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_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_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
1258 kiblnd_nl_set(int cmd, struct nlattr *attr, int type, void *data)
1260 struct lnet_lnd_tunables *tunables = data;
1263 if (cmd != LNET_CMD_NETS)
1266 if (nla_type(attr) != LN_SCALAR_ATTR_INT_VALUE)
1270 case LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS:
1271 tunables->lnd_tun_u.lnd_o2ib.lnd_peercredits_hiw = nla_get_s64(attr);
1273 case LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND:
1274 tunables->lnd_tun_u.lnd_o2ib.lnd_map_on_demand = nla_get_s64(attr);
1276 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS:
1277 tunables->lnd_tun_u.lnd_o2ib.lnd_concurrent_sends = nla_get_s64(attr);
1279 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE:
1280 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_pool_size = nla_get_s64(attr);
1282 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER:
1283 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_flush_trigger = nla_get_s64(attr);
1285 case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE:
1286 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_cache = nla_get_s64(attr);
1288 case LNET_NET_O2IBLND_TUNABLES_ATTR_NTX:
1289 tunables->lnd_tun_u.lnd_o2ib.lnd_ntx = nla_get_s64(attr);
1291 case LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER:
1292 num = nla_get_s64(attr);
1293 clamp_t(s64, num, 1, 127);
1294 tunables->lnd_tun_u.lnd_o2ib.lnd_conns_per_peer = num;
1304 kiblnd_free_pages(struct kib_pages *p)
1306 int npages = p->ibp_npages;
1309 for (i = 0; i < npages; i++) {
1310 if (p->ibp_pages[i] != NULL)
1311 __free_page(p->ibp_pages[i]);
1314 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1318 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1320 struct kib_pages *p;
1323 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1324 offsetof(struct kib_pages, ibp_pages[npages]));
1326 CERROR("Can't allocate descriptor for %d pages\n", npages);
1330 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1331 p->ibp_npages = npages;
1333 for (i = 0; i < npages; i++) {
1334 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1336 if (p->ibp_pages[i] == NULL) {
1337 CERROR("Can't allocate page %d of %d\n", i, npages);
1338 kiblnd_free_pages(p);
1348 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1353 LASSERT (conn->ibc_rxs != NULL);
1354 LASSERT (conn->ibc_hdev != NULL);
1356 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1357 rx = &conn->ibc_rxs[i];
1359 LASSERT(rx->rx_nob >= 0); /* not posted */
1361 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1362 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1364 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1367 kiblnd_free_pages(conn->ibc_rx_pages);
1369 conn->ibc_rx_pages = NULL;
1373 kiblnd_map_rx_descs(struct kib_conn *conn)
1381 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1382 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1383 rx = &conn->ibc_rxs[i];
1386 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1389 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1390 rx->rx_msg, IBLND_MSG_SIZE,
1392 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1394 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1396 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1397 i, rx->rx_msg, rx->rx_msgaddr,
1398 (__u64)(page_to_phys(pg) + pg_off));
1400 pg_off += IBLND_MSG_SIZE;
1401 LASSERT(pg_off <= PAGE_SIZE);
1403 if (pg_off == PAGE_SIZE) {
1406 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1412 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1414 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1418 LASSERT (tpo->tpo_pool.po_allocated == 0);
1423 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1424 tx = &tpo->tpo_tx_descs[i];
1425 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1426 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1428 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1431 kiblnd_hdev_decref(hdev);
1432 tpo->tpo_hdev = NULL;
1435 static struct kib_hca_dev *
1436 kiblnd_current_hdev(struct kib_dev *dev)
1438 struct kib_hca_dev *hdev;
1439 unsigned long flags;
1442 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1443 while (dev->ibd_failover) {
1444 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1446 CDEBUG(D_NET, "%s: Wait for failover\n",
1448 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1450 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1453 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1454 hdev = dev->ibd_hdev;
1456 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1462 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1464 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1465 struct kib_pool *pool = &tpo->tpo_pool;
1466 struct kib_net *net = pool->po_owner->ps_net;
1467 struct kib_dev *dev;
1474 LASSERT (net != NULL);
1478 /* pre-mapped messages are not bigger than 1 page */
1479 BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1481 /* No fancy arithmetic when we do the buffer calculations */
1482 BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1484 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1486 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1487 page = txpgs->ibp_pages[ipage];
1488 tx = &tpo->tpo_tx_descs[i];
1490 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1493 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1497 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1499 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1501 list_add(&tx->tx_list, &pool->po_free_list);
1503 page_offset += IBLND_MSG_SIZE;
1504 LASSERT(page_offset <= PAGE_SIZE);
1506 if (page_offset == PAGE_SIZE) {
1509 LASSERT(ipage <= txpgs->ibp_npages);
1515 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1517 LASSERT(fpo->fpo_map_count == 0);
1519 #ifdef HAVE_FMR_POOL_API
1520 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1521 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1523 #endif /* HAVE_FMR_POOL_API */
1525 struct kib_fast_reg_descriptor *frd, *tmp;
1528 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1530 list_del(&frd->frd_list);
1531 #ifndef HAVE_IB_MAP_MR_SG
1532 ib_free_fast_reg_page_list(frd->frd_frpl);
1534 ib_dereg_mr(frd->frd_mr);
1535 LIBCFS_FREE(frd, sizeof(*frd));
1538 if (i < fpo->fast_reg.fpo_pool_size)
1539 CERROR("FastReg pool still has %d regions registered\n",
1540 fpo->fast_reg.fpo_pool_size - i);
1544 kiblnd_hdev_decref(fpo->fpo_hdev);
1546 LIBCFS_FREE(fpo, sizeof(*fpo));
1550 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1552 struct kib_fmr_pool *fpo, *tmp;
1554 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1555 list_del(&fpo->fpo_list);
1556 kiblnd_destroy_fmr_pool(fpo);
1561 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1564 int size = tunables->lnd_fmr_pool_size / ncpts;
1566 return max(IBLND_FMR_POOL, size);
1570 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1573 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1575 return max(IBLND_FMR_POOL_FLUSH, size);
1578 #ifdef HAVE_FMR_POOL_API
1579 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1580 struct kib_fmr_pool *fpo)
1582 struct ib_fmr_pool_param param = {
1583 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1584 .page_shift = PAGE_SHIFT,
1585 .access = (IB_ACCESS_LOCAL_WRITE |
1586 IB_ACCESS_REMOTE_WRITE),
1587 .pool_size = fps->fps_pool_size,
1588 .dirty_watermark = fps->fps_flush_trigger,
1589 .flush_function = NULL,
1591 .cache = !!fps->fps_cache };
1594 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1596 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1597 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1599 CERROR("Failed to create FMR pool: %d\n", rc);
1601 CERROR("FMRs are not supported\n");
1603 fpo->fpo_is_fmr = true;
1607 #endif /* HAVE_FMR_POOL_API */
1609 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1610 struct kib_fmr_pool *fpo,
1611 enum kib_dev_caps dev_caps)
1613 struct kib_fast_reg_descriptor *frd, *tmp;
1616 #ifdef HAVE_FMR_POOL_API
1617 fpo->fpo_is_fmr = false;
1620 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1621 fpo->fast_reg.fpo_pool_size = 0;
1622 for (i = 0; i < fps->fps_pool_size; i++) {
1623 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1626 CERROR("Failed to allocate a new fast_reg descriptor\n");
1632 #ifndef HAVE_IB_MAP_MR_SG
1633 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1634 IBLND_MAX_RDMA_FRAGS);
1635 if (IS_ERR(frd->frd_frpl)) {
1636 rc = PTR_ERR(frd->frd_frpl);
1637 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1639 frd->frd_frpl = NULL;
1644 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1645 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1646 IBLND_MAX_RDMA_FRAGS);
1649 * it is expected to get here if this is an MLX-5 card.
1650 * MLX-4 cards will always use FMR and MLX-5 cards will
1651 * always use fast_reg. It turns out that some MLX-5 cards
1652 * (possibly due to older FW versions) do not natively support
1653 * gaps. So we will need to track them here.
1655 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1656 #ifdef IB_MR_TYPE_SG_GAPS
1657 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1658 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1659 IB_MR_TYPE_SG_GAPS :
1664 IBLND_MAX_RDMA_FRAGS);
1665 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1666 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1667 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1669 if (IS_ERR(frd->frd_mr)) {
1670 rc = PTR_ERR(frd->frd_mr);
1671 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1676 /* indicate that the local invalidate needs to be generated */
1677 frd->frd_valid = false;
1679 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1680 fpo->fast_reg.fpo_pool_size++;
1687 ib_dereg_mr(frd->frd_mr);
1688 #ifndef HAVE_IB_MAP_MR_SG
1690 ib_free_fast_reg_page_list(frd->frd_frpl);
1692 LIBCFS_FREE(frd, sizeof(*frd));
1695 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1697 list_del(&frd->frd_list);
1698 #ifndef HAVE_IB_MAP_MR_SG
1699 ib_free_fast_reg_page_list(frd->frd_frpl);
1701 ib_dereg_mr(frd->frd_mr);
1702 LIBCFS_FREE(frd, sizeof(*frd));
1708 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1709 struct kib_fmr_pool **pp_fpo)
1711 struct kib_dev *dev = fps->fps_net->ibn_dev;
1712 struct kib_fmr_pool *fpo;
1715 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1719 memset(fpo, 0, sizeof(*fpo));
1721 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1723 #ifdef HAVE_FMR_POOL_API
1724 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1725 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1727 #endif /* HAVE_FMR_POOL_API */
1728 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1732 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1733 fpo->fpo_owner = fps;
1739 kiblnd_hdev_decref(fpo->fpo_hdev);
1740 LIBCFS_FREE(fpo, sizeof(*fpo));
1745 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1747 struct kib_fmr_pool *fpo;
1749 if (fps->fps_net == NULL) /* intialized? */
1752 spin_lock(&fps->fps_lock);
1754 while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1755 struct kib_fmr_pool,
1756 fpo_list)) != NULL) {
1757 fpo->fpo_failed = 1;
1758 if (fpo->fpo_map_count == 0)
1759 list_move(&fpo->fpo_list, zombies);
1761 list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1764 spin_unlock(&fps->fps_lock);
1768 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1770 if (fps->fps_net != NULL) { /* initialized? */
1771 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1772 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1777 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1778 struct kib_net *net,
1779 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1781 struct kib_fmr_pool *fpo;
1784 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1789 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1790 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1791 fps->fps_cache = tunables->lnd_fmr_cache;
1793 spin_lock_init(&fps->fps_lock);
1794 INIT_LIST_HEAD(&fps->fps_pool_list);
1795 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1797 rc = kiblnd_create_fmr_pool(fps, &fpo);
1799 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1805 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1807 if (fpo->fpo_map_count != 0) /* still in use */
1809 if (fpo->fpo_failed)
1811 return now >= fpo->fpo_deadline;
1814 #if defined(HAVE_FMR_POOL_API) || !defined(HAVE_IB_MAP_MR_SG)
1816 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1818 struct kib_hca_dev *hdev;
1819 __u64 *pages = tx->tx_pages;
1824 hdev = tx->tx_pool->tpo_hdev;
1826 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1827 for (size = 0; size < rd->rd_frags[i].rf_nob;
1828 size += hdev->ibh_page_size) {
1829 pages[npages++] = (rd->rd_frags[i].rf_addr &
1830 hdev->ibh_page_mask) + size;
1839 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1842 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1843 struct kib_fmr_poolset *fps;
1844 time64_t now = ktime_get_seconds();
1845 struct kib_fmr_pool *tmp;
1850 fps = fpo->fpo_owner;
1852 #ifdef HAVE_FMR_POOL_API
1853 if (fpo->fpo_is_fmr) {
1854 if (fmr->fmr_pfmr) {
1855 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1856 fmr->fmr_pfmr = NULL;
1860 int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1864 #endif /* HAVE_FMR_POOL_API */
1866 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1868 frd->frd_posted = false;
1869 fmr->fmr_frd = NULL;
1870 spin_lock(&fps->fps_lock);
1871 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1872 spin_unlock(&fps->fps_lock);
1875 fmr->fmr_pool = NULL;
1877 spin_lock(&fps->fps_lock);
1878 fpo->fpo_map_count--; /* decref the pool */
1880 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1881 /* the first pool is persistent */
1882 if (fps->fps_pool_list.next == &fpo->fpo_list)
1885 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1886 list_move(&fpo->fpo_list, &zombies);
1890 spin_unlock(&fps->fps_lock);
1892 if (!list_empty(&zombies))
1893 kiblnd_destroy_fmr_pool_list(&zombies);
1896 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1897 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1898 struct kib_fmr *fmr)
1900 struct kib_fmr_pool *fpo;
1902 bool is_rx = (rd != tx->tx_rd);
1903 #ifdef HAVE_FMR_POOL_API
1904 __u64 *pages = tx->tx_pages;
1905 bool tx_pages_mapped = false;
1911 spin_lock(&fps->fps_lock);
1912 version = fps->fps_version;
1913 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1914 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1915 fpo->fpo_map_count++;
1917 #ifdef HAVE_FMR_POOL_API
1918 fmr->fmr_pfmr = NULL;
1919 if (fpo->fpo_is_fmr) {
1920 struct ib_pool_fmr *pfmr;
1922 spin_unlock(&fps->fps_lock);
1924 if (!tx_pages_mapped) {
1925 npages = kiblnd_map_tx_pages(tx, rd);
1926 tx_pages_mapped = true;
1929 pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1930 pages, npages, iov);
1931 if (likely(!IS_ERR(pfmr))) {
1932 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1934 fmr->fmr_frd = NULL;
1935 fmr->fmr_pfmr = pfmr;
1936 fmr->fmr_pool = fpo;
1941 #endif /* HAVE_FMR_POOL_API */
1943 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1944 struct kib_fast_reg_descriptor *frd;
1945 #ifdef HAVE_IB_MAP_MR_SG
1946 struct ib_reg_wr *wr;
1949 struct ib_rdma_wr *wr;
1950 struct ib_fast_reg_page_list *frpl;
1954 frd = list_first_entry(
1955 &fpo->fast_reg.fpo_pool_list,
1956 struct kib_fast_reg_descriptor,
1958 list_del(&frd->frd_list);
1959 spin_unlock(&fps->fps_lock);
1961 #ifndef HAVE_IB_MAP_MR_SG
1962 frpl = frd->frd_frpl;
1966 if (!frd->frd_valid) {
1967 struct ib_rdma_wr *inv_wr;
1968 __u32 key = is_rx ? mr->rkey : mr->lkey;
1970 frd->frd_valid = true;
1971 inv_wr = &frd->frd_inv_wr;
1972 memset(inv_wr, 0, sizeof(*inv_wr));
1974 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1975 inv_wr->wr.wr_id = IBLND_WID_MR;
1976 inv_wr->wr.ex.invalidate_rkey = key;
1979 key = ib_inc_rkey(key);
1980 ib_update_fast_reg_key(mr, key);
1983 #ifdef HAVE_IB_MAP_MR_SG
1984 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1985 n = ib_map_mr_sg(mr, tx->tx_frags,
1986 rd->rd_nfrags, NULL, PAGE_SIZE);
1988 n = ib_map_mr_sg(mr, tx->tx_frags,
1989 rd->rd_nfrags, PAGE_SIZE);
1990 #endif /* HAVE_IB_MAP_MR_SG_5ARGS */
1991 if (unlikely(n != rd->rd_nfrags)) {
1992 CERROR("Failed to map mr %d/%d elements\n",
1994 return n < 0 ? n : -EINVAL;
1997 wr = &frd->frd_fastreg_wr;
1998 memset(wr, 0, sizeof(*wr));
2000 wr->wr.opcode = IB_WR_REG_MR;
2001 wr->wr.wr_id = IBLND_WID_MR;
2003 wr->wr.send_flags = 0;
2005 wr->key = is_rx ? mr->rkey : mr->lkey;
2006 wr->access = (IB_ACCESS_LOCAL_WRITE |
2007 IB_ACCESS_REMOTE_WRITE);
2008 #else /* HAVE_IB_MAP_MR_SG */
2009 if (!tx_pages_mapped) {
2010 npages = kiblnd_map_tx_pages(tx, rd);
2011 tx_pages_mapped = true;
2014 LASSERT(npages <= frpl->max_page_list_len);
2015 memcpy(frpl->page_list, pages,
2016 sizeof(*pages) * npages);
2018 /* Prepare FastReg WR */
2019 wr = &frd->frd_fastreg_wr;
2020 memset(wr, 0, sizeof(*wr));
2022 wr->wr.opcode = IB_WR_FAST_REG_MR;
2023 wr->wr.wr_id = IBLND_WID_MR;
2025 wr->wr.wr.fast_reg.iova_start = iov;
2026 wr->wr.wr.fast_reg.page_list = frpl;
2027 wr->wr.wr.fast_reg.page_list_len = npages;
2028 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2029 wr->wr.wr.fast_reg.length = nob;
2030 wr->wr.wr.fast_reg.rkey =
2031 is_rx ? mr->rkey : mr->lkey;
2032 wr->wr.wr.fast_reg.access_flags =
2033 (IB_ACCESS_LOCAL_WRITE |
2034 IB_ACCESS_REMOTE_WRITE);
2035 #endif /* HAVE_IB_MAP_MR_SG */
2037 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
2039 fmr->fmr_pool = fpo;
2040 frd->frd_posted = false;
2043 spin_unlock(&fps->fps_lock);
2047 spin_lock(&fps->fps_lock);
2048 fpo->fpo_map_count--;
2049 if (rc != -EAGAIN) {
2050 spin_unlock(&fps->fps_lock);
2054 /* EAGAIN and ... */
2055 if (version != fps->fps_version) {
2056 spin_unlock(&fps->fps_lock);
2061 if (fps->fps_increasing) {
2062 spin_unlock(&fps->fps_lock);
2063 CDEBUG(D_NET, "Another thread is allocating new "
2064 "FMR pool, waiting for her to complete\n");
2065 wait_var_event(fps, !fps->fps_increasing);
2070 if (ktime_get_seconds() < fps->fps_next_retry) {
2071 /* someone failed recently */
2072 spin_unlock(&fps->fps_lock);
2076 fps->fps_increasing = 1;
2077 spin_unlock(&fps->fps_lock);
2079 CDEBUG(D_NET, "Allocate new FMR pool\n");
2080 rc = kiblnd_create_fmr_pool(fps, &fpo);
2081 spin_lock(&fps->fps_lock);
2082 fps->fps_increasing = 0;
2086 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2088 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2090 spin_unlock(&fps->fps_lock);
2096 kiblnd_fini_pool(struct kib_pool *pool)
2098 LASSERT(list_empty(&pool->po_free_list));
2099 LASSERT(pool->po_allocated == 0);
2101 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2105 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2107 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2109 memset(pool, 0, sizeof(struct kib_pool));
2110 INIT_LIST_HEAD(&pool->po_free_list);
2111 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2112 pool->po_owner = ps;
2113 pool->po_size = size;
2117 kiblnd_destroy_pool_list(struct list_head *head)
2119 struct kib_pool *pool;
2121 while ((pool = list_first_entry_or_null(head,
2123 po_list)) != NULL) {
2124 list_del(&pool->po_list);
2126 LASSERT(pool->po_owner != NULL);
2127 pool->po_owner->ps_pool_destroy(pool);
2132 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2134 struct kib_pool *po;
2136 if (ps->ps_net == NULL) /* intialized? */
2139 spin_lock(&ps->ps_lock);
2140 while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2142 po_list)) != NULL) {
2144 if (po->po_allocated == 0)
2145 list_move(&po->po_list, zombies);
2147 list_move(&po->po_list, &ps->ps_failed_pool_list);
2149 spin_unlock(&ps->ps_lock);
2153 kiblnd_fini_poolset(struct kib_poolset *ps)
2155 if (ps->ps_net != NULL) { /* initialized? */
2156 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2157 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2162 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2163 struct kib_net *net, char *name, int size,
2164 kib_ps_pool_create_t po_create,
2165 kib_ps_pool_destroy_t po_destroy,
2166 kib_ps_node_init_t nd_init,
2167 kib_ps_node_fini_t nd_fini)
2169 struct kib_pool *pool;
2172 memset(ps, 0, sizeof(struct kib_poolset));
2176 ps->ps_pool_create = po_create;
2177 ps->ps_pool_destroy = po_destroy;
2178 ps->ps_node_init = nd_init;
2179 ps->ps_node_fini = nd_fini;
2180 ps->ps_pool_size = size;
2181 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2182 >= sizeof(ps->ps_name))
2184 spin_lock_init(&ps->ps_lock);
2185 INIT_LIST_HEAD(&ps->ps_pool_list);
2186 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2188 rc = ps->ps_pool_create(ps, size, &pool);
2190 list_add(&pool->po_list, &ps->ps_pool_list);
2192 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2198 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2200 if (pool->po_allocated != 0) /* still in use */
2202 if (pool->po_failed)
2204 return now >= pool->po_deadline;
2208 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2211 struct kib_poolset *ps = pool->po_owner;
2212 struct kib_pool *tmp;
2213 time64_t now = ktime_get_seconds();
2215 spin_lock(&ps->ps_lock);
2217 if (ps->ps_node_fini != NULL)
2218 ps->ps_node_fini(pool, node);
2220 LASSERT(pool->po_allocated > 0);
2221 list_add(node, &pool->po_free_list);
2222 pool->po_allocated--;
2224 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2225 /* the first pool is persistent */
2226 if (ps->ps_pool_list.next == &pool->po_list)
2229 if (kiblnd_pool_is_idle(pool, now))
2230 list_move(&pool->po_list, &zombies);
2232 spin_unlock(&ps->ps_lock);
2234 if (!list_empty(&zombies))
2235 kiblnd_destroy_pool_list(&zombies);
2239 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2241 struct list_head *node;
2242 struct kib_pool *pool;
2244 unsigned int interval = 1;
2245 ktime_t time_before;
2246 unsigned int trips = 0;
2249 spin_lock(&ps->ps_lock);
2250 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2251 if (list_empty(&pool->po_free_list))
2254 pool->po_allocated++;
2255 pool->po_deadline = ktime_get_seconds() +
2256 IBLND_POOL_DEADLINE;
2257 node = pool->po_free_list.next;
2260 if (ps->ps_node_init != NULL) {
2261 /* still hold the lock */
2262 ps->ps_node_init(pool, node);
2264 spin_unlock(&ps->ps_lock);
2268 /* no available tx pool and ... */
2269 if (ps->ps_increasing) {
2270 /* another thread is allocating a new pool */
2271 spin_unlock(&ps->ps_lock);
2274 "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2275 ps->ps_name, interval, trips);
2277 schedule_timeout_interruptible(interval);
2278 if (interval < cfs_time_seconds(1))
2284 if (ktime_get_seconds() < ps->ps_next_retry) {
2285 /* someone failed recently */
2286 spin_unlock(&ps->ps_lock);
2290 ps->ps_increasing = 1;
2291 spin_unlock(&ps->ps_lock);
2293 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2294 time_before = ktime_get();
2295 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2296 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete\n",
2297 ktime_ms_delta(ktime_get(), time_before));
2299 spin_lock(&ps->ps_lock);
2300 ps->ps_increasing = 0;
2302 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2304 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2305 CERROR("Can't allocate new %s pool because out of memory\n",
2308 spin_unlock(&ps->ps_lock);
2314 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2316 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2320 LASSERT (pool->po_allocated == 0);
2322 if (tpo->tpo_tx_pages != NULL) {
2323 kiblnd_unmap_tx_pool(tpo);
2324 kiblnd_free_pages(tpo->tpo_tx_pages);
2327 if (tpo->tpo_tx_descs == NULL)
2330 for (i = 0; i < pool->po_size; i++) {
2331 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2332 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2334 list_del(&tx->tx_list);
2335 if (tx->tx_pages != NULL)
2336 CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2337 if (tx->tx_frags != NULL)
2338 CFS_FREE_PTR_ARRAY(tx->tx_frags,
2339 IBLND_MAX_RDMA_FRAGS);
2340 if (tx->tx_wrq != NULL)
2341 CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2342 IBLND_MAX_RDMA_FRAGS);
2343 if (tx->tx_sge != NULL) {
2344 /* +1 is for the lnet header/message itself */
2345 CFS_FREE_PTR_ARRAY(tx->tx_sge,
2346 (IBLND_MAX_RDMA_FRAGS *
2349 if (tx->tx_rd != NULL)
2350 LIBCFS_FREE(tx->tx_rd,
2351 offsetof(struct kib_rdma_desc,
2352 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2355 CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2357 kiblnd_fini_pool(pool);
2361 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2363 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2366 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2367 ntx = tunables->lnd_ntx / ncpts;
2369 return max(IBLND_TX_POOL, ntx);
2373 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2377 struct kib_pool *pool;
2378 struct kib_tx_pool *tpo;
2380 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2382 CERROR("Failed to allocate TX pool\n");
2386 pool = &tpo->tpo_pool;
2387 kiblnd_init_pool(ps, pool, size);
2388 tpo->tpo_tx_descs = NULL;
2389 tpo->tpo_tx_pages = NULL;
2391 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2392 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2393 CERROR("Can't allocate tx pages: %d\n", npg);
2398 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2399 size * sizeof(struct kib_tx));
2400 if (tpo->tpo_tx_descs == NULL) {
2401 CERROR("Can't allocate %d tx descriptors\n", size);
2402 ps->ps_pool_destroy(pool);
2406 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2408 for (i = 0; i < size; i++) {
2409 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2410 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2413 if (ps->ps_net->ibn_fmr_ps != NULL) {
2414 LIBCFS_CPT_ALLOC(tx->tx_pages,
2415 lnet_cpt_table(), ps->ps_cpt,
2416 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2417 if (tx->tx_pages == NULL)
2421 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2422 IBLND_MAX_RDMA_FRAGS *
2423 sizeof(*tx->tx_frags));
2424 if (tx->tx_frags == NULL)
2427 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2429 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2430 IBLND_MAX_RDMA_FRAGS *
2431 sizeof(*tx->tx_wrq));
2432 if (tx->tx_wrq == NULL)
2435 /* +1 is for the lnet header/message itself */
2436 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2437 (IBLND_MAX_RDMA_FRAGS * wrq_sge + 1) *
2438 sizeof(*tx->tx_sge));
2439 if (tx->tx_sge == NULL)
2442 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2443 offsetof(struct kib_rdma_desc,
2444 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2445 if (tx->tx_rd == NULL)
2450 kiblnd_map_tx_pool(tpo);
2455 ps->ps_pool_destroy(pool);
2460 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2462 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2463 struct kib_tx_poolset,
2465 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2467 tx->tx_cookie = tps->tps_next_tx_cookie++;
2471 kiblnd_net_fini_pools(struct kib_net *net)
2475 cfs_cpt_for_each(i, lnet_cpt_table()) {
2476 struct kib_tx_poolset *tps;
2477 struct kib_fmr_poolset *fps;
2479 if (net->ibn_tx_ps != NULL) {
2480 tps = net->ibn_tx_ps[i];
2481 kiblnd_fini_poolset(&tps->tps_poolset);
2484 if (net->ibn_fmr_ps != NULL) {
2485 fps = net->ibn_fmr_ps[i];
2486 kiblnd_fini_fmr_poolset(fps);
2490 if (net->ibn_tx_ps != NULL) {
2491 cfs_percpt_free(net->ibn_tx_ps);
2492 net->ibn_tx_ps = NULL;
2495 if (net->ibn_fmr_ps != NULL) {
2496 cfs_percpt_free(net->ibn_fmr_ps);
2497 net->ibn_fmr_ps = NULL;
2502 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2505 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2506 #ifdef HAVE_IB_GET_DMA_MR
2507 unsigned long flags;
2513 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2515 #ifdef HAVE_IB_GET_DMA_MR
2516 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2518 * if lnd_map_on_demand is zero then we have effectively disabled
2519 * FMR or FastReg and we're using global memory regions
2522 if (!tunables->lnd_map_on_demand) {
2523 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2525 goto create_tx_pool;
2528 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2531 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2532 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2533 tunables->lnd_fmr_pool_size,
2534 tunables->lnd_ntx / 4);
2539 /* TX pool must be created later than FMR, see LU-2268
2541 LASSERT(net->ibn_tx_ps == NULL);
2543 /* premapping can fail if ibd_nmr > 1, so we always create
2544 * FMR pool and map-on-demand if premapping failed */
2546 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2547 sizeof(struct kib_fmr_poolset));
2548 if (net->ibn_fmr_ps == NULL) {
2549 CERROR("Failed to allocate FMR pool array\n");
2554 for (i = 0; i < ncpts; i++) {
2555 cpt = (cpts == NULL) ? i : cpts[i];
2556 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2559 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2566 LASSERT(i == ncpts);
2568 #ifdef HAVE_IB_GET_DMA_MR
2571 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2572 sizeof(struct kib_tx_poolset));
2573 if (net->ibn_tx_ps == NULL) {
2574 CERROR("Failed to allocate tx pool array\n");
2579 for (i = 0; i < ncpts; i++) {
2580 cpt = (cpts == NULL) ? i : cpts[i];
2581 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2583 kiblnd_tx_pool_size(ni, ncpts),
2584 kiblnd_create_tx_pool,
2585 kiblnd_destroy_tx_pool,
2586 kiblnd_tx_init, NULL);
2588 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2596 kiblnd_net_fini_pools(net);
2602 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2604 struct ib_port_attr *port_attr;
2606 unsigned long flags;
2607 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2609 LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2610 if (port_attr == NULL) {
2611 CDEBUG(D_NETERROR, "Out of memory\n");
2615 rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2617 write_lock_irqsave(g_lock, flags);
2620 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2621 ? IBLND_DEV_PORT_ACTIVE
2622 : IBLND_DEV_PORT_DOWN;
2624 write_unlock_irqrestore(g_lock, flags);
2625 LIBCFS_FREE(port_attr, sizeof(*port_attr));
2628 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2635 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2637 struct kib_net *net;
2638 __u32 ni_state_before;
2639 bool update_ping_buf = false;
2641 /* for health check */
2642 list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2644 CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2645 libcfs_nidstr(&net->ibn_ni->ni_nid));
2646 ni_state_before = atomic_xchg(&net->ibn_ni->ni_fatal_error_on,
2648 if (!update_ping_buf && (val != ni_state_before))
2649 update_ping_buf = true;
2652 if (update_ping_buf)
2653 lnet_update_ping_buffer();
2657 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2659 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2660 struct kib_hca_dev *hdev;
2661 unsigned long flags;
2663 hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2665 write_lock_irqsave(g_lock, flags);
2667 switch (event->event) {
2668 case IB_EVENT_DEVICE_FATAL:
2669 CDEBUG(D_NET, "IB device fatal\n");
2670 hdev->ibh_state = IBLND_DEV_FATAL;
2671 kiblnd_set_ni_fatal_on(hdev, 1);
2673 case IB_EVENT_PORT_ACTIVE:
2674 CDEBUG(D_NET, "IB port active\n");
2675 if (event->element.port_num == hdev->ibh_port) {
2676 hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2677 kiblnd_set_ni_fatal_on(hdev, 0);
2680 case IB_EVENT_PORT_ERR:
2681 CDEBUG(D_NET, "IB port err\n");
2682 if (event->element.port_num == hdev->ibh_port) {
2683 hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2684 kiblnd_set_ni_fatal_on(hdev, 1);
2690 write_unlock_irqrestore(g_lock, flags);
2694 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2696 struct ib_device_attr *dev_attr;
2700 /* It's safe to assume a HCA can handle a page size
2701 * matching that of the native system */
2702 hdev->ibh_page_shift = PAGE_SHIFT;
2703 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2704 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2706 #ifndef HAVE_IB_DEVICE_ATTRS
2707 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2708 if (dev_attr == NULL) {
2709 CERROR("Out of memory\n");
2713 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2715 CERROR("Failed to query IB device: %d\n", rc);
2716 goto out_clean_attr;
2719 dev_attr = &hdev->ibh_ibdev->attrs;
2722 hdev->ibh_mr_size = dev_attr->max_mr_size;
2723 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2725 /* Setup device Memory Registration capabilities */
2726 #ifdef HAVE_FMR_POOL_API
2727 #ifdef HAVE_IB_DEVICE_OPS
2728 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2729 hdev->ibh_ibdev->ops.dealloc_fmr &&
2730 hdev->ibh_ibdev->ops.map_phys_fmr &&
2731 hdev->ibh_ibdev->ops.unmap_fmr) {
2733 if (hdev->ibh_ibdev->alloc_fmr &&
2734 hdev->ibh_ibdev->dealloc_fmr &&
2735 hdev->ibh_ibdev->map_phys_fmr &&
2736 hdev->ibh_ibdev->unmap_fmr) {
2738 LCONSOLE_INFO("Using FMR for registration\n");
2739 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2741 #endif /* HAVE_FMR_POOL_API */
2742 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2743 LCONSOLE_INFO("Using FastReg for registration\n");
2744 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2745 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2746 #ifdef IB_DEVICE_SG_GAPS_REG
2747 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2748 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2755 rc2 = kiblnd_port_get_attr(hdev);
2762 #ifndef HAVE_IB_DEVICE_ATTRS
2764 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2768 CERROR("IB device does not support FMRs nor FastRegs, can't "
2769 "register memory: %d\n", rc);
2770 else if (rc == -EINVAL)
2771 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2775 #ifdef HAVE_IB_GET_DMA_MR
2777 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2779 if (hdev->ibh_mrs == NULL)
2782 ib_dereg_mr(hdev->ibh_mrs);
2784 hdev->ibh_mrs = NULL;
2789 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2791 if (hdev->ibh_event_handler.device != NULL)
2792 ib_unregister_event_handler(&hdev->ibh_event_handler);
2794 #ifdef HAVE_IB_GET_DMA_MR
2795 kiblnd_hdev_cleanup_mrs(hdev);
2798 if (hdev->ibh_pd != NULL)
2799 ib_dealloc_pd(hdev->ibh_pd);
2801 if (hdev->ibh_cmid != NULL)
2802 rdma_destroy_id(hdev->ibh_cmid);
2804 LIBCFS_FREE(hdev, sizeof(*hdev));
2807 #ifdef HAVE_IB_GET_DMA_MR
2809 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2812 int acflags = IB_ACCESS_LOCAL_WRITE |
2813 IB_ACCESS_REMOTE_WRITE;
2815 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2817 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2818 kiblnd_hdev_cleanup_mrs(hdev);
2829 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2834 static int kiblnd_get_link_status(struct net_device *dev)
2840 if (!netif_running(dev))
2842 /* Some devices may not be providing link settings */
2843 else if (dev->ethtool_ops->get_link)
2844 ret = dev->ethtool_ops->get_link(dev);
2850 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2852 struct rdma_cm_id *cmid;
2853 struct sockaddr_in srcaddr;
2854 struct sockaddr_in dstaddr;
2857 if (dev->ibd_hdev == NULL || /* initializing */
2858 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2859 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2862 /* XXX: it's UGLY, but I don't have better way to find
2863 * ib-bonding HCA failover because:
2865 * a. no reliable CM event for HCA failover...
2866 * b. no OFED API to get ib_device for current net_device...
2868 * We have only two choices at this point:
2870 * a. rdma_bind_addr(), it will conflict with listener cmid
2871 * b. rdma_resolve_addr() to zero addr */
2872 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2873 RDMA_PS_TCP, IB_QPT_RC);
2876 CERROR("Failed to create cmid for failover: %d\n", rc);
2880 memset(&srcaddr, 0, sizeof(srcaddr));
2881 srcaddr.sin_family = AF_INET;
2882 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2884 memset(&dstaddr, 0, sizeof(dstaddr));
2885 dstaddr.sin_family = AF_INET;
2886 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2887 (struct sockaddr *)&dstaddr, 1);
2888 if (rc != 0 || cmid->device == NULL) {
2889 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2890 dev->ibd_ifname, &dev->ibd_ifip,
2892 rdma_destroy_id(cmid);
2896 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2897 rdma_destroy_id(cmid);
2902 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2904 LIST_HEAD(zombie_tpo);
2905 LIST_HEAD(zombie_ppo);
2906 LIST_HEAD(zombie_fpo);
2907 struct rdma_cm_id *cmid = NULL;
2908 struct kib_hca_dev *hdev = NULL;
2909 struct kib_hca_dev *old;
2911 struct kib_net *net;
2912 struct sockaddr_in addr;
2913 struct net_device *netdev;
2914 unsigned long flags;
2917 bool set_fatal = true;
2919 LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2920 dev->ibd_can_failover ||
2921 dev->ibd_hdev == NULL);
2923 rc = kiblnd_dev_need_failover(dev, ns);
2927 if (dev->ibd_hdev != NULL &&
2928 dev->ibd_hdev->ibh_cmid != NULL) {
2929 /* XXX it's not good to close old listener at here,
2930 * because we can fail to create new listener.
2931 * But we have to close it now, otherwise rdma_bind_addr
2932 * will return EADDRINUSE... How crap! */
2933 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2935 cmid = dev->ibd_hdev->ibh_cmid;
2936 /* make next schedule of kiblnd_dev_need_failover()
2937 * return 1 for me */
2938 dev->ibd_hdev->ibh_cmid = NULL;
2939 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2941 rdma_destroy_id(cmid);
2944 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2948 CERROR("Failed to create cmid for failover: %d\n", rc);
2952 memset(&addr, 0, sizeof(addr));
2953 addr.sin_family = AF_INET;
2954 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2955 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2957 /* Bind to failover device or port */
2958 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2959 if (rc != 0 || cmid->device == NULL) {
2960 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2961 dev->ibd_ifname, &dev->ibd_ifip,
2963 if (!rc && !cmid->device)
2965 rdma_destroy_id(cmid);
2969 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2971 CERROR("Failed to allocate kib_hca_dev\n");
2972 rdma_destroy_id(cmid);
2977 atomic_set(&hdev->ibh_ref, 1);
2978 hdev->ibh_dev = dev;
2979 hdev->ibh_cmid = cmid;
2980 hdev->ibh_ibdev = cmid->device;
2981 hdev->ibh_port = cmid->port_num;
2983 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2984 pd = ib_alloc_pd(cmid->device, 0);
2986 pd = ib_alloc_pd(cmid->device);
2990 CERROR("Can't allocate PD: %d\n", rc);
2996 rc = rdma_listen(cmid, 0);
2998 CERROR("Can't start new listener: %d\n", rc);
3002 rc = kiblnd_hdev_get_attr(hdev);
3004 CERROR("Can't get device attributes: %d\n", rc);
3008 #ifdef HAVE_IB_GET_DMA_MR
3009 rc = kiblnd_hdev_setup_mrs(hdev);
3011 CERROR("Can't setup device: %d\n", rc);
3016 INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
3017 hdev->ibh_ibdev, kiblnd_event_handler);
3018 ib_register_event_handler(&hdev->ibh_event_handler);
3020 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3022 old = dev->ibd_hdev;
3023 dev->ibd_hdev = hdev; /* take over the refcount */
3026 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
3027 cfs_cpt_for_each(i, lnet_cpt_table()) {
3028 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
3031 if (net->ibn_fmr_ps != NULL)
3032 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
3037 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3039 if (!list_empty(&zombie_tpo))
3040 kiblnd_destroy_pool_list(&zombie_tpo);
3041 if (!list_empty(&zombie_ppo))
3042 kiblnd_destroy_pool_list(&zombie_ppo);
3043 if (!list_empty(&zombie_fpo))
3044 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
3046 kiblnd_hdev_decref(hdev);
3049 dev->ibd_failed_failover++;
3051 dev->ibd_failed_failover = 0;
3055 netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
3056 if (netdev && (kiblnd_get_link_status(netdev) == 1))
3057 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
3066 kiblnd_destroy_dev(struct kib_dev *dev)
3068 LASSERT(dev->ibd_nnets == 0);
3069 LASSERT(list_empty(&dev->ibd_nets));
3071 list_del(&dev->ibd_fail_list);
3072 list_del(&dev->ibd_list);
3074 if (dev->ibd_hdev != NULL)
3075 kiblnd_hdev_decref(dev->ibd_hdev);
3077 LIBCFS_FREE(dev, sizeof(*dev));
3080 static struct kib_dev *
3081 kiblnd_dev_search(char *ifname)
3083 struct kib_dev *alias = NULL;
3084 struct kib_dev *dev;
3088 colon = strchr(ifname, ':');
3089 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3090 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3096 colon2 = strchr(dev->ibd_ifname, ':');
3102 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3114 kiblnd_handle_link_state_change(struct net_device *dev,
3115 unsigned char operstate)
3117 struct lnet_ni *ni = NULL;
3118 struct kib_dev *event_kibdev;
3119 struct kib_net *net;
3120 struct kib_net *cnxt;
3121 bool link_down = !(operstate == IF_OPER_UP);
3122 struct in_device *in_dev;
3123 bool found_ip = false;
3124 __u32 ni_state_before;
3125 bool update_ping_buf = false;
3126 DECLARE_CONST_IN_IFADDR(ifa);
3128 event_kibdev = kiblnd_dev_search(dev->name);
3133 list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets, ibn_list) {
3137 in_dev = __in_dev_get_rtnl(dev);
3139 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
3141 CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3142 libcfs_nidstr(&net->ibn_ni->ni_nid));
3143 ni_state_before = atomic_xchg(&ni->ni_fatal_error_on,
3147 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
3148 if (htonl(event_kibdev->ibd_ifip) == ifa->ifa_local)
3154 CDEBUG(D_NET, "Interface %s has no matching ip\n",
3156 CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3157 libcfs_nidstr(&net->ibn_ni->ni_nid));
3158 ni_state_before = atomic_xchg(&ni->ni_fatal_error_on,
3164 CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3165 libcfs_nidstr(&net->ibn_ni->ni_nid));
3166 ni_state_before = atomic_xchg(&ni->ni_fatal_error_on,
3169 CDEBUG(D_NET, "%s: set link fatal state to %u\n",
3170 libcfs_nidstr(&net->ibn_ni->ni_nid),
3171 (kiblnd_get_link_status(dev) == 0));
3172 ni_state_before = atomic_xchg(&ni->ni_fatal_error_on,
3173 (kiblnd_get_link_status(dev) == 0));
3176 if (!update_ping_buf &&
3177 (atomic_read(&ni->ni_fatal_error_on) != ni_state_before))
3178 update_ping_buf = true;
3181 if (update_ping_buf)
3182 lnet_update_ping_buffer();
3188 kiblnd_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
3190 struct kib_dev *event_kibdev;
3191 struct kib_net *net;
3192 struct kib_net *cnxt;
3193 struct net_device *event_netdev = ifa->ifa_dev->dev;
3194 __u32 ni_state_before;
3195 bool update_ping_buf = false;
3197 event_kibdev = kiblnd_dev_search(event_netdev->name);
3202 if (htonl(event_kibdev->ibd_ifip) != ifa->ifa_local)
3205 list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets,
3207 CDEBUG(D_NET, "%s: set link fatal state to %u\n",
3208 libcfs_nidstr(&net->ibn_ni->ni_nid),
3209 (event == NETDEV_DOWN));
3210 ni_state_before = atomic_xchg(&net->ibn_ni->ni_fatal_error_on,
3211 (event == NETDEV_DOWN));
3212 if (!update_ping_buf &&
3213 ((event == NETDEV_DOWN) != ni_state_before))
3214 update_ping_buf = true;
3217 if (update_ping_buf)
3218 lnet_update_ping_buffer();
3224 /************************************
3225 * Net device notifier event handler
3226 ************************************/
3227 static int kiblnd_device_event(struct notifier_block *unused,
3228 unsigned long event, void *ptr)
3230 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
3231 unsigned char operstate;
3233 operstate = dev->operstate;
3235 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
3236 event, dev->name, dev->ifindex, operstate);
3242 kiblnd_handle_link_state_change(dev, operstate);
3249 /************************************
3250 * Inetaddr notifier event handler
3251 ************************************/
3252 static int kiblnd_inetaddr_event(struct notifier_block *unused,
3253 unsigned long event, void *ptr)
3255 struct in_ifaddr *ifa = ptr;
3257 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
3258 event, &ifa->ifa_address, &ifa->ifa_mask);
3264 kiblnd_handle_inetaddr_change(ifa, event);
3271 static struct notifier_block kiblnd_dev_notifier_block = {
3272 .notifier_call = kiblnd_device_event,
3275 static struct notifier_block kiblnd_inetaddr_notifier_block = {
3276 .notifier_call = kiblnd_inetaddr_event,
3280 kiblnd_base_shutdown(void)
3282 struct kib_sched_info *sched;
3283 struct kib_peer_ni *peer_ni;
3286 LASSERT(list_empty(&kiblnd_data.kib_devs));
3288 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
3289 libcfs_kmem_read());
3291 if (kiblnd_data.kib_init == IBLND_INIT_ALL) {
3292 unregister_netdevice_notifier(&kiblnd_dev_notifier_block);
3293 unregister_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3296 switch (kiblnd_data.kib_init) {
3300 case IBLND_INIT_ALL:
3301 case IBLND_INIT_DATA:
3302 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
3304 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
3305 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
3306 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
3307 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
3309 /* flag threads to terminate; wake and wait for them to die */
3310 kiblnd_data.kib_shutdown = 1;
3312 /* NB: we really want to stop scheduler threads net by net
3313 * instead of the whole module, this should be improved
3314 * with dynamic configuration LNet.
3316 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
3317 wake_up_all(&sched->ibs_waitq);
3319 wake_up(&kiblnd_data.kib_connd_waitq);
3320 wake_up(&kiblnd_data.kib_failover_waitq);
3322 wait_var_event_warning(&kiblnd_data.kib_nthreads,
3323 !atomic_read(&kiblnd_data.kib_nthreads),
3324 "Waiting for %d threads to terminate\n",
3325 atomic_read(&kiblnd_data.kib_nthreads));
3328 case IBLND_INIT_NOTHING:
3332 if (kiblnd_data.kib_scheds != NULL)
3333 cfs_percpt_free(kiblnd_data.kib_scheds);
3335 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3336 libcfs_kmem_read());
3338 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3339 module_put(THIS_MODULE);
3343 kiblnd_shutdown(struct lnet_ni *ni)
3345 struct kib_net *net = ni->ni_data;
3346 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
3347 unsigned long flags;
3349 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3354 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3355 libcfs_kmem_read());
3357 write_lock_irqsave(g_lock, flags);
3358 net->ibn_shutdown = 1;
3359 write_unlock_irqrestore(g_lock, flags);
3361 switch (net->ibn_init) {
3365 case IBLND_INIT_ALL:
3366 /* nuke all existing peers within this net */
3367 kiblnd_del_peer(ni, LNET_NID_ANY);
3369 /* Wait for all peer_ni state to clean up */
3370 wait_var_event_warning(&net->ibn_npeers,
3371 atomic_read(&net->ibn_npeers) == 0,
3372 "%s: waiting for %d peers to disconnect\n",
3373 libcfs_nidstr(&ni->ni_nid),
3374 atomic_read(&net->ibn_npeers));
3376 kiblnd_net_fini_pools(net);
3378 write_lock_irqsave(g_lock, flags);
3379 LASSERT(net->ibn_dev->ibd_nnets > 0);
3380 net->ibn_dev->ibd_nnets--;
3381 list_del(&net->ibn_list);
3382 write_unlock_irqrestore(g_lock, flags);
3384 wake_up_all(&kiblnd_data.kib_connd_waitq);
3385 wait_var_event_warning(&net->ibn_nconns,
3386 atomic_read(&net->ibn_nconns) == 0,
3387 "%s: waiting for %d conns to clean\n",
3388 libcfs_nidstr(&ni->ni_nid),
3389 atomic_read(&net->ibn_nconns));
3392 case IBLND_INIT_NOTHING:
3393 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3395 if (net->ibn_dev != NULL &&
3396 net->ibn_dev->ibd_nnets == 0)
3397 kiblnd_destroy_dev(net->ibn_dev);
3402 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3403 libcfs_kmem_read());
3405 net->ibn_init = IBLND_INIT_NOTHING;
3408 LIBCFS_FREE(net, sizeof(*net));
3411 if (list_empty(&kiblnd_data.kib_devs))
3412 kiblnd_base_shutdown();
3416 kiblnd_base_startup(struct net *ns)
3418 struct kib_sched_info *sched;
3422 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3424 if (!try_module_get(THIS_MODULE))
3427 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3429 rwlock_init(&kiblnd_data.kib_global_lock);
3431 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3432 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3434 hash_init(kiblnd_data.kib_peers);
3436 spin_lock_init(&kiblnd_data.kib_connd_lock);
3437 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3438 INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3439 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3440 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3441 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3443 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3444 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3446 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3448 if (kiblnd_data.kib_scheds == NULL)
3451 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3454 spin_lock_init(&sched->ibs_lock);
3455 INIT_LIST_HEAD(&sched->ibs_conns);
3456 init_waitqueue_head(&sched->ibs_waitq);
3458 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3459 if (*kiblnd_tunables.kib_nscheds > 0) {
3460 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3462 /* max to half of CPUs, another half is reserved for
3463 * upper layer modules */
3464 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3467 sched->ibs_nthreads_max = nthrs;
3471 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3473 /* lists/ptrs/locks initialised */
3474 kiblnd_data.kib_init = IBLND_INIT_DATA;
3475 /*****************************************************/
3477 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3479 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3483 if (*kiblnd_tunables.kib_dev_failover != 0)
3484 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3488 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3492 register_netdevice_notifier(&kiblnd_dev_notifier_block);
3493 register_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3495 /* flag everything initialised */
3496 kiblnd_data.kib_init = IBLND_INIT_ALL;
3497 /*****************************************************/
3502 kiblnd_base_shutdown();
3507 kiblnd_start_schedulers(struct kib_sched_info *sched)
3513 if (sched->ibs_nthreads == 0) {
3514 if (*kiblnd_tunables.kib_nscheds > 0) {
3515 nthrs = sched->ibs_nthreads_max;
3517 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3519 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3520 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3523 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3524 /* increase one thread if there is new interface */
3525 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3528 for (i = 0; i < nthrs; i++) {
3529 long id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3531 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3532 "kiblnd_sd_%02ld_%02ld",
3533 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3537 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3538 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3542 sched->ibs_nthreads += i;
3546 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3553 for (i = 0; i < ncpts; i++) {
3554 struct kib_sched_info *sched;
3556 cpt = (cpts == NULL) ? i : cpts[i];
3557 sched = kiblnd_data.kib_scheds[cpt];
3559 if (!newdev && sched->ibs_nthreads > 0)
3562 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3564 CERROR("Failed to start scheduler threads for %s\n",
3573 kiblnd_startup(struct lnet_ni *ni)
3575 char *ifname = NULL;
3576 struct lnet_inetdev *ifaces = NULL;
3577 struct kib_dev *ibdev = NULL;
3578 struct kib_net *net = NULL;
3579 unsigned long flags;
3584 LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3586 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3587 rc = kiblnd_base_startup(ni->ni_net_ns);
3592 LIBCFS_ALLOC(net, sizeof(*net));
3600 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3602 kiblnd_tunables_setup(ni);
3604 /* Multi-Rail wants each secondary
3605 * IP to be treated as an unique 'struct ni' interface.
3607 if (ni->ni_interface != NULL) {
3608 /* Use the IPoIB interface specified in 'networks=' */
3609 ifname = ni->ni_interface;
3611 ifname = *kiblnd_tunables.kib_default_ipif;
3612 rc = libcfs_strnid(&ni->ni_nid, ifname);
3613 if (rc < 0 || ni->ni_nid.nid_type != O2IBLND)
3614 memset(&ni->ni_nid, 0, sizeof(ni->ni_nid));
3617 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3618 CERROR("IPoIB interface name too long: %s\n", ifname);
3623 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, false);
3627 i = lnet_inet_select(ni, ifaces, rc);
3631 if (nid_addr_is_set(&ni->ni_nid)) {
3632 strscpy(ifname, ifaces[i].li_name, sizeof(ifname));
3633 } else if (strcmp(ifname, ifaces[i].li_name) != 0) {
3634 CERROR("ko2iblnd: No matching interfaces\n");
3639 ibdev = kiblnd_dev_search(ifname);
3640 newdev = ibdev == NULL;
3641 /* hmm...create kib_dev even for alias */
3642 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3643 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3649 ibdev->ibd_ifip = ntohl(ifaces[i].li_ipaddr);
3650 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3651 sizeof(ibdev->ibd_ifname));
3652 ibdev->ibd_can_failover = ifaces[i].li_iff_master;
3654 INIT_LIST_HEAD(&ibdev->ibd_nets);
3655 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3656 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3658 /* initialize the device */
3659 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3661 CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3666 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3669 net->ibn_dev = ibdev;
3670 ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3671 if (!ni->ni_interface) {
3672 rc = lnet_ni_add_interface(ni, ifaces[i].li_name);
3674 CWARN("ko2iblnd failed to allocate ni_interface\n");
3676 ni->ni_dev_cpt = ifaces[i].li_cpt;
3678 rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3682 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3684 CERROR("Failed to initialize NI pools: %d\n", rc);
3688 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3690 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3691 /* for health check */
3692 if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3693 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3694 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3696 net->ibn_init = IBLND_INIT_ALL;
3701 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3702 kiblnd_destroy_dev(ibdev);
3705 kiblnd_shutdown(ni);
3707 CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3708 ifname ? ifname : "", rc);
3713 static const struct lnet_lnd the_o2iblnd = {
3714 .lnd_type = O2IBLND,
3715 .lnd_startup = kiblnd_startup,
3716 .lnd_shutdown = kiblnd_shutdown,
3717 .lnd_ctl = kiblnd_ctl,
3718 .lnd_send = kiblnd_send,
3719 .lnd_recv = kiblnd_recv,
3720 .lnd_get_dev_prio = kiblnd_get_dev_prio,
3721 .lnd_nl_set = kiblnd_nl_set,
3722 .lnd_keys = &kiblnd_tunables_keys,
3725 static void ko2inlnd_assert_wire_constants(void)
3727 BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3728 BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3729 BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3730 BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3732 BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3733 BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3734 BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3735 BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3736 BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3737 BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3738 BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3739 BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3740 BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3741 BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3743 BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3744 BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3745 BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3746 BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3747 BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3748 BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3749 BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3750 BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3752 BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3753 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3754 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3755 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3756 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3757 BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3758 BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3760 BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3761 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3762 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3763 BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3764 BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3766 BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3767 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3768 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3769 BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3770 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3772 BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3773 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3774 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3775 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3776 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3777 BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3778 BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3780 BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3781 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3782 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3783 BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3784 BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3786 BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3787 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3788 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3789 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3790 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3791 BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3792 BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3794 BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3795 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3796 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3797 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3798 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3799 BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3800 BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3802 BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3803 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3804 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3805 BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3806 BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3808 /* Checks for struct kib_msg */
3809 //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3810 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3811 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3812 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3813 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3814 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3815 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3816 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3817 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3818 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3819 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3820 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3821 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3822 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3823 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3824 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3825 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3826 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3827 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3828 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3829 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3832 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3833 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3834 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3835 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3836 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3837 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3839 /* Immediate message */
3840 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3841 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3842 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3843 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3845 /* PUT req message */
3846 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3847 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3848 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3849 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3852 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3853 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3854 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3855 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3856 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3857 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3860 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3861 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3862 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3863 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3864 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3865 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3867 /* Completion message */
3868 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3869 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3870 BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3871 BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3874 BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3875 BUILD_BUG_ON(offsetof(struct kib_msg,
3876 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3878 BUILD_BUG_ON(offsetof(struct kib_msg,
3879 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3883 static void __exit ko2iblnd_exit(void)
3885 lnet_unregister_lnd(&the_o2iblnd);
3888 static int __init ko2iblnd_init(void)
3892 ko2inlnd_assert_wire_constants();
3894 rc = kiblnd_tunables_init();
3898 lnet_register_lnd(&the_o2iblnd);
3903 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3904 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3905 MODULE_VERSION("2.8.0");
3906 MODULE_LICENSE("GPL");
3908 module_init(ko2iblnd_init);
3909 module_exit(ko2iblnd_exit);