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
737 unsigned int ret = 1 + conn->ibc_max_frags;
738 enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
740 /* FastReg needs two extra WRs for map and invalidate */
741 if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
744 /* account for a maximum of ibc_queue_depth in-flight transfers */
745 ret *= conn->ibc_queue_depth;
750 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
751 int state, int version)
754 * If the new conn is created successfully it takes over the caller's
755 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
756 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
757 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
758 * to destroy 'cmid' here since I'm called from the CM which still has
759 * its ref on 'cmid'). */
760 rwlock_t *glock = &kiblnd_data.kib_global_lock;
761 struct kib_net *net = peer_ni->ibp_ni->ni_data;
763 struct ib_qp_init_attr *init_qp_attr;
764 struct kib_sched_info *sched;
765 #ifdef HAVE_IB_CQ_INIT_ATTR
766 struct ib_cq_init_attr cq_attr = {};
768 struct kib_conn *conn;
775 LASSERT(net != NULL);
776 LASSERT(!in_interrupt());
780 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
781 sched = kiblnd_get_scheduler(cpt);
784 CERROR("no schedulers available. node is unhealthy\n");
789 * The cpt might have changed if we ended up selecting a non cpt
790 * native scheduler. So use the scheduler's cpt instead.
792 cpt = sched->ibs_cpt;
794 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
795 sizeof(*init_qp_attr));
796 if (init_qp_attr == NULL) {
797 CERROR("Can't allocate qp_attr for %s\n",
798 libcfs_nid2str(peer_ni->ibp_nid));
802 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
804 CERROR("Can't allocate connection for %s\n",
805 libcfs_nid2str(peer_ni->ibp_nid));
809 conn->ibc_state = IBLND_CONN_INIT;
810 conn->ibc_version = version;
811 conn->ibc_peer = peer_ni; /* I take the caller's ref */
812 cmid->context = conn; /* for future CM callbacks */
813 conn->ibc_cmid = cmid;
814 conn->ibc_max_frags = peer_ni->ibp_max_frags;
815 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
816 conn->ibc_rxs = NULL;
817 conn->ibc_rx_pages = NULL;
819 INIT_LIST_HEAD(&conn->ibc_early_rxs);
820 INIT_LIST_HEAD(&conn->ibc_tx_noops);
821 INIT_LIST_HEAD(&conn->ibc_tx_queue);
822 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
823 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
824 INIT_LIST_HEAD(&conn->ibc_active_txs);
825 INIT_LIST_HEAD(&conn->ibc_zombie_txs);
826 spin_lock_init(&conn->ibc_lock);
828 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
829 sizeof(*conn->ibc_connvars));
830 if (conn->ibc_connvars == NULL) {
831 CERROR("Can't allocate in-progress connection state\n");
835 write_lock_irqsave(glock, flags);
836 if (dev->ibd_failover) {
837 write_unlock_irqrestore(glock, flags);
838 CERROR("%s: failover in progress\n", dev->ibd_ifname);
842 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
843 /* wakeup failover thread and teardown connection */
844 if (kiblnd_dev_can_failover(dev)) {
845 list_add_tail(&dev->ibd_fail_list,
846 &kiblnd_data.kib_failed_devs);
847 wake_up(&kiblnd_data.kib_failover_waitq);
850 write_unlock_irqrestore(glock, flags);
851 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
852 cmid->device->name, dev->ibd_ifname);
856 kiblnd_hdev_addref_locked(dev->ibd_hdev);
857 conn->ibc_hdev = dev->ibd_hdev;
859 kiblnd_setup_mtu_locked(cmid);
861 write_unlock_irqrestore(glock, flags);
863 #ifdef HAVE_IB_CQ_INIT_ATTR
864 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
865 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
866 cq = ib_create_cq(cmid->device,
867 kiblnd_cq_completion, kiblnd_cq_event, conn,
870 cq = ib_create_cq(cmid->device,
871 kiblnd_cq_completion, kiblnd_cq_event, conn,
872 IBLND_CQ_ENTRIES(conn),
873 kiblnd_get_completion_vector(conn, cpt));
877 * on MLX-5 (possibly MLX-4 as well) this error could be
878 * hit if the concurrent_sends and/or peer_tx_credits is set
879 * too high. Or due to an MLX-5 bug which tries to
880 * allocate 256kb via kmalloc for WR cookie array
882 CERROR("Failed to create CQ with %d CQEs: %ld\n",
883 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
889 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
891 CERROR("Can't request completion notification: %d\n", rc);
895 init_qp_attr->event_handler = kiblnd_qp_event;
896 init_qp_attr->qp_context = conn;
897 init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
898 init_qp_attr->cap.max_recv_sge = 1;
899 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
900 init_qp_attr->qp_type = IB_QPT_RC;
901 init_qp_attr->send_cq = cq;
902 init_qp_attr->recv_cq = cq;
904 conn->ibc_sched = sched;
907 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
908 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
910 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
911 if (!rc || conn->ibc_queue_depth < 2)
914 conn->ibc_queue_depth--;
918 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
919 "send_sge: %d, recv_sge: %d\n",
920 rc, init_qp_attr->cap.max_send_wr,
921 init_qp_attr->cap.max_recv_wr,
922 init_qp_attr->cap.max_send_sge,
923 init_qp_attr->cap.max_recv_sge);
927 if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
928 CWARN("peer %s - queue depth reduced from %u to %u"
929 " to allow for qp creation\n",
930 libcfs_nid2str(peer_ni->ibp_nid),
931 peer_ni->ibp_queue_depth,
932 conn->ibc_queue_depth);
934 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
935 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
936 if (conn->ibc_rxs == NULL) {
937 CERROR("Cannot allocate RX buffers\n");
941 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
942 IBLND_RX_MSG_PAGES(conn));
946 kiblnd_map_rx_descs(conn);
948 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
950 /* 1 ref for caller and each rxmsg */
951 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
952 conn->ibc_nrx = IBLND_RX_MSGS(conn);
955 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
956 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
958 CERROR("Can't post rxmsg: %d\n", rc);
960 /* Make posted receives complete */
961 kiblnd_abort_receives(conn);
963 /* correct # of posted buffers
964 * NB locking needed now I'm racing with completion */
965 spin_lock_irqsave(&sched->ibs_lock, flags);
966 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
967 spin_unlock_irqrestore(&sched->ibs_lock, flags);
969 /* cmid will be destroyed by CM(ofed) after cm_callback
970 * returned, so we can't refer it anymore
971 * (by kiblnd_connd()->kiblnd_destroy_conn) */
972 rdma_destroy_qp(conn->ibc_cmid);
973 conn->ibc_cmid = NULL;
975 /* Drop my own and unused rxbuffer refcounts */
976 while (i++ <= IBLND_RX_MSGS(conn))
977 kiblnd_conn_decref(conn);
983 /* Init successful! */
984 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
985 state == IBLND_CONN_PASSIVE_WAIT);
986 conn->ibc_state = state;
989 atomic_inc(&net->ibn_nconns);
993 kiblnd_destroy_conn(conn);
994 LIBCFS_FREE(conn, sizeof(*conn));
996 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
1002 kiblnd_destroy_conn(struct kib_conn *conn)
1004 struct rdma_cm_id *cmid = conn->ibc_cmid;
1005 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1008 LASSERT (!in_interrupt());
1009 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1010 LASSERT(list_empty(&conn->ibc_early_rxs));
1011 LASSERT(list_empty(&conn->ibc_tx_noops));
1012 LASSERT(list_empty(&conn->ibc_tx_queue));
1013 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1014 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1015 LASSERT(list_empty(&conn->ibc_active_txs));
1016 LASSERT (conn->ibc_noops_posted == 0);
1017 LASSERT (conn->ibc_nsends_posted == 0);
1019 switch (conn->ibc_state) {
1021 /* conn must be completely disengaged from the network */
1024 case IBLND_CONN_DISCONNECTED:
1025 /* connvars should have been freed already */
1026 LASSERT (conn->ibc_connvars == NULL);
1029 case IBLND_CONN_INIT:
1033 /* conn->ibc_cmid might be destroyed by CM already */
1034 if (cmid != NULL && cmid->qp != NULL)
1035 rdma_destroy_qp(cmid);
1037 if (conn->ibc_cq != NULL) {
1038 rc = ib_destroy_cq(conn->ibc_cq);
1040 CWARN("Error destroying CQ: %d\n", rc);
1043 kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1044 LNET_MSG_STATUS_OK);
1046 if (conn->ibc_rx_pages != NULL)
1047 kiblnd_unmap_rx_descs(conn);
1049 if (conn->ibc_rxs != NULL) {
1050 LIBCFS_FREE(conn->ibc_rxs,
1051 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
1054 if (conn->ibc_connvars != NULL)
1055 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1057 if (conn->ibc_hdev != NULL)
1058 kiblnd_hdev_decref(conn->ibc_hdev);
1060 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1061 if (conn->ibc_state != IBLND_CONN_INIT) {
1062 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1064 kiblnd_peer_decref(peer_ni);
1065 rdma_destroy_id(cmid);
1066 atomic_dec(&net->ibn_nconns);
1071 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1073 struct kib_conn *conn;
1074 struct list_head *ctmp;
1075 struct list_head *cnxt;
1078 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1079 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1081 CDEBUG(D_NET, "Closing conn -> %s, "
1082 "version: %x, reason: %d\n",
1083 libcfs_nid2str(peer_ni->ibp_nid),
1084 conn->ibc_version, why);
1086 kiblnd_close_conn_locked(conn, why);
1094 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1095 int version, __u64 incarnation)
1097 struct kib_conn *conn;
1098 struct list_head *ctmp;
1099 struct list_head *cnxt;
1102 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1103 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1105 if (conn->ibc_version == version &&
1106 conn->ibc_incarnation == incarnation)
1109 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1110 "incarnation:%#llx(%x, %#llx)\n",
1111 libcfs_nid2str(peer_ni->ibp_nid),
1112 conn->ibc_version, conn->ibc_incarnation,
1113 version, incarnation);
1115 kiblnd_close_conn_locked(conn, -ESTALE);
1123 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1125 struct kib_peer_ni *peer_ni;
1126 struct list_head *ptmp;
1127 struct list_head *pnxt;
1131 unsigned long flags;
1134 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1136 if (nid != LNET_NID_ANY)
1137 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1140 hi = kiblnd_data.kib_peer_hash_size - 1;
1143 for (i = lo; i <= hi; i++) {
1144 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1146 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
1147 LASSERT(!kiblnd_peer_idle(peer_ni));
1149 if (peer_ni->ibp_ni != ni)
1152 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1155 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1159 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1161 /* wildcards always succeed */
1162 if (nid == LNET_NID_ANY)
1165 return (count == 0) ? -ENOENT : 0;
1169 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1171 struct libcfs_ioctl_data *data = arg;
1175 case IOC_LIBCFS_GET_PEER: {
1179 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1181 data->ioc_nid = nid;
1182 data->ioc_count = count;
1186 case IOC_LIBCFS_DEL_PEER: {
1187 rc = kiblnd_del_peer(ni, data->ioc_nid);
1190 case IOC_LIBCFS_GET_CONN: {
1191 struct kib_conn *conn;
1194 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1200 LASSERT(conn->ibc_cmid != NULL);
1201 data->ioc_nid = conn->ibc_peer->ibp_nid;
1202 if (conn->ibc_cmid->route.path_rec == NULL)
1203 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1206 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1207 kiblnd_conn_decref(conn);
1210 case IOC_LIBCFS_CLOSE_CONNECTION: {
1211 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1223 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1225 time64_t last_alive = 0;
1226 time64_t now = ktime_get_seconds();
1227 rwlock_t *glock = &kiblnd_data.kib_global_lock;
1228 struct kib_peer_ni *peer_ni;
1229 unsigned long flags;
1231 read_lock_irqsave(glock, flags);
1233 peer_ni = kiblnd_find_peer_locked(ni, nid);
1234 if (peer_ni != NULL)
1235 last_alive = peer_ni->ibp_last_alive;
1237 read_unlock_irqrestore(glock, flags);
1239 if (last_alive != 0)
1242 /* peer_ni is not persistent in hash, trigger peer_ni creation
1243 * and connection establishment with a NULL tx */
1244 if (peer_ni == NULL)
1245 kiblnd_launch_tx(ni, NULL, nid);
1247 CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n",
1248 libcfs_nid2str(nid), peer_ni,
1249 last_alive ? now - last_alive : -1);
1254 kiblnd_free_pages(struct kib_pages *p)
1256 int npages = p->ibp_npages;
1259 for (i = 0; i < npages; i++) {
1260 if (p->ibp_pages[i] != NULL)
1261 __free_page(p->ibp_pages[i]);
1264 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1268 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1270 struct kib_pages *p;
1273 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1274 offsetof(struct kib_pages, ibp_pages[npages]));
1276 CERROR("Can't allocate descriptor for %d pages\n", npages);
1280 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1281 p->ibp_npages = npages;
1283 for (i = 0; i < npages; i++) {
1284 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1286 if (p->ibp_pages[i] == NULL) {
1287 CERROR("Can't allocate page %d of %d\n", i, npages);
1288 kiblnd_free_pages(p);
1298 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1303 LASSERT (conn->ibc_rxs != NULL);
1304 LASSERT (conn->ibc_hdev != NULL);
1306 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1307 rx = &conn->ibc_rxs[i];
1309 LASSERT(rx->rx_nob >= 0); /* not posted */
1311 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1312 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1314 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1317 kiblnd_free_pages(conn->ibc_rx_pages);
1319 conn->ibc_rx_pages = NULL;
1323 kiblnd_map_rx_descs(struct kib_conn *conn)
1331 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1332 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1333 rx = &conn->ibc_rxs[i];
1336 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1339 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1340 rx->rx_msg, IBLND_MSG_SIZE,
1342 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1344 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1346 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1347 i, rx->rx_msg, rx->rx_msgaddr,
1348 (__u64)(page_to_phys(pg) + pg_off));
1350 pg_off += IBLND_MSG_SIZE;
1351 LASSERT(pg_off <= PAGE_SIZE);
1353 if (pg_off == PAGE_SIZE) {
1356 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1362 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1364 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1368 LASSERT (tpo->tpo_pool.po_allocated == 0);
1373 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1374 tx = &tpo->tpo_tx_descs[i];
1375 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1376 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1378 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1381 kiblnd_hdev_decref(hdev);
1382 tpo->tpo_hdev = NULL;
1385 static struct kib_hca_dev *
1386 kiblnd_current_hdev(struct kib_dev *dev)
1388 struct kib_hca_dev *hdev;
1389 unsigned long flags;
1392 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1393 while (dev->ibd_failover) {
1394 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1396 CDEBUG(D_NET, "%s: Wait for failover\n",
1398 set_current_state(TASK_INTERRUPTIBLE);
1399 schedule_timeout(cfs_time_seconds(1) / 100);
1401 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1404 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1405 hdev = dev->ibd_hdev;
1407 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1413 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1415 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1416 struct kib_pool *pool = &tpo->tpo_pool;
1417 struct kib_net *net = pool->po_owner->ps_net;
1418 struct kib_dev *dev;
1425 LASSERT (net != NULL);
1429 /* pre-mapped messages are not bigger than 1 page */
1430 CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1432 /* No fancy arithmetic when we do the buffer calculations */
1433 CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1435 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1437 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1438 page = txpgs->ibp_pages[ipage];
1439 tx = &tpo->tpo_tx_descs[i];
1441 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1444 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1448 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1450 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1452 list_add(&tx->tx_list, &pool->po_free_list);
1454 page_offset += IBLND_MSG_SIZE;
1455 LASSERT(page_offset <= PAGE_SIZE);
1457 if (page_offset == PAGE_SIZE) {
1460 LASSERT(ipage <= txpgs->ibp_npages);
1466 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1468 LASSERT(fpo->fpo_map_count == 0);
1470 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1471 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1473 struct kib_fast_reg_descriptor *frd, *tmp;
1476 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1478 list_del(&frd->frd_list);
1479 #ifndef HAVE_IB_MAP_MR_SG
1480 ib_free_fast_reg_page_list(frd->frd_frpl);
1482 ib_dereg_mr(frd->frd_mr);
1483 LIBCFS_FREE(frd, sizeof(*frd));
1486 if (i < fpo->fast_reg.fpo_pool_size)
1487 CERROR("FastReg pool still has %d regions registered\n",
1488 fpo->fast_reg.fpo_pool_size - i);
1492 kiblnd_hdev_decref(fpo->fpo_hdev);
1494 LIBCFS_FREE(fpo, sizeof(*fpo));
1498 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1500 struct kib_fmr_pool *fpo, *tmp;
1502 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1503 list_del(&fpo->fpo_list);
1504 kiblnd_destroy_fmr_pool(fpo);
1509 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1512 int size = tunables->lnd_fmr_pool_size / ncpts;
1514 return max(IBLND_FMR_POOL, size);
1518 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1521 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1523 return max(IBLND_FMR_POOL_FLUSH, size);
1526 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1527 struct kib_fmr_pool *fpo)
1529 struct ib_fmr_pool_param param = {
1530 .max_pages_per_fmr = LNET_MAX_IOV,
1531 .page_shift = PAGE_SHIFT,
1532 .access = (IB_ACCESS_LOCAL_WRITE |
1533 IB_ACCESS_REMOTE_WRITE),
1534 .pool_size = fps->fps_pool_size,
1535 .dirty_watermark = fps->fps_flush_trigger,
1536 .flush_function = NULL,
1538 .cache = !!fps->fps_cache };
1541 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1543 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1544 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1546 CERROR("Failed to create FMR pool: %d\n", rc);
1548 CERROR("FMRs are not supported\n");
1550 fpo->fpo_is_fmr = true;
1555 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1556 struct kib_fmr_pool *fpo,
1557 enum kib_dev_caps dev_caps)
1559 struct kib_fast_reg_descriptor *frd, *tmp;
1562 fpo->fpo_is_fmr = false;
1564 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1565 fpo->fast_reg.fpo_pool_size = 0;
1566 for (i = 0; i < fps->fps_pool_size; i++) {
1567 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1570 CERROR("Failed to allocate a new fast_reg descriptor\n");
1576 #ifndef HAVE_IB_MAP_MR_SG
1577 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1579 if (IS_ERR(frd->frd_frpl)) {
1580 rc = PTR_ERR(frd->frd_frpl);
1581 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1583 frd->frd_frpl = NULL;
1588 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1589 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1593 * it is expected to get here if this is an MLX-5 card.
1594 * MLX-4 cards will always use FMR and MLX-5 cards will
1595 * always use fast_reg. It turns out that some MLX-5 cards
1596 * (possibly due to older FW versions) do not natively support
1597 * gaps. So we will need to track them here.
1599 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1600 #ifdef IB_MR_TYPE_SG_GAPS
1601 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1602 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1603 IB_MR_TYPE_SG_GAPS :
1609 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1610 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1611 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1613 if (IS_ERR(frd->frd_mr)) {
1614 rc = PTR_ERR(frd->frd_mr);
1615 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1620 /* There appears to be a bug in MLX5 code where you must
1621 * invalidate the rkey of a new FastReg pool before first
1622 * using it. Thus, I am marking the FRD invalid here. */
1623 frd->frd_valid = false;
1625 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1626 fpo->fast_reg.fpo_pool_size++;
1633 ib_dereg_mr(frd->frd_mr);
1634 #ifndef HAVE_IB_MAP_MR_SG
1636 ib_free_fast_reg_page_list(frd->frd_frpl);
1638 LIBCFS_FREE(frd, sizeof(*frd));
1641 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1643 list_del(&frd->frd_list);
1644 #ifndef HAVE_IB_MAP_MR_SG
1645 ib_free_fast_reg_page_list(frd->frd_frpl);
1647 ib_dereg_mr(frd->frd_mr);
1648 LIBCFS_FREE(frd, sizeof(*frd));
1654 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1655 struct kib_fmr_pool **pp_fpo)
1657 struct kib_dev *dev = fps->fps_net->ibn_dev;
1658 struct kib_fmr_pool *fpo;
1661 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1665 memset(fpo, 0, sizeof(*fpo));
1667 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1669 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1670 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1672 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1676 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1677 fpo->fpo_owner = fps;
1683 kiblnd_hdev_decref(fpo->fpo_hdev);
1684 LIBCFS_FREE(fpo, sizeof(*fpo));
1689 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1691 if (fps->fps_net == NULL) /* intialized? */
1694 spin_lock(&fps->fps_lock);
1696 while (!list_empty(&fps->fps_pool_list)) {
1697 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1698 struct kib_fmr_pool,
1701 fpo->fpo_failed = 1;
1702 list_del(&fpo->fpo_list);
1703 if (fpo->fpo_map_count == 0)
1704 list_add(&fpo->fpo_list, zombies);
1706 list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1709 spin_unlock(&fps->fps_lock);
1713 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1715 if (fps->fps_net != NULL) { /* initialized? */
1716 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1717 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1722 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1723 struct kib_net *net,
1724 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1726 struct kib_fmr_pool *fpo;
1729 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1734 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1735 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1736 fps->fps_cache = tunables->lnd_fmr_cache;
1738 spin_lock_init(&fps->fps_lock);
1739 INIT_LIST_HEAD(&fps->fps_pool_list);
1740 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1742 rc = kiblnd_create_fmr_pool(fps, &fpo);
1744 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1750 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1752 if (fpo->fpo_map_count != 0) /* still in use */
1754 if (fpo->fpo_failed)
1756 return now >= fpo->fpo_deadline;
1760 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1762 struct kib_hca_dev *hdev;
1763 __u64 *pages = tx->tx_pages;
1768 hdev = tx->tx_pool->tpo_hdev;
1770 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1771 for (size = 0; size < rd->rd_frags[i].rf_nob;
1772 size += hdev->ibh_page_size) {
1773 pages[npages++] = (rd->rd_frags[i].rf_addr &
1774 hdev->ibh_page_mask) + size;
1782 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1784 struct list_head zombies = LIST_HEAD_INIT(zombies);
1785 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1786 struct kib_fmr_poolset *fps;
1787 time64_t now = ktime_get_seconds();
1788 struct kib_fmr_pool *tmp;
1794 fps = fpo->fpo_owner;
1795 if (fpo->fpo_is_fmr) {
1796 if (fmr->fmr_pfmr) {
1797 rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1799 fmr->fmr_pfmr = NULL;
1803 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1807 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1810 frd->frd_valid = false;
1811 spin_lock(&fps->fps_lock);
1812 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1813 spin_unlock(&fps->fps_lock);
1814 fmr->fmr_frd = NULL;
1817 fmr->fmr_pool = NULL;
1819 spin_lock(&fps->fps_lock);
1820 fpo->fpo_map_count--; /* decref the pool */
1822 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1823 /* the first pool is persistent */
1824 if (fps->fps_pool_list.next == &fpo->fpo_list)
1827 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1828 list_move(&fpo->fpo_list, &zombies);
1832 spin_unlock(&fps->fps_lock);
1834 if (!list_empty(&zombies))
1835 kiblnd_destroy_fmr_pool_list(&zombies);
1838 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1839 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1840 struct kib_fmr *fmr)
1842 struct kib_fmr_pool *fpo;
1843 __u64 *pages = tx->tx_pages;
1845 bool is_rx = (rd != tx->tx_rd);
1846 bool tx_pages_mapped = 0;
1851 spin_lock(&fps->fps_lock);
1852 version = fps->fps_version;
1853 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1854 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1855 fpo->fpo_map_count++;
1857 if (fpo->fpo_is_fmr) {
1858 struct ib_pool_fmr *pfmr;
1860 spin_unlock(&fps->fps_lock);
1862 if (!tx_pages_mapped) {
1863 npages = kiblnd_map_tx_pages(tx, rd);
1864 tx_pages_mapped = 1;
1867 pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1868 pages, npages, iov);
1869 if (likely(!IS_ERR(pfmr))) {
1870 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1872 fmr->fmr_frd = NULL;
1873 fmr->fmr_pfmr = pfmr;
1874 fmr->fmr_pool = fpo;
1879 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1880 struct kib_fast_reg_descriptor *frd;
1881 #ifdef HAVE_IB_MAP_MR_SG
1882 struct ib_reg_wr *wr;
1885 struct ib_rdma_wr *wr;
1886 struct ib_fast_reg_page_list *frpl;
1890 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1891 struct kib_fast_reg_descriptor,
1893 list_del(&frd->frd_list);
1894 spin_unlock(&fps->fps_lock);
1896 #ifndef HAVE_IB_MAP_MR_SG
1897 frpl = frd->frd_frpl;
1901 if (!frd->frd_valid) {
1902 struct ib_rdma_wr *inv_wr;
1903 __u32 key = is_rx ? mr->rkey : mr->lkey;
1905 inv_wr = &frd->frd_inv_wr;
1906 memset(inv_wr, 0, sizeof(*inv_wr));
1908 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1909 inv_wr->wr.wr_id = IBLND_WID_MR;
1910 inv_wr->wr.ex.invalidate_rkey = key;
1913 key = ib_inc_rkey(key);
1914 ib_update_fast_reg_key(mr, key);
1917 #ifdef HAVE_IB_MAP_MR_SG
1918 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1919 n = ib_map_mr_sg(mr, tx->tx_frags,
1920 tx->tx_nfrags, NULL, PAGE_SIZE);
1922 n = ib_map_mr_sg(mr, tx->tx_frags,
1923 tx->tx_nfrags, PAGE_SIZE);
1925 if (unlikely(n != tx->tx_nfrags)) {
1926 CERROR("Failed to map mr %d/%d "
1927 "elements\n", n, tx->tx_nfrags);
1928 return n < 0 ? n : -EINVAL;
1931 wr = &frd->frd_fastreg_wr;
1932 memset(wr, 0, sizeof(*wr));
1934 wr->wr.opcode = IB_WR_REG_MR;
1935 wr->wr.wr_id = IBLND_WID_MR;
1937 wr->wr.send_flags = 0;
1939 wr->key = is_rx ? mr->rkey : mr->lkey;
1940 wr->access = (IB_ACCESS_LOCAL_WRITE |
1941 IB_ACCESS_REMOTE_WRITE);
1943 if (!tx_pages_mapped) {
1944 npages = kiblnd_map_tx_pages(tx, rd);
1945 tx_pages_mapped = 1;
1948 LASSERT(npages <= frpl->max_page_list_len);
1949 memcpy(frpl->page_list, pages,
1950 sizeof(*pages) * npages);
1952 /* Prepare FastReg WR */
1953 wr = &frd->frd_fastreg_wr;
1954 memset(wr, 0, sizeof(*wr));
1956 wr->wr.opcode = IB_WR_FAST_REG_MR;
1957 wr->wr.wr_id = IBLND_WID_MR;
1959 wr->wr.wr.fast_reg.iova_start = iov;
1960 wr->wr.wr.fast_reg.page_list = frpl;
1961 wr->wr.wr.fast_reg.page_list_len = npages;
1962 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1963 wr->wr.wr.fast_reg.length = nob;
1964 wr->wr.wr.fast_reg.rkey =
1965 is_rx ? mr->rkey : mr->lkey;
1966 wr->wr.wr.fast_reg.access_flags =
1967 (IB_ACCESS_LOCAL_WRITE |
1968 IB_ACCESS_REMOTE_WRITE);
1971 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1973 fmr->fmr_pfmr = NULL;
1974 fmr->fmr_pool = fpo;
1977 spin_unlock(&fps->fps_lock);
1981 spin_lock(&fps->fps_lock);
1982 fpo->fpo_map_count--;
1983 if (rc != -EAGAIN) {
1984 spin_unlock(&fps->fps_lock);
1988 /* EAGAIN and ... */
1989 if (version != fps->fps_version) {
1990 spin_unlock(&fps->fps_lock);
1995 if (fps->fps_increasing) {
1996 spin_unlock(&fps->fps_lock);
1997 CDEBUG(D_NET, "Another thread is allocating new "
1998 "FMR pool, waiting for her to complete\n");
2004 if (ktime_get_seconds() < fps->fps_next_retry) {
2005 /* someone failed recently */
2006 spin_unlock(&fps->fps_lock);
2010 fps->fps_increasing = 1;
2011 spin_unlock(&fps->fps_lock);
2013 CDEBUG(D_NET, "Allocate new FMR pool\n");
2014 rc = kiblnd_create_fmr_pool(fps, &fpo);
2015 spin_lock(&fps->fps_lock);
2016 fps->fps_increasing = 0;
2019 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2021 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2023 spin_unlock(&fps->fps_lock);
2029 kiblnd_fini_pool(struct kib_pool *pool)
2031 LASSERT(list_empty(&pool->po_free_list));
2032 LASSERT(pool->po_allocated == 0);
2034 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2038 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2040 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2042 memset(pool, 0, sizeof(struct kib_pool));
2043 INIT_LIST_HEAD(&pool->po_free_list);
2044 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2045 pool->po_owner = ps;
2046 pool->po_size = size;
2050 kiblnd_destroy_pool_list(struct list_head *head)
2052 struct kib_pool *pool;
2054 while (!list_empty(head)) {
2055 pool = list_entry(head->next, struct kib_pool, po_list);
2056 list_del(&pool->po_list);
2058 LASSERT(pool->po_owner != NULL);
2059 pool->po_owner->ps_pool_destroy(pool);
2064 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2066 if (ps->ps_net == NULL) /* intialized? */
2069 spin_lock(&ps->ps_lock);
2070 while (!list_empty(&ps->ps_pool_list)) {
2071 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
2072 struct kib_pool, po_list);
2075 list_del(&po->po_list);
2076 if (po->po_allocated == 0)
2077 list_add(&po->po_list, zombies);
2079 list_add(&po->po_list, &ps->ps_failed_pool_list);
2081 spin_unlock(&ps->ps_lock);
2085 kiblnd_fini_poolset(struct kib_poolset *ps)
2087 if (ps->ps_net != NULL) { /* initialized? */
2088 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2089 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2094 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2095 struct kib_net *net, char *name, int size,
2096 kib_ps_pool_create_t po_create,
2097 kib_ps_pool_destroy_t po_destroy,
2098 kib_ps_node_init_t nd_init,
2099 kib_ps_node_fini_t nd_fini)
2101 struct kib_pool *pool;
2104 memset(ps, 0, sizeof(struct kib_poolset));
2108 ps->ps_pool_create = po_create;
2109 ps->ps_pool_destroy = po_destroy;
2110 ps->ps_node_init = nd_init;
2111 ps->ps_node_fini = nd_fini;
2112 ps->ps_pool_size = size;
2113 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2114 >= sizeof(ps->ps_name))
2116 spin_lock_init(&ps->ps_lock);
2117 INIT_LIST_HEAD(&ps->ps_pool_list);
2118 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2120 rc = ps->ps_pool_create(ps, size, &pool);
2122 list_add(&pool->po_list, &ps->ps_pool_list);
2124 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2130 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2132 if (pool->po_allocated != 0) /* still in use */
2134 if (pool->po_failed)
2136 return now >= pool->po_deadline;
2140 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2142 struct list_head zombies = LIST_HEAD_INIT(zombies);
2143 struct kib_poolset *ps = pool->po_owner;
2144 struct kib_pool *tmp;
2145 time64_t now = ktime_get_seconds();
2147 spin_lock(&ps->ps_lock);
2149 if (ps->ps_node_fini != NULL)
2150 ps->ps_node_fini(pool, node);
2152 LASSERT(pool->po_allocated > 0);
2153 list_add(node, &pool->po_free_list);
2154 pool->po_allocated--;
2156 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2157 /* the first pool is persistent */
2158 if (ps->ps_pool_list.next == &pool->po_list)
2161 if (kiblnd_pool_is_idle(pool, now))
2162 list_move(&pool->po_list, &zombies);
2164 spin_unlock(&ps->ps_lock);
2166 if (!list_empty(&zombies))
2167 kiblnd_destroy_pool_list(&zombies);
2171 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2173 struct list_head *node;
2174 struct kib_pool *pool;
2176 unsigned int interval = 1;
2177 ktime_t time_before;
2178 unsigned int trips = 0;
2181 spin_lock(&ps->ps_lock);
2182 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2183 if (list_empty(&pool->po_free_list))
2186 pool->po_allocated++;
2187 pool->po_deadline = ktime_get_seconds() +
2188 IBLND_POOL_DEADLINE;
2189 node = pool->po_free_list.next;
2192 if (ps->ps_node_init != NULL) {
2193 /* still hold the lock */
2194 ps->ps_node_init(pool, node);
2196 spin_unlock(&ps->ps_lock);
2200 /* no available tx pool and ... */
2201 if (ps->ps_increasing) {
2202 /* another thread is allocating a new pool */
2203 spin_unlock(&ps->ps_lock);
2205 CDEBUG(D_NET, "Another thread is allocating new "
2206 "%s pool, waiting %d HZs for her to complete."
2208 ps->ps_name, interval, trips);
2210 set_current_state(TASK_INTERRUPTIBLE);
2211 schedule_timeout(interval);
2212 if (interval < cfs_time_seconds(1))
2218 if (ktime_get_seconds() < ps->ps_next_retry) {
2219 /* someone failed recently */
2220 spin_unlock(&ps->ps_lock);
2224 ps->ps_increasing = 1;
2225 spin_unlock(&ps->ps_lock);
2227 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2228 time_before = ktime_get();
2229 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2230 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2231 ktime_ms_delta(ktime_get(), time_before));
2233 spin_lock(&ps->ps_lock);
2234 ps->ps_increasing = 0;
2236 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2238 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2239 CERROR("Can't allocate new %s pool because out of memory\n",
2242 spin_unlock(&ps->ps_lock);
2248 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2250 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2254 LASSERT (pool->po_allocated == 0);
2256 if (tpo->tpo_tx_pages != NULL) {
2257 kiblnd_unmap_tx_pool(tpo);
2258 kiblnd_free_pages(tpo->tpo_tx_pages);
2261 if (tpo->tpo_tx_descs == NULL)
2264 for (i = 0; i < pool->po_size; i++) {
2265 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2266 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2268 list_del(&tx->tx_list);
2269 if (tx->tx_pages != NULL)
2270 LIBCFS_FREE(tx->tx_pages,
2272 sizeof(*tx->tx_pages));
2273 if (tx->tx_frags != NULL)
2274 LIBCFS_FREE(tx->tx_frags,
2275 (1 + IBLND_MAX_RDMA_FRAGS) *
2276 sizeof(*tx->tx_frags));
2277 if (tx->tx_wrq != NULL)
2278 LIBCFS_FREE(tx->tx_wrq,
2279 (1 + IBLND_MAX_RDMA_FRAGS) *
2280 sizeof(*tx->tx_wrq));
2281 if (tx->tx_sge != NULL)
2282 LIBCFS_FREE(tx->tx_sge,
2283 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2284 sizeof(*tx->tx_sge));
2285 if (tx->tx_rd != NULL)
2286 LIBCFS_FREE(tx->tx_rd,
2287 offsetof(struct kib_rdma_desc,
2288 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2291 LIBCFS_FREE(tpo->tpo_tx_descs,
2292 pool->po_size * sizeof(struct kib_tx));
2294 kiblnd_fini_pool(pool);
2295 LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool));
2298 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2300 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2303 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2304 ntx = tunables->lnd_ntx / ncpts;
2306 return max(IBLND_TX_POOL, ntx);
2310 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2314 struct kib_pool *pool;
2315 struct kib_tx_pool *tpo;
2317 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2319 CERROR("Failed to allocate TX pool\n");
2323 pool = &tpo->tpo_pool;
2324 kiblnd_init_pool(ps, pool, size);
2325 tpo->tpo_tx_descs = NULL;
2326 tpo->tpo_tx_pages = NULL;
2328 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2329 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2330 CERROR("Can't allocate tx pages: %d\n", npg);
2331 LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool));
2335 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2336 size * sizeof(struct kib_tx));
2337 if (tpo->tpo_tx_descs == NULL) {
2338 CERROR("Can't allocate %d tx descriptors\n", size);
2339 ps->ps_pool_destroy(pool);
2343 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2345 for (i = 0; i < size; i++) {
2346 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2347 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2350 if (ps->ps_net->ibn_fmr_ps != NULL) {
2351 LIBCFS_CPT_ALLOC(tx->tx_pages,
2352 lnet_cpt_table(), ps->ps_cpt,
2353 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2354 if (tx->tx_pages == NULL)
2358 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2359 (1 + IBLND_MAX_RDMA_FRAGS) *
2360 sizeof(*tx->tx_frags));
2361 if (tx->tx_frags == NULL)
2364 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2366 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2367 (1 + IBLND_MAX_RDMA_FRAGS) *
2368 sizeof(*tx->tx_wrq));
2369 if (tx->tx_wrq == NULL)
2372 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2373 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2374 sizeof(*tx->tx_sge));
2375 if (tx->tx_sge == NULL)
2378 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2379 offsetof(struct kib_rdma_desc,
2380 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2381 if (tx->tx_rd == NULL)
2386 kiblnd_map_tx_pool(tpo);
2391 ps->ps_pool_destroy(pool);
2396 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2398 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2399 struct kib_tx_poolset,
2401 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2403 tx->tx_cookie = tps->tps_next_tx_cookie++;
2407 kiblnd_net_fini_pools(struct kib_net *net)
2411 cfs_cpt_for_each(i, lnet_cpt_table()) {
2412 struct kib_tx_poolset *tps;
2413 struct kib_fmr_poolset *fps;
2415 if (net->ibn_tx_ps != NULL) {
2416 tps = net->ibn_tx_ps[i];
2417 kiblnd_fini_poolset(&tps->tps_poolset);
2420 if (net->ibn_fmr_ps != NULL) {
2421 fps = net->ibn_fmr_ps[i];
2422 kiblnd_fini_fmr_poolset(fps);
2426 if (net->ibn_tx_ps != NULL) {
2427 cfs_percpt_free(net->ibn_tx_ps);
2428 net->ibn_tx_ps = NULL;
2431 if (net->ibn_fmr_ps != NULL) {
2432 cfs_percpt_free(net->ibn_fmr_ps);
2433 net->ibn_fmr_ps = NULL;
2438 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2441 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2442 #ifdef HAVE_IB_GET_DMA_MR
2443 unsigned long flags;
2449 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2451 #ifdef HAVE_IB_GET_DMA_MR
2452 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2454 * if lnd_map_on_demand is zero then we have effectively disabled
2455 * FMR or FastReg and we're using global memory regions
2458 if (!tunables->lnd_map_on_demand) {
2459 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2461 goto create_tx_pool;
2464 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2467 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2468 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2469 tunables->lnd_fmr_pool_size,
2470 tunables->lnd_ntx / 4);
2475 /* TX pool must be created later than FMR, see LU-2268
2477 LASSERT(net->ibn_tx_ps == NULL);
2479 /* premapping can fail if ibd_nmr > 1, so we always create
2480 * FMR pool and map-on-demand if premapping failed */
2482 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2483 sizeof(struct kib_fmr_poolset));
2484 if (net->ibn_fmr_ps == NULL) {
2485 CERROR("Failed to allocate FMR pool array\n");
2490 for (i = 0; i < ncpts; i++) {
2491 cpt = (cpts == NULL) ? i : cpts[i];
2492 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2495 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2502 LASSERT(i == ncpts);
2504 #ifdef HAVE_IB_GET_DMA_MR
2507 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2508 sizeof(struct kib_tx_poolset));
2509 if (net->ibn_tx_ps == NULL) {
2510 CERROR("Failed to allocate tx pool array\n");
2515 for (i = 0; i < ncpts; i++) {
2516 cpt = (cpts == NULL) ? i : cpts[i];
2517 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2519 kiblnd_tx_pool_size(ni, ncpts),
2520 kiblnd_create_tx_pool,
2521 kiblnd_destroy_tx_pool,
2522 kiblnd_tx_init, NULL);
2524 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2532 kiblnd_net_fini_pools(net);
2538 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2540 struct ib_device_attr *dev_attr;
2543 /* It's safe to assume a HCA can handle a page size
2544 * matching that of the native system */
2545 hdev->ibh_page_shift = PAGE_SHIFT;
2546 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2547 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2549 #ifndef HAVE_IB_DEVICE_ATTRS
2550 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2551 if (dev_attr == NULL) {
2552 CERROR("Out of memory\n");
2556 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2558 CERROR("Failed to query IB device: %d\n", rc);
2559 goto out_clean_attr;
2562 dev_attr = &hdev->ibh_ibdev->attrs;
2565 hdev->ibh_mr_size = dev_attr->max_mr_size;
2567 /* Setup device Memory Registration capabilities */
2568 if (hdev->ibh_ibdev->alloc_fmr &&
2569 hdev->ibh_ibdev->dealloc_fmr &&
2570 hdev->ibh_ibdev->map_phys_fmr &&
2571 hdev->ibh_ibdev->unmap_fmr) {
2572 LCONSOLE_INFO("Using FMR for registration\n");
2573 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2574 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2575 LCONSOLE_INFO("Using FastReg for registration\n");
2576 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2577 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2578 #ifdef IB_DEVICE_SG_GAPS_REG
2579 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2580 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2587 if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2588 hdev->ibh_mr_shift = 64;
2592 #ifndef HAVE_IB_DEVICE_ATTRS
2594 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2598 CERROR("IB device does not support FMRs nor FastRegs, can't "
2599 "register memory: %d\n", rc);
2600 else if (rc == -EINVAL)
2601 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2605 #ifdef HAVE_IB_GET_DMA_MR
2607 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2609 if (hdev->ibh_mrs == NULL)
2612 ib_dereg_mr(hdev->ibh_mrs);
2614 hdev->ibh_mrs = NULL;
2619 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2621 #ifdef HAVE_IB_GET_DMA_MR
2622 kiblnd_hdev_cleanup_mrs(hdev);
2625 if (hdev->ibh_pd != NULL)
2626 ib_dealloc_pd(hdev->ibh_pd);
2628 if (hdev->ibh_cmid != NULL)
2629 rdma_destroy_id(hdev->ibh_cmid);
2631 LIBCFS_FREE(hdev, sizeof(*hdev));
2634 #ifdef HAVE_IB_GET_DMA_MR
2636 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2639 int acflags = IB_ACCESS_LOCAL_WRITE |
2640 IB_ACCESS_REMOTE_WRITE;
2642 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2644 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2645 kiblnd_hdev_cleanup_mrs(hdev);
2656 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2662 kiblnd_dev_need_failover(struct kib_dev *dev)
2664 struct rdma_cm_id *cmid;
2665 struct sockaddr_in srcaddr;
2666 struct sockaddr_in dstaddr;
2669 if (dev->ibd_hdev == NULL || /* initializing */
2670 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2671 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2674 /* XXX: it's UGLY, but I don't have better way to find
2675 * ib-bonding HCA failover because:
2677 * a. no reliable CM event for HCA failover...
2678 * b. no OFED API to get ib_device for current net_device...
2680 * We have only two choices at this point:
2682 * a. rdma_bind_addr(), it will conflict with listener cmid
2683 * b. rdma_resolve_addr() to zero addr */
2684 cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2688 CERROR("Failed to create cmid for failover: %d\n", rc);
2692 memset(&srcaddr, 0, sizeof(srcaddr));
2693 srcaddr.sin_family = AF_INET;
2694 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2696 memset(&dstaddr, 0, sizeof(dstaddr));
2697 dstaddr.sin_family = AF_INET;
2698 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2699 (struct sockaddr *)&dstaddr, 1);
2700 if (rc != 0 || cmid->device == NULL) {
2701 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2702 dev->ibd_ifname, &dev->ibd_ifip,
2704 rdma_destroy_id(cmid);
2708 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2709 rdma_destroy_id(cmid);
2714 kiblnd_dev_failover(struct kib_dev *dev)
2716 struct list_head zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2717 struct list_head zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2718 struct list_head zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2719 struct rdma_cm_id *cmid = NULL;
2720 struct kib_hca_dev *hdev = NULL;
2721 struct kib_hca_dev *old;
2723 struct kib_net *net;
2724 struct sockaddr_in addr;
2725 unsigned long flags;
2729 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2730 dev->ibd_can_failover ||
2731 dev->ibd_hdev == NULL);
2733 rc = kiblnd_dev_need_failover(dev);
2737 if (dev->ibd_hdev != NULL &&
2738 dev->ibd_hdev->ibh_cmid != NULL) {
2739 /* XXX it's not good to close old listener at here,
2740 * because we can fail to create new listener.
2741 * But we have to close it now, otherwise rdma_bind_addr
2742 * will return EADDRINUSE... How crap! */
2743 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2745 cmid = dev->ibd_hdev->ibh_cmid;
2746 /* make next schedule of kiblnd_dev_need_failover()
2747 * return 1 for me */
2748 dev->ibd_hdev->ibh_cmid = NULL;
2749 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2751 rdma_destroy_id(cmid);
2754 cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2758 CERROR("Failed to create cmid for failover: %d\n", rc);
2762 memset(&addr, 0, sizeof(addr));
2763 addr.sin_family = AF_INET;
2764 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2765 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2767 /* Bind to failover device or port */
2768 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2769 if (rc != 0 || cmid->device == NULL) {
2770 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2771 dev->ibd_ifname, &dev->ibd_ifip,
2773 rdma_destroy_id(cmid);
2777 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2779 CERROR("Failed to allocate kib_hca_dev\n");
2780 rdma_destroy_id(cmid);
2785 atomic_set(&hdev->ibh_ref, 1);
2786 hdev->ibh_dev = dev;
2787 hdev->ibh_cmid = cmid;
2788 hdev->ibh_ibdev = cmid->device;
2790 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2791 pd = ib_alloc_pd(cmid->device, 0);
2793 pd = ib_alloc_pd(cmid->device);
2797 CERROR("Can't allocate PD: %d\n", rc);
2803 rc = rdma_listen(cmid, 0);
2805 CERROR("Can't start new listener: %d\n", rc);
2809 rc = kiblnd_hdev_get_attr(hdev);
2811 CERROR("Can't get device attributes: %d\n", rc);
2815 #ifdef HAVE_IB_GET_DMA_MR
2816 rc = kiblnd_hdev_setup_mrs(hdev);
2818 CERROR("Can't setup device: %d\n", rc);
2823 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2825 old = dev->ibd_hdev;
2826 dev->ibd_hdev = hdev; /* take over the refcount */
2829 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2830 cfs_cpt_for_each(i, lnet_cpt_table()) {
2831 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2834 if (net->ibn_fmr_ps != NULL)
2835 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2840 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2842 if (!list_empty(&zombie_tpo))
2843 kiblnd_destroy_pool_list(&zombie_tpo);
2844 if (!list_empty(&zombie_ppo))
2845 kiblnd_destroy_pool_list(&zombie_ppo);
2846 if (!list_empty(&zombie_fpo))
2847 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2849 kiblnd_hdev_decref(hdev);
2852 dev->ibd_failed_failover++;
2854 dev->ibd_failed_failover = 0;
2860 kiblnd_destroy_dev(struct kib_dev *dev)
2862 LASSERT(dev->ibd_nnets == 0);
2863 LASSERT(list_empty(&dev->ibd_nets));
2865 list_del(&dev->ibd_fail_list);
2866 list_del(&dev->ibd_list);
2868 if (dev->ibd_hdev != NULL)
2869 kiblnd_hdev_decref(dev->ibd_hdev);
2871 LIBCFS_FREE(dev, sizeof(*dev));
2874 static struct kib_dev *
2875 kiblnd_create_dev(char *ifname)
2877 struct net_device *netdev;
2878 struct kib_dev *dev = NULL;
2883 for_each_netdev(&init_net, netdev) {
2884 struct in_device *in_dev;
2886 if (strcmp(netdev->name, "lo") == 0) /* skip the loopback IF */
2889 flags = dev_get_flags(netdev);
2890 if (!(flags & IFF_UP)) {
2891 CWARN("Can't query IPoIB interface %s: it's down\n",
2896 in_dev = __in_dev_get_rtnl(netdev);
2898 CWARN("Interface %s has no IPv4 status.\n",
2904 if (strcmp(ifname, ifa->ifa_label) == 0) {
2905 LIBCFS_ALLOC(dev, sizeof(*dev));
2909 dev->ibd_can_failover = !!(flags & IFF_MASTER);
2910 dev->ibd_ifip = ntohl(ifa->ifa_local);
2912 INIT_LIST_HEAD(&dev->ibd_nets);
2913 INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2914 INIT_LIST_HEAD(&dev->ibd_fail_list);
2922 CERROR("Can't find any usable interfaces\n");
2926 if (dev->ibd_ifip == 0) {
2927 CERROR("Can't initialize device: no IP address\n");
2930 strcpy(&dev->ibd_ifname[0], ifname);
2932 /* initialize the device */
2933 rc = kiblnd_dev_failover(dev);
2935 CERROR("Can't initialize device: %d\n", rc);
2939 list_add_tail(&dev->ibd_list, &kiblnd_data.kib_devs);
2944 LIBCFS_FREE(dev, sizeof(*dev));
2949 kiblnd_base_shutdown(void)
2951 struct kib_sched_info *sched;
2954 LASSERT(list_empty(&kiblnd_data.kib_devs));
2956 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2957 atomic_read(&libcfs_kmemory));
2959 switch (kiblnd_data.kib_init) {
2963 case IBLND_INIT_ALL:
2964 case IBLND_INIT_DATA:
2965 LASSERT (kiblnd_data.kib_peers != NULL);
2966 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2967 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2969 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2970 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2971 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2972 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2974 /* flag threads to terminate; wake and wait for them to die */
2975 kiblnd_data.kib_shutdown = 1;
2977 /* NB: we really want to stop scheduler threads net by net
2978 * instead of the whole module, this should be improved
2979 * with dynamic configuration LNet */
2980 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2981 wake_up_all(&sched->ibs_waitq);
2983 wake_up_all(&kiblnd_data.kib_connd_waitq);
2984 wake_up_all(&kiblnd_data.kib_failover_waitq);
2987 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2990 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2991 "Waiting for %d threads to terminate\n",
2992 atomic_read(&kiblnd_data.kib_nthreads));
2993 set_current_state(TASK_UNINTERRUPTIBLE);
2994 schedule_timeout(cfs_time_seconds(1));
2999 case IBLND_INIT_NOTHING:
3003 if (kiblnd_data.kib_peers != NULL) {
3004 LIBCFS_FREE(kiblnd_data.kib_peers,
3005 sizeof(struct list_head) *
3006 kiblnd_data.kib_peer_hash_size);
3009 if (kiblnd_data.kib_scheds != NULL)
3010 cfs_percpt_free(kiblnd_data.kib_scheds);
3012 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
3013 atomic_read(&libcfs_kmemory));
3015 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3016 module_put(THIS_MODULE);
3020 kiblnd_shutdown(struct lnet_ni *ni)
3022 struct kib_net *net = ni->ni_data;
3023 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
3025 unsigned long flags;
3027 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3032 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
3033 atomic_read(&libcfs_kmemory));
3035 write_lock_irqsave(g_lock, flags);
3036 net->ibn_shutdown = 1;
3037 write_unlock_irqrestore(g_lock, flags);
3039 switch (net->ibn_init) {
3043 case IBLND_INIT_ALL:
3044 /* nuke all existing peers within this net */
3045 kiblnd_del_peer(ni, LNET_NID_ANY);
3047 /* Wait for all peer_ni state to clean up */
3049 while (atomic_read(&net->ibn_npeers) != 0) {
3052 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3053 "%s: waiting for %d peers to disconnect\n",
3054 libcfs_nid2str(ni->ni_nid),
3055 atomic_read(&net->ibn_npeers));
3056 set_current_state(TASK_UNINTERRUPTIBLE);
3057 schedule_timeout(cfs_time_seconds(1));
3060 kiblnd_net_fini_pools(net);
3062 write_lock_irqsave(g_lock, flags);
3063 LASSERT(net->ibn_dev->ibd_nnets > 0);
3064 net->ibn_dev->ibd_nnets--;
3065 list_del(&net->ibn_list);
3066 write_unlock_irqrestore(g_lock, flags);
3070 case IBLND_INIT_NOTHING:
3071 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3073 if (net->ibn_dev != NULL &&
3074 net->ibn_dev->ibd_nnets == 0)
3075 kiblnd_destroy_dev(net->ibn_dev);
3080 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3081 atomic_read(&libcfs_kmemory));
3083 net->ibn_init = IBLND_INIT_NOTHING;
3086 LIBCFS_FREE(net, sizeof(*net));
3089 if (list_empty(&kiblnd_data.kib_devs))
3090 kiblnd_base_shutdown();
3095 kiblnd_base_startup(void)
3097 struct kib_sched_info *sched;
3101 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3103 try_module_get(THIS_MODULE);
3104 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3106 rwlock_init(&kiblnd_data.kib_global_lock);
3108 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3109 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3111 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3112 LIBCFS_ALLOC(kiblnd_data.kib_peers,
3113 sizeof(struct list_head) *
3114 kiblnd_data.kib_peer_hash_size);
3115 if (kiblnd_data.kib_peers == NULL)
3118 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3119 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3121 spin_lock_init(&kiblnd_data.kib_connd_lock);
3122 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3123 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3124 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3125 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3127 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3128 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3130 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3132 if (kiblnd_data.kib_scheds == NULL)
3135 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3138 spin_lock_init(&sched->ibs_lock);
3139 INIT_LIST_HEAD(&sched->ibs_conns);
3140 init_waitqueue_head(&sched->ibs_waitq);
3142 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3143 if (*kiblnd_tunables.kib_nscheds > 0) {
3144 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3146 /* max to half of CPUs, another half is reserved for
3147 * upper layer modules */
3148 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3151 sched->ibs_nthreads_max = nthrs;
3155 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3157 /* lists/ptrs/locks initialised */
3158 kiblnd_data.kib_init = IBLND_INIT_DATA;
3159 /*****************************************************/
3161 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3163 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3167 if (*kiblnd_tunables.kib_dev_failover != 0)
3168 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3172 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3176 /* flag everything initialised */
3177 kiblnd_data.kib_init = IBLND_INIT_ALL;
3178 /*****************************************************/
3183 kiblnd_base_shutdown();
3188 kiblnd_start_schedulers(struct kib_sched_info *sched)
3194 if (sched->ibs_nthreads == 0) {
3195 if (*kiblnd_tunables.kib_nscheds > 0) {
3196 nthrs = sched->ibs_nthreads_max;
3198 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3200 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3201 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3204 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3205 /* increase one thread if there is new interface */
3206 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3209 for (i = 0; i < nthrs; i++) {
3212 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3213 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3214 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3215 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3219 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3220 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3224 sched->ibs_nthreads += i;
3229 kiblnd_dev_start_threads(struct kib_dev *dev, int newdev, u32 *cpts, int ncpts)
3235 for (i = 0; i < ncpts; i++) {
3236 struct kib_sched_info *sched;
3238 cpt = (cpts == NULL) ? i : cpts[i];
3239 sched = kiblnd_data.kib_scheds[cpt];
3241 if (!newdev && sched->ibs_nthreads > 0)
3244 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3246 CERROR("Failed to start scheduler threads for %s\n",
3254 static struct kib_dev *
3255 kiblnd_dev_search(char *ifname)
3257 struct kib_dev *alias = NULL;
3258 struct kib_dev *dev;
3262 colon = strchr(ifname, ':');
3263 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3264 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3270 colon2 = strchr(dev->ibd_ifname, ':');
3276 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3288 kiblnd_startup(struct lnet_ni *ni)
3291 struct kib_dev *ibdev = NULL;
3292 struct kib_net *net;
3293 unsigned long flags;
3298 LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3300 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3301 rc = kiblnd_base_startup();
3306 LIBCFS_ALLOC(net, sizeof(*net));
3311 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3313 kiblnd_tunables_setup(ni);
3316 * ni_interfaces is only to support legacy pre Multi-Rail
3317 * tcp bonding for ksocklnd. Multi-Rail wants each secondary
3318 * IP to be treated as an unique 'struct ni' interfaces instead.
3320 if (ni->ni_interfaces[0] != NULL) {
3321 /* Use the IPoIB interface specified in 'networks=' */
3323 CLASSERT(LNET_INTERFACES_NUM > 1);
3324 if (ni->ni_interfaces[1] != NULL) {
3325 CERROR("Multiple interfaces not supported\n");
3329 ifname = ni->ni_interfaces[0];
3331 ifname = *kiblnd_tunables.kib_default_ipif;
3334 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3335 CERROR("IPoIB interface name too long: %s\n", ifname);
3339 ibdev = kiblnd_dev_search(ifname);
3341 newdev = ibdev == NULL;
3342 /* hmm...create kib_dev even for alias */
3343 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3344 ibdev = kiblnd_create_dev(ifname);
3349 node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3350 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3352 net->ibn_dev = ibdev;
3353 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3355 rc = kiblnd_dev_start_threads(ibdev, newdev,
3356 ni->ni_cpts, ni->ni_ncpts);
3360 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3362 CERROR("Failed to initialize NI pools: %d\n", rc);
3366 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3368 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3369 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3371 net->ibn_init = IBLND_INIT_ALL;
3376 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3377 kiblnd_destroy_dev(ibdev);
3379 kiblnd_shutdown(ni);
3381 CDEBUG(D_NET, "kiblnd_startup failed\n");
3385 static struct lnet_lnd the_o2iblnd = {
3386 .lnd_type = O2IBLND,
3387 .lnd_startup = kiblnd_startup,
3388 .lnd_shutdown = kiblnd_shutdown,
3389 .lnd_ctl = kiblnd_ctl,
3390 .lnd_query = kiblnd_query,
3391 .lnd_send = kiblnd_send,
3392 .lnd_recv = kiblnd_recv,
3395 static void __exit ko2iblnd_exit(void)
3397 lnet_unregister_lnd(&the_o2iblnd);
3400 static int __init ko2iblnd_init(void)
3404 CLASSERT(sizeof(struct kib_msg) <= IBLND_MSG_SIZE);
3405 CLASSERT(offsetof(struct kib_msg,
3406 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3408 CLASSERT(offsetof(struct kib_msg,
3409 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3412 rc = kiblnd_tunables_init();
3416 lnet_register_lnd(&the_o2iblnd);
3421 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3422 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3423 MODULE_VERSION("2.8.0");
3424 MODULE_LICENSE("GPL");
3426 module_init(ko2iblnd_init);
3427 module_exit(ko2iblnd_exit);