X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=lib%2Fbond.c;h=85bb2da948c9d5c19bd91c045094beabf1c63129;hb=739e1050dcc320af88e7ab227a15d26eb63e5726;hp=5a370158d8c9bbbccad315aeacddd5214691789f;hpb=8bb6f31da62408faf4bcf479faa25640f428a151;p=openvswitch diff --git a/lib/bond.c b/lib/bond.c index 5a370158..85bb2da9 100644 --- a/lib/bond.c +++ b/lib/bond.c @@ -63,9 +63,8 @@ struct bond_slave { /* 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(). */ @@ -74,10 +73,7 @@ struct bond_slave { 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 @@ -95,6 +91,8 @@ struct bond { 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. */ @@ -103,17 +101,10 @@ struct bond { 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. */ @@ -128,17 +119,15 @@ 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, 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); @@ -184,34 +173,6 @@ bond_mode_to_string(enum bond_mode balance) { 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(); -} /* Creates and returns a new bond whose configuration is initially taken from * 's'. @@ -226,8 +187,9 @@ bond_create(const struct bond_settings *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->monitor = netdev_monitor_create(); bond_reconfigure(bond, s); @@ -287,8 +249,6 @@ bond_reconfigure(struct bond *bond, const struct bond_settings *s) 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; @@ -298,23 +258,9 @@ bond_reconfigure(struct bond *bond, const struct bond_settings *s) 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) { @@ -325,25 +271,9 @@ bond_reconfigure(struct bond *bond, const struct bond_settings *s) 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; } if (bond->balance == BM_AB || !bond->hash || revalidate) { @@ -353,6 +283,19 @@ bond_reconfigure(struct bond *bond, const struct bond_settings *s) return revalidate; } +static void +bond_slave_set_netdev__(struct bond *bond, struct bond_slave *slave, + struct netdev *netdev) +{ + if (slave->netdev != netdev) { + 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 @@ -367,7 +310,7 @@ bond_reconfigure(struct bond *bond, const struct bond_settings *s) * '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_); @@ -379,23 +322,38 @@ bond_slave_register(struct bond *bond, void *slave_, uint16_t stb_id, 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__(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. * @@ -410,6 +368,7 @@ bond_slave_unregister(struct bond *bond, const void *slave_) return; } + netdev_monitor_remove(bond->monitor, slave->netdev); bond_enable_slave(slave, false, NULL); del_active = bond->active_slave == slave; @@ -437,13 +396,15 @@ bond_slave_unregister(struct bond *bond, const void *slave_) } } -/* 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 @@ -458,14 +419,8 @@ bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated) 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; + if (bond->monitor) { + netdev_monitor_flush(bond->monitor); } /* Enable slaves based on link status and LACP feedback. */ @@ -482,7 +437,12 @@ bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated) bond->next_fake_iface_update = time_msec() + 1000; } - if (bond_stb_sort(bond) || is_tcp_hash != bond_is_tcp_hash(bond)) { + 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); if (bond->balance != BM_STABLE) { @@ -494,6 +454,7 @@ bond_run(struct bond *bond, struct tag_set *tags, bool lacp_negotiated) } else { tag_set_add(tags, bond->stb_tag); } + tag_set_add(tags, bond->no_slaves_tag); } /* Invalidate any tags required by */ @@ -507,11 +468,7 @@ bond_wait(struct bond *bond) { struct bond_slave *slave; - if (bond->detect == BLSM_CARRIER) { - netdev_monitor_poll_wait(bond->monitor); - } else { - poll_timer_wait_until(bond->miimon_next_update); - } + netdev_monitor_poll_wait(bond->monitor); HMAP_FOR_EACH (slave, hmap_node, &bond->slaves) { if (slave->delay_expires != LLONG_MAX) { @@ -981,13 +938,7 @@ bond_unixctl_show(struct unixctl_conn *conn, 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); @@ -1002,7 +953,6 @@ bond_unixctl_show(struct unixctl_conn *conn, 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", @@ -1016,15 +966,14 @@ bond_unixctl_show(struct unixctl_conn *conn, 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; @@ -1209,11 +1158,13 @@ bond_unixctl_hash(struct unixctl_conn *conn, const char *args_, 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) { @@ -1224,9 +1175,18 @@ bond_unixctl_hash(struct unixctl_conn *conn, const char *args_, 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); @@ -1284,77 +1244,10 @@ bond_slave_lookup(struct bond *bond, const void *slave_) 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 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->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; @@ -1367,7 +1260,10 @@ bond_enable_slave(struct bond_slave *slave, bool enable, struct tag_set *tags) 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; + } } } @@ -1377,7 +1273,7 @@ bond_link_status_update(struct bond_slave *slave, struct tag_set *tags) 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", @@ -1409,26 +1305,26 @@ bond_link_status_update(struct bond_slave *slave, struct tag_set *tags) 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 @@ -1437,8 +1333,8 @@ 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_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 * @@ -1448,6 +1344,37 @@ lookup_bond_entry(const struct bond *bond, const struct flow *flow, 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) @@ -1459,13 +1386,7 @@ choose_output_slave(const struct bond *bond, const struct flow *flow, 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); @@ -1501,7 +1422,7 @@ bond_choose_slave(const struct bond *bond) 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; }