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 uint32_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. */
96 uint32_t basis; /* Basis for flow hash function. */
98 /* SLB specific bonding info. */
99 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
100 int rebalance_interval; /* Interval between rebalances, in ms. */
101 long long int next_rebalance; /* Next rebalancing time. */
102 bool send_learning_packets;
104 /* BM_STABLE specific bonding info. */
105 tag_type stb_tag; /* Tag associated with this bond. */
108 enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
109 struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
110 long long int miimon_interval; /* Miimon status refresh interval. */
111 long long int miimon_next_update; /* Time of next miimon update. */
113 /* Legacy compatibility. */
114 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
116 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
117 * where we can't otherwise provide revalidation feedback to the client.
118 * That's only unixctl commands now; I hope no other cases will arise. */
119 struct tag_set unixctl_tags;
122 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
124 static void bond_entry_reset(struct bond *);
125 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
126 static bool bond_is_link_up(struct bond *, struct netdev *);
127 static void bond_enable_slave(struct bond_slave *, bool enable,
129 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
130 static void bond_choose_active_slave(struct bond *, struct tag_set *);
131 static bool bond_is_tcp_hash(const struct bond *);
132 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
133 uint16_t vlan, uint32_t basis);
134 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan,
136 static struct bond_entry *lookup_bond_entry(const struct bond *,
139 static tag_type bond_get_active_slave_tag(const struct bond *);
140 static struct bond_slave *choose_output_slave(const struct bond *,
143 static void bond_update_fake_slave_stats(struct bond *);
145 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
146 * stores the mode in '*balance' and returns true. Otherwise returns false
147 * without modifying '*balance'. */
149 bond_mode_from_string(enum bond_mode *balance, const char *s)
151 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
153 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
155 } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
156 *balance = BM_STABLE;
157 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
165 /* Returns a string representing 'balance'. */
167 bond_mode_to_string(enum bond_mode balance) {
170 return "balance-tcp";
172 return "balance-slb";
176 return "active-backup";
181 /* Attempts to parse 's' as the name of a bond link status detection mode. If
182 * successful, stores the mode in '*detect' and returns true. Otherwise
183 * returns false without modifying '*detect'. */
185 bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
187 if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
188 *detect = BLSM_CARRIER;
189 } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
190 *detect = BLSM_MIIMON;
197 /* Returns a string representing 'detect'. */
199 bond_detect_mode_to_string(enum bond_detect_mode detect)
210 /* Creates and returns a new bond whose configuration is initially taken from
213 * The caller should register each slave on the new bond by calling
214 * bond_slave_register(). */
216 bond_create(const struct bond_settings *s)
220 bond = xzalloc(sizeof *bond);
221 hmap_init(&bond->slaves);
222 bond->no_slaves_tag = tag_create_random();
223 bond->stb_tag = tag_create_random();
224 bond->miimon_next_update = LLONG_MAX;
225 bond->next_fake_iface_update = LLONG_MAX;
227 bond_reconfigure(bond, s);
229 tag_set_init(&bond->unixctl_tags);
236 bond_destroy(struct bond *bond)
238 struct bond_slave *slave, *next_slave;
244 hmap_remove(&all_bonds, &bond->hmap_node);
246 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
247 hmap_remove(&bond->slaves, &slave->hmap_node);
248 /* Client owns 'slave->netdev'. */
252 hmap_destroy(&bond->slaves);
256 netdev_monitor_destroy(bond->monitor);
262 /* Updates 'bond''s overall configuration to 's'.
264 * The caller should register each slave on 'bond' by calling
265 * bond_slave_register(). This is optional if none of the slaves'
266 * configuration has changed. In any case it can't hurt.
268 * Returns true if the configuration has changed in such a way that requires
272 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
274 bool revalidate = false;
276 if (!bond->name || strcmp(bond->name, s->name)) {
278 hmap_remove(&all_bonds, &bond->hmap_node);
281 bond->name = xstrdup(s->name);
282 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
285 bond->detect = s->detect;
286 bond->miimon_interval = s->miimon_interval;
287 bond->updelay = s->up_delay;
288 bond->downdelay = s->down_delay;
289 bond->rebalance_interval = s->rebalance_interval;
291 if (bond->balance != s->balance) {
292 bond->balance = s->balance;
296 if (bond->basis != s->basis) {
297 bond->basis = s->basis;
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->bond_revalidate) {
330 bond->bond_revalidate = false;
333 if (bond->balance == BM_AB || !bond->hash || revalidate) {
334 bond_entry_reset(bond);
341 bond_slave_set_netdev__(struct bond *bond, struct bond_slave *slave,
342 struct netdev *netdev)
344 if (slave->netdev != netdev) {
347 netdev_monitor_remove(bond->monitor, slave->netdev);
349 netdev_monitor_add(bond->monitor, netdev);
351 slave->netdev = netdev;
355 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
356 * arbitrary client-provided pointer that uniquely identifies a slave within a
357 * bond. If 'slave_' already exists within 'bond' then this function
358 * reconfigures the existing slave.
360 * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
361 * across restarts and distributed vswitch instances. It should be unique per
362 * slave, and preferably consistent across restarts and reconfigurations.
364 * 'netdev' must be the network device that 'slave_' represents. It is owned
365 * by the client, so the client must not close it before either unregistering
366 * 'slave_' or destroying 'bond'.
369 bond_slave_register(struct bond *bond, void *slave_, uint32_t stb_id,
370 struct netdev *netdev)
372 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
375 slave = xzalloc(sizeof *slave);
377 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
380 slave->delay_expires = LLONG_MAX;
381 slave->up = bond_is_link_up(bond, netdev);
382 slave->name = xstrdup(netdev_get_name(netdev));
383 bond->bond_revalidate = true;
385 slave->enabled = false;
386 bond_enable_slave(slave, slave->up, NULL);
389 if (slave->stb_id != stb_id) {
390 slave->stb_id = stb_id;
391 bond->bond_revalidate = true;
394 bond_slave_set_netdev__(bond, slave, netdev);
397 slave->name = xstrdup(netdev_get_name(netdev));
400 /* Updates the network device to be used with 'slave_' to 'netdev'.
402 * This is useful if the caller closes and re-opens the network device
403 * registered with bond_slave_register() but doesn't need to change anything
406 bond_slave_set_netdev(struct bond *bond, void *slave_, struct netdev *netdev)
408 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
410 bond_slave_set_netdev__(bond, slave, netdev);
414 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
415 * then this function has no effect.
417 * Unregistering a slave invalidates all flows. */
419 bond_slave_unregister(struct bond *bond, const void *slave_)
421 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
429 netdev_monitor_remove(bond->monitor, slave->netdev);
432 bond_enable_slave(slave, false, NULL);
434 del_active = bond->active_slave == slave;
436 struct bond_entry *e;
437 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
438 if (e->slave == slave) {
446 hmap_remove(&bond->slaves, &slave->hmap_node);
447 /* Client owns 'slave->netdev'. */
454 bond_choose_active_slave(bond, &tags);
455 bond->send_learning_packets = true;
459 /* Should be called on each slave in 'bond' before bond_run() to indicate the
460 * results of lacp_slave_may_enable() on 'slave_'. */
462 bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
465 bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
468 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
469 * allow tagged flows to be invalidated.
471 * The caller should check bond_should_send_learning_packets() afterward. */
473 bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
475 struct bond_slave *slave;
476 bool is_tcp_hash = bond_is_tcp_hash(bond);
478 bond->lacp_negotiated = lacp_negotiated;
480 /* Update link status. */
481 if (bond->detect == BLSM_CARRIER
482 || time_msec() >= bond->miimon_next_update)
484 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
485 slave->up = bond_is_link_up(bond, slave->netdev);
487 bond->miimon_next_update = time_msec() + bond->miimon_interval;
490 /* Enable slaves based on link status and LACP feedback. */
491 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
492 bond_link_status_update(slave, tags);
494 if (!bond->active_slave || !bond->active_slave->enabled) {
495 bond_choose_active_slave(bond, tags);
498 /* Update fake bond interface stats. */
499 if (time_msec() >= bond->next_fake_iface_update) {
500 bond_update_fake_slave_stats(bond);
501 bond->next_fake_iface_update = time_msec() + 1000;
504 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
505 bond->bond_revalidate = true;
508 if (bond->bond_revalidate) {
509 bond->bond_revalidate = false;
511 bond_entry_reset(bond);
512 if (bond->balance != BM_STABLE) {
513 struct bond_slave *slave;
515 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
516 tag_set_add(tags, slave->tag);
519 tag_set_add(tags, bond->stb_tag);
521 tag_set_add(tags, bond->no_slaves_tag);
524 /* Invalidate any tags required by */
525 tag_set_union(tags, &bond->unixctl_tags);
526 tag_set_init(&bond->unixctl_tags);
529 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
531 bond_wait(struct bond *bond)
533 struct bond_slave *slave;
535 if (bond->detect == BLSM_CARRIER) {
536 netdev_monitor_poll_wait(bond->monitor);
538 poll_timer_wait_until(bond->miimon_next_update);
541 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
542 if (slave->delay_expires != LLONG_MAX) {
543 poll_timer_wait_until(slave->delay_expires);
547 if (bond->next_fake_iface_update != LLONG_MAX) {
548 poll_timer_wait_until(bond->next_fake_iface_update);
551 /* Ensure that any saved tags get revalidated right away. */
552 if (!tag_set_is_empty(&bond->unixctl_tags)) {
553 poll_immediate_wake();
556 /* We don't wait for bond->next_rebalance because rebalancing can only run
557 * at a flow account checkpoint. ofproto does checkpointing on its own
558 * schedule and bond_rebalance() gets called afterward, so we'd just be
559 * waking up for no purpose. */
562 /* MAC learning table interaction. */
565 may_send_learning_packets(const struct bond *bond)
567 return !bond->lacp_negotiated && bond->balance != BM_AB;
570 /* Returns true if 'bond' needs the client to send out packets to assist with
571 * MAC learning on 'bond'. If this function returns true, then the client
572 * should iterate through its MAC learning table for the bridge on which 'bond'
573 * is located. For each MAC that has been learned on a port other than 'bond',
574 * it should call bond_send_learning_packet().
576 * This function will only return true if 'bond' is in SLB mode and LACP is not
577 * negotiated. Otherwise sending learning packets isn't necessary.
579 * Calling this function resets the state that it checks. */
581 bond_should_send_learning_packets(struct bond *bond)
583 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
584 bond->send_learning_packets = false;
588 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
590 * See bond_should_send_learning_packets() for description of usage. */
592 bond_send_learning_packet(struct bond *bond,
593 const uint8_t eth_src[ETH_ADDR_LEN],
596 struct bond_slave *slave;
597 struct ofpbuf packet;
601 assert(may_send_learning_packets(bond));
602 if (!bond->active_slave) {
603 /* Nowhere to send the learning packet. */
607 memset(&flow, 0, sizeof flow);
608 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
609 slave = choose_output_slave(bond, &flow, vlan);
611 ofpbuf_init(&packet, 0);
612 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
615 eth_set_vlan_tci(&packet, htons(vlan));
617 error = netdev_send(slave->netdev, &packet);
618 ofpbuf_uninit(&packet);
623 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
624 * Ethernet destination address of 'eth_dst', should be admitted.
626 * The return value is one of the following:
628 * - BV_ACCEPT: Admit the packet.
630 * - BV_DROP: Drop the packet.
632 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
633 * Ethernet source address and VLAN. If there is none, or if the packet
634 * is on the learned port, then admit the packet. If a different port has
635 * been learned, however, drop the packet (and do not use it for MAC
639 bond_check_admissibility(struct bond *bond, const void *slave_,
640 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
642 /* Admit all packets if LACP has been negotiated, because that means that
643 * the remote switch is aware of the bond and will "do the right thing". */
644 if (bond->lacp_negotiated) {
648 /* Drop all multicast packets on inactive slaves. */
649 if (eth_addr_is_multicast(eth_dst)) {
650 *tags |= bond_get_active_slave_tag(bond);
651 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
656 /* Drop all packets for which we have learned a different input port,
657 * because we probably sent the packet on one slave and got it back on the
658 * other. Gratuitous ARP packets are an exception to this rule: the host
659 * has moved to another switch. The exception to the exception is if we
660 * locked the learning table to avoid reflections on bond slaves. */
661 return BV_DROP_IF_MOVED;
664 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
665 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
666 * NULL if the packet should be dropped because no slaves are enabled.
668 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
669 * should be a VID only (i.e. excluding the PCP bits). Second,
670 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
671 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
672 * packet belongs to (so for an access port it will be the access port's VLAN).
674 * Adds a tag to '*tags' that associates the flow with the returned slave.
677 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
678 uint16_t vlan, tag_type *tags)
680 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
682 *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
685 *tags |= bond->no_slaves_tag;
693 bond_is_balanced(const struct bond *bond)
695 return bond->balance == BM_SLB || bond->balance == BM_TCP;
698 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
700 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
704 if (bond_is_balanced(bond)) {
705 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
709 static struct bond_slave *
710 bond_slave_from_bal_node(struct list *bal)
712 return CONTAINER_OF(bal, struct bond_slave, bal_node);
716 log_bals(struct bond *bond, const struct list *bals)
718 if (VLOG_IS_DBG_ENABLED()) {
719 struct ds ds = DS_EMPTY_INITIALIZER;
720 const struct bond_slave *slave;
722 LIST_FOR_EACH (slave, bal_node, bals) {
724 ds_put_char(&ds, ',');
726 ds_put_format(&ds, " %s %"PRIu64"kB",
727 slave->name, slave->tx_bytes / 1024);
729 if (!slave->enabled) {
730 ds_put_cstr(&ds, " (disabled)");
732 if (!list_is_empty(&slave->entries)) {
733 struct bond_entry *e;
735 ds_put_cstr(&ds, " (");
736 LIST_FOR_EACH (e, list_node, &slave->entries) {
737 if (&e->list_node != list_front(&slave->entries)) {
738 ds_put_cstr(&ds, " + ");
740 ds_put_format(&ds, "h%td: %"PRIu64"kB",
741 e - bond->hash, e->tx_bytes / 1024);
743 ds_put_cstr(&ds, ")");
746 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
751 /* Shifts 'hash' from its current slave to 'to'. */
753 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
756 struct bond_slave *from = hash->slave;
757 struct bond *bond = from->bond;
758 uint64_t delta = hash->tx_bytes;
760 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
761 "from %s to %s (now carrying %"PRIu64"kB and "
762 "%"PRIu64"kB load, respectively)",
763 bond->name, delta / 1024, hash - bond->hash,
764 from->name, to->name,
765 (from->tx_bytes - delta) / 1024,
766 (to->tx_bytes + delta) / 1024);
768 /* Shift load away from 'from' to 'to'. */
769 from->tx_bytes -= delta;
770 to->tx_bytes += delta;
772 /* Arrange for flows to be revalidated. */
773 tag_set_add(set, hash->tag);
775 hash->tag = tag_create_random();
778 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
779 * given that doing so must decrease the ratio of the load on the two slaves by
780 * at least 0.1. Returns NULL if there is no appropriate entry.
782 * The list of entries isn't sorted. I don't know of a reason to prefer to
783 * shift away small hashes or large hashes. */
784 static struct bond_entry *
785 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
787 struct bond_entry *e;
789 if (list_is_short(&from->entries)) {
790 /* 'from' carries no more than one MAC hash, so shifting load away from
791 * it would be pointless. */
795 LIST_FOR_EACH (e, list_node, &from->entries) {
796 double old_ratio, new_ratio;
799 if (to_tx_bytes == 0) {
800 /* Nothing on the new slave, move it. */
805 old_ratio = (double)from->tx_bytes / to_tx_bytes;
806 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
807 if (old_ratio - new_ratio > 0.1) {
808 /* Would decrease the ratio, move it. */
816 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
819 insert_bal(struct list *bals, struct bond_slave *slave)
821 struct bond_slave *pos;
823 LIST_FOR_EACH (pos, bal_node, bals) {
824 if (slave->tx_bytes > pos->tx_bytes) {
828 list_insert(&pos->bal_node, &slave->bal_node);
831 /* Removes 'slave' from its current list and then inserts it into 'bals' so
832 * that descending order of 'tx_bytes' is maintained. */
834 reinsert_bal(struct list *bals, struct bond_slave *slave)
836 list_remove(&slave->bal_node);
837 insert_bal(bals, slave);
840 /* If 'bond' needs rebalancing, does so.
842 * The caller should have called bond_account() for each active flow, to ensure
843 * that flow data is consistently accounted at this point. */
845 bond_rebalance(struct bond *bond, struct tag_set *tags)
847 struct bond_slave *slave;
848 struct bond_entry *e;
851 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
854 bond->next_rebalance = time_msec() + bond->rebalance_interval;
856 /* Add each bond_entry to its slave's 'entries' list.
857 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
858 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
860 list_init(&slave->entries);
862 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
863 if (e->slave && e->tx_bytes) {
864 e->slave->tx_bytes += e->tx_bytes;
865 list_push_back(&e->slave->entries, &e->list_node);
869 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
871 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
872 * with a proper list sort algorithm. */
874 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
875 if (slave->enabled) {
876 insert_bal(&bals, slave);
879 log_bals(bond, &bals);
881 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
882 while (!list_is_short(&bals)) {
883 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
884 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
887 overload = from->tx_bytes - to->tx_bytes;
888 if (overload < to->tx_bytes >> 5 || overload < 100000) {
889 /* The extra load on 'from' (and all less-loaded slaves), compared
890 * to that of 'to' (the least-loaded slave), is less than ~3%, or
891 * it is less than ~1Mbps. No point in rebalancing. */
895 /* 'from' is carrying significantly more load than 'to', and that load
896 * is split across at least two different hashes. */
897 e = choose_entry_to_migrate(from, to->tx_bytes);
899 bond_shift_load(e, to, tags);
901 /* Delete element from from->entries.
903 * We don't add the element to to->hashes. That would only allow
904 * 'e' to be migrated to another slave in this rebalancing run, and
905 * there is no point in doing that. */
906 list_remove(&e->list_node);
908 /* Re-sort 'bals'. */
909 reinsert_bal(&bals, from);
910 reinsert_bal(&bals, to);
912 /* Can't usefully migrate anything away from 'from'.
913 * Don't reconsider it. */
914 list_remove(&from->bal_node);
918 /* Implement exponentially weighted moving average. A weight of 1/2 causes
919 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
920 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
921 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
929 /* Bonding unixctl user interface functions. */
932 bond_find(const char *name)
936 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
938 if (!strcmp(bond->name, name)) {
945 static struct bond_slave *
946 bond_lookup_slave(struct bond *bond, const char *slave_name)
948 struct bond_slave *slave;
950 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
951 if (!strcmp(slave->name, slave_name)) {
959 bond_unixctl_list(struct unixctl_conn *conn,
960 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
962 struct ds ds = DS_EMPTY_INITIALIZER;
963 const struct bond *bond;
965 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
967 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
968 const struct bond_slave *slave;
971 ds_put_format(&ds, "%s\t%s\t",
972 bond->name, bond_mode_to_string(bond->balance));
975 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
977 ds_put_cstr(&ds, ", ");
979 ds_put_cstr(&ds, slave->name);
981 ds_put_char(&ds, '\n');
983 unixctl_command_reply(conn, 200, ds_cstr(&ds));
988 bond_unixctl_show(struct unixctl_conn *conn,
989 const char *args, void *aux OVS_UNUSED)
991 struct ds ds = DS_EMPTY_INITIALIZER;
992 const struct bond_slave *slave;
993 const struct bond *bond;
995 bond = bond_find(args);
997 unixctl_command_reply(conn, 501, "no such bond");
1001 ds_put_format(&ds, "bond_mode: %s\n",
1002 bond_mode_to_string(bond->balance));
1004 if (bond->balance != BM_AB) {
1005 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
1006 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
1009 ds_put_format(&ds, "bond-hash-basis: %"PRIu32"\n", bond->basis);
1011 ds_put_format(&ds, "bond-detect-mode: %s\n",
1012 bond->monitor ? "carrier" : "miimon");
1014 if (!bond->monitor) {
1015 ds_put_format(&ds, "bond-miimon-interval: %lld\n",
1016 bond->miimon_interval);
1019 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
1020 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
1022 if (bond_is_balanced(bond)) {
1023 ds_put_format(&ds, "next rebalance: %lld ms\n",
1024 bond->next_rebalance - time_msec());
1027 ds_put_format(&ds, "lacp_negotiated: %s\n",
1028 bond->lacp_negotiated ? "true" : "false");
1030 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1031 struct bond_entry *be;
1035 ds_put_format(&ds, "\nslave %s: %s\n",
1036 slave->name, slave->enabled ? "enabled" : "disabled");
1037 if (slave == bond->active_slave) {
1038 ds_put_cstr(&ds, "\tactive slave\n");
1040 if (slave->delay_expires != LLONG_MAX) {
1041 ds_put_format(&ds, "\t%s expires in %lld ms\n",
1042 slave->enabled ? "downdelay" : "updelay",
1043 slave->delay_expires - time_msec());
1046 ds_put_format(&ds, "\tlacp_may_enable: %s\n",
1047 slave->lacp_may_enable ? "true" : "false");
1049 if (!bond_is_balanced(bond)) {
1054 memset(&flow, 0, sizeof flow);
1055 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
1056 int hash = be - bond->hash;
1058 if (be->slave != slave) {
1062 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
1063 hash, be->tx_bytes / 1024);
1065 if (bond->balance != BM_SLB) {
1069 /* XXX How can we list the MACs assigned to hashes? */
1072 unixctl_command_reply(conn, 200, ds_cstr(&ds));
1077 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1078 void *aux OVS_UNUSED)
1080 char *args = (char *) args_;
1081 char *save_ptr = NULL;
1082 char *bond_s, *hash_s, *slave_s;
1084 struct bond_slave *slave;
1085 struct bond_entry *entry;
1088 bond_s = strtok_r(args, " ", &save_ptr);
1089 hash_s = strtok_r(NULL, " ", &save_ptr);
1090 slave_s = strtok_r(NULL, " ", &save_ptr);
1092 unixctl_command_reply(conn, 501,
1093 "usage: bond/migrate BOND HASH SLAVE");
1097 bond = bond_find(bond_s);
1099 unixctl_command_reply(conn, 501, "no such bond");
1103 if (bond->balance != BM_SLB) {
1104 unixctl_command_reply(conn, 501, "not an SLB bond");
1108 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1109 hash = atoi(hash_s) & BOND_MASK;
1111 unixctl_command_reply(conn, 501, "bad hash");
1115 slave = bond_lookup_slave(bond, slave_s);
1117 unixctl_command_reply(conn, 501, "no such slave");
1121 if (!slave->enabled) {
1122 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1126 entry = &bond->hash[hash];
1127 tag_set_add(&bond->unixctl_tags, entry->tag);
1128 entry->slave = slave;
1129 entry->tag = tag_create_random();
1130 unixctl_command_reply(conn, 200, "migrated");
1134 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1135 void *aux OVS_UNUSED)
1137 char *args = (char *) args_;
1138 char *save_ptr = NULL;
1139 char *bond_s, *slave_s;
1141 struct bond_slave *slave;
1143 bond_s = strtok_r(args, " ", &save_ptr);
1144 slave_s = strtok_r(NULL, " ", &save_ptr);
1146 unixctl_command_reply(conn, 501,
1147 "usage: bond/set-active-slave BOND SLAVE");
1151 bond = bond_find(bond_s);
1153 unixctl_command_reply(conn, 501, "no such bond");
1157 slave = bond_lookup_slave(bond, slave_s);
1159 unixctl_command_reply(conn, 501, "no such slave");
1163 if (!slave->enabled) {
1164 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1168 if (bond->active_slave != slave) {
1169 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1170 bond->active_slave = slave;
1171 bond->active_slave->tag = tag_create_random();
1172 VLOG_INFO("bond %s: active interface is now %s",
1173 bond->name, slave->name);
1174 bond->send_learning_packets = true;
1175 unixctl_command_reply(conn, 200, "done");
1177 unixctl_command_reply(conn, 200, "no change");
1182 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1184 char *args = (char *) args_;
1185 char *save_ptr = NULL;
1186 char *bond_s, *slave_s;
1188 struct bond_slave *slave;
1190 bond_s = strtok_r(args, " ", &save_ptr);
1191 slave_s = strtok_r(NULL, " ", &save_ptr);
1193 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1194 enable ? "enable" : "disable");
1195 unixctl_command_reply(conn, 501, usage);
1200 bond = bond_find(bond_s);
1202 unixctl_command_reply(conn, 501, "no such bond");
1206 slave = bond_lookup_slave(bond, slave_s);
1208 unixctl_command_reply(conn, 501, "no such slave");
1212 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1213 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1217 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1218 void *aux OVS_UNUSED)
1220 enable_slave(conn, args, true);
1224 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1225 void *aux OVS_UNUSED)
1227 enable_slave(conn, args, false);
1231 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1232 void *aux OVS_UNUSED)
1234 char *args = (char *) args_;
1235 uint8_t mac[ETH_ADDR_LEN];
1240 char *mac_s, *vlan_s, *basis_s;
1241 char *save_ptr = NULL;
1243 mac_s = strtok_r(args, " ", &save_ptr);
1244 vlan_s = strtok_r(NULL, " ", &save_ptr);
1245 basis_s = strtok_r(NULL, " ", &save_ptr);
1248 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1249 unixctl_command_reply(conn, 501, "invalid vlan");
1253 vlan = OFP_VLAN_NONE;
1257 if (sscanf(basis_s, "%"PRIu32, &basis) != 1) {
1258 unixctl_command_reply(conn, 501, "invalid basis");
1265 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1266 == ETH_ADDR_SCAN_COUNT) {
1267 hash = bond_hash_src(mac, vlan, basis) & BOND_MASK;
1269 hash_cstr = xasprintf("%u", hash);
1270 unixctl_command_reply(conn, 200, hash_cstr);
1273 unixctl_command_reply(conn, 501, "invalid mac");
1280 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1281 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1282 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1283 unixctl_command_register("bond/set-active-slave",
1284 bond_unixctl_set_active_slave, NULL);
1285 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1287 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1289 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1293 bond_entry_reset(struct bond *bond)
1295 if (bond->balance != BM_AB) {
1296 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1299 bond->hash = xmalloc(hash_len);
1301 memset(bond->hash, 0, hash_len);
1303 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1310 static struct bond_slave *
1311 bond_slave_lookup(struct bond *bond, const void *slave_)
1313 struct bond_slave *slave;
1315 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1317 if (slave->aux == slave_) {
1326 bond_is_link_up(struct bond *bond, struct netdev *netdev)
1328 return (bond->detect == BLSM_CARRIER
1329 ? netdev_get_carrier(netdev)
1330 : netdev_get_miimon(netdev));
1334 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1336 struct bond *bond = slave->bond;
1337 slave->delay_expires = LLONG_MAX;
1338 if (enable != slave->enabled) {
1339 slave->enabled = enable;
1340 if (!slave->enabled) {
1341 VLOG_WARN("interface %s: disabled", slave->name);
1343 tag_set_add(tags, slave->tag);
1346 VLOG_WARN("interface %s: enabled", slave->name);
1347 slave->tag = tag_create_random();
1350 if (bond->balance == BM_STABLE) {
1351 bond->bond_revalidate = true;
1357 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1359 struct bond *bond = slave->bond;
1362 up = slave->up && slave->lacp_may_enable;
1363 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1364 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1365 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1366 slave->name, up ? "up" : "down");
1367 if (up == slave->enabled) {
1368 slave->delay_expires = LLONG_MAX;
1369 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1370 slave->name, up ? "disabled" : "enabled");
1372 int delay = (bond->lacp_negotiated ? 0
1373 : up ? bond->updelay : bond->downdelay);
1374 slave->delay_expires = time_msec() + delay;
1376 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1379 up ? "enabled" : "disabled",
1386 if (time_msec() >= slave->delay_expires) {
1387 bond_enable_slave(slave, up, tags);
1392 bond_is_tcp_hash(const struct bond *bond)
1394 return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
1395 && bond->lacp_negotiated;
1399 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan, uint32_t basis)
1401 return hash_3words(hash_bytes(mac, ETH_ADDR_LEN, 0), vlan, basis);
1405 bond_hash_tcp(const struct flow *flow, uint16_t vlan, uint32_t basis)
1407 struct flow hash_flow = *flow;
1408 hash_flow.vlan_tci = vlan;
1410 /* The symmetric quality of this hash function is not required, but
1411 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1412 * purposes, so we use it out of convenience. */
1413 return flow_hash_symmetric_l4(&hash_flow, basis);
1417 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1419 assert(bond->balance != BM_AB);
1421 return (bond_is_tcp_hash(bond)
1422 ? bond_hash_tcp(flow, vlan, bond->basis)
1423 : bond_hash_src(flow->dl_src, vlan, bond->basis));
1426 static struct bond_entry *
1427 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1430 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1433 /* This function uses Highest Random Weight hashing to choose an output slave.
1434 * This approach only reassigns a minimal number of flows when slaves are
1435 * enabled or disabled. Unfortunately, it has O(n) performance against the
1436 * number of slaves. There exist algorithms which are O(1), but have slightly
1437 * more complex implementations and require the use of memory. This may need
1438 * to be reimplemented if it becomes a performance bottleneck. */
1439 static struct bond_slave *
1440 choose_stb_slave(const struct bond *bond, const struct flow *flow,
1443 struct bond_slave *best, *slave;
1444 uint32_t best_hash, flow_hash;
1448 flow_hash = bond_hash(bond, flow, vlan);
1449 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1450 if (slave->enabled) {
1453 hash = hash_2words(flow_hash, slave->stb_id);
1454 if (!best || hash > best_hash) {
1464 static struct bond_slave *
1465 choose_output_slave(const struct bond *bond, const struct flow *flow,
1468 struct bond_entry *e;
1470 switch (bond->balance) {
1472 return bond->active_slave;
1475 return choose_stb_slave(bond, flow, vlan);
1478 e = lookup_bond_entry(bond, flow, vlan);
1479 if (!e->slave || !e->slave->enabled) {
1480 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1481 struct bond_slave, hmap_node);
1482 if (!e->slave->enabled) {
1483 e->slave = bond->active_slave;
1485 e->tag = tag_create_random();
1494 static struct bond_slave *
1495 bond_choose_slave(const struct bond *bond)
1497 struct bond_slave *slave, *best;
1499 /* Find an enabled slave. */
1500 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1501 if (slave->enabled) {
1506 /* All interfaces are disabled. Find an interface that will be enabled
1507 * after its updelay expires. */
1509 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1510 if (slave->delay_expires != LLONG_MAX
1511 && slave->lacp_may_enable
1512 && (!best || slave->delay_expires < best->delay_expires)) {
1520 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1522 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1523 struct bond_slave *old_active_slave = bond->active_slave;
1525 bond->active_slave = bond_choose_slave(bond);
1526 if (bond->active_slave) {
1527 if (bond->active_slave->enabled) {
1528 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1529 bond->name, bond->active_slave->name);
1531 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1532 "remaining %lld ms updelay (since no interface was "
1533 "enabled)", bond->name, bond->active_slave->name,
1534 bond->active_slave->delay_expires - time_msec());
1535 bond_enable_slave(bond->active_slave, true, tags);
1538 if (!old_active_slave) {
1539 tag_set_add(tags, bond->no_slaves_tag);
1542 bond->send_learning_packets = true;
1543 } else if (old_active_slave) {
1544 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1548 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1549 * there is no active slave. */
1551 bond_get_active_slave_tag(const struct bond *bond)
1553 return (bond->active_slave
1554 ? bond->active_slave->tag
1555 : bond->no_slaves_tag);
1558 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1559 * bond interface. */
1561 bond_update_fake_slave_stats(struct bond *bond)
1563 struct netdev_stats bond_stats;
1564 struct bond_slave *slave;
1565 struct netdev *bond_dev;
1567 memset(&bond_stats, 0, sizeof bond_stats);
1569 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1570 struct netdev_stats slave_stats;
1572 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1573 /* XXX: We swap the stats here because they are swapped back when
1574 * reported by the internal device. The reason for this is
1575 * internal devices normally represent packets going into the
1576 * system but when used as fake bond device they represent packets
1577 * leaving the system. We really should do this in the internal
1578 * device itself because changing it here reverses the counts from
1579 * the perspective of the switch. However, the internal device
1580 * doesn't know what type of device it represents so we have to do
1581 * it here for now. */
1582 bond_stats.tx_packets += slave_stats.rx_packets;
1583 bond_stats.tx_bytes += slave_stats.rx_bytes;
1584 bond_stats.rx_packets += slave_stats.tx_packets;
1585 bond_stats.rx_bytes += slave_stats.tx_bytes;
1589 if (!netdev_open_default(bond->name, &bond_dev)) {
1590 netdev_set_stats(bond_dev, &bond_stats);
1591 netdev_close(bond_dev);