X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Frouter.c;h=a380035e94198fda87210f9495f965fd0ee08f75;hp=a8f816c92fb64387cb38a8e5e50fd4af9cadd486;hb=2f5ad4e8cf26406a7aa732efa4d813dbcced55d9;hpb=14a611ad680e89523abbcab0a3310511ab808ba8 diff --git a/lnet/lnet/router.c b/lnet/lnet/router.c index a8f816c..a380035 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 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2011, Whamcloud, Inc. * * This file is part of Portals * http://sourceforge.net/projects/sandiaportals/ @@ -26,19 +26,26 @@ #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_LARGE_MIN 256 /* min value for each CPT */ +#define LNET_NRB_LARGE (LNET_NRB_LARGE_MIN * 4) + 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"); @@ -78,19 +85,19 @@ CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444, "Assume routers are down and ping them before use"); static int avoid_asym_router_failure = 0; -CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0444, +CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644, "Avoid asymmetrical failures: reserved, use at your own risk"); -int dead_router_check_interval = 0; -CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0444, +static int dead_router_check_interval = 0; +CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644, "Seconds between dead router health checks (<= 0 to disable)"); -int live_router_check_interval = 0; -CFS_MODULE_PARM(live_router_check_interval, "i", int, 0444, +static int live_router_check_interval = 0; +CFS_MODULE_PARM(live_router_check_interval, "i", int, 0644, "Seconds between live router health checks (<= 0 to disable)"); -int router_ping_timeout = 50; -CFS_MODULE_PARM(router_ping_timeout, "i", int, 0444, +static int router_ping_timeout = 50; +CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644, "Seconds to wait for the reply to a router health query"); int @@ -122,27 +129,24 @@ lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_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; @@ -154,43 +158,42 @@ lnet_do_notify (lnet_peer_t *lp) lp->lp_notify = 0; if (notifylnd && ni->ni_lnd->lnd_notify != NULL) { - LNET_UNLOCK(); - - /* A new notification could happen now; I'll handle it - * when control returns to me */ + lnet_net_unlock(lp->lp_cpt); - (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive); + /* A new notification could happen now; I'll handle it + * when control returns to me */ - LNET_LOCK(); - } - } + (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive); - lp->lp_notifying = 0; + lnet_net_lock(lp->lp_cpt); + } + } - LNET_UNLOCK(); + lp->lp_notifying = 0; } static void lnet_rtr_addref_locked(lnet_peer_t *lp) { - LASSERT (lp->lp_refcount > 0); - LASSERT (lp->lp_rtr_refcount >= 0); + 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; + cfs_list_t *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); + 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; } - list_add(&lp->lp_rtr_list, pos); + cfs_list_add(&lp->lp_rtr_list, pos); /* addref for the_lnet.ln_routers */ lnet_peer_addref_locked(lp); the_lnet.ln_routers_version++; @@ -200,18 +203,21 @@ lnet_rtr_addref_locked(lnet_peer_t *lp) 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) { - if (lp->lp_rcd != NULL) { - list_add(&lp->lp_rcd->rcd_list, - &the_lnet.ln_zombie_rcd); + 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(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; } - list_del(&lp->lp_rtr_list); + cfs_list_del(&lp->lp_rtr_list); /* decref for the_lnet.ln_routers */ lnet_peer_decref_locked(lp); the_lnet.ln_routers_version++; @@ -222,12 +228,12 @@ lnet_remotenet_t * lnet_find_net_locked (__u32 net) { lnet_remotenet_t *rnet; - struct list_head *tmp; + cfs_list_t *tmp; LASSERT (!the_lnet.ln_shutdown); - list_for_each (tmp, &the_lnet.ln_remote_nets) { - rnet = list_entry(tmp, lnet_remotenet_t, lrn_list); + cfs_list_for_each (tmp, &the_lnet.ln_remote_nets) { + rnet = cfs_list_entry(tmp, lnet_remotenet_t, lrn_list); if (rnet->lrn_net == net) return rnet; @@ -235,10 +241,67 @@ lnet_find_net_locked (__u32 net) 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; +} + +/* 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; + 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); +} + int lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway) { - struct list_head *e; + cfs_list_t *e; lnet_remotenet_t *rnet; lnet_remotenet_t *rnet2; lnet_route_t *route; @@ -276,21 +339,24 @@ lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway) CFS_INIT_LIST_HEAD(&rnet->lrn_routes); rnet->lrn_net = net; route->lr_hops = hops; + route->lr_net = net; - LNET_LOCK(); - - rc = lnet_nid2peer_locked(&route->lr_gateway, gateway); - if (rc != 0) { - LNET_UNLOCK(); + lnet_net_lock(LNET_LOCK_EX); - LIBCFS_FREE(route, sizeof(*route)); - LIBCFS_FREE(rnet, sizeof(*rnet)); + rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX); + if (rc != 0) { + lnet_net_unlock(LNET_LOCK_EX); - if (rc == -EHOSTUNREACH) /* gateway is not on a local net */ - return 0; /* ignore the route entry */ + LIBCFS_FREE(route, sizeof(*route)); + LIBCFS_FREE(rnet, sizeof(*rnet)); - CERROR("Error %d creating route %s %d %s\n", rc, - libcfs_net2str(net), hops, libcfs_nid2str(gateway)); + 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)); + } return rc; } @@ -299,14 +365,14 @@ lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway) rnet2 = lnet_find_net_locked(net); if (rnet2 == NULL) { /* new network */ - list_add_tail(&rnet->lrn_list, &the_lnet.ln_remote_nets); + 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; - list_for_each (e, &rnet2->lrn_routes) { - lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list); + 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; @@ -318,77 +384,89 @@ lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway) } if (add_route) { - ni = route->lr_gateway->lp_ni; - lnet_ni_addref_locked(ni); + lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */ + lnet_add_route_to_rnet(rnet2, route); - list_add_tail(&route->lr_list, &rnet2->lrn_routes); - the_lnet.ln_remote_nets_version++; + ni = route->lr_gateway->lp_ni; + lnet_net_unlock(LNET_LOCK_EX); - lnet_rtr_addref_locked(route->lr_gateway); + /* XXX Assume alive */ + if (ni->ni_lnd->lnd_notify != NULL) + (ni->ni_lnd->lnd_notify)(ni, gateway, 1); - LNET_UNLOCK(); + lnet_net_lock(LNET_LOCK_EX); + } - /* XXX Assume alive */ - if (ni->ni_lnd->lnd_notify != NULL) - (ni->ni_lnd->lnd_notify)(ni, gateway, 1); + /* -1 for notify or !add_route */ + lnet_peer_decref_locked(route->lr_gateway); + lnet_net_unlock(LNET_LOCK_EX); - lnet_ni_decref(ni); - } else { - lnet_peer_decref_locked(route->lr_gateway); - LNET_UNLOCK(); - LIBCFS_FREE(route, sizeof(*route)); - } + if (!add_route) + LIBCFS_FREE(route, sizeof(*route)); - if (rnet != rnet2) - LIBCFS_FREE(rnet, sizeof(*rnet)); + if (rnet != rnet2) + LIBCFS_FREE(rnet, sizeof(*rnet)); - return 0; + return 0; } int -lnet_check_routes (void) +lnet_check_routes(void) { - 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; + lnet_remotenet_t *rnet; + lnet_route_t *route; + lnet_route_t *route2; + cfs_list_t *e1; + cfs_list_t *e2; + int cpt; + + 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); + + route2 = NULL; + cfs_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); + + 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; + } + } + + lnet_net_unlock(cpt); + return 0; } 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; - struct list_head *e1; - struct list_head *e2; + cfs_list_t *e1; + cfs_list_t *e2; int rc = -ENOENT; CDEBUG(D_NET, "Del route: net %s : gw %s\n", @@ -398,33 +476,36 @@ lnet_del_route (__u32 net, lnet_nid_t gw_nid) * or a specific route entry actual NIDs) */ again: - LNET_LOCK(); + lnet_net_lock(LNET_LOCK_EX); - list_for_each (e1, &the_lnet.ln_remote_nets) { - rnet = list_entry(e1, lnet_remotenet_t, lrn_list); + cfs_list_for_each (e1, &the_lnet.ln_remote_nets) { + rnet = cfs_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); + cfs_list_for_each (e2, &rnet->lrn_routes) { + route = cfs_list_entry(e2, lnet_route_t, lr_list); - if (!(gw_nid == LNET_NID_ANY || - gw_nid == route->lr_gateway->lp_nid)) - continue; + gateway = route->lr_gateway; + if (!(gw_nid == LNET_NID_ANY || + gw_nid == gateway->lp_nid)) + continue; - list_del(&route->lr_list); + cfs_list_del(&route->lr_list); + cfs_list_del(&route->lr_gwlist); the_lnet.ln_remote_nets_version++; - if (list_empty(&rnet->lrn_routes)) - list_del(&rnet->lrn_list); + if (cfs_list_empty(&rnet->lrn_routes)) + cfs_list_del(&rnet->lrn_list); else rnet = NULL; - lnet_rtr_decref_locked(route->lr_gateway); - lnet_peer_decref_locked(route->lr_gateway); - LNET_UNLOCK(); + lnet_rtr_decref_locked(gateway); + lnet_peer_decref_locked(gateway); + + lnet_net_unlock(LNET_LOCK_EX); LIBCFS_FREE(route, sizeof (*route)); @@ -436,8 +517,8 @@ lnet_del_route (__u32 net, lnet_nid_t gw_nid) } } - LNET_UNLOCK(); - return rc; + lnet_net_unlock(LNET_LOCK_EX); + return rc; } void @@ -447,44 +528,46 @@ lnet_destroy_routes (void) } int -lnet_get_route (int idx, __u32 *net, __u32 *hops, - lnet_nid_t *gateway, __u32 *alive) +lnet_get_route(int idx, __u32 *net, __u32 *hops, + lnet_nid_t *gateway, __u32 *alive) { - struct list_head *e1; - struct list_head *e2; - lnet_remotenet_t *rnet; - lnet_route_t *route; + cfs_list_t *e1; + cfs_list_t *e2; + lnet_remotenet_t *rnet; + lnet_route_t *route; + int cpt; - LNET_LOCK(); + cpt = lnet_net_lock_current(); - list_for_each (e1, &the_lnet.ln_remote_nets) { - rnet = list_entry(e1, lnet_remotenet_t, lrn_list); + cfs_list_for_each (e1, &the_lnet.ln_remote_nets) { + rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list); - list_for_each (e2, &rnet->lrn_routes) { - route = list_entry(e2, lnet_route_t, lr_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_UNLOCK(); - return 0; - } - } - } - - LNET_UNLOCK(); - return -ENOENT; + lnet_net_unlock(cpt); + return 0; + } + } + } + + lnet_net_unlock(cpt); + return -ENOENT; } void lnet_swap_pinginfo(lnet_ping_info_t *info) { - int i; - lnet_ni_status_t *stat; + int i; + lnet_ni_status_t *stat; - __swab32s(&info->pi_version); + __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++) { @@ -495,101 +578,163 @@ lnet_swap_pinginfo(lnet_ping_info_t *info) return; } -/* Returns # of down NIs, or negative error codes; ignore downed NIs - * if a NI in 'net' is up */ -int -lnet_router_down_ni(lnet_peer_t *rtr, __u32 net) +/** + * 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) { - int i; - int down = 0; - int ptl_up = 0; - int ptl_down = 0; - lnet_ping_info_t *info; - - if (!avoid_asym_router_failure) - return -ENOENT; - - if (rtr->lp_rcd == NULL) - return -EINVAL; - - if (!rtr->lp_alive) - return -EINVAL; /* stale lp_rcd */ - - info = rtr->lp_rcd->rcd_pinginfo; - LASSERT (info != NULL); - - /* NB always racing with network! */ - if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) { - lnet_swap_pinginfo(info); - } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) { - CDEBUG(D_NETERROR, "%s: Unexpected magic %08x\n", - libcfs_nid2str(rtr->lp_nid), info->pi_magic); - return -EPROTO; - } - - if (info->pi_version == LNET_PROTO_PING_VERSION1) - return -ENOENT; /* v1 doesn't carry NI status info */ - - if (info->pi_version != LNET_PROTO_PING_VERSION) { - CDEBUG(D_NETERROR, "%s: Unexpected version 0x%x\n", - libcfs_nid2str(rtr->lp_nid), info->pi_version); - return -EPROTO; - } - - 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_NETERROR, "%s: unexpected LNET_NID_ANY\n", - libcfs_nid2str(rtr->lp_nid)); - return -EPROTO; - } - - if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND) - continue; - - if (stat->ns_status == LNET_NI_STATUS_DOWN) { - if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND) - ptl_down = 1; - else - down++; - continue; - } - - if (stat->ns_status != LNET_NI_STATUS_UP) { - CDEBUG(D_NETERROR, "%s: Unexpected status 0x%x\n", - libcfs_nid2str(rtr->lp_nid), stat->ns_status); - return -EPROTO; - } - - /* ignore downed NIs if there's a NI up for dest network */ - if (LNET_NIDNET(nid) == net) - return 0; - - if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND) - ptl_up = 1; - } + lnet_ping_info_t *info = rcd->rcd_pinginfo; + struct lnet_peer *gw = rcd->rcd_gateway; + lnet_route_t *rtr; + + if (!gw->lp_alive) + return; + + if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) + lnet_swap_pinginfo(info); + + /* NB always racing with network! */ + if (info->pi_magic != LNET_PROTO_PING_MAGIC) { + CDEBUG(D_NET, "%s: Unexpected magic %08x\n", + libcfs_nid2str(gw->lp_nid), info->pi_magic); + gw->lp_ping_feats = LNET_PING_FEAT_INVAL; + return; + } + + gw->lp_ping_feats = info->pi_features; + if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) { + CDEBUG(D_NET, "%s: Unexpected features 0x%x\n", + libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats); + return; /* nothing I can understand */ + } + + if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0) + return; /* can't carry NI status info */ + + cfs_list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) { + int ptl_status = LNET_NI_STATUS_INVALID; + int down = 0; + int up = 0; + int i; + + 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) { + if (LNET_NETTYP(LNET_NIDNET(nid)) != PTLLND) + down++; + else if (ptl_status != LNET_NI_STATUS_UP) + ptl_status = LNET_NI_STATUS_DOWN; + continue; + } + + if (stat->ns_status == LNET_NI_STATUS_UP) { + if (LNET_NIDNET(nid) == rtr->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_feats = LNET_PING_FEAT_INVAL; + return; + } + + if (up) { /* ignore downed NIs if NI for dest network is up */ + rtr->lr_downis = 0; + continue; + } + rtr->lr_downis = down + (ptl_status == LNET_NI_STATUS_DOWN); + } +} - /* ptl NIs are considered down only when they're all down */ - return down + (ptl_up ? 0 : ptl_down); +static void +lnet_router_checker_event(lnet_event_t *event) +{ + lnet_rc_data_t *rcd = event->md.user_ptr; + struct lnet_peer *lp; + + LASSERT(rcd != NULL); + + if (event->unlinked) { + LNetInvalidateHandle(&rcd->rcd_mdh); + return; + } + + LASSERT(event->type == LNET_EVENT_SEND || + event->type == LNET_EVENT_REPLY); + + lp = rcd->rcd_gateway; + LASSERT(lp != NULL); + + /* NB: it's called with holding lnet_res_lock, we have a few + * places need to hold both locks at the same time, please take + * care of lock ordering */ + lnet_net_lock(lp->lp_cpt); + if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) { + /* ignore if no longer a router or rcd is replaced */ + goto out; + } + + if (event->type == LNET_EVENT_SEND) { + lp->lp_ping_notsent = 0; + if (event->status == 0) + goto out; + } + + /* LNET_EVENT_REPLY */ + /* A successful REPLY means the router is up. If _any_ comms + * to the router fail I assume it's down (this will happen if + * we ping alive routers to try to detect router death before + * apps get burned). */ + + lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current()); + /* The router checker will wake up very shortly and do the + * actual notification. + * XXX If 'lp' stops being a router before then, it will still + * have the notification pending!!! */ + + if (avoid_asym_router_failure && event->status == 0) + lnet_parse_rc_info(rcd); + + out: + lnet_net_unlock(lp->lp_cpt); } void lnet_wait_known_routerstate(void) { lnet_peer_t *rtr; - struct list_head *entry; + cfs_list_t *entry; int all_known; LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); for (;;) { - LNET_LOCK(); + 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); + cfs_list_for_each (entry, &the_lnet.ln_routers) { + rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list); if (rtr->lp_alive_count == 0) { all_known = 0; @@ -597,7 +742,7 @@ lnet_wait_known_routerstate(void) } } - LNET_UNLOCK(); + lnet_net_unlock(cpt); if (all_known) return; @@ -609,140 +754,87 @@ lnet_wait_known_routerstate(void) } } -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_rc_data_t *rcd = event->md.user_ptr; - lnet_peer_t *lp; - lnet_nid_t nid; - - if (event->unlinked) { - if (rcd != NULL) { - LNetInvalidateHandle(&rcd->rcd_mdh); - return; - } - - /* The router checker thread has unlinked the default rc_md - * and exited. */ - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING); - the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKED; -#ifdef __KERNEL__ - mutex_up(&the_lnet.ln_rc_signal); -#endif - return; - } - - LASSERT (event->type == LNET_EVENT_SEND || - event->type == LNET_EVENT_REPLY); - - nid = (event->type == LNET_EVENT_SEND) ? - event->target.nid : event->initiator.nid; - - lp = lnet_find_peer_locked(nid); - if (lp == NULL) { - /* router may have been removed */ - CDEBUG(D_NET, "Router %s not found\n", libcfs_nid2str(nid)); - return; - } - - if (event->type == LNET_EVENT_SEND) /* re-enable another ping */ - lp->lp_ping_notsent = 0; - - if (lnet_isrouter(lp) && /* ignore if no longer a router */ - (event->status != 0 || - event->type == LNET_EVENT_REPLY)) { - - /* A successful REPLY means the router is up. If _any_ comms - * to the router fail I assume it's down (this will happen if - * we ping alive routers to try to detect router death before - * apps get burned). */ - - lnet_notify_locked(lp, 1, (event->status == 0), - cfs_time_current()); - - /* The router checker will wake up very shortly and do the - * actual notification. - * XXX If 'lp' stops being a router before then, it will still - * have the notification pending!!! */ - } - - /* This decref will NOT drop LNET_LOCK (it had to have 1 ref when it - * was in the peer table and I've not dropped the lock, so no-one else - * can have reduced the refcount) */ - LASSERT(lp->lp_refcount > 1); - - lnet_peer_decref_locked(lp); + 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(); + cfs_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); + } } void -lnet_update_ni_status(void) +lnet_destroy_rc_data(lnet_rc_data_t *rcd) { - cfs_time_t now = cfs_time_current(); - lnet_ni_t *ni; - int status; - int timeout; + LASSERT(cfs_list_empty(&rcd->rcd_list)); + /* detached from network */ + LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh)); - LASSERT (the_lnet.ln_routing); + if (rcd->rcd_gateway != NULL) { + int cpt = rcd->rcd_gateway->lp_cpt; - timeout = router_ping_timeout + - MAX(live_router_check_interval, dead_router_check_interval); + lnet_net_lock(cpt); + lnet_peer_decref_locked(rcd->rcd_gateway); + lnet_net_unlock(cpt); + } - LNET_LOCK(); + if (rcd->rcd_pinginfo != NULL) + LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE); - list_for_each_entry (ni, &the_lnet.ln_nis, ni_list) { - lnet_ni_status_t *ns = ni->ni_status; - - LASSERT (ns != NULL); - - status = LNET_NI_STATUS_UP; - if (ni->ni_lnd->lnd_type != LOLND && /* @lo forever alive */ - cfs_time_after(now, cfs_time_add(ni->ni_last_alive, - cfs_time_seconds(timeout)))) - status = LNET_NI_STATUS_DOWN; - - if (ns->ns_status != status) { - ns->ns_status = status; - CDEBUG(D_NET, "NI(%s:%d) status changed to %s\n", - libcfs_nid2str(ni->ni_nid), timeout, - status == LNET_NI_STATUS_UP ? "up" : "down"); - } - } - - LNET_UNLOCK(); + LIBCFS_FREE(rcd, sizeof(*rcd)); } -void -lnet_destroy_rc_data (lnet_rc_data_t *rcd) +lnet_rc_data_t * +lnet_create_rc_data_locked(lnet_peer_t *gateway) { - LASSERT (list_empty(&rcd->rcd_list)); - /* detached from network */ - LASSERT (LNetHandleIsInvalid(rcd->rcd_mdh)); + lnet_rc_data_t *rcd = NULL; + lnet_ping_info_t *pi; + int rc; + int i; - LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE); - LIBCFS_FREE(rcd, sizeof(*rcd)); - return; -} + lnet_net_unlock(gateway->lp_cpt); -lnet_rc_data_t * -lnet_create_rc_data (void) -{ - int i; - int rc; - lnet_ping_info_t *pi; - lnet_rc_data_t *rcd; + LIBCFS_ALLOC(rcd, sizeof(*rcd)); + if (rcd == NULL) + goto out; - LIBCFS_ALLOC(rcd, sizeof(*rcd)); - if (rcd == NULL) - return NULL; + LNetInvalidateHandle(&rcd->rcd_mdh); + CFS_INIT_LIST_HEAD(&rcd->rcd_list); - LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE); - if (pi == NULL) { - LIBCFS_FREE(rcd, sizeof(*rcd)); - return NULL; - } + 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++) { @@ -750,8 +842,6 @@ lnet_create_rc_data (void) pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID; } rcd->rcd_pinginfo = pi; - LNetInvalidateHandle(&rcd->rcd_mdh); - CFS_INIT_LIST_HEAD(&rcd->rcd_list); LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh)); rc = LNetMDBind((lnet_md_t){.start = pi, @@ -764,11 +854,35 @@ lnet_create_rc_data (void) &rcd->rcd_mdh); if (rc < 0) { CERROR("Can't bind MD: %d\n", rc); - lnet_destroy_rc_data(rcd); - return NULL; - } - LASSERT (rc == 0); - return rcd; + goto out; + } + LASSERT(rc == 0); + + lnet_net_lock(gateway->lp_cpt); + /* router table changed or someone has created rcd for this gateway */ + if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) { + lnet_net_unlock(gateway->lp_cpt); + goto out; + } + + lnet_peer_addref_locked(gateway); + rcd->rcd_gateway = gateway; + gateway->lp_rcd = rcd; + gateway->lp_ping_notsent = 0; + + return rcd; + + out: + if (rcd != NULL) { + if (!LNetHandleIsInvalid(rcd->rcd_mdh)) { + rc = LNetMDUnlink(rcd->rcd_mdh); + LASSERT(rc == 0); + } + lnet_destroy_rc_data(rcd); + } + + lnet_net_lock(gateway->lp_cpt); + return gateway->lp_rcd; } static int @@ -787,7 +901,6 @@ lnet_router_check_interval (lnet_peer_t *rtr) static void lnet_ping_router_locked (lnet_peer_t *rtr) { - int newrcd = 0; lnet_rc_data_t *rcd = NULL; cfs_time_t now = cfs_time_current(); int secs; @@ -798,30 +911,21 @@ lnet_ping_router_locked (lnet_peer_t *rtr) cfs_time_after(now, rtr->lp_ping_deadline)) lnet_notify_locked(rtr, 1, 0, now); - if (avoid_asym_router_failure && rtr->lp_rcd == NULL) - newrcd = 1; - - LNET_UNLOCK(); + /* Run any outstanding notifications */ + lnet_ni_notify_locked(rtr->lp_ni, rtr); - /* Run any outstanding notifications */ - lnet_do_notify(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); + return; + } - if (newrcd) - rcd = lnet_create_rc_data(); + rcd = rtr->lp_rcd != NULL ? + rtr->lp_rcd : lnet_create_rc_data_locked(rtr); - LNET_LOCK(); - - if (!lnet_isrouter(rtr)) { - lnet_peer_decref_locked(rtr); - if (rcd != NULL) - list_add(&rcd->rcd_list, &the_lnet.ln_zombie_rcd); - return; /* router table changed! */ - } - - if (rcd != NULL) { - LASSERT (rtr->lp_rcd == NULL); - rtr->lp_rcd = rcd; - } + if (rcd == NULL) + return; secs = lnet_router_check_interval(rtr); @@ -845,18 +949,20 @@ lnet_ping_router_locked (lnet_peer_t *rtr) rtr->lp_ping_notsent = 1; rtr->lp_ping_timestamp = now; - mdh = (rtr->lp_rcd == NULL) ? the_lnet.ln_rc_mdh : - rtr->lp_rcd->rcd_mdh; - if (rtr->lp_ping_deadline == 0) - rtr->lp_ping_deadline = cfs_time_shift(router_ping_timeout); + mdh = rcd->rcd_mdh; + + if (rtr->lp_ping_deadline == 0) { + rtr->lp_ping_deadline = + cfs_time_shift(router_ping_timeout); + } - LNET_UNLOCK(); + lnet_net_unlock(rtr->lp_cpt); - rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL, - LNET_PROTO_PING_MATCHBITS, 0); + rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL, + LNET_PROTO_PING_MATCHBITS, 0); - LNET_LOCK(); + lnet_net_lock(rtr->lp_cpt); if (rc != 0) rtr->lp_ping_notsent = 0; /* no event pending */ } @@ -868,9 +974,6 @@ lnet_ping_router_locked (lnet_peer_t *rtr) int lnet_router_checker_start(void) { - static lnet_ping_info_t pinginfo; - - lnet_md_t md; int rc; int eqsz; #ifndef __KERNEL__ @@ -882,13 +985,13 @@ lnet_router_checker_start(void) LASSERT (check_routers_before_use); LASSERT (dead_router_check_interval > 0); - LNET_LOCK(); + 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) { + 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; @@ -899,7 +1002,7 @@ lnet_router_checker_start(void) id.nid = rtr->lp_nid; id.pid = LUSTRE_SRV_LNET_PID; - LNET_UNLOCK(); + lnet_net_unlock(0); rc = LNetSetAsync(id, 1); if (rc != 0) { @@ -908,12 +1011,12 @@ lnet_router_checker_start(void) return rc; } - LNET_LOCK(); - /* NB router list doesn't change in userspace */ - LASSERT (version == the_lnet.ln_routers_version); - } + lnet_net_lock(0); + /* NB router list doesn't change in userspace */ + LASSERT(version == the_lnet.ln_routers_version); + } - LNET_UNLOCK(); + lnet_net_unlock(0); if (nrtr == 0) { CDEBUG(D_NET, @@ -946,10 +1049,10 @@ lnet_router_checker_start(void) return 0; #ifdef __KERNEL__ - init_mutex_locked(&the_lnet.ln_rc_signal); + cfs_sema_init(&the_lnet.ln_rc_signal, 0); /* EQ size doesn't matter; the callback is guaranteed to get every * event */ - eqsz = 1; + eqsz = 0; rc = LNetEQAlloc(eqsz, lnet_router_checker_event, &the_lnet.ln_rc_eqh); #else @@ -961,32 +1064,13 @@ lnet_router_checker_start(void) return -ENOMEM; } - memset(&md, 0, sizeof(md)); - md.user_ptr = NULL; - md.start = &pinginfo; - md.length = sizeof(pinginfo); - md.options = LNET_MD_TRUNCATE; - md.threshold = LNET_MD_THRESH_INF; - md.eq_handle = the_lnet.ln_rc_eqh; - rc = LNetMDBind(md, LNET_UNLINK, &the_lnet.ln_rc_mdh); - if (rc < 0) { - CERROR("Can't bind MD: %d\n", rc); - rc = LNetEQFree(the_lnet.ln_rc_eqh); - LASSERT (rc == 0); - return -ENOMEM; - } - LASSERT (rc == 0); - the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING; #ifdef __KERNEL__ - rc = (int)cfs_kernel_thread(lnet_router_checker, NULL, 0); + rc = cfs_create_thread(lnet_router_checker, NULL, 0); if (rc < 0) { CERROR("Can't start router checker thread: %d\n", rc); - the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING; - rc = LNetMDUnlink(the_lnet.ln_rc_mdh); - LASSERT (rc == 0); /* block until event callback signals exit */ - mutex_down(&the_lnet.ln_rc_signal); + 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; @@ -1013,113 +1097,137 @@ lnet_router_checker_stop (void) return; LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); - the_lnet.ln_rc_state = LNET_RC_STATE_STOPTHREAD; + the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING; #ifdef __KERNEL__ - /* block until event callback signals exit */ - mutex_down(&the_lnet.ln_rc_signal); + /* block until event callback signals exit */ + cfs_down(&the_lnet.ln_rc_signal); #else - while (the_lnet.ln_rc_state != LNET_RC_STATE_UNLINKED) { - lnet_router_checker(); - cfs_pause(cfs_time_seconds(1)); - } + lnet_router_checker(); #endif - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED); + LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN); rc = LNetEQFree(the_lnet.ln_rc_eqh); LASSERT (rc == 0); - the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN; return; } -#if defined(__KERNEL__) && defined(LNET_ROUTER) - static void -lnet_prune_zombie_rcd (int wait_unlink) +lnet_prune_rc_data(int wait_unlink) { - lnet_rc_data_t *rcd; - lnet_rc_data_t *tmp; - struct list_head free_rcd; - int i; - __u64 version; + lnet_rc_data_t *rcd; + lnet_rc_data_t *tmp; + lnet_peer_t *lp; + cfs_list_t head; + int i = 2; - CFS_INIT_LIST_HEAD(&free_rcd); + 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))) + return; - LNET_LOCK(); -rescan: - version = the_lnet.ln_routers_version; - list_for_each_entry_safe (rcd, tmp, &the_lnet.ln_zombie_rcd, rcd_list) { - if (LNetHandleIsInvalid(rcd->rcd_mdh)) { - list_del(&rcd->rcd_list); - list_add(&rcd->rcd_list, &free_rcd); - continue; - } + CFS_INIT_LIST_HEAD(&head); - LNET_UNLOCK(); + lnet_net_lock(LNET_LOCK_EX); - LNetMDUnlink(rcd->rcd_mdh); + 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) + continue; - LNET_LOCK(); - if (version != the_lnet.ln_routers_version) - goto rescan; - } + 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; + } + } - i = 2; - while (wait_unlink && !list_empty(&the_lnet.ln_zombie_rcd)) { - rcd = list_entry(the_lnet.ln_zombie_rcd.next, - lnet_rc_data_t, rcd_list); - if (LNetHandleIsInvalid(rcd->rcd_mdh)) { - list_del(&rcd->rcd_list); - list_add(&rcd->rcd_list, &free_rcd); - continue; - } + /* unlink all RCDs on deathrow list */ + cfs_list_splice_init(&the_lnet.ln_rcd_deathrow, &head); - LNET_UNLOCK(); + if (!cfs_list_empty(&head)) { + lnet_net_unlock(LNET_LOCK_EX); - LNetMDUnlink(rcd->rcd_mdh); + cfs_list_for_each_entry(rcd, &head, rcd_list) + LNetMDUnlink(rcd->rcd_mdh); - i++; - CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, - "Waiting for rc buffers to unlink\n"); - cfs_pause(cfs_time_seconds(1)); + lnet_net_lock(LNET_LOCK_EX); + } - LNET_LOCK(); - } + cfs_list_splice_init(&head, &the_lnet.ln_rcd_zombie); - LNET_UNLOCK(); + /* 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(&free_rcd)) { - rcd = list_entry(free_rcd.next, lnet_rc_data_t, rcd_list); - list_del_init(&rcd->rcd_list); - lnet_destroy_rc_data(rcd); - } - return; + wait_unlink = wait_unlink && + !cfs_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); + 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) { - int rc; lnet_peer_t *rtr; - struct list_head *entry; - lnet_process_id_t rtr_id; + cfs_list_t *entry; cfs_daemonize("router_checker"); cfs_block_allsigs(); - rtr_id.pid = LUSTRE_SRV_LNET_PID; - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING); while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) { - __u64 version; + __u64 version; + int cpt; + int cpt2; - LNET_LOCK(); + cpt = lnet_net_lock_current(); rescan: - version = the_lnet.ln_routers_version; + 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); + + cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid); + if (cpt != cpt2) { + lnet_net_unlock(cpt); + cpt = cpt2; + lnet_net_lock(cpt); + /* the routers list has changed */ + if (version != the_lnet.ln_routers_version) + goto rescan; + } - list_for_each (entry, &the_lnet.ln_routers) { - rtr = list_entry(entry, lnet_peer_t, lp_rtr_list); lnet_ping_router_locked(rtr); /* NB dropped lock */ @@ -1129,45 +1237,28 @@ rescan: } } - LNET_UNLOCK(); + if (the_lnet.ln_routing) + lnet_update_ni_status_locked(); - if (the_lnet.ln_routing) - lnet_update_ni_status(); + lnet_net_unlock(cpt); - lnet_prune_zombie_rcd(0); /* don't wait for UNLINK */ + 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(CFS_TASK_INTERRUPTIBLE, - cfs_time_seconds(1)); + cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE, + cfs_time_seconds(1)); } - LNET_LOCK(); + LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING); - list_for_each (entry, &the_lnet.ln_routers) { - rtr = list_entry(entry, lnet_peer_t, lp_rtr_list); + lnet_prune_rc_data(1); /* wait for UNLINK */ - if (rtr->lp_rcd == NULL) - continue; - - LASSERT (list_empty(&rtr->lp_rcd->rcd_list)); - list_add(&rtr->lp_rcd->rcd_list, &the_lnet.ln_zombie_rcd); - rtr->lp_rcd = NULL; - } - - LNET_UNLOCK(); - - lnet_prune_zombie_rcd(1); /* wait for UNLINK */ - - LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD); - the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING; - - rc = LNetMDUnlink(the_lnet.ln_rc_mdh); - LASSERT (rc == 0); - - /* The unlink event callback will signal final completion */ - return 0; + 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; } void @@ -1182,7 +1273,7 @@ lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages) } 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]); @@ -1190,14 +1281,15 @@ lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp) 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, + CFS_ALLOC_ZERO | CFS_ALLOC_STD); if (page == NULL) { while (--i >= 0) cfs_free_page(rb->rb_kiov[i].kiov_page); @@ -1217,19 +1309,22 @@ lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp) void lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp) { - int npages = rbp->rbp_npages; - int nbuffers = 0; - lnet_rtrbuf_t *rb; + int npages = rbp->rbp_npages; + int nbuffers = 0; + lnet_rtrbuf_t *rb; - LASSERT (list_empty(&rbp->rbp_msgs)); + 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 (!list_empty(&rbp->rbp_bufs)) { + while (!cfs_list_empty(&rbp->rbp_bufs)) { LASSERT (rbp->rbp_credits > 0); - rb = list_entry(rbp->rbp_bufs.next, - lnet_rtrbuf_t, rb_list); - list_del(&rb->rb_list); + 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++; } @@ -1241,7 +1336,7 @@ lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp) } int -lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs) +lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt) { lnet_rtrbuf_t *rb; int i; @@ -1252,7 +1347,7 @@ lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs) } for (i = 0; i < nbufs; i++) { - rb = lnet_new_rtrbuf(rbp); + rb = lnet_new_rtrbuf(rbp, cpt); if (rb == NULL) { CERROR("Failed to allocate %d router bufs of %d pages\n", @@ -1263,7 +1358,7 @@ lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs) rbp->rbp_nbuffers++; rbp->rbp_credits++; rbp->rbp_mincredits++; - list_add(&rb->rb_list, &rbp->rbp_bufs); + cfs_list_add(&rb->rb_list, &rbp->rbp_bufs); /* No allocation "under fire" */ /* Otherwise we'd need code to schedule blocked msgs etc */ @@ -1286,29 +1381,92 @@ lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages) } void -lnet_free_rtrpools(void) +lnet_rtrpools_free(void) +{ + 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[0]); + lnet_rtrpool_free_bufs(&rtrp[1]); + lnet_rtrpool_free_bufs(&rtrp[2]); + } + + cfs_percpt_free(the_lnet.ln_rtrpools); + the_lnet.ln_rtrpools = NULL; +} + +static int +lnet_nrb_tiny_calculate(int npages) { - 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(int npages) { - int small_pages = 1; - int large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT; + 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; + } - lnet_rtrpool_init(&the_lnet.ln_rtrpools[0], 0); - lnet_rtrpool_init(&the_lnet.ln_rtrpools[1], small_pages); - lnet_rtrpool_init(&the_lnet.ln_rtrpools[2], large_pages); + if (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(int npages) +{ + 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; + + nrbs /= LNET_CPT_NUMBER; + return max(nrbs, LNET_NRB_LARGE_MIN); +} int -lnet_alloc_rtrpools(int im_a_router) +lnet_rtrpools_alloc(int im_a_router) { - int rc; + 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 */ @@ -1325,60 +1483,63 @@ lnet_alloc_rtrpools(int im_a_router) return -EINVAL; } - if (tiny_router_buffers <= 0) { - LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when " - "routing enabled\n", tiny_router_buffers); - rc = -EINVAL; - goto failed; - } - - rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[0], - tiny_router_buffers); - if (rc != 0) - goto failed; - - if (small_router_buffers <= 0) { - LCONSOLE_ERROR_MSG(0x10d, "small_router_buffers=%d invalid when" - " routing enabled\n", small_router_buffers); - rc = -EINVAL; - goto failed; - } - - rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[1], - small_router_buffers); - if (rc != 0) - goto failed; - - if (large_router_buffers <= 0) { - LCONSOLE_ERROR_MSG(0x10e, "large_router_buffers=%d invalid when" - " routing enabled\n", large_router_buffers); - rc = -EINVAL; - goto failed; - } - - rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[2], - large_router_buffers); - if (rc != 0) - goto failed; - - LNET_LOCK(); - the_lnet.ln_routing = 1; - LNET_UNLOCK(); - - return 0; + nrb_tiny = lnet_nrb_tiny_calculate(0); + if (nrb_tiny < 0) + return -EINVAL; + + nrb_small = lnet_nrb_small_calculate(small_pages); + if (nrb_small < 0) + return -EINVAL; + + nrb_large = lnet_nrb_large_calculate(large_pages); + 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[0], 0); + rc = lnet_rtrpool_alloc_bufs(&rtrp[0], 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); + if (rc != 0) + goto failed; + + lnet_rtrpool_init(&rtrp[2], large_pages); + rc = lnet_rtrpool_alloc_bufs(&rtrp[2], 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_free_rtrpools(); - return rc; + lnet_rtrpools_free(); + return rc; } int -lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) +lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) { - lnet_peer_t *lp = NULL; - cfs_time_t now = cfs_time_current(); + struct lnet_peer *lp = NULL; + cfs_time_t now = cfs_time_current(); + int cpt = lnet_cpt_of_nid(nid); - LASSERT (!in_interrupt ()); + LASSERT (!cfs_in_interrupt ()); CDEBUG (D_NET, "%s notifying %s: %s\n", (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid), @@ -1409,12 +1570,17 @@ lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) return 0; } - LNET_LOCK(); + lnet_net_lock(cpt); - lp = lnet_find_peer_locked(nid); - if (lp == NULL) { - /* nid not found */ - LNET_UNLOCK(); + if (the_lnet.ln_shutdown) { + lnet_net_unlock(cpt); + return -ESHUTDOWN; + } + + lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid); + if (lp == NULL) { + /* nid not found */ + lnet_net_unlock(cpt); CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid)); return 0; } @@ -1428,16 +1594,12 @@ lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when) lnet_notify_locked(lp, ni == NULL, alive, when); - LNET_UNLOCK(); - - lnet_do_notify(lp); + lnet_ni_notify_locked(ni, lp); - LNET_LOCK(); + lnet_peer_decref_locked(lp); - lnet_peer_decref_locked(lp); - - LNET_UNLOCK(); - return 0; + lnet_net_unlock(cpt); + return 0; } EXPORT_SYMBOL(lnet_notify); @@ -1475,22 +1637,21 @@ lnet_router_checker (void) if (last != 0 && interval > MAX(live_router_check_interval, dead_router_check_interval)) - CDEBUG(D_NETERROR, "Checker(%d/%d) not called for %d seconds\n", - live_router_check_interval, dead_router_check_interval, - interval); + CNETERR("Checker(%d/%d) not called for %d seconds\n", + live_router_check_interval, dead_router_check_interval, + interval); - LNET_LOCK(); - LASSERT (!running); /* recursion check */ - running = 1; - LNET_UNLOCK(); + LASSERT(LNET_CPT_NUMBER == 1); - last = now; + lnet_net_lock(0); + LASSERT(!running); /* recursion check */ + running = 1; + lnet_net_unlock(0); - if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD) { - the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING; - rc = LNetMDUnlink(the_lnet.ln_rc_mdh); - LASSERT (rc == 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) { @@ -1511,31 +1672,30 @@ lnet_router_checker (void) LASSERT (rc == 1); - LNET_LOCK(); lnet_router_checker_event(&ev); - LNET_UNLOCK(); } - if (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED || - the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING) { + 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_LOCK(); + lnet_net_lock(0); version = the_lnet.ln_routers_version; - list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) { + 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); } - LNET_UNLOCK(); + lnet_net_unlock(0); - running = 0; /* lock only needed for the recursion check */ - return; + running = 0; /* lock only needed for the recursion check */ + return; } /* NB lnet_peers_start_down depends on me, @@ -1561,17 +1721,12 @@ lnet_get_tunables (void) } void -lnet_free_rtrpools (void) -{ -} - -void -lnet_init_rtrpools (void) +lnet_rtrpools_free(void) { } int -lnet_alloc_rtrpools (int im_a_arouter) +lnet_rtrpools_alloc(int im_a_arouter) { return 0; }