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) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 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/lnet/lib-move.c
34 * Data movement routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/pagemap.h>
41 #include <lnet/lib-lnet.h>
42 #include <linux/nsproxy.h>
43 #include <net/net_namespace.h>
45 static int local_nid_dist_zero = 1;
46 module_param(local_nid_dist_zero, int, 0444);
47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
49 static inline struct lnet_comm_count *
50 get_stats_counts(struct lnet_element_stats *stats,
51 enum lnet_stats_type stats_type)
54 case LNET_STATS_TYPE_SEND:
55 return &stats->el_send_stats;
56 case LNET_STATS_TYPE_RECV:
57 return &stats->el_recv_stats;
58 case LNET_STATS_TYPE_DROP:
59 return &stats->el_drop_stats;
61 CERROR("Unknown stats type\n");
67 void lnet_incr_stats(struct lnet_element_stats *stats,
68 enum lnet_msg_type msg_type,
69 enum lnet_stats_type stats_type)
71 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
77 atomic_inc(&counts->co_ack_count);
80 atomic_inc(&counts->co_put_count);
83 atomic_inc(&counts->co_get_count);
86 atomic_inc(&counts->co_reply_count);
89 atomic_inc(&counts->co_hello_count);
92 CERROR("There is a BUG in the code. Unknown message type\n");
97 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
98 enum lnet_stats_type stats_type)
100 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
104 return (atomic_read(&counts->co_ack_count) +
105 atomic_read(&counts->co_put_count) +
106 atomic_read(&counts->co_get_count) +
107 atomic_read(&counts->co_reply_count) +
108 atomic_read(&counts->co_hello_count));
111 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
112 struct lnet_comm_count *counts)
114 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
115 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
116 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
117 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
118 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
121 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
122 struct lnet_element_stats *stats)
124 struct lnet_comm_count *counts;
129 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
132 assign_stats(&msg_stats->im_send_stats, counts);
134 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
137 assign_stats(&msg_stats->im_recv_stats, counts);
139 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
142 assign_stats(&msg_stats->im_drop_stats, counts);
146 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
148 struct lnet_test_peer *tp;
149 struct list_head *el;
150 struct list_head *next;
151 struct list_head cull;
153 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
154 if (threshold != 0) {
155 /* Adding a new entry */
156 LIBCFS_ALLOC(tp, sizeof(*tp));
161 tp->tp_threshold = threshold;
164 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
169 /* removing entries */
170 INIT_LIST_HEAD(&cull);
174 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
175 tp = list_entry(el, struct lnet_test_peer, tp_list);
177 if (tp->tp_threshold == 0 || /* needs culling anyway */
178 nid == LNET_NID_ANY || /* removing all entries */
179 tp->tp_nid == nid) { /* matched this one */
180 list_del(&tp->tp_list);
181 list_add(&tp->tp_list, &cull);
187 while (!list_empty(&cull)) {
188 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
190 list_del(&tp->tp_list);
191 LIBCFS_FREE(tp, sizeof(*tp));
197 fail_peer (lnet_nid_t nid, int outgoing)
199 struct lnet_test_peer *tp;
200 struct list_head *el;
201 struct list_head *next;
202 struct list_head cull;
205 INIT_LIST_HEAD(&cull);
207 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
210 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
211 tp = list_entry(el, struct lnet_test_peer, tp_list);
213 if (tp->tp_threshold == 0) {
216 /* only cull zombies on outgoing tests,
217 * since we may be at interrupt priority on
218 * incoming messages. */
219 list_del(&tp->tp_list);
220 list_add(&tp->tp_list, &cull);
225 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
226 nid == tp->tp_nid) { /* fail this peer */
229 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
232 tp->tp_threshold == 0) {
234 list_del(&tp->tp_list);
235 list_add(&tp->tp_list, &cull);
244 while (!list_empty(&cull)) {
245 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
246 list_del(&tp->tp_list);
248 LIBCFS_FREE(tp, sizeof(*tp));
255 lnet_iov_nob(unsigned int niov, struct kvec *iov)
257 unsigned int nob = 0;
259 LASSERT(niov == 0 || iov != NULL);
261 nob += (iov++)->iov_len;
265 EXPORT_SYMBOL(lnet_iov_nob);
268 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
269 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
272 /* NB diov, siov are READ-ONLY */
273 unsigned int this_nob;
278 /* skip complete frags before 'doffset' */
280 while (doffset >= diov->iov_len) {
281 doffset -= diov->iov_len;
287 /* skip complete frags before 'soffset' */
289 while (soffset >= siov->iov_len) {
290 soffset -= siov->iov_len;
299 this_nob = MIN(diov->iov_len - doffset,
300 siov->iov_len - soffset);
301 this_nob = MIN(this_nob, nob);
303 memcpy((char *)diov->iov_base + doffset,
304 (char *)siov->iov_base + soffset, this_nob);
307 if (diov->iov_len > doffset + this_nob) {
315 if (siov->iov_len > soffset + this_nob) {
324 EXPORT_SYMBOL(lnet_copy_iov2iov);
327 lnet_extract_iov(int dst_niov, struct kvec *dst,
328 int src_niov, struct kvec *src,
329 unsigned int offset, unsigned int len)
331 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
332 * for exactly 'len' bytes, and return the number of entries.
333 * NB not destructive to 'src' */
334 unsigned int frag_len;
337 if (len == 0) /* no data => */
338 return (0); /* no frags */
340 LASSERT(src_niov > 0);
341 while (offset >= src->iov_len) { /* skip initial frags */
342 offset -= src->iov_len;
345 LASSERT(src_niov > 0);
350 LASSERT(src_niov > 0);
351 LASSERT((int)niov <= dst_niov);
353 frag_len = src->iov_len - offset;
354 dst->iov_base = ((char *)src->iov_base) + offset;
356 if (len <= frag_len) {
361 dst->iov_len = frag_len;
371 EXPORT_SYMBOL(lnet_extract_iov);
375 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
377 unsigned int nob = 0;
379 LASSERT(niov == 0 || kiov != NULL);
381 nob += (kiov++)->kiov_len;
385 EXPORT_SYMBOL(lnet_kiov_nob);
388 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
389 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
392 /* NB diov, siov are READ-ONLY */
393 unsigned int this_nob;
400 LASSERT (!in_interrupt ());
403 while (doffset >= diov->kiov_len) {
404 doffset -= diov->kiov_len;
411 while (soffset >= siov->kiov_len) {
412 soffset -= siov->kiov_len;
421 this_nob = MIN(diov->kiov_len - doffset,
422 siov->kiov_len - soffset);
423 this_nob = MIN(this_nob, nob);
426 daddr = ((char *)kmap(diov->kiov_page)) +
427 diov->kiov_offset + doffset;
429 saddr = ((char *)kmap(siov->kiov_page)) +
430 siov->kiov_offset + soffset;
432 /* Vanishing risk of kmap deadlock when mapping 2 pages.
433 * However in practice at least one of the kiovs will be mapped
434 * kernel pages and the map/unmap will be NOOPs */
436 memcpy (daddr, saddr, this_nob);
439 if (diov->kiov_len > doffset + this_nob) {
443 kunmap(diov->kiov_page);
450 if (siov->kiov_len > soffset + this_nob) {
454 kunmap(siov->kiov_page);
463 kunmap(diov->kiov_page);
465 kunmap(siov->kiov_page);
467 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
470 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
471 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
474 /* NB iov, kiov are READ-ONLY */
475 unsigned int this_nob;
481 LASSERT (!in_interrupt ());
484 while (iovoffset >= iov->iov_len) {
485 iovoffset -= iov->iov_len;
492 while (kiovoffset >= kiov->kiov_len) {
493 kiovoffset -= kiov->kiov_len;
502 this_nob = MIN(iov->iov_len - iovoffset,
503 kiov->kiov_len - kiovoffset);
504 this_nob = MIN(this_nob, nob);
507 addr = ((char *)kmap(kiov->kiov_page)) +
508 kiov->kiov_offset + kiovoffset;
510 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
513 if (iov->iov_len > iovoffset + this_nob) {
514 iovoffset += this_nob;
521 if (kiov->kiov_len > kiovoffset + this_nob) {
523 kiovoffset += this_nob;
525 kunmap(kiov->kiov_page);
535 kunmap(kiov->kiov_page);
537 EXPORT_SYMBOL(lnet_copy_kiov2iov);
540 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
541 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
544 /* NB kiov, iov are READ-ONLY */
545 unsigned int this_nob;
551 LASSERT (!in_interrupt ());
554 while (kiovoffset >= kiov->kiov_len) {
555 kiovoffset -= kiov->kiov_len;
562 while (iovoffset >= iov->iov_len) {
563 iovoffset -= iov->iov_len;
572 this_nob = MIN(kiov->kiov_len - kiovoffset,
573 iov->iov_len - iovoffset);
574 this_nob = MIN(this_nob, nob);
577 addr = ((char *)kmap(kiov->kiov_page)) +
578 kiov->kiov_offset + kiovoffset;
580 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
583 if (kiov->kiov_len > kiovoffset + this_nob) {
585 kiovoffset += this_nob;
587 kunmap(kiov->kiov_page);
594 if (iov->iov_len > iovoffset + this_nob) {
595 iovoffset += this_nob;
604 kunmap(kiov->kiov_page);
606 EXPORT_SYMBOL(lnet_copy_iov2kiov);
609 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
610 int src_niov, lnet_kiov_t *src,
611 unsigned int offset, unsigned int len)
613 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
614 * for exactly 'len' bytes, and return the number of entries.
615 * NB not destructive to 'src' */
616 unsigned int frag_len;
619 if (len == 0) /* no data => */
620 return (0); /* no frags */
622 LASSERT(src_niov > 0);
623 while (offset >= src->kiov_len) { /* skip initial frags */
624 offset -= src->kiov_len;
627 LASSERT(src_niov > 0);
632 LASSERT(src_niov > 0);
633 LASSERT((int)niov <= dst_niov);
635 frag_len = src->kiov_len - offset;
636 dst->kiov_page = src->kiov_page;
637 dst->kiov_offset = src->kiov_offset + offset;
639 if (len <= frag_len) {
641 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
645 dst->kiov_len = frag_len;
646 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
656 EXPORT_SYMBOL(lnet_extract_kiov);
659 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
660 int delayed, unsigned int offset, unsigned int mlen,
663 unsigned int niov = 0;
664 struct kvec *iov = NULL;
665 lnet_kiov_t *kiov = NULL;
668 LASSERT (!in_interrupt ());
669 LASSERT (mlen == 0 || msg != NULL);
672 LASSERT(msg->msg_receiving);
673 LASSERT(!msg->msg_sending);
674 LASSERT(rlen == msg->msg_len);
675 LASSERT(mlen <= msg->msg_len);
676 LASSERT(msg->msg_offset == offset);
677 LASSERT(msg->msg_wanted == mlen);
679 msg->msg_receiving = 0;
682 niov = msg->msg_niov;
684 kiov = msg->msg_kiov;
687 LASSERT ((iov == NULL) != (kiov == NULL));
691 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
692 niov, iov, kiov, offset, mlen,
695 lnet_finalize(msg, rc);
699 lnet_setpayloadbuffer(struct lnet_msg *msg)
701 struct lnet_libmd *md = msg->msg_md;
703 LASSERT(msg->msg_len > 0);
704 LASSERT(!msg->msg_routing);
706 LASSERT(msg->msg_niov == 0);
707 LASSERT(msg->msg_iov == NULL);
708 LASSERT(msg->msg_kiov == NULL);
710 msg->msg_niov = md->md_niov;
711 if ((md->md_options & LNET_MD_KIOV) != 0)
712 msg->msg_kiov = md->md_iov.kiov;
714 msg->msg_iov = md->md_iov.iov;
718 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
719 unsigned int offset, unsigned int len)
721 msg->msg_type = type;
722 msg->msg_target = target;
724 msg->msg_offset = offset;
727 lnet_setpayloadbuffer(msg);
729 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
730 msg->msg_hdr.type = cpu_to_le32(type);
731 /* dest_nid will be overwritten by lnet_select_pathway() */
732 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
733 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
734 /* src_nid will be set later */
735 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
736 msg->msg_hdr.payload_length = cpu_to_le32(len);
740 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
742 void *priv = msg->msg_private;
745 LASSERT (!in_interrupt ());
746 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
747 (msg->msg_txcredit && msg->msg_peertxcredit));
749 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
751 lnet_finalize(msg, rc);
755 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
759 LASSERT(!msg->msg_sending);
760 LASSERT(msg->msg_receiving);
761 LASSERT(!msg->msg_rx_ready_delay);
762 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
764 msg->msg_rx_ready_delay = 1;
765 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
768 CERROR("recv from %s / send to %s aborted: "
769 "eager_recv failed %d\n",
770 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
771 libcfs_id2str(msg->msg_target), rc);
772 LASSERT(rc < 0); /* required by my callers */
779 * This function can be called from two paths:
780 * 1. when sending a message
781 * 2. when decommiting a message (lnet_msg_decommit_tx())
782 * In both these cases the peer_ni should have it's reference count
783 * acquired by the caller and therefore it is safe to drop the spin
784 * lock before calling lnd_query()
787 lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
789 time64_t last_alive = 0;
790 int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
792 LASSERT(lnet_peer_aliveness_enabled(lp));
793 LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
795 lnet_net_unlock(cpt);
796 (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
799 lp->lpni_last_query = ktime_get_seconds();
801 if (last_alive != 0) /* NI has updated timestamp */
802 lp->lpni_last_alive = last_alive;
805 /* NB: always called with lnet_net_lock held */
807 lnet_peer_is_alive(struct lnet_peer_ni *lp, time64_t now)
812 LASSERT (lnet_peer_aliveness_enabled(lp));
815 * Trust lnet_notify() if it has more recent aliveness news, but
816 * ignore the initial assumed death (see lnet_peers_start_down()).
818 spin_lock(&lp->lpni_lock);
819 if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
820 lp->lpni_timestamp >= lp->lpni_last_alive) {
821 spin_unlock(&lp->lpni_lock);
825 deadline = lp->lpni_last_alive +
826 lp->lpni_net->net_tunables.lct_peer_timeout;
827 alive = deadline > now;
830 * Update obsolete lp_alive except for routers assumed to be dead
831 * initially, because router checker would update aliveness in this
832 * case, and moreover lpni_last_alive at peer creation is assumed.
834 if (alive && !lp->lpni_alive &&
835 !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
836 spin_unlock(&lp->lpni_lock);
837 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
839 spin_unlock(&lp->lpni_lock);
846 /* NB: returns 1 when alive, 0 when dead, negative when error;
847 * may drop the lnet_net_lock */
849 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
851 time64_t now = ktime_get_seconds();
853 if (!lnet_peer_aliveness_enabled(lp))
856 if (lnet_peer_is_alive(lp, now))
860 * Peer appears dead, but we should avoid frequent NI queries (at
861 * most once per lnet_queryinterval seconds).
863 if (lp->lpni_last_query != 0) {
864 static const int lnet_queryinterval = 1;
867 next_query = lp->lpni_last_query + lnet_queryinterval;
869 if (now < next_query) {
871 CWARN("Unexpected aliveness of peer %s: "
872 "%lld < %lld (%d/%d)\n",
873 libcfs_nid2str(lp->lpni_nid),
876 lp->lpni_net->net_tunables.lct_peer_timeout);
881 /* query NI for latest aliveness news */
882 lnet_ni_query_locked(ni, lp);
884 if (lnet_peer_is_alive(lp, now))
887 lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
892 * \param msg The message to be sent.
893 * \param do_send True if lnet_ni_send() should be called in this function.
894 * lnet_send() is going to lnet_net_unlock immediately after this, so
895 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
897 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
898 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
899 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
900 * \retval -ECANCELED If the MD of the message has been unlinked.
903 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
905 struct lnet_peer_ni *lp = msg->msg_txpeer;
906 struct lnet_ni *ni = msg->msg_txni;
907 int cpt = msg->msg_tx_cpt;
908 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
910 /* non-lnet_send() callers have checked before */
911 LASSERT(!do_send || msg->msg_tx_delayed);
912 LASSERT(!msg->msg_receiving);
913 LASSERT(msg->msg_tx_committed);
915 /* NB 'lp' is always the next hop */
916 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
917 lnet_peer_alive_locked(ni, lp) == 0) {
918 the_lnet.ln_counters[cpt]->drop_count++;
919 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
920 lnet_net_unlock(cpt);
922 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
924 LNET_STATS_TYPE_DROP);
926 lnet_incr_stats(&msg->msg_txni->ni_stats,
928 LNET_STATS_TYPE_DROP);
930 CNETERR("Dropping message for %s: peer not alive\n",
931 libcfs_id2str(msg->msg_target));
933 lnet_finalize(msg, -EHOSTUNREACH);
936 return -EHOSTUNREACH;
939 if (msg->msg_md != NULL &&
940 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
941 lnet_net_unlock(cpt);
943 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
944 "called on the MD/ME.\n",
945 libcfs_id2str(msg->msg_target));
947 lnet_finalize(msg, -ECANCELED);
953 if (!msg->msg_peertxcredit) {
954 spin_lock(&lp->lpni_lock);
955 LASSERT((lp->lpni_txcredits < 0) ==
956 !list_empty(&lp->lpni_txq));
958 msg->msg_peertxcredit = 1;
959 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
960 lp->lpni_txcredits--;
962 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
963 lp->lpni_mintxcredits = lp->lpni_txcredits;
965 if (lp->lpni_txcredits < 0) {
966 msg->msg_tx_delayed = 1;
967 list_add_tail(&msg->msg_list, &lp->lpni_txq);
968 spin_unlock(&lp->lpni_lock);
969 return LNET_CREDIT_WAIT;
971 spin_unlock(&lp->lpni_lock);
974 if (!msg->msg_txcredit) {
975 LASSERT((tq->tq_credits < 0) ==
976 !list_empty(&tq->tq_delayed));
978 msg->msg_txcredit = 1;
980 atomic_dec(&ni->ni_tx_credits);
982 if (tq->tq_credits < tq->tq_credits_min)
983 tq->tq_credits_min = tq->tq_credits;
985 if (tq->tq_credits < 0) {
986 msg->msg_tx_delayed = 1;
987 list_add_tail(&msg->msg_list, &tq->tq_delayed);
988 return LNET_CREDIT_WAIT;
993 lnet_net_unlock(cpt);
994 lnet_ni_send(ni, msg);
997 return LNET_CREDIT_OK;
1001 static struct lnet_rtrbufpool *
1002 lnet_msg2bufpool(struct lnet_msg *msg)
1004 struct lnet_rtrbufpool *rbp;
1007 LASSERT(msg->msg_rx_committed);
1009 cpt = msg->msg_rx_cpt;
1010 rbp = &the_lnet.ln_rtrpools[cpt][0];
1012 LASSERT(msg->msg_len <= LNET_MTU);
1013 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
1015 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
1022 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
1024 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
1025 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
1026 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
1027 * received or OK to receive */
1028 struct lnet_peer_ni *lp = msg->msg_rxpeer;
1029 struct lnet_rtrbufpool *rbp;
1030 struct lnet_rtrbuf *rb;
1032 LASSERT (msg->msg_iov == NULL);
1033 LASSERT (msg->msg_kiov == NULL);
1034 LASSERT (msg->msg_niov == 0);
1035 LASSERT (msg->msg_routing);
1036 LASSERT (msg->msg_receiving);
1037 LASSERT (!msg->msg_sending);
1039 /* non-lnet_parse callers only receive delayed messages */
1040 LASSERT(!do_recv || msg->msg_rx_delayed);
1042 if (!msg->msg_peerrtrcredit) {
1043 spin_lock(&lp->lpni_lock);
1044 LASSERT((lp->lpni_rtrcredits < 0) ==
1045 !list_empty(&lp->lpni_rtrq));
1047 msg->msg_peerrtrcredit = 1;
1048 lp->lpni_rtrcredits--;
1049 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
1050 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
1052 if (lp->lpni_rtrcredits < 0) {
1053 /* must have checked eager_recv before here */
1054 LASSERT(msg->msg_rx_ready_delay);
1055 msg->msg_rx_delayed = 1;
1056 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
1057 spin_unlock(&lp->lpni_lock);
1058 return LNET_CREDIT_WAIT;
1060 spin_unlock(&lp->lpni_lock);
1063 rbp = lnet_msg2bufpool(msg);
1065 if (!msg->msg_rtrcredit) {
1066 msg->msg_rtrcredit = 1;
1068 if (rbp->rbp_credits < rbp->rbp_mincredits)
1069 rbp->rbp_mincredits = rbp->rbp_credits;
1071 if (rbp->rbp_credits < 0) {
1072 /* must have checked eager_recv before here */
1073 LASSERT(msg->msg_rx_ready_delay);
1074 msg->msg_rx_delayed = 1;
1075 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1076 return LNET_CREDIT_WAIT;
1080 LASSERT(!list_empty(&rbp->rbp_bufs));
1081 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1082 list_del(&rb->rb_list);
1084 msg->msg_niov = rbp->rbp_npages;
1085 msg->msg_kiov = &rb->rb_kiov[0];
1088 int cpt = msg->msg_rx_cpt;
1090 lnet_net_unlock(cpt);
1091 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1092 0, msg->msg_len, msg->msg_len);
1095 return LNET_CREDIT_OK;
1099 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1101 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1102 struct lnet_ni *txni = msg->msg_txni;
1103 struct lnet_msg *msg2;
1105 if (msg->msg_txcredit) {
1106 struct lnet_ni *ni = msg->msg_txni;
1107 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1109 /* give back NI txcredits */
1110 msg->msg_txcredit = 0;
1112 LASSERT((tq->tq_credits < 0) ==
1113 !list_empty(&tq->tq_delayed));
1116 atomic_inc(&ni->ni_tx_credits);
1117 if (tq->tq_credits <= 0) {
1118 msg2 = list_entry(tq->tq_delayed.next,
1119 struct lnet_msg, msg_list);
1120 list_del(&msg2->msg_list);
1122 LASSERT(msg2->msg_txni == ni);
1123 LASSERT(msg2->msg_tx_delayed);
1124 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1126 (void) lnet_post_send_locked(msg2, 1);
1130 if (msg->msg_peertxcredit) {
1131 /* give back peer txcredits */
1132 msg->msg_peertxcredit = 0;
1134 spin_lock(&txpeer->lpni_lock);
1135 LASSERT((txpeer->lpni_txcredits < 0) ==
1136 !list_empty(&txpeer->lpni_txq));
1138 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1139 LASSERT(txpeer->lpni_txqnob >= 0);
1141 txpeer->lpni_txcredits++;
1142 if (txpeer->lpni_txcredits <= 0) {
1145 msg2 = list_entry(txpeer->lpni_txq.next,
1146 struct lnet_msg, msg_list);
1147 list_del(&msg2->msg_list);
1148 spin_unlock(&txpeer->lpni_lock);
1150 LASSERT(msg2->msg_txpeer == txpeer);
1151 LASSERT(msg2->msg_tx_delayed);
1153 msg2_cpt = msg2->msg_tx_cpt;
1156 * The msg_cpt can be different from the msg2_cpt
1157 * so we need to make sure we lock the correct cpt
1159 * Once we call lnet_post_send_locked() it is no
1160 * longer safe to access msg2, since it could've
1161 * been freed by lnet_finalize(), but we still
1162 * need to relock the correct cpt, so we cache the
1163 * msg2_cpt for the purpose of the check that
1164 * follows the call to lnet_pose_send_locked().
1166 if (msg2_cpt != msg->msg_tx_cpt) {
1167 lnet_net_unlock(msg->msg_tx_cpt);
1168 lnet_net_lock(msg2_cpt);
1170 (void) lnet_post_send_locked(msg2, 1);
1171 if (msg2_cpt != msg->msg_tx_cpt) {
1172 lnet_net_unlock(msg2_cpt);
1173 lnet_net_lock(msg->msg_tx_cpt);
1176 spin_unlock(&txpeer->lpni_lock);
1181 msg->msg_txni = NULL;
1182 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1185 if (txpeer != NULL) {
1188 * Once the patch for the health comes in we need to set
1189 * the health of the peer ni to bad when we fail to send
1191 * int status = msg->msg_ev.status;
1193 * lnet_set_peer_ni_health_locked(txpeer, false)
1195 msg->msg_txpeer = NULL;
1196 lnet_peer_ni_decref_locked(txpeer);
1201 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1203 struct lnet_msg *msg;
1205 if (list_empty(&rbp->rbp_msgs))
1207 msg = list_entry(rbp->rbp_msgs.next,
1208 struct lnet_msg, msg_list);
1209 list_del(&msg->msg_list);
1211 (void)lnet_post_routed_recv_locked(msg, 1);
1215 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1217 struct lnet_msg *msg;
1218 struct lnet_msg *tmp;
1220 lnet_net_unlock(cpt);
1222 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1223 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1224 0, 0, 0, msg->msg_hdr.payload_length);
1225 list_del_init(&msg->msg_list);
1226 lnet_finalize(msg, -ECANCELED);
1233 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1235 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1236 struct lnet_ni *rxni = msg->msg_rxni;
1237 struct lnet_msg *msg2;
1239 if (msg->msg_rtrcredit) {
1240 /* give back global router credits */
1241 struct lnet_rtrbuf *rb;
1242 struct lnet_rtrbufpool *rbp;
1244 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1245 * there until it gets one allocated, or aborts the wait
1247 LASSERT(msg->msg_kiov != NULL);
1249 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1252 msg->msg_kiov = NULL;
1253 msg->msg_rtrcredit = 0;
1255 LASSERT(rbp == lnet_msg2bufpool(msg));
1257 LASSERT((rbp->rbp_credits > 0) ==
1258 !list_empty(&rbp->rbp_bufs));
1260 /* If routing is now turned off, we just drop this buffer and
1261 * don't bother trying to return credits. */
1262 if (!the_lnet.ln_routing) {
1263 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1267 /* It is possible that a user has lowered the desired number of
1268 * buffers in this pool. Make sure we never put back
1269 * more buffers than the stated number. */
1270 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1271 /* Discard this buffer so we don't have too
1273 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1274 rbp->rbp_nbuffers--;
1276 list_add(&rb->rb_list, &rbp->rbp_bufs);
1278 if (rbp->rbp_credits <= 0)
1279 lnet_schedule_blocked_locked(rbp);
1284 if (msg->msg_peerrtrcredit) {
1285 /* give back peer router credits */
1286 msg->msg_peerrtrcredit = 0;
1288 spin_lock(&rxpeer->lpni_lock);
1289 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1290 !list_empty(&rxpeer->lpni_rtrq));
1292 rxpeer->lpni_rtrcredits++;
1294 /* drop all messages which are queued to be routed on that
1296 if (!the_lnet.ln_routing) {
1297 struct list_head drop;
1298 INIT_LIST_HEAD(&drop);
1299 list_splice_init(&rxpeer->lpni_rtrq, &drop);
1300 spin_unlock(&rxpeer->lpni_lock);
1301 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1302 } else if (rxpeer->lpni_rtrcredits <= 0) {
1303 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1304 struct lnet_msg, msg_list);
1305 list_del(&msg2->msg_list);
1306 spin_unlock(&rxpeer->lpni_lock);
1307 (void) lnet_post_routed_recv_locked(msg2, 1);
1309 spin_unlock(&rxpeer->lpni_lock);
1313 msg->msg_rxni = NULL;
1314 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1316 if (rxpeer != NULL) {
1317 msg->msg_rxpeer = NULL;
1318 lnet_peer_ni_decref_locked(rxpeer);
1323 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1325 if (p1->lpni_txqnob < p2->lpni_txqnob)
1328 if (p1->lpni_txqnob > p2->lpni_txqnob)
1331 if (p1->lpni_txcredits > p2->lpni_txcredits)
1334 if (p1->lpni_txcredits < p2->lpni_txcredits)
1341 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1343 struct lnet_peer_ni *p1 = r1->lr_gateway;
1344 struct lnet_peer_ni *p2 = r2->lr_gateway;
1345 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1346 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1349 if (r1->lr_priority < r2->lr_priority)
1352 if (r1->lr_priority > r2->lr_priority)
1355 if (r1_hops < r2_hops)
1358 if (r1_hops > r2_hops)
1361 rc = lnet_compare_peers(p1, p2);
1365 if (r1->lr_seq - r2->lr_seq <= 0)
1371 static struct lnet_peer_ni *
1372 lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
1375 struct lnet_remotenet *rnet;
1376 struct lnet_route *route;
1377 struct lnet_route *best_route;
1378 struct lnet_route *last_route;
1379 struct lnet_peer_ni *lpni_best;
1380 struct lnet_peer_ni *lp;
1383 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1384 * rtr_nid nid, otherwise find the best gateway I can use */
1386 rnet = lnet_find_rnet_locked(LNET_NIDNET(target));
1391 best_route = last_route = NULL;
1392 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1393 lp = route->lr_gateway;
1395 if (!lnet_is_route_alive(route))
1398 if (net != NULL && lp->lpni_net != net)
1401 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1404 if (lpni_best == NULL) {
1405 best_route = last_route = route;
1410 /* no protection on below fields, but it's harmless */
1411 if (last_route->lr_seq - route->lr_seq < 0)
1414 rc = lnet_compare_routes(route, best_route);
1422 /* set sequence number on the best router to the latest sequence + 1
1423 * so we can round-robin all routers, it's race and inaccurate but
1424 * harmless and functional */
1425 if (best_route != NULL)
1426 best_route->lr_seq = last_route->lr_seq + 1;
1430 static struct lnet_ni *
1431 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *cur_ni,
1434 struct lnet_ni *ni = NULL, *best_ni = cur_ni;
1435 unsigned int shortest_distance;
1438 if (best_ni == NULL) {
1439 shortest_distance = UINT_MAX;
1440 best_credits = INT_MIN;
1442 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1443 best_ni->ni_dev_cpt);
1444 best_credits = atomic_read(&best_ni->ni_tx_credits);
1447 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1448 unsigned int distance;
1451 if (!lnet_is_ni_healthy_locked(ni))
1454 ni_credits = atomic_read(&ni->ni_tx_credits);
1457 * calculate the distance from the CPT on which
1458 * the message memory is allocated to the CPT of
1459 * the NI's physical device
1461 distance = cfs_cpt_distance(lnet_cpt_table(),
1466 * All distances smaller than the NUMA range
1467 * are treated equally.
1469 if (distance < lnet_numa_range)
1470 distance = lnet_numa_range;
1473 * Select on shorter distance, then available
1474 * credits, then round-robin.
1476 if (distance > shortest_distance) {
1478 } else if (distance < shortest_distance) {
1479 shortest_distance = distance;
1480 } else if (ni_credits < best_credits) {
1482 } else if (ni_credits == best_credits) {
1483 if (best_ni && (best_ni)->ni_seq <= ni->ni_seq)
1487 best_credits = ni_credits;
1494 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1495 * because such traffic is required to perform discovery. We therefore
1496 * exclude all GET and PUT on that portal. We also exclude all ACK and
1497 * REPLY traffic, but that is because the portal is not tracked in the
1498 * message structure for these message types. We could restrict this
1499 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1502 lnet_msg_discovery(struct lnet_msg *msg)
1504 if (msg->msg_type == LNET_MSG_PUT) {
1505 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1507 } else if (msg->msg_type == LNET_MSG_GET) {
1508 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1515 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
1516 struct lnet_msg *msg, lnet_nid_t rtr_nid)
1518 struct lnet_ni *best_ni;
1519 struct lnet_peer_ni *best_lpni;
1520 struct lnet_peer_ni *best_gw;
1521 struct lnet_peer_ni *lpni;
1522 struct lnet_peer_ni *final_dst;
1523 struct lnet_peer *peer;
1524 struct lnet_peer_net *peer_net;
1525 struct lnet_net *local_net;
1532 int best_lpni_credits;
1536 * get an initial CPT to use for locking. The idea here is not to
1537 * serialize the calls to select_pathway, so that as many
1538 * operations can run concurrently as possible. To do that we use
1539 * the CPT where this call is being executed. Later on when we
1540 * determine the CPT to use in lnet_message_commit, we switch the
1541 * lock and check if there was any configuration change. If none,
1542 * then we proceed, if there is, then we restart the operation.
1544 cpt = lnet_net_lock_current();
1546 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
1547 if (md_cpt == CFS_CPT_ANY)
1558 local_found = false;
1561 * lnet_nid2peerni_locked() is the path that will find an
1562 * existing peer_ni, or create one and mark it as having been
1563 * created due to network traffic.
1565 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
1567 lnet_net_unlock(cpt);
1568 return PTR_ERR(lpni);
1572 * If we're being asked to send to the loopback interface, there
1573 * is no need to go through any selection. We can just shortcut
1574 * the entire process and send over lolnd
1576 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
1577 lnet_peer_ni_decref_locked(lpni);
1578 best_ni = the_lnet.ln_loni;
1583 * Now that we have a peer_ni, check if we want to discover
1584 * the peer. Traffic to the LNET_RESERVED_PORTAL should not
1585 * trigger discovery.
1587 peer = lpni->lpni_peer_net->lpn_peer;
1588 if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
1589 rc = lnet_discover_peer_locked(lpni, cpt, false);
1591 lnet_peer_ni_decref_locked(lpni);
1592 lnet_net_unlock(cpt);
1595 /* The peer may have changed. */
1596 peer = lpni->lpni_peer_net->lpn_peer;
1597 /* queue message and return */
1598 msg->msg_src_nid_param = src_nid;
1599 msg->msg_rtr_nid_param = rtr_nid;
1600 msg->msg_sending = 0;
1601 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1602 CDEBUG(D_NET, "%s pending discovery\n",
1603 libcfs_nid2str(peer->lp_primary_nid));
1604 lnet_peer_ni_decref_locked(lpni);
1605 lnet_net_unlock(cpt);
1607 return LNET_DC_WAIT;
1609 lnet_peer_ni_decref_locked(lpni);
1611 /* If peer is not healthy then can not send anything to it */
1612 if (!lnet_is_peer_healthy_locked(peer)) {
1613 lnet_net_unlock(cpt);
1614 return -EHOSTUNREACH;
1618 * STEP 1: first jab at determining best_ni
1619 * if src_nid is explicitly specified, then best_ni is already
1620 * pre-determiend for us. Otherwise we need to select the best
1621 * one to use later on
1623 if (src_nid != LNET_NID_ANY) {
1624 best_ni = lnet_nid2ni_locked(src_nid, cpt);
1626 lnet_net_unlock(cpt);
1627 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1628 "local nid\n", libcfs_nid2str(dst_nid),
1629 libcfs_nid2str(src_nid));
1634 if (msg->msg_type == LNET_MSG_REPLY ||
1635 msg->msg_type == LNET_MSG_ACK ||
1636 !lnet_peer_is_multi_rail(peer) ||
1639 * for replies we want to respond on the same peer_ni we
1640 * received the message on if possible. If not, then pick
1641 * a peer_ni to send to
1643 * if the peer is non-multi-rail then you want to send to
1644 * the dst_nid provided as well.
1646 * If the best_ni has already been determined, IE the
1647 * src_nid has been specified, then use the
1648 * destination_nid provided as well, since we're
1649 * continuing a series of related messages for the same
1652 * It is expected to find the lpni using dst_nid, since we
1653 * created it earlier.
1655 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1657 lnet_peer_ni_decref_locked(best_lpni);
1659 if (best_lpni && !lnet_get_net_locked(LNET_NIDNET(dst_nid))) {
1661 * this lpni is not on a local network so we need
1662 * to route this reply.
1664 best_gw = lnet_find_route_locked(NULL,
1665 best_lpni->lpni_nid,
1669 * RULE: Each node considers only the next-hop
1671 * We're going to route the message, so change the peer to
1674 LASSERT(best_gw->lpni_peer_net);
1675 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1676 peer = best_gw->lpni_peer_net->lpn_peer;
1679 * if the router is not multi-rail then use the best_gw
1680 * found to send the message to
1682 if (!lnet_peer_is_multi_rail(peer))
1683 best_lpni = best_gw;
1691 } else if (!best_lpni) {
1692 lnet_net_unlock(cpt);
1693 CERROR("unable to send msg_type %d to "
1694 "originating %s. Destination NID not in DB\n",
1695 msg->msg_type, libcfs_nid2str(dst_nid));
1701 * We must use a consistent source address when sending to a
1702 * non-MR peer. However, a non-MR peer can have multiple NIDs
1703 * on multiple networks, and we may even need to talk to this
1704 * peer on multiple networks -- certain types of
1705 * load-balancing configuration do this.
1707 * So we need to pick the NI the peer prefers for this
1708 * particular network.
1710 if (!lnet_peer_is_multi_rail(peer)) {
1712 lnet_net_unlock(cpt);
1713 CERROR("no route to %s\n",
1714 libcfs_nid2str(dst_nid));
1715 return -EHOSTUNREACH;
1718 /* best ni is already set if src_nid was provided */
1720 /* Get the target peer_ni */
1721 peer_net = lnet_peer_get_net_locked(peer,
1722 LNET_NIDNET(best_lpni->lpni_nid));
1723 LASSERT(peer_net != NULL);
1724 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
1726 if (lpni->lpni_pref_nnids == 0)
1728 LASSERT(lpni->lpni_pref_nnids == 1);
1729 best_ni = lnet_nid2ni_locked(
1730 lpni->lpni_pref.nid, cpt);
1734 /* if best_ni is still not set just pick one */
1736 best_ni = lnet_net2ni_locked(
1737 best_lpni->lpni_net->net_id, cpt);
1738 /* If there is no best_ni we don't have a route */
1740 CERROR("no path to %s from net %s\n",
1741 libcfs_nid2str(best_lpni->lpni_nid),
1742 libcfs_net2str(best_lpni->lpni_net->net_id));
1743 lnet_net_unlock(cpt);
1744 return -EHOSTUNREACH;
1746 lpni = list_entry(peer_net->lpn_peer_nis.next,
1747 struct lnet_peer_ni,
1750 /* Set preferred NI if necessary. */
1751 if (lpni->lpni_pref_nnids == 0)
1752 lnet_peer_ni_set_non_mr_pref_nid(lpni, best_ni->ni_nid);
1756 * if we already found a best_ni because src_nid is specified and
1757 * best_lpni because we are replying to a message then just send
1760 if (best_ni && best_lpni)
1764 * If we already found a best_ni because src_nid is specified then
1765 * pick the peer then send the message
1771 * pick the best_ni by going through all the possible networks of
1772 * that peer and see which local NI is best suited to talk to that
1775 * Locally connected networks will always be preferred over
1776 * a routed network. If there are only routed paths to the peer,
1777 * then the best route is chosen. If all routes are equal then
1778 * they are used in round robin.
1780 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1781 if (!lnet_is_peer_net_healthy_locked(peer_net))
1784 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1785 if (!local_net && !routing && !local_found) {
1786 struct lnet_peer_ni *net_gw;
1788 lpni = list_entry(peer_net->lpn_peer_nis.next,
1789 struct lnet_peer_ni,
1792 net_gw = lnet_find_route_locked(NULL,
1800 * lnet_find_route_locked() call
1801 * will return the best_Gw on the
1802 * lpni->lpni_nid network.
1803 * However, best_gw and net_gw can
1804 * be on different networks.
1805 * Therefore need to compare them
1806 * to pick the better of either.
1808 if (lnet_compare_peers(best_gw, net_gw) > 0)
1810 if (best_gw->lpni_gw_seq <= net_gw->lpni_gw_seq)
1825 * a gw on this network is found, but there could be
1826 * other better gateways on other networks. So don't pick
1827 * the best_ni until we determine the best_gw.
1832 /* if no local_net found continue */
1837 * Iterate through the NIs in this local Net and select
1838 * the NI to send from. The selection is determined by
1839 * these 3 criterion in the following priority:
1841 * 2. NI available credits
1844 best_ni = lnet_get_best_ni(local_net, best_ni, md_cpt);
1847 if (!best_ni && !best_gw) {
1848 lnet_net_unlock(cpt);
1849 LCONSOLE_WARN("No local ni found to send from to %s\n",
1850 libcfs_nid2str(dst_nid));
1855 best_ni = lnet_get_best_ni(best_gw->lpni_net, best_ni, md_cpt);
1856 LASSERT(best_gw && best_ni);
1859 * We're going to route the message, so change the peer to
1862 LASSERT(best_gw->lpni_peer_net);
1863 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1864 best_gw->lpni_gw_seq++;
1865 peer = best_gw->lpni_peer_net->lpn_peer;
1869 * Now that we selected the NI to use increment its sequence
1870 * number so the Round Robin algorithm will detect that it has
1871 * been used and pick the next NI.
1877 * At this point the best_ni is on a local network on which
1878 * the peer has a peer_ni as well
1880 peer_net = lnet_peer_get_net_locked(peer,
1881 best_ni->ni_net->net_id);
1883 * peer_net is not available or the src_nid is explicitly defined
1884 * and the peer_net for that src_nid is unhealthy. find a route to
1885 * the destination nid.
1888 (src_nid != LNET_NID_ANY &&
1889 !lnet_is_peer_net_healthy_locked(peer_net))) {
1890 best_gw = lnet_find_route_locked(best_ni->ni_net,
1894 * if no route is found for that network then
1895 * move onto the next peer_ni in the peer
1898 LCONSOLE_WARN("No route to peer from %s\n",
1899 libcfs_nid2str(best_ni->ni_nid));
1900 lnet_net_unlock(cpt);
1901 return -EHOSTUNREACH;
1904 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1905 libcfs_nid2str(dst_nid),
1906 libcfs_nid2str(best_gw->lpni_nid),
1907 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1911 * RULE: Each node considers only the next-hop
1913 * We're going to route the message, so change the peer to
1916 LASSERT(best_gw->lpni_peer_net);
1917 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1918 peer = best_gw->lpni_peer_net->lpn_peer;
1919 } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
1921 * this peer_net is unhealthy but we still have an opportunity
1922 * to find another peer_net that we can use
1924 __u32 net_id = peer_net->lpn_net_id;
1925 LCONSOLE_WARN("peer net %s unhealthy\n",
1926 libcfs_net2str(net_id));
1931 * Look at the peer NIs for the destination peer that connect
1932 * to the chosen net. If a peer_ni is preferred when using the
1933 * best_ni to communicate, we use that one. If there is no
1934 * preferred peer_ni, or there are multiple preferred peer_ni,
1935 * the available transmit credits are used. If the transmit
1936 * credits are equal, we round-robin over the peer_ni.
1939 best_lpni_credits = INT_MIN;
1942 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1944 * if this peer ni is not healthy just skip it, no point in
1945 * examining it further
1947 if (!lnet_is_peer_ni_healthy_locked(lpni))
1949 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1952 /* if this is a preferred peer use it */
1953 if (!preferred && ni_is_pref) {
1955 } else if (preferred && !ni_is_pref) {
1957 * this is not the preferred peer so let's ignore
1961 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1963 * We already have a peer that has more credits
1964 * available than this one. No need to consider
1965 * this peer further.
1968 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1970 * The best peer found so far and the current peer
1971 * have the same number of available credits let's
1972 * make sure to select between them using Round
1976 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1982 best_lpni_credits = lpni->lpni_txcredits;
1985 /* if we still can't find a peer ni then we can't reach it */
1987 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1988 LNET_NIDNET(dst_nid);
1989 lnet_net_unlock(cpt);
1990 LCONSOLE_WARN("no peer_ni found on peer net %s\n",
1991 libcfs_net2str(net_id));
1992 return -EHOSTUNREACH;
1997 /* Shortcut for loopback. */
1998 if (best_ni == the_lnet.ln_loni) {
1999 /* No send credit hassles with LOLND */
2000 lnet_ni_addref_locked(best_ni, cpt);
2001 msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
2002 if (!msg->msg_routing)
2003 msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
2004 msg->msg_target.nid = best_ni->ni_nid;
2005 lnet_msg_commit(msg, cpt);
2006 msg->msg_txni = best_ni;
2007 lnet_net_unlock(cpt);
2009 return LNET_CREDIT_OK;
2012 routing = routing || routing2;
2015 * Increment sequence number of the peer selected so that we
2016 * pick the next one in Round Robin.
2018 best_lpni->lpni_seq++;
2021 * grab a reference on the peer_ni so it sticks around even if
2022 * we need to drop and relock the lnet_net_lock below.
2024 lnet_peer_ni_addref_locked(best_lpni);
2027 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
2028 * message. This ensures that we get a CPT that is correct for
2029 * the NI when the NI has been restricted to a subset of all CPTs.
2030 * If the selected CPT differs from the one currently locked, we
2031 * must unlock and relock the lnet_net_lock(), and then check whether
2032 * the configuration has changed. We don't have a hold on the best_ni
2033 * yet, and it may have vanished.
2035 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
2037 __u32 seq = lnet_get_dlc_seq_locked();
2038 lnet_net_unlock(cpt);
2041 if (seq != lnet_get_dlc_seq_locked()) {
2042 lnet_peer_ni_decref_locked(best_lpni);
2048 * store the best_lpni in the message right away to avoid having
2049 * to do the same operation under different conditions
2051 msg->msg_txpeer = best_lpni;
2052 msg->msg_txni = best_ni;
2055 * grab a reference for the best_ni since now it's in use in this
2056 * send. the reference will need to be dropped when the message is
2057 * finished in lnet_finalize()
2059 lnet_ni_addref_locked(msg->msg_txni, cpt);
2062 * Always set the target.nid to the best peer picked. Either the
2063 * nid will be one of the preconfigured NIDs, or the same NID as
2064 * what was originally set in the target or it will be the NID of
2065 * a router if this message should be routed
2067 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
2070 * lnet_msg_commit assigns the correct cpt to the message, which
2071 * is used to decrement the correct refcount on the ni when it's
2072 * time to return the credits
2074 lnet_msg_commit(msg, cpt);
2077 * If we are routing the message then we don't need to overwrite
2078 * the src_nid since it would've been set at the origin. Otherwise
2079 * we are the originator so we need to set it.
2081 if (!msg->msg_routing)
2082 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
2085 msg->msg_target_is_router = 1;
2086 msg->msg_target.pid = LNET_PID_LUSTRE;
2088 * since we're routing we want to ensure that the
2089 * msg_hdr.dest_nid is set to the final destination. When
2090 * the router receives this message it knows how to route
2093 msg->msg_hdr.dest_nid =
2094 cpu_to_le64(final_dst ? final_dst->lpni_nid : dst_nid);
2097 * if we're not routing set the dest_nid to the best peer
2098 * ni that we picked earlier in the algorithm.
2100 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
2103 rc = lnet_post_send_locked(msg, 0);
2106 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s\n",
2107 libcfs_nid2str(msg->msg_hdr.src_nid),
2108 libcfs_nid2str(msg->msg_txni->ni_nid),
2109 libcfs_nid2str(src_nid),
2110 libcfs_nid2str(msg->msg_hdr.dest_nid),
2111 libcfs_nid2str(dst_nid),
2112 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
2113 lnet_msgtyp2str(msg->msg_type));
2115 lnet_net_unlock(cpt);
2121 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2123 lnet_nid_t dst_nid = msg->msg_target.nid;
2127 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2128 * but we might want to use pre-determined router for ACK/REPLY
2131 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2132 LASSERT (msg->msg_txpeer == NULL);
2133 LASSERT (!msg->msg_sending);
2134 LASSERT (!msg->msg_target_is_router);
2135 LASSERT (!msg->msg_receiving);
2137 msg->msg_sending = 1;
2139 LASSERT(!msg->msg_tx_committed);
2141 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2145 if (rc == LNET_CREDIT_OK)
2146 lnet_ni_send(msg->msg_txni, msg);
2148 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2153 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
2157 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
2158 the_lnet.ln_counters[cpt]->drop_count++;
2159 the_lnet.ln_counters[cpt]->drop_length += nob;
2160 lnet_net_unlock(cpt);
2162 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
2166 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
2168 struct lnet_hdr *hdr = &msg->msg_hdr;
2170 if (msg->msg_wanted != 0)
2171 lnet_setpayloadbuffer(msg);
2173 lnet_build_msg_event(msg, LNET_EVENT_PUT);
2175 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
2176 * it back into the ACK during lnet_finalize() */
2177 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
2178 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
2180 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
2181 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
2185 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
2187 struct lnet_hdr *hdr = &msg->msg_hdr;
2188 struct lnet_match_info info;
2192 /* Convert put fields to host byte order */
2193 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
2194 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
2195 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
2197 /* Primary peer NID. */
2198 info.mi_id.nid = msg->msg_initiator;
2199 info.mi_id.pid = hdr->src_pid;
2200 info.mi_opc = LNET_MD_OP_PUT;
2201 info.mi_portal = hdr->msg.put.ptl_index;
2202 info.mi_rlength = hdr->payload_length;
2203 info.mi_roffset = hdr->msg.put.offset;
2204 info.mi_mbits = hdr->msg.put.match_bits;
2205 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
2207 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
2208 ready_delay = msg->msg_rx_ready_delay;
2211 rc = lnet_ptl_match_md(&info, msg);
2216 case LNET_MATCHMD_OK:
2217 lnet_recv_put(ni, msg);
2220 case LNET_MATCHMD_NONE:
2222 /* no eager_recv or has already called it, should
2223 * have been attached on delayed list */
2226 rc = lnet_ni_eager_recv(ni, msg);
2233 case LNET_MATCHMD_DROP:
2234 CNETERR("Dropping PUT from %s portal %d match %llu"
2235 " offset %d length %d: %d\n",
2236 libcfs_id2str(info.mi_id), info.mi_portal,
2237 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
2239 return -ENOENT; /* -ve: OK but no match */
2244 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
2246 struct lnet_match_info info;
2247 struct lnet_hdr *hdr = &msg->msg_hdr;
2248 struct lnet_process_id source_id;
2249 struct lnet_handle_wire reply_wmd;
2252 /* Convert get fields to host byte order */
2253 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
2254 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
2255 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
2256 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
2258 source_id.nid = hdr->src_nid;
2259 source_id.pid = hdr->src_pid;
2260 /* Primary peer NID */
2261 info.mi_id.nid = msg->msg_initiator;
2262 info.mi_id.pid = hdr->src_pid;
2263 info.mi_opc = LNET_MD_OP_GET;
2264 info.mi_portal = hdr->msg.get.ptl_index;
2265 info.mi_rlength = hdr->msg.get.sink_length;
2266 info.mi_roffset = hdr->msg.get.src_offset;
2267 info.mi_mbits = hdr->msg.get.match_bits;
2268 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
2270 rc = lnet_ptl_match_md(&info, msg);
2271 if (rc == LNET_MATCHMD_DROP) {
2272 CNETERR("Dropping GET from %s portal %d match %llu"
2273 " offset %d length %d\n",
2274 libcfs_id2str(info.mi_id), info.mi_portal,
2275 info.mi_mbits, info.mi_roffset, info.mi_rlength);
2276 return -ENOENT; /* -ve: OK but no match */
2279 LASSERT(rc == LNET_MATCHMD_OK);
2281 lnet_build_msg_event(msg, LNET_EVENT_GET);
2283 reply_wmd = hdr->msg.get.return_wmd;
2285 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
2286 msg->msg_offset, msg->msg_wanted);
2288 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
2291 /* The LND completes the REPLY from her recv procedure */
2292 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2293 msg->msg_offset, msg->msg_len, msg->msg_len);
2297 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
2298 msg->msg_receiving = 0;
2300 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
2302 /* didn't get as far as lnet_ni_send() */
2303 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
2304 libcfs_nid2str(ni->ni_nid),
2305 libcfs_id2str(info.mi_id), rc);
2307 lnet_finalize(msg, rc);
2314 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
2316 void *private = msg->msg_private;
2317 struct lnet_hdr *hdr = &msg->msg_hdr;
2318 struct lnet_process_id src = {0};
2319 struct lnet_libmd *md;
2324 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
2327 src.nid = hdr->src_nid;
2328 src.pid = hdr->src_pid;
2330 /* NB handles only looked up by creator (no flips) */
2331 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
2332 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2333 CNETERR("%s: Dropping REPLY from %s for %s "
2335 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2336 (md == NULL) ? "invalid" : "inactive",
2337 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2338 hdr->msg.reply.dst_wmd.wh_object_cookie);
2339 if (md != NULL && md->md_me != NULL)
2340 CERROR("REPLY MD also attached to portal %d\n",
2341 md->md_me->me_portal);
2343 lnet_res_unlock(cpt);
2344 return -ENOENT; /* -ve: OK but no match */
2347 LASSERT(md->md_offset == 0);
2349 rlength = hdr->payload_length;
2350 mlength = MIN(rlength, (int)md->md_length);
2352 if (mlength < rlength &&
2353 (md->md_options & LNET_MD_TRUNCATE) == 0) {
2354 CNETERR("%s: Dropping REPLY from %s length %d "
2355 "for MD %#llx would overflow (%d)\n",
2356 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2357 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2359 lnet_res_unlock(cpt);
2360 return -ENOENT; /* -ve: OK but no match */
2363 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
2364 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2365 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2367 lnet_msg_attach_md(msg, md, 0, mlength);
2370 lnet_setpayloadbuffer(msg);
2372 lnet_res_unlock(cpt);
2374 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2376 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2381 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
2383 struct lnet_hdr *hdr = &msg->msg_hdr;
2384 struct lnet_process_id src = {0};
2385 struct lnet_libmd *md;
2388 src.nid = hdr->src_nid;
2389 src.pid = hdr->src_pid;
2391 /* Convert ack fields to host byte order */
2392 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2393 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2395 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
2398 /* NB handles only looked up by creator (no flips) */
2399 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2400 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2401 /* Don't moan; this is expected */
2403 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
2404 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2405 (md == NULL) ? "invalid" : "inactive",
2406 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2407 hdr->msg.ack.dst_wmd.wh_object_cookie);
2408 if (md != NULL && md->md_me != NULL)
2409 CERROR("Source MD also attached to portal %d\n",
2410 md->md_me->me_portal);
2412 lnet_res_unlock(cpt);
2413 return -ENOENT; /* -ve! */
2416 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
2417 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2418 hdr->msg.ack.dst_wmd.wh_object_cookie);
2420 lnet_msg_attach_md(msg, md, 0, 0);
2422 lnet_res_unlock(cpt);
2424 lnet_build_msg_event(msg, LNET_EVENT_ACK);
2426 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2431 * \retval LNET_CREDIT_OK If \a msg is forwarded
2432 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
2433 * \retval -ve error code
2436 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
2440 if (!the_lnet.ln_routing)
2443 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
2444 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2445 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
2446 msg->msg_rx_ready_delay = 1;
2448 lnet_net_unlock(msg->msg_rx_cpt);
2449 rc = lnet_ni_eager_recv(ni, msg);
2450 lnet_net_lock(msg->msg_rx_cpt);
2455 rc = lnet_post_routed_recv_locked(msg, 0);
2460 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
2464 switch (msg->msg_type) {
2466 rc = lnet_parse_ack(ni, msg);
2469 rc = lnet_parse_put(ni, msg);
2472 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
2474 case LNET_MSG_REPLY:
2475 rc = lnet_parse_reply(ni, msg);
2477 default: /* prevent an unused label if !kernel */
2482 LASSERT(rc == 0 || rc == -ENOENT);
2487 lnet_msgtyp2str (int type)
2496 case LNET_MSG_REPLY:
2498 case LNET_MSG_HELLO:
2501 return ("<UNKNOWN>");
2506 lnet_print_hdr(struct lnet_hdr *hdr)
2508 struct lnet_process_id src = {
2509 .nid = hdr->src_nid,
2510 .pid = hdr->src_pid,
2512 struct lnet_process_id dst = {
2513 .nid = hdr->dest_nid,
2514 .pid = hdr->dest_pid,
2516 char *type_str = lnet_msgtyp2str(hdr->type);
2518 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2519 CWARN(" From %s\n", libcfs_id2str(src));
2520 CWARN(" To %s\n", libcfs_id2str(dst));
2522 switch (hdr->type) {
2527 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
2528 "match bits %llu\n",
2529 hdr->msg.put.ptl_index,
2530 hdr->msg.put.ack_wmd.wh_interface_cookie,
2531 hdr->msg.put.ack_wmd.wh_object_cookie,
2532 hdr->msg.put.match_bits);
2533 CWARN(" Length %d, offset %d, hdr data %#llx\n",
2534 hdr->payload_length, hdr->msg.put.offset,
2535 hdr->msg.put.hdr_data);
2539 CWARN(" Ptl index %d, return md %#llx.%#llx, "
2540 "match bits %llu\n", hdr->msg.get.ptl_index,
2541 hdr->msg.get.return_wmd.wh_interface_cookie,
2542 hdr->msg.get.return_wmd.wh_object_cookie,
2543 hdr->msg.get.match_bits);
2544 CWARN(" Length %d, src offset %d\n",
2545 hdr->msg.get.sink_length,
2546 hdr->msg.get.src_offset);
2550 CWARN(" dst md %#llx.%#llx, "
2551 "manipulated length %d\n",
2552 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2553 hdr->msg.ack.dst_wmd.wh_object_cookie,
2554 hdr->msg.ack.mlength);
2557 case LNET_MSG_REPLY:
2558 CWARN(" dst md %#llx.%#llx, "
2560 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2561 hdr->msg.reply.dst_wmd.wh_object_cookie,
2562 hdr->payload_length);
2568 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
2569 void *private, int rdma_req)
2574 struct lnet_msg *msg;
2575 lnet_pid_t dest_pid;
2576 lnet_nid_t dest_nid;
2578 struct lnet_peer_ni *lpni;
2579 __u32 payload_length;
2582 LASSERT (!in_interrupt ());
2584 type = le32_to_cpu(hdr->type);
2585 src_nid = le64_to_cpu(hdr->src_nid);
2586 dest_nid = le64_to_cpu(hdr->dest_nid);
2587 dest_pid = le32_to_cpu(hdr->dest_pid);
2588 payload_length = le32_to_cpu(hdr->payload_length);
2590 for_me = (ni->ni_nid == dest_nid);
2591 cpt = lnet_cpt_of_nid(from_nid, ni);
2593 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
2594 libcfs_nid2str(dest_nid),
2595 libcfs_nid2str(ni->ni_nid),
2596 libcfs_nid2str(src_nid),
2597 lnet_msgtyp2str(type),
2598 (for_me) ? "for me" : "routed");
2603 if (payload_length > 0) {
2604 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2605 libcfs_nid2str(from_nid),
2606 libcfs_nid2str(src_nid),
2607 lnet_msgtyp2str(type), payload_length);
2613 case LNET_MSG_REPLY:
2614 if (payload_length >
2615 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2616 CERROR("%s, src %s: bad %s payload %d "
2617 "(%d max expected)\n",
2618 libcfs_nid2str(from_nid),
2619 libcfs_nid2str(src_nid),
2620 lnet_msgtyp2str(type),
2622 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2628 CERROR("%s, src %s: Bad message type 0x%x\n",
2629 libcfs_nid2str(from_nid),
2630 libcfs_nid2str(src_nid), type);
2634 if (the_lnet.ln_routing &&
2635 ni->ni_last_alive != ktime_get_real_seconds()) {
2636 /* NB: so far here is the only place to set NI status to "up */
2638 ni->ni_last_alive = ktime_get_real_seconds();
2639 if (ni->ni_status != NULL &&
2640 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
2641 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
2645 /* Regard a bad destination NID as a protocol error. Senders should
2646 * know what they're doing; if they don't they're misconfigured, buggy
2647 * or malicious so we chop them off at the knees :) */
2650 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2651 /* should have gone direct */
2652 CERROR("%s, src %s: Bad dest nid %s "
2653 "(should have been sent direct)\n",
2654 libcfs_nid2str(from_nid),
2655 libcfs_nid2str(src_nid),
2656 libcfs_nid2str(dest_nid));
2660 if (lnet_islocalnid(dest_nid)) {
2661 /* dest is another local NI; sender should have used
2662 * this node's NID on its own network */
2663 CERROR("%s, src %s: Bad dest nid %s "
2664 "(it's my nid but on a different network)\n",
2665 libcfs_nid2str(from_nid),
2666 libcfs_nid2str(src_nid),
2667 libcfs_nid2str(dest_nid));
2671 if (rdma_req && type == LNET_MSG_GET) {
2672 CERROR("%s, src %s: Bad optimized GET for %s "
2673 "(final destination must be me)\n",
2674 libcfs_nid2str(from_nid),
2675 libcfs_nid2str(src_nid),
2676 libcfs_nid2str(dest_nid));
2680 if (!the_lnet.ln_routing) {
2681 CERROR("%s, src %s: Dropping message for %s "
2682 "(routing not enabled)\n",
2683 libcfs_nid2str(from_nid),
2684 libcfs_nid2str(src_nid),
2685 libcfs_nid2str(dest_nid));
2690 /* Message looks OK; we're not going to return an error, so we MUST
2691 * call back lnd_recv() come what may... */
2693 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2694 fail_peer(src_nid, 0)) { /* shall we now? */
2695 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2696 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2697 lnet_msgtyp2str(type));
2701 if (!list_empty(&the_lnet.ln_drop_rules) &&
2702 lnet_drop_rule_match(hdr)) {
2703 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
2704 "silent message loss\n",
2705 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2706 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
2711 msg = lnet_msg_alloc();
2713 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2714 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2715 lnet_msgtyp2str(type));
2719 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
2720 * pointers NULL etc */
2722 msg->msg_type = type;
2723 msg->msg_private = private;
2724 msg->msg_receiving = 1;
2725 msg->msg_rdma_get = rdma_req;
2726 msg->msg_len = msg->msg_wanted = payload_length;
2727 msg->msg_offset = 0;
2728 msg->msg_hdr = *hdr;
2729 /* for building message event */
2730 msg->msg_from = from_nid;
2732 msg->msg_target.pid = dest_pid;
2733 msg->msg_target.nid = dest_nid;
2734 msg->msg_routing = 1;
2737 /* convert common msg->hdr fields to host byteorder */
2738 msg->msg_hdr.type = type;
2739 msg->msg_hdr.src_nid = src_nid;
2740 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2741 msg->msg_hdr.dest_nid = dest_nid;
2742 msg->msg_hdr.dest_pid = dest_pid;
2743 msg->msg_hdr.payload_length = payload_length;
2747 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
2749 lnet_net_unlock(cpt);
2750 CERROR("%s, src %s: Dropping %s "
2751 "(error %ld looking up sender)\n",
2752 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2753 lnet_msgtyp2str(type), PTR_ERR(lpni));
2755 if (rc == -ESHUTDOWN)
2756 /* We are shutting down. Don't do anything more */
2760 msg->msg_rxpeer = lpni;
2762 lnet_ni_addref_locked(ni, cpt);
2763 /* Multi-Rail: Primary NID of source. */
2764 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
2766 if (lnet_isrouter(msg->msg_rxpeer)) {
2767 lnet_peer_set_alive(msg->msg_rxpeer);
2768 if (avoid_asym_router_failure &&
2769 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
2770 /* received a remote message from router, update
2771 * remote NI status on this router.
2772 * NB: multi-hop routed message will be ignored.
2774 lnet_router_ni_update_locked(msg->msg_rxpeer,
2775 LNET_NIDNET(src_nid));
2779 lnet_msg_commit(msg, cpt);
2781 /* message delay simulation */
2782 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
2783 lnet_delay_rule_match_locked(hdr, msg))) {
2784 lnet_net_unlock(cpt);
2789 rc = lnet_parse_forward_locked(ni, msg);
2790 lnet_net_unlock(cpt);
2795 if (rc == LNET_CREDIT_OK) {
2796 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2797 0, payload_length, payload_length);
2802 lnet_net_unlock(cpt);
2804 rc = lnet_parse_local(ni, msg);
2810 LASSERT(msg->msg_md == NULL);
2811 lnet_finalize(msg, rc);
2814 lnet_drop_message(ni, cpt, private, payload_length, type);
2817 EXPORT_SYMBOL(lnet_parse);
2820 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
2822 while (!list_empty(head)) {
2823 struct lnet_process_id id = {0};
2824 struct lnet_msg *msg;
2826 msg = list_entry(head->next, struct lnet_msg, msg_list);
2827 list_del(&msg->msg_list);
2829 id.nid = msg->msg_hdr.src_nid;
2830 id.pid = msg->msg_hdr.src_pid;
2832 LASSERT(msg->msg_md == NULL);
2833 LASSERT(msg->msg_rx_delayed);
2834 LASSERT(msg->msg_rxpeer != NULL);
2835 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2837 CWARN("Dropping delayed PUT from %s portal %d match %llu"
2838 " offset %d length %d: %s\n",
2840 msg->msg_hdr.msg.put.ptl_index,
2841 msg->msg_hdr.msg.put.match_bits,
2842 msg->msg_hdr.msg.put.offset,
2843 msg->msg_hdr.payload_length, reason);
2845 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2846 * called lnet_drop_message(), so I just hang onto msg as well
2847 * until that's done */
2849 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
2850 msg->msg_private, msg->msg_len,
2853 * NB: message will not generate event because w/o attached MD,
2854 * but we still should give error code so lnet_msg_decommit()
2855 * can skip counters operations and other checks.
2857 lnet_finalize(msg, -ENOENT);
2862 lnet_recv_delayed_msg_list(struct list_head *head)
2864 while (!list_empty(head)) {
2865 struct lnet_msg *msg;
2866 struct lnet_process_id id;
2868 msg = list_entry(head->next, struct lnet_msg, msg_list);
2869 list_del(&msg->msg_list);
2871 /* md won't disappear under me, since each msg
2872 * holds a ref on it */
2874 id.nid = msg->msg_hdr.src_nid;
2875 id.pid = msg->msg_hdr.src_pid;
2877 LASSERT(msg->msg_rx_delayed);
2878 LASSERT(msg->msg_md != NULL);
2879 LASSERT(msg->msg_rxpeer != NULL);
2880 LASSERT(msg->msg_rxni != NULL);
2881 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2883 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2884 "match %llu offset %d length %d.\n",
2885 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2886 msg->msg_hdr.msg.put.match_bits,
2887 msg->msg_hdr.msg.put.offset,
2888 msg->msg_hdr.payload_length);
2890 lnet_recv_put(msg->msg_rxni, msg);
2895 * Initiate an asynchronous PUT operation.
2897 * There are several events associated with a PUT: completion of the send on
2898 * the initiator node (LNET_EVENT_SEND), and when the send completes
2899 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2900 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2901 * used at the target node to indicate the completion of incoming data
2904 * The local events will be logged in the EQ associated with the MD pointed to
2905 * by \a mdh handle. Using a MD without an associated EQ results in these
2906 * events being discarded. In this case, the caller must have another
2907 * mechanism (e.g., a higher level protocol) for determining when it is safe
2908 * to modify the memory region associated with the MD.
2910 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2911 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2913 * \param self Indicates the NID of a local interface through which to send
2914 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2915 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2916 * must be "free floating" (See LNetMDBind()).
2917 * \param ack Controls whether an acknowledgment is requested.
2918 * Acknowledgments are only sent when they are requested by the initiating
2919 * process and the target MD enables them.
2920 * \param target A process identifier for the target process.
2921 * \param portal The index in the \a target's portal table.
2922 * \param match_bits The match bits to use for MD selection at the target
2924 * \param offset The offset into the target MD (only used when the target
2925 * MD has the LNET_MD_MANAGE_REMOTE option set).
2926 * \param hdr_data 64 bits of user data that can be included in the message
2927 * header. This data is written to an event queue entry at the target if an
2928 * EQ is present on the matching MD.
2930 * \retval 0 Success, and only in this case events will be generated
2931 * and logged to EQ (if it exists).
2932 * \retval -EIO Simulated failure.
2933 * \retval -ENOMEM Memory allocation failure.
2934 * \retval -ENOENT Invalid MD object.
2936 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
2939 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
2940 struct lnet_process_id target, unsigned int portal,
2941 __u64 match_bits, unsigned int offset,
2944 struct lnet_msg *msg;
2945 struct lnet_libmd *md;
2949 LASSERT(the_lnet.ln_refcount > 0);
2951 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2952 fail_peer(target.nid, 1)) { /* shall we now? */
2953 CERROR("Dropping PUT to %s: simulated failure\n",
2954 libcfs_id2str(target));
2958 msg = lnet_msg_alloc();
2960 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
2961 libcfs_id2str(target));
2964 msg->msg_vmflush = !!memory_pressure_get();
2966 cpt = lnet_cpt_of_cookie(mdh.cookie);
2969 md = lnet_handle2md(&mdh);
2970 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2971 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
2972 match_bits, portal, libcfs_id2str(target),
2973 md == NULL ? -1 : md->md_threshold);
2974 if (md != NULL && md->md_me != NULL)
2975 CERROR("Source MD also attached to portal %d\n",
2976 md->md_me->me_portal);
2977 lnet_res_unlock(cpt);
2983 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2985 lnet_msg_attach_md(msg, md, 0, 0);
2987 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2989 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2990 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2991 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2992 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2994 /* NB handles only looked up by creator (no flips) */
2995 if (ack == LNET_ACK_REQ) {
2996 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2997 the_lnet.ln_interface_cookie;
2998 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2999 md->md_lh.lh_cookie;
3001 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
3002 LNET_WIRE_HANDLE_COOKIE_NONE;
3003 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
3004 LNET_WIRE_HANDLE_COOKIE_NONE;
3007 lnet_res_unlock(cpt);
3009 lnet_build_msg_event(msg, LNET_EVENT_SEND);
3011 rc = lnet_send(self, msg, LNET_NID_ANY);
3013 CNETERR("Error sending PUT to %s: %d\n",
3014 libcfs_id2str(target), rc);
3015 lnet_finalize(msg, rc);
3018 /* completion will be signalled by an event */
3021 EXPORT_SYMBOL(LNetPut);
3024 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
3025 * returns a msg for the LND to pass to lnet_finalize() when the sink
3026 * data has been received.
3028 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
3029 * lnet_finalize() is called on it, so the LND must call this first
3032 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
3034 struct lnet_msg *msg = lnet_msg_alloc();
3035 struct lnet_libmd *getmd = getmsg->msg_md;
3036 struct lnet_process_id peer_id = getmsg->msg_target;
3039 LASSERT(!getmsg->msg_target_is_router);
3040 LASSERT(!getmsg->msg_routing);
3043 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
3044 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
3048 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
3051 LASSERT(getmd->md_refcount > 0);
3053 if (getmd->md_threshold == 0) {
3054 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
3055 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
3057 lnet_res_unlock(cpt);
3061 LASSERT(getmd->md_offset == 0);
3063 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
3064 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
3066 /* setup information for lnet_build_msg_event */
3067 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
3068 msg->msg_from = peer_id.nid;
3069 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
3070 msg->msg_hdr.src_nid = peer_id.nid;
3071 msg->msg_hdr.payload_length = getmd->md_length;
3072 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
3074 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
3075 lnet_res_unlock(cpt);
3077 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
3080 lnet_msg_commit(msg, cpt);
3081 lnet_net_unlock(cpt);
3083 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3088 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
3091 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
3092 the_lnet.ln_counters[cpt]->drop_count++;
3093 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
3094 lnet_net_unlock(cpt);
3101 EXPORT_SYMBOL(lnet_create_reply_msg);
3104 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
3107 /* Set the REPLY length, now the RDMA that elides the REPLY message has
3108 * completed and I know it. */
3109 LASSERT(reply != NULL);
3110 LASSERT(reply->msg_type == LNET_MSG_GET);
3111 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
3113 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
3114 * the end of my buffer, I might as well be dead. */
3115 LASSERT(len <= reply->msg_ev.mlength);
3117 reply->msg_ev.mlength = len;
3119 EXPORT_SYMBOL(lnet_set_reply_msg_len);
3122 * Initiate an asynchronous GET operation.
3124 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
3125 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
3126 * the target node in the REPLY has been written to local MD.
3128 * On the target node, an LNET_EVENT_GET is logged when the GET request
3129 * arrives and is accepted into a MD.
3131 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
3132 * \param mdh A handle for the MD that describes the memory into which the
3133 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
3135 * \retval 0 Success, and only in this case events will be generated
3136 * and logged to EQ (if it exists) of the MD.
3137 * \retval -EIO Simulated failure.
3138 * \retval -ENOMEM Memory allocation failure.
3139 * \retval -ENOENT Invalid MD object.
3142 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
3143 struct lnet_process_id target, unsigned int portal,
3144 __u64 match_bits, unsigned int offset)
3146 struct lnet_msg *msg;
3147 struct lnet_libmd *md;
3151 LASSERT(the_lnet.ln_refcount > 0);
3153 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
3154 fail_peer(target.nid, 1)) /* shall we now? */
3156 CERROR("Dropping GET to %s: simulated failure\n",
3157 libcfs_id2str(target));
3161 msg = lnet_msg_alloc();
3163 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
3164 libcfs_id2str(target));
3168 cpt = lnet_cpt_of_cookie(mdh.cookie);
3171 md = lnet_handle2md(&mdh);
3172 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3173 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
3174 match_bits, portal, libcfs_id2str(target),
3175 md == NULL ? -1 : md->md_threshold);
3176 if (md != NULL && md->md_me != NULL)
3177 CERROR("REPLY MD also attached to portal %d\n",
3178 md->md_me->me_portal);
3180 lnet_res_unlock(cpt);
3186 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
3188 lnet_msg_attach_md(msg, md, 0, 0);
3190 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
3192 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
3193 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
3194 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
3195 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
3197 /* NB handles only looked up by creator (no flips) */
3198 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
3199 the_lnet.ln_interface_cookie;
3200 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
3201 md->md_lh.lh_cookie;
3203 lnet_res_unlock(cpt);
3205 lnet_build_msg_event(msg, LNET_EVENT_SEND);
3207 rc = lnet_send(self, msg, LNET_NID_ANY);
3209 CNETERR("Error sending GET to %s: %d\n",
3210 libcfs_id2str(target), rc);
3211 lnet_finalize(msg, rc);
3214 /* completion will be signalled by an event */
3217 EXPORT_SYMBOL(LNetGet);
3220 * Calculate distance to node at \a dstnid.
3222 * \param dstnid Target NID.
3223 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
3225 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
3228 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
3229 * local_nid_dist_zero is set, which is the default.
3230 * \retval positives Distance to target NID, i.e. number of hops plus one.
3231 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
3234 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
3236 struct list_head *e;
3237 struct lnet_ni *ni = NULL;
3238 struct lnet_remotenet *rnet;
3239 __u32 dstnet = LNET_NIDNET(dstnid);
3243 struct list_head *rn_list;
3245 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
3246 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
3247 * keep order 0 free for 0@lo and order 1 free for a local NID
3250 LASSERT(the_lnet.ln_refcount > 0);
3252 cpt = lnet_net_lock_current();
3254 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
3255 if (ni->ni_nid == dstnid) {
3256 if (srcnidp != NULL)
3258 if (orderp != NULL) {
3259 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
3264 lnet_net_unlock(cpt);
3266 return local_nid_dist_zero ? 0 : 1;
3269 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
3270 /* Check if ni was originally created in
3271 * current net namespace.
3272 * If not, assign order above 0xffff0000,
3273 * to make this ni not a priority. */
3274 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
3275 order += 0xffff0000;
3277 if (srcnidp != NULL)
3278 *srcnidp = ni->ni_nid;
3281 lnet_net_unlock(cpt);
3288 rn_list = lnet_net2rnethash(dstnet);
3289 list_for_each(e, rn_list) {
3290 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
3292 if (rnet->lrn_net == dstnet) {
3293 struct lnet_route *route;
3294 struct lnet_route *shortest = NULL;
3295 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
3298 LASSERT(!list_empty(&rnet->lrn_routes));
3300 list_for_each_entry(route, &rnet->lrn_routes,
3302 route_hops = route->lr_hops;
3303 if (route_hops == LNET_UNDEFINED_HOPS)
3305 if (shortest == NULL ||
3306 route_hops < shortest_hops) {
3308 shortest_hops = route_hops;
3312 LASSERT(shortest != NULL);
3313 hops = shortest_hops;
3314 if (srcnidp != NULL) {
3315 ni = lnet_get_next_ni_locked(
3316 shortest->lr_gateway->lpni_net,
3318 *srcnidp = ni->ni_nid;
3322 lnet_net_unlock(cpt);
3328 lnet_net_unlock(cpt);
3329 return -EHOSTUNREACH;
3331 EXPORT_SYMBOL(LNetDist);