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 status %d msg_type %x cred %d\n",
574 rx, rx->rx_status, rx->rx_msg->ibm_type,
575 rx->rx_msg->ibm_credits);
579 kiblnd_debug_tx(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;
829 INIT_LIST_HEAD(&conn->ibc_early_rxs);
830 INIT_LIST_HEAD(&conn->ibc_tx_noops);
831 INIT_LIST_HEAD(&conn->ibc_tx_queue);
832 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
833 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
834 INIT_LIST_HEAD(&conn->ibc_active_txs);
835 INIT_LIST_HEAD(&conn->ibc_zombie_txs);
836 spin_lock_init(&conn->ibc_lock);
838 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
839 sizeof(*conn->ibc_connvars));
840 if (conn->ibc_connvars == NULL) {
841 CERROR("Can't allocate in-progress connection state\n");
845 write_lock_irqsave(glock, flags);
846 if (dev->ibd_failover) {
847 write_unlock_irqrestore(glock, flags);
848 CERROR("%s: failover in progress\n", dev->ibd_ifname);
852 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
853 /* wakeup failover thread and teardown connection */
854 if (kiblnd_dev_can_failover(dev)) {
855 list_add_tail(&dev->ibd_fail_list,
856 &kiblnd_data.kib_failed_devs);
857 wake_up(&kiblnd_data.kib_failover_waitq);
860 write_unlock_irqrestore(glock, flags);
861 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
862 cmid->device->name, dev->ibd_ifname);
866 kiblnd_hdev_addref_locked(dev->ibd_hdev);
867 conn->ibc_hdev = dev->ibd_hdev;
869 kiblnd_setup_mtu_locked(cmid);
871 write_unlock_irqrestore(glock, flags);
873 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
874 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
875 if (conn->ibc_rxs == NULL) {
876 CERROR("Cannot allocate RX buffers\n");
880 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
881 IBLND_RX_MSG_PAGES(conn));
885 kiblnd_map_rx_descs(conn);
887 #ifdef HAVE_IB_CQ_INIT_ATTR
888 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
889 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
890 cq = ib_create_cq(cmid->device,
891 kiblnd_cq_completion, kiblnd_cq_event, conn,
894 cq = ib_create_cq(cmid->device,
895 kiblnd_cq_completion, kiblnd_cq_event, conn,
896 IBLND_CQ_ENTRIES(conn),
897 kiblnd_get_completion_vector(conn, cpt));
901 * on MLX-5 (possibly MLX-4 as well) this error could be
902 * hit if the concurrent_sends and/or peer_tx_credits is set
903 * too high. Or due to an MLX-5 bug which tries to
904 * allocate 256kb via kmalloc for WR cookie array
906 CERROR("Failed to create CQ with %d CQEs: %ld\n",
907 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
913 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
915 CERROR("Can't request completion notification: %d\n", rc);
919 init_qp_attr->event_handler = kiblnd_qp_event;
920 init_qp_attr->qp_context = conn;
921 init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
922 init_qp_attr->cap.max_recv_sge = 1;
923 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
924 init_qp_attr->qp_type = IB_QPT_RC;
925 init_qp_attr->send_cq = cq;
926 init_qp_attr->recv_cq = cq;
928 * kiblnd_send_wrs() can change the connection's queue depth if
929 * the maximum work requests for the device is maxed out
931 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
932 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
934 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
936 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
937 "send_sge: %d, recv_sge: %d\n",
938 rc, init_qp_attr->cap.max_send_wr,
939 init_qp_attr->cap.max_recv_wr,
940 init_qp_attr->cap.max_send_sge,
941 init_qp_attr->cap.max_recv_sge);
945 conn->ibc_sched = sched;
947 if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
948 CWARN("peer %s - queue depth reduced from %u to %u"
949 " to allow for qp creation\n",
950 libcfs_nid2str(peer_ni->ibp_nid),
951 peer_ni->ibp_queue_depth,
952 conn->ibc_queue_depth);
954 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
956 /* 1 ref for caller and each rxmsg */
957 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
958 conn->ibc_nrx = IBLND_RX_MSGS(conn);
961 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
962 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
964 CERROR("Can't post rxmsg: %d\n", rc);
966 /* Make posted receives complete */
967 kiblnd_abort_receives(conn);
969 /* correct # of posted buffers
970 * NB locking needed now I'm racing with completion */
971 spin_lock_irqsave(&sched->ibs_lock, flags);
972 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
973 spin_unlock_irqrestore(&sched->ibs_lock, flags);
975 /* cmid will be destroyed by CM(ofed) after cm_callback
976 * returned, so we can't refer it anymore
977 * (by kiblnd_connd()->kiblnd_destroy_conn) */
978 rdma_destroy_qp(conn->ibc_cmid);
979 conn->ibc_cmid = NULL;
981 /* Drop my own and unused rxbuffer refcounts */
982 while (i++ <= IBLND_RX_MSGS(conn))
983 kiblnd_conn_decref(conn);
989 /* Init successful! */
990 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
991 state == IBLND_CONN_PASSIVE_WAIT);
992 conn->ibc_state = state;
995 atomic_inc(&net->ibn_nconns);
999 kiblnd_destroy_conn(conn);
1000 LIBCFS_FREE(conn, sizeof(*conn));
1002 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
1008 kiblnd_destroy_conn(struct kib_conn *conn)
1010 struct rdma_cm_id *cmid = conn->ibc_cmid;
1011 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1013 LASSERT (!in_interrupt());
1014 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1015 LASSERT(list_empty(&conn->ibc_early_rxs));
1016 LASSERT(list_empty(&conn->ibc_tx_noops));
1017 LASSERT(list_empty(&conn->ibc_tx_queue));
1018 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1019 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1020 LASSERT(list_empty(&conn->ibc_active_txs));
1021 LASSERT (conn->ibc_noops_posted == 0);
1022 LASSERT (conn->ibc_nsends_posted == 0);
1024 switch (conn->ibc_state) {
1026 /* conn must be completely disengaged from the network */
1029 case IBLND_CONN_DISCONNECTED:
1030 /* connvars should have been freed already */
1031 LASSERT (conn->ibc_connvars == NULL);
1034 case IBLND_CONN_INIT:
1038 /* conn->ibc_cmid might be destroyed by CM already */
1039 if (cmid != NULL && cmid->qp != NULL)
1040 rdma_destroy_qp(cmid);
1043 ib_destroy_cq(conn->ibc_cq);
1045 kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1046 LNET_MSG_STATUS_OK);
1048 if (conn->ibc_rx_pages != NULL)
1049 kiblnd_unmap_rx_descs(conn);
1051 if (conn->ibc_rxs != NULL) {
1052 LIBCFS_FREE(conn->ibc_rxs,
1053 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
1056 if (conn->ibc_connvars != NULL)
1057 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1059 if (conn->ibc_hdev != NULL)
1060 kiblnd_hdev_decref(conn->ibc_hdev);
1062 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1063 if (conn->ibc_state != IBLND_CONN_INIT) {
1064 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1066 kiblnd_peer_decref(peer_ni);
1067 rdma_destroy_id(cmid);
1068 atomic_dec(&net->ibn_nconns);
1073 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1075 struct kib_conn *conn;
1076 struct list_head *ctmp;
1077 struct list_head *cnxt;
1080 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1081 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1083 CDEBUG(D_NET, "Closing conn -> %s, "
1084 "version: %x, reason: %d\n",
1085 libcfs_nid2str(peer_ni->ibp_nid),
1086 conn->ibc_version, why);
1088 kiblnd_close_conn_locked(conn, why);
1096 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1097 int version, __u64 incarnation)
1099 struct kib_conn *conn;
1100 struct list_head *ctmp;
1101 struct list_head *cnxt;
1104 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1105 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1107 if (conn->ibc_version == version &&
1108 conn->ibc_incarnation == incarnation)
1111 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1112 "incarnation:%#llx(%x, %#llx)\n",
1113 libcfs_nid2str(peer_ni->ibp_nid),
1114 conn->ibc_version, conn->ibc_incarnation,
1115 version, incarnation);
1117 kiblnd_close_conn_locked(conn, -ESTALE);
1125 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1127 struct kib_peer_ni *peer_ni;
1128 struct list_head *ptmp;
1129 struct list_head *pnxt;
1133 unsigned long flags;
1136 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1138 if (nid != LNET_NID_ANY)
1139 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1142 hi = kiblnd_data.kib_peer_hash_size - 1;
1145 for (i = lo; i <= hi; i++) {
1146 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1148 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
1149 LASSERT(!kiblnd_peer_idle(peer_ni));
1151 if (peer_ni->ibp_ni != ni)
1154 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1157 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1161 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1163 /* wildcards always succeed */
1164 if (nid == LNET_NID_ANY)
1167 return (count == 0) ? -ENOENT : 0;
1171 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1173 struct libcfs_ioctl_data *data = arg;
1177 case IOC_LIBCFS_GET_PEER: {
1181 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1183 data->ioc_nid = nid;
1184 data->ioc_count = count;
1188 case IOC_LIBCFS_DEL_PEER: {
1189 rc = kiblnd_del_peer(ni, data->ioc_nid);
1192 case IOC_LIBCFS_GET_CONN: {
1193 struct kib_conn *conn;
1196 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1202 LASSERT(conn->ibc_cmid != NULL);
1203 data->ioc_nid = conn->ibc_peer->ibp_nid;
1204 if (conn->ibc_cmid->route.path_rec == NULL)
1205 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1208 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1209 kiblnd_conn_decref(conn);
1212 case IOC_LIBCFS_CLOSE_CONNECTION: {
1213 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1225 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1227 time64_t last_alive = 0;
1228 time64_t now = ktime_get_seconds();
1229 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1230 struct kib_peer_ni *peer_ni;
1231 unsigned long flags;
1233 read_lock_irqsave(glock, flags);
1235 peer_ni = kiblnd_find_peer_locked(ni, nid);
1236 if (peer_ni != NULL)
1237 last_alive = peer_ni->ibp_last_alive;
1239 read_unlock_irqrestore(glock, flags);
1241 if (last_alive != 0)
1244 /* peer_ni is not persistent in hash, trigger peer_ni creation
1245 * and connection establishment with a NULL tx */
1246 if (peer_ni == NULL)
1247 kiblnd_launch_tx(ni, NULL, nid);
1249 CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n",
1250 libcfs_nid2str(nid), peer_ni,
1251 last_alive ? now - last_alive : -1);
1256 kiblnd_free_pages(struct kib_pages *p)
1258 int npages = p->ibp_npages;
1261 for (i = 0; i < npages; i++) {
1262 if (p->ibp_pages[i] != NULL)
1263 __free_page(p->ibp_pages[i]);
1266 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1270 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1272 struct kib_pages *p;
1275 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1276 offsetof(struct kib_pages, ibp_pages[npages]));
1278 CERROR("Can't allocate descriptor for %d pages\n", npages);
1282 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1283 p->ibp_npages = npages;
1285 for (i = 0; i < npages; i++) {
1286 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1288 if (p->ibp_pages[i] == NULL) {
1289 CERROR("Can't allocate page %d of %d\n", i, npages);
1290 kiblnd_free_pages(p);
1300 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1305 LASSERT (conn->ibc_rxs != NULL);
1306 LASSERT (conn->ibc_hdev != NULL);
1308 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1309 rx = &conn->ibc_rxs[i];
1311 LASSERT(rx->rx_nob >= 0); /* not posted */
1313 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1314 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1316 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1319 kiblnd_free_pages(conn->ibc_rx_pages);
1321 conn->ibc_rx_pages = NULL;
1325 kiblnd_map_rx_descs(struct kib_conn *conn)
1333 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1334 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1335 rx = &conn->ibc_rxs[i];
1338 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1341 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1342 rx->rx_msg, IBLND_MSG_SIZE,
1344 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1346 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1348 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1349 i, rx->rx_msg, rx->rx_msgaddr,
1350 (__u64)(page_to_phys(pg) + pg_off));
1352 pg_off += IBLND_MSG_SIZE;
1353 LASSERT(pg_off <= PAGE_SIZE);
1355 if (pg_off == PAGE_SIZE) {
1358 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1364 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1366 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1370 LASSERT (tpo->tpo_pool.po_allocated == 0);
1375 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1376 tx = &tpo->tpo_tx_descs[i];
1377 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1378 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1380 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1383 kiblnd_hdev_decref(hdev);
1384 tpo->tpo_hdev = NULL;
1387 static struct kib_hca_dev *
1388 kiblnd_current_hdev(struct kib_dev *dev)
1390 struct kib_hca_dev *hdev;
1391 unsigned long flags;
1394 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1395 while (dev->ibd_failover) {
1396 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1398 CDEBUG(D_NET, "%s: Wait for failover\n",
1400 set_current_state(TASK_INTERRUPTIBLE);
1401 schedule_timeout(cfs_time_seconds(1) / 100);
1403 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1406 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1407 hdev = dev->ibd_hdev;
1409 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1415 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1417 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1418 struct kib_pool *pool = &tpo->tpo_pool;
1419 struct kib_net *net = pool->po_owner->ps_net;
1420 struct kib_dev *dev;
1427 LASSERT (net != NULL);
1431 /* pre-mapped messages are not bigger than 1 page */
1432 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1434 /* No fancy arithmetic when we do the buffer calculations */
1435 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1437 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1439 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1440 page = txpgs->ibp_pages[ipage];
1441 tx = &tpo->tpo_tx_descs[i];
1443 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1446 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1450 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1452 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1454 list_add(&tx->tx_list, &pool->po_free_list);
1456 page_offset += IBLND_MSG_SIZE;
1457 LASSERT(page_offset <= PAGE_SIZE);
1459 if (page_offset == PAGE_SIZE) {
1462 LASSERT(ipage <= txpgs->ibp_npages);
1468 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1470 LASSERT(fpo->fpo_map_count == 0);
1472 #ifdef HAVE_FMR_POOL_API
1473 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1474 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1476 #endif /* HAVE_FMR_POOL_API */
1478 struct kib_fast_reg_descriptor *frd, *tmp;
1481 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1483 list_del(&frd->frd_list);
1484 #ifndef HAVE_IB_MAP_MR_SG
1485 ib_free_fast_reg_page_list(frd->frd_frpl);
1487 ib_dereg_mr(frd->frd_mr);
1488 LIBCFS_FREE(frd, sizeof(*frd));
1491 if (i < fpo->fast_reg.fpo_pool_size)
1492 CERROR("FastReg pool still has %d regions registered\n",
1493 fpo->fast_reg.fpo_pool_size - i);
1497 kiblnd_hdev_decref(fpo->fpo_hdev);
1499 LIBCFS_FREE(fpo, sizeof(*fpo));
1503 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1505 struct kib_fmr_pool *fpo, *tmp;
1507 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1508 list_del(&fpo->fpo_list);
1509 kiblnd_destroy_fmr_pool(fpo);
1514 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1517 int size = tunables->lnd_fmr_pool_size / ncpts;
1519 return max(IBLND_FMR_POOL, size);
1523 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1526 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1528 return max(IBLND_FMR_POOL_FLUSH, size);
1531 #ifdef HAVE_FMR_POOL_API
1532 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1533 struct kib_fmr_pool *fpo)
1535 struct ib_fmr_pool_param param = {
1536 .max_pages_per_fmr = LNET_MAX_IOV,
1537 .page_shift = PAGE_SHIFT,
1538 .access = (IB_ACCESS_LOCAL_WRITE |
1539 IB_ACCESS_REMOTE_WRITE),
1540 .pool_size = fps->fps_pool_size,
1541 .dirty_watermark = fps->fps_flush_trigger,
1542 .flush_function = NULL,
1544 .cache = !!fps->fps_cache };
1547 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1549 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1550 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1552 CERROR("Failed to create FMR pool: %d\n", rc);
1554 CERROR("FMRs are not supported\n");
1556 fpo->fpo_is_fmr = true;
1560 #endif /* HAVE_FMR_POOL_API */
1562 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1563 struct kib_fmr_pool *fpo,
1564 enum kib_dev_caps dev_caps)
1566 struct kib_fast_reg_descriptor *frd, *tmp;
1569 #ifdef HAVE_FMR_POOL_API
1570 fpo->fpo_is_fmr = false;
1573 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1574 fpo->fast_reg.fpo_pool_size = 0;
1575 for (i = 0; i < fps->fps_pool_size; i++) {
1576 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1579 CERROR("Failed to allocate a new fast_reg descriptor\n");
1585 #ifndef HAVE_IB_MAP_MR_SG
1586 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1588 if (IS_ERR(frd->frd_frpl)) {
1589 rc = PTR_ERR(frd->frd_frpl);
1590 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1592 frd->frd_frpl = NULL;
1597 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1598 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1602 * it is expected to get here if this is an MLX-5 card.
1603 * MLX-4 cards will always use FMR and MLX-5 cards will
1604 * always use fast_reg. It turns out that some MLX-5 cards
1605 * (possibly due to older FW versions) do not natively support
1606 * gaps. So we will need to track them here.
1608 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1609 #ifdef IB_MR_TYPE_SG_GAPS
1610 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1611 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1612 IB_MR_TYPE_SG_GAPS :
1618 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1619 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1620 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1622 if (IS_ERR(frd->frd_mr)) {
1623 rc = PTR_ERR(frd->frd_mr);
1624 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1629 /* There appears to be a bug in MLX5 code where you must
1630 * invalidate the rkey of a new FastReg pool before first
1631 * using it. Thus, I am marking the FRD invalid here. */
1632 frd->frd_valid = false;
1634 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1635 fpo->fast_reg.fpo_pool_size++;
1642 ib_dereg_mr(frd->frd_mr);
1643 #ifndef HAVE_IB_MAP_MR_SG
1645 ib_free_fast_reg_page_list(frd->frd_frpl);
1647 LIBCFS_FREE(frd, sizeof(*frd));
1650 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1652 list_del(&frd->frd_list);
1653 #ifndef HAVE_IB_MAP_MR_SG
1654 ib_free_fast_reg_page_list(frd->frd_frpl);
1656 ib_dereg_mr(frd->frd_mr);
1657 LIBCFS_FREE(frd, sizeof(*frd));
1663 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1664 struct kib_fmr_pool **pp_fpo)
1666 struct kib_dev *dev = fps->fps_net->ibn_dev;
1667 struct kib_fmr_pool *fpo;
1670 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1674 memset(fpo, 0, sizeof(*fpo));
1676 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1678 #ifdef HAVE_FMR_POOL_API
1679 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1680 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1682 #endif /* HAVE_FMR_POOL_API */
1683 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1687 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1688 fpo->fpo_owner = fps;
1694 kiblnd_hdev_decref(fpo->fpo_hdev);
1695 LIBCFS_FREE(fpo, sizeof(*fpo));
1700 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1702 if (fps->fps_net == NULL) /* intialized? */
1705 spin_lock(&fps->fps_lock);
1707 while (!list_empty(&fps->fps_pool_list)) {
1708 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1709 struct kib_fmr_pool,
1712 fpo->fpo_failed = 1;
1713 list_del(&fpo->fpo_list);
1714 if (fpo->fpo_map_count == 0)
1715 list_add(&fpo->fpo_list, zombies);
1717 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1720 spin_unlock(&fps->fps_lock);
1724 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1726 if (fps->fps_net != NULL) { /* initialized? */
1727 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1728 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1733 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1734 struct kib_net *net,
1735 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1737 struct kib_fmr_pool *fpo;
1740 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1745 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1746 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1747 fps->fps_cache = tunables->lnd_fmr_cache;
1749 spin_lock_init(&fps->fps_lock);
1750 INIT_LIST_HEAD(&fps->fps_pool_list);
1751 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1753 rc = kiblnd_create_fmr_pool(fps, &fpo);
1755 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1761 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1763 if (fpo->fpo_map_count != 0) /* still in use */
1765 if (fpo->fpo_failed)
1767 return now >= fpo->fpo_deadline;
1770 #if defined(HAVE_FMR_POOL_API) || !defined(HAVE_IB_MAP_MR_SG)
1772 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1774 struct kib_hca_dev *hdev;
1775 __u64 *pages = tx->tx_pages;
1780 hdev = tx->tx_pool->tpo_hdev;
1782 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1783 for (size = 0; size < rd->rd_frags[i].rf_nob;
1784 size += hdev->ibh_page_size) {
1785 pages[npages++] = (rd->rd_frags[i].rf_addr &
1786 hdev->ibh_page_mask) + size;
1795 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1797 struct list_head zombies = LIST_HEAD_INIT(zombies);
1798 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1799 struct kib_fmr_poolset *fps;
1800 time64_t now = ktime_get_seconds();
1801 struct kib_fmr_pool *tmp;
1806 fps = fpo->fpo_owner;
1808 #ifdef HAVE_FMR_POOL_API
1809 if (fpo->fpo_is_fmr) {
1810 if (fmr->fmr_pfmr) {
1811 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1812 fmr->fmr_pfmr = NULL;
1816 int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1820 #endif /* HAVE_FMR_POOL_API */
1822 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1825 frd->frd_valid = false;
1826 fmr->fmr_frd = NULL;
1827 spin_lock(&fps->fps_lock);
1828 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1829 spin_unlock(&fps->fps_lock);
1832 fmr->fmr_pool = NULL;
1834 spin_lock(&fps->fps_lock);
1835 fpo->fpo_map_count--; /* decref the pool */
1837 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1838 /* the first pool is persistent */
1839 if (fps->fps_pool_list.next == &fpo->fpo_list)
1842 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1843 list_move(&fpo->fpo_list, &zombies);
1847 spin_unlock(&fps->fps_lock);
1849 if (!list_empty(&zombies))
1850 kiblnd_destroy_fmr_pool_list(&zombies);
1853 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1854 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1855 struct kib_fmr *fmr)
1857 struct kib_fmr_pool *fpo;
1859 bool is_rx = (rd != tx->tx_rd);
1860 #ifdef HAVE_FMR_POOL_API
1861 __u64 *pages = tx->tx_pages;
1862 bool tx_pages_mapped = 0;
1868 spin_lock(&fps->fps_lock);
1869 version = fps->fps_version;
1870 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1871 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1872 fpo->fpo_map_count++;
1874 #ifdef HAVE_FMR_POOL_API
1875 fmr->fmr_pfmr = NULL;
1876 if (fpo->fpo_is_fmr) {
1877 struct ib_pool_fmr *pfmr;
1879 spin_unlock(&fps->fps_lock);
1881 if (!tx_pages_mapped) {
1882 npages = kiblnd_map_tx_pages(tx, rd);
1883 tx_pages_mapped = 1;
1886 pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1887 pages, npages, iov);
1888 if (likely(!IS_ERR(pfmr))) {
1889 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1891 fmr->fmr_frd = NULL;
1892 fmr->fmr_pfmr = pfmr;
1893 fmr->fmr_pool = fpo;
1898 #endif /* HAVE_FMR_POOL_API */
1900 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1901 struct kib_fast_reg_descriptor *frd;
1902 #ifdef HAVE_IB_MAP_MR_SG
1903 struct ib_reg_wr *wr;
1906 struct ib_rdma_wr *wr;
1907 struct ib_fast_reg_page_list *frpl;
1911 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1912 struct kib_fast_reg_descriptor,
1914 list_del(&frd->frd_list);
1915 spin_unlock(&fps->fps_lock);
1917 #ifndef HAVE_IB_MAP_MR_SG
1918 frpl = frd->frd_frpl;
1922 if (!frd->frd_valid) {
1923 struct ib_rdma_wr *inv_wr;
1924 __u32 key = is_rx ? mr->rkey : mr->lkey;
1926 inv_wr = &frd->frd_inv_wr;
1927 memset(inv_wr, 0, sizeof(*inv_wr));
1929 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1930 inv_wr->wr.wr_id = IBLND_WID_MR;
1931 inv_wr->wr.ex.invalidate_rkey = key;
1934 key = ib_inc_rkey(key);
1935 ib_update_fast_reg_key(mr, key);
1938 #ifdef HAVE_IB_MAP_MR_SG
1939 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1940 n = ib_map_mr_sg(mr, tx->tx_frags,
1941 rd->rd_nfrags, NULL, PAGE_SIZE);
1943 n = ib_map_mr_sg(mr, tx->tx_frags,
1944 rd->rd_nfrags, PAGE_SIZE);
1945 #endif /* HAVE_IB_MAP_MR_SG_5ARGS */
1946 if (unlikely(n != rd->rd_nfrags)) {
1947 CERROR("Failed to map mr %d/%d "
1948 "elements\n", n, rd->rd_nfrags);
1949 return n < 0 ? n : -EINVAL;
1952 wr = &frd->frd_fastreg_wr;
1953 memset(wr, 0, sizeof(*wr));
1955 wr->wr.opcode = IB_WR_REG_MR;
1956 wr->wr.wr_id = IBLND_WID_MR;
1958 wr->wr.send_flags = 0;
1960 wr->key = is_rx ? mr->rkey : mr->lkey;
1961 wr->access = (IB_ACCESS_LOCAL_WRITE |
1962 IB_ACCESS_REMOTE_WRITE);
1963 #else /* HAVE_IB_MAP_MR_SG */
1964 if (!tx_pages_mapped) {
1965 npages = kiblnd_map_tx_pages(tx, rd);
1966 tx_pages_mapped = 1;
1969 LASSERT(npages <= frpl->max_page_list_len);
1970 memcpy(frpl->page_list, pages,
1971 sizeof(*pages) * npages);
1973 /* Prepare FastReg WR */
1974 wr = &frd->frd_fastreg_wr;
1975 memset(wr, 0, sizeof(*wr));
1977 wr->wr.opcode = IB_WR_FAST_REG_MR;
1978 wr->wr.wr_id = IBLND_WID_MR;
1980 wr->wr.wr.fast_reg.iova_start = iov;
1981 wr->wr.wr.fast_reg.page_list = frpl;
1982 wr->wr.wr.fast_reg.page_list_len = npages;
1983 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1984 wr->wr.wr.fast_reg.length = nob;
1985 wr->wr.wr.fast_reg.rkey =
1986 is_rx ? mr->rkey : mr->lkey;
1987 wr->wr.wr.fast_reg.access_flags =
1988 (IB_ACCESS_LOCAL_WRITE |
1989 IB_ACCESS_REMOTE_WRITE);
1990 #endif /* HAVE_IB_MAP_MR_SG */
1992 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1994 fmr->fmr_pool = fpo;
1997 spin_unlock(&fps->fps_lock);
2001 spin_lock(&fps->fps_lock);
2002 fpo->fpo_map_count--;
2003 if (rc != -EAGAIN) {
2004 spin_unlock(&fps->fps_lock);
2008 /* EAGAIN and ... */
2009 if (version != fps->fps_version) {
2010 spin_unlock(&fps->fps_lock);
2015 if (fps->fps_increasing) {
2016 spin_unlock(&fps->fps_lock);
2017 CDEBUG(D_NET, "Another thread is allocating new "
2018 "FMR pool, waiting for her to complete\n");
2024 if (ktime_get_seconds() < fps->fps_next_retry) {
2025 /* someone failed recently */
2026 spin_unlock(&fps->fps_lock);
2030 fps->fps_increasing = 1;
2031 spin_unlock(&fps->fps_lock);
2033 CDEBUG(D_NET, "Allocate new FMR pool\n");
2034 rc = kiblnd_create_fmr_pool(fps, &fpo);
2035 spin_lock(&fps->fps_lock);
2036 fps->fps_increasing = 0;
2039 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2041 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2043 spin_unlock(&fps->fps_lock);
2049 kiblnd_fini_pool(struct kib_pool *pool)
2051 LASSERT(list_empty(&pool->po_free_list));
2052 LASSERT(pool->po_allocated == 0);
2054 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2058 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2060 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2062 memset(pool, 0, sizeof(struct kib_pool));
2063 INIT_LIST_HEAD(&pool->po_free_list);
2064 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2065 pool->po_owner = ps;
2066 pool->po_size = size;
2070 kiblnd_destroy_pool_list(struct list_head *head)
2072 struct kib_pool *pool;
2074 while (!list_empty(head)) {
2075 pool = list_entry(head->next, struct kib_pool, po_list);
2076 list_del(&pool->po_list);
2078 LASSERT(pool->po_owner != NULL);
2079 pool->po_owner->ps_pool_destroy(pool);
2084 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2086 if (ps->ps_net == NULL) /* intialized? */
2089 spin_lock(&ps->ps_lock);
2090 while (!list_empty(&ps->ps_pool_list)) {
2091 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
2092 struct kib_pool, po_list);
2095 list_del(&po->po_list);
2096 if (po->po_allocated == 0)
2097 list_add(&po->po_list, zombies);
2099 list_add(&po->po_list, &ps->ps_failed_pool_list);
2101 spin_unlock(&ps->ps_lock);
2105 kiblnd_fini_poolset(struct kib_poolset *ps)
2107 if (ps->ps_net != NULL) { /* initialized? */
2108 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2109 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2114 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2115 struct kib_net *net, char *name, int size,
2116 kib_ps_pool_create_t po_create,
2117 kib_ps_pool_destroy_t po_destroy,
2118 kib_ps_node_init_t nd_init,
2119 kib_ps_node_fini_t nd_fini)
2121 struct kib_pool *pool;
2124 memset(ps, 0, sizeof(struct kib_poolset));
2128 ps->ps_pool_create = po_create;
2129 ps->ps_pool_destroy = po_destroy;
2130 ps->ps_node_init = nd_init;
2131 ps->ps_node_fini = nd_fini;
2132 ps->ps_pool_size = size;
2133 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2134 >= sizeof(ps->ps_name))
2136 spin_lock_init(&ps->ps_lock);
2137 INIT_LIST_HEAD(&ps->ps_pool_list);
2138 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2140 rc = ps->ps_pool_create(ps, size, &pool);
2142 list_add(&pool->po_list, &ps->ps_pool_list);
2144 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2150 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2152 if (pool->po_allocated != 0) /* still in use */
2154 if (pool->po_failed)
2156 return now >= pool->po_deadline;
2160 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2162 struct list_head zombies = LIST_HEAD_INIT(zombies);
2163 struct kib_poolset *ps = pool->po_owner;
2164 struct kib_pool *tmp;
2165 time64_t now = ktime_get_seconds();
2167 spin_lock(&ps->ps_lock);
2169 if (ps->ps_node_fini != NULL)
2170 ps->ps_node_fini(pool, node);
2172 LASSERT(pool->po_allocated > 0);
2173 list_add(node, &pool->po_free_list);
2174 pool->po_allocated--;
2176 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2177 /* the first pool is persistent */
2178 if (ps->ps_pool_list.next == &pool->po_list)
2181 if (kiblnd_pool_is_idle(pool, now))
2182 list_move(&pool->po_list, &zombies);
2184 spin_unlock(&ps->ps_lock);
2186 if (!list_empty(&zombies))
2187 kiblnd_destroy_pool_list(&zombies);
2191 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2193 struct list_head *node;
2194 struct kib_pool *pool;
2196 unsigned int interval = 1;
2197 ktime_t time_before;
2198 unsigned int trips = 0;
2201 spin_lock(&ps->ps_lock);
2202 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2203 if (list_empty(&pool->po_free_list))
2206 pool->po_allocated++;
2207 pool->po_deadline = ktime_get_seconds() +
2208 IBLND_POOL_DEADLINE;
2209 node = pool->po_free_list.next;
2212 if (ps->ps_node_init != NULL) {
2213 /* still hold the lock */
2214 ps->ps_node_init(pool, node);
2216 spin_unlock(&ps->ps_lock);
2220 /* no available tx pool and ... */
2221 if (ps->ps_increasing) {
2222 /* another thread is allocating a new pool */
2223 spin_unlock(&ps->ps_lock);
2225 CDEBUG(D_NET, "Another thread is allocating new "
2226 "%s pool, waiting %d HZs for her to complete."
2228 ps->ps_name, interval, trips);
2230 set_current_state(TASK_INTERRUPTIBLE);
2231 schedule_timeout(interval);
2232 if (interval < cfs_time_seconds(1))
2238 if (ktime_get_seconds() < ps->ps_next_retry) {
2239 /* someone failed recently */
2240 spin_unlock(&ps->ps_lock);
2244 ps->ps_increasing = 1;
2245 spin_unlock(&ps->ps_lock);
2247 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2248 time_before = ktime_get();
2249 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2250 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2251 ktime_ms_delta(ktime_get(), time_before));
2253 spin_lock(&ps->ps_lock);
2254 ps->ps_increasing = 0;
2256 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2258 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2259 CERROR("Can't allocate new %s pool because out of memory\n",
2262 spin_unlock(&ps->ps_lock);
2268 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2270 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2274 LASSERT (pool->po_allocated == 0);
2276 if (tpo->tpo_tx_pages != NULL) {
2277 kiblnd_unmap_tx_pool(tpo);
2278 kiblnd_free_pages(tpo->tpo_tx_pages);
2281 if (tpo->tpo_tx_descs == NULL)
2284 for (i = 0; i < pool->po_size; i++) {
2285 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2286 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2288 list_del(&tx->tx_list);
2289 if (tx->tx_pages != NULL)
2290 LIBCFS_FREE(tx->tx_pages,
2292 sizeof(*tx->tx_pages));
2293 if (tx->tx_frags != NULL)
2294 LIBCFS_FREE(tx->tx_frags,
2295 (1 + IBLND_MAX_RDMA_FRAGS) *
2296 sizeof(*tx->tx_frags));
2297 if (tx->tx_wrq != NULL)
2298 LIBCFS_FREE(tx->tx_wrq,
2299 (1 + IBLND_MAX_RDMA_FRAGS) *
2300 sizeof(*tx->tx_wrq));
2301 if (tx->tx_sge != NULL)
2302 LIBCFS_FREE(tx->tx_sge,
2303 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2304 sizeof(*tx->tx_sge));
2305 if (tx->tx_rd != NULL)
2306 LIBCFS_FREE(tx->tx_rd,
2307 offsetof(struct kib_rdma_desc,
2308 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2311 LIBCFS_FREE(tpo->tpo_tx_descs,
2312 pool->po_size * sizeof(struct kib_tx));
2314 kiblnd_fini_pool(pool);
2315 LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool));
2318 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2320 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2323 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2324 ntx = tunables->lnd_ntx / ncpts;
2326 return max(IBLND_TX_POOL, ntx);
2330 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2334 struct kib_pool *pool;
2335 struct kib_tx_pool *tpo;
2337 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2339 CERROR("Failed to allocate TX pool\n");
2343 pool = &tpo->tpo_pool;
2344 kiblnd_init_pool(ps, pool, size);
2345 tpo->tpo_tx_descs = NULL;
2346 tpo->tpo_tx_pages = NULL;
2348 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2349 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2350 CERROR("Can't allocate tx pages: %d\n", npg);
2351 LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool));
2355 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2356 size * sizeof(struct kib_tx));
2357 if (tpo->tpo_tx_descs == NULL) {
2358 CERROR("Can't allocate %d tx descriptors\n", size);
2359 ps->ps_pool_destroy(pool);
2363 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2365 for (i = 0; i < size; i++) {
2366 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2367 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2370 if (ps->ps_net->ibn_fmr_ps != NULL) {
2371 LIBCFS_CPT_ALLOC(tx->tx_pages,
2372 lnet_cpt_table(), ps->ps_cpt,
2373 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2374 if (tx->tx_pages == NULL)
2378 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2379 (1 + IBLND_MAX_RDMA_FRAGS) *
2380 sizeof(*tx->tx_frags));
2381 if (tx->tx_frags == NULL)
2384 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2386 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2387 (1 + IBLND_MAX_RDMA_FRAGS) *
2388 sizeof(*tx->tx_wrq));
2389 if (tx->tx_wrq == NULL)
2392 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2393 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2394 sizeof(*tx->tx_sge));
2395 if (tx->tx_sge == NULL)
2398 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2399 offsetof(struct kib_rdma_desc,
2400 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2401 if (tx->tx_rd == NULL)
2406 kiblnd_map_tx_pool(tpo);
2411 ps->ps_pool_destroy(pool);
2416 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2418 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2419 struct kib_tx_poolset,
2421 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2423 tx->tx_cookie = tps->tps_next_tx_cookie++;
2427 kiblnd_net_fini_pools(struct kib_net *net)
2431 cfs_cpt_for_each(i, lnet_cpt_table()) {
2432 struct kib_tx_poolset *tps;
2433 struct kib_fmr_poolset *fps;
2435 if (net->ibn_tx_ps != NULL) {
2436 tps = net->ibn_tx_ps[i];
2437 kiblnd_fini_poolset(&tps->tps_poolset);
2440 if (net->ibn_fmr_ps != NULL) {
2441 fps = net->ibn_fmr_ps[i];
2442 kiblnd_fini_fmr_poolset(fps);
2446 if (net->ibn_tx_ps != NULL) {
2447 cfs_percpt_free(net->ibn_tx_ps);
2448 net->ibn_tx_ps = NULL;
2451 if (net->ibn_fmr_ps != NULL) {
2452 cfs_percpt_free(net->ibn_fmr_ps);
2453 net->ibn_fmr_ps = NULL;
2458 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2461 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2462 #ifdef HAVE_IB_GET_DMA_MR
2463 unsigned long flags;
2469 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2471 #ifdef HAVE_IB_GET_DMA_MR
2472 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2474 * if lnd_map_on_demand is zero then we have effectively disabled
2475 * FMR or FastReg and we're using global memory regions
2478 if (!tunables->lnd_map_on_demand) {
2479 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2481 goto create_tx_pool;
2484 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2487 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2488 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2489 tunables->lnd_fmr_pool_size,
2490 tunables->lnd_ntx / 4);
2495 /* TX pool must be created later than FMR, see LU-2268
2497 LASSERT(net->ibn_tx_ps == NULL);
2499 /* premapping can fail if ibd_nmr > 1, so we always create
2500 * FMR pool and map-on-demand if premapping failed */
2502 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2503 sizeof(struct kib_fmr_poolset));
2504 if (net->ibn_fmr_ps == NULL) {
2505 CERROR("Failed to allocate FMR pool array\n");
2510 for (i = 0; i < ncpts; i++) {
2511 cpt = (cpts == NULL) ? i : cpts[i];
2512 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2515 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2522 LASSERT(i == ncpts);
2524 #ifdef HAVE_IB_GET_DMA_MR
2527 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2528 sizeof(struct kib_tx_poolset));
2529 if (net->ibn_tx_ps == NULL) {
2530 CERROR("Failed to allocate tx pool array\n");
2535 for (i = 0; i < ncpts; i++) {
2536 cpt = (cpts == NULL) ? i : cpts[i];
2537 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2539 kiblnd_tx_pool_size(ni, ncpts),
2540 kiblnd_create_tx_pool,
2541 kiblnd_destroy_tx_pool,
2542 kiblnd_tx_init, NULL);
2544 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2552 kiblnd_net_fini_pools(net);
2558 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2560 struct ib_device_attr *dev_attr;
2563 /* It's safe to assume a HCA can handle a page size
2564 * matching that of the native system */
2565 hdev->ibh_page_shift = PAGE_SHIFT;
2566 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2567 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2569 #ifndef HAVE_IB_DEVICE_ATTRS
2570 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2571 if (dev_attr == NULL) {
2572 CERROR("Out of memory\n");
2576 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2578 CERROR("Failed to query IB device: %d\n", rc);
2579 goto out_clean_attr;
2582 dev_attr = &hdev->ibh_ibdev->attrs;
2585 hdev->ibh_mr_size = dev_attr->max_mr_size;
2586 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2588 /* Setup device Memory Registration capabilities */
2589 #ifdef HAVE_FMR_POOL_API
2590 #ifdef HAVE_IB_DEVICE_OPS
2591 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2592 hdev->ibh_ibdev->ops.dealloc_fmr &&
2593 hdev->ibh_ibdev->ops.map_phys_fmr &&
2594 hdev->ibh_ibdev->ops.unmap_fmr) {
2596 if (hdev->ibh_ibdev->alloc_fmr &&
2597 hdev->ibh_ibdev->dealloc_fmr &&
2598 hdev->ibh_ibdev->map_phys_fmr &&
2599 hdev->ibh_ibdev->unmap_fmr) {
2601 LCONSOLE_INFO("Using FMR for registration\n");
2602 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2604 #endif /* HAVE_FMR_POOL_API */
2605 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2606 LCONSOLE_INFO("Using FastReg for registration\n");
2607 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2608 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2609 #ifdef IB_DEVICE_SG_GAPS_REG
2610 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2611 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2618 if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2619 hdev->ibh_mr_shift = 64;
2623 #ifndef HAVE_IB_DEVICE_ATTRS
2625 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2629 CERROR("IB device does not support FMRs nor FastRegs, can't "
2630 "register memory: %d\n", rc);
2631 else if (rc == -EINVAL)
2632 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2636 #ifdef HAVE_IB_GET_DMA_MR
2638 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2640 if (hdev->ibh_mrs == NULL)
2643 ib_dereg_mr(hdev->ibh_mrs);
2645 hdev->ibh_mrs = NULL;
2650 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2652 #ifdef HAVE_IB_GET_DMA_MR
2653 kiblnd_hdev_cleanup_mrs(hdev);
2656 if (hdev->ibh_pd != NULL)
2657 ib_dealloc_pd(hdev->ibh_pd);
2659 if (hdev->ibh_cmid != NULL)
2660 rdma_destroy_id(hdev->ibh_cmid);
2662 LIBCFS_FREE(hdev, sizeof(*hdev));
2665 #ifdef HAVE_IB_GET_DMA_MR
2667 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2670 int acflags = IB_ACCESS_LOCAL_WRITE |
2671 IB_ACCESS_REMOTE_WRITE;
2673 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2675 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2676 kiblnd_hdev_cleanup_mrs(hdev);
2687 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2693 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2695 struct rdma_cm_id *cmid;
2696 struct sockaddr_in srcaddr;
2697 struct sockaddr_in dstaddr;
2700 if (dev->ibd_hdev == NULL || /* initializing */
2701 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2702 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2705 /* XXX: it's UGLY, but I don't have better way to find
2706 * ib-bonding HCA failover because:
2708 * a. no reliable CM event for HCA failover...
2709 * b. no OFED API to get ib_device for current net_device...
2711 * We have only two choices at this point:
2713 * a. rdma_bind_addr(), it will conflict with listener cmid
2714 * b. rdma_resolve_addr() to zero addr */
2715 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2716 RDMA_PS_TCP, IB_QPT_RC);
2719 CERROR("Failed to create cmid for failover: %d\n", rc);
2723 memset(&srcaddr, 0, sizeof(srcaddr));
2724 srcaddr.sin_family = AF_INET;
2725 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2727 memset(&dstaddr, 0, sizeof(dstaddr));
2728 dstaddr.sin_family = AF_INET;
2729 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2730 (struct sockaddr *)&dstaddr, 1);
2731 if (rc != 0 || cmid->device == NULL) {
2732 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2733 dev->ibd_ifname, &dev->ibd_ifip,
2735 rdma_destroy_id(cmid);
2739 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2740 rdma_destroy_id(cmid);
2745 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2747 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2748 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2749 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2750 struct rdma_cm_id *cmid = NULL;
2751 struct kib_hca_dev *hdev = NULL;
2752 struct kib_hca_dev *old;
2754 struct kib_net *net;
2755 struct sockaddr_in addr;
2756 unsigned long flags;
2760 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2761 dev->ibd_can_failover ||
2762 dev->ibd_hdev == NULL);
2764 rc = kiblnd_dev_need_failover(dev, ns);
2768 if (dev->ibd_hdev != NULL &&
2769 dev->ibd_hdev->ibh_cmid != NULL) {
2770 /* XXX it's not good to close old listener at here,
2771 * because we can fail to create new listener.
2772 * But we have to close it now, otherwise rdma_bind_addr
2773 * will return EADDRINUSE... How crap! */
2774 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2776 cmid = dev->ibd_hdev->ibh_cmid;
2777 /* make next schedule of kiblnd_dev_need_failover()
2778 * return 1 for me */
2779 dev->ibd_hdev->ibh_cmid = NULL;
2780 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2782 rdma_destroy_id(cmid);
2785 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2789 CERROR("Failed to create cmid for failover: %d\n", rc);
2793 memset(&addr, 0, sizeof(addr));
2794 addr.sin_family = AF_INET;
2795 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2796 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2798 /* Bind to failover device or port */
2799 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2800 if (rc != 0 || cmid->device == NULL) {
2801 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2802 dev->ibd_ifname, &dev->ibd_ifip,
2804 rdma_destroy_id(cmid);
2808 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2810 CERROR("Failed to allocate kib_hca_dev\n");
2811 rdma_destroy_id(cmid);
2816 atomic_set(&hdev->ibh_ref, 1);
2817 hdev->ibh_dev = dev;
2818 hdev->ibh_cmid = cmid;
2819 hdev->ibh_ibdev = cmid->device;
2821 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2822 pd = ib_alloc_pd(cmid->device, 0);
2824 pd = ib_alloc_pd(cmid->device);
2828 CERROR("Can't allocate PD: %d\n", rc);
2834 rc = rdma_listen(cmid, 0);
2836 CERROR("Can't start new listener: %d\n", rc);
2840 rc = kiblnd_hdev_get_attr(hdev);
2842 CERROR("Can't get device attributes: %d\n", rc);
2846 #ifdef HAVE_IB_GET_DMA_MR
2847 rc = kiblnd_hdev_setup_mrs(hdev);
2849 CERROR("Can't setup device: %d\n", rc);
2854 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2856 old = dev->ibd_hdev;
2857 dev->ibd_hdev = hdev; /* take over the refcount */
2860 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2861 cfs_cpt_for_each(i, lnet_cpt_table()) {
2862 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2865 if (net->ibn_fmr_ps != NULL)
2866 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2871 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2873 if (!list_empty(&zombie_tpo))
2874 kiblnd_destroy_pool_list(&zombie_tpo);
2875 if (!list_empty(&zombie_ppo))
2876 kiblnd_destroy_pool_list(&zombie_ppo);
2877 if (!list_empty(&zombie_fpo))
2878 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2880 kiblnd_hdev_decref(hdev);
2883 dev->ibd_failed_failover++;
2885 dev->ibd_failed_failover = 0;
2891 kiblnd_destroy_dev(struct kib_dev *dev)
2893 LASSERT(dev->ibd_nnets == 0);
2894 LASSERT(list_empty(&dev->ibd_nets));
2896 list_del(&dev->ibd_fail_list);
2897 list_del(&dev->ibd_list);
2899 if (dev->ibd_hdev != NULL)
2900 kiblnd_hdev_decref(dev->ibd_hdev);
2902 LIBCFS_FREE(dev, sizeof(*dev));
2906 kiblnd_base_shutdown(void)
2908 struct kib_sched_info *sched;
2911 LASSERT(list_empty(&kiblnd_data.kib_devs));
2913 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2914 atomic_read(&libcfs_kmemory));
2916 switch (kiblnd_data.kib_init) {
2920 case IBLND_INIT_ALL:
2921 case IBLND_INIT_DATA:
2922 LASSERT (kiblnd_data.kib_peers != NULL);
2923 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2924 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2926 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2927 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2928 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2929 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2931 /* flag threads to terminate; wake and wait for them to die */
2932 kiblnd_data.kib_shutdown = 1;
2934 /* NB: we really want to stop scheduler threads net by net
2935 * instead of the whole module, this should be improved
2936 * with dynamic configuration LNet */
2937 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2938 wake_up_all(&sched->ibs_waitq);
2940 wake_up_all(&kiblnd_data.kib_connd_waitq);
2941 wake_up_all(&kiblnd_data.kib_failover_waitq);
2944 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2947 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2948 "Waiting for %d threads to terminate\n",
2949 atomic_read(&kiblnd_data.kib_nthreads));
2950 set_current_state(TASK_UNINTERRUPTIBLE);
2951 schedule_timeout(cfs_time_seconds(1));
2956 case IBLND_INIT_NOTHING:
2960 if (kiblnd_data.kib_peers != NULL) {
2961 LIBCFS_FREE(kiblnd_data.kib_peers,
2962 sizeof(struct list_head) *
2963 kiblnd_data.kib_peer_hash_size);
2966 if (kiblnd_data.kib_scheds != NULL)
2967 cfs_percpt_free(kiblnd_data.kib_scheds);
2969 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2970 atomic_read(&libcfs_kmemory));
2972 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2973 module_put(THIS_MODULE);
2977 kiblnd_shutdown(struct lnet_ni *ni)
2979 struct kib_net *net = ni->ni_data;
2980 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2982 unsigned long flags;
2984 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2989 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2990 atomic_read(&libcfs_kmemory));
2992 write_lock_irqsave(g_lock, flags);
2993 net->ibn_shutdown = 1;
2994 write_unlock_irqrestore(g_lock, flags);
2996 switch (net->ibn_init) {
3000 case IBLND_INIT_ALL:
3001 /* nuke all existing peers within this net */
3002 kiblnd_del_peer(ni, LNET_NID_ANY);
3004 /* Wait for all peer_ni state to clean up */
3006 while (atomic_read(&net->ibn_npeers) != 0) {
3009 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3010 "%s: waiting for %d peers to disconnect\n",
3011 libcfs_nid2str(ni->ni_nid),
3012 atomic_read(&net->ibn_npeers));
3013 set_current_state(TASK_UNINTERRUPTIBLE);
3014 schedule_timeout(cfs_time_seconds(1));
3017 kiblnd_net_fini_pools(net);
3019 write_lock_irqsave(g_lock, flags);
3020 LASSERT(net->ibn_dev->ibd_nnets > 0);
3021 net->ibn_dev->ibd_nnets--;
3022 list_del(&net->ibn_list);
3023 write_unlock_irqrestore(g_lock, flags);
3027 case IBLND_INIT_NOTHING:
3028 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3030 if (net->ibn_dev != NULL &&
3031 net->ibn_dev->ibd_nnets == 0)
3032 kiblnd_destroy_dev(net->ibn_dev);
3037 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3038 atomic_read(&libcfs_kmemory));
3040 net->ibn_init = IBLND_INIT_NOTHING;
3043 LIBCFS_FREE(net, sizeof(*net));
3046 if (list_empty(&kiblnd_data.kib_devs))
3047 kiblnd_base_shutdown();
3052 kiblnd_base_startup(struct net *ns)
3054 struct kib_sched_info *sched;
3058 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3060 try_module_get(THIS_MODULE);
3061 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3063 rwlock_init(&kiblnd_data.kib_global_lock);
3065 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3066 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3068 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3069 LIBCFS_ALLOC(kiblnd_data.kib_peers,
3070 sizeof(struct list_head) *
3071 kiblnd_data.kib_peer_hash_size);
3072 if (kiblnd_data.kib_peers == NULL)
3075 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3076 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3078 spin_lock_init(&kiblnd_data.kib_connd_lock);
3079 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3080 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3081 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3082 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3084 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3085 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3087 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3089 if (kiblnd_data.kib_scheds == NULL)
3092 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3095 spin_lock_init(&sched->ibs_lock);
3096 INIT_LIST_HEAD(&sched->ibs_conns);
3097 init_waitqueue_head(&sched->ibs_waitq);
3099 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3100 if (*kiblnd_tunables.kib_nscheds > 0) {
3101 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3103 /* max to half of CPUs, another half is reserved for
3104 * upper layer modules */
3105 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3108 sched->ibs_nthreads_max = nthrs;
3112 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3114 /* lists/ptrs/locks initialised */
3115 kiblnd_data.kib_init = IBLND_INIT_DATA;
3116 /*****************************************************/
3118 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3120 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3124 if (*kiblnd_tunables.kib_dev_failover != 0)
3125 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3129 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3133 /* flag everything initialised */
3134 kiblnd_data.kib_init = IBLND_INIT_ALL;
3135 /*****************************************************/
3140 kiblnd_base_shutdown();
3145 kiblnd_start_schedulers(struct kib_sched_info *sched)
3151 if (sched->ibs_nthreads == 0) {
3152 if (*kiblnd_tunables.kib_nscheds > 0) {
3153 nthrs = sched->ibs_nthreads_max;
3155 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3157 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3158 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3161 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3162 /* increase one thread if there is new interface */
3163 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3166 for (i = 0; i < nthrs; i++) {
3169 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3170 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3171 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3172 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3176 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3177 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3181 sched->ibs_nthreads += i;
3185 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3192 for (i = 0; i < ncpts; i++) {
3193 struct kib_sched_info *sched;
3195 cpt = (cpts == NULL) ? i : cpts[i];
3196 sched = kiblnd_data.kib_scheds[cpt];
3198 if (!newdev && sched->ibs_nthreads > 0)
3201 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3203 CERROR("Failed to start scheduler threads for %s\n",
3211 static struct kib_dev *
3212 kiblnd_dev_search(char *ifname)
3214 struct kib_dev *alias = NULL;
3215 struct kib_dev *dev;
3219 colon = strchr(ifname, ':');
3220 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3221 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3227 colon2 = strchr(dev->ibd_ifname, ':');
3233 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3245 kiblnd_startup(struct lnet_ni *ni)
3247 char *ifname = NULL;
3248 struct lnet_inetdev *ifaces = NULL;
3249 struct kib_dev *ibdev = NULL;
3250 struct kib_net *net = NULL;
3251 unsigned long flags;
3256 LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3258 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3259 rc = kiblnd_base_startup(ni->ni_net_ns);
3264 LIBCFS_ALLOC(net, sizeof(*net));
3271 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3273 kiblnd_tunables_setup(ni);
3276 * ni_interfaces is only to support legacy pre Multi-Rail
3277 * tcp bonding for ksocklnd. Multi-Rail wants each secondary
3278 * IP to be treated as an unique 'struct ni' interfaces instead.
3280 if (ni->ni_interfaces[0] != NULL) {
3281 /* Use the IPoIB interface specified in 'networks=' */
3282 if (ni->ni_interfaces[1] != NULL) {
3283 CERROR("ko2iblnd: Multiple interfaces not supported\n");
3288 ifname = ni->ni_interfaces[0];
3290 ifname = *kiblnd_tunables.kib_default_ipif;
3293 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3294 CERROR("IPoIB interface name too long: %s\n", ifname);
3299 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3303 for (i = 0; i < rc; i++) {
3304 if (strcmp(ifname, ifaces[i].li_name) == 0)
3309 CERROR("ko2iblnd: No matching interfaces\n");
3314 ibdev = kiblnd_dev_search(ifname);
3315 newdev = ibdev == NULL;
3316 /* hmm...create kib_dev even for alias */
3317 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3318 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3324 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3325 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3326 sizeof(ibdev->ibd_ifname));
3327 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3329 INIT_LIST_HEAD(&ibdev->ibd_nets);
3330 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3331 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3333 /* initialize the device */
3334 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3336 CERROR("ko2iblnd: Can't initialize device: rc = %d\n", rc);
3340 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3343 net->ibn_dev = ibdev;
3344 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3346 ni->ni_dev_cpt = ifaces[i].li_cpt;
3348 rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3352 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3354 CERROR("Failed to initialize NI pools: %d\n", rc);
3358 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3360 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3361 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3363 net->ibn_init = IBLND_INIT_ALL;
3368 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3369 kiblnd_destroy_dev(ibdev);
3372 kiblnd_shutdown(ni);
3374 CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3375 ifname ? ifname : "", rc);
3380 static struct lnet_lnd the_o2iblnd = {
3381 .lnd_type = O2IBLND,
3382 .lnd_startup = kiblnd_startup,
3383 .lnd_shutdown = kiblnd_shutdown,
3384 .lnd_ctl = kiblnd_ctl,
3385 .lnd_query = kiblnd_query,
3386 .lnd_send = kiblnd_send,
3387 .lnd_recv = kiblnd_recv,
3390 static void __exit ko2iblnd_exit(void)
3392 lnet_unregister_lnd(&the_o2iblnd);
3395 static int __init ko2iblnd_init(void)
3399 CLASSERT(sizeof(struct kib_msg) <= IBLND_MSG_SIZE);
3400 CLASSERT(offsetof(struct kib_msg,
3401 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3403 CLASSERT(offsetof(struct kib_msg,
3404 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3407 rc = kiblnd_tunables_init();
3411 lnet_register_lnd(&the_o2iblnd);
3416 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3417 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3418 MODULE_VERSION("2.8.0");
3419 MODULE_LICENSE("GPL");
3421 module_init(ko2iblnd_init);
3422 module_exit(ko2iblnd_exit);