X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lnet%2Flnet%2Frouter.c;h=71903cf24bc7fc915c975e4317cbd3326e369d2a;hb=362e52efec1a6f6fd43b16ccd555b714a30642da;hp=01b635aa8343a74dcfa2923b5c1a883748481845;hpb=83f755685e50da38cb916ae4fb83f62429b5d568;p=fs%2Flustre-release.git diff --git a/lnet/lnet/router.c b/lnet/lnet/router.c index 01b635a..71903cf 100644 --- a/lnet/lnet/router.c +++ b/lnet/lnet/router.c @@ -1,10 +1,9 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * - * Copyright (C) 2002 Cluster File Systems, Inc. + * Copyright (c) 2011, 2016, Intel Corporation. * - * This file is part of Portals - * http://sourceforge.net/projects/sandiaportals/ + * This file is part of Lustre, https://wiki.hpdd.intel.com/ * * Portals is free software; you can redistribute it and/or * modify it under the terms of version 2 of the GNU General Public @@ -24,1053 +23,1825 @@ #define DEBUG_SUBSYSTEM S_LNET #include -#if defined(__KERNEL__) && defined(LNET_ROUTER) +#define LNET_NRB_TINY_MIN 512 /* min value for each CPT */ +#define LNET_NRB_TINY (LNET_NRB_TINY_MIN * 4) +#define LNET_NRB_SMALL_MIN 4096 /* min value for each CPT */ +#define LNET_NRB_SMALL (LNET_NRB_SMALL_MIN * 4) +#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_SIZE - 1) >> \ + PAGE_SHIFT) static char *forwarding = ""; -CFS_MODULE_PARM(forwarding, "s", charp, 0444, - "Explicitly enable/disable forwarding between networks"); - -static int tiny_router_buffers = 1024; -CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444, - "# of 0 payload messages to buffer in the router"); -static int small_router_buffers = 8192; -CFS_MODULE_PARM(small_router_buffers, "i", int, 0444, - "# of small (1 page) messages to buffer in the router"); -static int large_router_buffers = 512; -CFS_MODULE_PARM(large_router_buffers, "i", int, 0444, - "# of large messages to buffer in the router"); +module_param(forwarding, charp, 0444); +MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks"); + +static int tiny_router_buffers; +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; +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; +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"); -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"); +int +lnet_peer_buffer_credits(struct lnet_net *net) +{ + /* 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"); -static int dead_router_check_interval = 0; -CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0444, - "Seconds between dead router health checks (<= 0 to disable)"); +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)"); -static int live_router_check_interval = 0; -CFS_MODULE_PARM(live_router_check_interval, "i", int, 0444, - "Seconds between live router health checks (<= 0 to disable)"); +static int dead_router_check_interval = 60; +module_param(dead_router_check_interval, int, 0644); +MODULE_PARM_DESC(dead_router_check_interval, "Seconds between dead router health checks (<= 0 to disable)"); + +static int live_router_check_interval = 60; +module_param(live_router_check_interval, int, 0644); +MODULE_PARM_DESC(live_router_check_interval, "Seconds between live router health checks (<= 0 to disable)"); static int router_ping_timeout = 50; -CFS_MODULE_PARM(router_ping_timeout, "i", int, 0444, - "Seconds to wait for the reply to a router health query"); +module_param(router_ping_timeout, int, 0644); +MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query"); int lnet_peers_start_down(void) { - return check_routers_before_use; + return check_routers_before_use; } void -lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, time_t when) +lnet_notify_locked(struct lnet_peer_ni *lp, int notifylnd, int alive, + cfs_time_t when) { - if (when < lp->lp_timestamp) { /* out of date information */ - CDEBUG(D_NET, "Out of date\n"); - return; - } + if (cfs_time_before(when, lp->lpni_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 */ + /* + * This function can be called with different cpt locks being + * held. lpni_alive_count modification needs to be properly protected. + * Significant reads to lpni_alive_count are also protected with + * the same lock + */ + spin_lock(&lp->lpni_lock); + + lp->lpni_timestamp = when; /* update timestamp */ + lp->lpni_ping_deadline = 0; /* disable ping timeout */ + + if (lp->lpni_alive_count != 0 && /* got old news */ + (!lp->lpni_alive) == (!alive)) { /* new date for old news */ + spin_unlock(&lp->lpni_lock); + CDEBUG(D_NET, "Old news\n"); + return; + } - 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 */ - /* Flag that notification is outstanding */ + lp->lpni_alive_count++; + lp->lpni_alive = (alive) ? 1 : 0; + lp->lpni_notify = 1; + lp->lpni_notifylnd = notifylnd; + if (lp->lpni_alive) + lp->lpni_ping_feats = LNET_PING_FEAT_INVAL; /* reset */ - lp->lp_alive_count++; - lp->lp_alive = !(!alive); /* 1 bit! */ - lp->lp_notify = 1; - lp->lp_notifylnd |= notifylnd; + spin_unlock(&lp->lpni_lock); - CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive); + CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive); } -void -lnet_do_notify (lnet_peer_t *lp) +/* + * This function will always be called with lp->lpni_cpt lock held. + */ +static void +lnet_ni_notify_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp) { - lnet_ni_t *ni = lp->lp_ni; - int alive; - int notifylnd; - - LNET_LOCK(); - - /* 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 */ + int alive; + int notifylnd; - if (lp->lp_notifying) { - LNET_UNLOCK(); - return; - } + /* 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 */ - lp->lp_notifying = 1; - - while (lp->lp_notify) { - alive = lp->lp_alive; - notifylnd = lp->lp_notifylnd; + spin_lock(&lp->lpni_lock); - lp->lp_notifylnd = 0; - lp->lp_notify = 0; + if (lp->lpni_notifying || ni == NULL) { + spin_unlock(&lp->lpni_lock); + return; + } - if (notifylnd && ni->ni_lnd->lnd_notify != NULL) { - LNET_UNLOCK(); + lp->lpni_notifying = 1; - /* A new notification could happen now; I'll handle it - * when control returns to me */ + /* + * lp->lpni_notify needs to be protected because it can be set in + * lnet_notify_locked(). + */ + while (lp->lpni_notify) { + alive = lp->lpni_alive; + notifylnd = lp->lpni_notifylnd; - (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive); + lp->lpni_notifylnd = 0; + lp->lpni_notify = 0; - LNET_LOCK(); - } - } + if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) { + spin_unlock(&lp->lpni_lock); + lnet_net_unlock(lp->lpni_cpt); - lp->lp_notifying = 0; + /* A new notification could happen now; I'll handle it + * when control returns to me */ - LNET_UNLOCK(); -} + (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lpni_nid, + alive); -int -lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, time_t when) -{ - lnet_peer_t *lp = NULL; - time_t now = cfs_time_current_sec(); - - 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 (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", - 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_LOCK(); - - lp = lnet_find_peer_locked(nid); - if (lp == NULL) { - /* nid not found */ - LNET_UNLOCK(); - CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid)); - return 0; - } - - lnet_notify_locked(lp, ni == NULL, alive, when); - - LNET_UNLOCK(); - - lnet_do_notify(lp); - - LNET_LOCK(); - - lnet_peer_decref_locked(lp); - - LNET_UNLOCK(); - return 0; -} -EXPORT_SYMBOL(lnet_notify); - -#else + lnet_net_lock(lp->lpni_cpt); + spin_lock(&lp->lpni_lock); + } + } -int -lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, time_t when) -{ - return -EOPNOTSUPP; + lp->lpni_notifying = 0; + spin_unlock(&lp->lpni_lock); } -#endif - static void -lnet_rtr_addref_locked(lnet_peer_t *lp) +lnet_rtr_addref_locked(struct lnet_peer_ni *lp) { - LASSERT (lp->lp_refcount > 0); - LASSERT (lp->lp_rtr_refcount >= 0); - - 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); - /* addref for the_lnet.ln_routers */ - lnet_peer_addref_locked(lp); - the_lnet.ln_routers_version++; - } + LASSERT(atomic_read(&lp->lpni_refcount) > 0); + LASSERT(lp->lpni_rtr_refcount >= 0); + + /* lnet_net_lock must be exclusively locked */ + lp->lpni_rtr_refcount++; + if (lp->lpni_rtr_refcount == 1) { + struct list_head *pos; + + /* a simple insertion sort */ + list_for_each_prev(pos, &the_lnet.ln_routers) { + struct lnet_peer_ni *rtr = + list_entry(pos, struct lnet_peer_ni, + lpni_rtr_list); + + if (rtr->lpni_nid < lp->lpni_nid) + break; + } + + list_add(&lp->lpni_rtr_list, pos); + /* addref for the_lnet.ln_routers */ + lnet_peer_ni_addref_locked(lp); + the_lnet.ln_routers_version++; + } } static void -lnet_rtr_decref_locked(lnet_peer_t *lp) +lnet_rtr_decref_locked(struct lnet_peer_ni *lp) { - LASSERT (lp->lp_refcount > 0); - LASSERT (lp->lp_rtr_refcount > 0); - - lp->lp_rtr_refcount--; - if (lp->lp_rtr_refcount == 0) { - list_del(&lp->lp_rtr_list); - /* decref for the_lnet.ln_routers */ - lnet_peer_decref_locked(lp); - the_lnet.ln_routers_version++; - } + LASSERT(atomic_read(&lp->lpni_refcount) > 0); + LASSERT(lp->lpni_rtr_refcount > 0); + + /* lnet_net_lock must be exclusively locked */ + lp->lpni_rtr_refcount--; + if (lp->lpni_rtr_refcount == 0) { + LASSERT(list_empty(&lp->lpni_routes)); + + if (lp->lpni_rcd != NULL) { + list_add(&lp->lpni_rcd->rcd_list, + &the_lnet.ln_rcd_deathrow); + lp->lpni_rcd = NULL; + } + + list_del(&lp->lpni_rtr_list); + /* decref for the_lnet.ln_routers */ + lnet_peer_ni_decref_locked(lp); + the_lnet.ln_routers_version++; + } } lnet_remotenet_t * -lnet_find_net_locked (__u32 net) +lnet_find_rnet_locked(__u32 net) { - lnet_remotenet_t *rnet; - struct list_head *tmp; + lnet_remotenet_t *rnet; + struct list_head *tmp; + struct list_head *rn_list; - LASSERT (!the_lnet.ln_shutdown); + LASSERT(!the_lnet.ln_shutdown); - list_for_each (tmp, &the_lnet.ln_remote_nets) { - rnet = list_entry(tmp, lnet_remotenet_t, lrn_list); + rn_list = lnet_net2rnethash(net); + list_for_each(tmp, rn_list) { + rnet = list_entry(tmp, lnet_remotenet_t, lrn_list); - if (rnet->lrn_net == net) - return rnet; - } - return NULL; + if (rnet->lrn_net == net) + return rnet; + } + return NULL; } -int -lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway) +static void lnet_shuffle_seed(void) { - struct list_head zombies; - struct list_head *e; - lnet_remotenet_t *rnet; - lnet_remotenet_t *rnet2; - lnet_route_t *route; - lnet_route_t *route2; - lnet_ni_t *ni; - int add_route; - int rc; - - CDEBUG(D_NET, "Add route: net %s hops %u gw %s\n", - libcfs_net2str(net), hops, libcfs_nid2str(gateway)); - - if (gateway == LNET_NID_ANY || - LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND || - net == LNET_NIDNET(LNET_NID_ANY) || - LNET_NETTYP(net) == LOLND || - LNET_NIDNET(gateway) == net || - hops < 1 || hops > 255) - return (-EINVAL); - - if (lnet_islocalnet(net)) /* it's a local network */ - return 0; /* ignore the route entry */ - - /* Assume net, route, all new */ - LIBCFS_ALLOC(route, sizeof(*route)); - LIBCFS_ALLOC(rnet, sizeof(*rnet)); - if (route == NULL || rnet == NULL) { - CERROR("Out of memory creating route %s %d %s\n", - libcfs_net2str(net), hops, libcfs_nid2str(gateway)); - if (route != NULL) - LIBCFS_FREE(route, sizeof(*route)); - if (rnet != NULL) - LIBCFS_FREE(rnet, sizeof(*rnet)); - return -ENOMEM; - } - - CFS_INIT_LIST_HEAD(&rnet->lrn_routes); - rnet->lrn_net = net; - rnet->lrn_hops = hops; - - LNET_LOCK(); - - rc = lnet_nid2peer_locked(&route->lr_gateway, gateway); - if (rc != 0) { - LNET_UNLOCK(); - - LIBCFS_FREE(route, sizeof(*route)); - LIBCFS_FREE(rnet, sizeof(*rnet)); - - if (rc == -EHOSTUNREACH) /* gateway is not on a local net */ - return 0; /* ignore the route entry */ - - CERROR("Error %d creating route %s %d %s\n", rc, - libcfs_net2str(net), hops, libcfs_nid2str(gateway)); - return rc; - } - - LASSERT (!the_lnet.ln_shutdown); - CFS_INIT_LIST_HEAD(&zombies); - - rnet2 = lnet_find_net_locked(net); - if (rnet2 == NULL) { - /* new network */ - list_add_tail(&rnet->lrn_list, &the_lnet.ln_remote_nets); - rnet2 = rnet; - } - - if (hops > rnet2->lrn_hops) { - /* New route is longer; ignore it */ - add_route = 0; - } else if (hops < rnet2->lrn_hops) { - /* new route supercedes all currently known routes to this - * net */ - list_add(&zombies, &rnet2->lrn_routes); - list_del_init(&rnet2->lrn_routes); - add_route = 1; - } else { - add_route = 1; - /* New route has the same hopcount as existing routes; search - * for a duplicate route (it's a NOOP if it is) */ - list_for_each (e, &rnet2->lrn_routes) { - route2 = list_entry(e, lnet_route_t, lr_list); - - if (route2->lr_gateway == route->lr_gateway) { - add_route = 0; - break; - } - - /* our loopups must be true */ - LASSERT (route2->lr_gateway->lp_nid != gateway); - } - } - - if (add_route) { - ni = route->lr_gateway->lp_ni; - lnet_ni_addref_locked(ni); - - LASSERT (rc == 0); - list_add_tail(&route->lr_list, &rnet2->lrn_routes); - the_lnet.ln_remote_nets_version++; - - lnet_rtr_addref_locked(route->lr_gateway); - - LNET_UNLOCK(); - - /* XXX Assume alive */ - if (ni->ni_lnd->lnd_notify != NULL) - (ni->ni_lnd->lnd_notify)(ni, gateway, 1); - - lnet_ni_decref(ni); - } else { - lnet_peer_decref_locked(route->lr_gateway); - LNET_UNLOCK(); - LIBCFS_FREE(route, sizeof(*route)); - } - - if (rnet != rnet2) - LIBCFS_FREE(rnet, sizeof(*rnet)); - - while (!list_empty(&zombies)) { - route = list_entry(zombies.next, lnet_route_t, lr_list); - list_del(&route->lr_list); - - LNET_LOCK(); - lnet_rtr_decref_locked(route->lr_gateway); - lnet_peer_decref_locked(route->lr_gateway); - LNET_UNLOCK(); - LIBCFS_FREE(route, sizeof(*route)); - } - - return rc; + static int seeded; + __u32 lnd_type; + __u32 seed[2]; + struct timespec64 ts; + lnet_ni_t *ni = NULL; + + if (seeded) + return; + + cfs_get_random_bytes(seed, sizeof(seed)); + + /* Nodes with small feet have little entropy + * the NID for this node gives the most entropy in the low bits */ + while ((ni = lnet_get_next_ni_locked(NULL, ni))) { + lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid)); + + if (lnd_type != LOLND) + seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type); + } + + ktime_get_ts64(&ts); + cfs_srand(ts.tv_sec ^ seed[0], ts.tv_nsec ^ seed[1]); + seeded = 1; + return; } -int -lnet_check_routes (void) +/* NB expects LNET_LOCK held */ +static void +lnet_add_route_to_rnet(lnet_remotenet_t *rnet, lnet_route_t *route) { - lnet_remotenet_t *rnet; - lnet_route_t *route; - lnet_route_t *route2; - struct list_head *e1; - struct list_head *e2; - - LNET_LOCK(); - - list_for_each (e1, &the_lnet.ln_remote_nets) { - rnet = list_entry(e1, lnet_remotenet_t, lrn_list); - - route2 = NULL; - list_for_each (e2, &rnet->lrn_routes) { - route = list_entry(e2, lnet_route_t, lr_list); - - if (route2 == NULL) - route2 = route; - else if (route->lr_gateway->lp_ni != - route2->lr_gateway->lp_ni) { - LNET_UNLOCK(); - - CERROR("Routes to %s via %s and %s not supported\n", - libcfs_net2str(rnet->lrn_net), - libcfs_nid2str(route->lr_gateway->lp_nid), - libcfs_nid2str(route2->lr_gateway->lp_nid)); - return -EINVAL; - } - } - } - - LNET_UNLOCK(); - return 0; + unsigned int len = 0; + unsigned int offset = 0; + struct list_head *e; + + lnet_shuffle_seed(); + + 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); + 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->lpni_routes); + + the_lnet.ln_remote_nets_version++; + lnet_rtr_addref_locked(route->lr_gateway); } int -lnet_del_route (__u32 net, lnet_nid_t gw_nid) +lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway, + unsigned int priority) { - lnet_remotenet_t *rnet; - lnet_route_t *route; - struct list_head *e1; - struct list_head *e2; - int rc = -ENOENT; + struct list_head *e; + lnet_remotenet_t *rnet; + lnet_remotenet_t *rnet2; + lnet_route_t *route; + lnet_ni_t *ni; + struct lnet_peer_ni *lpni; + int add_route; + int rc; + + CDEBUG(D_NET, "Add route: 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) || + LNET_NETTYP(net) == LOLND || + LNET_NIDNET(gateway) == net || + (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255))) + return -EINVAL; + + if (lnet_islocalnet(net)) /* it's a local network */ + return -EEXIST; + + /* Assume net, route, all new */ + LIBCFS_ALLOC(route, sizeof(*route)); + LIBCFS_ALLOC(rnet, sizeof(*rnet)); + if (route == NULL || rnet == NULL) { + CERROR("Out of memory creating route %s %d %s\n", + libcfs_net2str(net), hops, libcfs_nid2str(gateway)); + if (route != NULL) + LIBCFS_FREE(route, sizeof(*route)); + if (rnet != NULL) + LIBCFS_FREE(rnet, sizeof(*rnet)); + return -ENOMEM; + } + + INIT_LIST_HEAD(&rnet->lrn_routes); + rnet->lrn_net = net; + route->lr_hops = hops; + route->lr_net = net; + route->lr_priority = priority; + + lnet_net_lock(LNET_LOCK_EX); + + lpni = lnet_nid2peerni_locked(gateway, LNET_LOCK_EX); + if (IS_ERR(lpni)) { + lnet_net_unlock(LNET_LOCK_EX); + + LIBCFS_FREE(route, sizeof(*route)); + LIBCFS_FREE(rnet, sizeof(*rnet)); + + rc = PTR_ERR(lpni); + if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */ + return rc; /* ignore the route entry */ + CERROR("Error %d creating route %s %d %s\n", rc, + libcfs_net2str(net), hops, + libcfs_nid2str(gateway)); + return rc; + } + route->lr_gateway = lpni; + LASSERT(!the_lnet.ln_shutdown); + + rnet2 = lnet_find_rnet_locked(net); + if (rnet2 == NULL) { + /* new network */ + list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net)); + rnet2 = rnet; + } - CDEBUG(D_NET, "Del route: net %s : gw %s\n", - libcfs_net2str(net), libcfs_nid2str(gw_nid)); + /* 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); - /* NB Caller may specify either all routes via the given gateway - * or a specific route entry actual NIDs) */ + if (route2->lr_gateway == route->lr_gateway) { + add_route = 0; + break; + } - again: - LNET_LOCK(); + /* our lookups must be true */ + LASSERT(route2->lr_gateway->lpni_nid != gateway); + } - list_for_each (e1, &the_lnet.ln_remote_nets) { - rnet = list_entry(e1, lnet_remotenet_t, lrn_list); + if (add_route) { + lnet_peer_ni_addref_locked(route->lr_gateway); /* +1 for notify */ + lnet_add_route_to_rnet(rnet2, route); - if (!(net == LNET_NIDNET(LNET_NID_ANY) || - net == rnet->lrn_net)) - continue; + ni = lnet_get_next_ni_locked(route->lr_gateway->lpni_net, NULL); + lnet_net_unlock(LNET_LOCK_EX); - list_for_each (e2, &rnet->lrn_routes) { - route = list_entry(e2, lnet_route_t, lr_list); + /* XXX Assume alive */ + if (ni->ni_net->net_lnd->lnd_notify != NULL) + (ni->ni_net->net_lnd->lnd_notify)(ni, gateway, 1); + + lnet_net_lock(LNET_LOCK_EX); + } + + /* -1 for notify or !add_route */ + lnet_peer_ni_decref_locked(route->lr_gateway); + lnet_net_unlock(LNET_LOCK_EX); + + rc = 0; + + if (!add_route) { + rc = -EEXIST; + LIBCFS_FREE(route, sizeof(*route)); + } - if (!(gw_nid == LNET_NID_ANY || - gw_nid == route->lr_gateway->lp_nid)) - continue; + if (rnet != rnet2) + LIBCFS_FREE(rnet, sizeof(*rnet)); - list_del(&route->lr_list); - the_lnet.ln_remote_nets_version++; + /* indicate to startup the router checker if configured */ + wake_up(&the_lnet.ln_rc_waitq); - if (list_empty(&rnet->lrn_routes)) - list_del(&rnet->lrn_list); - else - rnet = NULL; + return rc; +} - lnet_rtr_decref_locked(route->lr_gateway); - lnet_peer_decref_locked(route->lr_gateway); - LNET_UNLOCK(); +int +lnet_check_routes(void) +{ + 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->lpni_net == + route2->lr_gateway->lpni_net) + continue; + + nid1 = route->lr_gateway->lpni_nid; + nid2 = route2->lr_gateway->lpni_nid; + net = rnet->lrn_net; + + lnet_net_unlock(cpt); + + CERROR("Routes to %s via %s and %s not " + "supported\n", + libcfs_net2str(net), + libcfs_nid2str(nid1), + libcfs_nid2str(nid2)); + return -EINVAL; + } + } + } - LIBCFS_FREE(route, sizeof (*route)); + lnet_net_unlock(cpt); + return 0; +} - if (rnet != NULL) - LIBCFS_FREE(rnet, sizeof(*rnet)); +int +lnet_del_route(__u32 net, lnet_nid_t gw_nid) +{ + struct lnet_peer_ni *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; + + CDEBUG(D_NET, "Del route: net %s : gw %s\n", + libcfs_net2str(net), libcfs_nid2str(gw_nid)); + + /* NB Caller may specify either all routes via the given gateway + * 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->lpni_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_ni_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; + } + } - rc = 0; - goto again; - } - } + 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); - LNET_UNLOCK(); - return rc; + return rc; } void lnet_destroy_routes (void) { - lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY); + 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, 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); + cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) { + if (i++ != idx) + continue; + + 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; + rc = 0; + break; + } + lnet_net_unlock(LNET_LOCK_EX); + } + + 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) +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->lpni_nid; + *alive = lnet_is_route_alive(route); + lnet_net_unlock(cpt); + return 0; + } + } + } + } + + lnet_net_unlock(cpt); + return -ENOENT; +} + +void +lnet_swap_pinginfo(struct lnet_ping_info *info) { - struct list_head *e1; - struct list_head *e2; - lnet_remotenet_t *rnet; - lnet_route_t *route; - - LNET_LOCK(); - - list_for_each (e1, &the_lnet.ln_remote_nets) { - 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 = rnet->lrn_hops; - *gateway = route->lr_gateway->lp_nid; - *alive = route->lr_gateway->lp_alive; - LNET_UNLOCK(); - return 0; - } - } - } - - LNET_UNLOCK(); - return -ENOENT; + int i; + struct lnet_ni_status *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; } -#if defined(__KERNEL__) && defined(LNET_ROUTER) +/** + * parse router-checker pinginfo, record number of down NIs for remote + * networks on that router. + */ static void -lnet_router_checker_event (lnet_event_t *event) +lnet_parse_rc_info(lnet_rc_data_t *rcd) { - /* CAVEAT EMPTOR: I'm called with LNET_LOCKed and I'm not allowed to - * drop it (that's how come I see _every_ event, even ones that would - * overflow my EQ) */ - lnet_peer_t *lp; - lnet_nid_t nid; - - if (event->unlinked) { - /* The router checker thread has unlinked the rc_md - * and exited. */ - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING); - the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKED; - mutex_up(&the_lnet.ln_rc_signal); - return; - } - - LASSERT (event->type == LNET_EVENT_SEND || - event->type == LNET_EVENT_REPLY); - - nid = (event->type == LNET_EVENT_SEND) ? - event->target.nid : event->initiator.nid; - - lp = lnet_find_peer_locked(nid); - if (lp == NULL) { - /* router may have been removed */ - CDEBUG(D_NET, "Router %s not found\n", libcfs_nid2str(nid)); - return; - } - - if (event->type == LNET_EVENT_SEND) /* re-enable another ping */ - lp->lp_ping_notsent = 0; - - if (lnet_isrouter(lp) && /* ignore if no longer a router */ - (event->status != 0 || - event->type == 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_sec()); - - /* 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!!! */ - } - - /* This decref will NOT drop LNET_LOCK (it had to have 1 ref when it - * was in the peer table and I've not dropped the lock, so no-one else - * can have reduced the refcount) */ - LASSERT(lp->lp_refcount > 1); - - lnet_peer_decref_locked(lp); + struct lnet_ping_info *info = rcd->rcd_pinginfo; + struct lnet_peer_ni *gw = rcd->rcd_gateway; + lnet_route_t *rte; + + if (!gw->lpni_alive) + return; + + /* + * Protect gw->lpni_ping_feats. This can be set from + * lnet_notify_locked with different locks being held + */ + spin_lock(&gw->lpni_lock); + + 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->lpni_nid), info->pi_magic); + gw->lpni_ping_feats = LNET_PING_FEAT_INVAL; + spin_unlock(&gw->lpni_lock); + return; + } + + gw->lpni_ping_feats = info->pi_features; + if ((gw->lpni_ping_feats & LNET_PING_FEAT_MASK) == 0) { + CDEBUG(D_NET, "%s: Unexpected features 0x%x\n", + libcfs_nid2str(gw->lpni_nid), gw->lpni_ping_feats); + spin_unlock(&gw->lpni_lock); + return; /* nothing I can understand */ + } + + if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0) { + spin_unlock(&gw->lpni_lock); + return; /* can't carry NI status info */ + } + + list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) { + int down = 0; + int up = 0; + int i; + + if ((gw->lpni_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++) { + struct lnet_ni_status *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->lpni_nid)); + gw->lpni_ping_feats = LNET_PING_FEAT_INVAL; + spin_unlock(&gw->lpni_lock); + 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->lpni_nid), stat->ns_status); + gw->lpni_ping_feats = LNET_PING_FEAT_INVAL; + spin_unlock(&gw->lpni_lock); + return; + } + + if (up) { /* ignore downed NIs if NI for dest network is up */ + rte->lr_downis = 0; + continue; + } + /* if @down is zero and this route is single-hop, it means + * we can't find NI for target network */ + if (down == 0 && rte->lr_hops == 1) + down = 1; + + rte->lr_downis = down; + } + + spin_unlock(&gw->lpni_lock); } -static int -lnet_router_checker(void *arg) +static void +lnet_router_checker_event(lnet_event_t *event) { - static lnet_ping_info_t pinginfo; + lnet_rc_data_t *rcd = event->md.user_ptr; + struct lnet_peer_ni *lp; - int rc; - lnet_handle_md_t mdh; - lnet_peer_t *rtr; - struct list_head *entry; - time_t now; - lnet_process_id_t rtr_id; - int secs; + LASSERT(rcd != NULL); - cfs_daemonize("router_checker"); - cfs_block_allsigs(); + if (event->unlinked) { + LNetInvalidateHandle(&rcd->rcd_mdh); + return; + } - rtr_id.pid = LUSTRE_SRV_LNET_PID; + LASSERT(event->type == LNET_EVENT_SEND || + event->type == LNET_EVENT_REPLY); - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); + lp = rcd->rcd_gateway; + LASSERT(lp != NULL); - rc = LNetMDBind((lnet_md_t){.start = &pinginfo, - .length = sizeof(pinginfo), - .threshold = LNET_MD_THRESH_INF, - .options = LNET_MD_TRUNCATE, - .eq_handle = the_lnet.ln_rc_eqh}, - LNET_UNLINK, - &mdh); + /* 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->lpni_cpt); + if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) { + /* ignore if no longer a router or rcd is replaced */ + goto out; + } - if (rc < 0) { - CERROR("Can't bind MD: %d\n", rc); - the_lnet.ln_rc_state = rc; - mutex_up(&the_lnet.ln_rc_signal); - return rc; - } + if (event->type == LNET_EVENT_SEND) { + lp->lpni_ping_notsent = 0; + if (event->status == 0) + goto out; + } - LASSERT (rc == 0); + /* 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). */ - the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING; - mutex_up(&the_lnet.ln_rc_signal); /* let my parent go */ + 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!!! */ - while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) { - __u64 version; + if (avoid_asym_router_failure && event->status == 0) + lnet_parse_rc_info(rcd); - LNET_LOCK(); -rescan: - version = the_lnet.ln_routers_version; + out: + lnet_net_unlock(lp->lpni_cpt); +} - list_for_each (entry, &the_lnet.ln_routers) { - rtr = list_entry(entry, lnet_peer_t, lp_rtr_list); +static void +lnet_wait_known_routerstate(void) +{ + struct lnet_peer_ni *rtr; + struct list_head *entry; + int all_known; - lnet_peer_addref_locked(rtr); + LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); - now = cfs_time_current_sec(); + for (;;) { + int cpt = lnet_net_lock_current(); - if (rtr->lp_ping_deadline != 0 && /* ping timed out? */ - now > rtr->lp_ping_deadline) - lnet_notify_locked(rtr, 1, 0, now); + all_known = 1; + list_for_each(entry, &the_lnet.ln_routers) { + rtr = list_entry(entry, struct lnet_peer_ni, + lpni_rtr_list); - LNET_UNLOCK(); + spin_lock(&rtr->lpni_lock); - /* Run any outstanding notificiations */ - lnet_do_notify(rtr); + if (rtr->lpni_alive_count == 0) { + all_known = 0; + spin_unlock(&rtr->lpni_lock); + break; + } + spin_unlock(&rtr->lpni_lock); + } - if (rtr->lp_alive) { - secs = live_router_check_interval; - } else { - secs = dead_router_check_interval; - } - if (secs <= 0) - secs = 0; - - if (secs != 0 && - !rtr->lp_ping_notsent && - now > rtr->lp_ping_timestamp + secs) { - CDEBUG(D_NET, "Check: %s\n", - libcfs_nid2str(rtr->lp_nid)); + lnet_net_unlock(cpt); - LNET_LOCK(); - rtr_id.nid = rtr->lp_nid; - rtr->lp_ping_notsent = 1; - rtr->lp_ping_timestamp = now; + if (all_known) + return; - if (rtr->lp_ping_deadline == 0) - rtr->lp_ping_deadline = - now + router_ping_timeout; + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1)); + } +} - LNET_UNLOCK(); +void +lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net) +{ + lnet_route_t *rte; + + if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) { + list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) { + if (rte->lr_net == net) { + rte->lr_downis = 0; + break; + } + } + } +} - LNetGet(LNET_NID_ANY, mdh, rtr_id, - LNET_RESERVED_PORTAL, - LNET_PROTO_PING_MATCHBITS, 0); - } - - LNET_LOCK(); - lnet_peer_decref_locked(rtr); +static void +lnet_update_ni_status_locked(void) +{ + lnet_ni_t *ni = NULL; + time64_t now; + int timeout; + + LASSERT(the_lnet.ln_routing); + + timeout = router_ping_timeout + + MAX(live_router_check_interval, dead_router_check_interval); + + now = ktime_get_real_seconds(); + while ((ni = lnet_get_next_ni_locked(NULL, ni))) { + if (ni->ni_net->net_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; + } + lnet_ni_unlock(ni); + } +} - if (version != the_lnet.ln_routers_version) { - /* the routers list has changed */ - goto rescan; - } - } +static void +lnet_destroy_rc_data(lnet_rc_data_t *rcd) +{ + LASSERT(list_empty(&rcd->rcd_list)); + /* detached from network */ + LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh)); - LNET_UNLOCK(); + if (rcd->rcd_gateway != NULL) { + int cpt = rcd->rcd_gateway->lpni_cpt; - /* Call cfs_pause() here always adds 1 to load average - * because kernel counts # active tasks as nr_running - * + nr_uninterruptible. */ - set_current_state(CFS_TASK_INTERRUPTIBLE); - cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE, - cfs_time_seconds(1)); + lnet_net_lock(cpt); + lnet_peer_ni_decref_locked(rcd->rcd_gateway); + lnet_net_unlock(cpt); } - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD); - the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING; - - rc = LNetMDUnlink(mdh); - LASSERT (rc == 0); - - /* The unlink event callback will signal final completion */ + if (rcd->rcd_pinginfo != NULL) + LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE); - return 0; + LIBCFS_FREE(rcd, sizeof(*rcd)); } +static lnet_rc_data_t * +lnet_create_rc_data_locked(struct lnet_peer_ni *gateway) +{ + lnet_rc_data_t *rcd = NULL; + struct lnet_ping_info *pi; + int rc; + int i; -void -lnet_wait_known_routerstate(void) + lnet_net_unlock(gateway->lpni_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->lpni_cpt); + /* router table changed or someone has created rcd for this gateway */ + if (!lnet_isrouter(gateway) || gateway->lpni_rcd != NULL) { + lnet_net_unlock(gateway->lpni_cpt); + goto out; + } + + lnet_peer_ni_addref_locked(gateway); + rcd->rcd_gateway = gateway; + gateway->lpni_rcd = rcd; + gateway->lpni_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); + } + + lnet_net_lock(gateway->lpni_cpt); + return gateway->lpni_rcd; +} + +static int +lnet_router_check_interval (struct lnet_peer_ni *rtr) { - lnet_peer_t *rtr; - struct list_head *entry; - int all_known; - - for (;;) { - LNET_LOCK(); - - 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_UNLOCK(); - - if (all_known) - return; - - cfs_pause(cfs_time_seconds(1)); - } + int secs; + + secs = rtr->lpni_alive ? live_router_check_interval : + dead_router_check_interval; + if (secs < 0) + secs = 0; + + return secs; } -void -lnet_router_checker_stop(void) +static void +lnet_ping_router_locked (struct lnet_peer_ni *rtr) { - int rc; + lnet_rc_data_t *rcd = NULL; + cfs_time_t now = cfs_time_current(); + int secs; + struct lnet_ni *ni; + + lnet_peer_ni_addref_locked(rtr); + + if (rtr->lpni_ping_deadline != 0 && /* ping timed out? */ + cfs_time_after(now, rtr->lpni_ping_deadline)) + lnet_notify_locked(rtr, 1, 0, now); + + /* Run any outstanding notifications */ + ni = lnet_get_next_ni_locked(rtr->lpni_net, NULL); + lnet_ni_notify_locked(ni, rtr); + + if (!lnet_isrouter(rtr) || + the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) { + /* router table changed or router checker is shutting down */ + lnet_peer_ni_decref_locked(rtr); + return; + } - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING || - the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); + rcd = rtr->lpni_rcd != NULL ? + rtr->lpni_rcd : lnet_create_rc_data_locked(rtr); - if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN) - return; + if (rcd == NULL) + return; - the_lnet.ln_rc_state = LNET_RC_STATE_STOPTHREAD; - /* block until event callback signals exit */ - mutex_down(&the_lnet.ln_rc_signal); + secs = lnet_router_check_interval(rtr); + + CDEBUG(D_NET, + "rtr %s %d: deadline %lu ping_notsent %d alive %d " + "alive_count %d lpni_ping_timestamp %lu\n", + libcfs_nid2str(rtr->lpni_nid), secs, + rtr->lpni_ping_deadline, rtr->lpni_ping_notsent, + rtr->lpni_alive, rtr->lpni_alive_count, rtr->lpni_ping_timestamp); + + if (secs != 0 && !rtr->lpni_ping_notsent && + cfs_time_after(now, cfs_time_add(rtr->lpni_ping_timestamp, + cfs_time_seconds(secs)))) { + int rc; + lnet_process_id_t id; + lnet_handle_md_t mdh; + + id.nid = rtr->lpni_nid; + id.pid = LNET_PID_LUSTRE; + CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id)); + + rtr->lpni_ping_notsent = 1; + rtr->lpni_ping_timestamp = now; + + mdh = rcd->rcd_mdh; + + if (rtr->lpni_ping_deadline == 0) { + rtr->lpni_ping_deadline = + cfs_time_shift(router_ping_timeout); + } - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED); + lnet_net_unlock(rtr->lpni_cpt); - rc = LNetEQFree(the_lnet.ln_rc_eqh); - LASSERT (rc == 0); - - the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN; + rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL, + LNET_PROTO_PING_MATCHBITS, 0); + + lnet_net_lock(rtr->lpni_cpt); + if (rc != 0) + rtr->lpni_ping_notsent = 0; /* no event pending */ + } + + lnet_peer_ni_decref_locked(rtr); + return; } int lnet_router_checker_start(void) { - int rc; - - 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 (live_router_check_interval <= 0 && - dead_router_check_interval <= 0) - return 0; - - init_mutex_locked(&the_lnet.ln_rc_signal); - - /* EQ size doesn't matter; the callback is guaranteed to get every - * event */ - rc = LNetEQAlloc(1, lnet_router_checker_event, - &the_lnet.ln_rc_eqh); - if (rc != 0) { - CERROR("Can't allocate EQ: %d\n", rc); - return -ENOMEM; - } - - rc = (int)cfs_kernel_thread(lnet_router_checker, NULL, 0); - if (rc < 0) { + int rc; + int eqsz = 0; + struct task_struct *task; + + 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; + } + + sema_init(&the_lnet.ln_rc_signal, 0); + + rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh); + if (rc != 0) { + CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc); + return -ENOMEM; + } + + the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING; + 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); - goto failed; - } - - mutex_down(&the_lnet.ln_rc_signal); /* wait for checker to startup */ - - rc = the_lnet.ln_rc_state; - if (rc < 0) { - the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN; - goto failed; - } - - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); - - 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; - - failed: - rc = LNetEQFree(the_lnet.ln_rc_eqh); - LASSERT (rc == 0); - return 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; + } + + 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_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages) +lnet_router_checker_stop (void) { - int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); + int rc; + + if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN) + return; - while (--npages >= 0) - cfs_free_page(rb->rb_kiov[npages].kiov_page); + LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); + the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING; + /* wakeup the RC thread if it's sleeping */ + wake_up(&the_lnet.ln_rc_waitq); + + /* block until event callback signals exit */ + down(&the_lnet.ln_rc_signal); + LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); - LIBCFS_FREE(rb, sz); + rc = LNetEQFree(the_lnet.ln_rc_eqh); + LASSERT(rc == 0); + return; } -lnet_rtrbuf_t * -lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp) +static void +lnet_prune_rc_data(int wait_unlink) { - int npages = rbp->rbp_npages; - int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); - struct page *page; - lnet_rtrbuf_t *rb; - int i; - - LIBCFS_ALLOC(rb, sz); - if (rb == NULL) - return NULL; - - rb->rb_pool = rbp; - - for (i = 0; i < npages; i++) { - page = cfs_alloc_page(CFS_ALLOC_ZERO | CFS_ALLOC_STD); - if (page == NULL) { - while (--i >= 0) - cfs_free_page(rb->rb_kiov[i].kiov_page); - - LIBCFS_FREE(rb, sz); - return NULL; - } - - rb->rb_kiov[i].kiov_len = CFS_PAGE_SIZE; - rb->rb_kiov[i].kiov_offset = 0; - rb->rb_kiov[i].kiov_page = page; - } - - return rb; + lnet_rc_data_t *rcd; + lnet_rc_data_t *tmp; + struct lnet_peer_ni *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, + lpni_rtr_list) { + if (lp->lpni_rcd == NULL) + continue; + + LASSERT(list_empty(&lp->lpni_rcd->rcd_list)); + list_add(&lp->lpni_rcd->rcd_list, + &the_lnet.ln_rcd_deathrow); + lp->lpni_rcd = NULL; + } + } + + /* unlink all RCDs on deathrow list */ + list_splice_init(&the_lnet.ln_rcd_deathrow, &head); + + if (!list_empty(&head)) { + lnet_net_unlock(LNET_LOCK_EX); + + list_for_each_entry(rcd, &head, rcd_list) + LNetMDUnlink(rcd->rcd_mdh); + + lnet_net_lock(LNET_LOCK_EX); + } + + list_splice_init(&head, &the_lnet.ln_rcd_zombie); + + /* 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); + } + + 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); + } + + if (!wait_unlink) + return; + + i++; + CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, + "Waiting for rc buffers to unlink\n"); + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1) / 4); + + lnet_net_lock(LNET_LOCK_EX); + } + + lnet_net_unlock(LNET_LOCK_EX); +} + +/* + * This function is called to check if the RC should block indefinitely. + * It's called from lnet_router_checker() as well as being passed to + * wait_event_interruptible() to avoid the lost wake_up problem. + * + * When it's called from wait_event_interruptible() it is necessary to + * also not sleep if the rc state is not running to avoid a deadlock + * when the system is shutting down + */ +static inline bool +lnet_router_checker_active(void) +{ + if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) + return true; + + /* 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; + + return !list_empty(&the_lnet.ln_routers) && + (live_router_check_interval > 0 || + dead_router_check_interval > 0); +} + +static int +lnet_router_checker(void *arg) +{ + struct lnet_peer_ni *rtr; + struct list_head *entry; + + cfs_block_allsigs(); + + while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) { + __u64 version; + int cpt; + int cpt2; + + cpt = lnet_net_lock_current(); +rescan: + version = the_lnet.ln_routers_version; + + list_for_each(entry, &the_lnet.ln_routers) { + rtr = list_entry(entry, struct lnet_peer_ni, + lpni_rtr_list); + + cpt2 = rtr->lpni_cpt; + 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; + } + + lnet_ping_router_locked(rtr); + + /* NB dropped lock */ + if (version != the_lnet.ln_routers_version) { + /* the routers list has changed */ + goto rescan; + } + } + + if (the_lnet.ln_routing) + lnet_update_ni_status_locked(); + + lnet_net_unlock(cpt); + + lnet_prune_rc_data(0); /* don't wait for UNLINK */ + + /* Call schedule_timeout() here always adds 1 to load average + * because kernel counts # active tasks as nr_running + * + nr_uninterruptible. */ + /* if there are any routes then wakeup every second. If + * there are no routes then sleep indefinitely until woken + * up by a user adding a route */ + if (!lnet_router_checker_active()) + wait_event_interruptible(the_lnet.ln_rc_waitq, + lnet_router_checker_active()); + else + wait_event_interruptible_timeout(the_lnet.ln_rc_waitq, + false, + cfs_time_seconds(1)); + } + + lnet_prune_rc_data(1); /* wait for UNLINK */ + + 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; } void -lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp) +lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages) { - int npages = rbp->rbp_npages; - int nbuffers = 0; - lnet_rtrbuf_t *rb; + int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); - LASSERT (list_empty(&rbp->rbp_msgs)); - LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers); + while (--npages >= 0) + __free_page(rb->rb_kiov[npages].kiov_page); - while (!list_empty(&rbp->rbp_bufs)) { - LASSERT (rbp->rbp_credits > 0); + LIBCFS_FREE(rb, sz); +} - rb = list_entry(rbp->rbp_bufs.next, - lnet_rtrbuf_t, rb_list); - list_del(&rb->rb_list); - lnet_destroy_rtrbuf(rb, npages); - nbuffers++; - } +static lnet_rtrbuf_t * +lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt) +{ + int npages = rbp->rbp_npages; + int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); + struct page *page; + lnet_rtrbuf_t *rb; + int i; + + LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz); + if (rb == NULL) + return NULL; + + rb->rb_pool = rbp; + + for (i = 0; i < npages; i++) { + page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, + GFP_KERNEL | __GFP_ZERO); + if (page == NULL) { + while (--i >= 0) + __free_page(rb->rb_kiov[i].kiov_page); + + LIBCFS_FREE(rb, sz); + return NULL; + } + + rb->rb_kiov[i].kiov_len = PAGE_SIZE; + rb->rb_kiov[i].kiov_offset = 0; + rb->rb_kiov[i].kiov_page = page; + } - LASSERT (rbp->rbp_nbuffers == nbuffers); - LASSERT (rbp->rbp_credits == nbuffers); + return rb; +} - rbp->rbp_nbuffers = rbp->rbp_credits = 0; +static void +lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt) +{ + int npages = rbp->rbp_npages; + lnet_rtrbuf_t *rb; + struct 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_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); + list_del(&rb->rb_list); + lnet_destroy_rtrbuf(rb, npages); + } } -int -lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs) +static int +lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt) { - lnet_rtrbuf_t *rb; - int i; - - if (rbp->rbp_nbuffers != 0) { - LASSERT (rbp->rbp_nbuffers == nbufs); - return 0; - } - - for (i = 0; i < nbufs; i++) { - rb = lnet_new_rtrbuf(rbp); - - if (rb == NULL) { - CERROR("Failed to allocate %d router bufs of %d pages\n", - nbufs, rbp->rbp_npages); - return -ENOMEM; - } - - rbp->rbp_nbuffers++; - rbp->rbp_credits++; - rbp->rbp_mincredits++; - list_add(&rb->rb_list, &rbp->rbp_bufs); - - /* No allocation "under fire" */ - /* Otherwise we'd need code to schedule blocked msgs etc */ - LASSERT (!the_lnet.ln_routing); - } - - LASSERT (rbp->rbp_credits == nbufs); - return 0; + struct list_head rb_list; + lnet_rtrbuf_t *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 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 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 + * allocated successfully then join this list to the rbp buffer + * list. If not then free all allocated buffers. */ + 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++; + } + + lnet_net_lock(cpt); + + list_splice_tail(&rb_list, &rbp->rbp_bufs); + rbp->rbp_nbuffers += num_buffers; + rbp->rbp_credits += num_buffers; + rbp->rbp_mincredits = rbp->rbp_credits; + /* We need to schedule blocked msg using the newly + * added buffers. */ + while (!list_empty(&rbp->rbp_bufs) && + !list_empty(&rbp->rbp_msgs)) + lnet_schedule_blocked_locked(rbp); + + lnet_net_unlock(cpt); + + return 0; + +failed: + while (!list_empty(&rb_list)) { + rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list); + list_del(&rb->rb_list); + lnet_destroy_rtrbuf(rb, npages); + } + + return -ENOMEM; } -void +static void lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages) { - CFS_INIT_LIST_HEAD(&rbp->rbp_msgs); - CFS_INIT_LIST_HEAD(&rbp->rbp_bufs); + 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_free_rtrpools(void) +lnet_rtrpools_free(int keep_pools) { - lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[0]); - lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[1]); - lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[2]); + lnet_rtrbufpool_t *rtrp; + int i; + + if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */ + return; + + cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) { + lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i); + lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i); + lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i); + } + + if (!keep_pools) { + cfs_percpt_free(the_lnet.ln_rtrpools); + the_lnet.ln_rtrpools = NULL; + } } -void -lnet_init_rtrpools(void) +static int +lnet_nrb_tiny_calculate(void) { - int small_pages = 1; - int large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT; + int nrbs = LNET_NRB_TINY; - lnet_rtrpool_init(&the_lnet.ln_rtrpools[0], 0); - lnet_rtrpool_init(&the_lnet.ln_rtrpools[1], small_pages); - lnet_rtrpool_init(&the_lnet.ln_rtrpools[2], large_pages); -} + if (tiny_router_buffers < 0) { + LCONSOLE_ERROR_MSG(0x10c, + "tiny_router_buffers=%d invalid when " + "routing enabled\n", tiny_router_buffers); + return -EINVAL; + } + if (tiny_router_buffers > 0) + nrbs = tiny_router_buffers; -int -lnet_alloc_rtrpools(int im_a_router) + nrbs /= LNET_CPT_NUMBER; + return max(nrbs, LNET_NRB_TINY_MIN); +} + +static int +lnet_nrb_small_calculate(void) { - int rc; - - if (!strcmp(forwarding, "")) { - /* not set either way */ - if (!im_a_router) - return 0; - } else if (!strcmp(forwarding, "disabled")) { - /* explicitly disabled */ - return 0; - } else if (!strcmp(forwarding, "enabled")) { - /* explicitly enabled */ - } else { - LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either " - "'enabled' or 'disabled'\n"); - return -EINVAL; - } - - if (tiny_router_buffers <= 0) { - LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when " - "routing enabled\n", tiny_router_buffers); - rc = -EINVAL; - goto failed; - } - - rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[0], - tiny_router_buffers); - if (rc != 0) - goto failed; - - if (small_router_buffers <= 0) { - LCONSOLE_ERROR_MSG(0x10d, "small_router_buffers=%d invalid when" - " routing enabled\n", small_router_buffers); - rc = -EINVAL; - goto failed; - } - - rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[1], - small_router_buffers); - if (rc != 0) - goto failed; - - if (large_router_buffers <= 0) { - LCONSOLE_ERROR_MSG(0x10e, "large_router_buffers=%d invalid when" - " routing enabled\n", large_router_buffers); - rc = -EINVAL; - goto failed; - } - - rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[2], - large_router_buffers); - if (rc != 0) - goto failed; - - LNET_LOCK(); - the_lnet.ln_routing = 1; - LNET_UNLOCK(); - - return 0; + int nrbs = LNET_NRB_SMALL; - failed: - lnet_free_rtrpools(); - return rc; + if (small_router_buffers < 0) { + LCONSOLE_ERROR_MSG(0x10c, + "small_router_buffers=%d invalid when " + "routing enabled\n", small_router_buffers); + return -EINVAL; + } + + if (small_router_buffers > 0) + nrbs = small_router_buffers; + + nrbs /= LNET_CPT_NUMBER; + return max(nrbs, LNET_NRB_SMALL_MIN); } -#else +static int +lnet_nrb_large_calculate(void) +{ + int nrbs = LNET_NRB_LARGE; + + if (large_router_buffers < 0) { + LCONSOLE_ERROR_MSG(0x10c, + "large_router_buffers=%d invalid when " + "routing enabled\n", large_router_buffers); + return -EINVAL; + } + + if (large_router_buffers > 0) + nrbs = large_router_buffers; + + nrbs /= LNET_CPT_NUMBER; + return max(nrbs, LNET_NRB_LARGE_MIN); +} int -lnet_peers_start_down(void) +lnet_rtrpools_alloc(int im_a_router) { - return 0; + lnet_rtrbufpool_t *rtrp; + int nrb_tiny; + int nrb_small; + int nrb_large; + int rc; + int i; + + if (!strcmp(forwarding, "")) { + /* not set either way */ + if (!im_a_router) + return 0; + } else if (!strcmp(forwarding, "disabled")) { + /* explicitly disabled */ + return 0; + } else if (!strcmp(forwarding, "enabled")) { + /* explicitly enabled */ + } else { + LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either " + "'enabled' or 'disabled'\n"); + return -EINVAL; + } + + nrb_tiny = lnet_nrb_tiny_calculate(); + if (nrb_tiny < 0) + return -EINVAL; + + nrb_small = lnet_nrb_small_calculate(); + if (nrb_small < 0) + return -EINVAL; + + nrb_large = lnet_nrb_large_calculate(); + if (nrb_large < 0) + return -EINVAL; + + the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(), + LNET_NRBPOOLS * + sizeof(lnet_rtrbufpool_t)); + if (the_lnet.ln_rtrpools == NULL) { + LCONSOLE_ERROR_MSG(0x10c, + "Failed to initialize router buffe pool\n"); + return -ENOMEM; + } + + cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) { + lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0); + rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX], + nrb_tiny, i); + if (rc != 0) + goto failed; + + lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX], + LNET_NRB_SMALL_PAGES); + rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX], + nrb_small, i); + if (rc != 0) + goto failed; + + lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX], + LNET_NRB_LARGE_PAGES); + rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX], + nrb_large, i); + if (rc != 0) + goto failed; + } + + lnet_net_lock(LNET_LOCK_EX); + the_lnet.ln_routing = 1; + lnet_net_unlock(LNET_LOCK_EX); + return 0; + + failed: + lnet_rtrpools_free(0); + return rc; } -void -lnet_router_checker_stop(void) +static int +lnet_rtrpools_adjust_helper(int tiny, int small, int large) { - return; + int nrb = 0; + int rc = 0; + int i; + lnet_rtrbufpool_t *rtrp; + + /* If the provided values for each buffer pool are different than the + * configured values, we need to take action. */ + if (tiny >= 0) { + tiny_router_buffers = tiny; + nrb = lnet_nrb_tiny_calculate(); + cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) { + rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX], + nrb, i); + if (rc != 0) + return rc; + } + } + if (small >= 0) { + small_router_buffers = small; + nrb = lnet_nrb_small_calculate(); + cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) { + rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX], + nrb, i); + if (rc != 0) + return rc; + } + } + if (large >= 0) { + large_router_buffers = large; + nrb = lnet_nrb_large_calculate(); + cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) { + rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX], + nrb, i); + if (rc != 0) + return rc; + } + } + + return 0; } int -lnet_router_checker_start(void) +lnet_rtrpools_adjust(int tiny, int small, int large) { - return 0; + /* this function doesn't revert the changes if adding new buffers + * failed. It's up to the user space caller to revert the + * changes. */ + + if (!the_lnet.ln_routing) + return 0; + + return lnet_rtrpools_adjust_helper(tiny, small, large); } -void -lnet_free_rtrpools (void) +int +lnet_rtrpools_enable(void) { + int rc = 0; + + if (the_lnet.ln_routing) + return 0; + + if (the_lnet.ln_rtrpools == NULL) + /* If routing is turned off, and we have never + * initialized the pools before, just call the + * standard buffer pool allocation routine as + * if we are just configuring this for the first + * time. */ + 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; + lnet_net_unlock(LNET_LOCK_EX); + + return rc; } void -lnet_init_rtrpools (void) +lnet_rtrpools_disable(void) { + if (!the_lnet.ln_routing) + return; + + lnet_net_lock(LNET_LOCK_EX); + the_lnet.ln_routing = 0; + the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED; + + tiny_router_buffers = 0; + small_router_buffers = 0; + large_router_buffers = 0; + lnet_net_unlock(LNET_LOCK_EX); + lnet_rtrpools_free(1); } int -lnet_alloc_rtrpools (int im_a_arouter) +lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) { - return 0; -} + struct lnet_peer_ni *lp = NULL; + cfs_time_t now = cfs_time_current(); + int cpt = lnet_cpt_of_nid(nid, ni); + + 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; + } -#endif + /* 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_ni_locked(nid); + if (lp == NULL) { + /* nid not found */ + lnet_net_unlock(cpt); + CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid)); + return 0; + } + + /* + * It is possible for this function to be called for the same peer + * but with different NIs. We want to synchronize the notification + * between the different calls. So we will use the lpni_cpt to + * grab the net lock. + */ + if (lp->lpni_cpt != cpt) { + lnet_net_unlock(cpt); + cpt = lp->lpni_cpt; + lnet_net_lock(cpt); + } + + /* 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->lpni_last_alive) + when = lp->lpni_last_alive; + + lnet_notify_locked(lp, ni == NULL, alive, when); + + if (ni != NULL) + lnet_ni_notify_locked(ni, lp); + + lnet_peer_ni_decref_locked(lp); + + lnet_net_unlock(cpt); + return 0; +} +EXPORT_SYMBOL(lnet_notify);