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 const 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 BUILD_BUG_ON(sizeof(msg->ibm_type) != 1);
260 BUILD_BUG_ON(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 if (atomic_dec_and_test(&net->ibn_npeers))
376 wake_up_var(&net->ibn_npeers);
380 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
382 /* the caller is responsible for accounting the additional reference
383 * that this creates */
384 struct list_head *peer_list = kiblnd_nid2peerlist(nid);
385 struct list_head *tmp;
386 struct kib_peer_ni *peer_ni;
388 list_for_each(tmp, peer_list) {
390 peer_ni = list_entry(tmp, struct kib_peer_ni, ibp_list);
391 LASSERT(!kiblnd_peer_idle(peer_ni));
394 * Match a peer if its NID and the NID of the local NI it
395 * communicates over are the same. Otherwise don't match
396 * the peer, which will result in a new lnd peer being
399 if (peer_ni->ibp_nid != nid ||
400 peer_ni->ibp_ni->ni_nid != ni->ni_nid)
403 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
404 peer_ni, libcfs_nid2str(nid),
405 atomic_read(&peer_ni->ibp_refcount),
406 peer_ni->ibp_version);
413 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
415 LASSERT(list_empty(&peer_ni->ibp_conns));
417 LASSERT (kiblnd_peer_active(peer_ni));
418 list_del_init(&peer_ni->ibp_list);
419 /* lose peerlist's ref */
420 kiblnd_peer_decref(peer_ni);
424 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
425 lnet_nid_t *nidp, int *count)
427 struct kib_peer_ni *peer_ni;
428 struct list_head *ptmp;
432 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
434 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
436 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
438 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
439 LASSERT(!kiblnd_peer_idle(peer_ni));
441 if (peer_ni->ibp_ni != ni)
447 *nidp = peer_ni->ibp_nid;
448 *count = atomic_read(&peer_ni->ibp_refcount);
450 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
456 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
461 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
463 struct kib_conn *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_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
471 kiblnd_close_conn_locked(conn, 0);
472 /* NB closing peer_ni's last conn unlinked it. */
474 /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
479 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
482 struct list_head *ptmp;
483 struct list_head *pnxt;
484 struct kib_peer_ni *peer_ni;
491 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
493 if (nid != LNET_NID_ANY) {
494 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
497 hi = kiblnd_data.kib_peer_hash_size - 1;
500 for (i = lo; i <= hi; i++) {
501 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
502 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
503 LASSERT(!kiblnd_peer_idle(peer_ni));
505 if (peer_ni->ibp_ni != ni)
508 if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
511 if (!list_empty(&peer_ni->ibp_tx_queue)) {
512 LASSERT(list_empty(&peer_ni->ibp_conns));
514 list_splice_init(&peer_ni->ibp_tx_queue,
518 kiblnd_del_peer_locked(peer_ni);
519 rc = 0; /* matched something */
523 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
525 kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
530 static struct kib_conn *
531 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
533 struct kib_peer_ni *peer_ni;
534 struct list_head *ptmp;
535 struct kib_conn *conn;
536 struct list_head *ctmp;
540 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
542 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
543 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
545 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
546 LASSERT(!kiblnd_peer_idle(peer_ni));
548 if (peer_ni->ibp_ni != ni)
551 list_for_each(ctmp, &peer_ni->ibp_conns) {
555 conn = list_entry(ctmp, struct kib_conn, ibc_list);
556 kiblnd_conn_addref(conn);
557 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
564 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
569 kiblnd_debug_rx(struct kib_rx *rx)
571 CDEBUG(D_CONSOLE, " %p msg_type %x cred %d\n",
572 rx, rx->rx_msg->ibm_type,
573 rx->rx_msg->ibm_credits);
577 kiblnd_debug_tx(struct kib_tx *tx)
579 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lld "
580 "cookie %#llx msg %s%s type %x cred %d\n",
581 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
582 tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
583 tx->tx_lntmsg[0] == NULL ? "-" : "!",
584 tx->tx_lntmsg[1] == NULL ? "-" : "!",
585 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
589 kiblnd_debug_conn(struct kib_conn *conn)
591 struct list_head *tmp;
594 spin_lock(&conn->ibc_lock);
596 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
597 atomic_read(&conn->ibc_refcount), conn,
598 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
599 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d "
600 " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
601 conn->ibc_nsends_posted, conn->ibc_credits,
602 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
603 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
605 CDEBUG(D_CONSOLE, " early_rxs:\n");
606 list_for_each(tmp, &conn->ibc_early_rxs)
607 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
609 CDEBUG(D_CONSOLE, " tx_noops:\n");
610 list_for_each(tmp, &conn->ibc_tx_noops)
611 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
613 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
614 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
615 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
617 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
618 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
619 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
621 CDEBUG(D_CONSOLE, " tx_queue:\n");
622 list_for_each(tmp, &conn->ibc_tx_queue)
623 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
625 CDEBUG(D_CONSOLE, " active_txs:\n");
626 list_for_each(tmp, &conn->ibc_active_txs)
627 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
629 CDEBUG(D_CONSOLE, " rxs:\n");
630 for (i = 0; i < IBLND_RX_MSGS(conn); i++)
631 kiblnd_debug_rx(&conn->ibc_rxs[i]);
633 spin_unlock(&conn->ibc_lock);
637 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
639 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
640 if (cmid->route.path_rec == NULL)
643 if (*kiblnd_tunables.kib_ib_mtu)
644 cmid->route.path_rec->mtu =
645 ib_mtu_int_to_enum(*kiblnd_tunables.kib_ib_mtu);
649 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
657 vectors = conn->ibc_cmid->device->num_comp_vectors;
661 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
663 /* hash NID to CPU id in this partition... */
664 ibp_nid = conn->ibc_peer->ibp_nid;
665 off = do_div(ibp_nid, cpumask_weight(*mask));
666 for_each_cpu(i, *mask) {
676 * Get the scheduler bound to this CPT. If the scheduler has no
677 * threads, which means that the CPT has no CPUs, then grab the
678 * next scheduler that we can use.
680 * This case would be triggered if a NUMA node is configured with
681 * no associated CPUs.
683 static struct kib_sched_info *
684 kiblnd_get_scheduler(int cpt)
686 struct kib_sched_info *sched;
689 sched = kiblnd_data.kib_scheds[cpt];
691 if (sched->ibs_nthreads > 0)
694 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
695 if (sched->ibs_nthreads > 0) {
696 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
697 cpt, sched->ibs_cpt);
705 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
708 * One WR for the LNet message
709 * And ibc_max_frags for the transfer WRs
712 int multiplier = 1 + conn->ibc_max_frags;
713 enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
715 /* FastReg needs two extra WRs for map and invalidate */
716 if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
719 /* account for a maximum of ibc_queue_depth in-flight transfers */
720 ret = multiplier * conn->ibc_queue_depth;
722 if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
723 CDEBUG(D_NET, "peer_credits %u will result in send work "
724 "request size %d larger than maximum %d device "
725 "can handle\n", conn->ibc_queue_depth, ret,
726 conn->ibc_hdev->ibh_max_qp_wr);
727 conn->ibc_queue_depth =
728 conn->ibc_hdev->ibh_max_qp_wr / multiplier;
731 /* don't go beyond the maximum the device can handle */
732 return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
736 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
737 int state, int version)
740 * If the new conn is created successfully it takes over the caller's
741 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
742 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
743 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
744 * to destroy 'cmid' here since I'm called from the CM which still has
745 * its ref on 'cmid'). */
746 rwlock_t *glock = &kiblnd_data.kib_global_lock;
747 struct kib_net *net = peer_ni->ibp_ni->ni_data;
749 struct ib_qp_init_attr init_qp_attr = {};
750 struct kib_sched_info *sched;
751 #ifdef HAVE_IB_CQ_INIT_ATTR
752 struct ib_cq_init_attr cq_attr = {};
754 struct kib_conn *conn;
761 LASSERT(net != NULL);
762 LASSERT(!in_interrupt());
766 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
767 sched = kiblnd_get_scheduler(cpt);
770 CERROR("no schedulers available. node is unhealthy\n");
775 * The cpt might have changed if we ended up selecting a non cpt
776 * native scheduler. So use the scheduler's cpt instead.
778 cpt = sched->ibs_cpt;
780 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
782 CERROR("Can't allocate connection for %s\n",
783 libcfs_nid2str(peer_ni->ibp_nid));
787 conn->ibc_state = IBLND_CONN_INIT;
788 conn->ibc_version = version;
789 conn->ibc_peer = peer_ni; /* I take the caller's ref */
790 cmid->context = conn; /* for future CM callbacks */
791 conn->ibc_cmid = cmid;
792 conn->ibc_max_frags = peer_ni->ibp_max_frags;
793 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
794 conn->ibc_rxs = NULL;
795 conn->ibc_rx_pages = NULL;
797 INIT_LIST_HEAD(&conn->ibc_early_rxs);
798 INIT_LIST_HEAD(&conn->ibc_tx_noops);
799 INIT_LIST_HEAD(&conn->ibc_tx_queue);
800 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
801 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
802 INIT_LIST_HEAD(&conn->ibc_active_txs);
803 INIT_LIST_HEAD(&conn->ibc_zombie_txs);
804 spin_lock_init(&conn->ibc_lock);
806 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
807 sizeof(*conn->ibc_connvars));
808 if (conn->ibc_connvars == NULL) {
809 CERROR("Can't allocate in-progress connection state\n");
813 write_lock_irqsave(glock, flags);
814 if (dev->ibd_failover) {
815 write_unlock_irqrestore(glock, flags);
816 CERROR("%s: failover in progress\n", dev->ibd_ifname);
820 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
821 /* wakeup failover thread and teardown connection */
822 if (kiblnd_dev_can_failover(dev)) {
823 list_add_tail(&dev->ibd_fail_list,
824 &kiblnd_data.kib_failed_devs);
825 wake_up(&kiblnd_data.kib_failover_waitq);
828 write_unlock_irqrestore(glock, flags);
829 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
830 cmid->device->name, dev->ibd_ifname);
834 kiblnd_hdev_addref_locked(dev->ibd_hdev);
835 conn->ibc_hdev = dev->ibd_hdev;
837 kiblnd_setup_mtu_locked(cmid);
839 write_unlock_irqrestore(glock, flags);
841 #ifdef HAVE_IB_CQ_INIT_ATTR
842 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
843 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
844 cq = ib_create_cq(cmid->device,
845 kiblnd_cq_completion, kiblnd_cq_event, conn,
848 cq = ib_create_cq(cmid->device,
849 kiblnd_cq_completion, kiblnd_cq_event, conn,
850 IBLND_CQ_ENTRIES(conn),
851 kiblnd_get_completion_vector(conn, cpt));
855 * on MLX-5 (possibly MLX-4 as well) this error could be
856 * hit if the concurrent_sends and/or peer_tx_credits is set
857 * too high. Or due to an MLX-5 bug which tries to
858 * allocate 256kb via kmalloc for WR cookie array
860 CERROR("Failed to create CQ with %d CQEs: %ld\n",
861 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
867 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
869 CERROR("Can't request completion notification: %d\n", rc);
873 init_qp_attr.event_handler = kiblnd_qp_event;
874 init_qp_attr.qp_context = conn;
875 init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
876 init_qp_attr.cap.max_recv_sge = 1;
877 init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
878 init_qp_attr.qp_type = IB_QPT_RC;
879 init_qp_attr.send_cq = cq;
880 init_qp_attr.recv_cq = cq;
882 * kiblnd_send_wrs() can change the connection's queue depth if
883 * the maximum work requests for the device is maxed out
885 init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
886 init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
888 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, &init_qp_attr);
890 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
891 "send_sge: %d, recv_sge: %d\n",
892 rc, init_qp_attr.cap.max_send_wr,
893 init_qp_attr.cap.max_recv_wr,
894 init_qp_attr.cap.max_send_sge,
895 init_qp_attr.cap.max_recv_sge);
899 conn->ibc_sched = sched;
901 if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
902 CWARN("peer %s - queue depth reduced from %u to %u"
903 " to allow for qp creation\n",
904 libcfs_nid2str(peer_ni->ibp_nid),
905 peer_ni->ibp_queue_depth,
906 conn->ibc_queue_depth);
908 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
909 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
910 if (conn->ibc_rxs == NULL) {
911 CERROR("Cannot allocate RX buffers\n");
915 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
916 IBLND_RX_MSG_PAGES(conn));
920 kiblnd_map_rx_descs(conn);
922 /* 1 ref for caller and each rxmsg */
923 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
924 conn->ibc_nrx = IBLND_RX_MSGS(conn);
927 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
928 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
930 CERROR("Can't post rxmsg: %d\n", rc);
932 /* Make posted receives complete */
933 kiblnd_abort_receives(conn);
935 /* correct # of posted buffers
936 * NB locking needed now I'm racing with completion */
937 spin_lock_irqsave(&sched->ibs_lock, flags);
938 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
939 spin_unlock_irqrestore(&sched->ibs_lock, flags);
941 /* cmid will be destroyed by CM(ofed) after cm_callback
942 * returned, so we can't refer it anymore
943 * (by kiblnd_connd()->kiblnd_destroy_conn) */
944 rdma_destroy_qp(conn->ibc_cmid);
945 conn->ibc_cmid = NULL;
947 /* Drop my own and unused rxbuffer refcounts */
948 while (i++ <= IBLND_RX_MSGS(conn))
949 kiblnd_conn_decref(conn);
955 /* Init successful! */
956 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
957 state == IBLND_CONN_PASSIVE_WAIT);
958 conn->ibc_state = state;
961 atomic_inc(&net->ibn_nconns);
965 kiblnd_destroy_conn(conn);
966 LIBCFS_FREE(conn, sizeof(*conn));
972 kiblnd_destroy_conn(struct kib_conn *conn)
974 struct rdma_cm_id *cmid = conn->ibc_cmid;
975 struct kib_peer_ni *peer_ni = conn->ibc_peer;
977 LASSERT (!in_interrupt());
978 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
979 LASSERT(list_empty(&conn->ibc_early_rxs));
980 LASSERT(list_empty(&conn->ibc_tx_noops));
981 LASSERT(list_empty(&conn->ibc_tx_queue));
982 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
983 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
984 LASSERT(list_empty(&conn->ibc_active_txs));
985 LASSERT (conn->ibc_noops_posted == 0);
986 LASSERT (conn->ibc_nsends_posted == 0);
988 switch (conn->ibc_state) {
990 /* conn must be completely disengaged from the network */
993 case IBLND_CONN_DISCONNECTED:
994 /* connvars should have been freed already */
995 LASSERT (conn->ibc_connvars == NULL);
998 case IBLND_CONN_INIT:
1002 /* conn->ibc_cmid might be destroyed by CM already */
1003 if (cmid != NULL && cmid->qp != NULL)
1004 rdma_destroy_qp(cmid);
1007 ib_destroy_cq(conn->ibc_cq);
1009 kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1010 LNET_MSG_STATUS_OK);
1012 if (conn->ibc_rx_pages != NULL)
1013 kiblnd_unmap_rx_descs(conn);
1015 if (conn->ibc_rxs != NULL)
1016 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1018 if (conn->ibc_connvars != NULL)
1019 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1021 if (conn->ibc_hdev != NULL)
1022 kiblnd_hdev_decref(conn->ibc_hdev);
1024 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1025 if (conn->ibc_state != IBLND_CONN_INIT) {
1026 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1028 kiblnd_peer_decref(peer_ni);
1029 rdma_destroy_id(cmid);
1030 atomic_dec(&net->ibn_nconns);
1035 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1037 struct kib_conn *conn;
1038 struct kib_conn *cnxt;
1041 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1043 CDEBUG(D_NET, "Closing conn -> %s, "
1044 "version: %x, reason: %d\n",
1045 libcfs_nid2str(peer_ni->ibp_nid),
1046 conn->ibc_version, why);
1048 kiblnd_close_conn_locked(conn, why);
1056 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1057 int version, __u64 incarnation)
1059 struct kib_conn *conn;
1060 struct kib_conn *cnxt;
1063 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1065 if (conn->ibc_version == version &&
1066 conn->ibc_incarnation == incarnation)
1069 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1070 "incarnation:%#llx(%x, %#llx)\n",
1071 libcfs_nid2str(peer_ni->ibp_nid),
1072 conn->ibc_version, conn->ibc_incarnation,
1073 version, incarnation);
1075 kiblnd_close_conn_locked(conn, -ESTALE);
1083 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1085 struct kib_peer_ni *peer_ni;
1086 struct list_head *ptmp;
1087 struct list_head *pnxt;
1091 unsigned long flags;
1094 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1096 if (nid != LNET_NID_ANY)
1097 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1100 hi = kiblnd_data.kib_peer_hash_size - 1;
1103 for (i = lo; i <= hi; i++) {
1104 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1106 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
1107 LASSERT(!kiblnd_peer_idle(peer_ni));
1109 if (peer_ni->ibp_ni != ni)
1112 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1115 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1119 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1121 /* wildcards always succeed */
1122 if (nid == LNET_NID_ANY)
1125 return (count == 0) ? -ENOENT : 0;
1129 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1131 struct libcfs_ioctl_data *data = arg;
1135 case IOC_LIBCFS_GET_PEER: {
1139 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1141 data->ioc_nid = nid;
1142 data->ioc_count = count;
1146 case IOC_LIBCFS_DEL_PEER: {
1147 rc = kiblnd_del_peer(ni, data->ioc_nid);
1150 case IOC_LIBCFS_GET_CONN: {
1151 struct kib_conn *conn;
1154 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1160 LASSERT(conn->ibc_cmid != NULL);
1161 data->ioc_nid = conn->ibc_peer->ibp_nid;
1162 if (conn->ibc_cmid->route.path_rec == NULL)
1163 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1166 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1167 kiblnd_conn_decref(conn);
1170 case IOC_LIBCFS_CLOSE_CONNECTION: {
1171 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1183 kiblnd_free_pages(struct kib_pages *p)
1185 int npages = p->ibp_npages;
1188 for (i = 0; i < npages; i++) {
1189 if (p->ibp_pages[i] != NULL)
1190 __free_page(p->ibp_pages[i]);
1193 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1197 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1199 struct kib_pages *p;
1202 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1203 offsetof(struct kib_pages, ibp_pages[npages]));
1205 CERROR("Can't allocate descriptor for %d pages\n", npages);
1209 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1210 p->ibp_npages = npages;
1212 for (i = 0; i < npages; i++) {
1213 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1215 if (p->ibp_pages[i] == NULL) {
1216 CERROR("Can't allocate page %d of %d\n", i, npages);
1217 kiblnd_free_pages(p);
1227 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1232 LASSERT (conn->ibc_rxs != NULL);
1233 LASSERT (conn->ibc_hdev != NULL);
1235 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1236 rx = &conn->ibc_rxs[i];
1238 LASSERT(rx->rx_nob >= 0); /* not posted */
1240 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1241 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1243 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1246 kiblnd_free_pages(conn->ibc_rx_pages);
1248 conn->ibc_rx_pages = NULL;
1252 kiblnd_map_rx_descs(struct kib_conn *conn)
1260 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1261 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1262 rx = &conn->ibc_rxs[i];
1265 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1268 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1269 rx->rx_msg, IBLND_MSG_SIZE,
1271 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1273 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1275 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1276 i, rx->rx_msg, rx->rx_msgaddr,
1277 (__u64)(page_to_phys(pg) + pg_off));
1279 pg_off += IBLND_MSG_SIZE;
1280 LASSERT(pg_off <= PAGE_SIZE);
1282 if (pg_off == PAGE_SIZE) {
1285 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1291 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1293 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1297 LASSERT (tpo->tpo_pool.po_allocated == 0);
1302 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1303 tx = &tpo->tpo_tx_descs[i];
1304 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1305 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1307 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1310 kiblnd_hdev_decref(hdev);
1311 tpo->tpo_hdev = NULL;
1314 static struct kib_hca_dev *
1315 kiblnd_current_hdev(struct kib_dev *dev)
1317 struct kib_hca_dev *hdev;
1318 unsigned long flags;
1321 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1322 while (dev->ibd_failover) {
1323 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1325 CDEBUG(D_NET, "%s: Wait for failover\n",
1327 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1329 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1332 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1333 hdev = dev->ibd_hdev;
1335 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1341 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1343 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1344 struct kib_pool *pool = &tpo->tpo_pool;
1345 struct kib_net *net = pool->po_owner->ps_net;
1346 struct kib_dev *dev;
1353 LASSERT (net != NULL);
1357 /* pre-mapped messages are not bigger than 1 page */
1358 BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1360 /* No fancy arithmetic when we do the buffer calculations */
1361 BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1363 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1365 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1366 page = txpgs->ibp_pages[ipage];
1367 tx = &tpo->tpo_tx_descs[i];
1369 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1372 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1376 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1378 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1380 list_add(&tx->tx_list, &pool->po_free_list);
1382 page_offset += IBLND_MSG_SIZE;
1383 LASSERT(page_offset <= PAGE_SIZE);
1385 if (page_offset == PAGE_SIZE) {
1388 LASSERT(ipage <= txpgs->ibp_npages);
1394 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1396 LASSERT(fpo->fpo_map_count == 0);
1398 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1399 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1401 struct kib_fast_reg_descriptor *frd, *tmp;
1404 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1406 list_del(&frd->frd_list);
1407 #ifndef HAVE_IB_MAP_MR_SG
1408 ib_free_fast_reg_page_list(frd->frd_frpl);
1410 ib_dereg_mr(frd->frd_mr);
1411 LIBCFS_FREE(frd, sizeof(*frd));
1414 if (i < fpo->fast_reg.fpo_pool_size)
1415 CERROR("FastReg pool still has %d regions registered\n",
1416 fpo->fast_reg.fpo_pool_size - i);
1420 kiblnd_hdev_decref(fpo->fpo_hdev);
1422 LIBCFS_FREE(fpo, sizeof(*fpo));
1426 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1428 struct kib_fmr_pool *fpo, *tmp;
1430 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1431 list_del(&fpo->fpo_list);
1432 kiblnd_destroy_fmr_pool(fpo);
1437 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1440 int size = tunables->lnd_fmr_pool_size / ncpts;
1442 return max(IBLND_FMR_POOL, size);
1446 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1449 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1451 return max(IBLND_FMR_POOL_FLUSH, size);
1454 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1455 struct kib_fmr_pool *fpo)
1457 struct ib_fmr_pool_param param = {
1458 .max_pages_per_fmr = LNET_MAX_IOV,
1459 .page_shift = PAGE_SHIFT,
1460 .access = (IB_ACCESS_LOCAL_WRITE |
1461 IB_ACCESS_REMOTE_WRITE),
1462 .pool_size = fps->fps_pool_size,
1463 .dirty_watermark = fps->fps_flush_trigger,
1464 .flush_function = NULL,
1466 .cache = !!fps->fps_cache };
1469 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1471 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1472 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1474 CERROR("Failed to create FMR pool: %d\n", rc);
1476 CERROR("FMRs are not supported\n");
1478 fpo->fpo_is_fmr = true;
1483 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1484 struct kib_fmr_pool *fpo,
1485 enum kib_dev_caps dev_caps)
1487 struct kib_fast_reg_descriptor *frd, *tmp;
1490 fpo->fpo_is_fmr = false;
1492 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1493 fpo->fast_reg.fpo_pool_size = 0;
1494 for (i = 0; i < fps->fps_pool_size; i++) {
1495 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1498 CERROR("Failed to allocate a new fast_reg descriptor\n");
1504 #ifndef HAVE_IB_MAP_MR_SG
1505 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1507 if (IS_ERR(frd->frd_frpl)) {
1508 rc = PTR_ERR(frd->frd_frpl);
1509 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1511 frd->frd_frpl = NULL;
1516 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1517 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1521 * it is expected to get here if this is an MLX-5 card.
1522 * MLX-4 cards will always use FMR and MLX-5 cards will
1523 * always use fast_reg. It turns out that some MLX-5 cards
1524 * (possibly due to older FW versions) do not natively support
1525 * gaps. So we will need to track them here.
1527 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1528 #ifdef IB_MR_TYPE_SG_GAPS
1529 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1530 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1531 IB_MR_TYPE_SG_GAPS :
1537 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1538 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1539 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1541 if (IS_ERR(frd->frd_mr)) {
1542 rc = PTR_ERR(frd->frd_mr);
1543 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1548 /* There appears to be a bug in MLX5 code where you must
1549 * invalidate the rkey of a new FastReg pool before first
1550 * using it. Thus, I am marking the FRD invalid here. */
1551 frd->frd_valid = false;
1553 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1554 fpo->fast_reg.fpo_pool_size++;
1561 ib_dereg_mr(frd->frd_mr);
1562 #ifndef HAVE_IB_MAP_MR_SG
1564 ib_free_fast_reg_page_list(frd->frd_frpl);
1566 LIBCFS_FREE(frd, sizeof(*frd));
1569 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1571 list_del(&frd->frd_list);
1572 #ifndef HAVE_IB_MAP_MR_SG
1573 ib_free_fast_reg_page_list(frd->frd_frpl);
1575 ib_dereg_mr(frd->frd_mr);
1576 LIBCFS_FREE(frd, sizeof(*frd));
1582 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1583 struct kib_fmr_pool **pp_fpo)
1585 struct kib_dev *dev = fps->fps_net->ibn_dev;
1586 struct kib_fmr_pool *fpo;
1589 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1593 memset(fpo, 0, sizeof(*fpo));
1595 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1597 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1598 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1600 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1604 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1605 fpo->fpo_owner = fps;
1611 kiblnd_hdev_decref(fpo->fpo_hdev);
1612 LIBCFS_FREE(fpo, sizeof(*fpo));
1617 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1619 if (fps->fps_net == NULL) /* intialized? */
1622 spin_lock(&fps->fps_lock);
1624 while (!list_empty(&fps->fps_pool_list)) {
1625 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1626 struct kib_fmr_pool,
1629 fpo->fpo_failed = 1;
1630 if (fpo->fpo_map_count == 0)
1631 list_move(&fpo->fpo_list, zombies);
1633 list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1636 spin_unlock(&fps->fps_lock);
1640 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1642 if (fps->fps_net != NULL) { /* initialized? */
1643 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1644 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1649 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1650 struct kib_net *net,
1651 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1653 struct kib_fmr_pool *fpo;
1656 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1661 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1662 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1663 fps->fps_cache = tunables->lnd_fmr_cache;
1665 spin_lock_init(&fps->fps_lock);
1666 INIT_LIST_HEAD(&fps->fps_pool_list);
1667 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1669 rc = kiblnd_create_fmr_pool(fps, &fpo);
1671 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1677 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1679 if (fpo->fpo_map_count != 0) /* still in use */
1681 if (fpo->fpo_failed)
1683 return now >= fpo->fpo_deadline;
1687 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1689 struct kib_hca_dev *hdev;
1690 __u64 *pages = tx->tx_pages;
1695 hdev = tx->tx_pool->tpo_hdev;
1697 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1698 for (size = 0; size < rd->rd_frags[i].rf_nob;
1699 size += hdev->ibh_page_size) {
1700 pages[npages++] = (rd->rd_frags[i].rf_addr &
1701 hdev->ibh_page_mask) + size;
1709 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1712 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1713 struct kib_fmr_poolset *fps;
1714 time64_t now = ktime_get_seconds();
1715 struct kib_fmr_pool *tmp;
1721 fps = fpo->fpo_owner;
1722 if (fpo->fpo_is_fmr) {
1723 if (fmr->fmr_pfmr) {
1724 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1725 fmr->fmr_pfmr = NULL;
1729 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1733 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1736 frd->frd_valid = false;
1737 spin_lock(&fps->fps_lock);
1738 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1739 spin_unlock(&fps->fps_lock);
1740 fmr->fmr_frd = NULL;
1743 fmr->fmr_pool = NULL;
1745 spin_lock(&fps->fps_lock);
1746 fpo->fpo_map_count--; /* decref the pool */
1748 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1749 /* the first pool is persistent */
1750 if (fps->fps_pool_list.next == &fpo->fpo_list)
1753 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1754 list_move(&fpo->fpo_list, &zombies);
1758 spin_unlock(&fps->fps_lock);
1760 if (!list_empty(&zombies))
1761 kiblnd_destroy_fmr_pool_list(&zombies);
1764 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1765 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1766 struct kib_fmr *fmr)
1768 struct kib_fmr_pool *fpo;
1769 __u64 *pages = tx->tx_pages;
1771 bool is_rx = (rd != tx->tx_rd);
1772 bool tx_pages_mapped = 0;
1777 spin_lock(&fps->fps_lock);
1778 version = fps->fps_version;
1779 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1780 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1781 fpo->fpo_map_count++;
1783 if (fpo->fpo_is_fmr) {
1784 struct ib_pool_fmr *pfmr;
1786 spin_unlock(&fps->fps_lock);
1788 if (!tx_pages_mapped) {
1789 npages = kiblnd_map_tx_pages(tx, rd);
1790 tx_pages_mapped = 1;
1793 pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1794 pages, npages, iov);
1795 if (likely(!IS_ERR(pfmr))) {
1796 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1798 fmr->fmr_frd = NULL;
1799 fmr->fmr_pfmr = pfmr;
1800 fmr->fmr_pool = fpo;
1805 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1806 struct kib_fast_reg_descriptor *frd;
1807 #ifdef HAVE_IB_MAP_MR_SG
1808 struct ib_reg_wr *wr;
1811 struct ib_rdma_wr *wr;
1812 struct ib_fast_reg_page_list *frpl;
1816 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1817 struct kib_fast_reg_descriptor,
1819 list_del(&frd->frd_list);
1820 spin_unlock(&fps->fps_lock);
1822 #ifndef HAVE_IB_MAP_MR_SG
1823 frpl = frd->frd_frpl;
1827 if (!frd->frd_valid) {
1828 struct ib_rdma_wr *inv_wr;
1829 __u32 key = is_rx ? mr->rkey : mr->lkey;
1831 inv_wr = &frd->frd_inv_wr;
1832 memset(inv_wr, 0, sizeof(*inv_wr));
1834 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1835 inv_wr->wr.wr_id = IBLND_WID_MR;
1836 inv_wr->wr.ex.invalidate_rkey = key;
1839 key = ib_inc_rkey(key);
1840 ib_update_fast_reg_key(mr, key);
1843 #ifdef HAVE_IB_MAP_MR_SG
1844 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1845 n = ib_map_mr_sg(mr, tx->tx_frags,
1846 rd->rd_nfrags, NULL, PAGE_SIZE);
1848 n = ib_map_mr_sg(mr, tx->tx_frags,
1849 rd->rd_nfrags, PAGE_SIZE);
1851 if (unlikely(n != rd->rd_nfrags)) {
1852 CERROR("Failed to map mr %d/%d "
1853 "elements\n", n, rd->rd_nfrags);
1854 return n < 0 ? n : -EINVAL;
1857 wr = &frd->frd_fastreg_wr;
1858 memset(wr, 0, sizeof(*wr));
1860 wr->wr.opcode = IB_WR_REG_MR;
1861 wr->wr.wr_id = IBLND_WID_MR;
1863 wr->wr.send_flags = 0;
1865 wr->key = is_rx ? mr->rkey : mr->lkey;
1866 wr->access = (IB_ACCESS_LOCAL_WRITE |
1867 IB_ACCESS_REMOTE_WRITE);
1869 if (!tx_pages_mapped) {
1870 npages = kiblnd_map_tx_pages(tx, rd);
1871 tx_pages_mapped = 1;
1874 LASSERT(npages <= frpl->max_page_list_len);
1875 memcpy(frpl->page_list, pages,
1876 sizeof(*pages) * npages);
1878 /* Prepare FastReg WR */
1879 wr = &frd->frd_fastreg_wr;
1880 memset(wr, 0, sizeof(*wr));
1882 wr->wr.opcode = IB_WR_FAST_REG_MR;
1883 wr->wr.wr_id = IBLND_WID_MR;
1885 wr->wr.wr.fast_reg.iova_start = iov;
1886 wr->wr.wr.fast_reg.page_list = frpl;
1887 wr->wr.wr.fast_reg.page_list_len = npages;
1888 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1889 wr->wr.wr.fast_reg.length = nob;
1890 wr->wr.wr.fast_reg.rkey =
1891 is_rx ? mr->rkey : mr->lkey;
1892 wr->wr.wr.fast_reg.access_flags =
1893 (IB_ACCESS_LOCAL_WRITE |
1894 IB_ACCESS_REMOTE_WRITE);
1897 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1899 fmr->fmr_pfmr = NULL;
1900 fmr->fmr_pool = fpo;
1903 spin_unlock(&fps->fps_lock);
1907 spin_lock(&fps->fps_lock);
1908 fpo->fpo_map_count--;
1909 if (rc != -EAGAIN) {
1910 spin_unlock(&fps->fps_lock);
1914 /* EAGAIN and ... */
1915 if (version != fps->fps_version) {
1916 spin_unlock(&fps->fps_lock);
1921 if (fps->fps_increasing) {
1922 spin_unlock(&fps->fps_lock);
1923 CDEBUG(D_NET, "Another thread is allocating new "
1924 "FMR pool, waiting for her to complete\n");
1925 wait_var_event(fps, !fps->fps_increasing);
1930 if (ktime_get_seconds() < fps->fps_next_retry) {
1931 /* someone failed recently */
1932 spin_unlock(&fps->fps_lock);
1936 fps->fps_increasing = 1;
1937 spin_unlock(&fps->fps_lock);
1939 CDEBUG(D_NET, "Allocate new FMR pool\n");
1940 rc = kiblnd_create_fmr_pool(fps, &fpo);
1941 spin_lock(&fps->fps_lock);
1942 fps->fps_increasing = 0;
1946 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1948 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
1950 spin_unlock(&fps->fps_lock);
1956 kiblnd_fini_pool(struct kib_pool *pool)
1958 LASSERT(list_empty(&pool->po_free_list));
1959 LASSERT(pool->po_allocated == 0);
1961 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1965 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
1967 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1969 memset(pool, 0, sizeof(struct kib_pool));
1970 INIT_LIST_HEAD(&pool->po_free_list);
1971 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1972 pool->po_owner = ps;
1973 pool->po_size = size;
1977 kiblnd_destroy_pool_list(struct list_head *head)
1979 struct kib_pool *pool;
1981 while (!list_empty(head)) {
1982 pool = list_entry(head->next, struct kib_pool, po_list);
1983 list_del(&pool->po_list);
1985 LASSERT(pool->po_owner != NULL);
1986 pool->po_owner->ps_pool_destroy(pool);
1991 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
1993 if (ps->ps_net == NULL) /* intialized? */
1996 spin_lock(&ps->ps_lock);
1997 while (!list_empty(&ps->ps_pool_list)) {
1998 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
1999 struct kib_pool, po_list);
2002 if (po->po_allocated == 0)
2003 list_move(&po->po_list, zombies);
2005 list_move(&po->po_list, &ps->ps_failed_pool_list);
2007 spin_unlock(&ps->ps_lock);
2011 kiblnd_fini_poolset(struct kib_poolset *ps)
2013 if (ps->ps_net != NULL) { /* initialized? */
2014 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2015 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2020 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2021 struct kib_net *net, char *name, int size,
2022 kib_ps_pool_create_t po_create,
2023 kib_ps_pool_destroy_t po_destroy,
2024 kib_ps_node_init_t nd_init,
2025 kib_ps_node_fini_t nd_fini)
2027 struct kib_pool *pool;
2030 memset(ps, 0, sizeof(struct kib_poolset));
2034 ps->ps_pool_create = po_create;
2035 ps->ps_pool_destroy = po_destroy;
2036 ps->ps_node_init = nd_init;
2037 ps->ps_node_fini = nd_fini;
2038 ps->ps_pool_size = size;
2039 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2040 >= sizeof(ps->ps_name))
2042 spin_lock_init(&ps->ps_lock);
2043 INIT_LIST_HEAD(&ps->ps_pool_list);
2044 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2046 rc = ps->ps_pool_create(ps, size, &pool);
2048 list_add(&pool->po_list, &ps->ps_pool_list);
2050 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2056 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2058 if (pool->po_allocated != 0) /* still in use */
2060 if (pool->po_failed)
2062 return now >= pool->po_deadline;
2066 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2069 struct kib_poolset *ps = pool->po_owner;
2070 struct kib_pool *tmp;
2071 time64_t now = ktime_get_seconds();
2073 spin_lock(&ps->ps_lock);
2075 if (ps->ps_node_fini != NULL)
2076 ps->ps_node_fini(pool, node);
2078 LASSERT(pool->po_allocated > 0);
2079 list_add(node, &pool->po_free_list);
2080 pool->po_allocated--;
2082 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2083 /* the first pool is persistent */
2084 if (ps->ps_pool_list.next == &pool->po_list)
2087 if (kiblnd_pool_is_idle(pool, now))
2088 list_move(&pool->po_list, &zombies);
2090 spin_unlock(&ps->ps_lock);
2092 if (!list_empty(&zombies))
2093 kiblnd_destroy_pool_list(&zombies);
2097 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2099 struct list_head *node;
2100 struct kib_pool *pool;
2102 unsigned int interval = 1;
2103 ktime_t time_before;
2104 unsigned int trips = 0;
2107 spin_lock(&ps->ps_lock);
2108 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2109 if (list_empty(&pool->po_free_list))
2112 pool->po_allocated++;
2113 pool->po_deadline = ktime_get_seconds() +
2114 IBLND_POOL_DEADLINE;
2115 node = pool->po_free_list.next;
2118 if (ps->ps_node_init != NULL) {
2119 /* still hold the lock */
2120 ps->ps_node_init(pool, node);
2122 spin_unlock(&ps->ps_lock);
2126 /* no available tx pool and ... */
2127 if (ps->ps_increasing) {
2128 /* another thread is allocating a new pool */
2129 spin_unlock(&ps->ps_lock);
2132 "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2133 ps->ps_name, interval, trips);
2135 schedule_timeout_interruptible(interval);
2136 if (interval < cfs_time_seconds(1))
2142 if (ktime_get_seconds() < ps->ps_next_retry) {
2143 /* someone failed recently */
2144 spin_unlock(&ps->ps_lock);
2148 ps->ps_increasing = 1;
2149 spin_unlock(&ps->ps_lock);
2151 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2152 time_before = ktime_get();
2153 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2154 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2155 ktime_ms_delta(ktime_get(), time_before));
2157 spin_lock(&ps->ps_lock);
2158 ps->ps_increasing = 0;
2160 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2162 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2163 CERROR("Can't allocate new %s pool because out of memory\n",
2166 spin_unlock(&ps->ps_lock);
2172 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2174 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2178 LASSERT (pool->po_allocated == 0);
2180 if (tpo->tpo_tx_pages != NULL) {
2181 kiblnd_unmap_tx_pool(tpo);
2182 kiblnd_free_pages(tpo->tpo_tx_pages);
2185 if (tpo->tpo_tx_descs == NULL)
2188 for (i = 0; i < pool->po_size; i++) {
2189 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2190 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2192 list_del(&tx->tx_list);
2193 if (tx->tx_pages != NULL)
2194 CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2195 if (tx->tx_frags != NULL)
2196 CFS_FREE_PTR_ARRAY(tx->tx_frags,
2197 (1 + IBLND_MAX_RDMA_FRAGS));
2198 if (tx->tx_wrq != NULL)
2199 CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2200 (1 + IBLND_MAX_RDMA_FRAGS));
2201 if (tx->tx_sge != NULL)
2202 CFS_FREE_PTR_ARRAY(tx->tx_sge,
2203 (1 + IBLND_MAX_RDMA_FRAGS) *
2205 if (tx->tx_rd != NULL)
2206 LIBCFS_FREE(tx->tx_rd,
2207 offsetof(struct kib_rdma_desc,
2208 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2211 CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2213 kiblnd_fini_pool(pool);
2217 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2219 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2222 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2223 ntx = tunables->lnd_ntx / ncpts;
2225 return max(IBLND_TX_POOL, ntx);
2229 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2233 struct kib_pool *pool;
2234 struct kib_tx_pool *tpo;
2236 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2238 CERROR("Failed to allocate TX pool\n");
2242 pool = &tpo->tpo_pool;
2243 kiblnd_init_pool(ps, pool, size);
2244 tpo->tpo_tx_descs = NULL;
2245 tpo->tpo_tx_pages = NULL;
2247 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2248 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2249 CERROR("Can't allocate tx pages: %d\n", npg);
2254 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2255 size * sizeof(struct kib_tx));
2256 if (tpo->tpo_tx_descs == NULL) {
2257 CERROR("Can't allocate %d tx descriptors\n", size);
2258 ps->ps_pool_destroy(pool);
2262 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2264 for (i = 0; i < size; i++) {
2265 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2266 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2269 if (ps->ps_net->ibn_fmr_ps != NULL) {
2270 LIBCFS_CPT_ALLOC(tx->tx_pages,
2271 lnet_cpt_table(), ps->ps_cpt,
2272 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2273 if (tx->tx_pages == NULL)
2277 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2278 (1 + IBLND_MAX_RDMA_FRAGS) *
2279 sizeof(*tx->tx_frags));
2280 if (tx->tx_frags == NULL)
2283 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2285 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2286 (1 + IBLND_MAX_RDMA_FRAGS) *
2287 sizeof(*tx->tx_wrq));
2288 if (tx->tx_wrq == NULL)
2291 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2292 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2293 sizeof(*tx->tx_sge));
2294 if (tx->tx_sge == NULL)
2297 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2298 offsetof(struct kib_rdma_desc,
2299 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2300 if (tx->tx_rd == NULL)
2305 kiblnd_map_tx_pool(tpo);
2310 ps->ps_pool_destroy(pool);
2315 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2317 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2318 struct kib_tx_poolset,
2320 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2322 tx->tx_cookie = tps->tps_next_tx_cookie++;
2326 kiblnd_net_fini_pools(struct kib_net *net)
2330 cfs_cpt_for_each(i, lnet_cpt_table()) {
2331 struct kib_tx_poolset *tps;
2332 struct kib_fmr_poolset *fps;
2334 if (net->ibn_tx_ps != NULL) {
2335 tps = net->ibn_tx_ps[i];
2336 kiblnd_fini_poolset(&tps->tps_poolset);
2339 if (net->ibn_fmr_ps != NULL) {
2340 fps = net->ibn_fmr_ps[i];
2341 kiblnd_fini_fmr_poolset(fps);
2345 if (net->ibn_tx_ps != NULL) {
2346 cfs_percpt_free(net->ibn_tx_ps);
2347 net->ibn_tx_ps = NULL;
2350 if (net->ibn_fmr_ps != NULL) {
2351 cfs_percpt_free(net->ibn_fmr_ps);
2352 net->ibn_fmr_ps = NULL;
2357 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2360 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2361 #ifdef HAVE_IB_GET_DMA_MR
2362 unsigned long flags;
2368 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2370 #ifdef HAVE_IB_GET_DMA_MR
2371 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2373 * if lnd_map_on_demand is zero then we have effectively disabled
2374 * FMR or FastReg and we're using global memory regions
2377 if (!tunables->lnd_map_on_demand) {
2378 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2380 goto create_tx_pool;
2383 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2386 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2387 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2388 tunables->lnd_fmr_pool_size,
2389 tunables->lnd_ntx / 4);
2394 /* TX pool must be created later than FMR, see LU-2268
2396 LASSERT(net->ibn_tx_ps == NULL);
2398 /* premapping can fail if ibd_nmr > 1, so we always create
2399 * FMR pool and map-on-demand if premapping failed */
2401 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2402 sizeof(struct kib_fmr_poolset));
2403 if (net->ibn_fmr_ps == NULL) {
2404 CERROR("Failed to allocate FMR pool array\n");
2409 for (i = 0; i < ncpts; i++) {
2410 cpt = (cpts == NULL) ? i : cpts[i];
2411 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2414 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2421 LASSERT(i == ncpts);
2423 #ifdef HAVE_IB_GET_DMA_MR
2426 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2427 sizeof(struct kib_tx_poolset));
2428 if (net->ibn_tx_ps == NULL) {
2429 CERROR("Failed to allocate tx pool array\n");
2434 for (i = 0; i < ncpts; i++) {
2435 cpt = (cpts == NULL) ? i : cpts[i];
2436 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2438 kiblnd_tx_pool_size(ni, ncpts),
2439 kiblnd_create_tx_pool,
2440 kiblnd_destroy_tx_pool,
2441 kiblnd_tx_init, NULL);
2443 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2451 kiblnd_net_fini_pools(net);
2457 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2459 struct ib_port_attr *port_attr;
2461 unsigned long flags;
2462 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2464 LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2465 if (port_attr == NULL) {
2466 CDEBUG(D_NETERROR, "Out of memory\n");
2470 rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2472 write_lock_irqsave(g_lock, flags);
2475 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2476 ? IBLND_DEV_PORT_ACTIVE
2477 : IBLND_DEV_PORT_DOWN;
2479 write_unlock_irqrestore(g_lock, flags);
2480 LIBCFS_FREE(port_attr, sizeof(*port_attr));
2483 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2490 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2492 struct kib_net *net;
2494 /* for health check */
2495 list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2497 CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2498 libcfs_nid2str(net->ibn_ni->ni_nid));
2499 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2504 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2506 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2507 struct kib_hca_dev *hdev;
2508 unsigned long flags;
2510 hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2512 write_lock_irqsave(g_lock, flags);
2514 switch (event->event) {
2515 case IB_EVENT_DEVICE_FATAL:
2516 CDEBUG(D_NET, "IB device fatal\n");
2517 hdev->ibh_state = IBLND_DEV_FATAL;
2518 kiblnd_set_ni_fatal_on(hdev, 1);
2520 case IB_EVENT_PORT_ACTIVE:
2521 CDEBUG(D_NET, "IB port active\n");
2522 if (event->element.port_num == hdev->ibh_port) {
2523 hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2524 kiblnd_set_ni_fatal_on(hdev, 0);
2527 case IB_EVENT_PORT_ERR:
2528 CDEBUG(D_NET, "IB port err\n");
2529 if (event->element.port_num == hdev->ibh_port) {
2530 hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2531 kiblnd_set_ni_fatal_on(hdev, 1);
2537 write_unlock_irqrestore(g_lock, flags);
2541 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2543 struct ib_device_attr *dev_attr;
2547 /* It's safe to assume a HCA can handle a page size
2548 * matching that of the native system */
2549 hdev->ibh_page_shift = PAGE_SHIFT;
2550 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2551 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2553 #ifndef HAVE_IB_DEVICE_ATTRS
2554 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2555 if (dev_attr == NULL) {
2556 CERROR("Out of memory\n");
2560 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2562 CERROR("Failed to query IB device: %d\n", rc);
2563 goto out_clean_attr;
2566 dev_attr = &hdev->ibh_ibdev->attrs;
2569 hdev->ibh_mr_size = dev_attr->max_mr_size;
2570 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2572 /* Setup device Memory Registration capabilities */
2573 #ifdef HAVE_IB_DEVICE_OPS
2574 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2575 hdev->ibh_ibdev->ops.dealloc_fmr &&
2576 hdev->ibh_ibdev->ops.map_phys_fmr &&
2577 hdev->ibh_ibdev->ops.unmap_fmr) {
2579 if (hdev->ibh_ibdev->alloc_fmr &&
2580 hdev->ibh_ibdev->dealloc_fmr &&
2581 hdev->ibh_ibdev->map_phys_fmr &&
2582 hdev->ibh_ibdev->unmap_fmr) {
2584 LCONSOLE_INFO("Using FMR for registration\n");
2585 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2586 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2587 LCONSOLE_INFO("Using FastReg for registration\n");
2588 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2589 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2590 #ifdef IB_DEVICE_SG_GAPS_REG
2591 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2592 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2599 rc2 = kiblnd_port_get_attr(hdev);
2606 #ifndef HAVE_IB_DEVICE_ATTRS
2608 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2612 CERROR("IB device does not support FMRs nor FastRegs, can't "
2613 "register memory: %d\n", rc);
2614 else if (rc == -EINVAL)
2615 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2619 #ifdef HAVE_IB_GET_DMA_MR
2621 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2623 if (hdev->ibh_mrs == NULL)
2626 ib_dereg_mr(hdev->ibh_mrs);
2628 hdev->ibh_mrs = NULL;
2633 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2635 if (hdev->ibh_event_handler.device != NULL)
2636 ib_unregister_event_handler(&hdev->ibh_event_handler);
2638 #ifdef HAVE_IB_GET_DMA_MR
2639 kiblnd_hdev_cleanup_mrs(hdev);
2642 if (hdev->ibh_pd != NULL)
2643 ib_dealloc_pd(hdev->ibh_pd);
2645 if (hdev->ibh_cmid != NULL)
2646 rdma_destroy_id(hdev->ibh_cmid);
2648 LIBCFS_FREE(hdev, sizeof(*hdev));
2651 #ifdef HAVE_IB_GET_DMA_MR
2653 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2656 int acflags = IB_ACCESS_LOCAL_WRITE |
2657 IB_ACCESS_REMOTE_WRITE;
2659 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2661 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2662 kiblnd_hdev_cleanup_mrs(hdev);
2673 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2679 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2681 struct rdma_cm_id *cmid;
2682 struct sockaddr_in srcaddr;
2683 struct sockaddr_in dstaddr;
2686 if (dev->ibd_hdev == NULL || /* initializing */
2687 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2688 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2691 /* XXX: it's UGLY, but I don't have better way to find
2692 * ib-bonding HCA failover because:
2694 * a. no reliable CM event for HCA failover...
2695 * b. no OFED API to get ib_device for current net_device...
2697 * We have only two choices at this point:
2699 * a. rdma_bind_addr(), it will conflict with listener cmid
2700 * b. rdma_resolve_addr() to zero addr */
2701 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2702 RDMA_PS_TCP, IB_QPT_RC);
2705 CERROR("Failed to create cmid for failover: %d\n", rc);
2709 memset(&srcaddr, 0, sizeof(srcaddr));
2710 srcaddr.sin_family = AF_INET;
2711 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2713 memset(&dstaddr, 0, sizeof(dstaddr));
2714 dstaddr.sin_family = AF_INET;
2715 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2716 (struct sockaddr *)&dstaddr, 1);
2717 if (rc != 0 || cmid->device == NULL) {
2718 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2719 dev->ibd_ifname, &dev->ibd_ifip,
2721 rdma_destroy_id(cmid);
2725 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2726 rdma_destroy_id(cmid);
2731 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2733 LIST_HEAD(zombie_tpo);
2734 LIST_HEAD(zombie_ppo);
2735 LIST_HEAD(zombie_fpo);
2736 struct rdma_cm_id *cmid = NULL;
2737 struct kib_hca_dev *hdev = NULL;
2738 struct kib_hca_dev *old;
2740 struct kib_net *net;
2741 struct sockaddr_in addr;
2742 unsigned long flags;
2746 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2747 dev->ibd_can_failover ||
2748 dev->ibd_hdev == NULL);
2750 rc = kiblnd_dev_need_failover(dev, ns);
2754 if (dev->ibd_hdev != NULL &&
2755 dev->ibd_hdev->ibh_cmid != NULL) {
2756 /* XXX it's not good to close old listener at here,
2757 * because we can fail to create new listener.
2758 * But we have to close it now, otherwise rdma_bind_addr
2759 * will return EADDRINUSE... How crap! */
2760 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2762 cmid = dev->ibd_hdev->ibh_cmid;
2763 /* make next schedule of kiblnd_dev_need_failover()
2764 * return 1 for me */
2765 dev->ibd_hdev->ibh_cmid = NULL;
2766 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2768 rdma_destroy_id(cmid);
2771 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2775 CERROR("Failed to create cmid for failover: %d\n", rc);
2779 memset(&addr, 0, sizeof(addr));
2780 addr.sin_family = AF_INET;
2781 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2782 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2784 /* Bind to failover device or port */
2785 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2786 if (rc != 0 || cmid->device == NULL) {
2787 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2788 dev->ibd_ifname, &dev->ibd_ifip,
2790 rdma_destroy_id(cmid);
2794 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2796 CERROR("Failed to allocate kib_hca_dev\n");
2797 rdma_destroy_id(cmid);
2802 atomic_set(&hdev->ibh_ref, 1);
2803 hdev->ibh_dev = dev;
2804 hdev->ibh_cmid = cmid;
2805 hdev->ibh_ibdev = cmid->device;
2806 hdev->ibh_port = cmid->port_num;
2808 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2809 pd = ib_alloc_pd(cmid->device, 0);
2811 pd = ib_alloc_pd(cmid->device);
2815 CERROR("Can't allocate PD: %d\n", rc);
2821 rc = rdma_listen(cmid, 0);
2823 CERROR("Can't start new listener: %d\n", rc);
2827 rc = kiblnd_hdev_get_attr(hdev);
2829 CERROR("Can't get device attributes: %d\n", rc);
2833 #ifdef HAVE_IB_GET_DMA_MR
2834 rc = kiblnd_hdev_setup_mrs(hdev);
2836 CERROR("Can't setup device: %d\n", rc);
2841 INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
2842 hdev->ibh_ibdev, kiblnd_event_handler);
2843 ib_register_event_handler(&hdev->ibh_event_handler);
2845 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2847 old = dev->ibd_hdev;
2848 dev->ibd_hdev = hdev; /* take over the refcount */
2851 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2852 cfs_cpt_for_each(i, lnet_cpt_table()) {
2853 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2856 if (net->ibn_fmr_ps != NULL)
2857 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2862 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2864 if (!list_empty(&zombie_tpo))
2865 kiblnd_destroy_pool_list(&zombie_tpo);
2866 if (!list_empty(&zombie_ppo))
2867 kiblnd_destroy_pool_list(&zombie_ppo);
2868 if (!list_empty(&zombie_fpo))
2869 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2871 kiblnd_hdev_decref(hdev);
2874 dev->ibd_failed_failover++;
2876 dev->ibd_failed_failover = 0;
2882 kiblnd_destroy_dev(struct kib_dev *dev)
2884 LASSERT(dev->ibd_nnets == 0);
2885 LASSERT(list_empty(&dev->ibd_nets));
2887 list_del(&dev->ibd_fail_list);
2888 list_del(&dev->ibd_list);
2890 if (dev->ibd_hdev != NULL)
2891 kiblnd_hdev_decref(dev->ibd_hdev);
2893 LIBCFS_FREE(dev, sizeof(*dev));
2897 kiblnd_base_shutdown(void)
2899 struct kib_sched_info *sched;
2902 LASSERT(list_empty(&kiblnd_data.kib_devs));
2904 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2905 atomic_read(&libcfs_kmemory));
2907 switch (kiblnd_data.kib_init) {
2911 case IBLND_INIT_ALL:
2912 case IBLND_INIT_DATA:
2913 LASSERT (kiblnd_data.kib_peers != NULL);
2914 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2915 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2917 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2918 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2919 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2920 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2922 /* flag threads to terminate; wake and wait for them to die */
2923 kiblnd_data.kib_shutdown = 1;
2925 /* NB: we really want to stop scheduler threads net by net
2926 * instead of the whole module, this should be improved
2927 * with dynamic configuration LNet */
2928 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2929 wake_up_all(&sched->ibs_waitq);
2931 wake_up_all(&kiblnd_data.kib_connd_waitq);
2932 wake_up_all(&kiblnd_data.kib_failover_waitq);
2934 wait_var_event_warning(&kiblnd_data.kib_nthreads,
2935 !atomic_read(&kiblnd_data.kib_nthreads),
2936 "Waiting for %d threads to terminate\n",
2937 atomic_read(&kiblnd_data.kib_nthreads));
2940 case IBLND_INIT_NOTHING:
2944 if (kiblnd_data.kib_peers)
2945 CFS_FREE_PTR_ARRAY(kiblnd_data.kib_peers,
2946 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;
2963 unsigned long flags;
2965 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2970 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2971 atomic_read(&libcfs_kmemory));
2973 write_lock_irqsave(g_lock, flags);
2974 net->ibn_shutdown = 1;
2975 write_unlock_irqrestore(g_lock, flags);
2977 switch (net->ibn_init) {
2981 case IBLND_INIT_ALL:
2982 /* nuke all existing peers within this net */
2983 kiblnd_del_peer(ni, LNET_NID_ANY);
2985 /* Wait for all peer_ni state to clean up */
2986 wait_var_event_warning(&net->ibn_npeers,
2987 atomic_read(&net->ibn_npeers) == 0,
2988 "%s: waiting for %d peers to disconnect\n",
2989 libcfs_nid2str(ni->ni_nid),
2990 atomic_read(&net->ibn_npeers));
2992 kiblnd_net_fini_pools(net);
2994 write_lock_irqsave(g_lock, flags);
2995 LASSERT(net->ibn_dev->ibd_nnets > 0);
2996 net->ibn_dev->ibd_nnets--;
2997 list_del(&net->ibn_list);
2998 write_unlock_irqrestore(g_lock, flags);
3002 case IBLND_INIT_NOTHING:
3003 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3005 if (net->ibn_dev != NULL &&
3006 net->ibn_dev->ibd_nnets == 0)
3007 kiblnd_destroy_dev(net->ibn_dev);
3012 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3013 atomic_read(&libcfs_kmemory));
3015 net->ibn_init = IBLND_INIT_NOTHING;
3018 LIBCFS_FREE(net, sizeof(*net));
3021 if (list_empty(&kiblnd_data.kib_devs))
3022 kiblnd_base_shutdown();
3026 kiblnd_base_startup(struct net *ns)
3028 struct kib_sched_info *sched;
3032 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3034 if (!try_module_get(THIS_MODULE))
3037 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3039 rwlock_init(&kiblnd_data.kib_global_lock);
3041 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3042 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3044 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3045 CFS_ALLOC_PTR_ARRAY(kiblnd_data.kib_peers,
3046 kiblnd_data.kib_peer_hash_size);
3047 if (kiblnd_data.kib_peers == NULL)
3050 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3051 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3053 spin_lock_init(&kiblnd_data.kib_connd_lock);
3054 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3055 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3056 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3057 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3059 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3060 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3062 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3064 if (kiblnd_data.kib_scheds == NULL)
3067 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3070 spin_lock_init(&sched->ibs_lock);
3071 INIT_LIST_HEAD(&sched->ibs_conns);
3072 init_waitqueue_head(&sched->ibs_waitq);
3074 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3075 if (*kiblnd_tunables.kib_nscheds > 0) {
3076 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3078 /* max to half of CPUs, another half is reserved for
3079 * upper layer modules */
3080 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3083 sched->ibs_nthreads_max = nthrs;
3087 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3089 /* lists/ptrs/locks initialised */
3090 kiblnd_data.kib_init = IBLND_INIT_DATA;
3091 /*****************************************************/
3093 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3095 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3099 if (*kiblnd_tunables.kib_dev_failover != 0)
3100 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3104 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3108 /* flag everything initialised */
3109 kiblnd_data.kib_init = IBLND_INIT_ALL;
3110 /*****************************************************/
3115 kiblnd_base_shutdown();
3120 kiblnd_start_schedulers(struct kib_sched_info *sched)
3126 if (sched->ibs_nthreads == 0) {
3127 if (*kiblnd_tunables.kib_nscheds > 0) {
3128 nthrs = sched->ibs_nthreads_max;
3130 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3132 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3133 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3136 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3137 /* increase one thread if there is new interface */
3138 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3141 for (i = 0; i < nthrs; i++) {
3144 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3145 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3146 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3147 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3151 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3152 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3156 sched->ibs_nthreads += i;
3160 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3167 for (i = 0; i < ncpts; i++) {
3168 struct kib_sched_info *sched;
3170 cpt = (cpts == NULL) ? i : cpts[i];
3171 sched = kiblnd_data.kib_scheds[cpt];
3173 if (!newdev && sched->ibs_nthreads > 0)
3176 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3178 CERROR("Failed to start scheduler threads for %s\n",
3186 static struct kib_dev *
3187 kiblnd_dev_search(char *ifname)
3189 struct kib_dev *alias = NULL;
3190 struct kib_dev *dev;
3194 colon = strchr(ifname, ':');
3195 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3196 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3202 colon2 = strchr(dev->ibd_ifname, ':');
3208 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3220 kiblnd_startup(struct lnet_ni *ni)
3222 char *ifname = NULL;
3223 struct lnet_inetdev *ifaces = NULL;
3224 struct kib_dev *ibdev = NULL;
3225 struct kib_net *net = NULL;
3226 unsigned long flags;
3231 LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3233 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3234 rc = kiblnd_base_startup(ni->ni_net_ns);
3239 LIBCFS_ALLOC(net, sizeof(*net));
3247 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3249 kiblnd_tunables_setup(ni);
3252 * ni_interfaces is only to support legacy pre Multi-Rail
3253 * tcp bonding for ksocklnd. Multi-Rail wants each secondary
3254 * IP to be treated as an unique 'struct ni' interfaces instead.
3256 if (ni->ni_interfaces[0] != NULL) {
3257 /* Use the IPoIB interface specified in 'networks=' */
3258 if (ni->ni_interfaces[1] != NULL) {
3259 CERROR("ko2iblnd: Multiple interfaces not supported\n");
3264 ifname = ni->ni_interfaces[0];
3266 ifname = *kiblnd_tunables.kib_default_ipif;
3269 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3270 CERROR("IPoIB interface name too long: %s\n", ifname);
3275 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3279 for (i = 0; i < rc; i++) {
3280 if (strcmp(ifname, ifaces[i].li_name) == 0)
3285 CERROR("ko2iblnd: No matching interfaces\n");
3290 ibdev = kiblnd_dev_search(ifname);
3291 newdev = ibdev == NULL;
3292 /* hmm...create kib_dev even for alias */
3293 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3294 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3300 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3301 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3302 sizeof(ibdev->ibd_ifname));
3303 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3305 INIT_LIST_HEAD(&ibdev->ibd_nets);
3306 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3307 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3309 /* initialize the device */
3310 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3312 CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3317 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3320 net->ibn_dev = ibdev;
3321 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3323 ni->ni_dev_cpt = ifaces[i].li_cpt;
3325 rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3329 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3331 CERROR("Failed to initialize NI pools: %d\n", rc);
3335 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3337 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3338 /* for health check */
3339 if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3340 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3341 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3343 net->ibn_init = IBLND_INIT_ALL;
3348 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3349 kiblnd_destroy_dev(ibdev);
3352 kiblnd_shutdown(ni);
3354 CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3355 ifname ? ifname : "", rc);
3360 static const struct lnet_lnd the_o2iblnd = {
3361 .lnd_type = O2IBLND,
3362 .lnd_startup = kiblnd_startup,
3363 .lnd_shutdown = kiblnd_shutdown,
3364 .lnd_ctl = kiblnd_ctl,
3365 .lnd_send = kiblnd_send,
3366 .lnd_recv = kiblnd_recv,
3369 static void __exit ko2iblnd_exit(void)
3371 lnet_unregister_lnd(&the_o2iblnd);
3374 static int __init ko2iblnd_init(void)
3378 BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3379 BUILD_BUG_ON(offsetof(struct kib_msg,
3380 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3382 BUILD_BUG_ON(offsetof(struct kib_msg,
3383 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3386 rc = kiblnd_tunables_init();
3390 lnet_register_lnd(&the_o2iblnd);
3395 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3396 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3397 MODULE_VERSION("2.8.0");
3398 MODULE_LICENSE("GPL");
3400 module_init(ko2iblnd_init);
3401 module_exit(ko2iblnd_exit);