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. */
98 bool bond_revalidate; /* True if flows need revalidation. */
100 /* SLB specific bonding info. */
101 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
102 int rebalance_interval; /* Interval between rebalances, in ms. */
103 long long int next_rebalance; /* Next rebalancing time. */
104 bool send_learning_packets;
106 /* BM_STABLE specific bonding info. */
107 struct bond_slave **stb_slaves; /* Ordered list of enabled slaves. */
108 size_t n_stb_slaves; /* Number of slaves in 'stb_slaves'. */
109 size_t len_stb_slaves; /* Slaves allocated in 'stb_slaves'. */
110 bool stb_need_sort; /* True if stb_slaves is not sorted. */
111 tag_type stb_tag; /* Tag associated with this bond. */
114 enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
115 struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
116 long long int miimon_interval; /* Miimon status refresh interval. */
117 long long int miimon_next_update; /* Time of next miimon update. */
119 /* Legacy compatibility. */
120 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
122 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
123 * where we can't otherwise provide revalidation feedback to the client.
124 * That's only unixctl commands now; I hope no other cases will arise. */
125 struct tag_set unixctl_tags;
128 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
130 static void bond_entry_reset(struct bond *);
131 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
132 static bool bond_is_link_up(struct bond *, struct netdev *);
133 static void bond_enable_slave(struct bond_slave *, bool enable,
135 static void bond_stb_sort(struct bond *);
136 static void bond_stb_enable_slave(struct bond_slave *);
137 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
138 static void bond_choose_active_slave(struct bond *, struct tag_set *);
139 static bool bond_is_tcp_hash(const struct bond *);
140 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
142 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan);
143 static struct bond_entry *lookup_bond_entry(const struct bond *,
146 static tag_type bond_get_active_slave_tag(const struct bond *);
147 static struct bond_slave *choose_output_slave(const struct bond *,
150 static void bond_update_fake_slave_stats(struct bond *);
152 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
153 * stores the mode in '*balance' and returns true. Otherwise returns false
154 * without modifying '*balance'. */
156 bond_mode_from_string(enum bond_mode *balance, const char *s)
158 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
160 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
162 } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
163 *balance = BM_STABLE;
164 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
172 /* Returns a string representing 'balance'. */
174 bond_mode_to_string(enum bond_mode balance) {
177 return "balance-tcp";
179 return "balance-slb";
183 return "active-backup";
188 /* Attempts to parse 's' as the name of a bond link status detection mode. If
189 * successful, stores the mode in '*detect' and returns true. Otherwise
190 * returns false without modifying '*detect'. */
192 bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
194 if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
195 *detect = BLSM_CARRIER;
196 } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
197 *detect = BLSM_MIIMON;
204 /* Returns a string representing 'detect'. */
206 bond_detect_mode_to_string(enum bond_detect_mode detect)
217 /* Creates and returns a new bond whose configuration is initially taken from
220 * The caller should register each slave on the new bond by calling
221 * bond_slave_register(). */
223 bond_create(const struct bond_settings *s)
227 bond = xzalloc(sizeof *bond);
228 hmap_init(&bond->slaves);
229 bond->no_slaves_tag = tag_create_random();
230 bond->miimon_next_update = LLONG_MAX;
231 bond->next_fake_iface_update = LLONG_MAX;
233 bond_reconfigure(bond, s);
235 tag_set_init(&bond->unixctl_tags);
242 bond_destroy(struct bond *bond)
244 struct bond_slave *slave, *next_slave;
250 hmap_remove(&all_bonds, &bond->hmap_node);
252 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
253 hmap_remove(&bond->slaves, &slave->hmap_node);
254 /* Client owns 'slave->netdev'. */
258 hmap_destroy(&bond->slaves);
262 netdev_monitor_destroy(bond->monitor);
268 /* Updates 'bond''s overall configuration to 's'.
270 * The caller should register each slave on 'bond' by calling
271 * bond_slave_register(). This is optional if none of the slaves'
272 * configuration has changed. In any case it can't hurt.
274 * Returns true if the configuration has changed in such a way that requires
278 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
280 bool revalidate = false;
282 if (!bond->name || strcmp(bond->name, s->name)) {
284 hmap_remove(&all_bonds, &bond->hmap_node);
287 bond->name = xstrdup(s->name);
288 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
291 bond->detect = s->detect;
292 bond->miimon_interval = s->miimon_interval;
293 bond->updelay = s->up_delay;
294 bond->downdelay = s->down_delay;
295 bond->rebalance_interval = s->rebalance_interval;
297 if (bond->balance != s->balance) {
298 bond->balance = s->balance;
302 if (bond->detect == BLSM_CARRIER) {
303 struct bond_slave *slave;
305 if (!bond->monitor) {
306 bond->monitor = netdev_monitor_create();
309 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
310 netdev_monitor_add(bond->monitor, slave->netdev);
313 netdev_monitor_destroy(bond->monitor);
314 bond->monitor = NULL;
316 if (bond->miimon_next_update == LLONG_MAX) {
317 bond->miimon_next_update = time_msec() + bond->miimon_interval;
322 if (bond->next_fake_iface_update == LLONG_MAX) {
323 bond->next_fake_iface_update = time_msec();
326 bond->next_fake_iface_update = LLONG_MAX;
329 if (bond->balance != BM_STABLE) {
330 free(bond->stb_slaves);
331 bond->stb_slaves = NULL;
333 } else if (!bond->stb_tag) {
334 struct bond_slave *slave;
336 bond->stb_tag = tag_create_random();
338 assert(!bond->stb_slaves);
339 bond->n_stb_slaves = 0;
340 bond->len_stb_slaves = 0;
341 bond->stb_slaves = NULL;
343 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
344 if (slave->enabled) {
345 bond_stb_enable_slave(slave);
350 if (bond->bond_revalidate) {
352 bond->bond_revalidate = false;
355 if (bond->balance == BM_AB || !bond->hash || revalidate) {
356 bond_entry_reset(bond);
362 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
363 * arbitrary client-provided pointer that uniquely identifies a slave within a
364 * bond. If 'slave_' already exists within 'bond' then this function
365 * reconfigures the existing slave.
367 * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
368 * across restarts and distributed vswitch instances. It should be unique per
369 * slave, and preferably consistent across restarts and reconfigurations.
371 * 'netdev' must be the network device that 'slave_' represents. It is owned
372 * by the client, so the client must not close it before either unregistering
373 * 'slave_' or destroying 'bond'.
376 bond_slave_register(struct bond *bond, void *slave_, uint16_t stb_id,
377 struct netdev *netdev)
379 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
382 slave = xzalloc(sizeof *slave);
384 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
387 slave->delay_expires = LLONG_MAX;
388 slave->up = bond_is_link_up(bond, netdev);
389 slave->name = xstrdup(netdev_get_name(netdev));
391 slave->enabled = false;
392 bond_enable_slave(slave, slave->up, NULL);
395 if (slave->stb_id != stb_id) {
396 bond->stb_need_sort = true;
397 slave->stb_id = stb_id;
400 slave->netdev = netdev;
402 slave->name = xstrdup(netdev_get_name(netdev));
405 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
406 * then this function has no effect.
408 * Unregistering a slave invalidates all flows. */
410 bond_slave_unregister(struct bond *bond, const void *slave_)
412 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
419 bond_enable_slave(slave, false, NULL);
421 del_active = bond->active_slave == slave;
423 struct bond_entry *e;
424 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
425 if (e->slave == slave) {
433 hmap_remove(&bond->slaves, &slave->hmap_node);
434 /* Client owns 'slave->netdev'. */
441 bond_choose_active_slave(bond, &tags);
442 bond->send_learning_packets = true;
446 /* Should be called on each slave in 'bond' before bond_run() to indicate the
447 * results of lacp_slave_may_enable() on 'slave_'. */
449 bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
452 bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
455 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
456 * allow tagged flows to be invalidated.
458 * The caller should check bond_should_send_learning_packets() afterward. */
460 bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
462 struct bond_slave *slave;
463 bool is_tcp_hash = bond_is_tcp_hash(bond);
465 bond->lacp_negotiated = lacp_negotiated;
467 /* Update link status. */
468 if (bond->detect == BLSM_CARRIER
469 || time_msec() >= bond->miimon_next_update)
471 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
472 slave->up = bond_is_link_up(bond, slave->netdev);
474 bond->miimon_next_update = time_msec() + bond->miimon_interval;
477 /* Enable slaves based on link status and LACP feedback. */
478 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
479 bond_link_status_update(slave, tags);
481 if (!bond->active_slave || !bond->active_slave->enabled) {
482 bond_choose_active_slave(bond, tags);
485 /* Update fake bond interface stats. */
486 if (time_msec() >= bond->next_fake_iface_update) {
487 bond_update_fake_slave_stats(bond);
488 bond->next_fake_iface_update = time_msec() + 1000;
493 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
494 bond->bond_revalidate = true;
497 if (bond->bond_revalidate) {
498 bond->bond_revalidate = false;
500 bond_entry_reset(bond);
501 if (bond->balance != BM_STABLE) {
502 struct bond_slave *slave;
504 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
505 tag_set_add(tags, slave->tag);
508 tag_set_add(tags, bond->stb_tag);
510 tag_set_add(tags, bond->no_slaves_tag);
513 /* Invalidate any tags required by */
514 tag_set_union(tags, &bond->unixctl_tags);
515 tag_set_init(&bond->unixctl_tags);
518 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
520 bond_wait(struct bond *bond)
522 struct bond_slave *slave;
524 if (bond->detect == BLSM_CARRIER) {
525 netdev_monitor_poll_wait(bond->monitor);
527 poll_timer_wait_until(bond->miimon_next_update);
530 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
531 if (slave->delay_expires != LLONG_MAX) {
532 poll_timer_wait_until(slave->delay_expires);
536 if (bond->next_fake_iface_update != LLONG_MAX) {
537 poll_timer_wait_until(bond->next_fake_iface_update);
540 /* Ensure that any saved tags get revalidated right away. */
541 if (!tag_set_is_empty(&bond->unixctl_tags)) {
542 poll_immediate_wake();
545 /* We don't wait for bond->next_rebalance because rebalancing can only run
546 * at a flow account checkpoint. ofproto does checkpointing on its own
547 * schedule and bond_rebalance() gets called afterward, so we'd just be
548 * waking up for no purpose. */
551 /* MAC learning table interaction. */
554 may_send_learning_packets(const struct bond *bond)
556 return !bond->lacp_negotiated && bond->balance != BM_AB;
559 /* Returns true if 'bond' needs the client to send out packets to assist with
560 * MAC learning on 'bond'. If this function returns true, then the client
561 * should iterate through its MAC learning table for the bridge on which 'bond'
562 * is located. For each MAC that has been learned on a port other than 'bond',
563 * it should call bond_send_learning_packet().
565 * This function will only return true if 'bond' is in SLB mode and LACP is not
566 * negotiated. Otherwise sending learning packets isn't necessary.
568 * Calling this function resets the state that it checks. */
570 bond_should_send_learning_packets(struct bond *bond)
572 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
573 bond->send_learning_packets = false;
577 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
579 * See bond_should_send_learning_packets() for description of usage. */
581 bond_send_learning_packet(struct bond *bond,
582 const uint8_t eth_src[ETH_ADDR_LEN],
585 struct bond_slave *slave;
586 struct ofpbuf packet;
590 assert(may_send_learning_packets(bond));
591 if (!bond->active_slave) {
592 /* Nowhere to send the learning packet. */
596 memset(&flow, 0, sizeof flow);
597 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
598 slave = choose_output_slave(bond, &flow, vlan);
600 ofpbuf_init(&packet, 0);
601 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
604 eth_set_vlan_tci(&packet, htons(vlan));
606 error = netdev_send(slave->netdev, &packet);
607 ofpbuf_uninit(&packet);
612 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
613 * Ethernet destination address of 'eth_dst', should be admitted.
615 * The return value is one of the following:
617 * - BV_ACCEPT: Admit the packet.
619 * - BV_DROP: Drop the packet.
621 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
622 * Ethernet source address and VLAN. If there is none, or if the packet
623 * is on the learned port, then admit the packet. If a different port has
624 * been learned, however, drop the packet (and do not use it for MAC
628 bond_check_admissibility(struct bond *bond, const void *slave_,
629 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
631 /* Admit all packets if LACP has been negotiated, because that means that
632 * the remote switch is aware of the bond and will "do the right thing". */
633 if (bond->lacp_negotiated) {
637 /* Drop all multicast packets on inactive slaves. */
638 if (eth_addr_is_multicast(eth_dst)) {
639 *tags |= bond_get_active_slave_tag(bond);
640 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
645 /* Drop all packets for which we have learned a different input port,
646 * because we probably sent the packet on one slave and got it back on the
647 * other. Gratuitous ARP packets are an exception to this rule: the host
648 * has moved to another switch. The exception to the exception is if we
649 * locked the learning table to avoid reflections on bond slaves. */
650 return BV_DROP_IF_MOVED;
653 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
654 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
655 * NULL if the packet should be dropped because no slaves are enabled.
657 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
658 * should be a VID only (i.e. excluding the PCP bits). Second,
659 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
660 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
661 * packet belongs to (so for an access port it will be the access port's VLAN).
663 * Adds a tag to '*tags' that associates the flow with the returned slave.
666 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
667 uint16_t vlan, tag_type *tags)
669 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
671 *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
674 *tags |= bond->no_slaves_tag;
682 bond_is_balanced(const struct bond *bond)
684 return bond->balance == BM_SLB || bond->balance == BM_TCP;
687 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
689 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
693 if (bond_is_balanced(bond)) {
694 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
698 static struct bond_slave *
699 bond_slave_from_bal_node(struct list *bal)
701 return CONTAINER_OF(bal, struct bond_slave, bal_node);
705 log_bals(struct bond *bond, const struct list *bals)
707 if (VLOG_IS_DBG_ENABLED()) {
708 struct ds ds = DS_EMPTY_INITIALIZER;
709 const struct bond_slave *slave;
711 LIST_FOR_EACH (slave, bal_node, bals) {
713 ds_put_char(&ds, ',');
715 ds_put_format(&ds, " %s %"PRIu64"kB",
716 slave->name, slave->tx_bytes / 1024);
718 if (!slave->enabled) {
719 ds_put_cstr(&ds, " (disabled)");
721 if (!list_is_empty(&slave->entries)) {
722 struct bond_entry *e;
724 ds_put_cstr(&ds, " (");
725 LIST_FOR_EACH (e, list_node, &slave->entries) {
726 if (&e->list_node != list_front(&slave->entries)) {
727 ds_put_cstr(&ds, " + ");
729 ds_put_format(&ds, "h%td: %"PRIu64"kB",
730 e - bond->hash, e->tx_bytes / 1024);
732 ds_put_cstr(&ds, ")");
735 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
740 /* Shifts 'hash' from its current slave to 'to'. */
742 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
745 struct bond_slave *from = hash->slave;
746 struct bond *bond = from->bond;
747 uint64_t delta = hash->tx_bytes;
749 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
750 "from %s to %s (now carrying %"PRIu64"kB and "
751 "%"PRIu64"kB load, respectively)",
752 bond->name, delta / 1024, hash - bond->hash,
753 from->name, to->name,
754 (from->tx_bytes - delta) / 1024,
755 (to->tx_bytes + delta) / 1024);
757 /* Shift load away from 'from' to 'to'. */
758 from->tx_bytes -= delta;
759 to->tx_bytes += delta;
761 /* Arrange for flows to be revalidated. */
762 tag_set_add(set, hash->tag);
764 hash->tag = tag_create_random();
767 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
768 * given that doing so must decrease the ratio of the load on the two slaves by
769 * at least 0.1. Returns NULL if there is no appropriate entry.
771 * The list of entries isn't sorted. I don't know of a reason to prefer to
772 * shift away small hashes or large hashes. */
773 static struct bond_entry *
774 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
776 struct bond_entry *e;
778 if (list_is_short(&from->entries)) {
779 /* 'from' carries no more than one MAC hash, so shifting load away from
780 * it would be pointless. */
784 LIST_FOR_EACH (e, list_node, &from->entries) {
785 double old_ratio, new_ratio;
788 if (to_tx_bytes == 0) {
789 /* Nothing on the new slave, move it. */
794 old_ratio = (double)from->tx_bytes / to_tx_bytes;
795 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
796 if (old_ratio - new_ratio > 0.1) {
797 /* Would decrease the ratio, move it. */
805 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
808 insert_bal(struct list *bals, struct bond_slave *slave)
810 struct bond_slave *pos;
812 LIST_FOR_EACH (pos, bal_node, bals) {
813 if (slave->tx_bytes > pos->tx_bytes) {
817 list_insert(&pos->bal_node, &slave->bal_node);
820 /* Removes 'slave' from its current list and then inserts it into 'bals' so
821 * that descending order of 'tx_bytes' is maintained. */
823 reinsert_bal(struct list *bals, struct bond_slave *slave)
825 list_remove(&slave->bal_node);
826 insert_bal(bals, slave);
829 /* If 'bond' needs rebalancing, does so.
831 * The caller should have called bond_account() for each active flow, to ensure
832 * that flow data is consistently accounted at this point. */
834 bond_rebalance(struct bond *bond, struct tag_set *tags)
836 struct bond_slave *slave;
837 struct bond_entry *e;
840 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
843 bond->next_rebalance = time_msec() + bond->rebalance_interval;
845 /* Add each bond_entry to its slave's 'entries' list.
846 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
847 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
849 list_init(&slave->entries);
851 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
852 if (e->slave && e->tx_bytes) {
853 e->slave->tx_bytes += e->tx_bytes;
854 list_push_back(&e->slave->entries, &e->list_node);
858 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
860 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
861 * with a proper list sort algorithm. */
863 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
864 if (slave->enabled) {
865 insert_bal(&bals, slave);
868 log_bals(bond, &bals);
870 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
871 while (!list_is_short(&bals)) {
872 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
873 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
876 overload = from->tx_bytes - to->tx_bytes;
877 if (overload < to->tx_bytes >> 5 || overload < 100000) {
878 /* The extra load on 'from' (and all less-loaded slaves), compared
879 * to that of 'to' (the least-loaded slave), is less than ~3%, or
880 * it is less than ~1Mbps. No point in rebalancing. */
884 /* 'from' is carrying significantly more load than 'to', and that load
885 * is split across at least two different hashes. */
886 e = choose_entry_to_migrate(from, to->tx_bytes);
888 bond_shift_load(e, to, tags);
890 /* Delete element from from->entries.
892 * We don't add the element to to->hashes. That would only allow
893 * 'e' to be migrated to another slave in this rebalancing run, and
894 * there is no point in doing that. */
895 list_remove(&e->list_node);
897 /* Re-sort 'bals'. */
898 reinsert_bal(&bals, from);
899 reinsert_bal(&bals, to);
901 /* Can't usefully migrate anything away from 'from'.
902 * Don't reconsider it. */
903 list_remove(&from->bal_node);
907 /* Implement exponentially weighted moving average. A weight of 1/2 causes
908 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
909 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
910 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
918 /* Bonding unixctl user interface functions. */
921 bond_find(const char *name)
925 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
927 if (!strcmp(bond->name, name)) {
934 static struct bond_slave *
935 bond_lookup_slave(struct bond *bond, const char *slave_name)
937 struct bond_slave *slave;
939 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
940 if (!strcmp(slave->name, slave_name)) {
948 bond_unixctl_list(struct unixctl_conn *conn,
949 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
951 struct ds ds = DS_EMPTY_INITIALIZER;
952 const struct bond *bond;
954 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
956 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
957 const struct bond_slave *slave;
960 ds_put_format(&ds, "%s\t%s\t",
961 bond->name, bond_mode_to_string(bond->balance));
964 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
966 ds_put_cstr(&ds, ", ");
968 ds_put_cstr(&ds, slave->name);
970 ds_put_char(&ds, '\n');
972 unixctl_command_reply(conn, 200, ds_cstr(&ds));
977 bond_unixctl_show(struct unixctl_conn *conn,
978 const char *args, void *aux OVS_UNUSED)
980 struct ds ds = DS_EMPTY_INITIALIZER;
981 const struct bond_slave *slave;
982 const struct bond *bond;
984 bond = bond_find(args);
986 unixctl_command_reply(conn, 501, "no such bond");
990 ds_put_format(&ds, "bond_mode: %s\n",
991 bond_mode_to_string(bond->balance));
993 if (bond->balance != BM_AB) {
994 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
995 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
998 ds_put_format(&ds, "bond-detect-mode: %s\n",
999 bond->monitor ? "carrier" : "miimon");
1001 if (!bond->monitor) {
1002 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
1003 bond->miimon_interval);
1006 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
1007 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
1009 if (bond_is_balanced(bond)) {
1010 ds_put_format(&ds, "next rebalance: %lld ms\n",
1011 bond->next_rebalance - time_msec());
1014 ds_put_format(&ds, "lacp_negotiated: %s\n",
1015 bond->lacp_negotiated ? "true" : "false");
1017 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1018 struct bond_entry *be;
1022 ds_put_format(&ds, "\nslave %s: %s\n",
1023 slave->name, slave->enabled ? "enabled" : "disabled");
1024 if (slave == bond->active_slave) {
1025 ds_put_cstr(&ds, "\tactive slave\n");
1027 if (slave->delay_expires != LLONG_MAX) {
1028 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1029 slave->enabled ? "downdelay" : "updelay",
1030 slave->delay_expires - time_msec());
1033 ds_put_format(&ds, "\tlacp_may_enable: %s\n",
1034 slave->lacp_may_enable ? "true" : "false");
1036 if (!bond_is_balanced(bond)) {
1041 memset(&flow, 0, sizeof flow);
1042 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1043 int hash = be - bond->hash;
1045 if (be->slave != slave) {
1049 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1050 hash, be->tx_bytes / 1024);
1052 if (bond->balance != BM_SLB) {
1056 /* XXX How can we list the MACs assigned to hashes? */
1059 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1064 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1065 void *aux OVS_UNUSED)
1067 char *args = (char *) args_;
1068 char *save_ptr = NULL;
1069 char *bond_s, *hash_s, *slave_s;
1071 struct bond_slave *slave;
1072 struct bond_entry *entry;
1075 bond_s = strtok_r(args, " ", &save_ptr);
1076 hash_s = strtok_r(NULL, " ", &save_ptr);
1077 slave_s = strtok_r(NULL, " ", &save_ptr);
1079 unixctl_command_reply(conn, 501,
1080 "usage: bond/migrate BOND HASH SLAVE");
1084 bond = bond_find(bond_s);
1086 unixctl_command_reply(conn, 501, "no such bond");
1090 if (bond->balance != BM_SLB) {
1091 unixctl_command_reply(conn, 501, "not an SLB bond");
1095 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1096 hash = atoi(hash_s) & BOND_MASK;
1098 unixctl_command_reply(conn, 501, "bad hash");
1102 slave = bond_lookup_slave(bond, slave_s);
1104 unixctl_command_reply(conn, 501, "no such slave");
1108 if (!slave->enabled) {
1109 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1113 entry = &bond->hash[hash];
1114 tag_set_add(&bond->unixctl_tags, entry->tag);
1115 entry->slave = slave;
1116 entry->tag = tag_create_random();
1117 unixctl_command_reply(conn, 200, "migrated");
1121 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1122 void *aux OVS_UNUSED)
1124 char *args = (char *) args_;
1125 char *save_ptr = NULL;
1126 char *bond_s, *slave_s;
1128 struct bond_slave *slave;
1130 bond_s = strtok_r(args, " ", &save_ptr);
1131 slave_s = strtok_r(NULL, " ", &save_ptr);
1133 unixctl_command_reply(conn, 501,
1134 "usage: bond/set-active-slave BOND SLAVE");
1138 bond = bond_find(bond_s);
1140 unixctl_command_reply(conn, 501, "no such bond");
1144 slave = bond_lookup_slave(bond, slave_s);
1146 unixctl_command_reply(conn, 501, "no such slave");
1150 if (!slave->enabled) {
1151 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1155 if (bond->active_slave != slave) {
1156 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1157 bond->active_slave = slave;
1158 bond->active_slave->tag = tag_create_random();
1159 VLOG_INFO("bond %s: active interface is now %s",
1160 bond->name, slave->name);
1161 bond->send_learning_packets = true;
1162 unixctl_command_reply(conn, 200, "done");
1164 unixctl_command_reply(conn, 200, "no change");
1169 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1171 char *args = (char *) args_;
1172 char *save_ptr = NULL;
1173 char *bond_s, *slave_s;
1175 struct bond_slave *slave;
1177 bond_s = strtok_r(args, " ", &save_ptr);
1178 slave_s = strtok_r(NULL, " ", &save_ptr);
1180 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1181 enable ? "enable" : "disable");
1182 unixctl_command_reply(conn, 501, usage);
1187 bond = bond_find(bond_s);
1189 unixctl_command_reply(conn, 501, "no such bond");
1193 slave = bond_lookup_slave(bond, slave_s);
1195 unixctl_command_reply(conn, 501, "no such slave");
1199 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1200 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1204 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1205 void *aux OVS_UNUSED)
1207 enable_slave(conn, args, true);
1211 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1212 void *aux OVS_UNUSED)
1214 enable_slave(conn, args, false);
1218 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1219 void *aux OVS_UNUSED)
1221 char *args = (char *) args_;
1222 uint8_t mac[ETH_ADDR_LEN];
1226 char *mac_s, *vlan_s;
1227 char *save_ptr = NULL;
1229 mac_s = strtok_r(args, " ", &save_ptr);
1230 vlan_s = strtok_r(NULL, " ", &save_ptr);
1233 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1234 unixctl_command_reply(conn, 501, "invalid vlan");
1238 vlan = OFP_VLAN_NONE;
1241 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1242 == ETH_ADDR_SCAN_COUNT) {
1243 hash = bond_hash_src(mac, vlan) & BOND_MASK;
1245 hash_cstr = xasprintf("%u", hash);
1246 unixctl_command_reply(conn, 200, hash_cstr);
1249 unixctl_command_reply(conn, 501, "invalid mac");
1256 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1257 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1258 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1259 unixctl_command_register("bond/set-active-slave",
1260 bond_unixctl_set_active_slave, NULL);
1261 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1263 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1265 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1269 bond_entry_reset(struct bond *bond)
1271 if (bond->balance != BM_AB) {
1272 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1275 bond->hash = xmalloc(hash_len);
1277 memset(bond->hash, 0, hash_len);
1279 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1286 static struct bond_slave *
1287 bond_slave_lookup(struct bond *bond, const void *slave_)
1289 struct bond_slave *slave;
1291 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1293 if (slave->aux == slave_) {
1302 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1304 return (bond->detect == BLSM_CARRIER
1305 ? netdev_get_carrier(netdev)
1306 : netdev_get_miimon(netdev));
1310 bond_stb_sort_cmp__(const void *a_, const void *b_)
1312 const struct bond_slave *const *ap = a_;
1313 const struct bond_slave *const *bp = b_;
1314 const struct bond_slave *a = *ap;
1315 const struct bond_slave *b = *bp;
1316 uint16_t aid = a->stb_id;
1317 uint16_t bid = b->stb_id;
1319 return aid < bid ? -1 : aid > bid;
1323 bond_stb_sort(struct bond *bond)
1327 if (!bond->stb_slaves || !bond->stb_need_sort) {
1331 bond->stb_need_sort = false;
1332 bond->bond_revalidate = true;
1334 qsort(bond->stb_slaves, bond->n_stb_slaves, sizeof *bond->stb_slaves,
1335 bond_stb_sort_cmp__);
1337 for (i = 0; i < bond->n_stb_slaves; i++) {
1338 bond->stb_slaves[i]->stb_idx = i;
1343 bond_stb_enable_slave(struct bond_slave *slave)
1345 struct bond *bond = slave->bond;
1347 if (bond->balance != BM_STABLE) {
1351 bond->stb_need_sort = true;
1353 if (slave->enabled) {
1354 if (bond->len_stb_slaves <= bond->n_stb_slaves) {
1355 bond->stb_slaves = x2nrealloc(bond->stb_slaves,
1356 &bond->len_stb_slaves,
1357 sizeof *bond->stb_slaves);
1360 slave->stb_idx = bond->n_stb_slaves++;
1361 bond->stb_slaves[slave->stb_idx] = slave;
1363 size_t index = slave->stb_idx;
1364 bond->stb_slaves[index] = bond->stb_slaves[--bond->n_stb_slaves];
1365 bond->stb_slaves[index]->stb_idx = index;
1370 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1372 slave->delay_expires = LLONG_MAX;
1373 if (enable != slave->enabled) {
1374 slave->enabled = enable;
1375 if (!slave->enabled) {
1376 VLOG_WARN("interface %s: disabled", slave->name);
1378 tag_set_add(tags, slave->tag);
1381 VLOG_WARN("interface %s: enabled", slave->name);
1382 slave->tag = tag_create_random();
1384 bond_stb_enable_slave(slave);
1389 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1391 struct bond *bond = slave->bond;
1394 up = slave->up && slave->lacp_may_enable;
1395 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1396 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1397 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1398 slave->name, up ? "up" : "down");
1399 if (up == slave->enabled) {
1400 slave->delay_expires = LLONG_MAX;
1401 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1402 slave->name, up ? "disabled" : "enabled");
1404 int delay = (bond->lacp_negotiated ? 0
1405 : up ? bond->updelay : bond->downdelay);
1406 slave->delay_expires = time_msec() + delay;
1408 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1411 up ? "enabled" : "disabled",
1418 if (time_msec() >= slave->delay_expires) {
1419 bond_enable_slave(slave, up, tags);
1424 bond_is_tcp_hash(const struct bond *bond)
1426 return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
1427 && bond->lacp_negotiated;
1431 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan)
1433 return hash_bytes(mac, ETH_ADDR_LEN, vlan);
1437 bond_hash_tcp(const struct flow *flow, uint16_t vlan)
1439 struct flow hash_flow = *flow;
1440 hash_flow.vlan_tci = vlan;
1442 /* The symmetric quality of this hash function is not required, but
1443 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1444 * purposes, so we use it out of convenience. */
1445 return flow_hash_symmetric_l4(&hash_flow, 0);
1449 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1451 assert(bond->balance != BM_AB);
1453 return (bond_is_tcp_hash(bond)
1454 ? bond_hash_tcp(flow, vlan)
1455 : bond_hash_src(flow->dl_src, vlan));
1458 static struct bond_entry *
1459 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1462 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1465 static struct bond_slave *
1466 choose_output_slave(const struct bond *bond, const struct flow *flow,
1469 struct bond_entry *e;
1471 switch (bond->balance) {
1473 return bond->active_slave;
1476 if (bond->n_stb_slaves) {
1477 return bond->stb_slaves[bond_hash(bond, flow, vlan)
1478 % bond->n_stb_slaves];
1480 return bond->active_slave;
1485 e = lookup_bond_entry(bond, flow, vlan);
1486 if (!e->slave || !e->slave->enabled) {
1487 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1488 struct bond_slave, hmap_node);
1489 if (!e->slave->enabled) {
1490 e->slave = bond->active_slave;
1492 e->tag = tag_create_random();
1501 static struct bond_slave *
1502 bond_choose_slave(const struct bond *bond)
1504 struct bond_slave *slave, *best;
1506 /* Find an enabled slave. */
1507 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1508 if (slave->enabled) {
1513 /* All interfaces are disabled. Find an interface that will be enabled
1514 * after its updelay expires. */
1516 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1517 if (slave->delay_expires != LLONG_MAX
1518 && slave->lacp_may_enable
1519 && (!best || slave->delay_expires < best->delay_expires)) {
1527 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1529 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1530 struct bond_slave *old_active_slave = bond->active_slave;
1532 bond->active_slave = bond_choose_slave(bond);
1533 if (bond->active_slave) {
1534 if (bond->active_slave->enabled) {
1535 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1536 bond->name, bond->active_slave->name);
1538 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1539 "remaining %lld ms updelay (since no interface was "
1540 "enabled)", bond->name, bond->active_slave->name,
1541 bond->active_slave->delay_expires - time_msec());
1542 bond_enable_slave(bond->active_slave, true, tags);
1545 if (!old_active_slave) {
1546 tag_set_add(tags, bond->no_slaves_tag);
1549 bond->send_learning_packets = true;
1550 } else if (old_active_slave) {
1551 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1555 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1556 * there is no active slave. */
1558 bond_get_active_slave_tag(const struct bond *bond)
1560 return (bond->active_slave
1561 ? bond->active_slave->tag
1562 : bond->no_slaves_tag);
1565 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1566 * bond interface. */
1568 bond_update_fake_slave_stats(struct bond *bond)
1570 struct netdev_stats bond_stats;
1571 struct bond_slave *slave;
1572 struct netdev *bond_dev;
1574 memset(&bond_stats, 0, sizeof bond_stats);
1576 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1577 struct netdev_stats slave_stats;
1579 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1580 /* XXX: We swap the stats here because they are swapped back when
1581 * reported by the internal device. The reason for this is
1582 * internal devices normally represent packets going into the
1583 * system but when used as fake bond device they represent packets
1584 * leaving the system. We really should do this in the internal
1585 * device itself because changing it here reverses the counts from
1586 * the perspective of the switch. However, the internal device
1587 * doesn't know what type of device it represents so we have to do
1588 * it here for now. */
1589 bond_stats.tx_packets += slave_stats.rx_packets;
1590 bond_stats.tx_bytes += slave_stats.rx_bytes;
1591 bond_stats.rx_packets += slave_stats.tx_packets;
1592 bond_stats.rx_bytes += slave_stats.tx_bytes;
1596 if (!netdev_open_default(bond->name, &bond_dev)) {
1597 netdev_set_stats(bond_dev, &bond_stats);
1598 netdev_close(bond_dev);