void *aux; /* Client-provided handle for this slave. */
struct netdev *netdev; /* Network device, owned by the client. */
+ unsigned int change_seq; /* Tracks changes in 'netdev'. */
char *name; /* Name (a copy of netdev_get_name(netdev)). */
/* Link status. */
long long delay_expires; /* Time after which 'enabled' may change. */
- bool up; /* Last link status read from netdev. */
bool enabled; /* May be chosen for flows? */
- bool lacp_may_enable; /* LACP considers this interface bondable. */
+ bool may_enable; /* Client considers this slave bondable. */
tag_type tag; /* Tag associated with this slave. */
/* Rebalancing info. Used only by bond_rebalance(). */
uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
/* BM_STABLE specific bonding info. */
- uint16_t stb_id; /* ID used for 'stb_slaves' ordering. */
- size_t stb_idx; /* Index in 'bond''s 'stb_slaves' array.
- Undefined value if participating in a
- BTM_STABLE bond or not enabled. */
+ uint32_t stb_id; /* ID used for 'stb_slaves' ordering. */
};
/* A bond, that is, a set of network devices grouped to improve performance or
int updelay, downdelay; /* Delay before slave goes up/down, in ms. */
bool lacp_negotiated; /* LACP negotiations were successful. */
bool bond_revalidate; /* True if flows need revalidation. */
+ uint32_t basis; /* Basis for flow hash function. */
/* SLB specific bonding info. */
struct bond_entry *hash; /* An array of (BOND_MASK + 1) elements. */
bool send_learning_packets;
/* BM_STABLE specific bonding info. */
- struct bond_slave **stb_slaves; /* Ordered list of enabled slaves. */
- size_t n_stb_slaves; /* Number of slaves in 'stb_slaves'. */
- size_t len_stb_slaves; /* Slaves allocated in 'stb_slaves'. */
- bool stb_need_sort; /* True if stb_slaves is not sorted. */
tag_type stb_tag; /* Tag associated with this bond. */
- /* Monitoring. */
- enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
- struct netdev_monitor *monitor; /* detect == BLSM_CARRIER only. */
- long long int miimon_interval; /* Miimon status refresh interval. */
- long long int miimon_next_update; /* Time of next miimon update. */
-
/* Legacy compatibility. */
long long int next_fake_iface_update; /* LLONG_MAX if disabled. */
static void bond_entry_reset(struct bond *);
static struct bond_slave *bond_slave_lookup(struct bond *, const void *slave_);
-static bool bond_is_link_up(struct bond *, struct netdev *);
static void bond_enable_slave(struct bond_slave *, bool enable,
struct tag_set *);
-static void bond_stb_sort(struct bond *);
-static void bond_stb_enable_slave(struct bond_slave *);
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);
-static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan);
+ uint16_t vlan, uint32_t basis);
+static unsigned int bond_hash_tcp(const struct flow *, uint16_t vlan,
+ uint32_t basis);
static struct bond_entry *lookup_bond_entry(const struct bond *,
const struct flow *,
uint16_t vlan);
NOT_REACHED();
}
-/* Attempts to parse 's' as the name of a bond link status detection mode. If
- * successful, stores the mode in '*detect' and returns true. Otherwise
- * returns false without modifying '*detect'. */
-bool
-bond_detect_mode_from_string(enum bond_detect_mode *detect, const char *s)
-{
- if (!strcmp(s, bond_detect_mode_to_string(BLSM_CARRIER))) {
- *detect = BLSM_CARRIER;
- } else if (!strcmp(s, bond_detect_mode_to_string(BLSM_MIIMON))) {
- *detect = BLSM_MIIMON;
- } else {
- return false;
- }
- return true;
-}
-
-/* Returns a string representing 'detect'. */
-const char *
-bond_detect_mode_to_string(enum bond_detect_mode detect)
-{
- switch (detect) {
- case BLSM_CARRIER:
- return "carrier";
- case BLSM_MIIMON:
- return "miimon";
- }
- NOT_REACHED();
-}
\f
/* Creates and returns a new bond whose configuration is initially taken from
* 's'.
bond = xzalloc(sizeof *bond);
hmap_init(&bond->slaves);
bond->no_slaves_tag = tag_create_random();
- bond->miimon_next_update = LLONG_MAX;
+ bond->stb_tag = tag_create_random();
bond->next_fake_iface_update = LLONG_MAX;
bond_reconfigure(bond, s);
hmap_destroy(&bond->slaves);
free(bond->hash);
-
- netdev_monitor_destroy(bond->monitor);
-
free(bond->name);
free(bond);
}
hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
}
- bond->detect = s->detect;
- bond->miimon_interval = s->miimon_interval;
bond->updelay = s->up_delay;
bond->downdelay = s->down_delay;
bond->rebalance_interval = s->rebalance_interval;
revalidate = true;
}
- if (bond->detect == BLSM_CARRIER) {
- struct bond_slave *slave;
-
- if (!bond->monitor) {
- bond->monitor = netdev_monitor_create();
- }
-
- HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
- netdev_monitor_add(bond->monitor, slave->netdev);
- }
- } else {
- netdev_monitor_destroy(bond->monitor);
- bond->monitor = NULL;
-
- if (bond->miimon_next_update == LLONG_MAX) {
- bond->miimon_next_update = time_msec() + bond->miimon_interval;
- }
+ if (bond->basis != s->basis) {
+ bond->basis = s->basis;
+ revalidate = true;
}
if (s->fake_iface) {
bond->next_fake_iface_update = LLONG_MAX;
}
- if (bond->balance != BM_STABLE) {
- free(bond->stb_slaves);
- bond->stb_slaves = NULL;
- bond->stb_tag = 0;
- } else if (!bond->stb_tag) {
- struct bond_slave *slave;
-
- bond->stb_tag = tag_create_random();
-
- assert(!bond->stb_slaves);
- bond->n_stb_slaves = 0;
- bond->len_stb_slaves = 0;
- bond->stb_slaves = NULL;
-
- HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
- if (slave->enabled) {
- bond_stb_enable_slave(slave);
- }
- }
- }
-
if (bond->bond_revalidate) {
revalidate = true;
bond->bond_revalidate = false;
return revalidate;
}
+static void
+bond_slave_set_netdev__(struct bond_slave *slave, struct netdev *netdev)
+{
+ if (slave->netdev != netdev) {
+ slave->netdev = netdev;
+ slave->change_seq = 0;
+ }
+}
+
/* Registers 'slave_' as a slave of 'bond'. The 'slave_' pointer is an
* arbitrary client-provided pointer that uniquely identifies a slave within a
* bond. If 'slave_' already exists within 'bond' then this function
* 'slave_' or destroying 'bond'.
*/
void
-bond_slave_register(struct bond *bond, void *slave_, uint16_t stb_id,
+bond_slave_register(struct bond *bond, void *slave_, uint32_t stb_id,
struct netdev *netdev)
{
struct bond_slave *slave = bond_slave_lookup(bond, slave_);
slave->bond = bond;
slave->aux = slave_;
slave->delay_expires = LLONG_MAX;
- slave->up = bond_is_link_up(bond, netdev);
slave->name = xstrdup(netdev_get_name(netdev));
+ bond->bond_revalidate = true;
slave->enabled = false;
- bond_enable_slave(slave, slave->up, NULL);
+ bond_enable_slave(slave, netdev_get_carrier(netdev), NULL);
}
if (slave->stb_id != stb_id) {
- bond->stb_need_sort = true;
slave->stb_id = stb_id;
+ bond->bond_revalidate = true;
}
- slave->netdev = netdev;
+ bond_slave_set_netdev__(slave, netdev);
+
free(slave->name);
slave->name = xstrdup(netdev_get_name(netdev));
}
+/* Updates the network device to be used with 'slave_' to 'netdev'.
+ *
+ * This is useful if the caller closes and re-opens the network device
+ * registered with bond_slave_register() but doesn't need to change anything
+ * else. */
+void
+bond_slave_set_netdev(struct bond *bond, void *slave_, struct netdev *netdev)
+{
+ struct bond_slave *slave = bond_slave_lookup(bond, slave_);
+ if (slave) {
+ bond_slave_set_netdev__(slave, netdev);
+ }
+}
+
/* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
* then this function has no effect.
*
}
}
-/* Should be called on each slave in 'bond' before bond_run() to indicate the
- * results of lacp_slave_may_enable() on 'slave_'. */
+/* Should be called on each slave in 'bond' before bond_run() to indicate
+ * whether or not 'slave_' may be enabled. This function is intended to allow
+ * other protocols to have some impact on bonding decisions. For example LACP
+ * or high level link monitoring protocols may decide that a given slave should
+ * not be able to send traffic. */
void
-bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
- bool may_enable)
+bond_slave_set_may_enable(struct bond *bond, void *slave_, bool may_enable)
{
- bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
+ bond_slave_lookup(bond, slave_)->may_enable = may_enable;
}
/* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
bond->lacp_negotiated = lacp_negotiated;
- /* Update link status. */
- if (bond->detect == BLSM_CARRIER
- || time_msec() >= bond->miimon_next_update)
- {
- HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
- slave->up = bond_is_link_up(bond, slave->netdev);
- }
- bond->miimon_next_update = time_msec() + bond->miimon_interval;
- }
-
/* Enable slaves based on link status and LACP feedback. */
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
bond_link_status_update(slave, tags);
+ slave->change_seq = netdev_change_seq(slave->netdev);
}
if (!bond->active_slave || !bond->active_slave->enabled) {
bond_choose_active_slave(bond, tags);
bond->next_fake_iface_update = time_msec() + 1000;
}
- bond_stb_sort(bond);
-
if (is_tcp_hash != bond_is_tcp_hash(bond)) {
bond->bond_revalidate = true;
}
{
struct bond_slave *slave;
- if (bond->detect == BLSM_CARRIER) {
- netdev_monitor_poll_wait(bond->monitor);
- } else {
- poll_timer_wait_until(bond->miimon_next_update);
- }
-
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
if (slave->delay_expires != LLONG_MAX) {
poll_timer_wait_until(slave->delay_expires);
}
+
+ if (slave->change_seq != netdev_change_seq(slave->netdev)) {
+ poll_immediate_wake();
+ }
}
if (bond->next_fake_iface_update != LLONG_MAX) {
compose_benign_packet(&packet, "Open vSwitch Bond Failover", 0xf177,
eth_src);
if (vlan) {
- eth_set_vlan_tci(&packet, htons(vlan));
+ eth_push_vlan(&packet, htons(vlan));
}
error = netdev_send(slave->netdev, &packet);
ofpbuf_uninit(&packet);
bond_check_admissibility(struct bond *bond, const void *slave_,
const uint8_t eth_dst[ETH_ADDR_LEN], tag_type *tags)
{
- /* Admit all packets if LACP has been negotiated, because that means that
- * the remote switch is aware of the bond and will "do the right thing". */
+ struct bond_slave *slave = bond_slave_lookup(bond, slave_);
+
+ /* 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 BV_ACCEPT;
+ return slave->enabled ? BV_ACCEPT : BV_DROP;
}
/* Drop all multicast packets on inactive slaves. */
}
}
+ /* Drop all packets which arrive on backup slaves. This is similar to how
+ * Linux bonding handles active-backup bonds. */
+ if (bond->balance == BM_AB) {
+ *tags |= bond_get_active_slave_tag(bond);
+ if (bond->active_slave != slave) {
+ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
+
+ VLOG_WARN_RL(&rl, "active-backup bond received packet on backup"
+ " slave (%s) destined for " ETH_ADDR_FMT,
+ slave->name, ETH_ADDR_ARGS(eth_dst));
+ return BV_DROP;
+ }
+ }
+
/* 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 the
* other. Gratuitous ARP packets are an exception to this rule: the host
bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
uint64_t n_bytes)
{
-
if (bond_is_balanced(bond)) {
lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
}
bond_is_tcp_hash(bond) ? "balance-tcp" : "balance-slb");
}
- ds_put_format(&ds, "bond-detect-mode: %s\n",
- bond->monitor ? "carrier" : "miimon");
-
- if (!bond->monitor) {
- ds_put_format(&ds, "bond-miimon-interval: %lld\n",
- bond->miimon_interval);
- }
+ ds_put_format(&ds, "bond-hash-basis: %"PRIu32"\n", bond->basis);
ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
struct bond_entry *be;
- struct flow flow;
/* Basic info. */
ds_put_format(&ds, "\nslave %s: %s\n",
slave->delay_expires - time_msec());
}
- ds_put_format(&ds, "\tlacp_may_enable: %s\n",
- slave->lacp_may_enable ? "true" : "false");
+ ds_put_format(&ds, "\tmay_enable: %s\n",
+ slave->may_enable ? "true" : "false");
if (!bond_is_balanced(bond)) {
continue;
}
/* Hashes. */
- memset(&flow, 0, sizeof flow);
for (be = bond->hash; be <= &bond->hash[BOND_MASK]; be++) {
int hash = be - bond->hash;
uint8_t hash;
char *hash_cstr;
unsigned int vlan;
- char *mac_s, *vlan_s;
+ uint32_t basis;
+ char *mac_s, *vlan_s, *basis_s;
char *save_ptr = NULL;
mac_s = strtok_r(args, " ", &save_ptr);
vlan_s = strtok_r(NULL, " ", &save_ptr);
+ basis_s = strtok_r(NULL, " ", &save_ptr);
if (vlan_s) {
if (sscanf(vlan_s, "%u", &vlan) != 1) {
vlan = OFP_VLAN_NONE;
}
+ if (basis_s) {
+ if (sscanf(basis_s, "%"PRIu32, &basis) != 1) {
+ unixctl_command_reply(conn, 501, "invalid basis");
+ return;
+ }
+ } else {
+ basis = 0;
+ }
+
if (sscanf(mac_s, ETH_ADDR_SCAN_FMT, ETH_ADDR_SCAN_ARGS(mac))
== ETH_ADDR_SCAN_COUNT) {
- hash = bond_hash_src(mac, vlan) & BOND_MASK;
+ hash = bond_hash_src(mac, vlan, basis) & BOND_MASK;
hash_cstr = xasprintf("%u", hash);
unixctl_command_reply(conn, 200, hash_cstr);
return NULL;
}
-static bool
-bond_is_link_up(struct bond *bond, struct netdev *netdev)
-{
- return (bond->detect == BLSM_CARRIER
- ? netdev_get_carrier(netdev)
- : netdev_get_miimon(netdev));
-}
-
-static int
-bond_stb_sort_cmp__(const void *a_, const void *b_)
-{
- const struct bond_slave *const *ap = a_;
- const struct bond_slave *const *bp = b_;
- const struct bond_slave *a = *ap;
- const struct bond_slave *b = *bp;
- uint16_t aid = a->stb_id;
- uint16_t bid = b->stb_id;
-
- return aid < bid ? -1 : aid > bid;
-}
-
-static void
-bond_stb_sort(struct bond *bond)
-{
- size_t i;
-
- if (!bond->stb_slaves || !bond->stb_need_sort) {
- return;
- }
-
- bond->stb_need_sort = false;
- bond->bond_revalidate = true;
-
- qsort(bond->stb_slaves, bond->n_stb_slaves, sizeof *bond->stb_slaves,
- bond_stb_sort_cmp__);
-
- for (i = 0; i < bond->n_stb_slaves; i++) {
- bond->stb_slaves[i]->stb_idx = i;
- }
-}
-
-static void
-bond_stb_enable_slave(struct bond_slave *slave)
-{
- struct bond *bond = slave->bond;
-
- if (bond->balance != BM_STABLE) {
- return;
- }
-
- bond->stb_need_sort = true;
-
- if (slave->enabled) {
- if (bond->len_stb_slaves <= bond->n_stb_slaves) {
- bond->stb_slaves = x2nrealloc(bond->stb_slaves,
- &bond->len_stb_slaves,
- sizeof *bond->stb_slaves);
- }
-
- slave->stb_idx = bond->n_stb_slaves++;
- bond->stb_slaves[slave->stb_idx] = slave;
- } else {
- size_t index = slave->stb_idx;
- bond->stb_slaves[index] = bond->stb_slaves[--bond->n_stb_slaves];
- bond->stb_slaves[index]->stb_idx = index;
- }
-}
-
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;
VLOG_WARN("interface %s: enabled", slave->name);
slave->tag = tag_create_random();
}
- bond_stb_enable_slave(slave);
+
+ if (bond->balance == BM_STABLE) {
+ bond->bond_revalidate = true;
+ }
}
}
struct bond *bond = slave->bond;
bool up;
- up = slave->up && slave->lacp_may_enable;
+ up = netdev_get_carrier(slave->netdev) && slave->may_enable;
if ((up == slave->enabled) != (slave->delay_expires == LLONG_MAX)) {
static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
VLOG_INFO_RL(&rl, "interface %s: link state %s",
static bool
bond_is_tcp_hash(const struct bond *bond)
{
- return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
- && bond->lacp_negotiated;
+ 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)
+bond_hash_src(const uint8_t mac[ETH_ADDR_LEN], uint16_t vlan, uint32_t basis)
{
- return hash_bytes(mac, ETH_ADDR_LEN, vlan);
+ return hash_3words(hash_bytes(mac, ETH_ADDR_LEN, 0), vlan, basis);
}
static unsigned int
-bond_hash_tcp(const struct flow *flow, uint16_t vlan)
+bond_hash_tcp(const struct flow *flow, uint16_t vlan, uint32_t basis)
{
struct flow hash_flow = *flow;
- hash_flow.vlan_tci = vlan;
+ hash_flow.vlan_tci = htons(vlan);
/* The symmetric quality of this hash function is not required, but
* flow_hash_symmetric_l4 already exists, and is sufficient for our
* purposes, so we use it out of convenience. */
- return flow_hash_symmetric_l4(&hash_flow, 0);
+ return flow_hash_symmetric_l4(&hash_flow, basis);
}
static unsigned int
assert(bond->balance != BM_AB);
return (bond_is_tcp_hash(bond)
- ? bond_hash_tcp(flow, vlan)
- : bond_hash_src(flow->dl_src, vlan));
+ ? bond_hash_tcp(flow, vlan, bond->basis)
+ : bond_hash_src(flow->dl_src, vlan, bond->basis));
}
static struct bond_entry *
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)
return bond->active_slave;
case BM_STABLE:
- if (bond->n_stb_slaves) {
- return bond->stb_slaves[bond_hash(bond, flow, vlan)
- % bond->n_stb_slaves];
- } else {
- return bond->active_slave;
- }
-
+ return choose_stb_slave(bond, flow, vlan);
case BM_SLB:
case BM_TCP:
e = lookup_bond_entry(bond, flow, vlan);
best = NULL;
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
if (slave->delay_expires != LLONG_MAX
- && slave->lacp_may_enable
+ && slave->may_enable
&& (!best || slave->delay_expires < best->delay_expires)) {
best = slave;
}
}
}
- if (!netdev_open_default(bond->name, &bond_dev)) {
+ if (!netdev_open(bond->name, "system", &bond_dev)) {
netdev_set_stats(bond_dev, &bond_stats);
netdev_close(bond_dev);
}