4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd.c
34 * Author: Eric Barton <eric@bartonsoftware.com>
40 static struct lnet_lnd the_o2iblnd;
42 kib_data_t kiblnd_data;
45 kiblnd_cksum (void *ptr, int nob)
51 sum = ((sum << 1) | (sum >> 31)) + *c++;
53 /* ensure I don't return 0 (== no checksum) */
54 return (sum == 0) ? 1 : sum;
58 kiblnd_msgtype2str(int type)
61 case IBLND_MSG_CONNREQ:
64 case IBLND_MSG_CONNACK:
70 case IBLND_MSG_IMMEDIATE:
73 case IBLND_MSG_PUT_REQ:
76 case IBLND_MSG_PUT_NAK:
79 case IBLND_MSG_PUT_ACK:
82 case IBLND_MSG_PUT_DONE:
85 case IBLND_MSG_GET_REQ:
88 case IBLND_MSG_GET_DONE:
97 kiblnd_msgtype2size(int type)
99 const int hdr_size = offsetof(kib_msg_t, ibm_u);
102 case IBLND_MSG_CONNREQ:
103 case IBLND_MSG_CONNACK:
104 return hdr_size + sizeof(kib_connparams_t);
109 case IBLND_MSG_IMMEDIATE:
110 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
112 case IBLND_MSG_PUT_REQ:
113 return hdr_size + sizeof(kib_putreq_msg_t);
115 case IBLND_MSG_PUT_ACK:
116 return hdr_size + sizeof(kib_putack_msg_t);
118 case IBLND_MSG_GET_REQ:
119 return hdr_size + sizeof(kib_get_msg_t);
121 case IBLND_MSG_PUT_NAK:
122 case IBLND_MSG_PUT_DONE:
123 case IBLND_MSG_GET_DONE:
124 return hdr_size + sizeof(kib_completion_msg_t);
131 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
138 LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
139 msg->ibm_type == IBLND_MSG_PUT_ACK);
141 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
142 &msg->ibm_u.get.ibgm_rd :
143 &msg->ibm_u.putack.ibpam_rd;
146 __swab32s(&rd->rd_key);
147 __swab32s(&rd->rd_nfrags);
152 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
153 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
154 n, IBLND_MAX_RDMA_FRAGS);
158 nob = offsetof (kib_msg_t, ibm_u) +
159 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
161 if (msg->ibm_nob < nob) {
162 CERROR("Short %s: %d(%d)\n",
163 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
170 for (i = 0; i < n; i++) {
171 __swab32s(&rd->rd_frags[i].rf_nob);
172 __swab64s(&rd->rd_frags[i].rf_addr);
179 kiblnd_pack_msg(struct lnet_ni *ni, kib_msg_t *msg, int version,
180 int credits, lnet_nid_t dstnid, __u64 dststamp)
182 kib_net_t *net = ni->ni_data;
184 /* CAVEAT EMPTOR! all message fields not set here should have been
185 * initialised previously. */
186 msg->ibm_magic = IBLND_MSG_MAGIC;
187 msg->ibm_version = version;
189 msg->ibm_credits = credits;
192 msg->ibm_srcnid = ni->ni_nid;
193 msg->ibm_srcstamp = net->ibn_incarnation;
194 msg->ibm_dstnid = dstnid;
195 msg->ibm_dststamp = dststamp;
197 if (*kiblnd_tunables.kib_cksum) {
198 /* NB ibm_cksum zero while computing cksum */
199 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
204 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
206 const int hdr_size = offsetof(kib_msg_t, ibm_u);
212 /* 6 bytes are enough to have received magic + version */
214 CERROR("Short message: %d\n", nob);
218 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
220 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
223 CERROR("Bad magic: %08x\n", msg->ibm_magic);
227 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
228 if (version != IBLND_MSG_VERSION &&
229 version != IBLND_MSG_VERSION_1) {
230 CERROR("Bad version: %x\n", version);
234 if (nob < hdr_size) {
235 CERROR("Short message: %d\n", nob);
239 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
241 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
245 /* checksum must be computed with ibm_cksum zero and BEFORE anything
247 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
249 if (msg_cksum != 0 &&
250 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
251 CERROR("Bad checksum\n");
255 msg->ibm_cksum = msg_cksum;
258 /* leave magic unflipped as a clue to peer_ni endianness */
259 msg->ibm_version = version;
260 CLASSERT (sizeof(msg->ibm_type) == 1);
261 CLASSERT (sizeof(msg->ibm_credits) == 1);
262 msg->ibm_nob = msg_nob;
263 __swab64s(&msg->ibm_srcnid);
264 __swab64s(&msg->ibm_srcstamp);
265 __swab64s(&msg->ibm_dstnid);
266 __swab64s(&msg->ibm_dststamp);
269 if (msg->ibm_srcnid == LNET_NID_ANY) {
270 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
274 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
275 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
276 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
280 switch (msg->ibm_type) {
282 CERROR("Unknown message type %x\n", msg->ibm_type);
286 case IBLND_MSG_IMMEDIATE:
287 case IBLND_MSG_PUT_REQ:
290 case IBLND_MSG_PUT_ACK:
291 case IBLND_MSG_GET_REQ:
292 if (kiblnd_unpack_rd(msg, flip))
296 case IBLND_MSG_PUT_NAK:
297 case IBLND_MSG_PUT_DONE:
298 case IBLND_MSG_GET_DONE:
300 __swab32s(&msg->ibm_u.completion.ibcm_status);
303 case IBLND_MSG_CONNREQ:
304 case IBLND_MSG_CONNACK:
306 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
307 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
308 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
316 kiblnd_create_peer(struct lnet_ni *ni, kib_peer_ni_t **peerp, lnet_nid_t nid)
318 kib_peer_ni_t *peer_ni;
319 kib_net_t *net = ni->ni_data;
320 int cpt = lnet_cpt_of_nid(nid, ni);
323 LASSERT(net != NULL);
324 LASSERT(nid != LNET_NID_ANY);
326 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
327 if (peer_ni == NULL) {
328 CERROR("Cannot allocate peer_ni\n");
332 peer_ni->ibp_ni = ni;
333 peer_ni->ibp_nid = nid;
334 peer_ni->ibp_error = 0;
335 peer_ni->ibp_last_alive = 0;
336 peer_ni->ibp_max_frags = kiblnd_cfg_rdma_frags(peer_ni->ibp_ni);
337 peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
338 atomic_set(&peer_ni->ibp_refcount, 1); /* 1 ref for caller */
340 INIT_LIST_HEAD(&peer_ni->ibp_list); /* not in the peer_ni table yet */
341 INIT_LIST_HEAD(&peer_ni->ibp_conns);
342 INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
344 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
346 /* always called with a ref on ni, which prevents ni being shutdown */
347 LASSERT(net->ibn_shutdown == 0);
349 /* npeers only grows with the global lock held */
350 atomic_inc(&net->ibn_npeers);
352 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
359 kiblnd_destroy_peer (kib_peer_ni_t *peer_ni)
361 kib_net_t *net = peer_ni->ibp_ni->ni_data;
363 LASSERT(net != NULL);
364 LASSERT (atomic_read(&peer_ni->ibp_refcount) == 0);
365 LASSERT(!kiblnd_peer_active(peer_ni));
366 LASSERT(kiblnd_peer_idle(peer_ni));
367 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
369 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
371 /* NB a peer_ni's connections keep a reference on their peer_ni until
372 * they are destroyed, so we can be assured that _all_ state to do
373 * with this peer_ni has been cleaned up when its refcount drops to
375 atomic_dec(&net->ibn_npeers);
379 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
381 /* the caller is responsible for accounting the additional reference
382 * that this creates */
383 struct list_head *peer_list = kiblnd_nid2peerlist(nid);
384 struct list_head *tmp;
385 kib_peer_ni_t *peer_ni;
387 list_for_each(tmp, peer_list) {
389 peer_ni = list_entry(tmp, kib_peer_ni_t, ibp_list);
390 LASSERT(!kiblnd_peer_idle(peer_ni));
393 * Match a peer if its NID and the NID of the local NI it
394 * communicates over are the same. Otherwise don't match
395 * the peer, which will result in a new lnd peer being
398 if (peer_ni->ibp_nid != nid ||
399 peer_ni->ibp_ni->ni_nid != ni->ni_nid)
402 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
403 peer_ni, libcfs_nid2str(nid),
404 atomic_read(&peer_ni->ibp_refcount),
405 peer_ni->ibp_version);
412 kiblnd_unlink_peer_locked (kib_peer_ni_t *peer_ni)
414 LASSERT(list_empty(&peer_ni->ibp_conns));
416 LASSERT (kiblnd_peer_active(peer_ni));
417 list_del_init(&peer_ni->ibp_list);
418 /* lose peerlist's ref */
419 kiblnd_peer_decref(peer_ni);
423 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
424 lnet_nid_t *nidp, int *count)
426 kib_peer_ni_t *peer_ni;
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_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
438 LASSERT(!kiblnd_peer_idle(peer_ni));
440 if (peer_ni->ibp_ni != ni)
446 *nidp = peer_ni->ibp_nid;
447 *count = atomic_read(&peer_ni->ibp_refcount);
449 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
455 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
460 kiblnd_del_peer_locked (kib_peer_ni_t *peer_ni)
462 struct list_head *ctmp;
463 struct list_head *cnxt;
466 if (list_empty(&peer_ni->ibp_conns)) {
467 kiblnd_unlink_peer_locked(peer_ni);
469 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
470 conn = list_entry(ctmp, kib_conn_t, ibc_list);
472 kiblnd_close_conn_locked(conn, 0);
474 /* NB closing peer_ni's last conn unlinked it. */
476 /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
481 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
483 struct list_head zombies = LIST_HEAD_INIT(zombies);
484 struct list_head *ptmp;
485 struct list_head *pnxt;
486 kib_peer_ni_t *peer_ni;
493 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
495 if (nid != LNET_NID_ANY) {
496 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
499 hi = kiblnd_data.kib_peer_hash_size - 1;
502 for (i = lo; i <= hi; i++) {
503 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
504 peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
505 LASSERT(!kiblnd_peer_idle(peer_ni));
507 if (peer_ni->ibp_ni != ni)
510 if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
513 if (!list_empty(&peer_ni->ibp_tx_queue)) {
514 LASSERT(list_empty(&peer_ni->ibp_conns));
516 list_splice_init(&peer_ni->ibp_tx_queue,
520 kiblnd_del_peer_locked(peer_ni);
521 rc = 0; /* matched something */
525 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
527 kiblnd_txlist_done(ni, &zombies, -EIO);
533 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
535 kib_peer_ni_t *peer_ni;
536 struct list_head *ptmp;
538 struct list_head *ctmp;
542 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
544 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
545 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
547 peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
548 LASSERT(!kiblnd_peer_idle(peer_ni));
550 if (peer_ni->ibp_ni != ni)
553 list_for_each(ctmp, &peer_ni->ibp_conns) {
557 conn = list_entry(ctmp, kib_conn_t, ibc_list);
558 kiblnd_conn_addref(conn);
559 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
566 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
571 kiblnd_debug_rx (kib_rx_t *rx)
573 CDEBUG(D_CONSOLE, " %p status %d msg_type %x cred %d\n",
574 rx, rx->rx_status, rx->rx_msg->ibm_type,
575 rx->rx_msg->ibm_credits);
579 kiblnd_debug_tx (kib_tx_t *tx)
581 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lx "
582 "cookie %#llx msg %s%s type %x cred %d\n",
583 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
584 tx->tx_status, tx->tx_deadline, tx->tx_cookie,
585 tx->tx_lntmsg[0] == NULL ? "-" : "!",
586 tx->tx_lntmsg[1] == NULL ? "-" : "!",
587 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
591 kiblnd_debug_conn (kib_conn_t *conn)
593 struct list_head *tmp;
596 spin_lock(&conn->ibc_lock);
598 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
599 atomic_read(&conn->ibc_refcount), conn,
600 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
601 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d "
602 " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
603 conn->ibc_nsends_posted, conn->ibc_credits,
604 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
605 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
607 CDEBUG(D_CONSOLE, " early_rxs:\n");
608 list_for_each(tmp, &conn->ibc_early_rxs)
609 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
611 CDEBUG(D_CONSOLE, " tx_noops:\n");
612 list_for_each(tmp, &conn->ibc_tx_noops)
613 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
615 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
616 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
617 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
619 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
620 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
621 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
623 CDEBUG(D_CONSOLE, " tx_queue:\n");
624 list_for_each(tmp, &conn->ibc_tx_queue)
625 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
627 CDEBUG(D_CONSOLE, " active_txs:\n");
628 list_for_each(tmp, &conn->ibc_active_txs)
629 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
631 CDEBUG(D_CONSOLE, " rxs:\n");
632 for (i = 0; i < IBLND_RX_MSGS(conn); i++)
633 kiblnd_debug_rx(&conn->ibc_rxs[i]);
635 spin_unlock(&conn->ibc_lock);
639 kiblnd_translate_mtu(int value)
660 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
664 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
665 if (cmid->route.path_rec == NULL)
668 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
671 cmid->route.path_rec->mtu = mtu;
675 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
683 vectors = conn->ibc_cmid->device->num_comp_vectors;
687 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
689 /* hash NID to CPU id in this partition... */
690 ibp_nid = conn->ibc_peer->ibp_nid;
691 off = do_div(ibp_nid, cpumask_weight(mask));
692 for_each_cpu(i, mask) {
702 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
703 int state, int version)
706 * If the new conn is created successfully it takes over the caller's
707 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
708 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
709 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
710 * to destroy 'cmid' here since I'm called from the CM which still has
711 * its ref on 'cmid'). */
712 rwlock_t *glock = &kiblnd_data.kib_global_lock;
713 kib_net_t *net = peer_ni->ibp_ni->ni_data;
715 struct ib_qp_init_attr *init_qp_attr;
716 struct kib_sched_info *sched;
717 #ifdef HAVE_IB_CQ_INIT_ATTR
718 struct ib_cq_init_attr cq_attr = {};
727 LASSERT(net != NULL);
728 LASSERT(!in_interrupt());
732 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
733 sched = kiblnd_data.kib_scheds[cpt];
735 LASSERT(sched->ibs_nthreads > 0);
737 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
738 sizeof(*init_qp_attr));
739 if (init_qp_attr == NULL) {
740 CERROR("Can't allocate qp_attr for %s\n",
741 libcfs_nid2str(peer_ni->ibp_nid));
745 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
747 CERROR("Can't allocate connection for %s\n",
748 libcfs_nid2str(peer_ni->ibp_nid));
752 conn->ibc_state = IBLND_CONN_INIT;
753 conn->ibc_version = version;
754 conn->ibc_peer = peer_ni; /* I take the caller's ref */
755 cmid->context = conn; /* for future CM callbacks */
756 conn->ibc_cmid = cmid;
757 conn->ibc_max_frags = peer_ni->ibp_max_frags;
758 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
760 INIT_LIST_HEAD(&conn->ibc_early_rxs);
761 INIT_LIST_HEAD(&conn->ibc_tx_noops);
762 INIT_LIST_HEAD(&conn->ibc_tx_queue);
763 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
764 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
765 INIT_LIST_HEAD(&conn->ibc_active_txs);
766 spin_lock_init(&conn->ibc_lock);
768 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
769 sizeof(*conn->ibc_connvars));
770 if (conn->ibc_connvars == NULL) {
771 CERROR("Can't allocate in-progress connection state\n");
775 write_lock_irqsave(glock, flags);
776 if (dev->ibd_failover) {
777 write_unlock_irqrestore(glock, flags);
778 CERROR("%s: failover in progress\n", dev->ibd_ifname);
782 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
783 /* wakeup failover thread and teardown connection */
784 if (kiblnd_dev_can_failover(dev)) {
785 list_add_tail(&dev->ibd_fail_list,
786 &kiblnd_data.kib_failed_devs);
787 wake_up(&kiblnd_data.kib_failover_waitq);
790 write_unlock_irqrestore(glock, flags);
791 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
792 cmid->device->name, dev->ibd_ifname);
796 kiblnd_hdev_addref_locked(dev->ibd_hdev);
797 conn->ibc_hdev = dev->ibd_hdev;
799 kiblnd_setup_mtu_locked(cmid);
801 write_unlock_irqrestore(glock, flags);
803 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
804 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
805 if (conn->ibc_rxs == NULL) {
806 CERROR("Cannot allocate RX buffers\n");
810 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
811 IBLND_RX_MSG_PAGES(conn));
815 kiblnd_map_rx_descs(conn);
817 #ifdef HAVE_IB_CQ_INIT_ATTR
818 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
819 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
820 cq = ib_create_cq(cmid->device,
821 kiblnd_cq_completion, kiblnd_cq_event, conn,
824 cq = ib_create_cq(cmid->device,
825 kiblnd_cq_completion, kiblnd_cq_event, conn,
826 IBLND_CQ_ENTRIES(conn),
827 kiblnd_get_completion_vector(conn, cpt));
830 CERROR("Failed to create CQ with %d CQEs: %ld\n",
831 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
837 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
839 CERROR("Can't request completion notification: %d\n", rc);
843 init_qp_attr->event_handler = kiblnd_qp_event;
844 init_qp_attr->qp_context = conn;
845 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
846 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
847 init_qp_attr->cap.max_send_sge = 1;
848 init_qp_attr->cap.max_recv_sge = 1;
849 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
850 init_qp_attr->qp_type = IB_QPT_RC;
851 init_qp_attr->send_cq = cq;
852 init_qp_attr->recv_cq = cq;
854 conn->ibc_sched = sched;
857 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
858 if (!rc || init_qp_attr->cap.max_send_wr < 16)
861 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
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 if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
872 CDEBUG(D_NET, "original send wr %d, created with %d\n",
873 IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
875 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
877 /* 1 ref for caller and each rxmsg */
878 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
879 conn->ibc_nrx = IBLND_RX_MSGS(conn);
882 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
883 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
885 CERROR("Can't post rxmsg: %d\n", rc);
887 /* Make posted receives complete */
888 kiblnd_abort_receives(conn);
890 /* correct # of posted buffers
891 * NB locking needed now I'm racing with completion */
892 spin_lock_irqsave(&sched->ibs_lock, flags);
893 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
894 spin_unlock_irqrestore(&sched->ibs_lock, flags);
896 /* cmid will be destroyed by CM(ofed) after cm_callback
897 * returned, so we can't refer it anymore
898 * (by kiblnd_connd()->kiblnd_destroy_conn) */
899 rdma_destroy_qp(conn->ibc_cmid);
900 conn->ibc_cmid = NULL;
902 /* Drop my own and unused rxbuffer refcounts */
903 while (i++ <= IBLND_RX_MSGS(conn))
904 kiblnd_conn_decref(conn);
910 /* Init successful! */
911 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
912 state == IBLND_CONN_PASSIVE_WAIT);
913 conn->ibc_state = state;
916 atomic_inc(&net->ibn_nconns);
920 kiblnd_destroy_conn(conn, true);
922 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
928 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
930 struct rdma_cm_id *cmid = conn->ibc_cmid;
931 kib_peer_ni_t *peer_ni = conn->ibc_peer;
934 LASSERT (!in_interrupt());
935 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
936 LASSERT(list_empty(&conn->ibc_early_rxs));
937 LASSERT(list_empty(&conn->ibc_tx_noops));
938 LASSERT(list_empty(&conn->ibc_tx_queue));
939 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
940 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
941 LASSERT(list_empty(&conn->ibc_active_txs));
942 LASSERT (conn->ibc_noops_posted == 0);
943 LASSERT (conn->ibc_nsends_posted == 0);
945 switch (conn->ibc_state) {
947 /* conn must be completely disengaged from the network */
950 case IBLND_CONN_DISCONNECTED:
951 /* connvars should have been freed already */
952 LASSERT (conn->ibc_connvars == NULL);
955 case IBLND_CONN_INIT:
959 /* conn->ibc_cmid might be destroyed by CM already */
960 if (cmid != NULL && cmid->qp != NULL)
961 rdma_destroy_qp(cmid);
963 if (conn->ibc_cq != NULL) {
964 rc = ib_destroy_cq(conn->ibc_cq);
966 CWARN("Error destroying CQ: %d\n", rc);
969 if (conn->ibc_rx_pages != NULL)
970 kiblnd_unmap_rx_descs(conn);
972 if (conn->ibc_rxs != NULL) {
973 LIBCFS_FREE(conn->ibc_rxs,
974 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
977 if (conn->ibc_connvars != NULL)
978 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
980 if (conn->ibc_hdev != NULL)
981 kiblnd_hdev_decref(conn->ibc_hdev);
983 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
984 if (conn->ibc_state != IBLND_CONN_INIT) {
985 kib_net_t *net = peer_ni->ibp_ni->ni_data;
987 kiblnd_peer_decref(peer_ni);
988 rdma_destroy_id(cmid);
989 atomic_dec(&net->ibn_nconns);
993 LIBCFS_FREE(conn, sizeof(*conn));
997 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1000 struct list_head *ctmp;
1001 struct list_head *cnxt;
1004 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1005 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1007 CDEBUG(D_NET, "Closing conn -> %s, "
1008 "version: %x, reason: %d\n",
1009 libcfs_nid2str(peer_ni->ibp_nid),
1010 conn->ibc_version, why);
1012 kiblnd_close_conn_locked(conn, why);
1020 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1021 int version, __u64 incarnation)
1024 struct list_head *ctmp;
1025 struct list_head *cnxt;
1028 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1029 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1031 if (conn->ibc_version == version &&
1032 conn->ibc_incarnation == incarnation)
1035 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1036 "incarnation:%#llx(%x, %#llx)\n",
1037 libcfs_nid2str(peer_ni->ibp_nid),
1038 conn->ibc_version, conn->ibc_incarnation,
1039 version, incarnation);
1041 kiblnd_close_conn_locked(conn, -ESTALE);
1049 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1051 kib_peer_ni_t *peer_ni;
1052 struct list_head *ptmp;
1053 struct list_head *pnxt;
1057 unsigned long flags;
1060 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1062 if (nid != LNET_NID_ANY)
1063 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1066 hi = kiblnd_data.kib_peer_hash_size - 1;
1069 for (i = lo; i <= hi; i++) {
1070 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1072 peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1073 LASSERT(!kiblnd_peer_idle(peer_ni));
1075 if (peer_ni->ibp_ni != ni)
1078 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1081 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1085 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1087 /* wildcards always succeed */
1088 if (nid == LNET_NID_ANY)
1091 return (count == 0) ? -ENOENT : 0;
1095 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1097 struct libcfs_ioctl_data *data = arg;
1101 case IOC_LIBCFS_GET_PEER: {
1105 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1107 data->ioc_nid = nid;
1108 data->ioc_count = count;
1112 case IOC_LIBCFS_DEL_PEER: {
1113 rc = kiblnd_del_peer(ni, data->ioc_nid);
1116 case IOC_LIBCFS_GET_CONN: {
1120 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1126 LASSERT (conn->ibc_cmid != NULL);
1127 data->ioc_nid = conn->ibc_peer->ibp_nid;
1128 if (conn->ibc_cmid->route.path_rec == NULL)
1129 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1132 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1133 kiblnd_conn_decref(conn);
1136 case IOC_LIBCFS_CLOSE_CONNECTION: {
1137 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1149 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when)
1151 cfs_time_t last_alive = 0;
1152 cfs_time_t now = cfs_time_current();
1153 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1154 kib_peer_ni_t *peer_ni;
1155 unsigned long flags;
1157 read_lock_irqsave(glock, flags);
1159 peer_ni = kiblnd_find_peer_locked(ni, nid);
1160 if (peer_ni != NULL)
1161 last_alive = peer_ni->ibp_last_alive;
1163 read_unlock_irqrestore(glock, flags);
1165 if (last_alive != 0)
1168 /* peer_ni is not persistent in hash, trigger peer_ni creation
1169 * and connection establishment with a NULL tx */
1170 if (peer_ni == NULL)
1171 kiblnd_launch_tx(ni, NULL, nid);
1173 CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago\n",
1174 libcfs_nid2str(nid), peer_ni,
1175 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1180 kiblnd_free_pages(kib_pages_t *p)
1182 int npages = p->ibp_npages;
1185 for (i = 0; i < npages; i++) {
1186 if (p->ibp_pages[i] != NULL)
1187 __free_page(p->ibp_pages[i]);
1190 LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1194 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1199 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1200 offsetof(kib_pages_t, ibp_pages[npages]));
1202 CERROR("Can't allocate descriptor for %d pages\n", npages);
1206 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1207 p->ibp_npages = npages;
1209 for (i = 0; i < npages; i++) {
1210 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1212 if (p->ibp_pages[i] == NULL) {
1213 CERROR("Can't allocate page %d of %d\n", i, npages);
1214 kiblnd_free_pages(p);
1224 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1229 LASSERT (conn->ibc_rxs != NULL);
1230 LASSERT (conn->ibc_hdev != NULL);
1232 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1233 rx = &conn->ibc_rxs[i];
1235 LASSERT(rx->rx_nob >= 0); /* not posted */
1237 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1238 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1240 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1243 kiblnd_free_pages(conn->ibc_rx_pages);
1245 conn->ibc_rx_pages = NULL;
1249 kiblnd_map_rx_descs(kib_conn_t *conn)
1257 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1258 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1259 rx = &conn->ibc_rxs[i];
1262 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1265 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1266 rx->rx_msg, IBLND_MSG_SIZE,
1268 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1270 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1272 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1273 i, rx->rx_msg, rx->rx_msgaddr,
1274 (__u64)(page_to_phys(pg) + pg_off));
1276 pg_off += IBLND_MSG_SIZE;
1277 LASSERT(pg_off <= PAGE_SIZE);
1279 if (pg_off == PAGE_SIZE) {
1282 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1288 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1290 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1294 LASSERT (tpo->tpo_pool.po_allocated == 0);
1299 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1300 tx = &tpo->tpo_tx_descs[i];
1301 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1302 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1304 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1307 kiblnd_hdev_decref(hdev);
1308 tpo->tpo_hdev = NULL;
1311 static kib_hca_dev_t *
1312 kiblnd_current_hdev(kib_dev_t *dev)
1314 kib_hca_dev_t *hdev;
1315 unsigned long flags;
1318 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1319 while (dev->ibd_failover) {
1320 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1322 CDEBUG(D_NET, "%s: Wait for failover\n",
1324 set_current_state(TASK_INTERRUPTIBLE);
1325 schedule_timeout(cfs_time_seconds(1) / 100);
1327 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1330 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1331 hdev = dev->ibd_hdev;
1333 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;
1351 LASSERT (net != NULL);
1355 /* pre-mapped messages are not bigger than 1 page */
1356 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1358 /* No fancy arithmetic when we do the buffer calculations */
1359 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1361 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1363 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1364 page = txpgs->ibp_pages[ipage];
1365 tx = &tpo->tpo_tx_descs[i];
1367 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1370 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1374 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1376 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1378 list_add(&tx->tx_list, &pool->po_free_list);
1380 page_offset += IBLND_MSG_SIZE;
1381 LASSERT(page_offset <= PAGE_SIZE);
1383 if (page_offset == PAGE_SIZE) {
1386 LASSERT(ipage <= txpgs->ibp_npages);
1392 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1393 int negotiated_nfrags)
1395 kib_net_t *net = ni->ni_data;
1396 kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
1397 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1401 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1402 mod = tunables->lnd_map_on_demand;
1403 nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1405 LASSERT(hdev->ibh_mrs != NULL);
1407 if (mod > 0 && nfrags <= rd->rd_nfrags)
1410 return hdev->ibh_mrs;
1414 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1416 LASSERT(fpo->fpo_map_count == 0);
1418 if (fpo->fpo_is_fmr) {
1419 if (fpo->fmr.fpo_fmr_pool)
1420 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1422 struct kib_fast_reg_descriptor *frd, *tmp;
1425 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1427 list_del(&frd->frd_list);
1428 #ifndef HAVE_IB_MAP_MR_SG
1429 ib_free_fast_reg_page_list(frd->frd_frpl);
1431 ib_dereg_mr(frd->frd_mr);
1432 LIBCFS_FREE(frd, sizeof(*frd));
1435 if (i < fpo->fast_reg.fpo_pool_size)
1436 CERROR("FastReg pool still has %d regions registered\n",
1437 fpo->fast_reg.fpo_pool_size - i);
1441 kiblnd_hdev_decref(fpo->fpo_hdev);
1443 LIBCFS_FREE(fpo, sizeof(*fpo));
1447 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1449 kib_fmr_pool_t *fpo, *tmp;
1451 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1452 list_del(&fpo->fpo_list);
1453 kiblnd_destroy_fmr_pool(fpo);
1458 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1461 int size = tunables->lnd_fmr_pool_size / ncpts;
1463 return max(IBLND_FMR_POOL, size);
1467 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1470 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1472 return max(IBLND_FMR_POOL_FLUSH, size);
1475 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1477 struct ib_fmr_pool_param param = {
1478 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1479 .page_shift = PAGE_SHIFT,
1480 .access = (IB_ACCESS_LOCAL_WRITE |
1481 IB_ACCESS_REMOTE_WRITE),
1482 .pool_size = fps->fps_pool_size,
1483 .dirty_watermark = fps->fps_flush_trigger,
1484 .flush_function = NULL,
1486 .cache = !!fps->fps_cache };
1489 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1491 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1492 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1494 CERROR("Failed to create FMR pool: %d\n", rc);
1496 CERROR("FMRs are not supported\n");
1502 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1504 struct kib_fast_reg_descriptor *frd, *tmp;
1507 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1508 fpo->fast_reg.fpo_pool_size = 0;
1509 for (i = 0; i < fps->fps_pool_size; i++) {
1510 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1513 CERROR("Failed to allocate a new fast_reg descriptor\n");
1519 #ifndef HAVE_IB_MAP_MR_SG
1520 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1521 LNET_MAX_PAYLOAD/PAGE_SIZE);
1522 if (IS_ERR(frd->frd_frpl)) {
1523 rc = PTR_ERR(frd->frd_frpl);
1524 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1526 frd->frd_frpl = NULL;
1531 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1532 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1533 LNET_MAX_PAYLOAD/PAGE_SIZE);
1535 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1537 LNET_MAX_PAYLOAD/PAGE_SIZE);
1539 if (IS_ERR(frd->frd_mr)) {
1540 rc = PTR_ERR(frd->frd_mr);
1541 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1546 /* There appears to be a bug in MLX5 code where you must
1547 * invalidate the rkey of a new FastReg pool before first
1548 * using it. Thus, I am marking the FRD invalid here. */
1549 frd->frd_valid = false;
1551 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1552 fpo->fast_reg.fpo_pool_size++;
1559 ib_dereg_mr(frd->frd_mr);
1560 #ifndef HAVE_IB_MAP_MR_SG
1562 ib_free_fast_reg_page_list(frd->frd_frpl);
1564 LIBCFS_FREE(frd, sizeof(*frd));
1567 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1569 list_del(&frd->frd_list);
1570 #ifndef HAVE_IB_MAP_MR_SG
1571 ib_free_fast_reg_page_list(frd->frd_frpl);
1573 ib_dereg_mr(frd->frd_mr);
1574 LIBCFS_FREE(frd, sizeof(*frd));
1581 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1583 struct ib_device_attr *dev_attr;
1584 kib_dev_t *dev = fps->fps_net->ibn_dev;
1585 kib_fmr_pool_t *fpo;
1588 #ifndef HAVE_IB_DEVICE_ATTRS
1589 dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1594 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1600 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1602 #ifdef HAVE_IB_DEVICE_ATTRS
1603 dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1605 rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1607 CERROR("Query device failed for %s: %d\n",
1608 fpo->fpo_hdev->ibh_ibdev->name, rc);
1613 /* Check for FMR or FastReg support */
1614 fpo->fpo_is_fmr = 0;
1615 if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1616 fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1617 fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1618 fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1619 LCONSOLE_INFO("Using FMR for registration\n");
1620 fpo->fpo_is_fmr = 1;
1621 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1622 LCONSOLE_INFO("Using FastReg for registration\n");
1625 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1629 if (fpo->fpo_is_fmr)
1630 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1632 rc = kiblnd_alloc_freg_pool(fps, fpo);
1636 #ifndef HAVE_IB_DEVICE_ATTRS
1639 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1640 fpo->fpo_owner = fps;
1646 kiblnd_hdev_decref(fpo->fpo_hdev);
1647 LIBCFS_FREE(fpo, sizeof(*fpo));
1650 #ifndef HAVE_IB_DEVICE_ATTRS
1658 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1660 if (fps->fps_net == NULL) /* intialized? */
1663 spin_lock(&fps->fps_lock);
1665 while (!list_empty(&fps->fps_pool_list)) {
1666 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1667 kib_fmr_pool_t, fpo_list);
1668 fpo->fpo_failed = 1;
1669 list_del(&fpo->fpo_list);
1670 if (fpo->fpo_map_count == 0)
1671 list_add(&fpo->fpo_list, zombies);
1673 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1676 spin_unlock(&fps->fps_lock);
1680 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1682 if (fps->fps_net != NULL) { /* initialized? */
1683 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1684 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1689 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1691 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1693 kib_fmr_pool_t *fpo;
1696 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1701 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1702 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1703 fps->fps_cache = tunables->lnd_fmr_cache;
1705 spin_lock_init(&fps->fps_lock);
1706 INIT_LIST_HEAD(&fps->fps_pool_list);
1707 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1709 rc = kiblnd_create_fmr_pool(fps, &fpo);
1711 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1717 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1719 if (fpo->fpo_map_count != 0) /* still in use */
1721 if (fpo->fpo_failed)
1723 return cfs_time_aftereq(now, fpo->fpo_deadline);
1727 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1729 kib_hca_dev_t *hdev;
1730 __u64 *pages = tx->tx_pages;
1735 hdev = tx->tx_pool->tpo_hdev;
1737 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1738 for (size = 0; size < rd->rd_frags[i].rf_nob;
1739 size += hdev->ibh_page_size) {
1740 pages[npages++] = (rd->rd_frags[i].rf_addr &
1741 hdev->ibh_page_mask) + size;
1749 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1751 struct list_head zombies = LIST_HEAD_INIT(zombies);
1752 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1753 kib_fmr_poolset_t *fps;
1754 cfs_time_t now = cfs_time_current();
1755 kib_fmr_pool_t *tmp;
1761 fps = fpo->fpo_owner;
1762 if (fpo->fpo_is_fmr) {
1763 if (fmr->fmr_pfmr) {
1764 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1766 fmr->fmr_pfmr = NULL;
1770 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1774 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1777 frd->frd_valid = false;
1778 spin_lock(&fps->fps_lock);
1779 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1780 spin_unlock(&fps->fps_lock);
1781 fmr->fmr_frd = NULL;
1784 fmr->fmr_pool = NULL;
1786 spin_lock(&fps->fps_lock);
1787 fpo->fpo_map_count--; /* decref the pool */
1789 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1790 /* the first pool is persistent */
1791 if (fps->fps_pool_list.next == &fpo->fpo_list)
1794 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1795 list_move(&fpo->fpo_list, &zombies);
1799 spin_unlock(&fps->fps_lock);
1801 if (!list_empty(&zombies))
1802 kiblnd_destroy_fmr_pool_list(&zombies);
1806 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1807 __u32 nob, __u64 iov, kib_fmr_t *fmr)
1809 kib_fmr_pool_t *fpo;
1810 __u64 *pages = tx->tx_pages;
1812 bool is_rx = (rd != tx->tx_rd);
1813 bool tx_pages_mapped = 0;
1818 spin_lock(&fps->fps_lock);
1819 version = fps->fps_version;
1820 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1821 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1822 fpo->fpo_map_count++;
1824 if (fpo->fpo_is_fmr) {
1825 struct ib_pool_fmr *pfmr;
1827 spin_unlock(&fps->fps_lock);
1829 if (!tx_pages_mapped) {
1830 npages = kiblnd_map_tx_pages(tx, rd);
1831 tx_pages_mapped = 1;
1834 pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1835 pages, npages, iov);
1836 if (likely(!IS_ERR(pfmr))) {
1837 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1839 fmr->fmr_frd = NULL;
1840 fmr->fmr_pfmr = pfmr;
1841 fmr->fmr_pool = fpo;
1846 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1847 struct kib_fast_reg_descriptor *frd;
1848 #ifdef HAVE_IB_MAP_MR_SG
1849 struct ib_reg_wr *wr;
1852 struct ib_rdma_wr *wr;
1853 struct ib_fast_reg_page_list *frpl;
1857 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1858 struct kib_fast_reg_descriptor,
1860 list_del(&frd->frd_list);
1861 spin_unlock(&fps->fps_lock);
1863 #ifndef HAVE_IB_MAP_MR_SG
1864 frpl = frd->frd_frpl;
1868 if (!frd->frd_valid) {
1869 struct ib_rdma_wr *inv_wr;
1870 __u32 key = is_rx ? mr->rkey : mr->lkey;
1872 inv_wr = &frd->frd_inv_wr;
1873 memset(inv_wr, 0, sizeof(*inv_wr));
1875 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1876 inv_wr->wr.wr_id = IBLND_WID_MR;
1877 inv_wr->wr.ex.invalidate_rkey = key;
1880 key = ib_inc_rkey(key);
1881 ib_update_fast_reg_key(mr, key);
1884 #ifdef HAVE_IB_MAP_MR_SG
1885 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1886 n = ib_map_mr_sg(mr, tx->tx_frags,
1887 tx->tx_nfrags, NULL, PAGE_SIZE);
1889 n = ib_map_mr_sg(mr, tx->tx_frags,
1890 tx->tx_nfrags, PAGE_SIZE);
1892 if (unlikely(n != tx->tx_nfrags)) {
1893 CERROR("Failed to map mr %d/%d "
1894 "elements\n", n, tx->tx_nfrags);
1895 return n < 0 ? n : -EINVAL;
1900 wr = &frd->frd_fastreg_wr;
1901 memset(wr, 0, sizeof(*wr));
1903 wr->wr.opcode = IB_WR_REG_MR;
1904 wr->wr.wr_id = IBLND_WID_MR;
1906 wr->wr.send_flags = 0;
1908 wr->key = is_rx ? mr->rkey : mr->lkey;
1909 wr->access = (IB_ACCESS_LOCAL_WRITE |
1910 IB_ACCESS_REMOTE_WRITE);
1912 if (!tx_pages_mapped) {
1913 npages = kiblnd_map_tx_pages(tx, rd);
1914 tx_pages_mapped = 1;
1917 LASSERT(npages <= frpl->max_page_list_len);
1918 memcpy(frpl->page_list, pages,
1919 sizeof(*pages) * npages);
1921 /* Prepare FastReg WR */
1922 wr = &frd->frd_fastreg_wr;
1923 memset(wr, 0, sizeof(*wr));
1925 wr->wr.opcode = IB_WR_FAST_REG_MR;
1926 wr->wr.wr_id = IBLND_WID_MR;
1928 wr->wr.wr.fast_reg.iova_start = iov;
1929 wr->wr.wr.fast_reg.page_list = frpl;
1930 wr->wr.wr.fast_reg.page_list_len = npages;
1931 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1932 wr->wr.wr.fast_reg.length = nob;
1933 wr->wr.wr.fast_reg.rkey =
1934 is_rx ? mr->rkey : mr->lkey;
1935 wr->wr.wr.fast_reg.access_flags =
1936 (IB_ACCESS_LOCAL_WRITE |
1937 IB_ACCESS_REMOTE_WRITE);
1940 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1942 fmr->fmr_pfmr = NULL;
1943 fmr->fmr_pool = fpo;
1946 spin_unlock(&fps->fps_lock);
1950 spin_lock(&fps->fps_lock);
1951 fpo->fpo_map_count--;
1952 if (rc != -EAGAIN) {
1953 spin_unlock(&fps->fps_lock);
1957 /* EAGAIN and ... */
1958 if (version != fps->fps_version) {
1959 spin_unlock(&fps->fps_lock);
1964 if (fps->fps_increasing) {
1965 spin_unlock(&fps->fps_lock);
1966 CDEBUG(D_NET, "Another thread is allocating new "
1967 "FMR pool, waiting for her to complete\n");
1973 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1974 /* someone failed recently */
1975 spin_unlock(&fps->fps_lock);
1979 fps->fps_increasing = 1;
1980 spin_unlock(&fps->fps_lock);
1982 CDEBUG(D_NET, "Allocate new FMR pool\n");
1983 rc = kiblnd_create_fmr_pool(fps, &fpo);
1984 spin_lock(&fps->fps_lock);
1985 fps->fps_increasing = 0;
1988 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1990 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1992 spin_unlock(&fps->fps_lock);
1998 kiblnd_fini_pool(kib_pool_t *pool)
2000 LASSERT(list_empty(&pool->po_free_list));
2001 LASSERT(pool->po_allocated == 0);
2003 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2007 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2009 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2011 memset(pool, 0, sizeof(kib_pool_t));
2012 INIT_LIST_HEAD(&pool->po_free_list);
2013 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2014 pool->po_owner = ps;
2015 pool->po_size = size;
2019 kiblnd_destroy_pool_list(struct list_head *head)
2023 while (!list_empty(head)) {
2024 pool = list_entry(head->next, kib_pool_t, po_list);
2025 list_del(&pool->po_list);
2027 LASSERT(pool->po_owner != NULL);
2028 pool->po_owner->ps_pool_destroy(pool);
2033 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2035 if (ps->ps_net == NULL) /* intialized? */
2038 spin_lock(&ps->ps_lock);
2039 while (!list_empty(&ps->ps_pool_list)) {
2040 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2041 kib_pool_t, po_list);
2043 list_del(&po->po_list);
2044 if (po->po_allocated == 0)
2045 list_add(&po->po_list, zombies);
2047 list_add(&po->po_list, &ps->ps_failed_pool_list);
2049 spin_unlock(&ps->ps_lock);
2053 kiblnd_fini_poolset(kib_poolset_t *ps)
2055 if (ps->ps_net != NULL) { /* initialized? */
2056 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2057 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2062 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2063 kib_net_t *net, char *name, int size,
2064 kib_ps_pool_create_t po_create,
2065 kib_ps_pool_destroy_t po_destroy,
2066 kib_ps_node_init_t nd_init,
2067 kib_ps_node_fini_t nd_fini)
2072 memset(ps, 0, sizeof(kib_poolset_t));
2076 ps->ps_pool_create = po_create;
2077 ps->ps_pool_destroy = po_destroy;
2078 ps->ps_node_init = nd_init;
2079 ps->ps_node_fini = nd_fini;
2080 ps->ps_pool_size = size;
2081 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2082 >= sizeof(ps->ps_name))
2084 spin_lock_init(&ps->ps_lock);
2085 INIT_LIST_HEAD(&ps->ps_pool_list);
2086 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2088 rc = ps->ps_pool_create(ps, size, &pool);
2090 list_add(&pool->po_list, &ps->ps_pool_list);
2092 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2098 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2100 if (pool->po_allocated != 0) /* still in use */
2102 if (pool->po_failed)
2104 return cfs_time_aftereq(now, pool->po_deadline);
2108 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2110 struct list_head zombies = LIST_HEAD_INIT(zombies);
2111 kib_poolset_t *ps = pool->po_owner;
2113 cfs_time_t now = cfs_time_current();
2115 spin_lock(&ps->ps_lock);
2117 if (ps->ps_node_fini != NULL)
2118 ps->ps_node_fini(pool, node);
2120 LASSERT(pool->po_allocated > 0);
2121 list_add(node, &pool->po_free_list);
2122 pool->po_allocated--;
2124 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2125 /* the first pool is persistent */
2126 if (ps->ps_pool_list.next == &pool->po_list)
2129 if (kiblnd_pool_is_idle(pool, now))
2130 list_move(&pool->po_list, &zombies);
2132 spin_unlock(&ps->ps_lock);
2134 if (!list_empty(&zombies))
2135 kiblnd_destroy_pool_list(&zombies);
2139 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2141 struct list_head *node;
2144 unsigned int interval = 1;
2145 cfs_time_t time_before;
2146 unsigned int trips = 0;
2149 spin_lock(&ps->ps_lock);
2150 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2151 if (list_empty(&pool->po_free_list))
2154 pool->po_allocated++;
2155 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2156 node = pool->po_free_list.next;
2159 if (ps->ps_node_init != NULL) {
2160 /* still hold the lock */
2161 ps->ps_node_init(pool, node);
2163 spin_unlock(&ps->ps_lock);
2167 /* no available tx pool and ... */
2168 if (ps->ps_increasing) {
2169 /* another thread is allocating a new pool */
2170 spin_unlock(&ps->ps_lock);
2172 CDEBUG(D_NET, "Another thread is allocating new "
2173 "%s pool, waiting %d HZs for her to complete."
2175 ps->ps_name, interval, trips);
2177 set_current_state(TASK_INTERRUPTIBLE);
2178 schedule_timeout(interval);
2179 if (interval < cfs_time_seconds(1))
2185 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2186 /* someone failed recently */
2187 spin_unlock(&ps->ps_lock);
2191 ps->ps_increasing = 1;
2192 spin_unlock(&ps->ps_lock);
2194 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2195 time_before = cfs_time_current();
2196 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2197 CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2198 cfs_time_current() - time_before);
2200 spin_lock(&ps->ps_lock);
2201 ps->ps_increasing = 0;
2203 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2205 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2206 CERROR("Can't allocate new %s pool because out of memory\n",
2209 spin_unlock(&ps->ps_lock);
2215 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2217 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2220 LASSERT (pool->po_allocated == 0);
2222 if (tpo->tpo_tx_pages != NULL) {
2223 kiblnd_unmap_tx_pool(tpo);
2224 kiblnd_free_pages(tpo->tpo_tx_pages);
2227 if (tpo->tpo_tx_descs == NULL)
2230 for (i = 0; i < pool->po_size; i++) {
2231 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2233 list_del(&tx->tx_list);
2234 if (tx->tx_pages != NULL)
2235 LIBCFS_FREE(tx->tx_pages,
2237 sizeof(*tx->tx_pages));
2238 if (tx->tx_frags != NULL)
2239 LIBCFS_FREE(tx->tx_frags,
2240 (1 + IBLND_MAX_RDMA_FRAGS) *
2241 sizeof(*tx->tx_frags));
2242 if (tx->tx_wrq != NULL)
2243 LIBCFS_FREE(tx->tx_wrq,
2244 (1 + IBLND_MAX_RDMA_FRAGS) *
2245 sizeof(*tx->tx_wrq));
2246 if (tx->tx_sge != NULL)
2247 LIBCFS_FREE(tx->tx_sge,
2248 (1 + IBLND_MAX_RDMA_FRAGS) *
2249 sizeof(*tx->tx_sge));
2250 if (tx->tx_rd != NULL)
2251 LIBCFS_FREE(tx->tx_rd,
2252 offsetof(kib_rdma_desc_t,
2253 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2256 LIBCFS_FREE(tpo->tpo_tx_descs,
2257 pool->po_size * sizeof(kib_tx_t));
2259 kiblnd_fini_pool(pool);
2260 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2263 static int kiblnd_tx_pool_size(int ncpts)
2265 int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2267 return max(IBLND_TX_POOL, ntx);
2271 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2278 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2280 CERROR("Failed to allocate TX pool\n");
2284 pool = &tpo->tpo_pool;
2285 kiblnd_init_pool(ps, pool, size);
2286 tpo->tpo_tx_descs = NULL;
2287 tpo->tpo_tx_pages = NULL;
2289 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2290 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2291 CERROR("Can't allocate tx pages: %d\n", npg);
2292 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2296 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2297 size * sizeof(kib_tx_t));
2298 if (tpo->tpo_tx_descs == NULL) {
2299 CERROR("Can't allocate %d tx descriptors\n", size);
2300 ps->ps_pool_destroy(pool);
2304 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2306 for (i = 0; i < size; i++) {
2307 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2310 if (ps->ps_net->ibn_fmr_ps != NULL) {
2311 LIBCFS_CPT_ALLOC(tx->tx_pages,
2312 lnet_cpt_table(), ps->ps_cpt,
2313 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2314 if (tx->tx_pages == NULL)
2318 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2319 (1 + IBLND_MAX_RDMA_FRAGS) *
2320 sizeof(*tx->tx_frags));
2321 if (tx->tx_frags == NULL)
2324 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2326 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2327 (1 + IBLND_MAX_RDMA_FRAGS) *
2328 sizeof(*tx->tx_wrq));
2329 if (tx->tx_wrq == NULL)
2332 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2333 (1 + IBLND_MAX_RDMA_FRAGS) *
2334 sizeof(*tx->tx_sge));
2335 if (tx->tx_sge == NULL)
2338 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2339 offsetof(kib_rdma_desc_t,
2340 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2341 if (tx->tx_rd == NULL)
2346 kiblnd_map_tx_pool(tpo);
2351 ps->ps_pool_destroy(pool);
2356 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2358 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2360 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
2362 tx->tx_cookie = tps->tps_next_tx_cookie++;
2366 kiblnd_net_fini_pools(kib_net_t *net)
2370 cfs_cpt_for_each(i, lnet_cpt_table()) {
2371 kib_tx_poolset_t *tps;
2372 kib_fmr_poolset_t *fps;
2374 if (net->ibn_tx_ps != NULL) {
2375 tps = net->ibn_tx_ps[i];
2376 kiblnd_fini_poolset(&tps->tps_poolset);
2379 if (net->ibn_fmr_ps != NULL) {
2380 fps = net->ibn_fmr_ps[i];
2381 kiblnd_fini_fmr_poolset(fps);
2385 if (net->ibn_tx_ps != NULL) {
2386 cfs_percpt_free(net->ibn_tx_ps);
2387 net->ibn_tx_ps = NULL;
2390 if (net->ibn_fmr_ps != NULL) {
2391 cfs_percpt_free(net->ibn_fmr_ps);
2392 net->ibn_fmr_ps = NULL;
2397 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2400 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2401 unsigned long flags;
2406 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2408 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2409 if (tunables->lnd_map_on_demand == 0) {
2410 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2412 goto create_tx_pool;
2415 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2417 if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2418 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2419 tunables->lnd_fmr_pool_size,
2420 *kiblnd_tunables.kib_ntx / 4);
2425 /* TX pool must be created later than FMR, see LU-2268
2427 LASSERT(net->ibn_tx_ps == NULL);
2429 /* premapping can fail if ibd_nmr > 1, so we always create
2430 * FMR pool and map-on-demand if premapping failed */
2432 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2433 sizeof(kib_fmr_poolset_t));
2434 if (net->ibn_fmr_ps == NULL) {
2435 CERROR("Failed to allocate FMR pool array\n");
2440 for (i = 0; i < ncpts; i++) {
2441 cpt = (cpts == NULL) ? i : cpts[i];
2442 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2445 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2452 LASSERT(i == ncpts);
2455 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2456 sizeof(kib_tx_poolset_t));
2457 if (net->ibn_tx_ps == NULL) {
2458 CERROR("Failed to allocate tx pool array\n");
2463 for (i = 0; i < ncpts; i++) {
2464 cpt = (cpts == NULL) ? i : cpts[i];
2465 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2467 kiblnd_tx_pool_size(ncpts),
2468 kiblnd_create_tx_pool,
2469 kiblnd_destroy_tx_pool,
2470 kiblnd_tx_init, NULL);
2472 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2480 kiblnd_net_fini_pools(net);
2486 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2488 #ifndef HAVE_IB_DEVICE_ATTRS
2489 struct ib_device_attr *attr;
2493 /* It's safe to assume a HCA can handle a page size
2494 * matching that of the native system */
2495 hdev->ibh_page_shift = PAGE_SHIFT;
2496 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2497 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2499 #ifdef HAVE_IB_DEVICE_ATTRS
2500 hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2502 LIBCFS_ALLOC(attr, sizeof(*attr));
2504 CERROR("Out of memory\n");
2508 rc = ib_query_device(hdev->ibh_ibdev, attr);
2510 hdev->ibh_mr_size = attr->max_mr_size;
2512 LIBCFS_FREE(attr, sizeof(*attr));
2515 CERROR("Failed to query IB device: %d\n", rc);
2520 if (hdev->ibh_mr_size == ~0ULL) {
2521 hdev->ibh_mr_shift = 64;
2525 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2530 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2532 if (hdev->ibh_mrs == NULL)
2535 ib_dereg_mr(hdev->ibh_mrs);
2537 hdev->ibh_mrs = NULL;
2541 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2543 kiblnd_hdev_cleanup_mrs(hdev);
2545 if (hdev->ibh_pd != NULL)
2546 ib_dealloc_pd(hdev->ibh_pd);
2548 if (hdev->ibh_cmid != NULL)
2549 rdma_destroy_id(hdev->ibh_cmid);
2551 LIBCFS_FREE(hdev, sizeof(*hdev));
2555 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2559 int acflags = IB_ACCESS_LOCAL_WRITE |
2560 IB_ACCESS_REMOTE_WRITE;
2562 rc = kiblnd_hdev_get_attr(hdev);
2566 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2568 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2569 kiblnd_hdev_cleanup_mrs(hdev);
2579 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2585 kiblnd_dev_need_failover(kib_dev_t *dev)
2587 struct rdma_cm_id *cmid;
2588 struct sockaddr_in srcaddr;
2589 struct sockaddr_in dstaddr;
2592 if (dev->ibd_hdev == NULL || /* initializing */
2593 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2594 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2597 /* XXX: it's UGLY, but I don't have better way to find
2598 * ib-bonding HCA failover because:
2600 * a. no reliable CM event for HCA failover...
2601 * b. no OFED API to get ib_device for current net_device...
2603 * We have only two choices at this point:
2605 * a. rdma_bind_addr(), it will conflict with listener cmid
2606 * b. rdma_resolve_addr() to zero addr */
2607 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2611 CERROR("Failed to create cmid for failover: %d\n", rc);
2615 memset(&srcaddr, 0, sizeof(srcaddr));
2616 srcaddr.sin_family = AF_INET;
2617 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2619 memset(&dstaddr, 0, sizeof(dstaddr));
2620 dstaddr.sin_family = AF_INET;
2621 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2622 (struct sockaddr *)&dstaddr, 1);
2623 if (rc != 0 || cmid->device == NULL) {
2624 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2625 dev->ibd_ifname, &dev->ibd_ifip,
2627 rdma_destroy_id(cmid);
2631 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2632 rdma_destroy_id(cmid);
2637 kiblnd_dev_failover(kib_dev_t *dev)
2639 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2640 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2641 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2642 struct rdma_cm_id *cmid = NULL;
2643 kib_hca_dev_t *hdev = NULL;
2647 struct sockaddr_in addr;
2648 unsigned long flags;
2652 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2653 dev->ibd_can_failover ||
2654 dev->ibd_hdev == NULL);
2656 rc = kiblnd_dev_need_failover(dev);
2660 if (dev->ibd_hdev != NULL &&
2661 dev->ibd_hdev->ibh_cmid != NULL) {
2662 /* XXX it's not good to close old listener at here,
2663 * because we can fail to create new listener.
2664 * But we have to close it now, otherwise rdma_bind_addr
2665 * will return EADDRINUSE... How crap! */
2666 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2668 cmid = dev->ibd_hdev->ibh_cmid;
2669 /* make next schedule of kiblnd_dev_need_failover()
2670 * return 1 for me */
2671 dev->ibd_hdev->ibh_cmid = NULL;
2672 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2674 rdma_destroy_id(cmid);
2677 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2681 CERROR("Failed to create cmid for failover: %d\n", rc);
2685 memset(&addr, 0, sizeof(addr));
2686 addr.sin_family = AF_INET;
2687 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2688 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2690 /* Bind to failover device or port */
2691 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2692 if (rc != 0 || cmid->device == NULL) {
2693 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2694 dev->ibd_ifname, &dev->ibd_ifip,
2696 rdma_destroy_id(cmid);
2700 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2702 CERROR("Failed to allocate kib_hca_dev\n");
2703 rdma_destroy_id(cmid);
2708 atomic_set(&hdev->ibh_ref, 1);
2709 hdev->ibh_dev = dev;
2710 hdev->ibh_cmid = cmid;
2711 hdev->ibh_ibdev = cmid->device;
2713 pd = ib_alloc_pd(cmid->device);
2716 CERROR("Can't allocate PD: %d\n", rc);
2722 rc = rdma_listen(cmid, 0);
2724 CERROR("Can't start new listener: %d\n", rc);
2728 rc = kiblnd_hdev_setup_mrs(hdev);
2730 CERROR("Can't setup device: %d\n", rc);
2734 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2736 old = dev->ibd_hdev;
2737 dev->ibd_hdev = hdev; /* take over the refcount */
2740 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2741 cfs_cpt_for_each(i, lnet_cpt_table()) {
2742 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2745 if (net->ibn_fmr_ps != NULL)
2746 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2751 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2753 if (!list_empty(&zombie_tpo))
2754 kiblnd_destroy_pool_list(&zombie_tpo);
2755 if (!list_empty(&zombie_ppo))
2756 kiblnd_destroy_pool_list(&zombie_ppo);
2757 if (!list_empty(&zombie_fpo))
2758 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2760 kiblnd_hdev_decref(hdev);
2763 dev->ibd_failed_failover++;
2765 dev->ibd_failed_failover = 0;
2771 kiblnd_destroy_dev (kib_dev_t *dev)
2773 LASSERT (dev->ibd_nnets == 0);
2774 LASSERT(list_empty(&dev->ibd_nets));
2776 list_del(&dev->ibd_fail_list);
2777 list_del(&dev->ibd_list);
2779 if (dev->ibd_hdev != NULL)
2780 kiblnd_hdev_decref(dev->ibd_hdev);
2782 LIBCFS_FREE(dev, sizeof(*dev));
2786 kiblnd_create_dev(char *ifname)
2788 struct net_device *netdev;
2795 rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2797 CERROR("Can't query IPoIB interface %s: %d\n",
2803 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2807 LIBCFS_ALLOC(dev, sizeof(*dev));
2811 netdev = dev_get_by_name(&init_net, ifname);
2812 if (netdev == NULL) {
2813 dev->ibd_can_failover = 0;
2815 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2819 INIT_LIST_HEAD(&dev->ibd_nets);
2820 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2821 INIT_LIST_HEAD(&dev->ibd_fail_list);
2823 strcpy(&dev->ibd_ifname[0], ifname);
2825 /* initialize the device */
2826 rc = kiblnd_dev_failover(dev);
2828 CERROR("Can't initialize device: %d\n", rc);
2829 LIBCFS_FREE(dev, sizeof(*dev));
2833 list_add_tail(&dev->ibd_list,
2834 &kiblnd_data.kib_devs);
2839 kiblnd_base_shutdown(void)
2841 struct kib_sched_info *sched;
2844 LASSERT(list_empty(&kiblnd_data.kib_devs));
2846 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2847 atomic_read(&libcfs_kmemory));
2849 switch (kiblnd_data.kib_init) {
2853 case IBLND_INIT_ALL:
2854 case IBLND_INIT_DATA:
2855 LASSERT (kiblnd_data.kib_peers != NULL);
2856 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2857 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2859 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2860 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2861 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2862 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2864 /* flag threads to terminate; wake and wait for them to die */
2865 kiblnd_data.kib_shutdown = 1;
2867 /* NB: we really want to stop scheduler threads net by net
2868 * instead of the whole module, this should be improved
2869 * with dynamic configuration LNet */
2870 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2871 wake_up_all(&sched->ibs_waitq);
2873 wake_up_all(&kiblnd_data.kib_connd_waitq);
2874 wake_up_all(&kiblnd_data.kib_failover_waitq);
2877 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2880 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2881 "Waiting for %d threads to terminate\n",
2882 atomic_read(&kiblnd_data.kib_nthreads));
2883 set_current_state(TASK_UNINTERRUPTIBLE);
2884 schedule_timeout(cfs_time_seconds(1));
2889 case IBLND_INIT_NOTHING:
2893 if (kiblnd_data.kib_peers != NULL) {
2894 LIBCFS_FREE(kiblnd_data.kib_peers,
2895 sizeof(struct list_head) *
2896 kiblnd_data.kib_peer_hash_size);
2899 if (kiblnd_data.kib_scheds != NULL)
2900 cfs_percpt_free(kiblnd_data.kib_scheds);
2902 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2903 atomic_read(&libcfs_kmemory));
2905 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2906 module_put(THIS_MODULE);
2910 kiblnd_shutdown(struct lnet_ni *ni)
2912 kib_net_t *net = ni->ni_data;
2913 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2915 unsigned long flags;
2917 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2922 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2923 atomic_read(&libcfs_kmemory));
2925 write_lock_irqsave(g_lock, flags);
2926 net->ibn_shutdown = 1;
2927 write_unlock_irqrestore(g_lock, flags);
2929 switch (net->ibn_init) {
2933 case IBLND_INIT_ALL:
2934 /* nuke all existing peers within this net */
2935 kiblnd_del_peer(ni, LNET_NID_ANY);
2937 /* Wait for all peer_ni state to clean up */
2939 while (atomic_read(&net->ibn_npeers) != 0) {
2942 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2943 "%s: waiting for %d peers to disconnect\n",
2944 libcfs_nid2str(ni->ni_nid),
2945 atomic_read(&net->ibn_npeers));
2946 set_current_state(TASK_UNINTERRUPTIBLE);
2947 schedule_timeout(cfs_time_seconds(1));
2950 kiblnd_net_fini_pools(net);
2952 write_lock_irqsave(g_lock, flags);
2953 LASSERT(net->ibn_dev->ibd_nnets > 0);
2954 net->ibn_dev->ibd_nnets--;
2955 list_del(&net->ibn_list);
2956 write_unlock_irqrestore(g_lock, flags);
2960 case IBLND_INIT_NOTHING:
2961 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2963 if (net->ibn_dev != NULL &&
2964 net->ibn_dev->ibd_nnets == 0)
2965 kiblnd_destroy_dev(net->ibn_dev);
2970 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2971 atomic_read(&libcfs_kmemory));
2973 net->ibn_init = IBLND_INIT_NOTHING;
2976 LIBCFS_FREE(net, sizeof(*net));
2979 if (list_empty(&kiblnd_data.kib_devs))
2980 kiblnd_base_shutdown();
2985 kiblnd_base_startup(void)
2987 struct kib_sched_info *sched;
2991 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2993 try_module_get(THIS_MODULE);
2994 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2996 rwlock_init(&kiblnd_data.kib_global_lock);
2998 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2999 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3001 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3002 LIBCFS_ALLOC(kiblnd_data.kib_peers,
3003 sizeof(struct list_head) *
3004 kiblnd_data.kib_peer_hash_size);
3005 if (kiblnd_data.kib_peers == NULL)
3008 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3009 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3011 spin_lock_init(&kiblnd_data.kib_connd_lock);
3012 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3013 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3014 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3015 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3017 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3018 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3020 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3022 if (kiblnd_data.kib_scheds == NULL)
3025 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3028 spin_lock_init(&sched->ibs_lock);
3029 INIT_LIST_HEAD(&sched->ibs_conns);
3030 init_waitqueue_head(&sched->ibs_waitq);
3032 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3033 if (*kiblnd_tunables.kib_nscheds > 0) {
3034 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3036 /* max to half of CPUs, another half is reserved for
3037 * upper layer modules */
3038 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3041 sched->ibs_nthreads_max = nthrs;
3045 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3047 /* lists/ptrs/locks initialised */
3048 kiblnd_data.kib_init = IBLND_INIT_DATA;
3049 /*****************************************************/
3051 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3053 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3057 if (*kiblnd_tunables.kib_dev_failover != 0)
3058 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3062 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3066 /* flag everything initialised */
3067 kiblnd_data.kib_init = IBLND_INIT_ALL;
3068 /*****************************************************/
3073 kiblnd_base_shutdown();
3078 kiblnd_start_schedulers(struct kib_sched_info *sched)
3084 if (sched->ibs_nthreads == 0) {
3085 if (*kiblnd_tunables.kib_nscheds > 0) {
3086 nthrs = sched->ibs_nthreads_max;
3088 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3090 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3091 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3094 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3095 /* increase one thread if there is new interface */
3096 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3099 for (i = 0; i < nthrs; i++) {
3102 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3103 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3104 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3105 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3109 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3110 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3114 sched->ibs_nthreads += i;
3119 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3125 for (i = 0; i < ncpts; i++) {
3126 struct kib_sched_info *sched;
3128 cpt = (cpts == NULL) ? i : cpts[i];
3129 sched = kiblnd_data.kib_scheds[cpt];
3131 if (!newdev && sched->ibs_nthreads > 0)
3134 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3136 CERROR("Failed to start scheduler threads for %s\n",
3145 kiblnd_dev_search(char *ifname)
3147 kib_dev_t *alias = NULL;
3152 colon = strchr(ifname, ':');
3153 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3154 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3160 colon2 = strchr(dev->ibd_ifname, ':');
3166 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3178 kiblnd_startup(struct lnet_ni *ni)
3181 kib_dev_t *ibdev = NULL;
3183 unsigned long flags;
3188 LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3190 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3191 rc = kiblnd_base_startup();
3196 LIBCFS_ALLOC(net, sizeof(*net));
3201 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3203 kiblnd_tunables_setup(ni);
3205 if (ni->ni_interfaces[0] != NULL) {
3206 /* Use the IPoIB interface specified in 'networks=' */
3208 CLASSERT (LNET_MAX_INTERFACES > 1);
3209 if (ni->ni_interfaces[1] != NULL) {
3210 CERROR("Multiple interfaces not supported\n");
3214 ifname = ni->ni_interfaces[0];
3216 ifname = *kiblnd_tunables.kib_default_ipif;
3219 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3220 CERROR("IPoIB interface name too long: %s\n", ifname);
3224 ibdev = kiblnd_dev_search(ifname);
3226 newdev = ibdev == NULL;
3227 /* hmm...create kib_dev even for alias */
3228 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3229 ibdev = kiblnd_create_dev(ifname);
3234 node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3235 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3237 net->ibn_dev = ibdev;
3238 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3240 rc = kiblnd_dev_start_threads(ibdev, newdev,
3241 ni->ni_cpts, ni->ni_ncpts);
3245 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3247 CERROR("Failed to initialize NI pools: %d\n", rc);
3251 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3253 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3254 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3256 net->ibn_init = IBLND_INIT_ALL;
3261 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3262 kiblnd_destroy_dev(ibdev);
3264 kiblnd_shutdown(ni);
3266 CDEBUG(D_NET, "kiblnd_startup failed\n");
3270 static struct lnet_lnd the_o2iblnd = {
3271 .lnd_type = O2IBLND,
3272 .lnd_startup = kiblnd_startup,
3273 .lnd_shutdown = kiblnd_shutdown,
3274 .lnd_ctl = kiblnd_ctl,
3275 .lnd_query = kiblnd_query,
3276 .lnd_send = kiblnd_send,
3277 .lnd_recv = kiblnd_recv,
3280 static void __exit ko2iblnd_exit(void)
3282 lnet_unregister_lnd(&the_o2iblnd);
3285 static int __init ko2iblnd_init(void)
3289 CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3290 CLASSERT(offsetof(kib_msg_t,
3291 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3293 CLASSERT(offsetof(kib_msg_t,
3294 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3297 rc = kiblnd_tunables_init();
3301 lnet_register_lnd(&the_o2iblnd);
3306 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3307 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3308 MODULE_VERSION("2.8.0");
3309 MODULE_LICENSE("GPL");
3311 module_init(ko2iblnd_init);
3312 module_exit(ko2iblnd_exit);