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);
342 atomic_set(&peer_ni->ibp_nconns, 0);
344 INIT_HLIST_NODE(&peer_ni->ibp_list);
345 INIT_LIST_HEAD(&peer_ni->ibp_conns);
346 INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
348 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
350 /* always called with a ref on ni, which prevents ni being shutdown */
351 LASSERT(net->ibn_shutdown == 0);
353 /* npeers only grows with the global lock held */
354 atomic_inc(&net->ibn_npeers);
356 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
363 kiblnd_destroy_peer(struct kref *kref)
365 struct kib_peer_ni *peer_ni = container_of(kref, struct kib_peer_ni,
367 struct kib_net *net = peer_ni->ibp_ni->ni_data;
369 LASSERT(net != NULL);
370 LASSERT(!kiblnd_peer_active(peer_ni));
371 LASSERT(kiblnd_peer_idle(peer_ni));
372 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
374 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
376 /* NB a peer_ni's connections keep a reference on their peer_ni until
377 * they are destroyed, so we can be assured that _all_ state to do
378 * with this peer_ni has been cleaned up when its refcount drops to
381 if (atomic_dec_and_test(&net->ibn_npeers))
382 wake_up_var(&net->ibn_npeers);
386 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
388 /* the caller is responsible for accounting the additional reference
391 struct kib_peer_ni *peer_ni;
393 hash_for_each_possible(kiblnd_data.kib_peers, peer_ni,
395 LASSERT(!kiblnd_peer_idle(peer_ni));
398 * Match a peer if its NID and the NID of the local NI it
399 * communicates over are the same. Otherwise don't match
400 * the peer, which will result in a new lnd peer being
403 if (peer_ni->ibp_nid != nid ||
404 !nid_same(&peer_ni->ibp_ni->ni_nid, &ni->ni_nid))
407 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
408 peer_ni, libcfs_nid2str(nid),
409 kref_read(&peer_ni->ibp_kref),
410 peer_ni->ibp_version);
417 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
419 LASSERT(list_empty(&peer_ni->ibp_conns));
421 LASSERT(kiblnd_peer_active(peer_ni));
422 hlist_del_init(&peer_ni->ibp_list);
423 /* lose peerlist's ref */
424 kiblnd_peer_decref(peer_ni);
429 kiblnd_debug_rx(struct kib_rx *rx)
431 CDEBUG(D_CONSOLE, " %p msg_type %x cred %d\n",
432 rx, rx->rx_msg->ibm_type,
433 rx->rx_msg->ibm_credits);
437 kiblnd_debug_tx(struct kib_tx *tx)
439 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lld "
440 "cookie %#llx msg %s%s type %x cred %d\n",
441 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
442 tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
443 tx->tx_lntmsg[0] == NULL ? "-" : "!",
444 tx->tx_lntmsg[1] == NULL ? "-" : "!",
445 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
449 kiblnd_debug_conn(struct kib_conn *conn)
451 struct list_head *tmp;
454 spin_lock(&conn->ibc_lock);
456 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
457 atomic_read(&conn->ibc_refcount), conn,
458 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
459 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d "
460 " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
461 conn->ibc_nsends_posted, conn->ibc_credits,
462 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
463 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
465 CDEBUG(D_CONSOLE, " early_rxs:\n");
466 list_for_each(tmp, &conn->ibc_early_rxs)
467 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
469 CDEBUG(D_CONSOLE, " tx_noops:\n");
470 list_for_each(tmp, &conn->ibc_tx_noops)
471 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
473 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
474 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
475 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
477 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
478 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
479 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
481 CDEBUG(D_CONSOLE, " tx_queue:\n");
482 list_for_each(tmp, &conn->ibc_tx_queue)
483 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
485 CDEBUG(D_CONSOLE, " active_txs:\n");
486 list_for_each(tmp, &conn->ibc_active_txs)
487 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
489 CDEBUG(D_CONSOLE, " rxs:\n");
490 for (i = 0; i < IBLND_RX_MSGS(conn); i++)
491 kiblnd_debug_rx(&conn->ibc_rxs[i]);
493 spin_unlock(&conn->ibc_lock);
497 kiblnd_dump_peer_debug_info(struct kib_peer_ni *peer_ni)
499 struct kib_conn *conn;
500 struct kib_conn *cnxt;
503 CDEBUG(D_CONSOLE, "[last_alive, races, reconnected, error]: %lld, %d, %d, %d\n",
504 peer_ni->ibp_last_alive,
506 peer_ni->ibp_reconnected,
508 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
510 CDEBUG(D_CONSOLE, "Conn %d:\n", count);
511 kiblnd_debug_conn(conn);
518 kiblnd_get_peer_info(struct lnet_ni *ni, lnet_nid_t nid, int index,
519 lnet_nid_t *nidp, int *count)
521 struct kib_peer_ni *peer_ni;
525 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
527 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
528 LASSERT(!kiblnd_peer_idle(peer_ni));
530 if (peer_ni->ibp_ni != ni)
533 if (peer_ni->ibp_nid == nid)
534 kiblnd_dump_peer_debug_info(peer_ni);
539 *nidp = peer_ni->ibp_nid;
540 *count = kref_read(&peer_ni->ibp_kref);
542 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
546 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
551 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
553 struct kib_conn *cnxt;
554 struct kib_conn *conn;
556 if (list_empty(&peer_ni->ibp_conns)) {
557 kiblnd_unlink_peer_locked(peer_ni);
559 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
561 kiblnd_close_conn_locked(conn, 0);
562 /* NB closing peer_ni's last conn unlinked it. */
564 /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
569 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
572 struct hlist_node *pnxt;
573 struct kib_peer_ni *peer_ni;
580 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
582 if (nid != LNET_NID_ANY) {
583 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
587 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
590 for (i = lo; i <= hi; i++) {
591 hlist_for_each_entry_safe(peer_ni, pnxt,
592 &kiblnd_data.kib_peers[i], ibp_list) {
593 LASSERT(!kiblnd_peer_idle(peer_ni));
595 if (peer_ni->ibp_ni != ni)
598 if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
601 if (!list_empty(&peer_ni->ibp_tx_queue)) {
602 LASSERT(list_empty(&peer_ni->ibp_conns));
604 list_splice_init(&peer_ni->ibp_tx_queue,
608 kiblnd_del_peer_locked(peer_ni);
609 rc = 0; /* matched something */
613 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
615 kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
620 static struct kib_conn *
621 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
623 struct kib_peer_ni *peer_ni;
624 struct kib_conn *conn;
628 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
630 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
631 LASSERT(!kiblnd_peer_idle(peer_ni));
633 if (peer_ni->ibp_ni != ni)
636 list_for_each_entry(conn, &peer_ni->ibp_conns,
641 kiblnd_conn_addref(conn);
642 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
648 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
653 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
655 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
656 if (cmid->route.path_rec == NULL)
659 if (*kiblnd_tunables.kib_ib_mtu)
660 cmid->route.path_rec->mtu =
661 ib_mtu_int_to_enum(*kiblnd_tunables.kib_ib_mtu);
665 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
673 vectors = conn->ibc_cmid->device->num_comp_vectors;
677 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
679 /* hash NID to CPU id in this partition... when targeting a single peer
680 * with multiple QPs, to engage more cores in CQ processing to a single
681 * peer, use ibp_nconns to salt the value the comp_vector value
683 ibp_nid = conn->ibc_peer->ibp_nid +
684 atomic_read(&conn->ibc_peer->ibp_nconns);
685 off = do_div(ibp_nid, cpumask_weight(*mask));
686 for_each_cpu(i, *mask) {
696 * Get the scheduler bound to this CPT. If the scheduler has no
697 * threads, which means that the CPT has no CPUs, then grab the
698 * next scheduler that we can use.
700 * This case would be triggered if a NUMA node is configured with
701 * no associated CPUs.
703 static struct kib_sched_info *
704 kiblnd_get_scheduler(int cpt)
706 struct kib_sched_info *sched;
709 sched = kiblnd_data.kib_scheds[cpt];
711 if (sched->ibs_nthreads > 0)
714 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
715 if (sched->ibs_nthreads > 0) {
716 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
717 cpt, sched->ibs_cpt);
725 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
728 * One WR for the LNet message
729 * And ibc_max_frags for the transfer WRs
732 int multiplier = 1 + conn->ibc_max_frags;
734 /* FastReg needs two extra WRs for map and invalidate */
735 if (IS_FAST_REG_DEV(conn->ibc_hdev->ibh_dev))
738 /* account for a maximum of ibc_queue_depth in-flight transfers */
739 ret = multiplier * conn->ibc_queue_depth;
741 if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
742 CDEBUG(D_NET, "peer_credits %u will result in send work "
743 "request size %d larger than maximum %d device "
744 "can handle\n", conn->ibc_queue_depth, ret,
745 conn->ibc_hdev->ibh_max_qp_wr);
746 conn->ibc_queue_depth =
747 conn->ibc_hdev->ibh_max_qp_wr / multiplier;
750 /* don't go beyond the maximum the device can handle */
751 return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
755 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
756 int state, int version)
759 * If the new conn is created successfully it takes over the caller's
760 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
761 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
762 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
763 * to destroy 'cmid' here since I'm called from the CM which still has
764 * its ref on 'cmid'). */
765 rwlock_t *glock = &kiblnd_data.kib_global_lock;
766 struct kib_net *net = peer_ni->ibp_ni->ni_data;
768 struct ib_qp_init_attr init_qp_attr = {};
769 struct kib_sched_info *sched;
770 #ifdef HAVE_OFED_IB_CQ_INIT_ATTR
771 struct ib_cq_init_attr cq_attr = {};
773 struct kib_conn *conn;
780 LASSERT(net != NULL);
781 LASSERT(!in_interrupt());
785 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
786 sched = kiblnd_get_scheduler(cpt);
789 CERROR("no schedulers available. node is unhealthy\n");
794 * The cpt might have changed if we ended up selecting a non cpt
795 * native scheduler. So use the scheduler's cpt instead.
797 cpt = sched->ibs_cpt;
799 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
801 CERROR("Can't allocate connection for %s\n",
802 libcfs_nid2str(peer_ni->ibp_nid));
806 conn->ibc_state = IBLND_CONN_INIT;
807 conn->ibc_version = version;
808 conn->ibc_peer = peer_ni; /* I take the caller's ref */
809 cmid->context = conn; /* for future CM callbacks */
810 conn->ibc_cmid = cmid;
811 conn->ibc_max_frags = peer_ni->ibp_max_frags;
812 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
813 conn->ibc_rxs = NULL;
814 conn->ibc_rx_pages = NULL;
816 INIT_LIST_HEAD(&conn->ibc_early_rxs);
817 INIT_LIST_HEAD(&conn->ibc_tx_noops);
818 INIT_LIST_HEAD(&conn->ibc_tx_queue);
819 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
820 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
821 INIT_LIST_HEAD(&conn->ibc_active_txs);
822 INIT_LIST_HEAD(&conn->ibc_zombie_txs);
823 spin_lock_init(&conn->ibc_lock);
825 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
826 sizeof(*conn->ibc_connvars));
827 if (conn->ibc_connvars == NULL) {
828 CERROR("Can't allocate in-progress connection state\n");
832 write_lock_irqsave(glock, flags);
833 if (dev->ibd_failover) {
834 write_unlock_irqrestore(glock, flags);
835 CERROR("%s: failover in progress\n", dev->ibd_ifname);
839 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
840 /* wakeup failover thread and teardown connection */
841 if (kiblnd_dev_can_failover(dev)) {
842 list_add_tail(&dev->ibd_fail_list,
843 &kiblnd_data.kib_failed_devs);
844 wake_up(&kiblnd_data.kib_failover_waitq);
847 write_unlock_irqrestore(glock, flags);
848 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
849 cmid->device->name, dev->ibd_ifname);
853 kiblnd_hdev_addref_locked(dev->ibd_hdev);
854 conn->ibc_hdev = dev->ibd_hdev;
856 kiblnd_setup_mtu_locked(cmid);
858 write_unlock_irqrestore(glock, flags);
860 #ifdef HAVE_OFED_IB_CQ_INIT_ATTR
861 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
862 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
863 cq = ib_create_cq(cmid->device,
864 kiblnd_cq_completion, kiblnd_cq_event, conn,
867 cq = ib_create_cq(cmid->device,
868 kiblnd_cq_completion, kiblnd_cq_event, conn,
869 IBLND_CQ_ENTRIES(conn),
870 kiblnd_get_completion_vector(conn, cpt));
874 * on MLX-5 (possibly MLX-4 as well) this error could be
875 * hit if the concurrent_sends and/or peer_tx_credits is set
876 * too high. Or due to an MLX-5 bug which tries to
877 * allocate 256kb via kmalloc for WR cookie array
879 CERROR("Failed to create CQ with %d CQEs: %ld\n",
880 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
886 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
888 CERROR("Can't request completion notification: %d\n", rc);
892 init_qp_attr.event_handler = kiblnd_qp_event;
893 init_qp_attr.qp_context = conn;
894 init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
895 init_qp_attr.cap.max_recv_sge = 1;
896 init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
897 init_qp_attr.qp_type = IB_QPT_RC;
898 init_qp_attr.send_cq = cq;
899 init_qp_attr.recv_cq = cq;
901 if (peer_ni->ibp_queue_depth_mod &&
902 peer_ni->ibp_queue_depth_mod < peer_ni->ibp_queue_depth) {
903 conn->ibc_queue_depth = peer_ni->ibp_queue_depth_mod;
904 CDEBUG(D_NET, "Use reduced queue depth %u (from %u)\n",
905 peer_ni->ibp_queue_depth_mod,
906 peer_ni->ibp_queue_depth);
910 /* kiblnd_send_wrs() can change the connection's queue depth if
911 * the maximum work requests for the device is maxed out
913 init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
914 init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
915 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd,
917 if (rc != -ENOMEM || conn->ibc_queue_depth < 2)
919 conn->ibc_queue_depth--;
923 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
924 "send_sge: %d, recv_sge: %d\n",
925 rc, init_qp_attr.cap.max_send_wr,
926 init_qp_attr.cap.max_recv_wr,
927 init_qp_attr.cap.max_send_sge,
928 init_qp_attr.cap.max_recv_sge);
932 conn->ibc_sched = sched;
934 if (!peer_ni->ibp_queue_depth_mod &&
935 conn->ibc_queue_depth != peer_ni->ibp_queue_depth) {
936 CWARN("peer %s - queue depth reduced from %u to %u"
937 " to allow for qp creation\n",
938 libcfs_nid2str(peer_ni->ibp_nid),
939 peer_ni->ibp_queue_depth,
940 conn->ibc_queue_depth);
941 peer_ni->ibp_queue_depth_mod = conn->ibc_queue_depth;
944 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
945 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
946 if (conn->ibc_rxs == NULL) {
947 CERROR("Cannot allocate RX buffers\n");
951 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
952 IBLND_RX_MSG_PAGES(conn));
956 kiblnd_map_rx_descs(conn);
958 /* 1 ref for caller and each rxmsg */
959 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
960 conn->ibc_nrx = IBLND_RX_MSGS(conn);
963 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
964 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
966 CERROR("Can't post rxmsg: %d\n", rc);
968 /* Make posted receives complete */
969 kiblnd_abort_receives(conn);
971 /* correct # of posted buffers
972 * NB locking needed now I'm racing with completion */
973 spin_lock_irqsave(&sched->ibs_lock, flags);
974 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
975 spin_unlock_irqrestore(&sched->ibs_lock, flags);
977 /* cmid will be destroyed by CM(ofed) after cm_callback
978 * returned, so we can't refer it anymore
979 * (by kiblnd_connd()->kiblnd_destroy_conn) */
980 rdma_destroy_qp(conn->ibc_cmid);
981 conn->ibc_cmid = NULL;
983 /* Drop my own and unused rxbuffer refcounts */
984 while (i++ <= IBLND_RX_MSGS(conn))
985 kiblnd_conn_decref(conn);
991 /* Init successful! */
992 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
993 state == IBLND_CONN_PASSIVE_WAIT);
994 conn->ibc_state = state;
997 atomic_inc(&peer_ni->ibp_nconns);
998 atomic_inc(&net->ibn_nconns);
1002 kiblnd_destroy_conn(conn);
1003 LIBCFS_FREE(conn, sizeof(*conn));
1009 kiblnd_destroy_conn(struct kib_conn *conn)
1011 struct rdma_cm_id *cmid = conn->ibc_cmid;
1012 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1014 LASSERT (!in_interrupt());
1015 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1016 LASSERT(list_empty(&conn->ibc_early_rxs));
1017 LASSERT(list_empty(&conn->ibc_tx_noops));
1018 LASSERT(list_empty(&conn->ibc_tx_queue));
1019 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1020 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1021 LASSERT(list_empty(&conn->ibc_active_txs));
1022 LASSERT (conn->ibc_noops_posted == 0);
1023 LASSERT (conn->ibc_nsends_posted == 0);
1025 switch (conn->ibc_state) {
1027 /* conn must be completely disengaged from the network */
1030 case IBLND_CONN_DISCONNECTED:
1031 /* connvars should have been freed already */
1032 LASSERT (conn->ibc_connvars == NULL);
1035 case IBLND_CONN_INIT:
1039 /* conn->ibc_cmid might be destroyed by CM already */
1040 if (cmid != NULL && cmid->qp != NULL)
1041 rdma_destroy_qp(cmid);
1044 ib_destroy_cq(conn->ibc_cq);
1046 kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1047 LNET_MSG_STATUS_OK);
1049 if (conn->ibc_rx_pages != NULL)
1050 kiblnd_unmap_rx_descs(conn);
1052 if (conn->ibc_rxs != NULL)
1053 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1055 if (conn->ibc_connvars != NULL)
1056 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1058 if (conn->ibc_hdev != NULL)
1059 kiblnd_hdev_decref(conn->ibc_hdev);
1061 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1062 if (conn->ibc_state != IBLND_CONN_INIT) {
1063 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1065 kiblnd_peer_decref(peer_ni);
1066 rdma_destroy_id(cmid);
1067 atomic_dec(&peer_ni->ibp_nconns);
1068 atomic_dec(&net->ibn_nconns);
1073 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1075 struct kib_conn *conn;
1076 struct kib_conn *cnxt;
1079 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1081 CDEBUG(D_NET, "Closing conn -> %s, "
1082 "version: %x, reason: %d\n",
1083 libcfs_nid2str(peer_ni->ibp_nid),
1084 conn->ibc_version, why);
1086 kiblnd_close_conn_locked(conn, why);
1094 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1095 int version, __u64 incarnation)
1097 struct kib_conn *conn;
1098 struct kib_conn *cnxt;
1101 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1103 if (conn->ibc_version == version &&
1104 conn->ibc_incarnation == incarnation)
1107 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1108 "incarnation:%#llx(%x, %#llx)\n",
1109 libcfs_nid2str(peer_ni->ibp_nid),
1110 conn->ibc_version, conn->ibc_incarnation,
1111 version, incarnation);
1113 kiblnd_close_conn_locked(conn, -ESTALE);
1121 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1123 struct kib_peer_ni *peer_ni;
1124 struct hlist_node *pnxt;
1128 unsigned long flags;
1131 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1133 if (nid != LNET_NID_ANY) {
1134 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
1138 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
1141 for (i = lo; i <= hi; i++) {
1142 hlist_for_each_entry_safe(peer_ni, pnxt,
1143 &kiblnd_data.kib_peers[i], ibp_list) {
1144 LASSERT(!kiblnd_peer_idle(peer_ni));
1146 if (peer_ni->ibp_ni != ni)
1149 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1152 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1156 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1158 /* wildcards always succeed */
1159 if (nid == LNET_NID_ANY)
1162 return (count == 0) ? -ENOENT : 0;
1166 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1168 struct libcfs_ioctl_data *data = arg;
1172 case IOC_LIBCFS_GET_PEER: {
1176 rc = kiblnd_get_peer_info(ni, data->ioc_nid, data->ioc_count,
1178 data->ioc_nid = nid;
1179 data->ioc_count = count;
1183 case IOC_LIBCFS_DEL_PEER: {
1184 rc = kiblnd_del_peer(ni, data->ioc_nid);
1187 case IOC_LIBCFS_GET_CONN: {
1188 struct kib_conn *conn;
1191 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1197 LASSERT(conn->ibc_cmid != NULL);
1198 data->ioc_nid = conn->ibc_peer->ibp_nid;
1199 if (conn->ibc_cmid->route.path_rec == NULL)
1200 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1203 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1204 kiblnd_conn_decref(conn);
1207 case IOC_LIBCFS_CLOSE_CONNECTION: {
1208 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1219 static const struct ln_key_list kiblnd_tunables_keys = {
1220 .lkl_maxattr = LNET_NET_O2IBLND_TUNABLES_ATTR_MAX,
1222 [LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS] = {
1223 .lkp_value = "peercredits_hiw",
1224 .lkp_data_type = NLA_U32
1226 [LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND] = {
1227 .lkp_value = "map_on_demand",
1228 .lkp_data_type = NLA_FLAG
1230 [LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS] = {
1231 .lkp_value = "concurrent_sends",
1232 .lkp_data_type = NLA_U32
1234 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE] = {
1235 .lkp_value = "fmr_pool_size",
1236 .lkp_data_type = NLA_U32
1238 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER] = {
1239 .lkp_value = "fmr_flush_trigger",
1240 .lkp_data_type = NLA_U32
1242 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE] = {
1243 .lkp_value = "fmr_cache",
1244 .lkp_data_type = NLA_U32
1246 [LNET_NET_O2IBLND_TUNABLES_ATTR_NTX] = {
1248 .lkp_data_type = NLA_U16
1250 [LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER] = {
1251 .lkp_value = "conns_per_peer",
1252 .lkp_data_type = NLA_U16
1254 [LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT] = {
1255 .lkp_value = "timeout",
1256 .lkp_data_type = NLA_U32,
1262 kiblnd_nl_get(int cmd, struct sk_buff *msg, int type, void *data)
1264 struct lnet_ioctl_config_o2iblnd_tunables *tuns;
1265 struct lnet_ni *ni = data;
1270 if (cmd != LNET_CMD_NETS || type != LNET_NET_LOCAL_NI_ATTR_LND_TUNABLES)
1273 tuns = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1274 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS,
1275 tuns->lnd_peercredits_hiw);
1276 if (tuns->lnd_map_on_demand) {
1278 LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND);
1280 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS,
1281 tuns->lnd_concurrent_sends);
1282 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE,
1283 tuns->lnd_fmr_pool_size);
1284 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER,
1285 tuns->lnd_fmr_flush_trigger);
1286 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE,
1287 tuns->lnd_fmr_cache);
1288 nla_put_u16(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_NTX, tuns->lnd_ntx);
1289 nla_put_u16(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER,
1290 tuns->lnd_conns_per_peer);
1291 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT,
1298 kiblnd_nl_set_default(int cmd, int type, void *data)
1300 struct lnet_lnd_tunables *tunables = data;
1301 struct lnet_ioctl_config_o2iblnd_tunables *lt;
1302 struct lnet_ioctl_config_o2iblnd_tunables *df;
1304 lt = &tunables->lnd_tun_u.lnd_o2ib;
1305 df = &kib_default_tunables;
1307 case LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS:
1308 lt->lnd_peercredits_hiw = df->lnd_peercredits_hiw;
1310 case LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND:
1311 lt->lnd_map_on_demand = df->lnd_map_on_demand;
1313 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS:
1314 lt->lnd_concurrent_sends = df->lnd_concurrent_sends;
1316 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE:
1317 lt->lnd_fmr_pool_size = df->lnd_fmr_pool_size;
1319 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER:
1320 lt->lnd_fmr_flush_trigger = df->lnd_fmr_flush_trigger;
1322 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE:
1323 lt->lnd_fmr_cache = df->lnd_fmr_cache;
1325 case LNET_NET_O2IBLND_TUNABLES_ATTR_NTX:
1326 lt->lnd_ntx = df->lnd_ntx;
1328 case LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT:
1329 lt->lnd_timeout = df->lnd_timeout;
1331 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER:
1332 lt->lnd_conns_per_peer = df->lnd_conns_per_peer;
1341 kiblnd_nl_set(int cmd, struct nlattr *attr, int type, void *data)
1343 struct lnet_lnd_tunables *tunables = data;
1347 if (cmd != LNET_CMD_NETS)
1351 kiblnd_nl_set_default(cmd, type, data);
1355 if (nla_type(attr) != LN_SCALAR_ATTR_INT_VALUE)
1359 case LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS:
1360 tunables->lnd_tun_u.lnd_o2ib.lnd_peercredits_hiw = nla_get_s64(attr);
1362 case LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND:
1363 tunables->lnd_tun_u.lnd_o2ib.lnd_map_on_demand = nla_get_s64(attr);
1365 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS:
1366 tunables->lnd_tun_u.lnd_o2ib.lnd_concurrent_sends = nla_get_s64(attr);
1368 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE:
1369 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_pool_size = nla_get_s64(attr);
1371 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER:
1372 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_flush_trigger = nla_get_s64(attr);
1374 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE:
1375 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_cache = nla_get_s64(attr);
1377 case LNET_NET_O2IBLND_TUNABLES_ATTR_NTX:
1378 tunables->lnd_tun_u.lnd_o2ib.lnd_ntx = nla_get_s64(attr);
1380 case LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT:
1381 tunables->lnd_tun_u.lnd_o2ib.lnd_timeout = nla_get_s64(attr);
1383 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER:
1384 num = nla_get_s64(attr);
1385 if (num >= 0 && num < 128)
1386 tunables->lnd_tun_u.lnd_o2ib.lnd_conns_per_peer = num;
1398 kiblnd_free_pages(struct kib_pages *p)
1400 int npages = p->ibp_npages;
1403 for (i = 0; i < npages; i++) {
1404 if (p->ibp_pages[i] != NULL)
1405 __free_page(p->ibp_pages[i]);
1408 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1412 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1414 struct kib_pages *p;
1417 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1418 offsetof(struct kib_pages, ibp_pages[npages]));
1420 CERROR("Can't allocate descriptor for %d pages\n", npages);
1424 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1425 p->ibp_npages = npages;
1427 for (i = 0; i < npages; i++) {
1428 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1430 if (p->ibp_pages[i] == NULL) {
1431 CERROR("Can't allocate page %d of %d\n", i, npages);
1432 kiblnd_free_pages(p);
1442 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1447 LASSERT (conn->ibc_rxs != NULL);
1448 LASSERT (conn->ibc_hdev != NULL);
1450 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1451 rx = &conn->ibc_rxs[i];
1453 LASSERT(rx->rx_nob >= 0); /* not posted */
1455 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1456 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1458 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1461 kiblnd_free_pages(conn->ibc_rx_pages);
1463 conn->ibc_rx_pages = NULL;
1467 kiblnd_map_rx_descs(struct kib_conn *conn)
1475 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1476 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1477 rx = &conn->ibc_rxs[i];
1480 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1483 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1484 rx->rx_msg, IBLND_MSG_SIZE,
1486 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1488 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1490 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1491 i, rx->rx_msg, rx->rx_msgaddr,
1492 (__u64)(page_to_phys(pg) + pg_off));
1494 pg_off += IBLND_MSG_SIZE;
1495 LASSERT(pg_off <= PAGE_SIZE);
1497 if (pg_off == PAGE_SIZE) {
1500 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1506 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1508 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1512 LASSERT (tpo->tpo_pool.po_allocated == 0);
1517 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1518 tx = &tpo->tpo_tx_descs[i];
1519 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1520 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1522 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1525 kiblnd_hdev_decref(hdev);
1526 tpo->tpo_hdev = NULL;
1529 static struct kib_hca_dev *
1530 kiblnd_current_hdev(struct kib_dev *dev)
1532 struct kib_hca_dev *hdev;
1533 unsigned long flags;
1536 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1537 while (dev->ibd_failover) {
1538 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1540 CDEBUG(D_NET, "%s: Wait for failover\n",
1542 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1544 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1547 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1548 hdev = dev->ibd_hdev;
1550 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1556 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1558 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1559 struct kib_pool *pool = &tpo->tpo_pool;
1560 struct kib_net *net = pool->po_owner->ps_net;
1561 struct kib_dev *dev;
1568 LASSERT (net != NULL);
1572 /* pre-mapped messages are not bigger than 1 page */
1573 BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1575 /* No fancy arithmetic when we do the buffer calculations */
1576 BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1578 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1580 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1581 page = txpgs->ibp_pages[ipage];
1582 tx = &tpo->tpo_tx_descs[i];
1584 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1587 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1591 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1593 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1595 list_add(&tx->tx_list, &pool->po_free_list);
1597 page_offset += IBLND_MSG_SIZE;
1598 LASSERT(page_offset <= PAGE_SIZE);
1600 if (page_offset == PAGE_SIZE) {
1603 LASSERT(ipage <= txpgs->ibp_npages);
1609 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1611 LASSERT(fpo->fpo_map_count == 0);
1613 #ifdef HAVE_OFED_FMR_POOL_API
1614 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1615 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1617 #endif /* HAVE_OFED_FMR_POOL_API */
1619 struct kib_fast_reg_descriptor *frd, *tmp;
1622 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1624 list_del(&frd->frd_list);
1625 #ifndef HAVE_OFED_IB_MAP_MR_SG
1626 ib_free_fast_reg_page_list(frd->frd_frpl);
1628 ib_dereg_mr(frd->frd_mr);
1629 LIBCFS_FREE(frd, sizeof(*frd));
1632 if (i < fpo->fast_reg.fpo_pool_size)
1633 CERROR("FastReg pool still has %d regions registered\n",
1634 fpo->fast_reg.fpo_pool_size - i);
1638 kiblnd_hdev_decref(fpo->fpo_hdev);
1640 LIBCFS_FREE(fpo, sizeof(*fpo));
1644 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1646 struct kib_fmr_pool *fpo, *tmp;
1648 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1649 list_del(&fpo->fpo_list);
1650 kiblnd_destroy_fmr_pool(fpo);
1655 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1658 int size = tunables->lnd_fmr_pool_size / ncpts;
1660 return max(IBLND_FMR_POOL, size);
1664 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1667 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1669 return max(IBLND_FMR_POOL_FLUSH, size);
1672 #ifdef HAVE_OFED_FMR_POOL_API
1673 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1674 struct kib_fmr_pool *fpo)
1676 struct ib_fmr_pool_param param = {
1677 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1678 .page_shift = PAGE_SHIFT,
1679 .access = (IB_ACCESS_LOCAL_WRITE |
1680 IB_ACCESS_REMOTE_WRITE),
1681 .pool_size = fps->fps_pool_size,
1682 .dirty_watermark = fps->fps_flush_trigger,
1683 .flush_function = NULL,
1685 .cache = !!fps->fps_cache };
1688 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1690 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1691 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1693 CERROR("Failed to create FMR pool: %d\n", rc);
1695 CERROR("FMRs are not supported\n");
1697 fpo->fpo_is_fmr = true;
1701 #endif /* HAVE_OFED_FMR_POOL_API */
1703 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1704 struct kib_fmr_pool *fpo,
1705 enum kib_dev_caps dev_caps)
1707 struct kib_fast_reg_descriptor *frd, *tmp;
1710 #ifdef HAVE_OFED_FMR_POOL_API
1711 fpo->fpo_is_fmr = false;
1714 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1715 fpo->fast_reg.fpo_pool_size = 0;
1716 for (i = 0; i < fps->fps_pool_size; i++) {
1717 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1720 CERROR("Failed to allocate a new fast_reg descriptor\n");
1726 #ifndef HAVE_OFED_IB_MAP_MR_SG
1727 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1728 IBLND_MAX_RDMA_FRAGS);
1729 if (IS_ERR(frd->frd_frpl)) {
1730 rc = PTR_ERR(frd->frd_frpl);
1731 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1733 frd->frd_frpl = NULL;
1738 #ifdef HAVE_OFED_IB_ALLOC_FAST_REG_MR
1739 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1740 IBLND_MAX_RDMA_FRAGS);
1743 * it is expected to get here if this is an MLX-5 card.
1744 * MLX-4 cards will always use FMR and MLX-5 cards will
1745 * always use fast_reg. It turns out that some MLX-5 cards
1746 * (possibly due to older FW versions) do not natively support
1747 * gaps. So we will need to track them here.
1749 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1750 #ifdef IB_MR_TYPE_SG_GAPS
1751 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1752 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1753 IB_MR_TYPE_SG_GAPS :
1758 IBLND_MAX_RDMA_FRAGS);
1759 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1760 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1761 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1763 if (IS_ERR(frd->frd_mr)) {
1764 rc = PTR_ERR(frd->frd_mr);
1765 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1770 /* indicate that the local invalidate needs to be generated */
1771 frd->frd_valid = false;
1773 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1774 fpo->fast_reg.fpo_pool_size++;
1781 ib_dereg_mr(frd->frd_mr);
1782 #ifndef HAVE_OFED_IB_MAP_MR_SG
1784 ib_free_fast_reg_page_list(frd->frd_frpl);
1786 LIBCFS_FREE(frd, sizeof(*frd));
1789 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1791 list_del(&frd->frd_list);
1792 #ifndef HAVE_OFED_IB_MAP_MR_SG
1793 ib_free_fast_reg_page_list(frd->frd_frpl);
1795 ib_dereg_mr(frd->frd_mr);
1796 LIBCFS_FREE(frd, sizeof(*frd));
1802 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1803 struct kib_fmr_pool **pp_fpo)
1805 struct kib_dev *dev = fps->fps_net->ibn_dev;
1806 struct kib_fmr_pool *fpo;
1809 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1813 memset(fpo, 0, sizeof(*fpo));
1815 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1817 #ifdef HAVE_OFED_FMR_POOL_API
1818 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1819 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1821 #endif /* HAVE_OFED_FMR_POOL_API */
1822 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1826 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1827 fpo->fpo_owner = fps;
1833 kiblnd_hdev_decref(fpo->fpo_hdev);
1834 LIBCFS_FREE(fpo, sizeof(*fpo));
1839 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1841 struct kib_fmr_pool *fpo;
1843 if (fps->fps_net == NULL) /* intialized? */
1846 spin_lock(&fps->fps_lock);
1848 while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1849 struct kib_fmr_pool,
1850 fpo_list)) != NULL) {
1851 fpo->fpo_failed = 1;
1852 if (fpo->fpo_map_count == 0)
1853 list_move(&fpo->fpo_list, zombies);
1855 list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1858 spin_unlock(&fps->fps_lock);
1862 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1864 if (fps->fps_net != NULL) { /* initialized? */
1865 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1866 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1871 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1872 struct kib_net *net,
1873 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1875 struct kib_fmr_pool *fpo;
1878 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1883 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1884 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1885 fps->fps_cache = tunables->lnd_fmr_cache;
1887 spin_lock_init(&fps->fps_lock);
1888 INIT_LIST_HEAD(&fps->fps_pool_list);
1889 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1891 rc = kiblnd_create_fmr_pool(fps, &fpo);
1893 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1899 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1901 if (fpo->fpo_map_count != 0) /* still in use */
1903 if (fpo->fpo_failed)
1905 return now >= fpo->fpo_deadline;
1908 #if defined(HAVE_OFED_FMR_POOL_API) || !defined(HAVE_OFED_IB_MAP_MR_SG)
1910 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1912 struct kib_hca_dev *hdev;
1913 __u64 *pages = tx->tx_pages;
1918 hdev = tx->tx_pool->tpo_hdev;
1920 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1921 for (size = 0; size < rd->rd_frags[i].rf_nob;
1922 size += hdev->ibh_page_size) {
1923 pages[npages++] = (rd->rd_frags[i].rf_addr &
1924 hdev->ibh_page_mask) + size;
1933 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1936 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1937 struct kib_fmr_poolset *fps;
1938 time64_t now = ktime_get_seconds();
1939 struct kib_fmr_pool *tmp;
1944 fps = fpo->fpo_owner;
1946 #ifdef HAVE_OFED_FMR_POOL_API
1947 if (fpo->fpo_is_fmr) {
1948 if (fmr->fmr_pfmr) {
1949 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1950 fmr->fmr_pfmr = NULL;
1954 int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1958 #endif /* HAVE_OFED_FMR_POOL_API */
1960 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1962 frd->frd_posted = false;
1963 fmr->fmr_frd = NULL;
1964 spin_lock(&fps->fps_lock);
1965 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1966 spin_unlock(&fps->fps_lock);
1969 fmr->fmr_pool = NULL;
1971 spin_lock(&fps->fps_lock);
1972 fpo->fpo_map_count--; /* decref the pool */
1974 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1975 /* the first pool is persistent */
1976 if (fps->fps_pool_list.next == &fpo->fpo_list)
1979 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1980 list_move(&fpo->fpo_list, &zombies);
1984 spin_unlock(&fps->fps_lock);
1986 if (!list_empty(&zombies))
1987 kiblnd_destroy_fmr_pool_list(&zombies);
1990 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1991 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1992 struct kib_fmr *fmr)
1994 struct kib_fmr_pool *fpo;
1996 bool is_rx = (rd != tx->tx_rd);
1997 #ifdef HAVE_OFED_FMR_POOL_API
1998 __u64 *pages = tx->tx_pages;
1999 bool tx_pages_mapped = false;
2005 spin_lock(&fps->fps_lock);
2006 version = fps->fps_version;
2007 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
2008 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2009 fpo->fpo_map_count++;
2011 #ifdef HAVE_OFED_FMR_POOL_API
2012 fmr->fmr_pfmr = NULL;
2013 if (fpo->fpo_is_fmr) {
2014 struct ib_pool_fmr *pfmr;
2016 spin_unlock(&fps->fps_lock);
2018 if (!tx_pages_mapped) {
2019 npages = kiblnd_map_tx_pages(tx, rd);
2020 tx_pages_mapped = true;
2023 pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
2024 pages, npages, iov);
2025 if (likely(!IS_ERR(pfmr))) {
2026 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
2028 fmr->fmr_frd = NULL;
2029 fmr->fmr_pfmr = pfmr;
2030 fmr->fmr_pool = fpo;
2035 #endif /* HAVE_OFED_FMR_POOL_API */
2037 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
2038 struct kib_fast_reg_descriptor *frd;
2039 #ifdef HAVE_OFED_IB_MAP_MR_SG
2040 struct ib_reg_wr *wr;
2043 struct ib_rdma_wr *wr;
2044 struct ib_fast_reg_page_list *frpl;
2048 frd = list_first_entry(
2049 &fpo->fast_reg.fpo_pool_list,
2050 struct kib_fast_reg_descriptor,
2052 list_del(&frd->frd_list);
2053 spin_unlock(&fps->fps_lock);
2055 #ifndef HAVE_OFED_IB_MAP_MR_SG
2056 frpl = frd->frd_frpl;
2060 if (!frd->frd_valid) {
2061 struct ib_rdma_wr *inv_wr;
2062 __u32 key = is_rx ? mr->rkey : mr->lkey;
2064 frd->frd_valid = true;
2065 inv_wr = &frd->frd_inv_wr;
2066 memset(inv_wr, 0, sizeof(*inv_wr));
2068 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
2069 inv_wr->wr.wr_id = IBLND_WID_MR;
2070 inv_wr->wr.ex.invalidate_rkey = key;
2073 key = ib_inc_rkey(key);
2074 ib_update_fast_reg_key(mr, key);
2077 #ifdef HAVE_OFED_IB_MAP_MR_SG
2078 #ifdef HAVE_OFED_IB_MAP_MR_SG_5ARGS
2079 n = ib_map_mr_sg(mr, tx->tx_frags,
2080 rd->rd_nfrags, NULL, PAGE_SIZE);
2082 n = ib_map_mr_sg(mr, tx->tx_frags,
2083 rd->rd_nfrags, PAGE_SIZE);
2084 #endif /* HAVE_OFED_IB_MAP_MR_SG_5ARGS */
2085 if (unlikely(n != rd->rd_nfrags)) {
2086 CERROR("Failed to map mr %d/%d elements\n",
2088 return n < 0 ? n : -EINVAL;
2091 wr = &frd->frd_fastreg_wr;
2092 memset(wr, 0, sizeof(*wr));
2094 wr->wr.opcode = IB_WR_REG_MR;
2095 wr->wr.wr_id = IBLND_WID_MR;
2097 wr->wr.send_flags = 0;
2099 wr->key = is_rx ? mr->rkey : mr->lkey;
2100 wr->access = (IB_ACCESS_LOCAL_WRITE |
2101 IB_ACCESS_REMOTE_WRITE);
2102 #else /* HAVE_OFED_IB_MAP_MR_SG */
2103 if (!tx_pages_mapped) {
2104 npages = kiblnd_map_tx_pages(tx, rd);
2105 tx_pages_mapped = true;
2108 LASSERT(npages <= frpl->max_page_list_len);
2109 memcpy(frpl->page_list, pages,
2110 sizeof(*pages) * npages);
2112 /* Prepare FastReg WR */
2113 wr = &frd->frd_fastreg_wr;
2114 memset(wr, 0, sizeof(*wr));
2116 wr->wr.opcode = IB_WR_FAST_REG_MR;
2117 wr->wr.wr_id = IBLND_WID_MR;
2119 wr->wr.wr.fast_reg.iova_start = iov;
2120 wr->wr.wr.fast_reg.page_list = frpl;
2121 wr->wr.wr.fast_reg.page_list_len = npages;
2122 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2123 wr->wr.wr.fast_reg.length = nob;
2124 wr->wr.wr.fast_reg.rkey =
2125 is_rx ? mr->rkey : mr->lkey;
2126 wr->wr.wr.fast_reg.access_flags =
2127 (IB_ACCESS_LOCAL_WRITE |
2128 IB_ACCESS_REMOTE_WRITE);
2129 #endif /* HAVE_OFED_IB_MAP_MR_SG */
2131 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
2133 fmr->fmr_pool = fpo;
2134 frd->frd_posted = false;
2137 spin_unlock(&fps->fps_lock);
2141 spin_lock(&fps->fps_lock);
2142 fpo->fpo_map_count--;
2143 if (rc != -EAGAIN) {
2144 spin_unlock(&fps->fps_lock);
2148 /* EAGAIN and ... */
2149 if (version != fps->fps_version) {
2150 spin_unlock(&fps->fps_lock);
2155 if (fps->fps_increasing) {
2156 spin_unlock(&fps->fps_lock);
2157 CDEBUG(D_NET, "Another thread is allocating new "
2158 "FMR pool, waiting for her to complete\n");
2159 wait_var_event(fps, !fps->fps_increasing);
2164 if (ktime_get_seconds() < fps->fps_next_retry) {
2165 /* someone failed recently */
2166 spin_unlock(&fps->fps_lock);
2170 fps->fps_increasing = 1;
2171 spin_unlock(&fps->fps_lock);
2173 CDEBUG(D_NET, "Allocate new FMR pool\n");
2174 rc = kiblnd_create_fmr_pool(fps, &fpo);
2175 spin_lock(&fps->fps_lock);
2176 fps->fps_increasing = 0;
2180 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2182 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2184 spin_unlock(&fps->fps_lock);
2190 kiblnd_fini_pool(struct kib_pool *pool)
2192 LASSERT(list_empty(&pool->po_free_list));
2193 LASSERT(pool->po_allocated == 0);
2195 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2199 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2201 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2203 memset(pool, 0, sizeof(struct kib_pool));
2204 INIT_LIST_HEAD(&pool->po_free_list);
2205 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2206 pool->po_owner = ps;
2207 pool->po_size = size;
2211 kiblnd_destroy_pool_list(struct list_head *head)
2213 struct kib_pool *pool;
2215 while ((pool = list_first_entry_or_null(head,
2217 po_list)) != NULL) {
2218 list_del(&pool->po_list);
2220 LASSERT(pool->po_owner != NULL);
2221 pool->po_owner->ps_pool_destroy(pool);
2226 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2228 struct kib_pool *po;
2230 if (ps->ps_net == NULL) /* intialized? */
2233 spin_lock(&ps->ps_lock);
2234 while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2236 po_list)) != NULL) {
2238 if (po->po_allocated == 0)
2239 list_move(&po->po_list, zombies);
2241 list_move(&po->po_list, &ps->ps_failed_pool_list);
2243 spin_unlock(&ps->ps_lock);
2247 kiblnd_fini_poolset(struct kib_poolset *ps)
2249 if (ps->ps_net != NULL) { /* initialized? */
2250 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2251 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2256 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2257 struct kib_net *net, char *name, int size,
2258 kib_ps_pool_create_t po_create,
2259 kib_ps_pool_destroy_t po_destroy,
2260 kib_ps_node_init_t nd_init,
2261 kib_ps_node_fini_t nd_fini)
2263 struct kib_pool *pool;
2266 memset(ps, 0, sizeof(struct kib_poolset));
2270 ps->ps_pool_create = po_create;
2271 ps->ps_pool_destroy = po_destroy;
2272 ps->ps_node_init = nd_init;
2273 ps->ps_node_fini = nd_fini;
2274 ps->ps_pool_size = size;
2275 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2276 >= sizeof(ps->ps_name))
2278 spin_lock_init(&ps->ps_lock);
2279 INIT_LIST_HEAD(&ps->ps_pool_list);
2280 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2282 rc = ps->ps_pool_create(ps, size, &pool);
2284 list_add(&pool->po_list, &ps->ps_pool_list);
2286 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2292 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2294 if (pool->po_allocated != 0) /* still in use */
2296 if (pool->po_failed)
2298 return now >= pool->po_deadline;
2302 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2305 struct kib_poolset *ps = pool->po_owner;
2306 struct kib_pool *tmp;
2307 time64_t now = ktime_get_seconds();
2309 spin_lock(&ps->ps_lock);
2311 if (ps->ps_node_fini != NULL)
2312 ps->ps_node_fini(pool, node);
2314 LASSERT(pool->po_allocated > 0);
2315 list_add(node, &pool->po_free_list);
2316 pool->po_allocated--;
2318 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2319 /* the first pool is persistent */
2320 if (ps->ps_pool_list.next == &pool->po_list)
2323 if (kiblnd_pool_is_idle(pool, now))
2324 list_move(&pool->po_list, &zombies);
2326 spin_unlock(&ps->ps_lock);
2328 if (!list_empty(&zombies))
2329 kiblnd_destroy_pool_list(&zombies);
2333 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2335 struct list_head *node;
2336 struct kib_pool *pool;
2338 unsigned int interval = 1;
2339 ktime_t time_before;
2340 unsigned int trips = 0;
2343 spin_lock(&ps->ps_lock);
2344 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2345 if (list_empty(&pool->po_free_list))
2348 pool->po_allocated++;
2349 pool->po_deadline = ktime_get_seconds() +
2350 IBLND_POOL_DEADLINE;
2351 node = pool->po_free_list.next;
2354 if (ps->ps_node_init != NULL) {
2355 /* still hold the lock */
2356 ps->ps_node_init(pool, node);
2358 spin_unlock(&ps->ps_lock);
2362 /* no available tx pool and ... */
2363 if (ps->ps_increasing) {
2364 /* another thread is allocating a new pool */
2365 spin_unlock(&ps->ps_lock);
2368 "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2369 ps->ps_name, interval, trips);
2371 schedule_timeout_interruptible(interval);
2372 if (interval < cfs_time_seconds(1))
2378 if (ktime_get_seconds() < ps->ps_next_retry) {
2379 /* someone failed recently */
2380 spin_unlock(&ps->ps_lock);
2384 ps->ps_increasing = 1;
2385 spin_unlock(&ps->ps_lock);
2387 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2388 time_before = ktime_get();
2389 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2390 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete\n",
2391 ktime_ms_delta(ktime_get(), time_before));
2393 spin_lock(&ps->ps_lock);
2394 ps->ps_increasing = 0;
2396 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2398 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2399 CERROR("Can't allocate new %s pool because out of memory\n",
2402 spin_unlock(&ps->ps_lock);
2408 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2410 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2414 LASSERT (pool->po_allocated == 0);
2416 if (tpo->tpo_tx_pages != NULL) {
2417 kiblnd_unmap_tx_pool(tpo);
2418 kiblnd_free_pages(tpo->tpo_tx_pages);
2421 if (tpo->tpo_tx_descs == NULL)
2424 for (i = 0; i < pool->po_size; i++) {
2425 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2426 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2428 list_del(&tx->tx_list);
2429 if (tx->tx_pages != NULL)
2430 CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2431 if (tx->tx_frags != NULL)
2432 CFS_FREE_PTR_ARRAY(tx->tx_frags,
2433 IBLND_MAX_RDMA_FRAGS);
2434 if (tx->tx_wrq != NULL)
2435 CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2436 IBLND_MAX_RDMA_FRAGS);
2437 if (tx->tx_sge != NULL) {
2438 /* +1 is for the lnet header/message itself */
2439 CFS_FREE_PTR_ARRAY(tx->tx_sge,
2440 (IBLND_MAX_RDMA_FRAGS *
2443 if (tx->tx_rd != NULL)
2444 LIBCFS_FREE(tx->tx_rd,
2445 offsetof(struct kib_rdma_desc,
2446 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2449 CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2451 kiblnd_fini_pool(pool);
2455 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2457 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2460 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2461 ntx = tunables->lnd_ntx / ncpts;
2463 return max(IBLND_TX_POOL, ntx);
2467 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2471 struct kib_pool *pool;
2472 struct kib_tx_pool *tpo;
2474 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2476 CERROR("Failed to allocate TX pool\n");
2480 pool = &tpo->tpo_pool;
2481 kiblnd_init_pool(ps, pool, size);
2482 tpo->tpo_tx_descs = NULL;
2483 tpo->tpo_tx_pages = NULL;
2485 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2486 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2487 CERROR("Can't allocate tx pages: %d\n", npg);
2492 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2493 size * sizeof(struct kib_tx));
2494 if (tpo->tpo_tx_descs == NULL) {
2495 CERROR("Can't allocate %d tx descriptors\n", size);
2496 ps->ps_pool_destroy(pool);
2500 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2502 for (i = 0; i < size; i++) {
2503 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2504 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2507 if (ps->ps_net->ibn_fmr_ps != NULL) {
2508 LIBCFS_CPT_ALLOC(tx->tx_pages,
2509 lnet_cpt_table(), ps->ps_cpt,
2510 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2511 if (tx->tx_pages == NULL)
2515 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2516 IBLND_MAX_RDMA_FRAGS *
2517 sizeof(*tx->tx_frags));
2518 if (tx->tx_frags == NULL)
2521 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2523 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2524 IBLND_MAX_RDMA_FRAGS *
2525 sizeof(*tx->tx_wrq));
2526 if (tx->tx_wrq == NULL)
2529 /* +1 is for the lnet header/message itself */
2530 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2531 (IBLND_MAX_RDMA_FRAGS * wrq_sge + 1) *
2532 sizeof(*tx->tx_sge));
2533 if (tx->tx_sge == NULL)
2536 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2537 offsetof(struct kib_rdma_desc,
2538 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2539 if (tx->tx_rd == NULL)
2544 kiblnd_map_tx_pool(tpo);
2549 ps->ps_pool_destroy(pool);
2554 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2556 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2557 struct kib_tx_poolset,
2559 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2561 tx->tx_cookie = tps->tps_next_tx_cookie++;
2565 kiblnd_net_fini_pools(struct kib_net *net)
2569 cfs_cpt_for_each(i, lnet_cpt_table()) {
2570 struct kib_tx_poolset *tps;
2571 struct kib_fmr_poolset *fps;
2573 if (net->ibn_tx_ps != NULL) {
2574 tps = net->ibn_tx_ps[i];
2575 kiblnd_fini_poolset(&tps->tps_poolset);
2578 if (net->ibn_fmr_ps != NULL) {
2579 fps = net->ibn_fmr_ps[i];
2580 kiblnd_fini_fmr_poolset(fps);
2584 if (net->ibn_tx_ps != NULL) {
2585 cfs_percpt_free(net->ibn_tx_ps);
2586 net->ibn_tx_ps = NULL;
2589 if (net->ibn_fmr_ps != NULL) {
2590 cfs_percpt_free(net->ibn_fmr_ps);
2591 net->ibn_fmr_ps = NULL;
2596 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2599 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2600 #ifdef HAVE_OFED_IB_GET_DMA_MR
2601 unsigned long flags;
2607 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2609 #ifdef HAVE_OFED_IB_GET_DMA_MR
2610 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2612 * if lnd_map_on_demand is zero then we have effectively disabled
2613 * FMR or FastReg and we're using global memory regions
2616 if (!tunables->lnd_map_on_demand) {
2617 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2619 goto create_tx_pool;
2622 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2625 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2626 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2627 tunables->lnd_fmr_pool_size,
2628 tunables->lnd_ntx / 4);
2633 /* TX pool must be created later than FMR, see LU-2268
2635 LASSERT(net->ibn_tx_ps == NULL);
2637 /* premapping can fail if ibd_nmr > 1, so we always create
2638 * FMR pool and map-on-demand if premapping failed */
2640 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2641 sizeof(struct kib_fmr_poolset));
2642 if (net->ibn_fmr_ps == NULL) {
2643 CERROR("Failed to allocate FMR pool array\n");
2648 for (i = 0; i < ncpts; i++) {
2649 cpt = (cpts == NULL) ? i : cpts[i];
2650 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2653 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2660 LASSERT(i == ncpts);
2662 #ifdef HAVE_OFED_IB_GET_DMA_MR
2665 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2666 sizeof(struct kib_tx_poolset));
2667 if (net->ibn_tx_ps == NULL) {
2668 CERROR("Failed to allocate tx pool array\n");
2673 for (i = 0; i < ncpts; i++) {
2674 cpt = (cpts == NULL) ? i : cpts[i];
2675 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2677 kiblnd_tx_pool_size(ni, ncpts),
2678 kiblnd_create_tx_pool,
2679 kiblnd_destroy_tx_pool,
2680 kiblnd_tx_init, NULL);
2682 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2690 kiblnd_net_fini_pools(net);
2696 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2698 struct ib_port_attr *port_attr;
2700 unsigned long flags;
2701 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2703 LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2704 if (port_attr == NULL) {
2705 CDEBUG(D_NETERROR, "Out of memory\n");
2709 rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2711 write_lock_irqsave(g_lock, flags);
2714 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2715 ? IBLND_DEV_PORT_ACTIVE
2716 : IBLND_DEV_PORT_DOWN;
2718 write_unlock_irqrestore(g_lock, flags);
2719 LIBCFS_FREE(port_attr, sizeof(*port_attr));
2722 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2729 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2731 struct kib_net *net;
2732 __u32 ni_state_before;
2733 bool update_ping_buf = false;
2734 struct lnet_ni *ni = NULL;
2736 /* for health check */
2737 list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2740 CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2741 libcfs_nidstr(&ni->ni_nid));
2742 ni_state_before = lnet_set_link_fatal_state(ni, val);
2744 if (!update_ping_buf &&
2745 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
2746 (val != ni_state_before) &&
2747 (net->ibn_init == IBLND_INIT_ALL))
2748 update_ping_buf = true;
2751 if (update_ping_buf)
2752 lnet_mark_ping_buffer_for_update();
2756 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2758 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2759 struct kib_hca_dev *hdev;
2760 unsigned long flags;
2762 hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2764 write_lock_irqsave(g_lock, flags);
2766 switch (event->event) {
2767 case IB_EVENT_DEVICE_FATAL:
2768 CDEBUG(D_NET, "IB device fatal\n");
2769 hdev->ibh_state = IBLND_DEV_FATAL;
2770 kiblnd_set_ni_fatal_on(hdev, 1);
2772 case IB_EVENT_PORT_ACTIVE:
2773 CDEBUG(D_NET, "IB port active\n");
2774 if (event->element.port_num == hdev->ibh_port) {
2775 hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2776 kiblnd_set_ni_fatal_on(hdev, 0);
2779 case IB_EVENT_PORT_ERR:
2780 CDEBUG(D_NET, "IB port err\n");
2781 if (event->element.port_num == hdev->ibh_port) {
2782 hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2783 kiblnd_set_ni_fatal_on(hdev, 1);
2789 write_unlock_irqrestore(g_lock, flags);
2793 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2795 struct ib_device_attr *dev_attr;
2799 /* It's safe to assume a HCA can handle a page size
2800 * matching that of the native system */
2801 hdev->ibh_page_shift = PAGE_SHIFT;
2802 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2803 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2805 #ifndef HAVE_OFED_IB_DEVICE_ATTRS
2806 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2807 if (dev_attr == NULL) {
2808 CERROR("Out of memory\n");
2812 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2814 CERROR("Failed to query IB device: %d\n", rc);
2815 goto out_clean_attr;
2818 dev_attr = &hdev->ibh_ibdev->attrs;
2821 hdev->ibh_mr_size = dev_attr->max_mr_size;
2822 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2824 /* Setup device Memory Registration capabilities */
2825 #ifdef HAVE_OFED_FMR_POOL_API
2826 #ifdef HAVE_OFED_IB_DEVICE_OPS
2827 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2828 hdev->ibh_ibdev->ops.dealloc_fmr &&
2829 hdev->ibh_ibdev->ops.map_phys_fmr &&
2830 hdev->ibh_ibdev->ops.unmap_fmr) {
2832 if (hdev->ibh_ibdev->alloc_fmr &&
2833 hdev->ibh_ibdev->dealloc_fmr &&
2834 hdev->ibh_ibdev->map_phys_fmr &&
2835 hdev->ibh_ibdev->unmap_fmr) {
2837 LCONSOLE_INFO("Using FMR for registration\n");
2838 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2840 #endif /* HAVE_OFED_FMR_POOL_API */
2841 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2842 LCONSOLE_INFO("Using FastReg for registration\n");
2843 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2844 #ifndef HAVE_OFED_IB_ALLOC_FAST_REG_MR
2845 #ifdef IB_DEVICE_SG_GAPS_REG
2846 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2847 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2854 rc2 = kiblnd_port_get_attr(hdev);
2861 #ifndef HAVE_OFED_IB_DEVICE_ATTRS
2863 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2867 CERROR("IB device does not support FMRs nor FastRegs, can't "
2868 "register memory: %d\n", rc);
2869 else if (rc == -EINVAL)
2870 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2874 #ifdef HAVE_OFED_IB_GET_DMA_MR
2876 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2878 if (hdev->ibh_mrs == NULL)
2881 ib_dereg_mr(hdev->ibh_mrs);
2883 hdev->ibh_mrs = NULL;
2888 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2890 if (hdev->ibh_event_handler.device != NULL)
2891 ib_unregister_event_handler(&hdev->ibh_event_handler);
2893 #ifdef HAVE_OFED_IB_GET_DMA_MR
2894 kiblnd_hdev_cleanup_mrs(hdev);
2897 if (hdev->ibh_pd != NULL)
2898 ib_dealloc_pd(hdev->ibh_pd);
2900 if (hdev->ibh_cmid != NULL)
2901 rdma_destroy_id(hdev->ibh_cmid);
2903 LIBCFS_FREE(hdev, sizeof(*hdev));
2906 #ifdef HAVE_OFED_IB_GET_DMA_MR
2908 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2911 int acflags = IB_ACCESS_LOCAL_WRITE |
2912 IB_ACCESS_REMOTE_WRITE;
2914 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2916 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2917 kiblnd_hdev_cleanup_mrs(hdev);
2928 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2934 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2936 struct rdma_cm_id *cmid;
2937 struct sockaddr_in srcaddr;
2938 struct sockaddr_in dstaddr;
2941 if (dev->ibd_hdev == NULL || /* initializing */
2942 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2943 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2946 /* XXX: it's UGLY, but I don't have better way to find
2947 * ib-bonding HCA failover because:
2949 * a. no reliable CM event for HCA failover...
2950 * b. no OFED API to get ib_device for current net_device...
2952 * We have only two choices at this point:
2954 * a. rdma_bind_addr(), it will conflict with listener cmid
2955 * b. rdma_resolve_addr() to zero addr */
2956 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2957 RDMA_PS_TCP, IB_QPT_RC);
2960 CERROR("Failed to create cmid for failover: %d\n", rc);
2964 memset(&srcaddr, 0, sizeof(srcaddr));
2965 srcaddr.sin_family = AF_INET;
2966 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2968 memset(&dstaddr, 0, sizeof(dstaddr));
2969 dstaddr.sin_family = AF_INET;
2970 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2971 (struct sockaddr *)&dstaddr, 1);
2972 if (rc != 0 || cmid->device == NULL) {
2973 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2974 dev->ibd_ifname, &dev->ibd_ifip,
2976 rdma_destroy_id(cmid);
2980 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2981 rdma_destroy_id(cmid);
2986 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2988 LIST_HEAD(zombie_tpo);
2989 LIST_HEAD(zombie_ppo);
2990 LIST_HEAD(zombie_fpo);
2991 struct rdma_cm_id *cmid = NULL;
2992 struct kib_hca_dev *hdev = NULL;
2993 struct kib_hca_dev *old;
2995 struct kib_net *net;
2996 struct sockaddr_in addr;
2997 struct net_device *netdev;
2998 unsigned long flags;
3001 bool set_fatal = true;
3003 LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
3004 dev->ibd_can_failover ||
3005 dev->ibd_hdev == NULL);
3007 rc = kiblnd_dev_need_failover(dev, ns);
3011 if (dev->ibd_hdev != NULL &&
3012 dev->ibd_hdev->ibh_cmid != NULL) {
3013 /* XXX it's not good to close old listener at here,
3014 * because we can fail to create new listener.
3015 * But we have to close it now, otherwise rdma_bind_addr
3016 * will return EADDRINUSE... How crap! */
3017 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3019 cmid = dev->ibd_hdev->ibh_cmid;
3020 /* make next schedule of kiblnd_dev_need_failover()
3021 * return 1 for me */
3022 dev->ibd_hdev->ibh_cmid = NULL;
3023 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3025 rdma_destroy_id(cmid);
3028 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
3032 CERROR("Failed to create cmid for failover: %d\n", rc);
3036 memset(&addr, 0, sizeof(addr));
3037 addr.sin_family = AF_INET;
3038 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
3039 addr.sin_port = htons(*kiblnd_tunables.kib_service);
3041 /* Bind to failover device or port */
3042 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
3043 if (rc != 0 || cmid->device == NULL) {
3044 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
3045 dev->ibd_ifname, &dev->ibd_ifip,
3047 if (!rc && !cmid->device)
3049 rdma_destroy_id(cmid);
3053 LIBCFS_ALLOC(hdev, sizeof(*hdev));
3055 CERROR("Failed to allocate kib_hca_dev\n");
3056 rdma_destroy_id(cmid);
3061 atomic_set(&hdev->ibh_ref, 1);
3062 hdev->ibh_dev = dev;
3063 hdev->ibh_cmid = cmid;
3064 hdev->ibh_ibdev = cmid->device;
3065 hdev->ibh_port = cmid->port_num;
3067 #ifdef HAVE_OFED_IB_ALLOC_PD_2ARGS
3068 pd = ib_alloc_pd(cmid->device, 0);
3070 pd = ib_alloc_pd(cmid->device);
3074 CERROR("Can't allocate PD: %d\n", rc);
3080 rc = rdma_listen(cmid, 0);
3082 CERROR("Can't start new listener: %d\n", rc);
3086 rc = kiblnd_hdev_get_attr(hdev);
3088 CERROR("Can't get device attributes: %d\n", rc);
3092 #ifdef HAVE_OFED_IB_GET_DMA_MR
3093 rc = kiblnd_hdev_setup_mrs(hdev);
3095 CERROR("Can't setup device: %d\n", rc);
3100 INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
3101 hdev->ibh_ibdev, kiblnd_event_handler);
3102 ib_register_event_handler(&hdev->ibh_event_handler);
3104 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3106 old = dev->ibd_hdev;
3107 dev->ibd_hdev = hdev; /* take over the refcount */
3110 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
3111 cfs_cpt_for_each(i, lnet_cpt_table()) {
3112 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
3115 if (net->ibn_fmr_ps != NULL)
3116 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
3121 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3123 if (!list_empty(&zombie_tpo))
3124 kiblnd_destroy_pool_list(&zombie_tpo);
3125 if (!list_empty(&zombie_ppo))
3126 kiblnd_destroy_pool_list(&zombie_ppo);
3127 if (!list_empty(&zombie_fpo))
3128 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
3130 kiblnd_hdev_decref(hdev);
3133 dev->ibd_failed_failover++;
3135 dev->ibd_failed_failover = 0;
3139 netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
3140 if (netdev && (lnet_get_link_status(netdev) == 1))
3141 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
3150 kiblnd_destroy_dev(struct kib_dev *dev)
3152 LASSERT(dev->ibd_nnets == 0);
3153 LASSERT(list_empty(&dev->ibd_nets));
3155 list_del(&dev->ibd_fail_list);
3156 list_del(&dev->ibd_list);
3158 if (dev->ibd_hdev != NULL)
3159 kiblnd_hdev_decref(dev->ibd_hdev);
3161 LIBCFS_FREE(dev, sizeof(*dev));
3164 static struct kib_dev *
3165 kiblnd_dev_search(char *ifname)
3167 struct kib_dev *alias = NULL;
3168 struct kib_dev *dev;
3172 colon = strchr(ifname, ':');
3173 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3174 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3180 colon2 = strchr(dev->ibd_ifname, ':');
3186 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3198 kiblnd_handle_link_state_change(struct net_device *dev,
3199 unsigned char operstate)
3201 struct lnet_ni *ni = NULL;
3202 struct kib_dev *event_kibdev;
3203 struct kib_net *net;
3204 struct kib_net *cnxt;
3205 bool link_down = !(operstate == IF_OPER_UP);
3206 struct in_device *in_dev;
3207 bool found_ip = false;
3208 __u32 ni_state_before;
3209 bool update_ping_buf = false;
3211 DECLARE_CONST_IN_IFADDR(ifa);
3213 event_kibdev = kiblnd_dev_search(dev->name);
3218 list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets, ibn_list) {
3222 in_dev = __in_dev_get_rtnl(dev);
3224 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
3226 ni_state_before = lnet_set_link_fatal_state(ni, 1);
3229 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
3230 if (htonl(event_kibdev->ibd_ifip) == ifa->ifa_local)
3236 CDEBUG(D_NET, "Interface %s has no matching ip\n",
3238 ni_state_before = lnet_set_link_fatal_state(ni, 1);
3243 ni_state_before = lnet_set_link_fatal_state(ni, 1);
3245 state = (lnet_get_link_status(dev) == 0);
3246 ni_state_before = lnet_set_link_fatal_state(ni,
3250 if (!update_ping_buf &&
3251 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
3252 (atomic_read(&ni->ni_fatal_error_on) != ni_state_before) &&
3253 (net->ibn_init == IBLND_INIT_ALL))
3254 update_ping_buf = true;
3257 if (update_ping_buf)
3258 lnet_mark_ping_buffer_for_update();
3264 kiblnd_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
3266 struct kib_dev *event_kibdev;
3267 struct kib_net *net;
3268 struct kib_net *cnxt;
3269 struct net_device *event_netdev = ifa->ifa_dev->dev;
3270 __u32 ni_state_before;
3271 bool update_ping_buf = false;
3272 struct lnet_ni *ni = NULL;
3275 event_kibdev = kiblnd_dev_search(event_netdev->name);
3280 if (htonl(event_kibdev->ibd_ifip) != ifa->ifa_local)
3283 list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets,
3286 link_down = (event == NETDEV_DOWN);
3287 ni_state_before = lnet_set_link_fatal_state(ni, link_down);
3288 if (!update_ping_buf &&
3289 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
3290 ((event == NETDEV_DOWN) != ni_state_before) &&
3291 (net->ibn_init == IBLND_INIT_ALL))
3292 update_ping_buf = true;
3295 if (update_ping_buf)
3296 lnet_mark_ping_buffer_for_update();
3302 /************************************
3303 * Net device notifier event handler
3304 ************************************/
3305 static int kiblnd_device_event(struct notifier_block *unused,
3306 unsigned long event, void *ptr)
3308 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
3309 unsigned char operstate;
3311 operstate = dev->operstate;
3313 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
3314 event, dev->name, dev->ifindex, operstate);
3320 kiblnd_handle_link_state_change(dev, operstate);
3327 /************************************
3328 * Inetaddr notifier event handler
3329 ************************************/
3330 static int kiblnd_inetaddr_event(struct notifier_block *unused,
3331 unsigned long event, void *ptr)
3333 struct in_ifaddr *ifa = ptr;
3335 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
3336 event, &ifa->ifa_address, &ifa->ifa_mask);
3342 kiblnd_handle_inetaddr_change(ifa, event);
3349 static struct notifier_block kiblnd_dev_notifier_block = {
3350 .notifier_call = kiblnd_device_event,
3353 static struct notifier_block kiblnd_inetaddr_notifier_block = {
3354 .notifier_call = kiblnd_inetaddr_event,
3358 kiblnd_base_shutdown(void)
3360 struct kib_sched_info *sched;
3361 struct kib_peer_ni *peer_ni;
3364 LASSERT(list_empty(&kiblnd_data.kib_devs));
3366 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
3367 libcfs_kmem_read());
3369 if (kiblnd_data.kib_init == IBLND_INIT_ALL) {
3370 unregister_netdevice_notifier(&kiblnd_dev_notifier_block);
3371 unregister_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3374 switch (kiblnd_data.kib_init) {
3378 case IBLND_INIT_ALL:
3379 case IBLND_INIT_DATA:
3380 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
3382 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
3383 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
3384 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
3385 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
3387 /* flag threads to terminate; wake and wait for them to die */
3388 kiblnd_data.kib_shutdown = 1;
3390 /* NB: we really want to stop scheduler threads net by net
3391 * instead of the whole module, this should be improved
3392 * with dynamic configuration LNet.
3394 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
3395 wake_up_all(&sched->ibs_waitq);
3397 wake_up(&kiblnd_data.kib_connd_waitq);
3398 wake_up(&kiblnd_data.kib_failover_waitq);
3400 wait_var_event_warning(&kiblnd_data.kib_nthreads,
3401 !atomic_read(&kiblnd_data.kib_nthreads),
3402 "Waiting for %d threads to terminate\n",
3403 atomic_read(&kiblnd_data.kib_nthreads));
3406 case IBLND_INIT_NOTHING:
3410 if (kiblnd_data.kib_scheds != NULL)
3411 cfs_percpt_free(kiblnd_data.kib_scheds);
3413 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3414 libcfs_kmem_read());
3416 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3417 module_put(THIS_MODULE);
3421 kiblnd_shutdown(struct lnet_ni *ni)
3423 struct kib_net *net = ni->ni_data;
3424 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
3425 unsigned long flags;
3427 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3432 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3433 libcfs_kmem_read());
3435 write_lock_irqsave(g_lock, flags);
3436 net->ibn_shutdown = 1;
3437 write_unlock_irqrestore(g_lock, flags);
3439 switch (net->ibn_init) {
3443 case IBLND_INIT_ALL:
3444 /* nuke all existing peers within this net */
3445 kiblnd_del_peer(ni, LNET_NID_ANY);
3447 /* Wait for all peer_ni state to clean up */
3448 wait_var_event_warning(&net->ibn_npeers,
3449 atomic_read(&net->ibn_npeers) == 0,
3450 "%s: waiting for %d peers to disconnect\n",
3451 libcfs_nidstr(&ni->ni_nid),
3452 atomic_read(&net->ibn_npeers));
3454 kiblnd_net_fini_pools(net);
3456 write_lock_irqsave(g_lock, flags);
3457 LASSERT(net->ibn_dev->ibd_nnets > 0);
3458 net->ibn_dev->ibd_nnets--;
3459 list_del(&net->ibn_list);
3460 write_unlock_irqrestore(g_lock, flags);
3462 wake_up_all(&kiblnd_data.kib_connd_waitq);
3463 wait_var_event_warning(&net->ibn_nconns,
3464 atomic_read(&net->ibn_nconns) == 0,
3465 "%s: waiting for %d conns to clean\n",
3466 libcfs_nidstr(&ni->ni_nid),
3467 atomic_read(&net->ibn_nconns));
3470 case IBLND_INIT_NOTHING:
3471 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3473 if (net->ibn_dev != NULL &&
3474 net->ibn_dev->ibd_nnets == 0)
3475 kiblnd_destroy_dev(net->ibn_dev);
3480 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3481 libcfs_kmem_read());
3483 net->ibn_init = IBLND_INIT_NOTHING;
3486 LIBCFS_FREE(net, sizeof(*net));
3489 if (list_empty(&kiblnd_data.kib_devs))
3490 kiblnd_base_shutdown();
3494 kiblnd_base_startup(struct net *ns)
3496 struct kib_sched_info *sched;
3500 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3502 if (!try_module_get(THIS_MODULE))
3505 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3507 rwlock_init(&kiblnd_data.kib_global_lock);
3509 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3510 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3512 hash_init(kiblnd_data.kib_peers);
3514 spin_lock_init(&kiblnd_data.kib_connd_lock);
3515 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3516 INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3517 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3518 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3519 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3521 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3522 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3524 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3526 if (kiblnd_data.kib_scheds == NULL)
3529 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3532 spin_lock_init(&sched->ibs_lock);
3533 INIT_LIST_HEAD(&sched->ibs_conns);
3534 init_waitqueue_head(&sched->ibs_waitq);
3536 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3537 if (*kiblnd_tunables.kib_nscheds > 0) {
3538 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3540 /* max to half of CPUs, another half is reserved for
3541 * upper layer modules */
3542 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3545 sched->ibs_nthreads_max = nthrs;
3549 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3551 /* lists/ptrs/locks initialised */
3552 kiblnd_data.kib_init = IBLND_INIT_DATA;
3553 /*****************************************************/
3555 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3557 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3561 if (*kiblnd_tunables.kib_dev_failover != 0)
3562 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3566 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3570 register_netdevice_notifier(&kiblnd_dev_notifier_block);
3571 register_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3573 /* flag everything initialised */
3574 kiblnd_data.kib_init = IBLND_INIT_ALL;
3575 /*****************************************************/
3580 kiblnd_base_shutdown();
3585 kiblnd_start_schedulers(struct kib_sched_info *sched)
3591 if (sched->ibs_nthreads == 0) {
3592 if (*kiblnd_tunables.kib_nscheds > 0) {
3593 nthrs = sched->ibs_nthreads_max;
3595 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3597 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3598 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3601 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3602 /* increase one thread if there is new interface */
3603 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3606 for (i = 0; i < nthrs; i++) {
3607 long id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3609 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3610 "kiblnd_sd_%02ld_%02ld",
3611 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3615 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3616 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3620 sched->ibs_nthreads += i;
3624 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3631 for (i = 0; i < ncpts; i++) {
3632 struct kib_sched_info *sched;
3634 cpt = (cpts == NULL) ? i : cpts[i];
3635 sched = kiblnd_data.kib_scheds[cpt];
3637 if (!newdev && sched->ibs_nthreads > 0)
3640 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3642 CERROR("Failed to start scheduler threads for %s\n",
3651 kiblnd_startup(struct lnet_ni *ni)
3653 char *ifname = NULL;
3654 struct lnet_inetdev *ifaces = NULL;
3655 struct kib_dev *ibdev = NULL;
3656 struct kib_net *net = NULL;
3657 unsigned long flags;
3661 struct net_device *netdev;
3663 LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3665 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3666 rc = kiblnd_base_startup(ni->ni_net_ns);
3671 LIBCFS_ALLOC(net, sizeof(*net));
3679 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3681 kiblnd_tunables_setup(ni);
3683 /* Multi-Rail wants each secondary
3684 * IP to be treated as an unique 'struct ni' interface.
3686 if (ni->ni_interface != NULL) {
3687 /* Use the IPoIB interface specified in 'networks=' */
3688 ifname = ni->ni_interface;
3690 ifname = *kiblnd_tunables.kib_default_ipif;
3691 rc = libcfs_strnid(&ni->ni_nid, ifname);
3692 if (rc < 0 || ni->ni_nid.nid_type != O2IBLND)
3693 memset(&ni->ni_nid, 0, sizeof(ni->ni_nid));
3696 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3697 CERROR("IPoIB interface name too long: %s\n", ifname);
3702 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, false);
3706 i = lnet_inet_select(ni, ifaces, rc);
3710 if (nid_addr_is_set(&ni->ni_nid)) {
3711 strscpy(ifname, ifaces[i].li_name, sizeof(ifname));
3712 } else if (strcmp(ifname, ifaces[i].li_name) != 0) {
3713 CERROR("ko2iblnd: No matching interfaces\n");
3718 ibdev = kiblnd_dev_search(ifname);
3719 newdev = ibdev == NULL;
3720 /* hmm...create kib_dev even for alias */
3721 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3722 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3728 ibdev->ibd_ifip = ntohl(ifaces[i].li_ipaddr);
3729 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3730 sizeof(ibdev->ibd_ifname));
3731 ibdev->ibd_can_failover = ifaces[i].li_iff_master;
3733 INIT_LIST_HEAD(&ibdev->ibd_nets);
3734 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3735 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3737 /* initialize the device */
3738 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3740 CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3745 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3748 net->ibn_dev = ibdev;
3749 ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3750 if (!ni->ni_interface) {
3751 rc = lnet_ni_add_interface(ni, ifaces[i].li_name);
3753 CWARN("ko2iblnd failed to allocate ni_interface\n");
3755 ni->ni_dev_cpt = ifaces[i].li_cpt;
3757 rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3761 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3763 CERROR("Failed to initialize NI pools: %d\n", rc);
3767 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3769 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3770 /* for health check */
3771 if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3772 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3775 netdev = dev_get_by_name_rcu(ni->ni_net_ns, net->ibn_dev->ibd_ifname);
3777 ((netdev->reg_state == NETREG_UNREGISTERING) ||
3778 (netdev->operstate != IF_OPER_UP) ||
3779 (lnet_get_link_status(netdev) == 0))) {
3780 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3784 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3786 net->ibn_init = IBLND_INIT_ALL;
3791 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3792 kiblnd_destroy_dev(ibdev);
3795 kiblnd_shutdown(ni);
3797 CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3798 ifname ? ifname : "", rc);
3803 static const struct lnet_lnd the_o2iblnd = {
3804 .lnd_type = O2IBLND,
3805 .lnd_startup = kiblnd_startup,
3806 .lnd_shutdown = kiblnd_shutdown,
3807 .lnd_ctl = kiblnd_ctl,
3808 .lnd_send = kiblnd_send,
3809 .lnd_recv = kiblnd_recv,
3810 .lnd_get_dev_prio = kiblnd_get_dev_prio,
3811 .lnd_nl_get = kiblnd_nl_get,
3812 .lnd_nl_set = kiblnd_nl_set,
3813 .lnd_keys = &kiblnd_tunables_keys,
3816 static void ko2inlnd_assert_wire_constants(void)
3818 BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3819 BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3820 BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3821 BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3823 BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3824 BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3825 BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3826 BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3827 BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3828 BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3829 BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3830 BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3831 BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3832 BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3834 BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3835 BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3836 BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3837 BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3838 BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3839 BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3840 BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3841 BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3843 BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3844 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3845 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3846 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3847 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3848 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3849 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3851 BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3852 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3853 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3854 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3855 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3857 BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3858 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3859 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3860 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3861 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3863 BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3864 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3865 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3866 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3867 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3868 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3869 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3871 BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3872 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3873 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3874 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3875 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3877 BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3878 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3879 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3880 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3881 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3882 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3883 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3885 BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3886 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3887 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3888 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3889 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3890 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3891 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3893 BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3894 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3895 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3896 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3897 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3899 /* Checks for struct kib_msg */
3900 //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3901 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3902 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3903 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3904 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3905 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3906 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3907 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3908 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3909 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3910 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3911 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3912 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3913 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3914 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3915 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3916 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3917 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3918 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3919 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3920 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3923 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3924 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3925 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3926 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3927 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3928 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3930 /* Immediate message */
3931 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3932 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3933 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3934 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3936 /* PUT req message */
3937 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3938 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3939 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3940 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3943 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3944 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3945 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3946 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3947 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3948 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3951 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3952 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3953 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3954 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3955 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3956 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3958 /* Completion message */
3959 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3960 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3961 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3962 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3965 BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3966 BUILD_BUG_ON(offsetof(struct kib_msg,
3967 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3969 BUILD_BUG_ON(offsetof(struct kib_msg,
3970 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3974 static void __exit ko2iblnd_exit(void)
3976 lnet_unregister_lnd(&the_o2iblnd);
3979 static int __init ko2iblnd_init(void)
3983 ko2inlnd_assert_wire_constants();
3985 rc = kiblnd_tunables_init();
3989 rc = libcfs_setup();
3993 lnet_register_lnd(&the_o2iblnd);
3998 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3999 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
4000 MODULE_VERSION("2.8.0");
4001 MODULE_LICENSE("GPL");
4003 module_init(ko2iblnd_init);
4004 module_exit(ko2iblnd_exit);