_slave *from = hash->slave;
721 struct bond *bond = from->bond;
722 uint64_t delta = hash->tx_bytes;
724 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
725 "from %s to %s (now carrying %"PRIu64"kB and "
726 "%"PRIu64"kB load, respectively)",
727 bond->name, delta / 1024, hash - bond->hash,
728 from->name, to->name,
729 (from->tx_bytes - delta) / 1024,
730 (to->tx_bytes + delta) / 1024);
732 /* Shift load away from 'from' to 'to'. */
733 from->tx_bytes -= delta;
734 to->tx_bytes += delta;
736 /* Arrange for flows to be revalidated. */
737 tag_set_add(set, hash->tag);
739 hash->tag = tag_create_random();
742 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
743 * given that doing so must decrease the ratio of the load on the two slaves by
744 * at least 0.1. Returns NULL if there is no appropriate entry.
746 * The list of entries isn't sorted. I don't know of a reason to prefer to
747 * shift away small hashes or large hashes. */
748 static struct bond_entry *
749 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
751 struct bond_entry *e;
753 if (list_is_short(&from->entries)) {
754 /* 'from' carries no more than one MAC hash, so shifting load away from
755 * it would be pointless. */
759 LIST_FOR_EACH (e, list_node, &from->entries) {
760 double old_ratio, new_ratio;
763 if (to_tx_bytes == 0) {
764 /* Nothing on the new slave, move it. */
769 old_ratio = (double)from->tx_bytes / to_tx_bytes;
770 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
771 if (old_ratio - new_ratio > 0.1) {
772 /* Would decrease the ratio, move it. */
780 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
783 insert_bal(struct list *bals, struct bond_slave *slave)
785 struct bond_slave *pos;
787 LIST_FOR_EACH (pos, bal_node, bals) {
788 if (slave->tx_bytes > pos->tx_bytes) {
792 list_insert(&pos->bal_node, &slave->bal_node);
795 /* Removes 'slave' from its current list and then inserts it into 'bals' so
796 * that descending order of 'tx_bytes' is maintained. */
798 reinsert_bal(struct list *bals, struct bond_slave *slave)
800 list_remove(&slave->bal_node);
801 insert_bal(bals, slave);
804 /* If 'bond' needs rebalancing, does so.
806 * The caller should have called bond_account() for each active flow, to ensure
807 * that flow data is consistently accounted at this point. */
809 bond_rebalance(struct bond *bond, struct tag_set *tags)
811 struct bond_slave *slave;
812 struct bond_entry *e;
815 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
818 bond->next_rebalance = time_msec() + bond->rebalance_interval;
820 /* Add each bond_entry to its slave's 'entries' list.
821 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
822 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
824 list_init(&slave->entries);
826 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
827 if (e->slave && e->tx_bytes) {
828 e->slave->tx_bytes += e->tx_bytes;
829 list_push_back(&e->slave->entries, &e->list_node);
833 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
835 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
836 * with a proper list sort algorithm. */
838 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
839 if (slave->enabled) {
840 insert_bal(&bals, slave);
843 log_bals(bond, &bals);
845 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
846 while (!list_is_short(&bals)) {
847 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
848 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
851 overload = from->tx_bytes - to->tx_bytes;
852 if (overload < to->tx_bytes >> 5 || overload < 100000) {
853 /* The extra load on 'from' (and all less-loaded slaves), compared
854 * to that of 'to' (the least-loaded slave), is less than ~3%, or
855 * it is less than ~1Mbps. No point in rebalancing. */
859 /* 'from' is carrying significantly more load than 'to', and that load
860 * is split across at least two different hashes. */
861 e = choose_entry_to_migrate(from, to->tx_bytes);
863 bond_shift_load(e, to, tags);
865 /* Delete element from from->entries.
867 * We don't add the element to to->hashes. That would only allow
868 * 'e' to be migrated to another slave in this rebalancing run, and
869 * there is no point in doing that. */
870 list_remove(&e->list_node);
872 /* Re-sort 'bals'. */
873 reinsert_bal(&bals, from);
874 reinsert_bal(&bals, to);
876 /* Can't usefully migrate anything away from 'from'.
877 * Don't reconsider it. */
878 list_remove(&from->bal_node);
882 /* Implement exponentially weighted moving average. A weight of 1/2 causes
883 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
884 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
885 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
893 /* Bonding unixctl user interface functions. */
896 bond_find(const char *name)
900 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
902 if (!strcmp(bond->name, name)) {
909 static struct bond_slave *
910 bond_lookup_slave(struct bond *bond, const char *slave_name)
912 struct bond_slave *slave;
914 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
915 if (!strcmp(slave->name, slave_name)) {
923 bond_unixctl_list(struct unixctl_conn *conn,
924 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
926 struct ds ds = DS_EMPTY_INITIALIZER;
927 const struct bond *bond;
929 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
931 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
932 const struct bond_slave *slave;
935 ds_put_format(&ds, "%s\t%s\t",
936 bond->name, bond_mode_to_string(bond->balance));
939 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
941 ds_put_cstr(&ds, ", ");
943 ds_put_cstr(&ds, slave->name);
945 ds_put_char(&ds, '\n');
947 unixctl_command_reply(conn, 200, ds_cstr(&ds));
952 bond_unixctl_show(struct unixctl_conn *conn,
953 const char *args, void *aux OVS_UNUSED)
955 struct ds ds = DS_EMPTY_INITIALIZER;
956 const struct bond_slave *slave;
957 const struct bond *bond;
959 bond = bond_find(args);
961 unixctl_command_reply(conn, 501, "no such bond");
965 ds_put_format(&ds, "bond_mode: %s\n",
966 bond_mode_to_string(bond->balance));
968 if (bond->balance != BM_AB) {
969 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
970 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
973 ds_put_format(&ds, "bond-detect-mode: %s\n",
974 bond->monitor ? "carrier" : "miimon");
976 if (!bond->monitor) {
977 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
978 bond->miimon_interval);
981 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
982 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
984 if (bond_is_balanced(bond)) {
985 ds_put_format(&ds, "next rebalance: %lld ms\n",
986 bond->next_rebalance - time_msec());
989 ds_put_format(&ds, "lacp_negotiated: %s\n",
990 bond->lacp_negotiated ? "true" : "false");
992 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
993 struct bond_entry *be;
997 ds_put_format(&ds, "\nslave %s: %s\n",
998 slave->name, slave->enabled ? "enabled" : "disabled");
999 if (slave == bond->active_slave) {
1000 ds_put_cstr(&ds, "\tactive slave\n");
1002 if (slave->delay_expires != LLONG_MAX) {
1003 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1004 slave->enabled ? "downdelay" : "updelay",
1005 slave->delay_expires - time_msec());
1008 ds_put_format(&ds, "lacp_may_enable: %s\n",
1009 slave->lacp_may_enable ? "true" : "false");
1011 if (!bond_is_balanced(bond)) {
1016 memset(&flow, 0, sizeof flow);
1017 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1018 int hash = be - bond->hash;
1020 if (be->slave != slave) {
1024 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1025 hash, be->tx_bytes / 1024);
1027 if (bond->balance != BM_SLB) {
1031 /* XXX How can we list the MACs assigned to hashes? */
1034 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1039 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1040 void *aux OVS_UNUSED)
1042 char *args = (char *) args_;
1043 char *save_ptr = NULL;
1044 char *bond_s, *hash_s, *slave_s;
1046 struct bond_slave *slave;
1047 struct bond_entry *entry;
1050 bond_s = strtok_r(args, " ", &save_ptr);
1051 hash_s = strtok_r(NULL, " ", &save_ptr);
1052 slave_s = strtok_r(NULL, " ", &save_ptr);
1054 unixctl_command_reply(conn, 501,
1055 "usage: bond/migrate BOND HASH SLAVE");
1059 bond = bond_find(bond_s);
1061 unixctl_command_reply(conn, 501, "no such bond");
1065 if (bond->balance != BM_SLB) {
1066 unixctl_command_reply(conn, 501, "not an SLB bond");
1070 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1071 hash = atoi(hash_s) & BOND_MASK;
1073 unixctl_command_reply(conn, 501, "bad hash");
1077 slave = bond_lookup_slave(bond, slave_s);
1079 unixctl_command_reply(conn, 501, "no such slave");
1083 if (!slave->enabled) {
1084 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1088 entry = &bond->hash[hash];
1089 tag_set_add(&bond->unixctl_tags, entry->tag);
1090 entry->slave = slave;
1091 entry->tag = tag_create_random();
1092 unixctl_command_reply(conn, 200, "migrated");
1096 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1097 void *aux OVS_UNUSED)
1099 char *args = (char *) args_;
1100 char *save_ptr = NULL;
1101 char *bond_s, *slave_s;
1103 struct bond_slave *slave;
1105 bond_s = strtok_r(args, " ", &save_ptr);
1106 slave_s = strtok_r(NULL, " ", &save_ptr);
1108 unixctl_command_reply(conn, 501,
1109 "usage: bond/set-active-slave BOND SLAVE");
1113 bond = bond_find(bond_s);
1115 unixctl_command_reply(conn, 501, "no such bond");
1119 slave = bond_lookup_slave(bond, slave_s);
1121 unixctl_command_reply(conn, 501, "no such slave");
1125 if (!slave->enabled) {
1126 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1130 if (bond->active_slave != slave) {
1131 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1132 bond->active_slave = slave;
1133 bond->active_slave->tag = tag_create_random();
1134 VLOG_INFO("bond %s: active interface is now %s",
1135 bond->name, slave->name);
1136 bond->send_learning_packets = true;
1137 unixctl_command_reply(conn, 200, "done");
1139 unixctl_command_reply(conn, 200, "no change");
1144 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1146 char *args = (char *) args_;
1147 char *save_ptr = NULL;
1148 char *bond_s, *slave_s;
1150 struct bond_slave *slave;
1152 bond_s = strtok_r(args, " ", &save_ptr);
1153 slave_s = strtok_r(NULL, " ", &save_ptr);
1155 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1156 enable ? "enable" : "disable");
1157 unixctl_command_reply(conn, 501, usage);
1162 bond = bond_find(bond_s);
1164 unixctl_command_reply(conn, 501, "no such bond");
1168 slave = bond_lookup_slave(bond, slave_s);
1170 unixctl_command_reply(conn, 501, "no such slave");
1174 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1175 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1179 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1180 void *aux OVS_UNUSED)
1182 enable_slave(conn, args, true);
1186 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1187 void *aux OVS_UNUSED)
1189 enable_slave(conn, args, false);
1193 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1194 void *aux OVS_UNUSED)
1196 char *args = (char *) args_;
1197 uint8_t mac[ETH_ADDR_LEN];
1201 char *mac_s, *vlan_s;
1202 char *save_ptr = NULL;
1204 mac_s = strtok_r(args, " ", &save_ptr);
1205 vlan_s = strtok_r(NULL, " ", &save_ptr);
1208 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1209 unixctl_command_reply(conn, 501, "invalid vlan");
1213 vlan = OFP_VLAN_NONE;
1216 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1217 == ETH_ADDR_SCAN_COUNT) {
1218 hash = bond_hash_src(mac, vlan) & BOND_MASK;
1220 hash_cstr = xasprintf("%u", hash);
1221 unixctl_command_reply(conn, 200, hash_cstr);
1224 unixctl_command_reply(conn, 501, "invalid mac");
1231 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1232 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1233 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1234 unixctl_command_register("bond/set-active-slave",
1235 bond_unixctl_set_active_slave, NULL);
1236 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1238 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1240 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1244 bond_entry_reset(struct bond *bond)
1246 if (bond->balance != BM_AB) {
1247 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1250 bond->hash = xmalloc(hash_len);
1252 memset(bond->hash, 0, hash_len);
1254 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1261 static struct bond_slave *
1262 bond_slave_lookup(struct bond *bond, const void *slave_)
1264 struct bond_slave *slave;
1266 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1268 if (slave->aux == slave_) {
1277 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1279 return (bond->detect == BLSM_CARRIER
1280 ? netdev_get_carrier(netdev)
1281 : netdev_get_miimon(netdev));
1285 bond_stb_sort_cmp__(const void *a_, const void *b_)
1287 const struct bond_slave *const *ap = a_;
1288 const struct bond_slave *const *bp = b_;
1289 const struct bond_slave *a = *ap;
1290 const struct bond_slave *b = *bp;
1291 uint16_t aid = a->stb_id;
1292 uint16_t bid = b->stb_id;
1294 return aid < bid ? -1 : aid > bid;
1298 bond_stb_sort(struct bond *bond)
1302 if (!bond->stb_slaves || !bond->stb_need_sort) {
1305 bond->stb_need_sort = false;
1307 qsort(bond->stb_slaves, bond->n_stb_slaves, sizeof *bond->stb_slaves,
1308 bond_stb_sort_cmp__);
1310 for (i = 0; i < bond->n_stb_slaves; i++) {
1311 bond->stb_slaves[i]->stb_idx = i;
1318 bond_stb_enable_slave(struct bond_slave *slave)
1320 struct bond *bond = slave->bond;
1322 if (!bond->stb_slaves) {
1326 bond->stb_need_sort = true;
1328 if (slave->enabled) {
1329 if (bond->len_stb_slaves <= bond->n_stb_slaves) {
1330 bond->stb_slaves = x2nrealloc(bond->stb_slaves,
1331 &bond->len_stb_slaves,
1332 sizeof *bond->stb_slaves);
1335 slave->stb_idx = bond->n_stb_slaves++;
1336 bond->stb_slaves[slave->stb_idx] = slave;
1338 size_t index = slave->stb_idx;
1339 bond->stb_slaves[index] = bond->stb_slaves[--bond->n_stb_slaves];
1340 bond->stb_slaves[index]->stb_idx = index;
1345 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1347 slave->delay_expires = LLONG_MAX;
1348 if (enable != slave->enabled) {
1349 slave->enabled = enable;
1350 if (!slave->enabled) {
1351 VLOG_WARN("interface %s: disabled", slave->name);
1353 tag_set_add(tags, slave->tag);
1356 VLOG_WARN("interface %s: enabled", slave->name);
1357 slave->tag = tag_create_random();
1359 bond_stb_enable_slave(slave);
1364 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1366 struct bond *bond = slave->bond;
1369 up = slave->up && slave->lacp_may_enable;
1370 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1371 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1372 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1373 slave->name, up ? "up" : "down");
1374 if (up == slave->enabled) {
1375 slave->delay_expires = LLONG_MAX;
1376 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1377 slave->name, up ? "disabled" : "enabled");
1379 int delay = (bond->lacp_negotiated ? 0
1380 : up ? bond->updelay : bond->downdelay);
1381 slave->delay_expires = time_msec() + delay;
1383 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1386 up ? "enabled" : "disabled",
1393 if (time_msec() >= slave->delay_expires) {
1394 bond_enable_slave(slave, up, tags);
1399 bond_is_tcp_hash(const struct bond *bond)
1401 return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
1402 && bond->lacp_negotiated;
1406 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan)
1408 return hash_bytes(mac, ETH_ADDR_LEN, vlan);
1412 bond_hash_tcp(const struct flow *flow, uint16_t vlan)
1414 struct flow hash_flow = *flow;
1415 hash_flow.vlan_tci = vlan;
1417 /* The symmetric quality of this hash function is not required, but
1418 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1419 * purposes, so we use it out of convenience. */
1420 return flow_hash_symmetric_l4(&hash_flow, 0);
1424 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1426 assert(bond->balance != BM_AB);
1428 return (bond_is_tcp_hash(bond)
1429 ? bond_hash_tcp(flow, vlan)
1430 : bond_hash_src(flow->dl_src, vlan));
1433 static struct bond_entry *
1434 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1437 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1440 static struct bond_slave *
1441 choose_output_slave(const struct bond *bond, const struct flow *flow,
1444 struct bond_entry *e;
1446 switch (bond->balance) {
1448 return bond->active_slave;
1451 if (bond->n_stb_slaves) {
1452 return bond->stb_slaves[bond_hash(bond, flow, vlan)
1453 % bond->n_stb_slaves];
1455 return bond->active_slave;
1460 e = lookup_bond_entry(bond, flow, vlan);
1461 if (!e->slave || !e->slave->enabled) {
1462 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1463 struct bond_slave, hmap_node);
1464 if (!e->slave->enabled) {
1465 e->slave = bond->active_slave;
1467 e->tag = tag_create_random();
1476 static struct bond_slave *
1477 bond_choose_slave(const struct bond *bond)
1479 struct bond_slave *slave, *best;
1481 /* Find an enabled slave. */
1482 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1483 if (slave->enabled) {
1488 /* All interfaces are disabled. Find an interface that will be enabled
1489 * after its updelay expires. */
1491 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1492 if (slave->delay_expires != LLONG_MAX
1493 && slave->lacp_may_enable
1494 && (!best || slave->delay_expires < best->delay_expires)) {
1502 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1504 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1505 struct bond_slave *old_active_slave = bond->active_slave;
1507 bond->active_slave = bond_choose_slave(bond);
1508 if (bond->active_slave) {
1509 if (bond->active_slave->enabled) {
1510 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1511 bond->name, bond->active_slave->name);
1513 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1514 "remaining %lld ms updelay (since no interface was "
1515 "enabled)", bond->name, bond->active_slave->name,
1516 bond->active_slave->delay_expires - time_msec());
1517 bond_enable_slave(bond->active_slave, true, tags);
1520 if (!old_active_slave) {
1521 tag_set_add(tags, bond->no_slaves_tag);
1524 bond->send_learning_packets = true;
1525 } else if (old_active_slave) {
1526 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1530 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1531 * there is no active slave. */
1533 bond_get_active_slave_tag(const struct bond *bond)
1535 return (bond->active_slave
1536 ? bond->active_slave->tag
1537 : bond->no_slaves_tag);
1540 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1541 * bond interface. */
1543 bond_update_fake_slave_stats(struct bond *bond)
1545 struct netdev_stats bond_stats;
1546 struct bond_slave *slave;
1547 struct netdev *bond_dev;
1549 memset(&bond_stats, 0, sizeof bond_stats);
1551 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1552 struct netdev_stats slave_stats;
1554 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1555 /* XXX: We swap the stats here because they are swapped back when
1556 * reported by the internal device. The reason for this is
1557 * internal devices normally represent packets going into the
1558 * system but when used as fake bond device they represent packets
1559 * leaving the system. We really should do this in the internal
1560 * device itself because changing it here reverses the counts from
1561 * the perspective of the switch. However, the internal device
1562 * doesn't know what type of device it represents so we have to do
1563 * it here for now. */
1564 bond_stats.tx_packets += slave_stats.rx_packets;
1565 bond_stats.tx_bytes += slave_stats.rx_bytes;
1566 bond_stats.rx_packets += slave_stats.tx_packets;
1567 bond_stats.rx_bytes += slave_stats.tx_bytes;
1571 if (!netdev_open_default(bond->name, &bond_dev)) {
1572 netdev_set_stats(bond_dev, &bond_stats);
1573 netdev_close(bond_dev);