X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Frouter.c;h=fb60c56d0fb40f56caff1bbda4effc5d8e288dd1;hp=45219a69c3a101a92b2dda320cd6958fc852bdaa;hb=2e07619477;hpb=bfaadd73b74da2aca82007ca78a6baf15ea2790c diff --git a/lnet/lnet/router.c b/lnet/lnet/router.c index 45219a6..fb60c56 100644 --- a/lnet/lnet/router.c +++ b/lnet/lnet/router.c @@ -1,10 +1,9 @@ /* * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * - * Copyright (c) 2011, 2014, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. * - * This file is part of Portals - * http://sourceforge.net/projects/sandiaportals/ + * This file is part of Lustre, https://wiki.whamcloud.com/ * * Portals is free software; you can redistribute it and/or * modify it under the terms of version 2 of the GNU General Public @@ -22,9 +21,9 @@ */ #define DEBUG_SUBSYSTEM S_LNET -#include -#if defined(__KERNEL__) && defined(LNET_ROUTER) +#include +#include #define LNET_NRB_TINY_MIN 512 /* min value for each CPT */ #define LNET_NRB_TINY (LNET_NRB_TINY_MIN * 4) @@ -33,170 +32,541 @@ #define LNET_NRB_SMALL_PAGES 1 #define LNET_NRB_LARGE_MIN 256 /* min value for each CPT */ #define LNET_NRB_LARGE (LNET_NRB_LARGE_MIN * 4) -#define LNET_NRB_LARGE_PAGES ((LNET_MTU + PAGE_CACHE_SIZE - 1) >> \ - PAGE_CACHE_SHIFT) +#define LNET_NRB_LARGE_PAGES ((LNET_MTU + PAGE_SIZE - 1) >> \ + PAGE_SHIFT) static char *forwarding = ""; -CFS_MODULE_PARM(forwarding, "s", charp, 0444, - "Explicitly enable/disable forwarding between networks"); +module_param(forwarding, charp, 0444); +MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks"); static int tiny_router_buffers; -CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444, - "# of 0 payload messages to buffer in the router"); +module_param(tiny_router_buffers, int, 0444); +MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router"); static int small_router_buffers; -CFS_MODULE_PARM(small_router_buffers, "i", int, 0444, - "# of small (1 page) messages to buffer in the router"); +module_param(small_router_buffers, int, 0444); +MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router"); static int large_router_buffers; -CFS_MODULE_PARM(large_router_buffers, "i", int, 0444, - "# of large messages to buffer in the router"); -static int peer_buffer_credits = 0; -CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444, - "# router buffer credits per peer"); +module_param(large_router_buffers, int, 0444); +MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router"); +static int peer_buffer_credits; +module_param(peer_buffer_credits, int, 0444); +MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer"); static int auto_down = 1; -CFS_MODULE_PARM(auto_down, "i", int, 0444, - "Automatically mark peers down on comms error"); +module_param(auto_down, int, 0444); +MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error"); int -lnet_peer_buffer_credits(lnet_ni_t *ni) +lnet_peer_buffer_credits(struct lnet_net *net) { - /* NI option overrides LNet default */ - if (ni->ni_peerrtrcredits > 0) - return ni->ni_peerrtrcredits; - if (peer_buffer_credits > 0) - return peer_buffer_credits; - - /* As an approximation, allow this peer the same number of router - * buffers as it is allowed outstanding sends */ - return ni->ni_peertxcredits; + /* NI option overrides LNet default */ + if (net->net_tunables.lct_peer_rtr_credits > 0) + return net->net_tunables.lct_peer_rtr_credits; + if (peer_buffer_credits > 0) + return peer_buffer_credits; + + /* As an approximation, allow this peer the same number of router + * buffers as it is allowed outstanding sends */ + return net->net_tunables.lct_peer_tx_credits; } -/* forward ref's */ -static int lnet_router_checker(void *); +static int check_routers_before_use; +module_param(check_routers_before_use, int, 0444); +MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use"); + +int avoid_asym_router_failure = 1; +module_param(avoid_asym_router_failure, int, 0644); +MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)"); + +int dead_router_check_interval = INT_MIN; +module_param(dead_router_check_interval, int, 0444); +MODULE_PARM_DESC(dead_router_check_interval, "(DEPRECATED - Use alive_router_check_interval)"); + +int live_router_check_interval = INT_MIN; +module_param(live_router_check_interval, int, 0444); +MODULE_PARM_DESC(live_router_check_interval, "(DEPRECATED - Use alive_router_check_interval)"); + +int alive_router_check_interval = 60; +module_param(alive_router_check_interval, int, 0644); +MODULE_PARM_DESC(alive_router_check_interval, "Seconds between live router health checks (<= 0 to disable)"); + +static int router_ping_timeout = 50; +module_param(router_ping_timeout, int, 0644); +MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query"); + +/* + * A value between 0 and 100. 0 meaning that even if router's interfaces + * have the worse health still consider the gateway usable. + * 100 means that at least one interface on the route's remote net is 100% + * healthy to consider the route alive. + * The default is set to 100 to ensure we maintain the original behavior. + */ +unsigned int router_sensitivity_percentage = 100; +static int rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp); +static struct kernel_param_ops param_ops_rtr_sensitivity = { + .set = rtr_sensitivity_set, + .get = param_get_int, +}; +#define param_check_rtr_sensitivity(name, p) \ + __param_check(name, p, int) +#ifdef HAVE_KERNEL_PARAM_OPS +module_param(router_sensitivity_percentage, rtr_sensitivity, S_IRUGO|S_IWUSR); #else +module_param_call(router_sensitivity_percentage, rtr_sensitivity_set, param_get_int, + &router_sensitivity_percentage, S_IRUGO|S_IWUSR); +#endif +MODULE_PARM_DESC(router_sensitivity_percentage, + "How healthy a gateway should be to be used in percent"); -int -lnet_peer_buffer_credits(lnet_ni_t *ni) +static void lnet_add_route_to_rnet(struct lnet_remotenet *rnet, + struct lnet_route *route); +static void lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list, + struct list_head *zombies); + +static int +rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp) { - return 0; + int rc; + unsigned *sen = (unsigned *)kp->arg; + unsigned long value; + + rc = kstrtoul(val, 0, &value); + if (rc) { + CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n"); + return rc; + } + + if (value < 0 || value > 100) { + CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value); + return -EINVAL; + } + + /* + * The purpose of locking the api_mutex here is to ensure that + * the correct value ends up stored properly. + */ + mutex_lock(&the_lnet.ln_api_mutex); + + *sen = value; + + mutex_unlock(&the_lnet.ln_api_mutex); + + return 0; } -#endif +void +lnet_move_route(struct lnet_route *route, struct lnet_peer *lp, + struct list_head *rt_list) +{ + struct lnet_remotenet *rnet; + struct list_head zombies; + struct list_head *l; -static int check_routers_before_use = 0; -CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444, - "Assume routers are down and ping them before use"); + INIT_LIST_HEAD(&zombies); -int avoid_asym_router_failure = 1; -CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644, - "Avoid asymmetrical router failures (0 to disable)"); + if (rt_list) + l = rt_list; + else + l = &zombies; + + rnet = lnet_find_rnet_locked(route->lr_net); + LASSERT(rnet); + + CDEBUG(D_NET, "deleting route %s->%s\n", + libcfs_net2str(route->lr_net), + libcfs_nid2str(route->lr_nid)); + + /* + * use the gateway's lp_primary_nid to delete the route as the + * lr_nid can be a constituent NID of the peer + */ + lnet_del_route_from_rnet(route->lr_gateway->lp_primary_nid, + &rnet->lrn_routes, l); + + if (lp) { + route = list_first_entry(l, struct lnet_route, + lr_list); + route->lr_gateway = lp; + lnet_add_route_to_rnet(rnet, route); + } else { + while (!list_empty(l) && !rt_list) { + route = list_first_entry(l, struct lnet_route, + lr_list); + list_del(&route->lr_list); + LIBCFS_FREE(route, sizeof(*route)); + } + } +} -static int dead_router_check_interval = 60; -CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644, - "Seconds between dead router health checks (<= 0 to disable)"); +void +lnet_rtr_transfer_to_peer(struct lnet_peer *src, struct lnet_peer *target) +{ + struct lnet_route *route; + struct lnet_route *tmp, *tmp2; + + lnet_net_lock(LNET_LOCK_EX); + CDEBUG(D_NET, "transfering routes from %s -> %s\n", + libcfs_nid2str(src->lp_primary_nid), + libcfs_nid2str(target->lp_primary_nid)); + list_for_each_entry(route, &src->lp_routes, lr_gwlist) { + CDEBUG(D_NET, "%s: %s->%s\n", libcfs_nid2str(src->lp_primary_nid), + libcfs_net2str(route->lr_net), + libcfs_nid2str(route->lr_nid)); + } + list_splice_init(&src->lp_rtrq, &target->lp_rtrq); + list_for_each_entry_safe(route, tmp, &src->lp_routes, lr_gwlist) { + struct lnet_route *r2; + bool present = false; + list_for_each_entry_safe(r2, tmp2, &target->lp_routes, lr_gwlist) { + if (route->lr_net == r2->lr_net) { + if (route->lr_priority >= r2->lr_priority) + present = true; + else if (route->lr_hops >= r2->lr_hops) + present = true; + else + lnet_move_route(r2, NULL, NULL); + } + } + if (present) + lnet_move_route(route, NULL, NULL); + else + lnet_move_route(route, target, NULL); + } -static int live_router_check_interval = 60; -CFS_MODULE_PARM(live_router_check_interval, "i", int, 0644, - "Seconds between live router health checks (<= 0 to disable)"); + if (list_empty(&target->lp_rtr_list)) { + lnet_peer_addref_locked(target); + list_add_tail(&target->lp_rtr_list, &the_lnet.ln_routers); + } -static int router_ping_timeout = 50; -CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644, - "Seconds to wait for the reply to a router health query"); + the_lnet.ln_routers_version++; + lnet_net_unlock(LNET_LOCK_EX); +} int lnet_peers_start_down(void) { - return check_routers_before_use; + return check_routers_before_use; +} + +/* + * The peer_net of a gateway is alive if at least one of the peer_ni's on + * that peer_net is alive. + */ +static bool +lnet_is_gateway_net_alive(struct lnet_peer_net *lpn) +{ + struct lnet_peer_ni *lpni; + + list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) { + if (lnet_is_peer_ni_alive(lpni)) + return true; + } + + return false; +} + +/* + * a gateway is alive only if all its nets are alive + * called with cpt lock held + */ +bool lnet_is_gateway_alive(struct lnet_peer *gw) +{ + struct lnet_peer_net *lpn; + + if (!gw->lp_alive) + return false; + + list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) { + if (!lnet_is_gateway_net_alive(lpn)) + return false; + } + + return true; +} + +/* + * lnet_is_route_alive() needs to be called with cpt lock held + * A route is alive if the gateway can route between the local network and + * the remote network of the route. + * This means at least one NI is alive on each of the local and remote + * networks of the gateway. + */ +bool lnet_is_route_alive(struct lnet_route *route) +{ + struct lnet_peer *gw = route->lr_gateway; + struct lnet_peer_net *llpn; + struct lnet_peer_net *rlpn; + + /* If the gateway is down then all routes are considered down */ + if (!gw->lp_alive) + return false; + + /* + * if discovery is disabled then rely on the cached aliveness + * information. This is handicapped information which we log when + * we receive the discovery ping response. The most uptodate + * aliveness information can only be obtained when discovery is + * enabled. + */ + if (lnet_is_discovery_disabled(gw)) + return route->lr_alive; + + /* + * check the gateway's interfaces on the local network + */ + llpn = lnet_peer_get_net_locked(gw, route->lr_lnet); + if (!llpn) + return false; + + if (!lnet_is_gateway_net_alive(llpn)) + return false; + + /* + * For single hop routes avoid_asym_router_failure dictates + * that the remote net must exist on the gateway. For multi-hop + * routes the next-hop will not have the remote net. + */ + if (avoid_asym_router_failure && + (route->lr_hops == 1 || route->lr_hops == LNET_UNDEFINED_HOPS)) { + rlpn = lnet_peer_get_net_locked(gw, route->lr_net); + if (!rlpn) + return false; + if (!lnet_is_gateway_net_alive(rlpn)) + return false; + } + + spin_lock(&gw->lp_lock); + if (!(gw->lp_state & LNET_PEER_ROUTER_ENABLED)) { + spin_unlock(&gw->lp_lock); + if (gw->lp_rtr_refcount > 0) + CERROR("peer %s is being used as a gateway but routing feature is not turned on\n", + libcfs_nid2str(gw->lp_primary_nid)); + return false; + } + spin_unlock(&gw->lp_lock); + + return true; } void -lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when) +lnet_consolidate_routes_locked(struct lnet_peer *orig_lp, + struct lnet_peer *new_lp) +{ + struct lnet_peer_ni *lpni; + struct lnet_route *route; + + /* + * Although a route is correlated with a peer, but when it's added + * a specific NID is used. That NID refers to a peer_ni within + * a peer. There could be other peer_nis on the same net, which + * can be used to send to that gateway. However when we are + * consolidating gateways because of discovery, the nid used to + * add the route might've moved between gateway peers. In this + * case we want to move the route to the new gateway as well. The + * intent here is not to confuse the user who added the route. + */ + list_for_each_entry(route, &orig_lp->lp_routes, lr_gwlist) { + lpni = lnet_peer_get_ni_locked(orig_lp, route->lr_nid); + if (!lpni) { + lnet_net_lock(LNET_LOCK_EX); + list_move(&route->lr_gwlist, &new_lp->lp_routes); + lnet_net_unlock(LNET_LOCK_EX); + } + } +} + +static inline void +lnet_check_route_inconsistency(struct lnet_route *route) +{ + if (!route->lr_single_hop && + (route->lr_hops == 1 || route->lr_hops == LNET_UNDEFINED_HOPS)) { + CWARN("route %s->%s is detected to be multi-hop but hop count is set to %d\n", + libcfs_net2str(route->lr_net), + libcfs_nid2str(route->lr_gateway->lp_primary_nid), + (int) route->lr_hops); + } +} + +static void +lnet_set_route_hop_type(struct lnet_peer *gw, struct lnet_route *route) { - if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */ - CDEBUG(D_NET, "Out of date\n"); - return; - } - - lp->lp_timestamp = when; /* update timestamp */ - lp->lp_ping_deadline = 0; /* disable ping timeout */ - - if (lp->lp_alive_count != 0 && /* got old news */ - (!lp->lp_alive) == (!alive)) { /* new date for old news */ - CDEBUG(D_NET, "Old news\n"); - return; - } - - /* Flag that notification is outstanding */ - - lp->lp_alive_count++; - lp->lp_alive = !(!alive); /* 1 bit! */ - lp->lp_notify = 1; - lp->lp_notifylnd |= notifylnd; - if (lp->lp_alive) - lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */ - - CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive); + struct lnet_peer_net *lpn; + bool single_hop = false; + + list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) { + if (route->lr_net == lpn->lpn_net_id) { + single_hop = true; + break; + } + } + route->lr_single_hop = single_hop; + lnet_check_route_inconsistency(route); +} + +/* Must hold net_lock/EX */ +static inline void +lnet_set_route_aliveness(struct lnet_route *route, bool alive) +{ + /* Log when there's a state change */ + if (route->lr_alive != alive) { + CERROR("route to %s through %s has gone from %s to %s\n", + libcfs_net2str(route->lr_net), + libcfs_nid2str(route->lr_gateway->lp_primary_nid), + (route->lr_alive) ? "up" : "down", + alive ? "up" : "down"); + route->lr_alive = alive; + } } +/* Must hold net_lock/EX */ void -lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp) +lnet_router_discovery_ping_reply(struct lnet_peer *lp) { - int alive; - int notifylnd; + struct lnet_ping_buffer *pbuf = lp->lp_data; + struct lnet_peer_net *llpn; + struct lnet_route *route; + bool single_hop = false; + bool net_up = false; + unsigned lp_state; + __u32 net; + int i; - /* Notify only in 1 thread at any time to ensure ordered notification. - * NB individual events can be missed; the only guarantee is that you - * always get the most recent news */ - if (lp->lp_notifying) - return; + spin_lock(&lp->lp_lock); + lp_state = lp->lp_state; - lp->lp_notifying = 1; + /* only handle replies if discovery is disabled. */ + if (!lnet_is_discovery_disabled_locked(lp)) { + spin_unlock(&lp->lp_lock); + return; + } - while (lp->lp_notify) { - alive = lp->lp_alive; - notifylnd = lp->lp_notifylnd; + spin_unlock(&lp->lp_lock); + + if (lp_state & LNET_PEER_PING_FAILED || + pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED) { + CDEBUG(D_NET, "Set routes down for gw %s because %s %d\n", + libcfs_nid2str(lp->lp_primary_nid), + lp_state & LNET_PEER_PING_FAILED ? "ping failed" : + "route feature is disabled", lp->lp_ping_error); + /* If the ping failed or the peer has routing disabled then + * mark the routes served by this peer down + */ + list_for_each_entry(route, &lp->lp_routes, lr_gwlist) + lnet_set_route_aliveness(route, false); + return; + } + + CDEBUG(D_NET, "Discovery is disabled. Processing reply for gw: %s:%d\n", + libcfs_nid2str(lp->lp_primary_nid), pbuf->pb_info.pi_nnis); + + /* + * examine the ping response to determine if the routes on that + * gateway should be declared alive. + * The route is alive if: + * 1. local network to reach the route is alive and + * 2. route is single hop, avoid_async_router_failure is set and + * there exists at least one NI on the route's remote net + */ + list_for_each_entry(route, &lp->lp_routes, lr_gwlist) { + llpn = lnet_peer_get_net_locked(lp, route->lr_lnet); + if (!llpn) { + lnet_set_route_aliveness(route, false); + continue; + } + + if (!lnet_is_gateway_net_alive(llpn)) { + lnet_set_route_aliveness(route, false); + continue; + } + + single_hop = net_up = false; + for (i = 1; i < pbuf->pb_info.pi_nnis; i++) { + net = LNET_NIDNET(pbuf->pb_info.pi_ni[i].ns_nid); + + if (route->lr_net == net) { + single_hop = true; + if (pbuf->pb_info.pi_ni[i].ns_status == + LNET_NI_STATUS_UP) { + net_up = true; + break; + } + } + } - lp->lp_notifylnd = 0; - lp->lp_notify = 0; + route->lr_single_hop = single_hop; + if (avoid_asym_router_failure && + (route->lr_hops == 1 || + route->lr_hops == LNET_UNDEFINED_HOPS)) + lnet_set_route_aliveness(route, net_up); + else + lnet_set_route_aliveness(route, true); + + /* + * warn that the route is configured as single-hop but it + * really is multi-hop as far as we can tell. + */ + lnet_check_route_inconsistency(route); + } +} - if (notifylnd && ni->ni_lnd->lnd_notify != NULL) { - lnet_net_unlock(lp->lp_cpt); +void +lnet_router_discovery_complete(struct lnet_peer *lp) +{ + struct lnet_peer_ni *lpni = NULL; + struct lnet_route *route; - /* A new notification could happen now; I'll handle it - * when control returns to me */ + spin_lock(&lp->lp_lock); + lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY; + lp->lp_state |= LNET_PEER_RTR_DISCOVERED; + lp->lp_alive = lp->lp_dc_error == 0; + spin_unlock(&lp->lp_lock); - (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive); + /* ping replies are being handled when discovery is disabled */ + if (lnet_is_discovery_disabled_locked(lp)) + return; - lnet_net_lock(lp->lp_cpt); + if (!lp->lp_dc_error) { + /* + * mark single-hop routes. If the remote net is not configured on + * the gateway we assume this is intentional and we mark the + * gateway as multi-hop + */ + list_for_each_entry(route, &lp->lp_routes, lr_gwlist) { + lnet_set_route_aliveness(route, true); + lnet_set_route_hop_type(lp, route); } + + return; } - lp->lp_notifying = 0; + /* + * We do not send messages directly to the remote interfaces + * of an LNet router. As such, we rely on the PING response + * to determine the up/down status of these interfaces. If + * a PING response is not receieved, or some other problem with + * discovery occurs that prevents us from getting this status, + * we assume all interfaces are down until we're able to + * determine otherwise. + */ + CDEBUG(D_NET, "%s: Router discovery failed %d\n", + libcfs_nid2str(lp->lp_primary_nid), lp->lp_dc_error); + while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) + lpni->lpni_ns_status = LNET_NI_STATUS_DOWN; + + list_for_each_entry(route, &lp->lp_routes, lr_gwlist) + lnet_set_route_aliveness(route, false); } - static void -lnet_rtr_addref_locked(lnet_peer_t *lp) +lnet_rtr_addref_locked(struct lnet_peer *lp) { - LASSERT(lp->lp_refcount > 0); LASSERT(lp->lp_rtr_refcount >= 0); /* lnet_net_lock must be exclusively locked */ lp->lp_rtr_refcount++; if (lp->lp_rtr_refcount == 1) { - struct list_head *pos; - - /* a simple insertion sort */ - list_for_each_prev(pos, &the_lnet.ln_routers) { - lnet_peer_t *rtr = list_entry(pos, lnet_peer_t, - lp_rtr_list); - - if (rtr->lp_nid < lp->lp_nid) - break; - } - - list_add(&lp->lp_rtr_list, pos); + list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers); /* addref for the_lnet.ln_routers */ lnet_peer_addref_locked(lp); the_lnet.ln_routers_version++; @@ -204,9 +574,9 @@ lnet_rtr_addref_locked(lnet_peer_t *lp) } static void -lnet_rtr_decref_locked(lnet_peer_t *lp) +lnet_rtr_decref_locked(struct lnet_peer *lp) { - LASSERT(lp->lp_refcount > 0); + LASSERT(atomic_read(&lp->lp_refcount) > 0); LASSERT(lp->lp_rtr_refcount > 0); /* lnet_net_lock must be exclusively locked */ @@ -214,12 +584,6 @@ lnet_rtr_decref_locked(lnet_peer_t *lp) if (lp->lp_rtr_refcount == 0) { LASSERT(list_empty(&lp->lp_routes)); - if (lp->lp_rcd != NULL) { - list_add(&lp->lp_rcd->rcd_list, - &the_lnet.ln_rcd_deathrow); - lp->lp_rcd = NULL; - } - list_del(&lp->lp_rtr_list); /* decref for the_lnet.ln_routers */ lnet_peer_decref_locked(lp); @@ -227,18 +591,18 @@ lnet_rtr_decref_locked(lnet_peer_t *lp) } } -lnet_remotenet_t * -lnet_find_net_locked (__u32 net) +struct lnet_remotenet * +lnet_find_rnet_locked(__u32 net) { - lnet_remotenet_t *rnet; + struct lnet_remotenet *rnet; struct list_head *tmp; struct list_head *rn_list; - LASSERT(!the_lnet.ln_shutdown); + LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING); rn_list = lnet_net2rnethash(net); list_for_each(tmp, rn_list) { - rnet = list_entry(tmp, lnet_remotenet_t, lrn_list); + rnet = list_entry(tmp, struct lnet_remotenet, lrn_list); if (rnet->lrn_net == net) return rnet; @@ -248,87 +612,99 @@ lnet_find_net_locked (__u32 net) static void lnet_shuffle_seed(void) { - static int seeded = 0; - int lnd_type, seed[2]; - struct timeval tv; - lnet_ni_t *ni; - struct list_head *tmp; + static int seeded; + struct lnet_ni *ni = NULL; - if (seeded) - return; - - cfs_get_random_bytes(seed, sizeof(seed)); + if (seeded) + return; /* Nodes with small feet have little entropy * the NID for this node gives the most entropy in the low bits */ - list_for_each(tmp, &the_lnet.ln_nis) { - ni = list_entry(tmp, lnet_ni_t, ni_list); - lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid)); + while ((ni = lnet_get_next_ni_locked(NULL, ni))) + add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid)); - if (lnd_type != LOLND) - seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type); - } - - do_gettimeofday(&tv); - cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]); seeded = 1; - return; } /* NB expects LNET_LOCK held */ -void -lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route) +static void +lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route) { - unsigned int len = 0; - unsigned int offset = 0; + struct lnet_peer_net *lpn; + unsigned int offset = 0; + unsigned int len = 0; struct list_head *e; lnet_shuffle_seed(); - list_for_each(e, &rnet->lrn_routes) { + list_for_each(e, &rnet->lrn_routes) len++; - } - /* len+1 positions to add a new entry, also prevents division by 0 */ - offset = cfs_rand() % (len + 1); + /* + * Randomly adding routes to the list is done to ensure that when + * different nodes are using the same list of routers, they end up + * preferring different routers. + */ + offset = prandom_u32_max(len + 1); list_for_each(e, &rnet->lrn_routes) { if (offset == 0) break; offset--; } list_add(&route->lr_list, e); - list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes); + /* + * force a router check on the gateway to make sure the route is + * alive + */ + list_for_each_entry(lpn, &route->lr_gateway->lp_peer_nets, + lpn_peer_nets) { + lpn->lpn_rtrcheck_timestamp = 0; + } the_lnet.ln_remote_nets_version++; + + /* add the route on the gateway list */ + list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes); + + /* take a router reference count on the gateway */ lnet_rtr_addref_locked(route->lr_gateway); } int -lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway, - unsigned int priority) +lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway, + __u32 priority, __u32 sensitivity) { - struct list_head *e; - lnet_remotenet_t *rnet; - lnet_remotenet_t *rnet2; - lnet_route_t *route; - lnet_ni_t *ni; - int add_route; - int rc; - - CDEBUG(D_NET, "Add route: net %s hops %u priority %u gw %s\n", + struct list_head *route_entry; + struct lnet_remotenet *rnet; + struct lnet_remotenet *rnet2; + struct lnet_route *route; + struct lnet_peer_ni *lpni; + struct lnet_peer *gw; + int add_route; + int rc; + + CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n", libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway)); if (gateway == LNET_NID_ANY || - LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND || - net == LNET_NIDNET(LNET_NID_ANY) || + gateway == LNET_NID_LO_0 || + net == LNET_NET_ANY || LNET_NETTYP(net) == LOLND || LNET_NIDNET(gateway) == net || - hops < 1 || hops > 255) + (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255))) return -EINVAL; - if (lnet_islocalnet(net)) /* it's a local network */ + /* it's a local network */ + if (lnet_islocalnet(net)) return -EEXIST; + if (!lnet_islocalnet(LNET_NIDNET(gateway))) { + CERROR("Cannot add route with gateway %s. There is no local interface configured on LNet %s\n", + libcfs_nid2str(gateway), + libcfs_net2str(LNET_NIDNET(gateway))); + return -EHOSTUNREACH; + } + /* Assume net, route, all new */ LIBCFS_ALLOC(route, sizeof(*route)); LIBCFS_ALLOC(rnet, sizeof(*rnet)); @@ -344,30 +720,39 @@ lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway, INIT_LIST_HEAD(&rnet->lrn_routes); rnet->lrn_net = net; - route->lr_hops = hops; + /* store the local and remote net that the route represents */ + route->lr_lnet = LNET_NIDNET(gateway); route->lr_net = net; + route->lr_nid = gateway; route->lr_priority = priority; + route->lr_hops = hops; lnet_net_lock(LNET_LOCK_EX); - rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX); - if (rc != 0) { + /* + * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to + * lose that once we're done + */ + lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX); + if (IS_ERR(lpni)) { lnet_net_unlock(LNET_LOCK_EX); LIBCFS_FREE(route, sizeof(*route)); LIBCFS_FREE(rnet, sizeof(*rnet)); - if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */ - return rc; /* ignore the route entry */ + rc = PTR_ERR(lpni); CERROR("Error %d creating route %s %d %s\n", rc, libcfs_net2str(net), hops, libcfs_nid2str(gateway)); return rc; } - LASSERT(!the_lnet.ln_shutdown); + LASSERT(lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer); + gw = lpni->lpni_peer_net->lpn_peer; + + route->lr_gateway = gw; - rnet2 = lnet_find_net_locked(net); + rnet2 = lnet_find_rnet_locked(net); if (rnet2 == NULL) { /* new network */ list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net)); @@ -376,34 +761,39 @@ lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway, /* Search for a duplicate route (it's a NOOP if it is) */ add_route = 1; - list_for_each(e, &rnet2->lrn_routes) { - lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list); + list_for_each(route_entry, &rnet2->lrn_routes) { + struct lnet_route *route2; + route2 = list_entry(route_entry, struct lnet_route, lr_list); if (route2->lr_gateway == route->lr_gateway) { add_route = 0; break; } /* our lookups must be true */ - LASSERT(route2->lr_gateway->lp_nid != gateway); + LASSERT(route2->lr_gateway->lp_primary_nid != gateway); } + /* + * It is possible to add multiple routes through the same peer, + * but it'll be using a different NID of that peer. When the + * gateway is discovered, discovery will consolidate the different + * peers into one peer. In this case the discovery code will have + * to move the routes from the peer that's being deleted to the + * consolidated peer lp_routes list + */ if (add_route) { - lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */ + gw->lp_health_sensitivity = sensitivity; lnet_add_route_to_rnet(rnet2, route); - - ni = route->lr_gateway->lp_ni; - lnet_net_unlock(LNET_LOCK_EX); - - /* XXX Assume alive */ - if (ni->ni_lnd->lnd_notify != NULL) - (ni->ni_lnd->lnd_notify)(ni, gateway, 1); - - lnet_net_lock(LNET_LOCK_EX); + if (lnet_peer_discovery_disabled) + CWARN("Consider turning discovery on to enable full " + "Multi-Rail routing functionality\n"); } - /* -1 for notify or !add_route */ - lnet_peer_decref_locked(route->lr_gateway); + /* + * get rid of the reference on the lpni. + */ + lnet_peer_ni_decref_locked(lpni); lnet_net_unlock(LNET_LOCK_EX); rc = 0; @@ -416,77 +806,54 @@ lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway, if (rnet != rnet2) LIBCFS_FREE(rnet, sizeof(*rnet)); + /* kick start the monitor thread to handle the added route */ + complete(&the_lnet.ln_mt_wait_complete); + return rc; } -int -lnet_check_routes(void) +void +lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list, + struct list_head *zombies) { - lnet_remotenet_t *rnet; - lnet_route_t *route; - lnet_route_t *route2; - struct list_head *e1; - struct list_head *e2; - int cpt; - struct list_head *rn_list; - int i; - - cpt = lnet_net_lock_current(); - - for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) { - rn_list = &the_lnet.ln_remote_nets_hash[i]; - list_for_each(e1, rn_list) { - rnet = list_entry(e1, lnet_remotenet_t, lrn_list); - - route2 = NULL; - list_for_each(e2, &rnet->lrn_routes) { - lnet_nid_t nid1; - lnet_nid_t nid2; - int net; - - route = list_entry(e2, lnet_route_t, - lr_list); - - if (route2 == NULL) { - route2 = route; - continue; - } - - if (route->lr_gateway->lp_ni == - route2->lr_gateway->lp_ni) - continue; - - nid1 = route->lr_gateway->lp_nid; - nid2 = route2->lr_gateway->lp_nid; - net = rnet->lrn_net; - - lnet_net_unlock(cpt); + struct lnet_peer *gateway; + struct lnet_route *route; + struct lnet_route *tmp; + + list_for_each_entry_safe(route, tmp, route_list, lr_list) { + gateway = route->lr_gateway; + if (gw_nid != LNET_NID_ANY && + gw_nid != gateway->lp_primary_nid) + continue; - CERROR("Routes to %s via %s and %s not " - "supported\n", - libcfs_net2str(net), - libcfs_nid2str(nid1), - libcfs_nid2str(nid2)); - return -EINVAL; - } - } + /* + * move to zombie to delete outside the lock + * Note that this function is called with the + * ln_api_mutex held as well as the exclusive net + * lock. Adding to the remote net list happens + * under the same conditions. Same goes for the + * gateway router list + */ + list_move(&route->lr_list, zombies); + the_lnet.ln_remote_nets_version++; + + list_del(&route->lr_gwlist); + lnet_rtr_decref_locked(gateway); } - - lnet_net_unlock(cpt); - return 0; } int lnet_del_route(__u32 net, lnet_nid_t gw_nid) { - struct lnet_peer *gateway; - lnet_remotenet_t *rnet; - lnet_route_t *route; - struct list_head *e1; - struct list_head *e2; - int rc = -ENOENT; - struct list_head *rn_list; - int idx = 0; + LIST_HEAD(rnet_zombies); + struct lnet_remotenet *rnet; + struct lnet_remotenet *tmp; + struct list_head *rn_list; + struct lnet_peer_ni *lpni; + struct lnet_route *route; + LIST_HEAD(zombies); + struct lnet_peer *lp = NULL; + int i = 0; CDEBUG(D_NET, "Del route: net %s : gw %s\n", libcfs_net2str(net), libcfs_nid2str(gw_nid)); @@ -495,879 +862,425 @@ lnet_del_route(__u32 net, lnet_nid_t gw_nid) * or a specific route entry actual NIDs) */ lnet_net_lock(LNET_LOCK_EX); - if (net == LNET_NIDNET(LNET_NID_ANY)) - rn_list = &the_lnet.ln_remote_nets_hash[0]; - else - rn_list = lnet_net2rnethash(net); - -again: - list_for_each(e1, rn_list) { - rnet = list_entry(e1, lnet_remotenet_t, lrn_list); - - if (!(net == LNET_NIDNET(LNET_NID_ANY) || - net == rnet->lrn_net)) - continue; - - list_for_each(e2, &rnet->lrn_routes) { - route = list_entry(e2, lnet_route_t, lr_list); - - gateway = route->lr_gateway; - if (!(gw_nid == LNET_NID_ANY || - gw_nid == gateway->lp_nid)) - continue; - - list_del(&route->lr_list); - list_del(&route->lr_gwlist); - the_lnet.ln_remote_nets_version++; - - if (list_empty(&rnet->lrn_routes)) - list_del(&rnet->lrn_list); - else - rnet = NULL; - lnet_rtr_decref_locked(gateway); - lnet_peer_decref_locked(gateway); - - lnet_net_unlock(LNET_LOCK_EX); - - LIBCFS_FREE(route, sizeof(*route)); - - if (rnet != NULL) - LIBCFS_FREE(rnet, sizeof(*rnet)); - - rc = 0; - lnet_net_lock(LNET_LOCK_EX); - goto again; - } + lpni = lnet_find_peer_ni_locked(gw_nid); + if (lpni) { + lp = lpni->lpni_peer_net->lpn_peer; + LASSERT(lp); + gw_nid = lp->lp_primary_nid; + lnet_peer_ni_decref_locked(lpni); } - if (net == LNET_NIDNET(LNET_NID_ANY) && - ++idx < LNET_REMOTE_NETS_HASH_SIZE) { - rn_list = &the_lnet.ln_remote_nets_hash[idx]; - goto again; - } - lnet_net_unlock(LNET_LOCK_EX); - - return rc; -} - -void -lnet_destroy_routes (void) -{ - lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY); -} - -int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg) -{ - int i, rc = -ENOENT, lidx, j; - - if (the_lnet.ln_rtrpools == NULL) - return rc; - - for (i = 0; i < LNET_NRBPOOLS; i++) { - lnet_rtrbufpool_t *rbp; - - lnet_net_lock(LNET_LOCK_EX); - lidx = idx; - cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) { - if (lidx-- == 0) { - rc = 0; - pool_cfg->pl_pools[i].pl_npages = - rbp[i].rbp_npages; - pool_cfg->pl_pools[i].pl_nbuffers = - rbp[i].rbp_nbuffers; - pool_cfg->pl_pools[i].pl_credits = - rbp[i].rbp_credits; - pool_cfg->pl_pools[i].pl_mincredits = - rbp[i].rbp_mincredits; - break; - } + if (net != LNET_NET_ANY) { + rnet = lnet_find_rnet_locked(net); + if (!rnet) { + lnet_net_unlock(LNET_LOCK_EX); + return -ENOENT; } - lnet_net_unlock(LNET_LOCK_EX); + lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes, + &zombies); + if (list_empty(&rnet->lrn_routes)) + list_move(&rnet->lrn_list, &rnet_zombies); + goto delete_zombies; } - lnet_net_lock(LNET_LOCK_EX); - pool_cfg->pl_routing = the_lnet.ln_routing; - lnet_net_unlock(LNET_LOCK_EX); - - return rc; -} - -int -lnet_get_route(int idx, __u32 *net, __u32 *hops, - lnet_nid_t *gateway, __u32 *alive, __u32 *priority) -{ - struct list_head *e1; - struct list_head *e2; - lnet_remotenet_t *rnet; - lnet_route_t *route; - int cpt; - int i; - struct list_head *rn_list; - - cpt = lnet_net_lock_current(); - for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) { rn_list = &the_lnet.ln_remote_nets_hash[i]; - list_for_each(e1, rn_list) { - rnet = list_entry(e1, lnet_remotenet_t, lrn_list); - - list_for_each(e2, &rnet->lrn_routes) { - route = list_entry(e2, lnet_route_t, - lr_list); - - if (idx-- == 0) { - *net = rnet->lrn_net; - *hops = route->lr_hops; - *priority = route->lr_priority; - *gateway = route->lr_gateway->lp_nid; - *alive = - route->lr_gateway->lp_alive && - !route->lr_downis; - lnet_net_unlock(cpt); - return 0; - } - } - } - } - - lnet_net_unlock(cpt); - return -ENOENT; -} - -void -lnet_swap_pinginfo(lnet_ping_info_t *info) -{ - int i; - lnet_ni_status_t *stat; - - __swab32s(&info->pi_magic); - __swab32s(&info->pi_features); - __swab32s(&info->pi_pid); - __swab32s(&info->pi_nnis); - for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) { - stat = &info->pi_ni[i]; - __swab64s(&stat->ns_nid); - __swab32s(&stat->ns_status); - } - return; -} - -/** - * parse router-checker pinginfo, record number of down NIs for remote - * networks on that router. - */ -static void -lnet_parse_rc_info(lnet_rc_data_t *rcd) -{ - lnet_ping_info_t *info = rcd->rcd_pinginfo; - struct lnet_peer *gw = rcd->rcd_gateway; - lnet_route_t *rte; - - if (!gw->lp_alive) - return; - - if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) - lnet_swap_pinginfo(info); - - /* NB always racing with network! */ - if (info->pi_magic != LNET_PROTO_PING_MAGIC) { - CDEBUG(D_NET, "%s: Unexpected magic %08x\n", - libcfs_nid2str(gw->lp_nid), info->pi_magic); - gw->lp_ping_feats = LNET_PING_FEAT_INVAL; - return; - } - - gw->lp_ping_feats = info->pi_features; - if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) { - CDEBUG(D_NET, "%s: Unexpected features 0x%x\n", - libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats); - return; /* nothing I can understand */ - } - - if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0) - return; /* can't carry NI status info */ - - list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) { - int down = 0; - int up = 0; - int i; - - if ((gw->lp_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) { - rte->lr_downis = 1; - continue; - } - - for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) { - lnet_ni_status_t *stat = &info->pi_ni[i]; - lnet_nid_t nid = stat->ns_nid; - - if (nid == LNET_NID_ANY) { - CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n", - libcfs_nid2str(gw->lp_nid)); - gw->lp_ping_feats = LNET_PING_FEAT_INVAL; - return; - } - - if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND) - continue; - - if (stat->ns_status == LNET_NI_STATUS_DOWN) { - down++; - continue; - } - - if (stat->ns_status == LNET_NI_STATUS_UP) { - if (LNET_NIDNET(nid) == rte->lr_net) { - up = 1; - break; - } - continue; - } - - CDEBUG(D_NET, "%s: Unexpected status 0x%x\n", - libcfs_nid2str(gw->lp_nid), stat->ns_status); - gw->lp_ping_feats = LNET_PING_FEAT_INVAL; - return; - } - - if (up) { /* ignore downed NIs if NI for dest network is up */ - rte->lr_downis = 0; - continue; - } - rte->lr_downis = down; - } -} - -static void -lnet_router_checker_event(lnet_event_t *event) -{ - lnet_rc_data_t *rcd = event->md.user_ptr; - struct lnet_peer *lp; - - LASSERT(rcd != NULL); - - if (event->unlinked) { - LNetInvalidateHandle(&rcd->rcd_mdh); - return; - } - - LASSERT(event->type == LNET_EVENT_SEND || - event->type == LNET_EVENT_REPLY); - - lp = rcd->rcd_gateway; - LASSERT(lp != NULL); - - /* NB: it's called with holding lnet_res_lock, we have a few - * places need to hold both locks at the same time, please take - * care of lock ordering */ - lnet_net_lock(lp->lp_cpt); - if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) { - /* ignore if no longer a router or rcd is replaced */ - goto out; - } - - if (event->type == LNET_EVENT_SEND) { - lp->lp_ping_notsent = 0; - if (event->status == 0) - goto out; - } - - /* LNET_EVENT_REPLY */ - /* A successful REPLY means the router is up. If _any_ comms - * to the router fail I assume it's down (this will happen if - * we ping alive routers to try to detect router death before - * apps get burned). */ - - lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current()); - /* The router checker will wake up very shortly and do the - * actual notification. - * XXX If 'lp' stops being a router before then, it will still - * have the notification pending!!! */ - - if (avoid_asym_router_failure && event->status == 0) - lnet_parse_rc_info(rcd); - - out: - lnet_net_unlock(lp->lp_cpt); -} - -void -lnet_wait_known_routerstate(void) -{ - lnet_peer_t *rtr; - struct list_head *entry; - int all_known; - - LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); - - for (;;) { - int cpt = lnet_net_lock_current(); - - all_known = 1; - list_for_each(entry, &the_lnet.ln_routers) { - rtr = list_entry(entry, lnet_peer_t, lp_rtr_list); - - if (rtr->lp_alive_count == 0) { - all_known = 0; - break; - } - } - - lnet_net_unlock(cpt); - - if (all_known) - return; - -#ifndef __KERNEL__ - lnet_router_checker(); -#endif - cfs_pause(cfs_time_seconds(1)); - } -} - -void -lnet_router_ni_update_locked(lnet_peer_t *gw, __u32 net) -{ - lnet_route_t *rte; - - if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) { - list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) { - if (rte->lr_net == net) { - rte->lr_downis = 0; - break; - } - } - } -} - -void -lnet_update_ni_status_locked(void) -{ - lnet_ni_t *ni; - long now; - int timeout; - - LASSERT(the_lnet.ln_routing); - - timeout = router_ping_timeout + - MAX(live_router_check_interval, dead_router_check_interval); - - now = cfs_time_current_sec(); - list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) { - if (ni->ni_lnd->lnd_type == LOLND) - continue; - - if (now < ni->ni_last_alive + timeout) - continue; - - lnet_ni_lock(ni); - /* re-check with lock */ - if (now < ni->ni_last_alive + timeout) { - lnet_ni_unlock(ni); - continue; - } - - LASSERT(ni->ni_status != NULL); - if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) { - CDEBUG(D_NET, "NI(%s:%d) status changed to down\n", - libcfs_nid2str(ni->ni_nid), timeout); - /* NB: so far, this is the only place to set - * NI status to "down" */ - ni->ni_status->ns_status = LNET_NI_STATUS_DOWN; + list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) { + lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes, + &zombies); + if (list_empty(&rnet->lrn_routes)) + list_move(&rnet->lrn_list, &rnet_zombies); } - lnet_ni_unlock(ni); } -} - -void -lnet_destroy_rc_data(lnet_rc_data_t *rcd) -{ - LASSERT(list_empty(&rcd->rcd_list)); - /* detached from network */ - LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh)); - - if (rcd->rcd_gateway != NULL) { - int cpt = rcd->rcd_gateway->lp_cpt; - lnet_net_lock(cpt); - lnet_peer_decref_locked(rcd->rcd_gateway); - lnet_net_unlock(cpt); +delete_zombies: + /* + * check if there are any routes remaining on the gateway + * If there are no more routes make sure to set the peer's + * lp_disc_net_id to 0 (invalid), in case we add more routes in + * the future on that gateway, then we start our discovery process + * from scratch + */ + if (lpni) { + if (list_empty(&lp->lp_routes)) + lp->lp_disc_net_id = 0; } - if (rcd->rcd_pinginfo != NULL) - LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE); + lnet_net_unlock(LNET_LOCK_EX); - LIBCFS_FREE(rcd, sizeof(*rcd)); -} + while (!list_empty(&zombies)) { + route = list_first_entry(&zombies, struct lnet_route, lr_list); + list_del(&route->lr_list); + LIBCFS_FREE(route, sizeof(*route)); + } -lnet_rc_data_t * -lnet_create_rc_data_locked(lnet_peer_t *gateway) -{ - lnet_rc_data_t *rcd = NULL; - lnet_ping_info_t *pi; - int rc; - int i; - - lnet_net_unlock(gateway->lp_cpt); - - LIBCFS_ALLOC(rcd, sizeof(*rcd)); - if (rcd == NULL) - goto out; - - LNetInvalidateHandle(&rcd->rcd_mdh); - INIT_LIST_HEAD(&rcd->rcd_list); - - LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE); - if (pi == NULL) - goto out; - - for (i = 0; i < LNET_MAX_RTR_NIS; i++) { - pi->pi_ni[i].ns_nid = LNET_NID_ANY; - pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID; - } - rcd->rcd_pinginfo = pi; - - LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh)); - rc = LNetMDBind((lnet_md_t){.start = pi, - .user_ptr = rcd, - .length = LNET_PINGINFO_SIZE, - .threshold = LNET_MD_THRESH_INF, - .options = LNET_MD_TRUNCATE, - .eq_handle = the_lnet.ln_rc_eqh}, - LNET_UNLINK, - &rcd->rcd_mdh); - if (rc < 0) { - CERROR("Can't bind MD: %d\n", rc); - goto out; - } - LASSERT(rc == 0); - - lnet_net_lock(gateway->lp_cpt); - /* router table changed or someone has created rcd for this gateway */ - if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) { - lnet_net_unlock(gateway->lp_cpt); - goto out; - } - - lnet_peer_addref_locked(gateway); - rcd->rcd_gateway = gateway; - gateway->lp_rcd = rcd; - gateway->lp_ping_notsent = 0; - - return rcd; - - out: - if (rcd != NULL) { - if (!LNetHandleIsInvalid(rcd->rcd_mdh)) { - rc = LNetMDUnlink(rcd->rcd_mdh); - LASSERT(rc == 0); - } - lnet_destroy_rc_data(rcd); + while (!list_empty(&rnet_zombies)) { + rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet, + lrn_list); + list_del(&rnet->lrn_list); + LIBCFS_FREE(rnet, sizeof(*rnet)); } - lnet_net_lock(gateway->lp_cpt); - return gateway->lp_rcd; + return 0; } -static int -lnet_router_check_interval (lnet_peer_t *rtr) +void +lnet_destroy_routes (void) { - int secs; - - secs = rtr->lp_alive ? live_router_check_interval : - dead_router_check_interval; - if (secs < 0) - secs = 0; - - return secs; + lnet_del_route(LNET_NET_ANY, LNET_NID_ANY); } -static void -lnet_ping_router_locked (lnet_peer_t *rtr) +int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg) { - lnet_rc_data_t *rcd = NULL; - cfs_time_t now = cfs_time_current(); - int secs; + struct lnet_rtrbufpool *rbp; + int i, rc = -ENOENT, j; - lnet_peer_addref_locked(rtr); + if (the_lnet.ln_rtrpools == NULL) + return rc; - if (rtr->lp_ping_deadline != 0 && /* ping timed out? */ - cfs_time_after(now, rtr->lp_ping_deadline)) - lnet_notify_locked(rtr, 1, 0, now); - /* Run any outstanding notifications */ - lnet_ni_notify_locked(rtr->lp_ni, rtr); + cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) { + if (i != cpt) + continue; - if (!lnet_isrouter(rtr) || - the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) { - /* router table changed or router checker is shutting down */ - lnet_peer_decref_locked(rtr); - return; + lnet_net_lock(i); + for (j = 0; j < LNET_NRBPOOLS; j++) { + pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages; + pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers; + pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits; + pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits; + } + lnet_net_unlock(i); + rc = 0; + break; } - rcd = rtr->lp_rcd != NULL ? - rtr->lp_rcd : lnet_create_rc_data_locked(rtr); - - if (rcd == NULL) - return; - - secs = lnet_router_check_interval(rtr); + lnet_net_lock(LNET_LOCK_EX); + pool_cfg->pl_routing = the_lnet.ln_routing; + lnet_net_unlock(LNET_LOCK_EX); - CDEBUG(D_NET, - "rtr %s %d: deadline %lu ping_notsent %d alive %d " - "alive_count %d lp_ping_timestamp %lu\n", - libcfs_nid2str(rtr->lp_nid), secs, - rtr->lp_ping_deadline, rtr->lp_ping_notsent, - rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp); + return rc; +} - if (secs != 0 && !rtr->lp_ping_notsent && - cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp, - cfs_time_seconds(secs)))) { - int rc; - lnet_process_id_t id; - lnet_handle_md_t mdh; +int +lnet_get_route(int idx, __u32 *net, __u32 *hops, + lnet_nid_t *gateway, __u32 *flags, __u32 *priority, __u32 *sensitivity) +{ + struct lnet_remotenet *rnet; + struct list_head *rn_list; + struct lnet_route *route; + struct list_head *e1; + struct list_head *e2; + int cpt; + int i; - id.nid = rtr->lp_nid; - id.pid = LNET_PID_LUSTRE; - CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id)); + cpt = lnet_net_lock_current(); - rtr->lp_ping_notsent = 1; - rtr->lp_ping_timestamp = now; + for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) { + rn_list = &the_lnet.ln_remote_nets_hash[i]; + list_for_each(e1, rn_list) { + rnet = list_entry(e1, struct lnet_remotenet, lrn_list); - mdh = rcd->rcd_mdh; + list_for_each(e2, &rnet->lrn_routes) { + route = list_entry(e2, struct lnet_route, + lr_list); - if (rtr->lp_ping_deadline == 0) { - rtr->lp_ping_deadline = - cfs_time_shift(router_ping_timeout); + if (idx-- == 0) { + *net = rnet->lrn_net; + *gateway = route->lr_nid; + *hops = route->lr_hops; + *priority = route->lr_priority; + *sensitivity = route->lr_gateway-> + lp_health_sensitivity; + if (lnet_is_route_alive(route)) + *flags |= LNET_RT_ALIVE; + else + *flags &= ~LNET_RT_ALIVE; + if (route->lr_single_hop) + *flags &= ~LNET_RT_MULTI_HOP; + else + *flags |= LNET_RT_MULTI_HOP; + lnet_net_unlock(cpt); + return 0; + } + } } + } - lnet_net_unlock(rtr->lp_cpt); - - rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL, - LNET_PROTO_PING_MATCHBITS, 0); - - lnet_net_lock(rtr->lp_cpt); - if (rc != 0) - rtr->lp_ping_notsent = 0; /* no event pending */ - } - - lnet_peer_decref_locked(rtr); - return; + lnet_net_unlock(cpt); + return -ENOENT; } -int -lnet_router_checker_start(void) +static void +lnet_wait_known_routerstate(void) { - int rc; - int eqsz; -#ifdef __KERNEL__ - struct task_struct *task; -#else /* __KERNEL__ */ - lnet_peer_t *rtr; - __u64 version; - int nrtr = 0; - int router_checker_max_eqsize = 10240; - - LASSERT (check_routers_before_use); - LASSERT (dead_router_check_interval > 0); - - lnet_net_lock(0); - - /* As an approximation, allow each router the same number of - * outstanding events as it is allowed outstanding sends */ - eqsz = 0; - version = the_lnet.ln_routers_version; - list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) { - lnet_ni_t *ni = rtr->lp_ni; - lnet_process_id_t id; - - nrtr++; - eqsz += ni->ni_peertxcredits; - - /* one async ping reply per router */ - id.nid = rtr->lp_nid; - id.pid = LNET_PID_LUSTRE; + struct lnet_peer *rtr; + struct list_head *entry; + int all_known; - lnet_net_unlock(0); + LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING); - rc = LNetSetAsync(id, 1); - if (rc != 0) { - CWARN("LNetSetAsync %s failed: %d\n", - libcfs_id2str(id), rc); - return rc; - } + for (;;) { + int cpt = lnet_net_lock_current(); - lnet_net_lock(0); - /* NB router list doesn't change in userspace */ - LASSERT(version == the_lnet.ln_routers_version); - } + all_known = 1; + list_for_each(entry, &the_lnet.ln_routers) { + rtr = list_entry(entry, struct lnet_peer, + lp_rtr_list); - lnet_net_unlock(0); + spin_lock(&rtr->lp_lock); - if (nrtr == 0) { - CDEBUG(D_NET, - "No router found, not starting router checker\n"); - return 0; - } + if ((rtr->lp_state & LNET_PEER_RTR_DISCOVERED) == 0) { + all_known = 0; + spin_unlock(&rtr->lp_lock); + break; + } + spin_unlock(&rtr->lp_lock); + } - /* at least allow a SENT and a REPLY per router */ - if (router_checker_max_eqsize < 2 * nrtr) - router_checker_max_eqsize = 2 * nrtr; + lnet_net_unlock(cpt); - LASSERT (eqsz > 0); - if (eqsz > router_checker_max_eqsize) - eqsz = router_checker_max_eqsize; -#endif + if (all_known) + return; - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); - - if (check_routers_before_use && - dead_router_check_interval <= 0) { - LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be" - " set if 'check_routers_before_use' is set" - "\n"); - return -EINVAL; - } - - if (!the_lnet.ln_routing && - live_router_check_interval <= 0 && - dead_router_check_interval <= 0) - return 0; - -#ifdef __KERNEL__ - sema_init(&the_lnet.ln_rc_signal, 0); - /* EQ size doesn't matter; the callback is guaranteed to get every - * event */ - eqsz = 0; - rc = LNetEQAlloc(eqsz, lnet_router_checker_event, - &the_lnet.ln_rc_eqh); -#else - rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE, - &the_lnet.ln_rc_eqh); -#endif - if (rc != 0) { - CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc); - return -ENOMEM; - } - - the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING; -#ifdef __KERNEL__ - task = kthread_run(lnet_router_checker, NULL, "router_checker"); - if (IS_ERR(task)) { - rc = PTR_ERR(task); - CERROR("Can't start router checker thread: %d\n", rc); - /* block until event callback signals exit */ - down(&the_lnet.ln_rc_signal); - rc = LNetEQFree(the_lnet.ln_rc_eqh); - LASSERT(rc == 0); - the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN; - return -ENOMEM; + schedule_timeout_uninterruptible(cfs_time_seconds(1)); } -#endif - - if (check_routers_before_use) { - /* Note that a helpful side-effect of pinging all known routers - * at startup is that it makes them drop stale connections they - * may have to a previous instance of me. */ - lnet_wait_known_routerstate(); - } - - return 0; } -void -lnet_router_checker_stop (void) +static inline bool +lnet_net_set_status_locked(struct lnet_net *net, __u32 status) { - int rc; + struct lnet_ni *ni; + bool update = false; - if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN) - return; - - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); - the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING; - -#ifdef __KERNEL__ - /* block until event callback signals exit */ - down(&the_lnet.ln_rc_signal); -#else - lnet_router_checker(); -#endif - LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); + list_for_each_entry(ni, &net->net_ni_list, ni_netlist) + if (lnet_ni_set_status(ni, status)) + update = true; - rc = LNetEQFree(the_lnet.ln_rc_eqh); - LASSERT (rc == 0); - return; + return update; } -static void -lnet_prune_rc_data(int wait_unlink) +static bool +lnet_update_ni_status_locked(void) { - lnet_rc_data_t *rcd; - lnet_rc_data_t *tmp; - lnet_peer_t *lp; - struct list_head head; - int i = 2; - - if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING && - list_empty(&the_lnet.ln_rcd_deathrow) && - list_empty(&the_lnet.ln_rcd_zombie))) - return; - - INIT_LIST_HEAD(&head); - - lnet_net_lock(LNET_LOCK_EX); - - if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) { - /* router checker is stopping, prune all */ - list_for_each_entry(lp, &the_lnet.ln_routers, - lp_rtr_list) { - if (lp->lp_rcd == NULL) - continue; - - LASSERT(list_empty(&lp->lp_rcd->rcd_list)); - list_add(&lp->lp_rcd->rcd_list, - &the_lnet.ln_rcd_deathrow); - lp->lp_rcd = NULL; - } - } - - /* unlink all RCDs on deathrow list */ - list_splice_init(&the_lnet.ln_rcd_deathrow, &head); + struct lnet_net *net; + struct lnet_ni *ni; + bool push = false; + time64_t now; + time64_t timeout; - if (!list_empty(&head)) { - lnet_net_unlock(LNET_LOCK_EX); + LASSERT(the_lnet.ln_routing); - list_for_each_entry(rcd, &head, rcd_list) - LNetMDUnlink(rcd->rcd_mdh); + timeout = router_ping_timeout + alive_router_check_interval; - lnet_net_lock(LNET_LOCK_EX); - } + now = ktime_get_real_seconds(); + list_for_each_entry(net, &the_lnet.ln_nets, net_list) { + if (net->net_lnd->lnd_type == LOLND) + continue; - list_splice_init(&head, &the_lnet.ln_rcd_zombie); + if (now < net->net_last_alive + timeout) + goto check_ni_fatal; - /* release all zombie RCDs */ - while (!list_empty(&the_lnet.ln_rcd_zombie)) { - list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie, - rcd_list) { - if (LNetHandleIsInvalid(rcd->rcd_mdh)) - list_move(&rcd->rcd_list, &head); + spin_lock(&net->net_lock); + /* re-check with lock */ + if (now < net->net_last_alive + timeout) { + spin_unlock(&net->net_lock); + goto check_ni_fatal; } - - wait_unlink = wait_unlink && - !list_empty(&the_lnet.ln_rcd_zombie); - - lnet_net_unlock(LNET_LOCK_EX); - - while (!list_empty(&head)) { - rcd = list_entry(head.next, - lnet_rc_data_t, rcd_list); - list_del_init(&rcd->rcd_list); - lnet_destroy_rc_data(rcd); + spin_unlock(&net->net_lock); + + /* + * if the net didn't receive any traffic for past the + * timeout on any of its constituent NIs, then mark all + * the NIs down. + */ + if (lnet_net_set_status_locked(net, LNET_NI_STATUS_DOWN)) { + push = true; + continue; } - if (!wait_unlink) - return; - - i++; - CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, - "Waiting for rc buffers to unlink\n"); - cfs_pause(cfs_time_seconds(1) / 4); - - lnet_net_lock(LNET_LOCK_EX); +check_ni_fatal: + list_for_each_entry(ni, &net->net_ni_list, ni_netlist) { + /* lnet_ni_set_status() will perform the same check of + * ni_status while holding the ni lock. We can safely + * check ni_status without that lock because it is only + * written to under net_lock/EX and our caller is + * holding a net lock. + */ + if (atomic_read(&ni->ni_fatal_error_on) && + ni->ni_status && + ni->ni_status->ns_status != LNET_NI_STATUS_DOWN && + lnet_ni_set_status(ni, LNET_NI_STATUS_DOWN)) + push = true; + } } - lnet_net_unlock(LNET_LOCK_EX); + return push; } - -#if defined(__KERNEL__) && defined(LNET_ROUTER) - -static int -lnet_router_checker(void *arg) +void lnet_wait_router_start(void) { - lnet_peer_t *rtr; - struct list_head *entry; + if (check_routers_before_use) { + /* Note that a helpful side-effect of pinging all known routers + * at startup is that it makes them drop stale connections they + * may have to a previous instance of me. */ + lnet_wait_known_routerstate(); + } +} - cfs_block_allsigs(); +/* + * This function is called from the monitor thread to check if there are + * any active routers that need to be checked. + */ +bool lnet_router_checker_active(void) +{ + /* Router Checker thread needs to run when routing is enabled in + * order to call lnet_update_ni_status_locked() */ + if (the_lnet.ln_routing) + return true; - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); + return !list_empty(&the_lnet.ln_routers) && + alive_router_check_interval > 0; +} - while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) { - __u64 version; - int cpt; - int cpt2; +void +lnet_check_routers(void) +{ + struct lnet_peer_net *first_lpn = NULL; + struct lnet_peer_net *lpn; + struct lnet_peer_ni *lpni; + struct list_head *entry; + struct lnet_peer *rtr; + bool push = false; + bool found_lpn; + __u64 version; + __u32 net_id; + time64_t now; + int cpt; + int rc; - cpt = lnet_net_lock_current(); + cpt = lnet_net_lock_current(); rescan: - version = the_lnet.ln_routers_version; + version = the_lnet.ln_routers_version; - list_for_each(entry, &the_lnet.ln_routers) { - rtr = list_entry(entry, lnet_peer_t, lp_rtr_list); - - cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid); - if (cpt != cpt2) { - lnet_net_unlock(cpt); - cpt = cpt2; - lnet_net_lock(cpt); - /* the routers list has changed */ - if (version != the_lnet.ln_routers_version) - goto rescan; + list_for_each(entry, &the_lnet.ln_routers) { + rtr = list_entry(entry, struct lnet_peer, + lp_rtr_list); + + now = ktime_get_real_seconds(); + + /* + * only discover the router if we've passed + * alive_router_check_interval seconds. Some of the router + * interfaces could be down and in that case they would be + * undergoing recovery separately from this discovery. + */ + /* find next peer net which is also local */ + net_id = rtr->lp_disc_net_id; + do { + lpn = lnet_get_next_peer_net_locked(rtr, net_id); + if (!lpn) { + CERROR("gateway %s has no networks\n", + libcfs_nid2str(rtr->lp_primary_nid)); + break; } + if (first_lpn == lpn) + break; + if (!first_lpn) + first_lpn = lpn; + found_lpn = lnet_islocalnet_locked(lpn->lpn_net_id); + net_id = lpn->lpn_net_id; + } while (!found_lpn); + + if (!found_lpn || !lpn) { + CERROR("no local network found for gateway %s\n", + libcfs_nid2str(rtr->lp_primary_nid)); + continue; + } - lnet_ping_router_locked(rtr); + if (now - lpn->lpn_rtrcheck_timestamp < + alive_router_check_interval / lnet_current_net_count) + continue; + + /* + * If we're currently discovering the peer then don't + * issue another discovery + */ + spin_lock(&rtr->lp_lock); + if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY) { + spin_unlock(&rtr->lp_lock); + continue; + } + /* make sure we fully discover the router */ + rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE; + rtr->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH | + LNET_PEER_RTR_DISCOVERY; + spin_unlock(&rtr->lp_lock); + + /* find the peer_ni associated with the primary NID */ + lpni = lnet_peer_get_ni_locked(rtr, rtr->lp_primary_nid); + if (!lpni) { + CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n", + libcfs_nid2str(rtr->lp_primary_nid)); + continue; + } + lnet_peer_ni_addref_locked(lpni); - /* NB dropped lock */ - if (version != the_lnet.ln_routers_version) { - /* the routers list has changed */ - goto rescan; - } - } + /* specify the net to use */ + rtr->lp_disc_net_id = lpn->lpn_net_id; - if (the_lnet.ln_routing) - lnet_update_ni_status_locked(); + /* discover the router */ + CDEBUG(D_NET, "discover %s, cpt = %d\n", + libcfs_nid2str(lpni->lpni_nid), cpt); + rc = lnet_discover_peer_locked(lpni, cpt, false); - lnet_net_unlock(cpt); + /* decrement ref count acquired by find_peer_ni_locked() */ + lnet_peer_ni_decref_locked(lpni); - lnet_prune_rc_data(0); /* don't wait for UNLINK */ + if (!rc) + lpn->lpn_rtrcheck_timestamp = now; + else + CERROR("Failed to discover router %s\n", + libcfs_nid2str(rtr->lp_primary_nid)); - /* Call cfs_pause() here always adds 1 to load average - * because kernel counts # active tasks as nr_running - * + nr_uninterruptible. */ - schedule_timeout_and_set_state(TASK_INTERRUPTIBLE, - cfs_time_seconds(1)); + /* NB dropped lock */ + if (version != the_lnet.ln_routers_version) { + /* the routers list has changed */ + goto rescan; + } } - LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING); + if (the_lnet.ln_routing) + push = lnet_update_ni_status_locked(); - lnet_prune_rc_data(1); /* wait for UNLINK */ + lnet_net_unlock(cpt); - the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN; - up(&the_lnet.ln_rc_signal); - /* The unlink event callback will signal final completion */ - return 0; + /* if the status of the ni changed update the peers */ + if (push) + lnet_push_update_to_peers(1); } void -lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages) +lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages) { - int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); + int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]); - while (--npages >= 0) - __free_page(rb->rb_kiov[npages].kiov_page); + while (--npages >= 0) + __free_page(rb->rb_kiov[npages].bv_page); - LIBCFS_FREE(rb, sz); + LIBCFS_FREE(rb, sz); } -lnet_rtrbuf_t * -lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt) +static struct lnet_rtrbuf * +lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt) { - int npages = rbp->rbp_npages; - int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); + int npages = rbp->rbp_npages; + int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]); struct page *page; - lnet_rtrbuf_t *rb; - int i; + struct lnet_rtrbuf *rb; + int i; LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz); if (rb == NULL) @@ -1377,88 +1290,99 @@ lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt) for (i = 0; i < npages; i++) { page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, - __GFP_ZERO | GFP_IOFS); - if (page == NULL) { - while (--i >= 0) - __free_page(rb->rb_kiov[i].kiov_page); + GFP_KERNEL | __GFP_ZERO); + if (page == NULL) { + while (--i >= 0) + __free_page(rb->rb_kiov[i].bv_page); - LIBCFS_FREE(rb, sz); - return NULL; - } + LIBCFS_FREE(rb, sz); + return NULL; + } - rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE; - rb->rb_kiov[i].kiov_offset = 0; - rb->rb_kiov[i].kiov_page = page; - } + rb->rb_kiov[i].bv_len = PAGE_SIZE; + rb->rb_kiov[i].bv_offset = 0; + rb->rb_kiov[i].bv_page = page; + } - return rb; + return rb; } -void -lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt) +static void +lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt) { - int npages = rbp->rbp_npages; - lnet_rtrbuf_t *rb; - struct list_head tmp; + int npages = rbp->rbp_npages; + struct lnet_rtrbuf *rb; + LIST_HEAD(tmp); if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */ return; - INIT_LIST_HEAD(&tmp); - lnet_net_lock(cpt); - lnet_drop_routed_msgs_locked(&rbp->rbp_msgs, cpt); + list_splice_init(&rbp->rbp_msgs, &tmp); + lnet_drop_routed_msgs_locked(&tmp, cpt); list_splice_init(&rbp->rbp_bufs, &tmp); + rbp->rbp_req_nbuffers = 0; rbp->rbp_nbuffers = rbp->rbp_credits = 0; rbp->rbp_mincredits = 0; lnet_net_unlock(cpt); /* Free buffers on the free list. */ while (!list_empty(&tmp)) { - rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list); + rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list); list_del(&rb->rb_list); lnet_destroy_rtrbuf(rb, npages); } } static int -lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt) +lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt) { - struct list_head rb_list; - lnet_rtrbuf_t *rb; + LIST_HEAD(rb_list); + struct lnet_rtrbuf *rb; int num_rb; int num_buffers = 0; + int old_req_nbufs; int npages = rbp->rbp_npages; + lnet_net_lock(cpt); /* If we are called for less buffers than already in the pool, we - * just lower the nbuffers number and excess buffers will be + * just lower the req_nbuffers number and excess buffers will be * thrown away as they are returned to the free list. Credits - * then get adjusted as well. */ - if (nbufs <= rbp->rbp_nbuffers) { - lnet_net_lock(cpt); - rbp->rbp_nbuffers = nbufs; + * then get adjusted as well. + * If we already have enough buffers allocated to serve the + * increase requested, then we can treat that the same way as we + * do the decrease. */ + num_rb = nbufs - rbp->rbp_nbuffers; + if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) { + rbp->rbp_req_nbuffers = nbufs; lnet_net_unlock(cpt); return 0; } + /* store the older value of rbp_req_nbuffers and then set it to + * the new request to prevent lnet_return_rx_credits_locked() from + * freeing buffers that we need to keep around */ + old_req_nbufs = rbp->rbp_req_nbuffers; + rbp->rbp_req_nbuffers = nbufs; + lnet_net_unlock(cpt); - INIT_LIST_HEAD(&rb_list); - - /* allocate the buffers on a local list first. If all buffers are + /* allocate the buffers on a local list first. If all buffers are * allocated successfully then join this list to the rbp buffer * list. If not then free all allocated buffers. */ - num_rb = rbp->rbp_nbuffers; - - while (num_rb < nbufs) { + while (num_rb-- > 0) { rb = lnet_new_rtrbuf(rbp, cpt); if (rb == NULL) { CERROR("Failed to allocate %d route bufs of %d pages\n", nbufs, npages); + + lnet_net_lock(cpt); + rbp->rbp_req_nbuffers = old_req_nbufs; + lnet_net_unlock(cpt); + goto failed; } list_add(&rb->rb_list, &rb_list); num_buffers++; - num_rb++; } lnet_net_lock(cpt); @@ -1479,7 +1403,7 @@ lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt) failed: while (!list_empty(&rb_list)) { - rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list); + rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list); list_del(&rb->rb_list); lnet_destroy_rtrbuf(rb, npages); } @@ -1487,21 +1411,21 @@ failed: return -ENOMEM; } -void -lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages) +static void +lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages) { INIT_LIST_HEAD(&rbp->rbp_msgs); INIT_LIST_HEAD(&rbp->rbp_bufs); - rbp->rbp_npages = npages; - rbp->rbp_credits = 0; - rbp->rbp_mincredits = 0; + rbp->rbp_npages = npages; + rbp->rbp_credits = 0; + rbp->rbp_mincredits = 0; } void lnet_rtrpools_free(int keep_pools) { - lnet_rtrbufpool_t *rtrp; + struct lnet_rtrbufpool *rtrp; int i; if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */ @@ -1528,7 +1452,7 @@ lnet_nrb_tiny_calculate(void) LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when " "routing enabled\n", tiny_router_buffers); - return -1; + return -EINVAL; } if (tiny_router_buffers > 0) @@ -1547,7 +1471,7 @@ lnet_nrb_small_calculate(void) LCONSOLE_ERROR_MSG(0x10c, "small_router_buffers=%d invalid when " "routing enabled\n", small_router_buffers); - return -1; + return -EINVAL; } if (small_router_buffers > 0) @@ -1566,7 +1490,7 @@ lnet_nrb_large_calculate(void) LCONSOLE_ERROR_MSG(0x10c, "large_router_buffers=%d invalid when " "routing enabled\n", large_router_buffers); - return -1; + return -EINVAL; } if (large_router_buffers > 0) @@ -1579,7 +1503,7 @@ lnet_nrb_large_calculate(void) int lnet_rtrpools_alloc(int im_a_router) { - lnet_rtrbufpool_t *rtrp; + struct lnet_rtrbufpool *rtrp; int nrb_tiny; int nrb_small; int nrb_large; @@ -1615,7 +1539,7 @@ lnet_rtrpools_alloc(int im_a_router) the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(), LNET_NRBPOOLS * - sizeof(lnet_rtrbufpool_t)); + sizeof(struct lnet_rtrbufpool)); if (the_lnet.ln_rtrpools == NULL) { LCONSOLE_ERROR_MSG(0x10c, "Failed to initialize router buffe pool\n"); @@ -1647,6 +1571,7 @@ lnet_rtrpools_alloc(int im_a_router) lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_routing = 1; lnet_net_unlock(LNET_LOCK_EX); + complete(&the_lnet.ln_mt_wait_complete); return 0; failed: @@ -1660,7 +1585,7 @@ lnet_rtrpools_adjust_helper(int tiny, int small, int large) int nrb = 0; int rc = 0; int i; - lnet_rtrbufpool_t *rtrp; + struct lnet_rtrbufpool *rtrp; /* If the provided values for each buffer pool are different than the * configured values, we need to take action. */ @@ -1714,7 +1639,7 @@ lnet_rtrpools_adjust(int tiny, int small, int large) int lnet_rtrpools_enable(void) { - int rc; + int rc = 0; if (the_lnet.ln_routing) return 0; @@ -1725,19 +1650,24 @@ lnet_rtrpools_enable(void) * standard buffer pool allocation routine as * if we are just configuring this for the first * time. */ - return lnet_rtrpools_alloc(1); - - rc = lnet_rtrpools_adjust_helper(0, 0, 0); + rc = lnet_rtrpools_alloc(1); + else + rc = lnet_rtrpools_adjust_helper(0, 0, 0); if (rc != 0) return rc; lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_routing = 1; - the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED; + the_lnet.ln_ping_target->pb_info.pi_features &= + ~LNET_PING_FEAT_RTE_DISABLED; lnet_net_unlock(LNET_LOCK_EX); - return 0; + if (lnet_peer_discovery_disabled) + CWARN("Consider turning discovery on to enable full " + "Multi-Rail routing functionality\n"); + + return rc; } void @@ -1748,7 +1678,8 @@ lnet_rtrpools_disable(void) lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_routing = 0; - the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED; + the_lnet.ln_ping_target->pb_info.pi_features |= + LNET_PING_FEAT_RTE_DISABLED; tiny_router_buffers = 0; small_router_buffers = 0; @@ -1757,204 +1688,135 @@ lnet_rtrpools_disable(void) lnet_rtrpools_free(1); } -int -lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) -{ - struct lnet_peer *lp = NULL; - cfs_time_t now = cfs_time_current(); - int cpt = lnet_cpt_of_nid(nid); - - LASSERT (!in_interrupt ()); - - CDEBUG (D_NET, "%s notifying %s: %s\n", - (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid), - libcfs_nid2str(nid), - alive ? "up" : "down"); - - if (ni != NULL && - LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) { - CWARN ("Ignoring notification of %s %s by %s (different net)\n", - libcfs_nid2str(nid), alive ? "birth" : "death", - libcfs_nid2str(ni->ni_nid)); - return -EINVAL; - } - - /* can't do predictions... */ - if (cfs_time_after(when, now)) { - CWARN ("Ignoring prediction from %s of %s %s " - "%ld seconds in the future\n", - (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid), - libcfs_nid2str(nid), alive ? "up" : "down", - cfs_duration_sec(cfs_time_sub(when, now))); - return -EINVAL; - } - - if (ni != NULL && !alive && /* LND telling me she's down */ - !auto_down) { /* auto-down disabled */ - CDEBUG(D_NET, "Auto-down disabled\n"); - return 0; - } - - lnet_net_lock(cpt); - - if (the_lnet.ln_shutdown) { - lnet_net_unlock(cpt); - return -ESHUTDOWN; - } - - lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid); - if (lp == NULL) { - /* nid not found */ - lnet_net_unlock(cpt); - CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid)); - return 0; - } - - /* We can't fully trust LND on reporting exact peer last_alive - * if he notifies us about dead peer. For example ksocklnd can - * call us with when == _time_when_the_node_was_booted_ if - * no connections were successfully established */ - if (ni != NULL && !alive && when < lp->lp_last_alive) - when = lp->lp_last_alive; - - lnet_notify_locked(lp, ni == NULL, alive, when); - - if (ni != NULL) - lnet_ni_notify_locked(ni, lp); - - lnet_peer_decref_locked(lp); - - lnet_net_unlock(cpt); - return 0; -} -EXPORT_SYMBOL(lnet_notify); - -void -lnet_get_tunables (void) +static inline void +lnet_notify_peer_down(struct lnet_ni *ni, lnet_nid_t nid) { - return; + if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL) + (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid); } -#else - +/* + * ni: local NI used to communicate with the peer + * nid: peer NID + * alive: true if peer is alive, false otherwise + * reset: reset health value. This is requested by the LND. + * when: notificaiton time. + */ int -lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) -{ - return -EOPNOTSUPP; -} - -void -lnet_router_checker (void) +lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, bool alive, bool reset, + time64_t when) { - static time_t last = 0; - static int running = 0; - - time_t now = cfs_time_current_sec(); - int interval = now - last; - int rc; - __u64 version; - lnet_peer_t *rtr; - - /* It's no use to call me again within a sec - all intervals and - * timeouts are measured in seconds */ - if (last != 0 && interval < 2) - return; + struct lnet_peer_ni *lpni = NULL; + struct lnet_route *route; + struct lnet_peer *lp; + time64_t now = ktime_get_seconds(); + int cpt; + + LASSERT(!in_interrupt()); + + CDEBUG(D_NET, "%s notifying %s: %s\n", + (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid), + libcfs_nid2str(nid), alive ? "up" : "down"); + + if (ni != NULL && + LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) { + CWARN("Ignoring notification of %s %s by %s (different net)\n", + libcfs_nid2str(nid), alive ? "birth" : "death", + libcfs_nid2str(ni->ni_nid)); + return -EINVAL; + } - if (last != 0 && - interval > MAX(live_router_check_interval, - dead_router_check_interval)) - CNETERR("Checker(%d/%d) not called for %d seconds\n", - live_router_check_interval, dead_router_check_interval, - interval); + /* can't do predictions... */ + if (when > now) { + CWARN("Ignoring prediction from %s of %s %s " + "%lld seconds in the future\n", + (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid), + libcfs_nid2str(nid), alive ? "up" : "down", when - now); + return -EINVAL; + } - LASSERT(LNET_CPT_NUMBER == 1); + if (ni != NULL && !alive && /* LND telling me she's down */ + !auto_down) { /* auto-down disabled */ + CDEBUG(D_NET, "Auto-down disabled\n"); + return 0; + } + /* must lock 0 since this is used for synchronization */ lnet_net_lock(0); - LASSERT(!running); /* recursion check */ - running = 1; - lnet_net_unlock(0); - - last = now; - - if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) - lnet_prune_rc_data(0); /* unlink all rcd and nowait */ - /* consume all pending events */ - while (1) { - int i; - lnet_event_t ev; - - /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the - * recursion breaker in LNetEQPoll would fail */ - rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i); - if (rc == 0) /* no event pending */ - break; - - /* NB a lost SENT prevents me from pinging a router again */ - if (rc == -EOVERFLOW) { - CERROR("Dropped an event!!!\n"); - abort(); - } - - LASSERT (rc == 1); - - lnet_router_checker_event(&ev); - } - - if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) { - lnet_prune_rc_data(1); /* release rcd */ - the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN; - running = 0; - return; - } + if (the_lnet.ln_state != LNET_STATE_RUNNING) { + lnet_net_unlock(0); + return -ESHUTDOWN; + } - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); + lpni = lnet_find_peer_ni_locked(nid); + if (lpni == NULL) { + /* nid not found */ + lnet_net_unlock(0); + CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid)); + return 0; + } - lnet_net_lock(0); + if (alive) { + if (reset) { + lpni->lpni_ns_status = LNET_NI_STATUS_UP; + lnet_set_lpni_healthv_locked(lpni, + LNET_MAX_HEALTH_VALUE); + } else { + __u32 sensitivity = lpni->lpni_peer_net-> + lpn_peer->lp_health_sensitivity; + + lnet_inc_lpni_healthv_locked(lpni, + (sensitivity) ? sensitivity : + lnet_health_sensitivity); + } + } else if (reset) { + lpni->lpni_ns_status = LNET_NI_STATUS_DOWN; + } - version = the_lnet.ln_routers_version; - list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) { - lnet_ping_router_locked(rtr); - LASSERT(version == the_lnet.ln_routers_version); + /* recalculate aliveness */ + alive = lnet_is_peer_ni_alive(lpni); + + lp = lpni->lpni_peer_net->lpn_peer; + /* If this is an LNet router then update route aliveness */ + if (lp->lp_rtr_refcount) { + if (reset) + /* reset flag indicates gateway peer went up or down */ + lp->lp_alive = alive; + + /* If discovery is disabled, locally or on the gateway, then + * any routes using lpni as next-hop need to be updated + * + * NB: We can get many notifications while a route is down, so + * we try and avoid the expensive net_lock/EX here for the + * common case of receiving duplicate lnet_notify() calls (i.e. + * only grab EX lock when we actually need to update the route + * aliveness). + */ + if (lnet_is_discovery_disabled(lp)) { + list_for_each_entry(route, &lp->lp_routes, lr_gwlist) { + if (route->lr_nid == lpni->lpni_nid && + route->lr_alive != alive) { + lnet_net_unlock(0); + lnet_net_lock(LNET_LOCK_EX); + lnet_set_route_aliveness(route, alive); + lnet_net_unlock(LNET_LOCK_EX); + lnet_net_lock(0); + } + } + } } lnet_net_unlock(0); - running = 0; /* lock only needed for the recursion check */ - return; -} - -/* NB lnet_peers_start_down depends on me, - * so must be called before any peer creation */ -void -lnet_get_tunables (void) -{ - char *s; - - s = getenv("LNET_ROUTER_PING_TIMEOUT"); - if (s != NULL) router_ping_timeout = atoi(s); + if (ni != NULL && !alive) + lnet_notify_peer_down(ni, lpni->lpni_nid); - s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL"); - if (s != NULL) live_router_check_interval = atoi(s); - - s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL"); - if (s != NULL) dead_router_check_interval = atoi(s); - - /* This replaces old lnd_notify mechanism */ - check_routers_before_use = 1; - if (dead_router_check_interval <= 0) - dead_router_check_interval = 30; -} - -void -lnet_rtrpools_free(int keep_pools) -{ -} + cpt = lpni->lpni_cpt; + lnet_net_lock(cpt); + lnet_peer_ni_decref_locked(lpni); + lnet_net_unlock(cpt); -int -lnet_rtrpools_alloc(int im_a_arouter) -{ - return 0; + return 0; } - -#endif +EXPORT_SYMBOL(lnet_notify);