X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Flib-move.c;h=615d755593c58490c374a7e61cb96c2c805c1ae3;hp=f3f47ffe4468a6ba095f4ff17407143ba8a2047a;hb=275016fcae729f8467dc09aa15aa2b4bb2690875;hpb=9ee453928ab854f43016624578b95c76e73a6a47 diff --git a/lnet/lnet/lib-move.c b/lnet/lnet/lib-move.c index f3f47ff..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; @@ -794,12 +791,32 @@ lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg) return rc; } +static bool +lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now) +{ + time64_t deadline; + + deadline = lpni->lpni_last_alive + + lpni->lpni_net->net_tunables.lct_peer_timeout; + + /* + * assume peer_ni is alive as long as we're within the configured + * peer timeout + */ + if (deadline > now) + return false; + + 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 *lpni, struct lnet_msg *msg) { + time64_t now = ktime_get_seconds(); + if (!lnet_peer_aliveness_enabled(lpni)) return -ENODEV; @@ -819,6 +836,9 @@ lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni, msg->msg_type == LNET_MSG_REPLY) return 1; + if (!lnet_is_peer_deadline_passed(lpni, now)) + return true; + return lnet_is_peer_ni_alive(lpni); } @@ -1294,9 +1314,8 @@ routing_off: } } -#if 0 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; @@ -1312,18 +1331,134 @@ lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2) return 0; } -#endif +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) { - /* TODO re-implement gateway comparison - 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; @@ -1337,49 +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; } -/* TODO: lnet_find_route_locked() needs to be reimplemented */ static struct lnet_route * -lnet_find_route_locked(struct lnet_net *net, __u32 remote_net, - lnet_nid_t rtr_nid, struct lnet_route **prev_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_peer_ni *lpni, *best_gw_ni = NULL; struct lnet_route *best_route; struct lnet_route *last_route; - struct lnet_peer *lp_best; - struct lnet_peer *lp; + struct lnet_route *route; int rc; + __u32 restrict_net; + __u32 any_net = LNET_NIDNET(LNET_NID_ANY); - /* 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; - - lp_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 (lp_best == NULL) { - best_route = last_route = route; - lp_best = 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)); + continue; } @@ -1388,14 +1517,38 @@ 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; + } + + if (rc == 1) { + best_route = route; + best_gw_ni = lpni; + continue; + } - best_route = route; - lp_best = lp; + 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; + } } *prev_route = last_route; + *gwni = best_gw_ni; return best_route; } @@ -1681,137 +1834,19 @@ lnet_handle_send(struct lnet_send_data *sd) rc = lnet_post_send_locked(msg, 0); if (!rc) - CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n", + 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 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. - */ - 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; - 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 - */ - ni_is_pref = lnet_peer_is_pref_nid_locked(lpni, - best_ni->ni_nid); - - lpni_healthv = atomic_read(&lpni->lpni_healthv); - - 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); - - /* 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_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); - - 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(sd, peer, peer_net); -} - static inline void lnet_set_non_mr_pref_nid(struct lnet_send_data *sd) { @@ -1861,8 +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) @@ -1875,17 +1913,6 @@ 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 && sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) return lnet_handle_lo_send(sd); @@ -1941,6 +1968,18 @@ lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, 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); @@ -1948,15 +1987,21 @@ lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, } /* 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; - spin_lock(&peer->lp_lock); 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); - primary_nid = peer->lp_primary_nid; CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n", msg, libcfs_nid2str(primary_nid)); @@ -1970,97 +2015,135 @@ lnet_handle_find_routed_path(struct lnet_send_data *sd, struct lnet_peer_ni **gw_lpni, struct lnet_peer **gw_peer) { + int rc; + __u32 local_lnet; struct lnet_peer *gw; - struct lnet_route *best_route; - struct lnet_route *last_route; + 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; - best_route = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid), - sd->sd_rtr_nid, &last_route); - if (!best_route) { - 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; - gw = best_route->lr_gateway; - *gw_peer = gw; + 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; + } + + 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; + + } /* * Discover this gateway if it hasn't already been discovered. * This means we might delay the message until discovery has * completed */ -#if 0 - /* TODO: disable discovey for now */ - if (lnet_msg_discovery(sd->sd_msg) && - !lnet_peer_is_uptodate(*gw_peer)) { - sd->sd_msg->msg_src_nid_param = sd->sd_src_nid; - return lnet_initiate_peer_discovery(gw, sd->sd_msg, - sd->sd_rtr_nid, sd->sd_cpt); - } -#endif + 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, lnet_peer_get_net_locked(gw, - best_route->lr_lnet), + 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(best_route->lr_lnet), + 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; } - /* - * if gw is MR let's find its best peer_ni - */ - if (lnet_peer_is_multi_rail(gw)) { - lpni = lnet_find_best_lpni_on_net(sd, gw, - 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 (!lpni) { - CERROR("Internal Error. Route expected to %s from %s\n", - libcfs_nid2str(dst_nid), - libcfs_nid2str(src_nid)); - return -EFAULT; - } - } else { - struct lnet_peer_net *lpn; - lpn = lnet_peer_get_net_locked(gw, best_route->lr_lnet); - if (!lpn) { - CERROR("Internal Error. Route expected to %s from %s\n", - libcfs_nid2str(dst_nid), - libcfs_nid2str(src_nid)); - return -EFAULT; - } - lpni = list_entry(lpn->lpn_peer_nis.next, struct lnet_peer_ni, - lpni_peer_nis); - if (!lpni) { - CERROR("Internal Error. Route expected to %s from %s\n", - libcfs_nid2str(dst_nid), - libcfs_nid2str(src_nid)); - return -EFAULT; - } - } - - *gw_lpni = lpni; + *gw_lpni = gwni; + *gw_peer = gw; /* - * increment the route sequence number since now we're sure we're - * going to use it + * increment the sequence numbers since now we're sure we're + * going to use this path */ - LASSERT(best_route && last_route); - best_route->lr_seq = last_route->lr_seq + 1; + 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++; + } return 0; } @@ -2120,7 +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; @@ -2142,6 +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) @@ -2311,10 +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); /* @@ -2421,11 +2515,11 @@ 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); @@ -2591,19 +2685,20 @@ 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)) { - rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt); + rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt); + if (rc) { lnet_peer_ni_decref_locked(lpni); lnet_net_unlock(cpt); return rc; } lnet_peer_ni_decref_locked(lpni); + peer = lpni->lpni_peer_net->lpn_peer; + /* * Identify the different send cases */ @@ -2684,8 +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); @@ -2719,25 +2819,57 @@ lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt) return; rspt = md->md_rspt_ptr; - md->md_rspt_ptr = NULL; /* debug code */ LASSERT(rspt->rspt_cpt == cpt); - /* - * 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); + 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(bool force) +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) @@ -2754,6 +2886,8 @@ lnet_finalize_expired_responses(bool force) 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 @@ -2769,41 +2903,74 @@ lnet_finalize_expired_responses(bool force) lnet_res_lock(i); if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) { lnet_res_unlock(i); - list_del_init(&rspt->rspt_on_list); + list_del(&rspt->rspt_on_list); lnet_rspt_free(rspt, i); continue; } - if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 || - force) { + 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); - list_del_init(&rspt->rspt_on_list); - lnet_rspt_free(rspt, i); continue; } LASSERT(md->md_rspt_ptr == rspt); md->md_rspt_ptr = NULL; lnet_res_unlock(i); - lnet_net_lock(i); - the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++; - lnet_net_unlock(i); - - list_del_init(&rspt->rspt_on_list); + LNetMDUnlink(rspt->rspt_mdh); nid = rspt->rspt_next_hop_nid; - CNETERR("Response timed out: md = %p: nid = %s\n", - md, libcfs_nid2str(nid)); - LNetMDUnlink(rspt->rspt_mdh); + 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 @@ -2822,10 +2989,11 @@ lnet_finalize_expired_responses(bool force) } } - lnet_net_lock(i); - if (!list_empty(&local_queue)) + if (!list_empty(&local_queue)) { + lnet_net_lock(i); list_splice(&local_queue, the_lnet.ln_mt_rstq[i]); - lnet_net_unlock(i); + lnet_net_unlock(i); + } } } @@ -3098,26 +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) { @@ -3360,6 +3508,7 @@ lnet_monitor_thread(void *arg) 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 @@ -3381,7 +3530,7 @@ lnet_monitor_thread(void *arg) lnet_resend_pending_msgs(); if (now >= rsp_timeout) { - lnet_finalize_expired_responses(false); + lnet_finalize_expired_responses(); rsp_timeout = now + (lnet_transaction_timeout / 2); } @@ -3399,12 +3548,22 @@ 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. */ interval = min(lnet_recovery_interval, - lnet_transaction_timeout / 2); - wait_event_interruptible_timeout(the_lnet.ln_mt_waitq, - false, - cfs_time_seconds(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); } /* Shutting down */ @@ -3559,6 +3718,7 @@ lnet_mt_event_handler(struct lnet_event *event) case LNET_EVENT_UNLINK: CDEBUG(D_NET, "%s recovery ping unlinked\n", libcfs_nid2str(ev_info->mt_nid)); + /* fallthrough */ case LNET_EVENT_REPLY: lnet_handle_recovery_reply(ev_info, event->status, event->type == LNET_EVENT_UNLINK); @@ -3597,7 +3757,7 @@ lnet_rsp_tracker_create(void) static void lnet_rsp_tracker_clean(void) { - lnet_finalize_expired_responses(true); + lnet_finalize_expired_responses(); cfs_percpt_free(the_lnet.ln_mt_rstq); the_lnet.ln_mt_rstq = NULL; @@ -3631,9 +3791,6 @@ int lnet_monitor_thr_start(void) goto clean_thread; } - /* post monitor thread start processing */ - lnet_router_post_mt_start(); - return 0; clean_thread: @@ -3671,7 +3828,7 @@ void lnet_monitor_thr_stop(void) 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); @@ -3682,8 +3839,6 @@ void lnet_monitor_thr_stop(void) lnet_clean_local_ni_recoveryq(); lnet_clean_peer_ni_recoveryq(); lnet_clean_resendqs(); - - return; } void @@ -4105,16 +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 ()); @@ -4169,16 +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 :) */ @@ -4236,7 +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, NULL)) { + 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), @@ -4348,6 +4510,10 @@ 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); @@ -4483,7 +4649,6 @@ lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt, struct lnet_libmd *md, struct lnet_handle_md mdh) { s64 timeout_ns; - bool new_entry = true; struct lnet_rsp_tracker *local_rspt; /* @@ -4502,8 +4667,7 @@ lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt, * we already have an rspt attached to the md, so we'll * update the deadline on that one. */ - LIBCFS_FREE(rspt, sizeof(*rspt)); - new_entry = false; + lnet_rspt_free(rspt, cpt); } else { /* new md */ rspt->rspt_mdh = mdh; @@ -4519,9 +4683,7 @@ lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt, * list in order to expire all the older entries first. */ lnet_net_lock(cpt); - if (!new_entry && !list_empty(&local_rspt->rspt_on_list)) - list_del_init(&local_rspt->rspt_on_list); - list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]); + list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]); lnet_net_unlock(cpt); lnet_res_unlock(cpt); } @@ -4623,7 +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); - LIBCFS_FREE(rspt, sizeof(*rspt)); + lnet_rspt_free(rspt, cpt); lnet_msg_free(msg); return -ENOENT; } @@ -4850,7 +5012,7 @@ LNetGet(lnet_nid_t self, struct lnet_handle_md mdh, lnet_res_unlock(cpt); lnet_msg_free(msg); - LIBCFS_FREE(rspt, sizeof(*rspt)); + lnet_rspt_free(rspt, cpt); return -ENOENT; } @@ -4945,9 +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)