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>
38 #include <linux/inetdevice.h>
42 static struct lnet_lnd the_o2iblnd;
44 struct kib_data kiblnd_data;
47 kiblnd_cksum (void *ptr, int nob)
53 sum = ((sum << 1) | (sum >> 31)) + *c++;
55 /* ensure I don't return 0 (== no checksum) */
56 return (sum == 0) ? 1 : sum;
60 kiblnd_msgtype2str(int type)
63 case IBLND_MSG_CONNREQ:
66 case IBLND_MSG_CONNACK:
72 case IBLND_MSG_IMMEDIATE:
75 case IBLND_MSG_PUT_REQ:
78 case IBLND_MSG_PUT_NAK:
81 case IBLND_MSG_PUT_ACK:
84 case IBLND_MSG_PUT_DONE:
87 case IBLND_MSG_GET_REQ:
90 case IBLND_MSG_GET_DONE:
99 kiblnd_msgtype2size(int type)
101 const int hdr_size = offsetof(struct kib_msg, ibm_u);
104 case IBLND_MSG_CONNREQ:
105 case IBLND_MSG_CONNACK:
106 return hdr_size + sizeof(struct kib_connparams);
111 case IBLND_MSG_IMMEDIATE:
112 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
114 case IBLND_MSG_PUT_REQ:
115 return hdr_size + sizeof(struct kib_putreq_msg);
117 case IBLND_MSG_PUT_ACK:
118 return hdr_size + sizeof(struct kib_putack_msg);
120 case IBLND_MSG_GET_REQ:
121 return hdr_size + sizeof(struct kib_get_msg);
123 case IBLND_MSG_PUT_NAK:
124 case IBLND_MSG_PUT_DONE:
125 case IBLND_MSG_GET_DONE:
126 return hdr_size + sizeof(struct kib_completion_msg);
132 static int kiblnd_unpack_rd(struct kib_msg *msg, int flip)
134 struct kib_rdma_desc *rd;
139 LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
140 msg->ibm_type == IBLND_MSG_PUT_ACK);
142 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
143 &msg->ibm_u.get.ibgm_rd :
144 &msg->ibm_u.putack.ibpam_rd;
147 __swab32s(&rd->rd_key);
148 __swab32s(&rd->rd_nfrags);
153 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
154 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
155 n, IBLND_MAX_RDMA_FRAGS);
159 nob = offsetof(struct kib_msg, ibm_u) +
160 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
162 if (msg->ibm_nob < nob) {
163 CERROR("Short %s: %d(%d)\n",
164 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
171 for (i = 0; i < n; i++) {
172 __swab32s(&rd->rd_frags[i].rf_nob);
173 __swab64s(&rd->rd_frags[i].rf_addr);
179 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
180 int credits, lnet_nid_t dstnid, __u64 dststamp)
182 struct kib_net *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);
203 int kiblnd_unpack_msg(struct kib_msg *msg, int nob)
205 const int hdr_size = offsetof(struct kib_msg, ibm_u);
211 /* 6 bytes are enough to have received magic + version */
213 CERROR("Short message: %d\n", nob);
217 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
219 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
222 CERROR("Bad magic: %08x\n", msg->ibm_magic);
226 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
227 if (version != IBLND_MSG_VERSION &&
228 version != IBLND_MSG_VERSION_1) {
229 CERROR("Bad version: %x\n", version);
233 if (nob < hdr_size) {
234 CERROR("Short message: %d\n", nob);
238 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
240 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
244 /* checksum must be computed with ibm_cksum zero and BEFORE anything
246 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
248 if (msg_cksum != 0 &&
249 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
250 CERROR("Bad checksum\n");
254 msg->ibm_cksum = msg_cksum;
257 /* leave magic unflipped as a clue to peer_ni endianness */
258 msg->ibm_version = version;
259 CLASSERT (sizeof(msg->ibm_type) == 1);
260 CLASSERT (sizeof(msg->ibm_credits) == 1);
261 msg->ibm_nob = msg_nob;
262 __swab64s(&msg->ibm_srcnid);
263 __swab64s(&msg->ibm_srcstamp);
264 __swab64s(&msg->ibm_dstnid);
265 __swab64s(&msg->ibm_dststamp);
268 if (msg->ibm_srcnid == LNET_NID_ANY) {
269 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
273 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
274 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
275 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
279 switch (msg->ibm_type) {
281 CERROR("Unknown message type %x\n", msg->ibm_type);
285 case IBLND_MSG_IMMEDIATE:
286 case IBLND_MSG_PUT_REQ:
289 case IBLND_MSG_PUT_ACK:
290 case IBLND_MSG_GET_REQ:
291 if (kiblnd_unpack_rd(msg, flip))
295 case IBLND_MSG_PUT_NAK:
296 case IBLND_MSG_PUT_DONE:
297 case IBLND_MSG_GET_DONE:
299 __swab32s(&msg->ibm_u.completion.ibcm_status);
302 case IBLND_MSG_CONNREQ:
303 case IBLND_MSG_CONNACK:
305 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
306 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
307 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
315 kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
318 struct kib_peer_ni *peer_ni;
319 struct kib_net *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(struct kib_peer_ni *peer_ni)
361 struct kib_net *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 struct kib_peer_ni *peer_ni;
387 list_for_each(tmp, peer_list) {
389 peer_ni = list_entry(tmp, struct kib_peer_ni, 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(struct kib_peer_ni *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 struct kib_peer_ni *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, struct kib_peer_ni, 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(struct kib_peer_ni *peer_ni)
462 struct list_head *ctmp;
463 struct list_head *cnxt;
464 struct kib_conn *conn;
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, struct kib_conn, 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 struct kib_peer_ni *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, struct kib_peer_ni, 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, LNET_MSG_STATUS_LOCAL_ERROR);
532 static struct kib_conn *
533 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
535 struct kib_peer_ni *peer_ni;
536 struct list_head *ptmp;
537 struct kib_conn *conn;
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, struct kib_peer_ni, 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, struct kib_conn, 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(struct kib_rx *rx)
573 CDEBUG(D_CONSOLE, " %p msg_type %x cred %d\n",
574 rx, rx->rx_msg->ibm_type,
575 rx->rx_msg->ibm_credits);
579 kiblnd_debug_tx(struct kib_tx *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(struct kib_conn *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, struct kib_rx, 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, struct kib_tx, 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, struct kib_tx, 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, struct kib_tx, 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, struct kib_tx, 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, struct kib_tx, 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(struct kib_conn *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
738 int multiplier = 1 + conn->ibc_max_frags;
739 enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
741 /* FastReg needs two extra WRs for map and invalidate */
742 if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
745 /* account for a maximum of ibc_queue_depth in-flight transfers */
746 ret = multiplier * conn->ibc_queue_depth;
748 if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
749 CDEBUG(D_NET, "peer_credits %u will result in send work "
750 "request size %d larger than maximum %d device "
751 "can handle\n", conn->ibc_queue_depth, ret,
752 conn->ibc_hdev->ibh_max_qp_wr);
753 conn->ibc_queue_depth =
754 conn->ibc_hdev->ibh_max_qp_wr / multiplier;
757 /* don't go beyond the maximum the device can handle */
758 return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
762 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
763 int state, int version)
766 * If the new conn is created successfully it takes over the caller's
767 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
768 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
769 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
770 * to destroy 'cmid' here since I'm called from the CM which still has
771 * its ref on 'cmid'). */
772 rwlock_t *glock = &kiblnd_data.kib_global_lock;
773 struct kib_net *net = peer_ni->ibp_ni->ni_data;
775 struct ib_qp_init_attr *init_qp_attr;
776 struct kib_sched_info *sched;
777 #ifdef HAVE_IB_CQ_INIT_ATTR
778 struct ib_cq_init_attr cq_attr = {};
780 struct kib_conn *conn;
787 LASSERT(net != NULL);
788 LASSERT(!in_interrupt());
792 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
793 sched = kiblnd_get_scheduler(cpt);
796 CERROR("no schedulers available. node is unhealthy\n");
801 * The cpt might have changed if we ended up selecting a non cpt
802 * native scheduler. So use the scheduler's cpt instead.
804 cpt = sched->ibs_cpt;
806 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
807 sizeof(*init_qp_attr));
808 if (init_qp_attr == NULL) {
809 CERROR("Can't allocate qp_attr for %s\n",
810 libcfs_nid2str(peer_ni->ibp_nid));
814 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
816 CERROR("Can't allocate connection for %s\n",
817 libcfs_nid2str(peer_ni->ibp_nid));
821 conn->ibc_state = IBLND_CONN_INIT;
822 conn->ibc_version = version;
823 conn->ibc_peer = peer_ni; /* I take the caller's ref */
824 cmid->context = conn; /* for future CM callbacks */
825 conn->ibc_cmid = cmid;
826 conn->ibc_max_frags = peer_ni->ibp_max_frags;
827 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
828 conn->ibc_rxs = NULL;
829 conn->ibc_rx_pages = NULL;
831 INIT_LIST_HEAD(&conn->ibc_early_rxs);
832 INIT_LIST_HEAD(&conn->ibc_tx_noops);
833 INIT_LIST_HEAD(&conn->ibc_tx_queue);
834 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
835 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
836 INIT_LIST_HEAD(&conn->ibc_active_txs);
837 INIT_LIST_HEAD(&conn->ibc_zombie_txs);
838 spin_lock_init(&conn->ibc_lock);
840 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
841 sizeof(*conn->ibc_connvars));
842 if (conn->ibc_connvars == NULL) {
843 CERROR("Can't allocate in-progress connection state\n");
847 write_lock_irqsave(glock, flags);
848 if (dev->ibd_failover) {
849 write_unlock_irqrestore(glock, flags);
850 CERROR("%s: failover in progress\n", dev->ibd_ifname);
854 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
855 /* wakeup failover thread and teardown connection */
856 if (kiblnd_dev_can_failover(dev)) {
857 list_add_tail(&dev->ibd_fail_list,
858 &kiblnd_data.kib_failed_devs);
859 wake_up(&kiblnd_data.kib_failover_waitq);
862 write_unlock_irqrestore(glock, flags);
863 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
864 cmid->device->name, dev->ibd_ifname);
868 kiblnd_hdev_addref_locked(dev->ibd_hdev);
869 conn->ibc_hdev = dev->ibd_hdev;
871 kiblnd_setup_mtu_locked(cmid);
873 write_unlock_irqrestore(glock, flags);
875 #ifdef HAVE_IB_CQ_INIT_ATTR
876 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
877 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
878 cq = ib_create_cq(cmid->device,
879 kiblnd_cq_completion, kiblnd_cq_event, conn,
882 cq = ib_create_cq(cmid->device,
883 kiblnd_cq_completion, kiblnd_cq_event, conn,
884 IBLND_CQ_ENTRIES(conn),
885 kiblnd_get_completion_vector(conn, cpt));
889 * on MLX-5 (possibly MLX-4 as well) this error could be
890 * hit if the concurrent_sends and/or peer_tx_credits is set
891 * too high. Or due to an MLX-5 bug which tries to
892 * allocate 256kb via kmalloc for WR cookie array
894 CERROR("Failed to create CQ with %d CQEs: %ld\n",
895 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
901 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
903 CERROR("Can't request completion notification: %d\n", rc);
907 init_qp_attr->event_handler = kiblnd_qp_event;
908 init_qp_attr->qp_context = conn;
909 init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
910 init_qp_attr->cap.max_recv_sge = 1;
911 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
912 init_qp_attr->qp_type = IB_QPT_RC;
913 init_qp_attr->send_cq = cq;
914 init_qp_attr->recv_cq = cq;
916 * kiblnd_send_wrs() can change the connection's queue depth if
917 * the maximum work requests for the device is maxed out
919 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
920 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
922 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
924 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
925 "send_sge: %d, recv_sge: %d\n",
926 rc, init_qp_attr->cap.max_send_wr,
927 init_qp_attr->cap.max_recv_wr,
928 init_qp_attr->cap.max_send_sge,
929 init_qp_attr->cap.max_recv_sge);
933 conn->ibc_sched = sched;
935 if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
936 CWARN("peer %s - queue depth reduced from %u to %u"
937 " to allow for qp creation\n",
938 libcfs_nid2str(peer_ni->ibp_nid),
939 peer_ni->ibp_queue_depth,
940 conn->ibc_queue_depth);
942 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
943 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
944 if (conn->ibc_rxs == NULL) {
945 CERROR("Cannot allocate RX buffers\n");
949 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
950 IBLND_RX_MSG_PAGES(conn));
954 kiblnd_map_rx_descs(conn);
956 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
958 /* 1 ref for caller and each rxmsg */
959 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
960 conn->ibc_nrx = IBLND_RX_MSGS(conn);
963 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
964 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
966 CERROR("Can't post rxmsg: %d\n", rc);
968 /* Make posted receives complete */
969 kiblnd_abort_receives(conn);
971 /* correct # of posted buffers
972 * NB locking needed now I'm racing with completion */
973 spin_lock_irqsave(&sched->ibs_lock, flags);
974 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
975 spin_unlock_irqrestore(&sched->ibs_lock, flags);
977 /* cmid will be destroyed by CM(ofed) after cm_callback
978 * returned, so we can't refer it anymore
979 * (by kiblnd_connd()->kiblnd_destroy_conn) */
980 rdma_destroy_qp(conn->ibc_cmid);
981 conn->ibc_cmid = NULL;
983 /* Drop my own and unused rxbuffer refcounts */
984 while (i++ <= IBLND_RX_MSGS(conn))
985 kiblnd_conn_decref(conn);
991 /* Init successful! */
992 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
993 state == IBLND_CONN_PASSIVE_WAIT);
994 conn->ibc_state = state;
997 atomic_inc(&net->ibn_nconns);
1001 kiblnd_destroy_conn(conn);
1002 LIBCFS_FREE(conn, sizeof(*conn));
1004 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
1010 kiblnd_destroy_conn(struct kib_conn *conn)
1012 struct rdma_cm_id *cmid = conn->ibc_cmid;
1013 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1016 LASSERT (!in_interrupt());
1017 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1018 LASSERT(list_empty(&conn->ibc_early_rxs));
1019 LASSERT(list_empty(&conn->ibc_tx_noops));
1020 LASSERT(list_empty(&conn->ibc_tx_queue));
1021 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1022 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1023 LASSERT(list_empty(&conn->ibc_active_txs));
1024 LASSERT (conn->ibc_noops_posted == 0);
1025 LASSERT (conn->ibc_nsends_posted == 0);
1027 switch (conn->ibc_state) {
1029 /* conn must be completely disengaged from the network */
1032 case IBLND_CONN_DISCONNECTED:
1033 /* connvars should have been freed already */
1034 LASSERT (conn->ibc_connvars == NULL);
1037 case IBLND_CONN_INIT:
1041 /* conn->ibc_cmid might be destroyed by CM already */
1042 if (cmid != NULL && cmid->qp != NULL)
1043 rdma_destroy_qp(cmid);
1045 if (conn->ibc_cq != NULL) {
1046 rc = ib_destroy_cq(conn->ibc_cq);
1048 CWARN("Error destroying CQ: %d\n", rc);
1051 kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1052 LNET_MSG_STATUS_OK);
1054 if (conn->ibc_rx_pages != NULL)
1055 kiblnd_unmap_rx_descs(conn);
1057 if (conn->ibc_rxs != NULL) {
1058 LIBCFS_FREE(conn->ibc_rxs,
1059 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
1062 if (conn->ibc_connvars != NULL)
1063 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1065 if (conn->ibc_hdev != NULL)
1066 kiblnd_hdev_decref(conn->ibc_hdev);
1068 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1069 if (conn->ibc_state != IBLND_CONN_INIT) {
1070 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1072 kiblnd_peer_decref(peer_ni);
1073 rdma_destroy_id(cmid);
1074 atomic_dec(&net->ibn_nconns);
1079 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1081 struct kib_conn *conn;
1082 struct list_head *ctmp;
1083 struct list_head *cnxt;
1086 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1087 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1089 CDEBUG(D_NET, "Closing conn -> %s, "
1090 "version: %x, reason: %d\n",
1091 libcfs_nid2str(peer_ni->ibp_nid),
1092 conn->ibc_version, why);
1094 kiblnd_close_conn_locked(conn, why);
1102 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1103 int version, __u64 incarnation)
1105 struct kib_conn *conn;
1106 struct list_head *ctmp;
1107 struct list_head *cnxt;
1110 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1111 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1113 if (conn->ibc_version == version &&
1114 conn->ibc_incarnation == incarnation)
1117 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1118 "incarnation:%#llx(%x, %#llx)\n",
1119 libcfs_nid2str(peer_ni->ibp_nid),
1120 conn->ibc_version, conn->ibc_incarnation,
1121 version, incarnation);
1123 kiblnd_close_conn_locked(conn, -ESTALE);
1131 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1133 struct kib_peer_ni *peer_ni;
1134 struct list_head *ptmp;
1135 struct list_head *pnxt;
1139 unsigned long flags;
1142 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1144 if (nid != LNET_NID_ANY)
1145 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1148 hi = kiblnd_data.kib_peer_hash_size - 1;
1151 for (i = lo; i <= hi; i++) {
1152 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1154 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
1155 LASSERT(!kiblnd_peer_idle(peer_ni));
1157 if (peer_ni->ibp_ni != ni)
1160 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1163 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1167 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1169 /* wildcards always succeed */
1170 if (nid == LNET_NID_ANY)
1173 return (count == 0) ? -ENOENT : 0;
1177 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1179 struct libcfs_ioctl_data *data = arg;
1183 case IOC_LIBCFS_GET_PEER: {
1187 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1189 data->ioc_nid = nid;
1190 data->ioc_count = count;
1194 case IOC_LIBCFS_DEL_PEER: {
1195 rc = kiblnd_del_peer(ni, data->ioc_nid);
1198 case IOC_LIBCFS_GET_CONN: {
1199 struct kib_conn *conn;
1202 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1208 LASSERT(conn->ibc_cmid != NULL);
1209 data->ioc_nid = conn->ibc_peer->ibp_nid;
1210 if (conn->ibc_cmid->route.path_rec == NULL)
1211 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1214 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1215 kiblnd_conn_decref(conn);
1218 case IOC_LIBCFS_CLOSE_CONNECTION: {
1219 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1231 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1233 time64_t last_alive = 0;
1234 time64_t now = ktime_get_seconds();
1235 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1236 struct kib_peer_ni *peer_ni;
1237 unsigned long flags;
1239 read_lock_irqsave(glock, flags);
1241 peer_ni = kiblnd_find_peer_locked(ni, nid);
1242 if (peer_ni != NULL)
1243 last_alive = peer_ni->ibp_last_alive;
1245 read_unlock_irqrestore(glock, flags);
1247 if (last_alive != 0)
1250 /* peer_ni is not persistent in hash, trigger peer_ni creation
1251 * and connection establishment with a NULL tx */
1252 if (peer_ni == NULL)
1253 kiblnd_launch_tx(ni, NULL, nid);
1255 CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n",
1256 libcfs_nid2str(nid), peer_ni,
1257 last_alive ? now - last_alive : -1);
1262 kiblnd_free_pages(struct kib_pages *p)
1264 int npages = p->ibp_npages;
1267 for (i = 0; i < npages; i++) {
1268 if (p->ibp_pages[i] != NULL)
1269 __free_page(p->ibp_pages[i]);
1272 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1276 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1278 struct kib_pages *p;
1281 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1282 offsetof(struct kib_pages, ibp_pages[npages]));
1284 CERROR("Can't allocate descriptor for %d pages\n", npages);
1288 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1289 p->ibp_npages = npages;
1291 for (i = 0; i < npages; i++) {
1292 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1294 if (p->ibp_pages[i] == NULL) {
1295 CERROR("Can't allocate page %d of %d\n", i, npages);
1296 kiblnd_free_pages(p);
1306 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1311 LASSERT (conn->ibc_rxs != NULL);
1312 LASSERT (conn->ibc_hdev != NULL);
1314 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1315 rx = &conn->ibc_rxs[i];
1317 LASSERT(rx->rx_nob >= 0); /* not posted */
1319 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1320 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1322 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1325 kiblnd_free_pages(conn->ibc_rx_pages);
1327 conn->ibc_rx_pages = NULL;
1331 kiblnd_map_rx_descs(struct kib_conn *conn)
1339 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1340 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1341 rx = &conn->ibc_rxs[i];
1344 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1347 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1348 rx->rx_msg, IBLND_MSG_SIZE,
1350 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1352 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1354 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1355 i, rx->rx_msg, rx->rx_msgaddr,
1356 (__u64)(page_to_phys(pg) + pg_off));
1358 pg_off += IBLND_MSG_SIZE;
1359 LASSERT(pg_off <= PAGE_SIZE);
1361 if (pg_off == PAGE_SIZE) {
1364 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1370 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1372 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1376 LASSERT (tpo->tpo_pool.po_allocated == 0);
1381 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1382 tx = &tpo->tpo_tx_descs[i];
1383 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1384 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1386 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1389 kiblnd_hdev_decref(hdev);
1390 tpo->tpo_hdev = NULL;
1393 static struct kib_hca_dev *
1394 kiblnd_current_hdev(struct kib_dev *dev)
1396 struct kib_hca_dev *hdev;
1397 unsigned long flags;
1400 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1401 while (dev->ibd_failover) {
1402 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1404 CDEBUG(D_NET, "%s: Wait for failover\n",
1406 set_current_state(TASK_INTERRUPTIBLE);
1407 schedule_timeout(cfs_time_seconds(1) / 100);
1409 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1412 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1413 hdev = dev->ibd_hdev;
1415 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1421 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1423 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1424 struct kib_pool *pool = &tpo->tpo_pool;
1425 struct kib_net *net = pool->po_owner->ps_net;
1426 struct kib_dev *dev;
1433 LASSERT (net != NULL);
1437 /* pre-mapped messages are not bigger than 1 page */
1438 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1440 /* No fancy arithmetic when we do the buffer calculations */
1441 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1443 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1445 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1446 page = txpgs->ibp_pages[ipage];
1447 tx = &tpo->tpo_tx_descs[i];
1449 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1452 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1456 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1458 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1460 list_add(&tx->tx_list, &pool->po_free_list);
1462 page_offset += IBLND_MSG_SIZE;
1463 LASSERT(page_offset <= PAGE_SIZE);
1465 if (page_offset == PAGE_SIZE) {
1468 LASSERT(ipage <= txpgs->ibp_npages);
1474 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1476 LASSERT(fpo->fpo_map_count == 0);
1478 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1479 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1481 struct kib_fast_reg_descriptor *frd, *tmp;
1484 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1486 list_del(&frd->frd_list);
1487 #ifndef HAVE_IB_MAP_MR_SG
1488 ib_free_fast_reg_page_list(frd->frd_frpl);
1490 ib_dereg_mr(frd->frd_mr);
1491 LIBCFS_FREE(frd, sizeof(*frd));
1494 if (i < fpo->fast_reg.fpo_pool_size)
1495 CERROR("FastReg pool still has %d regions registered\n",
1496 fpo->fast_reg.fpo_pool_size - i);
1500 kiblnd_hdev_decref(fpo->fpo_hdev);
1502 LIBCFS_FREE(fpo, sizeof(*fpo));
1506 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1508 struct kib_fmr_pool *fpo, *tmp;
1510 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1511 list_del(&fpo->fpo_list);
1512 kiblnd_destroy_fmr_pool(fpo);
1517 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1520 int size = tunables->lnd_fmr_pool_size / ncpts;
1522 return max(IBLND_FMR_POOL, size);
1526 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1529 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1531 return max(IBLND_FMR_POOL_FLUSH, size);
1534 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1535 struct kib_fmr_pool *fpo)
1537 struct ib_fmr_pool_param param = {
1538 .max_pages_per_fmr = LNET_MAX_IOV,
1539 .page_shift = PAGE_SHIFT,
1540 .access = (IB_ACCESS_LOCAL_WRITE |
1541 IB_ACCESS_REMOTE_WRITE),
1542 .pool_size = fps->fps_pool_size,
1543 .dirty_watermark = fps->fps_flush_trigger,
1544 .flush_function = NULL,
1546 .cache = !!fps->fps_cache };
1549 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1551 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1552 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1554 CERROR("Failed to create FMR pool: %d\n", rc);
1556 CERROR("FMRs are not supported\n");
1558 fpo->fpo_is_fmr = true;
1563 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1564 struct kib_fmr_pool *fpo,
1565 enum kib_dev_caps dev_caps)
1567 struct kib_fast_reg_descriptor *frd, *tmp;
1570 fpo->fpo_is_fmr = false;
1572 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1573 fpo->fast_reg.fpo_pool_size = 0;
1574 for (i = 0; i < fps->fps_pool_size; i++) {
1575 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1578 CERROR("Failed to allocate a new fast_reg descriptor\n");
1584 #ifndef HAVE_IB_MAP_MR_SG
1585 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1587 if (IS_ERR(frd->frd_frpl)) {
1588 rc = PTR_ERR(frd->frd_frpl);
1589 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1591 frd->frd_frpl = NULL;
1596 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1597 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1601 * it is expected to get here if this is an MLX-5 card.
1602 * MLX-4 cards will always use FMR and MLX-5 cards will
1603 * always use fast_reg. It turns out that some MLX-5 cards
1604 * (possibly due to older FW versions) do not natively support
1605 * gaps. So we will need to track them here.
1607 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1608 #ifdef IB_MR_TYPE_SG_GAPS
1609 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1610 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1611 IB_MR_TYPE_SG_GAPS :
1617 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1618 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1619 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1621 if (IS_ERR(frd->frd_mr)) {
1622 rc = PTR_ERR(frd->frd_mr);
1623 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1628 /* There appears to be a bug in MLX5 code where you must
1629 * invalidate the rkey of a new FastReg pool before first
1630 * using it. Thus, I am marking the FRD invalid here. */
1631 frd->frd_valid = false;
1633 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1634 fpo->fast_reg.fpo_pool_size++;
1641 ib_dereg_mr(frd->frd_mr);
1642 #ifndef HAVE_IB_MAP_MR_SG
1644 ib_free_fast_reg_page_list(frd->frd_frpl);
1646 LIBCFS_FREE(frd, sizeof(*frd));
1649 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1651 list_del(&frd->frd_list);
1652 #ifndef HAVE_IB_MAP_MR_SG
1653 ib_free_fast_reg_page_list(frd->frd_frpl);
1655 ib_dereg_mr(frd->frd_mr);
1656 LIBCFS_FREE(frd, sizeof(*frd));
1662 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1663 struct kib_fmr_pool **pp_fpo)
1665 struct kib_dev *dev = fps->fps_net->ibn_dev;
1666 struct kib_fmr_pool *fpo;
1669 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1673 memset(fpo, 0, sizeof(*fpo));
1675 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1677 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1678 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1680 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1684 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1685 fpo->fpo_owner = fps;
1691 kiblnd_hdev_decref(fpo->fpo_hdev);
1692 LIBCFS_FREE(fpo, sizeof(*fpo));
1697 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1699 if (fps->fps_net == NULL) /* intialized? */
1702 spin_lock(&fps->fps_lock);
1704 while (!list_empty(&fps->fps_pool_list)) {
1705 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1706 struct kib_fmr_pool,
1709 fpo->fpo_failed = 1;
1710 list_del(&fpo->fpo_list);
1711 if (fpo->fpo_map_count == 0)
1712 list_add(&fpo->fpo_list, zombies);
1714 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1717 spin_unlock(&fps->fps_lock);
1721 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1723 if (fps->fps_net != NULL) { /* initialized? */
1724 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1725 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1730 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1731 struct kib_net *net,
1732 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1734 struct kib_fmr_pool *fpo;
1737 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1742 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1743 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1744 fps->fps_cache = tunables->lnd_fmr_cache;
1746 spin_lock_init(&fps->fps_lock);
1747 INIT_LIST_HEAD(&fps->fps_pool_list);
1748 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1750 rc = kiblnd_create_fmr_pool(fps, &fpo);
1752 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1758 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1760 if (fpo->fpo_map_count != 0) /* still in use */
1762 if (fpo->fpo_failed)
1764 return now >= fpo->fpo_deadline;
1768 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1770 struct kib_hca_dev *hdev;
1771 __u64 *pages = tx->tx_pages;
1776 hdev = tx->tx_pool->tpo_hdev;
1778 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1779 for (size = 0; size < rd->rd_frags[i].rf_nob;
1780 size += hdev->ibh_page_size) {
1781 pages[npages++] = (rd->rd_frags[i].rf_addr &
1782 hdev->ibh_page_mask) + size;
1790 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1792 struct list_head zombies = LIST_HEAD_INIT(zombies);
1793 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1794 struct kib_fmr_poolset *fps;
1795 time64_t now = ktime_get_seconds();
1796 struct kib_fmr_pool *tmp;
1802 fps = fpo->fpo_owner;
1803 if (fpo->fpo_is_fmr) {
1804 if (fmr->fmr_pfmr) {
1805 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1806 fmr->fmr_pfmr = NULL;
1810 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1814 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1817 frd->frd_valid = false;
1818 spin_lock(&fps->fps_lock);
1819 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1820 spin_unlock(&fps->fps_lock);
1821 fmr->fmr_frd = NULL;
1824 fmr->fmr_pool = NULL;
1826 spin_lock(&fps->fps_lock);
1827 fpo->fpo_map_count--; /* decref the pool */
1829 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1830 /* the first pool is persistent */
1831 if (fps->fps_pool_list.next == &fpo->fpo_list)
1834 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1835 list_move(&fpo->fpo_list, &zombies);
1839 spin_unlock(&fps->fps_lock);
1841 if (!list_empty(&zombies))
1842 kiblnd_destroy_fmr_pool_list(&zombies);
1845 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1846 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1847 struct kib_fmr *fmr)
1849 struct kib_fmr_pool *fpo;
1850 __u64 *pages = tx->tx_pages;
1852 bool is_rx = (rd != tx->tx_rd);
1853 bool tx_pages_mapped = 0;
1858 spin_lock(&fps->fps_lock);
1859 version = fps->fps_version;
1860 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1861 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1862 fpo->fpo_map_count++;
1864 if (fpo->fpo_is_fmr) {
1865 struct ib_pool_fmr *pfmr;
1867 spin_unlock(&fps->fps_lock);
1869 if (!tx_pages_mapped) {
1870 npages = kiblnd_map_tx_pages(tx, rd);
1871 tx_pages_mapped = 1;
1874 pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1875 pages, npages, iov);
1876 if (likely(!IS_ERR(pfmr))) {
1877 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1879 fmr->fmr_frd = NULL;
1880 fmr->fmr_pfmr = pfmr;
1881 fmr->fmr_pool = fpo;
1886 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1887 struct kib_fast_reg_descriptor *frd;
1888 #ifdef HAVE_IB_MAP_MR_SG
1889 struct ib_reg_wr *wr;
1892 struct ib_rdma_wr *wr;
1893 struct ib_fast_reg_page_list *frpl;
1897 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1898 struct kib_fast_reg_descriptor,
1900 list_del(&frd->frd_list);
1901 spin_unlock(&fps->fps_lock);
1903 #ifndef HAVE_IB_MAP_MR_SG
1904 frpl = frd->frd_frpl;
1908 if (!frd->frd_valid) {
1909 struct ib_rdma_wr *inv_wr;
1910 __u32 key = is_rx ? mr->rkey : mr->lkey;
1912 inv_wr = &frd->frd_inv_wr;
1913 memset(inv_wr, 0, sizeof(*inv_wr));
1915 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1916 inv_wr->wr.wr_id = IBLND_WID_MR;
1917 inv_wr->wr.ex.invalidate_rkey = key;
1920 key = ib_inc_rkey(key);
1921 ib_update_fast_reg_key(mr, key);
1924 #ifdef HAVE_IB_MAP_MR_SG
1925 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1926 n = ib_map_mr_sg(mr, tx->tx_frags,
1927 tx->tx_nfrags, NULL, PAGE_SIZE);
1929 n = ib_map_mr_sg(mr, tx->tx_frags,
1930 tx->tx_nfrags, PAGE_SIZE);
1932 if (unlikely(n != tx->tx_nfrags)) {
1933 CERROR("Failed to map mr %d/%d "
1934 "elements\n", n, tx->tx_nfrags);
1935 return n < 0 ? n : -EINVAL;
1938 wr = &frd->frd_fastreg_wr;
1939 memset(wr, 0, sizeof(*wr));
1941 wr->wr.opcode = IB_WR_REG_MR;
1942 wr->wr.wr_id = IBLND_WID_MR;
1944 wr->wr.send_flags = 0;
1946 wr->key = is_rx ? mr->rkey : mr->lkey;
1947 wr->access = (IB_ACCESS_LOCAL_WRITE |
1948 IB_ACCESS_REMOTE_WRITE);
1950 if (!tx_pages_mapped) {
1951 npages = kiblnd_map_tx_pages(tx, rd);
1952 tx_pages_mapped = 1;
1955 LASSERT(npages <= frpl->max_page_list_len);
1956 memcpy(frpl->page_list, pages,
1957 sizeof(*pages) * npages);
1959 /* Prepare FastReg WR */
1960 wr = &frd->frd_fastreg_wr;
1961 memset(wr, 0, sizeof(*wr));
1963 wr->wr.opcode = IB_WR_FAST_REG_MR;
1964 wr->wr.wr_id = IBLND_WID_MR;
1966 wr->wr.wr.fast_reg.iova_start = iov;
1967 wr->wr.wr.fast_reg.page_list = frpl;
1968 wr->wr.wr.fast_reg.page_list_len = npages;
1969 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1970 wr->wr.wr.fast_reg.length = nob;
1971 wr->wr.wr.fast_reg.rkey =
1972 is_rx ? mr->rkey : mr->lkey;
1973 wr->wr.wr.fast_reg.access_flags =
1974 (IB_ACCESS_LOCAL_WRITE |
1975 IB_ACCESS_REMOTE_WRITE);
1978 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1980 fmr->fmr_pfmr = NULL;
1981 fmr->fmr_pool = fpo;
1984 spin_unlock(&fps->fps_lock);
1988 spin_lock(&fps->fps_lock);
1989 fpo->fpo_map_count--;
1990 if (rc != -EAGAIN) {
1991 spin_unlock(&fps->fps_lock);
1995 /* EAGAIN and ... */
1996 if (version != fps->fps_version) {
1997 spin_unlock(&fps->fps_lock);
2002 if (fps->fps_increasing) {
2003 spin_unlock(&fps->fps_lock);
2004 CDEBUG(D_NET, "Another thread is allocating new "
2005 "FMR pool, waiting for her to complete\n");
2011 if (ktime_get_seconds() < fps->fps_next_retry) {
2012 /* someone failed recently */
2013 spin_unlock(&fps->fps_lock);
2017 fps->fps_increasing = 1;
2018 spin_unlock(&fps->fps_lock);
2020 CDEBUG(D_NET, "Allocate new FMR pool\n");
2021 rc = kiblnd_create_fmr_pool(fps, &fpo);
2022 spin_lock(&fps->fps_lock);
2023 fps->fps_increasing = 0;
2026 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2028 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2030 spin_unlock(&fps->fps_lock);
2036 kiblnd_fini_pool(struct kib_pool *pool)
2038 LASSERT(list_empty(&pool->po_free_list));
2039 LASSERT(pool->po_allocated == 0);
2041 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2045 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2047 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2049 memset(pool, 0, sizeof(struct kib_pool));
2050 INIT_LIST_HEAD(&pool->po_free_list);
2051 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2052 pool->po_owner = ps;
2053 pool->po_size = size;
2057 kiblnd_destroy_pool_list(struct list_head *head)
2059 struct kib_pool *pool;
2061 while (!list_empty(head)) {
2062 pool = list_entry(head->next, struct kib_pool, po_list);
2063 list_del(&pool->po_list);
2065 LASSERT(pool->po_owner != NULL);
2066 pool->po_owner->ps_pool_destroy(pool);
2071 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2073 if (ps->ps_net == NULL) /* intialized? */
2076 spin_lock(&ps->ps_lock);
2077 while (!list_empty(&ps->ps_pool_list)) {
2078 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
2079 struct kib_pool, po_list);
2082 list_del(&po->po_list);
2083 if (po->po_allocated == 0)
2084 list_add(&po->po_list, zombies);
2086 list_add(&po->po_list, &ps->ps_failed_pool_list);
2088 spin_unlock(&ps->ps_lock);
2092 kiblnd_fini_poolset(struct kib_poolset *ps)
2094 if (ps->ps_net != NULL) { /* initialized? */
2095 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2096 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2101 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2102 struct kib_net *net, char *name, int size,
2103 kib_ps_pool_create_t po_create,
2104 kib_ps_pool_destroy_t po_destroy,
2105 kib_ps_node_init_t nd_init,
2106 kib_ps_node_fini_t nd_fini)
2108 struct kib_pool *pool;
2111 memset(ps, 0, sizeof(struct kib_poolset));
2115 ps->ps_pool_create = po_create;
2116 ps->ps_pool_destroy = po_destroy;
2117 ps->ps_node_init = nd_init;
2118 ps->ps_node_fini = nd_fini;
2119 ps->ps_pool_size = size;
2120 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2121 >= sizeof(ps->ps_name))
2123 spin_lock_init(&ps->ps_lock);
2124 INIT_LIST_HEAD(&ps->ps_pool_list);
2125 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2127 rc = ps->ps_pool_create(ps, size, &pool);
2129 list_add(&pool->po_list, &ps->ps_pool_list);
2131 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2137 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2139 if (pool->po_allocated != 0) /* still in use */
2141 if (pool->po_failed)
2143 return now >= pool->po_deadline;
2147 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2149 struct list_head zombies = LIST_HEAD_INIT(zombies);
2150 struct kib_poolset *ps = pool->po_owner;
2151 struct kib_pool *tmp;
2152 time64_t now = ktime_get_seconds();
2154 spin_lock(&ps->ps_lock);
2156 if (ps->ps_node_fini != NULL)
2157 ps->ps_node_fini(pool, node);
2159 LASSERT(pool->po_allocated > 0);
2160 list_add(node, &pool->po_free_list);
2161 pool->po_allocated--;
2163 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2164 /* the first pool is persistent */
2165 if (ps->ps_pool_list.next == &pool->po_list)
2168 if (kiblnd_pool_is_idle(pool, now))
2169 list_move(&pool->po_list, &zombies);
2171 spin_unlock(&ps->ps_lock);
2173 if (!list_empty(&zombies))
2174 kiblnd_destroy_pool_list(&zombies);
2178 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2180 struct list_head *node;
2181 struct kib_pool *pool;
2183 unsigned int interval = 1;
2184 ktime_t time_before;
2185 unsigned int trips = 0;
2188 spin_lock(&ps->ps_lock);
2189 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2190 if (list_empty(&pool->po_free_list))
2193 pool->po_allocated++;
2194 pool->po_deadline = ktime_get_seconds() +
2195 IBLND_POOL_DEADLINE;
2196 node = pool->po_free_list.next;
2199 if (ps->ps_node_init != NULL) {
2200 /* still hold the lock */
2201 ps->ps_node_init(pool, node);
2203 spin_unlock(&ps->ps_lock);
2207 /* no available tx pool and ... */
2208 if (ps->ps_increasing) {
2209 /* another thread is allocating a new pool */
2210 spin_unlock(&ps->ps_lock);
2212 CDEBUG(D_NET, "Another thread is allocating new "
2213 "%s pool, waiting %d HZs for her to complete."
2215 ps->ps_name, interval, trips);
2217 set_current_state(TASK_INTERRUPTIBLE);
2218 schedule_timeout(interval);
2219 if (interval < cfs_time_seconds(1))
2225 if (ktime_get_seconds() < ps->ps_next_retry) {
2226 /* someone failed recently */
2227 spin_unlock(&ps->ps_lock);
2231 ps->ps_increasing = 1;
2232 spin_unlock(&ps->ps_lock);
2234 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2235 time_before = ktime_get();
2236 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2237 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2238 ktime_ms_delta(ktime_get(), time_before));
2240 spin_lock(&ps->ps_lock);
2241 ps->ps_increasing = 0;
2243 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2245 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2246 CERROR("Can't allocate new %s pool because out of memory\n",
2249 spin_unlock(&ps->ps_lock);
2255 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2257 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2261 LASSERT (pool->po_allocated == 0);
2263 if (tpo->tpo_tx_pages != NULL) {
2264 kiblnd_unmap_tx_pool(tpo);
2265 kiblnd_free_pages(tpo->tpo_tx_pages);
2268 if (tpo->tpo_tx_descs == NULL)
2271 for (i = 0; i < pool->po_size; i++) {
2272 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2273 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2275 list_del(&tx->tx_list);
2276 if (tx->tx_pages != NULL)
2277 LIBCFS_FREE(tx->tx_pages,
2279 sizeof(*tx->tx_pages));
2280 if (tx->tx_frags != NULL)
2281 LIBCFS_FREE(tx->tx_frags,
2282 (1 + IBLND_MAX_RDMA_FRAGS) *
2283 sizeof(*tx->tx_frags));
2284 if (tx->tx_wrq != NULL)
2285 LIBCFS_FREE(tx->tx_wrq,
2286 (1 + IBLND_MAX_RDMA_FRAGS) *
2287 sizeof(*tx->tx_wrq));
2288 if (tx->tx_sge != NULL)
2289 LIBCFS_FREE(tx->tx_sge,
2290 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2291 sizeof(*tx->tx_sge));
2292 if (tx->tx_rd != NULL)
2293 LIBCFS_FREE(tx->tx_rd,
2294 offsetof(struct kib_rdma_desc,
2295 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2298 LIBCFS_FREE(tpo->tpo_tx_descs,
2299 pool->po_size * sizeof(struct kib_tx));
2301 kiblnd_fini_pool(pool);
2302 LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool));
2305 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2307 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2310 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2311 ntx = tunables->lnd_ntx / ncpts;
2313 return max(IBLND_TX_POOL, ntx);
2317 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2321 struct kib_pool *pool;
2322 struct kib_tx_pool *tpo;
2324 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2326 CERROR("Failed to allocate TX pool\n");
2330 pool = &tpo->tpo_pool;
2331 kiblnd_init_pool(ps, pool, size);
2332 tpo->tpo_tx_descs = NULL;
2333 tpo->tpo_tx_pages = NULL;
2335 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2336 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2337 CERROR("Can't allocate tx pages: %d\n", npg);
2338 LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool));
2342 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2343 size * sizeof(struct kib_tx));
2344 if (tpo->tpo_tx_descs == NULL) {
2345 CERROR("Can't allocate %d tx descriptors\n", size);
2346 ps->ps_pool_destroy(pool);
2350 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2352 for (i = 0; i < size; i++) {
2353 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2354 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2357 if (ps->ps_net->ibn_fmr_ps != NULL) {
2358 LIBCFS_CPT_ALLOC(tx->tx_pages,
2359 lnet_cpt_table(), ps->ps_cpt,
2360 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2361 if (tx->tx_pages == NULL)
2365 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2366 (1 + IBLND_MAX_RDMA_FRAGS) *
2367 sizeof(*tx->tx_frags));
2368 if (tx->tx_frags == NULL)
2371 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2373 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2374 (1 + IBLND_MAX_RDMA_FRAGS) *
2375 sizeof(*tx->tx_wrq));
2376 if (tx->tx_wrq == NULL)
2379 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2380 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2381 sizeof(*tx->tx_sge));
2382 if (tx->tx_sge == NULL)
2385 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2386 offsetof(struct kib_rdma_desc,
2387 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2388 if (tx->tx_rd == NULL)
2393 kiblnd_map_tx_pool(tpo);
2398 ps->ps_pool_destroy(pool);
2403 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2405 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2406 struct kib_tx_poolset,
2408 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2410 tx->tx_cookie = tps->tps_next_tx_cookie++;
2414 kiblnd_net_fini_pools(struct kib_net *net)
2418 cfs_cpt_for_each(i, lnet_cpt_table()) {
2419 struct kib_tx_poolset *tps;
2420 struct kib_fmr_poolset *fps;
2422 if (net->ibn_tx_ps != NULL) {
2423 tps = net->ibn_tx_ps[i];
2424 kiblnd_fini_poolset(&tps->tps_poolset);
2427 if (net->ibn_fmr_ps != NULL) {
2428 fps = net->ibn_fmr_ps[i];
2429 kiblnd_fini_fmr_poolset(fps);
2433 if (net->ibn_tx_ps != NULL) {
2434 cfs_percpt_free(net->ibn_tx_ps);
2435 net->ibn_tx_ps = NULL;
2438 if (net->ibn_fmr_ps != NULL) {
2439 cfs_percpt_free(net->ibn_fmr_ps);
2440 net->ibn_fmr_ps = NULL;
2445 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2448 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2449 #ifdef HAVE_IB_GET_DMA_MR
2450 unsigned long flags;
2456 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2458 #ifdef HAVE_IB_GET_DMA_MR
2459 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2461 * if lnd_map_on_demand is zero then we have effectively disabled
2462 * FMR or FastReg and we're using global memory regions
2465 if (!tunables->lnd_map_on_demand) {
2466 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2468 goto create_tx_pool;
2471 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2474 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2475 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2476 tunables->lnd_fmr_pool_size,
2477 tunables->lnd_ntx / 4);
2482 /* TX pool must be created later than FMR, see LU-2268
2484 LASSERT(net->ibn_tx_ps == NULL);
2486 /* premapping can fail if ibd_nmr > 1, so we always create
2487 * FMR pool and map-on-demand if premapping failed */
2489 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2490 sizeof(struct kib_fmr_poolset));
2491 if (net->ibn_fmr_ps == NULL) {
2492 CERROR("Failed to allocate FMR pool array\n");
2497 for (i = 0; i < ncpts; i++) {
2498 cpt = (cpts == NULL) ? i : cpts[i];
2499 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2502 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2509 LASSERT(i == ncpts);
2511 #ifdef HAVE_IB_GET_DMA_MR
2514 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2515 sizeof(struct kib_tx_poolset));
2516 if (net->ibn_tx_ps == NULL) {
2517 CERROR("Failed to allocate tx pool array\n");
2522 for (i = 0; i < ncpts; i++) {
2523 cpt = (cpts == NULL) ? i : cpts[i];
2524 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2526 kiblnd_tx_pool_size(ni, ncpts),
2527 kiblnd_create_tx_pool,
2528 kiblnd_destroy_tx_pool,
2529 kiblnd_tx_init, NULL);
2531 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2539 kiblnd_net_fini_pools(net);
2545 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2547 struct ib_device_attr *dev_attr;
2550 /* It's safe to assume a HCA can handle a page size
2551 * matching that of the native system */
2552 hdev->ibh_page_shift = PAGE_SHIFT;
2553 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2554 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2556 #ifndef HAVE_IB_DEVICE_ATTRS
2557 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2558 if (dev_attr == NULL) {
2559 CERROR("Out of memory\n");
2563 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2565 CERROR("Failed to query IB device: %d\n", rc);
2566 goto out_clean_attr;
2569 dev_attr = &hdev->ibh_ibdev->attrs;
2572 hdev->ibh_mr_size = dev_attr->max_mr_size;
2573 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2575 /* Setup device Memory Registration capabilities */
2576 #ifdef HAVE_IB_DEVICE_OPS
2577 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2578 hdev->ibh_ibdev->ops.dealloc_fmr &&
2579 hdev->ibh_ibdev->ops.map_phys_fmr &&
2580 hdev->ibh_ibdev->ops.unmap_fmr) {
2582 if (hdev->ibh_ibdev->alloc_fmr &&
2583 hdev->ibh_ibdev->dealloc_fmr &&
2584 hdev->ibh_ibdev->map_phys_fmr &&
2585 hdev->ibh_ibdev->unmap_fmr) {
2587 LCONSOLE_INFO("Using FMR for registration\n");
2588 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2589 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2590 LCONSOLE_INFO("Using FastReg for registration\n");
2591 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2592 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2593 #ifdef IB_DEVICE_SG_GAPS_REG
2594 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2595 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2605 #ifndef HAVE_IB_DEVICE_ATTRS
2607 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2611 CERROR("IB device does not support FMRs nor FastRegs, can't "
2612 "register memory: %d\n", rc);
2613 else if (rc == -EINVAL)
2614 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2618 #ifdef HAVE_IB_GET_DMA_MR
2620 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2622 if (hdev->ibh_mrs == NULL)
2625 ib_dereg_mr(hdev->ibh_mrs);
2627 hdev->ibh_mrs = NULL;
2632 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2634 #ifdef HAVE_IB_GET_DMA_MR
2635 kiblnd_hdev_cleanup_mrs(hdev);
2638 if (hdev->ibh_pd != NULL)
2639 ib_dealloc_pd(hdev->ibh_pd);
2641 if (hdev->ibh_cmid != NULL)
2642 rdma_destroy_id(hdev->ibh_cmid);
2644 LIBCFS_FREE(hdev, sizeof(*hdev));
2647 #ifdef HAVE_IB_GET_DMA_MR
2649 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2652 int acflags = IB_ACCESS_LOCAL_WRITE |
2653 IB_ACCESS_REMOTE_WRITE;
2655 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2657 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2658 kiblnd_hdev_cleanup_mrs(hdev);
2669 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2675 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2677 struct rdma_cm_id *cmid;
2678 struct sockaddr_in srcaddr;
2679 struct sockaddr_in dstaddr;
2682 if (dev->ibd_hdev == NULL || /* initializing */
2683 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2684 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2687 /* XXX: it's UGLY, but I don't have better way to find
2688 * ib-bonding HCA failover because:
2690 * a. no reliable CM event for HCA failover...
2691 * b. no OFED API to get ib_device for current net_device...
2693 * We have only two choices at this point:
2695 * a. rdma_bind_addr(), it will conflict with listener cmid
2696 * b. rdma_resolve_addr() to zero addr */
2697 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2698 RDMA_PS_TCP, IB_QPT_RC);
2701 CERROR("Failed to create cmid for failover: %d\n", rc);
2705 memset(&srcaddr, 0, sizeof(srcaddr));
2706 srcaddr.sin_family = AF_INET;
2707 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2709 memset(&dstaddr, 0, sizeof(dstaddr));
2710 dstaddr.sin_family = AF_INET;
2711 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2712 (struct sockaddr *)&dstaddr, 1);
2713 if (rc != 0 || cmid->device == NULL) {
2714 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2715 dev->ibd_ifname, &dev->ibd_ifip,
2717 rdma_destroy_id(cmid);
2721 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2722 rdma_destroy_id(cmid);
2727 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2729 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2730 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2731 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2732 struct rdma_cm_id *cmid = NULL;
2733 struct kib_hca_dev *hdev = NULL;
2734 struct kib_hca_dev *old;
2736 struct kib_net *net;
2737 struct sockaddr_in addr;
2738 unsigned long flags;
2742 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2743 dev->ibd_can_failover ||
2744 dev->ibd_hdev == NULL);
2746 rc = kiblnd_dev_need_failover(dev, ns);
2750 if (dev->ibd_hdev != NULL &&
2751 dev->ibd_hdev->ibh_cmid != NULL) {
2752 /* XXX it's not good to close old listener at here,
2753 * because we can fail to create new listener.
2754 * But we have to close it now, otherwise rdma_bind_addr
2755 * will return EADDRINUSE... How crap! */
2756 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2758 cmid = dev->ibd_hdev->ibh_cmid;
2759 /* make next schedule of kiblnd_dev_need_failover()
2760 * return 1 for me */
2761 dev->ibd_hdev->ibh_cmid = NULL;
2762 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2764 rdma_destroy_id(cmid);
2767 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2771 CERROR("Failed to create cmid for failover: %d\n", rc);
2775 memset(&addr, 0, sizeof(addr));
2776 addr.sin_family = AF_INET;
2777 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2778 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2780 /* Bind to failover device or port */
2781 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2782 if (rc != 0 || cmid->device == NULL) {
2783 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2784 dev->ibd_ifname, &dev->ibd_ifip,
2786 rdma_destroy_id(cmid);
2790 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2792 CERROR("Failed to allocate kib_hca_dev\n");
2793 rdma_destroy_id(cmid);
2798 atomic_set(&hdev->ibh_ref, 1);
2799 hdev->ibh_dev = dev;
2800 hdev->ibh_cmid = cmid;
2801 hdev->ibh_ibdev = cmid->device;
2803 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2804 pd = ib_alloc_pd(cmid->device, 0);
2806 pd = ib_alloc_pd(cmid->device);
2810 CERROR("Can't allocate PD: %d\n", rc);
2816 rc = rdma_listen(cmid, 0);
2818 CERROR("Can't start new listener: %d\n", rc);
2822 rc = kiblnd_hdev_get_attr(hdev);
2824 CERROR("Can't get device attributes: %d\n", rc);
2828 #ifdef HAVE_IB_GET_DMA_MR
2829 rc = kiblnd_hdev_setup_mrs(hdev);
2831 CERROR("Can't setup device: %d\n", rc);
2836 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2838 old = dev->ibd_hdev;
2839 dev->ibd_hdev = hdev; /* take over the refcount */
2842 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2843 cfs_cpt_for_each(i, lnet_cpt_table()) {
2844 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2847 if (net->ibn_fmr_ps != NULL)
2848 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2853 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2855 if (!list_empty(&zombie_tpo))
2856 kiblnd_destroy_pool_list(&zombie_tpo);
2857 if (!list_empty(&zombie_ppo))
2858 kiblnd_destroy_pool_list(&zombie_ppo);
2859 if (!list_empty(&zombie_fpo))
2860 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2862 kiblnd_hdev_decref(hdev);
2865 dev->ibd_failed_failover++;
2867 dev->ibd_failed_failover = 0;
2873 kiblnd_destroy_dev(struct kib_dev *dev)
2875 LASSERT(dev->ibd_nnets == 0);
2876 LASSERT(list_empty(&dev->ibd_nets));
2878 list_del(&dev->ibd_fail_list);
2879 list_del(&dev->ibd_list);
2881 if (dev->ibd_hdev != NULL)
2882 kiblnd_hdev_decref(dev->ibd_hdev);
2884 LIBCFS_FREE(dev, sizeof(*dev));
2888 kiblnd_base_shutdown(void)
2890 struct kib_sched_info *sched;
2893 LASSERT(list_empty(&kiblnd_data.kib_devs));
2895 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2896 atomic_read(&libcfs_kmemory));
2898 switch (kiblnd_data.kib_init) {
2902 case IBLND_INIT_ALL:
2903 case IBLND_INIT_DATA:
2904 LASSERT (kiblnd_data.kib_peers != NULL);
2905 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2906 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2908 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2909 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2910 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2911 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2913 /* flag threads to terminate; wake and wait for them to die */
2914 kiblnd_data.kib_shutdown = 1;
2916 /* NB: we really want to stop scheduler threads net by net
2917 * instead of the whole module, this should be improved
2918 * with dynamic configuration LNet */
2919 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2920 wake_up_all(&sched->ibs_waitq);
2922 wake_up_all(&kiblnd_data.kib_connd_waitq);
2923 wake_up_all(&kiblnd_data.kib_failover_waitq);
2926 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2929 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2930 "Waiting for %d threads to terminate\n",
2931 atomic_read(&kiblnd_data.kib_nthreads));
2932 set_current_state(TASK_UNINTERRUPTIBLE);
2933 schedule_timeout(cfs_time_seconds(1));
2938 case IBLND_INIT_NOTHING:
2942 if (kiblnd_data.kib_peers != NULL) {
2943 LIBCFS_FREE(kiblnd_data.kib_peers,
2944 sizeof(struct list_head) *
2945 kiblnd_data.kib_peer_hash_size);
2948 if (kiblnd_data.kib_scheds != NULL)
2949 cfs_percpt_free(kiblnd_data.kib_scheds);
2951 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2952 atomic_read(&libcfs_kmemory));
2954 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2955 module_put(THIS_MODULE);
2959 kiblnd_shutdown(struct lnet_ni *ni)
2961 struct kib_net *net = ni->ni_data;
2962 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2964 unsigned long flags;
2966 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2971 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2972 atomic_read(&libcfs_kmemory));
2974 write_lock_irqsave(g_lock, flags);
2975 net->ibn_shutdown = 1;
2976 write_unlock_irqrestore(g_lock, flags);
2978 switch (net->ibn_init) {
2982 case IBLND_INIT_ALL:
2983 /* nuke all existing peers within this net */
2984 kiblnd_del_peer(ni, LNET_NID_ANY);
2986 /* Wait for all peer_ni state to clean up */
2988 while (atomic_read(&net->ibn_npeers) != 0) {
2991 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2992 "%s: waiting for %d peers to disconnect\n",
2993 libcfs_nid2str(ni->ni_nid),
2994 atomic_read(&net->ibn_npeers));
2995 set_current_state(TASK_UNINTERRUPTIBLE);
2996 schedule_timeout(cfs_time_seconds(1));
2999 kiblnd_net_fini_pools(net);
3001 write_lock_irqsave(g_lock, flags);
3002 LASSERT(net->ibn_dev->ibd_nnets > 0);
3003 net->ibn_dev->ibd_nnets--;
3004 list_del(&net->ibn_list);
3005 write_unlock_irqrestore(g_lock, flags);
3009 case IBLND_INIT_NOTHING:
3010 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3012 if (net->ibn_dev != NULL &&
3013 net->ibn_dev->ibd_nnets == 0)
3014 kiblnd_destroy_dev(net->ibn_dev);
3019 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3020 atomic_read(&libcfs_kmemory));
3022 net->ibn_init = IBLND_INIT_NOTHING;
3025 LIBCFS_FREE(net, sizeof(*net));
3028 if (list_empty(&kiblnd_data.kib_devs))
3029 kiblnd_base_shutdown();
3034 kiblnd_base_startup(struct net *ns)
3036 struct kib_sched_info *sched;
3040 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3042 try_module_get(THIS_MODULE);
3043 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3045 rwlock_init(&kiblnd_data.kib_global_lock);
3047 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3048 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3050 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3051 LIBCFS_ALLOC(kiblnd_data.kib_peers,
3052 sizeof(struct list_head) *
3053 kiblnd_data.kib_peer_hash_size);
3054 if (kiblnd_data.kib_peers == NULL)
3057 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3058 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3060 spin_lock_init(&kiblnd_data.kib_connd_lock);
3061 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3062 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3063 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3064 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3066 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3067 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3069 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3071 if (kiblnd_data.kib_scheds == NULL)
3074 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3077 spin_lock_init(&sched->ibs_lock);
3078 INIT_LIST_HEAD(&sched->ibs_conns);
3079 init_waitqueue_head(&sched->ibs_waitq);
3081 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3082 if (*kiblnd_tunables.kib_nscheds > 0) {
3083 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3085 /* max to half of CPUs, another half is reserved for
3086 * upper layer modules */
3087 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3090 sched->ibs_nthreads_max = nthrs;
3094 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3096 /* lists/ptrs/locks initialised */
3097 kiblnd_data.kib_init = IBLND_INIT_DATA;
3098 /*****************************************************/
3100 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3102 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3106 if (*kiblnd_tunables.kib_dev_failover != 0)
3107 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3111 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3115 /* flag everything initialised */
3116 kiblnd_data.kib_init = IBLND_INIT_ALL;
3117 /*****************************************************/
3122 kiblnd_base_shutdown();
3127 kiblnd_start_schedulers(struct kib_sched_info *sched)
3133 if (sched->ibs_nthreads == 0) {
3134 if (*kiblnd_tunables.kib_nscheds > 0) {
3135 nthrs = sched->ibs_nthreads_max;
3137 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3139 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3140 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3143 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3144 /* increase one thread if there is new interface */
3145 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3148 for (i = 0; i < nthrs; i++) {
3151 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3152 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3153 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3154 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3158 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3159 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3163 sched->ibs_nthreads += i;
3167 static int kiblnd_dev_start_threads(struct kib_dev *dev, u32 *cpts, int ncpts)
3173 for (i = 0; i < ncpts; i++) {
3174 struct kib_sched_info *sched;
3176 cpt = (cpts == NULL) ? i : cpts[i];
3177 sched = kiblnd_data.kib_scheds[cpt];
3179 if (sched->ibs_nthreads > 0)
3182 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3184 CERROR("Failed to start scheduler threads for %s\n",
3193 kiblnd_startup(struct lnet_ni *ni)
3196 struct lnet_inetdev *ifaces = NULL;
3197 struct kib_dev *ibdev = NULL;
3198 struct kib_net *net;
3199 unsigned long flags;
3203 LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3205 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3206 rc = kiblnd_base_startup(ni->ni_net_ns);
3211 LIBCFS_ALLOC(net, sizeof(*net));
3216 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3218 kiblnd_tunables_setup(ni);
3221 * ni_interfaces is only to support legacy pre Multi-Rail
3222 * tcp bonding for ksocklnd. Multi-Rail wants each secondary
3223 * IP to be treated as an unique 'struct ni' interfaces instead.
3225 if (ni->ni_interfaces[0] != NULL) {
3226 /* Use the IPoIB interface specified in 'networks=' */
3227 if (ni->ni_interfaces[1] != NULL) {
3228 CERROR("ko2iblnd: Multiple interfaces not supported\n");
3232 ifname = ni->ni_interfaces[0];
3234 ifname = *kiblnd_tunables.kib_default_ipif;
3237 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3238 CERROR("IPoIB interface name too long: %s\n", ifname);
3242 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3246 for (i = 0; i < rc; i++) {
3247 if (strcmp(ifname, ifaces[i].li_name) == 0)
3252 CERROR("ko2iblnd: No matching interfaces\n");
3257 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3263 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3264 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3265 sizeof(ibdev->ibd_ifname));
3266 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3268 INIT_LIST_HEAD(&ibdev->ibd_nets);
3269 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3270 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3272 /* initialize the device */
3273 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3275 CERROR("ko2iblnd: Can't initialize device: rc = %d\n", rc);
3279 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3281 net->ibn_dev = ibdev;
3282 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3284 ni->ni_dev_cpt = ifaces[i].li_cpt;
3286 rc = kiblnd_dev_start_threads(ibdev, ni->ni_cpts, ni->ni_ncpts);
3290 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3292 CERROR("Failed to initialize NI pools: %d\n", rc);
3296 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3298 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3299 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3301 net->ibn_init = IBLND_INIT_ALL;
3306 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3307 kiblnd_destroy_dev(ibdev);
3310 kiblnd_shutdown(ni);
3312 CDEBUG(D_NET, "kiblnd_startup failed\n");
3316 static struct lnet_lnd the_o2iblnd = {
3317 .lnd_type = O2IBLND,
3318 .lnd_startup = kiblnd_startup,
3319 .lnd_shutdown = kiblnd_shutdown,
3320 .lnd_ctl = kiblnd_ctl,
3321 .lnd_query = kiblnd_query,
3322 .lnd_send = kiblnd_send,
3323 .lnd_recv = kiblnd_recv,
3326 static void __exit ko2iblnd_exit(void)
3328 lnet_unregister_lnd(&the_o2iblnd);
3331 static int __init ko2iblnd_init(void)
3335 CLASSERT(sizeof(struct kib_msg) <= IBLND_MSG_SIZE);
3336 CLASSERT(offsetof(struct kib_msg,
3337 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3339 CLASSERT(offsetof(struct kib_msg,
3340 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3343 rc = kiblnd_tunables_init();
3347 lnet_register_lnd(&the_o2iblnd);
3352 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3353 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3354 MODULE_VERSION("2.8.0");
3355 MODULE_LICENSE("GPL");
3357 module_init(ko2iblnd_init);
3358 module_exit(ko2iblnd_exit);