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 %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);
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,
1595 IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) ?
1596 IB_MR_TYPE_SG_GAPS :
1598 LNET_MAX_PAYLOAD/PAGE_SIZE);
1600 if (IS_ERR(frd->frd_mr)) {
1601 rc = PTR_ERR(frd->frd_mr);
1602 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1607 /* There appears to be a bug in MLX5 code where you must
1608 * invalidate the rkey of a new FastReg pool before first
1609 * using it. Thus, I am marking the FRD invalid here. */
1610 frd->frd_valid = false;
1612 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1613 fpo->fast_reg.fpo_pool_size++;
1620 ib_dereg_mr(frd->frd_mr);
1621 #ifndef HAVE_IB_MAP_MR_SG
1623 ib_free_fast_reg_page_list(frd->frd_frpl);
1625 LIBCFS_FREE(frd, sizeof(*frd));
1628 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1630 list_del(&frd->frd_list);
1631 #ifndef HAVE_IB_MAP_MR_SG
1632 ib_free_fast_reg_page_list(frd->frd_frpl);
1634 ib_dereg_mr(frd->frd_mr);
1635 LIBCFS_FREE(frd, sizeof(*frd));
1642 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1644 kib_dev_t *dev = fps->fps_net->ibn_dev;
1645 kib_fmr_pool_t *fpo;
1648 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1652 memset(fpo, 0, sizeof(*fpo));
1654 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1656 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1657 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1659 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1663 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1664 fpo->fpo_owner = fps;
1670 kiblnd_hdev_decref(fpo->fpo_hdev);
1671 LIBCFS_FREE(fpo, sizeof(*fpo));
1676 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1678 if (fps->fps_net == NULL) /* intialized? */
1681 spin_lock(&fps->fps_lock);
1683 while (!list_empty(&fps->fps_pool_list)) {
1684 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1685 kib_fmr_pool_t, fpo_list);
1686 fpo->fpo_failed = 1;
1687 list_del(&fpo->fpo_list);
1688 if (fpo->fpo_map_count == 0)
1689 list_add(&fpo->fpo_list, zombies);
1691 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1694 spin_unlock(&fps->fps_lock);
1698 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1700 if (fps->fps_net != NULL) { /* initialized? */
1701 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1702 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1707 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1709 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1711 kib_fmr_pool_t *fpo;
1714 memset(fps, 0, sizeof(kib_fmr_poolset_t));
1719 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1720 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1721 fps->fps_cache = tunables->lnd_fmr_cache;
1723 spin_lock_init(&fps->fps_lock);
1724 INIT_LIST_HEAD(&fps->fps_pool_list);
1725 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1727 rc = kiblnd_create_fmr_pool(fps, &fpo);
1729 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1735 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1737 if (fpo->fpo_map_count != 0) /* still in use */
1739 if (fpo->fpo_failed)
1741 return cfs_time_aftereq(now, fpo->fpo_deadline);
1745 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1747 kib_hca_dev_t *hdev;
1748 __u64 *pages = tx->tx_pages;
1753 hdev = tx->tx_pool->tpo_hdev;
1755 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1756 for (size = 0; size < rd->rd_frags[i].rf_nob;
1757 size += hdev->ibh_page_size) {
1758 pages[npages++] = (rd->rd_frags[i].rf_addr &
1759 hdev->ibh_page_mask) + size;
1767 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1769 struct list_head zombies = LIST_HEAD_INIT(zombies);
1770 kib_fmr_pool_t *fpo = fmr->fmr_pool;
1771 kib_fmr_poolset_t *fps;
1772 cfs_time_t now = cfs_time_current();
1773 kib_fmr_pool_t *tmp;
1779 fps = fpo->fpo_owner;
1780 if (fpo->fpo_is_fmr) {
1781 if (fmr->fmr_pfmr) {
1782 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1784 fmr->fmr_pfmr = NULL;
1788 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1792 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1795 frd->frd_valid = false;
1796 spin_lock(&fps->fps_lock);
1797 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1798 spin_unlock(&fps->fps_lock);
1799 fmr->fmr_frd = NULL;
1802 fmr->fmr_pool = NULL;
1804 spin_lock(&fps->fps_lock);
1805 fpo->fpo_map_count--; /* decref the pool */
1807 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1808 /* the first pool is persistent */
1809 if (fps->fps_pool_list.next == &fpo->fpo_list)
1812 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1813 list_move(&fpo->fpo_list, &zombies);
1817 spin_unlock(&fps->fps_lock);
1819 if (!list_empty(&zombies))
1820 kiblnd_destroy_fmr_pool_list(&zombies);
1824 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1825 __u32 nob, __u64 iov, kib_fmr_t *fmr)
1827 kib_fmr_pool_t *fpo;
1828 __u64 *pages = tx->tx_pages;
1830 bool is_rx = (rd != tx->tx_rd);
1831 bool tx_pages_mapped = 0;
1836 spin_lock(&fps->fps_lock);
1837 version = fps->fps_version;
1838 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1839 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1840 fpo->fpo_map_count++;
1842 if (fpo->fpo_is_fmr) {
1843 struct ib_pool_fmr *pfmr;
1845 spin_unlock(&fps->fps_lock);
1847 if (!tx_pages_mapped) {
1848 npages = kiblnd_map_tx_pages(tx, rd);
1849 tx_pages_mapped = 1;
1852 pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1853 pages, npages, iov);
1854 if (likely(!IS_ERR(pfmr))) {
1855 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1857 fmr->fmr_frd = NULL;
1858 fmr->fmr_pfmr = pfmr;
1859 fmr->fmr_pool = fpo;
1864 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1865 struct kib_fast_reg_descriptor *frd;
1866 #ifdef HAVE_IB_MAP_MR_SG
1867 struct ib_reg_wr *wr;
1870 struct ib_rdma_wr *wr;
1871 struct ib_fast_reg_page_list *frpl;
1875 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1876 struct kib_fast_reg_descriptor,
1878 list_del(&frd->frd_list);
1879 spin_unlock(&fps->fps_lock);
1881 #ifndef HAVE_IB_MAP_MR_SG
1882 frpl = frd->frd_frpl;
1886 if (!frd->frd_valid) {
1887 struct ib_rdma_wr *inv_wr;
1888 __u32 key = is_rx ? mr->rkey : mr->lkey;
1890 inv_wr = &frd->frd_inv_wr;
1891 memset(inv_wr, 0, sizeof(*inv_wr));
1893 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1894 inv_wr->wr.wr_id = IBLND_WID_MR;
1895 inv_wr->wr.ex.invalidate_rkey = key;
1898 key = ib_inc_rkey(key);
1899 ib_update_fast_reg_key(mr, key);
1902 #ifdef HAVE_IB_MAP_MR_SG
1903 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1904 n = ib_map_mr_sg(mr, tx->tx_frags,
1905 tx->tx_nfrags, NULL, PAGE_SIZE);
1907 n = ib_map_mr_sg(mr, tx->tx_frags,
1908 tx->tx_nfrags, PAGE_SIZE);
1910 if (unlikely(n != tx->tx_nfrags)) {
1911 CERROR("Failed to map mr %d/%d "
1912 "elements\n", n, tx->tx_nfrags);
1913 return n < 0 ? n : -EINVAL;
1916 wr = &frd->frd_fastreg_wr;
1917 memset(wr, 0, sizeof(*wr));
1919 wr->wr.opcode = IB_WR_REG_MR;
1920 wr->wr.wr_id = IBLND_WID_MR;
1922 wr->wr.send_flags = 0;
1924 wr->key = is_rx ? mr->rkey : mr->lkey;
1925 wr->access = (IB_ACCESS_LOCAL_WRITE |
1926 IB_ACCESS_REMOTE_WRITE);
1928 if (!tx_pages_mapped) {
1929 npages = kiblnd_map_tx_pages(tx, rd);
1930 tx_pages_mapped = 1;
1933 LASSERT(npages <= frpl->max_page_list_len);
1934 memcpy(frpl->page_list, pages,
1935 sizeof(*pages) * npages);
1937 /* Prepare FastReg WR */
1938 wr = &frd->frd_fastreg_wr;
1939 memset(wr, 0, sizeof(*wr));
1941 wr->wr.opcode = IB_WR_FAST_REG_MR;
1942 wr->wr.wr_id = IBLND_WID_MR;
1944 wr->wr.wr.fast_reg.iova_start = iov;
1945 wr->wr.wr.fast_reg.page_list = frpl;
1946 wr->wr.wr.fast_reg.page_list_len = npages;
1947 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1948 wr->wr.wr.fast_reg.length = nob;
1949 wr->wr.wr.fast_reg.rkey =
1950 is_rx ? mr->rkey : mr->lkey;
1951 wr->wr.wr.fast_reg.access_flags =
1952 (IB_ACCESS_LOCAL_WRITE |
1953 IB_ACCESS_REMOTE_WRITE);
1956 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1958 fmr->fmr_pfmr = NULL;
1959 fmr->fmr_pool = fpo;
1962 spin_unlock(&fps->fps_lock);
1966 spin_lock(&fps->fps_lock);
1967 fpo->fpo_map_count--;
1968 if (rc != -EAGAIN) {
1969 spin_unlock(&fps->fps_lock);
1973 /* EAGAIN and ... */
1974 if (version != fps->fps_version) {
1975 spin_unlock(&fps->fps_lock);
1980 if (fps->fps_increasing) {
1981 spin_unlock(&fps->fps_lock);
1982 CDEBUG(D_NET, "Another thread is allocating new "
1983 "FMR pool, waiting for her to complete\n");
1989 if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1990 /* someone failed recently */
1991 spin_unlock(&fps->fps_lock);
1995 fps->fps_increasing = 1;
1996 spin_unlock(&fps->fps_lock);
1998 CDEBUG(D_NET, "Allocate new FMR pool\n");
1999 rc = kiblnd_create_fmr_pool(fps, &fpo);
2000 spin_lock(&fps->fps_lock);
2001 fps->fps_increasing = 0;
2004 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2006 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2008 spin_unlock(&fps->fps_lock);
2014 kiblnd_fini_pool(kib_pool_t *pool)
2016 LASSERT(list_empty(&pool->po_free_list));
2017 LASSERT(pool->po_allocated == 0);
2019 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2023 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2025 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2027 memset(pool, 0, sizeof(kib_pool_t));
2028 INIT_LIST_HEAD(&pool->po_free_list);
2029 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2030 pool->po_owner = ps;
2031 pool->po_size = size;
2035 kiblnd_destroy_pool_list(struct list_head *head)
2039 while (!list_empty(head)) {
2040 pool = list_entry(head->next, kib_pool_t, po_list);
2041 list_del(&pool->po_list);
2043 LASSERT(pool->po_owner != NULL);
2044 pool->po_owner->ps_pool_destroy(pool);
2049 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2051 if (ps->ps_net == NULL) /* intialized? */
2054 spin_lock(&ps->ps_lock);
2055 while (!list_empty(&ps->ps_pool_list)) {
2056 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2057 kib_pool_t, po_list);
2059 list_del(&po->po_list);
2060 if (po->po_allocated == 0)
2061 list_add(&po->po_list, zombies);
2063 list_add(&po->po_list, &ps->ps_failed_pool_list);
2065 spin_unlock(&ps->ps_lock);
2069 kiblnd_fini_poolset(kib_poolset_t *ps)
2071 if (ps->ps_net != NULL) { /* initialized? */
2072 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2073 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2078 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2079 kib_net_t *net, char *name, int size,
2080 kib_ps_pool_create_t po_create,
2081 kib_ps_pool_destroy_t po_destroy,
2082 kib_ps_node_init_t nd_init,
2083 kib_ps_node_fini_t nd_fini)
2088 memset(ps, 0, sizeof(kib_poolset_t));
2092 ps->ps_pool_create = po_create;
2093 ps->ps_pool_destroy = po_destroy;
2094 ps->ps_node_init = nd_init;
2095 ps->ps_node_fini = nd_fini;
2096 ps->ps_pool_size = size;
2097 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2098 >= sizeof(ps->ps_name))
2100 spin_lock_init(&ps->ps_lock);
2101 INIT_LIST_HEAD(&ps->ps_pool_list);
2102 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2104 rc = ps->ps_pool_create(ps, size, &pool);
2106 list_add(&pool->po_list, &ps->ps_pool_list);
2108 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2114 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2116 if (pool->po_allocated != 0) /* still in use */
2118 if (pool->po_failed)
2120 return cfs_time_aftereq(now, pool->po_deadline);
2124 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2126 struct list_head zombies = LIST_HEAD_INIT(zombies);
2127 kib_poolset_t *ps = pool->po_owner;
2129 cfs_time_t now = cfs_time_current();
2131 spin_lock(&ps->ps_lock);
2133 if (ps->ps_node_fini != NULL)
2134 ps->ps_node_fini(pool, node);
2136 LASSERT(pool->po_allocated > 0);
2137 list_add(node, &pool->po_free_list);
2138 pool->po_allocated--;
2140 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2141 /* the first pool is persistent */
2142 if (ps->ps_pool_list.next == &pool->po_list)
2145 if (kiblnd_pool_is_idle(pool, now))
2146 list_move(&pool->po_list, &zombies);
2148 spin_unlock(&ps->ps_lock);
2150 if (!list_empty(&zombies))
2151 kiblnd_destroy_pool_list(&zombies);
2155 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2157 struct list_head *node;
2160 unsigned int interval = 1;
2161 cfs_time_t time_before;
2162 unsigned int trips = 0;
2165 spin_lock(&ps->ps_lock);
2166 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2167 if (list_empty(&pool->po_free_list))
2170 pool->po_allocated++;
2171 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2172 node = pool->po_free_list.next;
2175 if (ps->ps_node_init != NULL) {
2176 /* still hold the lock */
2177 ps->ps_node_init(pool, node);
2179 spin_unlock(&ps->ps_lock);
2183 /* no available tx pool and ... */
2184 if (ps->ps_increasing) {
2185 /* another thread is allocating a new pool */
2186 spin_unlock(&ps->ps_lock);
2188 CDEBUG(D_NET, "Another thread is allocating new "
2189 "%s pool, waiting %d HZs for her to complete."
2191 ps->ps_name, interval, trips);
2193 set_current_state(TASK_INTERRUPTIBLE);
2194 schedule_timeout(interval);
2195 if (interval < cfs_time_seconds(1))
2201 if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2202 /* someone failed recently */
2203 spin_unlock(&ps->ps_lock);
2207 ps->ps_increasing = 1;
2208 spin_unlock(&ps->ps_lock);
2210 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2211 time_before = cfs_time_current();
2212 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2213 CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2214 cfs_time_current() - time_before);
2216 spin_lock(&ps->ps_lock);
2217 ps->ps_increasing = 0;
2219 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2221 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2222 CERROR("Can't allocate new %s pool because out of memory\n",
2225 spin_unlock(&ps->ps_lock);
2231 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2233 kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2236 LASSERT (pool->po_allocated == 0);
2238 if (tpo->tpo_tx_pages != NULL) {
2239 kiblnd_unmap_tx_pool(tpo);
2240 kiblnd_free_pages(tpo->tpo_tx_pages);
2243 if (tpo->tpo_tx_descs == NULL)
2246 for (i = 0; i < pool->po_size; i++) {
2247 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2248 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2250 list_del(&tx->tx_list);
2251 if (tx->tx_pages != NULL)
2252 LIBCFS_FREE(tx->tx_pages,
2254 sizeof(*tx->tx_pages));
2255 if (tx->tx_frags != NULL)
2256 LIBCFS_FREE(tx->tx_frags,
2257 (1 + IBLND_MAX_RDMA_FRAGS) *
2258 sizeof(*tx->tx_frags));
2259 if (tx->tx_wrq != NULL)
2260 LIBCFS_FREE(tx->tx_wrq,
2261 (1 + IBLND_MAX_RDMA_FRAGS) *
2262 sizeof(*tx->tx_wrq));
2263 if (tx->tx_sge != NULL)
2264 LIBCFS_FREE(tx->tx_sge,
2265 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2266 sizeof(*tx->tx_sge));
2267 if (tx->tx_rd != NULL)
2268 LIBCFS_FREE(tx->tx_rd,
2269 offsetof(kib_rdma_desc_t,
2270 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2273 LIBCFS_FREE(tpo->tpo_tx_descs,
2274 pool->po_size * sizeof(kib_tx_t));
2276 kiblnd_fini_pool(pool);
2277 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2280 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2282 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2285 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2286 ntx = tunables->lnd_ntx / ncpts;
2288 return max(IBLND_TX_POOL, ntx);
2292 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2299 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2301 CERROR("Failed to allocate TX pool\n");
2305 pool = &tpo->tpo_pool;
2306 kiblnd_init_pool(ps, pool, size);
2307 tpo->tpo_tx_descs = NULL;
2308 tpo->tpo_tx_pages = NULL;
2310 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2311 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2312 CERROR("Can't allocate tx pages: %d\n", npg);
2313 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2317 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2318 size * sizeof(kib_tx_t));
2319 if (tpo->tpo_tx_descs == NULL) {
2320 CERROR("Can't allocate %d tx descriptors\n", size);
2321 ps->ps_pool_destroy(pool);
2325 memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2327 for (i = 0; i < size; i++) {
2328 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2329 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2332 if (ps->ps_net->ibn_fmr_ps != NULL) {
2333 LIBCFS_CPT_ALLOC(tx->tx_pages,
2334 lnet_cpt_table(), ps->ps_cpt,
2335 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2336 if (tx->tx_pages == NULL)
2340 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2341 (1 + IBLND_MAX_RDMA_FRAGS) *
2342 sizeof(*tx->tx_frags));
2343 if (tx->tx_frags == NULL)
2346 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2348 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2349 (1 + IBLND_MAX_RDMA_FRAGS) *
2350 sizeof(*tx->tx_wrq));
2351 if (tx->tx_wrq == NULL)
2354 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2355 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2356 sizeof(*tx->tx_sge));
2357 if (tx->tx_sge == NULL)
2360 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2361 offsetof(kib_rdma_desc_t,
2362 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2363 if (tx->tx_rd == NULL)
2368 kiblnd_map_tx_pool(tpo);
2373 ps->ps_pool_destroy(pool);
2378 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2380 kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2382 kib_tx_t *tx = list_entry(node, kib_tx_t, tx_list);
2384 tx->tx_cookie = tps->tps_next_tx_cookie++;
2388 kiblnd_net_fini_pools(kib_net_t *net)
2392 cfs_cpt_for_each(i, lnet_cpt_table()) {
2393 kib_tx_poolset_t *tps;
2394 kib_fmr_poolset_t *fps;
2396 if (net->ibn_tx_ps != NULL) {
2397 tps = net->ibn_tx_ps[i];
2398 kiblnd_fini_poolset(&tps->tps_poolset);
2401 if (net->ibn_fmr_ps != NULL) {
2402 fps = net->ibn_fmr_ps[i];
2403 kiblnd_fini_fmr_poolset(fps);
2407 if (net->ibn_tx_ps != NULL) {
2408 cfs_percpt_free(net->ibn_tx_ps);
2409 net->ibn_tx_ps = NULL;
2412 if (net->ibn_fmr_ps != NULL) {
2413 cfs_percpt_free(net->ibn_fmr_ps);
2414 net->ibn_fmr_ps = NULL;
2419 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2422 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2423 #ifdef HAVE_IB_GET_DMA_MR
2424 unsigned long flags;
2430 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2432 #ifdef HAVE_IB_GET_DMA_MR
2433 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2435 * if lnd_map_on_demand is zero then we have effectively disabled
2436 * FMR or FastReg and we're using global memory regions
2439 if (!tunables->lnd_map_on_demand) {
2440 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2442 goto create_tx_pool;
2445 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2448 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2449 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2450 tunables->lnd_fmr_pool_size,
2451 tunables->lnd_ntx / 4);
2456 /* TX pool must be created later than FMR, see LU-2268
2458 LASSERT(net->ibn_tx_ps == NULL);
2460 /* premapping can fail if ibd_nmr > 1, so we always create
2461 * FMR pool and map-on-demand if premapping failed */
2463 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2464 sizeof(kib_fmr_poolset_t));
2465 if (net->ibn_fmr_ps == NULL) {
2466 CERROR("Failed to allocate FMR pool array\n");
2471 for (i = 0; i < ncpts; i++) {
2472 cpt = (cpts == NULL) ? i : cpts[i];
2473 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2476 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2483 LASSERT(i == ncpts);
2485 #ifdef HAVE_IB_GET_DMA_MR
2488 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2489 sizeof(kib_tx_poolset_t));
2490 if (net->ibn_tx_ps == NULL) {
2491 CERROR("Failed to allocate tx pool array\n");
2496 for (i = 0; i < ncpts; i++) {
2497 cpt = (cpts == NULL) ? i : cpts[i];
2498 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2500 kiblnd_tx_pool_size(ni, ncpts),
2501 kiblnd_create_tx_pool,
2502 kiblnd_destroy_tx_pool,
2503 kiblnd_tx_init, NULL);
2505 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2513 kiblnd_net_fini_pools(net);
2519 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2521 struct ib_device_attr *dev_attr;
2524 /* It's safe to assume a HCA can handle a page size
2525 * matching that of the native system */
2526 hdev->ibh_page_shift = PAGE_SHIFT;
2527 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2528 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2530 #ifndef HAVE_IB_DEVICE_ATTRS
2531 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2532 if (dev_attr == NULL) {
2533 CERROR("Out of memory\n");
2537 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2539 CERROR("Failed to query IB device: %d\n", rc);
2540 goto out_clean_attr;
2543 dev_attr = &hdev->ibh_ibdev->attrs;
2546 hdev->ibh_mr_size = dev_attr->max_mr_size;
2548 /* Setup device Memory Registration capabilities */
2549 if (hdev->ibh_ibdev->alloc_fmr &&
2550 hdev->ibh_ibdev->dealloc_fmr &&
2551 hdev->ibh_ibdev->map_phys_fmr &&
2552 hdev->ibh_ibdev->unmap_fmr) {
2553 LCONSOLE_INFO("Using FMR for registration\n");
2554 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2555 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2556 LCONSOLE_INFO("Using FastReg for registration\n");
2557 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2558 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2559 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2560 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2566 if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2567 hdev->ibh_mr_shift = 64;
2571 #ifndef HAVE_IB_DEVICE_ATTRS
2573 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2577 CERROR("IB device does not support FMRs nor FastRegs, can't "
2578 "register memory: %d\n", rc);
2579 else if (rc == -EINVAL)
2580 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2584 #ifdef HAVE_IB_GET_DMA_MR
2586 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2588 if (hdev->ibh_mrs == NULL)
2591 ib_dereg_mr(hdev->ibh_mrs);
2593 hdev->ibh_mrs = NULL;
2598 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2600 #ifdef HAVE_IB_GET_DMA_MR
2601 kiblnd_hdev_cleanup_mrs(hdev);
2604 if (hdev->ibh_pd != NULL)
2605 ib_dealloc_pd(hdev->ibh_pd);
2607 if (hdev->ibh_cmid != NULL)
2608 rdma_destroy_id(hdev->ibh_cmid);
2610 LIBCFS_FREE(hdev, sizeof(*hdev));
2613 #ifdef HAVE_IB_GET_DMA_MR
2615 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2618 int acflags = IB_ACCESS_LOCAL_WRITE |
2619 IB_ACCESS_REMOTE_WRITE;
2621 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2623 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2624 kiblnd_hdev_cleanup_mrs(hdev);
2635 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2641 kiblnd_dev_need_failover(kib_dev_t *dev)
2643 struct rdma_cm_id *cmid;
2644 struct sockaddr_in srcaddr;
2645 struct sockaddr_in dstaddr;
2648 if (dev->ibd_hdev == NULL || /* initializing */
2649 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2650 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2653 /* XXX: it's UGLY, but I don't have better way to find
2654 * ib-bonding HCA failover because:
2656 * a. no reliable CM event for HCA failover...
2657 * b. no OFED API to get ib_device for current net_device...
2659 * We have only two choices at this point:
2661 * a. rdma_bind_addr(), it will conflict with listener cmid
2662 * b. rdma_resolve_addr() to zero addr */
2663 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2667 CERROR("Failed to create cmid for failover: %d\n", rc);
2671 memset(&srcaddr, 0, sizeof(srcaddr));
2672 srcaddr.sin_family = AF_INET;
2673 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2675 memset(&dstaddr, 0, sizeof(dstaddr));
2676 dstaddr.sin_family = AF_INET;
2677 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2678 (struct sockaddr *)&dstaddr, 1);
2679 if (rc != 0 || cmid->device == NULL) {
2680 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2681 dev->ibd_ifname, &dev->ibd_ifip,
2683 rdma_destroy_id(cmid);
2687 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2688 rdma_destroy_id(cmid);
2693 kiblnd_dev_failover(kib_dev_t *dev)
2695 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2696 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2697 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2698 struct rdma_cm_id *cmid = NULL;
2699 kib_hca_dev_t *hdev = NULL;
2703 struct sockaddr_in addr;
2704 unsigned long flags;
2708 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2709 dev->ibd_can_failover ||
2710 dev->ibd_hdev == NULL);
2712 rc = kiblnd_dev_need_failover(dev);
2716 if (dev->ibd_hdev != NULL &&
2717 dev->ibd_hdev->ibh_cmid != NULL) {
2718 /* XXX it's not good to close old listener at here,
2719 * because we can fail to create new listener.
2720 * But we have to close it now, otherwise rdma_bind_addr
2721 * will return EADDRINUSE... How crap! */
2722 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2724 cmid = dev->ibd_hdev->ibh_cmid;
2725 /* make next schedule of kiblnd_dev_need_failover()
2726 * return 1 for me */
2727 dev->ibd_hdev->ibh_cmid = NULL;
2728 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2730 rdma_destroy_id(cmid);
2733 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2737 CERROR("Failed to create cmid for failover: %d\n", rc);
2741 memset(&addr, 0, sizeof(addr));
2742 addr.sin_family = AF_INET;
2743 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2744 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2746 /* Bind to failover device or port */
2747 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2748 if (rc != 0 || cmid->device == NULL) {
2749 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2750 dev->ibd_ifname, &dev->ibd_ifip,
2752 rdma_destroy_id(cmid);
2756 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2758 CERROR("Failed to allocate kib_hca_dev\n");
2759 rdma_destroy_id(cmid);
2764 atomic_set(&hdev->ibh_ref, 1);
2765 hdev->ibh_dev = dev;
2766 hdev->ibh_cmid = cmid;
2767 hdev->ibh_ibdev = cmid->device;
2769 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2770 pd = ib_alloc_pd(cmid->device, 0);
2772 pd = ib_alloc_pd(cmid->device);
2776 CERROR("Can't allocate PD: %d\n", rc);
2782 rc = rdma_listen(cmid, 0);
2784 CERROR("Can't start new listener: %d\n", rc);
2788 rc = kiblnd_hdev_get_attr(hdev);
2790 CERROR("Can't get device attributes: %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);
2802 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2804 old = dev->ibd_hdev;
2805 dev->ibd_hdev = hdev; /* take over the refcount */
2808 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2809 cfs_cpt_for_each(i, lnet_cpt_table()) {
2810 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2813 if (net->ibn_fmr_ps != NULL)
2814 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2819 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2821 if (!list_empty(&zombie_tpo))
2822 kiblnd_destroy_pool_list(&zombie_tpo);
2823 if (!list_empty(&zombie_ppo))
2824 kiblnd_destroy_pool_list(&zombie_ppo);
2825 if (!list_empty(&zombie_fpo))
2826 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2828 kiblnd_hdev_decref(hdev);
2831 dev->ibd_failed_failover++;
2833 dev->ibd_failed_failover = 0;
2839 kiblnd_destroy_dev (kib_dev_t *dev)
2841 LASSERT (dev->ibd_nnets == 0);
2842 LASSERT(list_empty(&dev->ibd_nets));
2844 list_del(&dev->ibd_fail_list);
2845 list_del(&dev->ibd_list);
2847 if (dev->ibd_hdev != NULL)
2848 kiblnd_hdev_decref(dev->ibd_hdev);
2850 LIBCFS_FREE(dev, sizeof(*dev));
2854 kiblnd_create_dev(char *ifname)
2856 struct net_device *netdev;
2863 rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2865 CERROR("Can't query IPoIB interface %s: %d\n",
2871 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2875 LIBCFS_ALLOC(dev, sizeof(*dev));
2879 netdev = dev_get_by_name(&init_net, ifname);
2880 if (netdev == NULL) {
2881 dev->ibd_can_failover = 0;
2883 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2887 INIT_LIST_HEAD(&dev->ibd_nets);
2888 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2889 INIT_LIST_HEAD(&dev->ibd_fail_list);
2891 strcpy(&dev->ibd_ifname[0], ifname);
2893 /* initialize the device */
2894 rc = kiblnd_dev_failover(dev);
2896 CERROR("Can't initialize device: %d\n", rc);
2897 LIBCFS_FREE(dev, sizeof(*dev));
2901 list_add_tail(&dev->ibd_list,
2902 &kiblnd_data.kib_devs);
2907 kiblnd_base_shutdown(void)
2909 struct kib_sched_info *sched;
2912 LASSERT(list_empty(&kiblnd_data.kib_devs));
2914 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2915 atomic_read(&libcfs_kmemory));
2917 switch (kiblnd_data.kib_init) {
2921 case IBLND_INIT_ALL:
2922 case IBLND_INIT_DATA:
2923 LASSERT (kiblnd_data.kib_peers != NULL);
2924 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2925 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2927 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2928 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2929 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2930 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2932 /* flag threads to terminate; wake and wait for them to die */
2933 kiblnd_data.kib_shutdown = 1;
2935 /* NB: we really want to stop scheduler threads net by net
2936 * instead of the whole module, this should be improved
2937 * with dynamic configuration LNet */
2938 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2939 wake_up_all(&sched->ibs_waitq);
2941 wake_up_all(&kiblnd_data.kib_connd_waitq);
2942 wake_up_all(&kiblnd_data.kib_failover_waitq);
2945 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2948 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2949 "Waiting for %d threads to terminate\n",
2950 atomic_read(&kiblnd_data.kib_nthreads));
2951 set_current_state(TASK_UNINTERRUPTIBLE);
2952 schedule_timeout(cfs_time_seconds(1));
2957 case IBLND_INIT_NOTHING:
2961 if (kiblnd_data.kib_peers != NULL) {
2962 LIBCFS_FREE(kiblnd_data.kib_peers,
2963 sizeof(struct list_head) *
2964 kiblnd_data.kib_peer_hash_size);
2967 if (kiblnd_data.kib_scheds != NULL)
2968 cfs_percpt_free(kiblnd_data.kib_scheds);
2970 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2971 atomic_read(&libcfs_kmemory));
2973 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2974 module_put(THIS_MODULE);
2978 kiblnd_shutdown(struct lnet_ni *ni)
2980 kib_net_t *net = ni->ni_data;
2981 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2983 unsigned long flags;
2985 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2990 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2991 atomic_read(&libcfs_kmemory));
2993 write_lock_irqsave(g_lock, flags);
2994 net->ibn_shutdown = 1;
2995 write_unlock_irqrestore(g_lock, flags);
2997 switch (net->ibn_init) {
3001 case IBLND_INIT_ALL:
3002 /* nuke all existing peers within this net */
3003 kiblnd_del_peer(ni, LNET_NID_ANY);
3005 /* Wait for all peer_ni state to clean up */
3007 while (atomic_read(&net->ibn_npeers) != 0) {
3010 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3011 "%s: waiting for %d peers to disconnect\n",
3012 libcfs_nid2str(ni->ni_nid),
3013 atomic_read(&net->ibn_npeers));
3014 set_current_state(TASK_UNINTERRUPTIBLE);
3015 schedule_timeout(cfs_time_seconds(1));
3018 kiblnd_net_fini_pools(net);
3020 write_lock_irqsave(g_lock, flags);
3021 LASSERT(net->ibn_dev->ibd_nnets > 0);
3022 net->ibn_dev->ibd_nnets--;
3023 list_del(&net->ibn_list);
3024 write_unlock_irqrestore(g_lock, flags);
3028 case IBLND_INIT_NOTHING:
3029 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3031 if (net->ibn_dev != NULL &&
3032 net->ibn_dev->ibd_nnets == 0)
3033 kiblnd_destroy_dev(net->ibn_dev);
3038 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3039 atomic_read(&libcfs_kmemory));
3041 net->ibn_init = IBLND_INIT_NOTHING;
3044 LIBCFS_FREE(net, sizeof(*net));
3047 if (list_empty(&kiblnd_data.kib_devs))
3048 kiblnd_base_shutdown();
3053 kiblnd_base_startup(void)
3055 struct kib_sched_info *sched;
3059 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3061 try_module_get(THIS_MODULE);
3062 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3064 rwlock_init(&kiblnd_data.kib_global_lock);
3066 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3067 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3069 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3070 LIBCFS_ALLOC(kiblnd_data.kib_peers,
3071 sizeof(struct list_head) *
3072 kiblnd_data.kib_peer_hash_size);
3073 if (kiblnd_data.kib_peers == NULL)
3076 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3077 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3079 spin_lock_init(&kiblnd_data.kib_connd_lock);
3080 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3081 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3082 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3083 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3085 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3086 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3088 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3090 if (kiblnd_data.kib_scheds == NULL)
3093 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3096 spin_lock_init(&sched->ibs_lock);
3097 INIT_LIST_HEAD(&sched->ibs_conns);
3098 init_waitqueue_head(&sched->ibs_waitq);
3100 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3101 if (*kiblnd_tunables.kib_nscheds > 0) {
3102 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3104 /* max to half of CPUs, another half is reserved for
3105 * upper layer modules */
3106 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3109 sched->ibs_nthreads_max = nthrs;
3113 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3115 /* lists/ptrs/locks initialised */
3116 kiblnd_data.kib_init = IBLND_INIT_DATA;
3117 /*****************************************************/
3119 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3121 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3125 if (*kiblnd_tunables.kib_dev_failover != 0)
3126 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3130 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3134 /* flag everything initialised */
3135 kiblnd_data.kib_init = IBLND_INIT_ALL;
3136 /*****************************************************/
3141 kiblnd_base_shutdown();
3146 kiblnd_start_schedulers(struct kib_sched_info *sched)
3152 if (sched->ibs_nthreads == 0) {
3153 if (*kiblnd_tunables.kib_nscheds > 0) {
3154 nthrs = sched->ibs_nthreads_max;
3156 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3158 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3159 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3162 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3163 /* increase one thread if there is new interface */
3164 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3167 for (i = 0; i < nthrs; i++) {
3170 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3171 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3172 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3173 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3177 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3178 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3182 sched->ibs_nthreads += i;
3187 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3193 for (i = 0; i < ncpts; i++) {
3194 struct kib_sched_info *sched;
3196 cpt = (cpts == NULL) ? i : cpts[i];
3197 sched = kiblnd_data.kib_scheds[cpt];
3199 if (!newdev && sched->ibs_nthreads > 0)
3202 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3204 CERROR("Failed to start scheduler threads for %s\n",
3213 kiblnd_dev_search(char *ifname)
3215 kib_dev_t *alias = NULL;
3220 colon = strchr(ifname, ':');
3221 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3222 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3228 colon2 = strchr(dev->ibd_ifname, ':');
3234 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3246 kiblnd_startup(struct lnet_ni *ni)
3249 kib_dev_t *ibdev = NULL;
3251 unsigned long flags;
3256 LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3258 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3259 rc = kiblnd_base_startup();
3264 LIBCFS_ALLOC(net, sizeof(*net));
3269 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3271 kiblnd_tunables_setup(ni);
3273 if (ni->ni_interfaces[0] != NULL) {
3274 /* Use the IPoIB interface specified in 'networks=' */
3276 CLASSERT(LNET_INTERFACES_NUM > 1);
3277 if (ni->ni_interfaces[1] != NULL) {
3278 CERROR("Multiple interfaces not supported\n");
3282 ifname = ni->ni_interfaces[0];
3284 ifname = *kiblnd_tunables.kib_default_ipif;
3287 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3288 CERROR("IPoIB interface name too long: %s\n", ifname);
3292 ibdev = kiblnd_dev_search(ifname);
3294 newdev = ibdev == NULL;
3295 /* hmm...create kib_dev even for alias */
3296 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3297 ibdev = kiblnd_create_dev(ifname);
3302 node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3303 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3305 net->ibn_dev = ibdev;
3306 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3308 rc = kiblnd_dev_start_threads(ibdev, newdev,
3309 ni->ni_cpts, ni->ni_ncpts);
3313 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3315 CERROR("Failed to initialize NI pools: %d\n", rc);
3319 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3321 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3322 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3324 net->ibn_init = IBLND_INIT_ALL;
3329 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3330 kiblnd_destroy_dev(ibdev);
3332 kiblnd_shutdown(ni);
3334 CDEBUG(D_NET, "kiblnd_startup failed\n");
3338 static struct lnet_lnd the_o2iblnd = {
3339 .lnd_type = O2IBLND,
3340 .lnd_startup = kiblnd_startup,
3341 .lnd_shutdown = kiblnd_shutdown,
3342 .lnd_ctl = kiblnd_ctl,
3343 .lnd_query = kiblnd_query,
3344 .lnd_send = kiblnd_send,
3345 .lnd_recv = kiblnd_recv,
3348 static void __exit ko2iblnd_exit(void)
3350 lnet_unregister_lnd(&the_o2iblnd);
3353 static int __init ko2iblnd_init(void)
3357 CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3358 CLASSERT(offsetof(kib_msg_t,
3359 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3361 CLASSERT(offsetof(kib_msg_t,
3362 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3365 rc = kiblnd_tunables_init();
3369 lnet_register_lnd(&the_o2iblnd);
3374 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3375 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3376 MODULE_VERSION("2.8.0");
3377 MODULE_LICENSE("GPL");
3379 module_init(ko2iblnd_init);
3380 module_exit(ko2iblnd_exit);