X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd.c;h=c07d4d18d983249348ab8979675df72e8d8905f0;hp=3dddc0e0d7b89e58510235d6996eb3cb3388a803;hb=fa7276592a0cc12d651a5b2bd39dd7ca45bc25bf;hpb=4de2841b30bae846092ed092d388a1b59f459b8b diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c index 3dddc0e..c07d4d18 100644 --- a/lnet/klnds/o2iblnd/o2iblnd.c +++ b/lnet/klnds/o2iblnd/o2iblnd.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -17,17 +15,15 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2016, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -38,21 +34,14 @@ * Author: Eric Barton */ +#include #include "o2iblnd.h" -lnd_t the_o2iblnd = { - .lnd_type = O2IBLND, - .lnd_startup = kiblnd_startup, - .lnd_shutdown = kiblnd_shutdown, - .lnd_ctl = kiblnd_ctl, - .lnd_query = kiblnd_query, - .lnd_send = kiblnd_send, - .lnd_recv = kiblnd_recv, -}; +static lnd_t the_o2iblnd; kib_data_t kiblnd_data; -__u32 +static __u32 kiblnd_cksum (void *ptr, int nob) { char *c = ptr; @@ -324,183 +313,184 @@ kiblnd_unpack_msg(kib_msg_t *msg, int nob) } int -kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid) +kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid) { - kib_peer_t *peer; - kib_net_t *net = ni->ni_data; - unsigned long flags; + kib_peer_t *peer; + kib_net_t *net = ni->ni_data; + int cpt = lnet_cpt_of_nid(nid, ni); + unsigned long flags; - LASSERT (net != NULL); - LASSERT (nid != LNET_NID_ANY); + LASSERT(net != NULL); + LASSERT(nid != LNET_NID_ANY); - LIBCFS_ALLOC(peer, sizeof(*peer)); + LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer)); if (peer == NULL) { CERROR("Cannot allocate peer\n"); return -ENOMEM; } - memset(peer, 0, sizeof(*peer)); /* zero flags etc */ + peer->ibp_ni = ni; + peer->ibp_nid = nid; + peer->ibp_error = 0; + peer->ibp_last_alive = 0; + peer->ibp_max_frags = kiblnd_cfg_rdma_frags(peer->ibp_ni); + peer->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits; + atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */ - peer->ibp_ni = ni; - peer->ibp_nid = nid; - peer->ibp_error = 0; - peer->ibp_last_alive = 0; - atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */ + INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */ + INIT_LIST_HEAD(&peer->ibp_conns); + INIT_LIST_HEAD(&peer->ibp_tx_queue); - INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */ - INIT_LIST_HEAD(&peer->ibp_conns); - INIT_LIST_HEAD(&peer->ibp_tx_queue); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + /* always called with a ref on ni, which prevents ni being shutdown */ + LASSERT(net->ibn_shutdown == 0); - /* always called with a ref on ni, which prevents ni being shutdown */ - LASSERT (net->ibn_shutdown == 0); + /* npeers only grows with the global lock held */ + atomic_inc(&net->ibn_npeers); - /* npeers only grows with the global lock held */ - atomic_inc(&net->ibn_npeers); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - - *peerp = peer; - return 0; + *peerp = peer; + return 0; } void kiblnd_destroy_peer (kib_peer_t *peer) { - kib_net_t *net = peer->ibp_ni->ni_data; + kib_net_t *net = peer->ibp_ni->ni_data; - LASSERT (net != NULL); - LASSERT (atomic_read(&peer->ibp_refcount) == 0); - LASSERT (!kiblnd_peer_active(peer)); - LASSERT (peer->ibp_connecting == 0); - LASSERT (peer->ibp_accepting == 0); - LASSERT (list_empty(&peer->ibp_conns)); - LASSERT (list_empty(&peer->ibp_tx_queue)); + LASSERT(net != NULL); + LASSERT (atomic_read(&peer->ibp_refcount) == 0); + LASSERT(!kiblnd_peer_active(peer)); + LASSERT(kiblnd_peer_idle(peer)); + LASSERT(list_empty(&peer->ibp_tx_queue)); - LIBCFS_FREE(peer, sizeof(*peer)); + LIBCFS_FREE(peer, sizeof(*peer)); - /* NB a peer's connections keep a reference on their peer until - * they are destroyed, so we can be assured that _all_ state to do - * with this peer has been cleaned up when its refcount drops to - * zero. */ - atomic_dec(&net->ibn_npeers); + /* NB a peer's connections keep a reference on their peer until + * they are destroyed, so we can be assured that _all_ state to do + * with this peer has been cleaned up when its refcount drops to + * zero. */ + atomic_dec(&net->ibn_npeers); } kib_peer_t * -kiblnd_find_peer_locked (lnet_nid_t nid) +kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid) { - /* the caller is responsible for accounting the additional reference - * that this creates */ - struct list_head *peer_list = kiblnd_nid2peerlist(nid); - struct list_head *tmp; - kib_peer_t *peer; - - list_for_each (tmp, peer_list) { - - peer = list_entry(tmp, kib_peer_t, ibp_list); - - LASSERT (peer->ibp_connecting > 0 || /* creating conns */ - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); /* active conn */ - - if (peer->ibp_nid != nid) - continue; - - CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n", - peer, libcfs_nid2str(nid), - atomic_read(&peer->ibp_refcount), - peer->ibp_version); - return peer; - } - return NULL; + /* the caller is responsible for accounting the additional reference + * that this creates */ + struct list_head *peer_list = kiblnd_nid2peerlist(nid); + struct list_head *tmp; + kib_peer_t *peer; + + list_for_each(tmp, peer_list) { + + peer = list_entry(tmp, kib_peer_t, ibp_list); + LASSERT(!kiblnd_peer_idle(peer)); + + /* + * Match a peer if its NID and the NID of the local NI it + * communicates over are the same. Otherwise don't match + * the peer, which will result in a new lnd peer being + * created. + */ + if (peer->ibp_nid != nid || + peer->ibp_ni->ni_nid != ni->ni_nid) + continue; + + CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n", + peer, libcfs_nid2str(nid), + atomic_read(&peer->ibp_refcount), + peer->ibp_version); + return peer; + } + return NULL; } void kiblnd_unlink_peer_locked (kib_peer_t *peer) { - LASSERT (list_empty(&peer->ibp_conns)); + LASSERT(list_empty(&peer->ibp_conns)); LASSERT (kiblnd_peer_active(peer)); - list_del_init(&peer->ibp_list); + list_del_init(&peer->ibp_list); /* lose peerlist's ref */ kiblnd_peer_decref(peer); } -int -kiblnd_get_peer_info (lnet_ni_t *ni, int index, - lnet_nid_t *nidp, int *count) +static int +kiblnd_get_peer_info(lnet_ni_t *ni, int index, + lnet_nid_t *nidp, int *count) { - kib_peer_t *peer; - struct list_head *ptmp; - int i; - unsigned long flags; + kib_peer_t *peer; + struct list_head *ptmp; + int i; + unsigned long flags; - read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) { - list_for_each (ptmp, &kiblnd_data.kib_peers[i]) { + list_for_each(ptmp, &kiblnd_data.kib_peers[i]) { - peer = list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); + peer = list_entry(ptmp, kib_peer_t, ibp_list); + LASSERT(!kiblnd_peer_idle(peer)); - if (peer->ibp_ni != ni) - continue; + if (peer->ibp_ni != ni) + continue; - if (index-- > 0) - continue; + if (index-- > 0) + continue; - *nidp = peer->ibp_nid; - *count = atomic_read(&peer->ibp_refcount); + *nidp = peer->ibp_nid; + *count = atomic_read(&peer->ibp_refcount); - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, - flags); - return 0; - } - } + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, + flags); + return 0; + } + } - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - return -ENOENT; + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + return -ENOENT; } -void +static void kiblnd_del_peer_locked (kib_peer_t *peer) { - struct list_head *ctmp; - struct list_head *cnxt; - kib_conn_t *conn; - - if (list_empty(&peer->ibp_conns)) { - kiblnd_unlink_peer_locked(peer); - } else { - list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { - conn = list_entry(ctmp, kib_conn_t, ibc_list); - - kiblnd_close_conn_locked(conn, 0); - } - /* NB closing peer's last conn unlinked it. */ - } - /* NB peer now unlinked; might even be freed if the peer table had the - * last ref on it. */ + struct list_head *ctmp; + struct list_head *cnxt; + kib_conn_t *conn; + + if (list_empty(&peer->ibp_conns)) { + kiblnd_unlink_peer_locked(peer); + } else { + list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry(ctmp, kib_conn_t, ibc_list); + + kiblnd_close_conn_locked(conn, 0); + } + /* NB closing peer's last conn unlinked it. */ + } + /* NB peer now unlinked; might even be freed if the peer table had the + * last ref on it. */ } -int +static int kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid) { - CFS_LIST_HEAD (zombies); - struct list_head *ptmp; - struct list_head *pnxt; - kib_peer_t *peer; - int lo; - int hi; - int i; - unsigned long flags; - int rc = -ENOENT; - - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + struct list_head zombies = LIST_HEAD_INIT(zombies); + struct list_head *ptmp; + struct list_head *pnxt; + kib_peer_t *peer; + int lo; + int hi; + int i; + unsigned long flags; + int rc = -ENOENT; + + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); if (nid != LNET_NID_ANY) { lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers; @@ -509,78 +499,75 @@ kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid) hi = kiblnd_data.kib_peer_hash_size - 1; } - for (i = lo; i <= hi; i++) { - list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) { - peer = list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); + for (i = lo; i <= hi; i++) { + list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) { + peer = list_entry(ptmp, kib_peer_t, ibp_list); + LASSERT(!kiblnd_peer_idle(peer)); - if (peer->ibp_ni != ni) - continue; + if (peer->ibp_ni != ni) + continue; - if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid)) - continue; + if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid)) + continue; - if (!list_empty(&peer->ibp_tx_queue)) { - LASSERT (list_empty(&peer->ibp_conns)); + if (!list_empty(&peer->ibp_tx_queue)) { + LASSERT(list_empty(&peer->ibp_conns)); - list_splice_init(&peer->ibp_tx_queue, &zombies); - } + list_splice_init(&peer->ibp_tx_queue, + &zombies); + } - kiblnd_del_peer_locked(peer); - rc = 0; /* matched something */ - } - } + kiblnd_del_peer_locked(peer); + rc = 0; /* matched something */ + } + } - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - kiblnd_txlist_done(ni, &zombies, -EIO); + kiblnd_txlist_done(ni, &zombies, -EIO); - return rc; + return rc; } -kib_conn_t * -kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index) +static kib_conn_t * +kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index) { - kib_peer_t *peer; - struct list_head *ptmp; - kib_conn_t *conn; - struct list_head *ctmp; - int i; - unsigned long flags; - - read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - - for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) { - list_for_each (ptmp, &kiblnd_data.kib_peers[i]) { - - peer = list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); - - if (peer->ibp_ni != ni) - continue; - - list_for_each (ctmp, &peer->ibp_conns) { - if (index-- > 0) - continue; - - conn = list_entry(ctmp, kib_conn_t, ibc_list); - kiblnd_conn_addref(conn); - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, - flags); - return conn; - } - } - } - - read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + kib_peer_t *peer; + struct list_head *ptmp; + kib_conn_t *conn; + struct list_head *ctmp; + int i; + unsigned long flags; + + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + + for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) { + list_for_each(ptmp, &kiblnd_data.kib_peers[i]) { + + peer = list_entry(ptmp, kib_peer_t, ibp_list); + LASSERT(!kiblnd_peer_idle(peer)); + + if (peer->ibp_ni != ni) + continue; + + list_for_each(ctmp, &peer->ibp_conns) { + if (index-- > 0) + continue; + + conn = list_entry(ctmp, kib_conn_t, ibc_list); + kiblnd_conn_addref(conn); + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, + flags); + return conn; + } + } + } + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); return NULL; } -void +static void kiblnd_debug_rx (kib_rx_t *rx) { CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n", @@ -588,11 +575,11 @@ kiblnd_debug_rx (kib_rx_t *rx) rx->rx_msg->ibm_credits); } -void +static void kiblnd_debug_tx (kib_tx_t *tx) { CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx " - "cookie "LPX64" msg %s%s type %x cred %d\n", + "cookie %#llx msg %s%s type %x cred %d\n", tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting, tx->tx_status, tx->tx_deadline, tx->tx_cookie, tx->tx_lntmsg[0] == NULL ? "-" : "!", @@ -603,45 +590,49 @@ kiblnd_debug_tx (kib_tx_t *tx) void kiblnd_debug_conn (kib_conn_t *conn) { - struct list_head *tmp; - int i; + struct list_head *tmp; + int i; - spin_lock(&conn->ibc_lock); + spin_lock(&conn->ibc_lock); - CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n", - atomic_read(&conn->ibc_refcount), conn, - conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid)); - CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d r_cred %d\n", - conn->ibc_state, conn->ibc_noops_posted, - conn->ibc_nsends_posted, conn->ibc_credits, - conn->ibc_outstanding_credits, conn->ibc_reserved_credits); - CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error); + CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n", + atomic_read(&conn->ibc_refcount), conn, + conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid)); + CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d " + " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted, + conn->ibc_nsends_posted, conn->ibc_credits, + conn->ibc_outstanding_credits, conn->ibc_reserved_credits); + CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error); - CDEBUG(D_CONSOLE, " early_rxs:\n"); - list_for_each(tmp, &conn->ibc_early_rxs) - kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list)); + CDEBUG(D_CONSOLE, " early_rxs:\n"); + list_for_each(tmp, &conn->ibc_early_rxs) + kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list)); - CDEBUG(D_CONSOLE, " tx_queue_nocred:\n"); - list_for_each(tmp, &conn->ibc_tx_queue_nocred) - kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " tx_noops:\n"); + list_for_each(tmp, &conn->ibc_tx_noops) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n"); - list_for_each(tmp, &conn->ibc_tx_queue_rsrvd) - kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " tx_queue_nocred:\n"); + list_for_each(tmp, &conn->ibc_tx_queue_nocred) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " tx_queue:\n"); - list_for_each(tmp, &conn->ibc_tx_queue) - kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n"); + list_for_each(tmp, &conn->ibc_tx_queue_rsrvd) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " active_txs:\n"); - list_for_each(tmp, &conn->ibc_active_txs) - kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); + CDEBUG(D_CONSOLE, " tx_queue:\n"); + list_for_each(tmp, &conn->ibc_tx_queue) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - CDEBUG(D_CONSOLE, " rxs:\n"); - for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) - kiblnd_debug_rx(&conn->ibc_rxs[i]); + CDEBUG(D_CONSOLE, " active_txs:\n"); + list_for_each(tmp, &conn->ibc_active_txs) + kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list)); - spin_unlock(&conn->ibc_lock); + CDEBUG(D_CONSOLE, " rxs:\n"); + for (i = 0; i < IBLND_RX_MSGS(conn); i++) + kiblnd_debug_rx(&conn->ibc_rxs[i]); + + spin_unlock(&conn->ibc_lock); } int @@ -666,127 +657,193 @@ kiblnd_translate_mtu(int value) } static void -kiblnd_setup_mtu(struct rdma_cm_id *cmid) +kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid) { - unsigned long flags; int mtu; /* XXX There is no path record for iWARP, set by netdev->change_mtu? */ if (cmid->route.path_rec == NULL) return; - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu); LASSERT (mtu >= 0); if (mtu != 0) cmid->route.path_rec->mtu = mtu; +} - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); +static int +kiblnd_get_completion_vector(kib_conn_t *conn, int cpt) +{ + cpumask_t *mask; + int vectors; + int off; + int i; + lnet_nid_t ibp_nid; + + vectors = conn->ibc_cmid->device->num_comp_vectors; + if (vectors <= 1) + return 0; + + mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt); + + /* hash NID to CPU id in this partition... */ + ibp_nid = conn->ibc_peer->ibp_nid; + off = do_div(ibp_nid, cpumask_weight(mask)); + for_each_cpu(i, mask) { + if (off-- == 0) + return i % vectors; + } + + LBUG(); + return 1; } kib_conn_t * kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, - int state, int version) -{ - /* CAVEAT EMPTOR: - * If the new conn is created successfully it takes over the caller's - * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself - * is destroyed. On failure, the caller's ref on 'peer' remains and - * she must dispose of 'cmid'. (Actually I'd block forever if I tried - * to destroy 'cmid' here since I'm called from the CM which still has - * its ref on 'cmid'). */ - kib_net_t *net = peer->ibp_ni->ni_data; - struct ib_qp_init_attr *init_qp_attr; - kib_conn_t *conn; - struct ib_cq *cq; - unsigned long flags; - int rc; - int i; - - LASSERT (net != NULL); - LASSERT (!in_interrupt()); - - LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr)); - if (init_qp_attr == NULL) { - CERROR("Can't allocate qp_attr for %s\n", - libcfs_nid2str(peer->ibp_nid)); - goto failed_0; - } - - LIBCFS_ALLOC(conn, sizeof(*conn)); - if (conn == NULL) { - CERROR("Can't allocate connection for %s\n", - libcfs_nid2str(peer->ibp_nid)); - goto failed_1; - } - - memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */ - - conn->ibc_state = IBLND_CONN_INIT; - conn->ibc_version = version; - conn->ibc_peer = peer; /* I take the caller's ref */ - cmid->context = conn; /* for future CM callbacks */ - conn->ibc_cmid = cmid; - - INIT_LIST_HEAD(&conn->ibc_early_rxs); - INIT_LIST_HEAD(&conn->ibc_tx_queue); - INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd); - INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred); - INIT_LIST_HEAD(&conn->ibc_active_txs); - spin_lock_init(&conn->ibc_lock); - - LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); - if (conn->ibc_connvars == NULL) { - CERROR("Can't allocate in-progress connection state\n"); - goto failed_2; - } - memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars)); - - LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t)); - if (conn->ibc_rxs == NULL) { - CERROR("Cannot allocate RX buffers\n"); - goto failed_2; - } - memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t)); - - rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, - IBLND_RX_MSG_PAGES(version)); - if (rc != 0) - goto failed_2; - - kiblnd_map_rx_descs(conn); - -#ifdef HAVE_OFED_IB_COMP_VECTOR - cq = ib_create_cq(cmid->device, - kiblnd_cq_completion, kiblnd_cq_event, conn, - IBLND_CQ_ENTRIES(version), 0); + int state, int version) +{ + /* CAVEAT EMPTOR: + * If the new conn is created successfully it takes over the caller's + * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself + * is destroyed. On failure, the caller's ref on 'peer' remains and + * she must dispose of 'cmid'. (Actually I'd block forever if I tried + * to destroy 'cmid' here since I'm called from the CM which still has + * its ref on 'cmid'). */ + rwlock_t *glock = &kiblnd_data.kib_global_lock; + kib_net_t *net = peer->ibp_ni->ni_data; + kib_dev_t *dev; + struct ib_qp_init_attr *init_qp_attr; + struct kib_sched_info *sched; +#ifdef HAVE_IB_CQ_INIT_ATTR + struct ib_cq_init_attr cq_attr = {}; +#endif + kib_conn_t *conn; + struct ib_cq *cq; + unsigned long flags; + int cpt; + int rc; + int i; + + LASSERT(net != NULL); + LASSERT(!in_interrupt()); + + dev = net->ibn_dev; + + cpt = lnet_cpt_of_nid(peer->ibp_nid, peer->ibp_ni); + sched = kiblnd_data.kib_scheds[cpt]; + + LASSERT(sched->ibs_nthreads > 0); + + LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt, + sizeof(*init_qp_attr)); + if (init_qp_attr == NULL) { + CERROR("Can't allocate qp_attr for %s\n", + libcfs_nid2str(peer->ibp_nid)); + goto failed_0; + } + + LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn)); + if (conn == NULL) { + CERROR("Can't allocate connection for %s\n", + libcfs_nid2str(peer->ibp_nid)); + goto failed_1; + } + + conn->ibc_state = IBLND_CONN_INIT; + conn->ibc_version = version; + conn->ibc_peer = peer; /* I take the caller's ref */ + cmid->context = conn; /* for future CM callbacks */ + conn->ibc_cmid = cmid; + conn->ibc_max_frags = peer->ibp_max_frags; + conn->ibc_queue_depth = peer->ibp_queue_depth; + + INIT_LIST_HEAD(&conn->ibc_early_rxs); + INIT_LIST_HEAD(&conn->ibc_tx_noops); + INIT_LIST_HEAD(&conn->ibc_tx_queue); + INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd); + INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred); + INIT_LIST_HEAD(&conn->ibc_active_txs); + spin_lock_init(&conn->ibc_lock); + + LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt, + sizeof(*conn->ibc_connvars)); + if (conn->ibc_connvars == NULL) { + CERROR("Can't allocate in-progress connection state\n"); + goto failed_2; + } + + write_lock_irqsave(glock, flags); + if (dev->ibd_failover) { + write_unlock_irqrestore(glock, flags); + CERROR("%s: failover in progress\n", dev->ibd_ifname); + goto failed_2; + } + + if (dev->ibd_hdev->ibh_ibdev != cmid->device) { + /* wakeup failover thread and teardown connection */ + if (kiblnd_dev_can_failover(dev)) { + list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + wake_up(&kiblnd_data.kib_failover_waitq); + } + + write_unlock_irqrestore(glock, flags); + CERROR("cmid HCA(%s), kib_dev(%s) need failover\n", + cmid->device->name, dev->ibd_ifname); + goto failed_2; + } + + kiblnd_hdev_addref_locked(dev->ibd_hdev); + conn->ibc_hdev = dev->ibd_hdev; + + kiblnd_setup_mtu_locked(cmid); + + write_unlock_irqrestore(glock, flags); + + LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt, + IBLND_RX_MSGS(conn) * sizeof(kib_rx_t)); + if (conn->ibc_rxs == NULL) { + CERROR("Cannot allocate RX buffers\n"); + goto failed_2; + } + + rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt, + IBLND_RX_MSG_PAGES(conn)); + if (rc != 0) + goto failed_2; + + kiblnd_map_rx_descs(conn); + +#ifdef HAVE_IB_CQ_INIT_ATTR + cq_attr.cqe = IBLND_CQ_ENTRIES(conn); + cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt); + cq = ib_create_cq(cmid->device, + kiblnd_cq_completion, kiblnd_cq_event, conn, + &cq_attr); #else - cq = ib_create_cq(cmid->device, - kiblnd_cq_completion, kiblnd_cq_event, conn, - IBLND_CQ_ENTRIES(version)); + cq = ib_create_cq(cmid->device, + kiblnd_cq_completion, kiblnd_cq_event, conn, + IBLND_CQ_ENTRIES(conn), + kiblnd_get_completion_vector(conn, cpt)); #endif - if (IS_ERR(cq)) { - CERROR("Can't create CQ: %ld, cqe: %d\n", - PTR_ERR(cq), IBLND_CQ_ENTRIES(version)); - goto failed_2; - } + if (IS_ERR(cq)) { + CERROR("Failed to create CQ with %d CQEs: %ld\n", + IBLND_CQ_ENTRIES(conn), PTR_ERR(cq)); + goto failed_2; + } conn->ibc_cq = cq; - rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP); - if (rc != 0) { - CERROR("Can't request completion notificiation: %d\n", rc); - goto failed_2; - } - - kiblnd_setup_mtu(cmid); + rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP); + if (rc != 0) { + CERROR("Can't request completion notification: %d\n", rc); + goto failed_2; + } - memset(init_qp_attr, 0, sizeof(*init_qp_attr)); init_qp_attr->event_handler = kiblnd_qp_event; init_qp_attr->qp_context = conn; - init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version); - init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version); + init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn); + init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn); init_qp_attr->cap.max_send_sge = 1; init_qp_attr->cap.max_recv_sge = 1; init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR; @@ -794,36 +851,47 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, init_qp_attr->send_cq = cq; init_qp_attr->recv_cq = cq; - rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr); - if (rc != 0) { - CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n", - rc, init_qp_attr->cap.max_send_wr, - init_qp_attr->cap.max_recv_wr); - goto failed_2; - } + conn->ibc_sched = sched; - LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); + do { + rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr); + if (!rc || init_qp_attr->cap.max_send_wr < 16) + break; + + init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4; + } while (rc); - /* 1 ref for caller and each rxmsg */ - atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version)); - conn->ibc_nrx = IBLND_RX_MSGS(version); + if (rc) { + CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n", + rc, init_qp_attr->cap.max_send_wr, + init_qp_attr->cap.max_recv_wr); + goto failed_2; + } - /* post receives */ - for (i = 0; i < IBLND_RX_MSGS(version); i++) { - rc = kiblnd_post_rx(&conn->ibc_rxs[i], - IBLND_POSTRX_NO_CREDIT); - if (rc != 0) { - CERROR("Can't post rxmsg: %d\n", rc); + if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn)) + CDEBUG(D_NET, "original send wr %d, created with %d\n", + IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr); - /* Make posted receives complete */ - kiblnd_abort_receives(conn); + LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); - /* correct # of posted buffers - * NB locking needed now I'm racing with completion */ - spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags); - conn->ibc_nrx -= IBLND_RX_MSGS(version) - i; - spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, - flags); + /* 1 ref for caller and each rxmsg */ + atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn)); + conn->ibc_nrx = IBLND_RX_MSGS(conn); + + /* post receives */ + for (i = 0; i < IBLND_RX_MSGS(conn); i++) { + rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT); + if (rc != 0) { + CERROR("Can't post rxmsg: %d\n", rc); + + /* Make posted receives complete */ + kiblnd_abort_receives(conn); + + /* correct # of posted buffers + * NB locking needed now I'm racing with completion */ + spin_lock_irqsave(&sched->ibs_lock, flags); + conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i; + spin_unlock_irqrestore(&sched->ibs_lock, flags); /* cmid will be destroyed by CM(ofed) after cm_callback * returned, so we can't refer it anymore @@ -831,9 +899,9 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, rdma_destroy_qp(conn->ibc_cmid); conn->ibc_cmid = NULL; - /* Drop my own and unused rxbuffer refcounts */ - while (i++ <= IBLND_RX_MSGS(version)) - kiblnd_conn_decref(conn); + /* Drop my own and unused rxbuffer refcounts */ + while (i++ <= IBLND_RX_MSGS(conn)) + kiblnd_conn_decref(conn); return NULL; } @@ -845,11 +913,11 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, conn->ibc_state = state; /* 1 more conn */ - atomic_inc(&net->ibn_nconns); + atomic_inc(&net->ibn_nconns); return conn; failed_2: - kiblnd_destroy_conn(conn); + kiblnd_destroy_conn(conn, true); failed_1: LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr)); failed_0: @@ -857,170 +925,173 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, } void -kiblnd_destroy_conn (kib_conn_t *conn) +kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn) { - struct rdma_cm_id *cmid = conn->ibc_cmid; - kib_peer_t *peer = conn->ibc_peer; - int rc; - - LASSERT (!in_interrupt()); - LASSERT (atomic_read(&conn->ibc_refcount) == 0); - LASSERT (list_empty(&conn->ibc_early_rxs)); - LASSERT (list_empty(&conn->ibc_tx_queue)); - LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd)); - LASSERT (list_empty(&conn->ibc_tx_queue_nocred)); - LASSERT (list_empty(&conn->ibc_active_txs)); - LASSERT (conn->ibc_noops_posted == 0); - LASSERT (conn->ibc_nsends_posted == 0); - - switch (conn->ibc_state) { - default: - /* conn must be completely disengaged from the network */ - LBUG(); - - case IBLND_CONN_DISCONNECTED: - /* connvars should have been freed already */ - LASSERT (conn->ibc_connvars == NULL); - break; - - case IBLND_CONN_INIT: - break; - } - - /* conn->ibc_cmid might be destroyed by CM already */ - if (cmid != NULL && cmid->qp != NULL) - rdma_destroy_qp(cmid); - - if (conn->ibc_cq != NULL) { - rc = ib_destroy_cq(conn->ibc_cq); - if (rc != 0) - CWARN("Error destroying CQ: %d\n", rc); - } - - if (conn->ibc_rx_pages != NULL) - kiblnd_unmap_rx_descs(conn); - - if (conn->ibc_rxs != NULL) { - LIBCFS_FREE(conn->ibc_rxs, - IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t)); - } - - if (conn->ibc_connvars != NULL) - LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); - - /* See CAVEAT EMPTOR above in kiblnd_create_conn */ - if (conn->ibc_state != IBLND_CONN_INIT) { - kib_net_t *net = peer->ibp_ni->ni_data; - - kiblnd_peer_decref(peer); - rdma_destroy_id(cmid); - atomic_dec(&net->ibn_nconns); - } - - LIBCFS_FREE(conn, sizeof(*conn)); + struct rdma_cm_id *cmid = conn->ibc_cmid; + kib_peer_t *peer = conn->ibc_peer; + int rc; + + LASSERT (!in_interrupt()); + LASSERT (atomic_read(&conn->ibc_refcount) == 0); + LASSERT(list_empty(&conn->ibc_early_rxs)); + LASSERT(list_empty(&conn->ibc_tx_noops)); + LASSERT(list_empty(&conn->ibc_tx_queue)); + LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd)); + LASSERT(list_empty(&conn->ibc_tx_queue_nocred)); + LASSERT(list_empty(&conn->ibc_active_txs)); + LASSERT (conn->ibc_noops_posted == 0); + LASSERT (conn->ibc_nsends_posted == 0); + + switch (conn->ibc_state) { + default: + /* conn must be completely disengaged from the network */ + LBUG(); + + case IBLND_CONN_DISCONNECTED: + /* connvars should have been freed already */ + LASSERT (conn->ibc_connvars == NULL); + break; + + case IBLND_CONN_INIT: + break; + } + + /* conn->ibc_cmid might be destroyed by CM already */ + if (cmid != NULL && cmid->qp != NULL) + rdma_destroy_qp(cmid); + + if (conn->ibc_cq != NULL) { + rc = ib_destroy_cq(conn->ibc_cq); + if (rc != 0) + CWARN("Error destroying CQ: %d\n", rc); + } + + if (conn->ibc_rx_pages != NULL) + kiblnd_unmap_rx_descs(conn); + + if (conn->ibc_rxs != NULL) { + LIBCFS_FREE(conn->ibc_rxs, + IBLND_RX_MSGS(conn) * sizeof(kib_rx_t)); + } + + if (conn->ibc_connvars != NULL) + LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); + + if (conn->ibc_hdev != NULL) + kiblnd_hdev_decref(conn->ibc_hdev); + + /* See CAVEAT EMPTOR above in kiblnd_create_conn */ + if (conn->ibc_state != IBLND_CONN_INIT) { + kib_net_t *net = peer->ibp_ni->ni_data; + + kiblnd_peer_decref(peer); + rdma_destroy_id(cmid); + atomic_dec(&net->ibn_nconns); + } + + if (free_conn) + LIBCFS_FREE(conn, sizeof(*conn)); } int -kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why) +kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why) { - kib_conn_t *conn; - struct list_head *ctmp; - struct list_head *cnxt; - int count = 0; + kib_conn_t *conn; + struct list_head *ctmp; + struct list_head *cnxt; + int count = 0; - list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { - conn = list_entry(ctmp, kib_conn_t, ibc_list); + list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry(ctmp, kib_conn_t, ibc_list); - CDEBUG(D_NET, "Closing conn -> %s, " - "version: %x, reason: %d\n", - libcfs_nid2str(peer->ibp_nid), - conn->ibc_version, why); + CDEBUG(D_NET, "Closing conn -> %s, " + "version: %x, reason: %d\n", + libcfs_nid2str(peer->ibp_nid), + conn->ibc_version, why); - kiblnd_close_conn_locked(conn, why); - count++; - } + kiblnd_close_conn_locked(conn, why); + count++; + } - return count; + return count; } int -kiblnd_close_stale_conns_locked (kib_peer_t *peer, - int version, __u64 incarnation) +kiblnd_close_stale_conns_locked(kib_peer_t *peer, + int version, __u64 incarnation) { - kib_conn_t *conn; - struct list_head *ctmp; - struct list_head *cnxt; - int count = 0; + kib_conn_t *conn; + struct list_head *ctmp; + struct list_head *cnxt; + int count = 0; - list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) { - conn = list_entry(ctmp, kib_conn_t, ibc_list); + list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) { + conn = list_entry(ctmp, kib_conn_t, ibc_list); - if (conn->ibc_version == version && - conn->ibc_incarnation == incarnation) - continue; + if (conn->ibc_version == version && + conn->ibc_incarnation == incarnation) + continue; - CDEBUG(D_NET, "Closing stale conn -> %s version: %x, " - "incarnation:"LPX64"(%x, "LPX64")\n", - libcfs_nid2str(peer->ibp_nid), - conn->ibc_version, conn->ibc_incarnation, - version, incarnation); + CDEBUG(D_NET, "Closing stale conn -> %s version: %x, " + "incarnation:%#llx(%x, %#llx)\n", + libcfs_nid2str(peer->ibp_nid), + conn->ibc_version, conn->ibc_incarnation, + version, incarnation); - kiblnd_close_conn_locked(conn, -ESTALE); - count++; - } + kiblnd_close_conn_locked(conn, -ESTALE); + count++; + } - return count; + return count; } -int -kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid) -{ - kib_peer_t *peer; - struct list_head *ptmp; - struct list_head *pnxt; - int lo; - int hi; - int i; - unsigned long flags; - int count = 0; +static int +kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid) +{ + kib_peer_t *peer; + struct list_head *ptmp; + struct list_head *pnxt; + int lo; + int hi; + int i; + unsigned long flags; + int count = 0; - write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - if (nid != LNET_NID_ANY) - lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers; - else { - lo = 0; - hi = kiblnd_data.kib_peer_hash_size - 1; - } + if (nid != LNET_NID_ANY) + lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers; + else { + lo = 0; + hi = kiblnd_data.kib_peer_hash_size - 1; + } - for (i = lo; i <= hi; i++) { - list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) { + for (i = lo; i <= hi; i++) { + list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) { - peer = list_entry(ptmp, kib_peer_t, ibp_list); - LASSERT (peer->ibp_connecting > 0 || - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); + peer = list_entry(ptmp, kib_peer_t, ibp_list); + LASSERT(!kiblnd_peer_idle(peer)); - if (peer->ibp_ni != ni) - continue; + if (peer->ibp_ni != ni) + continue; - if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid)) - continue; + if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid)) + continue; - count += kiblnd_close_peer_conns_locked(peer, 0); - } - } + count += kiblnd_close_peer_conns_locked(peer, 0); + } + } - write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - /* wildcards always succeed */ - if (nid == LNET_NID_ANY) - return 0; + /* wildcards always succeed */ + if (nid == LNET_NID_ANY) + return 0; - return (count == 0) ? -ENOENT : 0; + return (count == 0) ? -ENOENT : 0; } -int +static int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) { struct libcfs_ioctl_data *data = arg; @@ -1074,59 +1145,59 @@ kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg) return rc; } -void -kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, time_t *when) +static void +kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when) { - cfs_time_t last_alive = 0; - rwlock_t *glock = &kiblnd_data.kib_global_lock; - kib_peer_t *peer; - unsigned long flags; + cfs_time_t last_alive = 0; + cfs_time_t now = cfs_time_current(); + rwlock_t *glock = &kiblnd_data.kib_global_lock; + kib_peer_t *peer; + unsigned long flags; - read_lock_irqsave(glock, flags); + read_lock_irqsave(glock, flags); - peer = kiblnd_find_peer_locked(nid); - if (peer != NULL) { - LASSERT (peer->ibp_connecting > 0 || /* creating conns */ - peer->ibp_accepting > 0 || - !list_empty(&peer->ibp_conns)); /* active conn */ - last_alive = peer->ibp_last_alive; - } + peer = kiblnd_find_peer_locked(ni, nid); + if (peer != NULL) + last_alive = peer->ibp_last_alive; - read_unlock_irqrestore(glock, flags); + read_unlock_irqrestore(glock, flags); - if (last_alive != 0) - *when = cfs_time_current_sec() - - cfs_duration_sec(cfs_time_current() - last_alive); + if (last_alive != 0) + *when = last_alive; - /* peer is not persistent in hash, trigger peer creation - * and connection establishment with a NULL tx */ - if (peer == NULL) - kiblnd_launch_tx(ni, NULL, nid); - return; + /* peer is not persistent in hash, trigger peer creation + * and connection establishment with a NULL tx */ + if (peer == NULL) + kiblnd_launch_tx(ni, NULL, nid); + + CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n", + libcfs_nid2str(nid), peer, + last_alive ? cfs_duration_sec(now - last_alive) : -1); + return; } -void -kiblnd_free_pages (kib_pages_t *p) +static void +kiblnd_free_pages(kib_pages_t *p) { - int npages = p->ibp_npages; - int i; + int npages = p->ibp_npages; + int i; - LASSERT (p->ibp_device == NULL); + for (i = 0; i < npages; i++) { + if (p->ibp_pages[i] != NULL) + __free_page(p->ibp_pages[i]); + } - for (i = 0; i < npages; i++) - if (p->ibp_pages[i] != NULL) - __free_page(p->ibp_pages[i]); - - LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages])); + LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages])); } int -kiblnd_alloc_pages (kib_pages_t **pp, int npages) +kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages) { - kib_pages_t *p; - int i; + kib_pages_t *p; + int i; - LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages])); + LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt, + offsetof(kib_pages_t, ibp_pages[npages])); if (p == NULL) { CERROR("Can't allocate descriptor for %d pages\n", npages); return -ENOMEM; @@ -1136,7 +1207,8 @@ kiblnd_alloc_pages (kib_pages_t **pp, int npages) p->ibp_npages = npages; for (i = 0; i < npages; i++) { - p->ibp_pages[i] = alloc_page(GFP_KERNEL); + p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, + GFP_NOFS); if (p->ibp_pages[i] == NULL) { CERROR("Can't allocate page %d of %d\n", i, npages); kiblnd_free_pages(p); @@ -1155,20 +1227,18 @@ kiblnd_unmap_rx_descs(kib_conn_t *conn) int i; LASSERT (conn->ibc_rxs != NULL); - LASSERT (conn->ibc_rx_pages->ibp_device != NULL); + LASSERT (conn->ibc_hdev != NULL); - for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) { - rx = &conn->ibc_rxs[i]; + for (i = 0; i < IBLND_RX_MSGS(conn); i++) { + rx = &conn->ibc_rxs[i]; - LASSERT (rx->rx_nob >= 0); /* not posted */ - - kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device, - KIBLND_UNMAP_ADDR(rx, rx_msgunmap, - rx->rx_msgaddr), - IBLND_MSG_SIZE, DMA_FROM_DEVICE); - } + LASSERT(rx->rx_nob >= 0); /* not posted */ - conn->ibc_rx_pages->ibp_device = NULL; + kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev, + KIBLND_UNMAP_ADDR(rx, rx_msgunmap, + rx->rx_msgaddr), + IBLND_MSG_SIZE, DMA_FROM_DEVICE); + } kiblnd_free_pages(conn->ibc_rx_pages); @@ -1184,64 +1254,94 @@ kiblnd_map_rx_descs(kib_conn_t *conn) int ipg; int i; - for (pg_off = ipg = i = 0; - i < IBLND_RX_MSGS(conn->ibc_version); i++) { - pg = conn->ibc_rx_pages->ibp_pages[ipg]; - rx = &conn->ibc_rxs[i]; - - rx->rx_conn = conn; - rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off); + for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) { + pg = conn->ibc_rx_pages->ibp_pages[ipg]; + rx = &conn->ibc_rxs[i]; + + rx->rx_conn = conn; + rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off); + + rx->rx_msgaddr = + kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev, + rx->rx_msg, IBLND_MSG_SIZE, + DMA_FROM_DEVICE); + LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev, + rx->rx_msgaddr)); + KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr); + + CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n", + i, rx->rx_msg, rx->rx_msgaddr, + (__u64)(page_to_phys(pg) + pg_off)); + + pg_off += IBLND_MSG_SIZE; + LASSERT(pg_off <= PAGE_SIZE); + + if (pg_off == PAGE_SIZE) { + pg_off = 0; + ipg++; + LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn)); + } + } +} - rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device, - rx->rx_msg, IBLND_MSG_SIZE, - DMA_FROM_DEVICE); - LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device, - rx->rx_msgaddr)); - KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr); +static void +kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo) +{ + kib_hca_dev_t *hdev = tpo->tpo_hdev; + kib_tx_t *tx; + int i; - CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n", - i, rx->rx_msg, rx->rx_msgaddr, - lnet_page2phys(pg) + pg_off); + LASSERT (tpo->tpo_pool.po_allocated == 0); - pg_off += IBLND_MSG_SIZE; - LASSERT (pg_off <= PAGE_SIZE); + if (hdev == NULL) + return; - if (pg_off == PAGE_SIZE) { - pg_off = 0; - ipg++; - LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version)); - } + for (i = 0; i < tpo->tpo_pool.po_size; i++) { + tx = &tpo->tpo_tx_descs[i]; + kiblnd_dma_unmap_single(hdev->ibh_ibdev, + KIBLND_UNMAP_ADDR(tx, tx_msgunmap, + tx->tx_msgaddr), + IBLND_MSG_SIZE, DMA_TO_DEVICE); } - conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device; + kiblnd_hdev_decref(hdev); + tpo->tpo_hdev = NULL; } -void -kiblnd_unmap_tx_descs(lnet_ni_t *ni) +static kib_hca_dev_t * +kiblnd_current_hdev(kib_dev_t *dev) { - int i; - kib_tx_t *tx; - kib_net_t *net = ni->ni_data; + kib_hca_dev_t *hdev; + unsigned long flags; + int i = 0; - LASSERT (net->ibn_tx_pages != NULL); - LASSERT (net->ibn_tx_pages->ibp_device != NULL); + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + while (dev->ibd_failover) { + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + if (i++ % 50 == 0) + CDEBUG(D_NET, "%s: Wait for failover\n", + dev->ibd_ifname); + set_current_state(TASK_INTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1) / 100); - for (i = 0; i < IBLND_TX_MSGS(); i++) { - tx = &net->ibn_tx_descs[i]; + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + } - kiblnd_dma_unmap_single(net->ibn_tx_pages->ibp_device, - KIBLND_UNMAP_ADDR(tx, tx_msgunmap, - tx->tx_msgaddr), - IBLND_MSG_SIZE, DMA_TO_DEVICE); - } + kiblnd_hdev_addref_locked(dev->ibd_hdev); + hdev = dev->ibd_hdev; + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - net->ibn_tx_pages->ibp_device = NULL; + return hdev; } -void -kiblnd_map_tx_descs (lnet_ni_t *ni) +static void +kiblnd_map_tx_pool(kib_tx_pool_t *tpo) { - kib_net_t *net = ni->ni_data; + kib_pages_t *txpgs = tpo->tpo_tx_pages; + kib_pool_t *pool = &tpo->tpo_pool; + kib_net_t *net = pool->po_owner->ps_net; + kib_dev_t *dev; struct page *page; kib_tx_t *tx; int page_offset; @@ -1250,427 +1350,1163 @@ kiblnd_map_tx_descs (lnet_ni_t *ni) LASSERT (net != NULL); + dev = net->ibn_dev; + /* pre-mapped messages are not bigger than 1 page */ CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE); /* No fancy arithmetic when we do the buffer calculations */ CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0); + tpo->tpo_hdev = kiblnd_current_hdev(dev); - for (ipage = page_offset = i = 0; i < IBLND_TX_MSGS(); i++) { - page = net->ibn_tx_pages->ibp_pages[ipage]; - tx = &net->ibn_tx_descs[i]; - - tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) + - page_offset); + for (ipage = page_offset = i = 0; i < pool->po_size; i++) { + page = txpgs->ibp_pages[ipage]; + tx = &tpo->tpo_tx_descs[i]; - tx->tx_msgaddr = kiblnd_dma_map_single( - net->ibn_dev->ibd_cmid->device, - tx->tx_msg, IBLND_MSG_SIZE, DMA_TO_DEVICE); - LASSERT (!kiblnd_dma_mapping_error(net->ibn_dev->ibd_cmid->device, - tx->tx_msgaddr)); - KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr); + tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) + + page_offset); - list_add(&tx->tx_list, &net->ibn_idle_txs); + tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev, + tx->tx_msg, + IBLND_MSG_SIZE, + DMA_TO_DEVICE); + LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev, + tx->tx_msgaddr)); + KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr); - page_offset += IBLND_MSG_SIZE; - LASSERT (page_offset <= PAGE_SIZE); + list_add(&tx->tx_list, &pool->po_free_list); - if (page_offset == PAGE_SIZE) { - page_offset = 0; - ipage++; - LASSERT (ipage <= IBLND_TX_MSG_PAGES()); - } - } + page_offset += IBLND_MSG_SIZE; + LASSERT(page_offset <= PAGE_SIZE); - net->ibn_tx_pages->ibp_device = net->ibn_dev->ibd_cmid->device; + if (page_offset == PAGE_SIZE) { + page_offset = 0; + ipage++; + LASSERT(ipage <= txpgs->ibp_npages); + } + } } -void -kiblnd_free_tx_descs (lnet_ni_t *ni) +struct ib_mr * +kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd, + int negotiated_nfrags) { - int i; - kib_net_t *net = ni->ni_data; + kib_net_t *net = ni->ni_data; + kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev; + struct lnet_ioctl_config_o2iblnd_tunables *tunables; + int mod; + __u16 nfrags; - LASSERT (net != NULL); + tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib; + mod = tunables->lnd_map_on_demand; + nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod; - if (net->ibn_tx_pages != NULL) - kiblnd_free_pages(net->ibn_tx_pages); + LASSERT(hdev->ibh_mrs != NULL); - if (net->ibn_tx_descs == NULL) - return; + if (mod > 0 && nfrags <= rd->rd_nfrags) + return NULL; - for (i = 0; i < IBLND_TX_MSGS(); i++) { - kib_tx_t *tx = &net->ibn_tx_descs[i]; + return hdev->ibh_mrs; +} - if (tx->tx_pages != NULL) - LIBCFS_FREE(tx->tx_pages, - LNET_MAX_IOV * - sizeof(*tx->tx_pages)); +static void +kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo) +{ + LASSERT(fpo->fpo_map_count == 0); + + if (fpo->fpo_is_fmr) { + if (fpo->fmr.fpo_fmr_pool) + ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool); + } else { + struct kib_fast_reg_descriptor *frd, *tmp; + int i = 0; + + list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list, + frd_list) { + list_del(&frd->frd_list); +#ifndef HAVE_IB_MAP_MR_SG + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + ib_dereg_mr(frd->frd_mr); + LIBCFS_FREE(frd, sizeof(*frd)); + i++; + } + if (i < fpo->fast_reg.fpo_pool_size) + CERROR("FastReg pool still has %d regions registered\n", + fpo->fast_reg.fpo_pool_size - i); + } + + if (fpo->fpo_hdev) + kiblnd_hdev_decref(fpo->fpo_hdev); + + LIBCFS_FREE(fpo, sizeof(*fpo)); +} - if (tx->tx_ipb != NULL) - LIBCFS_FREE(tx->tx_ipb, - IBLND_MAX_RDMA_FRAGS * - sizeof(*tx->tx_ipb)); +static void +kiblnd_destroy_fmr_pool_list(struct list_head *head) +{ + kib_fmr_pool_t *fpo, *tmp; - if (tx->tx_frags != NULL) - LIBCFS_FREE(tx->tx_frags, - IBLND_MAX_RDMA_FRAGS * - sizeof(*tx->tx_frags)); + list_for_each_entry_safe(fpo, tmp, head, fpo_list) { + list_del(&fpo->fpo_list); + kiblnd_destroy_fmr_pool(fpo); + } +} - if (tx->tx_wrq != NULL) - LIBCFS_FREE(tx->tx_wrq, - (1 + IBLND_MAX_RDMA_FRAGS) * - sizeof(*tx->tx_wrq)); +static int +kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables, + int ncpts) +{ + int size = tunables->lnd_fmr_pool_size / ncpts; - if (tx->tx_sge != NULL) - LIBCFS_FREE(tx->tx_sge, - (1 + IBLND_MAX_RDMA_FRAGS) * - sizeof(*tx->tx_sge)); + return max(IBLND_FMR_POOL, size); +} - if (tx->tx_rd != NULL) - LIBCFS_FREE(tx->tx_rd, - offsetof(kib_rdma_desc_t, - rd_frags[IBLND_MAX_RDMA_FRAGS])); - } +static int +kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables, + int ncpts) +{ + int size = tunables->lnd_fmr_flush_trigger / ncpts; - LIBCFS_FREE(net->ibn_tx_descs, - IBLND_TX_MSGS() * sizeof(kib_tx_t)); + return max(IBLND_FMR_POOL_FLUSH, size); } -int -kiblnd_alloc_tx_descs (lnet_ni_t *ni) +static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo) { - int i; - int rc; - kib_net_t *net = ni->ni_data; + struct ib_fmr_pool_param param = { + .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE, + .page_shift = PAGE_SHIFT, + .access = (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE), + .pool_size = fps->fps_pool_size, + .dirty_watermark = fps->fps_flush_trigger, + .flush_function = NULL, + .flush_arg = NULL, + .cache = !!fps->fps_cache }; + int rc = 0; + + fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, + ¶m); + if (IS_ERR(fpo->fmr.fpo_fmr_pool)) { + rc = PTR_ERR(fpo->fmr.fpo_fmr_pool); + if (rc != -ENOSYS) + CERROR("Failed to create FMR pool: %d\n", rc); + else + CERROR("FMRs are not supported\n"); + } + + return rc; +} - LASSERT (net != NULL); +static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo) +{ + struct kib_fast_reg_descriptor *frd, *tmp; + int i, rc; + + INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list); + fpo->fast_reg.fpo_pool_size = 0; + for (i = 0; i < fps->fps_pool_size; i++) { + LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt, + sizeof(*frd)); + if (!frd) { + CERROR("Failed to allocate a new fast_reg descriptor\n"); + rc = -ENOMEM; + goto out; + } + frd->frd_mr = NULL; + +#ifndef HAVE_IB_MAP_MR_SG + frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev, + LNET_MAX_PAYLOAD/PAGE_SIZE); + if (IS_ERR(frd->frd_frpl)) { + rc = PTR_ERR(frd->frd_frpl); + CERROR("Failed to allocate ib_fast_reg_page_list: %d\n", + rc); + frd->frd_frpl = NULL; + goto out_middle; + } +#endif - rc = kiblnd_alloc_pages(&net->ibn_tx_pages, IBLND_TX_MSG_PAGES()); +#ifdef HAVE_IB_ALLOC_FAST_REG_MR + frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd, + LNET_MAX_PAYLOAD/PAGE_SIZE); +#else + frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd, + IB_MR_TYPE_MEM_REG, + LNET_MAX_PAYLOAD/PAGE_SIZE); +#endif + if (IS_ERR(frd->frd_mr)) { + rc = PTR_ERR(frd->frd_mr); + CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc); + frd->frd_mr = NULL; + goto out_middle; + } + + /* There appears to be a bug in MLX5 code where you must + * invalidate the rkey of a new FastReg pool before first + * using it. Thus, I am marking the FRD invalid here. */ + frd->frd_valid = false; + + list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list); + fpo->fast_reg.fpo_pool_size++; + } + + return 0; + +out_middle: + if (frd->frd_mr) + ib_dereg_mr(frd->frd_mr); +#ifndef HAVE_IB_MAP_MR_SG + if (frd->frd_frpl) + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + LIBCFS_FREE(frd, sizeof(*frd)); - if (rc != 0) { - CERROR("Can't allocate tx pages\n"); - return rc; - } +out: + list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list, + frd_list) { + list_del(&frd->frd_list); +#ifndef HAVE_IB_MAP_MR_SG + ib_free_fast_reg_page_list(frd->frd_frpl); +#endif + ib_dereg_mr(frd->frd_mr); + LIBCFS_FREE(frd, sizeof(*frd)); + } - LIBCFS_ALLOC (net->ibn_tx_descs, - IBLND_TX_MSGS() * sizeof(kib_tx_t)); - if (net->ibn_tx_descs == NULL) { - CERROR("Can't allocate %d tx descriptors\n", IBLND_TX_MSGS()); - return -ENOMEM; - } + return rc; +} - memset(net->ibn_tx_descs, 0, - IBLND_TX_MSGS() * sizeof(kib_tx_t)); +static int +kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) +{ + struct ib_device_attr *dev_attr; + kib_dev_t *dev = fps->fps_net->ibn_dev; + kib_fmr_pool_t *fpo; + int rc; + +#ifndef HAVE_IB_DEVICE_ATTRS + dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL); + if (!dev_attr) + return -ENOMEM; +#endif - for (i = 0; i < IBLND_TX_MSGS(); i++) { - kib_tx_t *tx = &net->ibn_tx_descs[i]; + LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo)); + if (!fpo) { + rc = -ENOMEM; + goto out_dev_attr; + } - if (net->ibn_fmrpool != NULL) { - LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV * - sizeof(*tx->tx_pages)); - if (tx->tx_pages == NULL) - return -ENOMEM; - } + fpo->fpo_hdev = kiblnd_current_hdev(dev); - if (net->ibn_pmrpool != NULL) { - LIBCFS_ALLOC(tx->tx_ipb, - IBLND_MAX_RDMA_FRAGS * - sizeof(*tx->tx_ipb)); - if (tx->tx_ipb == NULL) - return -ENOMEM; - } +#ifdef HAVE_IB_DEVICE_ATTRS + dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs; +#else + rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr); + if (rc) { + CERROR("Query device failed for %s: %d\n", + fpo->fpo_hdev->ibh_ibdev->name, rc); + goto out_dev_attr; + } +#endif - LIBCFS_ALLOC(tx->tx_frags, - IBLND_MAX_RDMA_FRAGS * - sizeof(*tx->tx_frags)); - if (tx->tx_frags == NULL) - return -ENOMEM; + /* Check for FMR or FastReg support */ + fpo->fpo_is_fmr = 0; + if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr && + fpo->fpo_hdev->ibh_ibdev->dealloc_fmr && + fpo->fpo_hdev->ibh_ibdev->map_phys_fmr && + fpo->fpo_hdev->ibh_ibdev->unmap_fmr) { + LCONSOLE_INFO("Using FMR for registration\n"); + fpo->fpo_is_fmr = 1; + } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { + LCONSOLE_INFO("Using FastReg for registration\n"); + } else { + rc = -ENOSYS; + LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n"); + goto out_dev_attr; + } + + if (fpo->fpo_is_fmr) + rc = kiblnd_alloc_fmr_pool(fps, fpo); + else + rc = kiblnd_alloc_freg_pool(fps, fpo); + if (rc) + goto out_fpo; + +#ifndef HAVE_IB_DEVICE_ATTRS + kfree(dev_attr); +#endif + fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); + fpo->fpo_owner = fps; + *pp_fpo = fpo; - LIBCFS_ALLOC(tx->tx_wrq, - (1 + IBLND_MAX_RDMA_FRAGS) * - sizeof(*tx->tx_wrq)); - if (tx->tx_wrq == NULL) - return -ENOMEM; + return 0; - LIBCFS_ALLOC(tx->tx_sge, - (1 + IBLND_MAX_RDMA_FRAGS) * - sizeof(*tx->tx_sge)); - if (tx->tx_sge == NULL) - return -ENOMEM; +out_fpo: + kiblnd_hdev_decref(fpo->fpo_hdev); + LIBCFS_FREE(fpo, sizeof(*fpo)); - LIBCFS_ALLOC(tx->tx_rd, - offsetof(kib_rdma_desc_t, - rd_frags[IBLND_MAX_RDMA_FRAGS])); - if (tx->tx_rd == NULL) - return -ENOMEM; - } +out_dev_attr: +#ifndef HAVE_IB_DEVICE_ATTRS + kfree(dev_attr); +#endif - return 0; + return rc; } -struct ib_mr * -kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size) +static void +kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies) { - __u64 index; - - LASSERT (net->ibn_dev->ibd_mrs[0] != NULL); - - if (net->ibn_dev->ibd_nmrs == 1) - return net->ibn_dev->ibd_mrs[0]; - - index = addr >> net->ibn_dev->ibd_mr_shift; - - if (index < net->ibn_dev->ibd_nmrs && - index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift)) - return net->ibn_dev->ibd_mrs[index]; + if (fps->fps_net == NULL) /* intialized? */ + return; + + spin_lock(&fps->fps_lock); + + while (!list_empty(&fps->fps_pool_list)) { + kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next, + kib_fmr_pool_t, fpo_list); + fpo->fpo_failed = 1; + list_del(&fpo->fpo_list); + if (fpo->fpo_map_count == 0) + list_add(&fpo->fpo_list, zombies); + else + list_add(&fpo->fpo_list, &fps->fps_failed_pool_list); + } + + spin_unlock(&fps->fps_lock); +} - return NULL; +static void +kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps) +{ + if (fps->fps_net != NULL) { /* initialized? */ + kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list); + kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list); + } } -struct ib_mr * -kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd) +static int +kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts, + kib_net_t *net, + struct lnet_ioctl_config_o2iblnd_tunables *tunables) { - struct ib_mr *prev_mr; - struct ib_mr *mr; - int i; + kib_fmr_pool_t *fpo; + int rc; - LASSERT (net->ibn_dev->ibd_mrs[0] != NULL); + memset(fps, 0, sizeof(kib_fmr_poolset_t)); - if (*kiblnd_tunables.kib_map_on_demand > 0 && - *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags) - return NULL; + fps->fps_net = net; + fps->fps_cpt = cpt; - if (net->ibn_dev->ibd_nmrs == 1) - return net->ibn_dev->ibd_mrs[0]; + fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts); + fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts); + fps->fps_cache = tunables->lnd_fmr_cache; - for (i = 0, mr = prev_mr = NULL; - i < rd->rd_nfrags; i++) { - mr = kiblnd_find_dma_mr(net, - rd->rd_frags[i].rf_addr, - rd->rd_frags[i].rf_nob); - if (prev_mr == NULL) - prev_mr = mr; + spin_lock_init(&fps->fps_lock); + INIT_LIST_HEAD(&fps->fps_pool_list); + INIT_LIST_HEAD(&fps->fps_failed_pool_list); - if (mr == NULL || prev_mr != mr) { - /* Can't covered by one single MR */ - mr = NULL; - break; - } - } + rc = kiblnd_create_fmr_pool(fps, &fpo); + if (rc == 0) + list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); - return mr; + return rc; } -void -kiblnd_dev_cleanup(kib_dev_t *ibdev) +static int +kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now) { - int i; - - if (ibdev->ibd_mrs == NULL) - return; - - for (i = 0; i < ibdev->ibd_nmrs; i++) { - if (ibdev->ibd_mrs[i] == NULL) - break; + if (fpo->fpo_map_count != 0) /* still in use */ + return 0; + if (fpo->fpo_failed) + return 1; + return cfs_time_aftereq(now, fpo->fpo_deadline); +} - ib_dereg_mr(ibdev->ibd_mrs[i]); - } +static int +kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd) +{ + kib_hca_dev_t *hdev; + __u64 *pages = tx->tx_pages; + int npages; + int size; + int i; + + hdev = tx->tx_pool->tpo_hdev; + + for (i = 0, npages = 0; i < rd->rd_nfrags; i++) { + for (size = 0; size < rd->rd_frags[i].rf_nob; + size += hdev->ibh_page_size) { + pages[npages++] = (rd->rd_frags[i].rf_addr & + hdev->ibh_page_mask) + size; + } + } + + return npages; +} - LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); - ibdev->ibd_mrs = NULL; +void +kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) +{ + struct list_head zombies = LIST_HEAD_INIT(zombies); + kib_fmr_pool_t *fpo = fmr->fmr_pool; + kib_fmr_poolset_t *fps; + cfs_time_t now = cfs_time_current(); + kib_fmr_pool_t *tmp; + int rc; + + if (!fpo) + return; + + fps = fpo->fpo_owner; + if (fpo->fpo_is_fmr) { + if (fmr->fmr_pfmr) { + rc = ib_fmr_pool_unmap(fmr->fmr_pfmr); + LASSERT(!rc); + fmr->fmr_pfmr = NULL; + } + + if (status) { + rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool); + LASSERT(!rc); + } + } else { + struct kib_fast_reg_descriptor *frd = fmr->fmr_frd; + + if (frd) { + frd->frd_valid = false; + spin_lock(&fps->fps_lock); + list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list); + spin_unlock(&fps->fps_lock); + fmr->fmr_frd = NULL; + } + } + fmr->fmr_pool = NULL; + + spin_lock(&fps->fps_lock); + fpo->fpo_map_count--; /* decref the pool */ + + list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) { + /* the first pool is persistent */ + if (fps->fps_pool_list.next == &fpo->fpo_list) + continue; + + if (kiblnd_fmr_pool_is_idle(fpo, now)) { + list_move(&fpo->fpo_list, &zombies); + fps->fps_version++; + } + } + spin_unlock(&fps->fps_lock); + + if (!list_empty(&zombies)) + kiblnd_destroy_fmr_pool_list(&zombies); } int -kiblnd_ib_create_fmr_pool(kib_dev_t *ibdev, struct ib_fmr_pool **fmrpp) -{ - /* FMR pool for RDMA */ - struct ib_fmr_pool *fmrpool; - struct ib_fmr_pool_param param = { - .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE, - .page_shift = PAGE_SHIFT, - .access = (IB_ACCESS_LOCAL_WRITE | - IB_ACCESS_REMOTE_WRITE), - .pool_size = *kiblnd_tunables.kib_fmr_pool_size, - .dirty_watermark = *kiblnd_tunables.kib_fmr_flush_trigger, - .flush_function = NULL, - .flush_arg = NULL, - .cache = *kiblnd_tunables.kib_fmr_cache}; - - if (*kiblnd_tunables.kib_fmr_pool_size < - *kiblnd_tunables.kib_ntx) { - CERROR("Can't set fmr pool size (%d) < ntx(%d)\n", - *kiblnd_tunables.kib_fmr_pool_size, - *kiblnd_tunables.kib_ntx); - return -EINVAL; - } - - fmrpool = ib_create_fmr_pool(ibdev->ibd_pd, ¶m); - if (IS_ERR(fmrpool)) - return PTR_ERR(fmrpool); +kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd, + __u32 nob, __u64 iov, kib_fmr_t *fmr) +{ + kib_fmr_pool_t *fpo; + __u64 *pages = tx->tx_pages; + __u64 version; + bool is_rx = (rd != tx->tx_rd); + bool tx_pages_mapped = 0; + int npages = 0; + int rc; + +again: + spin_lock(&fps->fps_lock); + version = fps->fps_version; + list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) { + fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); + fpo->fpo_map_count++; + + if (fpo->fpo_is_fmr) { + struct ib_pool_fmr *pfmr; + + spin_unlock(&fps->fps_lock); + + if (!tx_pages_mapped) { + npages = kiblnd_map_tx_pages(tx, rd); + tx_pages_mapped = 1; + } + + pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool, + pages, npages, iov); + if (likely(!IS_ERR(pfmr))) { + fmr->fmr_key = is_rx ? pfmr->fmr->rkey + : pfmr->fmr->lkey; + fmr->fmr_frd = NULL; + fmr->fmr_pfmr = pfmr; + fmr->fmr_pool = fpo; + return 0; + } + rc = PTR_ERR(pfmr); + } else { + if (!list_empty(&fpo->fast_reg.fpo_pool_list)) { + struct kib_fast_reg_descriptor *frd; +#ifdef HAVE_IB_MAP_MR_SG + struct ib_reg_wr *wr; + int n; +#else + struct ib_rdma_wr *wr; + struct ib_fast_reg_page_list *frpl; +#endif + struct ib_mr *mr; - *fmrpp = fmrpool; + frd = list_first_entry(&fpo->fast_reg.fpo_pool_list, + struct kib_fast_reg_descriptor, + frd_list); + list_del(&frd->frd_list); + spin_unlock(&fps->fps_lock); - return 0; +#ifndef HAVE_IB_MAP_MR_SG + frpl = frd->frd_frpl; +#endif + mr = frd->frd_mr; + + if (!frd->frd_valid) { + struct ib_rdma_wr *inv_wr; + __u32 key = is_rx ? mr->rkey : mr->lkey; + + inv_wr = &frd->frd_inv_wr; + memset(inv_wr, 0, sizeof(*inv_wr)); + + inv_wr->wr.opcode = IB_WR_LOCAL_INV; + inv_wr->wr.wr_id = IBLND_WID_MR; + inv_wr->wr.ex.invalidate_rkey = key; + + /* Bump the key */ + key = ib_inc_rkey(key); + ib_update_fast_reg_key(mr, key); + } + +#ifdef HAVE_IB_MAP_MR_SG +#ifdef HAVE_IB_MAP_MR_SG_5ARGS + n = ib_map_mr_sg(mr, tx->tx_frags, + tx->tx_nfrags, NULL, PAGE_SIZE); +#else + n = ib_map_mr_sg(mr, tx->tx_frags, + tx->tx_nfrags, PAGE_SIZE); +#endif + if (unlikely(n != tx->tx_nfrags)) { + CERROR("Failed to map mr %d/%d " + "elements\n", n, tx->tx_nfrags); + return n < 0 ? n : -EINVAL; + } + + mr->iova = iov; + + wr = &frd->frd_fastreg_wr; + memset(wr, 0, sizeof(*wr)); + + wr->wr.opcode = IB_WR_REG_MR; + wr->wr.wr_id = IBLND_WID_MR; + wr->wr.num_sge = 0; + wr->wr.send_flags = 0; + wr->mr = mr; + wr->key = is_rx ? mr->rkey : mr->lkey; + wr->access = (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE); +#else + if (!tx_pages_mapped) { + npages = kiblnd_map_tx_pages(tx, rd); + tx_pages_mapped = 1; + } + + LASSERT(npages <= frpl->max_page_list_len); + memcpy(frpl->page_list, pages, + sizeof(*pages) * npages); + + /* Prepare FastReg WR */ + wr = &frd->frd_fastreg_wr; + memset(wr, 0, sizeof(*wr)); + + wr->wr.opcode = IB_WR_FAST_REG_MR; + wr->wr.wr_id = IBLND_WID_MR; + + wr->wr.wr.fast_reg.iova_start = iov; + wr->wr.wr.fast_reg.page_list = frpl; + wr->wr.wr.fast_reg.page_list_len = npages; + wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT; + wr->wr.wr.fast_reg.length = nob; + wr->wr.wr.fast_reg.rkey = + is_rx ? mr->rkey : mr->lkey; + wr->wr.wr.fast_reg.access_flags = + (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE); +#endif + + fmr->fmr_key = is_rx ? mr->rkey : mr->lkey; + fmr->fmr_frd = frd; + fmr->fmr_pfmr = NULL; + fmr->fmr_pool = fpo; + return 0; + } + spin_unlock(&fps->fps_lock); + rc = -EBUSY; + } + + spin_lock(&fps->fps_lock); + fpo->fpo_map_count--; + if (rc != -EAGAIN) { + spin_unlock(&fps->fps_lock); + return rc; + } + + /* EAGAIN and ... */ + if (version != fps->fps_version) { + spin_unlock(&fps->fps_lock); + goto again; + } + } + + if (fps->fps_increasing) { + spin_unlock(&fps->fps_lock); + CDEBUG(D_NET, "Another thread is allocating new " + "FMR pool, waiting for her to complete\n"); + schedule(); + goto again; + + } + + if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) { + /* someone failed recently */ + spin_unlock(&fps->fps_lock); + return -EAGAIN; + } + + fps->fps_increasing = 1; + spin_unlock(&fps->fps_lock); + + CDEBUG(D_NET, "Allocate new FMR pool\n"); + rc = kiblnd_create_fmr_pool(fps, &fpo); + spin_lock(&fps->fps_lock); + fps->fps_increasing = 0; + if (rc == 0) { + fps->fps_version++; + list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); + } else { + fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY); + } + spin_unlock(&fps->fps_lock); + + goto again; } -void -kiblnd_phys_mr_unmap(kib_net_t *net, kib_phys_mr_t *pmr) +static void +kiblnd_fini_pool(kib_pool_t *pool) { - kib_phys_mr_pool_t *pool = net->ibn_pmrpool; - struct ib_mr *mr; + LASSERT(list_empty(&pool->po_free_list)); + LASSERT(pool->po_allocated == 0); + + CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name); +} - spin_lock(&pool->ibmp_lock); +static void +kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size) +{ + CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name); - mr = pmr->ibpm_mr; + memset(pool, 0, sizeof(kib_pool_t)); + INIT_LIST_HEAD(&pool->po_free_list); + pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); + pool->po_owner = ps; + pool->po_size = size; +} - list_add(&pmr->ibpm_link, &pool->ibmp_free_list); - pool->ibmp_allocated --; +static void +kiblnd_destroy_pool_list(struct list_head *head) +{ + kib_pool_t *pool; - spin_unlock(&pool->ibmp_lock); + while (!list_empty(head)) { + pool = list_entry(head->next, kib_pool_t, po_list); + list_del(&pool->po_list); - if (mr != NULL) - ib_dereg_mr(mr); + LASSERT(pool->po_owner != NULL); + pool->po_owner->ps_pool_destroy(pool); + } } -kib_phys_mr_t * -kiblnd_phys_mr_map(kib_net_t *net, kib_rdma_desc_t *rd, - struct ib_phys_buf *ipb, __u64 *iova) +static void +kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies) { - kib_phys_mr_pool_t *pool = net->ibn_pmrpool; - kib_phys_mr_t *pmr; - int i; - - LASSERT (ipb != NULL); + if (ps->ps_net == NULL) /* intialized? */ + return; + + spin_lock(&ps->ps_lock); + while (!list_empty(&ps->ps_pool_list)) { + kib_pool_t *po = list_entry(ps->ps_pool_list.next, + kib_pool_t, po_list); + po->po_failed = 1; + list_del(&po->po_list); + if (po->po_allocated == 0) + list_add(&po->po_list, zombies); + else + list_add(&po->po_list, &ps->ps_failed_pool_list); + } + spin_unlock(&ps->ps_lock); +} - spin_lock(&pool->ibmp_lock); - if (list_empty(&pool->ibmp_free_list)) { - spin_unlock(&pool->ibmp_lock); - CERROR("pre-allocated MRs is not enough\n"); +static void +kiblnd_fini_poolset(kib_poolset_t *ps) +{ + if (ps->ps_net != NULL) { /* initialized? */ + kiblnd_destroy_pool_list(&ps->ps_failed_pool_list); + kiblnd_destroy_pool_list(&ps->ps_pool_list); + } +} - return NULL; - } +static int +kiblnd_init_poolset(kib_poolset_t *ps, int cpt, + kib_net_t *net, char *name, int size, + kib_ps_pool_create_t po_create, + kib_ps_pool_destroy_t po_destroy, + kib_ps_node_init_t nd_init, + kib_ps_node_fini_t nd_fini) +{ + kib_pool_t *pool; + int rc; + + memset(ps, 0, sizeof(kib_poolset_t)); + + ps->ps_cpt = cpt; + ps->ps_net = net; + ps->ps_pool_create = po_create; + ps->ps_pool_destroy = po_destroy; + ps->ps_node_init = nd_init; + ps->ps_node_fini = nd_fini; + ps->ps_pool_size = size; + if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name)) + >= sizeof(ps->ps_name)) + return -E2BIG; + spin_lock_init(&ps->ps_lock); + INIT_LIST_HEAD(&ps->ps_pool_list); + INIT_LIST_HEAD(&ps->ps_failed_pool_list); + + rc = ps->ps_pool_create(ps, size, &pool); + if (rc == 0) + list_add(&pool->po_list, &ps->ps_pool_list); + else + CERROR("Failed to create the first pool for %s\n", ps->ps_name); + + return rc; +} - pmr = list_entry(pool->ibmp_free_list.next, - kib_phys_mr_t, ibpm_link); - list_del_init(&pmr->ibpm_link); - pool->ibmp_allocated ++; +static int +kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now) +{ + if (pool->po_allocated != 0) /* still in use */ + return 0; + if (pool->po_failed) + return 1; + return cfs_time_aftereq(now, pool->po_deadline); +} - spin_unlock(&pool->ibmp_lock); +void +kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node) +{ + struct list_head zombies = LIST_HEAD_INIT(zombies); + kib_poolset_t *ps = pool->po_owner; + kib_pool_t *tmp; + cfs_time_t now = cfs_time_current(); - for (i = 0; i < rd->rd_nfrags; i ++) { - ipb[i].addr = rd->rd_frags[i].rf_addr; - ipb[i].size = rd->rd_frags[i].rf_nob; - } + spin_lock(&ps->ps_lock); - pmr->ibpm_mr = ib_reg_phys_mr(net->ibn_dev->ibd_pd, ipb, - rd->rd_nfrags, - IB_ACCESS_LOCAL_WRITE | - IB_ACCESS_REMOTE_WRITE, - iova); - if (!IS_ERR(pmr->ibpm_mr)) { - pmr->ibpm_iova = *iova; - return pmr; - } + if (ps->ps_node_fini != NULL) + ps->ps_node_fini(pool, node); - CERROR("Failed ib_reg_phys_mr: %ld\n", PTR_ERR(pmr->ibpm_mr)); - pmr->ibpm_mr = NULL; + LASSERT(pool->po_allocated > 0); + list_add(node, &pool->po_free_list); + pool->po_allocated--; - spin_lock(&pool->ibmp_lock); + list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) { + /* the first pool is persistent */ + if (ps->ps_pool_list.next == &pool->po_list) + continue; - list_add(&pmr->ibpm_link, &pool->ibmp_free_list); - pool->ibmp_allocated --; + if (kiblnd_pool_is_idle(pool, now)) + list_move(&pool->po_list, &zombies); + } + spin_unlock(&ps->ps_lock); - spin_unlock(&pool->ibmp_lock); + if (!list_empty(&zombies)) + kiblnd_destroy_pool_list(&zombies); +} - return NULL; +struct list_head * +kiblnd_pool_alloc_node(kib_poolset_t *ps) +{ + struct list_head *node; + kib_pool_t *pool; + int rc; + unsigned int interval = 1; + cfs_time_t time_before; + unsigned int trips = 0; + +again: + spin_lock(&ps->ps_lock); + list_for_each_entry(pool, &ps->ps_pool_list, po_list) { + if (list_empty(&pool->po_free_list)) + continue; + + pool->po_allocated++; + pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE); + node = pool->po_free_list.next; + list_del(node); + + if (ps->ps_node_init != NULL) { + /* still hold the lock */ + ps->ps_node_init(pool, node); + } + spin_unlock(&ps->ps_lock); + return node; + } + + /* no available tx pool and ... */ + if (ps->ps_increasing) { + /* another thread is allocating a new pool */ + spin_unlock(&ps->ps_lock); + trips++; + CDEBUG(D_NET, "Another thread is allocating new " + "%s pool, waiting %d HZs for her to complete." + "trips = %d\n", + ps->ps_name, interval, trips); + + set_current_state(TASK_INTERRUPTIBLE); + schedule_timeout(interval); + if (interval < cfs_time_seconds(1)) + interval *= 2; + + goto again; + } + + if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) { + /* someone failed recently */ + spin_unlock(&ps->ps_lock); + return NULL; + } + + ps->ps_increasing = 1; + spin_unlock(&ps->ps_lock); + + CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name); + time_before = cfs_time_current(); + rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool); + CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete", + cfs_time_current() - time_before); + + spin_lock(&ps->ps_lock); + ps->ps_increasing = 0; + if (rc == 0) { + list_add_tail(&pool->po_list, &ps->ps_pool_list); + } else { + ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY); + CERROR("Can't allocate new %s pool because out of memory\n", + ps->ps_name); + } + spin_unlock(&ps->ps_lock); + + goto again; } -void -kiblnd_destroy_pmr_pool(kib_phys_mr_pool_t *pool) +static void +kiblnd_destroy_tx_pool(kib_pool_t *pool) { - kib_phys_mr_t *pmr; + kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool); + int i; - LASSERT (pool->ibmp_allocated == 0); + LASSERT (pool->po_allocated == 0); - while (!list_empty(&pool->ibmp_free_list)) { - pmr = list_entry(pool->ibmp_free_list.next, - kib_phys_mr_t, ibpm_link); + if (tpo->tpo_tx_pages != NULL) { + kiblnd_unmap_tx_pool(tpo); + kiblnd_free_pages(tpo->tpo_tx_pages); + } - LASSERT (pmr->ibpm_mr == NULL); + if (tpo->tpo_tx_descs == NULL) + goto out; - list_del(&pmr->ibpm_link); + for (i = 0; i < pool->po_size; i++) { + kib_tx_t *tx = &tpo->tpo_tx_descs[i]; - LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t)); + list_del(&tx->tx_list); + if (tx->tx_pages != NULL) + LIBCFS_FREE(tx->tx_pages, + LNET_MAX_IOV * + sizeof(*tx->tx_pages)); + if (tx->tx_frags != NULL) + LIBCFS_FREE(tx->tx_frags, + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_frags)); + if (tx->tx_wrq != NULL) + LIBCFS_FREE(tx->tx_wrq, + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_wrq)); + if (tx->tx_sge != NULL) + LIBCFS_FREE(tx->tx_sge, + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_sge)); + if (tx->tx_rd != NULL) + LIBCFS_FREE(tx->tx_rd, + offsetof(kib_rdma_desc_t, + rd_frags[IBLND_MAX_RDMA_FRAGS])); } - LIBCFS_FREE(pool, sizeof(kib_phys_mr_pool_t)); + LIBCFS_FREE(tpo->tpo_tx_descs, + pool->po_size * sizeof(kib_tx_t)); +out: + kiblnd_fini_pool(pool); + LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t)); } -int -kiblnd_create_pmr_pool(kib_dev_t *ibdev, kib_phys_mr_pool_t **poolpp) -{ - kib_phys_mr_pool_t *pool; - kib_phys_mr_t *pmr; - int i; - - if (*kiblnd_tunables.kib_pmr_pool_size < - *kiblnd_tunables.kib_ntx) { - CERROR("Can't set pmr pool size (%d) < ntx(%d)\n", - *kiblnd_tunables.kib_pmr_pool_size, - *kiblnd_tunables.kib_ntx); - return -EINVAL; +static int kiblnd_tx_pool_size(int ncpts) +{ + int ntx = *kiblnd_tunables.kib_ntx / ncpts; + + return max(IBLND_TX_POOL, ntx); +} + +static int +kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) +{ + int i; + int npg; + kib_pool_t *pool; + kib_tx_pool_t *tpo; + + LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo)); + if (tpo == NULL) { + CERROR("Failed to allocate TX pool\n"); + return -ENOMEM; } - LIBCFS_ALLOC(pool, sizeof(kib_phys_mr_pool_t)); - if (pool == NULL) + pool = &tpo->tpo_pool; + kiblnd_init_pool(ps, pool, size); + tpo->tpo_tx_descs = NULL; + tpo->tpo_tx_pages = NULL; + + npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE; + if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) { + CERROR("Can't allocate tx pages: %d\n", npg); + LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t)); + return -ENOMEM; + } + + LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt, + size * sizeof(kib_tx_t)); + if (tpo->tpo_tx_descs == NULL) { + CERROR("Can't allocate %d tx descriptors\n", size); + ps->ps_pool_destroy(pool); return -ENOMEM; + } - spin_lock_init(&pool->ibmp_lock); + memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t)); - pool->ibmp_allocated = 0; - CFS_INIT_LIST_HEAD(&pool->ibmp_free_list); + for (i = 0; i < size; i++) { + kib_tx_t *tx = &tpo->tpo_tx_descs[i]; - for (i = 0; i < *kiblnd_tunables.kib_pmr_pool_size; i++) { - LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t)); + tx->tx_pool = tpo; + if (ps->ps_net->ibn_fmr_ps != NULL) { + LIBCFS_CPT_ALLOC(tx->tx_pages, + lnet_cpt_table(), ps->ps_cpt, + LNET_MAX_IOV * sizeof(*tx->tx_pages)); + if (tx->tx_pages == NULL) + break; + } - if (pmr == NULL) { - kiblnd_destroy_pmr_pool(pool); - return -ENOMEM; - } + LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt, + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_frags)); + if (tx->tx_frags == NULL) + break; + + sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1); - memset(pmr, 0, sizeof(kib_phys_mr_t)); + LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt, + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_wrq)); + if (tx->tx_wrq == NULL) + break; - list_add(&pmr->ibpm_link, &pool->ibmp_free_list); + LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt, + (1 + IBLND_MAX_RDMA_FRAGS) * + sizeof(*tx->tx_sge)); + if (tx->tx_sge == NULL) + break; + + LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt, + offsetof(kib_rdma_desc_t, + rd_frags[IBLND_MAX_RDMA_FRAGS])); + if (tx->tx_rd == NULL) + break; } - *poolpp = pool; + if (i == size) { + kiblnd_map_tx_pool(tpo); + *pp_po = pool; + return 0; + } - return 0; + ps->ps_pool_destroy(pool); + return -ENOMEM; +} + +static void +kiblnd_tx_init(kib_pool_t *pool, struct list_head *node) +{ + kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t, + tps_poolset); + kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list); + + tx->tx_cookie = tps->tps_next_tx_cookie++; +} + +static void +kiblnd_net_fini_pools(kib_net_t *net) +{ + int i; + + cfs_cpt_for_each(i, lnet_cpt_table()) { + kib_tx_poolset_t *tps; + kib_fmr_poolset_t *fps; + + if (net->ibn_tx_ps != NULL) { + tps = net->ibn_tx_ps[i]; + kiblnd_fini_poolset(&tps->tps_poolset); + } + + if (net->ibn_fmr_ps != NULL) { + fps = net->ibn_fmr_ps[i]; + kiblnd_fini_fmr_poolset(fps); + } + } + + if (net->ibn_tx_ps != NULL) { + cfs_percpt_free(net->ibn_tx_ps); + net->ibn_tx_ps = NULL; + } + + if (net->ibn_fmr_ps != NULL) { + cfs_percpt_free(net->ibn_fmr_ps); + net->ibn_fmr_ps = NULL; + } +} + +static int +kiblnd_net_init_pools(kib_net_t *net, lnet_ni_t *ni, __u32 *cpts, int ncpts) +{ + struct lnet_ioctl_config_o2iblnd_tunables *tunables; + unsigned long flags; + int cpt; + int rc; + int i; + + tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib; + + read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + if (tunables->lnd_map_on_demand == 0) { + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, + flags); + goto create_tx_pool; + } + + read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + + if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) { + CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n", + tunables->lnd_fmr_pool_size, + *kiblnd_tunables.kib_ntx / 4); + rc = -EINVAL; + goto failed; + } + + /* TX pool must be created later than FMR, see LU-2268 + * for details */ + LASSERT(net->ibn_tx_ps == NULL); + + /* premapping can fail if ibd_nmr > 1, so we always create + * FMR pool and map-on-demand if premapping failed */ + + net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(), + sizeof(kib_fmr_poolset_t)); + if (net->ibn_fmr_ps == NULL) { + CERROR("Failed to allocate FMR pool array\n"); + rc = -ENOMEM; + goto failed; + } + + for (i = 0; i < ncpts; i++) { + cpt = (cpts == NULL) ? i : cpts[i]; + rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts, + net, tunables); + if (rc != 0) { + CERROR("Can't initialize FMR pool for CPT %d: %d\n", + cpt, rc); + goto failed; + } + } + + if (i > 0) + LASSERT(i == ncpts); + + create_tx_pool: + net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(), + sizeof(kib_tx_poolset_t)); + if (net->ibn_tx_ps == NULL) { + CERROR("Failed to allocate tx pool array\n"); + rc = -ENOMEM; + goto failed; + } + + for (i = 0; i < ncpts; i++) { + cpt = (cpts == NULL) ? i : cpts[i]; + rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset, + cpt, net, "TX", + kiblnd_tx_pool_size(ncpts), + kiblnd_create_tx_pool, + kiblnd_destroy_tx_pool, + kiblnd_tx_init, NULL); + if (rc != 0) { + CERROR("Can't initialize TX pool for CPT %d: %d\n", + cpt, rc); + goto failed; + } + } + + return 0; + failed: + kiblnd_net_fini_pools(net); + LASSERT(rc != 0); + return rc; } static int -kiblnd_dev_get_attr(kib_dev_t *ibdev) +kiblnd_hdev_get_attr(kib_hca_dev_t *hdev) { - struct ib_device_attr *attr; - int rc; +#ifndef HAVE_IB_DEVICE_ATTRS + struct ib_device_attr *attr; + int rc; +#endif /* It's safe to assume a HCA can handle a page size * matching that of the native system */ - ibdev->ibd_page_shift = PAGE_SHIFT; - ibdev->ibd_page_size = 1 << PAGE_SHIFT; - ibdev->ibd_page_mask = ~((__u64)ibdev->ibd_page_size - 1); + hdev->ibh_page_shift = PAGE_SHIFT; + hdev->ibh_page_size = 1 << PAGE_SHIFT; + hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1); +#ifdef HAVE_IB_DEVICE_ATTRS + hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size; +#else LIBCFS_ALLOC(attr, sizeof(*attr)); if (attr == NULL) { CERROR("Out of memory\n"); return -ENOMEM; } - rc = ib_query_device(ibdev->ibd_cmid->device, attr); + rc = ib_query_device(hdev->ibh_ibdev, attr); if (rc == 0) - ibdev->ibd_mr_size = attr->max_mr_size; + hdev->ibh_mr_size = attr->max_mr_size; LIBCFS_FREE(attr, sizeof(*attr)); @@ -1678,150 +2514,336 @@ kiblnd_dev_get_attr(kib_dev_t *ibdev) CERROR("Failed to query IB device: %d\n", rc); return rc; } - -#ifdef HAVE_OFED_TRANSPORT_IWARP - /* XXX We can't trust this value returned by Chelsio driver, it's wrong - * and we have reported the bug, remove these in the future when Chelsio - * bug got fixed. */ - if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) == - RDMA_TRANSPORT_IWARP) - ibdev->ibd_mr_size = (1ULL << 32) - 1; #endif - if (ibdev->ibd_mr_size == ~0ULL) { - ibdev->ibd_mr_shift = 64; + if (hdev->ibh_mr_size == ~0ULL) { + hdev->ibh_mr_shift = 64; return 0; } - for (ibdev->ibd_mr_shift = 0; - ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) { - if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) || - ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1) - return 0; - } - - CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size); + CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size); return -EINVAL; } -int -kiblnd_dev_setup(kib_dev_t *ibdev) -{ - struct ib_mr *mr; - int i; - int rc; - __u64 mm_size; - __u64 mr_size; - int acflags = IB_ACCESS_LOCAL_WRITE | - IB_ACCESS_REMOTE_WRITE; - - rc = kiblnd_dev_get_attr(ibdev); - if (rc != 0) +static void +kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev) +{ + if (hdev->ibh_mrs == NULL) + return; + + ib_dereg_mr(hdev->ibh_mrs); + + hdev->ibh_mrs = NULL; +} + +void +kiblnd_hdev_destroy(kib_hca_dev_t *hdev) +{ + kiblnd_hdev_cleanup_mrs(hdev); + + if (hdev->ibh_pd != NULL) + ib_dealloc_pd(hdev->ibh_pd); + + if (hdev->ibh_cmid != NULL) + rdma_destroy_id(hdev->ibh_cmid); + + LIBCFS_FREE(hdev, sizeof(*hdev)); +} + +static int +kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev) +{ + struct ib_mr *mr; + int rc; + int acflags = IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_WRITE; + + rc = kiblnd_hdev_get_attr(hdev); + if (rc != 0) + return rc; + + mr = ib_get_dma_mr(hdev->ibh_pd, acflags); + if (IS_ERR(mr)) { + CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr)); + kiblnd_hdev_cleanup_mrs(hdev); + return PTR_ERR(mr); + } + + hdev->ibh_mrs = mr; + + return 0; +} + +static int +kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) +{ /* DUMMY */ + return 0; +} + +static int +kiblnd_dev_need_failover(kib_dev_t *dev) +{ + struct rdma_cm_id *cmid; + struct sockaddr_in srcaddr; + struct sockaddr_in dstaddr; + int rc; + + if (dev->ibd_hdev == NULL || /* initializing */ + dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */ + *kiblnd_tunables.kib_dev_failover > 1) /* debugging */ + return 1; + + /* XXX: it's UGLY, but I don't have better way to find + * ib-bonding HCA failover because: + * + * a. no reliable CM event for HCA failover... + * b. no OFED API to get ib_device for current net_device... + * + * We have only two choices at this point: + * + * a. rdma_bind_addr(), it will conflict with listener cmid + * b. rdma_resolve_addr() to zero addr */ + cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP, + IB_QPT_RC); + if (IS_ERR(cmid)) { + rc = PTR_ERR(cmid); + CERROR("Failed to create cmid for failover: %d\n", rc); return rc; + } - if (ibdev->ibd_mr_shift == 64) { - LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs)); - if (ibdev->ibd_mrs == NULL) { - CERROR("Failed to allocate MRs table\n"); - return -ENOMEM; - } + memset(&srcaddr, 0, sizeof(srcaddr)); + srcaddr.sin_family = AF_INET; + srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip); - ibdev->ibd_mrs[0] = NULL; - ibdev->ibd_nmrs = 1; + memset(&dstaddr, 0, sizeof(dstaddr)); + dstaddr.sin_family = AF_INET; + rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr, + (struct sockaddr *)&dstaddr, 1); + if (rc != 0 || cmid->device == NULL) { + CERROR("Failed to bind %s:%pI4h to device(%p): %d\n", + dev->ibd_ifname, &dev->ibd_ifip, + cmid->device, rc); + rdma_destroy_id(cmid); + return rc; + } - mr = ib_get_dma_mr(ibdev->ibd_pd, acflags); - if (IS_ERR(mr)) { - CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr)); - kiblnd_dev_cleanup(ibdev); - return PTR_ERR(mr); - } + rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */ + rdma_destroy_id(cmid); + return rc; +} + +int +kiblnd_dev_failover(kib_dev_t *dev) +{ + struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo); + struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo); + struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo); + struct rdma_cm_id *cmid = NULL; + kib_hca_dev_t *hdev = NULL; + kib_hca_dev_t *old; + struct ib_pd *pd; + kib_net_t *net; + struct sockaddr_in addr; + unsigned long flags; + int rc = 0; + int i; - ibdev->ibd_mrs[0] = mr; + LASSERT (*kiblnd_tunables.kib_dev_failover > 1 || + dev->ibd_can_failover || + dev->ibd_hdev == NULL); + rc = kiblnd_dev_need_failover(dev); + if (rc <= 0) goto out; - } - mr_size = (1ULL << ibdev->ibd_mr_shift); - mm_size = (unsigned long)high_memory - PAGE_OFFSET; + if (dev->ibd_hdev != NULL && + dev->ibd_hdev->ibh_cmid != NULL) { + /* XXX it's not good to close old listener at here, + * because we can fail to create new listener. + * But we have to close it now, otherwise rdma_bind_addr + * will return EADDRINUSE... How crap! */ + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift); + cmid = dev->ibd_hdev->ibh_cmid; + /* make next schedule of kiblnd_dev_need_failover() + * return 1 for me */ + dev->ibd_hdev->ibh_cmid = NULL; + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) { - /* it's 4T..., assume we will re-code at that time */ - CERROR("Can't support memory size: x"LPX64 - " with MR size: x"LPX64"\n", mm_size, mr_size); - return -EINVAL; + rdma_destroy_id(cmid); } - /* create an array of MRs to cover all memory */ - LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); - if (ibdev->ibd_mrs == NULL) { - CERROR("Failed to allocate MRs' table\n"); - return -ENOMEM; + cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP, + IB_QPT_RC); + if (IS_ERR(cmid)) { + rc = PTR_ERR(cmid); + CERROR("Failed to create cmid for failover: %d\n", rc); + goto out; } - memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); + memset(&addr, 0, sizeof(addr)); + addr.sin_family = AF_INET; + addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip); + addr.sin_port = htons(*kiblnd_tunables.kib_service); - for (i = 0; i < ibdev->ibd_nmrs; i++) { - struct ib_phys_buf ipb; - __u64 iova; + /* Bind to failover device or port */ + rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr); + if (rc != 0 || cmid->device == NULL) { + CERROR("Failed to bind %s:%pI4h to device(%p): %d\n", + dev->ibd_ifname, &dev->ibd_ifip, + cmid->device, rc); + rdma_destroy_id(cmid); + goto out; + } - ipb.size = ibdev->ibd_mr_size; - ipb.addr = i * mr_size; - iova = ipb.addr; + LIBCFS_ALLOC(hdev, sizeof(*hdev)); + if (hdev == NULL) { + CERROR("Failed to allocate kib_hca_dev\n"); + rdma_destroy_id(cmid); + rc = -ENOMEM; + goto out; + } - mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova); - if (IS_ERR(mr)) { - CERROR("Failed ib_reg_phys_mr addr "LPX64 - " size "LPX64" : %ld\n", - ipb.addr, ipb.size, PTR_ERR(mr)); - kiblnd_dev_cleanup(ibdev); - return PTR_ERR(mr); - } + atomic_set(&hdev->ibh_ref, 1); + hdev->ibh_dev = dev; + hdev->ibh_cmid = cmid; + hdev->ibh_ibdev = cmid->device; - LASSERT (iova == ipb.addr); + pd = ib_alloc_pd(cmid->device); + if (IS_ERR(pd)) { + rc = PTR_ERR(pd); + CERROR("Can't allocate PD: %d\n", rc); + goto out; + } + + hdev->ibh_pd = pd; - ibdev->ibd_mrs[i] = mr; + rc = rdma_listen(cmid, 0); + if (rc != 0) { + CERROR("Can't start new listener: %d\n", rc); + goto out; } -out: - CDEBUG(D_CONSOLE, "Register global MR array, MR size: " - LPX64", array size: %d\n", - ibdev->ibd_mr_size, ibdev->ibd_nmrs); + rc = kiblnd_hdev_setup_mrs(hdev); + if (rc != 0) { + CERROR("Can't setup device: %d\n", rc); + goto out; + } - list_add_tail(&ibdev->ibd_list, - &kiblnd_data.kib_devs); - return 0; + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + + old = dev->ibd_hdev; + dev->ibd_hdev = hdev; /* take over the refcount */ + hdev = old; + + list_for_each_entry(net, &dev->ibd_nets, ibn_list) { + cfs_cpt_for_each(i, lnet_cpt_table()) { + kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset, + &zombie_tpo); + + if (net->ibn_fmr_ps != NULL) + kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i], + &zombie_fpo); + } + } + + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + out: + if (!list_empty(&zombie_tpo)) + kiblnd_destroy_pool_list(&zombie_tpo); + if (!list_empty(&zombie_ppo)) + kiblnd_destroy_pool_list(&zombie_ppo); + if (!list_empty(&zombie_fpo)) + kiblnd_destroy_fmr_pool_list(&zombie_fpo); + if (hdev != NULL) + kiblnd_hdev_decref(hdev); + + if (rc != 0) + dev->ibd_failed_failover++; + else + dev->ibd_failed_failover = 0; + + return rc; } void kiblnd_destroy_dev (kib_dev_t *dev) { LASSERT (dev->ibd_nnets == 0); + LASSERT(list_empty(&dev->ibd_nets)); + + list_del(&dev->ibd_fail_list); + list_del(&dev->ibd_list); + + if (dev->ibd_hdev != NULL) + kiblnd_hdev_decref(dev->ibd_hdev); + + LIBCFS_FREE(dev, sizeof(*dev)); +} + +static kib_dev_t * +kiblnd_create_dev(char *ifname) +{ + struct net_device *netdev; + kib_dev_t *dev; + __u32 netmask; + __u32 ip; + int up; + int rc; + + rc = lnet_ipif_query(ifname, &up, &ip, &netmask); + if (rc != 0) { + CERROR("Can't query IPoIB interface %s: %d\n", + ifname, rc); + return NULL; + } - if (!list_empty(&dev->ibd_list)) /* on kib_devs? */ - list_del_init(&dev->ibd_list); + if (!up) { + CERROR("Can't query IPoIB interface %s: it's down\n", ifname); + return NULL; + } - kiblnd_dev_cleanup(dev); + LIBCFS_ALLOC(dev, sizeof(*dev)); + if (dev == NULL) + return NULL; - if (dev->ibd_pd != NULL) - ib_dealloc_pd(dev->ibd_pd); + netdev = dev_get_by_name(&init_net, ifname); + if (netdev == NULL) { + dev->ibd_can_failover = 0; + } else { + dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER); + dev_put(netdev); + } - if (dev->ibd_cmid != NULL) - rdma_destroy_id(dev->ibd_cmid); + INIT_LIST_HEAD(&dev->ibd_nets); + INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */ + INIT_LIST_HEAD(&dev->ibd_fail_list); + dev->ibd_ifip = ip; + strcpy(&dev->ibd_ifname[0], ifname); - LIBCFS_FREE(dev, sizeof(*dev)); + /* initialize the device */ + rc = kiblnd_dev_failover(dev); + if (rc != 0) { + CERROR("Can't initialize device: %d\n", rc); + LIBCFS_FREE(dev, sizeof(*dev)); + return NULL; + } + + list_add_tail(&dev->ibd_list, + &kiblnd_data.kib_devs); + return dev; } -void -kiblnd_base_shutdown (void) +static void +kiblnd_base_shutdown(void) { - int i; + struct kib_sched_info *sched; + int i; - LASSERT (list_empty(&kiblnd_data.kib_devs)); + LASSERT(list_empty(&kiblnd_data.kib_devs)); CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n", - atomic_read(&libcfs_kmemory)); + atomic_read(&libcfs_kmemory)); switch (kiblnd_data.kib_init) { default: @@ -1831,24 +2853,35 @@ kiblnd_base_shutdown (void) case IBLND_INIT_DATA: LASSERT (kiblnd_data.kib_peers != NULL); for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) { - LASSERT (list_empty(&kiblnd_data.kib_peers[i])); - } - LASSERT (list_empty(&kiblnd_data.kib_connd_zombies)); - LASSERT (list_empty(&kiblnd_data.kib_connd_conns)); - - /* flag threads to terminate; wake and wait for them to die */ - kiblnd_data.kib_shutdown = 1; - wake_up_all(&kiblnd_data.kib_sched_waitq); - wake_up_all(&kiblnd_data.kib_connd_waitq); - - i = 2; - while (atomic_read(&kiblnd_data.kib_nthreads) != 0) { - i++; - CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */ - "Waiting for %d threads to terminate\n", - atomic_read(&kiblnd_data.kib_nthreads)); - cfs_pause(cfs_time_seconds(1)); + LASSERT(list_empty(&kiblnd_data.kib_peers[i])); } + LASSERT(list_empty(&kiblnd_data.kib_connd_zombies)); + LASSERT(list_empty(&kiblnd_data.kib_connd_conns)); + LASSERT(list_empty(&kiblnd_data.kib_reconn_list)); + LASSERT(list_empty(&kiblnd_data.kib_reconn_wait)); + + /* flag threads to terminate; wake and wait for them to die */ + kiblnd_data.kib_shutdown = 1; + + /* NB: we really want to stop scheduler threads net by net + * instead of the whole module, this should be improved + * with dynamic configuration LNet */ + cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) + wake_up_all(&sched->ibs_waitq); + + wake_up_all(&kiblnd_data.kib_connd_waitq); + wake_up_all(&kiblnd_data.kib_failover_waitq); + + i = 2; + while (atomic_read(&kiblnd_data.kib_nthreads) != 0) { + i++; + /* power of 2? */ + CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, + "Waiting for %d threads to terminate\n", + atomic_read(&kiblnd_data.kib_nthreads)); + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1)); + } /* fall through */ @@ -1856,23 +2889,27 @@ kiblnd_base_shutdown (void) break; } - if (kiblnd_data.kib_peers != NULL) - LIBCFS_FREE(kiblnd_data.kib_peers, - sizeof(struct list_head) * - kiblnd_data.kib_peer_hash_size); + if (kiblnd_data.kib_peers != NULL) { + LIBCFS_FREE(kiblnd_data.kib_peers, + sizeof(struct list_head) * + kiblnd_data.kib_peer_hash_size); + } + + if (kiblnd_data.kib_scheds != NULL) + cfs_percpt_free(kiblnd_data.kib_scheds); CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n", - atomic_read(&libcfs_kmemory)); + atomic_read(&libcfs_kmemory)); - kiblnd_data.kib_init = IBLND_INIT_NOTHING; - PORTAL_MODULE_UNUSE; + kiblnd_data.kib_init = IBLND_INIT_NOTHING; + module_put(THIS_MODULE); } -void +static void kiblnd_shutdown (lnet_ni_t *ni) { kib_net_t *net = ni->ni_data; - rwlock_t *g_lock = &kiblnd_data.kib_global_lock; + rwlock_t *g_lock = &kiblnd_data.kib_global_lock; int i; unsigned long flags; @@ -1882,11 +2919,11 @@ kiblnd_shutdown (lnet_ni_t *ni) goto out; CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n", - atomic_read(&libcfs_kmemory)); + atomic_read(&libcfs_kmemory)); - write_lock_irqsave(g_lock, flags); - net->ibn_shutdown = 1; - write_unlock_irqrestore(g_lock, flags); + write_lock_irqsave(g_lock, flags); + net->ibn_shutdown = 1; + write_unlock_irqrestore(g_lock, flags); switch (net->ibn_init) { default: @@ -1896,31 +2933,31 @@ kiblnd_shutdown (lnet_ni_t *ni) /* nuke all existing peers within this net */ kiblnd_del_peer(ni, LNET_NID_ANY); - /* Wait for all peer state to clean up */ - i = 2; - while (atomic_read(&net->ibn_npeers) != 0) { - i++; - CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */ - "%s: waiting for %d peers to disconnect\n", - libcfs_nid2str(ni->ni_nid), - atomic_read(&net->ibn_npeers)); - cfs_pause(cfs_time_seconds(1)); - } - - kiblnd_unmap_tx_descs(ni); - - LASSERT (net->ibn_dev->ibd_nnets > 0); - net->ibn_dev->ibd_nnets--; + /* Wait for all peer state to clean up */ + i = 2; + while (atomic_read(&net->ibn_npeers) != 0) { + i++; + /* power of 2? */ + CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, + "%s: waiting for %d peers to disconnect\n", + libcfs_nid2str(ni->ni_nid), + atomic_read(&net->ibn_npeers)); + set_current_state(TASK_UNINTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1)); + } + + kiblnd_net_fini_pools(net); + + write_lock_irqsave(g_lock, flags); + LASSERT(net->ibn_dev->ibd_nnets > 0); + net->ibn_dev->ibd_nnets--; + list_del(&net->ibn_list); + write_unlock_irqrestore(g_lock, flags); /* fall through */ case IBLND_INIT_NOTHING: - LASSERT (atomic_read(&net->ibn_nconns) == 0); - - if (net->ibn_fmrpool != NULL) - ib_destroy_fmr_pool(net->ibn_fmrpool); - if (net->ibn_pmrpool != NULL) - kiblnd_destroy_pmr_pool(net->ibn_pmrpool); + LASSERT (atomic_read(&net->ibn_nconns) == 0); if (net->ibn_dev != NULL && net->ibn_dev->ibd_nnets == 0) @@ -1929,54 +2966,80 @@ kiblnd_shutdown (lnet_ni_t *ni) break; } - kiblnd_free_tx_descs(ni); - CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n", - atomic_read(&libcfs_kmemory)); + atomic_read(&libcfs_kmemory)); net->ibn_init = IBLND_INIT_NOTHING; ni->ni_data = NULL; - + LIBCFS_FREE(net, sizeof(*net)); out: - if (list_empty(&kiblnd_data.kib_devs)) + if (list_empty(&kiblnd_data.kib_devs)) kiblnd_base_shutdown(); return; } -int -kiblnd_base_startup (void) +static int +kiblnd_base_startup(void) { - int i; - int rc; + struct kib_sched_info *sched; + int rc; + int i; - LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING); + LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING); - PORTAL_MODULE_USE; - memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */ + try_module_get(THIS_MODULE); + memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */ - rwlock_init(&kiblnd_data.kib_global_lock); + rwlock_init(&kiblnd_data.kib_global_lock); - INIT_LIST_HEAD(&kiblnd_data.kib_devs); + INIT_LIST_HEAD(&kiblnd_data.kib_devs); + INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs); - kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE; - LIBCFS_ALLOC(kiblnd_data.kib_peers, - sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size); - if (kiblnd_data.kib_peers == NULL) { - goto failed; - } - for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) - INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]); + kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE; + LIBCFS_ALLOC(kiblnd_data.kib_peers, + sizeof(struct list_head) * + kiblnd_data.kib_peer_hash_size); + if (kiblnd_data.kib_peers == NULL) + goto failed; - spin_lock_init(&kiblnd_data.kib_connd_lock); - INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns); - INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies); - init_waitqueue_head(&kiblnd_data.kib_connd_waitq); + for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) + INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]); - spin_lock_init(&kiblnd_data.kib_sched_lock); - INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns); - init_waitqueue_head(&kiblnd_data.kib_sched_waitq); + spin_lock_init(&kiblnd_data.kib_connd_lock); + INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns); + INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies); + INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list); + INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait); + + init_waitqueue_head(&kiblnd_data.kib_connd_waitq); + init_waitqueue_head(&kiblnd_data.kib_failover_waitq); + + kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(), + sizeof(*sched)); + if (kiblnd_data.kib_scheds == NULL) + goto failed; + + cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) { + int nthrs; + + spin_lock_init(&sched->ibs_lock); + INIT_LIST_HEAD(&sched->ibs_conns); + init_waitqueue_head(&sched->ibs_waitq); + + nthrs = cfs_cpt_weight(lnet_cpt_table(), i); + if (*kiblnd_tunables.kib_nscheds > 0) { + nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds); + } else { + /* max to half of CPUs, another half is reserved for + * upper layer modules */ + nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs); + } + + sched->ibs_nthreads_max = nthrs; + sched->ibs_cpt = i; + } kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR; @@ -1984,18 +3047,18 @@ kiblnd_base_startup (void) kiblnd_data.kib_init = IBLND_INIT_DATA; /*****************************************************/ - for (i = 0; i < IBLND_N_SCHED; i++) { - rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i)); - if (rc != 0) { - CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n", - i, rc); - goto failed; - } + rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd"); + if (rc != 0) { + CERROR("Can't spawn o2iblnd connd: %d\n", rc); + goto failed; } - rc = kiblnd_thread_start(kiblnd_connd, NULL); + if (*kiblnd_tunables.kib_dev_failover != 0) + rc = kiblnd_thread_start(kiblnd_failover_thread, NULL, + "kiblnd_failover"); + if (rc != 0) { - CERROR("Can't spawn o2iblnd connd: %d\n", rc); + CERROR("Can't spawn o2iblnd failover thread: %d\n", rc); goto failed; } @@ -2010,17 +3073,119 @@ kiblnd_base_startup (void) return -ENETDOWN; } -int +static int +kiblnd_start_schedulers(struct kib_sched_info *sched) +{ + int rc = 0; + int nthrs; + int i; + + if (sched->ibs_nthreads == 0) { + if (*kiblnd_tunables.kib_nscheds > 0) { + nthrs = sched->ibs_nthreads_max; + } else { + nthrs = cfs_cpt_weight(lnet_cpt_table(), + sched->ibs_cpt); + nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs); + nthrs = min(IBLND_N_SCHED_HIGH, nthrs); + } + } else { + LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max); + /* increase one thread if there is new interface */ + nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max); + } + + for (i = 0; i < nthrs; i++) { + long id; + char name[20]; + id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i); + snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld", + KIB_THREAD_CPT(id), KIB_THREAD_TID(id)); + rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name); + if (rc == 0) + continue; + + CERROR("Can't spawn thread %d for scheduler[%d]: %d\n", + sched->ibs_cpt, sched->ibs_nthreads + i, rc); + break; + } + + sched->ibs_nthreads += i; + return rc; +} + +static int +kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts) +{ + int cpt; + int rc; + int i; + + for (i = 0; i < ncpts; i++) { + struct kib_sched_info *sched; + + cpt = (cpts == NULL) ? i : cpts[i]; + sched = kiblnd_data.kib_scheds[cpt]; + + if (!newdev && sched->ibs_nthreads > 0) + continue; + + rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]); + if (rc != 0) { + CERROR("Failed to start scheduler threads for %s\n", + dev->ibd_ifname); + return rc; + } + } + return 0; +} + +static kib_dev_t * +kiblnd_dev_search(char *ifname) +{ + kib_dev_t *alias = NULL; + kib_dev_t *dev; + char *colon; + char *colon2; + + colon = strchr(ifname, ':'); + list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) { + if (strcmp(&dev->ibd_ifname[0], ifname) == 0) + return dev; + + if (alias != NULL) + continue; + + colon2 = strchr(dev->ibd_ifname, ':'); + if (colon != NULL) + *colon = 0; + if (colon2 != NULL) + *colon2 = 0; + + if (strcmp(&dev->ibd_ifname[0], ifname) == 0) + alias = dev; + + if (colon != NULL) + *colon = ':'; + if (colon2 != NULL) + *colon2 = ':'; + } + return alias; +} + +static int kiblnd_startup (lnet_ni_t *ni) { char *ifname; kib_dev_t *ibdev = NULL; kib_net_t *net; - struct list_head *tmp; struct timeval tv; + unsigned long flags; int rc; + int newdev; + int node_id; - LASSERT (ni->ni_lnd == &the_o2iblnd); + LASSERT (ni->ni_net->net_lnd == &the_o2iblnd); if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) { rc = kiblnd_base_startup(); @@ -2033,18 +3198,10 @@ kiblnd_startup (lnet_ni_t *ni) if (net == NULL) goto failed; - memset(net, 0, sizeof(*net)); - - do_gettimeofday(&tv); - net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec; + do_gettimeofday(&tv); + net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec; - ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout; - ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits; - ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits; - ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits; - - spin_lock_init(&net->ibn_tx_lock); - INIT_LIST_HEAD(&net->ibn_idle_txs); + kiblnd_tunables_setup(ni); if (ni->ni_interfaces[0] != NULL) { /* Use the IPoIB interface specified in 'networks=' */ @@ -2065,130 +3222,44 @@ kiblnd_startup (lnet_ni_t *ni) goto failed; } - list_for_each (tmp, &kiblnd_data.kib_devs) { - ibdev = list_entry(tmp, kib_dev_t, ibd_list); - - if (!strcmp(&ibdev->ibd_ifname[0], ifname)) - break; - - ibdev = NULL; - } - - if (ibdev == NULL) { - __u32 ip; - __u32 netmask; - int up; - struct rdma_cm_id *id; - struct ib_pd *pd; - struct sockaddr_in addr; - - rc = libcfs_ipif_query(ifname, &up, &ip, &netmask); - if (rc != 0) { - CERROR("Can't query IPoIB interface %s: %d\n", - ifname, rc); - goto failed; - } - - if (!up) { - CERROR("Can't query IPoIB interface %s: it's down\n", - ifname); - goto failed; - } - - LIBCFS_ALLOC(ibdev, sizeof(*ibdev)); - if (ibdev == NULL) - goto failed; - - memset(ibdev, 0, sizeof(*ibdev)); - - CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */ - ibdev->ibd_ifip = ip; - strcpy(&ibdev->ibd_ifname[0], ifname); - - id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP); - if (IS_ERR(id)) { - CERROR("Can't create listen ID: %ld\n", PTR_ERR(id)); - goto failed; - } - - ibdev->ibd_cmid = id; + ibdev = kiblnd_dev_search(ifname); - memset(&addr, 0, sizeof(addr)); - addr.sin_family = AF_INET; - addr.sin_port = htons(*kiblnd_tunables.kib_service); - addr.sin_addr.s_addr = htonl(ip); + newdev = ibdev == NULL; + /* hmm...create kib_dev even for alias */ + if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) + ibdev = kiblnd_create_dev(ifname); - rc = rdma_bind_addr(id, (struct sockaddr *)&addr); - if (rc != 0) { - CERROR("Can't bind to %s: %d\n", ifname, rc); - goto failed; - } + if (ibdev == NULL) + goto failed; - /* Binding should have assigned me an IB device */ - LASSERT (id->device != NULL); - CDEBUG(D_CONSOLE, "Listener bound to %s:%u.%u.%u.%u:%d:%s\n", - ifname, HIPQUAD(ip), *kiblnd_tunables.kib_service, - id->device->name); + node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device); + ni->dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id); - pd = ib_alloc_pd(id->device); - if (IS_ERR(pd)) { - CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd)); - goto failed; - } - - ibdev->ibd_pd = pd; + net->ibn_dev = ibdev; + ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip); - rc = rdma_listen(id, 256); - if (rc != 0) { - CERROR("Can't start listener: %d\n", rc); - goto failed; - } + rc = kiblnd_dev_start_threads(ibdev, newdev, + ni->ni_cpts, ni->ni_ncpts); + if (rc != 0) + goto failed; - rc = kiblnd_dev_setup(ibdev); - if (rc != 0) { - CERROR("Can't setup device: %d\n", rc); - goto failed; - } - } - - ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip); - net->ibn_dev = ibdev; - - if (*kiblnd_tunables.kib_map_on_demand > 0 || - ibdev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1, - * so we always create FMR/PMR pool and - * map-on-demand if premapping failed */ - /* Map on demand */ - rc = kiblnd_ib_create_fmr_pool(ibdev, &net->ibn_fmrpool); - if (rc == -ENOSYS) { - CDEBUG(D_CONSOLE, "No FMR, creating physical mapping\n"); - - rc = kiblnd_create_pmr_pool(ibdev, &net->ibn_pmrpool); - } - - if (rc != 0) { - CERROR("Can't create FMR or physical mapping pool: %d, " - "please disable map_on_demand and retry\n", rc); - goto failed; - } - - } - - rc = kiblnd_alloc_tx_descs(ni); + rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts); if (rc != 0) { - CERROR("Can't allocate tx descs\n"); + CERROR("Failed to initialize NI pools: %d\n", rc); goto failed; } - kiblnd_map_tx_descs(ni); + write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + ibdev->ibd_nnets++; + list_add_tail(&net->ibn_list, &ibdev->ibd_nets); + write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); - ibdev->ibd_nnets++; net->ibn_init = IBLND_INIT_ALL; return 0; failed: - if (net->ibn_dev == NULL && ibdev != NULL) + if (net != NULL && net->ibn_dev == NULL && ibdev != NULL) kiblnd_destroy_dev(ibdev); kiblnd_shutdown(ni); @@ -2197,36 +3268,46 @@ failed: return -ENETDOWN; } -void __exit -kiblnd_module_fini (void) +static lnd_t the_o2iblnd = { + .lnd_type = O2IBLND, + .lnd_startup = kiblnd_startup, + .lnd_shutdown = kiblnd_shutdown, + .lnd_ctl = kiblnd_ctl, + .lnd_query = kiblnd_query, + .lnd_send = kiblnd_send, + .lnd_recv = kiblnd_recv, +}; + +static void __exit ko2iblnd_exit(void) { - lnet_unregister_lnd(&the_o2iblnd); - kiblnd_tunables_fini(); + lnet_unregister_lnd(&the_o2iblnd); } -int __init -kiblnd_module_init (void) +static int __init ko2iblnd_init(void) { - int rc; + int rc; - CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE); - CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) - <= IBLND_MSG_SIZE); - CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) - <= IBLND_MSG_SIZE); + CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE); + CLASSERT(offsetof(kib_msg_t, + ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <= + IBLND_MSG_SIZE); + CLASSERT(offsetof(kib_msg_t, + ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) + <= IBLND_MSG_SIZE); - rc = kiblnd_tunables_init(); - if (rc != 0) - return rc; + rc = kiblnd_tunables_init(); + if (rc != 0) + return rc; - lnet_register_lnd(&the_o2iblnd); + lnet_register_lnd(&the_o2iblnd); - return 0; + return 0; } -MODULE_AUTHOR("Sun Microsystems, Inc. "); -MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00"); +MODULE_AUTHOR("OpenSFS, Inc. "); +MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver"); +MODULE_VERSION("2.8.0"); MODULE_LICENSE("GPL"); -module_init(kiblnd_module_init); -module_exit(kiblnd_module_fini); +module_init(ko2iblnd_init); +module_exit(ko2iblnd_exit);