X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fgnilnd%2Fgnilnd.c;h=bbc3daa4d44e3598a2d6f04a54dfb79da04fafbb;hp=fe99809a15311e3cdb364e57384e8806092e3598;hb=d50617709841a37afd9db00ea6c4ace94828a3c2;hpb=e8bf4e3eadf1cec9a0c9dca609a0b023fc5a397d diff --git a/lnet/klnds/gnilnd/gnilnd.c b/lnet/klnds/gnilnd/gnilnd.c index fe99809..bbc3daa 100644 --- a/lnet/klnds/gnilnd/gnilnd.c +++ b/lnet/klnds/gnilnd/gnilnd.c @@ -1,6 +1,8 @@ /* * Copyright (C) 2012 Cray, Inc. * + * Copyright (c) 2013, 2017, Intel Corporation. + * * Author: Nic Henke * Author: James Shimek * @@ -23,23 +25,64 @@ #include "gnilnd.h" /* Primary entry points from LNET. There are no guarantees against reentrance. */ -lnd_t the_kgnilnd = { -#ifdef CONFIG_CRAY_XT +const struct lnet_lnd the_kgnilnd = { .lnd_type = GNILND, -#else - .lnd_type = GNIIPLND, -#endif .lnd_startup = kgnilnd_startup, .lnd_shutdown = kgnilnd_shutdown, .lnd_ctl = kgnilnd_ctl, .lnd_send = kgnilnd_send, .lnd_recv = kgnilnd_recv, .lnd_eager_recv = kgnilnd_eager_recv, - .lnd_query = kgnilnd_query, }; kgn_data_t kgnilnd_data; +int +kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id) +{ + struct task_struct *thrd; + + thrd = kthread_run(fn, arg, "%s_%02d", name, id); + if (IS_ERR(thrd)) + return PTR_ERR(thrd); + + atomic_inc(&kgnilnd_data.kgn_nthreads); + return 0; +} + +/* bind scheduler threads to cpus */ +int +kgnilnd_start_sd_threads(void) +{ + int cpu; + int i = 0; + struct task_struct *task; + + for_each_online_cpu(cpu) { + /* don't bind to cpu 0 - all interrupts are processed here */ + if (cpu == 0) + continue; + + task = kthread_create(kgnilnd_scheduler, (void *)((long)i), + "%s_%02d", "kgnilnd_sd", i); + if (!IS_ERR(task)) { + kthread_bind(task, cpu); + wake_up_process(task); + } else { + CERROR("Can't spawn gnilnd scheduler[%d] %ld\n", i, + PTR_ERR(task)); + return PTR_ERR(task); + } + atomic_inc(&kgnilnd_data.kgn_nthreads); + + if (++i >= *kgnilnd_tunables.kgn_sched_threads) { + break; + } + } + + return 0; +} + /* needs write_lock on kgn_peer_conn_lock */ int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn) @@ -69,8 +112,8 @@ 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, @@ -80,25 +123,25 @@ 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); } @@ -127,8 +170,8 @@ 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, @@ -200,12 +243,15 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev) return -ENOMEM; } - LIBCFS_ALLOC(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *)); + conn->gnc_tx_ref_table = + kgnilnd_vzalloc(GNILND_MAX_MSG_ID * sizeof(void *)); if (conn->gnc_tx_ref_table == NULL) { CERROR("Can't allocate conn tx_ref_table\n"); GOTO(failed, rc = -ENOMEM); } + mutex_init(&conn->gnc_smsg_mutex); + mutex_init(&conn->gnc_rdma_mutex); atomic_set(&conn->gnc_refcount, 1); atomic_set(&conn->gnc_reaper_noop, 0); atomic_set(&conn->gnc_sched_noop, 0); @@ -215,6 +261,7 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev) INIT_LIST_HEAD(&conn->gnc_schedlist); INIT_LIST_HEAD(&conn->gnc_fmaq); INIT_LIST_HEAD(&conn->gnc_mdd_list); + INIT_LIST_HEAD(&conn->gnc_delaylist); spin_lock_init(&conn->gnc_list_lock); spin_lock_init(&conn->gnc_tx_lock); conn->gnc_magic = GNILND_CONN_MAGIC; @@ -224,7 +271,7 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev) conn->gnc_next_tx = (int) GNILND_MAX_MSG_ID - 10; /* if this fails, we have conflicts and MAX_TX is too large */ - CLASSERT(GNILND_MAX_MSG_ID < GNILND_MSGID_CLOSE); + BUILD_BUG_ON(GNILND_MAX_MSG_ID >= GNILND_MSGID_CLOSE); /* get a new unique CQ id for this conn */ write_lock(&kgnilnd_data.kgn_peer_conn_lock); @@ -244,8 +291,8 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev) * check context */ conn->gnc_device = dev; - conn->gnc_timeout = MAX(*kgnilnd_tunables.kgn_timeout, - GNILND_MIN_TIMEOUT); + conn->gnc_timeout = max(*kgnilnd_tunables.kgn_timeout, + GNILND_MIN_TIMEOUT); kgnilnd_update_reaper_timeout(conn->gnc_timeout); /* this is the ep_handle for doing SMSG & BTE */ @@ -268,7 +315,8 @@ 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; } @@ -399,8 +447,9 @@ kgnilnd_destroy_conn(kgn_conn_t *conn) list_empty(&conn->gnc_hashlist) && list_empty(&conn->gnc_schedlist) && list_empty(&conn->gnc_mdd_list) && + list_empty(&conn->gnc_delaylist) && conn->gnc_magic == GNILND_CONN_MAGIC, - "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d\n", + "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d/%d\n", conn, conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid) : "", !!in_interrupt(), conn->gnc_scheduled, @@ -410,7 +459,8 @@ kgnilnd_destroy_conn(kgn_conn_t *conn) list_empty(&conn->gnc_list), list_empty(&conn->gnc_hashlist), list_empty(&conn->gnc_schedlist), - list_empty(&conn->gnc_mdd_list)); + list_empty(&conn->gnc_mdd_list), + list_empty(&conn->gnc_delaylist)); /* Tripping these is especially bad, as it means we have items on the * lists that didn't keep their refcount on the connection - or @@ -442,8 +492,8 @@ 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)); @@ -454,11 +504,13 @@ kgnilnd_destroy_conn(kgn_conn_t *conn) void kgnilnd_peer_alive(kgn_peer_t *peer) { - set_mb(peer->gnp_last_alive, jiffies); + time64_t now = ktime_get_seconds(); + + set_mb(peer->gnp_last_alive, now); } void -kgnilnd_peer_notify(kgn_peer_t *peer, int error) +kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive) { int tell_lnet = 0; int nnets = 0; @@ -487,10 +539,10 @@ kgnilnd_peer_notify(kgn_peer_t *peer, int error) peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting, conn, kgnilnd_data.kgn_in_reset, error); - if ((peer->gnp_connecting == GNILND_PEER_IDLE) && + if (((peer->gnp_connecting == GNILND_PEER_IDLE) && (conn == NULL) && (!kgnilnd_data.kgn_in_reset) && - (!kgnilnd_conn_clean_errno(error))) { + (!kgnilnd_conn_clean_errno(error))) || alive) { tell_lnet = 1; } @@ -524,7 +576,7 @@ kgnilnd_peer_notify(kgn_peer_t *peer, int error) return; } - LIBCFS_ALLOC(nets, nnets * sizeof(*nets)); + CFS_ALLOC_PTR_ARRAY(nets, nnets); if (nets == NULL) { up_read(&kgnilnd_data.kgn_net_rw_sem); @@ -550,17 +602,18 @@ kgnilnd_peer_notify(kgn_peer_t *peer, int error) peer_nid = kgnilnd_lnd2lnetnid(net->gnn_ni->ni_nid, peer->gnp_nid); - CDEBUG(D_NET, "peer 0x%p->%s last_alive %lu (%lus ago)\n", + CDEBUG(D_NET, "peer 0x%p->%s last_alive %lld (%llds ago)\n", peer, libcfs_nid2str(peer_nid), peer->gnp_last_alive, - cfs_duration_sec(jiffies - peer->gnp_last_alive)); - - lnet_notify(net->gnn_ni, peer_nid, 0, peer->gnp_last_alive); + ktime_get_seconds() - peer->gnp_last_alive); + lnet_notify(net->gnn_ni, peer_nid, alive, + (alive) ? true : false, + peer->gnp_last_alive); kgnilnd_net_decref(net); } - LIBCFS_FREE(nets, nnets * sizeof(*nets)); + CFS_FREE_PTR_ARRAY(nets, nnets); } } @@ -584,7 +637,7 @@ 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 { @@ -620,6 +673,10 @@ 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 */ @@ -682,6 +739,11 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn) kgnilnd_conn_state2str(conn)); LASSERT(list_empty(&conn->gnc_hashlist)); + /* We shouldnt be on the delay list, the conn can + * get added to this list during a retransmit, and retransmits + * only occur within scheduler threads. + */ + LASSERT(list_empty(&conn->gnc_delaylist)); /* we've sent the close, start nuking */ if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SCHEDULE_COMPLETE)) @@ -753,19 +815,13 @@ 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); @@ -804,8 +860,8 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn) /* I'm telling Mommy! - use peer_error if they initiated close */ kgnilnd_peer_notify(conn->gnc_peer, - conn->gnc_error == -ECONNRESET ? conn->gnc_peer_error - : conn->gnc_error); + conn->gnc_error == -ECONNRESET ? + conn->gnc_peer_error : conn->gnc_error, 0); EXIT; } @@ -823,7 +879,7 @@ kgnilnd_set_conn_params(kgn_dgram_t *dgram) /* set timeout vals in conn early so we can use them for the NAK */ /* use max of the requested and our timeout, peer will do the same */ - conn->gnc_timeout = MAX(conn->gnc_timeout, connreq->gncr_timeout); + conn->gnc_timeout = max(conn->gnc_timeout, connreq->gncr_timeout); /* only ep_bind really mucks around with the CQ */ /* only ep bind if we are not connecting to ourself and the dstnid is not a wildcard. this check @@ -869,8 +925,8 @@ 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, @@ -949,7 +1005,7 @@ 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); @@ -1164,12 +1220,12 @@ kgnilnd_release_purgatory_list(struct list_head *conn_list) * make sure we tell LNet - if this is from other context, * the checks in the function will prevent an errant * notification */ - kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error); + kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error, 0); 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); @@ -1205,8 +1261,8 @@ kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer) current_to += *kgnilnd_tunables.kgn_min_reconnect_interval / 2; } - current_to = MIN(current_to, - *kgnilnd_tunables.kgn_max_reconnect_interval); + current_to = min(current_to, + *kgnilnd_tunables.kgn_max_reconnect_interval); peer->gnp_reconnect_interval = current_to; CDEBUG(D_NET, "peer %s can reconnect at %lu interval %lu\n", @@ -1537,19 +1593,14 @@ kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, return rc; } - i = 4; - while (atomic_read(&kgnilnd_data.kgn_npending_conns) || - atomic_read(&kgnilnd_data.kgn_npending_detach) || - atomic_read(&kgnilnd_data.kgn_npending_unlink)) { - - cfs_pause(cfs_time_seconds(1)); - i++; - - CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting on %d peers %d closes %d detaches\n", + wait_var_event_warning(&kgnilnd_data, + !atomic_read(&kgnilnd_data.kgn_npending_conns) && + !atomic_read(&kgnilnd_data.kgn_npending_detach) && + !atomic_read(&kgnilnd_data.kgn_npending_unlink), + "Waiting on %d peers %d closes %d detaches\n", atomic_read(&kgnilnd_data.kgn_npending_unlink), atomic_read(&kgnilnd_data.kgn_npending_conns), atomic_read(&kgnilnd_data.kgn_npending_detach)); - } return rc; } @@ -1612,8 +1663,8 @@ kgnilnd_get_conn_info(kgn_peer_t *peer, *device_id = conn->gnc_device->gnd_host_id; *peerstamp = conn->gnc_peerstamp; - *tx_seq = conn->gnc_tx_seq; - *rx_seq = conn->gnc_rx_seq; + *tx_seq = atomic_read(&conn->gnc_tx_seq); + *rx_seq = atomic_read(&conn->gnc_rx_seq); *fmaq_len = kgnilnd_count_list(&conn->gnc_fmaq); *nfma = atomic_read(&conn->gnc_nlive_fma); *nrdma = atomic_read(&conn->gnc_nlive_rdma); @@ -1668,9 +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); @@ -1714,9 +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; @@ -1732,23 +1782,21 @@ 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 */ kgnilnd_txlist_done(&zombies, -ENETRESET); - - if (*kgnilnd_tunables.kgn_peer_health) { - kgnilnd_peer_notify(peer, -ECONNRESET); - } + kgnilnd_peer_notify(peer, -ECONNRESET, 0); + LCONSOLE_INFO("Received down event for nid %d\n", + LNET_NIDADDR(nid)); } - CDEBUG(D_INFO, "marking nid %lld %s\n", nid, down ? "down" : "up"); return 0; } int -kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) +kgnilnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg) { struct libcfs_ioctl_data *data = arg; kgn_net_t *net = ni->ni_data; @@ -1867,74 +1915,6 @@ kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) return rc; } -void -kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when) -{ - kgn_net_t *net = ni->ni_data; - kgn_tx_t *tx; - kgn_peer_t *peer = NULL; - kgn_conn_t *conn = NULL; - lnet_process_id_t id = {.nid = nid, .pid = LUSTRE_SRV_LNET_PID}; - ENTRY; - - /* I expect to find him, so only take a read lock */ - read_lock(&kgnilnd_data.kgn_peer_conn_lock); - peer = kgnilnd_find_peer_locked(nid); - if (peer != NULL) { - /* LIE if in a quiesce - we will update the timeouts after, - * but we don't want sends failing during it */ - if (kgnilnd_data.kgn_quiesce_trigger) { - *when = jiffies; - read_unlock(&kgnilnd_data.kgn_peer_conn_lock); - GOTO(out, 0); - } - - /* Update to best guess, might refine on later checks */ - *when = peer->gnp_last_alive; - - /* we have a peer, how about a conn? */ - conn = kgnilnd_find_conn_locked(peer); - - if (conn == NULL) { - /* if there is no conn, check peer last errno to see if clean disconnect - * - if it was, we lie to LNet because we believe a TX would complete - * on reconnect */ - if (kgnilnd_conn_clean_errno(peer->gnp_last_errno)) { - *when = jiffies; - } - /* we still want to fire a TX and new conn in this case */ - } else { - /* gnp_last_alive is valid, run for the hills */ - read_unlock(&kgnilnd_data.kgn_peer_conn_lock); - GOTO(out, 0); - } - } - /* if we get here, either we have no peer or no conn for him, so fire off - * new TX to trigger conn setup */ - read_unlock(&kgnilnd_data.kgn_peer_conn_lock); - - /* if we couldn't find him, we'll fire up a TX and get connected - - * if we don't do this, after ni_peer_timeout, LNet will declare him dead. - * So really we treat kgnilnd_query as a bit of a 'connect now' type - * event because it'll only do this when it wants to send - * - * Use a real TX for this to get the proper gnp_tx_queue behavior, etc - * normally we'd use kgnilnd_send_ctlmsg for this, but we don't really - * care that this goes out quickly since we already know we need a new conn - * formed */ - if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND)) - return; - - tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, ni->ni_nid); - if (tx != NULL) { - kgnilnd_launch_tx(tx, net, &id); - } -out: - CDEBUG(D_NETTRACE, "peer 0x%p->%s when %lu\n", peer, - libcfs_nid2str(nid), *when); - EXIT; -} - int kgnilnd_dev_init(kgn_device_t *dev) { @@ -1950,7 +1930,7 @@ kgnilnd_dev_init(kgn_device_t *dev) cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3; rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag, - GNILND_COOKIE, 0, + *kgnilnd_tunables.kgn_pkey, 0, &dev->gnd_domain); if (rrc != GNI_RC_SUCCESS) { CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc); @@ -1995,12 +1975,6 @@ kgnilnd_dev_init(kgn_device_t *dev) } } - rrc = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_IP, &kgnilnd_data.kgn_sock); - if (rrc < 0) { - CERROR("sock_create returned %d\n", rrc); - GOTO(failed, rrc); - } - rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid); if (rc < 0) { /* log messages during startup */ @@ -2012,7 +1986,7 @@ kgnilnd_dev_init(kgn_device_t *dev) } CDEBUG(D_NET, "NIC %x -> NID %d\n", dev->gnd_host_id, dev->gnd_nid); - rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size, + rrc = kgnilnd_cq_create(dev->gnd_handle, *kgnilnd_tunables.kgn_credits, 0, kgnilnd_device_callback, dev->gnd_id, &dev->gnd_snd_rdma_cqh); if (rrc != GNI_RC_SUCCESS) { @@ -2042,6 +2016,11 @@ kgnilnd_dev_init(kgn_device_t *dev) GOTO(failed, rc = -EINVAL); } + rrc = kgnilnd_register_smdd_buf(dev); + if (rrc != GNI_RC_SUCCESS) { + GOTO(failed, rc = -EINVAL); + } + RETURN(0); failed: @@ -2058,9 +2037,12 @@ kgnilnd_dev_fini(kgn_device_t *dev) /* At quiesce or rest time, need to loop through and clear gnd_ready_conns ?*/ LASSERTF(list_empty(&dev->gnd_ready_conns) && list_empty(&dev->gnd_map_tx) && - list_empty(&dev->gnd_rdmaq), - "dev 0x%p ready_conns %d@0x%p map_tx %d@0x%p rdmaq %d@0x%p\n", + list_empty(&dev->gnd_rdmaq) && + list_empty(&dev->gnd_delay_conns), + "dev 0x%p ready_conns %d@0x%p delay_conns %d@0x%p" + "map_tx %d@0x%p rdmaq %d@0x%p\n", dev, kgnilnd_count_list(&dev->gnd_ready_conns), &dev->gnd_ready_conns, + kgnilnd_count_list(&dev->gnd_delay_conns), &dev->gnd_delay_conns, kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx, kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq); @@ -2069,10 +2051,6 @@ kgnilnd_dev_fini(kgn_device_t *dev) "%d physical mappings of %d pages still mapped\n", 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", - dev->gnd_map_nvirt, dev->gnd_map_virtnob); - LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 && atomic_read(&dev->gnd_n_mdd_held) == 0 && atomic64_read(&dev->gnd_nbytes_map) == 0, @@ -2092,6 +2070,13 @@ kgnilnd_dev_fini(kgn_device_t *dev) atomic_read(&kgnilnd_data.kgn_nthreads) == 0, "tried to shutdown with threads active\n"); + if (dev->gnd_smdd_hold_buf) { + rrc = kgnilnd_deregister_smdd_buf(dev); + LASSERTF(rrc == GNI_RC_SUCCESS, + "bad rc from deregistion of sMDD buffer: %d\n", rrc); + dev->gnd_smdd_hold_buf = NULL; + } + if (dev->gnd_rcv_fma_cqh) { rrc = kgnilnd_cq_destroy(dev->gnd_rcv_fma_cqh); LASSERTF(rrc == GNI_RC_SUCCESS, @@ -2127,12 +2112,9 @@ kgnilnd_dev_fini(kgn_device_t *dev) dev->gnd_domain = NULL; } - sock_release(kgnilnd_data.kgn_sock); - EXIT; } - int kgnilnd_base_startup(void) { struct timeval tv; @@ -2141,6 +2123,15 @@ int kgnilnd_base_startup(void) int i; kgn_device_t *dev; struct task_struct *thrd; + +#if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE) + /* limit how much memory can be allocated for fma blocks in + * instances where many nodes need to reconnects at the same time */ + struct sysinfo si; + si_meminfo(&si); + kgnilnd_data.free_pages_limit = si.totalram/4; +#endif + ENTRY; LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_NOTHING, @@ -2148,6 +2139,7 @@ int kgnilnd_base_startup(void) /* zero pointers, flags etc */ memset(&kgnilnd_data, 0, sizeof(kgnilnd_data)); + kgnilnd_check_kgni_version(); /* CAVEAT EMPTOR: Every 'Fma' message includes the sender's NID and * a unique (for all time) connstamp so we can uniquely identify @@ -2167,10 +2159,11 @@ int kgnilnd_base_startup(void) dev->gnd_id = i; INIT_LIST_HEAD(&dev->gnd_ready_conns); + INIT_LIST_HEAD(&dev->gnd_delay_conns); INIT_LIST_HEAD(&dev->gnd_map_tx); INIT_LIST_HEAD(&dev->gnd_fma_buffs); mutex_init(&dev->gnd_cq_mutex); - sema_init(&dev->gnd_fmablk_sem, 1); + mutex_init(&dev->gnd_fmablk_mutex); spin_lock_init(&dev->gnd_fmablk_lock); init_waitqueue_head(&dev->gnd_waitq); init_waitqueue_head(&dev->gnd_dgram_waitq); @@ -2190,8 +2183,8 @@ int kgnilnd_base_startup(void) init_rwsem(&dev->gnd_conn_sem); /* alloc & setup nid based dgram table */ - LIBCFS_ALLOC(dev->gnd_dgrams, - sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size); + CFS_ALLOC_PTR_ARRAY(dev->gnd_dgrams, + *kgnilnd_tunables.kgn_peer_hash_size); if (dev->gnd_dgrams == NULL) GOTO(failed, rc = -ENOMEM); @@ -2218,7 +2211,7 @@ int kgnilnd_base_startup(void) init_waitqueue_head(&kgnilnd_data.kgn_ruhroh_waitq); spin_lock_init(&kgnilnd_data.kgn_reaper_lock); - sema_init(&kgnilnd_data.kgn_quiesce_sem, 1); + mutex_init(&kgnilnd_data.kgn_quiesce_mutex); atomic_set(&kgnilnd_data.kgn_nquiesce, 0); atomic_set(&kgnilnd_data.kgn_npending_conns, 0); atomic_set(&kgnilnd_data.kgn_npending_unlink, 0); @@ -2229,12 +2222,13 @@ int kgnilnd_base_startup(void) /* OK to call kgnilnd_api_shutdown() to cleanup now */ kgnilnd_data.kgn_init = GNILND_INIT_DATA; - try_module_get(THIS_MODULE); + if (!try_module_get(THIS_MODULE)) + GOTO(failed, rc = -ENOENT); rwlock_init(&kgnilnd_data.kgn_peer_conn_lock); - LIBCFS_ALLOC(kgnilnd_data.kgn_peers, - sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size); + CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_peers, + *kgnilnd_tunables.kgn_peer_hash_size); if (kgnilnd_data.kgn_peers == NULL) GOTO(failed, rc = -ENOMEM); @@ -2243,8 +2237,8 @@ int kgnilnd_base_startup(void) INIT_LIST_HEAD(&kgnilnd_data.kgn_peers[i]); } - LIBCFS_ALLOC(kgnilnd_data.kgn_conns, - sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size); + CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_conns, + *kgnilnd_tunables.kgn_peer_hash_size); if (kgnilnd_data.kgn_conns == NULL) GOTO(failed, rc = -ENOMEM); @@ -2253,8 +2247,8 @@ int kgnilnd_base_startup(void) INIT_LIST_HEAD(&kgnilnd_data.kgn_conns[i]); } - LIBCFS_ALLOC(kgnilnd_data.kgn_nets, - sizeof(struct list_head) * *kgnilnd_tunables.kgn_net_hash_size); + CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_nets, + *kgnilnd_tunables.kgn_net_hash_size); if (kgnilnd_data.kgn_nets == NULL) GOTO(failed, rc = -ENOMEM); @@ -2264,7 +2258,7 @@ int kgnilnd_base_startup(void) } kgnilnd_data.kgn_mbox_cache = - kmem_cache_create("kgn_mbox_block", KMALLOC_MAX_SIZE, 0, + kmem_cache_create("kgn_mbox_block", GNILND_MBOX_SIZE, 0, SLAB_HWCACHE_ALIGN, NULL); if (kgnilnd_data.kgn_mbox_cache == NULL) { CERROR("Can't create slab for physical mbox blocks\n"); @@ -2369,13 +2363,20 @@ int kgnilnd_base_startup(void) } /* threads will load balance across devs as they are available */ - for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) { - rc = kgnilnd_thread_start(kgnilnd_scheduler, (void *)((long)i), - "kgnilnd_sd", i); - if (rc != 0) { - CERROR("Can't spawn gnilnd scheduler[%d]: %d\n", - i, rc); + if (*kgnilnd_tunables.kgn_thread_affinity) { + rc = kgnilnd_start_sd_threads(); + if (rc != 0) GOTO(failed, rc); + } else { + for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) { + rc = kgnilnd_thread_start(kgnilnd_scheduler, + (void *)((long)i), + "kgnilnd_sd", i); + if (rc != 0) { + CERROR("Can't spawn gnilnd scheduler[%d]: %d\n", + i, rc); + GOTO(failed, rc); + } } } @@ -2406,8 +2407,6 @@ int kgnilnd_base_startup(void) } } - - /* flag everything initialised */ kgnilnd_data.kgn_init = GNILND_INIT_ALL; /*****************************************************/ @@ -2455,7 +2454,7 @@ kgnilnd_base_shutdown(void) CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting for conns to be cleaned up %d\n",atomic_read(&kgnilnd_data.kgn_nconns)); - cfs_pause(cfs_time_seconds(1)); + schedule_timeout_uninterruptible(cfs_time_seconds(1)); } /* Peer state all cleaned up BEFORE setting shutdown, so threads don't * have to worry about shutdown races. NB connections may be created @@ -2474,7 +2473,7 @@ kgnilnd_base_shutdown(void) i++; CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting for ruhroh thread to terminate\n"); - cfs_pause(cfs_time_seconds(1)); + schedule_timeout_uninterruptible(cfs_time_seconds(1)); } /* Flag threads to terminate */ @@ -2496,7 +2495,8 @@ kgnilnd_base_shutdown(void) wake_up_all(&kgnilnd_data.kgn_reaper_waitq); spin_unlock(&kgnilnd_data.kgn_reaper_lock); - kgnilnd_wakeup_rca_thread(); + if (atomic_read(&kgnilnd_data.kgn_nthreads)) + kgnilnd_wakeup_rca_thread(); /* Wait for threads to exit */ i = 2; @@ -2505,7 +2505,7 @@ kgnilnd_base_shutdown(void) CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */ "Waiting for %d threads to terminate\n", atomic_read(&kgnilnd_data.kgn_nthreads)); - cfs_pause(cfs_time_seconds(1)); + schedule_timeout_uninterruptible(cfs_time_seconds(1)); } LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0, @@ -2515,9 +2515,8 @@ kgnilnd_base_shutdown(void) for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) LASSERT(list_empty(&kgnilnd_data.kgn_peers[i])); - LIBCFS_FREE(kgnilnd_data.kgn_peers, - sizeof (struct list_head) * - *kgnilnd_tunables.kgn_peer_hash_size); + CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_peers, + *kgnilnd_tunables.kgn_peer_hash_size); } down_write(&kgnilnd_data.kgn_net_rw_sem); @@ -2525,9 +2524,8 @@ kgnilnd_base_shutdown(void) for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) LASSERT(list_empty(&kgnilnd_data.kgn_nets[i])); - LIBCFS_FREE(kgnilnd_data.kgn_nets, - sizeof (struct list_head) * - *kgnilnd_tunables.kgn_net_hash_size); + CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_nets, + *kgnilnd_tunables.kgn_net_hash_size); } up_write(&kgnilnd_data.kgn_net_rw_sem); @@ -2538,9 +2536,8 @@ kgnilnd_base_shutdown(void) for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) LASSERT(list_empty(&kgnilnd_data.kgn_conns[i])); - LIBCFS_FREE(kgnilnd_data.kgn_conns, - sizeof (struct list_head) * - *kgnilnd_tunables.kgn_peer_hash_size); + CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_conns, + *kgnilnd_tunables.kgn_peer_hash_size); } for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) { @@ -2551,12 +2548,12 @@ kgnilnd_base_shutdown(void) "dgrams left %d\n", atomic_read(&dev->gnd_ndgrams)); if (dev->gnd_dgrams != NULL) { - for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) + for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; + i++) LASSERT(list_empty(&dev->gnd_dgrams[i])); - LIBCFS_FREE(dev->gnd_dgrams, - sizeof (struct list_head) * - *kgnilnd_tunables.kgn_peer_hash_size); + CFS_FREE_PTR_ARRAY(dev->gnd_dgrams, + *kgnilnd_tunables.kgn_peer_hash_size); } kgnilnd_free_phys_fmablk(dev); @@ -2596,15 +2593,15 @@ kgnilnd_base_shutdown(void) } int -kgnilnd_startup(lnet_ni_t *ni) +kgnilnd_startup(struct lnet_ni *ni) { int rc, devno; 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(); @@ -2613,7 +2610,7 @@ kgnilnd_startup(lnet_ni_t *ni) } /* Serialize with shutdown. */ - down(&kgnilnd_data.kgn_quiesce_sem); + mutex_lock(&kgnilnd_data.kgn_quiesce_mutex); LIBCFS_ALLOC(net, sizeof(*net)); if (net == NULL) { @@ -2624,8 +2621,12 @@ 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; @@ -2635,9 +2636,10 @@ 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); @@ -2645,10 +2647,10 @@ 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); @@ -2680,16 +2682,16 @@ kgnilnd_startup(lnet_ni_t *ni) /* we need a separate thread to call probe_wait_by_id until * we get a function callback notifier from kgni */ - up(&kgnilnd_data.kgn_quiesce_sem); + mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex); RETURN(0); failed: - up(&kgnilnd_data.kgn_quiesce_sem); + mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex); kgnilnd_shutdown(ni); RETURN(rc); } void -kgnilnd_shutdown(lnet_ni_t *ni) +kgnilnd_shutdown(struct lnet_ni *ni) { kgn_net_t *net = ni->ni_data; int i; @@ -2702,7 +2704,7 @@ kgnilnd_shutdown(lnet_ni_t *ni) "init %d\n", kgnilnd_data.kgn_init); /* Serialize with startup. */ - down(&kgnilnd_data.kgn_quiesce_sem); + mutex_lock(&kgnilnd_data.kgn_quiesce_mutex); CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n", atomic_read(&libcfs_kmemory)); @@ -2751,7 +2753,7 @@ kgnilnd_shutdown(lnet_ni_t *ni) "Waiting for %d references to clear on net %d\n", atomic_read(&net->gnn_refcount), net->gnn_netnum); - cfs_pause(cfs_time_seconds(1)); + schedule_timeout_uninterruptible(cfs_time_seconds(1)); } /* release ref from kgnilnd_startup */ @@ -2785,21 +2787,18 @@ out: CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n", atomic_read(&libcfs_kmemory)); - up(&kgnilnd_data.kgn_quiesce_sem); + mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex); 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; @@ -2807,7 +2806,7 @@ kgnilnd_module_init(void) if (rc != 0) return rc; - printk(KERN_INFO "Lustre: kgnilnd build version: "KGNILND_BUILD_REV"\n"); + LCONSOLE_INFO("Lustre: kgnilnd build version: "LUSTRE_VERSION_STRING"\n"); kgnilnd_insert_sysctl(); kgnilnd_proc_init(); @@ -2818,8 +2817,9 @@ 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(LUSTRE_VERSION_STRING); MODULE_LICENSE("GPL"); -module_init(kgnilnd_module_init); -module_exit(kgnilnd_module_fini); +module_init(kgnilnd_init); +module_exit(kgnilnd_exit);