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 enabled; /* May be chosen for flows? */
67 bool may_enable; /* Client considers this slave bondable. */
68 tag_type tag; /* Tag associated with this slave. */
70 /* Rebalancing info. Used only by bond_rebalance(). */
71 struct list bal_node; /* In bond_rebalance()'s 'bals' list. */
72 struct list entries; /* 'struct bond_entry's assigned here. */
73 uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
75 /* BM_STABLE specific bonding info. */
76 uint32_t stb_id; /* ID used for 'stb_slaves' ordering. */
79 /* A bond, that is, a set of network devices grouped to improve performance or
82 struct hmap_node hmap_node; /* In 'all_bonds' hmap. */
83 char *name; /* Name provided by client. */
89 enum bond_mode balance; /* Balancing mode, one of BM_*. */
90 struct bond_slave *active_slave;
91 tag_type no_slaves_tag; /* Tag for flows when all slaves disabled. */
92 int updelay, downdelay; /* Delay before slave goes up/down, in ms. */
93 bool lacp_negotiated; /* LACP negotiations were successful. */
94 bool bond_revalidate; /* True if flows need revalidation. */
95 uint32_t basis; /* Basis for flow hash function. */
97 /* SLB specific bonding info. */
98 struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
99 int rebalance_interval; /* Interval between rebalances, in ms. */
100 long long int next_rebalance; /* Next rebalancing time. */
101 bool send_learning_packets;
103 /* BM_STABLE specific bonding info. */
104 tag_type stb_tag; /* Tag associated with this bond. */
107 struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
109 /* Legacy compatibility. */
110 long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
112 /* Tag set saved for next bond_run(). This tag set is a kluge for cases
113 * where we can't otherwise provide revalidation feedback to the client.
114 * That's only unixctl commands now; I hope no other cases will arise. */
115 struct tag_set unixctl_tags;
118 static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
120 static void bond_entry_reset(struct bond *);
121 static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
122 static void bond_enable_slave(struct bond_slave *, bool enable,
124 static void bond_link_status_update(struct bond_slave *, struct tag_set *);
125 static void bond_choose_active_slave(struct bond *, struct tag_set *);
126 static bool bond_is_tcp_hash(const struct bond *);
127 static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
128 uint16_t vlan, uint32_t basis);
129 static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan,
131 static struct bond_entry *lookup_bond_entry(const struct bond *,
134 static tag_type bond_get_active_slave_tag(const struct bond *);
135 static struct bond_slave *choose_output_slave(const struct bond *,
138 static void bond_update_fake_slave_stats(struct bond *);
140 /* Attempts to parse 's' as the name of a bond balancing mode. If successful,
141 * stores the mode in '*balance' and returns true. Otherwise returns false
142 * without modifying '*balance'. */
144 bond_mode_from_string(enum bond_mode *balance, const char *s)
146 if (!strcmp(s, bond_mode_to_string(BM_TCP))) {
148 } else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
150 } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
151 *balance = BM_STABLE;
152 } else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
160 /* Returns a string representing 'balance'. */
162 bond_mode_to_string(enum bond_mode balance) {
165 return "balance-tcp";
167 return "balance-slb";
171 return "active-backup";
177 /* Creates and returns a new bond whose configuration is initially taken from
180 * The caller should register each slave on the new bond by calling
181 * bond_slave_register(). */
183 bond_create(const struct bond_settings *s)
187 bond = xzalloc(sizeof *bond);
188 hmap_init(&bond->slaves);
189 bond->no_slaves_tag = tag_create_random();
190 bond->stb_tag = tag_create_random();
191 bond->next_fake_iface_update = LLONG_MAX;
192 bond->monitor = netdev_monitor_create();
194 bond_reconfigure(bond, s);
196 tag_set_init(&bond->unixctl_tags);
203 bond_destroy(struct bond *bond)
205 struct bond_slave *slave, *next_slave;
211 hmap_remove(&all_bonds, &bond->hmap_node);
213 HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
214 hmap_remove(&bond->slaves, &slave->hmap_node);
215 /* Client owns 'slave->netdev'. */
219 hmap_destroy(&bond->slaves);
223 netdev_monitor_destroy(bond->monitor);
229 /* Updates 'bond''s overall configuration to 's'.
231 * The caller should register each slave on 'bond' by calling
232 * bond_slave_register(). This is optional if none of the slaves'
233 * configuration has changed. In any case it can't hurt.
235 * Returns true if the configuration has changed in such a way that requires
239 bond_reconfigure(struct bond *bond, const struct bond_settings *s)
241 bool revalidate = false;
243 if (!bond->name || strcmp(bond->name, s->name)) {
245 hmap_remove(&all_bonds, &bond->hmap_node);
248 bond->name = xstrdup(s->name);
249 hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
252 bond->updelay = s->up_delay;
253 bond->downdelay = s->down_delay;
254 bond->rebalance_interval = s->rebalance_interval;
256 if (bond->balance != s->balance) {
257 bond->balance = s->balance;
261 if (bond->basis != s->basis) {
262 bond->basis = s->basis;
267 if (bond->next_fake_iface_update == LLONG_MAX) {
268 bond->next_fake_iface_update = time_msec();
271 bond->next_fake_iface_update = LLONG_MAX;
274 if (bond->bond_revalidate) {
276 bond->bond_revalidate = false;
279 if (bond->balance == BM_AB || !bond->hash || revalidate) {
280 bond_entry_reset(bond);
287 bond_slave_set_netdev__(struct bond *bond, struct bond_slave *slave,
288 struct netdev *netdev)
290 if (slave->netdev != netdev) {
292 netdev_monitor_remove(bond->monitor, slave->netdev);
294 netdev_monitor_add(bond->monitor, netdev);
295 slave->netdev = netdev;
299 /* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
300 * arbitrary client-provided pointer that uniquely identifies a slave within a
301 * bond. If 'slave_' already exists within 'bond' then this function
302 * reconfigures the existing slave.
304 * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
305 * across restarts and distributed vswitch instances. It should be unique per
306 * slave, and preferably consistent across restarts and reconfigurations.
308 * 'netdev' must be the network device that 'slave_' represents. It is owned
309 * by the client, so the client must not close it before either unregistering
310 * 'slave_' or destroying 'bond'.
313 bond_slave_register(struct bond *bond, void *slave_, uint32_t stb_id,
314 struct netdev *netdev)
316 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
319 slave = xzalloc(sizeof *slave);
321 hmap_insert(&bond->slaves, &slave->hmap_node, hash_pointer(slave_, 0));
324 slave->delay_expires = LLONG_MAX;
325 slave->name = xstrdup(netdev_get_name(netdev));
326 bond->bond_revalidate = true;
328 slave->enabled = false;
329 bond_enable_slave(slave, netdev_get_carrier(netdev), NULL);
332 if (slave->stb_id != stb_id) {
333 slave->stb_id = stb_id;
334 bond->bond_revalidate = true;
337 bond_slave_set_netdev__(bond, slave, netdev);
340 slave->name = xstrdup(netdev_get_name(netdev));
343 /* Updates the network device to be used with 'slave_' to 'netdev'.
345 * This is useful if the caller closes and re-opens the network device
346 * registered with bond_slave_register() but doesn't need to change anything
349 bond_slave_set_netdev(struct bond *bond, void *slave_, struct netdev *netdev)
351 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
353 bond_slave_set_netdev__(bond, slave, netdev);
357 /* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
358 * then this function has no effect.
360 * Unregistering a slave invalidates all flows. */
362 bond_slave_unregister(struct bond *bond, const void *slave_)
364 struct bond_slave *slave = bond_slave_lookup(bond, slave_);
371 netdev_monitor_remove(bond->monitor, slave->netdev);
372 bond_enable_slave(slave, false, NULL);
374 del_active = bond->active_slave == slave;
376 struct bond_entry *e;
377 for (e = bond->hash; e <= &bond->hash[BOND_MASK]; e++) {
378 if (e->slave == slave) {
386 hmap_remove(&bond->slaves, &slave->hmap_node);
387 /* Client owns 'slave->netdev'. */
394 bond_choose_active_slave(bond, &tags);
395 bond->send_learning_packets = true;
399 /* Should be called on each slave in 'bond' before bond_run() to indicate
400 * whether or not 'slave_' may be enabled. This function is intended to allow
401 * other protocols to have some impact on bonding decisions. For example LACP
402 * or high level link monitoring protocols may decide that a given slave should
403 * not be able to send traffic. */
405 bond_slave_set_may_enable(struct bond *bond, void *slave_, bool may_enable)
407 bond_slave_lookup(bond, slave_)->may_enable = may_enable;
410 /* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
411 * allow tagged flows to be invalidated.
413 * The caller should check bond_should_send_learning_packets() afterward. */
415 bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
417 struct bond_slave *slave;
418 bool is_tcp_hash = bond_is_tcp_hash(bond);
420 bond->lacp_negotiated = lacp_negotiated;
423 netdev_monitor_flush(bond->monitor);
426 /* Enable slaves based on link status and LACP feedback. */
427 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
428 bond_link_status_update(slave, tags);
430 if (!bond->active_slave || !bond->active_slave->enabled) {
431 bond_choose_active_slave(bond, tags);
434 /* Update fake bond interface stats. */
435 if (time_msec() >= bond->next_fake_iface_update) {
436 bond_update_fake_slave_stats(bond);
437 bond->next_fake_iface_update = time_msec() + 1000;
440 if (is_tcp_hash != bond_is_tcp_hash(bond)) {
441 bond->bond_revalidate = true;
444 if (bond->bond_revalidate) {
445 bond->bond_revalidate = false;
447 bond_entry_reset(bond);
448 if (bond->balance != BM_STABLE) {
449 struct bond_slave *slave;
451 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
452 tag_set_add(tags, slave->tag);
455 tag_set_add(tags, bond->stb_tag);
457 tag_set_add(tags, bond->no_slaves_tag);
460 /* Invalidate any tags required by */
461 tag_set_union(tags, &bond->unixctl_tags);
462 tag_set_init(&bond->unixctl_tags);
465 /* Causes poll_block() to wake up when 'bond' needs something to be done. */
467 bond_wait(struct bond *bond)
469 struct bond_slave *slave;
471 netdev_monitor_poll_wait(bond->monitor);
473 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
474 if (slave->delay_expires != LLONG_MAX) {
475 poll_timer_wait_until(slave->delay_expires);
479 if (bond->next_fake_iface_update != LLONG_MAX) {
480 poll_timer_wait_until(bond->next_fake_iface_update);
483 /* Ensure that any saved tags get revalidated right away. */
484 if (!tag_set_is_empty(&bond->unixctl_tags)) {
485 poll_immediate_wake();
488 /* We don't wait for bond->next_rebalance because rebalancing can only run
489 * at a flow account checkpoint. ofproto does checkpointing on its own
490 * schedule and bond_rebalance() gets called afterward, so we'd just be
491 * waking up for no purpose. */
494 /* MAC learning table interaction. */
497 may_send_learning_packets(const struct bond *bond)
499 return !bond->lacp_negotiated && bond->balance != BM_AB;
502 /* Returns true if 'bond' needs the client to send out packets to assist with
503 * MAC learning on 'bond'. If this function returns true, then the client
504 * should iterate through its MAC learning table for the bridge on which 'bond'
505 * is located. For each MAC that has been learned on a port other than 'bond',
506 * it should call bond_send_learning_packet().
508 * This function will only return true if 'bond' is in SLB mode and LACP is not
509 * negotiated. Otherwise sending learning packets isn't necessary.
511 * Calling this function resets the state that it checks. */
513 bond_should_send_learning_packets(struct bond *bond)
515 bool send = bond->send_learning_packets && may_send_learning_packets(bond);
516 bond->send_learning_packets = false;
520 /* Sends a gratuitous learning packet on 'bond' from 'eth_src' on 'vlan'.
522 * See bond_should_send_learning_packets() for description of usage. */
524 bond_send_learning_packet(struct bond *bond,
525 const uint8_t eth_src[ETH_ADDR_LEN],
528 struct bond_slave *slave;
529 struct ofpbuf packet;
533 assert(may_send_learning_packets(bond));
534 if (!bond->active_slave) {
535 /* Nowhere to send the learning packet. */
539 memset(&flow, 0, sizeof flow);
540 memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
541 slave = choose_output_slave(bond, &flow, vlan);
543 ofpbuf_init(&packet, 0);
544 compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
547 eth_set_vlan_tci(&packet, htons(vlan));
549 error = netdev_send(slave->netdev, &packet);
550 ofpbuf_uninit(&packet);
555 /* Checks whether a packet that arrived on 'slave_' within 'bond', with an
556 * Ethernet destination address of 'eth_dst', should be admitted.
558 * The return value is one of the following:
560 * - BV_ACCEPT: Admit the packet.
562 * - BV_DROP: Drop the packet.
564 * - BV_DROP_IF_MOVED: Consult the MAC learning table for the packet's
565 * Ethernet source address and VLAN. If there is none, or if the packet
566 * is on the learned port, then admit the packet. If a different port has
567 * been learned, however, drop the packet (and do not use it for MAC
571 bond_check_admissibility(struct bond *bond, const void *slave_,
572 const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
574 /* Admit all packets if LACP has been negotiated, because that means that
575 * the remote switch is aware of the bond and will "do the right thing". */
576 if (bond->lacp_negotiated) {
580 /* Drop all multicast packets on inactive slaves. */
581 if (eth_addr_is_multicast(eth_dst)) {
582 *tags |= bond_get_active_slave_tag(bond);
583 if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
588 /* Drop all packets for which we have learned a different input port,
589 * because we probably sent the packet on one slave and got it back on the
590 * other. Gratuitous ARP packets are an exception to this rule: the host
591 * has moved to another switch. The exception to the exception is if we
592 * locked the learning table to avoid reflections on bond slaves. */
593 return BV_DROP_IF_MOVED;
596 /* Returns the slave (registered on 'bond' by bond_slave_register()) to which
597 * a packet with the given 'flow' and 'vlan' should be forwarded. Returns
598 * NULL if the packet should be dropped because no slaves are enabled.
600 * 'vlan' is not necessarily the same as 'flow->vlan_tci'. First, 'vlan'
601 * should be a VID only (i.e. excluding the PCP bits). Second,
602 * 'flow->vlan_tci' is the VLAN TCI that appeared on the packet (so it will be
603 * nonzero only for trunk ports), whereas 'vlan' is the logical VLAN that the
604 * packet belongs to (so for an access port it will be the access port's VLAN).
606 * Adds a tag to '*tags' that associates the flow with the returned slave.
609 bond_choose_output_slave(struct bond *bond, const struct flow *flow,
610 uint16_t vlan, tag_type *tags)
612 struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
614 *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
617 *tags |= bond->no_slaves_tag;
625 bond_is_balanced(const struct bond *bond)
627 return bond->balance == BM_SLB || bond->balance == BM_TCP;
630 /* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
632 bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
636 if (bond_is_balanced(bond)) {
637 lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
641 static struct bond_slave *
642 bond_slave_from_bal_node(struct list *bal)
644 return CONTAINER_OF(bal, struct bond_slave, bal_node);
648 log_bals(struct bond *bond, const struct list *bals)
650 if (VLOG_IS_DBG_ENABLED()) {
651 struct ds ds = DS_EMPTY_INITIALIZER;
652 const struct bond_slave *slave;
654 LIST_FOR_EACH (slave, bal_node, bals) {
656 ds_put_char(&ds, ',');
658 ds_put_format(&ds, " %s %"PRIu64"kB",
659 slave->name, slave->tx_bytes / 1024);
661 if (!slave->enabled) {
662 ds_put_cstr(&ds, " (disabled)");
664 if (!list_is_empty(&slave->entries)) {
665 struct bond_entry *e;
667 ds_put_cstr(&ds, " (");
668 LIST_FOR_EACH (e, list_node, &slave->entries) {
669 if (&e->list_node != list_front(&slave->entries)) {
670 ds_put_cstr(&ds, " + ");
672 ds_put_format(&ds, "h%td: %"PRIu64"kB",
673 e - bond->hash, e->tx_bytes / 1024);
675 ds_put_cstr(&ds, ")");
678 VLOG_DBG("bond %s:%s", bond->name, ds_cstr(&ds));
683 /* Shifts 'hash' from its current slave to 'to'. */
685 bond_shift_load(struct bond_entry *hash, struct bond_slave *to,
688 struct bond_slave *from = hash->slave;
689 struct bond *bond = from->bond;
690 uint64_t delta = hash->tx_bytes;
692 VLOG_INFO("bond %s: shift %"PRIu64"kB of load (with hash %td) "
693 "from %s to %s (now carrying %"PRIu64"kB and "
694 "%"PRIu64"kB load, respectively)",
695 bond->name, delta / 1024, hash - bond->hash,
696 from->name, to->name,
697 (from->tx_bytes - delta) / 1024,
698 (to->tx_bytes + delta) / 1024);
700 /* Shift load away from 'from' to 'to'. */
701 from->tx_bytes -= delta;
702 to->tx_bytes += delta;
704 /* Arrange for flows to be revalidated. */
705 tag_set_add(set, hash->tag);
707 hash->tag = tag_create_random();
710 /* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
711 * given that doing so must decrease the ratio of the load on the two slaves by
712 * at least 0.1. Returns NULL if there is no appropriate entry.
714 * The list of entries isn't sorted. I don't know of a reason to prefer to
715 * shift away small hashes or large hashes. */
716 static struct bond_entry *
717 choose_entry_to_migrate(const struct bond_slave *from, uint64_t to_tx_bytes)
719 struct bond_entry *e;
721 if (list_is_short(&from->entries)) {
722 /* 'from' carries no more than one MAC hash, so shifting load away from
723 * it would be pointless. */
727 LIST_FOR_EACH (e, list_node, &from->entries) {
728 double old_ratio, new_ratio;
731 if (to_tx_bytes == 0) {
732 /* Nothing on the new slave, move it. */
737 old_ratio = (double)from->tx_bytes / to_tx_bytes;
738 new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
739 if (old_ratio - new_ratio > 0.1) {
740 /* Would decrease the ratio, move it. */
748 /* Inserts 'slave' into 'bals' so that descending order of 'tx_bytes' is
751 insert_bal(struct list *bals, struct bond_slave *slave)
753 struct bond_slave *pos;
755 LIST_FOR_EACH (pos, bal_node, bals) {
756 if (slave->tx_bytes > pos->tx_bytes) {
760 list_insert(&pos->bal_node, &slave->bal_node);
763 /* Removes 'slave' from its current list and then inserts it into 'bals' so
764 * that descending order of 'tx_bytes' is maintained. */
766 reinsert_bal(struct list *bals, struct bond_slave *slave)
768 list_remove(&slave->bal_node);
769 insert_bal(bals, slave);
772 /* If 'bond' needs rebalancing, does so.
774 * The caller should have called bond_account() for each active flow, to ensure
775 * that flow data is consistently accounted at this point. */
777 bond_rebalance(struct bond *bond, struct tag_set *tags)
779 struct bond_slave *slave;
780 struct bond_entry *e;
783 if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
786 bond->next_rebalance = time_msec() + bond->rebalance_interval;
788 /* Add each bond_entry to its slave's 'entries' list.
789 * Compute each slave's tx_bytes as the sum of its entries' tx_bytes. */
790 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
792 list_init(&slave->entries);
794 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
795 if (e->slave && e->tx_bytes) {
796 e->slave->tx_bytes += e->tx_bytes;
797 list_push_back(&e->slave->entries, &e->list_node);
801 /* Add enabled slaves to 'bals' in descending order of tx_bytes.
803 * XXX This is O(n**2) in the number of slaves but it could be O(n lg n)
804 * with a proper list sort algorithm. */
806 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
807 if (slave->enabled) {
808 insert_bal(&bals, slave);
811 log_bals(bond, &bals);
813 /* Shift load from the most-loaded slaves to the least-loaded slaves. */
814 while (!list_is_short(&bals)) {
815 struct bond_slave *from = bond_slave_from_bal_node(list_front(&bals));
816 struct bond_slave *to = bond_slave_from_bal_node(list_back(&bals));
819 overload = from->tx_bytes - to->tx_bytes;
820 if (overload < to->tx_bytes >> 5 || overload < 100000) {
821 /* The extra load on 'from' (and all less-loaded slaves), compared
822 * to that of 'to' (the least-loaded slave), is less than ~3%, or
823 * it is less than ~1Mbps. No point in rebalancing. */
827 /* 'from' is carrying significantly more load than 'to', and that load
828 * is split across at least two different hashes. */
829 e = choose_entry_to_migrate(from, to->tx_bytes);
831 bond_shift_load(e, to, tags);
833 /* Delete element from from->entries.
835 * We don't add the element to to->hashes. That would only allow
836 * 'e' to be migrated to another slave in this rebalancing run, and
837 * there is no point in doing that. */
838 list_remove(&e->list_node);
840 /* Re-sort 'bals'. */
841 reinsert_bal(&bals, from);
842 reinsert_bal(&bals, to);
844 /* Can't usefully migrate anything away from 'from'.
845 * Don't reconsider it. */
846 list_remove(&from->bal_node);
850 /* Implement exponentially weighted moving average. A weight of 1/2 causes
851 * historical data to decay to <1% in 7 rebalancing runs. 1,000,000 bytes
852 * take 20 rebalancing runs to decay to 0 and get deleted entirely. */
853 for (e = &bond->hash[0]; e <= &bond->hash[BOND_MASK]; e++) {
861 /* Bonding unixctl user interface functions. */
864 bond_find(const char *name)
868 HMAP_FOR_EACH_WITH_HASH (bond, hmap_node, hash_string(name, 0),
870 if (!strcmp(bond->name, name)) {
877 static struct bond_slave *
878 bond_lookup_slave(struct bond *bond, const char *slave_name)
880 struct bond_slave *slave;
882 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
883 if (!strcmp(slave->name, slave_name)) {
891 bond_unixctl_list(struct unixctl_conn *conn,
892 const char *args OVS_UNUSED, void *aux OVS_UNUSED)
894 struct ds ds = DS_EMPTY_INITIALIZER;
895 const struct bond *bond;
897 ds_put_cstr(&ds, "bond\ttype\tslaves\n");
899 HMAP_FOR_EACH (bond, hmap_node, &all_bonds) {
900 const struct bond_slave *slave;
903 ds_put_format(&ds, "%s\t%s\t",
904 bond->name, bond_mode_to_string(bond->balance));
907 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
909 ds_put_cstr(&ds, ", ");
911 ds_put_cstr(&ds, slave->name);
913 ds_put_char(&ds, '\n');
915 unixctl_command_reply(conn, 200, ds_cstr(&ds));
920 bond_unixctl_show(struct unixctl_conn *conn,
921 const char *args, void *aux OVS_UNUSED)
923 struct ds ds = DS_EMPTY_INITIALIZER;
924 const struct bond_slave *slave;
925 const struct bond *bond;
927 bond = bond_find(args);
929 unixctl_command_reply(conn, 501, "no such bond");
933 ds_put_format(&ds, "bond_mode: %s\n",
934 bond_mode_to_string(bond->balance));
936 if (bond->balance != BM_AB) {
937 ds_put_format(&ds, "bond-hash-algorithm: %s\n",
938 bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
941 ds_put_format(&ds, "bond-hash-basis: %"PRIu32"\n", bond->basis);
943 ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
944 ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
946 if (bond_is_balanced(bond)) {
947 ds_put_format(&ds, "next rebalance: %lld ms\n",
948 bond->next_rebalance - time_msec());
951 ds_put_format(&ds, "lacp_negotiated: %s\n",
952 bond->lacp_negotiated ? "true" : "false");
954 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
955 struct bond_entry *be;
958 ds_put_format(&ds, "\nslave %s: %s\n",
959 slave->name, slave->enabled ? "enabled" : "disabled");
960 if (slave == bond->active_slave) {
961 ds_put_cstr(&ds, "\tactive slave\n");
963 if (slave->delay_expires != LLONG_MAX) {
964 ds_put_format(&ds, "\t%s expires in %lld ms\n",
965 slave->enabled ? "downdelay" : "updelay",
966 slave->delay_expires - time_msec());
969 ds_put_format(&ds, "\tmay_enable: %s\n",
970 slave->may_enable ? "true" : "false");
972 if (!bond_is_balanced(bond)) {
977 for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
978 int hash = be - bond->hash;
980 if (be->slave != slave) {
984 ds_put_format(&ds, "\thash %d: %"PRIu64" kB load\n",
985 hash, be->tx_bytes / 1024);
987 if (bond->balance != BM_SLB) {
991 /* XXX How can we list the MACs assigned to hashes? */
994 unixctl_command_reply(conn, 200, ds_cstr(&ds));
999 bond_unixctl_migrate(struct unixctl_conn *conn, const char *args_,
1000 void *aux OVS_UNUSED)
1002 char *args = (char *) args_;
1003 char *save_ptr = NULL;
1004 char *bond_s, *hash_s, *slave_s;
1006 struct bond_slave *slave;
1007 struct bond_entry *entry;
1010 bond_s = strtok_r(args, " ", &save_ptr);
1011 hash_s = strtok_r(NULL, " ", &save_ptr);
1012 slave_s = strtok_r(NULL, " ", &save_ptr);
1014 unixctl_command_reply(conn, 501,
1015 "usage: bond/migrate BOND HASH SLAVE");
1019 bond = bond_find(bond_s);
1021 unixctl_command_reply(conn, 501, "no such bond");
1025 if (bond->balance != BM_SLB) {
1026 unixctl_command_reply(conn, 501, "not an SLB bond");
1030 if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
1031 hash = atoi(hash_s) & BOND_MASK;
1033 unixctl_command_reply(conn, 501, "bad hash");
1037 slave = bond_lookup_slave(bond, slave_s);
1039 unixctl_command_reply(conn, 501, "no such slave");
1043 if (!slave->enabled) {
1044 unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
1048 entry = &bond->hash[hash];
1049 tag_set_add(&bond->unixctl_tags, entry->tag);
1050 entry->slave = slave;
1051 entry->tag = tag_create_random();
1052 unixctl_command_reply(conn, 200, "migrated");
1056 bond_unixctl_set_active_slave(struct unixctl_conn *conn, const char *args_,
1057 void *aux OVS_UNUSED)
1059 char *args = (char *) args_;
1060 char *save_ptr = NULL;
1061 char *bond_s, *slave_s;
1063 struct bond_slave *slave;
1065 bond_s = strtok_r(args, " ", &save_ptr);
1066 slave_s = strtok_r(NULL, " ", &save_ptr);
1068 unixctl_command_reply(conn, 501,
1069 "usage: bond/set-active-slave BOND SLAVE");
1073 bond = bond_find(bond_s);
1075 unixctl_command_reply(conn, 501, "no such bond");
1079 slave = bond_lookup_slave(bond, slave_s);
1081 unixctl_command_reply(conn, 501, "no such slave");
1085 if (!slave->enabled) {
1086 unixctl_command_reply(conn, 501, "cannot make disabled slave active");
1090 if (bond->active_slave != slave) {
1091 tag_set_add(&bond->unixctl_tags, bond_get_active_slave_tag(bond));
1092 bond->active_slave = slave;
1093 bond->active_slave->tag = tag_create_random();
1094 VLOG_INFO("bond %s: active interface is now %s",
1095 bond->name, slave->name);
1096 bond->send_learning_packets = true;
1097 unixctl_command_reply(conn, 200, "done");
1099 unixctl_command_reply(conn, 200, "no change");
1104 enable_slave(struct unixctl_conn *conn, const char *args_, bool enable)
1106 char *args = (char *) args_;
1107 char *save_ptr = NULL;
1108 char *bond_s, *slave_s;
1110 struct bond_slave *slave;
1112 bond_s = strtok_r(args, " ", &save_ptr);
1113 slave_s = strtok_r(NULL, " ", &save_ptr);
1115 char *usage = xasprintf("usage: bond/%s-slave BOND SLAVE",
1116 enable ? "enable" : "disable");
1117 unixctl_command_reply(conn, 501, usage);
1122 bond = bond_find(bond_s);
1124 unixctl_command_reply(conn, 501, "no such bond");
1128 slave = bond_lookup_slave(bond, slave_s);
1130 unixctl_command_reply(conn, 501, "no such slave");
1134 bond_enable_slave(slave, enable, &bond->unixctl_tags);
1135 unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
1139 bond_unixctl_enable_slave(struct unixctl_conn *conn, const char *args,
1140 void *aux OVS_UNUSED)
1142 enable_slave(conn, args, true);
1146 bond_unixctl_disable_slave(struct unixctl_conn *conn, const char *args,
1147 void *aux OVS_UNUSED)
1149 enable_slave(conn, args, false);
1153 bond_unixctl_hash(struct unixctl_conn *conn, const char *args_,
1154 void *aux OVS_UNUSED)
1156 char *args = (char *) args_;
1157 uint8_t mac[ETH_ADDR_LEN];
1162 char *mac_s, *vlan_s, *basis_s;
1163 char *save_ptr = NULL;
1165 mac_s = strtok_r(args, " ", &save_ptr);
1166 vlan_s = strtok_r(NULL, " ", &save_ptr);
1167 basis_s = strtok_r(NULL, " ", &save_ptr);
1170 if (sscanf(vlan_s, "%u", &vlan) != 1) {
1171 unixctl_command_reply(conn, 501, "invalid vlan");
1175 vlan = OFP_VLAN_NONE;
1179 if (sscanf(basis_s, "%"PRIu32, &basis) != 1) {
1180 unixctl_command_reply(conn, 501, "invalid basis");
1187 if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
1188 == ETH_ADDR_SCAN_COUNT) {
1189 hash = bond_hash_src(mac, vlan, basis) & BOND_MASK;
1191 hash_cstr = xasprintf("%u", hash);
1192 unixctl_command_reply(conn, 200, hash_cstr);
1195 unixctl_command_reply(conn, 501, "invalid mac");
1202 unixctl_command_register("bond/list", bond_unixctl_list, NULL);
1203 unixctl_command_register("bond/show", bond_unixctl_show, NULL);
1204 unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
1205 unixctl_command_register("bond/set-active-slave",
1206 bond_unixctl_set_active_slave, NULL);
1207 unixctl_command_register("bond/enable-slave", bond_unixctl_enable_slave,
1209 unixctl_command_register("bond/disable-slave", bond_unixctl_disable_slave,
1211 unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
1215 bond_entry_reset(struct bond *bond)
1217 if (bond->balance != BM_AB) {
1218 size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
1221 bond->hash = xmalloc(hash_len);
1223 memset(bond->hash, 0, hash_len);
1225 bond->next_rebalance = time_msec() + bond->rebalance_interval;
1232 static struct bond_slave *
1233 bond_slave_lookup(struct bond *bond, const void *slave_)
1235 struct bond_slave *slave;
1237 HMAP_FOR_EACH_IN_BUCKET (slave, hmap_node, hash_pointer(slave_, 0),
1239 if (slave->aux == slave_) {
1248 bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
1250 struct bond *bond = slave->bond;
1251 slave->delay_expires = LLONG_MAX;
1252 if (enable != slave->enabled) {
1253 slave->enabled = enable;
1254 if (!slave->enabled) {
1255 VLOG_WARN("interface %s: disabled", slave->name);
1257 tag_set_add(tags, slave->tag);
1260 VLOG_WARN("interface %s: enabled", slave->name);
1261 slave->tag = tag_create_random();
1264 if (bond->balance == BM_STABLE) {
1265 bond->bond_revalidate = true;
1271 bond_link_status_update(struct bond_slave *slave, struct tag_set *tags)
1273 struct bond *bond = slave->bond;
1276 up = netdev_get_carrier(slave->netdev) && slave->may_enable;
1277 if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
1278 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1279 VLOG_INFO_RL(&rl, "interface %s: link state %s",
1280 slave->name, up ? "up" : "down");
1281 if (up == slave->enabled) {
1282 slave->delay_expires = LLONG_MAX;
1283 VLOG_INFO_RL(&rl, "interface %s: will not be %s",
1284 slave->name, up ? "disabled" : "enabled");
1286 int delay = (bond->lacp_negotiated ? 0
1287 : up ? bond->updelay : bond->downdelay);
1288 slave->delay_expires = time_msec() + delay;
1290 VLOG_INFO_RL(&rl, "interface %s: will be %s if it stays %s "
1293 up ? "enabled" : "disabled",
1300 if (time_msec() >= slave->delay_expires) {
1301 bond_enable_slave(slave, up, tags);
1306 bond_is_tcp_hash(const struct bond *bond)
1308 return (bond->balance == BM_TCP && bond->lacp_negotiated)
1309 || bond->balance == BM_STABLE;
1313 bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan, uint32_t basis)
1315 return hash_3words(hash_bytes(mac, ETH_ADDR_LEN, 0), vlan, basis);
1319 bond_hash_tcp(const struct flow *flow, uint16_t vlan, uint32_t basis)
1321 struct flow hash_flow = *flow;
1322 hash_flow.vlan_tci = htons(vlan);
1324 /* The symmetric quality of this hash function is not required, but
1325 * flow_hash_symmetric_l4 already exists, and is sufficient for our
1326 * purposes, so we use it out of convenience. */
1327 return flow_hash_symmetric_l4(&hash_flow, basis);
1331 bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
1333 assert(bond->balance != BM_AB);
1335 return (bond_is_tcp_hash(bond)
1336 ? bond_hash_tcp(flow, vlan, bond->basis)
1337 : bond_hash_src(flow->dl_src, vlan, bond->basis));
1340 static struct bond_entry *
1341 lookup_bond_entry(const struct bond *bond, const struct flow *flow,
1344 return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
1347 /* This function uses Highest Random Weight hashing to choose an output slave.
1348 * This approach only reassigns a minimal number of flows when slaves are
1349 * enabled or disabled. Unfortunately, it has O(n) performance against the
1350 * number of slaves. There exist algorithms which are O(1), but have slightly
1351 * more complex implementations and require the use of memory. This may need
1352 * to be reimplemented if it becomes a performance bottleneck. */
1353 static struct bond_slave *
1354 choose_stb_slave(const struct bond *bond, const struct flow *flow,
1357 struct bond_slave *best, *slave;
1358 uint32_t best_hash, flow_hash;
1362 flow_hash = bond_hash(bond, flow, vlan);
1363 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1364 if (slave->enabled) {
1367 hash = hash_2words(flow_hash, slave->stb_id);
1368 if (!best || hash > best_hash) {
1378 static struct bond_slave *
1379 choose_output_slave(const struct bond *bond, const struct flow *flow,
1382 struct bond_entry *e;
1384 switch (bond->balance) {
1386 return bond->active_slave;
1389 return choose_stb_slave(bond, flow, vlan);
1392 e = lookup_bond_entry(bond, flow, vlan);
1393 if (!e->slave || !e->slave->enabled) {
1394 e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
1395 struct bond_slave, hmap_node);
1396 if (!e->slave->enabled) {
1397 e->slave = bond->active_slave;
1399 e->tag = tag_create_random();
1408 static struct bond_slave *
1409 bond_choose_slave(const struct bond *bond)
1411 struct bond_slave *slave, *best;
1413 /* Find an enabled slave. */
1414 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1415 if (slave->enabled) {
1420 /* All interfaces are disabled. Find an interface that will be enabled
1421 * after its updelay expires. */
1423 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1424 if (slave->delay_expires != LLONG_MAX
1425 && slave->may_enable
1426 && (!best || slave->delay_expires < best->delay_expires)) {
1434 bond_choose_active_slave(struct bond *bond, struct tag_set *tags)
1436 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
1437 struct bond_slave *old_active_slave = bond->active_slave;
1439 bond->active_slave = bond_choose_slave(bond);
1440 if (bond->active_slave) {
1441 if (bond->active_slave->enabled) {
1442 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s",
1443 bond->name, bond->active_slave->name);
1445 VLOG_INFO_RL(&rl, "bond %s: active interface is now %s, skipping "
1446 "remaining %lld ms updelay (since no interface was "
1447 "enabled)", bond->name, bond->active_slave->name,
1448 bond->active_slave->delay_expires - time_msec());
1449 bond_enable_slave(bond->active_slave, true, tags);
1452 if (!old_active_slave) {
1453 tag_set_add(tags, bond->no_slaves_tag);
1456 bond->send_learning_packets = true;
1457 } else if (old_active_slave) {
1458 VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
1462 /* Returns the tag for 'bond''s active slave, or 'bond''s no_slaves_tag if
1463 * there is no active slave. */
1465 bond_get_active_slave_tag(const struct bond *bond)
1467 return (bond->active_slave
1468 ? bond->active_slave->tag
1469 : bond->no_slaves_tag);
1472 /* Attempts to make the sum of the bond slaves' statistics appear on the fake
1473 * bond interface. */
1475 bond_update_fake_slave_stats(struct bond *bond)
1477 struct netdev_stats bond_stats;
1478 struct bond_slave *slave;
1479 struct netdev *bond_dev;
1481 memset(&bond_stats, 0, sizeof bond_stats);
1483 HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
1484 struct netdev_stats slave_stats;
1486 if (!netdev_get_stats(slave->netdev, &slave_stats)) {
1487 /* XXX: We swap the stats here because they are swapped back when
1488 * reported by the internal device. The reason for this is
1489 * internal devices normally represent packets going into the
1490 * system but when used as fake bond device they represent packets
1491 * leaving the system. We really should do this in the internal
1492 * device itself because changing it here reverses the counts from
1493 * the perspective of the switch. However, the internal device
1494 * doesn't know what type of device it represents so we have to do
1495 * it here for now. */
1496 bond_stats.tx_packets += slave_stats.rx_packets;
1497 bond_stats.tx_bytes += slave_stats.rx_bytes;
1498 bond_stats.rx_packets += slave_stats.tx_packets;
1499 bond_stats.rx_bytes += slave_stats.tx_bytes;
1503 if (!netdev_open_default(bond->name, &bond_dev)) {
1504 netdev_set_stats(bond_dev, &bond_stats);
1505 netdev_close(bond_dev);