4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lnet/klnds/o2iblnd/o2iblnd.c
33 * Author: Eric Barton <eric@bartonsoftware.com>
37 #include <linux/inetdevice.h>
41 static const struct lnet_lnd the_o2iblnd;
43 struct kib_data kiblnd_data;
46 kiblnd_cksum (void *ptr, int nob)
52 sum = ((sum << 1) | (sum >> 31)) + *c++;
54 /* ensure I don't return 0 (== no checksum) */
55 return (sum == 0) ? 1 : sum;
59 kiblnd_msgtype2str(int type)
62 case IBLND_MSG_CONNREQ:
65 case IBLND_MSG_CONNACK:
71 case IBLND_MSG_IMMEDIATE:
74 case IBLND_MSG_PUT_REQ:
77 case IBLND_MSG_PUT_NAK:
80 case IBLND_MSG_PUT_ACK:
83 case IBLND_MSG_PUT_DONE:
86 case IBLND_MSG_GET_REQ:
89 case IBLND_MSG_GET_DONE:
98 kiblnd_msgtype2size(int type)
100 const int hdr_size = offsetof(struct kib_msg, ibm_u);
103 case IBLND_MSG_CONNREQ:
104 case IBLND_MSG_CONNACK:
105 return hdr_size + sizeof(struct kib_connparams);
110 case IBLND_MSG_IMMEDIATE:
111 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
113 case IBLND_MSG_PUT_REQ:
114 return hdr_size + sizeof(struct kib_putreq_msg);
116 case IBLND_MSG_PUT_ACK:
117 return hdr_size + sizeof(struct kib_putack_msg);
119 case IBLND_MSG_GET_REQ:
120 return hdr_size + sizeof(struct kib_get_msg);
122 case IBLND_MSG_PUT_NAK:
123 case IBLND_MSG_PUT_DONE:
124 case IBLND_MSG_GET_DONE:
125 return hdr_size + sizeof(struct kib_completion_msg);
131 static int kiblnd_unpack_rd(struct kib_msg *msg, bool flip)
133 struct kib_rdma_desc *rd;
138 LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
139 msg->ibm_type == IBLND_MSG_PUT_ACK);
141 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
142 &msg->ibm_u.get.ibgm_rd :
143 &msg->ibm_u.putack.ibpam_rd;
146 __swab32s(&rd->rd_key);
147 __swab32s(&rd->rd_nfrags);
152 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
153 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
154 n, IBLND_MAX_RDMA_FRAGS);
158 nob = offsetof(struct kib_msg, ibm_u) +
159 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
161 if (msg->ibm_nob < nob) {
162 CERROR("Short %s: %d(%d)\n",
163 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
170 for (i = 0; i < n; i++) {
171 __swab32s(&rd->rd_frags[i].rf_nob);
172 __swab64s(&rd->rd_frags[i].rf_addr);
178 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
179 int credits, lnet_nid_t dstnid, __u64 dststamp)
181 struct kib_net *net = ni->ni_data;
183 /* CAVEAT EMPTOR! all message fields not set here should have been
184 * initialised previously. */
185 msg->ibm_magic = IBLND_MSG_MAGIC;
186 msg->ibm_version = version;
188 msg->ibm_credits = credits;
191 msg->ibm_srcnid = ni->ni_nid;
192 msg->ibm_srcstamp = net->ibn_incarnation;
193 msg->ibm_dstnid = dstnid;
194 msg->ibm_dststamp = dststamp;
196 if (*kiblnd_tunables.kib_cksum) {
197 /* NB ibm_cksum zero while computing cksum */
198 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
202 int kiblnd_unpack_msg(struct kib_msg *msg, int nob)
204 const int hdr_size = offsetof(struct kib_msg, ibm_u);
210 /* 6 bytes are enough to have received magic + version */
212 CERROR("Short message: %d\n", nob);
216 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
218 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
221 CERROR("Bad magic: %08x\n", msg->ibm_magic);
225 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
226 if (version != IBLND_MSG_VERSION &&
227 version != IBLND_MSG_VERSION_1) {
228 CERROR("Bad version: %x\n", version);
232 if (nob < hdr_size) {
233 CERROR("Short message: %d\n", nob);
237 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
239 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
243 /* checksum must be computed with ibm_cksum zero and BEFORE anything
246 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
248 if (msg_cksum != 0 &&
249 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
250 CERROR("Bad checksum\n");
254 msg->ibm_cksum = msg_cksum;
257 /* leave magic unflipped as a clue to peer_ni endianness */
258 msg->ibm_version = version;
259 BUILD_BUG_ON(sizeof(msg->ibm_type) != 1);
260 BUILD_BUG_ON(sizeof(msg->ibm_credits) != 1);
261 msg->ibm_nob = msg_nob;
262 __swab64s(&msg->ibm_srcnid);
263 __swab64s(&msg->ibm_srcstamp);
264 __swab64s(&msg->ibm_dstnid);
265 __swab64s(&msg->ibm_dststamp);
268 if (msg->ibm_srcnid == LNET_NID_ANY) {
269 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
273 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
274 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
275 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
279 switch (msg->ibm_type) {
281 CERROR("Unknown message type %x\n", msg->ibm_type);
285 case IBLND_MSG_IMMEDIATE:
286 case IBLND_MSG_PUT_REQ:
289 case IBLND_MSG_PUT_ACK:
290 case IBLND_MSG_GET_REQ:
291 if (kiblnd_unpack_rd(msg, flip))
295 case IBLND_MSG_PUT_NAK:
296 case IBLND_MSG_PUT_DONE:
297 case IBLND_MSG_GET_DONE:
299 __swab32s(&msg->ibm_u.completion.ibcm_status);
302 case IBLND_MSG_CONNREQ:
303 case IBLND_MSG_CONNACK:
305 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
306 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
307 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
315 kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
318 struct kib_peer_ni *peer_ni;
319 struct kib_net *net = ni->ni_data;
320 int cpt = lnet_cpt_of_nid(nid, ni);
323 LASSERT(net != NULL);
324 LASSERT(nid != LNET_NID_ANY);
326 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
328 CERROR("Cannot allocate peer_ni\n");
332 peer_ni->ibp_ni = ni;
333 peer_ni->ibp_nid = nid;
334 peer_ni->ibp_error = 0;
335 peer_ni->ibp_last_alive = 0;
336 peer_ni->ibp_max_frags = IBLND_MAX_RDMA_FRAGS;
337 peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
338 peer_ni->ibp_queue_depth_mod = 0; /* try to use the default */
339 atomic_set(&peer_ni->ibp_refcount, 1); /* 1 ref for caller */
341 INIT_HLIST_NODE(&peer_ni->ibp_list);
342 INIT_LIST_HEAD(&peer_ni->ibp_conns);
343 INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
345 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
347 /* always called with a ref on ni, which prevents ni being shutdown */
348 LASSERT(net->ibn_shutdown == 0);
350 /* npeers only grows with the global lock held */
351 atomic_inc(&net->ibn_npeers);
353 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
360 kiblnd_destroy_peer(struct kib_peer_ni *peer_ni)
362 struct kib_net *net = peer_ni->ibp_ni->ni_data;
364 LASSERT(net != NULL);
365 LASSERT (atomic_read(&peer_ni->ibp_refcount) == 0);
366 LASSERT(!kiblnd_peer_active(peer_ni));
367 LASSERT(kiblnd_peer_idle(peer_ni));
368 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
370 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
372 /* NB a peer_ni's connections keep a reference on their peer_ni until
373 * they are destroyed, so we can be assured that _all_ state to do
374 * with this peer_ni has been cleaned up when its refcount drops to
377 if (atomic_dec_and_test(&net->ibn_npeers))
378 wake_up_var(&net->ibn_npeers);
382 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
384 /* the caller is responsible for accounting the additional reference
387 struct kib_peer_ni *peer_ni;
389 hash_for_each_possible(kiblnd_data.kib_peers, peer_ni,
391 LASSERT(!kiblnd_peer_idle(peer_ni));
394 * Match a peer if its NID and the NID of the local NI it
395 * communicates over are the same. Otherwise don't match
396 * the peer, which will result in a new lnd peer being
399 if (peer_ni->ibp_nid != nid ||
400 peer_ni->ibp_ni->ni_nid != ni->ni_nid)
403 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
404 peer_ni, libcfs_nid2str(nid),
405 atomic_read(&peer_ni->ibp_refcount),
406 peer_ni->ibp_version);
413 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
415 LASSERT(list_empty(&peer_ni->ibp_conns));
417 LASSERT(kiblnd_peer_active(peer_ni));
418 hlist_del_init(&peer_ni->ibp_list);
419 /* lose peerlist's ref */
420 kiblnd_peer_decref(peer_ni);
424 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
425 lnet_nid_t *nidp, int *count)
427 struct kib_peer_ni *peer_ni;
431 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
433 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
434 LASSERT(!kiblnd_peer_idle(peer_ni));
436 if (peer_ni->ibp_ni != ni)
442 *nidp = peer_ni->ibp_nid;
443 *count = atomic_read(&peer_ni->ibp_refcount);
445 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
449 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
454 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
456 struct kib_conn *cnxt;
457 struct kib_conn *conn;
459 if (list_empty(&peer_ni->ibp_conns)) {
460 kiblnd_unlink_peer_locked(peer_ni);
462 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
464 kiblnd_close_conn_locked(conn, 0);
465 /* NB closing peer_ni's last conn unlinked it. */
467 /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
472 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
475 struct hlist_node *pnxt;
476 struct kib_peer_ni *peer_ni;
483 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
485 if (nid != LNET_NID_ANY) {
486 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
490 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
493 for (i = lo; i <= hi; i++) {
494 hlist_for_each_entry_safe(peer_ni, pnxt,
495 &kiblnd_data.kib_peers[i], ibp_list) {
496 LASSERT(!kiblnd_peer_idle(peer_ni));
498 if (peer_ni->ibp_ni != ni)
501 if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
504 if (!list_empty(&peer_ni->ibp_tx_queue)) {
505 LASSERT(list_empty(&peer_ni->ibp_conns));
507 list_splice_init(&peer_ni->ibp_tx_queue,
511 kiblnd_del_peer_locked(peer_ni);
512 rc = 0; /* matched something */
516 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
518 kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
523 static struct kib_conn *
524 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
526 struct kib_peer_ni *peer_ni;
527 struct kib_conn *conn;
531 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
533 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
534 LASSERT(!kiblnd_peer_idle(peer_ni));
536 if (peer_ni->ibp_ni != ni)
539 list_for_each_entry(conn, &peer_ni->ibp_conns,
544 kiblnd_conn_addref(conn);
545 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
551 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
556 kiblnd_debug_rx(struct kib_rx *rx)
558 CDEBUG(D_CONSOLE, " %p msg_type %x cred %d\n",
559 rx, rx->rx_msg->ibm_type,
560 rx->rx_msg->ibm_credits);
564 kiblnd_debug_tx(struct kib_tx *tx)
566 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lld "
567 "cookie %#llx msg %s%s type %x cred %d\n",
568 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
569 tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
570 tx->tx_lntmsg[0] == NULL ? "-" : "!",
571 tx->tx_lntmsg[1] == NULL ? "-" : "!",
572 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
576 kiblnd_debug_conn(struct kib_conn *conn)
578 struct list_head *tmp;
581 spin_lock(&conn->ibc_lock);
583 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
584 atomic_read(&conn->ibc_refcount), conn,
585 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
586 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d "
587 " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
588 conn->ibc_nsends_posted, conn->ibc_credits,
589 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
590 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
592 CDEBUG(D_CONSOLE, " early_rxs:\n");
593 list_for_each(tmp, &conn->ibc_early_rxs)
594 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
596 CDEBUG(D_CONSOLE, " tx_noops:\n");
597 list_for_each(tmp, &conn->ibc_tx_noops)
598 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
600 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
601 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
602 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
604 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
605 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
606 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
608 CDEBUG(D_CONSOLE, " tx_queue:\n");
609 list_for_each(tmp, &conn->ibc_tx_queue)
610 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
612 CDEBUG(D_CONSOLE, " active_txs:\n");
613 list_for_each(tmp, &conn->ibc_active_txs)
614 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
616 CDEBUG(D_CONSOLE, " rxs:\n");
617 for (i = 0; i < IBLND_RX_MSGS(conn); i++)
618 kiblnd_debug_rx(&conn->ibc_rxs[i]);
620 spin_unlock(&conn->ibc_lock);
624 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
626 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
627 if (cmid->route.path_rec == NULL)
630 if (*kiblnd_tunables.kib_ib_mtu)
631 cmid->route.path_rec->mtu =
632 ib_mtu_int_to_enum(*kiblnd_tunables.kib_ib_mtu);
636 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
644 vectors = conn->ibc_cmid->device->num_comp_vectors;
648 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
650 /* hash NID to CPU id in this partition... */
651 ibp_nid = conn->ibc_peer->ibp_nid;
652 off = do_div(ibp_nid, cpumask_weight(*mask));
653 for_each_cpu(i, *mask) {
663 * Get the scheduler bound to this CPT. If the scheduler has no
664 * threads, which means that the CPT has no CPUs, then grab the
665 * next scheduler that we can use.
667 * This case would be triggered if a NUMA node is configured with
668 * no associated CPUs.
670 static struct kib_sched_info *
671 kiblnd_get_scheduler(int cpt)
673 struct kib_sched_info *sched;
676 sched = kiblnd_data.kib_scheds[cpt];
678 if (sched->ibs_nthreads > 0)
681 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
682 if (sched->ibs_nthreads > 0) {
683 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
684 cpt, sched->ibs_cpt);
692 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
695 * One WR for the LNet message
696 * And ibc_max_frags for the transfer WRs
699 int multiplier = 1 + conn->ibc_max_frags;
700 enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
702 /* FastReg needs two extra WRs for map and invalidate */
703 if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
706 /* account for a maximum of ibc_queue_depth in-flight transfers */
707 ret = multiplier * conn->ibc_queue_depth;
709 if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
710 CDEBUG(D_NET, "peer_credits %u will result in send work "
711 "request size %d larger than maximum %d device "
712 "can handle\n", conn->ibc_queue_depth, ret,
713 conn->ibc_hdev->ibh_max_qp_wr);
714 conn->ibc_queue_depth =
715 conn->ibc_hdev->ibh_max_qp_wr / multiplier;
718 /* don't go beyond the maximum the device can handle */
719 return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
723 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
724 int state, int version)
727 * If the new conn is created successfully it takes over the caller's
728 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
729 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
730 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
731 * to destroy 'cmid' here since I'm called from the CM which still has
732 * its ref on 'cmid'). */
733 rwlock_t *glock = &kiblnd_data.kib_global_lock;
734 struct kib_net *net = peer_ni->ibp_ni->ni_data;
736 struct ib_qp_init_attr init_qp_attr = {};
737 struct kib_sched_info *sched;
738 #ifdef HAVE_IB_CQ_INIT_ATTR
739 struct ib_cq_init_attr cq_attr = {};
741 struct kib_conn *conn;
748 LASSERT(net != NULL);
749 LASSERT(!in_interrupt());
753 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
754 sched = kiblnd_get_scheduler(cpt);
757 CERROR("no schedulers available. node is unhealthy\n");
762 * The cpt might have changed if we ended up selecting a non cpt
763 * native scheduler. So use the scheduler's cpt instead.
765 cpt = sched->ibs_cpt;
767 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
769 CERROR("Can't allocate connection for %s\n",
770 libcfs_nid2str(peer_ni->ibp_nid));
774 conn->ibc_state = IBLND_CONN_INIT;
775 conn->ibc_version = version;
776 conn->ibc_peer = peer_ni; /* I take the caller's ref */
777 cmid->context = conn; /* for future CM callbacks */
778 conn->ibc_cmid = cmid;
779 conn->ibc_max_frags = peer_ni->ibp_max_frags;
780 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
781 conn->ibc_rxs = NULL;
782 conn->ibc_rx_pages = NULL;
784 INIT_LIST_HEAD(&conn->ibc_early_rxs);
785 INIT_LIST_HEAD(&conn->ibc_tx_noops);
786 INIT_LIST_HEAD(&conn->ibc_tx_queue);
787 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
788 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
789 INIT_LIST_HEAD(&conn->ibc_active_txs);
790 INIT_LIST_HEAD(&conn->ibc_zombie_txs);
791 spin_lock_init(&conn->ibc_lock);
793 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
794 sizeof(*conn->ibc_connvars));
795 if (conn->ibc_connvars == NULL) {
796 CERROR("Can't allocate in-progress connection state\n");
800 write_lock_irqsave(glock, flags);
801 if (dev->ibd_failover) {
802 write_unlock_irqrestore(glock, flags);
803 CERROR("%s: failover in progress\n", dev->ibd_ifname);
807 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
808 /* wakeup failover thread and teardown connection */
809 if (kiblnd_dev_can_failover(dev)) {
810 list_add_tail(&dev->ibd_fail_list,
811 &kiblnd_data.kib_failed_devs);
812 wake_up(&kiblnd_data.kib_failover_waitq);
815 write_unlock_irqrestore(glock, flags);
816 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
817 cmid->device->name, dev->ibd_ifname);
821 kiblnd_hdev_addref_locked(dev->ibd_hdev);
822 conn->ibc_hdev = dev->ibd_hdev;
824 kiblnd_setup_mtu_locked(cmid);
826 write_unlock_irqrestore(glock, flags);
828 #ifdef HAVE_IB_CQ_INIT_ATTR
829 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
830 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
831 cq = ib_create_cq(cmid->device,
832 kiblnd_cq_completion, kiblnd_cq_event, conn,
835 cq = ib_create_cq(cmid->device,
836 kiblnd_cq_completion, kiblnd_cq_event, conn,
837 IBLND_CQ_ENTRIES(conn),
838 kiblnd_get_completion_vector(conn, cpt));
842 * on MLX-5 (possibly MLX-4 as well) this error could be
843 * hit if the concurrent_sends and/or peer_tx_credits is set
844 * too high. Or due to an MLX-5 bug which tries to
845 * allocate 256kb via kmalloc for WR cookie array
847 CERROR("Failed to create CQ with %d CQEs: %ld\n",
848 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
854 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
856 CERROR("Can't request completion notification: %d\n", rc);
860 init_qp_attr.event_handler = kiblnd_qp_event;
861 init_qp_attr.qp_context = conn;
862 init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
863 init_qp_attr.cap.max_recv_sge = 1;
864 init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
865 init_qp_attr.qp_type = IB_QPT_RC;
866 init_qp_attr.send_cq = cq;
867 init_qp_attr.recv_cq = cq;
869 if (peer_ni->ibp_queue_depth_mod &&
870 peer_ni->ibp_queue_depth_mod < peer_ni->ibp_queue_depth) {
871 conn->ibc_queue_depth = peer_ni->ibp_queue_depth_mod;
872 CDEBUG(D_NET, "Use reduced queue depth %u (from %u)\n",
873 peer_ni->ibp_queue_depth_mod,
874 peer_ni->ibp_queue_depth);
878 /* kiblnd_send_wrs() can change the connection's queue depth if
879 * the maximum work requests for the device is maxed out
881 init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
882 init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
883 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd,
885 if (rc != -ENOMEM || conn->ibc_queue_depth < 2)
887 conn->ibc_queue_depth--;
891 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
892 "send_sge: %d, recv_sge: %d\n",
893 rc, init_qp_attr.cap.max_send_wr,
894 init_qp_attr.cap.max_recv_wr,
895 init_qp_attr.cap.max_send_sge,
896 init_qp_attr.cap.max_recv_sge);
900 conn->ibc_sched = sched;
902 if (!peer_ni->ibp_queue_depth_mod &&
903 conn->ibc_queue_depth != peer_ni->ibp_queue_depth) {
904 CWARN("peer %s - queue depth reduced from %u to %u"
905 " to allow for qp creation\n",
906 libcfs_nid2str(peer_ni->ibp_nid),
907 peer_ni->ibp_queue_depth,
908 conn->ibc_queue_depth);
909 peer_ni->ibp_queue_depth_mod = conn->ibc_queue_depth;
912 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
913 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
914 if (conn->ibc_rxs == NULL) {
915 CERROR("Cannot allocate RX buffers\n");
919 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
920 IBLND_RX_MSG_PAGES(conn));
924 kiblnd_map_rx_descs(conn);
926 /* 1 ref for caller and each rxmsg */
927 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
928 conn->ibc_nrx = IBLND_RX_MSGS(conn);
931 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
932 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
934 CERROR("Can't post rxmsg: %d\n", rc);
936 /* Make posted receives complete */
937 kiblnd_abort_receives(conn);
939 /* correct # of posted buffers
940 * NB locking needed now I'm racing with completion */
941 spin_lock_irqsave(&sched->ibs_lock, flags);
942 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
943 spin_unlock_irqrestore(&sched->ibs_lock, flags);
945 /* cmid will be destroyed by CM(ofed) after cm_callback
946 * returned, so we can't refer it anymore
947 * (by kiblnd_connd()->kiblnd_destroy_conn) */
948 rdma_destroy_qp(conn->ibc_cmid);
949 conn->ibc_cmid = NULL;
951 /* Drop my own and unused rxbuffer refcounts */
952 while (i++ <= IBLND_RX_MSGS(conn))
953 kiblnd_conn_decref(conn);
959 /* Init successful! */
960 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
961 state == IBLND_CONN_PASSIVE_WAIT);
962 conn->ibc_state = state;
965 atomic_inc(&net->ibn_nconns);
969 kiblnd_destroy_conn(conn);
970 LIBCFS_FREE(conn, sizeof(*conn));
976 kiblnd_destroy_conn(struct kib_conn *conn)
978 struct rdma_cm_id *cmid = conn->ibc_cmid;
979 struct kib_peer_ni *peer_ni = conn->ibc_peer;
981 LASSERT (!in_interrupt());
982 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
983 LASSERT(list_empty(&conn->ibc_early_rxs));
984 LASSERT(list_empty(&conn->ibc_tx_noops));
985 LASSERT(list_empty(&conn->ibc_tx_queue));
986 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
987 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
988 LASSERT(list_empty(&conn->ibc_active_txs));
989 LASSERT (conn->ibc_noops_posted == 0);
990 LASSERT (conn->ibc_nsends_posted == 0);
992 switch (conn->ibc_state) {
994 /* conn must be completely disengaged from the network */
997 case IBLND_CONN_DISCONNECTED:
998 /* connvars should have been freed already */
999 LASSERT (conn->ibc_connvars == NULL);
1002 case IBLND_CONN_INIT:
1006 /* conn->ibc_cmid might be destroyed by CM already */
1007 if (cmid != NULL && cmid->qp != NULL)
1008 rdma_destroy_qp(cmid);
1011 ib_destroy_cq(conn->ibc_cq);
1013 kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1014 LNET_MSG_STATUS_OK);
1016 if (conn->ibc_rx_pages != NULL)
1017 kiblnd_unmap_rx_descs(conn);
1019 if (conn->ibc_rxs != NULL)
1020 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1022 if (conn->ibc_connvars != NULL)
1023 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1025 if (conn->ibc_hdev != NULL)
1026 kiblnd_hdev_decref(conn->ibc_hdev);
1028 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1029 if (conn->ibc_state != IBLND_CONN_INIT) {
1030 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1032 kiblnd_peer_decref(peer_ni);
1033 rdma_destroy_id(cmid);
1034 atomic_dec(&net->ibn_nconns);
1039 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1041 struct kib_conn *conn;
1042 struct kib_conn *cnxt;
1045 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1047 CDEBUG(D_NET, "Closing conn -> %s, "
1048 "version: %x, reason: %d\n",
1049 libcfs_nid2str(peer_ni->ibp_nid),
1050 conn->ibc_version, why);
1052 kiblnd_close_conn_locked(conn, why);
1060 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1061 int version, __u64 incarnation)
1063 struct kib_conn *conn;
1064 struct kib_conn *cnxt;
1067 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1069 if (conn->ibc_version == version &&
1070 conn->ibc_incarnation == incarnation)
1073 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1074 "incarnation:%#llx(%x, %#llx)\n",
1075 libcfs_nid2str(peer_ni->ibp_nid),
1076 conn->ibc_version, conn->ibc_incarnation,
1077 version, incarnation);
1079 kiblnd_close_conn_locked(conn, -ESTALE);
1087 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1089 struct kib_peer_ni *peer_ni;
1090 struct hlist_node *pnxt;
1094 unsigned long flags;
1097 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1099 if (nid != LNET_NID_ANY) {
1100 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
1104 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
1107 for (i = lo; i <= hi; i++) {
1108 hlist_for_each_entry_safe(peer_ni, pnxt,
1109 &kiblnd_data.kib_peers[i], ibp_list) {
1110 LASSERT(!kiblnd_peer_idle(peer_ni));
1112 if (peer_ni->ibp_ni != ni)
1115 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1118 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1122 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1124 /* wildcards always succeed */
1125 if (nid == LNET_NID_ANY)
1128 return (count == 0) ? -ENOENT : 0;
1132 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1134 struct libcfs_ioctl_data *data = arg;
1138 case IOC_LIBCFS_GET_PEER: {
1142 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1144 data->ioc_nid = nid;
1145 data->ioc_count = count;
1149 case IOC_LIBCFS_DEL_PEER: {
1150 rc = kiblnd_del_peer(ni, data->ioc_nid);
1153 case IOC_LIBCFS_GET_CONN: {
1154 struct kib_conn *conn;
1157 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1163 LASSERT(conn->ibc_cmid != NULL);
1164 data->ioc_nid = conn->ibc_peer->ibp_nid;
1165 if (conn->ibc_cmid->route.path_rec == NULL)
1166 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1169 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1170 kiblnd_conn_decref(conn);
1173 case IOC_LIBCFS_CLOSE_CONNECTION: {
1174 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1186 kiblnd_free_pages(struct kib_pages *p)
1188 int npages = p->ibp_npages;
1191 for (i = 0; i < npages; i++) {
1192 if (p->ibp_pages[i] != NULL)
1193 __free_page(p->ibp_pages[i]);
1196 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1200 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1202 struct kib_pages *p;
1205 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1206 offsetof(struct kib_pages, ibp_pages[npages]));
1208 CERROR("Can't allocate descriptor for %d pages\n", npages);
1212 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1213 p->ibp_npages = npages;
1215 for (i = 0; i < npages; i++) {
1216 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1218 if (p->ibp_pages[i] == NULL) {
1219 CERROR("Can't allocate page %d of %d\n", i, npages);
1220 kiblnd_free_pages(p);
1230 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1235 LASSERT (conn->ibc_rxs != NULL);
1236 LASSERT (conn->ibc_hdev != NULL);
1238 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1239 rx = &conn->ibc_rxs[i];
1241 LASSERT(rx->rx_nob >= 0); /* not posted */
1243 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1244 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1246 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1249 kiblnd_free_pages(conn->ibc_rx_pages);
1251 conn->ibc_rx_pages = NULL;
1255 kiblnd_map_rx_descs(struct kib_conn *conn)
1263 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1264 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1265 rx = &conn->ibc_rxs[i];
1268 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1271 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1272 rx->rx_msg, IBLND_MSG_SIZE,
1274 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1276 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1278 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1279 i, rx->rx_msg, rx->rx_msgaddr,
1280 (__u64)(page_to_phys(pg) + pg_off));
1282 pg_off += IBLND_MSG_SIZE;
1283 LASSERT(pg_off <= PAGE_SIZE);
1285 if (pg_off == PAGE_SIZE) {
1288 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1294 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1296 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1300 LASSERT (tpo->tpo_pool.po_allocated == 0);
1305 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1306 tx = &tpo->tpo_tx_descs[i];
1307 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1308 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1310 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1313 kiblnd_hdev_decref(hdev);
1314 tpo->tpo_hdev = NULL;
1317 static struct kib_hca_dev *
1318 kiblnd_current_hdev(struct kib_dev *dev)
1320 struct kib_hca_dev *hdev;
1321 unsigned long flags;
1324 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1325 while (dev->ibd_failover) {
1326 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1328 CDEBUG(D_NET, "%s: Wait for failover\n",
1330 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1332 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1335 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1336 hdev = dev->ibd_hdev;
1338 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1344 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1346 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1347 struct kib_pool *pool = &tpo->tpo_pool;
1348 struct kib_net *net = pool->po_owner->ps_net;
1349 struct kib_dev *dev;
1356 LASSERT (net != NULL);
1360 /* pre-mapped messages are not bigger than 1 page */
1361 BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1363 /* No fancy arithmetic when we do the buffer calculations */
1364 BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1366 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1368 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1369 page = txpgs->ibp_pages[ipage];
1370 tx = &tpo->tpo_tx_descs[i];
1372 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1375 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1379 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1381 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1383 list_add(&tx->tx_list, &pool->po_free_list);
1385 page_offset += IBLND_MSG_SIZE;
1386 LASSERT(page_offset <= PAGE_SIZE);
1388 if (page_offset == PAGE_SIZE) {
1391 LASSERT(ipage <= txpgs->ibp_npages);
1397 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1399 LASSERT(fpo->fpo_map_count == 0);
1401 #ifdef HAVE_FMR_POOL_API
1402 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1403 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1405 #endif /* HAVE_FMR_POOL_API */
1407 struct kib_fast_reg_descriptor *frd, *tmp;
1410 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1412 list_del(&frd->frd_list);
1413 #ifndef HAVE_IB_MAP_MR_SG
1414 ib_free_fast_reg_page_list(frd->frd_frpl);
1416 ib_dereg_mr(frd->frd_mr);
1417 LIBCFS_FREE(frd, sizeof(*frd));
1420 if (i < fpo->fast_reg.fpo_pool_size)
1421 CERROR("FastReg pool still has %d regions registered\n",
1422 fpo->fast_reg.fpo_pool_size - i);
1426 kiblnd_hdev_decref(fpo->fpo_hdev);
1428 LIBCFS_FREE(fpo, sizeof(*fpo));
1432 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1434 struct kib_fmr_pool *fpo, *tmp;
1436 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1437 list_del(&fpo->fpo_list);
1438 kiblnd_destroy_fmr_pool(fpo);
1443 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1446 int size = tunables->lnd_fmr_pool_size / ncpts;
1448 return max(IBLND_FMR_POOL, size);
1452 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1455 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1457 return max(IBLND_FMR_POOL_FLUSH, size);
1460 #ifdef HAVE_FMR_POOL_API
1461 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1462 struct kib_fmr_pool *fpo)
1464 struct ib_fmr_pool_param param = {
1465 .max_pages_per_fmr = LNET_MAX_IOV,
1466 .page_shift = PAGE_SHIFT,
1467 .access = (IB_ACCESS_LOCAL_WRITE |
1468 IB_ACCESS_REMOTE_WRITE),
1469 .pool_size = fps->fps_pool_size,
1470 .dirty_watermark = fps->fps_flush_trigger,
1471 .flush_function = NULL,
1473 .cache = !!fps->fps_cache };
1476 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1478 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1479 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1481 CERROR("Failed to create FMR pool: %d\n", rc);
1483 CERROR("FMRs are not supported\n");
1485 fpo->fpo_is_fmr = true;
1489 #endif /* HAVE_FMR_POOL_API */
1491 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1492 struct kib_fmr_pool *fpo,
1493 enum kib_dev_caps dev_caps)
1495 struct kib_fast_reg_descriptor *frd, *tmp;
1498 #ifdef HAVE_FMR_POOL_API
1499 fpo->fpo_is_fmr = false;
1502 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1503 fpo->fast_reg.fpo_pool_size = 0;
1504 for (i = 0; i < fps->fps_pool_size; i++) {
1505 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1508 CERROR("Failed to allocate a new fast_reg descriptor\n");
1514 #ifndef HAVE_IB_MAP_MR_SG
1515 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1517 if (IS_ERR(frd->frd_frpl)) {
1518 rc = PTR_ERR(frd->frd_frpl);
1519 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1521 frd->frd_frpl = NULL;
1526 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1527 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1531 * it is expected to get here if this is an MLX-5 card.
1532 * MLX-4 cards will always use FMR and MLX-5 cards will
1533 * always use fast_reg. It turns out that some MLX-5 cards
1534 * (possibly due to older FW versions) do not natively support
1535 * gaps. So we will need to track them here.
1537 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1538 #ifdef IB_MR_TYPE_SG_GAPS
1539 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1540 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1541 IB_MR_TYPE_SG_GAPS :
1547 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1548 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1549 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1551 if (IS_ERR(frd->frd_mr)) {
1552 rc = PTR_ERR(frd->frd_mr);
1553 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1558 /* There appears to be a bug in MLX5 code where you must
1559 * invalidate the rkey of a new FastReg pool before first
1560 * using it. Thus, I am marking the FRD invalid here. */
1561 frd->frd_valid = false;
1563 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1564 fpo->fast_reg.fpo_pool_size++;
1571 ib_dereg_mr(frd->frd_mr);
1572 #ifndef HAVE_IB_MAP_MR_SG
1574 ib_free_fast_reg_page_list(frd->frd_frpl);
1576 LIBCFS_FREE(frd, sizeof(*frd));
1579 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1581 list_del(&frd->frd_list);
1582 #ifndef HAVE_IB_MAP_MR_SG
1583 ib_free_fast_reg_page_list(frd->frd_frpl);
1585 ib_dereg_mr(frd->frd_mr);
1586 LIBCFS_FREE(frd, sizeof(*frd));
1592 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1593 struct kib_fmr_pool **pp_fpo)
1595 struct kib_dev *dev = fps->fps_net->ibn_dev;
1596 struct kib_fmr_pool *fpo;
1599 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1603 memset(fpo, 0, sizeof(*fpo));
1605 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1607 #ifdef HAVE_FMR_POOL_API
1608 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1609 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1611 #endif /* HAVE_FMR_POOL_API */
1612 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1616 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1617 fpo->fpo_owner = fps;
1623 kiblnd_hdev_decref(fpo->fpo_hdev);
1624 LIBCFS_FREE(fpo, sizeof(*fpo));
1629 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1631 struct kib_fmr_pool *fpo;
1633 if (fps->fps_net == NULL) /* intialized? */
1636 spin_lock(&fps->fps_lock);
1638 while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1639 struct kib_fmr_pool,
1640 fpo_list)) != NULL) {
1641 fpo->fpo_failed = 1;
1642 if (fpo->fpo_map_count == 0)
1643 list_move(&fpo->fpo_list, zombies);
1645 list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1648 spin_unlock(&fps->fps_lock);
1652 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1654 if (fps->fps_net != NULL) { /* initialized? */
1655 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1656 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1661 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1662 struct kib_net *net,
1663 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1665 struct kib_fmr_pool *fpo;
1668 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1673 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1674 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1675 fps->fps_cache = tunables->lnd_fmr_cache;
1677 spin_lock_init(&fps->fps_lock);
1678 INIT_LIST_HEAD(&fps->fps_pool_list);
1679 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1681 rc = kiblnd_create_fmr_pool(fps, &fpo);
1683 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1689 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1691 if (fpo->fpo_map_count != 0) /* still in use */
1693 if (fpo->fpo_failed)
1695 return now >= fpo->fpo_deadline;
1698 #if defined(HAVE_FMR_POOL_API) || !defined(HAVE_IB_MAP_MR_SG)
1700 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1702 struct kib_hca_dev *hdev;
1703 __u64 *pages = tx->tx_pages;
1708 hdev = tx->tx_pool->tpo_hdev;
1710 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1711 for (size = 0; size < rd->rd_frags[i].rf_nob;
1712 size += hdev->ibh_page_size) {
1713 pages[npages++] = (rd->rd_frags[i].rf_addr &
1714 hdev->ibh_page_mask) + size;
1723 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1726 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1727 struct kib_fmr_poolset *fps;
1728 time64_t now = ktime_get_seconds();
1729 struct kib_fmr_pool *tmp;
1734 fps = fpo->fpo_owner;
1736 #ifdef HAVE_FMR_POOL_API
1737 if (fpo->fpo_is_fmr) {
1738 if (fmr->fmr_pfmr) {
1739 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1740 fmr->fmr_pfmr = NULL;
1744 int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1748 #endif /* HAVE_FMR_POOL_API */
1750 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1753 frd->frd_valid = false;
1754 fmr->fmr_frd = NULL;
1755 spin_lock(&fps->fps_lock);
1756 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1757 spin_unlock(&fps->fps_lock);
1760 fmr->fmr_pool = NULL;
1762 spin_lock(&fps->fps_lock);
1763 fpo->fpo_map_count--; /* decref the pool */
1765 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1766 /* the first pool is persistent */
1767 if (fps->fps_pool_list.next == &fpo->fpo_list)
1770 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1771 list_move(&fpo->fpo_list, &zombies);
1775 spin_unlock(&fps->fps_lock);
1777 if (!list_empty(&zombies))
1778 kiblnd_destroy_fmr_pool_list(&zombies);
1781 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1782 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1783 struct kib_fmr *fmr)
1785 struct kib_fmr_pool *fpo;
1787 bool is_rx = (rd != tx->tx_rd);
1788 #ifdef HAVE_FMR_POOL_API
1789 __u64 *pages = tx->tx_pages;
1790 bool tx_pages_mapped = false;
1796 spin_lock(&fps->fps_lock);
1797 version = fps->fps_version;
1798 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1799 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1800 fpo->fpo_map_count++;
1802 #ifdef HAVE_FMR_POOL_API
1803 fmr->fmr_pfmr = NULL;
1804 if (fpo->fpo_is_fmr) {
1805 struct ib_pool_fmr *pfmr;
1807 spin_unlock(&fps->fps_lock);
1809 if (!tx_pages_mapped) {
1810 npages = kiblnd_map_tx_pages(tx, rd);
1811 tx_pages_mapped = true;
1814 pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1815 pages, npages, iov);
1816 if (likely(!IS_ERR(pfmr))) {
1817 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1819 fmr->fmr_frd = NULL;
1820 fmr->fmr_pfmr = pfmr;
1821 fmr->fmr_pool = fpo;
1826 #endif /* HAVE_FMR_POOL_API */
1828 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1829 struct kib_fast_reg_descriptor *frd;
1830 #ifdef HAVE_IB_MAP_MR_SG
1831 struct ib_reg_wr *wr;
1834 struct ib_rdma_wr *wr;
1835 struct ib_fast_reg_page_list *frpl;
1839 frd = list_first_entry(
1840 &fpo->fast_reg.fpo_pool_list,
1841 struct kib_fast_reg_descriptor,
1843 list_del(&frd->frd_list);
1844 spin_unlock(&fps->fps_lock);
1846 #ifndef HAVE_IB_MAP_MR_SG
1847 frpl = frd->frd_frpl;
1851 if (!frd->frd_valid) {
1852 struct ib_rdma_wr *inv_wr;
1853 __u32 key = is_rx ? mr->rkey : mr->lkey;
1855 inv_wr = &frd->frd_inv_wr;
1856 memset(inv_wr, 0, sizeof(*inv_wr));
1858 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1859 inv_wr->wr.wr_id = IBLND_WID_MR;
1860 inv_wr->wr.ex.invalidate_rkey = key;
1863 key = ib_inc_rkey(key);
1864 ib_update_fast_reg_key(mr, key);
1867 #ifdef HAVE_IB_MAP_MR_SG
1868 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1869 n = ib_map_mr_sg(mr, tx->tx_frags,
1870 rd->rd_nfrags, NULL, PAGE_SIZE);
1872 n = ib_map_mr_sg(mr, tx->tx_frags,
1873 rd->rd_nfrags, PAGE_SIZE);
1874 #endif /* HAVE_IB_MAP_MR_SG_5ARGS */
1875 if (unlikely(n != rd->rd_nfrags)) {
1876 CERROR("Failed to map mr %d/%d elements\n",
1878 return n < 0 ? n : -EINVAL;
1881 wr = &frd->frd_fastreg_wr;
1882 memset(wr, 0, sizeof(*wr));
1884 wr->wr.opcode = IB_WR_REG_MR;
1885 wr->wr.wr_id = IBLND_WID_MR;
1887 wr->wr.send_flags = 0;
1889 wr->key = is_rx ? mr->rkey : mr->lkey;
1890 wr->access = (IB_ACCESS_LOCAL_WRITE |
1891 IB_ACCESS_REMOTE_WRITE);
1892 #else /* HAVE_IB_MAP_MR_SG */
1893 if (!tx_pages_mapped) {
1894 npages = kiblnd_map_tx_pages(tx, rd);
1895 tx_pages_mapped = true;
1898 LASSERT(npages <= frpl->max_page_list_len);
1899 memcpy(frpl->page_list, pages,
1900 sizeof(*pages) * npages);
1902 /* Prepare FastReg WR */
1903 wr = &frd->frd_fastreg_wr;
1904 memset(wr, 0, sizeof(*wr));
1906 wr->wr.opcode = IB_WR_FAST_REG_MR;
1907 wr->wr.wr_id = IBLND_WID_MR;
1909 wr->wr.wr.fast_reg.iova_start = iov;
1910 wr->wr.wr.fast_reg.page_list = frpl;
1911 wr->wr.wr.fast_reg.page_list_len = npages;
1912 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1913 wr->wr.wr.fast_reg.length = nob;
1914 wr->wr.wr.fast_reg.rkey =
1915 is_rx ? mr->rkey : mr->lkey;
1916 wr->wr.wr.fast_reg.access_flags =
1917 (IB_ACCESS_LOCAL_WRITE |
1918 IB_ACCESS_REMOTE_WRITE);
1919 #endif /* HAVE_IB_MAP_MR_SG */
1921 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1923 fmr->fmr_pool = fpo;
1926 spin_unlock(&fps->fps_lock);
1930 spin_lock(&fps->fps_lock);
1931 fpo->fpo_map_count--;
1932 if (rc != -EAGAIN) {
1933 spin_unlock(&fps->fps_lock);
1937 /* EAGAIN and ... */
1938 if (version != fps->fps_version) {
1939 spin_unlock(&fps->fps_lock);
1944 if (fps->fps_increasing) {
1945 spin_unlock(&fps->fps_lock);
1946 CDEBUG(D_NET, "Another thread is allocating new "
1947 "FMR pool, waiting for her to complete\n");
1948 wait_var_event(fps, !fps->fps_increasing);
1953 if (ktime_get_seconds() < fps->fps_next_retry) {
1954 /* someone failed recently */
1955 spin_unlock(&fps->fps_lock);
1959 fps->fps_increasing = 1;
1960 spin_unlock(&fps->fps_lock);
1962 CDEBUG(D_NET, "Allocate new FMR pool\n");
1963 rc = kiblnd_create_fmr_pool(fps, &fpo);
1964 spin_lock(&fps->fps_lock);
1965 fps->fps_increasing = 0;
1969 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1971 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
1973 spin_unlock(&fps->fps_lock);
1979 kiblnd_fini_pool(struct kib_pool *pool)
1981 LASSERT(list_empty(&pool->po_free_list));
1982 LASSERT(pool->po_allocated == 0);
1984 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1988 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
1990 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1992 memset(pool, 0, sizeof(struct kib_pool));
1993 INIT_LIST_HEAD(&pool->po_free_list);
1994 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1995 pool->po_owner = ps;
1996 pool->po_size = size;
2000 kiblnd_destroy_pool_list(struct list_head *head)
2002 struct kib_pool *pool;
2004 while ((pool = list_first_entry_or_null(head,
2006 po_list)) != NULL) {
2007 list_del(&pool->po_list);
2009 LASSERT(pool->po_owner != NULL);
2010 pool->po_owner->ps_pool_destroy(pool);
2015 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2017 struct kib_pool *po;
2019 if (ps->ps_net == NULL) /* intialized? */
2022 spin_lock(&ps->ps_lock);
2023 while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2025 po_list)) != NULL) {
2027 if (po->po_allocated == 0)
2028 list_move(&po->po_list, zombies);
2030 list_move(&po->po_list, &ps->ps_failed_pool_list);
2032 spin_unlock(&ps->ps_lock);
2036 kiblnd_fini_poolset(struct kib_poolset *ps)
2038 if (ps->ps_net != NULL) { /* initialized? */
2039 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2040 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2045 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2046 struct kib_net *net, char *name, int size,
2047 kib_ps_pool_create_t po_create,
2048 kib_ps_pool_destroy_t po_destroy,
2049 kib_ps_node_init_t nd_init,
2050 kib_ps_node_fini_t nd_fini)
2052 struct kib_pool *pool;
2055 memset(ps, 0, sizeof(struct kib_poolset));
2059 ps->ps_pool_create = po_create;
2060 ps->ps_pool_destroy = po_destroy;
2061 ps->ps_node_init = nd_init;
2062 ps->ps_node_fini = nd_fini;
2063 ps->ps_pool_size = size;
2064 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2065 >= sizeof(ps->ps_name))
2067 spin_lock_init(&ps->ps_lock);
2068 INIT_LIST_HEAD(&ps->ps_pool_list);
2069 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2071 rc = ps->ps_pool_create(ps, size, &pool);
2073 list_add(&pool->po_list, &ps->ps_pool_list);
2075 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2081 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2083 if (pool->po_allocated != 0) /* still in use */
2085 if (pool->po_failed)
2087 return now >= pool->po_deadline;
2091 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2094 struct kib_poolset *ps = pool->po_owner;
2095 struct kib_pool *tmp;
2096 time64_t now = ktime_get_seconds();
2098 spin_lock(&ps->ps_lock);
2100 if (ps->ps_node_fini != NULL)
2101 ps->ps_node_fini(pool, node);
2103 LASSERT(pool->po_allocated > 0);
2104 list_add(node, &pool->po_free_list);
2105 pool->po_allocated--;
2107 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2108 /* the first pool is persistent */
2109 if (ps->ps_pool_list.next == &pool->po_list)
2112 if (kiblnd_pool_is_idle(pool, now))
2113 list_move(&pool->po_list, &zombies);
2115 spin_unlock(&ps->ps_lock);
2117 if (!list_empty(&zombies))
2118 kiblnd_destroy_pool_list(&zombies);
2122 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2124 struct list_head *node;
2125 struct kib_pool *pool;
2127 unsigned int interval = 1;
2128 ktime_t time_before;
2129 unsigned int trips = 0;
2132 spin_lock(&ps->ps_lock);
2133 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2134 if (list_empty(&pool->po_free_list))
2137 pool->po_allocated++;
2138 pool->po_deadline = ktime_get_seconds() +
2139 IBLND_POOL_DEADLINE;
2140 node = pool->po_free_list.next;
2143 if (ps->ps_node_init != NULL) {
2144 /* still hold the lock */
2145 ps->ps_node_init(pool, node);
2147 spin_unlock(&ps->ps_lock);
2151 /* no available tx pool and ... */
2152 if (ps->ps_increasing) {
2153 /* another thread is allocating a new pool */
2154 spin_unlock(&ps->ps_lock);
2157 "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2158 ps->ps_name, interval, trips);
2160 schedule_timeout_interruptible(interval);
2161 if (interval < cfs_time_seconds(1))
2167 if (ktime_get_seconds() < ps->ps_next_retry) {
2168 /* someone failed recently */
2169 spin_unlock(&ps->ps_lock);
2173 ps->ps_increasing = 1;
2174 spin_unlock(&ps->ps_lock);
2176 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2177 time_before = ktime_get();
2178 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2179 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2180 ktime_ms_delta(ktime_get(), time_before));
2182 spin_lock(&ps->ps_lock);
2183 ps->ps_increasing = 0;
2185 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2187 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2188 CERROR("Can't allocate new %s pool because out of memory\n",
2191 spin_unlock(&ps->ps_lock);
2197 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2199 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2203 LASSERT (pool->po_allocated == 0);
2205 if (tpo->tpo_tx_pages != NULL) {
2206 kiblnd_unmap_tx_pool(tpo);
2207 kiblnd_free_pages(tpo->tpo_tx_pages);
2210 if (tpo->tpo_tx_descs == NULL)
2213 for (i = 0; i < pool->po_size; i++) {
2214 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2215 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2217 list_del(&tx->tx_list);
2218 if (tx->tx_pages != NULL)
2219 CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2220 if (tx->tx_frags != NULL)
2221 CFS_FREE_PTR_ARRAY(tx->tx_frags,
2222 (1 + IBLND_MAX_RDMA_FRAGS));
2223 if (tx->tx_wrq != NULL)
2224 CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2225 (1 + IBLND_MAX_RDMA_FRAGS));
2226 if (tx->tx_sge != NULL)
2227 CFS_FREE_PTR_ARRAY(tx->tx_sge,
2228 (1 + IBLND_MAX_RDMA_FRAGS) *
2230 if (tx->tx_rd != NULL)
2231 LIBCFS_FREE(tx->tx_rd,
2232 offsetof(struct kib_rdma_desc,
2233 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2236 CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2238 kiblnd_fini_pool(pool);
2242 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2244 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2247 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2248 ntx = tunables->lnd_ntx / ncpts;
2250 return max(IBLND_TX_POOL, ntx);
2254 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2258 struct kib_pool *pool;
2259 struct kib_tx_pool *tpo;
2261 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2263 CERROR("Failed to allocate TX pool\n");
2267 pool = &tpo->tpo_pool;
2268 kiblnd_init_pool(ps, pool, size);
2269 tpo->tpo_tx_descs = NULL;
2270 tpo->tpo_tx_pages = NULL;
2272 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2273 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2274 CERROR("Can't allocate tx pages: %d\n", npg);
2279 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2280 size * sizeof(struct kib_tx));
2281 if (tpo->tpo_tx_descs == NULL) {
2282 CERROR("Can't allocate %d tx descriptors\n", size);
2283 ps->ps_pool_destroy(pool);
2287 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2289 for (i = 0; i < size; i++) {
2290 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2291 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2294 if (ps->ps_net->ibn_fmr_ps != NULL) {
2295 LIBCFS_CPT_ALLOC(tx->tx_pages,
2296 lnet_cpt_table(), ps->ps_cpt,
2297 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2298 if (tx->tx_pages == NULL)
2302 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2303 (1 + IBLND_MAX_RDMA_FRAGS) *
2304 sizeof(*tx->tx_frags));
2305 if (tx->tx_frags == NULL)
2308 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2310 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2311 (1 + IBLND_MAX_RDMA_FRAGS) *
2312 sizeof(*tx->tx_wrq));
2313 if (tx->tx_wrq == NULL)
2316 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2317 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2318 sizeof(*tx->tx_sge));
2319 if (tx->tx_sge == NULL)
2322 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2323 offsetof(struct kib_rdma_desc,
2324 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2325 if (tx->tx_rd == NULL)
2330 kiblnd_map_tx_pool(tpo);
2335 ps->ps_pool_destroy(pool);
2340 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2342 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2343 struct kib_tx_poolset,
2345 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2347 tx->tx_cookie = tps->tps_next_tx_cookie++;
2351 kiblnd_net_fini_pools(struct kib_net *net)
2355 cfs_cpt_for_each(i, lnet_cpt_table()) {
2356 struct kib_tx_poolset *tps;
2357 struct kib_fmr_poolset *fps;
2359 if (net->ibn_tx_ps != NULL) {
2360 tps = net->ibn_tx_ps[i];
2361 kiblnd_fini_poolset(&tps->tps_poolset);
2364 if (net->ibn_fmr_ps != NULL) {
2365 fps = net->ibn_fmr_ps[i];
2366 kiblnd_fini_fmr_poolset(fps);
2370 if (net->ibn_tx_ps != NULL) {
2371 cfs_percpt_free(net->ibn_tx_ps);
2372 net->ibn_tx_ps = NULL;
2375 if (net->ibn_fmr_ps != NULL) {
2376 cfs_percpt_free(net->ibn_fmr_ps);
2377 net->ibn_fmr_ps = NULL;
2382 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2385 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2386 #ifdef HAVE_IB_GET_DMA_MR
2387 unsigned long flags;
2393 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2395 #ifdef HAVE_IB_GET_DMA_MR
2396 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2398 * if lnd_map_on_demand is zero then we have effectively disabled
2399 * FMR or FastReg and we're using global memory regions
2402 if (!tunables->lnd_map_on_demand) {
2403 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2405 goto create_tx_pool;
2408 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2411 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2412 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2413 tunables->lnd_fmr_pool_size,
2414 tunables->lnd_ntx / 4);
2419 /* TX pool must be created later than FMR, see LU-2268
2421 LASSERT(net->ibn_tx_ps == NULL);
2423 /* premapping can fail if ibd_nmr > 1, so we always create
2424 * FMR pool and map-on-demand if premapping failed */
2426 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2427 sizeof(struct kib_fmr_poolset));
2428 if (net->ibn_fmr_ps == NULL) {
2429 CERROR("Failed to allocate FMR pool array\n");
2434 for (i = 0; i < ncpts; i++) {
2435 cpt = (cpts == NULL) ? i : cpts[i];
2436 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2439 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2446 LASSERT(i == ncpts);
2448 #ifdef HAVE_IB_GET_DMA_MR
2451 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2452 sizeof(struct kib_tx_poolset));
2453 if (net->ibn_tx_ps == NULL) {
2454 CERROR("Failed to allocate tx pool array\n");
2459 for (i = 0; i < ncpts; i++) {
2460 cpt = (cpts == NULL) ? i : cpts[i];
2461 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2463 kiblnd_tx_pool_size(ni, ncpts),
2464 kiblnd_create_tx_pool,
2465 kiblnd_destroy_tx_pool,
2466 kiblnd_tx_init, NULL);
2468 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2476 kiblnd_net_fini_pools(net);
2482 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2484 struct ib_port_attr *port_attr;
2486 unsigned long flags;
2487 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2489 LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2490 if (port_attr == NULL) {
2491 CDEBUG(D_NETERROR, "Out of memory\n");
2495 rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2497 write_lock_irqsave(g_lock, flags);
2500 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2501 ? IBLND_DEV_PORT_ACTIVE
2502 : IBLND_DEV_PORT_DOWN;
2504 write_unlock_irqrestore(g_lock, flags);
2505 LIBCFS_FREE(port_attr, sizeof(*port_attr));
2508 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2515 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2517 struct kib_net *net;
2519 /* for health check */
2520 list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2522 CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2523 libcfs_nid2str(net->ibn_ni->ni_nid));
2524 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2529 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2531 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2532 struct kib_hca_dev *hdev;
2533 unsigned long flags;
2535 hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2537 write_lock_irqsave(g_lock, flags);
2539 switch (event->event) {
2540 case IB_EVENT_DEVICE_FATAL:
2541 CDEBUG(D_NET, "IB device fatal\n");
2542 hdev->ibh_state = IBLND_DEV_FATAL;
2543 kiblnd_set_ni_fatal_on(hdev, 1);
2545 case IB_EVENT_PORT_ACTIVE:
2546 CDEBUG(D_NET, "IB port active\n");
2547 if (event->element.port_num == hdev->ibh_port) {
2548 hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2549 kiblnd_set_ni_fatal_on(hdev, 0);
2552 case IB_EVENT_PORT_ERR:
2553 CDEBUG(D_NET, "IB port err\n");
2554 if (event->element.port_num == hdev->ibh_port) {
2555 hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2556 kiblnd_set_ni_fatal_on(hdev, 1);
2562 write_unlock_irqrestore(g_lock, flags);
2566 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2568 struct ib_device_attr *dev_attr;
2572 /* It's safe to assume a HCA can handle a page size
2573 * matching that of the native system */
2574 hdev->ibh_page_shift = PAGE_SHIFT;
2575 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2576 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2578 #ifndef HAVE_IB_DEVICE_ATTRS
2579 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2580 if (dev_attr == NULL) {
2581 CERROR("Out of memory\n");
2585 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2587 CERROR("Failed to query IB device: %d\n", rc);
2588 goto out_clean_attr;
2591 dev_attr = &hdev->ibh_ibdev->attrs;
2594 hdev->ibh_mr_size = dev_attr->max_mr_size;
2595 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2597 /* Setup device Memory Registration capabilities */
2598 #ifdef HAVE_FMR_POOL_API
2599 #ifdef HAVE_IB_DEVICE_OPS
2600 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2601 hdev->ibh_ibdev->ops.dealloc_fmr &&
2602 hdev->ibh_ibdev->ops.map_phys_fmr &&
2603 hdev->ibh_ibdev->ops.unmap_fmr) {
2605 if (hdev->ibh_ibdev->alloc_fmr &&
2606 hdev->ibh_ibdev->dealloc_fmr &&
2607 hdev->ibh_ibdev->map_phys_fmr &&
2608 hdev->ibh_ibdev->unmap_fmr) {
2610 LCONSOLE_INFO("Using FMR for registration\n");
2611 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2613 #endif /* HAVE_FMR_POOL_API */
2614 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2615 LCONSOLE_INFO("Using FastReg for registration\n");
2616 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2617 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2618 #ifdef IB_DEVICE_SG_GAPS_REG
2619 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2620 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2627 rc2 = kiblnd_port_get_attr(hdev);
2634 #ifndef HAVE_IB_DEVICE_ATTRS
2636 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2640 CERROR("IB device does not support FMRs nor FastRegs, can't "
2641 "register memory: %d\n", rc);
2642 else if (rc == -EINVAL)
2643 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2647 #ifdef HAVE_IB_GET_DMA_MR
2649 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2651 if (hdev->ibh_mrs == NULL)
2654 ib_dereg_mr(hdev->ibh_mrs);
2656 hdev->ibh_mrs = NULL;
2661 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2663 if (hdev->ibh_event_handler.device != NULL)
2664 ib_unregister_event_handler(&hdev->ibh_event_handler);
2666 #ifdef HAVE_IB_GET_DMA_MR
2667 kiblnd_hdev_cleanup_mrs(hdev);
2670 if (hdev->ibh_pd != NULL)
2671 ib_dealloc_pd(hdev->ibh_pd);
2673 if (hdev->ibh_cmid != NULL)
2674 rdma_destroy_id(hdev->ibh_cmid);
2676 LIBCFS_FREE(hdev, sizeof(*hdev));
2679 #ifdef HAVE_IB_GET_DMA_MR
2681 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2684 int acflags = IB_ACCESS_LOCAL_WRITE |
2685 IB_ACCESS_REMOTE_WRITE;
2687 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2689 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2690 kiblnd_hdev_cleanup_mrs(hdev);
2701 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2707 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2709 struct rdma_cm_id *cmid;
2710 struct sockaddr_in srcaddr;
2711 struct sockaddr_in dstaddr;
2714 if (dev->ibd_hdev == NULL || /* initializing */
2715 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2716 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2719 /* XXX: it's UGLY, but I don't have better way to find
2720 * ib-bonding HCA failover because:
2722 * a. no reliable CM event for HCA failover...
2723 * b. no OFED API to get ib_device for current net_device...
2725 * We have only two choices at this point:
2727 * a. rdma_bind_addr(), it will conflict with listener cmid
2728 * b. rdma_resolve_addr() to zero addr */
2729 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2730 RDMA_PS_TCP, IB_QPT_RC);
2733 CERROR("Failed to create cmid for failover: %d\n", rc);
2737 memset(&srcaddr, 0, sizeof(srcaddr));
2738 srcaddr.sin_family = AF_INET;
2739 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2741 memset(&dstaddr, 0, sizeof(dstaddr));
2742 dstaddr.sin_family = AF_INET;
2743 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2744 (struct sockaddr *)&dstaddr, 1);
2745 if (rc != 0 || cmid->device == NULL) {
2746 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2747 dev->ibd_ifname, &dev->ibd_ifip,
2749 rdma_destroy_id(cmid);
2753 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2754 rdma_destroy_id(cmid);
2759 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2761 LIST_HEAD(zombie_tpo);
2762 LIST_HEAD(zombie_ppo);
2763 LIST_HEAD(zombie_fpo);
2764 struct rdma_cm_id *cmid = NULL;
2765 struct kib_hca_dev *hdev = NULL;
2766 struct kib_hca_dev *old;
2768 struct kib_net *net;
2769 struct sockaddr_in addr;
2770 unsigned long flags;
2774 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2775 dev->ibd_can_failover ||
2776 dev->ibd_hdev == NULL);
2778 rc = kiblnd_dev_need_failover(dev, ns);
2782 if (dev->ibd_hdev != NULL &&
2783 dev->ibd_hdev->ibh_cmid != NULL) {
2784 /* XXX it's not good to close old listener at here,
2785 * because we can fail to create new listener.
2786 * But we have to close it now, otherwise rdma_bind_addr
2787 * will return EADDRINUSE... How crap! */
2788 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2790 cmid = dev->ibd_hdev->ibh_cmid;
2791 /* make next schedule of kiblnd_dev_need_failover()
2792 * return 1 for me */
2793 dev->ibd_hdev->ibh_cmid = NULL;
2794 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2796 rdma_destroy_id(cmid);
2799 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2803 CERROR("Failed to create cmid for failover: %d\n", rc);
2807 memset(&addr, 0, sizeof(addr));
2808 addr.sin_family = AF_INET;
2809 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2810 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2812 /* Bind to failover device or port */
2813 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2814 if (rc != 0 || cmid->device == NULL) {
2815 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2816 dev->ibd_ifname, &dev->ibd_ifip,
2818 rdma_destroy_id(cmid);
2822 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2824 CERROR("Failed to allocate kib_hca_dev\n");
2825 rdma_destroy_id(cmid);
2830 atomic_set(&hdev->ibh_ref, 1);
2831 hdev->ibh_dev = dev;
2832 hdev->ibh_cmid = cmid;
2833 hdev->ibh_ibdev = cmid->device;
2834 hdev->ibh_port = cmid->port_num;
2836 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2837 pd = ib_alloc_pd(cmid->device, 0);
2839 pd = ib_alloc_pd(cmid->device);
2843 CERROR("Can't allocate PD: %d\n", rc);
2849 rc = rdma_listen(cmid, 0);
2851 CERROR("Can't start new listener: %d\n", rc);
2855 rc = kiblnd_hdev_get_attr(hdev);
2857 CERROR("Can't get device attributes: %d\n", rc);
2861 #ifdef HAVE_IB_GET_DMA_MR
2862 rc = kiblnd_hdev_setup_mrs(hdev);
2864 CERROR("Can't setup device: %d\n", rc);
2869 INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
2870 hdev->ibh_ibdev, kiblnd_event_handler);
2871 ib_register_event_handler(&hdev->ibh_event_handler);
2873 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2875 old = dev->ibd_hdev;
2876 dev->ibd_hdev = hdev; /* take over the refcount */
2879 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2880 cfs_cpt_for_each(i, lnet_cpt_table()) {
2881 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2884 if (net->ibn_fmr_ps != NULL)
2885 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2890 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2892 if (!list_empty(&zombie_tpo))
2893 kiblnd_destroy_pool_list(&zombie_tpo);
2894 if (!list_empty(&zombie_ppo))
2895 kiblnd_destroy_pool_list(&zombie_ppo);
2896 if (!list_empty(&zombie_fpo))
2897 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2899 kiblnd_hdev_decref(hdev);
2902 dev->ibd_failed_failover++;
2904 dev->ibd_failed_failover = 0;
2910 kiblnd_destroy_dev(struct kib_dev *dev)
2912 LASSERT(dev->ibd_nnets == 0);
2913 LASSERT(list_empty(&dev->ibd_nets));
2915 list_del(&dev->ibd_fail_list);
2916 list_del(&dev->ibd_list);
2918 if (dev->ibd_hdev != NULL)
2919 kiblnd_hdev_decref(dev->ibd_hdev);
2921 LIBCFS_FREE(dev, sizeof(*dev));
2925 kiblnd_base_shutdown(void)
2927 struct kib_sched_info *sched;
2928 struct kib_peer_ni *peer_ni;
2931 LASSERT(list_empty(&kiblnd_data.kib_devs));
2933 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
2934 libcfs_kmem_read());
2936 switch (kiblnd_data.kib_init) {
2940 case IBLND_INIT_ALL:
2941 case IBLND_INIT_DATA:
2942 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
2944 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2945 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2946 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2947 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2949 /* flag threads to terminate; wake and wait for them to die */
2950 kiblnd_data.kib_shutdown = 1;
2952 /* NB: we really want to stop scheduler threads net by net
2953 * instead of the whole module, this should be improved
2954 * with dynamic configuration LNet.
2956 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2957 wake_up_all(&sched->ibs_waitq);
2959 wake_up(&kiblnd_data.kib_connd_waitq);
2960 wake_up(&kiblnd_data.kib_failover_waitq);
2962 wait_var_event_warning(&kiblnd_data.kib_nthreads,
2963 !atomic_read(&kiblnd_data.kib_nthreads),
2964 "Waiting for %d threads to terminate\n",
2965 atomic_read(&kiblnd_data.kib_nthreads));
2968 case IBLND_INIT_NOTHING:
2972 if (kiblnd_data.kib_scheds != NULL)
2973 cfs_percpt_free(kiblnd_data.kib_scheds);
2975 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
2976 libcfs_kmem_read());
2978 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2979 module_put(THIS_MODULE);
2983 kiblnd_shutdown(struct lnet_ni *ni)
2985 struct kib_net *net = ni->ni_data;
2986 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2987 unsigned long flags;
2989 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2994 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
2995 libcfs_kmem_read());
2997 write_lock_irqsave(g_lock, flags);
2998 net->ibn_shutdown = 1;
2999 write_unlock_irqrestore(g_lock, flags);
3001 switch (net->ibn_init) {
3005 case IBLND_INIT_ALL:
3006 /* nuke all existing peers within this net */
3007 kiblnd_del_peer(ni, LNET_NID_ANY);
3009 /* Wait for all peer_ni state to clean up */
3010 wait_var_event_warning(&net->ibn_npeers,
3011 atomic_read(&net->ibn_npeers) == 0,
3012 "%s: waiting for %d peers to disconnect\n",
3013 libcfs_nid2str(ni->ni_nid),
3014 atomic_read(&net->ibn_npeers));
3016 kiblnd_net_fini_pools(net);
3018 write_lock_irqsave(g_lock, flags);
3019 LASSERT(net->ibn_dev->ibd_nnets > 0);
3020 net->ibn_dev->ibd_nnets--;
3021 list_del(&net->ibn_list);
3022 write_unlock_irqrestore(g_lock, flags);
3026 case IBLND_INIT_NOTHING:
3027 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3029 if (net->ibn_dev != NULL &&
3030 net->ibn_dev->ibd_nnets == 0)
3031 kiblnd_destroy_dev(net->ibn_dev);
3036 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3037 libcfs_kmem_read());
3039 net->ibn_init = IBLND_INIT_NOTHING;
3042 LIBCFS_FREE(net, sizeof(*net));
3045 if (list_empty(&kiblnd_data.kib_devs))
3046 kiblnd_base_shutdown();
3050 kiblnd_base_startup(struct net *ns)
3052 struct kib_sched_info *sched;
3056 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3058 if (!try_module_get(THIS_MODULE))
3061 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3063 rwlock_init(&kiblnd_data.kib_global_lock);
3065 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3066 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3068 hash_init(kiblnd_data.kib_peers);
3070 spin_lock_init(&kiblnd_data.kib_connd_lock);
3071 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3072 INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3073 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3074 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3075 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3077 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3078 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3080 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3082 if (kiblnd_data.kib_scheds == NULL)
3085 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3088 spin_lock_init(&sched->ibs_lock);
3089 INIT_LIST_HEAD(&sched->ibs_conns);
3090 init_waitqueue_head(&sched->ibs_waitq);
3092 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3093 if (*kiblnd_tunables.kib_nscheds > 0) {
3094 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3096 /* max to half of CPUs, another half is reserved for
3097 * upper layer modules */
3098 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3101 sched->ibs_nthreads_max = nthrs;
3105 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3107 /* lists/ptrs/locks initialised */
3108 kiblnd_data.kib_init = IBLND_INIT_DATA;
3109 /*****************************************************/
3111 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3113 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3117 if (*kiblnd_tunables.kib_dev_failover != 0)
3118 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3122 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3126 /* flag everything initialised */
3127 kiblnd_data.kib_init = IBLND_INIT_ALL;
3128 /*****************************************************/
3133 kiblnd_base_shutdown();
3138 kiblnd_start_schedulers(struct kib_sched_info *sched)
3144 if (sched->ibs_nthreads == 0) {
3145 if (*kiblnd_tunables.kib_nscheds > 0) {
3146 nthrs = sched->ibs_nthreads_max;
3148 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3150 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3151 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3154 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3155 /* increase one thread if there is new interface */
3156 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3159 for (i = 0; i < nthrs; i++) {
3162 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3163 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3164 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3165 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3169 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3170 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3174 sched->ibs_nthreads += i;
3178 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3185 for (i = 0; i < ncpts; i++) {
3186 struct kib_sched_info *sched;
3188 cpt = (cpts == NULL) ? i : cpts[i];
3189 sched = kiblnd_data.kib_scheds[cpt];
3191 if (!newdev && sched->ibs_nthreads > 0)
3194 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3196 CERROR("Failed to start scheduler threads for %s\n",
3204 static struct kib_dev *
3205 kiblnd_dev_search(char *ifname)
3207 struct kib_dev *alias = NULL;
3208 struct kib_dev *dev;
3212 colon = strchr(ifname, ':');
3213 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3214 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3220 colon2 = strchr(dev->ibd_ifname, ':');
3226 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3238 kiblnd_startup(struct lnet_ni *ni)
3240 char *ifname = NULL;
3241 struct lnet_inetdev *ifaces = NULL;
3242 struct kib_dev *ibdev = NULL;
3243 struct kib_net *net = NULL;
3244 unsigned long flags;
3249 LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3251 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3252 rc = kiblnd_base_startup(ni->ni_net_ns);
3257 LIBCFS_ALLOC(net, sizeof(*net));
3265 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3267 kiblnd_tunables_setup(ni);
3270 * Multi-Rail wants each secondary
3271 * IP to be treated as an unique 'struct ni' interface.
3273 if (ni->ni_interface != NULL) {
3274 /* Use the IPoIB interface specified in 'networks=' */
3275 ifname = ni->ni_interface;
3277 ifname = *kiblnd_tunables.kib_default_ipif;
3280 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3281 CERROR("IPoIB interface name too long: %s\n", ifname);
3286 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3290 for (i = 0; i < rc; i++) {
3291 if (strcmp(ifname, ifaces[i].li_name) == 0)
3296 CERROR("ko2iblnd: No matching interfaces\n");
3301 ibdev = kiblnd_dev_search(ifname);
3302 newdev = ibdev == NULL;
3303 /* hmm...create kib_dev even for alias */
3304 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3305 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3311 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3312 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3313 sizeof(ibdev->ibd_ifname));
3314 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3316 INIT_LIST_HEAD(&ibdev->ibd_nets);
3317 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3318 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3320 /* initialize the device */
3321 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3323 CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3328 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3331 net->ibn_dev = ibdev;
3332 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3334 ni->ni_dev_cpt = ifaces[i].li_cpt;
3336 rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3340 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3342 CERROR("Failed to initialize NI pools: %d\n", rc);
3346 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3348 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3349 /* for health check */
3350 if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3351 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3352 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3354 net->ibn_init = IBLND_INIT_ALL;
3359 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3360 kiblnd_destroy_dev(ibdev);
3363 kiblnd_shutdown(ni);
3365 CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3366 ifname ? ifname : "", rc);
3371 static const struct lnet_lnd the_o2iblnd = {
3372 .lnd_type = O2IBLND,
3373 .lnd_startup = kiblnd_startup,
3374 .lnd_shutdown = kiblnd_shutdown,
3375 .lnd_ctl = kiblnd_ctl,
3376 .lnd_send = kiblnd_send,
3377 .lnd_recv = kiblnd_recv,
3380 static void ko2inlnd_assert_wire_constants(void)
3382 BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3383 BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3384 BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3385 BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3387 BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3388 BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3389 BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3390 BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3391 BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3392 BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3393 BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3394 BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3395 BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3396 BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3398 BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3399 BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3400 BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3401 BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3402 BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3403 BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3404 BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3405 BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3407 BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3408 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3409 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3410 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3411 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3412 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3413 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3415 BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3416 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3417 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3418 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3419 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3421 BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3422 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3423 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3424 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3425 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3427 BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3428 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3429 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3430 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3431 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3432 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3433 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3435 BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3436 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3437 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3438 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3439 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3441 BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3442 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3443 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3444 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3445 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3446 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3447 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3449 BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3450 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3451 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3452 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3453 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3454 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3455 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3457 BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3458 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3459 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3460 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3461 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3463 /* Checks for struct kib_msg */
3464 //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3465 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3466 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3467 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3468 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3469 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3470 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3471 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3472 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3473 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3474 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3475 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3476 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3477 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3478 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3479 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3480 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3481 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3482 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3483 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3484 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3487 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3488 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3489 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3490 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3491 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3492 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3494 /* Immediate message */
3495 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3496 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3497 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3498 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3500 /* PUT req message */
3501 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3502 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3503 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3504 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3507 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3508 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3509 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3510 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3511 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3512 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3515 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3516 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3517 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3518 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3519 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3520 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3522 /* Completion message */
3523 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3524 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3525 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3526 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3529 BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3530 BUILD_BUG_ON(offsetof(struct kib_msg,
3531 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3533 BUILD_BUG_ON(offsetof(struct kib_msg,
3534 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3538 static void __exit ko2iblnd_exit(void)
3540 lnet_unregister_lnd(&the_o2iblnd);
3543 static int __init ko2iblnd_init(void)
3547 ko2inlnd_assert_wire_constants();
3549 rc = kiblnd_tunables_init();
3553 lnet_register_lnd(&the_o2iblnd);
3558 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3559 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3560 MODULE_VERSION("2.8.0");
3561 MODULE_LICENSE("GPL");
3563 module_init(ko2iblnd_init);
3564 module_exit(ko2iblnd_exit);