X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Flib-move.c;h=615d755593c58490c374a7e61cb96c2c805c1ae3;hp=969cb0763ac6ec8c4a0056769ec3de29b27c7ff4;hb=275016fcae729f8467dc09aa15aa2b4bb2690875;hpb=70616605dd44be37068f4e1a4745a2f8b90eb1f5 diff --git a/lnet/lnet/lib-move.c b/lnet/lnet/lib-move.c index 969cb07..615d755 100644 --- a/lnet/lnet/lib-move.c +++ b/lnet/lnet/lib-move.c @@ -194,8 +194,7 @@ lnet_fail_nid(lnet_nid_t nid, unsigned int threshold) if (tp->tp_threshold == 0 || /* needs culling anyway */ nid == LNET_NID_ANY || /* removing all entries */ tp->tp_nid == nid) { /* matched this one */ - list_del(&tp->tp_list); - list_add(&tp->tp_list, &cull); + list_move(&tp->tp_list, &cull); } } @@ -233,8 +232,7 @@ fail_peer (lnet_nid_t nid, int outgoing) /* only cull zombies on outgoing tests, * since we may be at interrupt priority on * incoming messages. */ - list_del(&tp->tp_list); - list_add(&tp->tp_list, &cull); + list_move(&tp->tp_list, &cull); } continue; } @@ -248,8 +246,7 @@ fail_peer (lnet_nid_t nid, int outgoing) if (outgoing && tp->tp_threshold == 0) { /* see above */ - list_del(&tp->tp_list); - list_add(&tp->tp_list, &cull); + list_move(&tp->tp_list, &cull); } } break; @@ -757,7 +754,7 @@ static void lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg) { void *priv = msg->msg_private; - int rc; + int rc; LASSERT (!in_interrupt ()); LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND || @@ -794,117 +791,55 @@ lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg) return rc; } -/* - * This function can be called from two paths: - * 1. when sending a message - * 2. when decommiting a message (lnet_msg_decommit_tx()) - * In both these cases the peer_ni should have it's reference count - * acquired by the caller and therefore it is safe to drop the spin - * lock before calling lnd_query() - */ -static void -lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp) -{ - time64_t last_alive = 0; - int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni); - - LASSERT(lnet_peer_aliveness_enabled(lp)); - LASSERT(ni->ni_net->net_lnd->lnd_query != NULL); - - lnet_net_unlock(cpt); - (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive); - lnet_net_lock(cpt); - - lp->lpni_last_query = ktime_get_seconds(); - - if (last_alive != 0) /* NI has updated timestamp */ - lp->lpni_last_alive = last_alive; -} - -/* NB: always called with lnet_net_lock held */ -static inline int -lnet_peer_is_alive(struct lnet_peer_ni *lp, time64_t now) +static bool +lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now) { - int alive; time64_t deadline; - LASSERT (lnet_peer_aliveness_enabled(lp)); + deadline = lpni->lpni_last_alive + + lpni->lpni_net->net_tunables.lct_peer_timeout; /* - * Trust lnet_notify() if it has more recent aliveness news, but - * ignore the initial assumed death (see lnet_peers_start_down()). + * assume peer_ni is alive as long as we're within the configured + * peer timeout */ - spin_lock(&lp->lpni_lock); - if (!lp->lpni_alive && lp->lpni_alive_count > 0 && - lp->lpni_timestamp >= lp->lpni_last_alive) { - spin_unlock(&lp->lpni_lock); - return 0; - } + if (deadline > now) + return false; - deadline = lp->lpni_last_alive + - lp->lpni_net->net_tunables.lct_peer_timeout; - alive = deadline > now; - - /* - * Update obsolete lp_alive except for routers assumed to be dead - * initially, because router checker would update aliveness in this - * case, and moreover lpni_last_alive at peer creation is assumed. - */ - if (alive && !lp->lpni_alive && - !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) { - spin_unlock(&lp->lpni_lock); - lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive); - } else { - spin_unlock(&lp->lpni_lock); - } - - return alive; + return true; } - /* NB: returns 1 when alive, 0 when dead, negative when error; * may drop the lnet_net_lock */ static int -lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp) +lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni, + struct lnet_msg *msg) { time64_t now = ktime_get_seconds(); - if (!lnet_peer_aliveness_enabled(lp)) + if (!lnet_peer_aliveness_enabled(lpni)) return -ENODEV; - if (lnet_peer_is_alive(lp, now)) - return 1; - /* - * Peer appears dead, but we should avoid frequent NI queries (at - * most once per lnet_queryinterval seconds). + * If we're resending a message, let's attempt to send it even if + * the peer is down to fulfill our resend quota on the message */ - if (lp->lpni_last_query != 0) { - static const int lnet_queryinterval = 1; - time64_t next_query; - - next_query = lp->lpni_last_query + lnet_queryinterval; - - if (now < next_query) { - if (lp->lpni_alive) - CWARN("Unexpected aliveness of peer %s: " - "%lld < %lld (%d/%d)\n", - libcfs_nid2str(lp->lpni_nid), - now, next_query, - lnet_queryinterval, - lp->lpni_net->net_tunables.lct_peer_timeout); - return 0; - } - } + if (msg->msg_retry_count > 0) + return 1; - /* query NI for latest aliveness news */ - lnet_ni_query_locked(ni, lp); + /* try and send recovery messages irregardless */ + if (msg->msg_recovery) + return 1; - if (lnet_peer_is_alive(lp, now)) + /* always send any responses */ + if (msg->msg_type == LNET_MSG_ACK || + msg->msg_type == LNET_MSG_REPLY) return 1; - lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive); - return 0; + if (!lnet_is_peer_deadline_passed(lpni, now)) + return true; + + return lnet_is_peer_ni_alive(lpni); } /** @@ -930,12 +865,15 @@ lnet_post_send_locked(struct lnet_msg *msg, int do_send) LASSERT(!do_send || msg->msg_tx_delayed); LASSERT(!msg->msg_receiving); LASSERT(msg->msg_tx_committed); + /* can't get here if we're sending to the loopback interface */ + LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid); /* NB 'lp' is always the next hop */ if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 && - lnet_peer_alive_locked(ni, lp) == 0) { - the_lnet.ln_counters[cpt]->drop_count++; - the_lnet.ln_counters[cpt]->drop_length += msg->msg_len; + lnet_peer_alive_locked(ni, lp, msg) == 0) { + the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++; + the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += + msg->msg_len; lnet_net_unlock(cpt); if (msg->msg_txpeer) lnet_incr_stats(&msg->msg_txpeer->lpni_stats, @@ -948,10 +886,9 @@ lnet_post_send_locked(struct lnet_msg *msg, int do_send) CNETERR("Dropping message for %s: peer not alive\n", libcfs_id2str(msg->msg_target)); - if (do_send) { - msg->msg_health_status = LNET_MSG_STATUS_LOCAL_DROPPED; + msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED; + if (do_send) lnet_finalize(msg, -EHOSTUNREACH); - } lnet_net_lock(cpt); return -EHOSTUNREACH; @@ -966,6 +903,8 @@ lnet_post_send_locked(struct lnet_msg *msg, int do_send) libcfs_id2str(msg->msg_target)); if (do_send) { msg->msg_no_resend = true; + CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n", + msg, libcfs_id2str(msg->msg_target)); lnet_finalize(msg, -ECANCELED); } @@ -1051,39 +990,47 @@ lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv) * sets do_recv FALSE and I don't do the unlock/send/lock bit. * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if * received or OK to receive */ - struct lnet_peer_ni *lp = msg->msg_rxpeer; + struct lnet_peer_ni *lpni = msg->msg_rxpeer; + struct lnet_peer *lp; struct lnet_rtrbufpool *rbp; struct lnet_rtrbuf *rb; - LASSERT (msg->msg_iov == NULL); - LASSERT (msg->msg_kiov == NULL); - LASSERT (msg->msg_niov == 0); - LASSERT (msg->msg_routing); - LASSERT (msg->msg_receiving); - LASSERT (!msg->msg_sending); + LASSERT(msg->msg_iov == NULL); + LASSERT(msg->msg_kiov == NULL); + LASSERT(msg->msg_niov == 0); + LASSERT(msg->msg_routing); + LASSERT(msg->msg_receiving); + LASSERT(!msg->msg_sending); + LASSERT(lpni->lpni_peer_net); + LASSERT(lpni->lpni_peer_net->lpn_peer); + + lp = lpni->lpni_peer_net->lpn_peer; /* non-lnet_parse callers only receive delayed messages */ LASSERT(!do_recv || msg->msg_rx_delayed); if (!msg->msg_peerrtrcredit) { - spin_lock(&lp->lpni_lock); - LASSERT((lp->lpni_rtrcredits < 0) == - !list_empty(&lp->lpni_rtrq)); + /* lpni_lock protects the credit manipulation */ + spin_lock(&lpni->lpni_lock); + /* lp_lock protects the lp_rtrq */ + spin_lock(&lp->lp_lock); msg->msg_peerrtrcredit = 1; - lp->lpni_rtrcredits--; - if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits) - lp->lpni_minrtrcredits = lp->lpni_rtrcredits; + lpni->lpni_rtrcredits--; + if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits) + lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits; - if (lp->lpni_rtrcredits < 0) { + if (lpni->lpni_rtrcredits < 0) { /* must have checked eager_recv before here */ LASSERT(msg->msg_rx_ready_delay); msg->msg_rx_delayed = 1; - list_add_tail(&msg->msg_list, &lp->lpni_rtrq); - spin_unlock(&lp->lpni_lock); + list_add_tail(&msg->msg_list, &lp->lp_rtrq); + spin_unlock(&lp->lp_lock); + spin_unlock(&lpni->lpni_lock); return LNET_CREDIT_WAIT; } - spin_unlock(&lp->lpni_lock); + spin_unlock(&lp->lp_lock); + spin_unlock(&lpni->lpni_lock); } rbp = lnet_msg2bufpool(msg); @@ -1212,15 +1159,6 @@ lnet_return_tx_credits_locked(struct lnet_msg *msg) } if (txpeer != NULL) { - /* - * TODO: - * Once the patch for the health comes in we need to set - * the health of the peer ni to bad when we fail to send - * a message. - * int status = msg->msg_ev.status; - * if (status != 0) - * lnet_set_peer_ni_health_locked(txpeer, false) - */ msg->msg_txpeer = NULL; lnet_peer_ni_decref_locked(txpeer); } @@ -1253,6 +1191,7 @@ lnet_drop_routed_msgs_locked(struct list_head *list, int cpt) 0, 0, 0, msg->msg_hdr.payload_length); list_del_init(&msg->msg_list); msg->msg_no_resend = true; + msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR; lnet_finalize(msg, -ECANCELED); } @@ -1262,7 +1201,8 @@ lnet_drop_routed_msgs_locked(struct list_head *list, int cpt) void lnet_return_rx_credits_locked(struct lnet_msg *msg) { - struct lnet_peer_ni *rxpeer = msg->msg_rxpeer; + struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer; + struct lnet_peer *lp; struct lnet_ni *rxni = msg->msg_rxni; struct lnet_msg *msg2; @@ -1312,45 +1252,70 @@ lnet_return_rx_credits_locked(struct lnet_msg *msg) routing_off: if (msg->msg_peerrtrcredit) { + LASSERT(rxpeerni); + LASSERT(rxpeerni->lpni_peer_net); + LASSERT(rxpeerni->lpni_peer_net->lpn_peer); + + lp = rxpeerni->lpni_peer_net->lpn_peer; + /* give back peer router credits */ msg->msg_peerrtrcredit = 0; - spin_lock(&rxpeer->lpni_lock); - LASSERT((rxpeer->lpni_rtrcredits < 0) == - !list_empty(&rxpeer->lpni_rtrq)); + spin_lock(&rxpeerni->lpni_lock); + spin_lock(&lp->lp_lock); - rxpeer->lpni_rtrcredits++; + rxpeerni->lpni_rtrcredits++; /* drop all messages which are queued to be routed on that * peer. */ if (!the_lnet.ln_routing) { struct list_head drop; INIT_LIST_HEAD(&drop); - list_splice_init(&rxpeer->lpni_rtrq, &drop); - spin_unlock(&rxpeer->lpni_lock); + list_splice_init(&lp->lp_rtrq, &drop); + spin_unlock(&lp->lp_lock); + spin_unlock(&rxpeerni->lpni_lock); lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt); - } else if (rxpeer->lpni_rtrcredits <= 0) { - msg2 = list_entry(rxpeer->lpni_rtrq.next, + } else if (!list_empty(&lp->lp_rtrq)) { + int msg2_cpt; + + msg2 = list_entry(lp->lp_rtrq.next, struct lnet_msg, msg_list); list_del(&msg2->msg_list); - spin_unlock(&rxpeer->lpni_lock); + msg2_cpt = msg2->msg_rx_cpt; + spin_unlock(&lp->lp_lock); + spin_unlock(&rxpeerni->lpni_lock); + /* + * messages on the lp_rtrq can be from any NID in + * the peer, which means they might have different + * cpts. We need to make sure we lock the right + * one. + */ + if (msg2_cpt != msg->msg_rx_cpt) { + lnet_net_unlock(msg->msg_rx_cpt); + lnet_net_lock(msg2_cpt); + } (void) lnet_post_routed_recv_locked(msg2, 1); + if (msg2_cpt != msg->msg_rx_cpt) { + lnet_net_unlock(msg2_cpt); + lnet_net_lock(msg->msg_rx_cpt); + } } else { - spin_unlock(&rxpeer->lpni_lock); + spin_unlock(&lp->lp_lock); + spin_unlock(&rxpeerni->lpni_lock); } } if (rxni != NULL) { msg->msg_rxni = NULL; lnet_ni_decref_locked(rxni, msg->msg_rx_cpt); } - if (rxpeer != NULL) { + if (rxpeerni != NULL) { msg->msg_rxpeer = NULL; - lnet_peer_ni_decref_locked(rxpeer); + lnet_peer_ni_decref_locked(rxpeerni); } } static int -lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2) +lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2) { if (p1->lpni_txqnob < p2->lpni_txqnob) return 1; @@ -1367,14 +1332,133 @@ lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2) return 0; } +static struct lnet_peer_ni * +lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid, + struct lnet_peer *peer, + struct lnet_peer_net *peer_net) +{ + /* + * Look at the peer NIs for the destination peer that connect + * to the chosen net. If a peer_ni is preferred when using the + * best_ni to communicate, we use that one. If there is no + * preferred peer_ni, or there are multiple preferred peer_ni, + * the available transmit credits are used. If the transmit + * credits are equal, we round-robin over the peer_ni. + */ + struct lnet_peer_ni *lpni = NULL; + struct lnet_peer_ni *best_lpni = NULL; + int best_lpni_credits = INT_MIN; + bool preferred = false; + bool ni_is_pref; + int best_lpni_healthv = 0; + int lpni_healthv; + + while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) { + /* + * if the best_ni we've chosen aleady has this lpni + * preferred, then let's use it + */ + if (best_ni) { + ni_is_pref = lnet_peer_is_pref_nid_locked(lpni, + best_ni->ni_nid); + CDEBUG(D_NET, "%s ni_is_pref = %d\n", + libcfs_nid2str(best_ni->ni_nid), ni_is_pref); + } else { + ni_is_pref = false; + } + + lpni_healthv = atomic_read(&lpni->lpni_healthv); + + if (best_lpni) + CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n", + libcfs_nid2str(lpni->lpni_nid), + lpni->lpni_txcredits, best_lpni_credits, + lpni->lpni_seq, best_lpni->lpni_seq); + + /* pick the healthiest peer ni */ + if (lpni_healthv < best_lpni_healthv) { + continue; + } else if (lpni_healthv > best_lpni_healthv) { + best_lpni_healthv = lpni_healthv; + /* if this is a preferred peer use it */ + } else if (!preferred && ni_is_pref) { + preferred = true; + } else if (preferred && !ni_is_pref) { + /* + * this is not the preferred peer so let's ignore + * it. + */ + continue; + } else if (lpni->lpni_txcredits < best_lpni_credits) { + /* + * We already have a peer that has more credits + * available than this one. No need to consider + * this peer further. + */ + continue; + } else if (lpni->lpni_txcredits == best_lpni_credits) { + /* + * The best peer found so far and the current peer + * have the same number of available credits let's + * make sure to select between them using Round + * Robin + */ + if (best_lpni) { + if (best_lpni->lpni_seq <= lpni->lpni_seq) + continue; + } + } + + best_lpni = lpni; + best_lpni_credits = lpni->lpni_txcredits; + } + + /* if we still can't find a peer ni then we can't reach it */ + if (!best_lpni) { + __u32 net_id = (peer_net) ? peer_net->lpn_net_id : + LNET_NIDNET(dst_nid); + CDEBUG(D_NET, "no peer_ni found on peer net %s\n", + libcfs_net2str(net_id)); + return NULL; + } + + CDEBUG(D_NET, "sd_best_lpni = %s\n", + libcfs_nid2str(best_lpni->lpni_nid)); + + return best_lpni; +} + +/* + * Prerequisite: the best_ni should already be set in the sd + */ +static inline struct lnet_peer_ni * +lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid, + struct lnet_peer *peer, __u32 net_id) +{ + struct lnet_peer_net *peer_net; + + /* + * The gateway is Multi-Rail capable so now we must select the + * proper peer_ni + */ + peer_net = lnet_peer_get_net_locked(peer, net_id); + + if (!peer_net) { + CERROR("gateway peer %s has no NI on net %s\n", + libcfs_nid2str(peer->lp_primary_nid), + libcfs_net2str(net_id)); + return NULL; + } + + return lnet_select_peer_ni(lni, dst_nid, peer, peer_net); +} + +/* Compare route priorities and hop counts */ static int lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2) { - struct lnet_peer_ni *p1 = r1->lr_gateway; - struct lnet_peer_ni *p2 = r2->lr_gateway; int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops; int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops; - int rc; if (r1->lr_priority < r2->lr_priority) return 1; @@ -1388,52 +1472,43 @@ lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2) if (r1_hops > r2_hops) return -1; - rc = lnet_compare_peers(p1, p2); - if (rc) - return rc; - - if (r1->lr_seq - r2->lr_seq <= 0) - return 1; - - return -1; + return 0; } -static struct lnet_peer_ni * -lnet_find_route_locked(struct lnet_net *net, __u32 remote_net, - lnet_nid_t rtr_nid) +static struct lnet_route * +lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net, + struct lnet_route **prev_route, + struct lnet_peer_ni **gwni) { - struct lnet_remotenet *rnet; - struct lnet_route *route; - struct lnet_route *best_route; - struct lnet_route *last_route; - struct lnet_peer_ni *lpni_best; - struct lnet_peer_ni *lp; - int rc; - - /* If @rtr_nid is not LNET_NID_ANY, return the gateway with - * rtr_nid nid, otherwise find the best gateway I can use */ - - rnet = lnet_find_rnet_locked(remote_net); - if (rnet == NULL) - return NULL; + struct lnet_peer_ni *lpni, *best_gw_ni = NULL; + struct lnet_route *best_route; + struct lnet_route *last_route; + struct lnet_route *route; + int rc; + __u32 restrict_net; + __u32 any_net = LNET_NIDNET(LNET_NID_ANY); - lpni_best = NULL; best_route = last_route = NULL; list_for_each_entry(route, &rnet->lrn_routes, lr_list) { - lp = route->lr_gateway; - if (!lnet_is_route_alive(route)) continue; - if (net != NULL && lp->lpni_net != net) - continue; - - if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */ - return lp; + /* If the src_net is specified then we need to find an lpni + * on that network + */ + restrict_net = src_net == any_net ? route->lr_lnet : src_net; + if (!best_route) { + lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY, + route->lr_gateway, + restrict_net); + if (lpni) { + best_route = last_route = route; + best_gw_ni = lpni; + } else + CERROR("Gateway %s does not have a peer NI on net %s\n", + libcfs_nid2str(route->lr_gateway->lp_primary_nid), + libcfs_net2str(restrict_net)); - if (lpni_best == NULL) { - best_route = last_route = route; - lpni_best = lp; continue; } @@ -1442,19 +1517,40 @@ lnet_find_route_locked(struct lnet_net *net, __u32 remote_net, last_route = route; rc = lnet_compare_routes(route, best_route); - if (rc < 0) + if (rc == -1) + continue; + + lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY, + route->lr_gateway, + restrict_net); + if (!lpni) { + CERROR("Gateway %s does not have a peer NI on net %s\n", + libcfs_nid2str(route->lr_gateway->lp_primary_nid), + libcfs_net2str(restrict_net)); continue; + } - best_route = route; - lpni_best = lp; + if (rc == 1) { + best_route = route; + best_gw_ni = lpni; + continue; + } + + rc = lnet_compare_gw_lpnis(lpni, best_gw_ni); + if (rc == -1) + continue; + + if (rc == 1 || route->lr_seq <= best_route->lr_seq) { + best_route = route; + best_gw_ni = lpni; + continue; + } } - /* set sequence number on the best router to the latest sequence + 1 - * so we can round-robin all routers, it's race and inaccurate but - * harmless and functional */ - if (best_route != NULL) - best_route->lr_seq = last_route->lr_seq + 1; - return lpni_best; + *prev_route = last_route; + *gwni = best_gw_ni; + + return best_route; } static struct lnet_ni * @@ -1489,9 +1585,11 @@ lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni, unsigned int distance; int ni_credits; int ni_healthv; + int ni_fatal; ni_credits = atomic_read(&ni->ni_tx_credits); ni_healthv = atomic_read(&ni->ni_healthv); + ni_fatal = atomic_read(&ni->ni_fatal_error_on); /* * calculate the distance from the CPT on which @@ -1519,7 +1617,9 @@ lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni, * Select on health, shorter distance, available * credits, then round-robin. */ - if (ni_healthv < best_healthv) { + if (ni_fatal) { + continue; + } else if (ni_healthv < best_healthv) { continue; } else if (ni_healthv > best_healthv) { best_healthv = ni_healthv; @@ -1596,6 +1696,25 @@ lnet_msg_discovery(struct lnet_msg *msg) #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST) static int +lnet_handle_lo_send(struct lnet_send_data *sd) +{ + struct lnet_msg *msg = sd->sd_msg; + int cpt = sd->sd_cpt; + + /* No send credit hassles with LOLND */ + lnet_ni_addref_locked(the_lnet.ln_loni, cpt); + msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid); + if (!msg->msg_routing) + msg->msg_hdr.src_nid = + cpu_to_le64(the_lnet.ln_loni->ni_nid); + msg->msg_target.nid = the_lnet.ln_loni->ni_nid; + lnet_msg_commit(msg, cpt); + msg->msg_txni = the_lnet.ln_loni; + + return LNET_CREDIT_OK; +} + +static int lnet_handle_send(struct lnet_send_data *sd) { struct lnet_ni *best_ni = sd->sd_best_ni; @@ -1606,6 +1725,7 @@ lnet_handle_send(struct lnet_send_data *sd) __u32 send_case = sd->sd_send_case; int rc; __u32 routing = send_case & REMOTE_DST; + struct lnet_rsp_tracker *rspt; /* * Increment sequence number of the selected peer so that we @@ -1698,129 +1818,33 @@ lnet_handle_send(struct lnet_send_data *sd) msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid); } - rc = lnet_post_send_locked(msg, 0); - - if (!rc) - CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s\n", - libcfs_nid2str(msg->msg_hdr.src_nid), - libcfs_nid2str(msg->msg_txni->ni_nid), - libcfs_nid2str(sd->sd_src_nid), - libcfs_nid2str(msg->msg_hdr.dest_nid), - libcfs_nid2str(sd->sd_dst_nid), - libcfs_nid2str(msg->msg_txpeer->lpni_nid), - lnet_msgtyp2str(msg->msg_type)); - - return rc; -} - -static struct lnet_peer_ni * -lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer, - struct lnet_peer_net *peer_net) -{ /* - * Look at the peer NIs for the destination peer that connect - * to the chosen net. If a peer_ni is preferred when using the - * best_ni to communicate, we use that one. If there is no - * preferred peer_ni, or there are multiple preferred peer_ni, - * the available transmit credits are used. If the transmit - * credits are equal, we round-robin over the peer_ni. + * if we have response tracker block update it with the next hop + * nid */ - struct lnet_peer_ni *lpni = NULL; - struct lnet_peer_ni *best_lpni = NULL; - struct lnet_ni *best_ni = sd->sd_best_ni; - lnet_nid_t dst_nid = sd->sd_dst_nid; - int best_lpni_credits = INT_MIN; - bool preferred = false; - bool ni_is_pref; - - while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) { - /* - * if the best_ni we've chosen aleady has this lpni - * preferred, then let's use it - */ - ni_is_pref = lnet_peer_is_pref_nid_locked(lpni, - best_ni->ni_nid); - - CDEBUG(D_NET, "%s ni_is_pref = %d\n", - libcfs_nid2str(best_ni->ni_nid), ni_is_pref); - - if (best_lpni) - CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n", - libcfs_nid2str(lpni->lpni_nid), - lpni->lpni_txcredits, best_lpni_credits, - lpni->lpni_seq, best_lpni->lpni_seq); - - /* if this is a preferred peer use it */ - if (!preferred && ni_is_pref) { - preferred = true; - } else if (preferred && !ni_is_pref) { - /* - * this is not the preferred peer so let's ignore - * it. - */ - continue; - } else if (lpni->lpni_txcredits < best_lpni_credits) { - /* - * We already have a peer that has more credits - * available than this one. No need to consider - * this peer further. - */ - continue; - } else if (lpni->lpni_txcredits == best_lpni_credits) { - /* - * The best peer found so far and the current peer - * have the same number of available credits let's - * make sure to select between them using Round - * Robin - */ - if (best_lpni) { - if (best_lpni->lpni_seq <= lpni->lpni_seq) - continue; - } + if (msg->msg_md) { + rspt = msg->msg_md->md_rspt_ptr; + if (rspt) { + rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid; + CDEBUG(D_NET, "rspt_next_hop_nid = %s\n", + libcfs_nid2str(rspt->rspt_next_hop_nid)); } - - best_lpni = lpni; - best_lpni_credits = lpni->lpni_txcredits; } - /* if we still can't find a peer ni then we can't reach it */ - if (!best_lpni) { - __u32 net_id = (peer_net) ? peer_net->lpn_net_id : - LNET_NIDNET(dst_nid); - CDEBUG(D_NET, "no peer_ni found on peer net %s\n", - libcfs_net2str(net_id)); - return NULL; - } - - CDEBUG(D_NET, "sd_best_lpni = %s\n", - libcfs_nid2str(best_lpni->lpni_nid)); - - return best_lpni; -} - -/* - * Prerequisite: the best_ni should already be set in the sd - */ -static inline struct lnet_peer_ni * -lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer, - __u32 net_id) -{ - struct lnet_peer_net *peer_net; - - /* - * The gateway is Multi-Rail capable so now we must select the - * proper peer_ni - */ - peer_net = lnet_peer_get_net_locked(peer, net_id); + rc = lnet_post_send_locked(msg, 0); - if (!peer_net) { - CERROR("gateway peer %s has no NI on net %s\n", - libcfs_nid2str(peer->lp_primary_nid), - libcfs_net2str(net_id)); - return NULL; - } + if (!rc) + CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n", + libcfs_nid2str(msg->msg_hdr.src_nid), + libcfs_nid2str(msg->msg_txni->ni_nid), + libcfs_nid2str(sd->sd_src_nid), + libcfs_nid2str(msg->msg_hdr.dest_nid), + libcfs_nid2str(sd->sd_dst_nid), + libcfs_nid2str(msg->msg_txpeer->lpni_nid), + libcfs_nid2str(sd->sd_rtr_nid), + lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count); - return lnet_select_peer_ni(sd, peer, peer_net); + return rc; } static inline void @@ -1872,8 +1896,11 @@ lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd) * Local Destination * MR Peer * - * Run the selection algorithm on the peer NIs unless we're sending - * a response, in this case just send to the destination + * Don't run the selection algorithm on the peer NIs. By specifying the + * local NID, we're also saying that we should always use the destination NID + * provided. This handles the case where we should be using the same + * destination NID for the all the messages which belong to the same RPC + * request. */ static int lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd) @@ -1886,18 +1913,10 @@ lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd) return -EINVAL; } - /* - * only run the selection algorithm to pick the peer_ni if we're - * sending a GET or a PUT. Responses are sent to the same - * destination NID provided. - */ - if (!(sd->sd_send_case & SND_RESP)) { - sd->sd_best_lpni = - lnet_find_best_lpni_on_net(sd, sd->sd_peer, - sd->sd_best_ni->ni_net->net_id); - } - - if (sd->sd_best_lpni) + if (sd->sd_best_lpni && + sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) + return lnet_handle_lo_send(sd); + else if (sd->sd_best_lpni) return lnet_handle_send(sd); CERROR("can't send to %s. no NI on %s\n", @@ -1939,63 +1958,193 @@ lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni, } static int +lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, + struct lnet_msg *msg, lnet_nid_t rtr_nid, + int cpt) +{ + struct lnet_peer *peer; + lnet_nid_t primary_nid; + int rc; + + lnet_peer_ni_addref_locked(lpni); + + peer = lpni->lpni_peer_net->lpn_peer; + + if (lnet_peer_gw_discovery(peer)) { + lnet_peer_ni_decref_locked(lpni); + return 0; + } + + if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) { + lnet_peer_ni_decref_locked(lpni); + return 0; + } + + rc = lnet_discover_peer_locked(lpni, cpt, false); + if (rc) { + lnet_peer_ni_decref_locked(lpni); + return rc; + } + /* The peer may have changed. */ + peer = lpni->lpni_peer_net->lpn_peer; + spin_lock(&peer->lp_lock); + if (lnet_peer_is_uptodate_locked(peer)) { + spin_unlock(&peer->lp_lock); + lnet_peer_ni_decref_locked(lpni); + return 0; + } + /* queue message and return */ + msg->msg_rtr_nid_param = rtr_nid; + msg->msg_sending = 0; + msg->msg_txpeer = NULL; + list_add_tail(&msg->msg_list, &peer->lp_dc_pendq); + primary_nid = peer->lp_primary_nid; + spin_unlock(&peer->lp_lock); + + lnet_peer_ni_decref_locked(lpni); + + CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n", + msg, libcfs_nid2str(primary_nid)); + + return LNET_DC_WAIT; +} + +static int lnet_handle_find_routed_path(struct lnet_send_data *sd, lnet_nid_t dst_nid, struct lnet_peer_ni **gw_lpni, struct lnet_peer **gw_peer) { - struct lnet_peer_ni *gw; + int rc; + __u32 local_lnet; + struct lnet_peer *gw; + struct lnet_peer *lp; + struct lnet_peer_net *lpn; + struct lnet_peer_net *best_lpn = NULL; + struct lnet_remotenet *rnet, *best_rnet = NULL; + struct lnet_route *best_route = NULL; + struct lnet_route *last_route = NULL; + struct lnet_peer_ni *lpni = NULL; + struct lnet_peer_ni *gwni = NULL; lnet_nid_t src_nid = sd->sd_src_nid; - gw = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid), - sd->sd_rtr_nid); - if (!gw) { - CERROR("no route to %s from %s\n", - libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid)); - return -EHOSTUNREACH; - } + /* If a router nid was specified then we are replying to a GET or + * sending an ACK. In this case we use the gateway associated with the + * specified router nid. + */ + if (sd->sd_rtr_nid != LNET_NID_ANY) { + gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid); + if (!gwni) { + CERROR("No peer NI for gateway %s\n", + libcfs_nid2str(sd->sd_rtr_nid)); + return -EHOSTUNREACH; + } + gw = gwni->lpni_peer_net->lpn_peer; + lnet_peer_ni_decref_locked(gwni); + local_lnet = LNET_NIDNET(sd->sd_rtr_nid); + } else { + /* we've already looked up the initial lpni using dst_nid */ + lpni = sd->sd_best_lpni; + /* the peer tree must be in existence */ + LASSERT(lpni && lpni->lpni_peer_net && + lpni->lpni_peer_net->lpn_peer); + lp = lpni->lpni_peer_net->lpn_peer; + + list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) { + /* is this remote network reachable? */ + rnet = lnet_find_rnet_locked(lpn->lpn_net_id); + if (!rnet) + continue; + + if (!best_lpn) { + best_lpn = lpn; + best_rnet = rnet; + } + + if (best_lpn->lpn_seq <= lpn->lpn_seq) + continue; + + best_lpn = lpn; + best_rnet = rnet; + } + + if (!best_lpn) { + CERROR("peer %s has no available nets\n", + libcfs_nid2str(sd->sd_dst_nid)); + return -EHOSTUNREACH; + } + + sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni, + sd->sd_dst_nid, + lp, + best_lpn->lpn_net_id); + if (!sd->sd_best_lpni) { + CERROR("peer %s down\n", + libcfs_nid2str(sd->sd_dst_nid)); + return -EHOSTUNREACH; + } - /* get the peer of the gw_ni */ - LASSERT(gw->lpni_peer_net); - LASSERT(gw->lpni_peer_net->lpn_peer); + best_route = lnet_find_route_locked(best_rnet, + LNET_NIDNET(src_nid), + &last_route, &gwni); + if (!best_route) { + CERROR("no route to %s from %s\n", + libcfs_nid2str(dst_nid), + libcfs_nid2str(src_nid)); + return -EHOSTUNREACH; + } + + if (!gwni) { + CERROR("Internal Error. Route expected to %s from %s\n", + libcfs_nid2str(dst_nid), + libcfs_nid2str(src_nid)); + return -EFAULT; + } + + gw = best_route->lr_gateway; + LASSERT(gw == gwni->lpni_peer_net->lpn_peer); + local_lnet = best_route->lr_lnet; + + } - *gw_peer = gw->lpni_peer_net->lpn_peer; + /* + * Discover this gateway if it hasn't already been discovered. + * This means we might delay the message until discovery has + * completed + */ + sd->sd_msg->msg_src_nid_param = sd->sd_src_nid; + rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid, + sd->sd_cpt); + if (rc) + return rc; if (!sd->sd_best_ni) - sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, *gw_peer, - gw->lpni_peer_net, + sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, + lnet_peer_get_net_locked(gw, + local_lnet), sd->sd_md_cpt, true); if (!sd->sd_best_ni) { - CERROR("Internal Error. Expected local ni on %s " - "but non found :%s\n", - libcfs_net2str(gw->lpni_peer_net->lpn_net_id), + CERROR("Internal Error. Expected local ni on %s but non found :%s\n", + libcfs_net2str(local_lnet), libcfs_nid2str(sd->sd_src_nid)); return -EFAULT; } + *gw_lpni = gwni; + *gw_peer = gw; + /* - * if gw is MR let's find its best peer_ni + * increment the sequence numbers since now we're sure we're + * going to use this path */ - if (lnet_peer_is_multi_rail(*gw_peer)) { - gw = lnet_find_best_lpni_on_net(sd, *gw_peer, - sd->sd_best_ni->ni_net->net_id); - /* - * We've already verified that the gw has an NI on that - * desired net, but we're not finding it. Something is - * wrong. - */ - if (!gw) { - CERROR("Internal Error. Route expected to %s from %s\n", - libcfs_nid2str(dst_nid), - libcfs_nid2str(src_nid)); - return -EFAULT; - } + if (sd->sd_rtr_nid == LNET_NID_ANY) { + LASSERT(best_route && last_route); + best_route->lr_seq = last_route->lr_seq + 1; + best_lpn->lpn_seq++; } - *gw_lpni = gw; - return 0; } @@ -2033,7 +2182,7 @@ lnet_handle_spec_router_dst(struct lnet_send_data *sd) rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni, &gw_peer); - if (rc < 0) + if (rc) return rc; if (sd->sd_send_case & NMR_DST) @@ -2054,7 +2203,8 @@ lnet_handle_spec_router_dst(struct lnet_send_data *sd) } struct lnet_ni * -lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt) +lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt, + bool discovery) { struct lnet_peer_net *peer_net = NULL; struct lnet_ni *best_ni = NULL; @@ -2076,6 +2226,14 @@ lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt) continue; best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer, peer_net, md_cpt, false); + + /* + * if this is a discovery message and lp_disc_net_id is + * specified then use that net to send the discovery on. + */ + if (peer->lp_disc_net_id == peer_net->lpn_net_id && + discovery) + break; } if (best_ni) @@ -2245,10 +2403,12 @@ lnet_handle_any_mr_dsta(struct lnet_send_data *sd) * networks. */ sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer, - sd->sd_md_cpt); + sd->sd_md_cpt, + lnet_msg_discovery(sd->sd_msg)); if (sd->sd_best_ni) { sd->sd_best_lpni = - lnet_find_best_lpni_on_net(sd, sd->sd_peer, + lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid, + sd->sd_peer, sd->sd_best_ni->ni_net->net_id); /* @@ -2257,7 +2417,16 @@ lnet_handle_any_mr_dsta(struct lnet_send_data *sd) * try and see if we can reach it over another routed * network */ - if (sd->sd_best_lpni) { + if (sd->sd_best_lpni && + sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) { + /* + * in case we initially started with a routed + * destination, let's reset to local + */ + sd->sd_send_case &= ~REMOTE_DST; + sd->sd_send_case |= LOCAL_DST; + return lnet_handle_lo_send(sd); + } else if (sd->sd_best_lpni) { /* * in case we initially started with a routed * destination, let's reset to local @@ -2323,6 +2492,8 @@ lnet_handle_any_mr_dst(struct lnet_send_data *sd) "No route available\n", libcfs_nid2str(sd->sd_dst_nid)); return -EHOSTUNREACH; + } else if (rc > 0) { + return rc; } sd->sd_best_lpni = gw; @@ -2344,15 +2515,15 @@ lnet_handle_any_mr_dst(struct lnet_send_data *sd) return rc; /* - * TODO; One possible enhancement is to run the selection - * algorithm on the peer. However for remote peers the credits are - * not decremented, so we'll be basically going over the peer NIs - * in round robin. An MR router will run the selection algorithm - * on the next-hop interfaces. + * Now that we must route to the destination, we must consider the + * MR case, where the destination has multiple interfaces, some of + * which we can route to and others we do not. For this reason we + * need to select the destination which we can route to and if + * there are multiple, we need to round robin. */ rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni, &gw_peer); - if (rc < 0) + if (rc) return rc; sd->sd_send_case &= ~LOCAL_DST; @@ -2391,7 +2562,7 @@ lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd) */ rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni, &gw_peer); - if (rc < 0) + if (rc) return rc; /* @@ -2485,19 +2656,12 @@ again: * is no need to go through any selection. We can just shortcut * the entire process and send over lolnd */ + send_data.sd_msg = msg; + send_data.sd_cpt = cpt; if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) { - /* No send credit hassles with LOLND */ - lnet_ni_addref_locked(the_lnet.ln_loni, cpt); - msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid); - if (!msg->msg_routing) - msg->msg_hdr.src_nid = - cpu_to_le64(the_lnet.ln_loni->ni_nid); - msg->msg_target.nid = the_lnet.ln_loni->ni_nid; - lnet_msg_commit(msg, cpt); - msg->msg_txni = the_lnet.ln_loni; + rc = lnet_handle_lo_send(&send_data); lnet_net_unlock(cpt); - - return LNET_CREDIT_OK; + return rc; } /* @@ -2521,41 +2685,19 @@ again: msg->msg_src_nid_param = src_nid; /* - * Now that we have a peer_ni, check if we want to discover - * the peer. Traffic to the LNET_RESERVED_PORTAL should not - * trigger discovery. + * If necessary, perform discovery on the peer that owns this peer_ni. + * Note, this can result in the ownership of this peer_ni changing + * to another peer object. */ - peer = lpni->lpni_peer_net->lpn_peer; - if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) { - lnet_nid_t primary_nid; - rc = lnet_discover_peer_locked(lpni, cpt, false); - if (rc) { - lnet_peer_ni_decref_locked(lpni); - lnet_net_unlock(cpt); - return rc; - } - /* The peer may have changed. */ - peer = lpni->lpni_peer_net->lpn_peer; - /* queue message and return */ - msg->msg_rtr_nid_param = rtr_nid; - msg->msg_sending = 0; - list_add_tail(&msg->msg_list, &peer->lp_dc_pendq); + rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt); + if (rc) { lnet_peer_ni_decref_locked(lpni); - primary_nid = peer->lp_primary_nid; lnet_net_unlock(cpt); - - CDEBUG(D_NET, "%s pending discovery\n", - libcfs_nid2str(primary_nid)); - - return LNET_DC_WAIT; + return rc; } lnet_peer_ni_decref_locked(lpni); - /* If peer is not healthy then can not send anything to it */ - if (!lnet_is_peer_healthy_locked(peer)) { - lnet_net_unlock(cpt); - return -EHOSTUNREACH; - } + peer = lpni->lpni_peer_net->lpn_peer; /* * Identify the different send cases @@ -2585,7 +2727,6 @@ again: send_case |= SND_RESP; /* assign parameters to the send_data */ - send_data.sd_msg = msg; send_data.sd_rtr_nid = rtr_nid; send_data.sd_src_nid = src_nid; send_data.sd_dst_nid = dst_nid; @@ -2597,15 +2738,20 @@ again: send_data.sd_final_dst_lpni = lpni; send_data.sd_peer = peer; send_data.sd_md_cpt = md_cpt; - send_data.sd_cpt = cpt; send_data.sd_send_case = send_case; rc = lnet_handle_send_case_locked(&send_data); + /* + * Update the local cpt since send_data.sd_cpt might've been + * updated as a result of calling lnet_handle_send_case_locked(). + */ + cpt = send_data.sd_cpt; + if (rc == REPEAT_SEND) goto again; - lnet_net_unlock(send_data.sd_cpt); + lnet_net_unlock(cpt); return rc; } @@ -2633,8 +2779,13 @@ lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid) LASSERT(!msg->msg_tx_committed); rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid); - if (rc < 0) + if (rc < 0) { + if (rc == -EHOSTUNREACH) + msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR; + else + msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR; return rc; + } if (rc == LNET_CREDIT_OK) lnet_ni_send(msg->msg_txni, msg); @@ -2643,6 +2794,209 @@ lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid) return 0; } +enum lnet_mt_event_type { + MT_TYPE_LOCAL_NI = 0, + MT_TYPE_PEER_NI +}; + +struct lnet_mt_event_info { + enum lnet_mt_event_type mt_type; + lnet_nid_t mt_nid; +}; + +/* called with res_lock held */ +void +lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt) +{ + struct lnet_rsp_tracker *rspt; + + /* + * msg has a refcount on the MD so the MD is not going away. + * The rspt queue for the cpt is protected by + * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie. + */ + if (!md->md_rspt_ptr) + return; + + rspt = md->md_rspt_ptr; + + /* debug code */ + LASSERT(rspt->rspt_cpt == cpt); + + md->md_rspt_ptr = NULL; + + if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) { + /* + * The monitor thread has invalidated this handle because the + * response timed out, but it failed to lookup the MD. That + * means this response tracker is on the zombie list. We can + * safely remove it under the resource lock (held by caller) and + * free the response tracker block. + */ + list_del(&rspt->rspt_on_list); + lnet_rspt_free(rspt, cpt); + } else { + /* + * invalidate the handle to indicate that a response has been + * received, which will then lead the monitor thread to clean up + * the rspt block. + */ + LNetInvalidateMDHandle(&rspt->rspt_mdh); + } +} + +void +lnet_clean_zombie_rstqs(void) +{ + struct lnet_rsp_tracker *rspt, *tmp; + int i; + + cfs_cpt_for_each(i, lnet_cpt_table()) { + list_for_each_entry_safe(rspt, tmp, + the_lnet.ln_mt_zombie_rstqs[i], + rspt_on_list) { + list_del(&rspt->rspt_on_list); + lnet_rspt_free(rspt, i); + } + } + + cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs); +} + +static void +lnet_finalize_expired_responses(void) +{ + struct lnet_libmd *md; + struct list_head local_queue; + struct lnet_rsp_tracker *rspt, *tmp; + ktime_t now; + int i; + + if (the_lnet.ln_mt_rstq == NULL) + return; + + cfs_cpt_for_each(i, lnet_cpt_table()) { + INIT_LIST_HEAD(&local_queue); + + lnet_net_lock(i); + if (!the_lnet.ln_mt_rstq[i]) { + lnet_net_unlock(i); + continue; + } + list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue); + lnet_net_unlock(i); + + now = ktime_get(); + + list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) { + /* + * The rspt mdh will be invalidated when a response + * is received or whenever we want to discard the + * block the monitor thread will walk the queue + * and clean up any rsts with an invalid mdh. + * The monitor thread will walk the queue until + * the first unexpired rspt block. This means that + * some rspt blocks which received their + * corresponding responses will linger in the + * queue until they are cleaned up eventually. + */ + lnet_res_lock(i); + if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) { + lnet_res_unlock(i); + list_del(&rspt->rspt_on_list); + lnet_rspt_free(rspt, i); + continue; + } + + if (ktime_compare(now, rspt->rspt_deadline) >= 0 || + the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) { + struct lnet_peer_ni *lpni; + lnet_nid_t nid; + + md = lnet_handle2md(&rspt->rspt_mdh); + if (!md) { + /* MD has been queued for unlink, but + * rspt hasn't been detached (Note we've + * checked above that the rspt_mdh is + * valid). Since we cannot lookup the MD + * we're unable to detach the rspt + * ourselves. Thus, move the rspt to the + * zombie list where we'll wait for + * either: + * 1. The remaining operations on the + * MD to complete. In this case the + * final operation will result in + * lnet_msg_detach_md()-> + * lnet_detach_rsp_tracker() where + * we will clean up this response + * tracker. + * 2. LNet to shutdown. In this case + * we'll wait until after all LND Nets + * have shutdown and then we can + * safely free any remaining response + * tracker blocks on the zombie list. + * Note: We need to hold the resource + * lock when adding to the zombie list + * because we may have concurrent access + * with lnet_detach_rsp_tracker(). + */ + LNetInvalidateMDHandle(&rspt->rspt_mdh); + list_move(&rspt->rspt_on_list, + the_lnet.ln_mt_zombie_rstqs[i]); + lnet_res_unlock(i); + continue; + } + LASSERT(md->md_rspt_ptr == rspt); + md->md_rspt_ptr = NULL; + lnet_res_unlock(i); + + LNetMDUnlink(rspt->rspt_mdh); + + nid = rspt->rspt_next_hop_nid; + + list_del(&rspt->rspt_on_list); + lnet_rspt_free(rspt, i); + + /* If we're shutting down we just want to clean + * up the rspt blocks + */ + if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) + continue; + + lnet_net_lock(i); + the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++; + lnet_net_unlock(i); + + CDEBUG(D_NET, + "Response timeout: md = %p: nid = %s\n", + md, libcfs_nid2str(nid)); + + /* + * If there is a timeout on the response + * from the next hop decrement its health + * value so that we don't use it + */ + lnet_net_lock(0); + lpni = lnet_find_peer_ni_locked(nid); + if (lpni) { + lnet_handle_remote_failure_locked(lpni); + lnet_peer_ni_decref_locked(lpni); + } + lnet_net_unlock(0); + } else { + lnet_res_unlock(i); + break; + } + } + + if (!list_empty(&local_queue)) { + lnet_net_lock(i); + list_splice(&local_queue, the_lnet.ln_mt_rstq[i]); + lnet_net_unlock(i); + } + } +} + static void lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt) { @@ -2692,6 +3046,12 @@ lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt) lnet_peer_ni_decref_locked(lpni); lnet_net_unlock(cpt); + CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n", + libcfs_nid2str(src_nid), + libcfs_id2str(msg->msg_target), + lnet_msgtyp2str(msg->msg_type), + msg->msg_recovery, + msg->msg_retry_count); rc = lnet_send(src_nid, msg, LNET_NID_ANY); if (rc) { CERROR("Error sending %s to %s: %d\n", @@ -2701,6 +3061,8 @@ lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt) lnet_finalize(msg, rc); } lnet_net_lock(cpt); + if (!rc) + the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++; } } } @@ -2719,13 +3081,14 @@ lnet_resend_pending_msgs(void) /* called with cpt and ni_lock held */ static void -lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt) +lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force) { struct lnet_handle_md recovery_mdh; LNetInvalidateMDHandle(&recovery_mdh); - if (ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING) { + if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING || + force) { recovery_mdh = ni->ni_ping_mdh; LNetInvalidateMDHandle(&ni->ni_ping_mdh); } @@ -2740,6 +3103,7 @@ lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt) static void lnet_recover_local_nis(void) { + struct lnet_mt_event_info *ev_info; struct list_head processed_list; struct list_head local_queue; struct lnet_handle_md mdh; @@ -2777,28 +3141,49 @@ lnet_recover_local_nis(void) lnet_net_lock(0); lnet_ni_lock(ni); - if (!(ni->ni_state & LNET_NI_STATE_ACTIVE) || + if (ni->ni_state != LNET_NI_STATE_ACTIVE || healthv == LNET_MAX_HEALTH_VALUE) { list_del_init(&ni->ni_recovery); - lnet_unlink_ni_recovery_mdh_locked(ni, 0); + lnet_unlink_ni_recovery_mdh_locked(ni, 0, false); lnet_ni_unlock(ni); lnet_ni_decref_locked(ni, 0); lnet_net_unlock(0); continue; } - lnet_ni_unlock(ni); - lnet_net_unlock(0); /* - * protect the ni->ni_state field. Once we call the - * lnet_send_ping function it's possible we receive - * a response before we check the rc. The lock ensures - * a stable value for the ni_state RECOVERY_PENDING bit + * if the local NI failed recovery we must unlink the md. + * But we want to keep the local_ni on the recovery queue + * so we can continue the attempts to recover it. */ + if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) { + lnet_unlink_ni_recovery_mdh_locked(ni, 0, true); + ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED; + } + + lnet_ni_unlock(ni); + lnet_net_unlock(0); + + + CDEBUG(D_NET, "attempting to recover local ni: %s\n", + libcfs_nid2str(ni->ni_nid)); + lnet_ni_lock(ni); - if (!(ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING)) { - ni->ni_state |= LNET_NI_STATE_RECOVERY_PENDING; + if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) { + ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING; lnet_ni_unlock(ni); + + LIBCFS_ALLOC(ev_info, sizeof(*ev_info)); + if (!ev_info) { + CERROR("out of memory. Can't recover %s\n", + libcfs_nid2str(ni->ni_nid)); + lnet_ni_lock(ni); + ni->ni_recovery_state &= + ~LNET_NI_RECOVERY_PENDING; + lnet_ni_unlock(ni); + continue; + } + mdh = ni->ni_ping_mdh; /* * Invalidate the ni mdh in case it's deleted. @@ -2829,9 +3214,10 @@ lnet_recover_local_nis(void) lnet_ni_decref_locked(ni, 0); lnet_net_unlock(0); - rc = lnet_send_ping(nid, &mdh, - LNET_INTERFACES_MIN, (void *)nid, - the_lnet.ln_mt_eqh, true); + ev_info->mt_type = MT_TYPE_LOCAL_NI; + ev_info->mt_nid = nid; + rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN, + ev_info, the_lnet.ln_mt_eqh, true); /* lookup the nid again */ lnet_net_lock(0); ni = lnet_nid2ni_locked(nid, 0); @@ -2865,7 +3251,7 @@ lnet_recover_local_nis(void) lnet_ni_lock(ni); if (rc) - ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING; + ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING; } lnet_ni_unlock(ni); } @@ -2880,26 +3266,6 @@ lnet_recover_local_nis(void) lnet_net_unlock(0); } -static struct list_head ** -lnet_create_array_of_queues(void) -{ - struct list_head **qs; - struct list_head *q; - int i; - - qs = cfs_percpt_alloc(lnet_cpt_table(), - sizeof(struct list_head)); - if (!qs) { - CERROR("Failed to allocate queues\n"); - return NULL; - } - - cfs_percpt_for_each(q, i, qs) - INIT_LIST_HEAD(q); - - return qs; -} - static int lnet_resendqs_create(void) { @@ -2924,17 +3290,56 @@ lnet_clean_local_ni_recoveryq(void) /* This is only called when the monitor thread has stopped */ lnet_net_lock(0); - while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) { - ni = list_entry(the_lnet.ln_mt_localNIRecovq.next, - struct lnet_ni, ni_recovery); - list_del_init(&ni->ni_recovery); - lnet_ni_lock(ni); - lnet_unlink_ni_recovery_mdh_locked(ni, 0); - lnet_ni_unlock(ni); - lnet_ni_decref_locked(ni, 0); + while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) { + ni = list_entry(the_lnet.ln_mt_localNIRecovq.next, + struct lnet_ni, ni_recovery); + list_del_init(&ni->ni_recovery); + lnet_ni_lock(ni); + lnet_unlink_ni_recovery_mdh_locked(ni, 0, true); + lnet_ni_unlock(ni); + lnet_ni_decref_locked(ni, 0); + } + + lnet_net_unlock(0); +} + +static void +lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt, + bool force) +{ + struct lnet_handle_md recovery_mdh; + + LNetInvalidateMDHandle(&recovery_mdh); + + if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) { + recovery_mdh = lpni->lpni_recovery_ping_mdh; + LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh); + } + spin_unlock(&lpni->lpni_lock); + lnet_net_unlock(cpt); + if (!LNetMDHandleIsInvalid(recovery_mdh)) + LNetMDUnlink(recovery_mdh); + lnet_net_lock(cpt); + spin_lock(&lpni->lpni_lock); +} + +static void +lnet_clean_peer_ni_recoveryq(void) +{ + struct lnet_peer_ni *lpni, *tmp; + + lnet_net_lock(LNET_LOCK_EX); + + list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq, + lpni_recovery) { + list_del_init(&lpni->lpni_recovery); + spin_lock(&lpni->lpni_lock); + lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true); + spin_unlock(&lpni->lpni_lock); + lnet_peer_ni_decref_locked(lpni); } - lnet_net_unlock(0); + lnet_net_unlock(LNET_LOCK_EX); } static void @@ -2960,9 +3365,150 @@ lnet_clean_resendqs(void) cfs_percpt_free(the_lnet.ln_mt_resendqs); } +static void +lnet_recover_peer_nis(void) +{ + struct lnet_mt_event_info *ev_info; + struct list_head processed_list; + struct list_head local_queue; + struct lnet_handle_md mdh; + struct lnet_peer_ni *lpni; + struct lnet_peer_ni *tmp; + lnet_nid_t nid; + int healthv; + int rc; + + INIT_LIST_HEAD(&local_queue); + INIT_LIST_HEAD(&processed_list); + + /* + * Always use cpt 0 for locking across all interactions with + * ln_mt_peerNIRecovq + */ + lnet_net_lock(0); + list_splice_init(&the_lnet.ln_mt_peerNIRecovq, + &local_queue); + lnet_net_unlock(0); + + list_for_each_entry_safe(lpni, tmp, &local_queue, + lpni_recovery) { + /* + * The same protection strategy is used here as is in the + * local recovery case. + */ + lnet_net_lock(0); + healthv = atomic_read(&lpni->lpni_healthv); + spin_lock(&lpni->lpni_lock); + if (lpni->lpni_state & LNET_PEER_NI_DELETING || + healthv == LNET_MAX_HEALTH_VALUE) { + list_del_init(&lpni->lpni_recovery); + lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false); + spin_unlock(&lpni->lpni_lock); + lnet_peer_ni_decref_locked(lpni); + lnet_net_unlock(0); + continue; + } + + /* + * If the peer NI has failed recovery we must unlink the + * md. But we want to keep the peer ni on the recovery + * queue so we can try to continue recovering it + */ + if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) { + lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true); + lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED; + } + + spin_unlock(&lpni->lpni_lock); + lnet_net_unlock(0); + + /* + * NOTE: we're racing with peer deletion from user space. + * It's possible that a peer is deleted after we check its + * state. In this case the recovery can create a new peer + */ + spin_lock(&lpni->lpni_lock); + if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) && + !(lpni->lpni_state & LNET_PEER_NI_DELETING)) { + lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING; + spin_unlock(&lpni->lpni_lock); + + LIBCFS_ALLOC(ev_info, sizeof(*ev_info)); + if (!ev_info) { + CERROR("out of memory. Can't recover %s\n", + libcfs_nid2str(lpni->lpni_nid)); + spin_lock(&lpni->lpni_lock); + lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING; + spin_unlock(&lpni->lpni_lock); + continue; + } + + /* look at the comments in lnet_recover_local_nis() */ + mdh = lpni->lpni_recovery_ping_mdh; + LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh); + nid = lpni->lpni_nid; + lnet_net_lock(0); + list_del_init(&lpni->lpni_recovery); + lnet_peer_ni_decref_locked(lpni); + lnet_net_unlock(0); + + ev_info->mt_type = MT_TYPE_PEER_NI; + ev_info->mt_nid = nid; + rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN, + ev_info, the_lnet.ln_mt_eqh, true); + lnet_net_lock(0); + /* + * lnet_find_peer_ni_locked() grabs a refcount for + * us. No need to take it explicitly. + */ + lpni = lnet_find_peer_ni_locked(nid); + if (!lpni) { + lnet_net_unlock(0); + LNetMDUnlink(mdh); + continue; + } + + lpni->lpni_recovery_ping_mdh = mdh; + /* + * While we're unlocked the lpni could've been + * readded on the recovery queue. In this case we + * don't need to add it to the local queue, since + * it's already on there and the thread that added + * it would've incremented the refcount on the + * peer, which means we need to decref the refcount + * that was implicitly grabbed by find_peer_ni_locked. + * Otherwise, if the lpni is still not on + * the recovery queue, then we'll add it to the + * processed list. + */ + if (list_empty(&lpni->lpni_recovery)) + list_add_tail(&lpni->lpni_recovery, &processed_list); + else + lnet_peer_ni_decref_locked(lpni); + lnet_net_unlock(0); + + spin_lock(&lpni->lpni_lock); + if (rc) + lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING; + } + spin_unlock(&lpni->lpni_lock); + } + + list_splice_init(&processed_list, &local_queue); + lnet_net_lock(0); + list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq); + lnet_net_unlock(0); +} + static int lnet_monitor_thread(void *arg) { + time64_t recovery_timeout = 0; + time64_t rsp_timeout = 0; + int interval; + time64_t now; + + wait_for_completion(&the_lnet.ln_started); /* * The monitor thread takes care of the following: * 1. Checks the aliveness of routers @@ -2976,12 +3522,23 @@ lnet_monitor_thread(void *arg) cfs_block_allsigs(); while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) { + now = ktime_get_real_seconds(); + if (lnet_router_checker_active()) lnet_check_routers(); lnet_resend_pending_msgs(); - lnet_recover_local_nis(); + if (now >= rsp_timeout) { + lnet_finalize_expired_responses(); + rsp_timeout = now + (lnet_transaction_timeout / 2); + } + + if (now >= recovery_timeout) { + lnet_recover_local_nis(); + lnet_recover_peer_nis(); + recovery_timeout = now + lnet_recovery_interval; + } /* * TODO do we need to check if we should sleep without @@ -2991,17 +3548,28 @@ lnet_monitor_thread(void *arg) * if we wake up every 1 second? Although, we've seen * cases where we get a complaint that an idle thread * is waking up unnecessarily. + * + * Take into account the current net_count when you wake + * up for alive router checking, since we need to check + * possibly as many networks as we have configured. */ - wait_event_interruptible_timeout(the_lnet.ln_mt_waitq, - false, - cfs_time_seconds(1)); + interval = min(lnet_recovery_interval, + min((unsigned int) alive_router_check_interval / + lnet_current_net_count, + lnet_transaction_timeout / 2)); + wait_for_completion_interruptible_timeout( + &the_lnet.ln_mt_wait_complete, + cfs_time_seconds(interval)); + /* Must re-init the completion before testing anything, + * including ln_mt_state. + */ + reinit_completion(&the_lnet.ln_mt_wait_complete); } - /* clean up the router checker */ - lnet_prune_rc_data(1); - /* Shutting down */ + lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN; + lnet_net_unlock(LNET_LOCK_EX); /* signal that the monitor thread is exiting */ up(&the_lnet.ln_mt_signal); @@ -3073,10 +3641,69 @@ fail_error: } static void +lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info, + int status, bool unlink_event) +{ + lnet_nid_t nid = ev_info->mt_nid; + + if (ev_info->mt_type == MT_TYPE_LOCAL_NI) { + struct lnet_ni *ni; + + lnet_net_lock(0); + ni = lnet_nid2ni_locked(nid, 0); + if (!ni) { + lnet_net_unlock(0); + return; + } + lnet_ni_lock(ni); + ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING; + if (status) + ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED; + lnet_ni_unlock(ni); + lnet_net_unlock(0); + + if (status != 0) { + CERROR("local NI (%s) recovery failed with %d\n", + libcfs_nid2str(nid), status); + return; + } + /* + * need to increment healthv for the ni here, because in + * the lnet_finalize() path we don't have access to this + * NI. And in order to get access to it, we'll need to + * carry forward too much information. + * In the peer case, it'll naturally be incremented + */ + if (!unlink_event) + lnet_inc_healthv(&ni->ni_healthv); + } else { + struct lnet_peer_ni *lpni; + int cpt; + + cpt = lnet_net_lock_current(); + lpni = lnet_find_peer_ni_locked(nid); + if (!lpni) { + lnet_net_unlock(cpt); + return; + } + spin_lock(&lpni->lpni_lock); + lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING; + if (status) + lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED; + spin_unlock(&lpni->lpni_lock); + lnet_peer_ni_decref_locked(lpni); + lnet_net_unlock(cpt); + + if (status != 0) + CERROR("peer NI (%s) recovery failed with %d\n", + libcfs_nid2str(nid), status); + } +} + +void lnet_mt_event_handler(struct lnet_event *event) { - lnet_nid_t nid = (lnet_nid_t) event->md.user_ptr; - struct lnet_ni *ni; + struct lnet_mt_event_info *ev_info = event->md.user_ptr; struct lnet_ping_buffer *pbuf; /* TODO: remove assert */ @@ -3088,43 +3715,54 @@ lnet_mt_event_handler(struct lnet_event *event) event->status); switch (event->type) { + case LNET_EVENT_UNLINK: + CDEBUG(D_NET, "%s recovery ping unlinked\n", + libcfs_nid2str(ev_info->mt_nid)); + /* fallthrough */ case LNET_EVENT_REPLY: - /* - * If the NI has been restored completely then remove from - * the recovery queue - */ - lnet_net_lock(0); - ni = lnet_nid2ni_locked(nid, 0); - if (!ni) { - lnet_net_unlock(0); - break; - } - lnet_ni_lock(ni); - ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING; - lnet_ni_unlock(ni); - lnet_net_unlock(0); + lnet_handle_recovery_reply(ev_info, event->status, + event->type == LNET_EVENT_UNLINK); break; case LNET_EVENT_SEND: CDEBUG(D_NET, "%s recovery message sent %s:%d\n", - libcfs_nid2str(nid), + libcfs_nid2str(ev_info->mt_nid), (event->status) ? "unsuccessfully" : "successfully", event->status); break; - case LNET_EVENT_UNLINK: - /* nothing to do */ - CDEBUG(D_NET, "%s recovery ping unlinked\n", - libcfs_nid2str(nid)); - break; default: CERROR("Unexpected event: %d\n", event->type); - return; + break; } if (event->unlinked) { + LIBCFS_FREE(ev_info, sizeof(*ev_info)); pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start); lnet_ping_buffer_decref(pbuf); } } +static int +lnet_rsp_tracker_create(void) +{ + struct list_head **rstqs; + rstqs = lnet_create_array_of_queues(); + + if (!rstqs) + return -ENOMEM; + + the_lnet.ln_mt_rstq = rstqs; + + return 0; +} + +static void +lnet_rsp_tracker_clean(void) +{ + lnet_finalize_expired_responses(); + + cfs_percpt_free(the_lnet.ln_mt_rstq); + the_lnet.ln_mt_rstq = NULL; +} + int lnet_monitor_thr_start(void) { int rc = 0; @@ -3137,20 +3775,15 @@ int lnet_monitor_thr_start(void) if (rc) return rc; - rc = LNetEQAlloc(0, lnet_mt_event_handler, &the_lnet.ln_mt_eqh); - if (rc != 0) { - CERROR("Can't allocate monitor thread EQ: %d\n", rc); - goto clean_queues; - } - - /* Pre monitor thread start processing */ - rc = lnet_router_pre_mt_start(); + rc = lnet_rsp_tracker_create(); if (rc) - goto free_mem; + goto clean_queues; sema_init(&the_lnet.ln_mt_signal, 0); + lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING; + lnet_net_unlock(LNET_LOCK_EX); task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread"); if (IS_ERR(task)) { rc = PTR_ERR(task); @@ -3158,54 +3791,54 @@ int lnet_monitor_thr_start(void) goto clean_thread; } - /* post monitor thread start processing */ - lnet_router_post_mt_start(); - return 0; clean_thread: + lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING; + lnet_net_unlock(LNET_LOCK_EX); /* block until event callback signals exit */ down(&the_lnet.ln_mt_signal); /* clean up */ - lnet_router_cleanup(); -free_mem: + lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN; - lnet_clean_resendqs(); + lnet_net_unlock(LNET_LOCK_EX); + lnet_rsp_tracker_clean(); lnet_clean_local_ni_recoveryq(); - LNetEQFree(the_lnet.ln_mt_eqh); + lnet_clean_peer_ni_recoveryq(); + lnet_clean_resendqs(); LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh); return rc; clean_queues: - lnet_clean_resendqs(); + lnet_rsp_tracker_clean(); lnet_clean_local_ni_recoveryq(); + lnet_clean_peer_ni_recoveryq(); + lnet_clean_resendqs(); return rc; } void lnet_monitor_thr_stop(void) { - int rc; - if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) return; LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING); + lnet_net_lock(LNET_LOCK_EX); the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING; + lnet_net_unlock(LNET_LOCK_EX); /* tell the monitor thread that we're shutting down */ - wake_up(&the_lnet.ln_mt_waitq); + complete(&the_lnet.ln_mt_wait_complete); /* block until monitor thread signals that it's done */ down(&the_lnet.ln_mt_signal); LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN); /* perform cleanup tasks */ - lnet_router_cleanup(); - lnet_clean_resendqs(); + lnet_rsp_tracker_clean(); lnet_clean_local_ni_recoveryq(); - rc = LNetEQFree(the_lnet.ln_mt_eqh); - LASSERT(rc == 0); - return; + lnet_clean_peer_ni_recoveryq(); + lnet_clean_resendqs(); } void @@ -3214,8 +3847,8 @@ lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob, { lnet_net_lock(cpt); lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP); - the_lnet.ln_counters[cpt]->drop_count++; - the_lnet.ln_counters[cpt]->drop_length += nob; + the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++; + the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob; lnet_net_unlock(cpt); lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob); @@ -3261,7 +3894,7 @@ lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg) info.mi_rlength = hdr->payload_length; info.mi_roffset = hdr->msg.put.offset; info.mi_mbits = hdr->msg.put.match_bits; - info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni); + info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni); msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL; ready_delay = msg->msg_rx_ready_delay; @@ -3324,7 +3957,7 @@ lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get) info.mi_rlength = hdr->msg.get.sink_length; info.mi_roffset = hdr->msg.get.src_offset; info.mi_mbits = hdr->msg.get.match_bits; - info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni); + info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni); rc = lnet_ptl_match_md(&info, msg); if (rc == LNET_MATCHMD_DROP) { @@ -3356,7 +3989,7 @@ lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get) lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0); msg->msg_receiving = 0; - rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY); + rc = lnet_send(ni->ni_nid, msg, msg->msg_from); if (rc < 0) { /* didn't get as far as lnet_ni_send() */ CERROR("%s: Unable to send REPLY for GET from %s: %d\n", @@ -3372,13 +4005,13 @@ lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get) static int lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg) { - void *private = msg->msg_private; - struct lnet_hdr *hdr = &msg->msg_hdr; + void *private = msg->msg_private; + struct lnet_hdr *hdr = &msg->msg_hdr; struct lnet_process_id src = {0}; - struct lnet_libmd *md; - int rlength; - int mlength; - int cpt; + struct lnet_libmd *md; + int rlength; + int mlength; + int cpt; cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie); lnet_res_lock(cpt); @@ -3439,10 +4072,10 @@ lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg) static int lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg) { - struct lnet_hdr *hdr = &msg->msg_hdr; + struct lnet_hdr *hdr = &msg->msg_hdr; struct lnet_process_id src = {0}; - struct lnet_libmd *md; - int cpt; + struct lnet_libmd *md; + int cpt; src.nid = hdr->src_nid; src.pid = hdr->src_pid; @@ -3627,16 +4260,17 @@ int lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid, void *private, int rdma_req) { - int rc = 0; - int cpt; - int for_me; - struct lnet_msg *msg; - lnet_pid_t dest_pid; - lnet_nid_t dest_nid; - lnet_nid_t src_nid; struct lnet_peer_ni *lpni; - __u32 payload_length; - __u32 type; + struct lnet_msg *msg; + __u32 payload_length; + lnet_pid_t dest_pid; + lnet_nid_t dest_nid; + lnet_nid_t src_nid; + bool push = false; + int for_me; + __u32 type; + int rc = 0; + int cpt; LASSERT (!in_interrupt ()); @@ -3691,16 +4325,22 @@ lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid, } if (the_lnet.ln_routing && - ni->ni_last_alive != ktime_get_real_seconds()) { - /* NB: so far here is the only place to set NI status to "up */ + ni->ni_net->net_last_alive != ktime_get_real_seconds()) { lnet_ni_lock(ni); - ni->ni_last_alive = ktime_get_real_seconds(); + spin_lock(&ni->ni_net->net_lock); + ni->ni_net->net_last_alive = ktime_get_real_seconds(); + spin_unlock(&ni->ni_net->net_lock); if (ni->ni_status != NULL && - ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) + ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) { ni->ni_status->ns_status = LNET_NI_STATUS_UP; + push = true; + } lnet_ni_unlock(ni); } + if (push) + lnet_push_update_to_peers(1); + /* Regard a bad destination NID as a protocol error. Senders should * know what they're doing; if they don't they're misconfigured, buggy * or malicious so we chop them off at the knees :) */ @@ -3758,7 +4398,7 @@ lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid, } if (!list_empty(&the_lnet.ln_drop_rules) && - lnet_drop_rule_match(hdr)) { + lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) { CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate" "silent message loss\n", libcfs_nid2str(from_nid), libcfs_nid2str(src_nid), @@ -3766,6 +4406,60 @@ lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid, goto drop; } + if (lnet_drop_asym_route && for_me && + LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) { + struct lnet_net *net; + struct lnet_remotenet *rnet; + bool found = true; + + /* we are dealing with a routed message, + * so see if route to reach src_nid goes through from_nid + */ + lnet_net_lock(cpt); + net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid)); + if (!net) { + lnet_net_unlock(cpt); + CERROR("net %s not found\n", + libcfs_net2str(LNET_NIDNET(ni->ni_nid))); + return -EPROTO; + } + + rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid)); + if (rnet) { + struct lnet_peer *gw = NULL; + struct lnet_peer_ni *lpni = NULL; + struct lnet_route *route; + + list_for_each_entry(route, &rnet->lrn_routes, lr_list) { + found = false; + gw = route->lr_gateway; + if (route->lr_lnet != net->net_id) + continue; + /* + * if the nid is one of the gateway's NIDs + * then this is a valid gateway + */ + while ((lpni = lnet_get_next_peer_ni_locked(gw, + NULL, lpni)) != NULL) { + if (lpni->lpni_nid == from_nid) { + found = true; + break; + } + } + } + } + lnet_net_unlock(cpt); + if (!found) { + /* we would not use from_nid to route a message to + * src_nid + * => asymmetric routing detected but forbidden + */ + CERROR("%s, src %s: Dropping asymmetrical route %s\n", + libcfs_nid2str(from_nid), + libcfs_nid2str(src_nid), lnet_msgtyp2str(type)); + goto drop; + } + } msg = lnet_msg_alloc(); if (msg == NULL) { @@ -3816,24 +4510,22 @@ lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid, return 0; goto drop; } + + if (the_lnet.ln_routing) + lpni->lpni_last_alive = ktime_get_seconds(); + msg->msg_rxpeer = lpni; msg->msg_rxni = ni; lnet_ni_addref_locked(ni, cpt); /* Multi-Rail: Primary NID of source. */ msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid); - if (lnet_isrouter(msg->msg_rxpeer)) { - lnet_peer_set_alive(msg->msg_rxpeer); - if (avoid_asym_router_failure && - LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) { - /* received a remote message from router, update - * remote NI status on this router. - * NB: multi-hop routed message will be ignored. - */ - lnet_router_ni_update_locked(msg->msg_rxpeer, - LNET_NIDNET(src_nid)); - } - } + /* + * mark the status of this lpni as UP since we received a message + * from it. The ping response reports back the ns_status which is + * marked on the remote as up or down and we cache it here. + */ + msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP; lnet_msg_commit(msg, cpt); @@ -3952,6 +4644,50 @@ lnet_recv_delayed_msg_list(struct list_head *head) } } +static void +lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt, + struct lnet_libmd *md, struct lnet_handle_md mdh) +{ + s64 timeout_ns; + struct lnet_rsp_tracker *local_rspt; + + /* + * MD has a refcount taken by message so it's not going away. + * The MD however can be looked up. We need to secure the access + * to the md_rspt_ptr by taking the res_lock. + * The rspt can be accessed without protection up to when it gets + * added to the list. + */ + + lnet_res_lock(cpt); + local_rspt = md->md_rspt_ptr; + timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC; + if (local_rspt != NULL) { + /* + * we already have an rspt attached to the md, so we'll + * update the deadline on that one. + */ + lnet_rspt_free(rspt, cpt); + } else { + /* new md */ + rspt->rspt_mdh = mdh; + rspt->rspt_cpt = cpt; + /* store the rspt so we can access it when we get the REPLY */ + md->md_rspt_ptr = rspt; + local_rspt = rspt; + } + local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns); + + /* + * add to the list of tracked responses. It's added to tail of the + * list in order to expire all the older entries first. + */ + lnet_net_lock(cpt); + list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]); + lnet_net_unlock(cpt); + lnet_res_unlock(cpt); +} + /** * Initiate an asynchronous PUT operation. * @@ -4002,10 +4738,11 @@ LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack, __u64 match_bits, unsigned int offset, __u64 hdr_data) { - struct lnet_msg *msg; - struct lnet_libmd *md; - int cpt; - int rc; + struct lnet_msg *msg; + struct lnet_libmd *md; + int cpt; + int rc; + struct lnet_rsp_tracker *rspt = NULL; LASSERT(the_lnet.ln_refcount > 0); @@ -4025,6 +4762,17 @@ LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack, msg->msg_vmflush = !!memory_pressure_get(); cpt = lnet_cpt_of_cookie(mdh.cookie); + + if (ack == LNET_ACK_REQ) { + rspt = lnet_rspt_alloc(cpt); + if (!rspt) { + CERROR("Dropping PUT to %s: ENOMEM on response tracker\n", + libcfs_id2str(target)); + return -ENOMEM; + } + INIT_LIST_HEAD(&rspt->rspt_on_list); + } + lnet_res_lock(cpt); md = lnet_handle2md(&mdh); @@ -4037,6 +4785,7 @@ LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack, md->md_me->me_portal); lnet_res_unlock(cpt); + lnet_rspt_free(rspt, cpt); lnet_msg_free(msg); return -ENOENT; } @@ -4069,7 +4818,15 @@ LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack, lnet_build_msg_event(msg, LNET_EVENT_SEND); - rc = lnet_send(self, msg, LNET_NID_ANY); + if (ack == LNET_ACK_REQ) + lnet_attach_rsp_tracker(rspt, cpt, md, mdh); + + if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2, + CFS_FAIL_ONCE)) + rc = -EIO; + else + rc = lnet_send(self, msg, LNET_NID_ANY); + if (rc != 0) { CNETERR("Error sending PUT to %s: %d\n", libcfs_id2str(target), rc); @@ -4151,8 +4908,9 @@ lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg) lnet_net_lock(cpt); lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP); - the_lnet.ln_counters[cpt]->drop_count++; - the_lnet.ln_counters[cpt]->drop_length += getmd->md_length; + the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++; + the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += + getmd->md_length; lnet_net_unlock(cpt); if (msg != NULL) @@ -4205,10 +4963,11 @@ LNetGet(lnet_nid_t self, struct lnet_handle_md mdh, struct lnet_process_id target, unsigned int portal, __u64 match_bits, unsigned int offset, bool recovery) { - struct lnet_msg *msg; - struct lnet_libmd *md; - int cpt; - int rc; + struct lnet_msg *msg; + struct lnet_libmd *md; + struct lnet_rsp_tracker *rspt; + int cpt; + int rc; LASSERT(the_lnet.ln_refcount > 0); @@ -4221,15 +4980,24 @@ LNetGet(lnet_nid_t self, struct lnet_handle_md mdh, } msg = lnet_msg_alloc(); - if (msg == NULL) { + if (!msg) { CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n", libcfs_id2str(target)); return -ENOMEM; } + cpt = lnet_cpt_of_cookie(mdh.cookie); + + rspt = lnet_rspt_alloc(cpt); + if (!rspt) { + CERROR("Dropping GET to %s: ENOMEM on response tracker\n", + libcfs_id2str(target)); + return -ENOMEM; + } + INIT_LIST_HEAD(&rspt->rspt_on_list); + msg->msg_recovery = recovery; - cpt = lnet_cpt_of_cookie(mdh.cookie); lnet_res_lock(cpt); md = lnet_handle2md(&mdh); @@ -4244,6 +5012,7 @@ LNetGet(lnet_nid_t self, struct lnet_handle_md mdh, lnet_res_unlock(cpt); lnet_msg_free(msg); + lnet_rspt_free(rspt, cpt); return -ENOENT; } @@ -4268,6 +5037,8 @@ LNetGet(lnet_nid_t self, struct lnet_handle_md mdh, lnet_build_msg_event(msg, LNET_EVENT_SEND); + lnet_attach_rsp_tracker(rspt, cpt, md, mdh); + rc = lnet_send(self, msg, LNET_NID_ANY); if (rc < 0) { CNETERR("Error sending GET to %s: %d\n", @@ -4336,9 +5107,9 @@ LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) * current net namespace. * If not, assign order above 0xffff0000, * to make this ni not a priority. */ - if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns)) - order += 0xffff0000; - + if (current->nsproxy && + !net_eq(ni->ni_net_ns, current->nsproxy->net_ns)) + order += 0xffff0000; if (srcnidp != NULL) *srcnidp = ni->ni_nid; if (orderp != NULL) @@ -4377,9 +5148,10 @@ LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) LASSERT(shortest != NULL); hops = shortest_hops; if (srcnidp != NULL) { - ni = lnet_get_next_ni_locked( - shortest->lr_gateway->lpni_net, - NULL); + struct lnet_net *net; + net = lnet_get_net_locked(shortest->lr_lnet); + LASSERT(net); + ni = lnet_get_next_ni_locked(net, NULL); *srcnidp = ni->ni_nid; } if (orderp != NULL)