#include "dynamic-string.h"
#include "flow.h"
#include "hmap.h"
-#include "lacp.h"
#include "list.h"
#include "netdev.h"
#include "odp-util.h"
VLOG_DEFINE_THIS_MODULE(bond);
-COVERAGE_DEFINE(bond_process_lacp);
-
/* Bit-mask for hashing a flow down to a bucket.
* There are (BOND_MASK + 1) buckets. */
#define BOND_MASK 0xff
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. */
tag_type tag; /* Tag associated with this slave. */
/* Rebalancing info. Used only by bond_rebalance(). */
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. */
+ 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. */
long long int next_rebalance; /* Next rebalancing time. */
bool send_learning_packets;
- /* LACP. */
- struct lacp *lacp; /* LACP object. NULL if LACP is disabled. */
+ /* BM_STABLE specific bonding info. */
+ tag_type stb_tag; /* Tag associated with this bond. */
/* Monitoring. */
enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
static struct hmap all_bonds = HMAP_INITIALIZER(&all_bonds);
+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,
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);
*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->miimon_next_update = LLONG_MAX;
bond->next_fake_iface_update = LLONG_MAX;
free(bond->hash);
- lacp_destroy(bond->lacp);
-
netdev_monitor_destroy(bond->monitor);
free(bond->name);
*
* The caller should register each slave on 'bond' by calling
* bond_slave_register(). This is optional if none of the slaves'
- * configuration has changed, except that it is mandatory if 's' enables LACP
- * and 'bond' previously didn't have LACP enabled. In any case it can't
- * hurt.
+ * configuration has changed. In any case it can't hurt.
*
* Returns true if the configuration has changed in such a way that requires
* flow revalidation.
revalidate = true;
}
- if (bond->balance != BM_AB) {
- if (!bond->hash) {
- bond->hash = xcalloc(BOND_MASK + 1, sizeof *bond->hash);
- bond->next_rebalance = time_msec() + bond->rebalance_interval;
- }
- } else {
- if (bond->hash) {
- free(bond->hash);
- bond->hash = NULL;
- }
+ if (bond->basis != s->basis) {
+ bond->basis = s->basis;
+ revalidate = true;
}
if (bond->detect == BLSM_CARRIER) {
}
}
- if (s->lacp) {
- if (!bond->lacp) {
- bond->lacp = lacp_create();
- }
- lacp_configure(bond->lacp, s->lacp);
- } else {
- lacp_destroy(bond->lacp);
- bond->lacp = NULL;
- }
-
if (s->fake_iface) {
if (bond->next_fake_iface_update == LLONG_MAX) {
bond->next_fake_iface_update = time_msec();
bond->next_fake_iface_update = LLONG_MAX;
}
+ if (bond->bond_revalidate) {
+ revalidate = true;
+ bond->bond_revalidate = false;
+ }
+
+ if (bond->balance == BM_AB || !bond->hash || revalidate) {
+ bond_entry_reset(bond);
+ }
+
return revalidate;
}
+static void
+bond_slave_set_netdev__(struct bond *bond, struct bond_slave *slave,
+ struct netdev *netdev)
+{
+ if (slave->netdev != netdev) {
+ if (bond->monitor) {
+ if (slave->netdev) {
+ netdev_monitor_remove(bond->monitor, slave->netdev);
+ }
+ netdev_monitor_add(bond->monitor, netdev);
+ }
+ slave->netdev = netdev;
+ }
+}
+
/* 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
* 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'.
- *
- * If 'bond' has a LACP configuration then 'lacp_settings' must point to LACP
- * settings for 'slave_'; otherwise 'lacp_settings' is ignored.
*/
void
-bond_slave_register(struct bond *bond, void *slave_, struct netdev *netdev,
- const struct lacp_slave_settings *lacp_settings)
+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->aux = slave_;
slave->delay_expires = LLONG_MAX;
slave->up = bond_is_link_up(bond, netdev);
- slave->enabled = slave->up;
+ slave->name = xstrdup(netdev_get_name(netdev));
+ bond->bond_revalidate = true;
+
+ slave->enabled = false;
+ bond_enable_slave(slave, slave->up, NULL);
}
- slave->netdev = netdev;
+ if (slave->stb_id != stb_id) {
+ slave->stb_id = stb_id;
+ bond->bond_revalidate = true;
+ }
+
+ bond_slave_set_netdev__(bond, slave, netdev);
+
free(slave->name);
slave->name = xstrdup(netdev_get_name(netdev));
+}
- if (bond->lacp) {
- assert(lacp_settings != NULL);
- lacp_slave_register(bond->lacp, slave, lacp_settings);
+/* 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__(bond, slave, netdev);
}
}
return;
}
+ if (bond->monitor) {
+ netdev_monitor_remove(bond->monitor, slave->netdev);
+ }
+
+ bond_enable_slave(slave, false, NULL);
+
del_active = bond->active_slave == slave;
if (bond->hash) {
struct bond_entry *e;
}
}
-/* Callback for lacp_run(). */
-static void
-bond_send_pdu_cb(void *slave_, const struct lacp_pdu *pdu)
+/* Should be called on each slave in 'bond' before bond_run() to indicate the
+ * results of lacp_slave_may_enable() on 'slave_'. */
+void
+bond_slave_set_lacp_may_enable(struct bond *bond, void *slave_,
+ bool may_enable)
{
- static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 10);
- struct bond_slave *slave = slave_;
- uint8_t ea[ETH_ADDR_LEN];
- int error;
-
- error = netdev_get_etheraddr(slave->netdev, ea);
- if (!error) {
- struct lacp_pdu *packet_pdu;
- struct ofpbuf packet;
-
- ofpbuf_init(&packet, 0);
- packet_pdu = eth_compose(&packet, eth_addr_lacp, ea, ETH_TYPE_LACP,
- sizeof *packet_pdu);
- *packet_pdu = *pdu;
- error = netdev_send(slave->netdev, &packet);
- if (error) {
- VLOG_WARN_RL(&rl, "bond %s: sending LACP PDU on slave %s failed "
- "(%s)",
- slave->bond->name, slave->name, strerror(error));
- }
- ofpbuf_uninit(&packet);
- } else {
- VLOG_ERR_RL(&rl, "bond %s: cannot obtain Ethernet address of slave "
- "%s (%s)",
- slave->bond->name, slave->name, strerror(error));
- }
+ bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
}
/* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
*
* The caller should check bond_should_send_learning_packets() afterward. */
void
-bond_run(struct bond *bond, struct tag_set *tags)
+bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated)
{
struct bond_slave *slave;
bool is_tcp_hash = bond_is_tcp_hash(bond);
+ bond->lacp_negotiated = lacp_negotiated;
+
/* Update link status. */
if (bond->detect == BLSM_CARRIER
|| time_msec() >= bond->miimon_next_update)
bond->miimon_next_update = time_msec() + bond->miimon_interval;
}
- /* Update LACP. */
- if (bond->lacp) {
- HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
- lacp_slave_enable(bond->lacp, slave, slave->enabled);
- }
-
- lacp_run(bond->lacp, bond_send_pdu_cb);
- }
-
/* Enable slaves based on link status and LACP feedback. */
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
bond_link_status_update(slave, tags);
}
if (is_tcp_hash != bond_is_tcp_hash(bond)) {
- struct bond_slave *slave;
+ bond->bond_revalidate = true;
+ }
- HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
- tag_set_add(tags, slave->tag);
+ if (bond->bond_revalidate) {
+ 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);
}
+ tag_set_add(tags, bond->no_slaves_tag);
}
/* Invalidate any tags required by */
static bool
may_send_learning_packets(const struct bond *bond)
{
- return !lacp_negotiated(bond->lacp) && bond->balance != BM_AB;
+ return !bond->lacp_negotiated && bond->balance != BM_AB;
}
/* Returns true if 'bond' needs the client to send out packets to assist with
{
/* 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". */
- if (lacp_negotiated(bond->lacp)) {
+ if (bond->lacp_negotiated) {
return BV_ACCEPT;
}
{
struct bond_slave *slave = choose_output_slave(bond, flow, vlan);
if (slave) {
- *tags |= slave->tag;
+ *tags |= bond->balance == BM_STABLE ? bond->stb_tag : slave->tag;
return slave->aux;
} else {
*tags |= bond->no_slaves_tag;
return NULL;
}
}
+\f
+/* Rebalancing. */
-/* Processes LACP packet 'packet', which was received on 'slave_' within
- * 'bond'.
- *
- * The client should use this function to pass along LACP messages received on
- * any of 'bond''s slaves. */
-void
-bond_process_lacp(struct bond *bond, void *slave_, const struct ofpbuf *packet)
+static bool
+bond_is_balanced(const struct bond *bond)
{
- if (bond->lacp) {
- struct bond_slave *slave = bond_slave_lookup(bond, slave_);
- const struct lacp_pdu *pdu = parse_lacp_packet(packet);
- if (slave && pdu) {
- COVERAGE_INC(bond_process_lacp);
- lacp_process_pdu(bond->lacp, slave, pdu);
- }
- }
+ return bond->balance == BM_SLB || bond->balance == BM_TCP;
}
-\f
-/* Rebalancing. */
/* Notifies 'bond' that 'n_bytes' bytes were sent in 'flow' within 'vlan'. */
void
bond_account(struct bond *bond, const struct flow *flow, uint16_t vlan,
uint64_t n_bytes)
{
- switch (bond->balance) {
- case BM_AB:
- /* Nothing to do. */
- break;
- case BM_SLB:
- case BM_TCP:
+ if (bond_is_balanced(bond)) {
lookup_bond_entry(bond, flow, vlan)->tx_bytes += n_bytes;
- break;
-
- default:
- NOT_REACHED();
}
}
struct bond_entry *e;
struct list bals;
- if (bond->balance == BM_AB || time_msec() < bond->next_rebalance) {
+ if (!bond_is_balanced(bond) || time_msec() < bond->next_rebalance) {
return;
}
bond->next_rebalance = time_msec() + bond->rebalance_interval;
ds_put_format(&ds, "bond_mode: %s\n",
bond_mode_to_string(bond->balance));
- if (bond->lacp) {
- ds_put_format(&ds, "lacp: %s\n",
- lacp_is_active(bond->lacp) ? "active" : "passive");
- } else {
- ds_put_cstr(&ds, "lacp: off\n");
- }
-
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, "bond-detect-mode: %s\n",
bond->monitor ? "carrier" : "miimon");
ds_put_format(&ds, "updelay: %d ms\n", bond->updelay);
ds_put_format(&ds, "downdelay: %d ms\n", bond->downdelay);
- if (bond->balance != BM_AB) {
+ if (bond_is_balanced(bond)) {
ds_put_format(&ds, "next rebalance: %lld ms\n",
bond->next_rebalance - time_msec());
}
+ ds_put_format(&ds, "lacp_negotiated: %s\n",
+ bond->lacp_negotiated ? "true" : "false");
+
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
struct bond_entry *be;
struct flow flow;
slave->delay_expires - time_msec());
}
- if (bond->balance == BM_AB) {
+ ds_put_format(&ds, "\tlacp_may_enable: %s\n",
+ slave->lacp_may_enable ? "true" : "false");
+
+ if (!bond_is_balanced(bond)) {
continue;
}
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);
void
bond_init(void)
{
- lacp_init();
-
unixctl_command_register("bond/list", bond_unixctl_list, NULL);
unixctl_command_register("bond/show", bond_unixctl_show, NULL);
unixctl_command_register("bond/migrate", bond_unixctl_migrate, NULL);
unixctl_command_register("bond/hash", bond_unixctl_hash, NULL);
}
\f
+static void
+bond_entry_reset(struct bond *bond)
+{
+ if (bond->balance != BM_AB) {
+ size_t hash_len = (BOND_MASK + 1) * sizeof *bond->hash;
+
+ if (!bond->hash) {
+ bond->hash = xmalloc(hash_len);
+ }
+ memset(bond->hash, 0, hash_len);
+
+ bond->next_rebalance = time_msec() + bond->rebalance_interval;
+ } else {
+ free(bond->hash);
+ bond->hash = NULL;
+ }
+}
+
static struct bond_slave *
bond_slave_lookup(struct bond *bond, const void *slave_)
{
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);
- tag_set_add(tags, slave->tag);
+ if (tags) {
+ tag_set_add(tags, slave->tag);
+ }
} else {
VLOG_WARN("interface %s: enabled", slave->name);
slave->tag = tag_create_random();
}
+
+ if (bond->balance == BM_STABLE) {
+ bond->bond_revalidate = true;
+ }
}
}
struct bond *bond = slave->bond;
bool up;
- up = slave->up && lacp_slave_may_enable(bond->lacp, slave);
+ up = slave->up && slave->lacp_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",
VLOG_INFO_RL(&rl, "interface %s: will not be %s",
slave->name, up ? "disabled" : "enabled");
} else {
- int delay = (lacp_negotiated(bond->lacp) ? 0
+ int delay = (bond->lacp_negotiated ? 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 && lacp_negotiated(bond->lacp);
+ return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
+ && bond->lacp_negotiated;
}
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;
/* 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
+bond_hash(const struct bond *bond, const struct flow *flow, uint16_t vlan)
+{
+ assert(bond->balance != BM_AB);
+
+ return (bond_is_tcp_hash(bond)
+ ? bond_hash_tcp(flow, vlan, bond->basis)
+ : bond_hash_src(flow->dl_src, vlan, bond->basis));
}
static struct bond_entry *
lookup_bond_entry(const struct bond *bond, const struct flow *flow,
uint16_t vlan)
{
- assert(bond->balance != BM_AB);
- return &bond->hash[(bond_is_tcp_hash(bond)
- ? bond_hash_tcp(flow, vlan)
- : bond_hash_src(flow->dl_src, vlan)) & BOND_MASK];
+ 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 *
case BM_AB:
return bond->active_slave;
+ case BM_STABLE:
+ 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
- && lacp_slave_may_enable(bond->lacp, slave)
+ && slave->lacp_may_enable
&& (!best || slave->delay_expires < best->delay_expires)) {
best = slave;
}
git://git.onelab.eu / sliver-openvswitch.git / blobdiff