X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fgnilnd%2Fgnilnd.c;h=74d0fa2b2dffe19334c12d0c1fa702c6be6994f7;hp=e110c27aecec9d65602a24216cc6f828850c4280;hb=254aec0ff702a856e688e0e61e773d63b3285171;hpb=15e06657921e3af9b1fb8dcaf80ff37f5c791b07 diff --git a/lnet/klnds/gnilnd/gnilnd.c b/lnet/klnds/gnilnd/gnilnd.c index e110c27..74d0fa2b 100644 --- a/lnet/klnds/gnilnd/gnilnd.c +++ b/lnet/klnds/gnilnd/gnilnd.c @@ -1,7 +1,7 @@ /* * Copyright (C) 2012 Cray, Inc. * - * Copyright (c) 2013, 2014, Intel Corporation. + * Copyright (c) 2013, 2015, Intel Corporation. * * Author: Nic Henke * Author: James Shimek @@ -117,8 +117,8 @@ kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn) newconn->gnc_peer_connstamp == conn->gnc_my_connstamp) { CDEBUG(D_NET, "skipping prune of %p, " "loopback and matching stamps" - " connstamp "LPU64"("LPU64")" - " peerstamp "LPU64"("LPU64")\n", + " connstamp %llu(%llu)" + " peerstamp %llu(%llu)\n", conn, newconn->gnc_my_connstamp, conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp, @@ -128,25 +128,25 @@ kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn) if (conn->gnc_peerstamp != newconn->gnc_peerstamp) { LASSERTF(conn->gnc_peerstamp < newconn->gnc_peerstamp, - "conn 0x%p peerstamp "LPU64" >= " - "newconn 0x%p peerstamp "LPU64"\n", + "conn 0x%p peerstamp %llu >= " + "newconn 0x%p peerstamp %llu\n", conn, conn->gnc_peerstamp, newconn, newconn->gnc_peerstamp); CDEBUG(D_NET, "Closing stale conn nid: %s " - " peerstamp:"LPX64"("LPX64")\n", + " peerstamp:%#llx(%#llx)\n", libcfs_nid2str(peer->gnp_nid), conn->gnc_peerstamp, newconn->gnc_peerstamp); } else { LASSERTF(conn->gnc_peer_connstamp < newconn->gnc_peer_connstamp, - "conn 0x%p peer_connstamp "LPU64" >= " - "newconn 0x%p peer_connstamp "LPU64"\n", + "conn 0x%p peer_connstamp %llu >= " + "newconn 0x%p peer_connstamp %llu\n", conn, conn->gnc_peer_connstamp, newconn, newconn->gnc_peer_connstamp); CDEBUG(D_NET, "Closing stale conn nid: %s" - " connstamp:"LPU64"("LPU64")\n", + " connstamp:%llu(%llu)\n", libcfs_nid2str(peer->gnp_nid), conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp); } @@ -175,8 +175,8 @@ kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn) list_for_each(tmp, &peer->gnp_conns) { conn = list_entry(tmp, kgn_conn_t, gnc_list); CDEBUG(D_NET, "checking conn 0x%p for peer %s" - " lo %d new "LPU64" existing "LPU64 - " new peer "LPU64" existing peer "LPU64 + " lo %d new %llu existing %llu" + " new peer %llu existing peer %llu" " new dev %p existing dev %p\n", conn, libcfs_nid2str(peer->gnp_nid), loopback, @@ -319,7 +319,8 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev) failed: atomic_dec(&kgnilnd_data.kgn_nconns); - LIBCFS_FREE(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *)); + kgnilnd_vfree(conn->gnc_tx_ref_table, + GNILND_MAX_MSG_ID * sizeof(void *)); LIBCFS_FREE(conn, sizeof(*conn)); return rc; } @@ -493,8 +494,8 @@ kgnilnd_destroy_conn(kgn_conn_t *conn) kgnilnd_peer_decref(conn->gnc_peer); if (conn->gnc_tx_ref_table != NULL) { - LIBCFS_FREE(conn->gnc_tx_ref_table, - GNILND_MAX_MSG_ID * sizeof(void *)); + kgnilnd_vfree(conn->gnc_tx_ref_table, + GNILND_MAX_MSG_ID * sizeof(void *)); } LIBCFS_FREE(conn, sizeof(*conn)); @@ -635,7 +636,7 @@ kgnilnd_close_conn_locked(kgn_conn_t *conn, int error) /* if we NETERROR, make sure it is rate limited */ if (!kgnilnd_conn_clean_errno(error) && - peer->gnp_down == GNILND_RCA_NODE_UP) { + peer->gnp_state != GNILND_PEER_DOWN) { CNETERR("closing conn to %s: error %d\n", libcfs_nid2str(peer->gnp_nid), error); } else { @@ -671,6 +672,10 @@ kgnilnd_close_conn_locked(kgn_conn_t *conn, int error) conn->gnc_state = GNILND_CONN_CLOSING; } + if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) { + msleep_interruptible(MSEC_PER_SEC); + } + /* leave on peer->gnp_conns to make sure we don't let the reaper * or others try to unlink this peer until the conn is fully * processed for closing */ @@ -804,19 +809,13 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn) logmsg = (nlive + nrdma + nq_rdma); if (logmsg) { - if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_UP) { - CNETERR("Closed conn 0x%p->%s (errno %d, peer errno %d): " - "canceled %d TX, %d/%d RDMA\n", - conn, libcfs_nid2str(conn->gnc_peer->gnp_nid), - conn->gnc_error, conn->gnc_peer_error, - nlive, nq_rdma, nrdma); - } else { - CDEBUG(D_NET, "Closed conn 0x%p->%s (errno %d," - " peer errno %d): canceled %d TX, %d/%d RDMA\n", - conn, libcfs_nid2str(conn->gnc_peer->gnp_nid), - conn->gnc_error, conn->gnc_peer_error, - nlive, nq_rdma, nrdma); - } + int level = conn->gnc_peer->gnp_state == GNILND_PEER_UP ? + D_NETERROR : D_NET; + CDEBUG(level, "Closed conn 0x%p->%s (errno %d," + " peer errno %d): canceled %d TX, %d/%d RDMA\n", + conn, libcfs_nid2str(conn->gnc_peer->gnp_nid), + conn->gnc_error, conn->gnc_peer_error, + nlive, nq_rdma, nrdma); } kgnilnd_destroy_conn_ep(conn); @@ -920,8 +919,8 @@ kgnilnd_set_conn_params(kgn_dgram_t *dgram) /* log this for help in debuggin SMSG buffer re-use */ CDEBUG(D_NET, "conn %p src %s dst %s smsg %p acquired" - " local cqid %u SMSG %p->%u hndl "LPX64"."LPX64 - " remote cqid %u SMSG %p->%u hndl "LPX64"."LPX64"\n", + " local cqid %u SMSG %p->%u hndl %#llx.%#llx" + " remote cqid %u SMSG %p->%u hndl %#llx.%#llx\n", conn, libcfs_nid2str(connreq->gncr_srcnid), libcfs_nid2str(connreq->gncr_dstnid), &conn->gnpr_smsg_attr, @@ -1000,7 +999,7 @@ kgnilnd_create_peer_safe(kgn_peer_t **peerp, return -ENOMEM; } peer->gnp_nid = nid; - peer->gnp_down = node_state; + peer->gnp_state = node_state; /* translate from nid to nic addr & store */ rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id); @@ -1220,7 +1219,7 @@ kgnilnd_release_purgatory_list(struct list_head *conn_list) list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list, gmp_list) { CDEBUG(D_NET, - "dev %p releasing held mdd "LPX64"."LPX64"\n", + "dev %p releasing held mdd %#llx.%#llx\n", conn->gnc_device, gmp->gmp_map_key.qword1, gmp->gmp_map_key.qword2); @@ -1720,9 +1719,8 @@ kgnilnd_report_node_state(lnet_nid_t nid, int down) write_unlock(&kgnilnd_data.kgn_peer_conn_lock); /* Don't add a peer for node up events */ - if (down == GNILND_RCA_NODE_UP) { + if (down == GNILND_PEER_UP) return 0; - } /* find any valid net - we don't care which one... */ down_read(&kgnilnd_data.kgn_net_rw_sem); @@ -1766,9 +1764,9 @@ kgnilnd_report_node_state(lnet_nid_t nid, int down) } } - peer->gnp_down = down; + peer->gnp_state = down; - if (down == GNILND_RCA_NODE_DOWN) { + if (down == GNILND_PEER_DOWN) { kgn_conn_t *conn; peer->gnp_down_event_time = jiffies; @@ -1784,7 +1782,7 @@ kgnilnd_report_node_state(lnet_nid_t nid, int down) write_unlock(&kgnilnd_data.kgn_peer_conn_lock); - if (down == GNILND_RCA_NODE_DOWN) { + if (down == GNILND_PEER_DOWN) { /* using ENETRESET so we don't get messages from * kgnilnd_tx_done */ @@ -2128,7 +2126,7 @@ kgnilnd_dev_fini(kgn_device_t *dev) dev->gnd_map_nphys, dev->gnd_map_physnop); LASSERTF(dev->gnd_map_nvirt == 0 && dev->gnd_map_virtnob == 0, - "%d virtual mappings of "LPU64" bytes still mapped\n", + "%d virtual mappings of %llu bytes still mapped\n", dev->gnd_map_nvirt, dev->gnd_map_virtnob); LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 && @@ -2686,9 +2684,9 @@ kgnilnd_startup(lnet_ni_t *ni) kgn_net_t *net; ENTRY; - LASSERTF(ni->ni_lnd == &the_kgnilnd, + LASSERTF(ni->ni_net->net_lnd == &the_kgnilnd, "bad LND 0x%p != the_kgnilnd @ 0x%p\n", - ni->ni_lnd, &the_kgnilnd); + ni->ni_net->net_lnd, &the_kgnilnd); if (kgnilnd_data.kgn_init == GNILND_INIT_NOTHING) { rc = kgnilnd_base_startup(); @@ -2708,8 +2706,12 @@ kgnilnd_startup(lnet_ni_t *ni) INIT_LIST_HEAD(&net->gnn_list); ni->ni_data = net; net->gnn_ni = ni; - ni->ni_maxtxcredits = *kgnilnd_tunables.kgn_credits; - ni->ni_peertxcredits = *kgnilnd_tunables.kgn_peer_credits; + if (!ni->ni_net->net_tunables_set) { + ni->ni_net->net_tunables.lct_max_tx_credits = + *kgnilnd_tunables.kgn_credits; + ni->ni_net->net_tunables.lct_peer_tx_credits = + *kgnilnd_tunables.kgn_peer_credits; + } if (*kgnilnd_tunables.kgn_peer_health) { int fudge; @@ -2719,9 +2721,10 @@ kgnilnd_startup(lnet_ni_t *ni) fudge = (GNILND_TO2KA(*kgnilnd_tunables.kgn_timeout) / GNILND_REAPER_NCHECKS); timeout = *kgnilnd_tunables.kgn_timeout + fudge; - if (*kgnilnd_tunables.kgn_peer_timeout >= timeout) - ni->ni_peertimeout = *kgnilnd_tunables.kgn_peer_timeout; - else if (*kgnilnd_tunables.kgn_peer_timeout > -1) { + if (*kgnilnd_tunables.kgn_peer_timeout >= timeout) { + ni->ni_net->net_tunables.lct_peer_timeout = + *kgnilnd_tunables.kgn_peer_timeout; + } else if (*kgnilnd_tunables.kgn_peer_timeout > -1) { LCONSOLE_ERROR("Peer_timeout is set to %d but needs to be >= %d\n", *kgnilnd_tunables.kgn_peer_timeout, timeout); @@ -2729,10 +2732,10 @@ kgnilnd_startup(lnet_ni_t *ni) LIBCFS_FREE(net, sizeof(*net)); GOTO(failed, rc = -EINVAL); } else - ni->ni_peertimeout = timeout; + ni->ni_net->net_tunables.lct_peer_timeout = timeout; LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n", - ni->ni_peertimeout); + ni->ni_net->net_tunables.lct_peer_timeout); } atomic_set(&net->gnn_refcount, 1); @@ -2874,17 +2877,14 @@ out: EXIT; } -void __exit -kgnilnd_module_fini(void) +static void __exit kgnilnd_exit(void) { lnet_unregister_lnd(&the_kgnilnd); kgnilnd_proc_fini(); kgnilnd_remove_sysctl(); - kgnilnd_tunables_fini(); } -int __init -kgnilnd_module_init(void) +static int __init kgnilnd_init(void) { int rc; @@ -2903,8 +2903,9 @@ kgnilnd_module_init(void) } MODULE_AUTHOR("Cray, Inc. "); -MODULE_DESCRIPTION("Kernel Gemini LND v"KGNILND_BUILD_REV); +MODULE_DESCRIPTION("Gemini LNet Network Driver"); +MODULE_VERSION(KGNILND_BUILD_REV); MODULE_LICENSE("GPL"); -module_init(kgnilnd_module_init); -module_exit(kgnilnd_module_fini); +module_init(kgnilnd_init); +module_exit(kgnilnd_exit);