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,
1258 [LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TOS] = {
1260 .lkp_data_type = NLA_S16,
1266 kiblnd_nl_get(int cmd, struct sk_buff *msg, int type, void *data)
1268 struct lnet_ioctl_config_o2iblnd_tunables *tuns;
1269 struct lnet_ni *ni = data;
1274 if (cmd != LNET_CMD_NETS || type != LNET_NET_LOCAL_NI_ATTR_LND_TUNABLES)
1277 tuns = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1278 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS,
1279 tuns->lnd_peercredits_hiw);
1280 if (tuns->lnd_map_on_demand) {
1282 LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND);
1284 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS,
1285 tuns->lnd_concurrent_sends);
1286 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE,
1287 tuns->lnd_fmr_pool_size);
1288 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER,
1289 tuns->lnd_fmr_flush_trigger);
1290 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE,
1291 tuns->lnd_fmr_cache);
1292 nla_put_u16(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_NTX, tuns->lnd_ntx);
1293 nla_put_u16(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER,
1294 tuns->lnd_conns_per_peer);
1295 nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT,
1297 nla_put_s16(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TOS,
1304 kiblnd_nl_set_default(int cmd, int type, void *data)
1306 struct lnet_lnd_tunables *tunables = data;
1307 struct lnet_ioctl_config_o2iblnd_tunables *lt;
1308 struct lnet_ioctl_config_o2iblnd_tunables *df;
1310 lt = &tunables->lnd_tun_u.lnd_o2ib;
1311 df = &kib_default_tunables;
1313 case LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS:
1314 lt->lnd_peercredits_hiw = df->lnd_peercredits_hiw;
1316 case LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND:
1317 lt->lnd_map_on_demand = df->lnd_map_on_demand;
1319 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS:
1320 lt->lnd_concurrent_sends = df->lnd_concurrent_sends;
1322 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE:
1323 lt->lnd_fmr_pool_size = df->lnd_fmr_pool_size;
1325 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER:
1326 lt->lnd_fmr_flush_trigger = df->lnd_fmr_flush_trigger;
1328 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE:
1329 lt->lnd_fmr_cache = df->lnd_fmr_cache;
1331 case LNET_NET_O2IBLND_TUNABLES_ATTR_NTX:
1332 lt->lnd_ntx = df->lnd_ntx;
1334 case LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT:
1335 lt->lnd_timeout = df->lnd_timeout;
1337 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER:
1338 lt->lnd_conns_per_peer = df->lnd_conns_per_peer;
1347 kiblnd_nl_set(int cmd, struct nlattr *attr, int type, void *data)
1349 struct lnet_lnd_tunables *tunables = data;
1353 if (cmd != LNET_CMD_NETS)
1357 kiblnd_nl_set_default(cmd, type, data);
1361 if (nla_type(attr) != LN_SCALAR_ATTR_INT_VALUE)
1365 case LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS:
1366 tunables->lnd_tun_u.lnd_o2ib.lnd_peercredits_hiw = nla_get_s64(attr);
1368 case LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND:
1369 tunables->lnd_tun_u.lnd_o2ib.lnd_map_on_demand = nla_get_s64(attr);
1371 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS:
1372 tunables->lnd_tun_u.lnd_o2ib.lnd_concurrent_sends = nla_get_s64(attr);
1374 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE:
1375 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_pool_size = nla_get_s64(attr);
1377 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER:
1378 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_flush_trigger = nla_get_s64(attr);
1380 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE:
1381 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_cache = nla_get_s64(attr);
1383 case LNET_NET_O2IBLND_TUNABLES_ATTR_NTX:
1384 tunables->lnd_tun_u.lnd_o2ib.lnd_ntx = nla_get_s64(attr);
1386 case LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT:
1387 tunables->lnd_tun_u.lnd_o2ib.lnd_timeout = nla_get_s64(attr);
1389 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER:
1390 num = nla_get_s64(attr);
1391 if (num >= 0 && num < 128)
1392 tunables->lnd_tun_u.lnd_o2ib.lnd_conns_per_peer = num;
1396 case LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TOS:
1397 num = nla_get_s64(attr);
1398 tunables->lnd_tun_u.lnd_o2ib.lnd_tos = num;
1408 kiblnd_free_pages(struct kib_pages *p)
1410 int npages = p->ibp_npages;
1413 for (i = 0; i < npages; i++) {
1414 if (p->ibp_pages[i] != NULL)
1415 __free_page(p->ibp_pages[i]);
1418 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1422 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1424 struct kib_pages *p;
1427 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1428 offsetof(struct kib_pages, ibp_pages[npages]));
1430 CERROR("Can't allocate descriptor for %d pages\n", npages);
1434 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1435 p->ibp_npages = npages;
1437 for (i = 0; i < npages; i++) {
1438 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1440 if (p->ibp_pages[i] == NULL) {
1441 CERROR("Can't allocate page %d of %d\n", i, npages);
1442 kiblnd_free_pages(p);
1452 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1457 LASSERT (conn->ibc_rxs != NULL);
1458 LASSERT (conn->ibc_hdev != NULL);
1460 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1461 rx = &conn->ibc_rxs[i];
1463 LASSERT(rx->rx_nob >= 0); /* not posted */
1465 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1466 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1468 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1471 kiblnd_free_pages(conn->ibc_rx_pages);
1473 conn->ibc_rx_pages = NULL;
1477 kiblnd_map_rx_descs(struct kib_conn *conn)
1485 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1486 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1487 rx = &conn->ibc_rxs[i];
1490 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1493 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1494 rx->rx_msg, IBLND_MSG_SIZE,
1496 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1498 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1500 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1501 i, rx->rx_msg, rx->rx_msgaddr,
1502 (__u64)(page_to_phys(pg) + pg_off));
1504 pg_off += IBLND_MSG_SIZE;
1505 LASSERT(pg_off <= PAGE_SIZE);
1507 if (pg_off == PAGE_SIZE) {
1510 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1516 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1518 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1522 LASSERT (tpo->tpo_pool.po_allocated == 0);
1527 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1528 tx = &tpo->tpo_tx_descs[i];
1529 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1530 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1532 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1535 kiblnd_hdev_decref(hdev);
1536 tpo->tpo_hdev = NULL;
1539 static struct kib_hca_dev *
1540 kiblnd_current_hdev(struct kib_dev *dev)
1542 struct kib_hca_dev *hdev;
1543 unsigned long flags;
1546 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1547 while (dev->ibd_failover) {
1548 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1550 CDEBUG(D_NET, "%s: Wait for failover\n",
1552 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1554 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1557 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1558 hdev = dev->ibd_hdev;
1560 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1566 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1568 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1569 struct kib_pool *pool = &tpo->tpo_pool;
1570 struct kib_net *net = pool->po_owner->ps_net;
1571 struct kib_dev *dev;
1578 LASSERT (net != NULL);
1582 /* pre-mapped messages are not bigger than 1 page */
1583 BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1585 /* No fancy arithmetic when we do the buffer calculations */
1586 BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1588 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1590 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1591 page = txpgs->ibp_pages[ipage];
1592 tx = &tpo->tpo_tx_descs[i];
1594 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1597 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1601 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1603 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1605 list_add(&tx->tx_list, &pool->po_free_list);
1607 page_offset += IBLND_MSG_SIZE;
1608 LASSERT(page_offset <= PAGE_SIZE);
1610 if (page_offset == PAGE_SIZE) {
1613 LASSERT(ipage <= txpgs->ibp_npages);
1619 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1621 LASSERT(fpo->fpo_map_count == 0);
1623 #ifdef HAVE_OFED_FMR_POOL_API
1624 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1625 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1627 #endif /* HAVE_OFED_FMR_POOL_API */
1629 struct kib_fast_reg_descriptor *frd, *tmp;
1632 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1634 list_del(&frd->frd_list);
1635 #ifndef HAVE_OFED_IB_MAP_MR_SG
1636 ib_free_fast_reg_page_list(frd->frd_frpl);
1638 ib_dereg_mr(frd->frd_mr);
1639 LIBCFS_FREE(frd, sizeof(*frd));
1642 if (i < fpo->fast_reg.fpo_pool_size)
1643 CERROR("FastReg pool still has %d regions registered\n",
1644 fpo->fast_reg.fpo_pool_size - i);
1648 kiblnd_hdev_decref(fpo->fpo_hdev);
1650 LIBCFS_FREE(fpo, sizeof(*fpo));
1654 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1656 struct kib_fmr_pool *fpo, *tmp;
1658 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1659 list_del(&fpo->fpo_list);
1660 kiblnd_destroy_fmr_pool(fpo);
1665 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1668 int size = tunables->lnd_fmr_pool_size / ncpts;
1670 return max(IBLND_FMR_POOL, size);
1674 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1677 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1679 return max(IBLND_FMR_POOL_FLUSH, size);
1682 #ifdef HAVE_OFED_FMR_POOL_API
1683 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1684 struct kib_fmr_pool *fpo)
1686 struct ib_fmr_pool_param param = {
1687 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1688 .page_shift = PAGE_SHIFT,
1689 .access = (IB_ACCESS_LOCAL_WRITE |
1690 IB_ACCESS_REMOTE_WRITE),
1691 .pool_size = fps->fps_pool_size,
1692 .dirty_watermark = fps->fps_flush_trigger,
1693 .flush_function = NULL,
1695 .cache = !!fps->fps_cache };
1698 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1700 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1701 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1703 CERROR("Failed to create FMR pool: %d\n", rc);
1705 CERROR("FMRs are not supported\n");
1707 fpo->fpo_is_fmr = true;
1711 #endif /* HAVE_OFED_FMR_POOL_API */
1713 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1714 struct kib_fmr_pool *fpo,
1715 enum kib_dev_caps dev_caps)
1717 struct kib_fast_reg_descriptor *frd, *tmp;
1720 #ifdef HAVE_OFED_FMR_POOL_API
1721 fpo->fpo_is_fmr = false;
1724 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1725 fpo->fast_reg.fpo_pool_size = 0;
1726 for (i = 0; i < fps->fps_pool_size; i++) {
1727 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1730 CERROR("Failed to allocate a new fast_reg descriptor\n");
1736 #ifndef HAVE_OFED_IB_MAP_MR_SG
1737 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1738 IBLND_MAX_RDMA_FRAGS);
1739 if (IS_ERR(frd->frd_frpl)) {
1740 rc = PTR_ERR(frd->frd_frpl);
1741 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1743 frd->frd_frpl = NULL;
1748 #ifdef HAVE_OFED_IB_ALLOC_FAST_REG_MR
1749 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1750 IBLND_MAX_RDMA_FRAGS);
1753 * it is expected to get here if this is an MLX-5 card.
1754 * MLX-4 cards will always use FMR and MLX-5 cards will
1755 * always use fast_reg. It turns out that some MLX-5 cards
1756 * (possibly due to older FW versions) do not natively support
1757 * gaps. So we will need to track them here.
1759 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1760 #ifdef IB_MR_TYPE_SG_GAPS
1761 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1762 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1763 IB_MR_TYPE_SG_GAPS :
1768 IBLND_MAX_RDMA_FRAGS);
1769 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1770 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1771 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1773 if (IS_ERR(frd->frd_mr)) {
1774 rc = PTR_ERR(frd->frd_mr);
1775 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1780 /* indicate that the local invalidate needs to be generated */
1781 frd->frd_valid = false;
1783 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1784 fpo->fast_reg.fpo_pool_size++;
1791 ib_dereg_mr(frd->frd_mr);
1792 #ifndef HAVE_OFED_IB_MAP_MR_SG
1794 ib_free_fast_reg_page_list(frd->frd_frpl);
1796 LIBCFS_FREE(frd, sizeof(*frd));
1799 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1801 list_del(&frd->frd_list);
1802 #ifndef HAVE_OFED_IB_MAP_MR_SG
1803 ib_free_fast_reg_page_list(frd->frd_frpl);
1805 ib_dereg_mr(frd->frd_mr);
1806 LIBCFS_FREE(frd, sizeof(*frd));
1812 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1813 struct kib_fmr_pool **pp_fpo)
1815 struct kib_dev *dev = fps->fps_net->ibn_dev;
1816 struct kib_fmr_pool *fpo;
1819 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1823 memset(fpo, 0, sizeof(*fpo));
1825 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1827 #ifdef HAVE_OFED_FMR_POOL_API
1828 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1829 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1831 #endif /* HAVE_OFED_FMR_POOL_API */
1832 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1836 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1837 fpo->fpo_owner = fps;
1843 kiblnd_hdev_decref(fpo->fpo_hdev);
1844 LIBCFS_FREE(fpo, sizeof(*fpo));
1849 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1851 struct kib_fmr_pool *fpo;
1853 if (fps->fps_net == NULL) /* intialized? */
1856 spin_lock(&fps->fps_lock);
1858 while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1859 struct kib_fmr_pool,
1860 fpo_list)) != NULL) {
1861 fpo->fpo_failed = 1;
1862 if (fpo->fpo_map_count == 0)
1863 list_move(&fpo->fpo_list, zombies);
1865 list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1868 spin_unlock(&fps->fps_lock);
1872 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1874 if (fps->fps_net != NULL) { /* initialized? */
1875 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1876 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1881 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1882 struct kib_net *net,
1883 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1885 struct kib_fmr_pool *fpo;
1888 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1893 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1894 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1895 fps->fps_cache = tunables->lnd_fmr_cache;
1897 spin_lock_init(&fps->fps_lock);
1898 INIT_LIST_HEAD(&fps->fps_pool_list);
1899 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1901 rc = kiblnd_create_fmr_pool(fps, &fpo);
1903 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1909 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1911 if (fpo->fpo_map_count != 0) /* still in use */
1913 if (fpo->fpo_failed)
1915 return now >= fpo->fpo_deadline;
1918 #if defined(HAVE_OFED_FMR_POOL_API) || !defined(HAVE_OFED_IB_MAP_MR_SG)
1920 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1922 struct kib_hca_dev *hdev;
1923 __u64 *pages = tx->tx_pages;
1928 hdev = tx->tx_pool->tpo_hdev;
1930 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1931 for (size = 0; size < rd->rd_frags[i].rf_nob;
1932 size += hdev->ibh_page_size) {
1933 pages[npages++] = (rd->rd_frags[i].rf_addr &
1934 hdev->ibh_page_mask) + size;
1943 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1946 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1947 struct kib_fmr_poolset *fps;
1948 time64_t now = ktime_get_seconds();
1949 struct kib_fmr_pool *tmp;
1954 fps = fpo->fpo_owner;
1956 #ifdef HAVE_OFED_FMR_POOL_API
1957 if (fpo->fpo_is_fmr) {
1958 if (fmr->fmr_pfmr) {
1959 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1960 fmr->fmr_pfmr = NULL;
1964 int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1968 #endif /* HAVE_OFED_FMR_POOL_API */
1970 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1972 frd->frd_posted = false;
1973 fmr->fmr_frd = NULL;
1974 spin_lock(&fps->fps_lock);
1975 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1976 spin_unlock(&fps->fps_lock);
1979 fmr->fmr_pool = NULL;
1981 spin_lock(&fps->fps_lock);
1982 fpo->fpo_map_count--; /* decref the pool */
1984 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1985 /* the first pool is persistent */
1986 if (fps->fps_pool_list.next == &fpo->fpo_list)
1989 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1990 list_move(&fpo->fpo_list, &zombies);
1994 spin_unlock(&fps->fps_lock);
1996 if (!list_empty(&zombies))
1997 kiblnd_destroy_fmr_pool_list(&zombies);
2000 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
2001 struct kib_rdma_desc *rd, u32 nob, u64 iov,
2002 struct kib_fmr *fmr)
2004 struct kib_fmr_pool *fpo;
2006 bool is_rx = (rd != tx->tx_rd);
2007 #ifdef HAVE_OFED_FMR_POOL_API
2008 __u64 *pages = tx->tx_pages;
2009 bool tx_pages_mapped = false;
2015 spin_lock(&fps->fps_lock);
2016 version = fps->fps_version;
2017 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
2018 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2019 fpo->fpo_map_count++;
2021 #ifdef HAVE_OFED_FMR_POOL_API
2022 fmr->fmr_pfmr = NULL;
2023 if (fpo->fpo_is_fmr) {
2024 struct ib_pool_fmr *pfmr;
2026 spin_unlock(&fps->fps_lock);
2028 if (!tx_pages_mapped) {
2029 npages = kiblnd_map_tx_pages(tx, rd);
2030 tx_pages_mapped = true;
2033 pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
2034 pages, npages, iov);
2035 if (likely(!IS_ERR(pfmr))) {
2036 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
2038 fmr->fmr_frd = NULL;
2039 fmr->fmr_pfmr = pfmr;
2040 fmr->fmr_pool = fpo;
2045 #endif /* HAVE_OFED_FMR_POOL_API */
2047 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
2048 struct kib_fast_reg_descriptor *frd;
2049 #ifdef HAVE_OFED_IB_MAP_MR_SG
2050 struct ib_reg_wr *wr;
2053 struct ib_rdma_wr *wr;
2054 struct ib_fast_reg_page_list *frpl;
2058 frd = list_first_entry(
2059 &fpo->fast_reg.fpo_pool_list,
2060 struct kib_fast_reg_descriptor,
2062 list_del(&frd->frd_list);
2063 spin_unlock(&fps->fps_lock);
2065 #ifndef HAVE_OFED_IB_MAP_MR_SG
2066 frpl = frd->frd_frpl;
2070 if (!frd->frd_valid) {
2071 struct ib_rdma_wr *inv_wr;
2072 __u32 key = is_rx ? mr->rkey : mr->lkey;
2074 frd->frd_valid = true;
2075 inv_wr = &frd->frd_inv_wr;
2076 memset(inv_wr, 0, sizeof(*inv_wr));
2078 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
2079 inv_wr->wr.wr_id = IBLND_WID_MR;
2080 inv_wr->wr.ex.invalidate_rkey = key;
2083 key = ib_inc_rkey(key);
2084 ib_update_fast_reg_key(mr, key);
2087 #ifdef HAVE_OFED_IB_MAP_MR_SG
2088 #ifdef HAVE_OFED_IB_MAP_MR_SG_5ARGS
2089 n = ib_map_mr_sg(mr, tx->tx_frags,
2090 rd->rd_nfrags, NULL, PAGE_SIZE);
2092 n = ib_map_mr_sg(mr, tx->tx_frags,
2093 rd->rd_nfrags, PAGE_SIZE);
2094 #endif /* HAVE_OFED_IB_MAP_MR_SG_5ARGS */
2095 if (unlikely(n != rd->rd_nfrags)) {
2096 CERROR("Failed to map mr %d/%d elements\n",
2098 return n < 0 ? n : -EINVAL;
2101 wr = &frd->frd_fastreg_wr;
2102 memset(wr, 0, sizeof(*wr));
2104 wr->wr.opcode = IB_WR_REG_MR;
2105 wr->wr.wr_id = IBLND_WID_MR;
2107 wr->wr.send_flags = 0;
2109 wr->key = is_rx ? mr->rkey : mr->lkey;
2110 wr->access = (IB_ACCESS_LOCAL_WRITE |
2111 IB_ACCESS_REMOTE_WRITE);
2112 #else /* HAVE_OFED_IB_MAP_MR_SG */
2113 if (!tx_pages_mapped) {
2114 npages = kiblnd_map_tx_pages(tx, rd);
2115 tx_pages_mapped = true;
2118 LASSERT(npages <= frpl->max_page_list_len);
2119 memcpy(frpl->page_list, pages,
2120 sizeof(*pages) * npages);
2122 /* Prepare FastReg WR */
2123 wr = &frd->frd_fastreg_wr;
2124 memset(wr, 0, sizeof(*wr));
2126 wr->wr.opcode = IB_WR_FAST_REG_MR;
2127 wr->wr.wr_id = IBLND_WID_MR;
2129 wr->wr.wr.fast_reg.iova_start = iov;
2130 wr->wr.wr.fast_reg.page_list = frpl;
2131 wr->wr.wr.fast_reg.page_list_len = npages;
2132 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2133 wr->wr.wr.fast_reg.length = nob;
2134 wr->wr.wr.fast_reg.rkey =
2135 is_rx ? mr->rkey : mr->lkey;
2136 wr->wr.wr.fast_reg.access_flags =
2137 (IB_ACCESS_LOCAL_WRITE |
2138 IB_ACCESS_REMOTE_WRITE);
2139 #endif /* HAVE_OFED_IB_MAP_MR_SG */
2141 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
2143 fmr->fmr_pool = fpo;
2144 frd->frd_posted = false;
2147 spin_unlock(&fps->fps_lock);
2151 spin_lock(&fps->fps_lock);
2152 fpo->fpo_map_count--;
2153 if (rc != -EAGAIN) {
2154 spin_unlock(&fps->fps_lock);
2158 /* EAGAIN and ... */
2159 if (version != fps->fps_version) {
2160 spin_unlock(&fps->fps_lock);
2165 if (fps->fps_increasing) {
2166 spin_unlock(&fps->fps_lock);
2167 CDEBUG(D_NET, "Another thread is allocating new "
2168 "FMR pool, waiting for her to complete\n");
2169 wait_var_event(fps, !fps->fps_increasing);
2174 if (ktime_get_seconds() < fps->fps_next_retry) {
2175 /* someone failed recently */
2176 spin_unlock(&fps->fps_lock);
2180 fps->fps_increasing = 1;
2181 spin_unlock(&fps->fps_lock);
2183 CDEBUG(D_NET, "Allocate new FMR pool\n");
2184 rc = kiblnd_create_fmr_pool(fps, &fpo);
2185 spin_lock(&fps->fps_lock);
2186 fps->fps_increasing = 0;
2190 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2192 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2194 spin_unlock(&fps->fps_lock);
2200 kiblnd_fini_pool(struct kib_pool *pool)
2202 LASSERT(list_empty(&pool->po_free_list));
2203 LASSERT(pool->po_allocated == 0);
2205 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2209 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2211 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2213 memset(pool, 0, sizeof(struct kib_pool));
2214 INIT_LIST_HEAD(&pool->po_free_list);
2215 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2216 pool->po_owner = ps;
2217 pool->po_size = size;
2221 kiblnd_destroy_pool_list(struct list_head *head)
2223 struct kib_pool *pool;
2225 while ((pool = list_first_entry_or_null(head,
2227 po_list)) != NULL) {
2228 list_del(&pool->po_list);
2230 LASSERT(pool->po_owner != NULL);
2231 pool->po_owner->ps_pool_destroy(pool);
2236 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2238 struct kib_pool *po;
2240 if (ps->ps_net == NULL) /* intialized? */
2243 spin_lock(&ps->ps_lock);
2244 while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2246 po_list)) != NULL) {
2248 if (po->po_allocated == 0)
2249 list_move(&po->po_list, zombies);
2251 list_move(&po->po_list, &ps->ps_failed_pool_list);
2253 spin_unlock(&ps->ps_lock);
2257 kiblnd_fini_poolset(struct kib_poolset *ps)
2259 if (ps->ps_net != NULL) { /* initialized? */
2260 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2261 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2266 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2267 struct kib_net *net, char *name, int size,
2268 kib_ps_pool_create_t po_create,
2269 kib_ps_pool_destroy_t po_destroy,
2270 kib_ps_node_init_t nd_init,
2271 kib_ps_node_fini_t nd_fini)
2273 struct kib_pool *pool;
2276 memset(ps, 0, sizeof(struct kib_poolset));
2280 ps->ps_pool_create = po_create;
2281 ps->ps_pool_destroy = po_destroy;
2282 ps->ps_node_init = nd_init;
2283 ps->ps_node_fini = nd_fini;
2284 ps->ps_pool_size = size;
2285 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2286 >= sizeof(ps->ps_name))
2288 spin_lock_init(&ps->ps_lock);
2289 INIT_LIST_HEAD(&ps->ps_pool_list);
2290 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2292 rc = ps->ps_pool_create(ps, size, &pool);
2294 list_add(&pool->po_list, &ps->ps_pool_list);
2296 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2302 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2304 if (pool->po_allocated != 0) /* still in use */
2306 if (pool->po_failed)
2308 return now >= pool->po_deadline;
2312 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2315 struct kib_poolset *ps = pool->po_owner;
2316 struct kib_pool *tmp;
2317 time64_t now = ktime_get_seconds();
2319 spin_lock(&ps->ps_lock);
2321 if (ps->ps_node_fini != NULL)
2322 ps->ps_node_fini(pool, node);
2324 LASSERT(pool->po_allocated > 0);
2325 list_add(node, &pool->po_free_list);
2326 pool->po_allocated--;
2328 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2329 /* the first pool is persistent */
2330 if (ps->ps_pool_list.next == &pool->po_list)
2333 if (kiblnd_pool_is_idle(pool, now))
2334 list_move(&pool->po_list, &zombies);
2336 spin_unlock(&ps->ps_lock);
2338 if (!list_empty(&zombies))
2339 kiblnd_destroy_pool_list(&zombies);
2343 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2345 struct list_head *node;
2346 struct kib_pool *pool;
2348 unsigned int interval = 1;
2349 ktime_t time_before;
2350 unsigned int trips = 0;
2353 spin_lock(&ps->ps_lock);
2354 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2355 if (list_empty(&pool->po_free_list))
2358 pool->po_allocated++;
2359 pool->po_deadline = ktime_get_seconds() +
2360 IBLND_POOL_DEADLINE;
2361 node = pool->po_free_list.next;
2364 if (ps->ps_node_init != NULL) {
2365 /* still hold the lock */
2366 ps->ps_node_init(pool, node);
2368 spin_unlock(&ps->ps_lock);
2372 /* no available tx pool and ... */
2373 if (ps->ps_increasing) {
2374 /* another thread is allocating a new pool */
2375 spin_unlock(&ps->ps_lock);
2378 "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2379 ps->ps_name, interval, trips);
2381 schedule_timeout_interruptible(interval);
2382 if (interval < cfs_time_seconds(1))
2388 if (ktime_get_seconds() < ps->ps_next_retry) {
2389 /* someone failed recently */
2390 spin_unlock(&ps->ps_lock);
2394 ps->ps_increasing = 1;
2395 spin_unlock(&ps->ps_lock);
2397 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2398 time_before = ktime_get();
2399 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2400 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete\n",
2401 ktime_ms_delta(ktime_get(), time_before));
2403 spin_lock(&ps->ps_lock);
2404 ps->ps_increasing = 0;
2406 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2408 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2409 CERROR("Can't allocate new %s pool because out of memory\n",
2412 spin_unlock(&ps->ps_lock);
2418 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2420 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2424 LASSERT (pool->po_allocated == 0);
2426 if (tpo->tpo_tx_pages != NULL) {
2427 kiblnd_unmap_tx_pool(tpo);
2428 kiblnd_free_pages(tpo->tpo_tx_pages);
2431 if (tpo->tpo_tx_descs == NULL)
2434 for (i = 0; i < pool->po_size; i++) {
2435 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2436 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2438 list_del(&tx->tx_list);
2439 if (tx->tx_pages != NULL)
2440 CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2441 if (tx->tx_frags != NULL)
2442 CFS_FREE_PTR_ARRAY(tx->tx_frags,
2443 IBLND_MAX_RDMA_FRAGS);
2444 if (tx->tx_wrq != NULL)
2445 CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2446 IBLND_MAX_RDMA_FRAGS);
2447 if (tx->tx_sge != NULL) {
2448 /* +1 is for the lnet header/message itself */
2449 CFS_FREE_PTR_ARRAY(tx->tx_sge,
2450 (IBLND_MAX_RDMA_FRAGS *
2453 if (tx->tx_rd != NULL)
2454 LIBCFS_FREE(tx->tx_rd,
2455 offsetof(struct kib_rdma_desc,
2456 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2459 CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2461 kiblnd_fini_pool(pool);
2465 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2467 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2470 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2471 ntx = tunables->lnd_ntx / ncpts;
2473 return max(IBLND_TX_POOL, ntx);
2477 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2481 struct kib_pool *pool;
2482 struct kib_tx_pool *tpo;
2484 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2486 CERROR("Failed to allocate TX pool\n");
2490 pool = &tpo->tpo_pool;
2491 kiblnd_init_pool(ps, pool, size);
2492 tpo->tpo_tx_descs = NULL;
2493 tpo->tpo_tx_pages = NULL;
2495 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2496 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2497 CERROR("Can't allocate tx pages: %d\n", npg);
2502 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2503 size * sizeof(struct kib_tx));
2504 if (tpo->tpo_tx_descs == NULL) {
2505 CERROR("Can't allocate %d tx descriptors\n", size);
2506 ps->ps_pool_destroy(pool);
2510 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2512 for (i = 0; i < size; i++) {
2513 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2514 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2517 if (ps->ps_net->ibn_fmr_ps != NULL) {
2518 LIBCFS_CPT_ALLOC(tx->tx_pages,
2519 lnet_cpt_table(), ps->ps_cpt,
2520 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2521 if (tx->tx_pages == NULL)
2525 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2526 IBLND_MAX_RDMA_FRAGS *
2527 sizeof(*tx->tx_frags));
2528 if (tx->tx_frags == NULL)
2531 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2533 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2534 IBLND_MAX_RDMA_FRAGS *
2535 sizeof(*tx->tx_wrq));
2536 if (tx->tx_wrq == NULL)
2539 /* +1 is for the lnet header/message itself */
2540 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2541 (IBLND_MAX_RDMA_FRAGS * wrq_sge + 1) *
2542 sizeof(*tx->tx_sge));
2543 if (tx->tx_sge == NULL)
2546 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2547 offsetof(struct kib_rdma_desc,
2548 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2549 if (tx->tx_rd == NULL)
2554 kiblnd_map_tx_pool(tpo);
2559 ps->ps_pool_destroy(pool);
2564 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2566 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2567 struct kib_tx_poolset,
2569 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2571 tx->tx_cookie = tps->tps_next_tx_cookie++;
2575 kiblnd_net_fini_pools(struct kib_net *net)
2579 cfs_cpt_for_each(i, lnet_cpt_table()) {
2580 struct kib_tx_poolset *tps;
2581 struct kib_fmr_poolset *fps;
2583 if (net->ibn_tx_ps != NULL) {
2584 tps = net->ibn_tx_ps[i];
2585 kiblnd_fini_poolset(&tps->tps_poolset);
2588 if (net->ibn_fmr_ps != NULL) {
2589 fps = net->ibn_fmr_ps[i];
2590 kiblnd_fini_fmr_poolset(fps);
2594 if (net->ibn_tx_ps != NULL) {
2595 cfs_percpt_free(net->ibn_tx_ps);
2596 net->ibn_tx_ps = NULL;
2599 if (net->ibn_fmr_ps != NULL) {
2600 cfs_percpt_free(net->ibn_fmr_ps);
2601 net->ibn_fmr_ps = NULL;
2606 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2609 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2610 #ifdef HAVE_OFED_IB_GET_DMA_MR
2611 unsigned long flags;
2617 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2619 #ifdef HAVE_OFED_IB_GET_DMA_MR
2620 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2622 * if lnd_map_on_demand is zero then we have effectively disabled
2623 * FMR or FastReg and we're using global memory regions
2626 if (!tunables->lnd_map_on_demand) {
2627 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2629 goto create_tx_pool;
2632 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2635 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2636 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2637 tunables->lnd_fmr_pool_size,
2638 tunables->lnd_ntx / 4);
2643 /* TX pool must be created later than FMR, see LU-2268
2645 LASSERT(net->ibn_tx_ps == NULL);
2647 /* premapping can fail if ibd_nmr > 1, so we always create
2648 * FMR pool and map-on-demand if premapping failed */
2650 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2651 sizeof(struct kib_fmr_poolset));
2652 if (net->ibn_fmr_ps == NULL) {
2653 CERROR("Failed to allocate FMR pool array\n");
2658 for (i = 0; i < ncpts; i++) {
2659 cpt = (cpts == NULL) ? i : cpts[i];
2660 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2663 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2670 LASSERT(i == ncpts);
2672 #ifdef HAVE_OFED_IB_GET_DMA_MR
2675 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2676 sizeof(struct kib_tx_poolset));
2677 if (net->ibn_tx_ps == NULL) {
2678 CERROR("Failed to allocate tx pool array\n");
2683 for (i = 0; i < ncpts; i++) {
2684 cpt = (cpts == NULL) ? i : cpts[i];
2685 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2687 kiblnd_tx_pool_size(ni, ncpts),
2688 kiblnd_create_tx_pool,
2689 kiblnd_destroy_tx_pool,
2690 kiblnd_tx_init, NULL);
2692 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2700 kiblnd_net_fini_pools(net);
2706 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2708 struct ib_port_attr *port_attr;
2710 unsigned long flags;
2711 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2713 LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2714 if (port_attr == NULL) {
2715 CDEBUG(D_NETERROR, "Out of memory\n");
2719 rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2721 write_lock_irqsave(g_lock, flags);
2724 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2725 ? IBLND_DEV_PORT_ACTIVE
2726 : IBLND_DEV_PORT_DOWN;
2728 write_unlock_irqrestore(g_lock, flags);
2729 LIBCFS_FREE(port_attr, sizeof(*port_attr));
2732 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2739 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2741 struct kib_net *net;
2742 __u32 ni_state_before;
2743 bool update_ping_buf = false;
2744 struct lnet_ni *ni = NULL;
2746 /* for health check */
2747 list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2750 CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2751 libcfs_nidstr(&ni->ni_nid));
2752 ni_state_before = lnet_set_link_fatal_state(ni, val);
2754 if (!update_ping_buf &&
2755 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
2756 (val != ni_state_before) &&
2757 (net->ibn_init == IBLND_INIT_ALL))
2758 update_ping_buf = true;
2761 if (update_ping_buf)
2762 lnet_mark_ping_buffer_for_update();
2766 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2768 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2769 struct kib_hca_dev *hdev;
2770 unsigned long flags;
2772 hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2774 write_lock_irqsave(g_lock, flags);
2776 switch (event->event) {
2777 case IB_EVENT_DEVICE_FATAL:
2778 CDEBUG(D_NET, "IB device fatal\n");
2779 hdev->ibh_state = IBLND_DEV_FATAL;
2780 kiblnd_set_ni_fatal_on(hdev, 1);
2782 case IB_EVENT_PORT_ACTIVE:
2783 CDEBUG(D_NET, "IB port active\n");
2784 if (event->element.port_num == hdev->ibh_port) {
2785 hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2786 kiblnd_set_ni_fatal_on(hdev, 0);
2789 case IB_EVENT_PORT_ERR:
2790 CDEBUG(D_NET, "IB port err\n");
2791 if (event->element.port_num == hdev->ibh_port) {
2792 hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2793 kiblnd_set_ni_fatal_on(hdev, 1);
2799 write_unlock_irqrestore(g_lock, flags);
2803 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2805 struct ib_device_attr *dev_attr;
2809 /* It's safe to assume a HCA can handle a page size
2810 * matching that of the native system */
2811 hdev->ibh_page_shift = PAGE_SHIFT;
2812 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2813 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2815 #ifndef HAVE_OFED_IB_DEVICE_ATTRS
2816 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2817 if (dev_attr == NULL) {
2818 CERROR("Out of memory\n");
2822 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2824 CERROR("Failed to query IB device: %d\n", rc);
2825 goto out_clean_attr;
2828 dev_attr = &hdev->ibh_ibdev->attrs;
2831 hdev->ibh_mr_size = dev_attr->max_mr_size;
2832 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2834 /* Setup device Memory Registration capabilities */
2835 #ifdef HAVE_OFED_FMR_POOL_API
2836 #ifdef HAVE_OFED_IB_DEVICE_OPS
2837 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2838 hdev->ibh_ibdev->ops.dealloc_fmr &&
2839 hdev->ibh_ibdev->ops.map_phys_fmr &&
2840 hdev->ibh_ibdev->ops.unmap_fmr) {
2842 if (hdev->ibh_ibdev->alloc_fmr &&
2843 hdev->ibh_ibdev->dealloc_fmr &&
2844 hdev->ibh_ibdev->map_phys_fmr &&
2845 hdev->ibh_ibdev->unmap_fmr) {
2847 LCONSOLE_INFO("Using FMR for registration\n");
2848 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2850 #endif /* HAVE_OFED_FMR_POOL_API */
2851 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2852 LCONSOLE_INFO("Using FastReg for registration\n");
2853 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2854 #ifndef HAVE_OFED_IB_ALLOC_FAST_REG_MR
2855 #ifdef IB_DEVICE_SG_GAPS_REG
2856 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2857 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2864 rc2 = kiblnd_port_get_attr(hdev);
2871 #ifndef HAVE_OFED_IB_DEVICE_ATTRS
2873 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2877 CERROR("IB device does not support FMRs nor FastRegs, can't "
2878 "register memory: %d\n", rc);
2879 else if (rc == -EINVAL)
2880 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2884 #ifdef HAVE_OFED_IB_GET_DMA_MR
2886 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2888 if (hdev->ibh_mrs == NULL)
2891 ib_dereg_mr(hdev->ibh_mrs);
2893 hdev->ibh_mrs = NULL;
2898 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2900 if (hdev->ibh_event_handler.device != NULL)
2901 ib_unregister_event_handler(&hdev->ibh_event_handler);
2903 #ifdef HAVE_OFED_IB_GET_DMA_MR
2904 kiblnd_hdev_cleanup_mrs(hdev);
2907 if (hdev->ibh_pd != NULL)
2908 ib_dealloc_pd(hdev->ibh_pd);
2910 if (hdev->ibh_cmid != NULL)
2911 rdma_destroy_id(hdev->ibh_cmid);
2913 LIBCFS_FREE(hdev, sizeof(*hdev));
2916 #ifdef HAVE_OFED_IB_GET_DMA_MR
2918 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2921 int acflags = IB_ACCESS_LOCAL_WRITE |
2922 IB_ACCESS_REMOTE_WRITE;
2924 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2926 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2927 kiblnd_hdev_cleanup_mrs(hdev);
2938 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2944 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2946 struct rdma_cm_id *cmid;
2947 struct sockaddr_in srcaddr;
2948 struct sockaddr_in dstaddr;
2951 if (dev->ibd_hdev == NULL || /* initializing */
2952 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2953 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2956 /* XXX: it's UGLY, but I don't have better way to find
2957 * ib-bonding HCA failover because:
2959 * a. no reliable CM event for HCA failover...
2960 * b. no OFED API to get ib_device for current net_device...
2962 * We have only two choices at this point:
2964 * a. rdma_bind_addr(), it will conflict with listener cmid
2965 * b. rdma_resolve_addr() to zero addr */
2966 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2967 RDMA_PS_TCP, IB_QPT_RC);
2970 CERROR("Failed to create cmid for failover: %d\n", rc);
2974 memset(&srcaddr, 0, sizeof(srcaddr));
2975 srcaddr.sin_family = AF_INET;
2976 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2978 memset(&dstaddr, 0, sizeof(dstaddr));
2979 dstaddr.sin_family = AF_INET;
2980 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2981 (struct sockaddr *)&dstaddr, 1);
2982 if (rc != 0 || cmid->device == NULL) {
2983 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2984 dev->ibd_ifname, &dev->ibd_ifip,
2986 rdma_destroy_id(cmid);
2990 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2991 rdma_destroy_id(cmid);
2996 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2998 LIST_HEAD(zombie_tpo);
2999 LIST_HEAD(zombie_ppo);
3000 LIST_HEAD(zombie_fpo);
3001 struct rdma_cm_id *cmid = NULL;
3002 struct kib_hca_dev *hdev = NULL;
3003 struct kib_hca_dev *old;
3005 struct kib_net *net;
3006 struct sockaddr_in addr;
3007 struct net_device *netdev;
3008 unsigned long flags;
3011 bool set_fatal = true;
3013 LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
3014 dev->ibd_can_failover ||
3015 dev->ibd_hdev == NULL);
3017 rc = kiblnd_dev_need_failover(dev, ns);
3021 if (dev->ibd_hdev != NULL &&
3022 dev->ibd_hdev->ibh_cmid != NULL) {
3023 /* XXX it's not good to close old listener at here,
3024 * because we can fail to create new listener.
3025 * But we have to close it now, otherwise rdma_bind_addr
3026 * will return EADDRINUSE... How crap! */
3027 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3029 cmid = dev->ibd_hdev->ibh_cmid;
3030 /* make next schedule of kiblnd_dev_need_failover()
3031 * return 1 for me */
3032 dev->ibd_hdev->ibh_cmid = NULL;
3033 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3035 rdma_destroy_id(cmid);
3038 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
3042 CERROR("Failed to create cmid for failover: %d\n", rc);
3046 memset(&addr, 0, sizeof(addr));
3047 addr.sin_family = AF_INET;
3048 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
3049 addr.sin_port = htons(*kiblnd_tunables.kib_service);
3051 /* Bind to failover device or port */
3052 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
3053 if (rc != 0 || cmid->device == NULL) {
3054 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
3055 dev->ibd_ifname, &dev->ibd_ifip,
3057 if (!rc && !cmid->device)
3059 rdma_destroy_id(cmid);
3063 LIBCFS_ALLOC(hdev, sizeof(*hdev));
3065 CERROR("Failed to allocate kib_hca_dev\n");
3066 rdma_destroy_id(cmid);
3071 atomic_set(&hdev->ibh_ref, 1);
3072 hdev->ibh_dev = dev;
3073 hdev->ibh_cmid = cmid;
3074 hdev->ibh_ibdev = cmid->device;
3075 hdev->ibh_port = cmid->port_num;
3077 #ifdef HAVE_OFED_IB_ALLOC_PD_2ARGS
3078 pd = ib_alloc_pd(cmid->device, 0);
3080 pd = ib_alloc_pd(cmid->device);
3084 CERROR("Can't allocate PD: %d\n", rc);
3090 rc = rdma_listen(cmid, 0);
3092 CERROR("Can't start new listener: %d\n", rc);
3096 rc = kiblnd_hdev_get_attr(hdev);
3098 CERROR("Can't get device attributes: %d\n", rc);
3102 #ifdef HAVE_OFED_IB_GET_DMA_MR
3103 rc = kiblnd_hdev_setup_mrs(hdev);
3105 CERROR("Can't setup device: %d\n", rc);
3110 INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
3111 hdev->ibh_ibdev, kiblnd_event_handler);
3112 ib_register_event_handler(&hdev->ibh_event_handler);
3114 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3116 old = dev->ibd_hdev;
3117 dev->ibd_hdev = hdev; /* take over the refcount */
3120 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
3121 cfs_cpt_for_each(i, lnet_cpt_table()) {
3122 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
3125 if (net->ibn_fmr_ps != NULL)
3126 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
3131 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3133 if (!list_empty(&zombie_tpo))
3134 kiblnd_destroy_pool_list(&zombie_tpo);
3135 if (!list_empty(&zombie_ppo))
3136 kiblnd_destroy_pool_list(&zombie_ppo);
3137 if (!list_empty(&zombie_fpo))
3138 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
3140 kiblnd_hdev_decref(hdev);
3143 dev->ibd_failed_failover++;
3145 dev->ibd_failed_failover = 0;
3149 netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
3150 if (netdev && (lnet_get_link_status(netdev) == 1))
3151 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
3160 kiblnd_destroy_dev(struct kib_dev *dev)
3162 LASSERT(dev->ibd_nnets == 0);
3163 LASSERT(list_empty(&dev->ibd_nets));
3165 list_del(&dev->ibd_fail_list);
3166 list_del(&dev->ibd_list);
3168 if (dev->ibd_hdev != NULL)
3169 kiblnd_hdev_decref(dev->ibd_hdev);
3171 LIBCFS_FREE(dev, sizeof(*dev));
3174 static struct kib_dev *
3175 kiblnd_dev_search(char *ifname)
3177 struct kib_dev *alias = NULL;
3178 struct kib_dev *dev;
3182 colon = strchr(ifname, ':');
3183 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3184 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3190 colon2 = strchr(dev->ibd_ifname, ':');
3196 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3208 kiblnd_handle_link_state_change(struct net_device *dev,
3209 unsigned char operstate)
3211 struct lnet_ni *ni = NULL;
3212 struct kib_dev *event_kibdev;
3213 struct kib_net *net;
3214 struct kib_net *cnxt;
3215 bool link_down = !(operstate == IF_OPER_UP);
3216 struct in_device *in_dev;
3217 bool found_ip = false;
3218 __u32 ni_state_before;
3219 bool update_ping_buf = false;
3221 DECLARE_CONST_IN_IFADDR(ifa);
3223 event_kibdev = kiblnd_dev_search(dev->name);
3228 list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets, ibn_list) {
3232 in_dev = __in_dev_get_rtnl(dev);
3234 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
3236 ni_state_before = lnet_set_link_fatal_state(ni, 1);
3239 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
3240 if (htonl(event_kibdev->ibd_ifip) == ifa->ifa_local)
3246 CDEBUG(D_NET, "Interface %s has no matching ip\n",
3248 ni_state_before = lnet_set_link_fatal_state(ni, 1);
3253 ni_state_before = lnet_set_link_fatal_state(ni, 1);
3255 state = (lnet_get_link_status(dev) == 0);
3256 ni_state_before = lnet_set_link_fatal_state(ni,
3260 if (!update_ping_buf &&
3261 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
3262 (atomic_read(&ni->ni_fatal_error_on) != ni_state_before) &&
3263 (net->ibn_init == IBLND_INIT_ALL))
3264 update_ping_buf = true;
3267 if (update_ping_buf)
3268 lnet_mark_ping_buffer_for_update();
3274 kiblnd_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
3276 struct kib_dev *event_kibdev;
3277 struct kib_net *net;
3278 struct kib_net *cnxt;
3279 struct net_device *event_netdev = ifa->ifa_dev->dev;
3280 __u32 ni_state_before;
3281 bool update_ping_buf = false;
3282 struct lnet_ni *ni = NULL;
3285 event_kibdev = kiblnd_dev_search(event_netdev->name);
3290 if (htonl(event_kibdev->ibd_ifip) != ifa->ifa_local)
3293 list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets,
3296 link_down = (event == NETDEV_DOWN);
3297 ni_state_before = lnet_set_link_fatal_state(ni, link_down);
3298 if (!update_ping_buf &&
3299 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
3300 ((event == NETDEV_DOWN) != ni_state_before) &&
3301 (net->ibn_init == IBLND_INIT_ALL))
3302 update_ping_buf = true;
3305 if (update_ping_buf)
3306 lnet_mark_ping_buffer_for_update();
3312 /************************************
3313 * Net device notifier event handler
3314 ************************************/
3315 static int kiblnd_device_event(struct notifier_block *unused,
3316 unsigned long event, void *ptr)
3318 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
3319 unsigned char operstate;
3321 operstate = dev->operstate;
3323 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
3324 event, dev->name, dev->ifindex, operstate);
3330 kiblnd_handle_link_state_change(dev, operstate);
3337 /************************************
3338 * Inetaddr notifier event handler
3339 ************************************/
3340 static int kiblnd_inetaddr_event(struct notifier_block *unused,
3341 unsigned long event, void *ptr)
3343 struct in_ifaddr *ifa = ptr;
3345 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
3346 event, &ifa->ifa_address, &ifa->ifa_mask);
3352 kiblnd_handle_inetaddr_change(ifa, event);
3359 static struct notifier_block kiblnd_dev_notifier_block = {
3360 .notifier_call = kiblnd_device_event,
3363 static struct notifier_block kiblnd_inetaddr_notifier_block = {
3364 .notifier_call = kiblnd_inetaddr_event,
3368 kiblnd_base_shutdown(void)
3370 struct kib_sched_info *sched;
3371 struct kib_peer_ni *peer_ni;
3374 LASSERT(list_empty(&kiblnd_data.kib_devs));
3376 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
3377 libcfs_kmem_read());
3379 if (kiblnd_data.kib_init == IBLND_INIT_ALL) {
3380 unregister_netdevice_notifier(&kiblnd_dev_notifier_block);
3381 unregister_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3384 switch (kiblnd_data.kib_init) {
3388 case IBLND_INIT_ALL:
3389 case IBLND_INIT_DATA:
3390 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
3392 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
3393 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
3394 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
3395 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
3397 /* flag threads to terminate; wake and wait for them to die */
3398 kiblnd_data.kib_shutdown = 1;
3400 /* NB: we really want to stop scheduler threads net by net
3401 * instead of the whole module, this should be improved
3402 * with dynamic configuration LNet.
3404 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
3405 wake_up_all(&sched->ibs_waitq);
3407 wake_up(&kiblnd_data.kib_connd_waitq);
3408 wake_up(&kiblnd_data.kib_failover_waitq);
3410 wait_var_event_warning(&kiblnd_data.kib_nthreads,
3411 !atomic_read(&kiblnd_data.kib_nthreads),
3412 "Waiting for %d threads to terminate\n",
3413 atomic_read(&kiblnd_data.kib_nthreads));
3416 case IBLND_INIT_NOTHING:
3420 if (kiblnd_data.kib_scheds != NULL)
3421 cfs_percpt_free(kiblnd_data.kib_scheds);
3423 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3424 libcfs_kmem_read());
3426 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3427 module_put(THIS_MODULE);
3431 kiblnd_shutdown(struct lnet_ni *ni)
3433 struct kib_net *net = ni->ni_data;
3434 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
3435 unsigned long flags;
3437 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3442 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3443 libcfs_kmem_read());
3445 write_lock_irqsave(g_lock, flags);
3446 net->ibn_shutdown = 1;
3447 write_unlock_irqrestore(g_lock, flags);
3449 switch (net->ibn_init) {
3453 case IBLND_INIT_ALL:
3454 /* nuke all existing peers within this net */
3455 kiblnd_del_peer(ni, LNET_NID_ANY);
3457 /* Wait for all peer_ni state to clean up */
3458 wait_var_event_warning(&net->ibn_npeers,
3459 atomic_read(&net->ibn_npeers) == 0,
3460 "%s: waiting for %d peers to disconnect\n",
3461 libcfs_nidstr(&ni->ni_nid),
3462 atomic_read(&net->ibn_npeers));
3464 kiblnd_net_fini_pools(net);
3466 write_lock_irqsave(g_lock, flags);
3467 LASSERT(net->ibn_dev->ibd_nnets > 0);
3468 net->ibn_dev->ibd_nnets--;
3469 list_del(&net->ibn_list);
3470 write_unlock_irqrestore(g_lock, flags);
3472 wake_up_all(&kiblnd_data.kib_connd_waitq);
3473 wait_var_event_warning(&net->ibn_nconns,
3474 atomic_read(&net->ibn_nconns) == 0,
3475 "%s: waiting for %d conns to clean\n",
3476 libcfs_nidstr(&ni->ni_nid),
3477 atomic_read(&net->ibn_nconns));
3480 case IBLND_INIT_NOTHING:
3481 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3483 if (net->ibn_dev != NULL &&
3484 net->ibn_dev->ibd_nnets == 0)
3485 kiblnd_destroy_dev(net->ibn_dev);
3490 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3491 libcfs_kmem_read());
3493 net->ibn_init = IBLND_INIT_NOTHING;
3496 LIBCFS_FREE(net, sizeof(*net));
3499 if (list_empty(&kiblnd_data.kib_devs))
3500 kiblnd_base_shutdown();
3504 kiblnd_base_startup(struct net *ns)
3506 struct kib_sched_info *sched;
3510 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3512 if (!try_module_get(THIS_MODULE))
3515 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3517 rwlock_init(&kiblnd_data.kib_global_lock);
3519 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3520 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3522 hash_init(kiblnd_data.kib_peers);
3524 spin_lock_init(&kiblnd_data.kib_connd_lock);
3525 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3526 INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3527 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3528 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3529 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3531 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3532 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3534 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3536 if (kiblnd_data.kib_scheds == NULL)
3539 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3542 spin_lock_init(&sched->ibs_lock);
3543 INIT_LIST_HEAD(&sched->ibs_conns);
3544 init_waitqueue_head(&sched->ibs_waitq);
3546 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3547 if (*kiblnd_tunables.kib_nscheds > 0) {
3548 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3550 /* max to half of CPUs, another half is reserved for
3551 * upper layer modules */
3552 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3555 sched->ibs_nthreads_max = nthrs;
3559 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3561 /* lists/ptrs/locks initialised */
3562 kiblnd_data.kib_init = IBLND_INIT_DATA;
3563 /*****************************************************/
3565 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3567 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3571 if (*kiblnd_tunables.kib_dev_failover != 0)
3572 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3576 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3580 register_netdevice_notifier(&kiblnd_dev_notifier_block);
3581 register_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3583 /* flag everything initialised */
3584 kiblnd_data.kib_init = IBLND_INIT_ALL;
3585 /*****************************************************/
3590 kiblnd_base_shutdown();
3595 kiblnd_start_schedulers(struct kib_sched_info *sched)
3601 if (sched->ibs_nthreads == 0) {
3602 if (*kiblnd_tunables.kib_nscheds > 0) {
3603 nthrs = sched->ibs_nthreads_max;
3605 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3607 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3608 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3611 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3612 /* increase one thread if there is new interface */
3613 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3616 for (i = 0; i < nthrs; i++) {
3617 long id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3619 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3620 "kiblnd_sd_%02ld_%02ld",
3621 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3625 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3626 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3630 sched->ibs_nthreads += i;
3634 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3641 for (i = 0; i < ncpts; i++) {
3642 struct kib_sched_info *sched;
3644 cpt = (cpts == NULL) ? i : cpts[i];
3645 sched = kiblnd_data.kib_scheds[cpt];
3647 if (!newdev && sched->ibs_nthreads > 0)
3650 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3652 CERROR("Failed to start scheduler threads for %s\n",
3661 kiblnd_startup(struct lnet_ni *ni)
3663 char *ifname = NULL;
3664 struct lnet_inetdev *ifaces = NULL;
3665 struct kib_dev *ibdev = NULL;
3666 struct kib_net *net = NULL;
3667 unsigned long flags;
3671 struct net_device *netdev;
3673 LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3675 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3676 rc = kiblnd_base_startup(ni->ni_net_ns);
3681 LIBCFS_ALLOC(net, sizeof(*net));
3689 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3691 kiblnd_tunables_setup(ni);
3693 /* Multi-Rail wants each secondary
3694 * IP to be treated as an unique 'struct ni' interface.
3696 if (ni->ni_interface != NULL) {
3697 /* Use the IPoIB interface specified in 'networks=' */
3698 ifname = ni->ni_interface;
3700 ifname = *kiblnd_tunables.kib_default_ipif;
3701 rc = libcfs_strnid(&ni->ni_nid, ifname);
3702 if (rc < 0 || ni->ni_nid.nid_type != O2IBLND)
3703 memset(&ni->ni_nid, 0, sizeof(ni->ni_nid));
3706 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3707 CERROR("IPoIB interface name too long: %s\n", ifname);
3712 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, false);
3716 i = lnet_inet_select(ni, ifaces, rc);
3720 if (nid_addr_is_set(&ni->ni_nid)) {
3721 strscpy(ifname, ifaces[i].li_name, sizeof(ifname));
3722 } else if (strcmp(ifname, ifaces[i].li_name) != 0) {
3723 CERROR("ko2iblnd: No matching interfaces\n");
3728 ibdev = kiblnd_dev_search(ifname);
3729 newdev = ibdev == NULL;
3730 /* hmm...create kib_dev even for alias */
3731 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3732 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3738 ibdev->ibd_ifip = ntohl(ifaces[i].li_ipaddr);
3739 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3740 sizeof(ibdev->ibd_ifname));
3741 ibdev->ibd_can_failover = ifaces[i].li_iff_master;
3743 INIT_LIST_HEAD(&ibdev->ibd_nets);
3744 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3745 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3747 /* initialize the device */
3748 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3750 CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3755 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3758 net->ibn_dev = ibdev;
3759 ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3760 if (!ni->ni_interface) {
3761 rc = lnet_ni_add_interface(ni, ifaces[i].li_name);
3763 CWARN("ko2iblnd failed to allocate ni_interface\n");
3765 ni->ni_dev_cpt = ifaces[i].li_cpt;
3767 rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3771 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3773 CERROR("Failed to initialize NI pools: %d\n", rc);
3777 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3779 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3780 /* for health check */
3781 if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3782 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3785 netdev = dev_get_by_name_rcu(ni->ni_net_ns, net->ibn_dev->ibd_ifname);
3787 ((netdev->reg_state == NETREG_UNREGISTERING) ||
3788 (netdev->operstate != IF_OPER_UP) ||
3789 (lnet_get_link_status(netdev) == 0))) {
3790 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3794 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3796 net->ibn_init = IBLND_INIT_ALL;
3802 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3803 kiblnd_destroy_dev(ibdev);
3806 kiblnd_shutdown(ni);
3808 CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3809 ifname ? ifname : "", rc);
3814 static const struct lnet_lnd the_o2iblnd = {
3815 .lnd_type = O2IBLND,
3816 .lnd_startup = kiblnd_startup,
3817 .lnd_shutdown = kiblnd_shutdown,
3818 .lnd_ctl = kiblnd_ctl,
3819 .lnd_send = kiblnd_send,
3820 .lnd_recv = kiblnd_recv,
3821 .lnd_get_dev_prio = kiblnd_get_dev_prio,
3822 .lnd_nl_get = kiblnd_nl_get,
3823 .lnd_nl_set = kiblnd_nl_set,
3824 .lnd_keys = &kiblnd_tunables_keys,
3827 static void ko2inlnd_assert_wire_constants(void)
3829 BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3830 BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3831 BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3832 BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3834 BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3835 BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3836 BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3837 BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3838 BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3839 BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3840 BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3841 BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3842 BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3843 BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3845 BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3846 BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3847 BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3848 BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3849 BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3850 BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3851 BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3852 BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3854 BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3855 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3856 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3857 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3858 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3859 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3860 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3862 BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3863 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3864 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3865 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3866 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3868 BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3869 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3870 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3871 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3872 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3874 BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3875 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3876 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3877 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3878 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3879 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3880 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3882 BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3883 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3884 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3885 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3886 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3888 BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3889 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3890 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3891 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3892 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3893 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3894 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3896 BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3897 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3898 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3899 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3900 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3901 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3902 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3904 BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3905 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3906 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3907 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3908 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3910 /* Checks for struct kib_msg */
3911 //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3912 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3913 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3914 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3915 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3916 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3917 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3918 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3919 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3920 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3921 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3922 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3923 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3924 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3925 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3926 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3927 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3928 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3929 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3930 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3931 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3934 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3935 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3936 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3937 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3938 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3939 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3941 /* Immediate message */
3942 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3943 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3944 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3945 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3947 /* PUT req message */
3948 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3949 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3950 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3951 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3954 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3955 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3956 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3957 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3958 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3959 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3962 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3963 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3964 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3965 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3966 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3967 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3969 /* Completion message */
3970 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3971 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3972 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3973 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3976 BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3977 BUILD_BUG_ON(offsetof(struct kib_msg,
3978 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3980 BUILD_BUG_ON(offsetof(struct kib_msg,
3981 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3985 static void __exit ko2iblnd_exit(void)
3987 lnet_unregister_lnd(&the_o2iblnd);
3990 static int __init ko2iblnd_init(void)
3994 ko2inlnd_assert_wire_constants();
3996 rc = kiblnd_tunables_init();
4000 rc = libcfs_setup();
4004 lnet_register_lnd(&the_o2iblnd);
4009 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
4010 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
4011 MODULE_VERSION("2.8.0");
4012 MODULE_LICENSE("GPL");
4014 module_init(ko2iblnd_init);
4015 module_exit(ko2iblnd_exit);