4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
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 = IBLND_MAX_RDMA_FRAGS;
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 %lld "
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, ktime_to_ns(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);
731 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
734 * One WR for the LNet message
735 * And ibc_max_frags for the transfer WRs
737 unsigned int ret = 1 + conn->ibc_max_frags;
738 __u32 dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
740 /* FastReg needs two extra WRs for map and invalidate */
741 if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
744 /* account for a maximum of ibc_queue_depth in-flight transfers */
745 ret *= conn->ibc_queue_depth;
750 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
751 int state, int version)
754 * If the new conn is created successfully it takes over the caller's
755 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
756 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
757 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
758 * to destroy 'cmid' here since I'm called from the CM which still has
759 * its ref on 'cmid'). */
760 rwlock_t *glock = &kiblnd_data.kib_global_lock;
761 kib_net_t *net = peer_ni->ibp_ni->ni_data;
763 struct ib_qp_init_attr *init_qp_attr;
764 struct kib_sched_info *sched;
765 #ifdef HAVE_IB_CQ_INIT_ATTR
766 struct ib_cq_init_attr cq_attr = {};
775 LASSERT(net != NULL);
776 LASSERT(!in_interrupt());
780 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
781 sched = kiblnd_get_scheduler(cpt);
784 CERROR("no schedulers available. node is unhealthy\n");
789 * The cpt might have changed if we ended up selecting a non cpt
790 * native scheduler. So use the scheduler's cpt instead.
792 cpt = sched->ibs_cpt;
794 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
795 sizeof(*init_qp_attr));
796 if (init_qp_attr == NULL) {
797 CERROR("Can't allocate qp_attr for %s\n",
798 libcfs_nid2str(peer_ni->ibp_nid));
802 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
804 CERROR("Can't allocate connection for %s\n",
805 libcfs_nid2str(peer_ni->ibp_nid));
809 conn->ibc_state = IBLND_CONN_INIT;
810 conn->ibc_version = version;
811 conn->ibc_peer = peer_ni; /* I take the caller's ref */
812 cmid->context = conn; /* for future CM callbacks */
813 conn->ibc_cmid = cmid;
814 conn->ibc_max_frags = peer_ni->ibp_max_frags;
815 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
817 INIT_LIST_HEAD(&conn->ibc_early_rxs);
818 INIT_LIST_HEAD(&conn->ibc_tx_noops);
819 INIT_LIST_HEAD(&conn->ibc_tx_queue);
820 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
821 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
822 INIT_LIST_HEAD(&conn->ibc_active_txs);
823 spin_lock_init(&conn->ibc_lock);
825 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
826 sizeof(*conn->ibc_connvars));
827 if (conn->ibc_connvars == NULL) {
828 CERROR("Can't allocate in-progress connection state\n");
832 write_lock_irqsave(glock, flags);
833 if (dev->ibd_failover) {
834 write_unlock_irqrestore(glock, flags);
835 CERROR("%s: failover in progress\n", dev->ibd_ifname);
839 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
840 /* wakeup failover thread and teardown connection */
841 if (kiblnd_dev_can_failover(dev)) {
842 list_add_tail(&dev->ibd_fail_list,
843 &kiblnd_data.kib_failed_devs);
844 wake_up(&kiblnd_data.kib_failover_waitq);
847 write_unlock_irqrestore(glock, flags);
848 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
849 cmid->device->name, dev->ibd_ifname);
853 kiblnd_hdev_addref_locked(dev->ibd_hdev);
854 conn->ibc_hdev = dev->ibd_hdev;
856 kiblnd_setup_mtu_locked(cmid);
858 write_unlock_irqrestore(glock, flags);
860 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
861 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
862 if (conn->ibc_rxs == NULL) {
863 CERROR("Cannot allocate RX buffers\n");
867 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
868 IBLND_RX_MSG_PAGES(conn));
872 kiblnd_map_rx_descs(conn);
874 #ifdef HAVE_IB_CQ_INIT_ATTR
875 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
876 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
877 cq = ib_create_cq(cmid->device,
878 kiblnd_cq_completion, kiblnd_cq_event, conn,
881 cq = ib_create_cq(cmid->device,
882 kiblnd_cq_completion, kiblnd_cq_event, conn,
883 IBLND_CQ_ENTRIES(conn),
884 kiblnd_get_completion_vector(conn, cpt));
888 * on MLX-5 (possibly MLX-4 as well) this error could be
889 * hit if the concurrent_sends and/or peer_tx_credits is set
890 * too high. Or due to an MLX-5 bug which tries to
891 * allocate 256kb via kmalloc for WR cookie array
893 CERROR("Failed to create CQ with %d CQEs: %ld\n",
894 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
900 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
902 CERROR("Can't request completion notification: %d\n", rc);
906 init_qp_attr->event_handler = kiblnd_qp_event;
907 init_qp_attr->qp_context = conn;
908 init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
909 init_qp_attr->cap.max_recv_sge = 1;
910 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
911 init_qp_attr->qp_type = IB_QPT_RC;
912 init_qp_attr->send_cq = cq;
913 init_qp_attr->recv_cq = cq;
915 conn->ibc_sched = sched;
918 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
919 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
921 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
922 if (!rc || conn->ibc_queue_depth < 2)
925 conn->ibc_queue_depth--;
929 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
930 "send_sge: %d, recv_sge: %d\n",
931 rc, init_qp_attr->cap.max_send_wr,
932 init_qp_attr->cap.max_recv_wr,
933 init_qp_attr->cap.max_send_sge,
934 init_qp_attr->cap.max_recv_sge);
938 if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
939 CWARN("peer %s - queue depth reduced from %u to %u"
940 " to allow for qp creation\n",
941 libcfs_nid2str(peer_ni->ibp_nid),
942 peer_ni->ibp_queue_depth,
943 conn->ibc_queue_depth);
945 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
947 /* 1 ref for caller and each rxmsg */
948 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
949 conn->ibc_nrx = IBLND_RX_MSGS(conn);
952 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
953 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
955 CERROR("Can't post rxmsg: %d\n", rc);
957 /* Make posted receives complete */
958 kiblnd_abort_receives(conn);
960 /* correct # of posted buffers
961 * NB locking needed now I'm racing with completion */
962 spin_lock_irqsave(&sched->ibs_lock, flags);
963 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
964 spin_unlock_irqrestore(&sched->ibs_lock, flags);
966 /* cmid will be destroyed by CM(ofed) after cm_callback
967 * returned, so we can't refer it anymore
968 * (by kiblnd_connd()->kiblnd_destroy_conn) */
969 rdma_destroy_qp(conn->ibc_cmid);
970 conn->ibc_cmid = NULL;
972 /* Drop my own and unused rxbuffer refcounts */
973 while (i++ <= IBLND_RX_MSGS(conn))
974 kiblnd_conn_decref(conn);
980 /* Init successful! */
981 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
982 state == IBLND_CONN_PASSIVE_WAIT);
983 conn->ibc_state = state;
986 atomic_inc(&net->ibn_nconns);
990 kiblnd_destroy_conn(conn, true);
992 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
998 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
1000 struct rdma_cm_id *cmid = conn->ibc_cmid;
1001 kib_peer_ni_t *peer_ni = conn->ibc_peer;
1004 LASSERT (!in_interrupt());
1005 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1006 LASSERT(list_empty(&conn->ibc_early_rxs));
1007 LASSERT(list_empty(&conn->ibc_tx_noops));
1008 LASSERT(list_empty(&conn->ibc_tx_queue));
1009 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1010 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1011 LASSERT(list_empty(&conn->ibc_active_txs));
1012 LASSERT (conn->ibc_noops_posted == 0);
1013 LASSERT (conn->ibc_nsends_posted == 0);
1015 switch (conn->ibc_state) {
1017 /* conn must be completely disengaged from the network */
1020 case IBLND_CONN_DISCONNECTED:
1021 /* connvars should have been freed already */
1022 LASSERT (conn->ibc_connvars == NULL);
1025 case IBLND_CONN_INIT:
1029 /* conn->ibc_cmid might be destroyed by CM already */
1030 if (cmid != NULL && cmid->qp != NULL)
1031 rdma_destroy_qp(cmid);
1033 if (conn->ibc_cq != NULL) {
1034 rc = ib_destroy_cq(conn->ibc_cq);
1036 CWARN("Error destroying CQ: %d\n", rc);
1039 if (conn->ibc_rx_pages != NULL)
1040 kiblnd_unmap_rx_descs(conn);
1042 if (conn->ibc_rxs != NULL) {
1043 LIBCFS_FREE(conn->ibc_rxs,
1044 IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
1047 if (conn->ibc_connvars != NULL)
1048 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1050 if (conn->ibc_hdev != NULL)
1051 kiblnd_hdev_decref(conn->ibc_hdev);
1053 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1054 if (conn->ibc_state != IBLND_CONN_INIT) {
1055 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1057 kiblnd_peer_decref(peer_ni);
1058 rdma_destroy_id(cmid);
1059 atomic_dec(&net->ibn_nconns);
1063 LIBCFS_FREE(conn, sizeof(*conn));
1067 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1070 struct list_head *ctmp;
1071 struct list_head *cnxt;
1074 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1075 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1077 CDEBUG(D_NET, "Closing conn -> %s, "
1078 "version: %x, reason: %d\n",
1079 libcfs_nid2str(peer_ni->ibp_nid),
1080 conn->ibc_version, why);
1082 kiblnd_close_conn_locked(conn, why);
1090 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1091 int version, __u64 incarnation)
1094 struct list_head *ctmp;
1095 struct list_head *cnxt;
1098 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1099 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1101 if (conn->ibc_version == version &&
1102 conn->ibc_incarnation == incarnation)
1105 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1106 "incarnation:%#llx(%x, %#llx)\n",
1107 libcfs_nid2str(peer_ni->ibp_nid),
1108 conn->ibc_version, conn->ibc_incarnation,
1109 version, incarnation);
1111 kiblnd_close_conn_locked(conn, -ESTALE);
1119 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1121 kib_peer_ni_t *peer_ni;
1122 struct list_head *ptmp;
1123 struct list_head *pnxt;
1127 unsigned long flags;
1130 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1132 if (nid != LNET_NID_ANY)
1133 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1136 hi = kiblnd_data.kib_peer_hash_size - 1;
1139 for (i = lo; i <= hi; i++) {
1140 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1142 peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1143 LASSERT(!kiblnd_peer_idle(peer_ni));
1145 if (peer_ni->ibp_ni != ni)
1148 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1151 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1155 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1157 /* wildcards always succeed */
1158 if (nid == LNET_NID_ANY)
1161 return (count == 0) ? -ENOENT : 0;
1165 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1167 struct libcfs_ioctl_data *data = arg;
1171 case IOC_LIBCFS_GET_PEER: {
1175 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1177 data->ioc_nid = nid;
1178 data->ioc_count = count;
1182 case IOC_LIBCFS_DEL_PEER: {
1183 rc = kiblnd_del_peer(ni, data->ioc_nid);
1186 case IOC_LIBCFS_GET_CONN: {
1190 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1196 LASSERT(conn->ibc_cmid != NULL);
1197 data->ioc_nid = conn->ibc_peer->ibp_nid;
1198 if (conn->ibc_cmid->route.path_rec == NULL)
1199 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1202 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1203 kiblnd_conn_decref(conn);
1206 case IOC_LIBCFS_CLOSE_CONNECTION: {
1207 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1219 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1221 time64_t last_alive = 0;
1222 time64_t now = ktime_get_seconds();
1223 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1224 kib_peer_ni_t *peer_ni;
1225 unsigned long flags;
1227 read_lock_irqsave(glock, flags);
1229 peer_ni = kiblnd_find_peer_locked(ni, nid);
1230 if (peer_ni != NULL)
1231 last_alive = peer_ni->ibp_last_alive;
1233 read_unlock_irqrestore(glock, flags);
1235 if (last_alive != 0)
1238 /* peer_ni is not persistent in hash, trigger peer_ni creation
1239 * and connection establishment with a NULL tx */
1240 if (peer_ni == NULL)
1241 kiblnd_launch_tx(ni, NULL, nid);
1243 CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n",
1244 libcfs_nid2str(nid), peer_ni,
1245 last_alive ? now - last_alive : -1);
1250 kiblnd_free_pages(kib_pages_t *p)
1252 int npages = p->ibp_npages;
1255 for (i = 0; i < npages; i++) {
1256 if (p->ibp_pages[i] != NULL)
1257 __free_page(p->ibp_pages[i]);
1260 LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1264 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1269 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1270 offsetof(kib_pages_t, ibp_pages[npages]));
1272 CERROR("Can't allocate descriptor for %d pages\n", npages);
1276 memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1277 p->ibp_npages = npages;
1279 for (i = 0; i < npages; i++) {
1280 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1282 if (p->ibp_pages[i] == NULL) {
1283 CERROR("Can't allocate page %d of %d\n", i, npages);
1284 kiblnd_free_pages(p);
1294 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1299 LASSERT (conn->ibc_rxs != NULL);
1300 LASSERT (conn->ibc_hdev != NULL);
1302 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1303 rx = &conn->ibc_rxs[i];
1305 LASSERT(rx->rx_nob >= 0); /* not posted */
1307 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1308 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1310 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1313 kiblnd_free_pages(conn->ibc_rx_pages);
1315 conn->ibc_rx_pages = NULL;
1319 kiblnd_map_rx_descs(kib_conn_t *conn)
1327 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1328 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1329 rx = &conn->ibc_rxs[i];
1332 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1335 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1336 rx->rx_msg, IBLND_MSG_SIZE,
1338 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1340 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1342 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1343 i, rx->rx_msg, rx->rx_msgaddr,
1344 (__u64)(page_to_phys(pg) + pg_off));
1346 pg_off += IBLND_MSG_SIZE;
1347 LASSERT(pg_off <= PAGE_SIZE);
1349 if (pg_off == PAGE_SIZE) {
1352 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1358 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1360 kib_hca_dev_t *hdev = tpo->tpo_hdev;
1364 LASSERT (tpo->tpo_pool.po_allocated == 0);
1369 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1370 tx = &tpo->tpo_tx_descs[i];
1371 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1372 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1374 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1377 kiblnd_hdev_decref(hdev);
1378 tpo->tpo_hdev = NULL;
1381 static kib_hca_dev_t *
1382 kiblnd_current_hdev(kib_dev_t *dev)
1384 kib_hca_dev_t *hdev;
1385 unsigned long flags;
1388 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1389 while (dev->ibd_failover) {
1390 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1392 CDEBUG(D_NET, "%s: Wait for failover\n",
1394 set_current_state(TASK_INTERRUPTIBLE);
1395 schedule_timeout(cfs_time_seconds(1) / 100);
1397 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1400 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1401 hdev = dev->ibd_hdev;
1403 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1409 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1411 kib_pages_t *txpgs = tpo->tpo_tx_pages;
1412 kib_pool_t *pool = &tpo->tpo_pool;
1413 kib_net_t *net = pool->po_owner->ps_net;
1421 LASSERT (net != NULL);
1425 /* pre-mapped messages are not bigger than 1 page */
1426 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1428 /* No fancy arithmetic when we do the buffer calculations */
1429 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1431 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1433 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1434 page = txpgs->ibp_pages[ipage];
1435 tx = &tpo->tpo_tx_descs[i];
1437 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1440 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1444 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1446 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1448 list_add(&tx->tx_list, &pool->po_free_list);
1450 page_offset += IBLND_MSG_SIZE;
1451 LASSERT(page_offset <= PAGE_SIZE);
1453 if (page_offset == PAGE_SIZE) {
1456 LASSERT(ipage <= txpgs->ibp_npages);
1462 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1464 LASSERT(fpo->fpo_map_count == 0);
1466 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1467 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1469 struct kib_fast_reg_descriptor *frd, *tmp;
1472 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1474 list_del(&frd->frd_list);
1475 #ifndef HAVE_IB_MAP_MR_SG
1476 ib_free_fast_reg_page_list(frd->frd_frpl);
1478 ib_dereg_mr(frd->frd_mr);
1479 LIBCFS_FREE(frd, sizeof(*frd));
1482 if (i < fpo->fast_reg.fpo_pool_size)
1483 CERROR("FastReg pool still has %d regions registered\n",
1484 fpo->fast_reg.fpo_pool_size - i);
1488 kiblnd_hdev_decref(fpo->fpo_hdev);
1490 LIBCFS_FREE(fpo, sizeof(*fpo));
1494 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1496 kib_fmr_pool_t *fpo, *tmp;
1498 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1499 list_del(&fpo->fpo_list);
1500 kiblnd_destroy_fmr_pool(fpo);
1505 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1508 int size = tunables->lnd_fmr_pool_size / ncpts;
1510 return max(IBLND_FMR_POOL, size);
1514 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1517 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1519 return max(IBLND_FMR_POOL_FLUSH, size);
1522 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1524 struct ib_fmr_pool_param param = {
1525 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1526 .page_shift = PAGE_SHIFT,
1527 .access = (IB_ACCESS_LOCAL_WRITE |
1528 IB_ACCESS_REMOTE_WRITE),
1529 .pool_size = fps->fps_pool_size,
1530 .dirty_watermark = fps->fps_flush_trigger,
1531 .flush_function = NULL,
1533 .cache = !!fps->fps_cache };
1536 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1538 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1539 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1541 CERROR("Failed to create FMR pool: %d\n", rc);
1543 CERROR("FMRs are not supported\n");
1545 fpo->fpo_is_fmr = true;
1550 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo,
1553 struct kib_fast_reg_descriptor *frd, *tmp;
1556 fpo->fpo_is_fmr = false;
1558 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1559 fpo->fast_reg.fpo_pool_size = 0;
1560 for (i = 0; i < fps->fps_pool_size; i++) {
1561 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1564 CERROR("Failed to allocate a new fast_reg descriptor\n");
1570 #ifndef HAVE_IB_MAP_MR_SG
1571 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1572 LNET_MAX_PAYLOAD/PAGE_SIZE);
1573 if (IS_ERR(frd->frd_frpl)) {
1574 rc = PTR_ERR(frd->frd_frpl);
1575 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1577 frd->frd_frpl = NULL;
1582 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1583 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1584 LNET_MAX_PAYLOAD/PAGE_SIZE);
1587 * it is expected to get here if this is an MLX-5 card.
1588 * MLX-4 cards will always use FMR and MLX-5 cards will
1589 * always use fast_reg. It turns out that some MLX-5 cards
1590 * (possibly due to older FW versions) do not natively support
1591 * gaps. So we will need to track them here.
1593 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1594 #ifdef IB_MR_TYPE_SG_GAPS
1595 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1596 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1597 IB_MR_TYPE_SG_GAPS :
1602 LNET_MAX_PAYLOAD/PAGE_SIZE);
1603 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1604 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1605 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1607 if (IS_ERR(frd->frd_mr)) {
1608 rc = PTR_ERR(frd->frd_mr);
1609 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1614 /* There appears to be a bug in MLX5 code where you must
1615 * invalidate the rkey of a new FastReg pool before first
1616 * using it. Thus, I am marking the FRD invalid here. */
1617 frd->frd_valid = false;
1619 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1620 fpo->fast_reg.fpo_pool_size++;
1627 ib_dereg_mr(frd->frd_mr);
1628 #ifndef HAVE_IB_MAP_MR_SG
1630 ib_free_fast_reg_page_list(frd->frd_frpl);
1632 LIBCFS_FREE(frd, sizeof(*frd));
1635 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1637 list_del(&frd->frd_list);
1638 #ifndef HAVE_IB_MAP_MR_SG
1639 ib_free_fast_reg_page_list(frd->frd_frpl);
1641 ib_dereg_mr(frd->frd_mr);
1642 LIBCFS_FREE(frd, sizeof(*frd));
1649 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1651 kib_dev_t *dev = fps->fps_net->ibn_dev;
1652 kib_fmr_pool_t *fpo;
1655 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1659 memset(fpo, 0, sizeof(*fpo));
1661 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1663 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1664 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1666 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1670 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1671 fpo->fpo_owner = fps;
1677 kiblnd_hdev_decref(fpo->fpo_hdev);
1678 LIBCFS_FREE(fpo, sizeof(*fpo));
1683 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1685 if (fps->fps_net == NULL) /* intialized? */
1688 spin_lock(&fps->fps_lock);
1690 while (!list_empty(&fps->fps_pool_list)) {
1691 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1692 kib_fmr_pool_t, fpo_list);
1693 fpo->fpo_failed = 1;
1694 list_del(&fpo->fpo_list);
1695 if (fpo->fpo_map_count == 0)
1696 list_add(&fpo->fpo_list, zombies);
1698 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1701 spin_unlock(&fps->fps_lock);
1705 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1707 if (fps->fps_net != NULL) { /* initialized? */
1708 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1709 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1714 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1716 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1718 kib_fmr_pool_t *fpo;
1721 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1726 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1727 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1728 fps->fps_cache = tunables->lnd_fmr_cache;
1730 spin_lock_init(&fps->fps_lock);
1731 INIT_LIST_HEAD(&fps->fps_pool_list);
1732 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1734 rc = kiblnd_create_fmr_pool(fps, &fpo);
1736 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1742 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, time64_t now)
1744 if (fpo->fpo_map_count != 0) /* still in use */
1746 if (fpo->fpo_failed)
1748 return now >= fpo->fpo_deadline;
1752 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1754 kib_hca_dev_t *hdev;
1755 __u64 *pages = tx->tx_pages;
1760 hdev = tx->tx_pool->tpo_hdev;
1762 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1763 for (size = 0; size < rd->rd_frags[i].rf_nob;
1764 size += hdev->ibh_page_size) {
1765 pages[npages++] = (rd->rd_frags[i].rf_addr &
1766 hdev->ibh_page_mask) + size;
1774 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1776 struct list_head zombies = LIST_HEAD_INIT(zombies);
1777 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1778 kib_fmr_poolset_t *fps;
1779 time64_t now = ktime_get_seconds();
1780 kib_fmr_pool_t *tmp;
1786 fps = fpo->fpo_owner;
1787 if (fpo->fpo_is_fmr) {
1788 if (fmr->fmr_pfmr) {
1789 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1791 fmr->fmr_pfmr = NULL;
1795 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1799 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1802 frd->frd_valid = false;
1803 spin_lock(&fps->fps_lock);
1804 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1805 spin_unlock(&fps->fps_lock);
1806 fmr->fmr_frd = NULL;
1809 fmr->fmr_pool = NULL;
1811 spin_lock(&fps->fps_lock);
1812 fpo->fpo_map_count--; /* decref the pool */
1814 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1815 /* the first pool is persistent */
1816 if (fps->fps_pool_list.next == &fpo->fpo_list)
1819 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1820 list_move(&fpo->fpo_list, &zombies);
1824 spin_unlock(&fps->fps_lock);
1826 if (!list_empty(&zombies))
1827 kiblnd_destroy_fmr_pool_list(&zombies);
1831 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1832 __u32 nob, __u64 iov, kib_fmr_t *fmr)
1834 kib_fmr_pool_t *fpo;
1835 __u64 *pages = tx->tx_pages;
1837 bool is_rx = (rd != tx->tx_rd);
1838 bool tx_pages_mapped = 0;
1843 spin_lock(&fps->fps_lock);
1844 version = fps->fps_version;
1845 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1846 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1847 fpo->fpo_map_count++;
1849 if (fpo->fpo_is_fmr) {
1850 struct ib_pool_fmr *pfmr;
1852 spin_unlock(&fps->fps_lock);
1854 if (!tx_pages_mapped) {
1855 npages = kiblnd_map_tx_pages(tx, rd);
1856 tx_pages_mapped = 1;
1859 pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1860 pages, npages, iov);
1861 if (likely(!IS_ERR(pfmr))) {
1862 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1864 fmr->fmr_frd = NULL;
1865 fmr->fmr_pfmr = pfmr;
1866 fmr->fmr_pool = fpo;
1871 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1872 struct kib_fast_reg_descriptor *frd;
1873 #ifdef HAVE_IB_MAP_MR_SG
1874 struct ib_reg_wr *wr;
1877 struct ib_rdma_wr *wr;
1878 struct ib_fast_reg_page_list *frpl;
1882 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1883 struct kib_fast_reg_descriptor,
1885 list_del(&frd->frd_list);
1886 spin_unlock(&fps->fps_lock);
1888 #ifndef HAVE_IB_MAP_MR_SG
1889 frpl = frd->frd_frpl;
1893 if (!frd->frd_valid) {
1894 struct ib_rdma_wr *inv_wr;
1895 __u32 key = is_rx ? mr->rkey : mr->lkey;
1897 inv_wr = &frd->frd_inv_wr;
1898 memset(inv_wr, 0, sizeof(*inv_wr));
1900 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1901 inv_wr->wr.wr_id = IBLND_WID_MR;
1902 inv_wr->wr.ex.invalidate_rkey = key;
1905 key = ib_inc_rkey(key);
1906 ib_update_fast_reg_key(mr, key);
1909 #ifdef HAVE_IB_MAP_MR_SG
1910 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1911 n = ib_map_mr_sg(mr, tx->tx_frags,
1912 tx->tx_nfrags, NULL, PAGE_SIZE);
1914 n = ib_map_mr_sg(mr, tx->tx_frags,
1915 tx->tx_nfrags, PAGE_SIZE);
1917 if (unlikely(n != tx->tx_nfrags)) {
1918 CERROR("Failed to map mr %d/%d "
1919 "elements\n", n, tx->tx_nfrags);
1920 return n < 0 ? n : -EINVAL;
1923 wr = &frd->frd_fastreg_wr;
1924 memset(wr, 0, sizeof(*wr));
1926 wr->wr.opcode = IB_WR_REG_MR;
1927 wr->wr.wr_id = IBLND_WID_MR;
1929 wr->wr.send_flags = 0;
1931 wr->key = is_rx ? mr->rkey : mr->lkey;
1932 wr->access = (IB_ACCESS_LOCAL_WRITE |
1933 IB_ACCESS_REMOTE_WRITE);
1935 if (!tx_pages_mapped) {
1936 npages = kiblnd_map_tx_pages(tx, rd);
1937 tx_pages_mapped = 1;
1940 LASSERT(npages <= frpl->max_page_list_len);
1941 memcpy(frpl->page_list, pages,
1942 sizeof(*pages) * npages);
1944 /* Prepare FastReg WR */
1945 wr = &frd->frd_fastreg_wr;
1946 memset(wr, 0, sizeof(*wr));
1948 wr->wr.opcode = IB_WR_FAST_REG_MR;
1949 wr->wr.wr_id = IBLND_WID_MR;
1951 wr->wr.wr.fast_reg.iova_start = iov;
1952 wr->wr.wr.fast_reg.page_list = frpl;
1953 wr->wr.wr.fast_reg.page_list_len = npages;
1954 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1955 wr->wr.wr.fast_reg.length = nob;
1956 wr->wr.wr.fast_reg.rkey =
1957 is_rx ? mr->rkey : mr->lkey;
1958 wr->wr.wr.fast_reg.access_flags =
1959 (IB_ACCESS_LOCAL_WRITE |
1960 IB_ACCESS_REMOTE_WRITE);
1963 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1965 fmr->fmr_pfmr = NULL;
1966 fmr->fmr_pool = fpo;
1969 spin_unlock(&fps->fps_lock);
1973 spin_lock(&fps->fps_lock);
1974 fpo->fpo_map_count--;
1975 if (rc != -EAGAIN) {
1976 spin_unlock(&fps->fps_lock);
1980 /* EAGAIN and ... */
1981 if (version != fps->fps_version) {
1982 spin_unlock(&fps->fps_lock);
1987 if (fps->fps_increasing) {
1988 spin_unlock(&fps->fps_lock);
1989 CDEBUG(D_NET, "Another thread is allocating new "
1990 "FMR pool, waiting for her to complete\n");
1996 if (ktime_get_seconds() < fps->fps_next_retry) {
1997 /* someone failed recently */
1998 spin_unlock(&fps->fps_lock);
2002 fps->fps_increasing = 1;
2003 spin_unlock(&fps->fps_lock);
2005 CDEBUG(D_NET, "Allocate new FMR pool\n");
2006 rc = kiblnd_create_fmr_pool(fps, &fpo);
2007 spin_lock(&fps->fps_lock);
2008 fps->fps_increasing = 0;
2011 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2013 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2015 spin_unlock(&fps->fps_lock);
2021 kiblnd_fini_pool(kib_pool_t *pool)
2023 LASSERT(list_empty(&pool->po_free_list));
2024 LASSERT(pool->po_allocated == 0);
2026 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2030 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2032 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2034 memset(pool, 0, sizeof(kib_pool_t));
2035 INIT_LIST_HEAD(&pool->po_free_list);
2036 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2037 pool->po_owner = ps;
2038 pool->po_size = size;
2042 kiblnd_destroy_pool_list(struct list_head *head)
2046 while (!list_empty(head)) {
2047 pool = list_entry(head->next, kib_pool_t, po_list);
2048 list_del(&pool->po_list);
2050 LASSERT(pool->po_owner != NULL);
2051 pool->po_owner->ps_pool_destroy(pool);
2056 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2058 if (ps->ps_net == NULL) /* intialized? */
2061 spin_lock(&ps->ps_lock);
2062 while (!list_empty(&ps->ps_pool_list)) {
2063 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2064 kib_pool_t, po_list);
2066 list_del(&po->po_list);
2067 if (po->po_allocated == 0)
2068 list_add(&po->po_list, zombies);
2070 list_add(&po->po_list, &ps->ps_failed_pool_list);
2072 spin_unlock(&ps->ps_lock);
2076 kiblnd_fini_poolset(kib_poolset_t *ps)
2078 if (ps->ps_net != NULL) { /* initialized? */
2079 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2080 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2085 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2086 kib_net_t *net, char *name, int size,
2087 kib_ps_pool_create_t po_create,
2088 kib_ps_pool_destroy_t po_destroy,
2089 kib_ps_node_init_t nd_init,
2090 kib_ps_node_fini_t nd_fini)
2095 memset(ps, 0, sizeof(kib_poolset_t));
2099 ps->ps_pool_create = po_create;
2100 ps->ps_pool_destroy = po_destroy;
2101 ps->ps_node_init = nd_init;
2102 ps->ps_node_fini = nd_fini;
2103 ps->ps_pool_size = size;
2104 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2105 >= sizeof(ps->ps_name))
2107 spin_lock_init(&ps->ps_lock);
2108 INIT_LIST_HEAD(&ps->ps_pool_list);
2109 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2111 rc = ps->ps_pool_create(ps, size, &pool);
2113 list_add(&pool->po_list, &ps->ps_pool_list);
2115 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2121 kiblnd_pool_is_idle(kib_pool_t *pool, time64_t now)
2123 if (pool->po_allocated != 0) /* still in use */
2125 if (pool->po_failed)
2127 return now >= pool->po_deadline;
2131 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2133 struct list_head zombies = LIST_HEAD_INIT(zombies);
2134 kib_poolset_t *ps = pool->po_owner;
2136 time64_t now = ktime_get_seconds();
2138 spin_lock(&ps->ps_lock);
2140 if (ps->ps_node_fini != NULL)
2141 ps->ps_node_fini(pool, node);
2143 LASSERT(pool->po_allocated > 0);
2144 list_add(node, &pool->po_free_list);
2145 pool->po_allocated--;
2147 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2148 /* the first pool is persistent */
2149 if (ps->ps_pool_list.next == &pool->po_list)
2152 if (kiblnd_pool_is_idle(pool, now))
2153 list_move(&pool->po_list, &zombies);
2155 spin_unlock(&ps->ps_lock);
2157 if (!list_empty(&zombies))
2158 kiblnd_destroy_pool_list(&zombies);
2162 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2164 struct list_head *node;
2167 unsigned int interval = 1;
2168 ktime_t time_before;
2169 unsigned int trips = 0;
2172 spin_lock(&ps->ps_lock);
2173 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2174 if (list_empty(&pool->po_free_list))
2177 pool->po_allocated++;
2178 pool->po_deadline = ktime_get_seconds() +
2179 IBLND_POOL_DEADLINE;
2180 node = pool->po_free_list.next;
2183 if (ps->ps_node_init != NULL) {
2184 /* still hold the lock */
2185 ps->ps_node_init(pool, node);
2187 spin_unlock(&ps->ps_lock);
2191 /* no available tx pool and ... */
2192 if (ps->ps_increasing) {
2193 /* another thread is allocating a new pool */
2194 spin_unlock(&ps->ps_lock);
2196 CDEBUG(D_NET, "Another thread is allocating new "
2197 "%s pool, waiting %d HZs for her to complete."
2199 ps->ps_name, interval, trips);
2201 set_current_state(TASK_INTERRUPTIBLE);
2202 schedule_timeout(interval);
2203 if (interval < cfs_time_seconds(1))
2209 if (ktime_get_seconds() < ps->ps_next_retry) {
2210 /* someone failed recently */
2211 spin_unlock(&ps->ps_lock);
2215 ps->ps_increasing = 1;
2216 spin_unlock(&ps->ps_lock);
2218 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2219 time_before = ktime_get();
2220 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2221 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2222 ktime_ms_delta(ktime_get(), time_before));
2224 spin_lock(&ps->ps_lock);
2225 ps->ps_increasing = 0;
2227 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2229 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2230 CERROR("Can't allocate new %s pool because out of memory\n",
2233 spin_unlock(&ps->ps_lock);
2239 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2241 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2244 LASSERT (pool->po_allocated == 0);
2246 if (tpo->tpo_tx_pages != NULL) {
2247 kiblnd_unmap_tx_pool(tpo);
2248 kiblnd_free_pages(tpo->tpo_tx_pages);
2251 if (tpo->tpo_tx_descs == NULL)
2254 for (i = 0; i < pool->po_size; i++) {
2255 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2256 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2258 list_del(&tx->tx_list);
2259 if (tx->tx_pages != NULL)
2260 LIBCFS_FREE(tx->tx_pages,
2262 sizeof(*tx->tx_pages));
2263 if (tx->tx_frags != NULL)
2264 LIBCFS_FREE(tx->tx_frags,
2265 (1 + IBLND_MAX_RDMA_FRAGS) *
2266 sizeof(*tx->tx_frags));
2267 if (tx->tx_wrq != NULL)
2268 LIBCFS_FREE(tx->tx_wrq,
2269 (1 + IBLND_MAX_RDMA_FRAGS) *
2270 sizeof(*tx->tx_wrq));
2271 if (tx->tx_sge != NULL)
2272 LIBCFS_FREE(tx->tx_sge,
2273 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2274 sizeof(*tx->tx_sge));
2275 if (tx->tx_rd != NULL)
2276 LIBCFS_FREE(tx->tx_rd,
2277 offsetof(kib_rdma_desc_t,
2278 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2281 LIBCFS_FREE(tpo->tpo_tx_descs,
2282 pool->po_size * sizeof(kib_tx_t));
2284 kiblnd_fini_pool(pool);
2285 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2288 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2290 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2293 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2294 ntx = tunables->lnd_ntx / ncpts;
2296 return max(IBLND_TX_POOL, ntx);
2300 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2307 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2309 CERROR("Failed to allocate TX pool\n");
2313 pool = &tpo->tpo_pool;
2314 kiblnd_init_pool(ps, pool, size);
2315 tpo->tpo_tx_descs = NULL;
2316 tpo->tpo_tx_pages = NULL;
2318 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2319 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2320 CERROR("Can't allocate tx pages: %d\n", npg);
2321 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2325 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2326 size * sizeof(kib_tx_t));
2327 if (tpo->tpo_tx_descs == NULL) {
2328 CERROR("Can't allocate %d tx descriptors\n", size);
2329 ps->ps_pool_destroy(pool);
2333 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2335 for (i = 0; i < size; i++) {
2336 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2337 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2340 if (ps->ps_net->ibn_fmr_ps != NULL) {
2341 LIBCFS_CPT_ALLOC(tx->tx_pages,
2342 lnet_cpt_table(), ps->ps_cpt,
2343 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2344 if (tx->tx_pages == NULL)
2348 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2349 (1 + IBLND_MAX_RDMA_FRAGS) *
2350 sizeof(*tx->tx_frags));
2351 if (tx->tx_frags == NULL)
2354 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2356 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2357 (1 + IBLND_MAX_RDMA_FRAGS) *
2358 sizeof(*tx->tx_wrq));
2359 if (tx->tx_wrq == NULL)
2362 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2363 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2364 sizeof(*tx->tx_sge));
2365 if (tx->tx_sge == NULL)
2368 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2369 offsetof(kib_rdma_desc_t,
2370 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2371 if (tx->tx_rd == NULL)
2376 kiblnd_map_tx_pool(tpo);
2381 ps->ps_pool_destroy(pool);
2386 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2388 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2390 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
2392 tx->tx_cookie = tps->tps_next_tx_cookie++;
2396 kiblnd_net_fini_pools(kib_net_t *net)
2400 cfs_cpt_for_each(i, lnet_cpt_table()) {
2401 kib_tx_poolset_t *tps;
2402 kib_fmr_poolset_t *fps;
2404 if (net->ibn_tx_ps != NULL) {
2405 tps = net->ibn_tx_ps[i];
2406 kiblnd_fini_poolset(&tps->tps_poolset);
2409 if (net->ibn_fmr_ps != NULL) {
2410 fps = net->ibn_fmr_ps[i];
2411 kiblnd_fini_fmr_poolset(fps);
2415 if (net->ibn_tx_ps != NULL) {
2416 cfs_percpt_free(net->ibn_tx_ps);
2417 net->ibn_tx_ps = NULL;
2420 if (net->ibn_fmr_ps != NULL) {
2421 cfs_percpt_free(net->ibn_fmr_ps);
2422 net->ibn_fmr_ps = NULL;
2427 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2430 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2431 #ifdef HAVE_IB_GET_DMA_MR
2432 unsigned long flags;
2438 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2440 #ifdef HAVE_IB_GET_DMA_MR
2441 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2443 * if lnd_map_on_demand is zero then we have effectively disabled
2444 * FMR or FastReg and we're using global memory regions
2447 if (!tunables->lnd_map_on_demand) {
2448 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2450 goto create_tx_pool;
2453 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2456 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2457 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2458 tunables->lnd_fmr_pool_size,
2459 tunables->lnd_ntx / 4);
2464 /* TX pool must be created later than FMR, see LU-2268
2466 LASSERT(net->ibn_tx_ps == NULL);
2468 /* premapping can fail if ibd_nmr > 1, so we always create
2469 * FMR pool and map-on-demand if premapping failed */
2471 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2472 sizeof(kib_fmr_poolset_t));
2473 if (net->ibn_fmr_ps == NULL) {
2474 CERROR("Failed to allocate FMR pool array\n");
2479 for (i = 0; i < ncpts; i++) {
2480 cpt = (cpts == NULL) ? i : cpts[i];
2481 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2484 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2491 LASSERT(i == ncpts);
2493 #ifdef HAVE_IB_GET_DMA_MR
2496 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2497 sizeof(kib_tx_poolset_t));
2498 if (net->ibn_tx_ps == NULL) {
2499 CERROR("Failed to allocate tx pool array\n");
2504 for (i = 0; i < ncpts; i++) {
2505 cpt = (cpts == NULL) ? i : cpts[i];
2506 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2508 kiblnd_tx_pool_size(ni, ncpts),
2509 kiblnd_create_tx_pool,
2510 kiblnd_destroy_tx_pool,
2511 kiblnd_tx_init, NULL);
2513 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2521 kiblnd_net_fini_pools(net);
2527 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2529 struct ib_device_attr *dev_attr;
2532 /* It's safe to assume a HCA can handle a page size
2533 * matching that of the native system */
2534 hdev->ibh_page_shift = PAGE_SHIFT;
2535 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2536 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2538 #ifndef HAVE_IB_DEVICE_ATTRS
2539 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2540 if (dev_attr == NULL) {
2541 CERROR("Out of memory\n");
2545 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2547 CERROR("Failed to query IB device: %d\n", rc);
2548 goto out_clean_attr;
2551 dev_attr = &hdev->ibh_ibdev->attrs;
2554 hdev->ibh_mr_size = dev_attr->max_mr_size;
2556 /* Setup device Memory Registration capabilities */
2557 if (hdev->ibh_ibdev->alloc_fmr &&
2558 hdev->ibh_ibdev->dealloc_fmr &&
2559 hdev->ibh_ibdev->map_phys_fmr &&
2560 hdev->ibh_ibdev->unmap_fmr) {
2561 LCONSOLE_INFO("Using FMR for registration\n");
2562 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2563 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2564 LCONSOLE_INFO("Using FastReg for registration\n");
2565 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2566 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2567 #ifdef IB_DEVICE_SG_GAPS_REG
2568 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2569 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2576 if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2577 hdev->ibh_mr_shift = 64;
2581 #ifndef HAVE_IB_DEVICE_ATTRS
2583 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2587 CERROR("IB device does not support FMRs nor FastRegs, can't "
2588 "register memory: %d\n", rc);
2589 else if (rc == -EINVAL)
2590 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2594 #ifdef HAVE_IB_GET_DMA_MR
2596 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2598 if (hdev->ibh_mrs == NULL)
2601 ib_dereg_mr(hdev->ibh_mrs);
2603 hdev->ibh_mrs = NULL;
2608 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2610 #ifdef HAVE_IB_GET_DMA_MR
2611 kiblnd_hdev_cleanup_mrs(hdev);
2614 if (hdev->ibh_pd != NULL)
2615 ib_dealloc_pd(hdev->ibh_pd);
2617 if (hdev->ibh_cmid != NULL)
2618 rdma_destroy_id(hdev->ibh_cmid);
2620 LIBCFS_FREE(hdev, sizeof(*hdev));
2623 #ifdef HAVE_IB_GET_DMA_MR
2625 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2628 int acflags = IB_ACCESS_LOCAL_WRITE |
2629 IB_ACCESS_REMOTE_WRITE;
2631 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2633 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2634 kiblnd_hdev_cleanup_mrs(hdev);
2645 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2651 kiblnd_dev_need_failover(kib_dev_t *dev)
2653 struct rdma_cm_id *cmid;
2654 struct sockaddr_in srcaddr;
2655 struct sockaddr_in dstaddr;
2658 if (dev->ibd_hdev == NULL || /* initializing */
2659 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2660 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2663 /* XXX: it's UGLY, but I don't have better way to find
2664 * ib-bonding HCA failover because:
2666 * a. no reliable CM event for HCA failover...
2667 * b. no OFED API to get ib_device for current net_device...
2669 * We have only two choices at this point:
2671 * a. rdma_bind_addr(), it will conflict with listener cmid
2672 * b. rdma_resolve_addr() to zero addr */
2673 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2677 CERROR("Failed to create cmid for failover: %d\n", rc);
2681 memset(&srcaddr, 0, sizeof(srcaddr));
2682 srcaddr.sin_family = AF_INET;
2683 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2685 memset(&dstaddr, 0, sizeof(dstaddr));
2686 dstaddr.sin_family = AF_INET;
2687 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2688 (struct sockaddr *)&dstaddr, 1);
2689 if (rc != 0 || cmid->device == NULL) {
2690 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2691 dev->ibd_ifname, &dev->ibd_ifip,
2693 rdma_destroy_id(cmid);
2697 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2698 rdma_destroy_id(cmid);
2703 kiblnd_dev_failover(kib_dev_t *dev)
2705 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2706 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2707 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2708 struct rdma_cm_id *cmid = NULL;
2709 kib_hca_dev_t *hdev = NULL;
2713 struct sockaddr_in addr;
2714 unsigned long flags;
2718 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2719 dev->ibd_can_failover ||
2720 dev->ibd_hdev == NULL);
2722 rc = kiblnd_dev_need_failover(dev);
2726 if (dev->ibd_hdev != NULL &&
2727 dev->ibd_hdev->ibh_cmid != NULL) {
2728 /* XXX it's not good to close old listener at here,
2729 * because we can fail to create new listener.
2730 * But we have to close it now, otherwise rdma_bind_addr
2731 * will return EADDRINUSE... How crap! */
2732 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2734 cmid = dev->ibd_hdev->ibh_cmid;
2735 /* make next schedule of kiblnd_dev_need_failover()
2736 * return 1 for me */
2737 dev->ibd_hdev->ibh_cmid = NULL;
2738 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2740 rdma_destroy_id(cmid);
2743 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2747 CERROR("Failed to create cmid for failover: %d\n", rc);
2751 memset(&addr, 0, sizeof(addr));
2752 addr.sin_family = AF_INET;
2753 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2754 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2756 /* Bind to failover device or port */
2757 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2758 if (rc != 0 || cmid->device == NULL) {
2759 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2760 dev->ibd_ifname, &dev->ibd_ifip,
2762 rdma_destroy_id(cmid);
2766 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2768 CERROR("Failed to allocate kib_hca_dev\n");
2769 rdma_destroy_id(cmid);
2774 atomic_set(&hdev->ibh_ref, 1);
2775 hdev->ibh_dev = dev;
2776 hdev->ibh_cmid = cmid;
2777 hdev->ibh_ibdev = cmid->device;
2779 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2780 pd = ib_alloc_pd(cmid->device, 0);
2782 pd = ib_alloc_pd(cmid->device);
2786 CERROR("Can't allocate PD: %d\n", rc);
2792 rc = rdma_listen(cmid, 0);
2794 CERROR("Can't start new listener: %d\n", rc);
2798 rc = kiblnd_hdev_get_attr(hdev);
2800 CERROR("Can't get device attributes: %d\n", rc);
2804 #ifdef HAVE_IB_GET_DMA_MR
2805 rc = kiblnd_hdev_setup_mrs(hdev);
2807 CERROR("Can't setup device: %d\n", rc);
2812 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2814 old = dev->ibd_hdev;
2815 dev->ibd_hdev = hdev; /* take over the refcount */
2818 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2819 cfs_cpt_for_each(i, lnet_cpt_table()) {
2820 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2823 if (net->ibn_fmr_ps != NULL)
2824 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2829 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2831 if (!list_empty(&zombie_tpo))
2832 kiblnd_destroy_pool_list(&zombie_tpo);
2833 if (!list_empty(&zombie_ppo))
2834 kiblnd_destroy_pool_list(&zombie_ppo);
2835 if (!list_empty(&zombie_fpo))
2836 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2838 kiblnd_hdev_decref(hdev);
2841 dev->ibd_failed_failover++;
2843 dev->ibd_failed_failover = 0;
2849 kiblnd_destroy_dev (kib_dev_t *dev)
2851 LASSERT (dev->ibd_nnets == 0);
2852 LASSERT(list_empty(&dev->ibd_nets));
2854 list_del(&dev->ibd_fail_list);
2855 list_del(&dev->ibd_list);
2857 if (dev->ibd_hdev != NULL)
2858 kiblnd_hdev_decref(dev->ibd_hdev);
2860 LIBCFS_FREE(dev, sizeof(*dev));
2864 kiblnd_create_dev(char *ifname)
2866 struct net_device *netdev;
2873 rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2875 CERROR("Can't query IPoIB interface %s: %d\n",
2881 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2885 LIBCFS_ALLOC(dev, sizeof(*dev));
2889 netdev = dev_get_by_name(&init_net, ifname);
2890 if (netdev == NULL) {
2891 dev->ibd_can_failover = 0;
2893 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2897 INIT_LIST_HEAD(&dev->ibd_nets);
2898 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2899 INIT_LIST_HEAD(&dev->ibd_fail_list);
2901 strcpy(&dev->ibd_ifname[0], ifname);
2903 /* initialize the device */
2904 rc = kiblnd_dev_failover(dev);
2906 CERROR("Can't initialize device: %d\n", rc);
2907 LIBCFS_FREE(dev, sizeof(*dev));
2911 list_add_tail(&dev->ibd_list,
2912 &kiblnd_data.kib_devs);
2917 kiblnd_base_shutdown(void)
2919 struct kib_sched_info *sched;
2922 LASSERT(list_empty(&kiblnd_data.kib_devs));
2924 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2925 atomic_read(&libcfs_kmemory));
2927 switch (kiblnd_data.kib_init) {
2931 case IBLND_INIT_ALL:
2932 case IBLND_INIT_DATA:
2933 LASSERT (kiblnd_data.kib_peers != NULL);
2934 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2935 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2937 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2938 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2939 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2940 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2942 /* flag threads to terminate; wake and wait for them to die */
2943 kiblnd_data.kib_shutdown = 1;
2945 /* NB: we really want to stop scheduler threads net by net
2946 * instead of the whole module, this should be improved
2947 * with dynamic configuration LNet */
2948 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2949 wake_up_all(&sched->ibs_waitq);
2951 wake_up_all(&kiblnd_data.kib_connd_waitq);
2952 wake_up_all(&kiblnd_data.kib_failover_waitq);
2955 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2958 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2959 "Waiting for %d threads to terminate\n",
2960 atomic_read(&kiblnd_data.kib_nthreads));
2961 set_current_state(TASK_UNINTERRUPTIBLE);
2962 schedule_timeout(cfs_time_seconds(1));
2967 case IBLND_INIT_NOTHING:
2971 if (kiblnd_data.kib_peers != NULL) {
2972 LIBCFS_FREE(kiblnd_data.kib_peers,
2973 sizeof(struct list_head) *
2974 kiblnd_data.kib_peer_hash_size);
2977 if (kiblnd_data.kib_scheds != NULL)
2978 cfs_percpt_free(kiblnd_data.kib_scheds);
2980 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2981 atomic_read(&libcfs_kmemory));
2983 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2984 module_put(THIS_MODULE);
2988 kiblnd_shutdown(struct lnet_ni *ni)
2990 kib_net_t *net = ni->ni_data;
2991 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2993 unsigned long flags;
2995 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3000 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
3001 atomic_read(&libcfs_kmemory));
3003 write_lock_irqsave(g_lock, flags);
3004 net->ibn_shutdown = 1;
3005 write_unlock_irqrestore(g_lock, flags);
3007 switch (net->ibn_init) {
3011 case IBLND_INIT_ALL:
3012 /* nuke all existing peers within this net */
3013 kiblnd_del_peer(ni, LNET_NID_ANY);
3015 /* Wait for all peer_ni state to clean up */
3017 while (atomic_read(&net->ibn_npeers) != 0) {
3020 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3021 "%s: waiting for %d peers to disconnect\n",
3022 libcfs_nid2str(ni->ni_nid),
3023 atomic_read(&net->ibn_npeers));
3024 set_current_state(TASK_UNINTERRUPTIBLE);
3025 schedule_timeout(cfs_time_seconds(1));
3028 kiblnd_net_fini_pools(net);
3030 write_lock_irqsave(g_lock, flags);
3031 LASSERT(net->ibn_dev->ibd_nnets > 0);
3032 net->ibn_dev->ibd_nnets--;
3033 list_del(&net->ibn_list);
3034 write_unlock_irqrestore(g_lock, flags);
3038 case IBLND_INIT_NOTHING:
3039 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3041 if (net->ibn_dev != NULL &&
3042 net->ibn_dev->ibd_nnets == 0)
3043 kiblnd_destroy_dev(net->ibn_dev);
3048 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3049 atomic_read(&libcfs_kmemory));
3051 net->ibn_init = IBLND_INIT_NOTHING;
3054 LIBCFS_FREE(net, sizeof(*net));
3057 if (list_empty(&kiblnd_data.kib_devs))
3058 kiblnd_base_shutdown();
3063 kiblnd_base_startup(void)
3065 struct kib_sched_info *sched;
3069 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3071 try_module_get(THIS_MODULE);
3072 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3074 rwlock_init(&kiblnd_data.kib_global_lock);
3076 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3077 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3079 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3080 LIBCFS_ALLOC(kiblnd_data.kib_peers,
3081 sizeof(struct list_head) *
3082 kiblnd_data.kib_peer_hash_size);
3083 if (kiblnd_data.kib_peers == NULL)
3086 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3087 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3089 spin_lock_init(&kiblnd_data.kib_connd_lock);
3090 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3091 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3092 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3093 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3095 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3096 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3098 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3100 if (kiblnd_data.kib_scheds == NULL)
3103 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3106 spin_lock_init(&sched->ibs_lock);
3107 INIT_LIST_HEAD(&sched->ibs_conns);
3108 init_waitqueue_head(&sched->ibs_waitq);
3110 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3111 if (*kiblnd_tunables.kib_nscheds > 0) {
3112 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3114 /* max to half of CPUs, another half is reserved for
3115 * upper layer modules */
3116 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3119 sched->ibs_nthreads_max = nthrs;
3123 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3125 /* lists/ptrs/locks initialised */
3126 kiblnd_data.kib_init = IBLND_INIT_DATA;
3127 /*****************************************************/
3129 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3131 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3135 if (*kiblnd_tunables.kib_dev_failover != 0)
3136 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3140 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3144 /* flag everything initialised */
3145 kiblnd_data.kib_init = IBLND_INIT_ALL;
3146 /*****************************************************/
3151 kiblnd_base_shutdown();
3156 kiblnd_start_schedulers(struct kib_sched_info *sched)
3162 if (sched->ibs_nthreads == 0) {
3163 if (*kiblnd_tunables.kib_nscheds > 0) {
3164 nthrs = sched->ibs_nthreads_max;
3166 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3168 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3169 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3172 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3173 /* increase one thread if there is new interface */
3174 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3177 for (i = 0; i < nthrs; i++) {
3180 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3181 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3182 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3183 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3187 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3188 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3192 sched->ibs_nthreads += i;
3197 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3203 for (i = 0; i < ncpts; i++) {
3204 struct kib_sched_info *sched;
3206 cpt = (cpts == NULL) ? i : cpts[i];
3207 sched = kiblnd_data.kib_scheds[cpt];
3209 if (!newdev && sched->ibs_nthreads > 0)
3212 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3214 CERROR("Failed to start scheduler threads for %s\n",
3223 kiblnd_dev_search(char *ifname)
3225 kib_dev_t *alias = NULL;
3230 colon = strchr(ifname, ':');
3231 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3232 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3238 colon2 = strchr(dev->ibd_ifname, ':');
3244 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3256 kiblnd_startup(struct lnet_ni *ni)
3259 kib_dev_t *ibdev = NULL;
3261 unsigned long flags;
3266 LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3268 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3269 rc = kiblnd_base_startup();
3274 LIBCFS_ALLOC(net, sizeof(*net));
3279 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3281 kiblnd_tunables_setup(ni);
3283 if (ni->ni_interfaces[0] != NULL) {
3284 /* Use the IPoIB interface specified in 'networks=' */
3286 CLASSERT(LNET_INTERFACES_NUM > 1);
3287 if (ni->ni_interfaces[1] != NULL) {
3288 CERROR("Multiple interfaces not supported\n");
3292 ifname = ni->ni_interfaces[0];
3294 ifname = *kiblnd_tunables.kib_default_ipif;
3297 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3298 CERROR("IPoIB interface name too long: %s\n", ifname);
3302 ibdev = kiblnd_dev_search(ifname);
3304 newdev = ibdev == NULL;
3305 /* hmm...create kib_dev even for alias */
3306 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3307 ibdev = kiblnd_create_dev(ifname);
3312 node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3313 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3315 net->ibn_dev = ibdev;
3316 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3318 rc = kiblnd_dev_start_threads(ibdev, newdev,
3319 ni->ni_cpts, ni->ni_ncpts);
3323 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3325 CERROR("Failed to initialize NI pools: %d\n", rc);
3329 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3331 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3332 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3334 net->ibn_init = IBLND_INIT_ALL;
3339 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3340 kiblnd_destroy_dev(ibdev);
3342 kiblnd_shutdown(ni);
3344 CDEBUG(D_NET, "kiblnd_startup failed\n");
3348 static struct lnet_lnd the_o2iblnd = {
3349 .lnd_type = O2IBLND,
3350 .lnd_startup = kiblnd_startup,
3351 .lnd_shutdown = kiblnd_shutdown,
3352 .lnd_ctl = kiblnd_ctl,
3353 .lnd_query = kiblnd_query,
3354 .lnd_send = kiblnd_send,
3355 .lnd_recv = kiblnd_recv,
3358 static void __exit ko2iblnd_exit(void)
3360 lnet_unregister_lnd(&the_o2iblnd);
3363 static int __init ko2iblnd_init(void)
3367 CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3368 CLASSERT(offsetof(kib_msg_t,
3369 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3371 CLASSERT(offsetof(kib_msg_t,
3372 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3375 rc = kiblnd_tunables_init();
3379 lnet_register_lnd(&the_o2iblnd);
3384 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3385 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3386 MODULE_VERSION("2.8.0");
3387 MODULE_LICENSE("GPL");
3389 module_init(ko2iblnd_init);
3390 module_exit(ko2iblnd_exit);