X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Frouter.c;h=091cfd967d608ef7339ebedd6f5b72185d6d4967;hp=df1bfc86685f9f70333497649eca1a39dbb57ebd;hb=fb259fe85813e0f28ac7f7410689e3856ef26316;hpb=682bb58318d7d5d101febfc6770c95887c750e2e diff --git a/lnet/lnet/router.c b/lnet/lnet/router.c index df1bfc8..091cfd9 100644 --- a/lnet/lnet/router.c +++ b/lnet/lnet/router.c @@ -1,7 +1,7 @@ -/* -*- 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, 2013, Intel Corporation. * * This file is part of Portals * http://sourceforge.net/projects/sandiaportals/ @@ -26,39 +26,82 @@ #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_CACHE_SIZE - 1) >> \ + PAGE_CACHE_SHIFT) + static char *forwarding = ""; CFS_MODULE_PARM(forwarding, "s", charp, 0444, "Explicitly enable/disable forwarding between networks"); -static int tiny_router_buffers = 1024; +static int tiny_router_buffers; CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444, - "# of 0 payload messages to buffer in the router"); -static int small_router_buffers = 8192; + "# 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"); -static int large_router_buffers = 512; + "# 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"); + "# 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"); static int auto_down = 1; CFS_MODULE_PARM(auto_down, "i", int, 0444, "Automatically mark peers down on comms error"); +int +lnet_peer_buffer_credits(lnet_ni_t *ni) +{ + /* 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; +} + +/* forward ref's */ +static int lnet_router_checker(void *); +#else + +int +lnet_peer_buffer_credits(lnet_ni_t *ni) +{ + return 0; +} + +#endif + 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"); + "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; +CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644, + "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; +CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644, + "Seconds between dead router health checks (<= 0 to disable)"); + +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)"); 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"); +CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644, + "Seconds to wait for the reply to a router health query"); int lnet_peers_start_down(void) @@ -67,9 +110,9 @@ lnet_peers_start_down(void) } void -lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, time_t when) +lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when) { - if (when < lp->lp_timestamp) { /* out of date information */ + if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */ CDEBUG(D_NET, "Out of date\n"); return; } @@ -89,30 +132,27 @@ lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, time_t when) 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); + CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive); } void -lnet_do_notify (lnet_peer_t *lp) +lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *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 */ - if (lp->lp_notifying) { - LNET_UNLOCK(); + if (lp->lp_notifying) return; - } lp->lp_notifying = 1; - + while (lp->lp_notify) { alive = lp->lp_alive; notifylnd = lp->lp_notifylnd; @@ -121,662 +161,961 @@ lnet_do_notify (lnet_peer_t *lp) lp->lp_notify = 0; if (notifylnd && ni->ni_lnd->lnd_notify != NULL) { - LNET_UNLOCK(); + lnet_net_unlock(lp->lp_cpt); - /* A new notification could happen now; I'll handle it - * when control returns to me */ + /* A new notification could happen now; I'll handle it + * when control returns to me */ - (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive); + (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive); - LNET_LOCK(); - } - } + lnet_net_lock(lp->lp_cpt); + } + } + + lp->lp_notifying = 0; +} - lp->lp_notifying = 0; - LNET_UNLOCK(); +static void +lnet_rtr_addref_locked(lnet_peer_t *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); + /* addref for the_lnet.ln_routers */ + lnet_peer_addref_locked(lp); + the_lnet.ln_routers_version++; + } } -int -lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, time_t when) +static void +lnet_rtr_decref_locked(lnet_peer_t *lp) { - lnet_peer_t *lp = NULL; - time_t now = cfs_time_current_sec(); + 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 == 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); + the_lnet.ln_routers_version++; + } +} - LASSERT (!in_interrupt ()); +lnet_remotenet_t * +lnet_find_net_locked (__u32 net) +{ + lnet_remotenet_t *rnet; + struct list_head *tmp; + struct list_head *rn_list; - CDEBUG (D_NET, "%s notifying %s: %s\n", - (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid), - libcfs_nid2str(nid), - alive ? "up" : "down"); + LASSERT(!the_lnet.ln_shutdown); - 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; - } + rn_list = lnet_net2rnethash(net); + list_for_each(tmp, rn_list) { + rnet = list_entry(tmp, lnet_remotenet_t, lrn_list); - /* 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 (rnet->lrn_net == net) + return rnet; + } + return NULL; +} - 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(); +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; - 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; - } + if (seeded) + return; - lnet_notify_locked(lp, ni == NULL, alive, when); + cfs_get_random_bytes(seed, sizeof(seed)); - LNET_UNLOCK(); - - lnet_do_notify(lp); - - LNET_LOCK(); + /* 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)); - lnet_peer_decref_locked(lp); + if (lnd_type != LOLND) + seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type); + } - LNET_UNLOCK(); - return 0; + do_gettimeofday(&tv); + cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]); + seeded = 1; + return; } -EXPORT_SYMBOL(lnet_notify); -#else +/* NB expects LNET_LOCK held */ +void +lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route) +{ + 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->lp_routes); + + the_lnet.ln_remote_nets_version++; + lnet_rtr_addref_locked(route->lr_gateway); +} int -lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, time_t when) +lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway, + unsigned int priority) { - return -EOPNOTSUPP; -} + 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", + 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 < 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; + } -#endif + INIT_LIST_HEAD(&rnet->lrn_routes); + rnet->lrn_net = net; + route->lr_hops = hops; + route->lr_net = net; + route->lr_priority = priority; -static void -lnet_rtr_addref_locked(lnet_peer_t *lp) -{ - LASSERT (lp->lp_refcount > 0); - LASSERT (lp->lp_rtr_refcount >= 0); + lnet_net_lock(LNET_LOCK_EX); - lp->lp_rtr_refcount++; - if (lp->lp_rtr_refcount == 1) { - struct list_head *pos; + rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX); + if (rc != 0) { + lnet_net_unlock(LNET_LOCK_EX); - /* 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); + LIBCFS_FREE(route, sizeof(*route)); + LIBCFS_FREE(rnet, sizeof(*rnet)); - if (rtr->lp_nid < lp->lp_nid) - break; - } + 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; + } - list_add(&lp->lp_rtr_list, pos); - /* addref for the_lnet.ln_routers */ - lnet_peer_addref_locked(lp); - the_lnet.ln_routers_version++; - } -} + LASSERT(!the_lnet.ln_shutdown); -static void -lnet_rtr_decref_locked(lnet_peer_t *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++; - } -} + rnet2 = lnet_find_net_locked(net); + if (rnet2 == NULL) { + /* new network */ + list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net)); + rnet2 = rnet; + } -lnet_remotenet_t * -lnet_find_net_locked (__u32 net) -{ - lnet_remotenet_t *rnet; - struct list_head *tmp; + /* 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); - LASSERT (!the_lnet.ln_shutdown); + if (route2->lr_gateway == route->lr_gateway) { + add_route = 0; + break; + } - list_for_each (tmp, &the_lnet.ln_remote_nets) { - rnet = list_entry(tmp, lnet_remotenet_t, lrn_list); + /* our lookups must be true */ + LASSERT(route2->lr_gateway->lp_nid != gateway); + } - if (rnet->lrn_net == net) - return rnet; - } - return NULL; -} + if (add_route) { + lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */ + lnet_add_route_to_rnet(rnet2, route); -int -lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway) -{ - 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; - } + ni = route->lr_gateway->lp_ni; + lnet_net_unlock(LNET_LOCK_EX); - INIT_LIST_HEAD(&rnet->lrn_routes); - rnet->lrn_net = net; - rnet->lrn_hops = hops; + /* XXX Assume alive */ + if (ni->ni_lnd->lnd_notify != NULL) + (ni->ni_lnd->lnd_notify)(ni, gateway, 1); - LNET_LOCK(); + lnet_net_lock(LNET_LOCK_EX); + } - rc = lnet_nid2peer_locked(&route->lr_gateway, gateway); - if (rc != 0) { - LNET_UNLOCK(); + /* -1 for notify or !add_route */ + lnet_peer_decref_locked(route->lr_gateway); + lnet_net_unlock(LNET_LOCK_EX); - LIBCFS_FREE(route, sizeof(*route)); - LIBCFS_FREE(rnet, sizeof(*rnet)); + if (!add_route) + LIBCFS_FREE(route, sizeof(*route)); - if (rc == -EHOSTUNREACH) /* gateway is not on a local net */ - return 0; /* ignore the route entry */ + if (rnet != rnet2) + LIBCFS_FREE(rnet, sizeof(*rnet)); - CERROR("Error %d creating route %s %d %s\n", rc, - libcfs_net2str(net), hops, libcfs_nid2str(gateway)); - return rc; - } + return 0; +} - LASSERT (!the_lnet.ln_shutdown); - CFS_INIT_LIST_HEAD(&zombies); +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->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); + + CERROR("Routes to %s via %s and %s not " + "supported\n", + libcfs_net2str(net), + libcfs_nid2str(nid1), + libcfs_nid2str(nid2)); + return -EINVAL; + } + } + } - rnet2 = lnet_find_net_locked(net); - if (rnet2 == NULL) { - /* new network */ - list_add_tail(&rnet->lrn_list, &the_lnet.ln_remote_nets); - rnet2 = rnet; - } + lnet_net_unlock(cpt); + return 0; +} - 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; - } +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; + + 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->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; + } + } - /* 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 (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); - 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_peer_decref_locked(route->lr_gateway); - LNET_UNLOCK(); - LIBCFS_FREE(route, sizeof(*route)); - } + return rc; +} - return rc; +void +lnet_destroy_routes (void) +{ + lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY); } -int -lnet_check_routes (void) +int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg) { - 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; + 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; + } + } + 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_del_route (__u32 net, lnet_nid_t gw_nid) +lnet_get_route(int idx, __u32 *net, __u32 *hops, + lnet_nid_t *gateway, __u32 *alive, __u32 *priority) { - lnet_remotenet_t *rnet; - lnet_route_t *route; - struct list_head *e1; - struct list_head *e2; - int rc = -ENOENT; + 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; + } + } + } + } - CDEBUG(D_NET, "Del route: net %s : gw %s\n", - libcfs_net2str(net), libcfs_nid2str(gw_nid)); + lnet_net_unlock(cpt); + return -ENOENT; +} - /* NB Caller may specify either all routes via the given gateway - * or a specific route entry actual NIDs) */ +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; +} - again: - LNET_LOCK(); +/** + * 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; + } - list_for_each (e1, &the_lnet.ln_remote_nets) { - rnet = list_entry(e1, lnet_remotenet_t, lrn_list); + 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 (!(net == LNET_NIDNET(LNET_NID_ANY) || - net == rnet->lrn_net)) - continue; + 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; + } +} - list_for_each (e2, &rnet->lrn_routes) { - route = list_entry(e2, lnet_route_t, lr_list); +static void +lnet_router_checker_event(lnet_event_t *event) +{ + lnet_rc_data_t *rcd = event->md.user_ptr; + struct lnet_peer *lp; - if (!(gw_nid == LNET_NID_ANY || - gw_nid == route->lr_gateway->lp_nid)) - continue; + LASSERT(rcd != NULL); - list_del(&route->lr_list); - the_lnet.ln_remote_nets_version++; + if (event->unlinked) { + LNetInvalidateHandle(&rcd->rcd_mdh); + return; + } - if (list_empty(&rnet->lrn_routes)) - list_del(&rnet->lrn_list); - else - rnet = NULL; + LASSERT(event->type == LNET_EVENT_SEND || + event->type == LNET_EVENT_REPLY); - lnet_rtr_decref_locked(route->lr_gateway); - lnet_peer_decref_locked(route->lr_gateway); - LNET_UNLOCK(); + lp = rcd->rcd_gateway; + LASSERT(lp != NULL); - LIBCFS_FREE(route, sizeof (*route)); + /* 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 (rnet != NULL) - LIBCFS_FREE(rnet, sizeof(*rnet)); + if (event->type == LNET_EVENT_SEND) { + lp->lp_ping_notsent = 0; + if (event->status == 0) + goto out; + } - rc = 0; - goto again; - } - } + /* 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!!! */ - LNET_UNLOCK(); - return rc; + if (avoid_asym_router_failure && event->status == 0) + lnet_parse_rc_info(rcd); + + out: + lnet_net_unlock(lp->lp_cpt); } void -lnet_destroy_routes (void) +lnet_wait_known_routerstate(void) { - lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY); -} + lnet_peer_t *rtr; + struct list_head *entry; + int all_known; -int -lnet_get_route (int idx, __u32 *net, __u32 *hops, - lnet_nid_t *gateway, __u32 *alive) -{ - 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; - } - } + 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)); } +} - LNET_UNLOCK(); - return -ENOENT; +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; + } + } + } } -#if defined(__KERNEL__) && defined(LNET_ROUTER) -static void -lnet_router_checker_event (lnet_event_t *event) +void +lnet_update_ni_status_locked(void) { - /* 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; - } + 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; + } + lnet_ni_unlock(ni); + } +} - LASSERT (event->type == LNET_EVENT_SEND || - event->type == LNET_EVENT_REPLY); - - nid = (event->type == LNET_EVENT_SEND) ? - event->target.nid : event->initiator.nid; +void +lnet_destroy_rc_data(lnet_rc_data_t *rcd) +{ + LASSERT(list_empty(&rcd->rcd_list)); + /* detached from network */ + LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh)); - 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 (rcd->rcd_gateway != NULL) { + int cpt = rcd->rcd_gateway->lp_cpt; - 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!!! */ - } + lnet_net_lock(cpt); + lnet_peer_decref_locked(rcd->rcd_gateway); + lnet_net_unlock(cpt); + } - /* 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); + if (rcd->rcd_pinginfo != NULL) + LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE); - lnet_peer_decref_locked(lp); + LIBCFS_FREE(rcd, sizeof(*rcd)); } -static int -lnet_router_checker(void *arg) +lnet_rc_data_t * +lnet_create_rc_data_locked(lnet_peer_t *gateway) { - static lnet_ping_info_t pinginfo; + lnet_rc_data_t *rcd = NULL; + lnet_ping_info_t *pi; + int rc; + int i; - 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; + lnet_net_unlock(gateway->lp_cpt); - cfs_daemonize("router_checker"); - cfs_block_allsigs(); + LIBCFS_ALLOC(rcd, sizeof(*rcd)); + if (rcd == NULL) + goto out; - rtr_id.pid = LUSTRE_SRV_LNET_PID; + LNetInvalidateHandle(&rcd->rcd_mdh); + INIT_LIST_HEAD(&rcd->rcd_list); - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); + 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; - rc = LNetMDBind((lnet_md_t){.start = &pinginfo, - .length = sizeof(pinginfo), + 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, - &mdh); - + &rcd->rcd_mdh); 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; - } + goto out; + } + LASSERT(rc == 0); - 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; + } - the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING; - mutex_up(&the_lnet.ln_rc_signal); /* let my parent go */ + lnet_peer_addref_locked(gateway); + rcd->rcd_gateway = gateway; + gateway->lp_rcd = rcd; + gateway->lp_ping_notsent = 0; - while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) { - __u64 version; + return rcd; - LNET_LOCK(); -rescan: - version = the_lnet.ln_routers_version; + out: + if (rcd != NULL) { + if (!LNetHandleIsInvalid(rcd->rcd_mdh)) { + rc = LNetMDUnlink(rcd->rcd_mdh); + LASSERT(rc == 0); + } + lnet_destroy_rc_data(rcd); + } - list_for_each (entry, &the_lnet.ln_routers) { - rtr = list_entry(entry, lnet_peer_t, lp_rtr_list); + lnet_net_lock(gateway->lp_cpt); + return gateway->lp_rcd; +} - lnet_peer_addref_locked(rtr); +static int +lnet_router_check_interval (lnet_peer_t *rtr) +{ + int secs; - now = cfs_time_current_sec(); + secs = rtr->lp_alive ? live_router_check_interval : + dead_router_check_interval; + if (secs < 0) + secs = 0; - if (rtr->lp_ping_deadline != 0 && /* ping timed out? */ - now > rtr->lp_ping_deadline) - lnet_notify_locked(rtr, 1, 0, now); + return secs; +} - LNET_UNLOCK(); +static void +lnet_ping_router_locked (lnet_peer_t *rtr) +{ + lnet_rc_data_t *rcd = NULL; + cfs_time_t now = cfs_time_current(); + int secs; - /* Run any outstanding notificiations */ - lnet_do_notify(rtr); + lnet_peer_addref_locked(rtr); - 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_LOCK(); - rtr_id.nid = rtr->lp_nid; - rtr->lp_ping_notsent = 1; - rtr->lp_ping_timestamp = now; - - if (rtr->lp_ping_deadline == 0) - rtr->lp_ping_deadline = - now + router_ping_timeout; - - LNET_UNLOCK(); - - LNetGet(LNET_NID_ANY, mdh, rtr_id, - LNET_RESERVED_PORTAL, - LNET_PROTO_PING_MATCHBITS, 0); - } - - LNET_LOCK(); - lnet_peer_decref_locked(rtr); + if (rtr->lp_ping_deadline != 0 && /* ping timed out? */ + cfs_time_after(now, rtr->lp_ping_deadline)) + lnet_notify_locked(rtr, 1, 0, now); - if (version != the_lnet.ln_routers_version) { - /* the routers list has changed */ - goto rescan; - } - } + /* Run any outstanding notifications */ + lnet_ni_notify_locked(rtr->lp_ni, rtr); - LNET_UNLOCK(); - - /* 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)); + 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; } - 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); + rcd = rtr->lp_rcd != NULL ? + rtr->lp_rcd : lnet_create_rc_data_locked(rtr); - /* The unlink event callback will signal final completion */ + if (rcd == NULL) + return; - return 0; -} + secs = lnet_router_check_interval(rtr); + 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); -void -lnet_wait_known_routerstate(void) -{ - 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; - } - } + 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; - LNET_UNLOCK(); + id.nid = rtr->lp_nid; + id.pid = LNET_PID_LUSTRE; + CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id)); - if (all_known) - return; - - cfs_pause(cfs_time_seconds(1)); - } -} + rtr->lp_ping_notsent = 1; + rtr->lp_ping_timestamp = now; -void -lnet_router_checker_stop(void) -{ - int rc; + mdh = rcd->rcd_mdh; - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING || - the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); + if (rtr->lp_ping_deadline == 0) { + rtr->lp_ping_deadline = + cfs_time_shift(router_ping_timeout); + } - if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN) - return; + lnet_net_unlock(rtr->lp_cpt); - the_lnet.ln_rc_state = LNET_RC_STATE_STOPTHREAD; - /* block until event callback signals exit */ - mutex_down(&the_lnet.ln_rc_signal); + rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL, + LNET_PROTO_PING_MATCHBITS, 0); - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED); + lnet_net_lock(rtr->lp_cpt); + if (rc != 0) + rtr->lp_ping_notsent = 0; /* no event pending */ + } - rc = LNetEQFree(the_lnet.ln_rc_eqh); - LASSERT (rc == 0); - - the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN; + lnet_peer_decref_locked(rtr); + return; } int lnet_router_checker_start(void) { - int rc; + 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; + + lnet_net_unlock(0); + + rc = LNetSetAsync(id, 1); + if (rc != 0) { + CWARN("LNetSetAsync %s failed: %d\n", + libcfs_id2str(id), rc); + return rc; + } + + lnet_net_lock(0); + /* NB router list doesn't change in userspace */ + LASSERT(version == the_lnet.ln_routers_version); + } + + lnet_net_unlock(0); + + if (nrtr == 0) { + CDEBUG(D_NET, + "No router found, not starting router checker\n"); + return 0; + } + + /* at least allow a SENT and a REPLY per router */ + if (router_checker_max_eqsize < 2 * nrtr) + router_checker_max_eqsize = 2 * nrtr; + + LASSERT (eqsz > 0); + if (eqsz > router_checker_max_eqsize) + eqsz = router_checker_max_eqsize; +#endif LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); @@ -787,37 +1126,42 @@ lnet_router_checker_start(void) "\n"); return -EINVAL; } - - if (live_router_check_interval <= 0 && + + if (!the_lnet.ln_routing && + live_router_check_interval <= 0 && dead_router_check_interval <= 0) return 0; - init_mutex_locked(&the_lnet.ln_rc_signal); - +#ifdef __KERNEL__ + sema_init(&the_lnet.ln_rc_signal, 0); /* EQ size doesn't matter; the callback is guaranteed to get every * event */ - rc = LNetEQAlloc(1, lnet_router_checker_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\n", rc); + CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc); return -ENOMEM; } - rc = (int)cfs_kernel_thread(lnet_router_checker, NULL, 0); - if (rc < 0) { + 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); - goto failed; + /* 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; } - - 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); +#endif if (check_routers_before_use) { /* Note that a helpful side-effect of pinging all known routers @@ -825,13 +1169,180 @@ lnet_router_checker_start(void) * may have to a previous instance of me. */ lnet_wait_known_routerstate(); } - + return 0; - - failed: +} + +void +lnet_router_checker_stop (void) +{ + int rc; + + 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); + rc = LNetEQFree(the_lnet.ln_rc_eqh); LASSERT (rc == 0); - return rc; + return; +} + +static void +lnet_prune_rc_data(int wait_unlink) +{ + 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); + + 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"); + cfs_pause(cfs_time_seconds(1) / 4); + + lnet_net_lock(LNET_LOCK_EX); + } + + lnet_net_unlock(LNET_LOCK_EX); +} + + +#if defined(__KERNEL__) && defined(LNET_ROUTER) + +static int +lnet_router_checker(void *arg) +{ + lnet_peer_t *rtr; + struct list_head *entry; + + cfs_block_allsigs(); + + LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); + + 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, 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; + } + + 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 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)); + } + + LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING); + + 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 @@ -840,37 +1351,38 @@ lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages) int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); while (--npages >= 0) - cfs_free_page(rb->rb_kiov[npages].kiov_page); + __free_page(rb->rb_kiov[npages].kiov_page); LIBCFS_FREE(rb, sz); } lnet_rtrbuf_t * -lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp) +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; + 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; + LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz); + if (rb == NULL) + return NULL; - rb->rb_pool = rbp; + rb->rb_pool = rbp; - for (i = 0; i < npages; i++) { - page = cfs_alloc_page(CFS_ALLOC_ZERO | CFS_ALLOC_STD); + 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) - cfs_free_page(rb->rb_kiov[i].kiov_page); + __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_len = PAGE_CACHE_SIZE; rb->rb_kiov[i].kiov_offset = 0; rb->rb_kiov[i].kiov_page = page; } @@ -879,70 +1391,103 @@ lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp) } void -lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp) +lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt) { - int npages = rbp->rbp_npages; - int nbuffers = 0; - lnet_rtrbuf_t *rb; + 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_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); + } +} - LASSERT (list_empty(&rbp->rbp_msgs)); - LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers); +static int +lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt) +{ + struct list_head rb_list; + lnet_rtrbuf_t *rb; + int num_rb; + int num_buffers = 0; + int npages = rbp->rbp_npages; + + /* If we are called for less buffers than already in the pool, we + * just lower the 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; + lnet_net_unlock(cpt); + return 0; + } - while (!list_empty(&rbp->rbp_bufs)) { - LASSERT (rbp->rbp_credits > 0); + INIT_LIST_HEAD(&rb_list); - rb = list_entry(rbp->rbp_bufs.next, - lnet_rtrbuf_t, rb_list); - list_del(&rb->rb_list); - lnet_destroy_rtrbuf(rb, npages); - nbuffers++; - } + /* 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; - LASSERT (rbp->rbp_nbuffers == nbuffers); - LASSERT (rbp->rbp_credits == nbuffers); + while (num_rb < nbufs) { + rb = lnet_new_rtrbuf(rbp, cpt); + if (rb == NULL) { + CERROR("Failed to allocate %d route bufs of %d pages\n", + nbufs, npages); + goto failed; + } - rbp->rbp_nbuffers = rbp->rbp_credits = 0; -} + list_add(&rb->rb_list, &rb_list); + num_buffers++; + num_rb++; + } -int -lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs) -{ - lnet_rtrbuf_t *rb; - int i; + lnet_net_lock(cpt); - 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; - } + 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); - rbp->rbp_nbuffers++; - rbp->rbp_credits++; - rbp->rbp_mincredits++; - list_add(&rb->rb_list, &rbp->rbp_bufs); + lnet_net_unlock(cpt); - /* No allocation "under fire" */ - /* Otherwise we'd need code to schedule blocked msgs etc */ - LASSERT (!the_lnet.ln_routing); - } + return 0; - LASSERT (rbp->rbp_credits == nbufs); - 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 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; @@ -950,124 +1495,460 @@ lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages) } void -lnet_free_rtrpools(void) +lnet_rtrpools_free(int keep_pools) +{ + 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; + } +} + +static int +lnet_nrb_tiny_calculate(void) { - 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]); + int nrbs = LNET_NRB_TINY; + + if (tiny_router_buffers < 0) { + LCONSOLE_ERROR_MSG(0x10c, + "tiny_router_buffers=%d invalid when " + "routing enabled\n", tiny_router_buffers); + return -1; + } + + if (tiny_router_buffers > 0) + nrbs = tiny_router_buffers; + + nrbs /= LNET_CPT_NUMBER; + return max(nrbs, LNET_NRB_TINY_MIN); } -void -lnet_init_rtrpools(void) +static int +lnet_nrb_small_calculate(void) +{ + int nrbs = LNET_NRB_SMALL; + + if (small_router_buffers < 0) { + LCONSOLE_ERROR_MSG(0x10c, + "small_router_buffers=%d invalid when " + "routing enabled\n", small_router_buffers); + return -1; + } + + if (small_router_buffers > 0) + nrbs = small_router_buffers; + + nrbs /= LNET_CPT_NUMBER; + return max(nrbs, LNET_NRB_SMALL_MIN); +} + +static int +lnet_nrb_large_calculate(void) { - int small_pages = 1; - int large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT; + 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 -1; + } + + if (large_router_buffers > 0) + nrbs = large_router_buffers; - 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); + nrbs /= LNET_CPT_NUMBER; + return max(nrbs, LNET_NRB_LARGE_MIN); } +int +lnet_rtrpools_alloc(int im_a_router) +{ + 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; +} + +static int +lnet_rtrpools_adjust_helper(int tiny, int small, int large) +{ + 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_alloc_rtrpools(int im_a_router) +lnet_rtrpools_adjust(int tiny, int small, int large) { - 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"); + /* 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); +} + +int +lnet_rtrpools_enable(void) +{ + int rc; + + 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. */ + return lnet_rtrpools_alloc(1); + + 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 0; +} + +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_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; } - - 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; + /* 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 (small_router_buffers <= 0) { - LCONSOLE_ERROR_MSG(0x10d, "small_router_buffers=%d invalid when" - " routing enabled\n", small_router_buffers); - rc = -EINVAL; - goto failed; + if (ni != NULL && !alive && /* LND telling me she's down */ + !auto_down) { /* auto-down disabled */ + CDEBUG(D_NET, "Auto-down disabled\n"); + return 0; } - rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[1], - small_router_buffers); - if (rc != 0) - goto failed; + lnet_net_lock(cpt); + + if (the_lnet.ln_shutdown) { + lnet_net_unlock(cpt); + return -ESHUTDOWN; + } - 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; + 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; } - rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[2], - large_router_buffers); - if (rc != 0) - goto failed; + /* 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_LOCK(); - the_lnet.ln_routing = 1; - LNET_UNLOCK(); - - return 0; + lnet_notify_locked(lp, ni == NULL, alive, when); - failed: - lnet_free_rtrpools(); - return rc; -} + if (ni != NULL) + lnet_ni_notify_locked(ni, lp); -#else + lnet_peer_decref_locked(lp); -int -lnet_peers_start_down(void) -{ - return 0; + lnet_net_unlock(cpt); + return 0; } +EXPORT_SYMBOL(lnet_notify); void -lnet_router_checker_stop(void) +lnet_get_tunables (void) { return; } +#else + int -lnet_router_checker_start(void) +lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) { - return 0; + return -EOPNOTSUPP; } void -lnet_free_rtrpools (void) +lnet_router_checker (void) { + 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; + + 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); + + LASSERT(LNET_CPT_NUMBER == 1); + + 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; + } + + LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); + + lnet_net_lock(0); + + 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); + } + + 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); + + 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_init_rtrpools (void) +lnet_rtrpools_free(int keep_pools) { } int -lnet_alloc_rtrpools (int im_a_arouter) +lnet_rtrpools_alloc(int im_a_arouter) { return 0; }