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(&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 * Get the scheduler bound to this CPT. If the scheduler has no
703 * threads, which means that the CPT has no CPUs, then grab the
704 * next scheduler that we can use.
706 * This case would be triggered if a NUMA node is configured with
707 * no associated CPUs.
709 static struct kib_sched_info *
710 kiblnd_get_scheduler(int cpt)
712 struct kib_sched_info *sched;
715 sched = kiblnd_data.kib_scheds[cpt];
717 if (sched->ibs_nthreads > 0)
720 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
721 if (sched->ibs_nthreads > 0) {
722 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
723 cpt, sched->ibs_cpt);
732 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
733 int state, int version)
736 * If the new conn is created successfully it takes over the caller's
737 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
738 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
739 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
740 * to destroy 'cmid' here since I'm called from the CM which still has
741 * its ref on 'cmid'). */
742 rwlock_t *glock = &kiblnd_data.kib_global_lock;
743 kib_net_t *net = peer_ni->ibp_ni->ni_data;
745 struct ib_qp_init_attr *init_qp_attr;
746 struct kib_sched_info *sched;
747 #ifdef HAVE_IB_CQ_INIT_ATTR
748 struct ib_cq_init_attr cq_attr = {};
757 LASSERT(net != NULL);
758 LASSERT(!in_interrupt());
762 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
763 sched = kiblnd_get_scheduler(cpt);
766 CERROR("no schedulers available. node is unhealthy\n");
771 * The cpt might have changed if we ended up selecting a non cpt
772 * native scheduler. So use the scheduler's cpt instead.
774 cpt = sched->ibs_cpt;
776 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
777 sizeof(*init_qp_attr));
778 if (init_qp_attr == NULL) {
779 CERROR("Can't allocate qp_attr for %s\n",
780 libcfs_nid2str(peer_ni->ibp_nid));
784 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
786 CERROR("Can't allocate connection for %s\n",
787 libcfs_nid2str(peer_ni->ibp_nid));
791 conn->ibc_state = IBLND_CONN_INIT;
792 conn->ibc_version = version;
793 conn->ibc_peer = peer_ni; /* I take the caller's ref */
794 cmid->context = conn; /* for future CM callbacks */
795 conn->ibc_cmid = cmid;
796 conn->ibc_max_frags = peer_ni->ibp_max_frags;
797 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
799 INIT_LIST_HEAD(&conn->ibc_early_rxs);
800 INIT_LIST_HEAD(&conn->ibc_tx_noops);
801 INIT_LIST_HEAD(&conn->ibc_tx_queue);
802 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
803 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
804 INIT_LIST_HEAD(&conn->ibc_active_txs);
805 spin_lock_init(&conn->ibc_lock);
807 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
808 sizeof(*conn->ibc_connvars));
809 if (conn->ibc_connvars == NULL) {
810 CERROR("Can't allocate in-progress connection state\n");
814 write_lock_irqsave(glock, flags);
815 if (dev->ibd_failover) {
816 write_unlock_irqrestore(glock, flags);
817 CERROR("%s: failover in progress\n", dev->ibd_ifname);
821 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
822 /* wakeup failover thread and teardown connection */
823 if (kiblnd_dev_can_failover(dev)) {
824 list_add_tail(&dev->ibd_fail_list,
825 &kiblnd_data.kib_failed_devs);
826 wake_up(&kiblnd_data.kib_failover_waitq);
829 write_unlock_irqrestore(glock, flags);
830 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
831 cmid->device->name, dev->ibd_ifname);
835 kiblnd_hdev_addref_locked(dev->ibd_hdev);
836 conn->ibc_hdev = dev->ibd_hdev;
838 kiblnd_setup_mtu_locked(cmid);
840 write_unlock_irqrestore(glock, flags);
842 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
843 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
844 if (conn->ibc_rxs == NULL) {
845 CERROR("Cannot allocate RX buffers\n");
849 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
850 IBLND_RX_MSG_PAGES(conn));
854 kiblnd_map_rx_descs(conn);
856 #ifdef HAVE_IB_CQ_INIT_ATTR
857 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
858 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
859 cq = ib_create_cq(cmid->device,
860 kiblnd_cq_completion, kiblnd_cq_event, conn,
863 cq = ib_create_cq(cmid->device,
864 kiblnd_cq_completion, kiblnd_cq_event, conn,
865 IBLND_CQ_ENTRIES(conn),
866 kiblnd_get_completion_vector(conn, cpt));
869 CERROR("Failed to create CQ with %d CQEs: %ld\n",
870 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
876 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
878 CERROR("Can't request completion notification: %d\n", rc);
882 init_qp_attr->event_handler = kiblnd_qp_event;
883 init_qp_attr->qp_context = conn;
884 init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
885 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
886 init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
887 init_qp_attr->cap.max_recv_sge = 1;
888 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
889 init_qp_attr->qp_type = IB_QPT_RC;
890 init_qp_attr->send_cq = cq;
891 init_qp_attr->recv_cq = cq;
893 conn->ibc_sched = sched;
896 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
897 if (!rc || init_qp_attr->cap.max_send_wr < 16)
900 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
904 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
905 "send_sge: %d, recv_sge: %d\n",
906 rc, init_qp_attr->cap.max_send_wr,
907 init_qp_attr->cap.max_recv_wr,
908 init_qp_attr->cap.max_send_sge,
909 init_qp_attr->cap.max_recv_sge);
913 if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
914 CDEBUG(D_NET, "original send wr %d, created with %d\n",
915 IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
917 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
919 /* 1 ref for caller and each rxmsg */
920 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
921 conn->ibc_nrx = IBLND_RX_MSGS(conn);
924 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
925 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
927 CERROR("Can't post rxmsg: %d\n", rc);
929 /* Make posted receives complete */
930 kiblnd_abort_receives(conn);
932 /* correct # of posted buffers
933 * NB locking needed now I'm racing with completion */
934 spin_lock_irqsave(&sched->ibs_lock, flags);
935 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
936 spin_unlock_irqrestore(&sched->ibs_lock, flags);
938 /* cmid will be destroyed by CM(ofed) after cm_callback
939 * returned, so we can't refer it anymore
940 * (by kiblnd_connd()->kiblnd_destroy_conn) */
941 rdma_destroy_qp(conn->ibc_cmid);
942 conn->ibc_cmid = NULL;
944 /* Drop my own and unused rxbuffer refcounts */
945 while (i++ <= IBLND_RX_MSGS(conn))
946 kiblnd_conn_decref(conn);
952 /* Init successful! */
953 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
954 state == IBLND_CONN_PASSIVE_WAIT);
955 conn->ibc_state = state;
958 atomic_inc(&net->ibn_nconns);
962 kiblnd_destroy_conn(conn, true);
964 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
970 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
972 struct rdma_cm_id *cmid = conn->ibc_cmid;
973 kib_peer_ni_t *peer_ni = conn->ibc_peer;
976 LASSERT (!in_interrupt());
977 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
978 LASSERT(list_empty(&conn->ibc_early_rxs));
979 LASSERT(list_empty(&conn->ibc_tx_noops));
980 LASSERT(list_empty(&conn->ibc_tx_queue));
981 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
982 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
983 LASSERT(list_empty(&conn->ibc_active_txs));
984 LASSERT (conn->ibc_noops_posted == 0);
985 LASSERT (conn->ibc_nsends_posted == 0);
987 switch (conn->ibc_state) {
989 /* conn must be completely disengaged from the network */
992 case IBLND_CONN_DISCONNECTED:
993 /* connvars should have been freed already */
994 LASSERT (conn->ibc_connvars == NULL);
997 case IBLND_CONN_INIT:
1001 /* conn->ibc_cmid might be destroyed by CM already */
1002 if (cmid != NULL && cmid->qp != NULL)
1003 rdma_destroy_qp(cmid);
1005 if (conn->ibc_cq != NULL) {
1006 rc = ib_destroy_cq(conn->ibc_cq);
1008 CWARN("Error destroying CQ: %d\n", rc);
1011 if (conn->ibc_rx_pages != NULL)
1012 kiblnd_unmap_rx_descs(conn);
1014 if (conn->ibc_rxs != NULL) {
1015 LIBCFS_FREE(conn->ibc_rxs,
1016 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
1019 if (conn->ibc_connvars != NULL)
1020 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1022 if (conn->ibc_hdev != NULL)
1023 kiblnd_hdev_decref(conn->ibc_hdev);
1025 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1026 if (conn->ibc_state != IBLND_CONN_INIT) {
1027 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1029 kiblnd_peer_decref(peer_ni);
1030 rdma_destroy_id(cmid);
1031 atomic_dec(&net->ibn_nconns);
1035 LIBCFS_FREE(conn, sizeof(*conn));
1039 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1042 struct list_head *ctmp;
1043 struct list_head *cnxt;
1046 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1047 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1049 CDEBUG(D_NET, "Closing conn -> %s, "
1050 "version: %x, reason: %d\n",
1051 libcfs_nid2str(peer_ni->ibp_nid),
1052 conn->ibc_version, why);
1054 kiblnd_close_conn_locked(conn, why);
1062 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1063 int version, __u64 incarnation)
1066 struct list_head *ctmp;
1067 struct list_head *cnxt;
1070 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1071 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1073 if (conn->ibc_version == version &&
1074 conn->ibc_incarnation == incarnation)
1077 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1078 "incarnation:%#llx(%x, %#llx)\n",
1079 libcfs_nid2str(peer_ni->ibp_nid),
1080 conn->ibc_version, conn->ibc_incarnation,
1081 version, incarnation);
1083 kiblnd_close_conn_locked(conn, -ESTALE);
1091 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1093 kib_peer_ni_t *peer_ni;
1094 struct list_head *ptmp;
1095 struct list_head *pnxt;
1099 unsigned long flags;
1102 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1104 if (nid != LNET_NID_ANY)
1105 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1108 hi = kiblnd_data.kib_peer_hash_size - 1;
1111 for (i = lo; i <= hi; i++) {
1112 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1114 peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1115 LASSERT(!kiblnd_peer_idle(peer_ni));
1117 if (peer_ni->ibp_ni != ni)
1120 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1123 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1127 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1129 /* wildcards always succeed */
1130 if (nid == LNET_NID_ANY)
1133 return (count == 0) ? -ENOENT : 0;
1137 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1139 struct libcfs_ioctl_data *data = arg;
1143 case IOC_LIBCFS_GET_PEER: {
1147 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1149 data->ioc_nid = nid;
1150 data->ioc_count = count;
1154 case IOC_LIBCFS_DEL_PEER: {
1155 rc = kiblnd_del_peer(ni, data->ioc_nid);
1158 case IOC_LIBCFS_GET_CONN: {
1162 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1168 LASSERT(conn->ibc_cmid != NULL);
1169 data->ioc_nid = conn->ibc_peer->ibp_nid;
1170 if (conn->ibc_cmid->route.path_rec == NULL)
1171 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1174 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1175 kiblnd_conn_decref(conn);
1178 case IOC_LIBCFS_CLOSE_CONNECTION: {
1179 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1191 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when)
1193 cfs_time_t last_alive = 0;
1194 cfs_time_t now = cfs_time_current();
1195 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1196 kib_peer_ni_t *peer_ni;
1197 unsigned long flags;
1199 read_lock_irqsave(glock, flags);
1201 peer_ni = kiblnd_find_peer_locked(ni, nid);
1202 if (peer_ni != NULL)
1203 last_alive = peer_ni->ibp_last_alive;
1205 read_unlock_irqrestore(glock, flags);
1207 if (last_alive != 0)
1210 /* peer_ni is not persistent in hash, trigger peer_ni creation
1211 * and connection establishment with a NULL tx */
1212 if (peer_ni == NULL)
1213 kiblnd_launch_tx(ni, NULL, nid);
1215 CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago\n",
1216 libcfs_nid2str(nid), peer_ni,
1217 last_alive ? cfs_duration_sec(now - last_alive) : -1);
1222 kiblnd_free_pages(kib_pages_t *p)
1224 int npages = p->ibp_npages;
1227 for (i = 0; i < npages; i++) {
1228 if (p->ibp_pages[i] != NULL)
1229 __free_page(p->ibp_pages[i]);
1232 LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1236 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1241 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1242 offsetof(kib_pages_t, ibp_pages[npages]));
1244 CERROR("Can't allocate descriptor for %d pages\n", npages);
1248 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1249 p->ibp_npages = npages;
1251 for (i = 0; i < npages; i++) {
1252 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1254 if (p->ibp_pages[i] == NULL) {
1255 CERROR("Can't allocate page %d of %d\n", i, npages);
1256 kiblnd_free_pages(p);
1266 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1271 LASSERT (conn->ibc_rxs != NULL);
1272 LASSERT (conn->ibc_hdev != NULL);
1274 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1275 rx = &conn->ibc_rxs[i];
1277 LASSERT(rx->rx_nob >= 0); /* not posted */
1279 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1280 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1282 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1285 kiblnd_free_pages(conn->ibc_rx_pages);
1287 conn->ibc_rx_pages = NULL;
1291 kiblnd_map_rx_descs(kib_conn_t *conn)
1299 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1300 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1301 rx = &conn->ibc_rxs[i];
1304 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1307 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1308 rx->rx_msg, IBLND_MSG_SIZE,
1310 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1312 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1314 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1315 i, rx->rx_msg, rx->rx_msgaddr,
1316 (__u64)(page_to_phys(pg) + pg_off));
1318 pg_off += IBLND_MSG_SIZE;
1319 LASSERT(pg_off <= PAGE_SIZE);
1321 if (pg_off == PAGE_SIZE) {
1324 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1330 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1332 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1336 LASSERT (tpo->tpo_pool.po_allocated == 0);
1341 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1342 tx = &tpo->tpo_tx_descs[i];
1343 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1344 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1346 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1349 kiblnd_hdev_decref(hdev);
1350 tpo->tpo_hdev = NULL;
1353 static kib_hca_dev_t *
1354 kiblnd_current_hdev(kib_dev_t *dev)
1356 kib_hca_dev_t *hdev;
1357 unsigned long flags;
1360 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1361 while (dev->ibd_failover) {
1362 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1364 CDEBUG(D_NET, "%s: Wait for failover\n",
1366 set_current_state(TASK_INTERRUPTIBLE);
1367 schedule_timeout(cfs_time_seconds(1) / 100);
1369 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1372 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1373 hdev = dev->ibd_hdev;
1375 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1381 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1383 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1384 kib_pool_t *pool = &tpo->tpo_pool;
1385 kib_net_t *net = pool->po_owner->ps_net;
1393 LASSERT (net != NULL);
1397 /* pre-mapped messages are not bigger than 1 page */
1398 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1400 /* No fancy arithmetic when we do the buffer calculations */
1401 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1403 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1405 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1406 page = txpgs->ibp_pages[ipage];
1407 tx = &tpo->tpo_tx_descs[i];
1409 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1412 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1416 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1418 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1420 list_add(&tx->tx_list, &pool->po_free_list);
1422 page_offset += IBLND_MSG_SIZE;
1423 LASSERT(page_offset <= PAGE_SIZE);
1425 if (page_offset == PAGE_SIZE) {
1428 LASSERT(ipage <= txpgs->ibp_npages);
1433 #ifdef HAVE_IB_GET_DMA_MR
1435 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1436 int negotiated_nfrags)
1438 kib_net_t *net = ni->ni_data;
1439 kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
1440 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1444 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1445 mod = tunables->lnd_map_on_demand;
1446 nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1448 LASSERT(hdev->ibh_mrs != NULL);
1450 if (mod > 0 && nfrags <= rd->rd_nfrags)
1453 return hdev->ibh_mrs;
1458 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1460 LASSERT(fpo->fpo_map_count == 0);
1462 if (fpo->fpo_is_fmr) {
1463 if (fpo->fmr.fpo_fmr_pool)
1464 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1466 struct kib_fast_reg_descriptor *frd, *tmp;
1469 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1471 list_del(&frd->frd_list);
1472 #ifndef HAVE_IB_MAP_MR_SG
1473 ib_free_fast_reg_page_list(frd->frd_frpl);
1475 ib_dereg_mr(frd->frd_mr);
1476 LIBCFS_FREE(frd, sizeof(*frd));
1479 if (i < fpo->fast_reg.fpo_pool_size)
1480 CERROR("FastReg pool still has %d regions registered\n",
1481 fpo->fast_reg.fpo_pool_size - i);
1485 kiblnd_hdev_decref(fpo->fpo_hdev);
1487 LIBCFS_FREE(fpo, sizeof(*fpo));
1491 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1493 kib_fmr_pool_t *fpo, *tmp;
1495 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1496 list_del(&fpo->fpo_list);
1497 kiblnd_destroy_fmr_pool(fpo);
1502 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1505 int size = tunables->lnd_fmr_pool_size / ncpts;
1507 return max(IBLND_FMR_POOL, size);
1511 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1514 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1516 return max(IBLND_FMR_POOL_FLUSH, size);
1519 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1521 struct ib_fmr_pool_param param = {
1522 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1523 .page_shift = PAGE_SHIFT,
1524 .access = (IB_ACCESS_LOCAL_WRITE |
1525 IB_ACCESS_REMOTE_WRITE),
1526 .pool_size = fps->fps_pool_size,
1527 .dirty_watermark = fps->fps_flush_trigger,
1528 .flush_function = NULL,
1530 .cache = !!fps->fps_cache };
1533 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1535 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1536 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1538 CERROR("Failed to create FMR pool: %d\n", rc);
1540 CERROR("FMRs are not supported\n");
1546 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1548 struct kib_fast_reg_descriptor *frd, *tmp;
1551 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1552 fpo->fast_reg.fpo_pool_size = 0;
1553 for (i = 0; i < fps->fps_pool_size; i++) {
1554 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1557 CERROR("Failed to allocate a new fast_reg descriptor\n");
1563 #ifndef HAVE_IB_MAP_MR_SG
1564 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1565 LNET_MAX_PAYLOAD/PAGE_SIZE);
1566 if (IS_ERR(frd->frd_frpl)) {
1567 rc = PTR_ERR(frd->frd_frpl);
1568 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1570 frd->frd_frpl = NULL;
1575 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1576 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1577 LNET_MAX_PAYLOAD/PAGE_SIZE);
1579 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1581 LNET_MAX_PAYLOAD/PAGE_SIZE);
1583 if (IS_ERR(frd->frd_mr)) {
1584 rc = PTR_ERR(frd->frd_mr);
1585 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1590 /* There appears to be a bug in MLX5 code where you must
1591 * invalidate the rkey of a new FastReg pool before first
1592 * using it. Thus, I am marking the FRD invalid here. */
1593 frd->frd_valid = false;
1595 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1596 fpo->fast_reg.fpo_pool_size++;
1603 ib_dereg_mr(frd->frd_mr);
1604 #ifndef HAVE_IB_MAP_MR_SG
1606 ib_free_fast_reg_page_list(frd->frd_frpl);
1608 LIBCFS_FREE(frd, sizeof(*frd));
1611 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1613 list_del(&frd->frd_list);
1614 #ifndef HAVE_IB_MAP_MR_SG
1615 ib_free_fast_reg_page_list(frd->frd_frpl);
1617 ib_dereg_mr(frd->frd_mr);
1618 LIBCFS_FREE(frd, sizeof(*frd));
1625 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1627 struct ib_device_attr *dev_attr;
1628 kib_dev_t *dev = fps->fps_net->ibn_dev;
1629 kib_fmr_pool_t *fpo;
1632 #ifndef HAVE_IB_DEVICE_ATTRS
1633 dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1638 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1644 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1646 #ifdef HAVE_IB_DEVICE_ATTRS
1647 dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1649 rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1651 CERROR("Query device failed for %s: %d\n",
1652 fpo->fpo_hdev->ibh_ibdev->name, rc);
1657 /* Check for FMR or FastReg support */
1658 fpo->fpo_is_fmr = 0;
1659 if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1660 fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1661 fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1662 fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1663 LCONSOLE_INFO("Using FMR for registration\n");
1664 fpo->fpo_is_fmr = 1;
1665 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1666 LCONSOLE_INFO("Using FastReg for registration\n");
1669 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1673 if (fpo->fpo_is_fmr)
1674 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1676 rc = kiblnd_alloc_freg_pool(fps, fpo);
1680 #ifndef HAVE_IB_DEVICE_ATTRS
1683 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1684 fpo->fpo_owner = fps;
1690 kiblnd_hdev_decref(fpo->fpo_hdev);
1691 LIBCFS_FREE(fpo, sizeof(*fpo));
1694 #ifndef HAVE_IB_DEVICE_ATTRS
1702 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1704 if (fps->fps_net == NULL) /* intialized? */
1707 spin_lock(&fps->fps_lock);
1709 while (!list_empty(&fps->fps_pool_list)) {
1710 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1711 kib_fmr_pool_t, fpo_list);
1712 fpo->fpo_failed = 1;
1713 list_del(&fpo->fpo_list);
1714 if (fpo->fpo_map_count == 0)
1715 list_add(&fpo->fpo_list, zombies);
1717 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1720 spin_unlock(&fps->fps_lock);
1724 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1726 if (fps->fps_net != NULL) { /* initialized? */
1727 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1728 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1733 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1735 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1737 kib_fmr_pool_t *fpo;
1740 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1745 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1746 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1747 fps->fps_cache = tunables->lnd_fmr_cache;
1749 spin_lock_init(&fps->fps_lock);
1750 INIT_LIST_HEAD(&fps->fps_pool_list);
1751 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1753 rc = kiblnd_create_fmr_pool(fps, &fpo);
1755 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1761 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1763 if (fpo->fpo_map_count != 0) /* still in use */
1765 if (fpo->fpo_failed)
1767 return cfs_time_aftereq(now, fpo->fpo_deadline);
1771 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1773 kib_hca_dev_t *hdev;
1774 __u64 *pages = tx->tx_pages;
1779 hdev = tx->tx_pool->tpo_hdev;
1781 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1782 for (size = 0; size < rd->rd_frags[i].rf_nob;
1783 size += hdev->ibh_page_size) {
1784 pages[npages++] = (rd->rd_frags[i].rf_addr &
1785 hdev->ibh_page_mask) + size;
1793 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1795 struct list_head zombies = LIST_HEAD_INIT(zombies);
1796 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1797 kib_fmr_poolset_t *fps;
1798 cfs_time_t now = cfs_time_current();
1799 kib_fmr_pool_t *tmp;
1805 fps = fpo->fpo_owner;
1806 if (fpo->fpo_is_fmr) {
1807 if (fmr->fmr_pfmr) {
1808 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1810 fmr->fmr_pfmr = NULL;
1814 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1818 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1821 frd->frd_valid = false;
1822 spin_lock(&fps->fps_lock);
1823 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1824 spin_unlock(&fps->fps_lock);
1825 fmr->fmr_frd = NULL;
1828 fmr->fmr_pool = NULL;
1830 spin_lock(&fps->fps_lock);
1831 fpo->fpo_map_count--; /* decref the pool */
1833 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1834 /* the first pool is persistent */
1835 if (fps->fps_pool_list.next == &fpo->fpo_list)
1838 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1839 list_move(&fpo->fpo_list, &zombies);
1843 spin_unlock(&fps->fps_lock);
1845 if (!list_empty(&zombies))
1846 kiblnd_destroy_fmr_pool_list(&zombies);
1850 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1851 __u32 nob, __u64 iov, kib_fmr_t *fmr, bool *is_fastreg)
1853 kib_fmr_pool_t *fpo;
1854 __u64 *pages = tx->tx_pages;
1856 bool is_rx = (rd != tx->tx_rd);
1857 bool tx_pages_mapped = 0;
1862 spin_lock(&fps->fps_lock);
1863 version = fps->fps_version;
1864 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1865 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1866 fpo->fpo_map_count++;
1868 if (fpo->fpo_is_fmr) {
1869 struct ib_pool_fmr *pfmr;
1872 spin_unlock(&fps->fps_lock);
1874 if (!tx_pages_mapped) {
1875 npages = kiblnd_map_tx_pages(tx, rd);
1876 tx_pages_mapped = 1;
1879 pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1880 pages, npages, iov);
1881 if (likely(!IS_ERR(pfmr))) {
1882 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1884 fmr->fmr_frd = NULL;
1885 fmr->fmr_pfmr = pfmr;
1886 fmr->fmr_pool = fpo;
1892 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1893 struct kib_fast_reg_descriptor *frd;
1894 #ifdef HAVE_IB_MAP_MR_SG
1895 struct ib_reg_wr *wr;
1898 struct ib_rdma_wr *wr;
1899 struct ib_fast_reg_page_list *frpl;
1903 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1904 struct kib_fast_reg_descriptor,
1906 list_del(&frd->frd_list);
1907 spin_unlock(&fps->fps_lock);
1909 #ifndef HAVE_IB_MAP_MR_SG
1910 frpl = frd->frd_frpl;
1914 if (!frd->frd_valid) {
1915 struct ib_rdma_wr *inv_wr;
1916 __u32 key = is_rx ? mr->rkey : mr->lkey;
1918 inv_wr = &frd->frd_inv_wr;
1919 memset(inv_wr, 0, sizeof(*inv_wr));
1921 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1922 inv_wr->wr.wr_id = IBLND_WID_MR;
1923 inv_wr->wr.ex.invalidate_rkey = key;
1926 key = ib_inc_rkey(key);
1927 ib_update_fast_reg_key(mr, key);
1930 #ifdef HAVE_IB_MAP_MR_SG
1931 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1932 n = ib_map_mr_sg(mr, tx->tx_frags,
1933 tx->tx_nfrags, NULL, PAGE_SIZE);
1935 n = ib_map_mr_sg(mr, tx->tx_frags,
1936 tx->tx_nfrags, PAGE_SIZE);
1938 if (unlikely(n != tx->tx_nfrags)) {
1939 CERROR("Failed to map mr %d/%d "
1940 "elements\n", n, tx->tx_nfrags);
1941 return n < 0 ? n : -EINVAL;
1944 wr = &frd->frd_fastreg_wr;
1945 memset(wr, 0, sizeof(*wr));
1947 wr->wr.opcode = IB_WR_REG_MR;
1948 wr->wr.wr_id = IBLND_WID_MR;
1950 wr->wr.send_flags = 0;
1952 wr->key = is_rx ? mr->rkey : mr->lkey;
1953 wr->access = (IB_ACCESS_LOCAL_WRITE |
1954 IB_ACCESS_REMOTE_WRITE);
1956 if (!tx_pages_mapped) {
1957 npages = kiblnd_map_tx_pages(tx, rd);
1958 tx_pages_mapped = 1;
1961 LASSERT(npages <= frpl->max_page_list_len);
1962 memcpy(frpl->page_list, pages,
1963 sizeof(*pages) * npages);
1965 /* Prepare FastReg WR */
1966 wr = &frd->frd_fastreg_wr;
1967 memset(wr, 0, sizeof(*wr));
1969 wr->wr.opcode = IB_WR_FAST_REG_MR;
1970 wr->wr.wr_id = IBLND_WID_MR;
1972 wr->wr.wr.fast_reg.iova_start = iov;
1973 wr->wr.wr.fast_reg.page_list = frpl;
1974 wr->wr.wr.fast_reg.page_list_len = npages;
1975 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1976 wr->wr.wr.fast_reg.length = nob;
1977 wr->wr.wr.fast_reg.rkey =
1978 is_rx ? mr->rkey : mr->lkey;
1979 wr->wr.wr.fast_reg.access_flags =
1980 (IB_ACCESS_LOCAL_WRITE |
1981 IB_ACCESS_REMOTE_WRITE);
1984 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1986 fmr->fmr_pfmr = NULL;
1987 fmr->fmr_pool = fpo;
1990 spin_unlock(&fps->fps_lock);
1994 spin_lock(&fps->fps_lock);
1995 fpo->fpo_map_count--;
1996 if (rc != -EAGAIN) {
1997 spin_unlock(&fps->fps_lock);
2001 /* EAGAIN and ... */
2002 if (version != fps->fps_version) {
2003 spin_unlock(&fps->fps_lock);
2008 if (fps->fps_increasing) {
2009 spin_unlock(&fps->fps_lock);
2010 CDEBUG(D_NET, "Another thread is allocating new "
2011 "FMR pool, waiting for her to complete\n");
2017 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
2018 /* someone failed recently */
2019 spin_unlock(&fps->fps_lock);
2023 fps->fps_increasing = 1;
2024 spin_unlock(&fps->fps_lock);
2026 CDEBUG(D_NET, "Allocate new FMR pool\n");
2027 rc = kiblnd_create_fmr_pool(fps, &fpo);
2028 spin_lock(&fps->fps_lock);
2029 fps->fps_increasing = 0;
2032 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2034 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2036 spin_unlock(&fps->fps_lock);
2042 kiblnd_fini_pool(kib_pool_t *pool)
2044 LASSERT(list_empty(&pool->po_free_list));
2045 LASSERT(pool->po_allocated == 0);
2047 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2051 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2053 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2055 memset(pool, 0, sizeof(kib_pool_t));
2056 INIT_LIST_HEAD(&pool->po_free_list);
2057 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2058 pool->po_owner = ps;
2059 pool->po_size = size;
2063 kiblnd_destroy_pool_list(struct list_head *head)
2067 while (!list_empty(head)) {
2068 pool = list_entry(head->next, kib_pool_t, po_list);
2069 list_del(&pool->po_list);
2071 LASSERT(pool->po_owner != NULL);
2072 pool->po_owner->ps_pool_destroy(pool);
2077 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2079 if (ps->ps_net == NULL) /* intialized? */
2082 spin_lock(&ps->ps_lock);
2083 while (!list_empty(&ps->ps_pool_list)) {
2084 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2085 kib_pool_t, po_list);
2087 list_del(&po->po_list);
2088 if (po->po_allocated == 0)
2089 list_add(&po->po_list, zombies);
2091 list_add(&po->po_list, &ps->ps_failed_pool_list);
2093 spin_unlock(&ps->ps_lock);
2097 kiblnd_fini_poolset(kib_poolset_t *ps)
2099 if (ps->ps_net != NULL) { /* initialized? */
2100 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2101 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2106 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2107 kib_net_t *net, char *name, int size,
2108 kib_ps_pool_create_t po_create,
2109 kib_ps_pool_destroy_t po_destroy,
2110 kib_ps_node_init_t nd_init,
2111 kib_ps_node_fini_t nd_fini)
2116 memset(ps, 0, sizeof(kib_poolset_t));
2120 ps->ps_pool_create = po_create;
2121 ps->ps_pool_destroy = po_destroy;
2122 ps->ps_node_init = nd_init;
2123 ps->ps_node_fini = nd_fini;
2124 ps->ps_pool_size = size;
2125 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2126 >= sizeof(ps->ps_name))
2128 spin_lock_init(&ps->ps_lock);
2129 INIT_LIST_HEAD(&ps->ps_pool_list);
2130 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2132 rc = ps->ps_pool_create(ps, size, &pool);
2134 list_add(&pool->po_list, &ps->ps_pool_list);
2136 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2142 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2144 if (pool->po_allocated != 0) /* still in use */
2146 if (pool->po_failed)
2148 return cfs_time_aftereq(now, pool->po_deadline);
2152 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2154 struct list_head zombies = LIST_HEAD_INIT(zombies);
2155 kib_poolset_t *ps = pool->po_owner;
2157 cfs_time_t now = cfs_time_current();
2159 spin_lock(&ps->ps_lock);
2161 if (ps->ps_node_fini != NULL)
2162 ps->ps_node_fini(pool, node);
2164 LASSERT(pool->po_allocated > 0);
2165 list_add(node, &pool->po_free_list);
2166 pool->po_allocated--;
2168 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2169 /* the first pool is persistent */
2170 if (ps->ps_pool_list.next == &pool->po_list)
2173 if (kiblnd_pool_is_idle(pool, now))
2174 list_move(&pool->po_list, &zombies);
2176 spin_unlock(&ps->ps_lock);
2178 if (!list_empty(&zombies))
2179 kiblnd_destroy_pool_list(&zombies);
2183 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2185 struct list_head *node;
2188 unsigned int interval = 1;
2189 cfs_time_t time_before;
2190 unsigned int trips = 0;
2193 spin_lock(&ps->ps_lock);
2194 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2195 if (list_empty(&pool->po_free_list))
2198 pool->po_allocated++;
2199 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2200 node = pool->po_free_list.next;
2203 if (ps->ps_node_init != NULL) {
2204 /* still hold the lock */
2205 ps->ps_node_init(pool, node);
2207 spin_unlock(&ps->ps_lock);
2211 /* no available tx pool and ... */
2212 if (ps->ps_increasing) {
2213 /* another thread is allocating a new pool */
2214 spin_unlock(&ps->ps_lock);
2216 CDEBUG(D_NET, "Another thread is allocating new "
2217 "%s pool, waiting %d HZs for her to complete."
2219 ps->ps_name, interval, trips);
2221 set_current_state(TASK_INTERRUPTIBLE);
2222 schedule_timeout(interval);
2223 if (interval < cfs_time_seconds(1))
2229 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2230 /* someone failed recently */
2231 spin_unlock(&ps->ps_lock);
2235 ps->ps_increasing = 1;
2236 spin_unlock(&ps->ps_lock);
2238 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2239 time_before = cfs_time_current();
2240 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2241 CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2242 cfs_time_current() - time_before);
2244 spin_lock(&ps->ps_lock);
2245 ps->ps_increasing = 0;
2247 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2249 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2250 CERROR("Can't allocate new %s pool because out of memory\n",
2253 spin_unlock(&ps->ps_lock);
2259 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2261 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2264 LASSERT (pool->po_allocated == 0);
2266 if (tpo->tpo_tx_pages != NULL) {
2267 kiblnd_unmap_tx_pool(tpo);
2268 kiblnd_free_pages(tpo->tpo_tx_pages);
2271 if (tpo->tpo_tx_descs == NULL)
2274 for (i = 0; i < pool->po_size; i++) {
2275 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2276 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2278 list_del(&tx->tx_list);
2279 if (tx->tx_pages != NULL)
2280 LIBCFS_FREE(tx->tx_pages,
2282 sizeof(*tx->tx_pages));
2283 if (tx->tx_frags != NULL)
2284 LIBCFS_FREE(tx->tx_frags,
2285 (1 + IBLND_MAX_RDMA_FRAGS) *
2286 sizeof(*tx->tx_frags));
2287 if (tx->tx_wrq != NULL)
2288 LIBCFS_FREE(tx->tx_wrq,
2289 (1 + IBLND_MAX_RDMA_FRAGS) *
2290 sizeof(*tx->tx_wrq));
2291 if (tx->tx_sge != NULL)
2292 LIBCFS_FREE(tx->tx_sge,
2293 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2294 sizeof(*tx->tx_sge));
2295 if (tx->tx_rd != NULL)
2296 LIBCFS_FREE(tx->tx_rd,
2297 offsetof(kib_rdma_desc_t,
2298 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2301 LIBCFS_FREE(tpo->tpo_tx_descs,
2302 pool->po_size * sizeof(kib_tx_t));
2304 kiblnd_fini_pool(pool);
2305 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2308 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2310 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2313 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2314 ntx = tunables->lnd_ntx / ncpts;
2316 return max(IBLND_TX_POOL, ntx);
2320 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2327 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2329 CERROR("Failed to allocate TX pool\n");
2333 pool = &tpo->tpo_pool;
2334 kiblnd_init_pool(ps, pool, size);
2335 tpo->tpo_tx_descs = NULL;
2336 tpo->tpo_tx_pages = NULL;
2338 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2339 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2340 CERROR("Can't allocate tx pages: %d\n", npg);
2341 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2345 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2346 size * sizeof(kib_tx_t));
2347 if (tpo->tpo_tx_descs == NULL) {
2348 CERROR("Can't allocate %d tx descriptors\n", size);
2349 ps->ps_pool_destroy(pool);
2353 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2355 for (i = 0; i < size; i++) {
2356 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2357 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2360 if (ps->ps_net->ibn_fmr_ps != NULL) {
2361 LIBCFS_CPT_ALLOC(tx->tx_pages,
2362 lnet_cpt_table(), ps->ps_cpt,
2363 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2364 if (tx->tx_pages == NULL)
2368 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2369 (1 + IBLND_MAX_RDMA_FRAGS) *
2370 sizeof(*tx->tx_frags));
2371 if (tx->tx_frags == NULL)
2374 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2376 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2377 (1 + IBLND_MAX_RDMA_FRAGS) *
2378 sizeof(*tx->tx_wrq));
2379 if (tx->tx_wrq == NULL)
2382 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2383 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2384 sizeof(*tx->tx_sge));
2385 if (tx->tx_sge == NULL)
2388 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2389 offsetof(kib_rdma_desc_t,
2390 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2391 if (tx->tx_rd == NULL)
2396 kiblnd_map_tx_pool(tpo);
2401 ps->ps_pool_destroy(pool);
2406 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2408 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2410 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
2412 tx->tx_cookie = tps->tps_next_tx_cookie++;
2416 kiblnd_net_fini_pools(kib_net_t *net)
2420 cfs_cpt_for_each(i, lnet_cpt_table()) {
2421 kib_tx_poolset_t *tps;
2422 kib_fmr_poolset_t *fps;
2424 if (net->ibn_tx_ps != NULL) {
2425 tps = net->ibn_tx_ps[i];
2426 kiblnd_fini_poolset(&tps->tps_poolset);
2429 if (net->ibn_fmr_ps != NULL) {
2430 fps = net->ibn_fmr_ps[i];
2431 kiblnd_fini_fmr_poolset(fps);
2435 if (net->ibn_tx_ps != NULL) {
2436 cfs_percpt_free(net->ibn_tx_ps);
2437 net->ibn_tx_ps = NULL;
2440 if (net->ibn_fmr_ps != NULL) {
2441 cfs_percpt_free(net->ibn_fmr_ps);
2442 net->ibn_fmr_ps = NULL;
2447 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2450 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2451 #ifdef HAVE_IB_GET_DMA_MR
2452 unsigned long flags;
2458 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2460 #ifdef HAVE_IB_GET_DMA_MR
2461 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2462 if (tunables->lnd_map_on_demand == 0) {
2463 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2465 goto create_tx_pool;
2468 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2471 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2472 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2473 tunables->lnd_fmr_pool_size,
2474 tunables->lnd_ntx / 4);
2479 /* TX pool must be created later than FMR, see LU-2268
2481 LASSERT(net->ibn_tx_ps == NULL);
2483 /* premapping can fail if ibd_nmr > 1, so we always create
2484 * FMR pool and map-on-demand if premapping failed */
2486 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2487 sizeof(kib_fmr_poolset_t));
2488 if (net->ibn_fmr_ps == NULL) {
2489 CERROR("Failed to allocate FMR pool array\n");
2494 for (i = 0; i < ncpts; i++) {
2495 cpt = (cpts == NULL) ? i : cpts[i];
2496 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2499 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2506 LASSERT(i == ncpts);
2508 #ifdef HAVE_IB_GET_DMA_MR
2511 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2512 sizeof(kib_tx_poolset_t));
2513 if (net->ibn_tx_ps == NULL) {
2514 CERROR("Failed to allocate tx pool array\n");
2519 for (i = 0; i < ncpts; i++) {
2520 cpt = (cpts == NULL) ? i : cpts[i];
2521 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2523 kiblnd_tx_pool_size(ni, ncpts),
2524 kiblnd_create_tx_pool,
2525 kiblnd_destroy_tx_pool,
2526 kiblnd_tx_init, NULL);
2528 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2536 kiblnd_net_fini_pools(net);
2542 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2544 #ifndef HAVE_IB_DEVICE_ATTRS
2545 struct ib_device_attr *attr;
2549 /* It's safe to assume a HCA can handle a page size
2550 * matching that of the native system */
2551 hdev->ibh_page_shift = PAGE_SHIFT;
2552 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2553 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2555 #ifdef HAVE_IB_DEVICE_ATTRS
2556 hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2558 LIBCFS_ALLOC(attr, sizeof(*attr));
2560 CERROR("Out of memory\n");
2564 rc = ib_query_device(hdev->ibh_ibdev, attr);
2566 hdev->ibh_mr_size = attr->max_mr_size;
2568 LIBCFS_FREE(attr, sizeof(*attr));
2571 CERROR("Failed to query IB device: %d\n", rc);
2576 if (hdev->ibh_mr_size == ~0ULL) {
2577 hdev->ibh_mr_shift = 64;
2581 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2585 #ifdef HAVE_IB_GET_DMA_MR
2587 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2589 if (hdev->ibh_mrs == NULL)
2592 ib_dereg_mr(hdev->ibh_mrs);
2594 hdev->ibh_mrs = NULL;
2599 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2601 #ifdef HAVE_IB_GET_DMA_MR
2602 kiblnd_hdev_cleanup_mrs(hdev);
2605 if (hdev->ibh_pd != NULL)
2606 ib_dealloc_pd(hdev->ibh_pd);
2608 if (hdev->ibh_cmid != NULL)
2609 rdma_destroy_id(hdev->ibh_cmid);
2611 LIBCFS_FREE(hdev, sizeof(*hdev));
2614 #ifdef HAVE_IB_GET_DMA_MR
2616 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2620 int acflags = IB_ACCESS_LOCAL_WRITE |
2621 IB_ACCESS_REMOTE_WRITE;
2623 rc = kiblnd_hdev_get_attr(hdev);
2627 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2629 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2630 kiblnd_hdev_cleanup_mrs(hdev);
2641 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2647 kiblnd_dev_need_failover(kib_dev_t *dev)
2649 struct rdma_cm_id *cmid;
2650 struct sockaddr_in srcaddr;
2651 struct sockaddr_in dstaddr;
2654 if (dev->ibd_hdev == NULL || /* initializing */
2655 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2656 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2659 /* XXX: it's UGLY, but I don't have better way to find
2660 * ib-bonding HCA failover because:
2662 * a. no reliable CM event for HCA failover...
2663 * b. no OFED API to get ib_device for current net_device...
2665 * We have only two choices at this point:
2667 * a. rdma_bind_addr(), it will conflict with listener cmid
2668 * b. rdma_resolve_addr() to zero addr */
2669 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2673 CERROR("Failed to create cmid for failover: %d\n", rc);
2677 memset(&srcaddr, 0, sizeof(srcaddr));
2678 srcaddr.sin_family = AF_INET;
2679 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2681 memset(&dstaddr, 0, sizeof(dstaddr));
2682 dstaddr.sin_family = AF_INET;
2683 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2684 (struct sockaddr *)&dstaddr, 1);
2685 if (rc != 0 || cmid->device == NULL) {
2686 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2687 dev->ibd_ifname, &dev->ibd_ifip,
2689 rdma_destroy_id(cmid);
2693 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2694 rdma_destroy_id(cmid);
2699 kiblnd_dev_failover(kib_dev_t *dev)
2701 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2702 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2703 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2704 struct rdma_cm_id *cmid = NULL;
2705 kib_hca_dev_t *hdev = NULL;
2709 struct sockaddr_in addr;
2710 unsigned long flags;
2714 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2715 dev->ibd_can_failover ||
2716 dev->ibd_hdev == NULL);
2718 rc = kiblnd_dev_need_failover(dev);
2722 if (dev->ibd_hdev != NULL &&
2723 dev->ibd_hdev->ibh_cmid != NULL) {
2724 /* XXX it's not good to close old listener at here,
2725 * because we can fail to create new listener.
2726 * But we have to close it now, otherwise rdma_bind_addr
2727 * will return EADDRINUSE... How crap! */
2728 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2730 cmid = dev->ibd_hdev->ibh_cmid;
2731 /* make next schedule of kiblnd_dev_need_failover()
2732 * return 1 for me */
2733 dev->ibd_hdev->ibh_cmid = NULL;
2734 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2736 rdma_destroy_id(cmid);
2739 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2743 CERROR("Failed to create cmid for failover: %d\n", rc);
2747 memset(&addr, 0, sizeof(addr));
2748 addr.sin_family = AF_INET;
2749 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2750 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2752 /* Bind to failover device or port */
2753 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2754 if (rc != 0 || cmid->device == NULL) {
2755 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2756 dev->ibd_ifname, &dev->ibd_ifip,
2758 rdma_destroy_id(cmid);
2762 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2764 CERROR("Failed to allocate kib_hca_dev\n");
2765 rdma_destroy_id(cmid);
2770 atomic_set(&hdev->ibh_ref, 1);
2771 hdev->ibh_dev = dev;
2772 hdev->ibh_cmid = cmid;
2773 hdev->ibh_ibdev = cmid->device;
2775 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2776 pd = ib_alloc_pd(cmid->device, 0);
2778 pd = ib_alloc_pd(cmid->device);
2782 CERROR("Can't allocate PD: %d\n", rc);
2788 rc = rdma_listen(cmid, 0);
2790 CERROR("Can't start new listener: %d\n", rc);
2794 #ifdef HAVE_IB_GET_DMA_MR
2795 rc = kiblnd_hdev_setup_mrs(hdev);
2797 CERROR("Can't setup device: %d\n", rc);
2801 rc = kiblnd_hdev_get_attr(hdev);
2803 CERROR("Can't get device attributes: %d\n", rc);
2808 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2810 old = dev->ibd_hdev;
2811 dev->ibd_hdev = hdev; /* take over the refcount */
2814 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2815 cfs_cpt_for_each(i, lnet_cpt_table()) {
2816 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2819 if (net->ibn_fmr_ps != NULL)
2820 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2825 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2827 if (!list_empty(&zombie_tpo))
2828 kiblnd_destroy_pool_list(&zombie_tpo);
2829 if (!list_empty(&zombie_ppo))
2830 kiblnd_destroy_pool_list(&zombie_ppo);
2831 if (!list_empty(&zombie_fpo))
2832 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2834 kiblnd_hdev_decref(hdev);
2837 dev->ibd_failed_failover++;
2839 dev->ibd_failed_failover = 0;
2845 kiblnd_destroy_dev (kib_dev_t *dev)
2847 LASSERT (dev->ibd_nnets == 0);
2848 LASSERT(list_empty(&dev->ibd_nets));
2850 list_del(&dev->ibd_fail_list);
2851 list_del(&dev->ibd_list);
2853 if (dev->ibd_hdev != NULL)
2854 kiblnd_hdev_decref(dev->ibd_hdev);
2856 LIBCFS_FREE(dev, sizeof(*dev));
2860 kiblnd_create_dev(char *ifname)
2862 struct net_device *netdev;
2869 rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2871 CERROR("Can't query IPoIB interface %s: %d\n",
2877 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2881 LIBCFS_ALLOC(dev, sizeof(*dev));
2885 netdev = dev_get_by_name(&init_net, ifname);
2886 if (netdev == NULL) {
2887 dev->ibd_can_failover = 0;
2889 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2893 INIT_LIST_HEAD(&dev->ibd_nets);
2894 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2895 INIT_LIST_HEAD(&dev->ibd_fail_list);
2897 strcpy(&dev->ibd_ifname[0], ifname);
2899 /* initialize the device */
2900 rc = kiblnd_dev_failover(dev);
2902 CERROR("Can't initialize device: %d\n", rc);
2903 LIBCFS_FREE(dev, sizeof(*dev));
2907 list_add_tail(&dev->ibd_list,
2908 &kiblnd_data.kib_devs);
2913 kiblnd_base_shutdown(void)
2915 struct kib_sched_info *sched;
2918 LASSERT(list_empty(&kiblnd_data.kib_devs));
2920 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2921 atomic_read(&libcfs_kmemory));
2923 switch (kiblnd_data.kib_init) {
2927 case IBLND_INIT_ALL:
2928 case IBLND_INIT_DATA:
2929 LASSERT (kiblnd_data.kib_peers != NULL);
2930 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2931 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2933 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2934 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2935 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2936 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2938 /* flag threads to terminate; wake and wait for them to die */
2939 kiblnd_data.kib_shutdown = 1;
2941 /* NB: we really want to stop scheduler threads net by net
2942 * instead of the whole module, this should be improved
2943 * with dynamic configuration LNet */
2944 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2945 wake_up_all(&sched->ibs_waitq);
2947 wake_up_all(&kiblnd_data.kib_connd_waitq);
2948 wake_up_all(&kiblnd_data.kib_failover_waitq);
2951 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2954 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2955 "Waiting for %d threads to terminate\n",
2956 atomic_read(&kiblnd_data.kib_nthreads));
2957 set_current_state(TASK_UNINTERRUPTIBLE);
2958 schedule_timeout(cfs_time_seconds(1));
2963 case IBLND_INIT_NOTHING:
2967 if (kiblnd_data.kib_peers != NULL) {
2968 LIBCFS_FREE(kiblnd_data.kib_peers,
2969 sizeof(struct list_head) *
2970 kiblnd_data.kib_peer_hash_size);
2973 if (kiblnd_data.kib_scheds != NULL)
2974 cfs_percpt_free(kiblnd_data.kib_scheds);
2976 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2977 atomic_read(&libcfs_kmemory));
2979 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2980 module_put(THIS_MODULE);
2984 kiblnd_shutdown(struct lnet_ni *ni)
2986 kib_net_t *net = ni->ni_data;
2987 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2989 unsigned long flags;
2991 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2996 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2997 atomic_read(&libcfs_kmemory));
2999 write_lock_irqsave(g_lock, flags);
3000 net->ibn_shutdown = 1;
3001 write_unlock_irqrestore(g_lock, flags);
3003 switch (net->ibn_init) {
3007 case IBLND_INIT_ALL:
3008 /* nuke all existing peers within this net */
3009 kiblnd_del_peer(ni, LNET_NID_ANY);
3011 /* Wait for all peer_ni state to clean up */
3013 while (atomic_read(&net->ibn_npeers) != 0) {
3016 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3017 "%s: waiting for %d peers to disconnect\n",
3018 libcfs_nid2str(ni->ni_nid),
3019 atomic_read(&net->ibn_npeers));
3020 set_current_state(TASK_UNINTERRUPTIBLE);
3021 schedule_timeout(cfs_time_seconds(1));
3024 kiblnd_net_fini_pools(net);
3026 write_lock_irqsave(g_lock, flags);
3027 LASSERT(net->ibn_dev->ibd_nnets > 0);
3028 net->ibn_dev->ibd_nnets--;
3029 list_del(&net->ibn_list);
3030 write_unlock_irqrestore(g_lock, flags);
3034 case IBLND_INIT_NOTHING:
3035 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3037 if (net->ibn_dev != NULL &&
3038 net->ibn_dev->ibd_nnets == 0)
3039 kiblnd_destroy_dev(net->ibn_dev);
3044 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3045 atomic_read(&libcfs_kmemory));
3047 net->ibn_init = IBLND_INIT_NOTHING;
3050 LIBCFS_FREE(net, sizeof(*net));
3053 if (list_empty(&kiblnd_data.kib_devs))
3054 kiblnd_base_shutdown();
3059 kiblnd_base_startup(void)
3061 struct kib_sched_info *sched;
3065 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3067 try_module_get(THIS_MODULE);
3068 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3070 rwlock_init(&kiblnd_data.kib_global_lock);
3072 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3073 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3075 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3076 LIBCFS_ALLOC(kiblnd_data.kib_peers,
3077 sizeof(struct list_head) *
3078 kiblnd_data.kib_peer_hash_size);
3079 if (kiblnd_data.kib_peers == NULL)
3082 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3083 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3085 spin_lock_init(&kiblnd_data.kib_connd_lock);
3086 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3087 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3088 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3089 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3091 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3092 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3094 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3096 if (kiblnd_data.kib_scheds == NULL)
3099 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3102 spin_lock_init(&sched->ibs_lock);
3103 INIT_LIST_HEAD(&sched->ibs_conns);
3104 init_waitqueue_head(&sched->ibs_waitq);
3106 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3107 if (*kiblnd_tunables.kib_nscheds > 0) {
3108 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3110 /* max to half of CPUs, another half is reserved for
3111 * upper layer modules */
3112 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3115 sched->ibs_nthreads_max = nthrs;
3119 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3121 /* lists/ptrs/locks initialised */
3122 kiblnd_data.kib_init = IBLND_INIT_DATA;
3123 /*****************************************************/
3125 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3127 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3131 if (*kiblnd_tunables.kib_dev_failover != 0)
3132 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3136 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3140 /* flag everything initialised */
3141 kiblnd_data.kib_init = IBLND_INIT_ALL;
3142 /*****************************************************/
3147 kiblnd_base_shutdown();
3152 kiblnd_start_schedulers(struct kib_sched_info *sched)
3158 if (sched->ibs_nthreads == 0) {
3159 if (*kiblnd_tunables.kib_nscheds > 0) {
3160 nthrs = sched->ibs_nthreads_max;
3162 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3164 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3165 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3168 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3169 /* increase one thread if there is new interface */
3170 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3173 for (i = 0; i < nthrs; i++) {
3176 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3177 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3178 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3179 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3183 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3184 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3188 sched->ibs_nthreads += i;
3193 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3199 for (i = 0; i < ncpts; i++) {
3200 struct kib_sched_info *sched;
3202 cpt = (cpts == NULL) ? i : cpts[i];
3203 sched = kiblnd_data.kib_scheds[cpt];
3205 if (!newdev && sched->ibs_nthreads > 0)
3208 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3210 CERROR("Failed to start scheduler threads for %s\n",
3219 kiblnd_dev_search(char *ifname)
3221 kib_dev_t *alias = NULL;
3226 colon = strchr(ifname, ':');
3227 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3228 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3234 colon2 = strchr(dev->ibd_ifname, ':');
3240 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3252 kiblnd_startup(struct lnet_ni *ni)
3255 kib_dev_t *ibdev = NULL;
3257 unsigned long flags;
3262 LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3264 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3265 rc = kiblnd_base_startup();
3270 LIBCFS_ALLOC(net, sizeof(*net));
3275 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3277 kiblnd_tunables_setup(ni);
3279 if (ni->ni_interfaces[0] != NULL) {
3280 /* Use the IPoIB interface specified in 'networks=' */
3282 CLASSERT(LNET_INTERFACES_NUM > 1);
3283 if (ni->ni_interfaces[1] != NULL) {
3284 CERROR("Multiple interfaces not supported\n");
3288 ifname = ni->ni_interfaces[0];
3290 ifname = *kiblnd_tunables.kib_default_ipif;
3293 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3294 CERROR("IPoIB interface name too long: %s\n", ifname);
3298 ibdev = kiblnd_dev_search(ifname);
3300 newdev = ibdev == NULL;
3301 /* hmm...create kib_dev even for alias */
3302 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3303 ibdev = kiblnd_create_dev(ifname);
3308 node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3309 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3311 net->ibn_dev = ibdev;
3312 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3314 rc = kiblnd_dev_start_threads(ibdev, newdev,
3315 ni->ni_cpts, ni->ni_ncpts);
3319 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3321 CERROR("Failed to initialize NI pools: %d\n", rc);
3325 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3327 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3328 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3330 net->ibn_init = IBLND_INIT_ALL;
3335 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3336 kiblnd_destroy_dev(ibdev);
3338 kiblnd_shutdown(ni);
3340 CDEBUG(D_NET, "kiblnd_startup failed\n");
3344 static struct lnet_lnd the_o2iblnd = {
3345 .lnd_type = O2IBLND,
3346 .lnd_startup = kiblnd_startup,
3347 .lnd_shutdown = kiblnd_shutdown,
3348 .lnd_ctl = kiblnd_ctl,
3349 .lnd_query = kiblnd_query,
3350 .lnd_send = kiblnd_send,
3351 .lnd_recv = kiblnd_recv,
3354 static void __exit ko2iblnd_exit(void)
3356 lnet_unregister_lnd(&the_o2iblnd);
3359 static int __init ko2iblnd_init(void)
3363 CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3364 CLASSERT(offsetof(kib_msg_t,
3365 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3367 CLASSERT(offsetof(kib_msg_t,
3368 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3371 rc = kiblnd_tunables_init();
3375 lnet_register_lnd(&the_o2iblnd);
3380 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3381 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3382 MODULE_VERSION("2.8.0");
3383 MODULE_LICENSE("GPL");
3385 module_init(ko2iblnd_init);
3386 module_exit(ko2iblnd_exit);