4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd.c
34 * Author: Eric Barton <eric@bartonsoftware.com>
38 #include <linux/inetdevice.h>
42 static const struct lnet_lnd the_o2iblnd;
44 struct kib_data kiblnd_data;
47 kiblnd_cksum (void *ptr, int nob)
53 sum = ((sum << 1) | (sum >> 31)) + *c++;
55 /* ensure I don't return 0 (== no checksum) */
56 return (sum == 0) ? 1 : sum;
60 kiblnd_msgtype2str(int type)
63 case IBLND_MSG_CONNREQ:
66 case IBLND_MSG_CONNACK:
72 case IBLND_MSG_IMMEDIATE:
75 case IBLND_MSG_PUT_REQ:
78 case IBLND_MSG_PUT_NAK:
81 case IBLND_MSG_PUT_ACK:
84 case IBLND_MSG_PUT_DONE:
87 case IBLND_MSG_GET_REQ:
90 case IBLND_MSG_GET_DONE:
99 kiblnd_msgtype2size(int type)
101 const int hdr_size = offsetof(struct kib_msg, ibm_u);
104 case IBLND_MSG_CONNREQ:
105 case IBLND_MSG_CONNACK:
106 return hdr_size + sizeof(struct kib_connparams);
111 case IBLND_MSG_IMMEDIATE:
112 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
114 case IBLND_MSG_PUT_REQ:
115 return hdr_size + sizeof(struct kib_putreq_msg);
117 case IBLND_MSG_PUT_ACK:
118 return hdr_size + sizeof(struct kib_putack_msg);
120 case IBLND_MSG_GET_REQ:
121 return hdr_size + sizeof(struct kib_get_msg);
123 case IBLND_MSG_PUT_NAK:
124 case IBLND_MSG_PUT_DONE:
125 case IBLND_MSG_GET_DONE:
126 return hdr_size + sizeof(struct kib_completion_msg);
132 static int kiblnd_unpack_rd(struct kib_msg *msg, int flip)
134 struct kib_rdma_desc *rd;
139 LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
140 msg->ibm_type == IBLND_MSG_PUT_ACK);
142 rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
143 &msg->ibm_u.get.ibgm_rd :
144 &msg->ibm_u.putack.ibpam_rd;
147 __swab32s(&rd->rd_key);
148 __swab32s(&rd->rd_nfrags);
153 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
154 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
155 n, IBLND_MAX_RDMA_FRAGS);
159 nob = offsetof(struct kib_msg, ibm_u) +
160 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
162 if (msg->ibm_nob < nob) {
163 CERROR("Short %s: %d(%d)\n",
164 kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
171 for (i = 0; i < n; i++) {
172 __swab32s(&rd->rd_frags[i].rf_nob);
173 __swab64s(&rd->rd_frags[i].rf_addr);
179 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
180 int credits, lnet_nid_t dstnid, __u64 dststamp)
182 struct kib_net *net = ni->ni_data;
184 /* CAVEAT EMPTOR! all message fields not set here should have been
185 * initialised previously. */
186 msg->ibm_magic = IBLND_MSG_MAGIC;
187 msg->ibm_version = version;
189 msg->ibm_credits = credits;
192 msg->ibm_srcnid = ni->ni_nid;
193 msg->ibm_srcstamp = net->ibn_incarnation;
194 msg->ibm_dstnid = dstnid;
195 msg->ibm_dststamp = dststamp;
197 if (*kiblnd_tunables.kib_cksum) {
198 /* NB ibm_cksum zero while computing cksum */
199 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
203 int kiblnd_unpack_msg(struct kib_msg *msg, int nob)
205 const int hdr_size = offsetof(struct kib_msg, ibm_u);
211 /* 6 bytes are enough to have received magic + version */
213 CERROR("Short message: %d\n", nob);
217 if (msg->ibm_magic == IBLND_MSG_MAGIC) {
219 } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
222 CERROR("Bad magic: %08x\n", msg->ibm_magic);
226 version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
227 if (version != IBLND_MSG_VERSION &&
228 version != IBLND_MSG_VERSION_1) {
229 CERROR("Bad version: %x\n", version);
233 if (nob < hdr_size) {
234 CERROR("Short message: %d\n", nob);
238 msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
240 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
244 /* checksum must be computed with ibm_cksum zero and BEFORE anything
246 msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
248 if (msg_cksum != 0 &&
249 msg_cksum != kiblnd_cksum(msg, msg_nob)) {
250 CERROR("Bad checksum\n");
254 msg->ibm_cksum = msg_cksum;
257 /* leave magic unflipped as a clue to peer_ni endianness */
258 msg->ibm_version = version;
259 BUILD_BUG_ON(sizeof(msg->ibm_type) != 1);
260 BUILD_BUG_ON(sizeof(msg->ibm_credits) != 1);
261 msg->ibm_nob = msg_nob;
262 __swab64s(&msg->ibm_srcnid);
263 __swab64s(&msg->ibm_srcstamp);
264 __swab64s(&msg->ibm_dstnid);
265 __swab64s(&msg->ibm_dststamp);
268 if (msg->ibm_srcnid == LNET_NID_ANY) {
269 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
273 if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
274 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
275 msg_nob, kiblnd_msgtype2size(msg->ibm_type));
279 switch (msg->ibm_type) {
281 CERROR("Unknown message type %x\n", msg->ibm_type);
285 case IBLND_MSG_IMMEDIATE:
286 case IBLND_MSG_PUT_REQ:
289 case IBLND_MSG_PUT_ACK:
290 case IBLND_MSG_GET_REQ:
291 if (kiblnd_unpack_rd(msg, flip))
295 case IBLND_MSG_PUT_NAK:
296 case IBLND_MSG_PUT_DONE:
297 case IBLND_MSG_GET_DONE:
299 __swab32s(&msg->ibm_u.completion.ibcm_status);
302 case IBLND_MSG_CONNREQ:
303 case IBLND_MSG_CONNACK:
305 __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
306 __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
307 __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
315 kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
318 struct kib_peer_ni *peer_ni;
319 struct kib_net *net = ni->ni_data;
320 int cpt = lnet_cpt_of_nid(nid, ni);
323 LASSERT(net != NULL);
324 LASSERT(nid != LNET_NID_ANY);
326 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
327 if (peer_ni == NULL) {
328 CERROR("Cannot allocate peer_ni\n");
332 peer_ni->ibp_ni = ni;
333 peer_ni->ibp_nid = nid;
334 peer_ni->ibp_error = 0;
335 peer_ni->ibp_last_alive = 0;
336 peer_ni->ibp_max_frags = IBLND_MAX_RDMA_FRAGS;
337 peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
338 atomic_set(&peer_ni->ibp_refcount, 1); /* 1 ref for caller */
340 INIT_LIST_HEAD(&peer_ni->ibp_list); /* not in the peer_ni table yet */
341 INIT_LIST_HEAD(&peer_ni->ibp_conns);
342 INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
344 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
346 /* always called with a ref on ni, which prevents ni being shutdown */
347 LASSERT(net->ibn_shutdown == 0);
349 /* npeers only grows with the global lock held */
350 atomic_inc(&net->ibn_npeers);
352 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
359 kiblnd_destroy_peer(struct kib_peer_ni *peer_ni)
361 struct kib_net *net = peer_ni->ibp_ni->ni_data;
363 LASSERT(net != NULL);
364 LASSERT (atomic_read(&peer_ni->ibp_refcount) == 0);
365 LASSERT(!kiblnd_peer_active(peer_ni));
366 LASSERT(kiblnd_peer_idle(peer_ni));
367 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
369 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
371 /* NB a peer_ni's connections keep a reference on their peer_ni until
372 * they are destroyed, so we can be assured that _all_ state to do
373 * with this peer_ni has been cleaned up when its refcount drops to
375 if (atomic_dec_and_test(&net->ibn_npeers))
376 wake_up_var(&net->ibn_npeers);
380 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
382 /* the caller is responsible for accounting the additional reference
383 * that this creates */
384 struct list_head *peer_list = kiblnd_nid2peerlist(nid);
385 struct list_head *tmp;
386 struct kib_peer_ni *peer_ni;
388 list_for_each(tmp, peer_list) {
390 peer_ni = list_entry(tmp, struct kib_peer_ni, ibp_list);
391 LASSERT(!kiblnd_peer_idle(peer_ni));
394 * Match a peer if its NID and the NID of the local NI it
395 * communicates over are the same. Otherwise don't match
396 * the peer, which will result in a new lnd peer being
399 if (peer_ni->ibp_nid != nid ||
400 peer_ni->ibp_ni->ni_nid != ni->ni_nid)
403 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
404 peer_ni, libcfs_nid2str(nid),
405 atomic_read(&peer_ni->ibp_refcount),
406 peer_ni->ibp_version);
413 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
415 LASSERT(list_empty(&peer_ni->ibp_conns));
417 LASSERT (kiblnd_peer_active(peer_ni));
418 list_del_init(&peer_ni->ibp_list);
419 /* lose peerlist's ref */
420 kiblnd_peer_decref(peer_ni);
424 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
425 lnet_nid_t *nidp, int *count)
427 struct kib_peer_ni *peer_ni;
428 struct list_head *ptmp;
432 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
434 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
436 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
438 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
439 LASSERT(!kiblnd_peer_idle(peer_ni));
441 if (peer_ni->ibp_ni != ni)
447 *nidp = peer_ni->ibp_nid;
448 *count = atomic_read(&peer_ni->ibp_refcount);
450 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
456 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
461 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
463 struct list_head *ctmp;
464 struct list_head *cnxt;
465 struct kib_conn *conn;
467 if (list_empty(&peer_ni->ibp_conns)) {
468 kiblnd_unlink_peer_locked(peer_ni);
470 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
471 conn = list_entry(ctmp, struct kib_conn, ibc_list);
473 kiblnd_close_conn_locked(conn, 0);
475 /* NB closing peer_ni's last conn unlinked it. */
477 /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
482 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
485 struct list_head *ptmp;
486 struct list_head *pnxt;
487 struct kib_peer_ni *peer_ni;
494 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
496 if (nid != LNET_NID_ANY) {
497 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
500 hi = kiblnd_data.kib_peer_hash_size - 1;
503 for (i = lo; i <= hi; i++) {
504 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
505 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
506 LASSERT(!kiblnd_peer_idle(peer_ni));
508 if (peer_ni->ibp_ni != ni)
511 if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
514 if (!list_empty(&peer_ni->ibp_tx_queue)) {
515 LASSERT(list_empty(&peer_ni->ibp_conns));
517 list_splice_init(&peer_ni->ibp_tx_queue,
521 kiblnd_del_peer_locked(peer_ni);
522 rc = 0; /* matched something */
526 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
528 kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
533 static struct kib_conn *
534 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
536 struct kib_peer_ni *peer_ni;
537 struct list_head *ptmp;
538 struct kib_conn *conn;
539 struct list_head *ctmp;
543 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
545 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
546 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
548 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
549 LASSERT(!kiblnd_peer_idle(peer_ni));
551 if (peer_ni->ibp_ni != ni)
554 list_for_each(ctmp, &peer_ni->ibp_conns) {
558 conn = list_entry(ctmp, struct kib_conn, ibc_list);
559 kiblnd_conn_addref(conn);
560 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
567 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
572 kiblnd_debug_rx(struct kib_rx *rx)
574 CDEBUG(D_CONSOLE, " %p msg_type %x cred %d\n",
575 rx, rx->rx_msg->ibm_type,
576 rx->rx_msg->ibm_credits);
580 kiblnd_debug_tx(struct kib_tx *tx)
582 CDEBUG(D_CONSOLE, " %p snd %d q %d w %d rc %d dl %lld "
583 "cookie %#llx msg %s%s type %x cred %d\n",
584 tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
585 tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
586 tx->tx_lntmsg[0] == NULL ? "-" : "!",
587 tx->tx_lntmsg[1] == NULL ? "-" : "!",
588 tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
592 kiblnd_debug_conn(struct kib_conn *conn)
594 struct list_head *tmp;
597 spin_lock(&conn->ibc_lock);
599 CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
600 atomic_read(&conn->ibc_refcount), conn,
601 conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
602 CDEBUG(D_CONSOLE, " state %d nposted %d/%d cred %d o_cred %d "
603 " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
604 conn->ibc_nsends_posted, conn->ibc_credits,
605 conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
606 CDEBUG(D_CONSOLE, " comms_err %d\n", conn->ibc_comms_error);
608 CDEBUG(D_CONSOLE, " early_rxs:\n");
609 list_for_each(tmp, &conn->ibc_early_rxs)
610 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
612 CDEBUG(D_CONSOLE, " tx_noops:\n");
613 list_for_each(tmp, &conn->ibc_tx_noops)
614 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
616 CDEBUG(D_CONSOLE, " tx_queue_nocred:\n");
617 list_for_each(tmp, &conn->ibc_tx_queue_nocred)
618 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
620 CDEBUG(D_CONSOLE, " tx_queue_rsrvd:\n");
621 list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
622 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
624 CDEBUG(D_CONSOLE, " tx_queue:\n");
625 list_for_each(tmp, &conn->ibc_tx_queue)
626 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
628 CDEBUG(D_CONSOLE, " active_txs:\n");
629 list_for_each(tmp, &conn->ibc_active_txs)
630 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
632 CDEBUG(D_CONSOLE, " rxs:\n");
633 for (i = 0; i < IBLND_RX_MSGS(conn); i++)
634 kiblnd_debug_rx(&conn->ibc_rxs[i]);
636 spin_unlock(&conn->ibc_lock);
640 kiblnd_translate_mtu(int value)
661 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
665 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
666 if (cmid->route.path_rec == NULL)
669 mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
672 cmid->route.path_rec->mtu = mtu;
676 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
684 vectors = conn->ibc_cmid->device->num_comp_vectors;
688 mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
690 /* hash NID to CPU id in this partition... */
691 ibp_nid = conn->ibc_peer->ibp_nid;
692 off = do_div(ibp_nid, cpumask_weight(*mask));
693 for_each_cpu(i, *mask) {
703 * Get the scheduler bound to this CPT. If the scheduler has no
704 * threads, which means that the CPT has no CPUs, then grab the
705 * next scheduler that we can use.
707 * This case would be triggered if a NUMA node is configured with
708 * no associated CPUs.
710 static struct kib_sched_info *
711 kiblnd_get_scheduler(int cpt)
713 struct kib_sched_info *sched;
716 sched = kiblnd_data.kib_scheds[cpt];
718 if (sched->ibs_nthreads > 0)
721 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
722 if (sched->ibs_nthreads > 0) {
723 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
724 cpt, sched->ibs_cpt);
732 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
735 * One WR for the LNet message
736 * And ibc_max_frags for the transfer WRs
739 int multiplier = 1 + conn->ibc_max_frags;
740 enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
742 /* FastReg needs two extra WRs for map and invalidate */
743 if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
746 /* account for a maximum of ibc_queue_depth in-flight transfers */
747 ret = multiplier * conn->ibc_queue_depth;
749 if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
750 CDEBUG(D_NET, "peer_credits %u will result in send work "
751 "request size %d larger than maximum %d device "
752 "can handle\n", conn->ibc_queue_depth, ret,
753 conn->ibc_hdev->ibh_max_qp_wr);
754 conn->ibc_queue_depth =
755 conn->ibc_hdev->ibh_max_qp_wr / multiplier;
758 /* don't go beyond the maximum the device can handle */
759 return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
763 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
764 int state, int version)
767 * If the new conn is created successfully it takes over the caller's
768 * ref on 'peer_ni'. It also "owns" 'cmid' and destroys it when it itself
769 * is destroyed. On failure, the caller's ref on 'peer_ni' remains and
770 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
771 * to destroy 'cmid' here since I'm called from the CM which still has
772 * its ref on 'cmid'). */
773 rwlock_t *glock = &kiblnd_data.kib_global_lock;
774 struct kib_net *net = peer_ni->ibp_ni->ni_data;
776 struct ib_qp_init_attr *init_qp_attr;
777 struct kib_sched_info *sched;
778 #ifdef HAVE_IB_CQ_INIT_ATTR
779 struct ib_cq_init_attr cq_attr = {};
781 struct kib_conn *conn;
788 LASSERT(net != NULL);
789 LASSERT(!in_interrupt());
793 cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
794 sched = kiblnd_get_scheduler(cpt);
797 CERROR("no schedulers available. node is unhealthy\n");
802 * The cpt might have changed if we ended up selecting a non cpt
803 * native scheduler. So use the scheduler's cpt instead.
805 cpt = sched->ibs_cpt;
807 LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
808 sizeof(*init_qp_attr));
809 if (init_qp_attr == NULL) {
810 CERROR("Can't allocate qp_attr for %s\n",
811 libcfs_nid2str(peer_ni->ibp_nid));
815 LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
817 CERROR("Can't allocate connection for %s\n",
818 libcfs_nid2str(peer_ni->ibp_nid));
822 conn->ibc_state = IBLND_CONN_INIT;
823 conn->ibc_version = version;
824 conn->ibc_peer = peer_ni; /* I take the caller's ref */
825 cmid->context = conn; /* for future CM callbacks */
826 conn->ibc_cmid = cmid;
827 conn->ibc_max_frags = peer_ni->ibp_max_frags;
828 conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
829 conn->ibc_rxs = NULL;
830 conn->ibc_rx_pages = NULL;
832 INIT_LIST_HEAD(&conn->ibc_early_rxs);
833 INIT_LIST_HEAD(&conn->ibc_tx_noops);
834 INIT_LIST_HEAD(&conn->ibc_tx_queue);
835 INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
836 INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
837 INIT_LIST_HEAD(&conn->ibc_active_txs);
838 INIT_LIST_HEAD(&conn->ibc_zombie_txs);
839 spin_lock_init(&conn->ibc_lock);
841 LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
842 sizeof(*conn->ibc_connvars));
843 if (conn->ibc_connvars == NULL) {
844 CERROR("Can't allocate in-progress connection state\n");
848 write_lock_irqsave(glock, flags);
849 if (dev->ibd_failover) {
850 write_unlock_irqrestore(glock, flags);
851 CERROR("%s: failover in progress\n", dev->ibd_ifname);
855 if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
856 /* wakeup failover thread and teardown connection */
857 if (kiblnd_dev_can_failover(dev)) {
858 list_add_tail(&dev->ibd_fail_list,
859 &kiblnd_data.kib_failed_devs);
860 wake_up(&kiblnd_data.kib_failover_waitq);
863 write_unlock_irqrestore(glock, flags);
864 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
865 cmid->device->name, dev->ibd_ifname);
869 kiblnd_hdev_addref_locked(dev->ibd_hdev);
870 conn->ibc_hdev = dev->ibd_hdev;
872 kiblnd_setup_mtu_locked(cmid);
874 write_unlock_irqrestore(glock, flags);
876 #ifdef HAVE_IB_CQ_INIT_ATTR
877 cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
878 cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
879 cq = ib_create_cq(cmid->device,
880 kiblnd_cq_completion, kiblnd_cq_event, conn,
883 cq = ib_create_cq(cmid->device,
884 kiblnd_cq_completion, kiblnd_cq_event, conn,
885 IBLND_CQ_ENTRIES(conn),
886 kiblnd_get_completion_vector(conn, cpt));
890 * on MLX-5 (possibly MLX-4 as well) this error could be
891 * hit if the concurrent_sends and/or peer_tx_credits is set
892 * too high. Or due to an MLX-5 bug which tries to
893 * allocate 256kb via kmalloc for WR cookie array
895 CERROR("Failed to create CQ with %d CQEs: %ld\n",
896 IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
902 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
904 CERROR("Can't request completion notification: %d\n", rc);
908 init_qp_attr->event_handler = kiblnd_qp_event;
909 init_qp_attr->qp_context = conn;
910 init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
911 init_qp_attr->cap.max_recv_sge = 1;
912 init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
913 init_qp_attr->qp_type = IB_QPT_RC;
914 init_qp_attr->send_cq = cq;
915 init_qp_attr->recv_cq = cq;
917 * kiblnd_send_wrs() can change the connection's queue depth if
918 * the maximum work requests for the device is maxed out
920 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
921 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
923 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
925 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
926 "send_sge: %d, recv_sge: %d\n",
927 rc, init_qp_attr->cap.max_send_wr,
928 init_qp_attr->cap.max_recv_wr,
929 init_qp_attr->cap.max_send_sge,
930 init_qp_attr->cap.max_recv_sge);
934 conn->ibc_sched = sched;
936 if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
937 CWARN("peer %s - queue depth reduced from %u to %u"
938 " to allow for qp creation\n",
939 libcfs_nid2str(peer_ni->ibp_nid),
940 peer_ni->ibp_queue_depth,
941 conn->ibc_queue_depth);
943 LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
944 IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
945 if (conn->ibc_rxs == NULL) {
946 CERROR("Cannot allocate RX buffers\n");
950 rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
951 IBLND_RX_MSG_PAGES(conn));
955 kiblnd_map_rx_descs(conn);
957 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
959 /* 1 ref for caller and each rxmsg */
960 atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
961 conn->ibc_nrx = IBLND_RX_MSGS(conn);
964 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
965 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
967 CERROR("Can't post rxmsg: %d\n", rc);
969 /* Make posted receives complete */
970 kiblnd_abort_receives(conn);
972 /* correct # of posted buffers
973 * NB locking needed now I'm racing with completion */
974 spin_lock_irqsave(&sched->ibs_lock, flags);
975 conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
976 spin_unlock_irqrestore(&sched->ibs_lock, flags);
978 /* cmid will be destroyed by CM(ofed) after cm_callback
979 * returned, so we can't refer it anymore
980 * (by kiblnd_connd()->kiblnd_destroy_conn) */
981 rdma_destroy_qp(conn->ibc_cmid);
982 conn->ibc_cmid = NULL;
984 /* Drop my own and unused rxbuffer refcounts */
985 while (i++ <= IBLND_RX_MSGS(conn))
986 kiblnd_conn_decref(conn);
992 /* Init successful! */
993 LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
994 state == IBLND_CONN_PASSIVE_WAIT);
995 conn->ibc_state = state;
998 atomic_inc(&net->ibn_nconns);
1002 kiblnd_destroy_conn(conn);
1003 LIBCFS_FREE(conn, sizeof(*conn));
1005 LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
1011 kiblnd_destroy_conn(struct kib_conn *conn)
1013 struct rdma_cm_id *cmid = conn->ibc_cmid;
1014 struct kib_peer_ni *peer_ni = conn->ibc_peer;
1016 LASSERT (!in_interrupt());
1017 LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1018 LASSERT(list_empty(&conn->ibc_early_rxs));
1019 LASSERT(list_empty(&conn->ibc_tx_noops));
1020 LASSERT(list_empty(&conn->ibc_tx_queue));
1021 LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1022 LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1023 LASSERT(list_empty(&conn->ibc_active_txs));
1024 LASSERT (conn->ibc_noops_posted == 0);
1025 LASSERT (conn->ibc_nsends_posted == 0);
1027 switch (conn->ibc_state) {
1029 /* conn must be completely disengaged from the network */
1032 case IBLND_CONN_DISCONNECTED:
1033 /* connvars should have been freed already */
1034 LASSERT (conn->ibc_connvars == NULL);
1037 case IBLND_CONN_INIT:
1041 /* conn->ibc_cmid might be destroyed by CM already */
1042 if (cmid != NULL && cmid->qp != NULL)
1043 rdma_destroy_qp(cmid);
1046 ib_destroy_cq(conn->ibc_cq);
1048 kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1049 LNET_MSG_STATUS_OK);
1051 if (conn->ibc_rx_pages != NULL)
1052 kiblnd_unmap_rx_descs(conn);
1054 if (conn->ibc_rxs != NULL)
1055 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1057 if (conn->ibc_connvars != NULL)
1058 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1060 if (conn->ibc_hdev != NULL)
1061 kiblnd_hdev_decref(conn->ibc_hdev);
1063 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1064 if (conn->ibc_state != IBLND_CONN_INIT) {
1065 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1067 kiblnd_peer_decref(peer_ni);
1068 rdma_destroy_id(cmid);
1069 atomic_dec(&net->ibn_nconns);
1074 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1076 struct kib_conn *conn;
1077 struct list_head *ctmp;
1078 struct list_head *cnxt;
1081 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1082 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1084 CDEBUG(D_NET, "Closing conn -> %s, "
1085 "version: %x, reason: %d\n",
1086 libcfs_nid2str(peer_ni->ibp_nid),
1087 conn->ibc_version, why);
1089 kiblnd_close_conn_locked(conn, why);
1097 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1098 int version, __u64 incarnation)
1100 struct kib_conn *conn;
1101 struct list_head *ctmp;
1102 struct list_head *cnxt;
1105 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1106 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1108 if (conn->ibc_version == version &&
1109 conn->ibc_incarnation == incarnation)
1112 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1113 "incarnation:%#llx(%x, %#llx)\n",
1114 libcfs_nid2str(peer_ni->ibp_nid),
1115 conn->ibc_version, conn->ibc_incarnation,
1116 version, incarnation);
1118 kiblnd_close_conn_locked(conn, -ESTALE);
1126 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1128 struct kib_peer_ni *peer_ni;
1129 struct list_head *ptmp;
1130 struct list_head *pnxt;
1134 unsigned long flags;
1137 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1139 if (nid != LNET_NID_ANY)
1140 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1143 hi = kiblnd_data.kib_peer_hash_size - 1;
1146 for (i = lo; i <= hi; i++) {
1147 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1149 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
1150 LASSERT(!kiblnd_peer_idle(peer_ni));
1152 if (peer_ni->ibp_ni != ni)
1155 if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1158 count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1162 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1164 /* wildcards always succeed */
1165 if (nid == LNET_NID_ANY)
1168 return (count == 0) ? -ENOENT : 0;
1172 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1174 struct libcfs_ioctl_data *data = arg;
1178 case IOC_LIBCFS_GET_PEER: {
1182 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1184 data->ioc_nid = nid;
1185 data->ioc_count = count;
1189 case IOC_LIBCFS_DEL_PEER: {
1190 rc = kiblnd_del_peer(ni, data->ioc_nid);
1193 case IOC_LIBCFS_GET_CONN: {
1194 struct kib_conn *conn;
1197 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1203 LASSERT(conn->ibc_cmid != NULL);
1204 data->ioc_nid = conn->ibc_peer->ibp_nid;
1205 if (conn->ibc_cmid->route.path_rec == NULL)
1206 data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1209 ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1210 kiblnd_conn_decref(conn);
1213 case IOC_LIBCFS_CLOSE_CONNECTION: {
1214 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1226 kiblnd_free_pages(struct kib_pages *p)
1228 int npages = p->ibp_npages;
1231 for (i = 0; i < npages; i++) {
1232 if (p->ibp_pages[i] != NULL)
1233 __free_page(p->ibp_pages[i]);
1236 LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1240 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1242 struct kib_pages *p;
1245 LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1246 offsetof(struct kib_pages, ibp_pages[npages]));
1248 CERROR("Can't allocate descriptor for %d pages\n", npages);
1252 memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1253 p->ibp_npages = npages;
1255 for (i = 0; i < npages; i++) {
1256 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1258 if (p->ibp_pages[i] == NULL) {
1259 CERROR("Can't allocate page %d of %d\n", i, npages);
1260 kiblnd_free_pages(p);
1270 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1275 LASSERT (conn->ibc_rxs != NULL);
1276 LASSERT (conn->ibc_hdev != NULL);
1278 for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1279 rx = &conn->ibc_rxs[i];
1281 LASSERT(rx->rx_nob >= 0); /* not posted */
1283 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1284 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1286 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1289 kiblnd_free_pages(conn->ibc_rx_pages);
1291 conn->ibc_rx_pages = NULL;
1295 kiblnd_map_rx_descs(struct kib_conn *conn)
1303 for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1304 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1305 rx = &conn->ibc_rxs[i];
1308 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1311 kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1312 rx->rx_msg, IBLND_MSG_SIZE,
1314 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1316 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1318 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1319 i, rx->rx_msg, rx->rx_msgaddr,
1320 (__u64)(page_to_phys(pg) + pg_off));
1322 pg_off += IBLND_MSG_SIZE;
1323 LASSERT(pg_off <= PAGE_SIZE);
1325 if (pg_off == PAGE_SIZE) {
1328 LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1334 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1336 struct kib_hca_dev *hdev = tpo->tpo_hdev;
1340 LASSERT (tpo->tpo_pool.po_allocated == 0);
1345 for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1346 tx = &tpo->tpo_tx_descs[i];
1347 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1348 KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1350 IBLND_MSG_SIZE, DMA_TO_DEVICE);
1353 kiblnd_hdev_decref(hdev);
1354 tpo->tpo_hdev = NULL;
1357 static struct kib_hca_dev *
1358 kiblnd_current_hdev(struct kib_dev *dev)
1360 struct kib_hca_dev *hdev;
1361 unsigned long flags;
1364 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1365 while (dev->ibd_failover) {
1366 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1368 CDEBUG(D_NET, "%s: Wait for failover\n",
1370 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1372 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1375 kiblnd_hdev_addref_locked(dev->ibd_hdev);
1376 hdev = dev->ibd_hdev;
1378 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1384 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1386 struct kib_pages *txpgs = tpo->tpo_tx_pages;
1387 struct kib_pool *pool = &tpo->tpo_pool;
1388 struct kib_net *net = pool->po_owner->ps_net;
1389 struct kib_dev *dev;
1396 LASSERT (net != NULL);
1400 /* pre-mapped messages are not bigger than 1 page */
1401 BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1403 /* No fancy arithmetic when we do the buffer calculations */
1404 BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1406 tpo->tpo_hdev = kiblnd_current_hdev(dev);
1408 for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1409 page = txpgs->ibp_pages[ipage];
1410 tx = &tpo->tpo_tx_descs[i];
1412 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1415 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1419 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1421 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1423 list_add(&tx->tx_list, &pool->po_free_list);
1425 page_offset += IBLND_MSG_SIZE;
1426 LASSERT(page_offset <= PAGE_SIZE);
1428 if (page_offset == PAGE_SIZE) {
1431 LASSERT(ipage <= txpgs->ibp_npages);
1437 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1439 LASSERT(fpo->fpo_map_count == 0);
1441 if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1442 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1444 struct kib_fast_reg_descriptor *frd, *tmp;
1447 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1449 list_del(&frd->frd_list);
1450 #ifndef HAVE_IB_MAP_MR_SG
1451 ib_free_fast_reg_page_list(frd->frd_frpl);
1453 ib_dereg_mr(frd->frd_mr);
1454 LIBCFS_FREE(frd, sizeof(*frd));
1457 if (i < fpo->fast_reg.fpo_pool_size)
1458 CERROR("FastReg pool still has %d regions registered\n",
1459 fpo->fast_reg.fpo_pool_size - i);
1463 kiblnd_hdev_decref(fpo->fpo_hdev);
1465 LIBCFS_FREE(fpo, sizeof(*fpo));
1469 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1471 struct kib_fmr_pool *fpo, *tmp;
1473 list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1474 list_del(&fpo->fpo_list);
1475 kiblnd_destroy_fmr_pool(fpo);
1480 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1483 int size = tunables->lnd_fmr_pool_size / ncpts;
1485 return max(IBLND_FMR_POOL, size);
1489 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1492 int size = tunables->lnd_fmr_flush_trigger / ncpts;
1494 return max(IBLND_FMR_POOL_FLUSH, size);
1497 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1498 struct kib_fmr_pool *fpo)
1500 struct ib_fmr_pool_param param = {
1501 .max_pages_per_fmr = LNET_MAX_IOV,
1502 .page_shift = PAGE_SHIFT,
1503 .access = (IB_ACCESS_LOCAL_WRITE |
1504 IB_ACCESS_REMOTE_WRITE),
1505 .pool_size = fps->fps_pool_size,
1506 .dirty_watermark = fps->fps_flush_trigger,
1507 .flush_function = NULL,
1509 .cache = !!fps->fps_cache };
1512 fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1514 if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1515 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1517 CERROR("Failed to create FMR pool: %d\n", rc);
1519 CERROR("FMRs are not supported\n");
1521 fpo->fpo_is_fmr = true;
1526 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1527 struct kib_fmr_pool *fpo,
1528 enum kib_dev_caps dev_caps)
1530 struct kib_fast_reg_descriptor *frd, *tmp;
1533 fpo->fpo_is_fmr = false;
1535 INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1536 fpo->fast_reg.fpo_pool_size = 0;
1537 for (i = 0; i < fps->fps_pool_size; i++) {
1538 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1541 CERROR("Failed to allocate a new fast_reg descriptor\n");
1547 #ifndef HAVE_IB_MAP_MR_SG
1548 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1550 if (IS_ERR(frd->frd_frpl)) {
1551 rc = PTR_ERR(frd->frd_frpl);
1552 CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1554 frd->frd_frpl = NULL;
1559 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1560 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1564 * it is expected to get here if this is an MLX-5 card.
1565 * MLX-4 cards will always use FMR and MLX-5 cards will
1566 * always use fast_reg. It turns out that some MLX-5 cards
1567 * (possibly due to older FW versions) do not natively support
1568 * gaps. So we will need to track them here.
1570 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1571 #ifdef IB_MR_TYPE_SG_GAPS
1572 ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1573 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1574 IB_MR_TYPE_SG_GAPS :
1580 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1581 (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1582 CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1584 if (IS_ERR(frd->frd_mr)) {
1585 rc = PTR_ERR(frd->frd_mr);
1586 CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1591 /* There appears to be a bug in MLX5 code where you must
1592 * invalidate the rkey of a new FastReg pool before first
1593 * using it. Thus, I am marking the FRD invalid here. */
1594 frd->frd_valid = false;
1596 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1597 fpo->fast_reg.fpo_pool_size++;
1604 ib_dereg_mr(frd->frd_mr);
1605 #ifndef HAVE_IB_MAP_MR_SG
1607 ib_free_fast_reg_page_list(frd->frd_frpl);
1609 LIBCFS_FREE(frd, sizeof(*frd));
1612 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1614 list_del(&frd->frd_list);
1615 #ifndef HAVE_IB_MAP_MR_SG
1616 ib_free_fast_reg_page_list(frd->frd_frpl);
1618 ib_dereg_mr(frd->frd_mr);
1619 LIBCFS_FREE(frd, sizeof(*frd));
1625 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1626 struct kib_fmr_pool **pp_fpo)
1628 struct kib_dev *dev = fps->fps_net->ibn_dev;
1629 struct kib_fmr_pool *fpo;
1632 LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1636 memset(fpo, 0, sizeof(*fpo));
1638 fpo->fpo_hdev = kiblnd_current_hdev(dev);
1640 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1641 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1643 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1647 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1648 fpo->fpo_owner = fps;
1654 kiblnd_hdev_decref(fpo->fpo_hdev);
1655 LIBCFS_FREE(fpo, sizeof(*fpo));
1660 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1662 if (fps->fps_net == NULL) /* intialized? */
1665 spin_lock(&fps->fps_lock);
1667 while (!list_empty(&fps->fps_pool_list)) {
1668 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1669 struct kib_fmr_pool,
1672 fpo->fpo_failed = 1;
1673 if (fpo->fpo_map_count == 0)
1674 list_move(&fpo->fpo_list, zombies);
1676 list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1679 spin_unlock(&fps->fps_lock);
1683 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1685 if (fps->fps_net != NULL) { /* initialized? */
1686 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1687 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1692 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1693 struct kib_net *net,
1694 struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1696 struct kib_fmr_pool *fpo;
1699 memset(fps, 0, sizeof(struct kib_fmr_poolset));
1704 fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1705 fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1706 fps->fps_cache = tunables->lnd_fmr_cache;
1708 spin_lock_init(&fps->fps_lock);
1709 INIT_LIST_HEAD(&fps->fps_pool_list);
1710 INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1712 rc = kiblnd_create_fmr_pool(fps, &fpo);
1714 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1720 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1722 if (fpo->fpo_map_count != 0) /* still in use */
1724 if (fpo->fpo_failed)
1726 return now >= fpo->fpo_deadline;
1730 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1732 struct kib_hca_dev *hdev;
1733 __u64 *pages = tx->tx_pages;
1738 hdev = tx->tx_pool->tpo_hdev;
1740 for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1741 for (size = 0; size < rd->rd_frags[i].rf_nob;
1742 size += hdev->ibh_page_size) {
1743 pages[npages++] = (rd->rd_frags[i].rf_addr &
1744 hdev->ibh_page_mask) + size;
1752 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1755 struct kib_fmr_pool *fpo = fmr->fmr_pool;
1756 struct kib_fmr_poolset *fps;
1757 time64_t now = ktime_get_seconds();
1758 struct kib_fmr_pool *tmp;
1764 fps = fpo->fpo_owner;
1765 if (fpo->fpo_is_fmr) {
1766 if (fmr->fmr_pfmr) {
1767 ib_fmr_pool_unmap(fmr->fmr_pfmr);
1768 fmr->fmr_pfmr = NULL;
1772 rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1776 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1779 frd->frd_valid = false;
1780 spin_lock(&fps->fps_lock);
1781 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1782 spin_unlock(&fps->fps_lock);
1783 fmr->fmr_frd = NULL;
1786 fmr->fmr_pool = NULL;
1788 spin_lock(&fps->fps_lock);
1789 fpo->fpo_map_count--; /* decref the pool */
1791 list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1792 /* the first pool is persistent */
1793 if (fps->fps_pool_list.next == &fpo->fpo_list)
1796 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1797 list_move(&fpo->fpo_list, &zombies);
1801 spin_unlock(&fps->fps_lock);
1803 if (!list_empty(&zombies))
1804 kiblnd_destroy_fmr_pool_list(&zombies);
1807 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1808 struct kib_rdma_desc *rd, u32 nob, u64 iov,
1809 struct kib_fmr *fmr)
1811 struct kib_fmr_pool *fpo;
1812 __u64 *pages = tx->tx_pages;
1814 bool is_rx = (rd != tx->tx_rd);
1815 bool tx_pages_mapped = 0;
1820 spin_lock(&fps->fps_lock);
1821 version = fps->fps_version;
1822 list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1823 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1824 fpo->fpo_map_count++;
1826 if (fpo->fpo_is_fmr) {
1827 struct ib_pool_fmr *pfmr;
1829 spin_unlock(&fps->fps_lock);
1831 if (!tx_pages_mapped) {
1832 npages = kiblnd_map_tx_pages(tx, rd);
1833 tx_pages_mapped = 1;
1836 pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1837 pages, npages, iov);
1838 if (likely(!IS_ERR(pfmr))) {
1839 fmr->fmr_key = is_rx ? pfmr->fmr->rkey
1841 fmr->fmr_frd = NULL;
1842 fmr->fmr_pfmr = pfmr;
1843 fmr->fmr_pool = fpo;
1848 if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1849 struct kib_fast_reg_descriptor *frd;
1850 #ifdef HAVE_IB_MAP_MR_SG
1851 struct ib_reg_wr *wr;
1854 struct ib_rdma_wr *wr;
1855 struct ib_fast_reg_page_list *frpl;
1859 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1860 struct kib_fast_reg_descriptor,
1862 list_del(&frd->frd_list);
1863 spin_unlock(&fps->fps_lock);
1865 #ifndef HAVE_IB_MAP_MR_SG
1866 frpl = frd->frd_frpl;
1870 if (!frd->frd_valid) {
1871 struct ib_rdma_wr *inv_wr;
1872 __u32 key = is_rx ? mr->rkey : mr->lkey;
1874 inv_wr = &frd->frd_inv_wr;
1875 memset(inv_wr, 0, sizeof(*inv_wr));
1877 inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1878 inv_wr->wr.wr_id = IBLND_WID_MR;
1879 inv_wr->wr.ex.invalidate_rkey = key;
1882 key = ib_inc_rkey(key);
1883 ib_update_fast_reg_key(mr, key);
1886 #ifdef HAVE_IB_MAP_MR_SG
1887 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1888 n = ib_map_mr_sg(mr, tx->tx_frags,
1889 tx->tx_nfrags, NULL, PAGE_SIZE);
1891 n = ib_map_mr_sg(mr, tx->tx_frags,
1892 tx->tx_nfrags, PAGE_SIZE);
1894 if (unlikely(n != tx->tx_nfrags)) {
1895 CERROR("Failed to map mr %d/%d "
1896 "elements\n", n, tx->tx_nfrags);
1897 return n < 0 ? n : -EINVAL;
1900 wr = &frd->frd_fastreg_wr;
1901 memset(wr, 0, sizeof(*wr));
1903 wr->wr.opcode = IB_WR_REG_MR;
1904 wr->wr.wr_id = IBLND_WID_MR;
1906 wr->wr.send_flags = 0;
1908 wr->key = is_rx ? mr->rkey : mr->lkey;
1909 wr->access = (IB_ACCESS_LOCAL_WRITE |
1910 IB_ACCESS_REMOTE_WRITE);
1912 if (!tx_pages_mapped) {
1913 npages = kiblnd_map_tx_pages(tx, rd);
1914 tx_pages_mapped = 1;
1917 LASSERT(npages <= frpl->max_page_list_len);
1918 memcpy(frpl->page_list, pages,
1919 sizeof(*pages) * npages);
1921 /* Prepare FastReg WR */
1922 wr = &frd->frd_fastreg_wr;
1923 memset(wr, 0, sizeof(*wr));
1925 wr->wr.opcode = IB_WR_FAST_REG_MR;
1926 wr->wr.wr_id = IBLND_WID_MR;
1928 wr->wr.wr.fast_reg.iova_start = iov;
1929 wr->wr.wr.fast_reg.page_list = frpl;
1930 wr->wr.wr.fast_reg.page_list_len = npages;
1931 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1932 wr->wr.wr.fast_reg.length = nob;
1933 wr->wr.wr.fast_reg.rkey =
1934 is_rx ? mr->rkey : mr->lkey;
1935 wr->wr.wr.fast_reg.access_flags =
1936 (IB_ACCESS_LOCAL_WRITE |
1937 IB_ACCESS_REMOTE_WRITE);
1940 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1942 fmr->fmr_pfmr = NULL;
1943 fmr->fmr_pool = fpo;
1946 spin_unlock(&fps->fps_lock);
1950 spin_lock(&fps->fps_lock);
1951 fpo->fpo_map_count--;
1952 if (rc != -EAGAIN) {
1953 spin_unlock(&fps->fps_lock);
1957 /* EAGAIN and ... */
1958 if (version != fps->fps_version) {
1959 spin_unlock(&fps->fps_lock);
1964 if (fps->fps_increasing) {
1965 spin_unlock(&fps->fps_lock);
1966 CDEBUG(D_NET, "Another thread is allocating new "
1967 "FMR pool, waiting for her to complete\n");
1973 if (ktime_get_seconds() < fps->fps_next_retry) {
1974 /* someone failed recently */
1975 spin_unlock(&fps->fps_lock);
1979 fps->fps_increasing = 1;
1980 spin_unlock(&fps->fps_lock);
1982 CDEBUG(D_NET, "Allocate new FMR pool\n");
1983 rc = kiblnd_create_fmr_pool(fps, &fpo);
1984 spin_lock(&fps->fps_lock);
1985 fps->fps_increasing = 0;
1988 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1990 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
1992 spin_unlock(&fps->fps_lock);
1998 kiblnd_fini_pool(struct kib_pool *pool)
2000 LASSERT(list_empty(&pool->po_free_list));
2001 LASSERT(pool->po_allocated == 0);
2003 CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2007 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2009 CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2011 memset(pool, 0, sizeof(struct kib_pool));
2012 INIT_LIST_HEAD(&pool->po_free_list);
2013 pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2014 pool->po_owner = ps;
2015 pool->po_size = size;
2019 kiblnd_destroy_pool_list(struct list_head *head)
2021 struct kib_pool *pool;
2023 while (!list_empty(head)) {
2024 pool = list_entry(head->next, struct kib_pool, po_list);
2025 list_del(&pool->po_list);
2027 LASSERT(pool->po_owner != NULL);
2028 pool->po_owner->ps_pool_destroy(pool);
2033 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2035 if (ps->ps_net == NULL) /* intialized? */
2038 spin_lock(&ps->ps_lock);
2039 while (!list_empty(&ps->ps_pool_list)) {
2040 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
2041 struct kib_pool, po_list);
2044 if (po->po_allocated == 0)
2045 list_move(&po->po_list, zombies);
2047 list_move(&po->po_list, &ps->ps_failed_pool_list);
2049 spin_unlock(&ps->ps_lock);
2053 kiblnd_fini_poolset(struct kib_poolset *ps)
2055 if (ps->ps_net != NULL) { /* initialized? */
2056 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2057 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2062 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2063 struct kib_net *net, char *name, int size,
2064 kib_ps_pool_create_t po_create,
2065 kib_ps_pool_destroy_t po_destroy,
2066 kib_ps_node_init_t nd_init,
2067 kib_ps_node_fini_t nd_fini)
2069 struct kib_pool *pool;
2072 memset(ps, 0, sizeof(struct kib_poolset));
2076 ps->ps_pool_create = po_create;
2077 ps->ps_pool_destroy = po_destroy;
2078 ps->ps_node_init = nd_init;
2079 ps->ps_node_fini = nd_fini;
2080 ps->ps_pool_size = size;
2081 if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2082 >= sizeof(ps->ps_name))
2084 spin_lock_init(&ps->ps_lock);
2085 INIT_LIST_HEAD(&ps->ps_pool_list);
2086 INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2088 rc = ps->ps_pool_create(ps, size, &pool);
2090 list_add(&pool->po_list, &ps->ps_pool_list);
2092 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2098 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2100 if (pool->po_allocated != 0) /* still in use */
2102 if (pool->po_failed)
2104 return now >= pool->po_deadline;
2108 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2111 struct kib_poolset *ps = pool->po_owner;
2112 struct kib_pool *tmp;
2113 time64_t now = ktime_get_seconds();
2115 spin_lock(&ps->ps_lock);
2117 if (ps->ps_node_fini != NULL)
2118 ps->ps_node_fini(pool, node);
2120 LASSERT(pool->po_allocated > 0);
2121 list_add(node, &pool->po_free_list);
2122 pool->po_allocated--;
2124 list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2125 /* the first pool is persistent */
2126 if (ps->ps_pool_list.next == &pool->po_list)
2129 if (kiblnd_pool_is_idle(pool, now))
2130 list_move(&pool->po_list, &zombies);
2132 spin_unlock(&ps->ps_lock);
2134 if (!list_empty(&zombies))
2135 kiblnd_destroy_pool_list(&zombies);
2139 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2141 struct list_head *node;
2142 struct kib_pool *pool;
2144 unsigned int interval = 1;
2145 ktime_t time_before;
2146 unsigned int trips = 0;
2149 spin_lock(&ps->ps_lock);
2150 list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2151 if (list_empty(&pool->po_free_list))
2154 pool->po_allocated++;
2155 pool->po_deadline = ktime_get_seconds() +
2156 IBLND_POOL_DEADLINE;
2157 node = pool->po_free_list.next;
2160 if (ps->ps_node_init != NULL) {
2161 /* still hold the lock */
2162 ps->ps_node_init(pool, node);
2164 spin_unlock(&ps->ps_lock);
2168 /* no available tx pool and ... */
2169 if (ps->ps_increasing) {
2170 /* another thread is allocating a new pool */
2171 spin_unlock(&ps->ps_lock);
2174 "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2175 ps->ps_name, interval, trips);
2177 schedule_timeout_interruptible(interval);
2178 if (interval < cfs_time_seconds(1))
2184 if (ktime_get_seconds() < ps->ps_next_retry) {
2185 /* someone failed recently */
2186 spin_unlock(&ps->ps_lock);
2190 ps->ps_increasing = 1;
2191 spin_unlock(&ps->ps_lock);
2193 CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2194 time_before = ktime_get();
2195 rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2196 CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2197 ktime_ms_delta(ktime_get(), time_before));
2199 spin_lock(&ps->ps_lock);
2200 ps->ps_increasing = 0;
2202 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2204 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2205 CERROR("Can't allocate new %s pool because out of memory\n",
2208 spin_unlock(&ps->ps_lock);
2214 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2216 struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2220 LASSERT (pool->po_allocated == 0);
2222 if (tpo->tpo_tx_pages != NULL) {
2223 kiblnd_unmap_tx_pool(tpo);
2224 kiblnd_free_pages(tpo->tpo_tx_pages);
2227 if (tpo->tpo_tx_descs == NULL)
2230 for (i = 0; i < pool->po_size; i++) {
2231 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2232 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2234 list_del(&tx->tx_list);
2235 if (tx->tx_pages != NULL)
2236 CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2237 if (tx->tx_frags != NULL)
2238 CFS_FREE_PTR_ARRAY(tx->tx_frags,
2239 (1 + IBLND_MAX_RDMA_FRAGS));
2240 if (tx->tx_wrq != NULL)
2241 CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2242 (1 + IBLND_MAX_RDMA_FRAGS));
2243 if (tx->tx_sge != NULL)
2244 CFS_FREE_PTR_ARRAY(tx->tx_sge,
2245 (1 + IBLND_MAX_RDMA_FRAGS) *
2247 if (tx->tx_rd != NULL)
2248 LIBCFS_FREE(tx->tx_rd,
2249 offsetof(struct kib_rdma_desc,
2250 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2253 CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2255 kiblnd_fini_pool(pool);
2259 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2261 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2264 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2265 ntx = tunables->lnd_ntx / ncpts;
2267 return max(IBLND_TX_POOL, ntx);
2271 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2275 struct kib_pool *pool;
2276 struct kib_tx_pool *tpo;
2278 LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2280 CERROR("Failed to allocate TX pool\n");
2284 pool = &tpo->tpo_pool;
2285 kiblnd_init_pool(ps, pool, size);
2286 tpo->tpo_tx_descs = NULL;
2287 tpo->tpo_tx_pages = NULL;
2289 npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2290 if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2291 CERROR("Can't allocate tx pages: %d\n", npg);
2296 LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2297 size * sizeof(struct kib_tx));
2298 if (tpo->tpo_tx_descs == NULL) {
2299 CERROR("Can't allocate %d tx descriptors\n", size);
2300 ps->ps_pool_destroy(pool);
2304 memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2306 for (i = 0; i < size; i++) {
2307 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2308 int wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2311 if (ps->ps_net->ibn_fmr_ps != NULL) {
2312 LIBCFS_CPT_ALLOC(tx->tx_pages,
2313 lnet_cpt_table(), ps->ps_cpt,
2314 LNET_MAX_IOV * sizeof(*tx->tx_pages));
2315 if (tx->tx_pages == NULL)
2319 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2320 (1 + IBLND_MAX_RDMA_FRAGS) *
2321 sizeof(*tx->tx_frags));
2322 if (tx->tx_frags == NULL)
2325 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2327 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2328 (1 + IBLND_MAX_RDMA_FRAGS) *
2329 sizeof(*tx->tx_wrq));
2330 if (tx->tx_wrq == NULL)
2333 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2334 (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2335 sizeof(*tx->tx_sge));
2336 if (tx->tx_sge == NULL)
2339 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2340 offsetof(struct kib_rdma_desc,
2341 rd_frags[IBLND_MAX_RDMA_FRAGS]));
2342 if (tx->tx_rd == NULL)
2347 kiblnd_map_tx_pool(tpo);
2352 ps->ps_pool_destroy(pool);
2357 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2359 struct kib_tx_poolset *tps = container_of(pool->po_owner,
2360 struct kib_tx_poolset,
2362 struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2364 tx->tx_cookie = tps->tps_next_tx_cookie++;
2368 kiblnd_net_fini_pools(struct kib_net *net)
2372 cfs_cpt_for_each(i, lnet_cpt_table()) {
2373 struct kib_tx_poolset *tps;
2374 struct kib_fmr_poolset *fps;
2376 if (net->ibn_tx_ps != NULL) {
2377 tps = net->ibn_tx_ps[i];
2378 kiblnd_fini_poolset(&tps->tps_poolset);
2381 if (net->ibn_fmr_ps != NULL) {
2382 fps = net->ibn_fmr_ps[i];
2383 kiblnd_fini_fmr_poolset(fps);
2387 if (net->ibn_tx_ps != NULL) {
2388 cfs_percpt_free(net->ibn_tx_ps);
2389 net->ibn_tx_ps = NULL;
2392 if (net->ibn_fmr_ps != NULL) {
2393 cfs_percpt_free(net->ibn_fmr_ps);
2394 net->ibn_fmr_ps = NULL;
2399 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2402 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2403 #ifdef HAVE_IB_GET_DMA_MR
2404 unsigned long flags;
2410 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2412 #ifdef HAVE_IB_GET_DMA_MR
2413 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2415 * if lnd_map_on_demand is zero then we have effectively disabled
2416 * FMR or FastReg and we're using global memory regions
2419 if (!tunables->lnd_map_on_demand) {
2420 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2422 goto create_tx_pool;
2425 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2428 if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2429 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2430 tunables->lnd_fmr_pool_size,
2431 tunables->lnd_ntx / 4);
2436 /* TX pool must be created later than FMR, see LU-2268
2438 LASSERT(net->ibn_tx_ps == NULL);
2440 /* premapping can fail if ibd_nmr > 1, so we always create
2441 * FMR pool and map-on-demand if premapping failed */
2443 net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2444 sizeof(struct kib_fmr_poolset));
2445 if (net->ibn_fmr_ps == NULL) {
2446 CERROR("Failed to allocate FMR pool array\n");
2451 for (i = 0; i < ncpts; i++) {
2452 cpt = (cpts == NULL) ? i : cpts[i];
2453 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2456 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2463 LASSERT(i == ncpts);
2465 #ifdef HAVE_IB_GET_DMA_MR
2468 net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2469 sizeof(struct kib_tx_poolset));
2470 if (net->ibn_tx_ps == NULL) {
2471 CERROR("Failed to allocate tx pool array\n");
2476 for (i = 0; i < ncpts; i++) {
2477 cpt = (cpts == NULL) ? i : cpts[i];
2478 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2480 kiblnd_tx_pool_size(ni, ncpts),
2481 kiblnd_create_tx_pool,
2482 kiblnd_destroy_tx_pool,
2483 kiblnd_tx_init, NULL);
2485 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2493 kiblnd_net_fini_pools(net);
2499 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2501 struct ib_port_attr *port_attr;
2503 unsigned long flags;
2504 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2506 LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2507 if (port_attr == NULL) {
2508 CDEBUG(D_NETERROR, "Out of memory\n");
2512 rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2514 write_lock_irqsave(g_lock, flags);
2517 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2518 ? IBLND_DEV_PORT_ACTIVE
2519 : IBLND_DEV_PORT_DOWN;
2521 write_unlock_irqrestore(g_lock, flags);
2522 LIBCFS_FREE(port_attr, sizeof(*port_attr));
2525 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2532 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2534 struct kib_net *net;
2536 /* for health check */
2537 list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2539 CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2540 libcfs_nid2str(net->ibn_ni->ni_nid));
2541 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2546 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2548 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2549 struct kib_hca_dev *hdev;
2550 unsigned long flags;
2552 hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2554 write_lock_irqsave(g_lock, flags);
2556 switch (event->event) {
2557 case IB_EVENT_DEVICE_FATAL:
2558 CDEBUG(D_NET, "IB device fatal\n");
2559 hdev->ibh_state = IBLND_DEV_FATAL;
2560 kiblnd_set_ni_fatal_on(hdev, 1);
2562 case IB_EVENT_PORT_ACTIVE:
2563 CDEBUG(D_NET, "IB port active\n");
2564 if (event->element.port_num == hdev->ibh_port) {
2565 hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2566 kiblnd_set_ni_fatal_on(hdev, 0);
2569 case IB_EVENT_PORT_ERR:
2570 CDEBUG(D_NET, "IB port err\n");
2571 if (event->element.port_num == hdev->ibh_port) {
2572 hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2573 kiblnd_set_ni_fatal_on(hdev, 1);
2579 write_unlock_irqrestore(g_lock, flags);
2583 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2585 struct ib_device_attr *dev_attr;
2589 /* It's safe to assume a HCA can handle a page size
2590 * matching that of the native system */
2591 hdev->ibh_page_shift = PAGE_SHIFT;
2592 hdev->ibh_page_size = 1 << PAGE_SHIFT;
2593 hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1);
2595 #ifndef HAVE_IB_DEVICE_ATTRS
2596 LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2597 if (dev_attr == NULL) {
2598 CERROR("Out of memory\n");
2602 rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2604 CERROR("Failed to query IB device: %d\n", rc);
2605 goto out_clean_attr;
2608 dev_attr = &hdev->ibh_ibdev->attrs;
2611 hdev->ibh_mr_size = dev_attr->max_mr_size;
2612 hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2614 /* Setup device Memory Registration capabilities */
2615 #ifdef HAVE_IB_DEVICE_OPS
2616 if (hdev->ibh_ibdev->ops.alloc_fmr &&
2617 hdev->ibh_ibdev->ops.dealloc_fmr &&
2618 hdev->ibh_ibdev->ops.map_phys_fmr &&
2619 hdev->ibh_ibdev->ops.unmap_fmr) {
2621 if (hdev->ibh_ibdev->alloc_fmr &&
2622 hdev->ibh_ibdev->dealloc_fmr &&
2623 hdev->ibh_ibdev->map_phys_fmr &&
2624 hdev->ibh_ibdev->unmap_fmr) {
2626 LCONSOLE_INFO("Using FMR for registration\n");
2627 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2628 } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2629 LCONSOLE_INFO("Using FastReg for registration\n");
2630 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2631 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2632 #ifdef IB_DEVICE_SG_GAPS_REG
2633 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2634 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2641 rc2 = kiblnd_port_get_attr(hdev);
2648 #ifndef HAVE_IB_DEVICE_ATTRS
2650 LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2654 CERROR("IB device does not support FMRs nor FastRegs, can't "
2655 "register memory: %d\n", rc);
2656 else if (rc == -EINVAL)
2657 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2661 #ifdef HAVE_IB_GET_DMA_MR
2663 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2665 if (hdev->ibh_mrs == NULL)
2668 ib_dereg_mr(hdev->ibh_mrs);
2670 hdev->ibh_mrs = NULL;
2675 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2677 if (hdev->ibh_event_handler.device != NULL)
2678 ib_unregister_event_handler(&hdev->ibh_event_handler);
2680 #ifdef HAVE_IB_GET_DMA_MR
2681 kiblnd_hdev_cleanup_mrs(hdev);
2684 if (hdev->ibh_pd != NULL)
2685 ib_dealloc_pd(hdev->ibh_pd);
2687 if (hdev->ibh_cmid != NULL)
2688 rdma_destroy_id(hdev->ibh_cmid);
2690 LIBCFS_FREE(hdev, sizeof(*hdev));
2693 #ifdef HAVE_IB_GET_DMA_MR
2695 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2698 int acflags = IB_ACCESS_LOCAL_WRITE |
2699 IB_ACCESS_REMOTE_WRITE;
2701 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2703 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2704 kiblnd_hdev_cleanup_mrs(hdev);
2715 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2721 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2723 struct rdma_cm_id *cmid;
2724 struct sockaddr_in srcaddr;
2725 struct sockaddr_in dstaddr;
2728 if (dev->ibd_hdev == NULL || /* initializing */
2729 dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2730 *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2733 /* XXX: it's UGLY, but I don't have better way to find
2734 * ib-bonding HCA failover because:
2736 * a. no reliable CM event for HCA failover...
2737 * b. no OFED API to get ib_device for current net_device...
2739 * We have only two choices at this point:
2741 * a. rdma_bind_addr(), it will conflict with listener cmid
2742 * b. rdma_resolve_addr() to zero addr */
2743 cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2744 RDMA_PS_TCP, IB_QPT_RC);
2747 CERROR("Failed to create cmid for failover: %d\n", rc);
2751 memset(&srcaddr, 0, sizeof(srcaddr));
2752 srcaddr.sin_family = AF_INET;
2753 srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2755 memset(&dstaddr, 0, sizeof(dstaddr));
2756 dstaddr.sin_family = AF_INET;
2757 rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2758 (struct sockaddr *)&dstaddr, 1);
2759 if (rc != 0 || cmid->device == NULL) {
2760 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2761 dev->ibd_ifname, &dev->ibd_ifip,
2763 rdma_destroy_id(cmid);
2767 rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2768 rdma_destroy_id(cmid);
2773 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2775 LIST_HEAD(zombie_tpo);
2776 LIST_HEAD(zombie_ppo);
2777 LIST_HEAD(zombie_fpo);
2778 struct rdma_cm_id *cmid = NULL;
2779 struct kib_hca_dev *hdev = NULL;
2780 struct kib_hca_dev *old;
2782 struct kib_net *net;
2783 struct sockaddr_in addr;
2784 unsigned long flags;
2788 LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2789 dev->ibd_can_failover ||
2790 dev->ibd_hdev == NULL);
2792 rc = kiblnd_dev_need_failover(dev, ns);
2796 if (dev->ibd_hdev != NULL &&
2797 dev->ibd_hdev->ibh_cmid != NULL) {
2798 /* XXX it's not good to close old listener at here,
2799 * because we can fail to create new listener.
2800 * But we have to close it now, otherwise rdma_bind_addr
2801 * will return EADDRINUSE... How crap! */
2802 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2804 cmid = dev->ibd_hdev->ibh_cmid;
2805 /* make next schedule of kiblnd_dev_need_failover()
2806 * return 1 for me */
2807 dev->ibd_hdev->ibh_cmid = NULL;
2808 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2810 rdma_destroy_id(cmid);
2813 cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2817 CERROR("Failed to create cmid for failover: %d\n", rc);
2821 memset(&addr, 0, sizeof(addr));
2822 addr.sin_family = AF_INET;
2823 addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2824 addr.sin_port = htons(*kiblnd_tunables.kib_service);
2826 /* Bind to failover device or port */
2827 rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2828 if (rc != 0 || cmid->device == NULL) {
2829 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2830 dev->ibd_ifname, &dev->ibd_ifip,
2832 rdma_destroy_id(cmid);
2836 LIBCFS_ALLOC(hdev, sizeof(*hdev));
2838 CERROR("Failed to allocate kib_hca_dev\n");
2839 rdma_destroy_id(cmid);
2844 atomic_set(&hdev->ibh_ref, 1);
2845 hdev->ibh_dev = dev;
2846 hdev->ibh_cmid = cmid;
2847 hdev->ibh_ibdev = cmid->device;
2848 hdev->ibh_port = cmid->port_num;
2850 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2851 pd = ib_alloc_pd(cmid->device, 0);
2853 pd = ib_alloc_pd(cmid->device);
2857 CERROR("Can't allocate PD: %d\n", rc);
2863 rc = rdma_listen(cmid, 0);
2865 CERROR("Can't start new listener: %d\n", rc);
2869 rc = kiblnd_hdev_get_attr(hdev);
2871 CERROR("Can't get device attributes: %d\n", rc);
2875 #ifdef HAVE_IB_GET_DMA_MR
2876 rc = kiblnd_hdev_setup_mrs(hdev);
2878 CERROR("Can't setup device: %d\n", rc);
2883 INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
2884 hdev->ibh_ibdev, kiblnd_event_handler);
2885 ib_register_event_handler(&hdev->ibh_event_handler);
2887 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2889 old = dev->ibd_hdev;
2890 dev->ibd_hdev = hdev; /* take over the refcount */
2893 list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2894 cfs_cpt_for_each(i, lnet_cpt_table()) {
2895 kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2898 if (net->ibn_fmr_ps != NULL)
2899 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2904 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2906 if (!list_empty(&zombie_tpo))
2907 kiblnd_destroy_pool_list(&zombie_tpo);
2908 if (!list_empty(&zombie_ppo))
2909 kiblnd_destroy_pool_list(&zombie_ppo);
2910 if (!list_empty(&zombie_fpo))
2911 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2913 kiblnd_hdev_decref(hdev);
2916 dev->ibd_failed_failover++;
2918 dev->ibd_failed_failover = 0;
2924 kiblnd_destroy_dev(struct kib_dev *dev)
2926 LASSERT(dev->ibd_nnets == 0);
2927 LASSERT(list_empty(&dev->ibd_nets));
2929 list_del(&dev->ibd_fail_list);
2930 list_del(&dev->ibd_list);
2932 if (dev->ibd_hdev != NULL)
2933 kiblnd_hdev_decref(dev->ibd_hdev);
2935 LIBCFS_FREE(dev, sizeof(*dev));
2939 kiblnd_base_shutdown(void)
2941 struct kib_sched_info *sched;
2944 LASSERT(list_empty(&kiblnd_data.kib_devs));
2946 CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2947 atomic_read(&libcfs_kmemory));
2949 switch (kiblnd_data.kib_init) {
2953 case IBLND_INIT_ALL:
2954 case IBLND_INIT_DATA:
2955 LASSERT (kiblnd_data.kib_peers != NULL);
2956 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2957 LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2959 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2960 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2961 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2962 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2964 /* flag threads to terminate; wake and wait for them to die */
2965 kiblnd_data.kib_shutdown = 1;
2967 /* NB: we really want to stop scheduler threads net by net
2968 * instead of the whole module, this should be improved
2969 * with dynamic configuration LNet */
2970 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2971 wake_up_all(&sched->ibs_waitq);
2973 wake_up_all(&kiblnd_data.kib_connd_waitq);
2974 wake_up_all(&kiblnd_data.kib_failover_waitq);
2976 wait_var_event_warning(&kiblnd_data.kib_nthreads,
2977 !atomic_read(&kiblnd_data.kib_nthreads),
2978 "Waiting for %d threads to terminate\n",
2979 atomic_read(&kiblnd_data.kib_nthreads));
2982 case IBLND_INIT_NOTHING:
2986 if (kiblnd_data.kib_peers)
2987 CFS_FREE_PTR_ARRAY(kiblnd_data.kib_peers,
2988 kiblnd_data.kib_peer_hash_size);
2990 if (kiblnd_data.kib_scheds != NULL)
2991 cfs_percpt_free(kiblnd_data.kib_scheds);
2993 CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2994 atomic_read(&libcfs_kmemory));
2996 kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2997 module_put(THIS_MODULE);
3001 kiblnd_shutdown(struct lnet_ni *ni)
3003 struct kib_net *net = ni->ni_data;
3004 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
3005 unsigned long flags;
3007 LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3012 CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
3013 atomic_read(&libcfs_kmemory));
3015 write_lock_irqsave(g_lock, flags);
3016 net->ibn_shutdown = 1;
3017 write_unlock_irqrestore(g_lock, flags);
3019 switch (net->ibn_init) {
3023 case IBLND_INIT_ALL:
3024 /* nuke all existing peers within this net */
3025 kiblnd_del_peer(ni, LNET_NID_ANY);
3027 /* Wait for all peer_ni state to clean up */
3028 wait_var_event_warning(&net->ibn_npeers,
3029 atomic_read(&net->ibn_npeers) == 0,
3030 "%s: waiting for %d peers to disconnect\n",
3031 libcfs_nid2str(ni->ni_nid),
3032 atomic_read(&net->ibn_npeers));
3034 kiblnd_net_fini_pools(net);
3036 write_lock_irqsave(g_lock, flags);
3037 LASSERT(net->ibn_dev->ibd_nnets > 0);
3038 net->ibn_dev->ibd_nnets--;
3039 list_del(&net->ibn_list);
3040 write_unlock_irqrestore(g_lock, flags);
3044 case IBLND_INIT_NOTHING:
3045 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3047 if (net->ibn_dev != NULL &&
3048 net->ibn_dev->ibd_nnets == 0)
3049 kiblnd_destroy_dev(net->ibn_dev);
3054 CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3055 atomic_read(&libcfs_kmemory));
3057 net->ibn_init = IBLND_INIT_NOTHING;
3060 LIBCFS_FREE(net, sizeof(*net));
3063 if (list_empty(&kiblnd_data.kib_devs))
3064 kiblnd_base_shutdown();
3068 kiblnd_base_startup(struct net *ns)
3070 struct kib_sched_info *sched;
3074 LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3076 if (!try_module_get(THIS_MODULE))
3079 memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3081 rwlock_init(&kiblnd_data.kib_global_lock);
3083 INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3084 INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3086 kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3087 CFS_ALLOC_PTR_ARRAY(kiblnd_data.kib_peers,
3088 kiblnd_data.kib_peer_hash_size);
3089 if (kiblnd_data.kib_peers == NULL)
3092 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3093 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3095 spin_lock_init(&kiblnd_data.kib_connd_lock);
3096 INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3097 INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3098 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3099 INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3101 init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3102 init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3104 kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3106 if (kiblnd_data.kib_scheds == NULL)
3109 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3112 spin_lock_init(&sched->ibs_lock);
3113 INIT_LIST_HEAD(&sched->ibs_conns);
3114 init_waitqueue_head(&sched->ibs_waitq);
3116 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3117 if (*kiblnd_tunables.kib_nscheds > 0) {
3118 nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3120 /* max to half of CPUs, another half is reserved for
3121 * upper layer modules */
3122 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3125 sched->ibs_nthreads_max = nthrs;
3129 kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3131 /* lists/ptrs/locks initialised */
3132 kiblnd_data.kib_init = IBLND_INIT_DATA;
3133 /*****************************************************/
3135 rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3137 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3141 if (*kiblnd_tunables.kib_dev_failover != 0)
3142 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3146 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3150 /* flag everything initialised */
3151 kiblnd_data.kib_init = IBLND_INIT_ALL;
3152 /*****************************************************/
3157 kiblnd_base_shutdown();
3162 kiblnd_start_schedulers(struct kib_sched_info *sched)
3168 if (sched->ibs_nthreads == 0) {
3169 if (*kiblnd_tunables.kib_nscheds > 0) {
3170 nthrs = sched->ibs_nthreads_max;
3172 nthrs = cfs_cpt_weight(lnet_cpt_table(),
3174 nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3175 nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3178 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3179 /* increase one thread if there is new interface */
3180 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3183 for (i = 0; i < nthrs; i++) {
3186 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3187 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3188 KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3189 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3193 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3194 sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3198 sched->ibs_nthreads += i;
3202 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3209 for (i = 0; i < ncpts; i++) {
3210 struct kib_sched_info *sched;
3212 cpt = (cpts == NULL) ? i : cpts[i];
3213 sched = kiblnd_data.kib_scheds[cpt];
3215 if (!newdev && sched->ibs_nthreads > 0)
3218 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3220 CERROR("Failed to start scheduler threads for %s\n",
3228 static struct kib_dev *
3229 kiblnd_dev_search(char *ifname)
3231 struct kib_dev *alias = NULL;
3232 struct kib_dev *dev;
3236 colon = strchr(ifname, ':');
3237 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3238 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3244 colon2 = strchr(dev->ibd_ifname, ':');
3250 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3262 kiblnd_startup(struct lnet_ni *ni)
3264 char *ifname = NULL;
3265 struct lnet_inetdev *ifaces = NULL;
3266 struct kib_dev *ibdev = NULL;
3267 struct kib_net *net = NULL;
3268 unsigned long flags;
3273 LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3275 if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3276 rc = kiblnd_base_startup(ni->ni_net_ns);
3281 LIBCFS_ALLOC(net, sizeof(*net));
3289 net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3291 kiblnd_tunables_setup(ni);
3294 * ni_interfaces is only to support legacy pre Multi-Rail
3295 * tcp bonding for ksocklnd. Multi-Rail wants each secondary
3296 * IP to be treated as an unique 'struct ni' interfaces instead.
3298 if (ni->ni_interfaces[0] != NULL) {
3299 /* Use the IPoIB interface specified in 'networks=' */
3300 if (ni->ni_interfaces[1] != NULL) {
3301 CERROR("ko2iblnd: Multiple interfaces not supported\n");
3306 ifname = ni->ni_interfaces[0];
3308 ifname = *kiblnd_tunables.kib_default_ipif;
3311 if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3312 CERROR("IPoIB interface name too long: %s\n", ifname);
3317 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3321 for (i = 0; i < rc; i++) {
3322 if (strcmp(ifname, ifaces[i].li_name) == 0)
3327 CERROR("ko2iblnd: No matching interfaces\n");
3332 ibdev = kiblnd_dev_search(ifname);
3333 newdev = ibdev == NULL;
3334 /* hmm...create kib_dev even for alias */
3335 if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3336 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3342 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3343 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3344 sizeof(ibdev->ibd_ifname));
3345 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3347 INIT_LIST_HEAD(&ibdev->ibd_nets);
3348 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3349 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3351 /* initialize the device */
3352 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3354 CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3359 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3362 net->ibn_dev = ibdev;
3363 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3365 ni->ni_dev_cpt = ifaces[i].li_cpt;
3367 rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3371 rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3373 CERROR("Failed to initialize NI pools: %d\n", rc);
3377 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3379 list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3380 /* for health check */
3381 if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3382 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3383 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3385 net->ibn_init = IBLND_INIT_ALL;
3390 if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3391 kiblnd_destroy_dev(ibdev);
3394 kiblnd_shutdown(ni);
3396 CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3397 ifname ? ifname : "", rc);
3402 static const struct lnet_lnd the_o2iblnd = {
3403 .lnd_type = O2IBLND,
3404 .lnd_startup = kiblnd_startup,
3405 .lnd_shutdown = kiblnd_shutdown,
3406 .lnd_ctl = kiblnd_ctl,
3407 .lnd_send = kiblnd_send,
3408 .lnd_recv = kiblnd_recv,
3411 static void __exit ko2iblnd_exit(void)
3413 lnet_unregister_lnd(&the_o2iblnd);
3416 static int __init ko2iblnd_init(void)
3420 BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3421 BUILD_BUG_ON(offsetof(struct kib_msg,
3422 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3424 BUILD_BUG_ON(offsetof(struct kib_msg,
3425 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3428 rc = kiblnd_tunables_init();
3432 lnet_register_lnd(&the_o2iblnd);
3437 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3438 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3439 MODULE_VERSION("2.8.0");
3440 MODULE_LICENSE("GPL");
3442 module_init(ko2iblnd_init);
3443 module_exit(ko2iblnd_exit);