#include "dynamic-string.h"
#include "flow.h"
#include "hmap.h"
-#include "lacp.h"
#include "list.h"
#include "netdev.h"
#include "odp-util.h"
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(). */
uint64_t tx_bytes; /* Sum across 'tx_bytes' of entries. */
/* BM_STABLE specific bonding info. */
- 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
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. */
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. */
-
- /* LACP. */
- const struct lacp *lacp; /* Client provided LACP handle. */
+ tag_type stb_tag; /* Tag associated with this bond. */
/* Monitoring. */
enum bond_detect_mode detect; /* Link status mode, one of BLSM_*. */
static bool bond_is_link_up(struct bond *, struct netdev *);
static void bond_enable_slave(struct bond_slave *, bool enable,
struct tag_set *);
-static bool 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);
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;
*
* 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.
hmap_insert(&all_bonds, &bond->hmap_node, hash_string(bond->name, 0));
}
- bond->lacp = s->lacp;
bond->detect = s->detect;
bond->miimon_interval = s->miimon_interval;
bond->updelay = s->up_delay;
revalidate = true;
}
+ if (bond->basis != s->basis) {
+ bond->basis = s->basis;
+ revalidate = true;
+ }
+
if (bond->detect == BLSM_CARRIER) {
struct bond_slave *slave;
bond->next_fake_iface_update = LLONG_MAX;
}
- if (bond->balance != BM_STABLE) {
- free(bond->stb_slaves);
- bond->stb_slaves = NULL;
- } else if (!bond->stb_slaves) {
- struct bond_slave *slave;
-
- 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;
}
if (bond->balance == BM_AB || !bond->hash || revalidate) {
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 'slave_' should be the same handle
- * used in the LACP module. */
+ */
void
-bond_slave_register(struct bond *bond, void *slave_, struct netdev *netdev)
+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->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));
}
+/* 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);
+ }
+}
+
/* Unregisters 'slave_' from 'bond'. If 'bond' does not contain such a slave
* then this function has no effect.
*
return;
}
+ if (bond->monitor) {
+ netdev_monitor_remove(bond->monitor, slave->netdev);
+ }
+
bond_enable_slave(slave, false, NULL);
del_active = bond->active_slave == slave;
}
}
+/* 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)
+{
+ bond_slave_lookup(bond, slave_)->lacp_may_enable = may_enable;
+}
+
/* Performs periodic maintenance on 'bond'. The caller must provide 'tags' to
* allow tagged flows to be invalidated.
*
* 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;
}
+ if (bond->monitor) {
+ netdev_monitor_flush(bond->monitor);
+ }
+
/* Enable slaves based on link status and LACP feedback. */
HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
bond_link_status_update(slave, tags);
bond->next_fake_iface_update = time_msec() + 1000;
}
- if (bond_stb_sort(bond) || is_tcp_hash != bond_is_tcp_hash(bond)) {
- struct bond_slave *slave;
+ if (is_tcp_hash != bond_is_tcp_hash(bond)) {
+ bond->bond_revalidate = true;
+ }
+
+ if (bond->bond_revalidate) {
+ bond->bond_revalidate = false;
bond_entry_reset(bond);
- HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) {
- tag_set_add(tags, slave->tag);
+ 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;
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");
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());
}
+ 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);
: 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;
- const struct lacp *lacp = a->bond->lacp;
- int a_id, b_id;
-
- if (lacp) {
- a_id = lacp_slave_get_port_id(lacp, a->aux);
- b_id = lacp_slave_get_port_id(lacp, b->aux);
- } else {
- a_id = netdev_get_ifindex(a->netdev);
- b_id = netdev_get_ifindex(b->netdev);
- }
-
- return (a_id == b_id ? 0 : (a_id < b_id ? -1 : 1));
-}
-
-static bool
-bond_stb_sort(struct bond *bond)
-{
- size_t i;
-
- if (!bond->stb_slaves || !bond->stb_need_sort) {
- return false;
- }
- bond->stb_need_sort = false;
-
- 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;
- }
-
- return true;
-}
-
-static void
-bond_stb_enable_slave(struct bond_slave *slave)
-{
- struct bond *bond = slave->bond;
-
- if (!bond->stb_slaves) {
- 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 && lacp_slave_may_enable(bond->lacp, slave->aux);
+ 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) {
bond_is_tcp_hash(const struct bond *bond)
{
return (bond->balance == BM_TCP || bond->balance == BM_STABLE)
- && lacp_negotiated(bond->lacp);
+ && 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;
+ 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
- && lacp_slave_may_enable(bond->lacp, slave->aux)
+ && slave->lacp_may_enable
&& (!best || slave->delay_expires < best->delay_expires)) {
best = slave;
}
git://git.onelab.eu / sliver-openvswitch.git / blobdiff