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/ethtool.h>
38 #include <linux/inetdevice.h>
42 static const struct lnet_lnd the_o2iblnd;
44 struct kib_data kiblnd_data;
47 kiblnd_cksum (void *ptr, int nob)
53 sum = ((sum << 1) | (sum >> 31)) + *c++;
55 /* ensure I don't return 0 (== no checksum) */
56 return (sum == 0) ? 1 : sum;
60 kiblnd_msgtype2str(int type)
63 case IBLND_MSG_CONNREQ:
66 case IBLND_MSG_CONNACK:
72 case IBLND_MSG_IMMEDIATE:
75 case IBLND_MSG_PUT_REQ:
78 case IBLND_MSG_PUT_NAK:
81 case IBLND_MSG_PUT_ACK:
84 case IBLND_MSG_PUT_DONE:
87 case IBLND_MSG_GET_REQ:
90 case IBLND_MSG_GET_DONE:
99 kiblnd_msgtype2size(int type)
101 const int hdr_size = offsetof(struct kib_msg, ibm_u);
104 case IBLND_MSG_CONNREQ:
105 case IBLND_MSG_CONNACK:
106 return hdr_size + sizeof(struct kib_connparams);
111 case IBLND_MSG_IMMEDIATE:
112 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
114 case IBLND_MSG_PUT_REQ:
115 return hdr_size + sizeof(struct kib_putreq_msg);
117 case IBLND_MSG_PUT_ACK:
118 return hdr_size + sizeof(struct kib_putack_msg);
120 case IBLND_MSG_GET_REQ:
121 return hdr_size + sizeof(struct kib_get_msg);
123 case IBLND_MSG_PUT_NAK:
124 case IBLND_MSG_PUT_DONE:
125 case IBLND_MSG_GET_DONE:
126 return hdr_size + sizeof(struct kib_completion_msg);
132 static int kiblnd_unpack_rd(struct kib_msg *msg, bool flip)
134 struct kib_rdma_desc *rd;
139 LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
140 msg->ibm_type == IBLND_MSG_PUT_ACK);
142 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
143 &msg->ibm_u.get.ibgm_rd :
144 &msg->ibm_u.putack.ibpam_rd;
147 __swab32s(&rd->rd_key);
148 __swab32s(&rd->rd_nfrags);
153 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
154 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
155 n, IBLND_MAX_RDMA_FRAGS);
159 nob = offsetof(struct kib_msg, ibm_u) +
160 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
162 if (msg->ibm_nob < nob) {
163 CERROR("Short %s: %d(%d)\n",
164 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
171 for (i = 0; i < n; i++) {
172 __swab32s(&rd->rd_frags[i].rf_nob);
173 __swab64s(&rd->rd_frags[i].rf_addr);
179 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
180 int credits, lnet_nid_t dstnid, __u64 dststamp)
182 struct kib_net *net = ni->ni_data;
184 /* CAVEAT EMPTOR! all message fields not set here should have been
185 * initialised previously.
187 msg->ibm_magic = IBLND_MSG_MAGIC;
188 msg->ibm_version = version;
190 msg->ibm_credits = credits;
193 msg->ibm_srcnid = lnet_nid_to_nid4(&ni->ni_nid);
194 msg->ibm_srcstamp = net->ibn_incarnation;
195 msg->ibm_dstnid = dstnid;
196 msg->ibm_dststamp = dststamp;
198 if (*kiblnd_tunables.kib_cksum) {
199 /* NB ibm_cksum zero while computing cksum */
200 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
204 int kiblnd_unpack_msg(struct kib_msg *msg, int nob)
206 const int hdr_size = offsetof(struct kib_msg, ibm_u);
212 /* 6 bytes are enough to have received magic + version */
214 CERROR("Short message: %d\n", nob);
218 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
220 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
223 CERROR("Bad magic: %08x\n", msg->ibm_magic);
227 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
228 if (version != IBLND_MSG_VERSION &&
229 version != IBLND_MSG_VERSION_1) {
230 CERROR("Bad version: %x\n", version);
234 if (nob < hdr_size) {
235 CERROR("Short message: %d\n", nob);
239 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
241 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
245 /* checksum must be computed with ibm_cksum zero and BEFORE anything
248 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
250 if (msg_cksum != 0 &&
251 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
252 CERROR("Bad checksum\n");
256 msg->ibm_cksum = msg_cksum;
259 /* leave magic unflipped as a clue to peer_ni endianness */
260 msg->ibm_version = version;
261 BUILD_BUG_ON(sizeof(msg->ibm_type) != 1);
262 BUILD_BUG_ON(sizeof(msg->ibm_credits) != 1);
263 msg->ibm_nob = msg_nob;
264 __swab64s(&msg->ibm_srcnid);
265 __swab64s(&msg->ibm_srcstamp);
266 __swab64s(&msg->ibm_dstnid);
267 __swab64s(&msg->ibm_dststamp);
270 if (msg->ibm_srcnid == LNET_NID_ANY) {
271 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
275 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
276 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
277 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
281 switch (msg->ibm_type) {
283 CERROR("Unknown message type %x\n", msg->ibm_type);
287 case IBLND_MSG_IMMEDIATE:
288 case IBLND_MSG_PUT_REQ:
291 case IBLND_MSG_PUT_ACK:
292 case IBLND_MSG_GET_REQ:
293 if (kiblnd_unpack_rd(msg, flip))
297 case IBLND_MSG_PUT_NAK:
298 case IBLND_MSG_PUT_DONE:
299 case IBLND_MSG_GET_DONE:
301 __swab32s(&msg->ibm_u.completion.ibcm_status);
304 case IBLND_MSG_CONNREQ:
305 case IBLND_MSG_CONNACK:
307 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
308 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
309 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
317 kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
320 struct kib_peer_ni *peer_ni;
321 struct kib_net *net = ni->ni_data;
322 int cpt = lnet_cpt_of_nid(nid, ni);
325 LASSERT(net != NULL);
326 LASSERT(nid != LNET_NID_ANY);
328 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
330 CERROR("Cannot allocate peer_ni\n");
334 peer_ni->ibp_ni = ni;
335 peer_ni->ibp_nid = nid;
336 peer_ni->ibp_error = 0;
337 peer_ni->ibp_last_alive = 0;
338 peer_ni->ibp_max_frags = IBLND_MAX_RDMA_FRAGS;
339 peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
340 peer_ni->ibp_queue_depth_mod = 0; /* try to use the default */
341 kref_init(&peer_ni->ibp_kref);
343 INIT_HLIST_NODE(&peer_ni->ibp_list);
344 INIT_LIST_HEAD(&peer_ni->ibp_conns);
345 INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
347 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
349 /* always called with a ref on ni, which prevents ni being shutdown */
350 LASSERT(net->ibn_shutdown == 0);
352 /* npeers only grows with the global lock held */
353 atomic_inc(&net->ibn_npeers);
355 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
362 kiblnd_destroy_peer(struct kref *kref)
364 struct kib_peer_ni *peer_ni = container_of(kref, struct kib_peer_ni,
366 struct kib_net *net = peer_ni->ibp_ni->ni_data;
368 LASSERT(net != NULL);
369 LASSERT(!kiblnd_peer_active(peer_ni));
370 LASSERT(kiblnd_peer_idle(peer_ni));
371 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
373 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
375 /* NB a peer_ni's connections keep a reference on their peer_ni until
376 * they are destroyed, so we can be assured that _all_ state to do
377 * with this peer_ni has been cleaned up when its refcount drops to
380 if (atomic_dec_and_test(&net->ibn_npeers))
381 wake_up_var(&net->ibn_npeers);
385 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
387 /* the caller is responsible for accounting the additional reference
390 struct kib_peer_ni *peer_ni;
392 hash_for_each_possible(kiblnd_data.kib_peers, peer_ni,
394 LASSERT(!kiblnd_peer_idle(peer_ni));
397 * Match a peer if its NID and the NID of the local NI it
398 * communicates over are the same. Otherwise don't match
399 * the peer, which will result in a new lnd peer being
402 if (peer_ni->ibp_nid != nid ||
403 !nid_same(&peer_ni->ibp_ni->ni_nid, &ni->ni_nid))
406 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
407 peer_ni, libcfs_nid2str(nid),
408 kref_read(&peer_ni->ibp_kref),
409 peer_ni->ibp_version);
416 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
418 LASSERT(list_empty(&peer_ni->ibp_conns));
420 LASSERT(kiblnd_peer_active(peer_ni));
421 hlist_del_init(&peer_ni->ibp_list);
422 /* lose peerlist's ref */
423 kiblnd_peer_decref(peer_ni);
427 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
428 lnet_nid_t *nidp, int *count)
430 struct kib_peer_ni *peer_ni;
434 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
436 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
437 LASSERT(!kiblnd_peer_idle(peer_ni));
439 if (peer_ni->ibp_ni != ni)
445 *nidp = peer_ni->ibp_nid;
446 *count = kref_read(&peer_ni->ibp_kref);
448 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
452 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
457 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
459 struct kib_conn *cnxt;
460 struct kib_conn *conn;
462 if (list_empty(&peer_ni->ibp_conns)) {
463 kiblnd_unlink_peer_locked(peer_ni);
465 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
467 kiblnd_close_conn_locked(conn, 0);
468 /* NB closing peer_ni's last conn unlinked it. */
470 /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
475 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
478 struct hlist_node *pnxt;
479 struct kib_peer_ni *peer_ni;
486 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
488 if (nid != LNET_NID_ANY) {
489 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
493 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
496 for (i = lo; i <= hi; i++) {
497 hlist_for_each_entry_safe(peer_ni, pnxt,
498 &kiblnd_data.kib_peers[i], ibp_list) {
499 LASSERT(!kiblnd_peer_idle(peer_ni));
501 if (peer_ni->ibp_ni != ni)
504 if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
507 if (!list_empty(&peer_ni->ibp_tx_queue)) {
508 LASSERT(list_empty(&peer_ni->ibp_conns));
510 list_splice_init(&peer_ni->ibp_tx_queue,
514 kiblnd_del_peer_locked(peer_ni);
515 rc = 0; /* matched something */
519 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
521 kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
526 static struct kib_conn *
527 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
529 struct kib_peer_ni *peer_ni;
530 struct kib_conn *conn;
534 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
536 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
537 LASSERT(!kiblnd_peer_idle(peer_ni));
539 if (peer_ni->ibp_ni != ni)
542 list_for_each_entry(conn, &peer_ni->ibp_conns,
547 kiblnd_conn_addref(conn);
548 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
554 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
559 kiblnd_debug_rx(struct kib_rx *rx)
561 CDEBUG(D_CONSOLE, " %p msg_type %x cred %d\n",
562 rx, rx->rx_msg->ibm_type,
563 rx->rx_msg->ibm_credits);
567 kiblnd_debug_tx(struct kib_tx *tx)
569 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lld "
570 "cookie %#llx msg %s%s type %x cred %d\n",
571 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
572 tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
573 tx->tx_lntmsg[0] == NULL ? "-" : "!",
574 tx->tx_lntmsg[1] == NULL ? "-" : "!",
575 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
579 kiblnd_debug_conn(struct kib_conn *conn)
581 struct list_head *tmp;
584 spin_lock(&conn->ibc_lock);
586 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
587 atomic_read(&conn->ibc_refcount), conn,
588 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
589 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d "
590 " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
591 conn->ibc_nsends_posted, conn->ibc_credits,
592 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
593 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
595 CDEBUG(D_CONSOLE, " early_rxs:\n");
596 list_for_each(tmp, &conn->ibc_early_rxs)
597 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
599 CDEBUG(D_CONSOLE, " tx_noops:\n");
600 list_for_each(tmp, &conn->ibc_tx_noops)
601 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
603 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
604 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
605 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
607 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
608 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
609 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
611 CDEBUG(D_CONSOLE, " tx_queue:\n");
612 list_for_each(tmp, &conn->ibc_tx_queue)
613 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
615 CDEBUG(D_CONSOLE, " active_txs:\n");
616 list_for_each(tmp, &conn->ibc_active_txs)
617 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
619 CDEBUG(D_CONSOLE, " rxs:\n");
620 for (i = 0; i < IBLND_RX_MSGS(conn); i++)
621 kiblnd_debug_rx(&conn->ibc_rxs[i]);
623 spin_unlock(&conn->ibc_lock);
627 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
629 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
630 if (cmid->route.path_rec == NULL)
633 if (*kiblnd_tunables.kib_ib_mtu)
634 cmid->route.path_rec->mtu =
635 ib_mtu_int_to_enum(*kiblnd_tunables.kib_ib_mtu);
639 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
647 vectors = conn->ibc_cmid->device->num_comp_vectors;
651 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
653 /* hash NID to CPU id in this partition... */
654 ibp_nid = conn->ibc_peer->ibp_nid;
655 off = do_div(ibp_nid, cpumask_weight(*mask));
656 for_each_cpu(i, *mask) {
666 * Get the scheduler bound to this CPT. If the scheduler has no
667 * threads, which means that the CPT has no CPUs, then grab the
668 * next scheduler that we can use.
670 * This case would be triggered if a NUMA node is configured with
671 * no associated CPUs.
673 static struct kib_sched_info *
674 kiblnd_get_scheduler(int cpt)
676 struct kib_sched_info *sched;
679 sched = kiblnd_data.kib_scheds[cpt];
681 if (sched->ibs_nthreads > 0)
684 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
685 if (sched->ibs_nthreads > 0) {
686 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
687 cpt, sched->ibs_cpt);
695 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
698 * One WR for the LNet message
699 * And ibc_max_frags for the transfer WRs
702 int multiplier = 1 + conn->ibc_max_frags;
703 enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
705 /* FastReg needs two extra WRs for map and invalidate */
706 if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
709 /* account for a maximum of ibc_queue_depth in-flight transfers */
710 ret = multiplier * conn->ibc_queue_depth;
712 if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
713 CDEBUG(D_NET, "peer_credits %u will result in send work "
714 "request size %d larger than maximum %d device "
715 "can handle\n", conn->ibc_queue_depth, ret,
716 conn->ibc_hdev->ibh_max_qp_wr);
717 conn->ibc_queue_depth =
718 conn->ibc_hdev->ibh_max_qp_wr / multiplier;
721 /* don't go beyond the maximum the device can handle */
722 return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
726 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
727 int state, int version)
730 * If the new conn is created successfully it takes over the caller's
731 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
732 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
733 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
734 * to destroy 'cmid' here since I'm called from the CM which still has
735 * its ref on 'cmid'). */
736 rwlock_t *glock = &kiblnd_data.kib_global_lock;
737 struct kib_net *net = peer_ni->ibp_ni->ni_data;
739 struct ib_qp_init_attr init_qp_attr = {};
740 struct kib_sched_info *sched;
741 #ifdef HAVE_IB_CQ_INIT_ATTR
742 struct ib_cq_init_attr cq_attr = {};
744 struct kib_conn *conn;
751 LASSERT(net != NULL);
752 LASSERT(!in_interrupt());
756 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
757 sched = kiblnd_get_scheduler(cpt);
760 CERROR("no schedulers available. node is unhealthy\n");
765 * The cpt might have changed if we ended up selecting a non cpt
766 * native scheduler. So use the scheduler's cpt instead.
768 cpt = sched->ibs_cpt;
770 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
772 CERROR("Can't allocate connection for %s\n",
773 libcfs_nid2str(peer_ni->ibp_nid));
777 conn->ibc_state = IBLND_CONN_INIT;
778 conn->ibc_version = version;
779 conn->ibc_peer = peer_ni; /* I take the caller's ref */
780 cmid->context = conn; /* for future CM callbacks */
781 conn->ibc_cmid = cmid;
782 conn->ibc_max_frags = peer_ni->ibp_max_frags;
783 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
784 conn->ibc_rxs = NULL;
785 conn->ibc_rx_pages = NULL;
787 INIT_LIST_HEAD(&conn->ibc_early_rxs);
788 INIT_LIST_HEAD(&conn->ibc_tx_noops);
789 INIT_LIST_HEAD(&conn->ibc_tx_queue);
790 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
791 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
792 INIT_LIST_HEAD(&conn->ibc_active_txs);
793 INIT_LIST_HEAD(&conn->ibc_zombie_txs);
794 spin_lock_init(&conn->ibc_lock);
796 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
797 sizeof(*conn->ibc_connvars));
798 if (conn->ibc_connvars == NULL) {
799 CERROR("Can't allocate in-progress connection state\n");
803 write_lock_irqsave(glock, flags);
804 if (dev->ibd_failover) {
805 write_unlock_irqrestore(glock, flags);
806 CERROR("%s: failover in progress\n", dev->ibd_ifname);
810 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
811 /* wakeup failover thread and teardown connection */
812 if (kiblnd_dev_can_failover(dev)) {
813 list_add_tail(&dev->ibd_fail_list,
814 &kiblnd_data.kib_failed_devs);
815 wake_up(&kiblnd_data.kib_failover_waitq);
818 write_unlock_irqrestore(glock, flags);
819 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
820 cmid->device->name, dev->ibd_ifname);
824 kiblnd_hdev_addref_locked(dev->ibd_hdev);
825 conn->ibc_hdev = dev->ibd_hdev;
827 kiblnd_setup_mtu_locked(cmid);
829 write_unlock_irqrestore(glock, flags);
831 #ifdef HAVE_IB_CQ_INIT_ATTR
832 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
833 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
834 cq = ib_create_cq(cmid->device,
835 kiblnd_cq_completion, kiblnd_cq_event, conn,
838 cq = ib_create_cq(cmid->device,
839 kiblnd_cq_completion, kiblnd_cq_event, conn,
840 IBLND_CQ_ENTRIES(conn),
841 kiblnd_get_completion_vector(conn, cpt));
845 * on MLX-5 (possibly MLX-4 as well) this error could be
846 * hit if the concurrent_sends and/or peer_tx_credits is set
847 * too high. Or due to an MLX-5 bug which tries to
848 * allocate 256kb via kmalloc for WR cookie array
850 CERROR("Failed to create CQ with %d CQEs: %ld\n",
851 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
857 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
859 CERROR("Can't request completion notification: %d\n", rc);
863 init_qp_attr.event_handler = kiblnd_qp_event;
864 init_qp_attr.qp_context = conn;
865 init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
866 init_qp_attr.cap.max_recv_sge = 1;
867 init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
868 init_qp_attr.qp_type = IB_QPT_RC;
869 init_qp_attr.send_cq = cq;
870 init_qp_attr.recv_cq = cq;
872 if (peer_ni->ibp_queue_depth_mod &&
873 peer_ni->ibp_queue_depth_mod < peer_ni->ibp_queue_depth) {
874 conn->ibc_queue_depth = peer_ni->ibp_queue_depth_mod;
875 CDEBUG(D_NET, "Use reduced queue depth %u (from %u)\n",
876 peer_ni->ibp_queue_depth_mod,
877 peer_ni->ibp_queue_depth);
881 /* kiblnd_send_wrs() can change the connection's queue depth if
882 * the maximum work requests for the device is maxed out
884 init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
885 init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
886 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd,
888 if (rc != -ENOMEM || conn->ibc_queue_depth < 2)
890 conn->ibc_queue_depth--;
894 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
895 "send_sge: %d, recv_sge: %d\n",
896 rc, init_qp_attr.cap.max_send_wr,
897 init_qp_attr.cap.max_recv_wr,
898 init_qp_attr.cap.max_send_sge,
899 init_qp_attr.cap.max_recv_sge);
903 conn->ibc_sched = sched;
905 if (!peer_ni->ibp_queue_depth_mod &&
906 conn->ibc_queue_depth != peer_ni->ibp_queue_depth) {
907 CWARN("peer %s - queue depth reduced from %u to %u"
908 " to allow for qp creation\n",
909 libcfs_nid2str(peer_ni->ibp_nid),
910 peer_ni->ibp_queue_depth,
911 conn->ibc_queue_depth);
912 peer_ni->ibp_queue_depth_mod = conn->ibc_queue_depth;
915 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
916 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
917 if (conn->ibc_rxs == NULL) {
918 CERROR("Cannot allocate RX buffers\n");
922 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
923 IBLND_RX_MSG_PAGES(conn));
927 kiblnd_map_rx_descs(conn);
929 /* 1 ref for caller and each rxmsg */
930 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
931 conn->ibc_nrx = IBLND_RX_MSGS(conn);
934 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
935 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
937 CERROR("Can't post rxmsg: %d\n", rc);
939 /* Make posted receives complete */
940 kiblnd_abort_receives(conn);
942 /* correct # of posted buffers
943 * NB locking needed now I'm racing with completion */
944 spin_lock_irqsave(&sched->ibs_lock, flags);
945 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
946 spin_unlock_irqrestore(&sched->ibs_lock, flags);
948 /* cmid will be destroyed by CM(ofed) after cm_callback
949 * returned, so we can't refer it anymore
950 * (by kiblnd_connd()->kiblnd_destroy_conn) */
951 rdma_destroy_qp(conn->ibc_cmid);
952 conn->ibc_cmid = NULL;
954 /* Drop my own and unused rxbuffer refcounts */
955 while (i++ <= IBLND_RX_MSGS(conn))
956 kiblnd_conn_decref(conn);
962 /* Init successful! */
963 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
964 state == IBLND_CONN_PASSIVE_WAIT);
965 conn->ibc_state = state;
968 atomic_inc(&net->ibn_nconns);
972 kiblnd_destroy_conn(conn);
973 LIBCFS_FREE(conn, sizeof(*conn));
979 kiblnd_destroy_conn(struct kib_conn *conn)
981 struct rdma_cm_id *cmid = conn->ibc_cmid;
982 struct kib_peer_ni *peer_ni = conn->ibc_peer;
984 LASSERT (!in_interrupt());
985 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
986 LASSERT(list_empty(&conn->ibc_early_rxs));
987 LASSERT(list_empty(&conn->ibc_tx_noops));
988 LASSERT(list_empty(&conn->ibc_tx_queue));
989 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
990 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
991 LASSERT(list_empty(&conn->ibc_active_txs));
992 LASSERT (conn->ibc_noops_posted == 0);
993 LASSERT (conn->ibc_nsends_posted == 0);
995 switch (conn->ibc_state) {
997 /* conn must be completely disengaged from the network */
1000 case IBLND_CONN_DISCONNECTED:
1001 /* connvars should have been freed already */
1002 LASSERT (conn->ibc_connvars == NULL);
1005 case IBLND_CONN_INIT:
1009 /* conn->ibc_cmid might be destroyed by CM already */
1010 if (cmid != NULL && cmid->qp != NULL)
1011 rdma_destroy_qp(cmid);
1014 ib_destroy_cq(conn->ibc_cq);
1016 kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1017 LNET_MSG_STATUS_OK);
1019 if (conn->ibc_rx_pages != NULL)
1020 kiblnd_unmap_rx_descs(conn);
1022 if (conn->ibc_rxs != NULL)
1023 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1025 if (conn->ibc_connvars != NULL)
1026 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1028 if (conn->ibc_hdev != NULL)
1029 kiblnd_hdev_decref(conn->ibc_hdev);
1031 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1032 if (conn->ibc_state != IBLND_CONN_INIT) {
1033 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1035 kiblnd_peer_decref(peer_ni);
1036 rdma_destroy_id(cmid);
1037 atomic_dec(&net->ibn_nconns);
1042 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1044 struct kib_conn *conn;
1045 struct kib_conn *cnxt;
1048 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1050 CDEBUG(D_NET, "Closing conn -> %s, "
1051 "version: %x, reason: %d\n",
1052 libcfs_nid2str(peer_ni->ibp_nid),
1053 conn->ibc_version, why);
1055 kiblnd_close_conn_locked(conn, why);
1063 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1064 int version, __u64 incarnation)
1066 struct kib_conn *conn;
1067 struct kib_conn *cnxt;
1070 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1072 if (conn->ibc_version == version &&
1073 conn->ibc_incarnation == incarnation)
1076 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1077 "incarnation:%#llx(%x, %#llx)\n",
1078 libcfs_nid2str(peer_ni->ibp_nid),
1079 conn->ibc_version, conn->ibc_incarnation,
1080 version, incarnation);
1082 kiblnd_close_conn_locked(conn, -ESTALE);
1090 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1092 struct kib_peer_ni *peer_ni;
1093 struct hlist_node *pnxt;
1097 unsigned long flags;
1100 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1102 if (nid != LNET_NID_ANY) {
1103 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
1107 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
1110 for (i = lo; i <= hi; i++) {
1111 hlist_for_each_entry_safe(peer_ni, pnxt,
1112 &kiblnd_data.kib_peers[i], ibp_list) {
1113 LASSERT(!kiblnd_peer_idle(peer_ni));
1115 if (peer_ni->ibp_ni != ni)
1118 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1121 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1125 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1127 /* wildcards always succeed */
1128 if (nid == LNET_NID_ANY)
1131 return (count == 0) ? -ENOENT : 0;
1135 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1137 struct libcfs_ioctl_data *data = arg;
1141 case IOC_LIBCFS_GET_PEER: {
1145 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1147 data->ioc_nid = nid;
1148 data->ioc_count = count;
1152 case IOC_LIBCFS_DEL_PEER: {
1153 rc = kiblnd_del_peer(ni, data->ioc_nid);
1156 case IOC_LIBCFS_GET_CONN: {
1157 struct kib_conn *conn;
1160 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1166 LASSERT(conn->ibc_cmid != NULL);
1167 data->ioc_nid = conn->ibc_peer->ibp_nid;
1168 if (conn->ibc_cmid->route.path_rec == NULL)
1169 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1172 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1173 kiblnd_conn_decref(conn);
1176 case IOC_LIBCFS_CLOSE_CONNECTION: {
1177 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1189 kiblnd_free_pages(struct kib_pages *p)
1191 int npages = p->ibp_npages;
1194 for (i = 0; i < npages; i++) {
1195 if (p->ibp_pages[i] != NULL)
1196 __free_page(p->ibp_pages[i]);
1199 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1203 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1205 struct kib_pages *p;
1208 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1209 offsetof(struct kib_pages, ibp_pages[npages]));
1211 CERROR("Can't allocate descriptor for %d pages\n", npages);
1215 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1216 p->ibp_npages = npages;
1218 for (i = 0; i < npages; i++) {
1219 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1221 if (p->ibp_pages[i] == NULL) {
1222 CERROR("Can't allocate page %d of %d\n", i, npages);
1223 kiblnd_free_pages(p);
1233 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1238 LASSERT (conn->ibc_rxs != NULL);
1239 LASSERT (conn->ibc_hdev != NULL);
1241 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1242 rx = &conn->ibc_rxs[i];
1244 LASSERT(rx->rx_nob >= 0); /* not posted */
1246 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1247 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1249 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1252 kiblnd_free_pages(conn->ibc_rx_pages);
1254 conn->ibc_rx_pages = NULL;
1258 kiblnd_map_rx_descs(struct kib_conn *conn)
1266 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1267 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1268 rx = &conn->ibc_rxs[i];
1271 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1274 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1275 rx->rx_msg, IBLND_MSG_SIZE,
1277 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1279 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1281 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1282 i, rx->rx_msg, rx->rx_msgaddr,
1283 (__u64)(page_to_phys(pg) + pg_off));
1285 pg_off += IBLND_MSG_SIZE;
1286 LASSERT(pg_off <= PAGE_SIZE);
1288 if (pg_off == PAGE_SIZE) {
1291 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1297 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1299 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1303 LASSERT (tpo->tpo_pool.po_allocated == 0);
1308 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1309 tx = &tpo->tpo_tx_descs[i];
1310 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1311 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1313 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1316 kiblnd_hdev_decref(hdev);
1317 tpo->tpo_hdev = NULL;
1320 static struct kib_hca_dev *
1321 kiblnd_current_hdev(struct kib_dev *dev)
1323 struct kib_hca_dev *hdev;
1324 unsigned long flags;
1327 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1328 while (dev->ibd_failover) {
1329 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1331 CDEBUG(D_NET, "%s: Wait for failover\n",
1333 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1335 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1338 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1339 hdev = dev->ibd_hdev;
1341 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1347 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1349 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1350 struct kib_pool *pool = &tpo->tpo_pool;
1351 struct kib_net *net = pool->po_owner->ps_net;
1352 struct kib_dev *dev;
1359 LASSERT (net != NULL);
1363 /* pre-mapped messages are not bigger than 1 page */
1364 BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1366 /* No fancy arithmetic when we do the buffer calculations */
1367 BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1369 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1371 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1372 page = txpgs->ibp_pages[ipage];
1373 tx = &tpo->tpo_tx_descs[i];
1375 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1378 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1382 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1384 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1386 list_add(&tx->tx_list, &pool->po_free_list);
1388 page_offset += IBLND_MSG_SIZE;
1389 LASSERT(page_offset <= PAGE_SIZE);
1391 if (page_offset == PAGE_SIZE) {
1394 LASSERT(ipage <= txpgs->ibp_npages);
1400 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1402 LASSERT(fpo->fpo_map_count == 0);
1404 #ifdef HAVE_FMR_POOL_API
1405 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1406 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1408 #endif /* HAVE_FMR_POOL_API */
1410 struct kib_fast_reg_descriptor *frd, *tmp;
1413 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1415 list_del(&frd->frd_list);
1416 #ifndef HAVE_IB_MAP_MR_SG
1417 ib_free_fast_reg_page_list(frd->frd_frpl);
1419 ib_dereg_mr(frd->frd_mr);
1420 LIBCFS_FREE(frd, sizeof(*frd));
1423 if (i < fpo->fast_reg.fpo_pool_size)
1424 CERROR("FastReg pool still has %d regions registered\n",
1425 fpo->fast_reg.fpo_pool_size - i);
1429 kiblnd_hdev_decref(fpo->fpo_hdev);
1431 LIBCFS_FREE(fpo, sizeof(*fpo));
1435 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1437 struct kib_fmr_pool *fpo, *tmp;
1439 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1440 list_del(&fpo->fpo_list);
1441 kiblnd_destroy_fmr_pool(fpo);
1446 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1449 int size = tunables->lnd_fmr_pool_size / ncpts;
1451 return max(IBLND_FMR_POOL, size);
1455 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1458 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1460 return max(IBLND_FMR_POOL_FLUSH, size);
1463 #ifdef HAVE_FMR_POOL_API
1464 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1465 struct kib_fmr_pool *fpo)
1467 struct ib_fmr_pool_param param = {
1468 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1469 .page_shift = PAGE_SHIFT,
1470 .access = (IB_ACCESS_LOCAL_WRITE |
1471 IB_ACCESS_REMOTE_WRITE),
1472 .pool_size = fps->fps_pool_size,
1473 .dirty_watermark = fps->fps_flush_trigger,
1474 .flush_function = NULL,
1476 .cache = !!fps->fps_cache };
1479 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1481 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1482 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1484 CERROR("Failed to create FMR pool: %d\n", rc);
1486 CERROR("FMRs are not supported\n");
1488 fpo->fpo_is_fmr = true;
1492 #endif /* HAVE_FMR_POOL_API */
1494 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1495 struct kib_fmr_pool *fpo,
1496 enum kib_dev_caps dev_caps)
1498 struct kib_fast_reg_descriptor *frd, *tmp;
1501 #ifdef HAVE_FMR_POOL_API
1502 fpo->fpo_is_fmr = false;
1505 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1506 fpo->fast_reg.fpo_pool_size = 0;
1507 for (i = 0; i < fps->fps_pool_size; i++) {
1508 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1511 CERROR("Failed to allocate a new fast_reg descriptor\n");
1517 #ifndef HAVE_IB_MAP_MR_SG
1518 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1519 IBLND_MAX_RDMA_FRAGS);
1520 if (IS_ERR(frd->frd_frpl)) {
1521 rc = PTR_ERR(frd->frd_frpl);
1522 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1524 frd->frd_frpl = NULL;
1529 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1530 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1531 IBLND_MAX_RDMA_FRAGS);
1534 * it is expected to get here if this is an MLX-5 card.
1535 * MLX-4 cards will always use FMR and MLX-5 cards will
1536 * always use fast_reg. It turns out that some MLX-5 cards
1537 * (possibly due to older FW versions) do not natively support
1538 * gaps. So we will need to track them here.
1540 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1541 #ifdef IB_MR_TYPE_SG_GAPS
1542 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1543 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1544 IB_MR_TYPE_SG_GAPS :
1549 IBLND_MAX_RDMA_FRAGS);
1550 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1551 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1552 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1554 if (IS_ERR(frd->frd_mr)) {
1555 rc = PTR_ERR(frd->frd_mr);
1556 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1561 /* There appears to be a bug in MLX5 code where you must
1562 * invalidate the rkey of a new FastReg pool before first
1563 * using it. Thus, I am marking the FRD invalid here. */
1564 frd->frd_valid = false;
1566 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1567 fpo->fast_reg.fpo_pool_size++;
1574 ib_dereg_mr(frd->frd_mr);
1575 #ifndef HAVE_IB_MAP_MR_SG
1577 ib_free_fast_reg_page_list(frd->frd_frpl);
1579 LIBCFS_FREE(frd, sizeof(*frd));
1582 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1584 list_del(&frd->frd_list);
1585 #ifndef HAVE_IB_MAP_MR_SG
1586 ib_free_fast_reg_page_list(frd->frd_frpl);
1588 ib_dereg_mr(frd->frd_mr);
1589 LIBCFS_FREE(frd, sizeof(*frd));
1595 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1596 struct kib_fmr_pool **pp_fpo)
1598 struct kib_dev *dev = fps->fps_net->ibn_dev;
1599 struct kib_fmr_pool *fpo;
1602 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1606 memset(fpo, 0, sizeof(*fpo));
1608 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1610 #ifdef HAVE_FMR_POOL_API
1611 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1612 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1614 #endif /* HAVE_FMR_POOL_API */
1615 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1619 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1620 fpo->fpo_owner = fps;
1626 kiblnd_hdev_decref(fpo->fpo_hdev);
1627 LIBCFS_FREE(fpo, sizeof(*fpo));
1632 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1634 struct kib_fmr_pool *fpo;
1636 if (fps->fps_net == NULL) /* intialized? */
1639 spin_lock(&fps->fps_lock);
1641 while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1642 struct kib_fmr_pool,
1643 fpo_list)) != NULL) {
1644 fpo->fpo_failed = 1;
1645 if (fpo->fpo_map_count == 0)
1646 list_move(&fpo->fpo_list, zombies);
1648 list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1651 spin_unlock(&fps->fps_lock);
1655 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1657 if (fps->fps_net != NULL) { /* initialized? */
1658 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1659 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1664 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1665 struct kib_net *net,
1666 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1668 struct kib_fmr_pool *fpo;
1671 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1676 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1677 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1678 fps->fps_cache = tunables->lnd_fmr_cache;
1680 spin_lock_init(&fps->fps_lock);
1681 INIT_LIST_HEAD(&fps->fps_pool_list);
1682 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1684 rc = kiblnd_create_fmr_pool(fps, &fpo);
1686 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1692 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1694 if (fpo->fpo_map_count != 0) /* still in use */
1696 if (fpo->fpo_failed)
1698 return now >= fpo->fpo_deadline;
1701 #if defined(HAVE_FMR_POOL_API) || !defined(HAVE_IB_MAP_MR_SG)
1703 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1705 struct kib_hca_dev *hdev;
1706 __u64 *pages = tx->tx_pages;
1711 hdev = tx->tx_pool->tpo_hdev;
1713 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1714 for (size = 0; size < rd->rd_frags[i].rf_nob;
1715 size += hdev->ibh_page_size) {
1716 pages[npages++] = (rd->rd_frags[i].rf_addr &
1717 hdev->ibh_page_mask) + size;
1726 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1729 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1730 struct kib_fmr_poolset *fps;
1731 time64_t now = ktime_get_seconds();
1732 struct kib_fmr_pool *tmp;
1737 fps = fpo->fpo_owner;
1739 #ifdef HAVE_FMR_POOL_API
1740 if (fpo->fpo_is_fmr) {
1741 if (fmr->fmr_pfmr) {
1742 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1743 fmr->fmr_pfmr = NULL;
1747 int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1751 #endif /* HAVE_FMR_POOL_API */
1753 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1756 frd->frd_valid = false;
1757 frd->frd_posted = false;
1758 fmr->fmr_frd = NULL;
1759 spin_lock(&fps->fps_lock);
1760 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1761 spin_unlock(&fps->fps_lock);
1764 fmr->fmr_pool = NULL;
1766 spin_lock(&fps->fps_lock);
1767 fpo->fpo_map_count--; /* decref the pool */
1769 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1770 /* the first pool is persistent */
1771 if (fps->fps_pool_list.next == &fpo->fpo_list)
1774 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1775 list_move(&fpo->fpo_list, &zombies);
1779 spin_unlock(&fps->fps_lock);
1781 if (!list_empty(&zombies))
1782 kiblnd_destroy_fmr_pool_list(&zombies);
1785 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1786 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1787 struct kib_fmr *fmr)
1789 struct kib_fmr_pool *fpo;
1791 bool is_rx = (rd != tx->tx_rd);
1792 #ifdef HAVE_FMR_POOL_API
1793 __u64 *pages = tx->tx_pages;
1794 bool tx_pages_mapped = false;
1800 spin_lock(&fps->fps_lock);
1801 version = fps->fps_version;
1802 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1803 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1804 fpo->fpo_map_count++;
1806 #ifdef HAVE_FMR_POOL_API
1807 fmr->fmr_pfmr = NULL;
1808 if (fpo->fpo_is_fmr) {
1809 struct ib_pool_fmr *pfmr;
1811 spin_unlock(&fps->fps_lock);
1813 if (!tx_pages_mapped) {
1814 npages = kiblnd_map_tx_pages(tx, rd);
1815 tx_pages_mapped = true;
1818 pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1819 pages, npages, iov);
1820 if (likely(!IS_ERR(pfmr))) {
1821 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1823 fmr->fmr_frd = NULL;
1824 fmr->fmr_pfmr = pfmr;
1825 fmr->fmr_pool = fpo;
1830 #endif /* HAVE_FMR_POOL_API */
1832 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1833 struct kib_fast_reg_descriptor *frd;
1834 #ifdef HAVE_IB_MAP_MR_SG
1835 struct ib_reg_wr *wr;
1838 struct ib_rdma_wr *wr;
1839 struct ib_fast_reg_page_list *frpl;
1843 frd = list_first_entry(
1844 &fpo->fast_reg.fpo_pool_list,
1845 struct kib_fast_reg_descriptor,
1847 list_del(&frd->frd_list);
1848 spin_unlock(&fps->fps_lock);
1850 #ifndef HAVE_IB_MAP_MR_SG
1851 frpl = frd->frd_frpl;
1855 if (!frd->frd_valid) {
1856 struct ib_rdma_wr *inv_wr;
1857 __u32 key = is_rx ? mr->rkey : mr->lkey;
1859 inv_wr = &frd->frd_inv_wr;
1860 memset(inv_wr, 0, sizeof(*inv_wr));
1862 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1863 inv_wr->wr.wr_id = IBLND_WID_MR;
1864 inv_wr->wr.ex.invalidate_rkey = key;
1867 key = ib_inc_rkey(key);
1868 ib_update_fast_reg_key(mr, key);
1871 #ifdef HAVE_IB_MAP_MR_SG
1872 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1873 n = ib_map_mr_sg(mr, tx->tx_frags,
1874 rd->rd_nfrags, NULL, PAGE_SIZE);
1876 n = ib_map_mr_sg(mr, tx->tx_frags,
1877 rd->rd_nfrags, PAGE_SIZE);
1878 #endif /* HAVE_IB_MAP_MR_SG_5ARGS */
1879 if (unlikely(n != rd->rd_nfrags)) {
1880 CERROR("Failed to map mr %d/%d elements\n",
1882 return n < 0 ? n : -EINVAL;
1885 wr = &frd->frd_fastreg_wr;
1886 memset(wr, 0, sizeof(*wr));
1888 wr->wr.opcode = IB_WR_REG_MR;
1889 wr->wr.wr_id = IBLND_WID_MR;
1891 wr->wr.send_flags = 0;
1893 wr->key = is_rx ? mr->rkey : mr->lkey;
1894 wr->access = (IB_ACCESS_LOCAL_WRITE |
1895 IB_ACCESS_REMOTE_WRITE);
1896 #else /* HAVE_IB_MAP_MR_SG */
1897 if (!tx_pages_mapped) {
1898 npages = kiblnd_map_tx_pages(tx, rd);
1899 tx_pages_mapped = true;
1902 LASSERT(npages <= frpl->max_page_list_len);
1903 memcpy(frpl->page_list, pages,
1904 sizeof(*pages) * npages);
1906 /* Prepare FastReg WR */
1907 wr = &frd->frd_fastreg_wr;
1908 memset(wr, 0, sizeof(*wr));
1910 wr->wr.opcode = IB_WR_FAST_REG_MR;
1911 wr->wr.wr_id = IBLND_WID_MR;
1913 wr->wr.wr.fast_reg.iova_start = iov;
1914 wr->wr.wr.fast_reg.page_list = frpl;
1915 wr->wr.wr.fast_reg.page_list_len = npages;
1916 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1917 wr->wr.wr.fast_reg.length = nob;
1918 wr->wr.wr.fast_reg.rkey =
1919 is_rx ? mr->rkey : mr->lkey;
1920 wr->wr.wr.fast_reg.access_flags =
1921 (IB_ACCESS_LOCAL_WRITE |
1922 IB_ACCESS_REMOTE_WRITE);
1923 #endif /* HAVE_IB_MAP_MR_SG */
1925 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1927 fmr->fmr_pool = fpo;
1928 frd->frd_posted = false;
1931 spin_unlock(&fps->fps_lock);
1935 spin_lock(&fps->fps_lock);
1936 fpo->fpo_map_count--;
1937 if (rc != -EAGAIN) {
1938 spin_unlock(&fps->fps_lock);
1942 /* EAGAIN and ... */
1943 if (version != fps->fps_version) {
1944 spin_unlock(&fps->fps_lock);
1949 if (fps->fps_increasing) {
1950 spin_unlock(&fps->fps_lock);
1951 CDEBUG(D_NET, "Another thread is allocating new "
1952 "FMR pool, waiting for her to complete\n");
1953 wait_var_event(fps, !fps->fps_increasing);
1958 if (ktime_get_seconds() < fps->fps_next_retry) {
1959 /* someone failed recently */
1960 spin_unlock(&fps->fps_lock);
1964 fps->fps_increasing = 1;
1965 spin_unlock(&fps->fps_lock);
1967 CDEBUG(D_NET, "Allocate new FMR pool\n");
1968 rc = kiblnd_create_fmr_pool(fps, &fpo);
1969 spin_lock(&fps->fps_lock);
1970 fps->fps_increasing = 0;
1974 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1976 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
1978 spin_unlock(&fps->fps_lock);
1984 kiblnd_fini_pool(struct kib_pool *pool)
1986 LASSERT(list_empty(&pool->po_free_list));
1987 LASSERT(pool->po_allocated == 0);
1989 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1993 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
1995 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1997 memset(pool, 0, sizeof(struct kib_pool));
1998 INIT_LIST_HEAD(&pool->po_free_list);
1999 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2000 pool->po_owner = ps;
2001 pool->po_size = size;
2005 kiblnd_destroy_pool_list(struct list_head *head)
2007 struct kib_pool *pool;
2009 while ((pool = list_first_entry_or_null(head,
2011 po_list)) != NULL) {
2012 list_del(&pool->po_list);
2014 LASSERT(pool->po_owner != NULL);
2015 pool->po_owner->ps_pool_destroy(pool);
2020 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2022 struct kib_pool *po;
2024 if (ps->ps_net == NULL) /* intialized? */
2027 spin_lock(&ps->ps_lock);
2028 while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2030 po_list)) != NULL) {
2032 if (po->po_allocated == 0)
2033 list_move(&po->po_list, zombies);
2035 list_move(&po->po_list, &ps->ps_failed_pool_list);
2037 spin_unlock(&ps->ps_lock);
2041 kiblnd_fini_poolset(struct kib_poolset *ps)
2043 if (ps->ps_net != NULL) { /* initialized? */
2044 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2045 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2050 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2051 struct kib_net *net, char *name, int size,
2052 kib_ps_pool_create_t po_create,
2053 kib_ps_pool_destroy_t po_destroy,
2054 kib_ps_node_init_t nd_init,
2055 kib_ps_node_fini_t nd_fini)
2057 struct kib_pool *pool;
2060 memset(ps, 0, sizeof(struct kib_poolset));
2064 ps->ps_pool_create = po_create;
2065 ps->ps_pool_destroy = po_destroy;
2066 ps->ps_node_init = nd_init;
2067 ps->ps_node_fini = nd_fini;
2068 ps->ps_pool_size = size;
2069 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2070 >= sizeof(ps->ps_name))
2072 spin_lock_init(&ps->ps_lock);
2073 INIT_LIST_HEAD(&ps->ps_pool_list);
2074 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2076 rc = ps->ps_pool_create(ps, size, &pool);
2078 list_add(&pool->po_list, &ps->ps_pool_list);
2080 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2086 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2088 if (pool->po_allocated != 0) /* still in use */
2090 if (pool->po_failed)
2092 return now >= pool->po_deadline;
2096 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2099 struct kib_poolset *ps = pool->po_owner;
2100 struct kib_pool *tmp;
2101 time64_t now = ktime_get_seconds();
2103 spin_lock(&ps->ps_lock);
2105 if (ps->ps_node_fini != NULL)
2106 ps->ps_node_fini(pool, node);
2108 LASSERT(pool->po_allocated > 0);
2109 list_add(node, &pool->po_free_list);
2110 pool->po_allocated--;
2112 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2113 /* the first pool is persistent */
2114 if (ps->ps_pool_list.next == &pool->po_list)
2117 if (kiblnd_pool_is_idle(pool, now))
2118 list_move(&pool->po_list, &zombies);
2120 spin_unlock(&ps->ps_lock);
2122 if (!list_empty(&zombies))
2123 kiblnd_destroy_pool_list(&zombies);
2127 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2129 struct list_head *node;
2130 struct kib_pool *pool;
2132 unsigned int interval = 1;
2133 ktime_t time_before;
2134 unsigned int trips = 0;
2137 spin_lock(&ps->ps_lock);
2138 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2139 if (list_empty(&pool->po_free_list))
2142 pool->po_allocated++;
2143 pool->po_deadline = ktime_get_seconds() +
2144 IBLND_POOL_DEADLINE;
2145 node = pool->po_free_list.next;
2148 if (ps->ps_node_init != NULL) {
2149 /* still hold the lock */
2150 ps->ps_node_init(pool, node);
2152 spin_unlock(&ps->ps_lock);
2156 /* no available tx pool and ... */
2157 if (ps->ps_increasing) {
2158 /* another thread is allocating a new pool */
2159 spin_unlock(&ps->ps_lock);
2162 "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2163 ps->ps_name, interval, trips);
2165 schedule_timeout_interruptible(interval);
2166 if (interval < cfs_time_seconds(1))
2172 if (ktime_get_seconds() < ps->ps_next_retry) {
2173 /* someone failed recently */
2174 spin_unlock(&ps->ps_lock);
2178 ps->ps_increasing = 1;
2179 spin_unlock(&ps->ps_lock);
2181 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2182 time_before = ktime_get();
2183 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2184 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2185 ktime_ms_delta(ktime_get(), time_before));
2187 spin_lock(&ps->ps_lock);
2188 ps->ps_increasing = 0;
2190 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2192 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2193 CERROR("Can't allocate new %s pool because out of memory\n",
2196 spin_unlock(&ps->ps_lock);
2202 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2204 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2208 LASSERT (pool->po_allocated == 0);
2210 if (tpo->tpo_tx_pages != NULL) {
2211 kiblnd_unmap_tx_pool(tpo);
2212 kiblnd_free_pages(tpo->tpo_tx_pages);
2215 if (tpo->tpo_tx_descs == NULL)
2218 for (i = 0; i < pool->po_size; i++) {
2219 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2220 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2222 list_del(&tx->tx_list);
2223 if (tx->tx_pages != NULL)
2224 CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2225 if (tx->tx_frags != NULL)
2226 CFS_FREE_PTR_ARRAY(tx->tx_frags,
2227 IBLND_MAX_RDMA_FRAGS);
2228 if (tx->tx_wrq != NULL)
2229 CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2230 IBLND_MAX_RDMA_FRAGS);
2231 if (tx->tx_sge != NULL)
2232 CFS_FREE_PTR_ARRAY(tx->tx_sge,
2233 IBLND_MAX_RDMA_FRAGS *
2235 if (tx->tx_rd != NULL)
2236 LIBCFS_FREE(tx->tx_rd,
2237 offsetof(struct kib_rdma_desc,
2238 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2241 CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2243 kiblnd_fini_pool(pool);
2247 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2249 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2252 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2253 ntx = tunables->lnd_ntx / ncpts;
2255 return max(IBLND_TX_POOL, ntx);
2259 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2263 struct kib_pool *pool;
2264 struct kib_tx_pool *tpo;
2266 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2268 CERROR("Failed to allocate TX pool\n");
2272 pool = &tpo->tpo_pool;
2273 kiblnd_init_pool(ps, pool, size);
2274 tpo->tpo_tx_descs = NULL;
2275 tpo->tpo_tx_pages = NULL;
2277 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2278 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2279 CERROR("Can't allocate tx pages: %d\n", npg);
2284 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2285 size * sizeof(struct kib_tx));
2286 if (tpo->tpo_tx_descs == NULL) {
2287 CERROR("Can't allocate %d tx descriptors\n", size);
2288 ps->ps_pool_destroy(pool);
2292 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2294 for (i = 0; i < size; i++) {
2295 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2296 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2299 if (ps->ps_net->ibn_fmr_ps != NULL) {
2300 LIBCFS_CPT_ALLOC(tx->tx_pages,
2301 lnet_cpt_table(), ps->ps_cpt,
2302 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2303 if (tx->tx_pages == NULL)
2307 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2308 IBLND_MAX_RDMA_FRAGS *
2309 sizeof(*tx->tx_frags));
2310 if (tx->tx_frags == NULL)
2313 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2315 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2316 IBLND_MAX_RDMA_FRAGS *
2317 sizeof(*tx->tx_wrq));
2318 if (tx->tx_wrq == NULL)
2321 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2322 IBLND_MAX_RDMA_FRAGS * wrq_sge *
2323 sizeof(*tx->tx_sge));
2324 if (tx->tx_sge == NULL)
2327 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2328 offsetof(struct kib_rdma_desc,
2329 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2330 if (tx->tx_rd == NULL)
2335 kiblnd_map_tx_pool(tpo);
2340 ps->ps_pool_destroy(pool);
2345 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2347 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2348 struct kib_tx_poolset,
2350 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2352 tx->tx_cookie = tps->tps_next_tx_cookie++;
2356 kiblnd_net_fini_pools(struct kib_net *net)
2360 cfs_cpt_for_each(i, lnet_cpt_table()) {
2361 struct kib_tx_poolset *tps;
2362 struct kib_fmr_poolset *fps;
2364 if (net->ibn_tx_ps != NULL) {
2365 tps = net->ibn_tx_ps[i];
2366 kiblnd_fini_poolset(&tps->tps_poolset);
2369 if (net->ibn_fmr_ps != NULL) {
2370 fps = net->ibn_fmr_ps[i];
2371 kiblnd_fini_fmr_poolset(fps);
2375 if (net->ibn_tx_ps != NULL) {
2376 cfs_percpt_free(net->ibn_tx_ps);
2377 net->ibn_tx_ps = NULL;
2380 if (net->ibn_fmr_ps != NULL) {
2381 cfs_percpt_free(net->ibn_fmr_ps);
2382 net->ibn_fmr_ps = NULL;
2387 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2390 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2391 #ifdef HAVE_IB_GET_DMA_MR
2392 unsigned long flags;
2398 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2400 #ifdef HAVE_IB_GET_DMA_MR
2401 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2403 * if lnd_map_on_demand is zero then we have effectively disabled
2404 * FMR or FastReg and we're using global memory regions
2407 if (!tunables->lnd_map_on_demand) {
2408 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2410 goto create_tx_pool;
2413 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2416 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2417 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2418 tunables->lnd_fmr_pool_size,
2419 tunables->lnd_ntx / 4);
2424 /* TX pool must be created later than FMR, see LU-2268
2426 LASSERT(net->ibn_tx_ps == NULL);
2428 /* premapping can fail if ibd_nmr > 1, so we always create
2429 * FMR pool and map-on-demand if premapping failed */
2431 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2432 sizeof(struct kib_fmr_poolset));
2433 if (net->ibn_fmr_ps == NULL) {
2434 CERROR("Failed to allocate FMR pool array\n");
2439 for (i = 0; i < ncpts; i++) {
2440 cpt = (cpts == NULL) ? i : cpts[i];
2441 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2444 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2451 LASSERT(i == ncpts);
2453 #ifdef HAVE_IB_GET_DMA_MR
2456 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2457 sizeof(struct kib_tx_poolset));
2458 if (net->ibn_tx_ps == NULL) {
2459 CERROR("Failed to allocate tx pool array\n");
2464 for (i = 0; i < ncpts; i++) {
2465 cpt = (cpts == NULL) ? i : cpts[i];
2466 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2468 kiblnd_tx_pool_size(ni, ncpts),
2469 kiblnd_create_tx_pool,
2470 kiblnd_destroy_tx_pool,
2471 kiblnd_tx_init, NULL);
2473 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2481 kiblnd_net_fini_pools(net);
2487 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2489 struct ib_port_attr *port_attr;
2491 unsigned long flags;
2492 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2494 LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2495 if (port_attr == NULL) {
2496 CDEBUG(D_NETERROR, "Out of memory\n");
2500 rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2502 write_lock_irqsave(g_lock, flags);
2505 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2506 ? IBLND_DEV_PORT_ACTIVE
2507 : IBLND_DEV_PORT_DOWN;
2509 write_unlock_irqrestore(g_lock, flags);
2510 LIBCFS_FREE(port_attr, sizeof(*port_attr));
2513 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2520 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2522 struct kib_net *net;
2524 /* for health check */
2525 list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2527 CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2528 libcfs_nidstr(&net->ibn_ni->ni_nid));
2529 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2534 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2536 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2537 struct kib_hca_dev *hdev;
2538 unsigned long flags;
2540 hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2542 write_lock_irqsave(g_lock, flags);
2544 switch (event->event) {
2545 case IB_EVENT_DEVICE_FATAL:
2546 CDEBUG(D_NET, "IB device fatal\n");
2547 hdev->ibh_state = IBLND_DEV_FATAL;
2548 kiblnd_set_ni_fatal_on(hdev, 1);
2550 case IB_EVENT_PORT_ACTIVE:
2551 CDEBUG(D_NET, "IB port active\n");
2552 if (event->element.port_num == hdev->ibh_port) {
2553 hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2554 kiblnd_set_ni_fatal_on(hdev, 0);
2557 case IB_EVENT_PORT_ERR:
2558 CDEBUG(D_NET, "IB port err\n");
2559 if (event->element.port_num == hdev->ibh_port) {
2560 hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2561 kiblnd_set_ni_fatal_on(hdev, 1);
2567 write_unlock_irqrestore(g_lock, flags);
2571 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2573 struct ib_device_attr *dev_attr;
2577 /* It's safe to assume a HCA can handle a page size
2578 * matching that of the native system */
2579 hdev->ibh_page_shift = PAGE_SHIFT;
2580 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2581 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2583 #ifndef HAVE_IB_DEVICE_ATTRS
2584 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2585 if (dev_attr == NULL) {
2586 CERROR("Out of memory\n");
2590 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2592 CERROR("Failed to query IB device: %d\n", rc);
2593 goto out_clean_attr;
2596 dev_attr = &hdev->ibh_ibdev->attrs;
2599 hdev->ibh_mr_size = dev_attr->max_mr_size;
2600 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2602 /* Setup device Memory Registration capabilities */
2603 #ifdef HAVE_FMR_POOL_API
2604 #ifdef HAVE_IB_DEVICE_OPS
2605 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2606 hdev->ibh_ibdev->ops.dealloc_fmr &&
2607 hdev->ibh_ibdev->ops.map_phys_fmr &&
2608 hdev->ibh_ibdev->ops.unmap_fmr) {
2610 if (hdev->ibh_ibdev->alloc_fmr &&
2611 hdev->ibh_ibdev->dealloc_fmr &&
2612 hdev->ibh_ibdev->map_phys_fmr &&
2613 hdev->ibh_ibdev->unmap_fmr) {
2615 LCONSOLE_INFO("Using FMR for registration\n");
2616 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2618 #endif /* HAVE_FMR_POOL_API */
2619 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2620 LCONSOLE_INFO("Using FastReg for registration\n");
2621 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2622 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2623 #ifdef IB_DEVICE_SG_GAPS_REG
2624 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2625 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2632 rc2 = kiblnd_port_get_attr(hdev);
2639 #ifndef HAVE_IB_DEVICE_ATTRS
2641 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2645 CERROR("IB device does not support FMRs nor FastRegs, can't "
2646 "register memory: %d\n", rc);
2647 else if (rc == -EINVAL)
2648 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2652 #ifdef HAVE_IB_GET_DMA_MR
2654 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2656 if (hdev->ibh_mrs == NULL)
2659 ib_dereg_mr(hdev->ibh_mrs);
2661 hdev->ibh_mrs = NULL;
2666 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2668 if (hdev->ibh_event_handler.device != NULL)
2669 ib_unregister_event_handler(&hdev->ibh_event_handler);
2671 #ifdef HAVE_IB_GET_DMA_MR
2672 kiblnd_hdev_cleanup_mrs(hdev);
2675 if (hdev->ibh_pd != NULL)
2676 ib_dealloc_pd(hdev->ibh_pd);
2678 if (hdev->ibh_cmid != NULL)
2679 rdma_destroy_id(hdev->ibh_cmid);
2681 LIBCFS_FREE(hdev, sizeof(*hdev));
2684 #ifdef HAVE_IB_GET_DMA_MR
2686 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2689 int acflags = IB_ACCESS_LOCAL_WRITE |
2690 IB_ACCESS_REMOTE_WRITE;
2692 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2694 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2695 kiblnd_hdev_cleanup_mrs(hdev);
2706 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2711 static int kiblnd_get_link_status(struct net_device *dev)
2717 if (!netif_running(dev))
2719 /* Some devices may not be providing link settings */
2720 else if (dev->ethtool_ops->get_link)
2721 ret = dev->ethtool_ops->get_link(dev);
2727 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2729 struct rdma_cm_id *cmid;
2730 struct sockaddr_in srcaddr;
2731 struct sockaddr_in dstaddr;
2734 if (dev->ibd_hdev == NULL || /* initializing */
2735 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2736 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2739 /* XXX: it's UGLY, but I don't have better way to find
2740 * ib-bonding HCA failover because:
2742 * a. no reliable CM event for HCA failover...
2743 * b. no OFED API to get ib_device for current net_device...
2745 * We have only two choices at this point:
2747 * a. rdma_bind_addr(), it will conflict with listener cmid
2748 * b. rdma_resolve_addr() to zero addr */
2749 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2750 RDMA_PS_TCP, IB_QPT_RC);
2753 CERROR("Failed to create cmid for failover: %d\n", rc);
2757 memset(&srcaddr, 0, sizeof(srcaddr));
2758 srcaddr.sin_family = AF_INET;
2759 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2761 memset(&dstaddr, 0, sizeof(dstaddr));
2762 dstaddr.sin_family = AF_INET;
2763 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2764 (struct sockaddr *)&dstaddr, 1);
2765 if (rc != 0 || cmid->device == NULL) {
2766 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2767 dev->ibd_ifname, &dev->ibd_ifip,
2769 rdma_destroy_id(cmid);
2773 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2774 rdma_destroy_id(cmid);
2779 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2781 LIST_HEAD(zombie_tpo);
2782 LIST_HEAD(zombie_ppo);
2783 LIST_HEAD(zombie_fpo);
2784 struct rdma_cm_id *cmid = NULL;
2785 struct kib_hca_dev *hdev = NULL;
2786 struct kib_hca_dev *old;
2788 struct kib_net *net;
2789 struct sockaddr_in addr;
2790 struct net_device *netdev;
2791 unsigned long flags;
2795 LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2796 dev->ibd_can_failover ||
2797 dev->ibd_hdev == NULL);
2799 rc = kiblnd_dev_need_failover(dev, ns);
2803 if (dev->ibd_hdev != NULL &&
2804 dev->ibd_hdev->ibh_cmid != NULL) {
2805 /* XXX it's not good to close old listener at here,
2806 * because we can fail to create new listener.
2807 * But we have to close it now, otherwise rdma_bind_addr
2808 * will return EADDRINUSE... How crap! */
2809 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2811 cmid = dev->ibd_hdev->ibh_cmid;
2812 /* make next schedule of kiblnd_dev_need_failover()
2813 * return 1 for me */
2814 dev->ibd_hdev->ibh_cmid = NULL;
2815 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2817 rdma_destroy_id(cmid);
2820 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2824 CERROR("Failed to create cmid for failover: %d\n", rc);
2828 memset(&addr, 0, sizeof(addr));
2829 addr.sin_family = AF_INET;
2830 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2831 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2833 /* Bind to failover device or port */
2834 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2835 if (rc != 0 || cmid->device == NULL) {
2836 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2837 dev->ibd_ifname, &dev->ibd_ifip,
2839 rdma_destroy_id(cmid);
2843 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2845 CERROR("Failed to allocate kib_hca_dev\n");
2846 rdma_destroy_id(cmid);
2851 atomic_set(&hdev->ibh_ref, 1);
2852 hdev->ibh_dev = dev;
2853 hdev->ibh_cmid = cmid;
2854 hdev->ibh_ibdev = cmid->device;
2855 hdev->ibh_port = cmid->port_num;
2857 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2858 pd = ib_alloc_pd(cmid->device, 0);
2860 pd = ib_alloc_pd(cmid->device);
2864 CERROR("Can't allocate PD: %d\n", rc);
2870 rc = rdma_listen(cmid, 0);
2872 CERROR("Can't start new listener: %d\n", rc);
2876 rc = kiblnd_hdev_get_attr(hdev);
2878 CERROR("Can't get device attributes: %d\n", rc);
2882 #ifdef HAVE_IB_GET_DMA_MR
2883 rc = kiblnd_hdev_setup_mrs(hdev);
2885 CERROR("Can't setup device: %d\n", rc);
2890 INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
2891 hdev->ibh_ibdev, kiblnd_event_handler);
2892 ib_register_event_handler(&hdev->ibh_event_handler);
2894 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2896 old = dev->ibd_hdev;
2897 dev->ibd_hdev = hdev; /* take over the refcount */
2900 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2901 cfs_cpt_for_each(i, lnet_cpt_table()) {
2902 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2905 if (net->ibn_fmr_ps != NULL)
2906 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2911 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2913 if (!list_empty(&zombie_tpo))
2914 kiblnd_destroy_pool_list(&zombie_tpo);
2915 if (!list_empty(&zombie_ppo))
2916 kiblnd_destroy_pool_list(&zombie_ppo);
2917 if (!list_empty(&zombie_fpo))
2918 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2920 kiblnd_hdev_decref(hdev);
2923 dev->ibd_failed_failover++;
2925 dev->ibd_failed_failover = 0;
2928 netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
2929 if (netdev && (kiblnd_get_link_status(netdev) == 1))
2930 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
2938 kiblnd_destroy_dev(struct kib_dev *dev)
2940 LASSERT(dev->ibd_nnets == 0);
2941 LASSERT(list_empty(&dev->ibd_nets));
2943 list_del(&dev->ibd_fail_list);
2944 list_del(&dev->ibd_list);
2946 if (dev->ibd_hdev != NULL)
2947 kiblnd_hdev_decref(dev->ibd_hdev);
2949 LIBCFS_FREE(dev, sizeof(*dev));
2953 kiblnd_base_shutdown(void)
2955 struct kib_sched_info *sched;
2956 struct kib_peer_ni *peer_ni;
2959 LASSERT(list_empty(&kiblnd_data.kib_devs));
2961 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
2962 libcfs_kmem_read());
2964 switch (kiblnd_data.kib_init) {
2968 case IBLND_INIT_ALL:
2969 case IBLND_INIT_DATA:
2970 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
2972 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2973 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2974 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2975 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2977 /* flag threads to terminate; wake and wait for them to die */
2978 kiblnd_data.kib_shutdown = 1;
2980 /* NB: we really want to stop scheduler threads net by net
2981 * instead of the whole module, this should be improved
2982 * with dynamic configuration LNet.
2984 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2985 wake_up_all(&sched->ibs_waitq);
2987 wake_up(&kiblnd_data.kib_connd_waitq);
2988 wake_up(&kiblnd_data.kib_failover_waitq);
2990 wait_var_event_warning(&kiblnd_data.kib_nthreads,
2991 !atomic_read(&kiblnd_data.kib_nthreads),
2992 "Waiting for %d threads to terminate\n",
2993 atomic_read(&kiblnd_data.kib_nthreads));
2996 case IBLND_INIT_NOTHING:
3000 if (kiblnd_data.kib_scheds != NULL)
3001 cfs_percpt_free(kiblnd_data.kib_scheds);
3003 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3004 libcfs_kmem_read());
3006 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3007 module_put(THIS_MODULE);
3011 kiblnd_shutdown(struct lnet_ni *ni)
3013 struct kib_net *net = ni->ni_data;
3014 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
3015 unsigned long flags;
3017 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3022 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3023 libcfs_kmem_read());
3025 write_lock_irqsave(g_lock, flags);
3026 net->ibn_shutdown = 1;
3027 write_unlock_irqrestore(g_lock, flags);
3029 switch (net->ibn_init) {
3033 case IBLND_INIT_ALL:
3034 /* nuke all existing peers within this net */
3035 kiblnd_del_peer(ni, LNET_NID_ANY);
3037 /* Wait for all peer_ni state to clean up */
3038 wait_var_event_warning(&net->ibn_npeers,
3039 atomic_read(&net->ibn_npeers) == 0,
3040 "%s: waiting for %d peers to disconnect\n",
3041 libcfs_nidstr(&ni->ni_nid),
3042 atomic_read(&net->ibn_npeers));
3044 kiblnd_net_fini_pools(net);
3046 write_lock_irqsave(g_lock, flags);
3047 LASSERT(net->ibn_dev->ibd_nnets > 0);
3048 net->ibn_dev->ibd_nnets--;
3049 list_del(&net->ibn_list);
3050 write_unlock_irqrestore(g_lock, flags);
3054 case IBLND_INIT_NOTHING:
3055 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3057 if (net->ibn_dev != NULL &&
3058 net->ibn_dev->ibd_nnets == 0)
3059 kiblnd_destroy_dev(net->ibn_dev);
3064 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3065 libcfs_kmem_read());
3067 net->ibn_init = IBLND_INIT_NOTHING;
3070 LIBCFS_FREE(net, sizeof(*net));
3073 if (list_empty(&kiblnd_data.kib_devs))
3074 kiblnd_base_shutdown();
3078 kiblnd_base_startup(struct net *ns)
3080 struct kib_sched_info *sched;
3084 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3086 if (!try_module_get(THIS_MODULE))
3089 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3091 rwlock_init(&kiblnd_data.kib_global_lock);
3093 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3094 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3096 hash_init(kiblnd_data.kib_peers);
3098 spin_lock_init(&kiblnd_data.kib_connd_lock);
3099 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3100 INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3101 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3102 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3103 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3105 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3106 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3108 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3110 if (kiblnd_data.kib_scheds == NULL)
3113 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3116 spin_lock_init(&sched->ibs_lock);
3117 INIT_LIST_HEAD(&sched->ibs_conns);
3118 init_waitqueue_head(&sched->ibs_waitq);
3120 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3121 if (*kiblnd_tunables.kib_nscheds > 0) {
3122 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3124 /* max to half of CPUs, another half is reserved for
3125 * upper layer modules */
3126 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3129 sched->ibs_nthreads_max = nthrs;
3133 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3135 /* lists/ptrs/locks initialised */
3136 kiblnd_data.kib_init = IBLND_INIT_DATA;
3137 /*****************************************************/
3139 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3141 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3145 if (*kiblnd_tunables.kib_dev_failover != 0)
3146 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3150 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3154 /* flag everything initialised */
3155 kiblnd_data.kib_init = IBLND_INIT_ALL;
3156 /*****************************************************/
3161 kiblnd_base_shutdown();
3166 kiblnd_start_schedulers(struct kib_sched_info *sched)
3172 if (sched->ibs_nthreads == 0) {
3173 if (*kiblnd_tunables.kib_nscheds > 0) {
3174 nthrs = sched->ibs_nthreads_max;
3176 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3178 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3179 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3182 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3183 /* increase one thread if there is new interface */
3184 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3187 for (i = 0; i < nthrs; i++) {
3188 long id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3190 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3191 "kiblnd_sd_%02ld_%02ld",
3192 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3196 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3197 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3201 sched->ibs_nthreads += i;
3205 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3212 for (i = 0; i < ncpts; i++) {
3213 struct kib_sched_info *sched;
3215 cpt = (cpts == NULL) ? i : cpts[i];
3216 sched = kiblnd_data.kib_scheds[cpt];
3218 if (!newdev && sched->ibs_nthreads > 0)
3221 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3223 CERROR("Failed to start scheduler threads for %s\n",
3231 static struct kib_dev *
3232 kiblnd_dev_search(char *ifname)
3234 struct kib_dev *alias = NULL;
3235 struct kib_dev *dev;
3239 colon = strchr(ifname, ':');
3240 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3241 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3247 colon2 = strchr(dev->ibd_ifname, ':');
3253 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3265 kiblnd_startup(struct lnet_ni *ni)
3267 char *ifname = NULL;
3268 struct lnet_inetdev *ifaces = NULL;
3269 struct kib_dev *ibdev = NULL;
3270 struct kib_net *net = NULL;
3271 unsigned long flags;
3276 LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3278 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3279 rc = kiblnd_base_startup(ni->ni_net_ns);
3284 LIBCFS_ALLOC(net, sizeof(*net));
3292 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3294 kiblnd_tunables_setup(ni);
3297 * Multi-Rail wants each secondary
3298 * IP to be treated as an unique 'struct ni' interface.
3300 if (ni->ni_interface != NULL) {
3301 /* Use the IPoIB interface specified in 'networks=' */
3302 ifname = ni->ni_interface;
3304 ifname = *kiblnd_tunables.kib_default_ipif;
3307 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3308 CERROR("IPoIB interface name too long: %s\n", ifname);
3313 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3317 for (i = 0; i < rc; i++) {
3318 if (strcmp(ifname, ifaces[i].li_name) == 0)
3323 CERROR("ko2iblnd: No matching interfaces\n");
3328 ibdev = kiblnd_dev_search(ifname);
3329 newdev = ibdev == NULL;
3330 /* hmm...create kib_dev even for alias */
3331 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3332 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3338 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3339 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3340 sizeof(ibdev->ibd_ifname));
3341 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3343 INIT_LIST_HEAD(&ibdev->ibd_nets);
3344 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3345 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3347 /* initialize the device */
3348 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3350 CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3355 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3358 net->ibn_dev = ibdev;
3359 ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3361 ni->ni_dev_cpt = ifaces[i].li_cpt;
3363 rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3367 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3369 CERROR("Failed to initialize NI pools: %d\n", rc);
3373 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3375 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3376 /* for health check */
3377 if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3378 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3379 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3381 net->ibn_init = IBLND_INIT_ALL;
3386 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3387 kiblnd_destroy_dev(ibdev);
3390 kiblnd_shutdown(ni);
3392 CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3393 ifname ? ifname : "", rc);
3398 static const struct lnet_lnd the_o2iblnd = {
3399 .lnd_type = O2IBLND,
3400 .lnd_startup = kiblnd_startup,
3401 .lnd_shutdown = kiblnd_shutdown,
3402 .lnd_ctl = kiblnd_ctl,
3403 .lnd_send = kiblnd_send,
3404 .lnd_recv = kiblnd_recv,
3405 .lnd_get_dev_prio = kiblnd_get_dev_prio,
3408 static void ko2inlnd_assert_wire_constants(void)
3410 BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3411 BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3412 BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3413 BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3415 BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3416 BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3417 BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3418 BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3419 BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3420 BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3421 BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3422 BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3423 BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3424 BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3426 BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3427 BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3428 BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3429 BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3430 BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3431 BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3432 BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3433 BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3435 BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3436 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3437 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3438 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3439 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3440 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3441 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3443 BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3444 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3445 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3446 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3447 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3449 BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3450 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3451 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3452 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3453 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3455 BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3456 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3457 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3458 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3459 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3460 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3461 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3463 BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3464 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3465 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3466 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3467 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3469 BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3470 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3471 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3472 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3473 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3474 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3475 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3477 BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3478 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3479 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3480 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3481 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3482 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3483 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3485 BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3486 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3487 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3488 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3489 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3491 /* Checks for struct kib_msg */
3492 //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3493 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3494 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3495 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3496 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3497 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3498 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3499 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3500 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3501 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3502 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3503 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3504 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3505 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3506 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3507 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3508 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3509 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3510 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3511 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3512 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3515 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3516 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3517 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3518 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3519 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3520 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3522 /* Immediate message */
3523 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3524 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3525 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3526 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3528 /* PUT req message */
3529 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3530 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3531 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3532 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3535 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3536 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3537 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3538 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3539 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3540 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3543 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3544 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3545 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3546 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3547 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3548 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3550 /* Completion message */
3551 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3552 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3553 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3554 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3557 BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3558 BUILD_BUG_ON(offsetof(struct kib_msg,
3559 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3561 BUILD_BUG_ON(offsetof(struct kib_msg,
3562 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3566 static void __exit ko2iblnd_exit(void)
3568 lnet_unregister_lnd(&the_o2iblnd);
3571 static int __init ko2iblnd_init(void)
3575 ko2inlnd_assert_wire_constants();
3577 rc = kiblnd_tunables_init();
3581 lnet_register_lnd(&the_o2iblnd);
3586 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3587 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3588 MODULE_VERSION("2.8.0");
3589 MODULE_LICENSE("GPL");
3591 module_init(ko2iblnd_init);
3592 module_exit(ko2iblnd_exit);