/*
- * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
+ * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 Nicira, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
#include <limits.h>
#include <stdint.h>
#include <stdlib.h>
+#include <math.h>
#include "coverage.h"
#include "dynamic-string.h"
#include "flow.h"
#include "hmap.h"
+#include "lacp.h"
#include "list.h"
#include "netdev.h"
#include "odp-util.h"
struct bond_entry {
struct bond_slave *slave; /* Assigned slave, NULL if unassigned. */
uint64_t tx_bytes; /* Count of bytes recently transmitted. */
- tag_type tag; /* Tag for entry<->slave association. */
+ tag_type tag; /* Tag for entry<->facet association. */
struct list list_node; /* In bond_slave's 'entries' list. */
};
struct list bal_node; /* In bond_rebalance()'s 'bals' list. */
struct list entries; /* 'struct bond_entry's assigned here. */
uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
-
- /* BM_STABLE specific bonding info. */
- uint32_t stb_id; /* ID used for 'stb_slaves' ordering. */
};
/* A bond, that is, a set of network devices grouped to improve performance or
struct bond_slave *active_slave;
tag_type no_slaves_tag; /* Tag for flows when all slaves disabled. */
int updelay, downdelay; /* Delay before slave goes up/down, in ms. */
- bool lacp_negotiated; /* LACP negotiations were successful. */
+ enum lacp_status lacp_status; /* Status of LACP negotiations. */
bool bond_revalidate; /* True if flows need revalidation. */
uint32_t basis; /* Basis for flow hash function. */
long long int next_rebalance; /* Next rebalancing time. */
bool send_learning_packets;
- /* BM_STABLE specific bonding info. */
- tag_type stb_tag; /* Tag associated with this bond. */
-
/* Legacy compatibility. */
long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
* where we can't otherwise provide revalidation feedback to the client.
* That's only unixctl commands now; I hope no other cases will arise. */
struct tag_set unixctl_tags;
+
+ int ref_cnt;
};
static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
struct tag_set *);
static void bond_link_status_update(struct bond_slave *, struct tag_set *);
static void bond_choose_active_slave(struct bond *, struct tag_set *);
-static bool bond_is_tcp_hash(const struct bond *);
static unsigned int bond_hash_src(const uint8_t mac[ETH_ADDR_LEN],
uint16_t vlan, uint32_t basis);
static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan,
static tag_type bond_get_active_slave_tag(const struct bond *);
static struct bond_slave *choose_output_slave(const struct bond *,
const struct flow *,
- uint16_t vlan);
+ struct flow_wildcards *,
+ uint16_t vlan, tag_type *tags);
static void bond_update_fake_slave_stats(struct bond *);
/* Attempts to parse 's' as the name of a bond balancing mode. If successful,
*balance = BM_TCP;
} else if (!strcmp(s, bond_mode_to_string(BM_SLB))) {
*balance = BM_SLB;
- } else if (!strcmp(s, bond_mode_to_string(BM_STABLE))) {
- *balance = BM_STABLE;
} else if (!strcmp(s, bond_mode_to_string(BM_AB))) {
*balance = BM_AB;
} else {
return "balance-tcp";
case BM_SLB:
return "balance-slb";
- case BM_STABLE:
- return "stable";
case BM_AB:
return "active-backup";
}
bond = xzalloc(sizeof *bond);
hmap_init(&bond->slaves);
bond->no_slaves_tag = tag_create_random();
- bond->stb_tag = tag_create_random();
bond->next_fake_iface_update = LLONG_MAX;
+ bond->ref_cnt = 1;
bond_reconfigure(bond, s);
return bond;
}
+struct bond *
+bond_ref(const struct bond *bond_)
+{
+ struct bond *bond = CONST_CAST(struct bond *, bond_);
+
+ ovs_assert(bond->ref_cnt > 0);
+ bond->ref_cnt++;
+ return bond;
+}
+
/* Frees 'bond'. */
void
-bond_destroy(struct bond *bond)
+bond_unref(struct bond *bond)
{
struct bond_slave *slave, *next_slave;
return;
}
+ ovs_assert(bond->ref_cnt > 0);
+ if (--bond->ref_cnt) {
+ return;
+ }
+
hmap_remove(&all_bonds, &bond->hmap_node);
HMAP_FOR_EACH_SAFE (slave, next_slave, hmap_node, &bond->slaves) {
bond->updelay = s->up_delay;
bond->downdelay = s->down_delay;
- bond->rebalance_interval = s->rebalance_interval;
+
+ if (bond->rebalance_interval != s->rebalance_interval) {
+ bond->rebalance_interval = s->rebalance_interval;
+ revalidate = true;
+ }
if (bond->balance != s->balance) {
bond->balance = s->balance;
* bond. If 'slave_' already exists within 'bond' then this function
* reconfigures the existing slave.
*
- * 'stb_id' is used in BM_STABLE bonds to guarantee consistent slave choices
- * across restarts and distributed vswitch instances. It should be unique per
- * slave, and preferably consistent across restarts and reconfigurations.
- *
* 'netdev' must be the network device that 'slave_' represents. It is owned
* by the client, so the client must not close it before either unregistering
* 'slave_' or destroying 'bond'.
*/
void
-bond_slave_register(struct bond *bond, void *slave_, uint32_t stb_id,
- struct netdev *netdev)
+bond_slave_register(struct bond *bond, void *slave_, struct netdev *netdev)
{
struct bond_slave *slave = bond_slave_lookup(bond, slave_);
bond_enable_slave(slave, netdev_get_carrier(netdev), NULL);
}
- if (slave->stb_id != stb_id) {
- slave->stb_id = stb_id;
- bond->bond_revalidate = true;
- }
-
bond_slave_set_netdev__(slave, netdev);
free(slave->name);
*
* The caller should check bond_should_send_learning_packets() afterward. */
void
-bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
+bond_run(struct bond *bond, struct tag_set *tags, enum lacp_status lacp_status)
{
struct bond_slave *slave;
- bool is_tcp_hash = bond_is_tcp_hash(bond);
- bond->lacp_negotiated = lacp_negotiated;
+ if (bond->lacp_status != lacp_status) {
+ bond->lacp_status = lacp_status;
+ bond->bond_revalidate = true;
+ }
/* Enable slaves based on link status and LACP feedback. */
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
bond->next_fake_iface_update = time_msec() + 1000;
}
- if (is_tcp_hash != bond_is_tcp_hash(bond)) {
- bond->bond_revalidate = true;
- }
-
if (bond->bond_revalidate) {
- bond->bond_revalidate = false;
+ struct bond_slave *slave;
+ bond->bond_revalidate = false;
bond_entry_reset(bond);
- if (bond->balance != BM_STABLE) {
- struct bond_slave *slave;
-
- HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
- tag_set_add(tags, slave->tag);
- }
- } else {
- tag_set_add(tags, bond->stb_tag);
+ HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
+ tag_set_add(tags, slave->tag);
}
tag_set_add(tags, bond->no_slaves_tag);
}
static bool
may_send_learning_packets(const struct bond *bond)
{
- return !bond->lacp_negotiated && bond->balance != BM_AB &&
- bond->active_slave;
+ return bond->lacp_status == LACP_DISABLED
+ && (bond->balance == BM_SLB || bond->balance == BM_AB)
+ && bond->active_slave;
}
/* Returns true if 'bond' needs the client to send out packets to assist with
* is located. For each MAC that has been learned on a port other than 'bond',
* it should call bond_compose_learning_packet().
*
- * This function will only return true if 'bond' is in SLB mode and LACP is not
- * negotiated. Otherwise sending learning packets isn't necessary.
+ * This function will only return true if 'bond' is in SLB or active-backup
+ * mode and LACP is not negotiated. Otherwise sending learning packets isn't
+ * necessary.
*
* Calling this function resets the state that it checks. */
bool
{
struct bond_slave *slave;
struct ofpbuf *packet;
+ tag_type tags = 0;
struct flow flow;
- assert(may_send_learning_packets(bond));
+ ovs_assert(may_send_learning_packets(bond));
memset(&flow, 0, sizeof flow);
memcpy(flow.dl_src, eth_src, ETH_ADDR_LEN);
- slave = choose_output_slave(bond, &flow, vlan);
+ slave = choose_output_slave(bond, &flow, NULL, vlan, &tags);
packet = ofpbuf_new(0);
- compose_benign_packet(packet, "Open vSwitch Bond Failover", 0xf177,
- eth_src);
+ compose_rarp(packet, eth_src);
if (vlan) {
eth_push_vlan(packet, htons(vlan));
}
{
struct bond_slave *slave = bond_slave_lookup(bond, slave_);
+ if (!slave) {
+ return BV_DROP;
+ }
+
/* LACP bonds have very loose admissibility restrictions because we can
* assume the remote switch is aware of the bond and will "do the right
* thing". However, as a precaution we drop packets on disabled slaves
* because no correctly implemented partner switch should be sending
- * packets to them. */
- if (bond->lacp_negotiated) {
- return slave->enabled ? BV_ACCEPT : BV_DROP;
+ * packets to them.
+ *
+ * If LACP is configured, but LACP negotiations have been unsuccessful, we
+ * drop all incoming traffic. */
+ switch (bond->lacp_status) {
+ case LACP_NEGOTIATED: return slave->enabled ? BV_ACCEPT : BV_DROP;
+ case LACP_CONFIGURED: return BV_DROP;
+ case LACP_DISABLED: break;
}
/* Drop all multicast packets on inactive slaves. */
if (eth_addr_is_multicast(eth_dst)) {
*tags |= bond_get_active_slave_tag(bond);
- if (bond->active_slave != bond_slave_lookup(bond, slave_)) {
+ if (bond->active_slave != slave) {
return BV_DROP;
}
}
return BV_ACCEPT;
case BM_TCP:
- /* TCP balancing has degraded to SLB (otherwise the
- * bond->lacp_negotiated check above would have processed this).
- *
- * Fall through. */
+ /* TCP balanced bonds require successful LACP negotiated. Based on the
+ * above check, LACP is off on this bond. Therfore, we drop all
+ * incoming traffic. */
+ return BV_DROP;
+
case BM_SLB:
/* Drop all packets for which we have learned a different input port,
* because we probably sent the packet on one slave and got it back on
* exception is if we locked the learning table to avoid reflections on
* bond slaves. */
return BV_DROP_IF_MOVED;
-
- case BM_STABLE:
- return BV_ACCEPT;
}
NOT_REACHED();
* packet belongs to (so for an access port it will be the access port's VLAN).
*
* Adds a tag to '*tags' that associates the flow with the returned slave.
+ *
+ * If 'wc' is non-NULL, bitwise-OR's 'wc' with the set of bits that were
+ * significant in the selection. At some point earlier, 'wc' should
+ * have been initialized (e.g., by flow_wildcards_init_catchall()).
*/
void *
bond_choose_output_slave(struct bond *bond, const struct flow *flow,
- uint16_t vlan, tag_type *tags)
+ struct flow_wildcards *wc, uint16_t vlan,
+ tag_type *tags)
{
- struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
+ struct bond_slave *slave = choose_output_slave(bond, flow, wc, vlan, tags);
if (slave) {
- *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
+ *tags |= slave->tag;
return slave->aux;
} else {
*tags |= bond->no_slaves_tag;
static bool
bond_is_balanced(const struct bond *bond)
{
- return bond->balance == BM_SLB || bond->balance == BM_TCP;
+ return bond->rebalance_interval
+ && (bond->balance == BM_SLB || bond->balance == BM_TCP);
}
/* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
hash->tag = tag_create_random();
}
-/* Pick and returns a bond_entry to migrate to 'to' (the least-loaded slave),
+/* Picks and returns a bond_entry to migrate from 'from' (the most heavily
+ * loaded bond slave) to a bond slave that has 'to_tx_bytes' bytes of load,
* given that doing so must decrease the ratio of the load on the two slaves by
* at least 0.1. Returns NULL if there is no appropriate entry.
*
delta = e->tx_bytes;
old_ratio = (double)from->tx_bytes / to_tx_bytes;
new_ratio = (double)(from->tx_bytes - delta) / (to_tx_bytes + delta);
- if (old_ratio - new_ratio > 0.1) {
- /* Would decrease the ratio, move it. */
+ if (old_ratio - new_ratio > 0.1
+ && fabs(new_ratio - 1.0) < fabs(old_ratio - 1.0)) {
+ /* We're aiming for an ideal ratio of 1, meaning both the 'from'
+ and 'to' slave have the same load. Therefore, we only move an
+ entry if it decreases the load on 'from', and brings us closer
+ to equal traffic load. */
return e;
}
}
break;
}
- /* 'from' is carrying significantly more load than 'to', and that load
- * is split across at least two different hashes. */
+ /* 'from' is carrying significantly more load than 'to'. Pick a hash
+ * to move from 'from' to 'to'. */
e = choose_entry_to_migrate(from, to->tx_bytes);
if (e) {
bond_shift_load(e, to, tags);
}
ds_put_char(&ds, '\n');
}
- unixctl_command_reply(conn, 200, ds_cstr(&ds));
+ unixctl_command_reply(conn, ds_cstr(&ds));
ds_destroy(&ds);
}
ds_put_format(ds, "bond_mode: %s\n",
bond_mode_to_string(bond->balance));
- if (bond->balance != BM_AB) {
- ds_put_format(ds, "bond-hash-algorithm: %s\n",
- bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
- }
-
ds_put_format(ds, "bond-hash-basis: %"PRIu32"\n", bond->basis);
ds_put_format(ds, "updelay: %d ms\n", bond->updelay);
bond->next_rebalance - time_msec());
}
- ds_put_format(ds, "lacp_negotiated: %s\n",
- bond->lacp_negotiated ? "true" : "false");
+ ds_put_cstr(ds, "lacp_status: ");
+ switch (bond->lacp_status) {
+ case LACP_NEGOTIATED:
+ ds_put_cstr(ds, "negotiated\n");
+ break;
+ case LACP_CONFIGURED:
+ ds_put_cstr(ds, "configured\n");
+ break;
+ case LACP_DISABLED:
+ ds_put_cstr(ds, "off\n");
+ break;
+ default:
+ ds_put_cstr(ds, "<unknown>\n");
+ break;
+ }
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
shash_add(&slave_shash, slave->name, slave);
const struct bond *bond = bond_find(argv[1]);
if (!bond) {
- unixctl_command_reply(conn, 501, "no such bond");
+ unixctl_command_reply_error(conn, "no such bond");
return;
}
bond_print_details(&ds, bond);
}
}
- unixctl_command_reply(conn, 200, ds_cstr(&ds));
+ unixctl_command_reply(conn, ds_cstr(&ds));
ds_destroy(&ds);
}
bond = bond_find(bond_s);
if (!bond) {
- unixctl_command_reply(conn, 501, "no such bond");
+ unixctl_command_reply_error(conn, "no such bond");
return;
}
if (bond->balance != BM_SLB) {
- unixctl_command_reply(conn, 501, "not an SLB bond");
+ unixctl_command_reply_error(conn, "not an SLB bond");
return;
}
if (strspn(hash_s, "0123456789") == strlen(hash_s)) {
hash = atoi(hash_s) & BOND_MASK;
} else {
- unixctl_command_reply(conn, 501, "bad hash");
+ unixctl_command_reply_error(conn, "bad hash");
return;
}
slave = bond_lookup_slave(bond, slave_s);
if (!slave) {
- unixctl_command_reply(conn, 501, "no such slave");
+ unixctl_command_reply_error(conn, "no such slave");
return;
}
if (!slave->enabled) {
- unixctl_command_reply(conn, 501, "cannot migrate to disabled slave");
+ unixctl_command_reply_error(conn, "cannot migrate to disabled slave");
return;
}
tag_set_add(&bond->unixctl_tags, entry->tag);
entry->slave = slave;
entry->tag = tag_create_random();
- unixctl_command_reply(conn, 200, "migrated");
+ unixctl_command_reply(conn, "migrated");
}
static void
bond = bond_find(bond_s);
if (!bond) {
- unixctl_command_reply(conn, 501, "no such bond");
+ unixctl_command_reply_error(conn, "no such bond");
return;
}
slave = bond_lookup_slave(bond, slave_s);
if (!slave) {
- unixctl_command_reply(conn, 501, "no such slave");
+ unixctl_command_reply_error(conn, "no such slave");
return;
}
if (!slave->enabled) {
- unixctl_command_reply(conn, 501, "cannot make disabled slave active");
+ unixctl_command_reply_error(conn, "cannot make disabled slave active");
return;
}
VLOG_INFO("bond %s: active interface is now %s",
bond->name, slave->name);
bond->send_learning_packets = true;
- unixctl_command_reply(conn, 200, "done");
+ unixctl_command_reply(conn, "done");
} else {
- unixctl_command_reply(conn, 200, "no change");
+ unixctl_command_reply(conn, "no change");
}
}
bond = bond_find(bond_s);
if (!bond) {
- unixctl_command_reply(conn, 501, "no such bond");
+ unixctl_command_reply_error(conn, "no such bond");
return;
}
slave = bond_lookup_slave(bond, slave_s);
if (!slave) {
- unixctl_command_reply(conn, 501, "no such slave");
+ unixctl_command_reply_error(conn, "no such slave");
return;
}
bond_enable_slave(slave, enable, &bond->unixctl_tags);
- unixctl_command_reply(conn, 501, enable ? "enabled" : "disabled");
+ unixctl_command_reply(conn, enable ? "enabled" : "disabled");
}
static void
if (vlan_s) {
if (sscanf(vlan_s, "%u", &vlan) != 1) {
- unixctl_command_reply(conn, 501, "invalid vlan");
+ unixctl_command_reply_error(conn, "invalid vlan");
return;
}
} else {
if (basis_s) {
if (sscanf(basis_s, "%"PRIu32, &basis) != 1) {
- unixctl_command_reply(conn, 501, "invalid basis");
+ unixctl_command_reply_error(conn, "invalid basis");
return;
}
} else {
hash = bond_hash_src(mac, vlan, basis) & BOND_MASK;
hash_cstr = xasprintf("%u", hash);
- unixctl_command_reply(conn, 200, hash_cstr);
+ unixctl_command_reply(conn, hash_cstr);
free(hash_cstr);
} else {
- unixctl_command_reply(conn, 501, "invalid mac");
+ unixctl_command_reply_error(conn, "invalid mac");
}
}
static void
bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags)
{
- struct bond *bond = slave->bond;
slave->delay_expires = LLONG_MAX;
if (enable != slave->enabled) {
slave->enabled = enable;
if (!slave->enabled) {
- VLOG_WARN("interface %s: disabled", slave->name);
+ VLOG_INFO("interface %s: disabled", slave->name);
if (tags) {
tag_set_add(tags, slave->tag);
}
} else {
- VLOG_WARN("interface %s: enabled", slave->name);
+ VLOG_INFO("interface %s: enabled", slave->name);
slave->tag = tag_create_random();
}
-
- if (bond->balance == BM_STABLE) {
- bond->bond_revalidate = true;
- }
}
}
VLOG_INFO_RL(&rl, "interface %s: will not be %s",
slave->name, up ? "disabled" : "enabled");
} else {
- int delay = (bond->lacp_negotiated ? 0
+ int delay = (bond->lacp_status != LACP_DISABLED ? 0
: up ? bond->updelay : bond->downdelay);
slave->delay_expires = time_msec() + delay;
if (delay) {
}
}
-static bool
-bond_is_tcp_hash(const struct bond *bond)
-{
- return (bond->balance == BM_TCP && bond->lacp_negotiated)
- || bond->balance == BM_STABLE;
-}
-
static unsigned int
bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan, uint32_t basis)
{
static unsigned int
bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
{
- assert(bond->balance != BM_AB);
+ ovs_assert(bond->balance == BM_TCP || bond->balance == BM_SLB);
- return (bond_is_tcp_hash(bond)
+ return (bond->balance == BM_TCP
? bond_hash_tcp(flow, vlan, bond->basis)
: bond_hash_src(flow->dl_src, vlan, bond->basis));
}
return &bond->hash[bond_hash(bond, flow, vlan) & BOND_MASK];
}
-/* This function uses Highest Random Weight hashing to choose an output slave.
- * This approach only reassigns a minimal number of flows when slaves are
- * enabled or disabled. Unfortunately, it has O(n) performance against the
- * number of slaves. There exist algorithms which are O(1), but have slightly
- * more complex implementations and require the use of memory. This may need
- * to be reimplemented if it becomes a performance bottleneck. */
-static struct bond_slave *
-choose_stb_slave(const struct bond *bond, const struct flow *flow,
- uint16_t vlan)
-{
- struct bond_slave *best, *slave;
- uint32_t best_hash, flow_hash;
-
- best = NULL;
- best_hash = 0;
- flow_hash = bond_hash(bond, flow, vlan);
- HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
- if (slave->enabled) {
- uint32_t hash;
-
- hash = hash_2words(flow_hash, slave->stb_id);
- if (!best || hash > best_hash) {
- best = slave;
- best_hash = hash;
- }
- }
- }
-
- return best;
-}
-
static struct bond_slave *
choose_output_slave(const struct bond *bond, const struct flow *flow,
- uint16_t vlan)
+ struct flow_wildcards *wc, uint16_t vlan, tag_type *tags)
{
struct bond_entry *e;
+ if (bond->lacp_status == LACP_CONFIGURED) {
+ /* LACP has been configured on this bond but negotiations were
+ * unsuccussful. Drop all traffic. */
+ return NULL;
+ }
+
switch (bond->balance) {
case BM_AB:
return bond->active_slave;
- case BM_STABLE:
- return choose_stb_slave(bond, flow, vlan);
- case BM_SLB:
case BM_TCP:
+ if (bond->lacp_status != LACP_NEGOTIATED) {
+ /* Must have LACP negotiations for TCP balanced bonds. */
+ return NULL;
+ }
+ if (wc) {
+ flow_mask_hash_fields(flow, wc, NX_HASH_FIELDS_SYMMETRIC_L4);
+ }
+ /* Fall Through. */
+ case BM_SLB:
+ if (wc) {
+ flow_mask_hash_fields(flow, wc, NX_HASH_FIELDS_ETH_SRC);
+ }
e = lookup_bond_entry(bond, flow, vlan);
if (!e->slave || !e->slave->enabled) {
e->slave = CONTAINER_OF(hmap_random_node(&bond->slaves),
}
e->tag = tag_create_random();
}
+ *tags |= e->tag;
return e->slave;
default:
bond->send_learning_packets = true;
} else if (old_active_slave) {
- VLOG_WARN_RL(&rl, "bond %s: all interfaces disabled", bond->name);
+ VLOG_INFO_RL(&rl, "bond %s: all interfaces disabled", bond->name);
}
}
git://git.onelab.eu / sliver-openvswitch.git / blobdiff