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, Whamcloud, Inc.
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>
45 .lnd_startup = kiblnd_startup,
46 .lnd_shutdown = kiblnd_shutdown,
47 .lnd_ctl = kiblnd_ctl,
48 .lnd_query = kiblnd_query,
49 .lnd_send = kiblnd_send,
50 .lnd_recv = kiblnd_recv,
53 kib_data_t kiblnd_data;
56 kiblnd_cksum (void *ptr, int nob)
62 sum = ((sum << 1) | (sum >> 31)) + *c++;
64 /* ensure I don't return 0 (== no checksum) */
65 return (sum == 0) ? 1 : sum;
69 kiblnd_msgtype2str(int type)
72 case IBLND_MSG_CONNREQ:
75 case IBLND_MSG_CONNACK:
81 case IBLND_MSG_IMMEDIATE:
84 case IBLND_MSG_PUT_REQ:
87 case IBLND_MSG_PUT_NAK:
90 case IBLND_MSG_PUT_ACK:
93 case IBLND_MSG_PUT_DONE:
96 case IBLND_MSG_GET_REQ:
99 case IBLND_MSG_GET_DONE:
108 kiblnd_msgtype2size(int type)
110 const int hdr_size = offsetof(kib_msg_t, ibm_u);
113 case IBLND_MSG_CONNREQ:
114 case IBLND_MSG_CONNACK:
115 return hdr_size + sizeof(kib_connparams_t);
120 case IBLND_MSG_IMMEDIATE:
121 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
123 case IBLND_MSG_PUT_REQ:
124 return hdr_size + sizeof(kib_putreq_msg_t);
126 case IBLND_MSG_PUT_ACK:
127 return hdr_size + sizeof(kib_putack_msg_t);
129 case IBLND_MSG_GET_REQ:
130 return hdr_size + sizeof(kib_get_msg_t);
132 case IBLND_MSG_PUT_NAK:
133 case IBLND_MSG_PUT_DONE:
134 case IBLND_MSG_GET_DONE:
135 return hdr_size + sizeof(kib_completion_msg_t);
142 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
149 LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
150 msg->ibm_type == IBLND_MSG_PUT_ACK);
152 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
153 &msg->ibm_u.get.ibgm_rd :
154 &msg->ibm_u.putack.ibpam_rd;
157 __swab32s(&rd->rd_key);
158 __swab32s(&rd->rd_nfrags);
163 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
164 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
165 n, IBLND_MAX_RDMA_FRAGS);
169 nob = offsetof (kib_msg_t, ibm_u) +
170 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
172 if (msg->ibm_nob < nob) {
173 CERROR("Short %s: %d(%d)\n",
174 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
181 for (i = 0; i < n; i++) {
182 __swab32s(&rd->rd_frags[i].rf_nob);
183 __swab64s(&rd->rd_frags[i].rf_addr);
190 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
191 int credits, lnet_nid_t dstnid, __u64 dststamp)
193 kib_net_t *net = ni->ni_data;
195 /* CAVEAT EMPTOR! all message fields not set here should have been
196 * initialised previously. */
197 msg->ibm_magic = IBLND_MSG_MAGIC;
198 msg->ibm_version = version;
200 msg->ibm_credits = credits;
203 msg->ibm_srcnid = ni->ni_nid;
204 msg->ibm_srcstamp = net->ibn_incarnation;
205 msg->ibm_dstnid = dstnid;
206 msg->ibm_dststamp = dststamp;
208 if (*kiblnd_tunables.kib_cksum) {
209 /* NB ibm_cksum zero while computing cksum */
210 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
215 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
217 const int hdr_size = offsetof(kib_msg_t, ibm_u);
223 /* 6 bytes are enough to have received magic + version */
225 CERROR("Short message: %d\n", nob);
229 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
231 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
234 CERROR("Bad magic: %08x\n", msg->ibm_magic);
238 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
239 if (version != IBLND_MSG_VERSION &&
240 version != IBLND_MSG_VERSION_1) {
241 CERROR("Bad version: %x\n", version);
245 if (nob < hdr_size) {
246 CERROR("Short message: %d\n", nob);
250 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
252 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
256 /* checksum must be computed with ibm_cksum zero and BEFORE anything
258 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
260 if (msg_cksum != 0 &&
261 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
262 CERROR("Bad checksum\n");
266 msg->ibm_cksum = msg_cksum;
269 /* leave magic unflipped as a clue to peer endianness */
270 msg->ibm_version = version;
271 CLASSERT (sizeof(msg->ibm_type) == 1);
272 CLASSERT (sizeof(msg->ibm_credits) == 1);
273 msg->ibm_nob = msg_nob;
274 __swab64s(&msg->ibm_srcnid);
275 __swab64s(&msg->ibm_srcstamp);
276 __swab64s(&msg->ibm_dstnid);
277 __swab64s(&msg->ibm_dststamp);
280 if (msg->ibm_srcnid == LNET_NID_ANY) {
281 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
285 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
286 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
287 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
291 switch (msg->ibm_type) {
293 CERROR("Unknown message type %x\n", msg->ibm_type);
297 case IBLND_MSG_IMMEDIATE:
298 case IBLND_MSG_PUT_REQ:
301 case IBLND_MSG_PUT_ACK:
302 case IBLND_MSG_GET_REQ:
303 if (kiblnd_unpack_rd(msg, flip))
307 case IBLND_MSG_PUT_NAK:
308 case IBLND_MSG_PUT_DONE:
309 case IBLND_MSG_GET_DONE:
311 __swab32s(&msg->ibm_u.completion.ibcm_status);
314 case IBLND_MSG_CONNREQ:
315 case IBLND_MSG_CONNACK:
317 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
318 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
319 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
327 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
330 kib_net_t *net = ni->ni_data;
331 int cpt = lnet_cpt_of_nid(nid);
334 LASSERT(net != NULL);
335 LASSERT(nid != LNET_NID_ANY);
337 LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
339 CERROR("Cannot allocate peer\n");
343 memset(peer, 0, sizeof(*peer)); /* zero flags etc */
348 peer->ibp_last_alive = 0;
349 cfs_atomic_set(&peer->ibp_refcount, 1); /* 1 ref for caller */
351 CFS_INIT_LIST_HEAD(&peer->ibp_list); /* not in the peer table yet */
352 CFS_INIT_LIST_HEAD(&peer->ibp_conns);
353 CFS_INIT_LIST_HEAD(&peer->ibp_tx_queue);
355 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
357 /* always called with a ref on ni, which prevents ni being shutdown */
358 LASSERT (net->ibn_shutdown == 0);
360 /* npeers only grows with the global lock held */
361 cfs_atomic_inc(&net->ibn_npeers);
363 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
370 kiblnd_destroy_peer (kib_peer_t *peer)
372 kib_net_t *net = peer->ibp_ni->ni_data;
374 LASSERT (net != NULL);
375 LASSERT (cfs_atomic_read(&peer->ibp_refcount) == 0);
376 LASSERT (!kiblnd_peer_active(peer));
377 LASSERT (peer->ibp_connecting == 0);
378 LASSERT (peer->ibp_accepting == 0);
379 LASSERT (cfs_list_empty(&peer->ibp_conns));
380 LASSERT (cfs_list_empty(&peer->ibp_tx_queue));
382 LIBCFS_FREE(peer, sizeof(*peer));
384 /* NB a peer's connections keep a reference on their peer until
385 * they are destroyed, so we can be assured that _all_ state to do
386 * with this peer has been cleaned up when its refcount drops to
388 cfs_atomic_dec(&net->ibn_npeers);
392 kiblnd_find_peer_locked (lnet_nid_t nid)
394 /* the caller is responsible for accounting the additional reference
395 * that this creates */
396 cfs_list_t *peer_list = kiblnd_nid2peerlist(nid);
400 cfs_list_for_each (tmp, peer_list) {
402 peer = cfs_list_entry(tmp, kib_peer_t, ibp_list);
404 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
405 peer->ibp_accepting > 0 ||
406 !cfs_list_empty(&peer->ibp_conns)); /* active conn */
408 if (peer->ibp_nid != nid)
411 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
412 peer, libcfs_nid2str(nid),
413 cfs_atomic_read(&peer->ibp_refcount),
421 kiblnd_unlink_peer_locked (kib_peer_t *peer)
423 LASSERT (cfs_list_empty(&peer->ibp_conns));
425 LASSERT (kiblnd_peer_active(peer));
426 cfs_list_del_init(&peer->ibp_list);
427 /* lose peerlist's ref */
428 kiblnd_peer_decref(peer);
432 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
433 lnet_nid_t *nidp, int *count)
440 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
442 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
444 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
446 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
447 LASSERT (peer->ibp_connecting > 0 ||
448 peer->ibp_accepting > 0 ||
449 !cfs_list_empty(&peer->ibp_conns));
451 if (peer->ibp_ni != ni)
457 *nidp = peer->ibp_nid;
458 *count = cfs_atomic_read(&peer->ibp_refcount);
460 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
466 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
471 kiblnd_del_peer_locked (kib_peer_t *peer)
477 if (cfs_list_empty(&peer->ibp_conns)) {
478 kiblnd_unlink_peer_locked(peer);
480 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
481 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
483 kiblnd_close_conn_locked(conn, 0);
485 /* NB closing peer's last conn unlinked it. */
487 /* NB peer now unlinked; might even be freed if the peer table had the
492 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
494 CFS_LIST_HEAD (zombies);
504 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
506 if (nid != LNET_NID_ANY) {
507 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
510 hi = kiblnd_data.kib_peer_hash_size - 1;
513 for (i = lo; i <= hi; i++) {
514 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
515 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
516 LASSERT (peer->ibp_connecting > 0 ||
517 peer->ibp_accepting > 0 ||
518 !cfs_list_empty(&peer->ibp_conns));
520 if (peer->ibp_ni != ni)
523 if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
526 if (!cfs_list_empty(&peer->ibp_tx_queue)) {
527 LASSERT (cfs_list_empty(&peer->ibp_conns));
529 cfs_list_splice_init(&peer->ibp_tx_queue,
533 kiblnd_del_peer_locked(peer);
534 rc = 0; /* matched something */
538 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
540 kiblnd_txlist_done(ni, &zombies, -EIO);
546 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
555 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
557 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
558 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
560 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
561 LASSERT (peer->ibp_connecting > 0 ||
562 peer->ibp_accepting > 0 ||
563 !cfs_list_empty(&peer->ibp_conns));
565 if (peer->ibp_ni != ni)
568 cfs_list_for_each (ctmp, &peer->ibp_conns) {
572 conn = cfs_list_entry(ctmp, kib_conn_t,
574 kiblnd_conn_addref(conn);
575 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
582 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
587 kiblnd_debug_rx (kib_rx_t *rx)
589 CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n",
590 rx, rx->rx_status, rx->rx_msg->ibm_type,
591 rx->rx_msg->ibm_credits);
595 kiblnd_debug_tx (kib_tx_t *tx)
597 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx "
598 "cookie "LPX64" msg %s%s type %x cred %d\n",
599 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
600 tx->tx_status, tx->tx_deadline, tx->tx_cookie,
601 tx->tx_lntmsg[0] == NULL ? "-" : "!",
602 tx->tx_lntmsg[1] == NULL ? "-" : "!",
603 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
607 kiblnd_debug_conn (kib_conn_t *conn)
612 cfs_spin_lock(&conn->ibc_lock);
614 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
615 cfs_atomic_read(&conn->ibc_refcount), conn,
616 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
617 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
618 conn->ibc_state, conn->ibc_noops_posted,
619 conn->ibc_nsends_posted, conn->ibc_credits,
620 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
621 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
623 CDEBUG(D_CONSOLE, " early_rxs:\n");
624 cfs_list_for_each(tmp, &conn->ibc_early_rxs)
625 kiblnd_debug_rx(cfs_list_entry(tmp, kib_rx_t, rx_list));
627 CDEBUG(D_CONSOLE, " tx_noops:\n");
628 cfs_list_for_each(tmp, &conn->ibc_tx_noops)
629 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
631 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
632 cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred)
633 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
635 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
636 cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
637 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
639 CDEBUG(D_CONSOLE, " tx_queue:\n");
640 cfs_list_for_each(tmp, &conn->ibc_tx_queue)
641 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
643 CDEBUG(D_CONSOLE, " active_txs:\n");
644 cfs_list_for_each(tmp, &conn->ibc_active_txs)
645 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
647 CDEBUG(D_CONSOLE, " rxs:\n");
648 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
649 kiblnd_debug_rx(&conn->ibc_rxs[i]);
651 cfs_spin_unlock(&conn->ibc_lock);
655 kiblnd_translate_mtu(int value)
676 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
680 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
681 if (cmid->route.path_rec == NULL)
684 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
687 cmid->route.path_rec->mtu = mtu;
690 #ifdef HAVE_OFED_IB_COMP_VECTOR
692 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
699 vectors = conn->ibc_cmid->device->num_comp_vectors;
703 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
705 /* hash NID to CPU id in this partition... */
706 off = conn->ibc_peer->ibp_nid % cpus_weight(*mask);
707 for_each_cpu_mask(i, *mask) {
718 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
719 int state, int version)
722 * If the new conn is created successfully it takes over the caller's
723 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
724 * is destroyed. On failure, the caller's ref on 'peer' remains and
725 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
726 * to destroy 'cmid' here since I'm called from the CM which still has
727 * its ref on 'cmid'). */
728 cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock;
729 kib_net_t *net = peer->ibp_ni->ni_data;
730 kib_dev_t *dev = net->ibn_dev;
731 struct ib_qp_init_attr *init_qp_attr;
732 struct kib_sched_info *sched;
740 LASSERT(net != NULL);
741 LASSERT(!cfs_in_interrupt());
743 cpt = lnet_cpt_of_nid(peer->ibp_nid);
744 sched = kiblnd_data.kib_scheds[cpt];
746 LASSERT(sched->ibs_nthreads > 0);
748 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
749 sizeof(*init_qp_attr));
750 if (init_qp_attr == NULL) {
751 CERROR("Can't allocate qp_attr for %s\n",
752 libcfs_nid2str(peer->ibp_nid));
756 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
758 CERROR("Can't allocate connection for %s\n",
759 libcfs_nid2str(peer->ibp_nid));
763 conn->ibc_state = IBLND_CONN_INIT;
764 conn->ibc_version = version;
765 conn->ibc_peer = peer; /* I take the caller's ref */
766 cmid->context = conn; /* for future CM callbacks */
767 conn->ibc_cmid = cmid;
769 CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
770 CFS_INIT_LIST_HEAD(&conn->ibc_tx_noops);
771 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
772 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
773 CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
774 CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
775 cfs_spin_lock_init(&conn->ibc_lock);
777 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
778 sizeof(*conn->ibc_connvars));
779 if (conn->ibc_connvars == NULL) {
780 CERROR("Can't allocate in-progress connection state\n");
784 cfs_write_lock_irqsave(glock, flags);
785 if (dev->ibd_failover) {
786 cfs_write_unlock_irqrestore(glock, flags);
787 CERROR("%s: failover in progress\n", dev->ibd_ifname);
791 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
792 /* wakeup failover thread and teardown connection */
793 if (kiblnd_dev_can_failover(dev)) {
794 cfs_list_add_tail(&dev->ibd_fail_list,
795 &kiblnd_data.kib_failed_devs);
796 cfs_waitq_signal(&kiblnd_data.kib_failover_waitq);
799 cfs_write_unlock_irqrestore(glock, flags);
800 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
801 cmid->device->name, dev->ibd_ifname);
805 kiblnd_hdev_addref_locked(dev->ibd_hdev);
806 conn->ibc_hdev = dev->ibd_hdev;
808 kiblnd_setup_mtu_locked(cmid);
810 cfs_write_unlock_irqrestore(glock, flags);
812 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
813 IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
814 if (conn->ibc_rxs == NULL) {
815 CERROR("Cannot allocate RX buffers\n");
819 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
820 IBLND_RX_MSG_PAGES(version));
824 kiblnd_map_rx_descs(conn);
826 #ifdef HAVE_OFED_IB_COMP_VECTOR
827 cq = ib_create_cq(cmid->device,
828 kiblnd_cq_completion, kiblnd_cq_event, conn,
829 IBLND_CQ_ENTRIES(version),
830 kiblnd_get_completion_vector(conn, cpt));
832 cq = ib_create_cq(cmid->device,
833 kiblnd_cq_completion, kiblnd_cq_event, conn,
834 IBLND_CQ_ENTRIES(version));
837 CERROR("Can't create CQ: %ld, cqe: %d\n",
838 PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
844 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
846 CERROR("Can't request completion notificiation: %d\n", rc);
850 init_qp_attr->event_handler = kiblnd_qp_event;
851 init_qp_attr->qp_context = conn;
852 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
853 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
854 init_qp_attr->cap.max_send_sge = 1;
855 init_qp_attr->cap.max_recv_sge = 1;
856 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
857 init_qp_attr->qp_type = IB_QPT_RC;
858 init_qp_attr->send_cq = cq;
859 init_qp_attr->recv_cq = cq;
861 conn->ibc_sched = sched;
863 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
865 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
866 rc, init_qp_attr->cap.max_send_wr,
867 init_qp_attr->cap.max_recv_wr);
871 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
873 /* 1 ref for caller and each rxmsg */
874 cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
875 conn->ibc_nrx = IBLND_RX_MSGS(version);
878 for (i = 0; i < IBLND_RX_MSGS(version); i++) {
879 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
880 IBLND_POSTRX_NO_CREDIT);
882 CERROR("Can't post rxmsg: %d\n", rc);
884 /* Make posted receives complete */
885 kiblnd_abort_receives(conn);
887 /* correct # of posted buffers
888 * NB locking needed now I'm racing with completion */
889 cfs_spin_lock_irqsave(&sched->ibs_lock, flags);
890 conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
891 cfs_spin_unlock_irqrestore(&sched->ibs_lock, flags);
893 /* cmid will be destroyed by CM(ofed) after cm_callback
894 * returned, so we can't refer it anymore
895 * (by kiblnd_connd()->kiblnd_destroy_conn) */
896 rdma_destroy_qp(conn->ibc_cmid);
897 conn->ibc_cmid = NULL;
899 /* Drop my own and unused rxbuffer refcounts */
900 while (i++ <= IBLND_RX_MSGS(version))
901 kiblnd_conn_decref(conn);
907 /* Init successful! */
908 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
909 state == IBLND_CONN_PASSIVE_WAIT);
910 conn->ibc_state = state;
913 cfs_atomic_inc(&net->ibn_nconns);
917 kiblnd_destroy_conn(conn);
919 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
925 kiblnd_destroy_conn (kib_conn_t *conn)
927 struct rdma_cm_id *cmid = conn->ibc_cmid;
928 kib_peer_t *peer = conn->ibc_peer;
931 LASSERT (!cfs_in_interrupt());
932 LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
933 LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
934 LASSERT (cfs_list_empty(&conn->ibc_tx_noops));
935 LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
936 LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
937 LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
938 LASSERT (cfs_list_empty(&conn->ibc_active_txs));
939 LASSERT (conn->ibc_noops_posted == 0);
940 LASSERT (conn->ibc_nsends_posted == 0);
942 switch (conn->ibc_state) {
944 /* conn must be completely disengaged from the network */
947 case IBLND_CONN_DISCONNECTED:
948 /* connvars should have been freed already */
949 LASSERT (conn->ibc_connvars == NULL);
952 case IBLND_CONN_INIT:
956 /* conn->ibc_cmid might be destroyed by CM already */
957 if (cmid != NULL && cmid->qp != NULL)
958 rdma_destroy_qp(cmid);
960 if (conn->ibc_cq != NULL) {
961 rc = ib_destroy_cq(conn->ibc_cq);
963 CWARN("Error destroying CQ: %d\n", rc);
966 if (conn->ibc_rx_pages != NULL)
967 kiblnd_unmap_rx_descs(conn);
969 if (conn->ibc_rxs != NULL) {
970 LIBCFS_FREE(conn->ibc_rxs,
971 IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
974 if (conn->ibc_connvars != NULL)
975 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
977 if (conn->ibc_hdev != NULL)
978 kiblnd_hdev_decref(conn->ibc_hdev);
980 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
981 if (conn->ibc_state != IBLND_CONN_INIT) {
982 kib_net_t *net = peer->ibp_ni->ni_data;
984 kiblnd_peer_decref(peer);
985 rdma_destroy_id(cmid);
986 cfs_atomic_dec(&net->ibn_nconns);
989 LIBCFS_FREE(conn, sizeof(*conn));
993 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
1000 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
1001 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
1003 CDEBUG(D_NET, "Closing conn -> %s, "
1004 "version: %x, reason: %d\n",
1005 libcfs_nid2str(peer->ibp_nid),
1006 conn->ibc_version, why);
1008 kiblnd_close_conn_locked(conn, why);
1016 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1017 int version, __u64 incarnation)
1024 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
1025 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
1027 if (conn->ibc_version == version &&
1028 conn->ibc_incarnation == incarnation)
1031 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1032 "incarnation:"LPX64"(%x, "LPX64")\n",
1033 libcfs_nid2str(peer->ibp_nid),
1034 conn->ibc_version, conn->ibc_incarnation,
1035 version, incarnation);
1037 kiblnd_close_conn_locked(conn, -ESTALE);
1045 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
1053 unsigned long flags;
1056 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1058 if (nid != LNET_NID_ANY)
1059 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1062 hi = kiblnd_data.kib_peer_hash_size - 1;
1065 for (i = lo; i <= hi; i++) {
1066 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1068 peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1069 LASSERT (peer->ibp_connecting > 0 ||
1070 peer->ibp_accepting > 0 ||
1071 !cfs_list_empty(&peer->ibp_conns));
1073 if (peer->ibp_ni != ni)
1076 if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1079 count += kiblnd_close_peer_conns_locked(peer, 0);
1083 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1085 /* wildcards always succeed */
1086 if (nid == LNET_NID_ANY)
1089 return (count == 0) ? -ENOENT : 0;
1093 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1095 struct libcfs_ioctl_data *data = arg;
1099 case IOC_LIBCFS_GET_PEER: {
1103 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1105 data->ioc_nid = nid;
1106 data->ioc_count = count;
1110 case IOC_LIBCFS_DEL_PEER: {
1111 rc = kiblnd_del_peer(ni, data->ioc_nid);
1114 case IOC_LIBCFS_GET_CONN: {
1118 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1124 LASSERT (conn->ibc_cmid != NULL);
1125 data->ioc_nid = conn->ibc_peer->ibp_nid;
1126 if (conn->ibc_cmid->route.path_rec == NULL)
1127 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1130 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1131 kiblnd_conn_decref(conn);
1134 case IOC_LIBCFS_CLOSE_CONNECTION: {
1135 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1147 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1149 cfs_time_t last_alive = 0;
1150 cfs_time_t now = cfs_time_current();
1151 cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock;
1153 unsigned long flags;
1155 cfs_read_lock_irqsave(glock, flags);
1157 peer = kiblnd_find_peer_locked(nid);
1159 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1160 peer->ibp_accepting > 0 ||
1161 !cfs_list_empty(&peer->ibp_conns)); /* active conn */
1162 last_alive = peer->ibp_last_alive;
1165 cfs_read_unlock_irqrestore(glock, flags);
1167 if (last_alive != 0)
1170 /* peer is not persistent in hash, trigger peer creation
1171 * and connection establishment with a NULL tx */
1173 kiblnd_launch_tx(ni, NULL, nid);
1175 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1176 libcfs_nid2str(nid), peer,
1177 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1182 kiblnd_free_pages(kib_pages_t *p)
1184 int npages = p->ibp_npages;
1187 for (i = 0; i < npages; i++) {
1188 if (p->ibp_pages[i] != NULL)
1189 cfs_free_page(p->ibp_pages[i]);
1192 LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1196 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1201 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1202 offsetof(kib_pages_t, ibp_pages[npages]));
1204 CERROR("Can't allocate descriptor for %d pages\n", npages);
1208 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1209 p->ibp_npages = npages;
1211 for (i = 0; i < npages; i++) {
1212 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1214 if (p->ibp_pages[i] == NULL) {
1215 CERROR("Can't allocate page %d of %d\n", i, npages);
1216 kiblnd_free_pages(p);
1226 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1231 LASSERT (conn->ibc_rxs != NULL);
1232 LASSERT (conn->ibc_hdev != NULL);
1234 for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1235 rx = &conn->ibc_rxs[i];
1237 LASSERT (rx->rx_nob >= 0); /* not posted */
1239 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1240 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1242 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1245 kiblnd_free_pages(conn->ibc_rx_pages);
1247 conn->ibc_rx_pages = NULL;
1251 kiblnd_map_rx_descs(kib_conn_t *conn)
1259 for (pg_off = ipg = i = 0;
1260 i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1261 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1262 rx = &conn->ibc_rxs[i];
1265 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1267 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1268 rx->rx_msg, IBLND_MSG_SIZE,
1270 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1272 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1274 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1275 i, rx->rx_msg, rx->rx_msgaddr,
1276 lnet_page2phys(pg) + pg_off);
1278 pg_off += IBLND_MSG_SIZE;
1279 LASSERT (pg_off <= PAGE_SIZE);
1281 if (pg_off == PAGE_SIZE) {
1284 LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1290 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1292 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1296 LASSERT (tpo->tpo_pool.po_allocated == 0);
1301 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1302 tx = &tpo->tpo_tx_descs[i];
1303 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1304 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1306 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1309 kiblnd_hdev_decref(hdev);
1310 tpo->tpo_hdev = NULL;
1313 static kib_hca_dev_t *
1314 kiblnd_current_hdev(kib_dev_t *dev)
1316 kib_hca_dev_t *hdev;
1317 unsigned long flags;
1320 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1321 while (dev->ibd_failover) {
1322 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1324 CDEBUG(D_NET, "Wait for dev(%s) failover\n", dev->ibd_ifname);
1325 cfs_schedule_timeout(cfs_time_seconds(1) / 100);
1327 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1330 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1331 hdev = dev->ibd_hdev;
1333 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1339 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1341 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1342 kib_pool_t *pool = &tpo->tpo_pool;
1343 kib_net_t *net = pool->po_owner->ps_net;
1344 kib_dev_t *dev = net->ibn_dev;
1351 LASSERT (net != NULL);
1353 /* pre-mapped messages are not bigger than 1 page */
1354 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1356 /* No fancy arithmetic when we do the buffer calculations */
1357 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1359 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1361 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1362 page = txpgs->ibp_pages[ipage];
1363 tx = &tpo->tpo_tx_descs[i];
1365 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1368 tx->tx_msgaddr = kiblnd_dma_map_single(
1369 tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1370 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1371 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1373 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1375 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1377 page_offset += IBLND_MSG_SIZE;
1378 LASSERT (page_offset <= PAGE_SIZE);
1380 if (page_offset == PAGE_SIZE) {
1383 LASSERT (ipage <= txpgs->ibp_npages);
1389 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1393 LASSERT (hdev->ibh_mrs[0] != NULL);
1395 if (hdev->ibh_nmrs == 1)
1396 return hdev->ibh_mrs[0];
1398 index = addr >> hdev->ibh_mr_shift;
1400 if (index < hdev->ibh_nmrs &&
1401 index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1402 return hdev->ibh_mrs[index];
1408 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1410 struct ib_mr *prev_mr;
1414 LASSERT (hdev->ibh_mrs[0] != NULL);
1416 if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1417 *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1420 if (hdev->ibh_nmrs == 1)
1421 return hdev->ibh_mrs[0];
1423 for (i = 0, mr = prev_mr = NULL;
1424 i < rd->rd_nfrags; i++) {
1425 mr = kiblnd_find_dma_mr(hdev,
1426 rd->rd_frags[i].rf_addr,
1427 rd->rd_frags[i].rf_nob);
1428 if (prev_mr == NULL)
1431 if (mr == NULL || prev_mr != mr) {
1432 /* Can't covered by one single MR */
1442 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1444 LASSERT (pool->fpo_map_count == 0);
1446 if (pool->fpo_fmr_pool != NULL)
1447 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1449 if (pool->fpo_hdev != NULL)
1450 kiblnd_hdev_decref(pool->fpo_hdev);
1452 LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1456 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1458 kib_fmr_pool_t *pool;
1460 while (!cfs_list_empty(head)) {
1461 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1462 cfs_list_del(&pool->fpo_list);
1463 kiblnd_destroy_fmr_pool(pool);
1467 static int kiblnd_fmr_pool_size(int ncpts)
1469 int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1471 return max(IBLND_FMR_POOL, size);
1474 static int kiblnd_fmr_flush_trigger(int ncpts)
1476 int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1478 return max(IBLND_FMR_POOL_FLUSH, size);
1482 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1484 /* FMR pool for RDMA */
1485 kib_dev_t *dev = fps->fps_net->ibn_dev;
1486 kib_fmr_pool_t *fpo;
1487 struct ib_fmr_pool_param param = {
1488 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1489 .page_shift = PAGE_SHIFT,
1490 .access = (IB_ACCESS_LOCAL_WRITE |
1491 IB_ACCESS_REMOTE_WRITE),
1492 .pool_size = fps->fps_pool_size,
1493 .dirty_watermark = fps->fps_flush_trigger,
1494 .flush_function = NULL,
1496 .cache = !!*kiblnd_tunables.kib_fmr_cache};
1499 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1503 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1505 fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, ¶m);
1506 if (IS_ERR(fpo->fpo_fmr_pool)) {
1507 rc = PTR_ERR(fpo->fpo_fmr_pool);
1508 CERROR("Failed to create FMR pool: %d\n", rc);
1510 kiblnd_hdev_decref(fpo->fpo_hdev);
1511 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1515 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1516 fpo->fpo_owner = fps;
1523 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, cfs_list_t *zombies)
1525 if (fps->fps_net == NULL) /* intialized? */
1528 cfs_spin_lock(&fps->fps_lock);
1530 while (!cfs_list_empty(&fps->fps_pool_list)) {
1531 kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next,
1532 kib_fmr_pool_t, fpo_list);
1533 fpo->fpo_failed = 1;
1534 cfs_list_del(&fpo->fpo_list);
1535 if (fpo->fpo_map_count == 0)
1536 cfs_list_add(&fpo->fpo_list, zombies);
1538 cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1541 cfs_spin_unlock(&fps->fps_lock);
1545 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1547 if (fps->fps_net != NULL) { /* initialized? */
1548 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1549 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1554 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1555 int pool_size, int flush_trigger)
1557 kib_fmr_pool_t *fpo;
1560 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1564 fps->fps_pool_size = pool_size;
1565 fps->fps_flush_trigger = flush_trigger;
1566 cfs_spin_lock_init(&fps->fps_lock);
1567 CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1568 CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1570 rc = kiblnd_create_fmr_pool(fps, &fpo);
1572 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1578 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1580 if (fpo->fpo_map_count != 0) /* still in use */
1582 if (fpo->fpo_failed)
1584 return cfs_time_aftereq(now, fpo->fpo_deadline);
1588 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1590 CFS_LIST_HEAD (zombies);
1591 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1592 kib_fmr_poolset_t *fps = fpo->fpo_owner;
1593 cfs_time_t now = cfs_time_current();
1594 kib_fmr_pool_t *tmp;
1597 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1601 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1605 fmr->fmr_pool = NULL;
1606 fmr->fmr_pfmr = NULL;
1608 cfs_spin_lock(&fps->fps_lock);
1609 fpo->fpo_map_count --; /* decref the pool */
1611 cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1612 /* the first pool is persistent */
1613 if (fps->fps_pool_list.next == &fpo->fpo_list)
1616 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1617 cfs_list_move(&fpo->fpo_list, &zombies);
1618 fps->fps_version ++;
1621 cfs_spin_unlock(&fps->fps_lock);
1623 if (!cfs_list_empty(&zombies))
1624 kiblnd_destroy_fmr_pool_list(&zombies);
1628 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1629 __u64 iov, kib_fmr_t *fmr)
1631 struct ib_pool_fmr *pfmr;
1632 kib_fmr_pool_t *fpo;
1637 cfs_spin_lock(&fps->fps_lock);
1638 version = fps->fps_version;
1639 cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1640 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1641 fpo->fpo_map_count ++;
1642 cfs_spin_unlock(&fps->fps_lock);
1644 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1645 pages, npages, iov);
1646 if (likely(!IS_ERR(pfmr))) {
1647 fmr->fmr_pool = fpo;
1648 fmr->fmr_pfmr = pfmr;
1652 cfs_spin_lock(&fps->fps_lock);
1653 fpo->fpo_map_count --;
1654 if (PTR_ERR(pfmr) != -EAGAIN) {
1655 cfs_spin_unlock(&fps->fps_lock);
1656 return PTR_ERR(pfmr);
1659 /* EAGAIN and ... */
1660 if (version != fps->fps_version) {
1661 cfs_spin_unlock(&fps->fps_lock);
1666 if (fps->fps_increasing) {
1667 cfs_spin_unlock(&fps->fps_lock);
1668 CDEBUG(D_NET, "Another thread is allocating new "
1669 "FMR pool, waiting for her to complete\n");
1675 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1676 /* someone failed recently */
1677 cfs_spin_unlock(&fps->fps_lock);
1681 fps->fps_increasing = 1;
1682 cfs_spin_unlock(&fps->fps_lock);
1684 CDEBUG(D_NET, "Allocate new FMR pool\n");
1685 rc = kiblnd_create_fmr_pool(fps, &fpo);
1686 cfs_spin_lock(&fps->fps_lock);
1687 fps->fps_increasing = 0;
1689 fps->fps_version ++;
1690 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1692 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1694 cfs_spin_unlock(&fps->fps_lock);
1700 kiblnd_fini_pool(kib_pool_t *pool)
1702 LASSERT (cfs_list_empty(&pool->po_free_list));
1703 LASSERT (pool->po_allocated == 0);
1705 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1709 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1711 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1713 memset(pool, 0, sizeof(kib_pool_t));
1714 CFS_INIT_LIST_HEAD(&pool->po_free_list);
1715 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1716 pool->po_owner = ps;
1717 pool->po_size = size;
1721 kiblnd_destroy_pool_list(cfs_list_t *head)
1725 while (!cfs_list_empty(head)) {
1726 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1727 cfs_list_del(&pool->po_list);
1729 LASSERT (pool->po_owner != NULL);
1730 pool->po_owner->ps_pool_destroy(pool);
1735 kiblnd_fail_poolset(kib_poolset_t *ps, cfs_list_t *zombies)
1737 if (ps->ps_net == NULL) /* intialized? */
1740 cfs_spin_lock(&ps->ps_lock);
1741 while (!cfs_list_empty(&ps->ps_pool_list)) {
1742 kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next,
1743 kib_pool_t, po_list);
1745 cfs_list_del(&po->po_list);
1746 if (po->po_allocated == 0)
1747 cfs_list_add(&po->po_list, zombies);
1749 cfs_list_add(&po->po_list, &ps->ps_failed_pool_list);
1751 cfs_spin_unlock(&ps->ps_lock);
1755 kiblnd_fini_poolset(kib_poolset_t *ps)
1757 if (ps->ps_net != NULL) { /* initialized? */
1758 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1759 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1764 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1765 kib_net_t *net, char *name, int size,
1766 kib_ps_pool_create_t po_create,
1767 kib_ps_pool_destroy_t po_destroy,
1768 kib_ps_node_init_t nd_init,
1769 kib_ps_node_fini_t nd_fini)
1774 memset(ps, 0, sizeof(kib_poolset_t));
1778 ps->ps_pool_create = po_create;
1779 ps->ps_pool_destroy = po_destroy;
1780 ps->ps_node_init = nd_init;
1781 ps->ps_node_fini = nd_fini;
1782 ps->ps_pool_size = size;
1783 strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1784 cfs_spin_lock_init(&ps->ps_lock);
1785 CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1786 CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1788 rc = ps->ps_pool_create(ps, size, &pool);
1790 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1792 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1798 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1800 if (pool->po_allocated != 0) /* still in use */
1802 if (pool->po_failed)
1804 return cfs_time_aftereq(now, pool->po_deadline);
1808 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1810 CFS_LIST_HEAD (zombies);
1811 kib_poolset_t *ps = pool->po_owner;
1813 cfs_time_t now = cfs_time_current();
1815 cfs_spin_lock(&ps->ps_lock);
1817 if (ps->ps_node_fini != NULL)
1818 ps->ps_node_fini(pool, node);
1820 LASSERT (pool->po_allocated > 0);
1821 cfs_list_add(node, &pool->po_free_list);
1822 pool->po_allocated --;
1824 cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1825 /* the first pool is persistent */
1826 if (ps->ps_pool_list.next == &pool->po_list)
1829 if (kiblnd_pool_is_idle(pool, now))
1830 cfs_list_move(&pool->po_list, &zombies);
1832 cfs_spin_unlock(&ps->ps_lock);
1834 if (!cfs_list_empty(&zombies))
1835 kiblnd_destroy_pool_list(&zombies);
1839 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1846 cfs_spin_lock(&ps->ps_lock);
1847 cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1848 if (cfs_list_empty(&pool->po_free_list))
1851 pool->po_allocated ++;
1852 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1853 node = pool->po_free_list.next;
1856 if (ps->ps_node_init != NULL) {
1857 /* still hold the lock */
1858 ps->ps_node_init(pool, node);
1860 cfs_spin_unlock(&ps->ps_lock);
1864 /* no available tx pool and ... */
1865 if (ps->ps_increasing) {
1866 /* another thread is allocating a new pool */
1867 cfs_spin_unlock(&ps->ps_lock);
1868 CDEBUG(D_NET, "Another thread is allocating new "
1869 "%s pool, waiting for her to complete\n",
1875 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1876 /* someone failed recently */
1877 cfs_spin_unlock(&ps->ps_lock);
1881 ps->ps_increasing = 1;
1882 cfs_spin_unlock(&ps->ps_lock);
1884 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1886 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1888 cfs_spin_lock(&ps->ps_lock);
1889 ps->ps_increasing = 0;
1891 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1893 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1894 CERROR("Can't allocate new %s pool because out of memory\n",
1897 cfs_spin_unlock(&ps->ps_lock);
1903 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1905 kib_pmr_pool_t *ppo = pmr->pmr_pool;
1906 struct ib_mr *mr = pmr->pmr_mr;
1909 kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1915 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1916 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1923 node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1925 CERROR("Failed to allocate PMR descriptor\n");
1929 pmr = container_of(node, kib_phys_mr_t, pmr_list);
1930 if (pmr->pmr_pool->ppo_hdev != hdev) {
1931 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1935 for (i = 0; i < rd->rd_nfrags; i ++) {
1936 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1937 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1940 pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1941 pmr->pmr_ipb, rd->rd_nfrags,
1942 IB_ACCESS_LOCAL_WRITE |
1943 IB_ACCESS_REMOTE_WRITE,
1945 if (!IS_ERR(pmr->pmr_mr)) {
1946 pmr->pmr_iova = *iova;
1951 rc = PTR_ERR(pmr->pmr_mr);
1952 CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1955 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1961 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1963 kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1966 LASSERT (pool->po_allocated == 0);
1968 while (!cfs_list_empty(&pool->po_free_list)) {
1969 pmr = cfs_list_entry(pool->po_free_list.next,
1970 kib_phys_mr_t, pmr_list);
1972 LASSERT (pmr->pmr_mr == NULL);
1973 cfs_list_del(&pmr->pmr_list);
1975 if (pmr->pmr_ipb != NULL) {
1976 LIBCFS_FREE(pmr->pmr_ipb,
1977 IBLND_MAX_RDMA_FRAGS *
1978 sizeof(struct ib_phys_buf));
1981 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1984 kiblnd_fini_pool(pool);
1985 if (ppo->ppo_hdev != NULL)
1986 kiblnd_hdev_decref(ppo->ppo_hdev);
1988 LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1991 static inline int kiblnd_pmr_pool_size(int ncpts)
1993 int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1995 return max(IBLND_PMR_POOL, size);
1999 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2001 struct kib_pmr_pool *ppo;
2002 struct kib_pool *pool;
2006 LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2007 ps->ps_cpt, sizeof(kib_pmr_pool_t));
2009 CERROR("Failed to allocate PMR pool\n");
2013 pool = &ppo->ppo_pool;
2014 kiblnd_init_pool(ps, pool, size);
2016 for (i = 0; i < size; i++) {
2017 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2018 ps->ps_cpt, sizeof(kib_phys_mr_t));
2022 pmr->pmr_pool = ppo;
2023 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2024 IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2025 if (pmr->pmr_ipb == NULL)
2028 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
2032 ps->ps_pool_destroy(pool);
2036 ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2042 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2044 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2047 LASSERT (pool->po_allocated == 0);
2049 if (tpo->tpo_tx_pages != NULL) {
2050 kiblnd_unmap_tx_pool(tpo);
2051 kiblnd_free_pages(tpo->tpo_tx_pages);
2054 if (tpo->tpo_tx_descs == NULL)
2057 for (i = 0; i < pool->po_size; i++) {
2058 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2060 cfs_list_del(&tx->tx_list);
2061 if (tx->tx_pages != NULL)
2062 LIBCFS_FREE(tx->tx_pages,
2064 sizeof(*tx->tx_pages));
2065 if (tx->tx_frags != NULL)
2066 LIBCFS_FREE(tx->tx_frags,
2067 IBLND_MAX_RDMA_FRAGS *
2068 sizeof(*tx->tx_frags));
2069 if (tx->tx_wrq != NULL)
2070 LIBCFS_FREE(tx->tx_wrq,
2071 (1 + IBLND_MAX_RDMA_FRAGS) *
2072 sizeof(*tx->tx_wrq));
2073 if (tx->tx_sge != NULL)
2074 LIBCFS_FREE(tx->tx_sge,
2075 (1 + IBLND_MAX_RDMA_FRAGS) *
2076 sizeof(*tx->tx_sge));
2077 if (tx->tx_rd != NULL)
2078 LIBCFS_FREE(tx->tx_rd,
2079 offsetof(kib_rdma_desc_t,
2080 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2083 LIBCFS_FREE(tpo->tpo_tx_descs,
2084 pool->po_size * sizeof(kib_tx_t));
2086 kiblnd_fini_pool(pool);
2087 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2090 static int kiblnd_tx_pool_size(int ncpts)
2092 int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2094 return max(IBLND_TX_POOL, ntx);
2098 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2105 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2107 CERROR("Failed to allocate TX pool\n");
2111 pool = &tpo->tpo_pool;
2112 kiblnd_init_pool(ps, pool, size);
2113 tpo->tpo_tx_descs = NULL;
2114 tpo->tpo_tx_pages = NULL;
2116 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2117 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2118 CERROR("Can't allocate tx pages: %d\n", npg);
2119 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2123 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2124 size * sizeof(kib_tx_t));
2125 if (tpo->tpo_tx_descs == NULL) {
2126 CERROR("Can't allocate %d tx descriptors\n", size);
2127 ps->ps_pool_destroy(pool);
2131 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2133 for (i = 0; i < size; i++) {
2134 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2137 if (ps->ps_net->ibn_fmr_ps != NULL) {
2138 LIBCFS_CPT_ALLOC(tx->tx_pages,
2139 lnet_cpt_table(), ps->ps_cpt,
2140 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2141 if (tx->tx_pages == NULL)
2145 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2146 IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2147 if (tx->tx_frags == NULL)
2150 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2151 (1 + IBLND_MAX_RDMA_FRAGS) *
2152 sizeof(*tx->tx_wrq));
2153 if (tx->tx_wrq == NULL)
2156 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2157 (1 + IBLND_MAX_RDMA_FRAGS) *
2158 sizeof(*tx->tx_sge));
2159 if (tx->tx_sge == NULL)
2162 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2163 offsetof(kib_rdma_desc_t,
2164 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2165 if (tx->tx_rd == NULL)
2170 kiblnd_map_tx_pool(tpo);
2175 ps->ps_pool_destroy(pool);
2180 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
2182 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2184 kib_tx_t *tx = cfs_list_entry(node, kib_tx_t, tx_list);
2186 tx->tx_cookie = tps->tps_next_tx_cookie ++;
2190 kiblnd_net_fini_pools(kib_net_t *net)
2194 cfs_cpt_for_each(i, lnet_cpt_table()) {
2195 kib_tx_poolset_t *tps;
2196 kib_fmr_poolset_t *fps;
2197 kib_pmr_poolset_t *pps;
2199 if (net->ibn_tx_ps != NULL) {
2200 tps = net->ibn_tx_ps[i];
2201 kiblnd_fini_poolset(&tps->tps_poolset);
2204 if (net->ibn_fmr_ps != NULL) {
2205 fps = net->ibn_fmr_ps[i];
2206 kiblnd_fini_fmr_poolset(fps);
2209 if (net->ibn_pmr_ps != NULL) {
2210 pps = net->ibn_pmr_ps[i];
2211 kiblnd_fini_poolset(&pps->pps_poolset);
2215 if (net->ibn_tx_ps != NULL) {
2216 cfs_percpt_free(net->ibn_tx_ps);
2217 net->ibn_tx_ps = NULL;
2220 if (net->ibn_fmr_ps != NULL) {
2221 cfs_percpt_free(net->ibn_fmr_ps);
2222 net->ibn_fmr_ps = NULL;
2225 if (net->ibn_pmr_ps != NULL) {
2226 cfs_percpt_free(net->ibn_pmr_ps);
2227 net->ibn_pmr_ps = NULL;
2232 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2234 unsigned long flags;
2239 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2240 sizeof(kib_tx_poolset_t));
2241 if (net->ibn_tx_ps == NULL) {
2242 CERROR("Failed to allocate tx pool array\n");
2246 for (i = 0; i < ncpts; i++) {
2247 cpt = (cpts == NULL) ? i : cpts[i];
2248 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2250 kiblnd_tx_pool_size(ncpts),
2251 kiblnd_create_tx_pool,
2252 kiblnd_destroy_tx_pool,
2253 kiblnd_tx_init, NULL);
2255 CERROR("Failed to initialize TX pool\n");
2260 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2261 if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2262 net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2263 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2268 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2270 if (*kiblnd_tunables.kib_fmr_pool_size <
2271 *kiblnd_tunables.kib_ntx / 4) {
2272 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2273 *kiblnd_tunables.kib_fmr_pool_size,
2274 *kiblnd_tunables.kib_ntx / 4);
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");
2288 for (i = 0; i < ncpts; i++) {
2289 cpt = (cpts == NULL) ? i : cpts[i];
2290 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2291 kiblnd_fmr_pool_size(ncpts),
2292 kiblnd_fmr_flush_trigger(ncpts));
2293 if (rc == -ENOSYS && i == 0) /* no FMR */
2294 break; /* create PMR pool */
2296 goto failed; /* a real error */
2299 cfs_percpt_free(net->ibn_fmr_ps);
2300 net->ibn_fmr_ps = NULL;
2302 if (*kiblnd_tunables.kib_pmr_pool_size <
2303 *kiblnd_tunables.kib_ntx / 4) {
2304 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2305 *kiblnd_tunables.kib_pmr_pool_size,
2306 *kiblnd_tunables.kib_ntx / 4);
2310 net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2311 sizeof(kib_pmr_poolset_t));
2312 if (net->ibn_pmr_ps == NULL) {
2313 CERROR("Failed to allocate PMR pool array\n");
2317 for (i = 0; i < ncpts; i++) {
2318 cpt = (cpts == NULL) ? i : cpts[i];
2319 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2321 kiblnd_pmr_pool_size(ncpts),
2322 kiblnd_create_pmr_pool,
2323 kiblnd_destroy_pmr_pool, NULL, NULL);
2331 kiblnd_net_fini_pools(net);
2336 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2338 struct ib_device_attr *attr;
2341 /* It's safe to assume a HCA can handle a page size
2342 * matching that of the native system */
2343 hdev->ibh_page_shift = PAGE_SHIFT;
2344 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2345 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2347 LIBCFS_ALLOC(attr, sizeof(*attr));
2349 CERROR("Out of memory\n");
2353 rc = ib_query_device(hdev->ibh_ibdev, attr);
2355 hdev->ibh_mr_size = attr->max_mr_size;
2357 LIBCFS_FREE(attr, sizeof(*attr));
2360 CERROR("Failed to query IB device: %d\n", rc);
2364 if (hdev->ibh_mr_size == ~0ULL) {
2365 hdev->ibh_mr_shift = 64;
2369 for (hdev->ibh_mr_shift = 0;
2370 hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2371 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2372 hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2376 CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2381 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2385 if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2388 for (i = 0; i < hdev->ibh_nmrs; i++) {
2389 if (hdev->ibh_mrs[i] == NULL)
2392 ib_dereg_mr(hdev->ibh_mrs[i]);
2395 LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2396 hdev->ibh_mrs = NULL;
2401 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2403 kiblnd_hdev_cleanup_mrs(hdev);
2405 if (hdev->ibh_pd != NULL)
2406 ib_dealloc_pd(hdev->ibh_pd);
2408 if (hdev->ibh_cmid != NULL)
2409 rdma_destroy_id(hdev->ibh_cmid);
2411 LIBCFS_FREE(hdev, sizeof(*hdev));
2415 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2422 int acflags = IB_ACCESS_LOCAL_WRITE |
2423 IB_ACCESS_REMOTE_WRITE;
2425 rc = kiblnd_hdev_get_attr(hdev);
2429 if (hdev->ibh_mr_shift == 64) {
2430 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2431 if (hdev->ibh_mrs == NULL) {
2432 CERROR("Failed to allocate MRs table\n");
2436 hdev->ibh_mrs[0] = NULL;
2439 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2441 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2442 kiblnd_hdev_cleanup_mrs(hdev);
2446 hdev->ibh_mrs[0] = mr;
2451 mr_size = (1ULL << hdev->ibh_mr_shift);
2452 mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2454 hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2456 if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2457 /* it's 4T..., assume we will re-code at that time */
2458 CERROR("Can't support memory size: x"LPX64
2459 " with MR size: x"LPX64"\n", mm_size, mr_size);
2463 /* create an array of MRs to cover all memory */
2464 LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2465 if (hdev->ibh_mrs == NULL) {
2466 CERROR("Failed to allocate MRs' table\n");
2470 memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2472 for (i = 0; i < hdev->ibh_nmrs; i++) {
2473 struct ib_phys_buf ipb;
2476 ipb.size = hdev->ibh_mr_size;
2477 ipb.addr = i * mr_size;
2480 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2482 CERROR("Failed ib_reg_phys_mr addr "LPX64
2483 " size "LPX64" : %ld\n",
2484 ipb.addr, ipb.size, PTR_ERR(mr));
2485 kiblnd_hdev_cleanup_mrs(hdev);
2489 LASSERT (iova == ipb.addr);
2491 hdev->ibh_mrs[i] = mr;
2495 if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2496 LCONSOLE_INFO("Register global MR array, MR size: "
2497 LPX64", array size: %d\n",
2498 hdev->ibh_mr_size, hdev->ibh_nmrs);
2503 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2509 kiblnd_dev_need_failover(kib_dev_t *dev)
2511 struct rdma_cm_id *cmid;
2512 struct sockaddr_in srcaddr;
2513 struct sockaddr_in dstaddr;
2516 if (dev->ibd_hdev == NULL || /* initializing */
2517 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2518 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2521 /* XXX: it's UGLY, but I don't have better way to find
2522 * ib-bonding HCA failover because:
2524 * a. no reliable CM event for HCA failover...
2525 * b. no OFED API to get ib_device for current net_device...
2527 * We have only two choices at this point:
2529 * a. rdma_bind_addr(), it will conflict with listener cmid
2530 * b. rdma_resolve_addr() to zero addr */
2531 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2535 CERROR("Failed to create cmid for failover: %d\n", rc);
2539 memset(&srcaddr, 0, sizeof(srcaddr));
2540 srcaddr.sin_family = AF_INET;
2541 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2543 memset(&dstaddr, 0, sizeof(dstaddr));
2544 dstaddr.sin_family = AF_INET;
2545 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2546 (struct sockaddr *)&dstaddr, 1);
2547 if (rc != 0 || cmid->device == NULL) {
2548 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2549 dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2551 rdma_destroy_id(cmid);
2555 if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2556 /* don't need device failover */
2557 rdma_destroy_id(cmid);
2565 kiblnd_dev_failover(kib_dev_t *dev)
2567 CFS_LIST_HEAD (zombie_tpo);
2568 CFS_LIST_HEAD (zombie_ppo);
2569 CFS_LIST_HEAD (zombie_fpo);
2570 struct rdma_cm_id *cmid = NULL;
2571 kib_hca_dev_t *hdev = NULL;
2575 struct sockaddr_in addr;
2576 unsigned long flags;
2580 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2581 dev->ibd_can_failover ||
2582 dev->ibd_hdev == NULL);
2584 rc = kiblnd_dev_need_failover(dev);
2588 if (dev->ibd_hdev != NULL &&
2589 dev->ibd_hdev->ibh_cmid != NULL) {
2590 /* XXX it's not good to close old listener at here,
2591 * because we can fail to create new listener.
2592 * But we have to close it now, otherwise rdma_bind_addr
2593 * will return EADDRINUSE... How crap! */
2594 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2596 cmid = dev->ibd_hdev->ibh_cmid;
2597 /* make next schedule of kiblnd_dev_need_failover
2598 * will return 1 for me */
2599 dev->ibd_hdev->ibh_cmid = NULL;
2600 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2602 rdma_destroy_id(cmid);
2605 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2609 CERROR("Failed to create cmid for failover: %d\n", rc);
2613 memset(&addr, 0, sizeof(addr));
2614 addr.sin_family = AF_INET;
2615 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2616 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2618 /* Bind to failover device or port */
2619 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2620 if (rc != 0 || cmid->device == NULL) {
2621 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2622 dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2624 rdma_destroy_id(cmid);
2628 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2630 CERROR("Failed to allocate kib_hca_dev\n");
2631 rdma_destroy_id(cmid);
2636 atomic_set(&hdev->ibh_ref, 1);
2637 hdev->ibh_dev = dev;
2638 hdev->ibh_cmid = cmid;
2639 hdev->ibh_ibdev = cmid->device;
2641 pd = ib_alloc_pd(cmid->device);
2644 CERROR("Can't allocate PD: %d\n", rc);
2650 rc = rdma_listen(cmid, 0);
2652 CERROR("Can't start new listener: %d\n", rc);
2656 rc = kiblnd_hdev_setup_mrs(hdev);
2658 CERROR("Can't setup device: %d\n", rc);
2662 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2664 old = dev->ibd_hdev;
2665 dev->ibd_hdev = hdev; /* take over the refcount */
2668 cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2669 cfs_cpt_for_each(i, lnet_cpt_table()) {
2670 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2673 if (net->ibn_fmr_ps != NULL) {
2674 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2677 } else if (net->ibn_pmr_ps != NULL) {
2678 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2679 pps_poolset, &zombie_ppo);
2684 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2686 if (!cfs_list_empty(&zombie_tpo))
2687 kiblnd_destroy_pool_list(&zombie_tpo);
2688 if (!cfs_list_empty(&zombie_ppo))
2689 kiblnd_destroy_pool_list(&zombie_ppo);
2690 if (!cfs_list_empty(&zombie_fpo))
2691 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2693 kiblnd_hdev_decref(hdev);
2696 dev->ibd_failed_failover++;
2698 dev->ibd_failed_failover = 0;
2704 kiblnd_destroy_dev (kib_dev_t *dev)
2706 LASSERT (dev->ibd_nnets == 0);
2707 LASSERT (cfs_list_empty(&dev->ibd_nets));
2709 cfs_list_del(&dev->ibd_fail_list);
2710 cfs_list_del(&dev->ibd_list);
2712 if (dev->ibd_hdev != NULL)
2713 kiblnd_hdev_decref(dev->ibd_hdev);
2715 LIBCFS_FREE(dev, sizeof(*dev));
2719 kiblnd_create_dev(char *ifname)
2721 struct net_device *netdev;
2728 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2730 CERROR("Can't query IPoIB interface %s: %d\n",
2736 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2740 LIBCFS_ALLOC(dev, sizeof(*dev));
2744 memset(dev, 0, sizeof(*dev));
2745 #ifdef HAVE_DEV_GET_BY_NAME_2ARG
2746 netdev = dev_get_by_name(&init_net, ifname);
2748 netdev = dev_get_by_name(ifname);
2750 if (netdev == NULL) {
2751 dev->ibd_can_failover = 0;
2753 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2757 CFS_INIT_LIST_HEAD(&dev->ibd_nets);
2758 CFS_INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2759 CFS_INIT_LIST_HEAD(&dev->ibd_fail_list);
2761 strcpy(&dev->ibd_ifname[0], ifname);
2763 /* initialize the device */
2764 rc = kiblnd_dev_failover(dev);
2766 CERROR("Can't initialize device: %d\n", rc);
2767 LIBCFS_FREE(dev, sizeof(*dev));
2771 cfs_list_add_tail(&dev->ibd_list,
2772 &kiblnd_data.kib_devs);
2777 kiblnd_base_shutdown(void)
2779 struct kib_sched_info *sched;
2782 LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2784 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2785 cfs_atomic_read(&libcfs_kmemory));
2787 switch (kiblnd_data.kib_init) {
2791 case IBLND_INIT_ALL:
2792 case IBLND_INIT_DATA:
2793 LASSERT (kiblnd_data.kib_peers != NULL);
2794 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2795 LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2797 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2798 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2800 /* flag threads to terminate; wake and wait for them to die */
2801 kiblnd_data.kib_shutdown = 1;
2803 /* NB: we really want to stop scheduler threads net by net
2804 * instead of the whole module, this should be improved
2805 * with dynamic configuration LNet */
2806 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2807 cfs_waitq_broadcast(&sched->ibs_waitq);
2809 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2810 cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq);
2813 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2815 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2816 "Waiting for %d threads to terminate\n",
2817 cfs_atomic_read(&kiblnd_data.kib_nthreads));
2818 cfs_pause(cfs_time_seconds(1));
2823 case IBLND_INIT_NOTHING:
2827 if (kiblnd_data.kib_peers != NULL) {
2828 LIBCFS_FREE(kiblnd_data.kib_peers,
2829 sizeof(cfs_list_t) *
2830 kiblnd_data.kib_peer_hash_size);
2833 if (kiblnd_data.kib_scheds != NULL)
2834 cfs_percpt_free(kiblnd_data.kib_scheds);
2836 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2837 cfs_atomic_read(&libcfs_kmemory));
2839 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2840 PORTAL_MODULE_UNUSE;
2844 kiblnd_shutdown (lnet_ni_t *ni)
2846 kib_net_t *net = ni->ni_data;
2847 cfs_rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2849 unsigned long flags;
2851 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2856 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2857 cfs_atomic_read(&libcfs_kmemory));
2859 cfs_write_lock_irqsave(g_lock, flags);
2860 net->ibn_shutdown = 1;
2861 cfs_write_unlock_irqrestore(g_lock, flags);
2863 switch (net->ibn_init) {
2867 case IBLND_INIT_ALL:
2868 /* nuke all existing peers within this net */
2869 kiblnd_del_peer(ni, LNET_NID_ANY);
2871 /* Wait for all peer state to clean up */
2873 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2875 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2876 "%s: waiting for %d peers to disconnect\n",
2877 libcfs_nid2str(ni->ni_nid),
2878 cfs_atomic_read(&net->ibn_npeers));
2879 cfs_pause(cfs_time_seconds(1));
2882 kiblnd_net_fini_pools(net);
2884 cfs_write_lock_irqsave(g_lock, flags);
2885 LASSERT (net->ibn_dev->ibd_nnets > 0);
2886 net->ibn_dev->ibd_nnets--;
2887 cfs_list_del(&net->ibn_list);
2888 cfs_write_unlock_irqrestore(g_lock, flags);
2892 case IBLND_INIT_NOTHING:
2893 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2895 if (net->ibn_dev != NULL &&
2896 net->ibn_dev->ibd_nnets == 0)
2897 kiblnd_destroy_dev(net->ibn_dev);
2902 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2903 cfs_atomic_read(&libcfs_kmemory));
2905 net->ibn_init = IBLND_INIT_NOTHING;
2908 LIBCFS_FREE(net, sizeof(*net));
2911 if (cfs_list_empty(&kiblnd_data.kib_devs))
2912 kiblnd_base_shutdown();
2917 kiblnd_base_startup(void)
2919 struct kib_sched_info *sched;
2923 LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2926 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2928 cfs_rwlock_init(&kiblnd_data.kib_global_lock);
2930 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2931 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2933 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2934 LIBCFS_ALLOC(kiblnd_data.kib_peers,
2935 sizeof(cfs_list_t) *
2936 kiblnd_data.kib_peer_hash_size);
2937 if (kiblnd_data.kib_peers == NULL) {
2940 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2941 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2943 cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
2944 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2945 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2946 cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2947 cfs_waitq_init(&kiblnd_data.kib_failover_waitq);
2949 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2951 if (kiblnd_data.kib_scheds == NULL)
2954 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2957 cfs_spin_lock_init(&sched->ibs_lock);
2958 CFS_INIT_LIST_HEAD(&sched->ibs_conns);
2959 cfs_waitq_init(&sched->ibs_waitq);
2961 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2962 if (*kiblnd_tunables.kib_nscheds > 0) {
2963 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2965 /* max to half of CPUs, another half is reserved for
2966 * upper layer modules */
2967 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2970 sched->ibs_nthreads_max = nthrs;
2974 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2976 /* lists/ptrs/locks initialised */
2977 kiblnd_data.kib_init = IBLND_INIT_DATA;
2978 /*****************************************************/
2980 rc = kiblnd_thread_start(kiblnd_connd, NULL);
2982 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2986 if (*kiblnd_tunables.kib_dev_failover != 0)
2987 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL);
2990 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2994 /* flag everything initialised */
2995 kiblnd_data.kib_init = IBLND_INIT_ALL;
2996 /*****************************************************/
3001 kiblnd_base_shutdown();
3006 kiblnd_start_schedulers(struct kib_sched_info *sched)
3012 if (sched->ibs_nthreads == 0) {
3013 if (*kiblnd_tunables.kib_nscheds > 0) {
3014 nthrs = sched->ibs_nthreads_max;
3016 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3018 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3019 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3022 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3023 /* increase one thread if there is new interface */
3024 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3027 for (i = 0; i < nthrs; i++) {
3030 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3031 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id);
3035 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3036 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3040 sched->ibs_nthreads += i;
3045 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3051 for (i = 0; i < ncpts; i++) {
3052 struct kib_sched_info *sched;
3054 cpt = (cpts == NULL) ? i : cpts[i];
3055 sched = kiblnd_data.kib_scheds[cpt];
3057 if (!newdev && sched->ibs_nthreads > 0)
3060 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3062 CERROR("Failed to start scheduler threads for %s\n",
3071 kiblnd_dev_search(char *ifname)
3073 kib_dev_t *alias = NULL;
3078 colon = strchr(ifname, ':');
3079 cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3080 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3086 colon2 = strchr(dev->ibd_ifname, ':');
3092 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3104 kiblnd_startup (lnet_ni_t *ni)
3107 kib_dev_t *ibdev = NULL;
3110 unsigned long flags;
3114 LASSERT (ni->ni_lnd == &the_o2iblnd);
3116 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3117 rc = kiblnd_base_startup();
3122 LIBCFS_ALLOC(net, sizeof(*net));
3127 memset(net, 0, sizeof(*net));
3129 cfs_gettimeofday(&tv);
3130 net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3132 ni->ni_peertimeout = *kiblnd_tunables.kib_peertimeout;
3133 ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
3134 ni->ni_peertxcredits = *kiblnd_tunables.kib_peertxcredits;
3135 ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3137 if (ni->ni_interfaces[0] != NULL) {
3138 /* Use the IPoIB interface specified in 'networks=' */
3140 CLASSERT (LNET_MAX_INTERFACES > 1);
3141 if (ni->ni_interfaces[1] != NULL) {
3142 CERROR("Multiple interfaces not supported\n");
3146 ifname = ni->ni_interfaces[0];
3148 ifname = *kiblnd_tunables.kib_default_ipif;
3151 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3152 CERROR("IPoIB interface name too long: %s\n", ifname);
3156 ibdev = kiblnd_dev_search(ifname);
3158 newdev = ibdev == NULL;
3159 /* hmm...create kib_dev even for alias */
3160 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3161 ibdev = kiblnd_create_dev(ifname);
3166 net->ibn_dev = ibdev;
3167 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3169 rc = kiblnd_dev_start_threads(ibdev, newdev,
3170 ni->ni_cpts, ni->ni_ncpts);
3174 rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3176 CERROR("Failed to initialize NI pools: %d\n", rc);
3180 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3182 cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3183 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3185 net->ibn_init = IBLND_INIT_ALL;
3190 if (net->ibn_dev == NULL && ibdev != NULL)
3191 kiblnd_destroy_dev(ibdev);
3193 kiblnd_shutdown(ni);
3195 CDEBUG(D_NET, "kiblnd_startup failed\n");
3200 kiblnd_module_fini (void)
3202 lnet_unregister_lnd(&the_o2iblnd);
3203 kiblnd_tunables_fini();
3207 kiblnd_module_init (void)
3211 CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3212 CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3214 CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3217 rc = kiblnd_tunables_init();
3221 lnet_register_lnd(&the_o2iblnd);
3226 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3227 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3228 MODULE_LICENSE("GPL");
3230 module_init(kiblnd_module_init);
3231 module_exit(kiblnd_module_fini);