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 struct lnet_send_data {
50 struct lnet_ni *sd_best_ni;
51 struct lnet_peer_ni *sd_best_lpni;
52 struct lnet_peer_ni *sd_final_dst_lpni;
53 struct lnet_peer *sd_peer;
54 struct lnet_peer *sd_gw_peer;
55 struct lnet_peer_ni *sd_gw_lpni;
56 struct lnet_peer_net *sd_peer_net;
57 struct lnet_msg *sd_msg;
58 lnet_nid_t sd_dst_nid;
59 lnet_nid_t sd_src_nid;
60 lnet_nid_t sd_rtr_nid;
66 static inline struct lnet_comm_count *
67 get_stats_counts(struct lnet_element_stats *stats,
68 enum lnet_stats_type stats_type)
71 case LNET_STATS_TYPE_SEND:
72 return &stats->el_send_stats;
73 case LNET_STATS_TYPE_RECV:
74 return &stats->el_recv_stats;
75 case LNET_STATS_TYPE_DROP:
76 return &stats->el_drop_stats;
78 CERROR("Unknown stats type\n");
84 void lnet_incr_stats(struct lnet_element_stats *stats,
85 enum lnet_msg_type msg_type,
86 enum lnet_stats_type stats_type)
88 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
94 atomic_inc(&counts->co_ack_count);
97 atomic_inc(&counts->co_put_count);
100 atomic_inc(&counts->co_get_count);
103 atomic_inc(&counts->co_reply_count);
106 atomic_inc(&counts->co_hello_count);
109 CERROR("There is a BUG in the code. Unknown message type\n");
114 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
115 enum lnet_stats_type stats_type)
117 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
121 return (atomic_read(&counts->co_ack_count) +
122 atomic_read(&counts->co_put_count) +
123 atomic_read(&counts->co_get_count) +
124 atomic_read(&counts->co_reply_count) +
125 atomic_read(&counts->co_hello_count));
128 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
129 struct lnet_comm_count *counts)
131 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
132 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
133 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
134 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
135 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
138 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
139 struct lnet_element_stats *stats)
141 struct lnet_comm_count *counts;
146 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
149 assign_stats(&msg_stats->im_send_stats, counts);
151 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
154 assign_stats(&msg_stats->im_recv_stats, counts);
156 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
159 assign_stats(&msg_stats->im_drop_stats, counts);
163 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
165 struct lnet_test_peer *tp;
166 struct list_head *el;
167 struct list_head *next;
168 struct list_head cull;
170 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
171 if (threshold != 0) {
172 /* Adding a new entry */
173 LIBCFS_ALLOC(tp, sizeof(*tp));
178 tp->tp_threshold = threshold;
181 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
186 /* removing entries */
187 INIT_LIST_HEAD(&cull);
191 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
192 tp = list_entry(el, struct lnet_test_peer, tp_list);
194 if (tp->tp_threshold == 0 || /* needs culling anyway */
195 nid == LNET_NID_ANY || /* removing all entries */
196 tp->tp_nid == nid) { /* matched this one */
197 list_del(&tp->tp_list);
198 list_add(&tp->tp_list, &cull);
204 while (!list_empty(&cull)) {
205 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
207 list_del(&tp->tp_list);
208 LIBCFS_FREE(tp, sizeof(*tp));
214 fail_peer (lnet_nid_t nid, int outgoing)
216 struct lnet_test_peer *tp;
217 struct list_head *el;
218 struct list_head *next;
219 struct list_head cull;
222 INIT_LIST_HEAD(&cull);
224 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
227 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
228 tp = list_entry(el, struct lnet_test_peer, tp_list);
230 if (tp->tp_threshold == 0) {
233 /* only cull zombies on outgoing tests,
234 * since we may be at interrupt priority on
235 * incoming messages. */
236 list_del(&tp->tp_list);
237 list_add(&tp->tp_list, &cull);
242 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
243 nid == tp->tp_nid) { /* fail this peer */
246 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
249 tp->tp_threshold == 0) {
251 list_del(&tp->tp_list);
252 list_add(&tp->tp_list, &cull);
261 while (!list_empty(&cull)) {
262 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
263 list_del(&tp->tp_list);
265 LIBCFS_FREE(tp, sizeof(*tp));
272 lnet_iov_nob(unsigned int niov, struct kvec *iov)
274 unsigned int nob = 0;
276 LASSERT(niov == 0 || iov != NULL);
278 nob += (iov++)->iov_len;
282 EXPORT_SYMBOL(lnet_iov_nob);
285 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
286 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
289 /* NB diov, siov are READ-ONLY */
290 unsigned int this_nob;
295 /* skip complete frags before 'doffset' */
297 while (doffset >= diov->iov_len) {
298 doffset -= diov->iov_len;
304 /* skip complete frags before 'soffset' */
306 while (soffset >= siov->iov_len) {
307 soffset -= siov->iov_len;
316 this_nob = MIN(diov->iov_len - doffset,
317 siov->iov_len - soffset);
318 this_nob = MIN(this_nob, nob);
320 memcpy((char *)diov->iov_base + doffset,
321 (char *)siov->iov_base + soffset, this_nob);
324 if (diov->iov_len > doffset + this_nob) {
332 if (siov->iov_len > soffset + this_nob) {
341 EXPORT_SYMBOL(lnet_copy_iov2iov);
344 lnet_extract_iov(int dst_niov, struct kvec *dst,
345 int src_niov, struct kvec *src,
346 unsigned int offset, unsigned int len)
348 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
349 * for exactly 'len' bytes, and return the number of entries.
350 * NB not destructive to 'src' */
351 unsigned int frag_len;
354 if (len == 0) /* no data => */
355 return (0); /* no frags */
357 LASSERT(src_niov > 0);
358 while (offset >= src->iov_len) { /* skip initial frags */
359 offset -= src->iov_len;
362 LASSERT(src_niov > 0);
367 LASSERT(src_niov > 0);
368 LASSERT((int)niov <= dst_niov);
370 frag_len = src->iov_len - offset;
371 dst->iov_base = ((char *)src->iov_base) + offset;
373 if (len <= frag_len) {
378 dst->iov_len = frag_len;
388 EXPORT_SYMBOL(lnet_extract_iov);
392 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
394 unsigned int nob = 0;
396 LASSERT(niov == 0 || kiov != NULL);
398 nob += (kiov++)->kiov_len;
402 EXPORT_SYMBOL(lnet_kiov_nob);
405 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
406 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
409 /* NB diov, siov are READ-ONLY */
410 unsigned int this_nob;
417 LASSERT (!in_interrupt ());
420 while (doffset >= diov->kiov_len) {
421 doffset -= diov->kiov_len;
428 while (soffset >= siov->kiov_len) {
429 soffset -= siov->kiov_len;
438 this_nob = MIN(diov->kiov_len - doffset,
439 siov->kiov_len - soffset);
440 this_nob = MIN(this_nob, nob);
443 daddr = ((char *)kmap(diov->kiov_page)) +
444 diov->kiov_offset + doffset;
446 saddr = ((char *)kmap(siov->kiov_page)) +
447 siov->kiov_offset + soffset;
449 /* Vanishing risk of kmap deadlock when mapping 2 pages.
450 * However in practice at least one of the kiovs will be mapped
451 * kernel pages and the map/unmap will be NOOPs */
453 memcpy (daddr, saddr, this_nob);
456 if (diov->kiov_len > doffset + this_nob) {
460 kunmap(diov->kiov_page);
467 if (siov->kiov_len > soffset + this_nob) {
471 kunmap(siov->kiov_page);
480 kunmap(diov->kiov_page);
482 kunmap(siov->kiov_page);
484 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
487 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
488 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
491 /* NB iov, kiov are READ-ONLY */
492 unsigned int this_nob;
498 LASSERT (!in_interrupt ());
501 while (iovoffset >= iov->iov_len) {
502 iovoffset -= iov->iov_len;
509 while (kiovoffset >= kiov->kiov_len) {
510 kiovoffset -= kiov->kiov_len;
519 this_nob = MIN(iov->iov_len - iovoffset,
520 kiov->kiov_len - kiovoffset);
521 this_nob = MIN(this_nob, nob);
524 addr = ((char *)kmap(kiov->kiov_page)) +
525 kiov->kiov_offset + kiovoffset;
527 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
530 if (iov->iov_len > iovoffset + this_nob) {
531 iovoffset += this_nob;
538 if (kiov->kiov_len > kiovoffset + this_nob) {
540 kiovoffset += this_nob;
542 kunmap(kiov->kiov_page);
552 kunmap(kiov->kiov_page);
554 EXPORT_SYMBOL(lnet_copy_kiov2iov);
557 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
558 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
561 /* NB kiov, iov are READ-ONLY */
562 unsigned int this_nob;
568 LASSERT (!in_interrupt ());
571 while (kiovoffset >= kiov->kiov_len) {
572 kiovoffset -= kiov->kiov_len;
579 while (iovoffset >= iov->iov_len) {
580 iovoffset -= iov->iov_len;
589 this_nob = MIN(kiov->kiov_len - kiovoffset,
590 iov->iov_len - iovoffset);
591 this_nob = MIN(this_nob, nob);
594 addr = ((char *)kmap(kiov->kiov_page)) +
595 kiov->kiov_offset + kiovoffset;
597 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
600 if (kiov->kiov_len > kiovoffset + this_nob) {
602 kiovoffset += this_nob;
604 kunmap(kiov->kiov_page);
611 if (iov->iov_len > iovoffset + this_nob) {
612 iovoffset += this_nob;
621 kunmap(kiov->kiov_page);
623 EXPORT_SYMBOL(lnet_copy_iov2kiov);
626 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
627 int src_niov, lnet_kiov_t *src,
628 unsigned int offset, unsigned int len)
630 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
631 * for exactly 'len' bytes, and return the number of entries.
632 * NB not destructive to 'src' */
633 unsigned int frag_len;
636 if (len == 0) /* no data => */
637 return (0); /* no frags */
639 LASSERT(src_niov > 0);
640 while (offset >= src->kiov_len) { /* skip initial frags */
641 offset -= src->kiov_len;
644 LASSERT(src_niov > 0);
649 LASSERT(src_niov > 0);
650 LASSERT((int)niov <= dst_niov);
652 frag_len = src->kiov_len - offset;
653 dst->kiov_page = src->kiov_page;
654 dst->kiov_offset = src->kiov_offset + offset;
656 if (len <= frag_len) {
658 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
662 dst->kiov_len = frag_len;
663 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
673 EXPORT_SYMBOL(lnet_extract_kiov);
676 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
677 int delayed, unsigned int offset, unsigned int mlen,
680 unsigned int niov = 0;
681 struct kvec *iov = NULL;
682 lnet_kiov_t *kiov = NULL;
685 LASSERT (!in_interrupt ());
686 LASSERT (mlen == 0 || msg != NULL);
689 LASSERT(msg->msg_receiving);
690 LASSERT(!msg->msg_sending);
691 LASSERT(rlen == msg->msg_len);
692 LASSERT(mlen <= msg->msg_len);
693 LASSERT(msg->msg_offset == offset);
694 LASSERT(msg->msg_wanted == mlen);
696 msg->msg_receiving = 0;
699 niov = msg->msg_niov;
701 kiov = msg->msg_kiov;
704 LASSERT ((iov == NULL) != (kiov == NULL));
708 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
709 niov, iov, kiov, offset, mlen,
712 lnet_finalize(msg, rc);
716 lnet_setpayloadbuffer(struct lnet_msg *msg)
718 struct lnet_libmd *md = msg->msg_md;
720 LASSERT(msg->msg_len > 0);
721 LASSERT(!msg->msg_routing);
723 LASSERT(msg->msg_niov == 0);
724 LASSERT(msg->msg_iov == NULL);
725 LASSERT(msg->msg_kiov == NULL);
727 msg->msg_niov = md->md_niov;
728 if ((md->md_options & LNET_MD_KIOV) != 0)
729 msg->msg_kiov = md->md_iov.kiov;
731 msg->msg_iov = md->md_iov.iov;
735 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
736 unsigned int offset, unsigned int len)
738 msg->msg_type = type;
739 msg->msg_target = target;
741 msg->msg_offset = offset;
744 lnet_setpayloadbuffer(msg);
746 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
747 msg->msg_hdr.type = cpu_to_le32(type);
748 /* dest_nid will be overwritten by lnet_select_pathway() */
749 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
750 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
751 /* src_nid will be set later */
752 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
753 msg->msg_hdr.payload_length = cpu_to_le32(len);
757 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
759 void *priv = msg->msg_private;
762 LASSERT (!in_interrupt ());
763 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
764 (msg->msg_txcredit && msg->msg_peertxcredit));
766 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
768 msg->msg_no_resend = true;
769 lnet_finalize(msg, rc);
774 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
778 LASSERT(!msg->msg_sending);
779 LASSERT(msg->msg_receiving);
780 LASSERT(!msg->msg_rx_ready_delay);
781 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
783 msg->msg_rx_ready_delay = 1;
784 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
787 CERROR("recv from %s / send to %s aborted: "
788 "eager_recv failed %d\n",
789 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
790 libcfs_id2str(msg->msg_target), rc);
791 LASSERT(rc < 0); /* required by my callers */
798 * This function can be called from two paths:
799 * 1. when sending a message
800 * 2. when decommiting a message (lnet_msg_decommit_tx())
801 * In both these cases the peer_ni should have it's reference count
802 * acquired by the caller and therefore it is safe to drop the spin
803 * lock before calling lnd_query()
806 lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
808 time64_t last_alive = 0;
809 int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
811 LASSERT(lnet_peer_aliveness_enabled(lp));
812 LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
814 lnet_net_unlock(cpt);
815 (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
818 lp->lpni_last_query = ktime_get_seconds();
820 if (last_alive != 0) /* NI has updated timestamp */
821 lp->lpni_last_alive = last_alive;
824 /* NB: always called with lnet_net_lock held */
826 lnet_peer_is_alive(struct lnet_peer_ni *lp, time64_t now)
831 LASSERT (lnet_peer_aliveness_enabled(lp));
834 * Trust lnet_notify() if it has more recent aliveness news, but
835 * ignore the initial assumed death (see lnet_peers_start_down()).
837 spin_lock(&lp->lpni_lock);
838 if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
839 lp->lpni_timestamp >= lp->lpni_last_alive) {
840 spin_unlock(&lp->lpni_lock);
844 deadline = lp->lpni_last_alive +
845 lp->lpni_net->net_tunables.lct_peer_timeout;
846 alive = deadline > now;
849 * Update obsolete lp_alive except for routers assumed to be dead
850 * initially, because router checker would update aliveness in this
851 * case, and moreover lpni_last_alive at peer creation is assumed.
853 if (alive && !lp->lpni_alive &&
854 !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
855 spin_unlock(&lp->lpni_lock);
856 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
858 spin_unlock(&lp->lpni_lock);
865 /* NB: returns 1 when alive, 0 when dead, negative when error;
866 * may drop the lnet_net_lock */
868 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
870 time64_t now = ktime_get_seconds();
872 if (!lnet_peer_aliveness_enabled(lp))
875 if (lnet_peer_is_alive(lp, now))
879 * Peer appears dead, but we should avoid frequent NI queries (at
880 * most once per lnet_queryinterval seconds).
882 if (lp->lpni_last_query != 0) {
883 static const int lnet_queryinterval = 1;
886 next_query = lp->lpni_last_query + lnet_queryinterval;
888 if (now < next_query) {
890 CWARN("Unexpected aliveness of peer %s: "
891 "%lld < %lld (%d/%d)\n",
892 libcfs_nid2str(lp->lpni_nid),
895 lp->lpni_net->net_tunables.lct_peer_timeout);
900 /* query NI for latest aliveness news */
901 lnet_ni_query_locked(ni, lp);
903 if (lnet_peer_is_alive(lp, now))
906 lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
911 * \param msg The message to be sent.
912 * \param do_send True if lnet_ni_send() should be called in this function.
913 * lnet_send() is going to lnet_net_unlock immediately after this, so
914 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
916 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
917 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
918 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
919 * \retval -ECANCELED If the MD of the message has been unlinked.
922 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
924 struct lnet_peer_ni *lp = msg->msg_txpeer;
925 struct lnet_ni *ni = msg->msg_txni;
926 int cpt = msg->msg_tx_cpt;
927 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
929 /* non-lnet_send() callers have checked before */
930 LASSERT(!do_send || msg->msg_tx_delayed);
931 LASSERT(!msg->msg_receiving);
932 LASSERT(msg->msg_tx_committed);
934 /* NB 'lp' is always the next hop */
935 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
936 lnet_peer_alive_locked(ni, lp) == 0) {
937 the_lnet.ln_counters[cpt]->drop_count++;
938 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
939 lnet_net_unlock(cpt);
941 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
943 LNET_STATS_TYPE_DROP);
945 lnet_incr_stats(&msg->msg_txni->ni_stats,
947 LNET_STATS_TYPE_DROP);
949 CNETERR("Dropping message for %s: peer not alive\n",
950 libcfs_id2str(msg->msg_target));
952 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_DROPPED;
953 lnet_finalize(msg, -EHOSTUNREACH);
957 return -EHOSTUNREACH;
960 if (msg->msg_md != NULL &&
961 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
962 lnet_net_unlock(cpt);
964 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
965 "called on the MD/ME.\n",
966 libcfs_id2str(msg->msg_target));
968 msg->msg_no_resend = true;
969 lnet_finalize(msg, -ECANCELED);
976 if (!msg->msg_peertxcredit) {
977 spin_lock(&lp->lpni_lock);
978 LASSERT((lp->lpni_txcredits < 0) ==
979 !list_empty(&lp->lpni_txq));
981 msg->msg_peertxcredit = 1;
982 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
983 lp->lpni_txcredits--;
985 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
986 lp->lpni_mintxcredits = lp->lpni_txcredits;
988 if (lp->lpni_txcredits < 0) {
989 msg->msg_tx_delayed = 1;
990 list_add_tail(&msg->msg_list, &lp->lpni_txq);
991 spin_unlock(&lp->lpni_lock);
992 return LNET_CREDIT_WAIT;
994 spin_unlock(&lp->lpni_lock);
997 if (!msg->msg_txcredit) {
998 LASSERT((tq->tq_credits < 0) ==
999 !list_empty(&tq->tq_delayed));
1001 msg->msg_txcredit = 1;
1003 atomic_dec(&ni->ni_tx_credits);
1005 if (tq->tq_credits < tq->tq_credits_min)
1006 tq->tq_credits_min = tq->tq_credits;
1008 if (tq->tq_credits < 0) {
1009 msg->msg_tx_delayed = 1;
1010 list_add_tail(&msg->msg_list, &tq->tq_delayed);
1011 return LNET_CREDIT_WAIT;
1015 /* unset the tx_delay flag as we're going to send it now */
1016 msg->msg_tx_delayed = 0;
1019 lnet_net_unlock(cpt);
1020 lnet_ni_send(ni, msg);
1023 return LNET_CREDIT_OK;
1027 static struct lnet_rtrbufpool *
1028 lnet_msg2bufpool(struct lnet_msg *msg)
1030 struct lnet_rtrbufpool *rbp;
1033 LASSERT(msg->msg_rx_committed);
1035 cpt = msg->msg_rx_cpt;
1036 rbp = &the_lnet.ln_rtrpools[cpt][0];
1038 LASSERT(msg->msg_len <= LNET_MTU);
1039 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
1041 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
1048 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
1050 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
1051 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
1052 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
1053 * received or OK to receive */
1054 struct lnet_peer_ni *lp = msg->msg_rxpeer;
1055 struct lnet_rtrbufpool *rbp;
1056 struct lnet_rtrbuf *rb;
1058 LASSERT (msg->msg_iov == NULL);
1059 LASSERT (msg->msg_kiov == NULL);
1060 LASSERT (msg->msg_niov == 0);
1061 LASSERT (msg->msg_routing);
1062 LASSERT (msg->msg_receiving);
1063 LASSERT (!msg->msg_sending);
1065 /* non-lnet_parse callers only receive delayed messages */
1066 LASSERT(!do_recv || msg->msg_rx_delayed);
1068 if (!msg->msg_peerrtrcredit) {
1069 spin_lock(&lp->lpni_lock);
1070 LASSERT((lp->lpni_rtrcredits < 0) ==
1071 !list_empty(&lp->lpni_rtrq));
1073 msg->msg_peerrtrcredit = 1;
1074 lp->lpni_rtrcredits--;
1075 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
1076 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
1078 if (lp->lpni_rtrcredits < 0) {
1079 /* must have checked eager_recv before here */
1080 LASSERT(msg->msg_rx_ready_delay);
1081 msg->msg_rx_delayed = 1;
1082 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
1083 spin_unlock(&lp->lpni_lock);
1084 return LNET_CREDIT_WAIT;
1086 spin_unlock(&lp->lpni_lock);
1089 rbp = lnet_msg2bufpool(msg);
1091 if (!msg->msg_rtrcredit) {
1092 msg->msg_rtrcredit = 1;
1094 if (rbp->rbp_credits < rbp->rbp_mincredits)
1095 rbp->rbp_mincredits = rbp->rbp_credits;
1097 if (rbp->rbp_credits < 0) {
1098 /* must have checked eager_recv before here */
1099 LASSERT(msg->msg_rx_ready_delay);
1100 msg->msg_rx_delayed = 1;
1101 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1102 return LNET_CREDIT_WAIT;
1106 LASSERT(!list_empty(&rbp->rbp_bufs));
1107 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1108 list_del(&rb->rb_list);
1110 msg->msg_niov = rbp->rbp_npages;
1111 msg->msg_kiov = &rb->rb_kiov[0];
1113 /* unset the msg-rx_delayed flag since we're receiving the message */
1114 msg->msg_rx_delayed = 0;
1117 int cpt = msg->msg_rx_cpt;
1119 lnet_net_unlock(cpt);
1120 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1121 0, msg->msg_len, msg->msg_len);
1124 return LNET_CREDIT_OK;
1128 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1130 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1131 struct lnet_ni *txni = msg->msg_txni;
1132 struct lnet_msg *msg2;
1134 if (msg->msg_txcredit) {
1135 struct lnet_ni *ni = msg->msg_txni;
1136 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1138 /* give back NI txcredits */
1139 msg->msg_txcredit = 0;
1141 LASSERT((tq->tq_credits < 0) ==
1142 !list_empty(&tq->tq_delayed));
1145 atomic_inc(&ni->ni_tx_credits);
1146 if (tq->tq_credits <= 0) {
1147 msg2 = list_entry(tq->tq_delayed.next,
1148 struct lnet_msg, msg_list);
1149 list_del(&msg2->msg_list);
1151 LASSERT(msg2->msg_txni == ni);
1152 LASSERT(msg2->msg_tx_delayed);
1153 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1155 (void) lnet_post_send_locked(msg2, 1);
1159 if (msg->msg_peertxcredit) {
1160 /* give back peer txcredits */
1161 msg->msg_peertxcredit = 0;
1163 spin_lock(&txpeer->lpni_lock);
1164 LASSERT((txpeer->lpni_txcredits < 0) ==
1165 !list_empty(&txpeer->lpni_txq));
1167 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1168 LASSERT(txpeer->lpni_txqnob >= 0);
1170 txpeer->lpni_txcredits++;
1171 if (txpeer->lpni_txcredits <= 0) {
1174 msg2 = list_entry(txpeer->lpni_txq.next,
1175 struct lnet_msg, msg_list);
1176 list_del(&msg2->msg_list);
1177 spin_unlock(&txpeer->lpni_lock);
1179 LASSERT(msg2->msg_txpeer == txpeer);
1180 LASSERT(msg2->msg_tx_delayed);
1182 msg2_cpt = msg2->msg_tx_cpt;
1185 * The msg_cpt can be different from the msg2_cpt
1186 * so we need to make sure we lock the correct cpt
1188 * Once we call lnet_post_send_locked() it is no
1189 * longer safe to access msg2, since it could've
1190 * been freed by lnet_finalize(), but we still
1191 * need to relock the correct cpt, so we cache the
1192 * msg2_cpt for the purpose of the check that
1193 * follows the call to lnet_pose_send_locked().
1195 if (msg2_cpt != msg->msg_tx_cpt) {
1196 lnet_net_unlock(msg->msg_tx_cpt);
1197 lnet_net_lock(msg2_cpt);
1199 (void) lnet_post_send_locked(msg2, 1);
1200 if (msg2_cpt != msg->msg_tx_cpt) {
1201 lnet_net_unlock(msg2_cpt);
1202 lnet_net_lock(msg->msg_tx_cpt);
1205 spin_unlock(&txpeer->lpni_lock);
1210 msg->msg_txni = NULL;
1211 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1214 if (txpeer != NULL) {
1215 msg->msg_txpeer = NULL;
1216 lnet_peer_ni_decref_locked(txpeer);
1221 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1223 struct lnet_msg *msg;
1225 if (list_empty(&rbp->rbp_msgs))
1227 msg = list_entry(rbp->rbp_msgs.next,
1228 struct lnet_msg, msg_list);
1229 list_del(&msg->msg_list);
1231 (void)lnet_post_routed_recv_locked(msg, 1);
1235 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1237 struct lnet_msg *msg;
1238 struct lnet_msg *tmp;
1240 lnet_net_unlock(cpt);
1242 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1243 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1244 0, 0, 0, msg->msg_hdr.payload_length);
1245 list_del_init(&msg->msg_list);
1246 msg->msg_no_resend = true;
1247 lnet_finalize(msg, -ECANCELED);
1254 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1256 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1257 struct lnet_ni *rxni = msg->msg_rxni;
1258 struct lnet_msg *msg2;
1260 if (msg->msg_rtrcredit) {
1261 /* give back global router credits */
1262 struct lnet_rtrbuf *rb;
1263 struct lnet_rtrbufpool *rbp;
1265 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1266 * there until it gets one allocated, or aborts the wait
1268 LASSERT(msg->msg_kiov != NULL);
1270 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1273 msg->msg_kiov = NULL;
1274 msg->msg_rtrcredit = 0;
1276 LASSERT(rbp == lnet_msg2bufpool(msg));
1278 LASSERT((rbp->rbp_credits > 0) ==
1279 !list_empty(&rbp->rbp_bufs));
1281 /* If routing is now turned off, we just drop this buffer and
1282 * don't bother trying to return credits. */
1283 if (!the_lnet.ln_routing) {
1284 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1288 /* It is possible that a user has lowered the desired number of
1289 * buffers in this pool. Make sure we never put back
1290 * more buffers than the stated number. */
1291 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1292 /* Discard this buffer so we don't have too
1294 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1295 rbp->rbp_nbuffers--;
1297 list_add(&rb->rb_list, &rbp->rbp_bufs);
1299 if (rbp->rbp_credits <= 0)
1300 lnet_schedule_blocked_locked(rbp);
1305 if (msg->msg_peerrtrcredit) {
1306 /* give back peer router credits */
1307 msg->msg_peerrtrcredit = 0;
1309 spin_lock(&rxpeer->lpni_lock);
1310 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1311 !list_empty(&rxpeer->lpni_rtrq));
1313 rxpeer->lpni_rtrcredits++;
1315 /* drop all messages which are queued to be routed on that
1317 if (!the_lnet.ln_routing) {
1318 struct list_head drop;
1319 INIT_LIST_HEAD(&drop);
1320 list_splice_init(&rxpeer->lpni_rtrq, &drop);
1321 spin_unlock(&rxpeer->lpni_lock);
1322 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1323 } else if (rxpeer->lpni_rtrcredits <= 0) {
1324 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1325 struct lnet_msg, msg_list);
1326 list_del(&msg2->msg_list);
1327 spin_unlock(&rxpeer->lpni_lock);
1328 (void) lnet_post_routed_recv_locked(msg2, 1);
1330 spin_unlock(&rxpeer->lpni_lock);
1334 msg->msg_rxni = NULL;
1335 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1337 if (rxpeer != NULL) {
1338 msg->msg_rxpeer = NULL;
1339 lnet_peer_ni_decref_locked(rxpeer);
1344 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1346 if (p1->lpni_txqnob < p2->lpni_txqnob)
1349 if (p1->lpni_txqnob > p2->lpni_txqnob)
1352 if (p1->lpni_txcredits > p2->lpni_txcredits)
1355 if (p1->lpni_txcredits < p2->lpni_txcredits)
1362 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1364 struct lnet_peer_ni *p1 = r1->lr_gateway;
1365 struct lnet_peer_ni *p2 = r2->lr_gateway;
1366 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1367 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1370 if (r1->lr_priority < r2->lr_priority)
1373 if (r1->lr_priority > r2->lr_priority)
1376 if (r1_hops < r2_hops)
1379 if (r1_hops > r2_hops)
1382 rc = lnet_compare_peers(p1, p2);
1386 if (r1->lr_seq - r2->lr_seq <= 0)
1392 static struct lnet_peer_ni *
1393 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1396 struct lnet_remotenet *rnet;
1397 struct lnet_route *route;
1398 struct lnet_route *best_route;
1399 struct lnet_route *last_route;
1400 struct lnet_peer_ni *lpni_best;
1401 struct lnet_peer_ni *lp;
1404 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1405 * rtr_nid nid, otherwise find the best gateway I can use */
1407 rnet = lnet_find_rnet_locked(remote_net);
1412 best_route = last_route = NULL;
1413 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1414 lp = route->lr_gateway;
1416 if (!lnet_is_route_alive(route))
1419 if (net != NULL && lp->lpni_net != net)
1422 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1425 if (lpni_best == NULL) {
1426 best_route = last_route = route;
1431 /* no protection on below fields, but it's harmless */
1432 if (last_route->lr_seq - route->lr_seq < 0)
1435 rc = lnet_compare_routes(route, best_route);
1443 /* set sequence number on the best router to the latest sequence + 1
1444 * so we can round-robin all routers, it's race and inaccurate but
1445 * harmless and functional */
1446 if (best_route != NULL)
1447 best_route->lr_seq = last_route->lr_seq + 1;
1451 static struct lnet_ni *
1452 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1453 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1456 struct lnet_ni *ni = NULL;
1457 unsigned int shortest_distance;
1462 * If there is no peer_ni that we can send to on this network,
1463 * then there is no point in looking for a new best_ni here.
1465 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1468 if (best_ni == NULL) {
1469 shortest_distance = UINT_MAX;
1470 best_credits = INT_MIN;
1473 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1474 best_ni->ni_dev_cpt);
1475 best_credits = atomic_read(&best_ni->ni_tx_credits);
1476 best_healthv = atomic_read(&best_ni->ni_healthv);
1479 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1480 unsigned int distance;
1485 ni_credits = atomic_read(&ni->ni_tx_credits);
1486 ni_healthv = atomic_read(&ni->ni_healthv);
1487 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1490 * calculate the distance from the CPT on which
1491 * the message memory is allocated to the CPT of
1492 * the NI's physical device
1494 distance = cfs_cpt_distance(lnet_cpt_table(),
1498 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1499 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1500 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1501 : "not seleced", best_credits, shortest_distance,
1502 (best_ni) ? best_ni->ni_seq : 0);
1505 * All distances smaller than the NUMA range
1506 * are treated equally.
1508 if (distance < lnet_numa_range)
1509 distance = lnet_numa_range;
1512 * Select on health, shorter distance, available
1513 * credits, then round-robin.
1517 } else if (ni_healthv < best_healthv) {
1519 } else if (ni_healthv > best_healthv) {
1520 best_healthv = ni_healthv;
1522 * If we're going to prefer this ni because it's
1523 * the healthiest, then we should set the
1524 * shortest_distance in the algorithm in case
1525 * there are multiple NIs with the same health but
1526 * different distances.
1528 if (distance < shortest_distance)
1529 shortest_distance = distance;
1530 } else if (distance > shortest_distance) {
1532 } else if (distance < shortest_distance) {
1533 shortest_distance = distance;
1534 } else if (ni_credits < best_credits) {
1536 } else if (ni_credits == best_credits) {
1537 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1541 best_credits = ni_credits;
1544 CDEBUG(D_NET, "selected best_ni %s\n",
1545 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1551 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1552 * because such traffic is required to perform discovery. We therefore
1553 * exclude all GET and PUT on that portal. We also exclude all ACK and
1554 * REPLY traffic, but that is because the portal is not tracked in the
1555 * message structure for these message types. We could restrict this
1556 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1559 lnet_msg_discovery(struct lnet_msg *msg)
1561 if (msg->msg_type == LNET_MSG_PUT) {
1562 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1564 } else if (msg->msg_type == LNET_MSG_GET) {
1565 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1571 #define SRC_SPEC 0x0001
1572 #define SRC_ANY 0x0002
1573 #define LOCAL_DST 0x0004
1574 #define REMOTE_DST 0x0008
1575 #define MR_DST 0x0010
1576 #define NMR_DST 0x0020
1577 #define SND_RESP 0x0040
1579 /* The following to defines are used for return codes */
1580 #define REPEAT_SEND 0x1000
1581 #define PASS_THROUGH 0x2000
1583 /* The different cases lnet_select pathway needs to handle */
1584 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1585 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1586 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1587 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1588 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1589 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1590 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1591 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1594 lnet_handle_send(struct lnet_send_data *sd)
1596 struct lnet_ni *best_ni = sd->sd_best_ni;
1597 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1598 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1599 struct lnet_msg *msg = sd->sd_msg;
1601 __u32 send_case = sd->sd_send_case;
1603 __u32 routing = send_case & REMOTE_DST;
1606 * Increment sequence number of the selected peer so that we
1607 * pick the next one in Round Robin.
1609 best_lpni->lpni_seq++;
1612 * grab a reference on the peer_ni so it sticks around even if
1613 * we need to drop and relock the lnet_net_lock below.
1615 lnet_peer_ni_addref_locked(best_lpni);
1618 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1619 * message. This ensures that we get a CPT that is correct for
1620 * the NI when the NI has been restricted to a subset of all CPTs.
1621 * If the selected CPT differs from the one currently locked, we
1622 * must unlock and relock the lnet_net_lock(), and then check whether
1623 * the configuration has changed. We don't have a hold on the best_ni
1624 * yet, and it may have vanished.
1626 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1627 if (sd->sd_cpt != cpt2) {
1628 __u32 seq = lnet_get_dlc_seq_locked();
1629 lnet_net_unlock(sd->sd_cpt);
1631 lnet_net_lock(sd->sd_cpt);
1632 if (seq != lnet_get_dlc_seq_locked()) {
1633 lnet_peer_ni_decref_locked(best_lpni);
1639 * store the best_lpni in the message right away to avoid having
1640 * to do the same operation under different conditions
1642 msg->msg_txpeer = best_lpni;
1643 msg->msg_txni = best_ni;
1646 * grab a reference for the best_ni since now it's in use in this
1647 * send. The reference will be dropped in lnet_finalize()
1649 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1652 * Always set the target.nid to the best peer picked. Either the
1653 * NID will be one of the peer NIDs selected, or the same NID as
1654 * what was originally set in the target or it will be the NID of
1655 * a router if this message should be routed
1657 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1660 * lnet_msg_commit assigns the correct cpt to the message, which
1661 * is used to decrement the correct refcount on the ni when it's
1662 * time to return the credits
1664 lnet_msg_commit(msg, sd->sd_cpt);
1667 * If we are routing the message then we keep the src_nid that was
1668 * set by the originator. If we are not routing then we are the
1669 * originator and set it here.
1671 if (!msg->msg_routing)
1672 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1675 msg->msg_target_is_router = 1;
1676 msg->msg_target.pid = LNET_PID_LUSTRE;
1678 * since we're routing we want to ensure that the
1679 * msg_hdr.dest_nid is set to the final destination. When
1680 * the router receives this message it knows how to route
1683 * final_dst_lpni is set at the beginning of the
1684 * lnet_select_pathway() function and is never changed.
1685 * It's safe to use it here.
1687 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1690 * if we're not routing set the dest_nid to the best peer
1691 * ni NID that we picked earlier in the algorithm.
1693 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1696 rc = lnet_post_send_locked(msg, 0);
1699 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s\n",
1700 libcfs_nid2str(msg->msg_hdr.src_nid),
1701 libcfs_nid2str(msg->msg_txni->ni_nid),
1702 libcfs_nid2str(sd->sd_src_nid),
1703 libcfs_nid2str(msg->msg_hdr.dest_nid),
1704 libcfs_nid2str(sd->sd_dst_nid),
1705 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1706 lnet_msgtyp2str(msg->msg_type));
1711 static struct lnet_peer_ni *
1712 lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer,
1713 struct lnet_peer_net *peer_net)
1716 * Look at the peer NIs for the destination peer that connect
1717 * to the chosen net. If a peer_ni is preferred when using the
1718 * best_ni to communicate, we use that one. If there is no
1719 * preferred peer_ni, or there are multiple preferred peer_ni,
1720 * the available transmit credits are used. If the transmit
1721 * credits are equal, we round-robin over the peer_ni.
1723 struct lnet_peer_ni *lpni = NULL;
1724 struct lnet_peer_ni *best_lpni = NULL;
1725 struct lnet_ni *best_ni = sd->sd_best_ni;
1726 lnet_nid_t dst_nid = sd->sd_dst_nid;
1727 int best_lpni_credits = INT_MIN;
1728 bool preferred = false;
1730 int best_lpni_healthv = 0;
1733 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1735 * if the best_ni we've chosen aleady has this lpni
1736 * preferred, then let's use it
1738 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1741 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1743 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1744 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1747 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1748 libcfs_nid2str(lpni->lpni_nid),
1749 lpni->lpni_txcredits, best_lpni_credits,
1750 lpni->lpni_seq, best_lpni->lpni_seq);
1752 /* pick the healthiest peer ni */
1753 if (lpni_healthv < best_lpni_healthv) {
1755 } else if (lpni_healthv > best_lpni_healthv) {
1756 best_lpni_healthv = lpni_healthv;
1757 /* if this is a preferred peer use it */
1758 } else if (!preferred && ni_is_pref) {
1760 } else if (preferred && !ni_is_pref) {
1762 * this is not the preferred peer so let's ignore
1766 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1768 * We already have a peer that has more credits
1769 * available than this one. No need to consider
1770 * this peer further.
1773 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1775 * The best peer found so far and the current peer
1776 * have the same number of available credits let's
1777 * make sure to select between them using Round
1781 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1787 best_lpni_credits = lpni->lpni_txcredits;
1790 /* if we still can't find a peer ni then we can't reach it */
1792 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1793 LNET_NIDNET(dst_nid);
1794 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1795 libcfs_net2str(net_id));
1799 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1800 libcfs_nid2str(best_lpni->lpni_nid));
1806 * Prerequisite: the best_ni should already be set in the sd
1808 static inline struct lnet_peer_ni *
1809 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1812 struct lnet_peer_net *peer_net;
1815 * The gateway is Multi-Rail capable so now we must select the
1818 peer_net = lnet_peer_get_net_locked(peer, net_id);
1821 CERROR("gateway peer %s has no NI on net %s\n",
1822 libcfs_nid2str(peer->lp_primary_nid),
1823 libcfs_net2str(net_id));
1827 return lnet_select_peer_ni(sd, peer, peer_net);
1831 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1833 if (sd->sd_send_case & NMR_DST &&
1834 sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1835 sd->sd_msg->msg_type != LNET_MSG_ACK &&
1836 sd->sd_best_lpni->lpni_pref_nnids == 0) {
1837 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1838 libcfs_nid2str(sd->sd_best_ni->ni_nid),
1839 libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1840 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1841 sd->sd_best_ni->ni_nid);
1850 * use the source and destination NIDs as the pathway
1853 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1855 /* the destination lpni is set before we get here. */
1858 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1859 if (!sd->sd_best_ni) {
1860 CERROR("Can't send to %s: src %s is not a "
1861 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1862 libcfs_nid2str(sd->sd_src_nid));
1867 * the preferred NID will only be set for NMR peers
1869 lnet_set_non_mr_pref_nid(sd);
1871 return lnet_handle_send(sd);
1879 * Run the selection algorithm on the peer NIs unless we're sending
1880 * a response, in this case just send to the destination
1883 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1885 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1886 if (!sd->sd_best_ni) {
1887 CERROR("Can't send to %s: src %s is not a "
1888 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1889 libcfs_nid2str(sd->sd_src_nid));
1894 * only run the selection algorithm to pick the peer_ni if we're
1895 * sending a GET or a PUT. Responses are sent to the same
1896 * destination NID provided.
1898 if (!(sd->sd_send_case & SND_RESP)) {
1900 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1901 sd->sd_best_ni->ni_net->net_id);
1904 if (sd->sd_best_lpni)
1905 return lnet_handle_send(sd);
1907 CERROR("can't send to %s. no NI on %s\n",
1908 libcfs_nid2str(sd->sd_dst_nid),
1909 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1911 return -EHOSTUNREACH;
1915 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1916 struct lnet_peer *peer,
1917 struct lnet_peer_net *peer_net,
1921 struct lnet_net *local_net;
1922 struct lnet_ni *best_ni;
1924 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1929 * Iterate through the NIs in this local Net and select
1930 * the NI to send from. The selection is determined by
1931 * these 3 criterion in the following priority:
1933 * 2. NI available credits
1936 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1937 peer, peer_net, cpt);
1939 if (incr_seq && best_ni)
1946 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1948 struct lnet_peer_ni **gw_lpni,
1949 struct lnet_peer **gw_peer)
1951 struct lnet_peer_ni *gw;
1952 lnet_nid_t src_nid = sd->sd_src_nid;
1954 gw = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
1957 CERROR("no route to %s from %s\n",
1958 libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1959 return -EHOSTUNREACH;
1962 /* get the peer of the gw_ni */
1963 LASSERT(gw->lpni_peer_net);
1964 LASSERT(gw->lpni_peer_net->lpn_peer);
1966 *gw_peer = gw->lpni_peer_net->lpn_peer;
1968 if (!sd->sd_best_ni)
1969 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, *gw_peer,
1974 if (!sd->sd_best_ni) {
1975 CERROR("Internal Error. Expected local ni on %s "
1976 "but non found :%s\n",
1977 libcfs_net2str(gw->lpni_peer_net->lpn_net_id),
1978 libcfs_nid2str(sd->sd_src_nid));
1983 * if gw is MR let's find its best peer_ni
1985 if (lnet_peer_is_multi_rail(*gw_peer)) {
1986 gw = lnet_find_best_lpni_on_net(sd, *gw_peer,
1987 sd->sd_best_ni->ni_net->net_id);
1989 * We've already verified that the gw has an NI on that
1990 * desired net, but we're not finding it. Something is
1994 CERROR("Internal Error. Route expected to %s from %s\n",
1995 libcfs_nid2str(dst_nid),
1996 libcfs_nid2str(src_nid));
2011 * Remote destination
2012 * Non-MR destination
2016 * Remote destination
2019 * The handling of these two cases is similar. Even though the destination
2020 * can be MR or non-MR, we'll deal directly with the router.
2023 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2026 struct lnet_peer_ni *gw_lpni = NULL;
2027 struct lnet_peer *gw_peer = NULL;
2030 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2031 if (!sd->sd_best_ni) {
2032 CERROR("Can't send to %s: src %s is not a "
2033 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2034 libcfs_nid2str(sd->sd_src_nid));
2038 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2043 if (sd->sd_send_case & NMR_DST)
2045 * since the final destination is non-MR let's set its preferred
2046 * NID before we send
2048 lnet_set_non_mr_pref_nid(sd);
2051 * We're going to send to the gw found so let's set its
2054 sd->sd_peer = gw_peer;
2055 sd->sd_best_lpni = gw_lpni;
2057 return lnet_handle_send(sd);
2061 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
2063 struct lnet_peer_net *peer_net = NULL;
2064 struct lnet_ni *best_ni = NULL;
2067 * The peer can have multiple interfaces, some of them can be on
2068 * the local network and others on a routed network. We should
2069 * prefer the local network. However if the local network is not
2070 * available then we need to try the routed network
2073 /* go through all the peer nets and find the best_ni */
2074 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2076 * The peer's list of nets can contain non-local nets. We
2077 * want to only examine the local ones.
2079 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2081 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2082 peer_net, md_cpt, false);
2086 /* increment sequence number so we can round robin */
2092 static struct lnet_ni *
2093 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2095 struct lnet_ni *best_ni = NULL;
2096 struct lnet_peer_net *peer_net;
2097 struct lnet_peer *peer = sd->sd_peer;
2098 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2099 struct lnet_peer_ni *lpni;
2100 int cpt = sd->sd_cpt;
2103 * We must use a consistent source address when sending to a
2104 * non-MR peer. However, a non-MR peer can have multiple NIDs
2105 * on multiple networks, and we may even need to talk to this
2106 * peer on multiple networks -- certain types of
2107 * load-balancing configuration do this.
2109 * So we need to pick the NI the peer prefers for this
2110 * particular network.
2113 /* Get the target peer_ni */
2114 peer_net = lnet_peer_get_net_locked(peer,
2115 LNET_NIDNET(best_lpni->lpni_nid));
2116 LASSERT(peer_net != NULL);
2117 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2119 if (lpni->lpni_pref_nnids == 0)
2121 LASSERT(lpni->lpni_pref_nnids == 1);
2122 best_ni = lnet_nid2ni_locked(
2123 lpni->lpni_pref.nid, cpt);
2130 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2132 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2134 struct lnet_ni *best_ni = NULL;
2135 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2138 * We must use a consistent source address when sending to a
2139 * non-MR peer. However, a non-MR peer can have multiple NIDs
2140 * on multiple networks, and we may even need to talk to this
2141 * peer on multiple networks -- certain types of
2142 * load-balancing configuration do this.
2144 * So we need to pick the NI the peer prefers for this
2145 * particular network.
2148 best_ni = lnet_find_existing_preferred_best_ni(sd);
2150 /* if best_ni is still not set just pick one */
2153 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2154 sd->sd_best_lpni->lpni_peer_net,
2155 sd->sd_md_cpt, true);
2156 /* If there is no best_ni we don't have a route */
2158 CERROR("no path to %s from net %s\n",
2159 libcfs_nid2str(best_lpni->lpni_nid),
2160 libcfs_net2str(best_lpni->lpni_net->net_id));
2161 return -EHOSTUNREACH;
2165 sd->sd_best_ni = best_ni;
2167 /* Set preferred NI if necessary. */
2168 lnet_set_non_mr_pref_nid(sd);
2175 * Source not specified
2179 * always use the same source NID for NMR peers
2180 * If we've talked to that peer before then we already have a preferred
2181 * source NI associated with it. Otherwise, we select a preferred local NI
2182 * and store it in the peer
2185 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2189 /* sd->sd_best_lpni is already set to the final destination */
2192 * At this point we should've created the peer ni and peer. If we
2193 * can't find it, then something went wrong. Instead of assert
2194 * output a relevant message and fail the send
2196 if (!sd->sd_best_lpni) {
2197 CERROR("Internal fault. Unable to send msg %s to %s. "
2199 lnet_msgtyp2str(sd->sd_msg->msg_type),
2200 libcfs_nid2str(sd->sd_dst_nid));
2204 rc = lnet_select_preferred_best_ni(sd);
2206 rc = lnet_handle_send(sd);
2212 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2215 * NOTE we've already handled the remote peer case. So we only
2216 * need to worry about the local case here.
2218 * if we're sending a response, ACK or reply, we need to send it
2219 * to the destination NID given to us. At this point we already
2220 * have the peer_ni we're suppose to send to, so just find the
2221 * best_ni on the peer net and use that. Since we're sending to an
2222 * MR peer then we can just run the selection algorithm on our
2223 * local NIs and pick the best one.
2225 if (sd->sd_send_case & SND_RESP) {
2227 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2228 sd->sd_best_lpni->lpni_peer_net,
2229 sd->sd_md_cpt, true);
2231 if (!sd->sd_best_ni) {
2233 * We're not going to deal with not able to send
2234 * a response to the provided final destination
2236 CERROR("Can't send response to %s. "
2237 "No local NI available\n",
2238 libcfs_nid2str(sd->sd_dst_nid));
2239 return -EHOSTUNREACH;
2242 return lnet_handle_send(sd);
2246 * If we get here that means we're sending a fresh request, PUT or
2247 * GET, so we need to run our standard selection algorithm.
2248 * First find the best local interface that's on any of the peer's
2251 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2253 if (sd->sd_best_ni) {
2255 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2256 sd->sd_best_ni->ni_net->net_id);
2259 * if we're successful in selecting a peer_ni on the local
2260 * network, then send to it. Otherwise fall through and
2261 * try and see if we can reach it over another routed
2264 if (sd->sd_best_lpni) {
2266 * in case we initially started with a routed
2267 * destination, let's reset to local
2269 sd->sd_send_case &= ~REMOTE_DST;
2270 sd->sd_send_case |= LOCAL_DST;
2271 return lnet_handle_send(sd);
2274 CERROR("Internal Error. Expected to have a best_lpni: "
2276 libcfs_nid2str(sd->sd_src_nid),
2277 libcfs_nid2str(sd->sd_dst_nid));
2283 * Peer doesn't have a local network. Let's see if there is
2284 * a remote network we can reach it on.
2286 return PASS_THROUGH;
2291 * Source NID not specified
2296 * Source NID not speified
2297 * Remote destination
2300 * In both of these cases if we're sending a response, ACK or REPLY, then
2301 * we need to send to the destination NID provided.
2303 * In the remote case let's deal with MR routers.
2308 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2311 struct lnet_peer *gw_peer = NULL;
2312 struct lnet_peer_ni *gw_lpni = NULL;
2315 * handle sending a response to a remote peer here so we don't
2316 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2318 if (sd->sd_send_case & REMOTE_DST &&
2319 sd->sd_send_case & SND_RESP) {
2320 struct lnet_peer_ni *gw;
2321 struct lnet_peer *gw_peer;
2323 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2326 CERROR("Can't send response to %s. "
2327 "No route available\n",
2328 libcfs_nid2str(sd->sd_dst_nid));
2329 return -EHOSTUNREACH;
2332 sd->sd_best_lpni = gw;
2333 sd->sd_peer = gw_peer;
2335 return lnet_handle_send(sd);
2339 * Even though the NID for the peer might not be on a local network,
2340 * since the peer is MR there could be other interfaces on the
2341 * local network. In that case we'd still like to prefer the local
2342 * network over the routed network. If we're unable to do that
2343 * then we select the best router among the different routed networks,
2344 * and if the router is MR then we can deal with it as such.
2346 rc = lnet_handle_any_mr_dsta(sd);
2347 if (rc != PASS_THROUGH)
2351 * TODO; One possible enhancement is to run the selection
2352 * algorithm on the peer. However for remote peers the credits are
2353 * not decremented, so we'll be basically going over the peer NIs
2354 * in round robin. An MR router will run the selection algorithm
2355 * on the next-hop interfaces.
2357 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2362 sd->sd_send_case &= ~LOCAL_DST;
2363 sd->sd_send_case |= REMOTE_DST;
2365 sd->sd_peer = gw_peer;
2366 sd->sd_best_lpni = gw_lpni;
2368 return lnet_handle_send(sd);
2372 * Source not specified
2373 * Remote destination
2376 * Must send to the specified peer NID using the same source NID that
2377 * we've used before. If it's the first time to talk to that peer then
2378 * find the source NI and assign it as preferred to that peer
2381 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2384 struct lnet_peer_ni *gw_lpni = NULL;
2385 struct lnet_peer *gw_peer = NULL;
2388 * Let's set if we have a preferred NI to talk to this NMR peer
2390 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2393 * find the router and that'll find the best NI if we didn't find
2396 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2402 * set the best_ni we've chosen as the preferred one for
2405 lnet_set_non_mr_pref_nid(sd);
2407 /* we'll be sending to the gw */
2408 sd->sd_best_lpni = gw_lpni;
2409 sd->sd_peer = gw_peer;
2411 return lnet_handle_send(sd);
2415 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2418 * turn off the SND_RESP bit.
2419 * It will be checked in the case handling
2421 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2423 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2424 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2425 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2426 (send_case & MR_DST) ? "MR: " : "NMR: ",
2427 libcfs_nid2str(sd->sd_dst_nid),
2428 (send_case & LOCAL_DST) ? "local" : "routed");
2430 switch (send_case) {
2432 * For all cases where the source is specified, we should always
2433 * use the destination NID, whether it's an MR destination or not,
2434 * since we're continuing a series of related messages for the
2437 case SRC_SPEC_LOCAL_NMR_DST:
2438 return lnet_handle_spec_local_nmr_dst(sd);
2439 case SRC_SPEC_LOCAL_MR_DST:
2440 return lnet_handle_spec_local_mr_dst(sd);
2441 case SRC_SPEC_ROUTER_NMR_DST:
2442 case SRC_SPEC_ROUTER_MR_DST:
2443 return lnet_handle_spec_router_dst(sd);
2444 case SRC_ANY_LOCAL_NMR_DST:
2445 return lnet_handle_any_local_nmr_dst(sd);
2446 case SRC_ANY_LOCAL_MR_DST:
2447 case SRC_ANY_ROUTER_MR_DST:
2448 return lnet_handle_any_mr_dst(sd);
2449 case SRC_ANY_ROUTER_NMR_DST:
2450 return lnet_handle_any_router_nmr_dst(sd);
2452 CERROR("Unknown send case\n");
2458 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2459 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2461 struct lnet_peer_ni *lpni;
2462 struct lnet_peer *peer;
2463 struct lnet_send_data send_data;
2466 __u32 send_case = 0;
2468 memset(&send_data, 0, sizeof(send_data));
2471 * get an initial CPT to use for locking. The idea here is not to
2472 * serialize the calls to select_pathway, so that as many
2473 * operations can run concurrently as possible. To do that we use
2474 * the CPT where this call is being executed. Later on when we
2475 * determine the CPT to use in lnet_message_commit, we switch the
2476 * lock and check if there was any configuration change. If none,
2477 * then we proceed, if there is, then we restart the operation.
2479 cpt = lnet_net_lock_current();
2481 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2482 if (md_cpt == CFS_CPT_ANY)
2488 * If we're being asked to send to the loopback interface, there
2489 * is no need to go through any selection. We can just shortcut
2490 * the entire process and send over lolnd
2492 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2493 /* No send credit hassles with LOLND */
2494 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
2495 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
2496 if (!msg->msg_routing)
2497 msg->msg_hdr.src_nid =
2498 cpu_to_le64(the_lnet.ln_loni->ni_nid);
2499 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
2500 lnet_msg_commit(msg, cpt);
2501 msg->msg_txni = the_lnet.ln_loni;
2502 lnet_net_unlock(cpt);
2504 return LNET_CREDIT_OK;
2508 * find an existing peer_ni, or create one and mark it as having been
2509 * created due to network traffic. This call will create the
2510 * peer->peer_net->peer_ni tree.
2512 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2514 lnet_net_unlock(cpt);
2515 return PTR_ERR(lpni);
2519 * Cache the original src_nid. If we need to resend the message
2520 * then we'll need to know whether the src_nid was originally
2521 * specified for this message. If it was originally specified,
2522 * then we need to keep using the same src_nid since it's
2523 * continuing the same sequence of messages.
2525 msg->msg_src_nid_param = src_nid;
2528 * Now that we have a peer_ni, check if we want to discover
2529 * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2530 * trigger discovery.
2532 peer = lpni->lpni_peer_net->lpn_peer;
2533 if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
2534 lnet_nid_t primary_nid;
2535 rc = lnet_discover_peer_locked(lpni, cpt, false);
2537 lnet_peer_ni_decref_locked(lpni);
2538 lnet_net_unlock(cpt);
2541 /* The peer may have changed. */
2542 peer = lpni->lpni_peer_net->lpn_peer;
2543 /* queue message and return */
2544 msg->msg_rtr_nid_param = rtr_nid;
2545 msg->msg_sending = 0;
2546 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2547 lnet_peer_ni_decref_locked(lpni);
2548 primary_nid = peer->lp_primary_nid;
2549 lnet_net_unlock(cpt);
2551 CDEBUG(D_NET, "%s pending discovery\n",
2552 libcfs_nid2str(primary_nid));
2554 return LNET_DC_WAIT;
2556 lnet_peer_ni_decref_locked(lpni);
2559 * Identify the different send cases
2561 if (src_nid == LNET_NID_ANY)
2562 send_case |= SRC_ANY;
2564 send_case |= SRC_SPEC;
2566 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2567 send_case |= LOCAL_DST;
2569 send_case |= REMOTE_DST;
2572 * if this is a non-MR peer or if we're recovering a peer ni then
2573 * let's consider this an NMR case so we can hit the destination
2576 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2577 send_case |= NMR_DST;
2579 send_case |= MR_DST;
2581 if (msg->msg_type == LNET_MSG_REPLY ||
2582 msg->msg_type == LNET_MSG_ACK)
2583 send_case |= SND_RESP;
2585 /* assign parameters to the send_data */
2586 send_data.sd_msg = msg;
2587 send_data.sd_rtr_nid = rtr_nid;
2588 send_data.sd_src_nid = src_nid;
2589 send_data.sd_dst_nid = dst_nid;
2590 send_data.sd_best_lpni = lpni;
2592 * keep a pointer to the final destination in case we're going to
2593 * route, so we'll need to access it later
2595 send_data.sd_final_dst_lpni = lpni;
2596 send_data.sd_peer = peer;
2597 send_data.sd_md_cpt = md_cpt;
2598 send_data.sd_cpt = cpt;
2599 send_data.sd_send_case = send_case;
2601 rc = lnet_handle_send_case_locked(&send_data);
2603 if (rc == REPEAT_SEND)
2606 lnet_net_unlock(send_data.sd_cpt);
2612 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2614 lnet_nid_t dst_nid = msg->msg_target.nid;
2618 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2619 * but we might want to use pre-determined router for ACK/REPLY
2622 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2623 LASSERT(msg->msg_txpeer == NULL);
2624 LASSERT(msg->msg_txni == NULL);
2625 LASSERT(!msg->msg_sending);
2626 LASSERT(!msg->msg_target_is_router);
2627 LASSERT(!msg->msg_receiving);
2629 msg->msg_sending = 1;
2631 LASSERT(!msg->msg_tx_committed);
2633 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2637 if (rc == LNET_CREDIT_OK)
2638 lnet_ni_send(msg->msg_txni, msg);
2640 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2644 enum lnet_mt_event_type {
2645 MT_TYPE_LOCAL_NI = 0,
2649 struct lnet_mt_event_info {
2650 enum lnet_mt_event_type mt_type;
2655 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2657 struct lnet_rsp_tracker *rspt;
2660 * msg has a refcount on the MD so the MD is not going away.
2661 * The rspt queue for the cpt is protected by
2662 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2665 if (!md->md_rspt_ptr) {
2666 lnet_res_unlock(cpt);
2669 rspt = md->md_rspt_ptr;
2670 md->md_rspt_ptr = NULL;
2673 LASSERT(rspt->rspt_cpt == cpt);
2676 * invalidate the handle to indicate that a response has been
2677 * received, which will then lead the monitor thread to clean up
2680 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2681 lnet_res_unlock(cpt);
2685 lnet_finalize_expired_responses(bool force)
2687 struct lnet_libmd *md;
2688 struct list_head local_queue;
2689 struct lnet_rsp_tracker *rspt, *tmp;
2692 if (the_lnet.ln_mt_rstq == NULL)
2695 cfs_cpt_for_each(i, lnet_cpt_table()) {
2696 INIT_LIST_HEAD(&local_queue);
2699 if (!the_lnet.ln_mt_rstq[i]) {
2703 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2706 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2708 * The rspt mdh will be invalidated when a response
2709 * is received or whenever we want to discard the
2710 * block the monitor thread will walk the queue
2711 * and clean up any rsts with an invalid mdh.
2712 * The monitor thread will walk the queue until
2713 * the first unexpired rspt block. This means that
2714 * some rspt blocks which received their
2715 * corresponding responses will linger in the
2716 * queue until they are cleaned up eventually.
2719 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2721 list_del_init(&rspt->rspt_on_list);
2722 lnet_rspt_free(rspt, i);
2726 if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
2728 md = lnet_handle2md(&rspt->rspt_mdh);
2730 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2732 list_del_init(&rspt->rspt_on_list);
2733 lnet_rspt_free(rspt, i);
2736 LASSERT(md->md_rspt_ptr == rspt);
2737 md->md_rspt_ptr = NULL;
2740 list_del_init(&rspt->rspt_on_list);
2742 CDEBUG(D_NET, "Response timed out: md = %p\n", md);
2743 LNetMDUnlink(rspt->rspt_mdh);
2744 lnet_rspt_free(rspt, i);
2752 if (!list_empty(&local_queue))
2753 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2759 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2761 struct lnet_msg *msg;
2763 while (!list_empty(resendq)) {
2764 struct lnet_peer_ni *lpni;
2766 msg = list_entry(resendq->next, struct lnet_msg,
2769 list_del_init(&msg->msg_list);
2771 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2773 lnet_net_unlock(cpt);
2774 CERROR("Expected that a peer is already created for %s\n",
2775 libcfs_nid2str(msg->msg_hdr.dest_nid));
2776 msg->msg_no_resend = true;
2777 lnet_finalize(msg, -EFAULT);
2780 struct lnet_peer *peer;
2782 lnet_nid_t src_nid = LNET_NID_ANY;
2785 * if this message is not being routed and the
2786 * peer is non-MR then we must use the same
2787 * src_nid that was used in the original send.
2788 * Otherwise if we're routing the message (IE
2789 * we're a router) then we can use any of our
2790 * local interfaces. It doesn't matter to the
2791 * final destination.
2793 peer = lpni->lpni_peer_net->lpn_peer;
2794 if (!msg->msg_routing &&
2795 !lnet_peer_is_multi_rail(peer))
2796 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
2799 * If we originally specified a src NID, then we
2800 * must attempt to reuse it in the resend as well.
2802 if (msg->msg_src_nid_param != LNET_NID_ANY)
2803 src_nid = msg->msg_src_nid_param;
2804 lnet_peer_ni_decref_locked(lpni);
2806 lnet_net_unlock(cpt);
2807 rc = lnet_send(src_nid, msg, LNET_NID_ANY);
2809 CERROR("Error sending %s to %s: %d\n",
2810 lnet_msgtyp2str(msg->msg_type),
2811 libcfs_id2str(msg->msg_target), rc);
2812 msg->msg_no_resend = true;
2813 lnet_finalize(msg, rc);
2821 lnet_resend_pending_msgs(void)
2825 cfs_cpt_for_each(i, lnet_cpt_table()) {
2827 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
2832 /* called with cpt and ni_lock held */
2834 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt)
2836 struct lnet_handle_md recovery_mdh;
2838 LNetInvalidateMDHandle(&recovery_mdh);
2840 if (ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING) {
2841 recovery_mdh = ni->ni_ping_mdh;
2842 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2845 lnet_net_unlock(cpt);
2846 if (!LNetMDHandleIsInvalid(recovery_mdh))
2847 LNetMDUnlink(recovery_mdh);
2853 lnet_recover_local_nis(void)
2855 struct lnet_mt_event_info *ev_info;
2856 struct list_head processed_list;
2857 struct list_head local_queue;
2858 struct lnet_handle_md mdh;
2859 struct lnet_ni *tmp;
2865 INIT_LIST_HEAD(&local_queue);
2866 INIT_LIST_HEAD(&processed_list);
2869 * splice the recovery queue on a local queue. We will iterate
2870 * through the local queue and update it as needed. Once we're
2871 * done with the traversal, we'll splice the local queue back on
2872 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
2873 * will be traversed in the next iteration.
2876 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
2880 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
2882 * if an NI is being deleted or it is now healthy, there
2883 * is no need to keep it around in the recovery queue.
2884 * The monitor thread is the only thread responsible for
2885 * removing the NI from the recovery queue.
2886 * Multiple threads can be adding NIs to the recovery
2889 healthv = atomic_read(&ni->ni_healthv);
2893 if (!(ni->ni_state & LNET_NI_STATE_ACTIVE) ||
2894 healthv == LNET_MAX_HEALTH_VALUE) {
2895 list_del_init(&ni->ni_recovery);
2896 lnet_unlink_ni_recovery_mdh_locked(ni, 0);
2898 lnet_ni_decref_locked(ni, 0);
2906 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
2907 libcfs_nid2str(ni->ni_nid));
2910 if (!(ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING)) {
2911 ni->ni_state |= LNET_NI_STATE_RECOVERY_PENDING;
2914 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
2916 CERROR("out of memory. Can't recover %s\n",
2917 libcfs_nid2str(ni->ni_nid));
2919 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
2924 mdh = ni->ni_ping_mdh;
2926 * Invalidate the ni mdh in case it's deleted.
2927 * We'll unlink the mdh in this case below.
2929 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2933 * remove the NI from the local queue and drop the
2934 * reference count to it while we're recovering
2935 * it. The reason for that, is that the NI could
2936 * be deleted, and the way the code is structured
2937 * is if we don't drop the NI, then the deletion
2938 * code will enter a loop waiting for the
2939 * reference count to be removed while holding the
2940 * ln_mutex_lock(). When we look up the peer to
2941 * send to in lnet_select_pathway() we will try to
2942 * lock the ln_mutex_lock() as well, leading to
2943 * a deadlock. By dropping the refcount and
2944 * removing it from the list, we allow for the NI
2945 * to be removed, then we use the cached NID to
2946 * look it up again. If it's gone, then we just
2947 * continue examining the rest of the queue.
2950 list_del_init(&ni->ni_recovery);
2951 lnet_ni_decref_locked(ni, 0);
2954 ev_info->mt_type = MT_TYPE_LOCAL_NI;
2955 ev_info->mt_nid = nid;
2956 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
2957 ev_info, the_lnet.ln_mt_eqh, true);
2958 /* lookup the nid again */
2960 ni = lnet_nid2ni_locked(nid, 0);
2963 * the NI has been deleted when we dropped
2971 * Same note as in lnet_recover_peer_nis(). When
2972 * we're sending the ping, the NI is free to be
2973 * deleted or manipulated. By this point it
2974 * could've been added back on the recovery queue,
2975 * and a refcount taken on it.
2976 * So we can't just add it blindly again or we'll
2977 * corrupt the queue. We must check under lock if
2978 * it's not on any list and if not then add it
2979 * to the processed list, which will eventually be
2980 * spliced back on to the recovery queue.
2982 ni->ni_ping_mdh = mdh;
2983 if (list_empty(&ni->ni_recovery)) {
2984 list_add_tail(&ni->ni_recovery, &processed_list);
2985 lnet_ni_addref_locked(ni, 0);
2991 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
2997 * put back the remaining NIs on the ln_mt_localNIRecovq to be
2998 * reexamined in the next iteration.
3000 list_splice_init(&processed_list, &local_queue);
3002 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3006 static struct list_head **
3007 lnet_create_array_of_queues(void)
3009 struct list_head **qs;
3010 struct list_head *q;
3013 qs = cfs_percpt_alloc(lnet_cpt_table(),
3014 sizeof(struct list_head));
3016 CERROR("Failed to allocate queues\n");
3020 cfs_percpt_for_each(q, i, qs)
3027 lnet_resendqs_create(void)
3029 struct list_head **resendqs;
3030 resendqs = lnet_create_array_of_queues();
3035 lnet_net_lock(LNET_LOCK_EX);
3036 the_lnet.ln_mt_resendqs = resendqs;
3037 lnet_net_unlock(LNET_LOCK_EX);
3043 lnet_clean_local_ni_recoveryq(void)
3047 /* This is only called when the monitor thread has stopped */
3050 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3051 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3052 struct lnet_ni, ni_recovery);
3053 list_del_init(&ni->ni_recovery);
3055 lnet_unlink_ni_recovery_mdh_locked(ni, 0);
3057 lnet_ni_decref_locked(ni, 0);
3064 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt)
3066 struct lnet_handle_md recovery_mdh;
3068 LNetInvalidateMDHandle(&recovery_mdh);
3070 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) {
3071 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3072 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3074 spin_unlock(&lpni->lpni_lock);
3075 lnet_net_unlock(cpt);
3076 if (!LNetMDHandleIsInvalid(recovery_mdh))
3077 LNetMDUnlink(recovery_mdh);
3079 spin_lock(&lpni->lpni_lock);
3083 lnet_clean_peer_ni_recoveryq(void)
3085 struct lnet_peer_ni *lpni, *tmp;
3087 lnet_net_lock(LNET_LOCK_EX);
3089 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3091 list_del_init(&lpni->lpni_recovery);
3092 spin_lock(&lpni->lpni_lock);
3093 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX);
3094 spin_unlock(&lpni->lpni_lock);
3095 lnet_peer_ni_decref_locked(lpni);
3098 lnet_net_unlock(LNET_LOCK_EX);
3102 lnet_clean_resendqs(void)
3104 struct lnet_msg *msg, *tmp;
3105 struct list_head msgs;
3108 INIT_LIST_HEAD(&msgs);
3110 cfs_cpt_for_each(i, lnet_cpt_table()) {
3112 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3114 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3115 list_del_init(&msg->msg_list);
3116 msg->msg_no_resend = true;
3117 lnet_finalize(msg, -ESHUTDOWN);
3121 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3125 lnet_recover_peer_nis(void)
3127 struct lnet_mt_event_info *ev_info;
3128 struct list_head processed_list;
3129 struct list_head local_queue;
3130 struct lnet_handle_md mdh;
3131 struct lnet_peer_ni *lpni;
3132 struct lnet_peer_ni *tmp;
3137 INIT_LIST_HEAD(&local_queue);
3138 INIT_LIST_HEAD(&processed_list);
3141 * Always use cpt 0 for locking across all interactions with
3142 * ln_mt_peerNIRecovq
3145 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3149 list_for_each_entry_safe(lpni, tmp, &local_queue,
3152 * The same protection strategy is used here as is in the
3153 * local recovery case.
3156 healthv = atomic_read(&lpni->lpni_healthv);
3157 spin_lock(&lpni->lpni_lock);
3158 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3159 healthv == LNET_MAX_HEALTH_VALUE) {
3160 list_del_init(&lpni->lpni_recovery);
3161 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0);
3162 spin_unlock(&lpni->lpni_lock);
3163 lnet_peer_ni_decref_locked(lpni);
3167 spin_unlock(&lpni->lpni_lock);
3171 * NOTE: we're racing with peer deletion from user space.
3172 * It's possible that a peer is deleted after we check its
3173 * state. In this case the recovery can create a new peer
3175 spin_lock(&lpni->lpni_lock);
3176 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3177 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3178 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3179 spin_unlock(&lpni->lpni_lock);
3181 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3183 CERROR("out of memory. Can't recover %s\n",
3184 libcfs_nid2str(lpni->lpni_nid));
3185 spin_lock(&lpni->lpni_lock);
3186 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3187 spin_unlock(&lpni->lpni_lock);
3191 /* look at the comments in lnet_recover_local_nis() */
3192 mdh = lpni->lpni_recovery_ping_mdh;
3193 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3194 nid = lpni->lpni_nid;
3196 list_del_init(&lpni->lpni_recovery);
3197 lnet_peer_ni_decref_locked(lpni);
3200 ev_info->mt_type = MT_TYPE_PEER_NI;
3201 ev_info->mt_nid = nid;
3202 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3203 ev_info, the_lnet.ln_mt_eqh, true);
3206 * lnet_find_peer_ni_locked() grabs a refcount for
3207 * us. No need to take it explicitly.
3209 lpni = lnet_find_peer_ni_locked(nid);
3216 lpni->lpni_recovery_ping_mdh = mdh;
3218 * While we're unlocked the lpni could've been
3219 * readded on the recovery queue. In this case we
3220 * don't need to add it to the local queue, since
3221 * it's already on there and the thread that added
3222 * it would've incremented the refcount on the
3223 * peer, which means we need to decref the refcount
3224 * that was implicitly grabbed by find_peer_ni_locked.
3225 * Otherwise, if the lpni is still not on
3226 * the recovery queue, then we'll add it to the
3229 if (list_empty(&lpni->lpni_recovery))
3230 list_add_tail(&lpni->lpni_recovery, &processed_list);
3232 lnet_peer_ni_decref_locked(lpni);
3235 spin_lock(&lpni->lpni_lock);
3237 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3239 spin_unlock(&lpni->lpni_lock);
3242 list_splice_init(&processed_list, &local_queue);
3244 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3249 lnet_monitor_thread(void *arg)
3251 int wakeup_counter = 0;
3254 * The monitor thread takes care of the following:
3255 * 1. Checks the aliveness of routers
3256 * 2. Checks if there are messages on the resend queue to resend
3258 * 3. Check if there are any NIs on the local recovery queue and
3260 * 4. Checks if there are any NIs on the remote recovery queue
3263 cfs_block_allsigs();
3265 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3266 if (lnet_router_checker_active())
3267 lnet_check_routers();
3269 lnet_resend_pending_msgs();
3272 if (wakeup_counter >= lnet_transaction_timeout / 2) {
3273 lnet_finalize_expired_responses(false);
3277 lnet_recover_local_nis();
3279 lnet_recover_peer_nis();
3282 * TODO do we need to check if we should sleep without
3283 * timeout? Technically, an active system will always
3284 * have messages in flight so this check will always
3285 * evaluate to false. And on an idle system do we care
3286 * if we wake up every 1 second? Although, we've seen
3287 * cases where we get a complaint that an idle thread
3288 * is waking up unnecessarily.
3290 wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
3292 cfs_time_seconds(1));
3295 /* clean up the router checker */
3296 lnet_prune_rc_data(1);
3299 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3301 /* signal that the monitor thread is exiting */
3302 up(&the_lnet.ln_mt_signal);
3310 * Returns == 0 if success
3311 * Returns > 0 if LNetMDBind or prior fails
3312 * Returns < 0 if LNetGet fails
3315 lnet_send_ping(lnet_nid_t dest_nid,
3316 struct lnet_handle_md *mdh, int nnis,
3317 void *user_data, struct lnet_handle_eq eqh, bool recovery)
3319 struct lnet_md md = { NULL };
3320 struct lnet_process_id id;
3321 struct lnet_ping_buffer *pbuf;
3324 if (dest_nid == LNET_NID_ANY) {
3329 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3335 /* initialize md content */
3336 md.start = &pbuf->pb_info;
3337 md.length = LNET_PING_INFO_SIZE(nnis);
3338 md.threshold = 2; /* GET/REPLY */
3340 md.options = LNET_MD_TRUNCATE;
3341 md.user_ptr = user_data;
3344 rc = LNetMDBind(md, LNET_UNLINK, mdh);
3346 lnet_ping_buffer_decref(pbuf);
3347 CERROR("Can't bind MD: %d\n", rc);
3348 rc = -rc; /* change the rc to positive */
3351 id.pid = LNET_PID_LUSTRE;
3354 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3355 LNET_RESERVED_PORTAL,
3356 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3359 goto fail_unlink_md;
3365 LNetInvalidateMDHandle(mdh);
3371 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3374 lnet_nid_t nid = ev_info->mt_nid;
3376 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3380 ni = lnet_nid2ni_locked(nid, 0);
3386 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
3391 CERROR("local NI recovery failed with %d\n", status);
3395 * need to increment healthv for the ni here, because in
3396 * the lnet_finalize() path we don't have access to this
3397 * NI. And in order to get access to it, we'll need to
3398 * carry forward too much information.
3399 * In the peer case, it'll naturally be incremented
3401 lnet_inc_healthv(&ni->ni_healthv);
3403 struct lnet_peer_ni *lpni;
3406 cpt = lnet_net_lock_current();
3407 lpni = lnet_find_peer_ni_locked(nid);
3409 lnet_net_unlock(cpt);
3412 spin_lock(&lpni->lpni_lock);
3413 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3414 spin_unlock(&lpni->lpni_lock);
3415 lnet_peer_ni_decref_locked(lpni);
3416 lnet_net_unlock(cpt);
3419 CERROR("peer NI recovery failed with %d\n", status);
3424 lnet_mt_event_handler(struct lnet_event *event)
3426 struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3427 struct lnet_ping_buffer *pbuf;
3429 /* TODO: remove assert */
3430 LASSERT(event->type == LNET_EVENT_REPLY ||
3431 event->type == LNET_EVENT_SEND ||
3432 event->type == LNET_EVENT_UNLINK);
3434 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3437 switch (event->type) {
3438 case LNET_EVENT_UNLINK:
3439 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3440 libcfs_nid2str(ev_info->mt_nid));
3441 case LNET_EVENT_REPLY:
3442 lnet_handle_recovery_reply(ev_info, event->status);
3444 case LNET_EVENT_SEND:
3445 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3446 libcfs_nid2str(ev_info->mt_nid),
3447 (event->status) ? "unsuccessfully" :
3448 "successfully", event->status);
3451 CERROR("Unexpected event: %d\n", event->type);
3454 if (event->unlinked) {
3455 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3456 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3457 lnet_ping_buffer_decref(pbuf);
3462 lnet_rsp_tracker_create(void)
3464 struct list_head **rstqs;
3465 rstqs = lnet_create_array_of_queues();
3470 the_lnet.ln_mt_rstq = rstqs;
3476 lnet_rsp_tracker_clean(void)
3478 lnet_finalize_expired_responses(true);
3480 cfs_percpt_free(the_lnet.ln_mt_rstq);
3481 the_lnet.ln_mt_rstq = NULL;
3484 int lnet_monitor_thr_start(void)
3487 struct task_struct *task;
3489 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3492 rc = lnet_resendqs_create();
3496 rc = lnet_rsp_tracker_create();
3500 rc = LNetEQAlloc(0, lnet_mt_event_handler, &the_lnet.ln_mt_eqh);
3502 CERROR("Can't allocate monitor thread EQ: %d\n", rc);
3506 /* Pre monitor thread start processing */
3507 rc = lnet_router_pre_mt_start();
3511 sema_init(&the_lnet.ln_mt_signal, 0);
3513 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3514 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3517 CERROR("Can't start monitor thread: %d\n", rc);
3521 /* post monitor thread start processing */
3522 lnet_router_post_mt_start();
3527 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3528 /* block until event callback signals exit */
3529 down(&the_lnet.ln_mt_signal);
3531 lnet_router_cleanup();
3533 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3534 lnet_rsp_tracker_clean();
3535 lnet_clean_local_ni_recoveryq();
3536 lnet_clean_peer_ni_recoveryq();
3537 lnet_clean_resendqs();
3538 LNetEQFree(the_lnet.ln_mt_eqh);
3539 LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3542 lnet_rsp_tracker_clean();
3543 lnet_clean_local_ni_recoveryq();
3544 lnet_clean_peer_ni_recoveryq();
3545 lnet_clean_resendqs();
3549 void lnet_monitor_thr_stop(void)
3553 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3556 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3557 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3559 /* tell the monitor thread that we're shutting down */
3560 wake_up(&the_lnet.ln_mt_waitq);
3562 /* block until monitor thread signals that it's done */
3563 down(&the_lnet.ln_mt_signal);
3564 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3566 /* perform cleanup tasks */
3567 lnet_router_cleanup();
3568 lnet_rsp_tracker_clean();
3569 lnet_clean_local_ni_recoveryq();
3570 lnet_clean_peer_ni_recoveryq();
3571 lnet_clean_resendqs();
3572 rc = LNetEQFree(the_lnet.ln_mt_eqh);
3578 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3582 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3583 the_lnet.ln_counters[cpt]->drop_count++;
3584 the_lnet.ln_counters[cpt]->drop_length += nob;
3585 lnet_net_unlock(cpt);
3587 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3591 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3593 struct lnet_hdr *hdr = &msg->msg_hdr;
3595 if (msg->msg_wanted != 0)
3596 lnet_setpayloadbuffer(msg);
3598 lnet_build_msg_event(msg, LNET_EVENT_PUT);
3600 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
3601 * it back into the ACK during lnet_finalize() */
3602 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3603 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3605 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3606 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3610 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3612 struct lnet_hdr *hdr = &msg->msg_hdr;
3613 struct lnet_match_info info;
3617 /* Convert put fields to host byte order */
3618 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3619 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
3620 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
3622 /* Primary peer NID. */
3623 info.mi_id.nid = msg->msg_initiator;
3624 info.mi_id.pid = hdr->src_pid;
3625 info.mi_opc = LNET_MD_OP_PUT;
3626 info.mi_portal = hdr->msg.put.ptl_index;
3627 info.mi_rlength = hdr->payload_length;
3628 info.mi_roffset = hdr->msg.put.offset;
3629 info.mi_mbits = hdr->msg.put.match_bits;
3630 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
3632 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3633 ready_delay = msg->msg_rx_ready_delay;
3636 rc = lnet_ptl_match_md(&info, msg);
3641 case LNET_MATCHMD_OK:
3642 lnet_recv_put(ni, msg);
3645 case LNET_MATCHMD_NONE:
3647 /* no eager_recv or has already called it, should
3648 * have been attached on delayed list */
3651 rc = lnet_ni_eager_recv(ni, msg);
3658 case LNET_MATCHMD_DROP:
3659 CNETERR("Dropping PUT from %s portal %d match %llu"
3660 " offset %d length %d: %d\n",
3661 libcfs_id2str(info.mi_id), info.mi_portal,
3662 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3664 return -ENOENT; /* -ve: OK but no match */
3669 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3671 struct lnet_match_info info;
3672 struct lnet_hdr *hdr = &msg->msg_hdr;
3673 struct lnet_process_id source_id;
3674 struct lnet_handle_wire reply_wmd;
3677 /* Convert get fields to host byte order */
3678 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
3679 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
3680 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
3681 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
3683 source_id.nid = hdr->src_nid;
3684 source_id.pid = hdr->src_pid;
3685 /* Primary peer NID */
3686 info.mi_id.nid = msg->msg_initiator;
3687 info.mi_id.pid = hdr->src_pid;
3688 info.mi_opc = LNET_MD_OP_GET;
3689 info.mi_portal = hdr->msg.get.ptl_index;
3690 info.mi_rlength = hdr->msg.get.sink_length;
3691 info.mi_roffset = hdr->msg.get.src_offset;
3692 info.mi_mbits = hdr->msg.get.match_bits;
3693 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
3695 rc = lnet_ptl_match_md(&info, msg);
3696 if (rc == LNET_MATCHMD_DROP) {
3697 CNETERR("Dropping GET from %s portal %d match %llu"
3698 " offset %d length %d\n",
3699 libcfs_id2str(info.mi_id), info.mi_portal,
3700 info.mi_mbits, info.mi_roffset, info.mi_rlength);
3701 return -ENOENT; /* -ve: OK but no match */
3704 LASSERT(rc == LNET_MATCHMD_OK);
3706 lnet_build_msg_event(msg, LNET_EVENT_GET);
3708 reply_wmd = hdr->msg.get.return_wmd;
3710 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3711 msg->msg_offset, msg->msg_wanted);
3713 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3716 /* The LND completes the REPLY from her recv procedure */
3717 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3718 msg->msg_offset, msg->msg_len, msg->msg_len);
3722 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3723 msg->msg_receiving = 0;
3725 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
3727 /* didn't get as far as lnet_ni_send() */
3728 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3729 libcfs_nid2str(ni->ni_nid),
3730 libcfs_id2str(info.mi_id), rc);
3732 lnet_finalize(msg, rc);
3739 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3741 void *private = msg->msg_private;
3742 struct lnet_hdr *hdr = &msg->msg_hdr;
3743 struct lnet_process_id src = {0};
3744 struct lnet_libmd *md;
3749 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3752 src.nid = hdr->src_nid;
3753 src.pid = hdr->src_pid;
3755 /* NB handles only looked up by creator (no flips) */
3756 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
3757 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3758 CNETERR("%s: Dropping REPLY from %s for %s "
3760 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3761 (md == NULL) ? "invalid" : "inactive",
3762 hdr->msg.reply.dst_wmd.wh_interface_cookie,
3763 hdr->msg.reply.dst_wmd.wh_object_cookie);
3764 if (md != NULL && md->md_me != NULL)
3765 CERROR("REPLY MD also attached to portal %d\n",
3766 md->md_me->me_portal);
3768 lnet_res_unlock(cpt);
3769 return -ENOENT; /* -ve: OK but no match */
3772 LASSERT(md->md_offset == 0);
3774 rlength = hdr->payload_length;
3775 mlength = MIN(rlength, (int)md->md_length);
3777 if (mlength < rlength &&
3778 (md->md_options & LNET_MD_TRUNCATE) == 0) {
3779 CNETERR("%s: Dropping REPLY from %s length %d "
3780 "for MD %#llx would overflow (%d)\n",
3781 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3782 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
3784 lnet_res_unlock(cpt);
3785 return -ENOENT; /* -ve: OK but no match */
3788 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
3789 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3790 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
3792 lnet_msg_attach_md(msg, md, 0, mlength);
3795 lnet_setpayloadbuffer(msg);
3797 lnet_res_unlock(cpt);
3799 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3801 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
3806 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
3808 struct lnet_hdr *hdr = &msg->msg_hdr;
3809 struct lnet_process_id src = {0};
3810 struct lnet_libmd *md;
3813 src.nid = hdr->src_nid;
3814 src.pid = hdr->src_pid;
3816 /* Convert ack fields to host byte order */
3817 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
3818 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
3820 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
3823 /* NB handles only looked up by creator (no flips) */
3824 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
3825 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3826 /* Don't moan; this is expected */
3828 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
3829 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3830 (md == NULL) ? "invalid" : "inactive",
3831 hdr->msg.ack.dst_wmd.wh_interface_cookie,
3832 hdr->msg.ack.dst_wmd.wh_object_cookie);
3833 if (md != NULL && md->md_me != NULL)
3834 CERROR("Source MD also attached to portal %d\n",
3835 md->md_me->me_portal);
3837 lnet_res_unlock(cpt);
3838 return -ENOENT; /* -ve! */
3841 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
3842 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3843 hdr->msg.ack.dst_wmd.wh_object_cookie);
3845 lnet_msg_attach_md(msg, md, 0, 0);
3847 lnet_res_unlock(cpt);
3849 lnet_build_msg_event(msg, LNET_EVENT_ACK);
3851 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
3856 * \retval LNET_CREDIT_OK If \a msg is forwarded
3857 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
3858 * \retval -ve error code
3861 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
3865 if (!the_lnet.ln_routing)
3868 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
3869 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
3870 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
3871 msg->msg_rx_ready_delay = 1;
3873 lnet_net_unlock(msg->msg_rx_cpt);
3874 rc = lnet_ni_eager_recv(ni, msg);
3875 lnet_net_lock(msg->msg_rx_cpt);
3880 rc = lnet_post_routed_recv_locked(msg, 0);
3885 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
3889 switch (msg->msg_type) {
3891 rc = lnet_parse_ack(ni, msg);
3894 rc = lnet_parse_put(ni, msg);
3897 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
3899 case LNET_MSG_REPLY:
3900 rc = lnet_parse_reply(ni, msg);
3902 default: /* prevent an unused label if !kernel */
3907 LASSERT(rc == 0 || rc == -ENOENT);
3912 lnet_msgtyp2str (int type)
3921 case LNET_MSG_REPLY:
3923 case LNET_MSG_HELLO:
3926 return ("<UNKNOWN>");
3931 lnet_print_hdr(struct lnet_hdr *hdr)
3933 struct lnet_process_id src = {
3934 .nid = hdr->src_nid,
3935 .pid = hdr->src_pid,
3937 struct lnet_process_id dst = {
3938 .nid = hdr->dest_nid,
3939 .pid = hdr->dest_pid,
3941 char *type_str = lnet_msgtyp2str(hdr->type);
3943 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
3944 CWARN(" From %s\n", libcfs_id2str(src));
3945 CWARN(" To %s\n", libcfs_id2str(dst));
3947 switch (hdr->type) {
3952 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
3953 "match bits %llu\n",
3954 hdr->msg.put.ptl_index,
3955 hdr->msg.put.ack_wmd.wh_interface_cookie,
3956 hdr->msg.put.ack_wmd.wh_object_cookie,
3957 hdr->msg.put.match_bits);
3958 CWARN(" Length %d, offset %d, hdr data %#llx\n",
3959 hdr->payload_length, hdr->msg.put.offset,
3960 hdr->msg.put.hdr_data);
3964 CWARN(" Ptl index %d, return md %#llx.%#llx, "
3965 "match bits %llu\n", hdr->msg.get.ptl_index,
3966 hdr->msg.get.return_wmd.wh_interface_cookie,
3967 hdr->msg.get.return_wmd.wh_object_cookie,
3968 hdr->msg.get.match_bits);
3969 CWARN(" Length %d, src offset %d\n",
3970 hdr->msg.get.sink_length,
3971 hdr->msg.get.src_offset);
3975 CWARN(" dst md %#llx.%#llx, "
3976 "manipulated length %d\n",
3977 hdr->msg.ack.dst_wmd.wh_interface_cookie,
3978 hdr->msg.ack.dst_wmd.wh_object_cookie,
3979 hdr->msg.ack.mlength);
3982 case LNET_MSG_REPLY:
3983 CWARN(" dst md %#llx.%#llx, "
3985 hdr->msg.reply.dst_wmd.wh_interface_cookie,
3986 hdr->msg.reply.dst_wmd.wh_object_cookie,
3987 hdr->payload_length);
3993 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
3994 void *private, int rdma_req)
3999 struct lnet_msg *msg;
4000 lnet_pid_t dest_pid;
4001 lnet_nid_t dest_nid;
4003 struct lnet_peer_ni *lpni;
4004 __u32 payload_length;
4007 LASSERT (!in_interrupt ());
4009 type = le32_to_cpu(hdr->type);
4010 src_nid = le64_to_cpu(hdr->src_nid);
4011 dest_nid = le64_to_cpu(hdr->dest_nid);
4012 dest_pid = le32_to_cpu(hdr->dest_pid);
4013 payload_length = le32_to_cpu(hdr->payload_length);
4015 for_me = (ni->ni_nid == dest_nid);
4016 cpt = lnet_cpt_of_nid(from_nid, ni);
4018 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4019 libcfs_nid2str(dest_nid),
4020 libcfs_nid2str(ni->ni_nid),
4021 libcfs_nid2str(src_nid),
4022 lnet_msgtyp2str(type),
4023 (for_me) ? "for me" : "routed");
4028 if (payload_length > 0) {
4029 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4030 libcfs_nid2str(from_nid),
4031 libcfs_nid2str(src_nid),
4032 lnet_msgtyp2str(type), payload_length);
4038 case LNET_MSG_REPLY:
4039 if (payload_length >
4040 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4041 CERROR("%s, src %s: bad %s payload %d "
4042 "(%d max expected)\n",
4043 libcfs_nid2str(from_nid),
4044 libcfs_nid2str(src_nid),
4045 lnet_msgtyp2str(type),
4047 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4053 CERROR("%s, src %s: Bad message type 0x%x\n",
4054 libcfs_nid2str(from_nid),
4055 libcfs_nid2str(src_nid), type);
4059 if (the_lnet.ln_routing &&
4060 ni->ni_last_alive != ktime_get_real_seconds()) {
4061 /* NB: so far here is the only place to set NI status to "up */
4063 ni->ni_last_alive = ktime_get_real_seconds();
4064 if (ni->ni_status != NULL &&
4065 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
4066 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4070 /* Regard a bad destination NID as a protocol error. Senders should
4071 * know what they're doing; if they don't they're misconfigured, buggy
4072 * or malicious so we chop them off at the knees :) */
4075 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4076 /* should have gone direct */
4077 CERROR("%s, src %s: Bad dest nid %s "
4078 "(should have been sent direct)\n",
4079 libcfs_nid2str(from_nid),
4080 libcfs_nid2str(src_nid),
4081 libcfs_nid2str(dest_nid));
4085 if (lnet_islocalnid(dest_nid)) {
4086 /* dest is another local NI; sender should have used
4087 * this node's NID on its own network */
4088 CERROR("%s, src %s: Bad dest nid %s "
4089 "(it's my nid but on a different network)\n",
4090 libcfs_nid2str(from_nid),
4091 libcfs_nid2str(src_nid),
4092 libcfs_nid2str(dest_nid));
4096 if (rdma_req && type == LNET_MSG_GET) {
4097 CERROR("%s, src %s: Bad optimized GET for %s "
4098 "(final destination must be me)\n",
4099 libcfs_nid2str(from_nid),
4100 libcfs_nid2str(src_nid),
4101 libcfs_nid2str(dest_nid));
4105 if (!the_lnet.ln_routing) {
4106 CERROR("%s, src %s: Dropping message for %s "
4107 "(routing not enabled)\n",
4108 libcfs_nid2str(from_nid),
4109 libcfs_nid2str(src_nid),
4110 libcfs_nid2str(dest_nid));
4115 /* Message looks OK; we're not going to return an error, so we MUST
4116 * call back lnd_recv() come what may... */
4118 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4119 fail_peer(src_nid, 0)) { /* shall we now? */
4120 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4121 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4122 lnet_msgtyp2str(type));
4126 if (!list_empty(&the_lnet.ln_drop_rules) &&
4127 lnet_drop_rule_match(hdr)) {
4128 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4129 "silent message loss\n",
4130 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4131 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4136 msg = lnet_msg_alloc();
4138 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4139 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4140 lnet_msgtyp2str(type));
4144 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4145 * pointers NULL etc */
4147 msg->msg_type = type;
4148 msg->msg_private = private;
4149 msg->msg_receiving = 1;
4150 msg->msg_rdma_get = rdma_req;
4151 msg->msg_len = msg->msg_wanted = payload_length;
4152 msg->msg_offset = 0;
4153 msg->msg_hdr = *hdr;
4154 /* for building message event */
4155 msg->msg_from = from_nid;
4157 msg->msg_target.pid = dest_pid;
4158 msg->msg_target.nid = dest_nid;
4159 msg->msg_routing = 1;
4162 /* convert common msg->hdr fields to host byteorder */
4163 msg->msg_hdr.type = type;
4164 msg->msg_hdr.src_nid = src_nid;
4165 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4166 msg->msg_hdr.dest_nid = dest_nid;
4167 msg->msg_hdr.dest_pid = dest_pid;
4168 msg->msg_hdr.payload_length = payload_length;
4172 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4174 lnet_net_unlock(cpt);
4175 CERROR("%s, src %s: Dropping %s "
4176 "(error %ld looking up sender)\n",
4177 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4178 lnet_msgtyp2str(type), PTR_ERR(lpni));
4180 if (rc == -ESHUTDOWN)
4181 /* We are shutting down. Don't do anything more */
4185 msg->msg_rxpeer = lpni;
4187 lnet_ni_addref_locked(ni, cpt);
4188 /* Multi-Rail: Primary NID of source. */
4189 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4191 if (lnet_isrouter(msg->msg_rxpeer)) {
4192 lnet_peer_set_alive(msg->msg_rxpeer);
4193 if (avoid_asym_router_failure &&
4194 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4195 /* received a remote message from router, update
4196 * remote NI status on this router.
4197 * NB: multi-hop routed message will be ignored.
4199 lnet_router_ni_update_locked(msg->msg_rxpeer,
4200 LNET_NIDNET(src_nid));
4204 lnet_msg_commit(msg, cpt);
4206 /* message delay simulation */
4207 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4208 lnet_delay_rule_match_locked(hdr, msg))) {
4209 lnet_net_unlock(cpt);
4214 rc = lnet_parse_forward_locked(ni, msg);
4215 lnet_net_unlock(cpt);
4220 if (rc == LNET_CREDIT_OK) {
4221 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4222 0, payload_length, payload_length);
4227 lnet_net_unlock(cpt);
4229 rc = lnet_parse_local(ni, msg);
4235 LASSERT(msg->msg_md == NULL);
4236 lnet_finalize(msg, rc);
4239 lnet_drop_message(ni, cpt, private, payload_length, type);
4242 EXPORT_SYMBOL(lnet_parse);
4245 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4247 while (!list_empty(head)) {
4248 struct lnet_process_id id = {0};
4249 struct lnet_msg *msg;
4251 msg = list_entry(head->next, struct lnet_msg, msg_list);
4252 list_del(&msg->msg_list);
4254 id.nid = msg->msg_hdr.src_nid;
4255 id.pid = msg->msg_hdr.src_pid;
4257 LASSERT(msg->msg_md == NULL);
4258 LASSERT(msg->msg_rx_delayed);
4259 LASSERT(msg->msg_rxpeer != NULL);
4260 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4262 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4263 " offset %d length %d: %s\n",
4265 msg->msg_hdr.msg.put.ptl_index,
4266 msg->msg_hdr.msg.put.match_bits,
4267 msg->msg_hdr.msg.put.offset,
4268 msg->msg_hdr.payload_length, reason);
4270 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4271 * called lnet_drop_message(), so I just hang onto msg as well
4272 * until that's done */
4274 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4275 msg->msg_private, msg->msg_len,
4278 msg->msg_no_resend = true;
4280 * NB: message will not generate event because w/o attached MD,
4281 * but we still should give error code so lnet_msg_decommit()
4282 * can skip counters operations and other checks.
4284 lnet_finalize(msg, -ENOENT);
4289 lnet_recv_delayed_msg_list(struct list_head *head)
4291 while (!list_empty(head)) {
4292 struct lnet_msg *msg;
4293 struct lnet_process_id id;
4295 msg = list_entry(head->next, struct lnet_msg, msg_list);
4296 list_del(&msg->msg_list);
4298 /* md won't disappear under me, since each msg
4299 * holds a ref on it */
4301 id.nid = msg->msg_hdr.src_nid;
4302 id.pid = msg->msg_hdr.src_pid;
4304 LASSERT(msg->msg_rx_delayed);
4305 LASSERT(msg->msg_md != NULL);
4306 LASSERT(msg->msg_rxpeer != NULL);
4307 LASSERT(msg->msg_rxni != NULL);
4308 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4310 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4311 "match %llu offset %d length %d.\n",
4312 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4313 msg->msg_hdr.msg.put.match_bits,
4314 msg->msg_hdr.msg.put.offset,
4315 msg->msg_hdr.payload_length);
4317 lnet_recv_put(msg->msg_rxni, msg);
4322 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4323 struct lnet_libmd *md, struct lnet_handle_md mdh)
4328 * MD has a refcount taken by message so it's not going away.
4329 * The MD however can be looked up. We need to secure the access
4330 * to the md_rspt_ptr by taking the res_lock.
4331 * The rspt can be accessed without protection up to when it gets
4332 * added to the list.
4336 LASSERT(md->md_rspt_ptr == NULL);
4338 /* we'll use that same event in case we never get a response */
4339 rspt->rspt_mdh = mdh;
4340 rspt->rspt_cpt = cpt;
4341 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4342 rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4345 /* store the rspt so we can access it when we get the REPLY */
4346 md->md_rspt_ptr = rspt;
4347 lnet_res_unlock(cpt);
4350 * add to the list of tracked responses. It's added to tail of the
4351 * list in order to expire all the older entries first.
4354 list_add_tail(&rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4355 lnet_net_unlock(cpt);
4359 * Initiate an asynchronous PUT operation.
4361 * There are several events associated with a PUT: completion of the send on
4362 * the initiator node (LNET_EVENT_SEND), and when the send completes
4363 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4364 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4365 * used at the target node to indicate the completion of incoming data
4368 * The local events will be logged in the EQ associated with the MD pointed to
4369 * by \a mdh handle. Using a MD without an associated EQ results in these
4370 * events being discarded. In this case, the caller must have another
4371 * mechanism (e.g., a higher level protocol) for determining when it is safe
4372 * to modify the memory region associated with the MD.
4374 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4375 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4377 * \param self Indicates the NID of a local interface through which to send
4378 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4379 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4380 * must be "free floating" (See LNetMDBind()).
4381 * \param ack Controls whether an acknowledgment is requested.
4382 * Acknowledgments are only sent when they are requested by the initiating
4383 * process and the target MD enables them.
4384 * \param target A process identifier for the target process.
4385 * \param portal The index in the \a target's portal table.
4386 * \param match_bits The match bits to use for MD selection at the target
4388 * \param offset The offset into the target MD (only used when the target
4389 * MD has the LNET_MD_MANAGE_REMOTE option set).
4390 * \param hdr_data 64 bits of user data that can be included in the message
4391 * header. This data is written to an event queue entry at the target if an
4392 * EQ is present on the matching MD.
4394 * \retval 0 Success, and only in this case events will be generated
4395 * and logged to EQ (if it exists).
4396 * \retval -EIO Simulated failure.
4397 * \retval -ENOMEM Memory allocation failure.
4398 * \retval -ENOENT Invalid MD object.
4400 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4403 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4404 struct lnet_process_id target, unsigned int portal,
4405 __u64 match_bits, unsigned int offset,
4408 struct lnet_msg *msg;
4409 struct lnet_libmd *md;
4412 struct lnet_rsp_tracker *rspt = NULL;
4414 LASSERT(the_lnet.ln_refcount > 0);
4416 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4417 fail_peer(target.nid, 1)) { /* shall we now? */
4418 CERROR("Dropping PUT to %s: simulated failure\n",
4419 libcfs_id2str(target));
4423 msg = lnet_msg_alloc();
4425 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4426 libcfs_id2str(target));
4429 msg->msg_vmflush = !!memory_pressure_get();
4431 cpt = lnet_cpt_of_cookie(mdh.cookie);
4433 if (ack == LNET_ACK_REQ) {
4434 rspt = lnet_rspt_alloc(cpt);
4436 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4437 libcfs_id2str(target));
4440 INIT_LIST_HEAD(&rspt->rspt_on_list);
4445 md = lnet_handle2md(&mdh);
4446 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4447 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4448 match_bits, portal, libcfs_id2str(target),
4449 md == NULL ? -1 : md->md_threshold);
4450 if (md != NULL && md->md_me != NULL)
4451 CERROR("Source MD also attached to portal %d\n",
4452 md->md_me->me_portal);
4453 lnet_res_unlock(cpt);
4455 LIBCFS_FREE(rspt, sizeof(*rspt));
4460 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4462 lnet_msg_attach_md(msg, md, 0, 0);
4464 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4466 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4467 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4468 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4469 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4471 /* NB handles only looked up by creator (no flips) */
4472 if (ack == LNET_ACK_REQ) {
4473 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4474 the_lnet.ln_interface_cookie;
4475 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4476 md->md_lh.lh_cookie;
4478 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4479 LNET_WIRE_HANDLE_COOKIE_NONE;
4480 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4481 LNET_WIRE_HANDLE_COOKIE_NONE;
4484 lnet_res_unlock(cpt);
4486 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4488 if (ack == LNET_ACK_REQ)
4489 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4491 rc = lnet_send(self, msg, LNET_NID_ANY);
4493 CNETERR("Error sending PUT to %s: %d\n",
4494 libcfs_id2str(target), rc);
4495 msg->msg_no_resend = true;
4496 lnet_detach_rsp_tracker(msg->msg_md, cpt);
4497 lnet_finalize(msg, rc);
4500 /* completion will be signalled by an event */
4503 EXPORT_SYMBOL(LNetPut);
4506 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
4507 * returns a msg for the LND to pass to lnet_finalize() when the sink
4508 * data has been received.
4510 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4511 * lnet_finalize() is called on it, so the LND must call this first
4514 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4516 struct lnet_msg *msg = lnet_msg_alloc();
4517 struct lnet_libmd *getmd = getmsg->msg_md;
4518 struct lnet_process_id peer_id = getmsg->msg_target;
4521 LASSERT(!getmsg->msg_target_is_router);
4522 LASSERT(!getmsg->msg_routing);
4525 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4526 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4530 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4533 LASSERT(getmd->md_refcount > 0);
4535 if (getmd->md_threshold == 0) {
4536 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4537 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4539 lnet_res_unlock(cpt);
4543 LASSERT(getmd->md_offset == 0);
4545 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4546 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4548 /* setup information for lnet_build_msg_event */
4549 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4550 msg->msg_from = peer_id.nid;
4551 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4552 msg->msg_hdr.src_nid = peer_id.nid;
4553 msg->msg_hdr.payload_length = getmd->md_length;
4554 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4556 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4557 lnet_res_unlock(cpt);
4559 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4562 lnet_msg_commit(msg, cpt);
4563 lnet_net_unlock(cpt);
4565 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4570 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4573 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4574 the_lnet.ln_counters[cpt]->drop_count++;
4575 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
4576 lnet_net_unlock(cpt);
4583 EXPORT_SYMBOL(lnet_create_reply_msg);
4586 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4589 /* Set the REPLY length, now the RDMA that elides the REPLY message has
4590 * completed and I know it. */
4591 LASSERT(reply != NULL);
4592 LASSERT(reply->msg_type == LNET_MSG_GET);
4593 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4595 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
4596 * the end of my buffer, I might as well be dead. */
4597 LASSERT(len <= reply->msg_ev.mlength);
4599 reply->msg_ev.mlength = len;
4601 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4604 * Initiate an asynchronous GET operation.
4606 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4607 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4608 * the target node in the REPLY has been written to local MD.
4610 * On the target node, an LNET_EVENT_GET is logged when the GET request
4611 * arrives and is accepted into a MD.
4613 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4614 * \param mdh A handle for the MD that describes the memory into which the
4615 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4617 * \retval 0 Success, and only in this case events will be generated
4618 * and logged to EQ (if it exists) of the MD.
4619 * \retval -EIO Simulated failure.
4620 * \retval -ENOMEM Memory allocation failure.
4621 * \retval -ENOENT Invalid MD object.
4624 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4625 struct lnet_process_id target, unsigned int portal,
4626 __u64 match_bits, unsigned int offset, bool recovery)
4628 struct lnet_msg *msg;
4629 struct lnet_libmd *md;
4630 struct lnet_rsp_tracker *rspt;
4634 LASSERT(the_lnet.ln_refcount > 0);
4636 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4637 fail_peer(target.nid, 1)) /* shall we now? */
4639 CERROR("Dropping GET to %s: simulated failure\n",
4640 libcfs_id2str(target));
4644 msg = lnet_msg_alloc();
4646 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4647 libcfs_id2str(target));
4651 cpt = lnet_cpt_of_cookie(mdh.cookie);
4653 rspt = lnet_rspt_alloc(cpt);
4655 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4656 libcfs_id2str(target));
4659 INIT_LIST_HEAD(&rspt->rspt_on_list);
4661 msg->msg_recovery = recovery;
4665 md = lnet_handle2md(&mdh);
4666 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4667 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4668 match_bits, portal, libcfs_id2str(target),
4669 md == NULL ? -1 : md->md_threshold);
4670 if (md != NULL && md->md_me != NULL)
4671 CERROR("REPLY MD also attached to portal %d\n",
4672 md->md_me->me_portal);
4674 lnet_res_unlock(cpt);
4677 LIBCFS_FREE(rspt, sizeof(*rspt));
4681 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4683 lnet_msg_attach_md(msg, md, 0, 0);
4685 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
4687 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
4688 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
4689 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
4690 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
4692 /* NB handles only looked up by creator (no flips) */
4693 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
4694 the_lnet.ln_interface_cookie;
4695 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
4696 md->md_lh.lh_cookie;
4698 lnet_res_unlock(cpt);
4700 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4702 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4704 rc = lnet_send(self, msg, LNET_NID_ANY);
4706 CNETERR("Error sending GET to %s: %d\n",
4707 libcfs_id2str(target), rc);
4708 msg->msg_no_resend = true;
4709 lnet_detach_rsp_tracker(msg->msg_md, cpt);
4710 lnet_finalize(msg, rc);
4713 /* completion will be signalled by an event */
4716 EXPORT_SYMBOL(LNetGet);
4719 * Calculate distance to node at \a dstnid.
4721 * \param dstnid Target NID.
4722 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
4724 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
4727 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
4728 * local_nid_dist_zero is set, which is the default.
4729 * \retval positives Distance to target NID, i.e. number of hops plus one.
4730 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
4733 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
4735 struct list_head *e;
4736 struct lnet_ni *ni = NULL;
4737 struct lnet_remotenet *rnet;
4738 __u32 dstnet = LNET_NIDNET(dstnid);
4742 struct list_head *rn_list;
4744 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
4745 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
4746 * keep order 0 free for 0@lo and order 1 free for a local NID
4749 LASSERT(the_lnet.ln_refcount > 0);
4751 cpt = lnet_net_lock_current();
4753 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
4754 if (ni->ni_nid == dstnid) {
4755 if (srcnidp != NULL)
4757 if (orderp != NULL) {
4758 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
4763 lnet_net_unlock(cpt);
4765 return local_nid_dist_zero ? 0 : 1;
4768 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
4769 /* Check if ni was originally created in
4770 * current net namespace.
4771 * If not, assign order above 0xffff0000,
4772 * to make this ni not a priority. */
4773 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
4774 order += 0xffff0000;
4776 if (srcnidp != NULL)
4777 *srcnidp = ni->ni_nid;
4780 lnet_net_unlock(cpt);
4787 rn_list = lnet_net2rnethash(dstnet);
4788 list_for_each(e, rn_list) {
4789 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
4791 if (rnet->lrn_net == dstnet) {
4792 struct lnet_route *route;
4793 struct lnet_route *shortest = NULL;
4794 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
4797 LASSERT(!list_empty(&rnet->lrn_routes));
4799 list_for_each_entry(route, &rnet->lrn_routes,
4801 route_hops = route->lr_hops;
4802 if (route_hops == LNET_UNDEFINED_HOPS)
4804 if (shortest == NULL ||
4805 route_hops < shortest_hops) {
4807 shortest_hops = route_hops;
4811 LASSERT(shortest != NULL);
4812 hops = shortest_hops;
4813 if (srcnidp != NULL) {
4814 ni = lnet_get_next_ni_locked(
4815 shortest->lr_gateway->lpni_net,
4817 *srcnidp = ni->ni_nid;
4821 lnet_net_unlock(cpt);
4827 lnet_net_unlock(cpt);
4828 return -EHOSTUNREACH;
4830 EXPORT_SYMBOL(LNetDist);