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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd.c
38 * Author: Eric Barton <eric@bartonsoftware.com>
44 static lnd_t the_o2iblnd;
46 kib_data_t kiblnd_data;
49 kiblnd_cksum (void *ptr, int nob)
55 sum = ((sum << 1) | (sum >> 31)) + *c++;
57 /* ensure I don't return 0 (== no checksum) */
58 return (sum == 0) ? 1 : sum;
62 kiblnd_msgtype2str(int type)
65 case IBLND_MSG_CONNREQ:
68 case IBLND_MSG_CONNACK:
74 case IBLND_MSG_IMMEDIATE:
77 case IBLND_MSG_PUT_REQ:
80 case IBLND_MSG_PUT_NAK:
83 case IBLND_MSG_PUT_ACK:
86 case IBLND_MSG_PUT_DONE:
89 case IBLND_MSG_GET_REQ:
92 case IBLND_MSG_GET_DONE:
101 kiblnd_msgtype2size(int type)
103 const int hdr_size = offsetof(kib_msg_t, ibm_u);
106 case IBLND_MSG_CONNREQ:
107 case IBLND_MSG_CONNACK:
108 return hdr_size + sizeof(kib_connparams_t);
113 case IBLND_MSG_IMMEDIATE:
114 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
116 case IBLND_MSG_PUT_REQ:
117 return hdr_size + sizeof(kib_putreq_msg_t);
119 case IBLND_MSG_PUT_ACK:
120 return hdr_size + sizeof(kib_putack_msg_t);
122 case IBLND_MSG_GET_REQ:
123 return hdr_size + sizeof(kib_get_msg_t);
125 case IBLND_MSG_PUT_NAK:
126 case IBLND_MSG_PUT_DONE:
127 case IBLND_MSG_GET_DONE:
128 return hdr_size + sizeof(kib_completion_msg_t);
135 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
142 LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
143 msg->ibm_type == IBLND_MSG_PUT_ACK);
145 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
146 &msg->ibm_u.get.ibgm_rd :
147 &msg->ibm_u.putack.ibpam_rd;
150 __swab32s(&rd->rd_key);
151 __swab32s(&rd->rd_nfrags);
156 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
157 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
158 n, IBLND_MAX_RDMA_FRAGS);
162 nob = offsetof (kib_msg_t, ibm_u) +
163 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
165 if (msg->ibm_nob < nob) {
166 CERROR("Short %s: %d(%d)\n",
167 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
174 for (i = 0; i < n; i++) {
175 __swab32s(&rd->rd_frags[i].rf_nob);
176 __swab64s(&rd->rd_frags[i].rf_addr);
183 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
184 int credits, lnet_nid_t dstnid, __u64 dststamp)
186 kib_net_t *net = ni->ni_data;
188 /* CAVEAT EMPTOR! all message fields not set here should have been
189 * initialised previously. */
190 msg->ibm_magic = IBLND_MSG_MAGIC;
191 msg->ibm_version = version;
193 msg->ibm_credits = credits;
196 msg->ibm_srcnid = ni->ni_nid;
197 msg->ibm_srcstamp = net->ibn_incarnation;
198 msg->ibm_dstnid = dstnid;
199 msg->ibm_dststamp = dststamp;
201 if (*kiblnd_tunables.kib_cksum) {
202 /* NB ibm_cksum zero while computing cksum */
203 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
208 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
210 const int hdr_size = offsetof(kib_msg_t, ibm_u);
216 /* 6 bytes are enough to have received magic + version */
218 CERROR("Short message: %d\n", nob);
222 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
224 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
227 CERROR("Bad magic: %08x\n", msg->ibm_magic);
231 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
232 if (version != IBLND_MSG_VERSION &&
233 version != IBLND_MSG_VERSION_1) {
234 CERROR("Bad version: %x\n", version);
238 if (nob < hdr_size) {
239 CERROR("Short message: %d\n", nob);
243 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
245 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
249 /* checksum must be computed with ibm_cksum zero and BEFORE anything
251 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
253 if (msg_cksum != 0 &&
254 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
255 CERROR("Bad checksum\n");
259 msg->ibm_cksum = msg_cksum;
262 /* leave magic unflipped as a clue to peer endianness */
263 msg->ibm_version = version;
264 CLASSERT (sizeof(msg->ibm_type) == 1);
265 CLASSERT (sizeof(msg->ibm_credits) == 1);
266 msg->ibm_nob = msg_nob;
267 __swab64s(&msg->ibm_srcnid);
268 __swab64s(&msg->ibm_srcstamp);
269 __swab64s(&msg->ibm_dstnid);
270 __swab64s(&msg->ibm_dststamp);
273 if (msg->ibm_srcnid == LNET_NID_ANY) {
274 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
278 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
279 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
280 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
284 switch (msg->ibm_type) {
286 CERROR("Unknown message type %x\n", msg->ibm_type);
290 case IBLND_MSG_IMMEDIATE:
291 case IBLND_MSG_PUT_REQ:
294 case IBLND_MSG_PUT_ACK:
295 case IBLND_MSG_GET_REQ:
296 if (kiblnd_unpack_rd(msg, flip))
300 case IBLND_MSG_PUT_NAK:
301 case IBLND_MSG_PUT_DONE:
302 case IBLND_MSG_GET_DONE:
304 __swab32s(&msg->ibm_u.completion.ibcm_status);
307 case IBLND_MSG_CONNREQ:
308 case IBLND_MSG_CONNACK:
310 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
311 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
312 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
320 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
323 kib_net_t *net = ni->ni_data;
324 int cpt = lnet_cpt_of_nid(nid);
327 LASSERT(net != NULL);
328 LASSERT(nid != LNET_NID_ANY);
330 LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
332 CERROR("Cannot allocate peer\n");
339 peer->ibp_last_alive = 0;
340 atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
342 INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */
343 INIT_LIST_HEAD(&peer->ibp_conns);
344 INIT_LIST_HEAD(&peer->ibp_tx_queue);
346 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
348 /* always called with a ref on ni, which prevents ni being shutdown */
349 LASSERT(net->ibn_shutdown == 0);
351 /* npeers only grows with the global lock held */
352 atomic_inc(&net->ibn_npeers);
354 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
361 kiblnd_destroy_peer (kib_peer_t *peer)
363 kib_net_t *net = peer->ibp_ni->ni_data;
365 LASSERT(net != NULL);
366 LASSERT (atomic_read(&peer->ibp_refcount) == 0);
367 LASSERT(!kiblnd_peer_active(peer));
368 LASSERT(peer->ibp_connecting == 0);
369 LASSERT(peer->ibp_accepting == 0);
370 LASSERT(list_empty(&peer->ibp_conns));
371 LASSERT(list_empty(&peer->ibp_tx_queue));
373 LIBCFS_FREE(peer, sizeof(*peer));
375 /* NB a peer's connections keep a reference on their peer until
376 * they are destroyed, so we can be assured that _all_ state to do
377 * with this peer has been cleaned up when its refcount drops to
379 atomic_dec(&net->ibn_npeers);
383 kiblnd_find_peer_locked (lnet_nid_t nid)
385 /* the caller is responsible for accounting the additional reference
386 * that this creates */
387 struct list_head *peer_list = kiblnd_nid2peerlist(nid);
388 struct list_head *tmp;
391 list_for_each(tmp, peer_list) {
393 peer = list_entry(tmp, kib_peer_t, ibp_list);
395 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
396 peer->ibp_accepting > 0 ||
397 !list_empty(&peer->ibp_conns)); /* active conn */
399 if (peer->ibp_nid != nid)
402 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
403 peer, libcfs_nid2str(nid),
404 atomic_read(&peer->ibp_refcount),
412 kiblnd_unlink_peer_locked (kib_peer_t *peer)
414 LASSERT(list_empty(&peer->ibp_conns));
416 LASSERT (kiblnd_peer_active(peer));
417 list_del_init(&peer->ibp_list);
418 /* lose peerlist's ref */
419 kiblnd_peer_decref(peer);
423 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
424 lnet_nid_t *nidp, int *count)
427 struct list_head *ptmp;
431 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
433 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
435 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
437 peer = list_entry(ptmp, kib_peer_t, ibp_list);
438 LASSERT(peer->ibp_connecting > 0 ||
439 peer->ibp_accepting > 0 ||
440 !list_empty(&peer->ibp_conns));
442 if (peer->ibp_ni != ni)
448 *nidp = peer->ibp_nid;
449 *count = atomic_read(&peer->ibp_refcount);
451 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
457 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
462 kiblnd_del_peer_locked (kib_peer_t *peer)
464 struct list_head *ctmp;
465 struct list_head *cnxt;
468 if (list_empty(&peer->ibp_conns)) {
469 kiblnd_unlink_peer_locked(peer);
471 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
472 conn = list_entry(ctmp, kib_conn_t, ibc_list);
474 kiblnd_close_conn_locked(conn, 0);
476 /* NB closing peer's last conn unlinked it. */
478 /* NB peer now unlinked; might even be freed if the peer table had the
483 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
485 struct list_head zombies = LIST_HEAD_INIT(zombies);
486 struct list_head *ptmp;
487 struct list_head *pnxt;
495 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
497 if (nid != LNET_NID_ANY) {
498 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
501 hi = kiblnd_data.kib_peer_hash_size - 1;
504 for (i = lo; i <= hi; i++) {
505 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
506 peer = list_entry(ptmp, kib_peer_t, ibp_list);
507 LASSERT(peer->ibp_connecting > 0 ||
508 peer->ibp_accepting > 0 ||
509 !list_empty(&peer->ibp_conns));
511 if (peer->ibp_ni != ni)
514 if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
517 if (!list_empty(&peer->ibp_tx_queue)) {
518 LASSERT(list_empty(&peer->ibp_conns));
520 list_splice_init(&peer->ibp_tx_queue,
524 kiblnd_del_peer_locked(peer);
525 rc = 0; /* matched something */
529 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
531 kiblnd_txlist_done(ni, &zombies, -EIO);
537 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
540 struct list_head *ptmp;
542 struct list_head *ctmp;
546 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
548 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
549 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
551 peer = list_entry(ptmp, kib_peer_t, ibp_list);
552 LASSERT(peer->ibp_connecting > 0 ||
553 peer->ibp_accepting > 0 ||
554 !list_empty(&peer->ibp_conns));
556 if (peer->ibp_ni != ni)
559 list_for_each(ctmp, &peer->ibp_conns) {
563 conn = list_entry(ctmp, kib_conn_t, ibc_list);
564 kiblnd_conn_addref(conn);
565 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
572 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
577 kiblnd_debug_rx (kib_rx_t *rx)
579 CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n",
580 rx, rx->rx_status, rx->rx_msg->ibm_type,
581 rx->rx_msg->ibm_credits);
585 kiblnd_debug_tx (kib_tx_t *tx)
587 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx "
588 "cookie "LPX64" msg %s%s type %x cred %d\n",
589 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
590 tx->tx_status, tx->tx_deadline, tx->tx_cookie,
591 tx->tx_lntmsg[0] == NULL ? "-" : "!",
592 tx->tx_lntmsg[1] == NULL ? "-" : "!",
593 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
597 kiblnd_debug_conn (kib_conn_t *conn)
599 struct list_head *tmp;
602 spin_lock(&conn->ibc_lock);
604 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
605 atomic_read(&conn->ibc_refcount), conn,
606 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
607 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d "
608 " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
609 conn->ibc_nsends_posted, conn->ibc_credits,
610 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
611 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
613 CDEBUG(D_CONSOLE, " early_rxs:\n");
614 list_for_each(tmp, &conn->ibc_early_rxs)
615 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
617 CDEBUG(D_CONSOLE, " tx_noops:\n");
618 list_for_each(tmp, &conn->ibc_tx_noops)
619 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
621 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
622 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
623 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
625 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
626 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
627 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
629 CDEBUG(D_CONSOLE, " tx_queue:\n");
630 list_for_each(tmp, &conn->ibc_tx_queue)
631 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
633 CDEBUG(D_CONSOLE, " active_txs:\n");
634 list_for_each(tmp, &conn->ibc_active_txs)
635 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
637 CDEBUG(D_CONSOLE, " rxs:\n");
638 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
639 kiblnd_debug_rx(&conn->ibc_rxs[i]);
641 spin_unlock(&conn->ibc_lock);
645 kiblnd_translate_mtu(int value)
666 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
670 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
671 if (cmid->route.path_rec == NULL)
674 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
677 cmid->route.path_rec->mtu = mtu;
681 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
688 vectors = conn->ibc_cmid->device->num_comp_vectors;
692 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
694 /* hash NID to CPU id in this partition... */
695 off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
696 for_each_cpu(i, mask) {
706 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
707 int state, int version)
710 * If the new conn is created successfully it takes over the caller's
711 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
712 * is destroyed. On failure, the caller's ref on 'peer' remains and
713 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
714 * to destroy 'cmid' here since I'm called from the CM which still has
715 * its ref on 'cmid'). */
716 rwlock_t *glock = &kiblnd_data.kib_global_lock;
717 kib_net_t *net = peer->ibp_ni->ni_data;
719 struct ib_qp_init_attr *init_qp_attr;
720 struct kib_sched_info *sched;
721 #ifdef HAVE_IB_CQ_INIT_ATTR
722 struct ib_cq_init_attr cq_attr = {};
731 LASSERT(net != NULL);
732 LASSERT(!in_interrupt());
736 cpt = lnet_cpt_of_nid(peer->ibp_nid);
737 sched = kiblnd_data.kib_scheds[cpt];
739 LASSERT(sched->ibs_nthreads > 0);
741 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
742 sizeof(*init_qp_attr));
743 if (init_qp_attr == NULL) {
744 CERROR("Can't allocate qp_attr for %s\n",
745 libcfs_nid2str(peer->ibp_nid));
749 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
751 CERROR("Can't allocate connection for %s\n",
752 libcfs_nid2str(peer->ibp_nid));
756 conn->ibc_state = IBLND_CONN_INIT;
757 conn->ibc_version = version;
758 conn->ibc_peer = peer; /* I take the caller's ref */
759 cmid->context = conn; /* for future CM callbacks */
760 conn->ibc_cmid = cmid;
762 INIT_LIST_HEAD(&conn->ibc_early_rxs);
763 INIT_LIST_HEAD(&conn->ibc_tx_noops);
764 INIT_LIST_HEAD(&conn->ibc_tx_queue);
765 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
766 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
767 INIT_LIST_HEAD(&conn->ibc_active_txs);
768 spin_lock_init(&conn->ibc_lock);
770 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
771 sizeof(*conn->ibc_connvars));
772 if (conn->ibc_connvars == NULL) {
773 CERROR("Can't allocate in-progress connection state\n");
777 write_lock_irqsave(glock, flags);
778 if (dev->ibd_failover) {
779 write_unlock_irqrestore(glock, flags);
780 CERROR("%s: failover in progress\n", dev->ibd_ifname);
784 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
785 /* wakeup failover thread and teardown connection */
786 if (kiblnd_dev_can_failover(dev)) {
787 list_add_tail(&dev->ibd_fail_list,
788 &kiblnd_data.kib_failed_devs);
789 wake_up(&kiblnd_data.kib_failover_waitq);
792 write_unlock_irqrestore(glock, flags);
793 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
794 cmid->device->name, dev->ibd_ifname);
798 kiblnd_hdev_addref_locked(dev->ibd_hdev);
799 conn->ibc_hdev = dev->ibd_hdev;
801 kiblnd_setup_mtu_locked(cmid);
803 write_unlock_irqrestore(glock, flags);
805 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
806 IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
807 if (conn->ibc_rxs == NULL) {
808 CERROR("Cannot allocate RX buffers\n");
812 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
813 IBLND_RX_MSG_PAGES(version));
817 kiblnd_map_rx_descs(conn);
819 #ifdef HAVE_IB_CQ_INIT_ATTR
820 cq_attr.cqe = IBLND_CQ_ENTRIES(version);
821 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
822 cq = ib_create_cq(cmid->device,
823 kiblnd_cq_completion, kiblnd_cq_event, conn,
826 cq = ib_create_cq(cmid->device,
827 kiblnd_cq_completion, kiblnd_cq_event, conn,
828 IBLND_CQ_ENTRIES(version),
829 kiblnd_get_completion_vector(conn, cpt));
832 CERROR("Can't create CQ: %ld, cqe: %d\n",
833 PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
839 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
841 CERROR("Can't request completion notification: %d\n", rc);
845 init_qp_attr->event_handler = kiblnd_qp_event;
846 init_qp_attr->qp_context = conn;
847 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
848 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
849 init_qp_attr->cap.max_send_sge = 1;
850 init_qp_attr->cap.max_recv_sge = 1;
851 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
852 init_qp_attr->qp_type = IB_QPT_RC;
853 init_qp_attr->send_cq = cq;
854 init_qp_attr->recv_cq = cq;
856 conn->ibc_sched = sched;
858 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
860 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
861 rc, init_qp_attr->cap.max_send_wr,
862 init_qp_attr->cap.max_recv_wr);
866 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
868 /* 1 ref for caller and each rxmsg */
869 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
870 conn->ibc_nrx = IBLND_RX_MSGS(version);
873 for (i = 0; i < IBLND_RX_MSGS(version); i++) {
874 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
875 IBLND_POSTRX_NO_CREDIT);
877 CERROR("Can't post rxmsg: %d\n", rc);
879 /* Make posted receives complete */
880 kiblnd_abort_receives(conn);
882 /* correct # of posted buffers
883 * NB locking needed now I'm racing with completion */
884 spin_lock_irqsave(&sched->ibs_lock, flags);
885 conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
886 spin_unlock_irqrestore(&sched->ibs_lock, flags);
888 /* cmid will be destroyed by CM(ofed) after cm_callback
889 * returned, so we can't refer it anymore
890 * (by kiblnd_connd()->kiblnd_destroy_conn) */
891 rdma_destroy_qp(conn->ibc_cmid);
892 conn->ibc_cmid = NULL;
894 /* Drop my own and unused rxbuffer refcounts */
895 while (i++ <= IBLND_RX_MSGS(version))
896 kiblnd_conn_decref(conn);
902 /* Init successful! */
903 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
904 state == IBLND_CONN_PASSIVE_WAIT);
905 conn->ibc_state = state;
908 atomic_inc(&net->ibn_nconns);
912 kiblnd_destroy_conn(conn);
914 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
920 kiblnd_destroy_conn (kib_conn_t *conn)
922 struct rdma_cm_id *cmid = conn->ibc_cmid;
923 kib_peer_t *peer = conn->ibc_peer;
926 LASSERT (!in_interrupt());
927 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
928 LASSERT(list_empty(&conn->ibc_early_rxs));
929 LASSERT(list_empty(&conn->ibc_tx_noops));
930 LASSERT(list_empty(&conn->ibc_tx_queue));
931 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
932 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
933 LASSERT(list_empty(&conn->ibc_active_txs));
934 LASSERT (conn->ibc_noops_posted == 0);
935 LASSERT (conn->ibc_nsends_posted == 0);
937 switch (conn->ibc_state) {
939 /* conn must be completely disengaged from the network */
942 case IBLND_CONN_DISCONNECTED:
943 /* connvars should have been freed already */
944 LASSERT (conn->ibc_connvars == NULL);
947 case IBLND_CONN_INIT:
951 /* conn->ibc_cmid might be destroyed by CM already */
952 if (cmid != NULL && cmid->qp != NULL)
953 rdma_destroy_qp(cmid);
955 if (conn->ibc_cq != NULL) {
956 rc = ib_destroy_cq(conn->ibc_cq);
958 CWARN("Error destroying CQ: %d\n", rc);
961 if (conn->ibc_rx_pages != NULL)
962 kiblnd_unmap_rx_descs(conn);
964 if (conn->ibc_rxs != NULL) {
965 LIBCFS_FREE(conn->ibc_rxs,
966 IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
969 if (conn->ibc_connvars != NULL)
970 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
972 if (conn->ibc_hdev != NULL)
973 kiblnd_hdev_decref(conn->ibc_hdev);
975 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
976 if (conn->ibc_state != IBLND_CONN_INIT) {
977 kib_net_t *net = peer->ibp_ni->ni_data;
979 rdma_destroy_id(cmid);
980 atomic_dec(&net->ibn_nconns);
981 if (conn->ibc_conn_race) {
982 if (peer->ibp_accepting == 0 &&
983 !list_empty(&peer->ibp_tx_queue)) {
984 kiblnd_connect_peer(peer);
986 rwlock_t *glock = &kiblnd_data.kib_global_lock;
989 write_lock_irqsave(glock, flags);
990 peer->ibp_connecting--;
991 write_unlock_irqrestore(glock, flags);
994 kiblnd_peer_decref(peer);
997 LIBCFS_FREE(conn, sizeof(*conn));
1001 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
1004 struct list_head *ctmp;
1005 struct list_head *cnxt;
1008 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1009 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1011 CDEBUG(D_NET, "Closing conn -> %s, "
1012 "version: %x, reason: %d\n",
1013 libcfs_nid2str(peer->ibp_nid),
1014 conn->ibc_version, why);
1016 kiblnd_close_conn_locked(conn, why);
1024 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1025 int version, __u64 incarnation)
1028 struct list_head *ctmp;
1029 struct list_head *cnxt;
1032 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1033 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1035 if (conn->ibc_version == version &&
1036 conn->ibc_incarnation == incarnation)
1039 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1040 "incarnation:"LPX64"(%x, "LPX64")\n",
1041 libcfs_nid2str(peer->ibp_nid),
1042 conn->ibc_version, conn->ibc_incarnation,
1043 version, incarnation);
1045 kiblnd_close_conn_locked(conn, -ESTALE);
1053 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1056 struct list_head *ptmp;
1057 struct list_head *pnxt;
1061 unsigned long flags;
1064 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1066 if (nid != LNET_NID_ANY)
1067 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1070 hi = kiblnd_data.kib_peer_hash_size - 1;
1073 for (i = lo; i <= hi; i++) {
1074 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1076 peer = list_entry(ptmp, kib_peer_t, ibp_list);
1077 LASSERT(peer->ibp_connecting > 0 ||
1078 peer->ibp_accepting > 0 ||
1079 !list_empty(&peer->ibp_conns));
1081 if (peer->ibp_ni != ni)
1084 if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1087 count += kiblnd_close_peer_conns_locked(peer, 0);
1091 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1093 /* wildcards always succeed */
1094 if (nid == LNET_NID_ANY)
1097 return (count == 0) ? -ENOENT : 0;
1101 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1103 struct libcfs_ioctl_data *data = arg;
1107 case IOC_LIBCFS_GET_PEER: {
1111 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1113 data->ioc_nid = nid;
1114 data->ioc_count = count;
1118 case IOC_LIBCFS_DEL_PEER: {
1119 rc = kiblnd_del_peer(ni, data->ioc_nid);
1122 case IOC_LIBCFS_GET_CONN: {
1126 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1132 LASSERT (conn->ibc_cmid != NULL);
1133 data->ioc_nid = conn->ibc_peer->ibp_nid;
1134 if (conn->ibc_cmid->route.path_rec == NULL)
1135 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1138 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1139 kiblnd_conn_decref(conn);
1142 case IOC_LIBCFS_CLOSE_CONNECTION: {
1143 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1155 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1157 cfs_time_t last_alive = 0;
1158 cfs_time_t now = cfs_time_current();
1159 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1161 unsigned long flags;
1163 read_lock_irqsave(glock, flags);
1165 peer = kiblnd_find_peer_locked(nid);
1167 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1168 peer->ibp_accepting > 0 ||
1169 !list_empty(&peer->ibp_conns)); /* active conn */
1170 last_alive = peer->ibp_last_alive;
1173 read_unlock_irqrestore(glock, flags);
1175 if (last_alive != 0)
1178 /* peer is not persistent in hash, trigger peer creation
1179 * and connection establishment with a NULL tx */
1181 kiblnd_launch_tx(ni, NULL, nid);
1183 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1184 libcfs_nid2str(nid), peer,
1185 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1190 kiblnd_free_pages(kib_pages_t *p)
1192 int npages = p->ibp_npages;
1195 for (i = 0; i < npages; i++) {
1196 if (p->ibp_pages[i] != NULL)
1197 __free_page(p->ibp_pages[i]);
1200 LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1204 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1209 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1210 offsetof(kib_pages_t, ibp_pages[npages]));
1212 CERROR("Can't allocate descriptor for %d pages\n", npages);
1216 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1217 p->ibp_npages = npages;
1219 for (i = 0; i < npages; i++) {
1220 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1222 if (p->ibp_pages[i] == NULL) {
1223 CERROR("Can't allocate page %d of %d\n", i, npages);
1224 kiblnd_free_pages(p);
1234 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1239 LASSERT (conn->ibc_rxs != NULL);
1240 LASSERT (conn->ibc_hdev != NULL);
1242 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1243 rx = &conn->ibc_rxs[i];
1245 LASSERT (rx->rx_nob >= 0); /* not posted */
1247 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1248 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1250 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1253 kiblnd_free_pages(conn->ibc_rx_pages);
1255 conn->ibc_rx_pages = NULL;
1259 kiblnd_map_rx_descs(kib_conn_t *conn)
1267 for (pg_off = ipg = i = 0;
1268 i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1269 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1270 rx = &conn->ibc_rxs[i];
1273 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1275 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1276 rx->rx_msg, IBLND_MSG_SIZE,
1278 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1280 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1282 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1283 i, rx->rx_msg, rx->rx_msgaddr,
1284 (__u64)(page_to_phys(pg) + pg_off));
1286 pg_off += IBLND_MSG_SIZE;
1287 LASSERT (pg_off <= PAGE_SIZE);
1289 if (pg_off == PAGE_SIZE) {
1292 LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1298 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1300 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1304 LASSERT (tpo->tpo_pool.po_allocated == 0);
1309 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1310 tx = &tpo->tpo_tx_descs[i];
1311 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1312 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1314 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1317 kiblnd_hdev_decref(hdev);
1318 tpo->tpo_hdev = NULL;
1321 static kib_hca_dev_t *
1322 kiblnd_current_hdev(kib_dev_t *dev)
1324 kib_hca_dev_t *hdev;
1325 unsigned long flags;
1328 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1329 while (dev->ibd_failover) {
1330 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1332 CDEBUG(D_NET, "%s: Wait for failover\n",
1334 schedule_timeout(cfs_time_seconds(1) / 100);
1336 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1339 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1340 hdev = dev->ibd_hdev;
1342 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1348 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1350 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1351 kib_pool_t *pool = &tpo->tpo_pool;
1352 kib_net_t *net = pool->po_owner->ps_net;
1360 LASSERT (net != NULL);
1364 /* pre-mapped messages are not bigger than 1 page */
1365 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1367 /* No fancy arithmetic when we do the buffer calculations */
1368 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1370 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1372 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1373 page = txpgs->ibp_pages[ipage];
1374 tx = &tpo->tpo_tx_descs[i];
1376 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1379 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1383 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1385 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1387 list_add(&tx->tx_list, &pool->po_free_list);
1389 page_offset += IBLND_MSG_SIZE;
1390 LASSERT(page_offset <= PAGE_SIZE);
1392 if (page_offset == PAGE_SIZE) {
1395 LASSERT(ipage <= txpgs->ibp_npages);
1401 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1405 LASSERT (hdev->ibh_mrs[0] != NULL);
1407 if (hdev->ibh_nmrs == 1)
1408 return hdev->ibh_mrs[0];
1410 index = addr >> hdev->ibh_mr_shift;
1412 if (index < hdev->ibh_nmrs &&
1413 index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1414 return hdev->ibh_mrs[index];
1420 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1422 struct ib_mr *prev_mr;
1426 LASSERT (hdev->ibh_mrs[0] != NULL);
1428 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1429 *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1432 if (hdev->ibh_nmrs == 1)
1433 return hdev->ibh_mrs[0];
1435 for (i = 0, mr = prev_mr = NULL;
1436 i < rd->rd_nfrags; i++) {
1437 mr = kiblnd_find_dma_mr(hdev,
1438 rd->rd_frags[i].rf_addr,
1439 rd->rd_frags[i].rf_nob);
1440 if (prev_mr == NULL)
1443 if (mr == NULL || prev_mr != mr) {
1444 /* Can't covered by one single MR */
1454 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1456 LASSERT (pool->fpo_map_count == 0);
1458 if (pool->fpo_fmr_pool != NULL)
1459 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1461 if (pool->fpo_hdev != NULL)
1462 kiblnd_hdev_decref(pool->fpo_hdev);
1464 LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1468 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1470 kib_fmr_pool_t *pool;
1472 while (!list_empty(head)) {
1473 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1474 list_del(&pool->fpo_list);
1475 kiblnd_destroy_fmr_pool(pool);
1479 static int kiblnd_fmr_pool_size(int ncpts)
1481 int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1483 return max(IBLND_FMR_POOL, size);
1486 static int kiblnd_fmr_flush_trigger(int ncpts)
1488 int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1490 return max(IBLND_FMR_POOL_FLUSH, size);
1494 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1496 /* FMR pool for RDMA */
1497 kib_dev_t *dev = fps->fps_net->ibn_dev;
1498 kib_fmr_pool_t *fpo;
1499 struct ib_fmr_pool_param param = {
1500 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1501 .page_shift = PAGE_SHIFT,
1502 .access = (IB_ACCESS_LOCAL_WRITE |
1503 IB_ACCESS_REMOTE_WRITE),
1504 .pool_size = fps->fps_pool_size,
1505 .dirty_watermark = fps->fps_flush_trigger,
1506 .flush_function = NULL,
1508 .cache = !!*kiblnd_tunables.kib_fmr_cache};
1511 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1515 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1517 fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, ¶m);
1518 if (IS_ERR(fpo->fpo_fmr_pool)) {
1519 rc = PTR_ERR(fpo->fpo_fmr_pool);
1520 CERROR("Failed to create FMR pool: %d\n", rc);
1522 kiblnd_hdev_decref(fpo->fpo_hdev);
1523 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1527 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1528 fpo->fpo_owner = fps;
1535 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1537 if (fps->fps_net == NULL) /* intialized? */
1540 spin_lock(&fps->fps_lock);
1542 while (!list_empty(&fps->fps_pool_list)) {
1543 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1544 kib_fmr_pool_t, fpo_list);
1545 fpo->fpo_failed = 1;
1546 list_del(&fpo->fpo_list);
1547 if (fpo->fpo_map_count == 0)
1548 list_add(&fpo->fpo_list, zombies);
1550 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1553 spin_unlock(&fps->fps_lock);
1557 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1559 if (fps->fps_net != NULL) { /* initialized? */
1560 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1561 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1566 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1567 int pool_size, int flush_trigger)
1569 kib_fmr_pool_t *fpo;
1572 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1576 fps->fps_pool_size = pool_size;
1577 fps->fps_flush_trigger = flush_trigger;
1578 spin_lock_init(&fps->fps_lock);
1579 INIT_LIST_HEAD(&fps->fps_pool_list);
1580 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1582 rc = kiblnd_create_fmr_pool(fps, &fpo);
1584 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1590 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1592 if (fpo->fpo_map_count != 0) /* still in use */
1594 if (fpo->fpo_failed)
1596 return cfs_time_aftereq(now, fpo->fpo_deadline);
1600 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1602 struct list_head zombies = LIST_HEAD_INIT(zombies);
1603 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1604 kib_fmr_poolset_t *fps = fpo->fpo_owner;
1605 cfs_time_t now = cfs_time_current();
1606 kib_fmr_pool_t *tmp;
1609 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1613 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1617 fmr->fmr_pool = NULL;
1618 fmr->fmr_pfmr = NULL;
1620 spin_lock(&fps->fps_lock);
1621 fpo->fpo_map_count--; /* decref the pool */
1623 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1624 /* the first pool is persistent */
1625 if (fps->fps_pool_list.next == &fpo->fpo_list)
1628 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1629 list_move(&fpo->fpo_list, &zombies);
1633 spin_unlock(&fps->fps_lock);
1635 if (!list_empty(&zombies))
1636 kiblnd_destroy_fmr_pool_list(&zombies);
1640 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1641 __u64 iov, kib_fmr_t *fmr)
1643 struct ib_pool_fmr *pfmr;
1644 kib_fmr_pool_t *fpo;
1649 spin_lock(&fps->fps_lock);
1650 version = fps->fps_version;
1651 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1652 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1653 fpo->fpo_map_count++;
1654 spin_unlock(&fps->fps_lock);
1656 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1657 pages, npages, iov);
1658 if (likely(!IS_ERR(pfmr))) {
1659 fmr->fmr_pool = fpo;
1660 fmr->fmr_pfmr = pfmr;
1664 spin_lock(&fps->fps_lock);
1665 fpo->fpo_map_count--;
1666 if (PTR_ERR(pfmr) != -EAGAIN) {
1667 spin_unlock(&fps->fps_lock);
1668 return PTR_ERR(pfmr);
1671 /* EAGAIN and ... */
1672 if (version != fps->fps_version) {
1673 spin_unlock(&fps->fps_lock);
1678 if (fps->fps_increasing) {
1679 spin_unlock(&fps->fps_lock);
1680 CDEBUG(D_NET, "Another thread is allocating new "
1681 "FMR pool, waiting for her to complete\n");
1687 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1688 /* someone failed recently */
1689 spin_unlock(&fps->fps_lock);
1693 fps->fps_increasing = 1;
1694 spin_unlock(&fps->fps_lock);
1696 CDEBUG(D_NET, "Allocate new FMR pool\n");
1697 rc = kiblnd_create_fmr_pool(fps, &fpo);
1698 spin_lock(&fps->fps_lock);
1699 fps->fps_increasing = 0;
1702 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1704 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1706 spin_unlock(&fps->fps_lock);
1712 kiblnd_fini_pool(kib_pool_t *pool)
1714 LASSERT(list_empty(&pool->po_free_list));
1715 LASSERT(pool->po_allocated == 0);
1717 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1721 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1723 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1725 memset(pool, 0, sizeof(kib_pool_t));
1726 INIT_LIST_HEAD(&pool->po_free_list);
1727 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1728 pool->po_owner = ps;
1729 pool->po_size = size;
1733 kiblnd_destroy_pool_list(struct list_head *head)
1737 while (!list_empty(head)) {
1738 pool = list_entry(head->next, kib_pool_t, po_list);
1739 list_del(&pool->po_list);
1741 LASSERT(pool->po_owner != NULL);
1742 pool->po_owner->ps_pool_destroy(pool);
1747 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1749 if (ps->ps_net == NULL) /* intialized? */
1752 spin_lock(&ps->ps_lock);
1753 while (!list_empty(&ps->ps_pool_list)) {
1754 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1755 kib_pool_t, po_list);
1757 list_del(&po->po_list);
1758 if (po->po_allocated == 0)
1759 list_add(&po->po_list, zombies);
1761 list_add(&po->po_list, &ps->ps_failed_pool_list);
1763 spin_unlock(&ps->ps_lock);
1767 kiblnd_fini_poolset(kib_poolset_t *ps)
1769 if (ps->ps_net != NULL) { /* initialized? */
1770 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1771 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1776 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1777 kib_net_t *net, char *name, int size,
1778 kib_ps_pool_create_t po_create,
1779 kib_ps_pool_destroy_t po_destroy,
1780 kib_ps_node_init_t nd_init,
1781 kib_ps_node_fini_t nd_fini)
1786 memset(ps, 0, sizeof(kib_poolset_t));
1790 ps->ps_pool_create = po_create;
1791 ps->ps_pool_destroy = po_destroy;
1792 ps->ps_node_init = nd_init;
1793 ps->ps_node_fini = nd_fini;
1794 ps->ps_pool_size = size;
1795 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1796 >= sizeof(ps->ps_name))
1798 spin_lock_init(&ps->ps_lock);
1799 INIT_LIST_HEAD(&ps->ps_pool_list);
1800 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1802 rc = ps->ps_pool_create(ps, size, &pool);
1804 list_add(&pool->po_list, &ps->ps_pool_list);
1806 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1812 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1814 if (pool->po_allocated != 0) /* still in use */
1816 if (pool->po_failed)
1818 return cfs_time_aftereq(now, pool->po_deadline);
1822 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1824 struct list_head zombies = LIST_HEAD_INIT(zombies);
1825 kib_poolset_t *ps = pool->po_owner;
1827 cfs_time_t now = cfs_time_current();
1829 spin_lock(&ps->ps_lock);
1831 if (ps->ps_node_fini != NULL)
1832 ps->ps_node_fini(pool, node);
1834 LASSERT(pool->po_allocated > 0);
1835 list_add(node, &pool->po_free_list);
1836 pool->po_allocated--;
1838 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1839 /* the first pool is persistent */
1840 if (ps->ps_pool_list.next == &pool->po_list)
1843 if (kiblnd_pool_is_idle(pool, now))
1844 list_move(&pool->po_list, &zombies);
1846 spin_unlock(&ps->ps_lock);
1848 if (!list_empty(&zombies))
1849 kiblnd_destroy_pool_list(&zombies);
1853 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1855 struct list_head *node;
1860 spin_lock(&ps->ps_lock);
1861 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1862 if (list_empty(&pool->po_free_list))
1865 pool->po_allocated++;
1866 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1867 node = pool->po_free_list.next;
1870 if (ps->ps_node_init != NULL) {
1871 /* still hold the lock */
1872 ps->ps_node_init(pool, node);
1874 spin_unlock(&ps->ps_lock);
1878 /* no available tx pool and ... */
1879 if (ps->ps_increasing) {
1880 /* another thread is allocating a new pool */
1881 spin_unlock(&ps->ps_lock);
1882 CDEBUG(D_NET, "Another thread is allocating new "
1883 "%s pool, waiting for her to complete\n",
1889 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1890 /* someone failed recently */
1891 spin_unlock(&ps->ps_lock);
1895 ps->ps_increasing = 1;
1896 spin_unlock(&ps->ps_lock);
1898 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1900 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1902 spin_lock(&ps->ps_lock);
1903 ps->ps_increasing = 0;
1905 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1907 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1908 CERROR("Can't allocate new %s pool because out of memory\n",
1911 spin_unlock(&ps->ps_lock);
1917 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1919 kib_pmr_pool_t *ppo = pmr->pmr_pool;
1920 struct ib_mr *mr = pmr->pmr_mr;
1923 kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1929 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1930 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1933 struct list_head *node;
1937 node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1939 CERROR("Failed to allocate PMR descriptor\n");
1943 pmr = container_of(node, kib_phys_mr_t, pmr_list);
1944 if (pmr->pmr_pool->ppo_hdev != hdev) {
1945 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1949 for (i = 0; i < rd->rd_nfrags; i ++) {
1950 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1951 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1954 pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1955 pmr->pmr_ipb, rd->rd_nfrags,
1956 IB_ACCESS_LOCAL_WRITE |
1957 IB_ACCESS_REMOTE_WRITE,
1959 if (!IS_ERR(pmr->pmr_mr)) {
1960 pmr->pmr_iova = *iova;
1965 rc = PTR_ERR(pmr->pmr_mr);
1966 CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1969 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1975 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1977 kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1980 LASSERT (pool->po_allocated == 0);
1982 while (!list_empty(&pool->po_free_list)) {
1983 pmr = list_entry(pool->po_free_list.next,
1984 kib_phys_mr_t, pmr_list);
1986 LASSERT (pmr->pmr_mr == NULL);
1987 list_del(&pmr->pmr_list);
1989 if (pmr->pmr_ipb != NULL) {
1990 LIBCFS_FREE(pmr->pmr_ipb,
1991 IBLND_MAX_RDMA_FRAGS *
1992 sizeof(struct ib_phys_buf));
1995 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1998 kiblnd_fini_pool(pool);
1999 if (ppo->ppo_hdev != NULL)
2000 kiblnd_hdev_decref(ppo->ppo_hdev);
2002 LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
2005 static inline int kiblnd_pmr_pool_size(int ncpts)
2007 int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
2009 return max(IBLND_PMR_POOL, size);
2013 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2015 struct kib_pmr_pool *ppo;
2016 struct kib_pool *pool;
2020 LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2021 ps->ps_cpt, sizeof(kib_pmr_pool_t));
2023 CERROR("Failed to allocate PMR pool\n");
2027 pool = &ppo->ppo_pool;
2028 kiblnd_init_pool(ps, pool, size);
2030 for (i = 0; i < size; i++) {
2031 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2032 ps->ps_cpt, sizeof(kib_phys_mr_t));
2036 pmr->pmr_pool = ppo;
2037 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2038 IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2039 if (pmr->pmr_ipb == NULL)
2042 list_add(&pmr->pmr_list, &pool->po_free_list);
2046 ps->ps_pool_destroy(pool);
2050 ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2056 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2058 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2061 LASSERT (pool->po_allocated == 0);
2063 if (tpo->tpo_tx_pages != NULL) {
2064 kiblnd_unmap_tx_pool(tpo);
2065 kiblnd_free_pages(tpo->tpo_tx_pages);
2068 if (tpo->tpo_tx_descs == NULL)
2071 for (i = 0; i < pool->po_size; i++) {
2072 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2074 list_del(&tx->tx_list);
2075 if (tx->tx_pages != NULL)
2076 LIBCFS_FREE(tx->tx_pages,
2078 sizeof(*tx->tx_pages));
2079 if (tx->tx_frags != NULL)
2080 LIBCFS_FREE(tx->tx_frags,
2081 IBLND_MAX_RDMA_FRAGS *
2082 sizeof(*tx->tx_frags));
2083 if (tx->tx_wrq != NULL)
2084 LIBCFS_FREE(tx->tx_wrq,
2085 (1 + IBLND_MAX_RDMA_FRAGS) *
2086 sizeof(*tx->tx_wrq));
2087 if (tx->tx_sge != NULL)
2088 LIBCFS_FREE(tx->tx_sge,
2089 (1 + IBLND_MAX_RDMA_FRAGS) *
2090 sizeof(*tx->tx_sge));
2091 if (tx->tx_rd != NULL)
2092 LIBCFS_FREE(tx->tx_rd,
2093 offsetof(kib_rdma_desc_t,
2094 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2097 LIBCFS_FREE(tpo->tpo_tx_descs,
2098 pool->po_size * sizeof(kib_tx_t));
2100 kiblnd_fini_pool(pool);
2101 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2104 static int kiblnd_tx_pool_size(int ncpts)
2106 int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2108 return max(IBLND_TX_POOL, ntx);
2112 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2119 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2121 CERROR("Failed to allocate TX pool\n");
2125 pool = &tpo->tpo_pool;
2126 kiblnd_init_pool(ps, pool, size);
2127 tpo->tpo_tx_descs = NULL;
2128 tpo->tpo_tx_pages = NULL;
2130 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2131 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2132 CERROR("Can't allocate tx pages: %d\n", npg);
2133 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2137 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2138 size * sizeof(kib_tx_t));
2139 if (tpo->tpo_tx_descs == NULL) {
2140 CERROR("Can't allocate %d tx descriptors\n", size);
2141 ps->ps_pool_destroy(pool);
2145 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2147 for (i = 0; i < size; i++) {
2148 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2151 if (ps->ps_net->ibn_fmr_ps != NULL) {
2152 LIBCFS_CPT_ALLOC(tx->tx_pages,
2153 lnet_cpt_table(), ps->ps_cpt,
2154 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2155 if (tx->tx_pages == NULL)
2159 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2160 IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2161 if (tx->tx_frags == NULL)
2164 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2166 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2167 (1 + IBLND_MAX_RDMA_FRAGS) *
2168 sizeof(*tx->tx_wrq));
2169 if (tx->tx_wrq == NULL)
2172 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2173 (1 + IBLND_MAX_RDMA_FRAGS) *
2174 sizeof(*tx->tx_sge));
2175 if (tx->tx_sge == NULL)
2178 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2179 offsetof(kib_rdma_desc_t,
2180 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2181 if (tx->tx_rd == NULL)
2186 kiblnd_map_tx_pool(tpo);
2191 ps->ps_pool_destroy(pool);
2196 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2198 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2200 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
2202 tx->tx_cookie = tps->tps_next_tx_cookie++;
2206 kiblnd_net_fini_pools(kib_net_t *net)
2210 cfs_cpt_for_each(i, lnet_cpt_table()) {
2211 kib_tx_poolset_t *tps;
2212 kib_fmr_poolset_t *fps;
2213 kib_pmr_poolset_t *pps;
2215 if (net->ibn_tx_ps != NULL) {
2216 tps = net->ibn_tx_ps[i];
2217 kiblnd_fini_poolset(&tps->tps_poolset);
2220 if (net->ibn_fmr_ps != NULL) {
2221 fps = net->ibn_fmr_ps[i];
2222 kiblnd_fini_fmr_poolset(fps);
2225 if (net->ibn_pmr_ps != NULL) {
2226 pps = net->ibn_pmr_ps[i];
2227 kiblnd_fini_poolset(&pps->pps_poolset);
2231 if (net->ibn_tx_ps != NULL) {
2232 cfs_percpt_free(net->ibn_tx_ps);
2233 net->ibn_tx_ps = NULL;
2236 if (net->ibn_fmr_ps != NULL) {
2237 cfs_percpt_free(net->ibn_fmr_ps);
2238 net->ibn_fmr_ps = NULL;
2241 if (net->ibn_pmr_ps != NULL) {
2242 cfs_percpt_free(net->ibn_pmr_ps);
2243 net->ibn_pmr_ps = NULL;
2248 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2250 unsigned long flags;
2255 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2256 if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2257 net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2258 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2260 goto create_tx_pool;
2263 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2265 if (*kiblnd_tunables.kib_fmr_pool_size <
2266 *kiblnd_tunables.kib_ntx / 4) {
2267 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2268 *kiblnd_tunables.kib_fmr_pool_size,
2269 *kiblnd_tunables.kib_ntx / 4);
2274 /* TX pool must be created later than FMR/PMR, see LU-2268
2276 LASSERT(net->ibn_tx_ps == NULL);
2278 /* premapping can fail if ibd_nmr > 1, so we always create
2279 * FMR/PMR pool and map-on-demand if premapping failed */
2281 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2282 sizeof(kib_fmr_poolset_t));
2283 if (net->ibn_fmr_ps == NULL) {
2284 CERROR("Failed to allocate FMR pool array\n");
2289 for (i = 0; i < ncpts; i++) {
2290 cpt = (cpts == NULL) ? i : cpts[i];
2291 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2292 kiblnd_fmr_pool_size(ncpts),
2293 kiblnd_fmr_flush_trigger(ncpts));
2294 if (rc == -ENOSYS && i == 0) /* no FMR */
2295 break; /* create PMR pool */
2297 if (rc != 0) { /* a real error */
2298 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2305 LASSERT(i == ncpts);
2306 goto create_tx_pool;
2309 cfs_percpt_free(net->ibn_fmr_ps);
2310 net->ibn_fmr_ps = NULL;
2312 CWARN("Device does not support FMR, failing back to PMR\n");
2314 if (*kiblnd_tunables.kib_pmr_pool_size <
2315 *kiblnd_tunables.kib_ntx / 4) {
2316 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2317 *kiblnd_tunables.kib_pmr_pool_size,
2318 *kiblnd_tunables.kib_ntx / 4);
2323 net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2324 sizeof(kib_pmr_poolset_t));
2325 if (net->ibn_pmr_ps == NULL) {
2326 CERROR("Failed to allocate PMR pool array\n");
2331 for (i = 0; i < ncpts; i++) {
2332 cpt = (cpts == NULL) ? i : cpts[i];
2333 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2335 kiblnd_pmr_pool_size(ncpts),
2336 kiblnd_create_pmr_pool,
2337 kiblnd_destroy_pmr_pool, NULL, NULL);
2339 CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2346 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2347 sizeof(kib_tx_poolset_t));
2348 if (net->ibn_tx_ps == NULL) {
2349 CERROR("Failed to allocate tx pool array\n");
2354 for (i = 0; i < ncpts; i++) {
2355 cpt = (cpts == NULL) ? i : cpts[i];
2356 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2358 kiblnd_tx_pool_size(ncpts),
2359 kiblnd_create_tx_pool,
2360 kiblnd_destroy_tx_pool,
2361 kiblnd_tx_init, NULL);
2363 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2371 kiblnd_net_fini_pools(net);
2377 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2379 struct ib_device_attr *attr;
2382 /* It's safe to assume a HCA can handle a page size
2383 * matching that of the native system */
2384 hdev->ibh_page_shift = PAGE_SHIFT;
2385 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2386 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2388 LIBCFS_ALLOC(attr, sizeof(*attr));
2390 CERROR("Out of memory\n");
2394 rc = ib_query_device(hdev->ibh_ibdev, attr);
2396 hdev->ibh_mr_size = attr->max_mr_size;
2398 LIBCFS_FREE(attr, sizeof(*attr));
2401 CERROR("Failed to query IB device: %d\n", rc);
2405 if (hdev->ibh_mr_size == ~0ULL) {
2406 hdev->ibh_mr_shift = 64;
2410 for (hdev->ibh_mr_shift = 0;
2411 hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2412 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2413 hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2417 CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2422 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2426 if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2429 for (i = 0; i < hdev->ibh_nmrs; i++) {
2430 if (hdev->ibh_mrs[i] == NULL)
2433 ib_dereg_mr(hdev->ibh_mrs[i]);
2436 LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2437 hdev->ibh_mrs = NULL;
2442 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2444 kiblnd_hdev_cleanup_mrs(hdev);
2446 if (hdev->ibh_pd != NULL)
2447 ib_dealloc_pd(hdev->ibh_pd);
2449 if (hdev->ibh_cmid != NULL)
2450 rdma_destroy_id(hdev->ibh_cmid);
2452 LIBCFS_FREE(hdev, sizeof(*hdev));
2456 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2463 int acflags = IB_ACCESS_LOCAL_WRITE |
2464 IB_ACCESS_REMOTE_WRITE;
2466 rc = kiblnd_hdev_get_attr(hdev);
2470 if (hdev->ibh_mr_shift == 64) {
2471 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2472 if (hdev->ibh_mrs == NULL) {
2473 CERROR("Failed to allocate MRs table\n");
2477 hdev->ibh_mrs[0] = NULL;
2480 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2482 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2483 kiblnd_hdev_cleanup_mrs(hdev);
2487 hdev->ibh_mrs[0] = mr;
2492 mr_size = (1ULL << hdev->ibh_mr_shift);
2493 mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2495 hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2497 if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2498 /* it's 4T..., assume we will re-code at that time */
2499 CERROR("Can't support memory size: x"LPX64
2500 " with MR size: x"LPX64"\n", mm_size, mr_size);
2504 /* create an array of MRs to cover all memory */
2505 LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2506 if (hdev->ibh_mrs == NULL) {
2507 CERROR("Failed to allocate MRs' table\n");
2511 memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2513 for (i = 0; i < hdev->ibh_nmrs; i++) {
2514 struct ib_phys_buf ipb;
2517 ipb.size = hdev->ibh_mr_size;
2518 ipb.addr = i * mr_size;
2521 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2523 CERROR("Failed ib_reg_phys_mr addr "LPX64
2524 " size "LPX64" : %ld\n",
2525 ipb.addr, ipb.size, PTR_ERR(mr));
2526 kiblnd_hdev_cleanup_mrs(hdev);
2530 LASSERT (iova == ipb.addr);
2532 hdev->ibh_mrs[i] = mr;
2536 if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2537 LCONSOLE_INFO("Register global MR array, MR size: "
2538 LPX64", array size: %d\n",
2539 hdev->ibh_mr_size, hdev->ibh_nmrs);
2544 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2550 kiblnd_dev_need_failover(kib_dev_t *dev)
2552 struct rdma_cm_id *cmid;
2553 struct sockaddr_in srcaddr;
2554 struct sockaddr_in dstaddr;
2557 if (dev->ibd_hdev == NULL || /* initializing */
2558 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2559 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2562 /* XXX: it's UGLY, but I don't have better way to find
2563 * ib-bonding HCA failover because:
2565 * a. no reliable CM event for HCA failover...
2566 * b. no OFED API to get ib_device for current net_device...
2568 * We have only two choices at this point:
2570 * a. rdma_bind_addr(), it will conflict with listener cmid
2571 * b. rdma_resolve_addr() to zero addr */
2572 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2576 CERROR("Failed to create cmid for failover: %d\n", rc);
2580 memset(&srcaddr, 0, sizeof(srcaddr));
2581 srcaddr.sin_family = AF_INET;
2582 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2584 memset(&dstaddr, 0, sizeof(dstaddr));
2585 dstaddr.sin_family = AF_INET;
2586 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2587 (struct sockaddr *)&dstaddr, 1);
2588 if (rc != 0 || cmid->device == NULL) {
2589 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2590 dev->ibd_ifname, &dev->ibd_ifip,
2592 rdma_destroy_id(cmid);
2596 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2597 rdma_destroy_id(cmid);
2602 kiblnd_dev_failover(kib_dev_t *dev)
2604 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2605 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2606 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2607 struct rdma_cm_id *cmid = NULL;
2608 kib_hca_dev_t *hdev = NULL;
2612 struct sockaddr_in addr;
2613 unsigned long flags;
2617 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2618 dev->ibd_can_failover ||
2619 dev->ibd_hdev == NULL);
2621 rc = kiblnd_dev_need_failover(dev);
2625 if (dev->ibd_hdev != NULL &&
2626 dev->ibd_hdev->ibh_cmid != NULL) {
2627 /* XXX it's not good to close old listener at here,
2628 * because we can fail to create new listener.
2629 * But we have to close it now, otherwise rdma_bind_addr
2630 * will return EADDRINUSE... How crap! */
2631 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2633 cmid = dev->ibd_hdev->ibh_cmid;
2634 /* make next schedule of kiblnd_dev_need_failover()
2635 * return 1 for me */
2636 dev->ibd_hdev->ibh_cmid = NULL;
2637 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2639 rdma_destroy_id(cmid);
2642 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2646 CERROR("Failed to create cmid for failover: %d\n", rc);
2650 memset(&addr, 0, sizeof(addr));
2651 addr.sin_family = AF_INET;
2652 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2653 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2655 /* Bind to failover device or port */
2656 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2657 if (rc != 0 || cmid->device == NULL) {
2658 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2659 dev->ibd_ifname, &dev->ibd_ifip,
2661 rdma_destroy_id(cmid);
2665 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2667 CERROR("Failed to allocate kib_hca_dev\n");
2668 rdma_destroy_id(cmid);
2673 atomic_set(&hdev->ibh_ref, 1);
2674 hdev->ibh_dev = dev;
2675 hdev->ibh_cmid = cmid;
2676 hdev->ibh_ibdev = cmid->device;
2678 pd = ib_alloc_pd(cmid->device);
2681 CERROR("Can't allocate PD: %d\n", rc);
2687 rc = rdma_listen(cmid, 0);
2689 CERROR("Can't start new listener: %d\n", rc);
2693 rc = kiblnd_hdev_setup_mrs(hdev);
2695 CERROR("Can't setup device: %d\n", rc);
2699 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2701 old = dev->ibd_hdev;
2702 dev->ibd_hdev = hdev; /* take over the refcount */
2705 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2706 cfs_cpt_for_each(i, lnet_cpt_table()) {
2707 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2710 if (net->ibn_fmr_ps != NULL) {
2711 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2714 } else if (net->ibn_pmr_ps != NULL) {
2715 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2716 pps_poolset, &zombie_ppo);
2721 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2723 if (!list_empty(&zombie_tpo))
2724 kiblnd_destroy_pool_list(&zombie_tpo);
2725 if (!list_empty(&zombie_ppo))
2726 kiblnd_destroy_pool_list(&zombie_ppo);
2727 if (!list_empty(&zombie_fpo))
2728 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2730 kiblnd_hdev_decref(hdev);
2733 dev->ibd_failed_failover++;
2735 dev->ibd_failed_failover = 0;
2741 kiblnd_destroy_dev (kib_dev_t *dev)
2743 LASSERT (dev->ibd_nnets == 0);
2744 LASSERT(list_empty(&dev->ibd_nets));
2746 list_del(&dev->ibd_fail_list);
2747 list_del(&dev->ibd_list);
2749 if (dev->ibd_hdev != NULL)
2750 kiblnd_hdev_decref(dev->ibd_hdev);
2752 LIBCFS_FREE(dev, sizeof(*dev));
2756 kiblnd_create_dev(char *ifname)
2758 struct net_device *netdev;
2765 rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2767 CERROR("Can't query IPoIB interface %s: %d\n",
2773 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2777 LIBCFS_ALLOC(dev, sizeof(*dev));
2781 memset(dev, 0, sizeof(*dev));
2782 netdev = dev_get_by_name(&init_net, ifname);
2783 if (netdev == NULL) {
2784 dev->ibd_can_failover = 0;
2786 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2790 INIT_LIST_HEAD(&dev->ibd_nets);
2791 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2792 INIT_LIST_HEAD(&dev->ibd_fail_list);
2794 strcpy(&dev->ibd_ifname[0], ifname);
2796 /* initialize the device */
2797 rc = kiblnd_dev_failover(dev);
2799 CERROR("Can't initialize device: %d\n", rc);
2800 LIBCFS_FREE(dev, sizeof(*dev));
2804 list_add_tail(&dev->ibd_list,
2805 &kiblnd_data.kib_devs);
2810 kiblnd_base_shutdown(void)
2812 struct kib_sched_info *sched;
2815 LASSERT(list_empty(&kiblnd_data.kib_devs));
2817 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2818 atomic_read(&libcfs_kmemory));
2820 switch (kiblnd_data.kib_init) {
2824 case IBLND_INIT_ALL:
2825 case IBLND_INIT_DATA:
2826 LASSERT (kiblnd_data.kib_peers != NULL);
2827 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2828 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2830 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2831 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2833 /* flag threads to terminate; wake and wait for them to die */
2834 kiblnd_data.kib_shutdown = 1;
2836 /* NB: we really want to stop scheduler threads net by net
2837 * instead of the whole module, this should be improved
2838 * with dynamic configuration LNet */
2839 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2840 wake_up_all(&sched->ibs_waitq);
2842 wake_up_all(&kiblnd_data.kib_connd_waitq);
2843 wake_up_all(&kiblnd_data.kib_failover_waitq);
2846 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2849 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2850 "Waiting for %d threads to terminate\n",
2851 atomic_read(&kiblnd_data.kib_nthreads));
2852 set_current_state(TASK_UNINTERRUPTIBLE);
2853 schedule_timeout(cfs_time_seconds(1));
2858 case IBLND_INIT_NOTHING:
2862 if (kiblnd_data.kib_peers != NULL) {
2863 LIBCFS_FREE(kiblnd_data.kib_peers,
2864 sizeof(struct list_head) *
2865 kiblnd_data.kib_peer_hash_size);
2868 if (kiblnd_data.kib_scheds != NULL)
2869 cfs_percpt_free(kiblnd_data.kib_scheds);
2871 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2872 atomic_read(&libcfs_kmemory));
2874 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2875 module_put(THIS_MODULE);
2879 kiblnd_shutdown (lnet_ni_t *ni)
2881 kib_net_t *net = ni->ni_data;
2882 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2884 unsigned long flags;
2886 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2891 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2892 atomic_read(&libcfs_kmemory));
2894 write_lock_irqsave(g_lock, flags);
2895 net->ibn_shutdown = 1;
2896 write_unlock_irqrestore(g_lock, flags);
2898 switch (net->ibn_init) {
2902 case IBLND_INIT_ALL:
2903 /* nuke all existing peers within this net */
2904 kiblnd_del_peer(ni, LNET_NID_ANY);
2906 /* Wait for all peer state to clean up */
2908 while (atomic_read(&net->ibn_npeers) != 0) {
2911 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2912 "%s: waiting for %d peers to disconnect\n",
2913 libcfs_nid2str(ni->ni_nid),
2914 atomic_read(&net->ibn_npeers));
2915 set_current_state(TASK_UNINTERRUPTIBLE);
2916 schedule_timeout(cfs_time_seconds(1));
2919 kiblnd_net_fini_pools(net);
2921 write_lock_irqsave(g_lock, flags);
2922 LASSERT(net->ibn_dev->ibd_nnets > 0);
2923 net->ibn_dev->ibd_nnets--;
2924 list_del(&net->ibn_list);
2925 write_unlock_irqrestore(g_lock, flags);
2929 case IBLND_INIT_NOTHING:
2930 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2932 if (net->ibn_dev != NULL &&
2933 net->ibn_dev->ibd_nnets == 0)
2934 kiblnd_destroy_dev(net->ibn_dev);
2939 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2940 atomic_read(&libcfs_kmemory));
2942 net->ibn_init = IBLND_INIT_NOTHING;
2945 LIBCFS_FREE(net, sizeof(*net));
2948 if (list_empty(&kiblnd_data.kib_devs))
2949 kiblnd_base_shutdown();
2954 kiblnd_base_startup(void)
2956 struct kib_sched_info *sched;
2960 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2962 try_module_get(THIS_MODULE);
2963 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2965 rwlock_init(&kiblnd_data.kib_global_lock);
2967 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2968 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2970 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2971 LIBCFS_ALLOC(kiblnd_data.kib_peers,
2972 sizeof(struct list_head) *
2973 kiblnd_data.kib_peer_hash_size);
2974 if (kiblnd_data.kib_peers == NULL)
2977 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2978 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2980 spin_lock_init(&kiblnd_data.kib_connd_lock);
2981 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2982 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2983 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2984 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2986 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2988 if (kiblnd_data.kib_scheds == NULL)
2991 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2994 spin_lock_init(&sched->ibs_lock);
2995 INIT_LIST_HEAD(&sched->ibs_conns);
2996 init_waitqueue_head(&sched->ibs_waitq);
2998 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2999 if (*kiblnd_tunables.kib_nscheds > 0) {
3000 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3002 /* max to half of CPUs, another half is reserved for
3003 * upper layer modules */
3004 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3007 sched->ibs_nthreads_max = nthrs;
3011 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3013 /* lists/ptrs/locks initialised */
3014 kiblnd_data.kib_init = IBLND_INIT_DATA;
3015 /*****************************************************/
3017 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3019 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3023 if (*kiblnd_tunables.kib_dev_failover != 0)
3024 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3028 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3032 /* flag everything initialised */
3033 kiblnd_data.kib_init = IBLND_INIT_ALL;
3034 /*****************************************************/
3039 kiblnd_base_shutdown();
3044 kiblnd_start_schedulers(struct kib_sched_info *sched)
3050 if (sched->ibs_nthreads == 0) {
3051 if (*kiblnd_tunables.kib_nscheds > 0) {
3052 nthrs = sched->ibs_nthreads_max;
3054 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3056 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3057 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3060 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3061 /* increase one thread if there is new interface */
3062 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3065 for (i = 0; i < nthrs; i++) {
3068 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3069 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3070 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3071 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3075 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3076 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3080 sched->ibs_nthreads += i;
3085 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3091 for (i = 0; i < ncpts; i++) {
3092 struct kib_sched_info *sched;
3094 cpt = (cpts == NULL) ? i : cpts[i];
3095 sched = kiblnd_data.kib_scheds[cpt];
3097 if (!newdev && sched->ibs_nthreads > 0)
3100 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3102 CERROR("Failed to start scheduler threads for %s\n",
3111 kiblnd_dev_search(char *ifname)
3113 kib_dev_t *alias = NULL;
3118 colon = strchr(ifname, ':');
3119 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3120 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3126 colon2 = strchr(dev->ibd_ifname, ':');
3132 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3144 kiblnd_startup (lnet_ni_t *ni)
3147 kib_dev_t *ibdev = NULL;
3150 unsigned long flags;
3154 LASSERT (ni->ni_lnd == &the_o2iblnd);
3156 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3157 rc = kiblnd_base_startup();
3162 LIBCFS_ALLOC(net, sizeof(*net));
3167 memset(net, 0, sizeof(*net));
3169 do_gettimeofday(&tv);
3170 net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3172 ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
3173 ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
3174 ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
3175 ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3177 if (ni->ni_interfaces[0] != NULL) {
3178 /* Use the IPoIB interface specified in 'networks=' */
3180 CLASSERT (LNET_MAX_INTERFACES > 1);
3181 if (ni->ni_interfaces[1] != NULL) {
3182 CERROR("Multiple interfaces not supported\n");
3186 ifname = ni->ni_interfaces[0];
3188 ifname = *kiblnd_tunables.kib_default_ipif;
3191 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3192 CERROR("IPoIB interface name too long: %s\n", ifname);
3196 ibdev = kiblnd_dev_search(ifname);
3198 newdev = ibdev == NULL;
3199 /* hmm...create kib_dev even for alias */
3200 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3201 ibdev = kiblnd_create_dev(ifname);
3206 net->ibn_dev = ibdev;
3207 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3209 rc = kiblnd_dev_start_threads(ibdev, newdev,
3210 ni->ni_cpts, ni->ni_ncpts);
3214 rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3216 CERROR("Failed to initialize NI pools: %d\n", rc);
3220 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3222 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3223 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3225 net->ibn_init = IBLND_INIT_ALL;
3230 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3231 kiblnd_destroy_dev(ibdev);
3233 kiblnd_shutdown(ni);
3235 CDEBUG(D_NET, "kiblnd_startup failed\n");
3239 static lnd_t the_o2iblnd = {
3240 .lnd_type = O2IBLND,
3241 .lnd_startup = kiblnd_startup,
3242 .lnd_shutdown = kiblnd_shutdown,
3243 .lnd_ctl = kiblnd_ctl,
3244 .lnd_query = kiblnd_query,
3245 .lnd_send = kiblnd_send,
3246 .lnd_recv = kiblnd_recv,
3250 kiblnd_module_fini (void)
3252 lnet_unregister_lnd(&the_o2iblnd);
3253 kiblnd_tunables_fini();
3257 kiblnd_module_init (void)
3261 CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3262 CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3264 CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3267 rc = kiblnd_tunables_init();
3271 lnet_register_lnd(&the_o2iblnd);
3276 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3277 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3278 MODULE_LICENSE("GPL");
3280 module_init(kiblnd_module_init);
3281 module_exit(kiblnd_module_fini);