X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Frouter.c;h=71903cf24bc7fc915c975e4317cbd3326e369d2a;hp=49fa60f413336641c6376979054994515c4cedc6;hb=362e52efec1a6f6fd43b16ccd555b714a30642da;hpb=e069296630240947f1815e505067fd48033909f7 diff --git a/lnet/lnet/router.c b/lnet/lnet/router.c index 49fa60f..71903cf 100644 --- a/lnet/lnet/router.c +++ b/lnet/lnet/router.c @@ -1,10 +1,9 @@ /* * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * - * Copyright (c) 2011, Whamcloud, 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 @@ -24,434 +23,471 @@ #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"); +module_param(forwarding, charp, 0444); +MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks"); static int tiny_router_buffers; -CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444, - "# of 0 payload messages to buffer in the router"); +module_param(tiny_router_buffers, int, 0444); +MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router"); static int small_router_buffers; -CFS_MODULE_PARM(small_router_buffers, "i", int, 0444, - "# of small (1 page) messages to buffer in the router"); +module_param(small_router_buffers, int, 0444); +MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router"); static int large_router_buffers; -CFS_MODULE_PARM(large_router_buffers, "i", int, 0444, - "# of large messages to buffer in the router"); -static int peer_buffer_credits = 0; -CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444, - "# router buffer credits per peer"); +module_param(large_router_buffers, int, 0444); +MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router"); +static int peer_buffer_credits; +module_param(peer_buffer_credits, int, 0444); +MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer"); static int auto_down = 1; -CFS_MODULE_PARM(auto_down, "i", int, 0444, - "Automatically mark peers down on comms error"); +module_param(auto_down, int, 0444); +MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error"); int -lnet_peer_buffer_credits(lnet_ni_t *ni) +lnet_peer_buffer_credits(struct lnet_net *net) { - /* NI option overrides LNet default */ - if (ni->ni_peerrtrcredits > 0) - return ni->ni_peerrtrcredits; - if (peer_buffer_credits > 0) - return peer_buffer_credits; - - /* As an approximation, allow this peer the same number of router - * buffers as it is allowed outstanding sends */ - return ni->ni_peertxcredits; + /* NI option overrides LNet default */ + if (net->net_tunables.lct_peer_rtr_credits > 0) + return net->net_tunables.lct_peer_rtr_credits; + if (peer_buffer_credits > 0) + return peer_buffer_credits; + + /* As an approximation, allow this peer the same number of router + * buffers as it is allowed outstanding sends */ + return net->net_tunables.lct_peer_tx_credits; } /* forward ref's */ static int lnet_router_checker(void *); -#else - -int -lnet_peer_buffer_credits(lnet_ni_t *ni) -{ - return 0; -} -#endif +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 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 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 avoid_asym_router_failure = 0; -CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0444, - "Avoid asymmetrical failures: reserved, use at your own risk"); +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 dead_router_check_interval = 0; -CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0444, - "Seconds between dead router health checks (<= 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 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, cfs_time_t when) +lnet_notify_locked(struct lnet_peer_ni *lp, int notifylnd, int alive, + cfs_time_t when) { - if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */ - CDEBUG(D_NET, "Out of date\n"); - return; - } - - lp->lp_timestamp = when; /* update timestamp */ - lp->lp_ping_deadline = 0; /* disable ping timeout */ - - if (lp->lp_alive_count != 0 && /* got old news */ - (!lp->lp_alive) == (!alive)) { /* new date for old news */ - CDEBUG(D_NET, "Old news\n"); - return; - } - - /* Flag that notification is outstanding */ - - lp->lp_alive_count++; - lp->lp_alive = !(!alive); /* 1 bit! */ - lp->lp_notify = 1; - lp->lp_notifylnd |= notifylnd; - if (lp->lp_alive) - lp->lp_ping_version = LNET_PROTO_PING_UNKNOWN; /* reset */ - - CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive); + if (cfs_time_before(when, lp->lpni_timestamp)) { /* out of date information */ + CDEBUG(D_NET, "Out of date\n"); + return; + } + + /* + * 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; + } + + /* 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 */ + + spin_unlock(&lp->lpni_lock); + + CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive); } -void -lnet_ni_notify_locked(lnet_ni_t *ni, 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) { - int alive; - int notifylnd; + int alive; + int notifylnd; - /* 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 */ + /* Notify only in 1 thread at any time to ensure ordered notification. + * NB individual events can be missed; the only guarantee is that you + * always get the most recent news */ - if (lp->lp_notifying) - return; + spin_lock(&lp->lpni_lock); - lp->lp_notifying = 1; + if (lp->lpni_notifying || ni == NULL) { + spin_unlock(&lp->lpni_lock); + return; + } + + lp->lpni_notifying = 1; - while (lp->lp_notify) { - alive = lp->lp_alive; - notifylnd = lp->lp_notifylnd; + /* + * 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; - lp->lp_notifylnd = 0; - lp->lp_notify = 0; + lp->lpni_notifylnd = 0; + lp->lpni_notify = 0; - if (notifylnd && ni->ni_lnd->lnd_notify != NULL) { - lnet_net_unlock(lp->lp_cpt); + if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) { + spin_unlock(&lp->lpni_lock); + lnet_net_unlock(lp->lpni_cpt); /* 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_net->net_lnd->lnd_notify)(ni, lp->lpni_nid, + alive); - lnet_net_lock(lp->lp_cpt); + lnet_net_lock(lp->lpni_cpt); + spin_lock(&lp->lpni_lock); } } - lp->lp_notifying = 0; + lp->lpni_notifying = 0; + spin_unlock(&lp->lpni_lock); } - 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); + LASSERT(atomic_read(&lp->lpni_refcount) > 0); + LASSERT(lp->lpni_rtr_refcount >= 0); /* lnet_net_lock must be exclusively locked */ - lp->lp_rtr_refcount++; - if (lp->lp_rtr_refcount == 1) { - cfs_list_t *pos; - - /* a simple insertion sort */ - cfs_list_for_each_prev(pos, &the_lnet.ln_routers) { - lnet_peer_t *rtr = cfs_list_entry(pos, lnet_peer_t, - lp_rtr_list); - - if (rtr->lp_nid < lp->lp_nid) - break; - } - - cfs_list_add(&lp->lp_rtr_list, pos); - /* addref for the_lnet.ln_routers */ - lnet_peer_addref_locked(lp); - the_lnet.ln_routers_version++; - } + 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); + LASSERT(atomic_read(&lp->lpni_refcount) > 0); + LASSERT(lp->lpni_rtr_refcount > 0); /* lnet_net_lock must be exclusively locked */ - lp->lp_rtr_refcount--; - if (lp->lp_rtr_refcount == 0) { - LASSERT(cfs_list_empty(&lp->lp_routes)); - - if (lp->lp_rcd != NULL) { - cfs_list_add(&lp->lp_rcd->rcd_list, - &the_lnet.ln_rcd_deathrow); - lp->lp_rcd = NULL; - } - - cfs_list_del(&lp->lp_rtr_list); - /* decref for the_lnet.ln_routers */ - lnet_peer_decref_locked(lp); - the_lnet.ln_routers_version++; - } + 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; - cfs_list_t *tmp; + lnet_remotenet_t *rnet; + struct list_head *tmp; + struct list_head *rn_list; - LASSERT (!the_lnet.ln_shutdown); + LASSERT(!the_lnet.ln_shutdown); - cfs_list_for_each (tmp, &the_lnet.ln_remote_nets) { - rnet = cfs_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; } static void lnet_shuffle_seed(void) { - static int seeded = 0; - int lnd_type, seed[2]; - struct timeval tv; - lnet_ni_t *ni; - cfs_list_t *tmp; - - 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 */ - cfs_list_for_each(tmp, &the_lnet.ln_nis) { - ni = cfs_list_entry(tmp, lnet_ni_t, ni_list); - lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid)); - - if (lnd_type != LOLND) - seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type); - } - - cfs_gettimeofday(&tv); - cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]); - seeded = 1; - return; + 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; } /* NB expects LNET_LOCK held */ -void -lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route) +static void +lnet_add_route_to_rnet(lnet_remotenet_t *rnet, lnet_route_t *route) { - unsigned int len = 0; - unsigned int offset = 0; - cfs_list_t *e; - - lnet_shuffle_seed(); - - cfs_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); - cfs_list_for_each (e, &rnet->lrn_routes) { - if (offset == 0) - break; - offset--; - } - cfs_list_add(&route->lr_list, e); - cfs_list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes); - - the_lnet.ln_remote_nets_version++; - lnet_rtr_addref_locked(route->lr_gateway); + 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_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway) +lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway, + unsigned int priority) { - cfs_list_t *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 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; - route->lr_hops = hops; + 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); - rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX); - if (rc != 0) { + 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)); - if (rc == -EHOSTUNREACH) { /* gateway is not on a local net */ - return 0; /* ignore the route entry */ - } else { - CERROR("Error %d creating route %s %d %s\n", rc, - libcfs_net2str(net), hops, - libcfs_nid2str(gateway)); + 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; + } + + /* 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); + + if (route2->lr_gateway == route->lr_gateway) { + add_route = 0; + break; } - return rc; - } - - LASSERT (!the_lnet.ln_shutdown); - - rnet2 = lnet_find_net_locked(net); - if (rnet2 == NULL) { - /* new network */ - cfs_list_add_tail(&rnet->lrn_list, &the_lnet.ln_remote_nets); - rnet2 = rnet; - } - - /* Search for a duplicate route (it's a NOOP if it is) */ - add_route = 1; - cfs_list_for_each (e, &rnet2->lrn_routes) { - lnet_route_t *route2 = cfs_list_entry(e, lnet_route_t, lr_list); - - if (route2->lr_gateway == route->lr_gateway) { - add_route = 0; - break; - } - - /* our lookups must be true */ - LASSERT (route2->lr_gateway->lp_nid != gateway); - } - - if (add_route) { - lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */ + + /* our lookups must be true */ + LASSERT(route2->lr_gateway->lpni_nid != gateway); + } + + if (add_route) { + lnet_peer_ni_addref_locked(route->lr_gateway); /* +1 for notify */ lnet_add_route_to_rnet(rnet2, route); - ni = route->lr_gateway->lp_ni; + ni = lnet_get_next_ni_locked(route->lr_gateway->lpni_net, NULL); lnet_net_unlock(LNET_LOCK_EX); /* XXX Assume alive */ - if (ni->ni_lnd->lnd_notify != NULL) - (ni->ni_lnd->lnd_notify)(ni, gateway, 1); + 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_decref_locked(route->lr_gateway); + lnet_peer_ni_decref_locked(route->lr_gateway); lnet_net_unlock(LNET_LOCK_EX); - if (!add_route) + rc = 0; + + if (!add_route) { + rc = -EEXIST; LIBCFS_FREE(route, sizeof(*route)); + } - if (rnet != rnet2) - LIBCFS_FREE(rnet, sizeof(*rnet)); + if (rnet != rnet2) + LIBCFS_FREE(rnet, sizeof(*rnet)); - return 0; + /* indicate to startup the router checker if configured */ + wake_up(&the_lnet.ln_rc_waitq); + + return rc; } int lnet_check_routes(void) { - lnet_remotenet_t *rnet; - lnet_route_t *route; - lnet_route_t *route2; - cfs_list_t *e1; - cfs_list_t *e2; - int cpt; + 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(); - cfs_list_for_each(e1, &the_lnet.ln_remote_nets) { - rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list); + 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; - cfs_list_for_each(e2, &rnet->lrn_routes) { - lnet_nid_t nid1; - lnet_nid_t nid2; - int net; + route2 = NULL; + list_for_each(e2, &rnet->lrn_routes) { + lnet_nid_t nid1; + lnet_nid_t nid2; + int net; - route = cfs_list_entry(e2, lnet_route_t, lr_list); + route = list_entry(e2, lnet_route_t, + lr_list); - if (route2 == NULL) { - route2 = route; - continue; - } + if (route2 == NULL) { + route2 = route; + continue; + } - if (route->lr_gateway->lp_ni == - route2->lr_gateway->lp_ni) - continue; + if (route->lr_gateway->lpni_net == + route2->lr_gateway->lpni_net) + continue; - nid1 = route->lr_gateway->lp_nid; - nid2 = route2->lr_gateway->lp_nid; - net = rnet->lrn_net; + nid1 = route->lr_gateway->lpni_nid; + nid2 = route2->lr_gateway->lpni_nid; + net = rnet->lrn_net; - lnet_net_unlock(cpt); + 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; + CERROR("Routes to %s via %s and %s not " + "supported\n", + libcfs_net2str(net), + libcfs_nid2str(nid1), + libcfs_nid2str(nid2)); + return -EINVAL; + } } } @@ -460,98 +496,150 @@ lnet_check_routes(void) } int -lnet_del_route (__u32 net, lnet_nid_t gw_nid) +lnet_del_route(__u32 net, lnet_nid_t gw_nid) { - struct lnet_peer *gateway; - lnet_remotenet_t *rnet; - lnet_route_t *route; - cfs_list_t *e1; - cfs_list_t *e2; - int rc = -ENOENT; + 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)); + 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) */ + /* NB Caller may specify either all routes via the given gateway + * or a specific route entry actual NIDs) */ - again: 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); - cfs_list_for_each (e1, &the_lnet.ln_remote_nets) { - rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list); +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; + if (!(net == LNET_NIDNET(LNET_NID_ANY) || + net == rnet->lrn_net)) + continue; - cfs_list_for_each (e2, &rnet->lrn_routes) { - route = cfs_list_entry(e2, lnet_route_t, lr_list); + 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)) + gw_nid == gateway->lpni_nid)) continue; - cfs_list_del(&route->lr_list); - cfs_list_del(&route->lr_gwlist); - the_lnet.ln_remote_nets_version++; + list_del(&route->lr_list); + list_del(&route->lr_gwlist); + the_lnet.ln_remote_nets_version++; - if (cfs_list_empty(&rnet->lrn_routes)) - cfs_list_del(&rnet->lrn_list); - else - rnet = NULL; + 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_peer_ni_decref_locked(gateway); lnet_net_unlock(LNET_LOCK_EX); - LIBCFS_FREE(route, sizeof (*route)); + LIBCFS_FREE(route, sizeof(*route)); - if (rnet != NULL) - LIBCFS_FREE(rnet, sizeof(*rnet)); + if (rnet != NULL) + LIBCFS_FREE(rnet, sizeof(*rnet)); - rc = 0; - goto again; - } - } + rc = 0; + lnet_net_lock(LNET_LOCK_EX); + 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); + 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_nid_t *gateway, __u32 *alive, __u32 *priority) { - cfs_list_t *e1; - cfs_list_t *e2; - lnet_remotenet_t *rnet; - lnet_route_t *route; - int cpt; + 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(); - cfs_list_for_each (e1, &the_lnet.ln_remote_nets) { - rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list); - - cfs_list_for_each (e2, &rnet->lrn_routes) { - route = cfs_list_entry(e2, lnet_route_t, lr_list); - - if (idx-- == 0) { - *net = rnet->lrn_net; - *hops = route->lr_hops; - *gateway = route->lr_gateway->lp_nid; - *alive = route->lr_gateway->lp_alive; - lnet_net_unlock(cpt); - return 0; + 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; + } } } } @@ -561,21 +649,21 @@ lnet_get_route(int idx, __u32 *net, __u32 *hops, } void -lnet_swap_pinginfo(lnet_ping_info_t *info) +lnet_swap_pinginfo(struct lnet_ping_info *info) { - int i; - lnet_ni_status_t *stat; + int i; + struct lnet_ni_status *stat; __swab32s(&info->pi_magic); - __swab32s(&info->pi_version); - __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; + __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; } /** @@ -585,49 +673,63 @@ lnet_swap_pinginfo(lnet_ping_info_t *info) 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 *rtr; + struct lnet_ping_info *info = rcd->rcd_pinginfo; + struct lnet_peer_ni *gw = rcd->rcd_gateway; + lnet_route_t *rte; - if (!gw->lp_alive) + 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->lp_nid), info->pi_magic); - gw->lp_ping_version = LNET_PROTO_PING_UNKNOWN; + libcfs_nid2str(gw->lpni_nid), info->pi_magic); + gw->lpni_ping_feats = LNET_PING_FEAT_INVAL; + spin_unlock(&gw->lpni_lock); return; } - gw->lp_ping_version = info->pi_version; - if (gw->lp_ping_version == LNET_PROTO_PING_VERSION_1) - return; /* v1 doesn't carry NI status info */ + 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->lp_ping_version != LNET_PROTO_PING_VERSION) { - CDEBUG(D_NET, "%s: Unexpected version 0x%x\n", - libcfs_nid2str(gw->lp_nid), gw->lp_ping_version); - gw->lp_ping_version = LNET_PROTO_PING_UNKNOWN; - return; + if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0) { + spin_unlock(&gw->lpni_lock); + return; /* can't carry NI status info */ } - cfs_list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) { - int ptl_status = LNET_NI_STATUS_INVALID; + 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++) { - lnet_ni_status_t *stat = &info->pi_ni[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->lp_nid)); - gw->lp_ping_version = LNET_PROTO_PING_UNKNOWN; + libcfs_nid2str(gw->lpni_nid)); + gw->lpni_ping_feats = LNET_PING_FEAT_INVAL; + spin_unlock(&gw->lpni_lock); return; } @@ -635,51 +737,52 @@ lnet_parse_rc_info(lnet_rc_data_t *rcd) continue; if (stat->ns_status == LNET_NI_STATUS_DOWN) { - if (LNET_NETTYP(LNET_NIDNET(nid)) != PTLLND) - down++; - else if (ptl_status != LNET_NI_STATUS_UP) - ptl_status = LNET_NI_STATUS_DOWN; + down++; continue; } if (stat->ns_status == LNET_NI_STATUS_UP) { - if (LNET_NIDNET(nid) == rtr->lr_net) { + if (LNET_NIDNET(nid) == rte->lr_net) { up = 1; break; } - /* ptl NIs are considered down only when - * they're all down */ - if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND) - ptl_status = LNET_NI_STATUS_UP; continue; } CDEBUG(D_NET, "%s: Unexpected status 0x%x\n", - libcfs_nid2str(gw->lp_nid), stat->ns_status); - gw->lp_ping_version = LNET_PROTO_PING_UNKNOWN; + 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 */ - rtr->lr_downis = 0; + rte->lr_downis = 0; continue; } - rtr->lr_downis = down + (ptl_status == LNET_NI_STATUS_DOWN); + /* 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 void lnet_router_checker_event(lnet_event_t *event) { - lnet_rc_data_t *rcd = event->md.user_ptr; - struct lnet_peer *lp; + lnet_rc_data_t *rcd = event->md.user_ptr; + struct lnet_peer_ni *lp; LASSERT(rcd != NULL); if (event->unlinked) { LNetInvalidateHandle(&rcd->rcd_mdh); return; - } + } LASSERT(event->type == LNET_EVENT_SEND || event->type == LNET_EVENT_REPLY); @@ -690,14 +793,14 @@ lnet_router_checker_event(lnet_event_t *event) /* 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) { + 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 (event->type == LNET_EVENT_SEND) { - lp->lp_ping_notsent = 0; + lp->lpni_ping_notsent = 0; if (event->status == 0) goto out; } @@ -718,48 +821,66 @@ lnet_router_checker_event(lnet_event_t *event) lnet_parse_rc_info(rcd); out: - lnet_net_unlock(lp->lp_cpt); + lnet_net_unlock(lp->lpni_cpt); } -void +static void lnet_wait_known_routerstate(void) { - lnet_peer_t *rtr; - cfs_list_t *entry; - int all_known; + struct lnet_peer_ni *rtr; + struct list_head *entry; + int all_known; + + LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); + for (;;) { + int cpt = lnet_net_lock_current(); - for (;;) { - int cpt = lnet_net_lock_current(); + all_known = 1; + list_for_each(entry, &the_lnet.ln_routers) { + rtr = list_entry(entry, struct lnet_peer_ni, + lpni_rtr_list); - all_known = 1; - cfs_list_for_each (entry, &the_lnet.ln_routers) { - rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list); + spin_lock(&rtr->lpni_lock); - if (rtr->lp_alive_count == 0) { - all_known = 0; - break; - } - } + if (rtr->lpni_alive_count == 0) { + all_known = 0; + spin_unlock(&rtr->lpni_lock); + break; + } + spin_unlock(&rtr->lpni_lock); + } lnet_net_unlock(cpt); - if (all_known) - return; + if (all_known) + return; -#ifndef __KERNEL__ - lnet_router_checker(); -#endif - cfs_pause(cfs_time_seconds(1)); - } + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1)); + } } 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; + } + } + } +} + +static void lnet_update_ni_status_locked(void) { - lnet_ni_t *ni; - long now; + lnet_ni_t *ni = NULL; + time64_t now; int timeout; LASSERT(the_lnet.ln_routing); @@ -767,9 +888,9 @@ lnet_update_ni_status_locked(void) timeout = router_ping_timeout + MAX(live_router_check_interval, dead_router_check_interval); - now = cfs_time_current_sec(); - cfs_list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) { - if (ni->ni_lnd->lnd_type == LOLND) + 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) @@ -795,18 +916,18 @@ lnet_update_ni_status_locked(void) } } -void +static void lnet_destroy_rc_data(lnet_rc_data_t *rcd) { - LASSERT(cfs_list_empty(&rcd->rcd_list)); + LASSERT(list_empty(&rcd->rcd_list)); /* detached from network */ LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh)); if (rcd->rcd_gateway != NULL) { - int cpt = rcd->rcd_gateway->lp_cpt; + int cpt = rcd->rcd_gateway->lpni_cpt; lnet_net_lock(cpt); - lnet_peer_decref_locked(rcd->rcd_gateway); + lnet_peer_ni_decref_locked(rcd->rcd_gateway); lnet_net_unlock(cpt); } @@ -816,64 +937,63 @@ lnet_destroy_rc_data(lnet_rc_data_t *rcd) LIBCFS_FREE(rcd, sizeof(*rcd)); } -lnet_rc_data_t * -lnet_create_rc_data_locked(lnet_peer_t *gateway) +static lnet_rc_data_t * +lnet_create_rc_data_locked(struct lnet_peer_ni *gateway) { lnet_rc_data_t *rcd = NULL; - lnet_ping_info_t *pi; + struct lnet_ping_info *pi; int rc; int i; - lnet_net_unlock(gateway->lp_cpt); + lnet_net_unlock(gateway->lpni_cpt); LIBCFS_ALLOC(rcd, sizeof(*rcd)); if (rcd == NULL) goto out; LNetInvalidateHandle(&rcd->rcd_mdh); - CFS_INIT_LIST_HEAD(&rcd->rcd_list); + INIT_LIST_HEAD(&rcd->rcd_list); LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE); if (pi == NULL) goto out; - memset(pi, 0, LNET_PINGINFO_SIZE); - 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); + for (i = 0; i < LNET_MAX_RTR_NIS; i++) { + pi->pi_ni[i].ns_nid = LNET_NID_ANY; + pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID; + } + rcd->rcd_pinginfo = pi; + + LASSERT(!LNetHandleIsInvalid(the_lnet.ln_rc_eqh)); + rc = LNetMDBind((lnet_md_t){.start = pi, + .user_ptr = rcd, + .length = LNET_PINGINFO_SIZE, + .threshold = LNET_MD_THRESH_INF, + .options = LNET_MD_TRUNCATE, + .eq_handle = the_lnet.ln_rc_eqh}, + LNET_UNLINK, + &rcd->rcd_mdh); + if (rc < 0) { + CERROR("Can't bind MD: %d\n", rc); goto out; } LASSERT(rc == 0); - lnet_net_lock(gateway->lp_cpt); + lnet_net_lock(gateway->lpni_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); + if (!lnet_isrouter(gateway) || gateway->lpni_rcd != NULL) { + lnet_net_unlock(gateway->lpni_cpt); goto out; } - lnet_peer_addref_locked(gateway); + lnet_peer_ni_addref_locked(gateway); rcd->rcd_gateway = gateway; - gateway->lp_rcd = rcd; - gateway->lp_ping_notsent = 0; + gateway->lpni_rcd = rcd; + gateway->lpni_ping_notsent = 0; return rcd; - out: +out: if (rcd != NULL) { if (!LNetHandleIsInvalid(rcd->rcd_mdh)) { rc = LNetMDUnlink(rcd->rcd_mdh); @@ -882,235 +1002,166 @@ lnet_create_rc_data_locked(lnet_peer_t *gateway) lnet_destroy_rc_data(rcd); } - lnet_net_lock(gateway->lp_cpt); - return gateway->lp_rcd; + lnet_net_lock(gateway->lpni_cpt); + return gateway->lpni_rcd; } static int -lnet_router_check_interval (lnet_peer_t *rtr) +lnet_router_check_interval (struct lnet_peer_ni *rtr) { - int secs; + int secs; - secs = rtr->lp_alive ? live_router_check_interval : - dead_router_check_interval; - if (secs < 0) - secs = 0; + secs = rtr->lpni_alive ? live_router_check_interval : + dead_router_check_interval; + if (secs < 0) + secs = 0; - return secs; + return secs; } static void -lnet_ping_router_locked (lnet_peer_t *rtr) +lnet_ping_router_locked (struct lnet_peer_ni *rtr) { - lnet_rc_data_t *rcd = NULL; - cfs_time_t now = cfs_time_current(); - int secs; + lnet_rc_data_t *rcd = NULL; + cfs_time_t now = cfs_time_current(); + int secs; + struct lnet_ni *ni; - lnet_peer_addref_locked(rtr); + lnet_peer_ni_addref_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 (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 */ - lnet_ni_notify_locked(rtr->lp_ni, rtr); + 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_decref_locked(rtr); + lnet_peer_ni_decref_locked(rtr); return; } - rcd = rtr->lp_rcd != NULL ? - rtr->lp_rcd : lnet_create_rc_data_locked(rtr); + rcd = rtr->lpni_rcd != NULL ? + rtr->lpni_rcd : lnet_create_rc_data_locked(rtr); if (rcd == NULL) return; - secs = lnet_router_check_interval(rtr); + 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); + 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->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; + 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->lp_nid; - id.pid = LUSTRE_SRV_LNET_PID; - CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id)); + id.nid = rtr->lpni_nid; + id.pid = LNET_PID_LUSTRE; + CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id)); - rtr->lp_ping_notsent = 1; - rtr->lp_ping_timestamp = now; + rtr->lpni_ping_notsent = 1; + rtr->lpni_ping_timestamp = now; mdh = rcd->rcd_mdh; - if (rtr->lp_ping_deadline == 0) { - rtr->lp_ping_deadline = \ + if (rtr->lpni_ping_deadline == 0) { + rtr->lpni_ping_deadline = cfs_time_shift(router_ping_timeout); } - lnet_net_unlock(rtr->lp_cpt); + lnet_net_unlock(rtr->lpni_cpt); rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL, LNET_PROTO_PING_MATCHBITS, 0); - lnet_net_lock(rtr->lp_cpt); - if (rc != 0) - rtr->lp_ping_notsent = 0; /* no event pending */ - } + lnet_net_lock(rtr->lpni_cpt); + if (rc != 0) + rtr->lpni_ping_notsent = 0; /* no event pending */ + } - lnet_peer_decref_locked(rtr); - return; + lnet_peer_ni_decref_locked(rtr); + return; } int lnet_router_checker_start(void) { - int rc; - int eqsz; -#ifndef __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; - cfs_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 = LUSTRE_SRV_LNET_PID; - - 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); - - if (check_routers_before_use && - dead_router_check_interval <= 0) { - LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be" - " set if 'check_routers_before_use' is set" - "\n"); - return -EINVAL; - } - - if (!the_lnet.ln_routing && - live_router_check_interval <= 0 && - dead_router_check_interval <= 0) - return 0; - -#ifdef __KERNEL__ - cfs_sema_init(&the_lnet.ln_rc_signal, 0); - /* EQ size doesn't matter; the callback is guaranteed to get every - * event */ - eqsz = 0; - rc = LNetEQAlloc(eqsz, lnet_router_checker_event, - &the_lnet.ln_rc_eqh); -#else - rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE, - &the_lnet.ln_rc_eqh); -#endif - if (rc != 0) { - CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc); - return -ENOMEM; - } - - the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING; -#ifdef __KERNEL__ - rc = cfs_create_thread(lnet_router_checker, NULL, 0); - if (rc < 0) { - CERROR("Can't start router checker thread: %d\n", rc); - /* block until event callback signals exit */ - cfs_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; - } -#endif - - if (check_routers_before_use) { - /* Note that a helpful side-effect of pinging all known routers - * at startup is that it makes them drop stale connections they - * may have to a previous instance of me. */ - lnet_wait_known_routerstate(); - } - - return 0; + 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); + /* 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_router_checker_stop (void) { - int rc; + int rc; - if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN) - return; + if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN) + return; - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); + 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); -#ifdef __KERNEL__ /* block until event callback signals exit */ - cfs_down(&the_lnet.ln_rc_signal); -#else - lnet_router_checker(); -#endif + down(&the_lnet.ln_rc_signal); LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); - rc = LNetEQFree(the_lnet.ln_rc_eqh); - LASSERT (rc == 0); - return; + rc = LNetEQFree(the_lnet.ln_rc_eqh); + LASSERT(rc == 0); + return; } static void @@ -1118,94 +1169,116 @@ lnet_prune_rc_data(int wait_unlink) { lnet_rc_data_t *rcd; lnet_rc_data_t *tmp; - lnet_peer_t *lp; - cfs_list_t head; - int i = 2; + struct lnet_peer_ni *lp; + struct list_head head; + int i = 2; if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING && - cfs_list_empty(&the_lnet.ln_rcd_deathrow) && - cfs_list_empty(&the_lnet.ln_rcd_zombie))) + list_empty(&the_lnet.ln_rcd_deathrow) && + list_empty(&the_lnet.ln_rcd_zombie))) return; - CFS_INIT_LIST_HEAD(&head); + 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 */ - cfs_list_for_each_entry(lp, &the_lnet.ln_routers, - lp_rtr_list) { - if (lp->lp_rcd == NULL) + list_for_each_entry(lp, &the_lnet.ln_routers, + lpni_rtr_list) { + if (lp->lpni_rcd == NULL) continue; - LASSERT(cfs_list_empty(&lp->lp_rcd->rcd_list)); - cfs_list_add(&lp->lp_rcd->rcd_list, - &the_lnet.ln_rcd_deathrow); - lp->lp_rcd = NULL; + 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 */ - cfs_list_splice_init(&the_lnet.ln_rcd_deathrow, &head); + list_splice_init(&the_lnet.ln_rcd_deathrow, &head); - if (!cfs_list_empty(&head)) { + if (!list_empty(&head)) { lnet_net_unlock(LNET_LOCK_EX); - cfs_list_for_each_entry(rcd, &head, rcd_list) + list_for_each_entry(rcd, &head, rcd_list) LNetMDUnlink(rcd->rcd_mdh); lnet_net_lock(LNET_LOCK_EX); - } + } - cfs_list_splice_init(&head, &the_lnet.ln_rcd_zombie); + list_splice_init(&head, &the_lnet.ln_rcd_zombie); /* release all zombie RCDs */ - while (!cfs_list_empty(&the_lnet.ln_rcd_zombie)) { - cfs_list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie, - rcd_list) { - if (!LNetHandleIsInvalid(rcd->rcd_mdh)) - cfs_list_move(&rcd->rcd_list, &head); - } + 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 && - !cfs_list_empty(&the_lnet.ln_rcd_zombie); + !list_empty(&the_lnet.ln_rcd_zombie); lnet_net_unlock(LNET_LOCK_EX); - while (!cfs_list_empty(&head)) { - rcd = cfs_list_entry(head.next, - lnet_rc_data_t, rcd_list); - cfs_list_del_init(&rcd->rcd_list); + 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) - break; + return; i++; CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting for rc buffers to unlink\n"); - cfs_pause(cfs_time_seconds(1) / 4); + 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; -#if defined(__KERNEL__) && defined(LNET_ROUTER) + return !list_empty(&the_lnet.ln_routers) && + (live_router_check_interval > 0 || + dead_router_check_interval > 0); +} static int lnet_router_checker(void *arg) { - lnet_peer_t *rtr; - cfs_list_t *entry; + struct lnet_peer_ni *rtr; + struct list_head *entry; - cfs_daemonize("router_checker"); - cfs_block_allsigs(); + cfs_block_allsigs(); - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); - - while (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; @@ -1214,10 +1287,11 @@ lnet_router_checker(void *arg) rescan: version = the_lnet.ln_routers_version; - cfs_list_for_each(entry, &the_lnet.ln_routers) { - rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list); + list_for_each(entry, &the_lnet.ln_routers) { + rtr = list_entry(entry, struct lnet_peer_ni, + lpni_rtr_list); - cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid); + cpt2 = rtr->lpni_cpt; if (cpt != cpt2) { lnet_net_unlock(cpt); cpt = cpt2; @@ -1227,14 +1301,14 @@ rescan: goto rescan; } - lnet_ping_router_locked(rtr); + lnet_ping_router_locked(rtr); - /* NB dropped lock */ - if (version != the_lnet.ln_routers_version) { - /* the routers list has changed */ - goto rescan; - } - } + /* 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(); @@ -1243,42 +1317,48 @@ rescan: 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. */ - cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE, - cfs_time_seconds(1)); - } - - LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING); + /* 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; - cfs_up(&the_lnet.ln_rc_signal); - /* The unlink event callback will signal final completion */ - return 0; + up(&the_lnet.ln_rc_signal); + /* The unlink event callback will signal final completion */ + return 0; } void lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages) { - int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); + int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]); - while (--npages >= 0) - cfs_free_page(rb->rb_kiov[npages].kiov_page); + while (--npages >= 0) + __free_page(rb->rb_kiov[npages].kiov_page); - LIBCFS_FREE(rb, sz); + LIBCFS_FREE(rb, sz); } -lnet_rtrbuf_t * +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; + 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) @@ -1288,118 +1368,164 @@ lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt) for (i = 0; i < npages; i++) { page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, - CFS_ALLOC_ZERO | CFS_ALLOC_STD); - if (page == NULL) { - while (--i >= 0) - cfs_free_page(rb->rb_kiov[i].kiov_page); + GFP_KERNEL | __GFP_ZERO); + if (page == NULL) { + while (--i >= 0) + __free_page(rb->rb_kiov[i].kiov_page); - LIBCFS_FREE(rb, sz); - return NULL; - } + 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; - } + rb->rb_kiov[i].kiov_len = PAGE_SIZE; + rb->rb_kiov[i].kiov_offset = 0; + rb->rb_kiov[i].kiov_page = page; + } - return rb; + return rb; } -void -lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp) +static void +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; - LASSERT (cfs_list_empty(&rbp->rbp_msgs)); - LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers); - - while (!cfs_list_empty(&rbp->rbp_bufs)) { - LASSERT (rbp->rbp_credits > 0); + INIT_LIST_HEAD(&tmp); - rb = cfs_list_entry(rbp->rbp_bufs.next, - lnet_rtrbuf_t, rb_list); - cfs_list_del(&rb->rb_list); - lnet_destroy_rtrbuf(rb, npages); - nbuffers++; - } - - LASSERT (rbp->rbp_nbuffers == nbuffers); - LASSERT (rbp->rbp_credits == nbuffers); + 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); - rbp->rbp_nbuffers = rbp->rbp_credits = 0; + /* 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, int cpt) +static int +lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt) { - lnet_rtrbuf_t *rb; - int i; + 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; - if (rbp->rbp_nbuffers != 0) { - LASSERT (rbp->rbp_nbuffers == nbufs); - return 0; - } + 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); - for (i = 0; i < nbufs; i++) { + 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); - if (rb == NULL) { - CERROR("Failed to allocate %d router bufs of %d pages\n", - nbufs, rbp->rbp_npages); - return -ENOMEM; - } + lnet_net_lock(cpt); + rbp->rbp_req_nbuffers = old_req_nbufs; + lnet_net_unlock(cpt); - rbp->rbp_nbuffers++; - rbp->rbp_credits++; - rbp->rbp_mincredits++; - cfs_list_add(&rb->rb_list, &rbp->rbp_bufs); + 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); - /* No allocation "under fire" */ - /* Otherwise we'd need code to schedule blocked msgs etc */ - LASSERT (!the_lnet.ln_routing); - } + 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); + } - LASSERT (rbp->rbp_credits == nbufs); - return 0; + 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_rtrpools_free(void) +lnet_rtrpools_free(int keep_pools) { lnet_rtrbufpool_t *rtrp; - int i; + 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[0]); - lnet_rtrpool_free_bufs(&rtrp[1]); - lnet_rtrpool_free_bufs(&rtrp[2]); + 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); } - cfs_percpt_free(the_lnet.ln_rtrpools); - the_lnet.ln_rtrpools = NULL; + if (!keep_pools) { + cfs_percpt_free(the_lnet.ln_rtrpools); + the_lnet.ln_rtrpools = NULL; + } } static int -lnet_nrb_tiny_calculate(int npages) +lnet_nrb_tiny_calculate(void) { int nrbs = LNET_NRB_TINY; @@ -1407,7 +1533,7 @@ lnet_nrb_tiny_calculate(int npages) LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when " "routing enabled\n", tiny_router_buffers); - return -1; + return -EINVAL; } if (tiny_router_buffers > 0) @@ -1418,7 +1544,7 @@ lnet_nrb_tiny_calculate(int npages) } static int -lnet_nrb_small_calculate(int npages) +lnet_nrb_small_calculate(void) { int nrbs = LNET_NRB_SMALL; @@ -1426,7 +1552,7 @@ lnet_nrb_small_calculate(int npages) LCONSOLE_ERROR_MSG(0x10c, "small_router_buffers=%d invalid when " "routing enabled\n", small_router_buffers); - return -1; + return -EINVAL; } if (small_router_buffers > 0) @@ -1437,7 +1563,7 @@ lnet_nrb_small_calculate(int npages) } static int -lnet_nrb_large_calculate(int npages) +lnet_nrb_large_calculate(void) { int nrbs = LNET_NRB_LARGE; @@ -1445,7 +1571,7 @@ lnet_nrb_large_calculate(int npages) LCONSOLE_ERROR_MSG(0x10c, "large_router_buffers=%d invalid when " "routing enabled\n", large_router_buffers); - return -1; + return -EINVAL; } if (large_router_buffers > 0) @@ -1459,38 +1585,36 @@ int lnet_rtrpools_alloc(int im_a_router) { lnet_rtrbufpool_t *rtrp; - int large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT; - int small_pages = 1; 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(0); + 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(small_pages); + nrb_small = lnet_nrb_small_calculate(); if (nrb_small < 0) return -EINVAL; - nrb_large = lnet_nrb_large_calculate(large_pages); + nrb_large = lnet_nrb_large_calculate(); if (nrb_large < 0) return -EINVAL; @@ -1504,18 +1628,23 @@ lnet_rtrpools_alloc(int im_a_router) } cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) { - lnet_rtrpool_init(&rtrp[0], 0); - rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny, i); + 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[1], small_pages); - rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small, i); + 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[2], large_pages); - rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large, i); + 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; } @@ -1523,206 +1652,196 @@ lnet_rtrpools_alloc(int im_a_router) lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_routing = 1; lnet_net_unlock(LNET_LOCK_EX); - return 0; failed: - lnet_rtrpools_free(); + lnet_rtrpools_free(0); return rc; } -int -lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) +static int +lnet_rtrpools_adjust_helper(int tiny, int small, int large) { - struct lnet_peer *lp = NULL; - cfs_time_t now = cfs_time_current(); - int cpt = lnet_cpt_of_nid(nid); - - LASSERT (!cfs_in_interrupt ()); - - CDEBUG (D_NET, "%s notifying %s: %s\n", - (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid), - libcfs_nid2str(nid), - alive ? "up" : "down"); - - if (ni != NULL && - LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) { - CWARN ("Ignoring notification of %s %s by %s (different net)\n", - libcfs_nid2str(nid), alive ? "birth" : "death", - libcfs_nid2str(ni->ni_nid)); - return -EINVAL; - } - - /* can't do predictions... */ - if (cfs_time_after(when, now)) { - CWARN ("Ignoring prediction from %s of %s %s " - "%ld seconds in the future\n", - (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid), - libcfs_nid2str(nid), alive ? "up" : "down", - cfs_duration_sec(cfs_time_sub(when, now))); - return -EINVAL; - } - - if (ni != NULL && !alive && /* LND telling me she's down */ - !auto_down) { /* auto-down disabled */ - CDEBUG(D_NET, "Auto-down disabled\n"); - return 0; - } + int nrb = 0; + int rc = 0; + int i; + lnet_rtrbufpool_t *rtrp; - lnet_net_lock(cpt); + /* 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; + } + } - 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; - } + return 0; +} - /* We can't fully trust LND on reporting exact peer last_alive - * if he notifies us about dead peer. For example ksocklnd can - * call us with when == _time_when_the_node_was_booted_ if - * no connections were successfully established */ - if (ni != NULL && !alive && when < lp->lp_last_alive) - when = lp->lp_last_alive; +int +lnet_rtrpools_adjust(int tiny, int small, int large) +{ + /* this function doesn't revert the changes if adding new buffers + * failed. It's up to the user space caller to revert the + * changes. */ - lnet_notify_locked(lp, ni == NULL, alive, when); + if (!the_lnet.ln_routing) + return 0; - lnet_ni_notify_locked(ni, lp); + return lnet_rtrpools_adjust_helper(tiny, small, large); +} - lnet_peer_decref_locked(lp); +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_unlock(cpt); - return 0; + 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; } -EXPORT_SYMBOL(lnet_notify); void -lnet_get_tunables (void) +lnet_rtrpools_disable(void) { - return; -} + if (!the_lnet.ln_routing) + return; -#else + lnet_net_lock(LNET_LOCK_EX); + the_lnet.ln_routing = 0; + the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED; -int -lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) -{ - return -EOPNOTSUPP; + tiny_router_buffers = 0; + small_router_buffers = 0; + large_router_buffers = 0; + lnet_net_unlock(LNET_LOCK_EX); + lnet_rtrpools_free(1); } -void -lnet_router_checker (void) +int +lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) { - static time_t last = 0; - static int running = 0; - - time_t now = cfs_time_current_sec(); - int interval = now - last; - int rc; - __u64 version; - lnet_peer_t *rtr; - - /* It's no use to call me again within a sec - all intervals and - * timeouts are measured in seconds */ - if (last != 0 && interval < 2) - return; - - 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); + 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; + } - lnet_net_lock(0); + /* 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; + } - version = the_lnet.ln_routers_version; - cfs_list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) { - lnet_ping_router_locked(rtr); - LASSERT (version == the_lnet.ln_routers_version); - } + 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_unlock(0); + lnet_net_lock(cpt); - running = 0; /* lock only needed for the recursion check */ - return; -} + if (the_lnet.ln_shutdown) { + lnet_net_unlock(cpt); + return -ESHUTDOWN; + } -/* NB lnet_peers_start_down depends on me, - * so must be called before any peer creation */ -void -lnet_get_tunables (void) -{ - char *s; + 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; + } - s = getenv("LNET_ROUTER_PING_TIMEOUT"); - if (s != NULL) router_ping_timeout = atoi(s); + /* + * 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); + } - s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL"); - if (s != NULL) live_router_check_interval = atoi(s); + /* 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; - s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL"); - if (s != NULL) dead_router_check_interval = atoi(s); + lnet_notify_locked(lp, ni == NULL, alive, when); - /* This replaces old lnd_notify mechanism */ - check_routers_before_use = 1; - if (dead_router_check_interval <= 0) - dead_router_check_interval = 30; -} + if (ni != NULL) + lnet_ni_notify_locked(ni, lp); -void -lnet_rtrpools_free(void) -{ -} + lnet_peer_ni_decref_locked(lp); -int -lnet_rtrpools_alloc(int im_a_arouter) -{ - return 0; + lnet_net_unlock(cpt); + return 0; } - -#endif +EXPORT_SYMBOL(lnet_notify);