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,
869 struct lnet_msg *msg)
871 time64_t now = ktime_get_seconds();
873 if (!lnet_peer_aliveness_enabled(lp))
876 if (lnet_peer_is_alive(lp, now))
880 * If we're resending a message, let's attempt to send it even if
881 * the peer is down to fulfill our resend quota on the message
883 if (msg->msg_retry_count > 0)
887 * Peer appears dead, but we should avoid frequent NI queries (at
888 * most once per lnet_queryinterval seconds).
890 if (lp->lpni_last_query != 0) {
891 static const int lnet_queryinterval = 1;
894 next_query = lp->lpni_last_query + lnet_queryinterval;
896 if (now < next_query) {
898 CWARN("Unexpected aliveness of peer %s: "
899 "%lld < %lld (%d/%d)\n",
900 libcfs_nid2str(lp->lpni_nid),
903 lp->lpni_net->net_tunables.lct_peer_timeout);
908 /* query NI for latest aliveness news */
909 lnet_ni_query_locked(ni, lp);
911 if (lnet_peer_is_alive(lp, now))
914 lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
919 * \param msg The message to be sent.
920 * \param do_send True if lnet_ni_send() should be called in this function.
921 * lnet_send() is going to lnet_net_unlock immediately after this, so
922 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
924 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
925 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
926 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
927 * \retval -ECANCELED If the MD of the message has been unlinked.
930 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
932 struct lnet_peer_ni *lp = msg->msg_txpeer;
933 struct lnet_ni *ni = msg->msg_txni;
934 int cpt = msg->msg_tx_cpt;
935 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
937 /* non-lnet_send() callers have checked before */
938 LASSERT(!do_send || msg->msg_tx_delayed);
939 LASSERT(!msg->msg_receiving);
940 LASSERT(msg->msg_tx_committed);
942 /* NB 'lp' is always the next hop */
943 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
944 lnet_peer_alive_locked(ni, lp, msg) == 0) {
945 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
946 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
948 lnet_net_unlock(cpt);
950 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
952 LNET_STATS_TYPE_DROP);
954 lnet_incr_stats(&msg->msg_txni->ni_stats,
956 LNET_STATS_TYPE_DROP);
958 CNETERR("Dropping message for %s: peer not alive\n",
959 libcfs_id2str(msg->msg_target));
960 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_DROPPED;
962 lnet_finalize(msg, -EHOSTUNREACH);
965 return -EHOSTUNREACH;
968 if (msg->msg_md != NULL &&
969 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
970 lnet_net_unlock(cpt);
972 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
973 "called on the MD/ME.\n",
974 libcfs_id2str(msg->msg_target));
976 msg->msg_no_resend = true;
977 lnet_finalize(msg, -ECANCELED);
984 if (!msg->msg_peertxcredit) {
985 spin_lock(&lp->lpni_lock);
986 LASSERT((lp->lpni_txcredits < 0) ==
987 !list_empty(&lp->lpni_txq));
989 msg->msg_peertxcredit = 1;
990 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
991 lp->lpni_txcredits--;
993 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
994 lp->lpni_mintxcredits = lp->lpni_txcredits;
996 if (lp->lpni_txcredits < 0) {
997 msg->msg_tx_delayed = 1;
998 list_add_tail(&msg->msg_list, &lp->lpni_txq);
999 spin_unlock(&lp->lpni_lock);
1000 return LNET_CREDIT_WAIT;
1002 spin_unlock(&lp->lpni_lock);
1005 if (!msg->msg_txcredit) {
1006 LASSERT((tq->tq_credits < 0) ==
1007 !list_empty(&tq->tq_delayed));
1009 msg->msg_txcredit = 1;
1011 atomic_dec(&ni->ni_tx_credits);
1013 if (tq->tq_credits < tq->tq_credits_min)
1014 tq->tq_credits_min = tq->tq_credits;
1016 if (tq->tq_credits < 0) {
1017 msg->msg_tx_delayed = 1;
1018 list_add_tail(&msg->msg_list, &tq->tq_delayed);
1019 return LNET_CREDIT_WAIT;
1023 /* unset the tx_delay flag as we're going to send it now */
1024 msg->msg_tx_delayed = 0;
1027 lnet_net_unlock(cpt);
1028 lnet_ni_send(ni, msg);
1031 return LNET_CREDIT_OK;
1035 static struct lnet_rtrbufpool *
1036 lnet_msg2bufpool(struct lnet_msg *msg)
1038 struct lnet_rtrbufpool *rbp;
1041 LASSERT(msg->msg_rx_committed);
1043 cpt = msg->msg_rx_cpt;
1044 rbp = &the_lnet.ln_rtrpools[cpt][0];
1046 LASSERT(msg->msg_len <= LNET_MTU);
1047 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
1049 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
1056 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
1058 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
1059 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
1060 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
1061 * received or OK to receive */
1062 struct lnet_peer_ni *lp = msg->msg_rxpeer;
1063 struct lnet_rtrbufpool *rbp;
1064 struct lnet_rtrbuf *rb;
1066 LASSERT (msg->msg_iov == NULL);
1067 LASSERT (msg->msg_kiov == NULL);
1068 LASSERT (msg->msg_niov == 0);
1069 LASSERT (msg->msg_routing);
1070 LASSERT (msg->msg_receiving);
1071 LASSERT (!msg->msg_sending);
1073 /* non-lnet_parse callers only receive delayed messages */
1074 LASSERT(!do_recv || msg->msg_rx_delayed);
1076 if (!msg->msg_peerrtrcredit) {
1077 spin_lock(&lp->lpni_lock);
1078 LASSERT((lp->lpni_rtrcredits < 0) ==
1079 !list_empty(&lp->lpni_rtrq));
1081 msg->msg_peerrtrcredit = 1;
1082 lp->lpni_rtrcredits--;
1083 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
1084 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
1086 if (lp->lpni_rtrcredits < 0) {
1087 /* must have checked eager_recv before here */
1088 LASSERT(msg->msg_rx_ready_delay);
1089 msg->msg_rx_delayed = 1;
1090 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
1091 spin_unlock(&lp->lpni_lock);
1092 return LNET_CREDIT_WAIT;
1094 spin_unlock(&lp->lpni_lock);
1097 rbp = lnet_msg2bufpool(msg);
1099 if (!msg->msg_rtrcredit) {
1100 msg->msg_rtrcredit = 1;
1102 if (rbp->rbp_credits < rbp->rbp_mincredits)
1103 rbp->rbp_mincredits = rbp->rbp_credits;
1105 if (rbp->rbp_credits < 0) {
1106 /* must have checked eager_recv before here */
1107 LASSERT(msg->msg_rx_ready_delay);
1108 msg->msg_rx_delayed = 1;
1109 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1110 return LNET_CREDIT_WAIT;
1114 LASSERT(!list_empty(&rbp->rbp_bufs));
1115 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1116 list_del(&rb->rb_list);
1118 msg->msg_niov = rbp->rbp_npages;
1119 msg->msg_kiov = &rb->rb_kiov[0];
1121 /* unset the msg-rx_delayed flag since we're receiving the message */
1122 msg->msg_rx_delayed = 0;
1125 int cpt = msg->msg_rx_cpt;
1127 lnet_net_unlock(cpt);
1128 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1129 0, msg->msg_len, msg->msg_len);
1132 return LNET_CREDIT_OK;
1136 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1138 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1139 struct lnet_ni *txni = msg->msg_txni;
1140 struct lnet_msg *msg2;
1142 if (msg->msg_txcredit) {
1143 struct lnet_ni *ni = msg->msg_txni;
1144 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1146 /* give back NI txcredits */
1147 msg->msg_txcredit = 0;
1149 LASSERT((tq->tq_credits < 0) ==
1150 !list_empty(&tq->tq_delayed));
1153 atomic_inc(&ni->ni_tx_credits);
1154 if (tq->tq_credits <= 0) {
1155 msg2 = list_entry(tq->tq_delayed.next,
1156 struct lnet_msg, msg_list);
1157 list_del(&msg2->msg_list);
1159 LASSERT(msg2->msg_txni == ni);
1160 LASSERT(msg2->msg_tx_delayed);
1161 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1163 (void) lnet_post_send_locked(msg2, 1);
1167 if (msg->msg_peertxcredit) {
1168 /* give back peer txcredits */
1169 msg->msg_peertxcredit = 0;
1171 spin_lock(&txpeer->lpni_lock);
1172 LASSERT((txpeer->lpni_txcredits < 0) ==
1173 !list_empty(&txpeer->lpni_txq));
1175 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1176 LASSERT(txpeer->lpni_txqnob >= 0);
1178 txpeer->lpni_txcredits++;
1179 if (txpeer->lpni_txcredits <= 0) {
1182 msg2 = list_entry(txpeer->lpni_txq.next,
1183 struct lnet_msg, msg_list);
1184 list_del(&msg2->msg_list);
1185 spin_unlock(&txpeer->lpni_lock);
1187 LASSERT(msg2->msg_txpeer == txpeer);
1188 LASSERT(msg2->msg_tx_delayed);
1190 msg2_cpt = msg2->msg_tx_cpt;
1193 * The msg_cpt can be different from the msg2_cpt
1194 * so we need to make sure we lock the correct cpt
1196 * Once we call lnet_post_send_locked() it is no
1197 * longer safe to access msg2, since it could've
1198 * been freed by lnet_finalize(), but we still
1199 * need to relock the correct cpt, so we cache the
1200 * msg2_cpt for the purpose of the check that
1201 * follows the call to lnet_pose_send_locked().
1203 if (msg2_cpt != msg->msg_tx_cpt) {
1204 lnet_net_unlock(msg->msg_tx_cpt);
1205 lnet_net_lock(msg2_cpt);
1207 (void) lnet_post_send_locked(msg2, 1);
1208 if (msg2_cpt != msg->msg_tx_cpt) {
1209 lnet_net_unlock(msg2_cpt);
1210 lnet_net_lock(msg->msg_tx_cpt);
1213 spin_unlock(&txpeer->lpni_lock);
1218 msg->msg_txni = NULL;
1219 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1222 if (txpeer != NULL) {
1223 msg->msg_txpeer = NULL;
1224 lnet_peer_ni_decref_locked(txpeer);
1229 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1231 struct lnet_msg *msg;
1233 if (list_empty(&rbp->rbp_msgs))
1235 msg = list_entry(rbp->rbp_msgs.next,
1236 struct lnet_msg, msg_list);
1237 list_del(&msg->msg_list);
1239 (void)lnet_post_routed_recv_locked(msg, 1);
1243 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1245 struct lnet_msg *msg;
1246 struct lnet_msg *tmp;
1248 lnet_net_unlock(cpt);
1250 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1251 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1252 0, 0, 0, msg->msg_hdr.payload_length);
1253 list_del_init(&msg->msg_list);
1254 msg->msg_no_resend = true;
1255 lnet_finalize(msg, -ECANCELED);
1262 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1264 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1265 struct lnet_ni *rxni = msg->msg_rxni;
1266 struct lnet_msg *msg2;
1268 if (msg->msg_rtrcredit) {
1269 /* give back global router credits */
1270 struct lnet_rtrbuf *rb;
1271 struct lnet_rtrbufpool *rbp;
1273 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1274 * there until it gets one allocated, or aborts the wait
1276 LASSERT(msg->msg_kiov != NULL);
1278 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1281 msg->msg_kiov = NULL;
1282 msg->msg_rtrcredit = 0;
1284 LASSERT(rbp == lnet_msg2bufpool(msg));
1286 LASSERT((rbp->rbp_credits > 0) ==
1287 !list_empty(&rbp->rbp_bufs));
1289 /* If routing is now turned off, we just drop this buffer and
1290 * don't bother trying to return credits. */
1291 if (!the_lnet.ln_routing) {
1292 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1296 /* It is possible that a user has lowered the desired number of
1297 * buffers in this pool. Make sure we never put back
1298 * more buffers than the stated number. */
1299 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1300 /* Discard this buffer so we don't have too
1302 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1303 rbp->rbp_nbuffers--;
1305 list_add(&rb->rb_list, &rbp->rbp_bufs);
1307 if (rbp->rbp_credits <= 0)
1308 lnet_schedule_blocked_locked(rbp);
1313 if (msg->msg_peerrtrcredit) {
1314 /* give back peer router credits */
1315 msg->msg_peerrtrcredit = 0;
1317 spin_lock(&rxpeer->lpni_lock);
1318 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1319 !list_empty(&rxpeer->lpni_rtrq));
1321 rxpeer->lpni_rtrcredits++;
1323 /* drop all messages which are queued to be routed on that
1325 if (!the_lnet.ln_routing) {
1326 struct list_head drop;
1327 INIT_LIST_HEAD(&drop);
1328 list_splice_init(&rxpeer->lpni_rtrq, &drop);
1329 spin_unlock(&rxpeer->lpni_lock);
1330 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1331 } else if (rxpeer->lpni_rtrcredits <= 0) {
1332 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1333 struct lnet_msg, msg_list);
1334 list_del(&msg2->msg_list);
1335 spin_unlock(&rxpeer->lpni_lock);
1336 (void) lnet_post_routed_recv_locked(msg2, 1);
1338 spin_unlock(&rxpeer->lpni_lock);
1342 msg->msg_rxni = NULL;
1343 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1345 if (rxpeer != NULL) {
1346 msg->msg_rxpeer = NULL;
1347 lnet_peer_ni_decref_locked(rxpeer);
1352 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1354 if (p1->lpni_txqnob < p2->lpni_txqnob)
1357 if (p1->lpni_txqnob > p2->lpni_txqnob)
1360 if (p1->lpni_txcredits > p2->lpni_txcredits)
1363 if (p1->lpni_txcredits < p2->lpni_txcredits)
1370 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1372 struct lnet_peer_ni *p1 = r1->lr_gateway;
1373 struct lnet_peer_ni *p2 = r2->lr_gateway;
1374 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1375 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1378 if (r1->lr_priority < r2->lr_priority)
1381 if (r1->lr_priority > r2->lr_priority)
1384 if (r1_hops < r2_hops)
1387 if (r1_hops > r2_hops)
1390 rc = lnet_compare_peers(p1, p2);
1394 if (r1->lr_seq - r2->lr_seq <= 0)
1400 static struct lnet_peer_ni *
1401 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1404 struct lnet_remotenet *rnet;
1405 struct lnet_route *route;
1406 struct lnet_route *best_route;
1407 struct lnet_route *last_route;
1408 struct lnet_peer_ni *lpni_best;
1409 struct lnet_peer_ni *lp;
1412 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1413 * rtr_nid nid, otherwise find the best gateway I can use */
1415 rnet = lnet_find_rnet_locked(remote_net);
1420 best_route = last_route = NULL;
1421 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1422 lp = route->lr_gateway;
1424 if (!lnet_is_route_alive(route))
1427 if (net != NULL && lp->lpni_net != net)
1430 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1433 if (lpni_best == NULL) {
1434 best_route = last_route = route;
1439 /* no protection on below fields, but it's harmless */
1440 if (last_route->lr_seq - route->lr_seq < 0)
1443 rc = lnet_compare_routes(route, best_route);
1451 /* set sequence number on the best router to the latest sequence + 1
1452 * so we can round-robin all routers, it's race and inaccurate but
1453 * harmless and functional */
1454 if (best_route != NULL)
1455 best_route->lr_seq = last_route->lr_seq + 1;
1459 static struct lnet_ni *
1460 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1461 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1464 struct lnet_ni *ni = NULL;
1465 unsigned int shortest_distance;
1470 * If there is no peer_ni that we can send to on this network,
1471 * then there is no point in looking for a new best_ni here.
1473 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1476 if (best_ni == NULL) {
1477 shortest_distance = UINT_MAX;
1478 best_credits = INT_MIN;
1481 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1482 best_ni->ni_dev_cpt);
1483 best_credits = atomic_read(&best_ni->ni_tx_credits);
1484 best_healthv = atomic_read(&best_ni->ni_healthv);
1487 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1488 unsigned int distance;
1493 ni_credits = atomic_read(&ni->ni_tx_credits);
1494 ni_healthv = atomic_read(&ni->ni_healthv);
1495 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1498 * calculate the distance from the CPT on which
1499 * the message memory is allocated to the CPT of
1500 * the NI's physical device
1502 distance = cfs_cpt_distance(lnet_cpt_table(),
1506 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1507 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1508 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1509 : "not seleced", best_credits, shortest_distance,
1510 (best_ni) ? best_ni->ni_seq : 0);
1513 * All distances smaller than the NUMA range
1514 * are treated equally.
1516 if (distance < lnet_numa_range)
1517 distance = lnet_numa_range;
1520 * Select on health, shorter distance, available
1521 * credits, then round-robin.
1525 } else if (ni_healthv < best_healthv) {
1527 } else if (ni_healthv > best_healthv) {
1528 best_healthv = ni_healthv;
1530 * If we're going to prefer this ni because it's
1531 * the healthiest, then we should set the
1532 * shortest_distance in the algorithm in case
1533 * there are multiple NIs with the same health but
1534 * different distances.
1536 if (distance < shortest_distance)
1537 shortest_distance = distance;
1538 } else if (distance > shortest_distance) {
1540 } else if (distance < shortest_distance) {
1541 shortest_distance = distance;
1542 } else if (ni_credits < best_credits) {
1544 } else if (ni_credits == best_credits) {
1545 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1549 best_credits = ni_credits;
1552 CDEBUG(D_NET, "selected best_ni %s\n",
1553 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1559 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1560 * because such traffic is required to perform discovery. We therefore
1561 * exclude all GET and PUT on that portal. We also exclude all ACK and
1562 * REPLY traffic, but that is because the portal is not tracked in the
1563 * message structure for these message types. We could restrict this
1564 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1567 lnet_msg_discovery(struct lnet_msg *msg)
1569 if (msg->msg_type == LNET_MSG_PUT) {
1570 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1572 } else if (msg->msg_type == LNET_MSG_GET) {
1573 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1579 #define SRC_SPEC 0x0001
1580 #define SRC_ANY 0x0002
1581 #define LOCAL_DST 0x0004
1582 #define REMOTE_DST 0x0008
1583 #define MR_DST 0x0010
1584 #define NMR_DST 0x0020
1585 #define SND_RESP 0x0040
1587 /* The following to defines are used for return codes */
1588 #define REPEAT_SEND 0x1000
1589 #define PASS_THROUGH 0x2000
1591 /* The different cases lnet_select pathway needs to handle */
1592 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1593 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1594 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1595 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1596 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1597 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1598 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1599 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1602 lnet_handle_send(struct lnet_send_data *sd)
1604 struct lnet_ni *best_ni = sd->sd_best_ni;
1605 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1606 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1607 struct lnet_msg *msg = sd->sd_msg;
1609 __u32 send_case = sd->sd_send_case;
1611 __u32 routing = send_case & REMOTE_DST;
1614 * Increment sequence number of the selected peer so that we
1615 * pick the next one in Round Robin.
1617 best_lpni->lpni_seq++;
1620 * grab a reference on the peer_ni so it sticks around even if
1621 * we need to drop and relock the lnet_net_lock below.
1623 lnet_peer_ni_addref_locked(best_lpni);
1626 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1627 * message. This ensures that we get a CPT that is correct for
1628 * the NI when the NI has been restricted to a subset of all CPTs.
1629 * If the selected CPT differs from the one currently locked, we
1630 * must unlock and relock the lnet_net_lock(), and then check whether
1631 * the configuration has changed. We don't have a hold on the best_ni
1632 * yet, and it may have vanished.
1634 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1635 if (sd->sd_cpt != cpt2) {
1636 __u32 seq = lnet_get_dlc_seq_locked();
1637 lnet_net_unlock(sd->sd_cpt);
1639 lnet_net_lock(sd->sd_cpt);
1640 if (seq != lnet_get_dlc_seq_locked()) {
1641 lnet_peer_ni_decref_locked(best_lpni);
1647 * store the best_lpni in the message right away to avoid having
1648 * to do the same operation under different conditions
1650 msg->msg_txpeer = best_lpni;
1651 msg->msg_txni = best_ni;
1654 * grab a reference for the best_ni since now it's in use in this
1655 * send. The reference will be dropped in lnet_finalize()
1657 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1660 * Always set the target.nid to the best peer picked. Either the
1661 * NID will be one of the peer NIDs selected, or the same NID as
1662 * what was originally set in the target or it will be the NID of
1663 * a router if this message should be routed
1665 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1668 * lnet_msg_commit assigns the correct cpt to the message, which
1669 * is used to decrement the correct refcount on the ni when it's
1670 * time to return the credits
1672 lnet_msg_commit(msg, sd->sd_cpt);
1675 * If we are routing the message then we keep the src_nid that was
1676 * set by the originator. If we are not routing then we are the
1677 * originator and set it here.
1679 if (!msg->msg_routing)
1680 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1683 msg->msg_target_is_router = 1;
1684 msg->msg_target.pid = LNET_PID_LUSTRE;
1686 * since we're routing we want to ensure that the
1687 * msg_hdr.dest_nid is set to the final destination. When
1688 * the router receives this message it knows how to route
1691 * final_dst_lpni is set at the beginning of the
1692 * lnet_select_pathway() function and is never changed.
1693 * It's safe to use it here.
1695 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1698 * if we're not routing set the dest_nid to the best peer
1699 * ni NID that we picked earlier in the algorithm.
1701 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1704 rc = lnet_post_send_locked(msg, 0);
1707 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1708 libcfs_nid2str(msg->msg_hdr.src_nid),
1709 libcfs_nid2str(msg->msg_txni->ni_nid),
1710 libcfs_nid2str(sd->sd_src_nid),
1711 libcfs_nid2str(msg->msg_hdr.dest_nid),
1712 libcfs_nid2str(sd->sd_dst_nid),
1713 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1714 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1719 static struct lnet_peer_ni *
1720 lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer,
1721 struct lnet_peer_net *peer_net)
1724 * Look at the peer NIs for the destination peer that connect
1725 * to the chosen net. If a peer_ni is preferred when using the
1726 * best_ni to communicate, we use that one. If there is no
1727 * preferred peer_ni, or there are multiple preferred peer_ni,
1728 * the available transmit credits are used. If the transmit
1729 * credits are equal, we round-robin over the peer_ni.
1731 struct lnet_peer_ni *lpni = NULL;
1732 struct lnet_peer_ni *best_lpni = NULL;
1733 struct lnet_ni *best_ni = sd->sd_best_ni;
1734 lnet_nid_t dst_nid = sd->sd_dst_nid;
1735 int best_lpni_credits = INT_MIN;
1736 bool preferred = false;
1738 int best_lpni_healthv = 0;
1741 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1743 * if the best_ni we've chosen aleady has this lpni
1744 * preferred, then let's use it
1746 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1749 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1751 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1752 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1755 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1756 libcfs_nid2str(lpni->lpni_nid),
1757 lpni->lpni_txcredits, best_lpni_credits,
1758 lpni->lpni_seq, best_lpni->lpni_seq);
1760 /* pick the healthiest peer ni */
1761 if (lpni_healthv < best_lpni_healthv) {
1763 } else if (lpni_healthv > best_lpni_healthv) {
1764 best_lpni_healthv = lpni_healthv;
1765 /* if this is a preferred peer use it */
1766 } else if (!preferred && ni_is_pref) {
1768 } else if (preferred && !ni_is_pref) {
1770 * this is not the preferred peer so let's ignore
1774 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1776 * We already have a peer that has more credits
1777 * available than this one. No need to consider
1778 * this peer further.
1781 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1783 * The best peer found so far and the current peer
1784 * have the same number of available credits let's
1785 * make sure to select between them using Round
1789 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1795 best_lpni_credits = lpni->lpni_txcredits;
1798 /* if we still can't find a peer ni then we can't reach it */
1800 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1801 LNET_NIDNET(dst_nid);
1802 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1803 libcfs_net2str(net_id));
1807 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1808 libcfs_nid2str(best_lpni->lpni_nid));
1814 * Prerequisite: the best_ni should already be set in the sd
1816 static inline struct lnet_peer_ni *
1817 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1820 struct lnet_peer_net *peer_net;
1823 * The gateway is Multi-Rail capable so now we must select the
1826 peer_net = lnet_peer_get_net_locked(peer, net_id);
1829 CERROR("gateway peer %s has no NI on net %s\n",
1830 libcfs_nid2str(peer->lp_primary_nid),
1831 libcfs_net2str(net_id));
1835 return lnet_select_peer_ni(sd, peer, peer_net);
1839 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1841 if (sd->sd_send_case & NMR_DST &&
1842 sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1843 sd->sd_msg->msg_type != LNET_MSG_ACK &&
1844 sd->sd_best_lpni->lpni_pref_nnids == 0) {
1845 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1846 libcfs_nid2str(sd->sd_best_ni->ni_nid),
1847 libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1848 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1849 sd->sd_best_ni->ni_nid);
1858 * use the source and destination NIDs as the pathway
1861 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1863 /* the destination lpni is set before we get here. */
1866 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1867 if (!sd->sd_best_ni) {
1868 CERROR("Can't send to %s: src %s is not a "
1869 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1870 libcfs_nid2str(sd->sd_src_nid));
1875 * the preferred NID will only be set for NMR peers
1877 lnet_set_non_mr_pref_nid(sd);
1879 return lnet_handle_send(sd);
1887 * Run the selection algorithm on the peer NIs unless we're sending
1888 * a response, in this case just send to the destination
1891 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1893 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1894 if (!sd->sd_best_ni) {
1895 CERROR("Can't send to %s: src %s is not a "
1896 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1897 libcfs_nid2str(sd->sd_src_nid));
1902 * only run the selection algorithm to pick the peer_ni if we're
1903 * sending a GET or a PUT. Responses are sent to the same
1904 * destination NID provided.
1906 if (!(sd->sd_send_case & SND_RESP)) {
1908 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1909 sd->sd_best_ni->ni_net->net_id);
1912 if (sd->sd_best_lpni)
1913 return lnet_handle_send(sd);
1915 CERROR("can't send to %s. no NI on %s\n",
1916 libcfs_nid2str(sd->sd_dst_nid),
1917 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1919 return -EHOSTUNREACH;
1923 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1924 struct lnet_peer *peer,
1925 struct lnet_peer_net *peer_net,
1929 struct lnet_net *local_net;
1930 struct lnet_ni *best_ni;
1932 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1937 * Iterate through the NIs in this local Net and select
1938 * the NI to send from. The selection is determined by
1939 * these 3 criterion in the following priority:
1941 * 2. NI available credits
1944 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1945 peer, peer_net, cpt);
1947 if (incr_seq && best_ni)
1954 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1956 struct lnet_peer_ni **gw_lpni,
1957 struct lnet_peer **gw_peer)
1959 struct lnet_peer_ni *gw;
1960 lnet_nid_t src_nid = sd->sd_src_nid;
1962 gw = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
1965 CERROR("no route to %s from %s\n",
1966 libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1967 return -EHOSTUNREACH;
1970 /* get the peer of the gw_ni */
1971 LASSERT(gw->lpni_peer_net);
1972 LASSERT(gw->lpni_peer_net->lpn_peer);
1974 *gw_peer = gw->lpni_peer_net->lpn_peer;
1976 if (!sd->sd_best_ni)
1977 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, *gw_peer,
1982 if (!sd->sd_best_ni) {
1983 CERROR("Internal Error. Expected local ni on %s "
1984 "but non found :%s\n",
1985 libcfs_net2str(gw->lpni_peer_net->lpn_net_id),
1986 libcfs_nid2str(sd->sd_src_nid));
1991 * if gw is MR let's find its best peer_ni
1993 if (lnet_peer_is_multi_rail(*gw_peer)) {
1994 gw = lnet_find_best_lpni_on_net(sd, *gw_peer,
1995 sd->sd_best_ni->ni_net->net_id);
1997 * We've already verified that the gw has an NI on that
1998 * desired net, but we're not finding it. Something is
2002 CERROR("Internal Error. Route expected to %s from %s\n",
2003 libcfs_nid2str(dst_nid),
2004 libcfs_nid2str(src_nid));
2019 * Remote destination
2020 * Non-MR destination
2024 * Remote destination
2027 * The handling of these two cases is similar. Even though the destination
2028 * can be MR or non-MR, we'll deal directly with the router.
2031 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2034 struct lnet_peer_ni *gw_lpni = NULL;
2035 struct lnet_peer *gw_peer = NULL;
2038 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2039 if (!sd->sd_best_ni) {
2040 CERROR("Can't send to %s: src %s is not a "
2041 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2042 libcfs_nid2str(sd->sd_src_nid));
2046 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2051 if (sd->sd_send_case & NMR_DST)
2053 * since the final destination is non-MR let's set its preferred
2054 * NID before we send
2056 lnet_set_non_mr_pref_nid(sd);
2059 * We're going to send to the gw found so let's set its
2062 sd->sd_peer = gw_peer;
2063 sd->sd_best_lpni = gw_lpni;
2065 return lnet_handle_send(sd);
2069 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
2071 struct lnet_peer_net *peer_net = NULL;
2072 struct lnet_ni *best_ni = NULL;
2075 * The peer can have multiple interfaces, some of them can be on
2076 * the local network and others on a routed network. We should
2077 * prefer the local network. However if the local network is not
2078 * available then we need to try the routed network
2081 /* go through all the peer nets and find the best_ni */
2082 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2084 * The peer's list of nets can contain non-local nets. We
2085 * want to only examine the local ones.
2087 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2089 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2090 peer_net, md_cpt, false);
2094 /* increment sequence number so we can round robin */
2100 static struct lnet_ni *
2101 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2103 struct lnet_ni *best_ni = NULL;
2104 struct lnet_peer_net *peer_net;
2105 struct lnet_peer *peer = sd->sd_peer;
2106 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2107 struct lnet_peer_ni *lpni;
2108 int cpt = sd->sd_cpt;
2111 * We must use a consistent source address when sending to a
2112 * non-MR peer. However, a non-MR peer can have multiple NIDs
2113 * on multiple networks, and we may even need to talk to this
2114 * peer on multiple networks -- certain types of
2115 * load-balancing configuration do this.
2117 * So we need to pick the NI the peer prefers for this
2118 * particular network.
2121 /* Get the target peer_ni */
2122 peer_net = lnet_peer_get_net_locked(peer,
2123 LNET_NIDNET(best_lpni->lpni_nid));
2124 LASSERT(peer_net != NULL);
2125 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2127 if (lpni->lpni_pref_nnids == 0)
2129 LASSERT(lpni->lpni_pref_nnids == 1);
2130 best_ni = lnet_nid2ni_locked(
2131 lpni->lpni_pref.nid, cpt);
2138 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2140 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2142 struct lnet_ni *best_ni = NULL;
2143 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2146 * We must use a consistent source address when sending to a
2147 * non-MR peer. However, a non-MR peer can have multiple NIDs
2148 * on multiple networks, and we may even need to talk to this
2149 * peer on multiple networks -- certain types of
2150 * load-balancing configuration do this.
2152 * So we need to pick the NI the peer prefers for this
2153 * particular network.
2156 best_ni = lnet_find_existing_preferred_best_ni(sd);
2158 /* if best_ni is still not set just pick one */
2161 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2162 sd->sd_best_lpni->lpni_peer_net,
2163 sd->sd_md_cpt, true);
2164 /* If there is no best_ni we don't have a route */
2166 CERROR("no path to %s from net %s\n",
2167 libcfs_nid2str(best_lpni->lpni_nid),
2168 libcfs_net2str(best_lpni->lpni_net->net_id));
2169 return -EHOSTUNREACH;
2173 sd->sd_best_ni = best_ni;
2175 /* Set preferred NI if necessary. */
2176 lnet_set_non_mr_pref_nid(sd);
2183 * Source not specified
2187 * always use the same source NID for NMR peers
2188 * If we've talked to that peer before then we already have a preferred
2189 * source NI associated with it. Otherwise, we select a preferred local NI
2190 * and store it in the peer
2193 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2197 /* sd->sd_best_lpni is already set to the final destination */
2200 * At this point we should've created the peer ni and peer. If we
2201 * can't find it, then something went wrong. Instead of assert
2202 * output a relevant message and fail the send
2204 if (!sd->sd_best_lpni) {
2205 CERROR("Internal fault. Unable to send msg %s to %s. "
2207 lnet_msgtyp2str(sd->sd_msg->msg_type),
2208 libcfs_nid2str(sd->sd_dst_nid));
2212 rc = lnet_select_preferred_best_ni(sd);
2214 rc = lnet_handle_send(sd);
2220 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2223 * NOTE we've already handled the remote peer case. So we only
2224 * need to worry about the local case here.
2226 * if we're sending a response, ACK or reply, we need to send it
2227 * to the destination NID given to us. At this point we already
2228 * have the peer_ni we're suppose to send to, so just find the
2229 * best_ni on the peer net and use that. Since we're sending to an
2230 * MR peer then we can just run the selection algorithm on our
2231 * local NIs and pick the best one.
2233 if (sd->sd_send_case & SND_RESP) {
2235 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2236 sd->sd_best_lpni->lpni_peer_net,
2237 sd->sd_md_cpt, true);
2239 if (!sd->sd_best_ni) {
2241 * We're not going to deal with not able to send
2242 * a response to the provided final destination
2244 CERROR("Can't send response to %s. "
2245 "No local NI available\n",
2246 libcfs_nid2str(sd->sd_dst_nid));
2247 return -EHOSTUNREACH;
2250 return lnet_handle_send(sd);
2254 * If we get here that means we're sending a fresh request, PUT or
2255 * GET, so we need to run our standard selection algorithm.
2256 * First find the best local interface that's on any of the peer's
2259 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2261 if (sd->sd_best_ni) {
2263 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2264 sd->sd_best_ni->ni_net->net_id);
2267 * if we're successful in selecting a peer_ni on the local
2268 * network, then send to it. Otherwise fall through and
2269 * try and see if we can reach it over another routed
2272 if (sd->sd_best_lpni) {
2274 * in case we initially started with a routed
2275 * destination, let's reset to local
2277 sd->sd_send_case &= ~REMOTE_DST;
2278 sd->sd_send_case |= LOCAL_DST;
2279 return lnet_handle_send(sd);
2282 CERROR("Internal Error. Expected to have a best_lpni: "
2284 libcfs_nid2str(sd->sd_src_nid),
2285 libcfs_nid2str(sd->sd_dst_nid));
2291 * Peer doesn't have a local network. Let's see if there is
2292 * a remote network we can reach it on.
2294 return PASS_THROUGH;
2299 * Source NID not specified
2304 * Source NID not speified
2305 * Remote destination
2308 * In both of these cases if we're sending a response, ACK or REPLY, then
2309 * we need to send to the destination NID provided.
2311 * In the remote case let's deal with MR routers.
2316 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2319 struct lnet_peer *gw_peer = NULL;
2320 struct lnet_peer_ni *gw_lpni = NULL;
2323 * handle sending a response to a remote peer here so we don't
2324 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2326 if (sd->sd_send_case & REMOTE_DST &&
2327 sd->sd_send_case & SND_RESP) {
2328 struct lnet_peer_ni *gw;
2329 struct lnet_peer *gw_peer;
2331 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2334 CERROR("Can't send response to %s. "
2335 "No route available\n",
2336 libcfs_nid2str(sd->sd_dst_nid));
2337 return -EHOSTUNREACH;
2340 sd->sd_best_lpni = gw;
2341 sd->sd_peer = gw_peer;
2343 return lnet_handle_send(sd);
2347 * Even though the NID for the peer might not be on a local network,
2348 * since the peer is MR there could be other interfaces on the
2349 * local network. In that case we'd still like to prefer the local
2350 * network over the routed network. If we're unable to do that
2351 * then we select the best router among the different routed networks,
2352 * and if the router is MR then we can deal with it as such.
2354 rc = lnet_handle_any_mr_dsta(sd);
2355 if (rc != PASS_THROUGH)
2359 * TODO; One possible enhancement is to run the selection
2360 * algorithm on the peer. However for remote peers the credits are
2361 * not decremented, so we'll be basically going over the peer NIs
2362 * in round robin. An MR router will run the selection algorithm
2363 * on the next-hop interfaces.
2365 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2370 sd->sd_send_case &= ~LOCAL_DST;
2371 sd->sd_send_case |= REMOTE_DST;
2373 sd->sd_peer = gw_peer;
2374 sd->sd_best_lpni = gw_lpni;
2376 return lnet_handle_send(sd);
2380 * Source not specified
2381 * Remote destination
2384 * Must send to the specified peer NID using the same source NID that
2385 * we've used before. If it's the first time to talk to that peer then
2386 * find the source NI and assign it as preferred to that peer
2389 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2392 struct lnet_peer_ni *gw_lpni = NULL;
2393 struct lnet_peer *gw_peer = NULL;
2396 * Let's set if we have a preferred NI to talk to this NMR peer
2398 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2401 * find the router and that'll find the best NI if we didn't find
2404 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2410 * set the best_ni we've chosen as the preferred one for
2413 lnet_set_non_mr_pref_nid(sd);
2415 /* we'll be sending to the gw */
2416 sd->sd_best_lpni = gw_lpni;
2417 sd->sd_peer = gw_peer;
2419 return lnet_handle_send(sd);
2423 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2426 * turn off the SND_RESP bit.
2427 * It will be checked in the case handling
2429 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2431 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2432 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2433 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2434 (send_case & MR_DST) ? "MR: " : "NMR: ",
2435 libcfs_nid2str(sd->sd_dst_nid),
2436 (send_case & LOCAL_DST) ? "local" : "routed");
2438 switch (send_case) {
2440 * For all cases where the source is specified, we should always
2441 * use the destination NID, whether it's an MR destination or not,
2442 * since we're continuing a series of related messages for the
2445 case SRC_SPEC_LOCAL_NMR_DST:
2446 return lnet_handle_spec_local_nmr_dst(sd);
2447 case SRC_SPEC_LOCAL_MR_DST:
2448 return lnet_handle_spec_local_mr_dst(sd);
2449 case SRC_SPEC_ROUTER_NMR_DST:
2450 case SRC_SPEC_ROUTER_MR_DST:
2451 return lnet_handle_spec_router_dst(sd);
2452 case SRC_ANY_LOCAL_NMR_DST:
2453 return lnet_handle_any_local_nmr_dst(sd);
2454 case SRC_ANY_LOCAL_MR_DST:
2455 case SRC_ANY_ROUTER_MR_DST:
2456 return lnet_handle_any_mr_dst(sd);
2457 case SRC_ANY_ROUTER_NMR_DST:
2458 return lnet_handle_any_router_nmr_dst(sd);
2460 CERROR("Unknown send case\n");
2466 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2467 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2469 struct lnet_peer_ni *lpni;
2470 struct lnet_peer *peer;
2471 struct lnet_send_data send_data;
2474 __u32 send_case = 0;
2476 memset(&send_data, 0, sizeof(send_data));
2479 * get an initial CPT to use for locking. The idea here is not to
2480 * serialize the calls to select_pathway, so that as many
2481 * operations can run concurrently as possible. To do that we use
2482 * the CPT where this call is being executed. Later on when we
2483 * determine the CPT to use in lnet_message_commit, we switch the
2484 * lock and check if there was any configuration change. If none,
2485 * then we proceed, if there is, then we restart the operation.
2487 cpt = lnet_net_lock_current();
2489 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2490 if (md_cpt == CFS_CPT_ANY)
2496 * If we're being asked to send to the loopback interface, there
2497 * is no need to go through any selection. We can just shortcut
2498 * the entire process and send over lolnd
2500 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2501 /* No send credit hassles with LOLND */
2502 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
2503 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
2504 if (!msg->msg_routing)
2505 msg->msg_hdr.src_nid =
2506 cpu_to_le64(the_lnet.ln_loni->ni_nid);
2507 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
2508 lnet_msg_commit(msg, cpt);
2509 msg->msg_txni = the_lnet.ln_loni;
2510 lnet_net_unlock(cpt);
2512 return LNET_CREDIT_OK;
2516 * find an existing peer_ni, or create one and mark it as having been
2517 * created due to network traffic. This call will create the
2518 * peer->peer_net->peer_ni tree.
2520 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2522 lnet_net_unlock(cpt);
2523 return PTR_ERR(lpni);
2527 * Cache the original src_nid. If we need to resend the message
2528 * then we'll need to know whether the src_nid was originally
2529 * specified for this message. If it was originally specified,
2530 * then we need to keep using the same src_nid since it's
2531 * continuing the same sequence of messages.
2533 msg->msg_src_nid_param = src_nid;
2536 * Now that we have a peer_ni, check if we want to discover
2537 * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2538 * trigger discovery.
2540 peer = lpni->lpni_peer_net->lpn_peer;
2541 if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
2542 lnet_nid_t primary_nid;
2543 rc = lnet_discover_peer_locked(lpni, cpt, false);
2545 lnet_peer_ni_decref_locked(lpni);
2546 lnet_net_unlock(cpt);
2549 /* The peer may have changed. */
2550 peer = lpni->lpni_peer_net->lpn_peer;
2551 /* queue message and return */
2552 msg->msg_rtr_nid_param = rtr_nid;
2553 msg->msg_sending = 0;
2554 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2555 lnet_peer_ni_decref_locked(lpni);
2556 primary_nid = peer->lp_primary_nid;
2557 lnet_net_unlock(cpt);
2559 CDEBUG(D_NET, "%s pending discovery\n",
2560 libcfs_nid2str(primary_nid));
2562 return LNET_DC_WAIT;
2564 lnet_peer_ni_decref_locked(lpni);
2567 * Identify the different send cases
2569 if (src_nid == LNET_NID_ANY)
2570 send_case |= SRC_ANY;
2572 send_case |= SRC_SPEC;
2574 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2575 send_case |= LOCAL_DST;
2577 send_case |= REMOTE_DST;
2580 * if this is a non-MR peer or if we're recovering a peer ni then
2581 * let's consider this an NMR case so we can hit the destination
2584 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2585 send_case |= NMR_DST;
2587 send_case |= MR_DST;
2589 if (msg->msg_type == LNET_MSG_REPLY ||
2590 msg->msg_type == LNET_MSG_ACK)
2591 send_case |= SND_RESP;
2593 /* assign parameters to the send_data */
2594 send_data.sd_msg = msg;
2595 send_data.sd_rtr_nid = rtr_nid;
2596 send_data.sd_src_nid = src_nid;
2597 send_data.sd_dst_nid = dst_nid;
2598 send_data.sd_best_lpni = lpni;
2600 * keep a pointer to the final destination in case we're going to
2601 * route, so we'll need to access it later
2603 send_data.sd_final_dst_lpni = lpni;
2604 send_data.sd_peer = peer;
2605 send_data.sd_md_cpt = md_cpt;
2606 send_data.sd_cpt = cpt;
2607 send_data.sd_send_case = send_case;
2609 rc = lnet_handle_send_case_locked(&send_data);
2611 if (rc == REPEAT_SEND)
2614 lnet_net_unlock(send_data.sd_cpt);
2620 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2622 lnet_nid_t dst_nid = msg->msg_target.nid;
2626 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2627 * but we might want to use pre-determined router for ACK/REPLY
2630 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2631 LASSERT(msg->msg_txpeer == NULL);
2632 LASSERT(msg->msg_txni == NULL);
2633 LASSERT(!msg->msg_sending);
2634 LASSERT(!msg->msg_target_is_router);
2635 LASSERT(!msg->msg_receiving);
2637 msg->msg_sending = 1;
2639 LASSERT(!msg->msg_tx_committed);
2641 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2645 if (rc == LNET_CREDIT_OK)
2646 lnet_ni_send(msg->msg_txni, msg);
2648 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2652 enum lnet_mt_event_type {
2653 MT_TYPE_LOCAL_NI = 0,
2657 struct lnet_mt_event_info {
2658 enum lnet_mt_event_type mt_type;
2663 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2665 struct lnet_rsp_tracker *rspt;
2668 * msg has a refcount on the MD so the MD is not going away.
2669 * The rspt queue for the cpt is protected by
2670 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2673 if (!md->md_rspt_ptr) {
2674 lnet_res_unlock(cpt);
2677 rspt = md->md_rspt_ptr;
2678 md->md_rspt_ptr = NULL;
2681 LASSERT(rspt->rspt_cpt == cpt);
2684 * invalidate the handle to indicate that a response has been
2685 * received, which will then lead the monitor thread to clean up
2688 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2689 lnet_res_unlock(cpt);
2693 lnet_finalize_expired_responses(bool force)
2695 struct lnet_libmd *md;
2696 struct list_head local_queue;
2697 struct lnet_rsp_tracker *rspt, *tmp;
2700 if (the_lnet.ln_mt_rstq == NULL)
2703 cfs_cpt_for_each(i, lnet_cpt_table()) {
2704 INIT_LIST_HEAD(&local_queue);
2707 if (!the_lnet.ln_mt_rstq[i]) {
2711 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2714 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2716 * The rspt mdh will be invalidated when a response
2717 * is received or whenever we want to discard the
2718 * block the monitor thread will walk the queue
2719 * and clean up any rsts with an invalid mdh.
2720 * The monitor thread will walk the queue until
2721 * the first unexpired rspt block. This means that
2722 * some rspt blocks which received their
2723 * corresponding responses will linger in the
2724 * queue until they are cleaned up eventually.
2727 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2729 list_del_init(&rspt->rspt_on_list);
2730 lnet_rspt_free(rspt, i);
2734 if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
2736 md = lnet_handle2md(&rspt->rspt_mdh);
2738 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2740 list_del_init(&rspt->rspt_on_list);
2741 lnet_rspt_free(rspt, i);
2744 LASSERT(md->md_rspt_ptr == rspt);
2745 md->md_rspt_ptr = NULL;
2749 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2752 list_del_init(&rspt->rspt_on_list);
2754 CNETERR("Response timed out: md = %p\n", md);
2755 LNetMDUnlink(rspt->rspt_mdh);
2756 lnet_rspt_free(rspt, i);
2764 if (!list_empty(&local_queue))
2765 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2771 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2773 struct lnet_msg *msg;
2775 while (!list_empty(resendq)) {
2776 struct lnet_peer_ni *lpni;
2778 msg = list_entry(resendq->next, struct lnet_msg,
2781 list_del_init(&msg->msg_list);
2783 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2785 lnet_net_unlock(cpt);
2786 CERROR("Expected that a peer is already created for %s\n",
2787 libcfs_nid2str(msg->msg_hdr.dest_nid));
2788 msg->msg_no_resend = true;
2789 lnet_finalize(msg, -EFAULT);
2792 struct lnet_peer *peer;
2794 lnet_nid_t src_nid = LNET_NID_ANY;
2797 * if this message is not being routed and the
2798 * peer is non-MR then we must use the same
2799 * src_nid that was used in the original send.
2800 * Otherwise if we're routing the message (IE
2801 * we're a router) then we can use any of our
2802 * local interfaces. It doesn't matter to the
2803 * final destination.
2805 peer = lpni->lpni_peer_net->lpn_peer;
2806 if (!msg->msg_routing &&
2807 !lnet_peer_is_multi_rail(peer))
2808 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
2811 * If we originally specified a src NID, then we
2812 * must attempt to reuse it in the resend as well.
2814 if (msg->msg_src_nid_param != LNET_NID_ANY)
2815 src_nid = msg->msg_src_nid_param;
2816 lnet_peer_ni_decref_locked(lpni);
2818 lnet_net_unlock(cpt);
2819 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
2820 libcfs_nid2str(src_nid),
2821 libcfs_id2str(msg->msg_target),
2822 lnet_msgtyp2str(msg->msg_type),
2824 msg->msg_retry_count);
2825 rc = lnet_send(src_nid, msg, LNET_NID_ANY);
2827 CERROR("Error sending %s to %s: %d\n",
2828 lnet_msgtyp2str(msg->msg_type),
2829 libcfs_id2str(msg->msg_target), rc);
2830 msg->msg_no_resend = true;
2831 lnet_finalize(msg, rc);
2835 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
2841 lnet_resend_pending_msgs(void)
2845 cfs_cpt_for_each(i, lnet_cpt_table()) {
2847 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
2852 /* called with cpt and ni_lock held */
2854 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
2856 struct lnet_handle_md recovery_mdh;
2858 LNetInvalidateMDHandle(&recovery_mdh);
2860 if (ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING || force) {
2861 recovery_mdh = ni->ni_ping_mdh;
2862 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2865 lnet_net_unlock(cpt);
2866 if (!LNetMDHandleIsInvalid(recovery_mdh))
2867 LNetMDUnlink(recovery_mdh);
2873 lnet_recover_local_nis(void)
2875 struct lnet_mt_event_info *ev_info;
2876 struct list_head processed_list;
2877 struct list_head local_queue;
2878 struct lnet_handle_md mdh;
2879 struct lnet_ni *tmp;
2885 INIT_LIST_HEAD(&local_queue);
2886 INIT_LIST_HEAD(&processed_list);
2889 * splice the recovery queue on a local queue. We will iterate
2890 * through the local queue and update it as needed. Once we're
2891 * done with the traversal, we'll splice the local queue back on
2892 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
2893 * will be traversed in the next iteration.
2896 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
2900 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
2902 * if an NI is being deleted or it is now healthy, there
2903 * is no need to keep it around in the recovery queue.
2904 * The monitor thread is the only thread responsible for
2905 * removing the NI from the recovery queue.
2906 * Multiple threads can be adding NIs to the recovery
2909 healthv = atomic_read(&ni->ni_healthv);
2913 if (!(ni->ni_state & LNET_NI_STATE_ACTIVE) ||
2914 healthv == LNET_MAX_HEALTH_VALUE) {
2915 list_del_init(&ni->ni_recovery);
2916 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
2918 lnet_ni_decref_locked(ni, 0);
2924 * if the local NI failed recovery we must unlink the md.
2925 * But we want to keep the local_ni on the recovery queue
2926 * so we can continue the attempts to recover it.
2928 if (ni->ni_state & LNET_NI_STATE_RECOVERY_FAILED) {
2929 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
2930 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_FAILED;
2937 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
2938 libcfs_nid2str(ni->ni_nid));
2941 if (!(ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING)) {
2942 ni->ni_state |= LNET_NI_STATE_RECOVERY_PENDING;
2945 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
2947 CERROR("out of memory. Can't recover %s\n",
2948 libcfs_nid2str(ni->ni_nid));
2950 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
2955 mdh = ni->ni_ping_mdh;
2957 * Invalidate the ni mdh in case it's deleted.
2958 * We'll unlink the mdh in this case below.
2960 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2964 * remove the NI from the local queue and drop the
2965 * reference count to it while we're recovering
2966 * it. The reason for that, is that the NI could
2967 * be deleted, and the way the code is structured
2968 * is if we don't drop the NI, then the deletion
2969 * code will enter a loop waiting for the
2970 * reference count to be removed while holding the
2971 * ln_mutex_lock(). When we look up the peer to
2972 * send to in lnet_select_pathway() we will try to
2973 * lock the ln_mutex_lock() as well, leading to
2974 * a deadlock. By dropping the refcount and
2975 * removing it from the list, we allow for the NI
2976 * to be removed, then we use the cached NID to
2977 * look it up again. If it's gone, then we just
2978 * continue examining the rest of the queue.
2981 list_del_init(&ni->ni_recovery);
2982 lnet_ni_decref_locked(ni, 0);
2985 ev_info->mt_type = MT_TYPE_LOCAL_NI;
2986 ev_info->mt_nid = nid;
2987 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
2988 ev_info, the_lnet.ln_mt_eqh, true);
2989 /* lookup the nid again */
2991 ni = lnet_nid2ni_locked(nid, 0);
2994 * the NI has been deleted when we dropped
3002 * Same note as in lnet_recover_peer_nis(). When
3003 * we're sending the ping, the NI is free to be
3004 * deleted or manipulated. By this point it
3005 * could've been added back on the recovery queue,
3006 * and a refcount taken on it.
3007 * So we can't just add it blindly again or we'll
3008 * corrupt the queue. We must check under lock if
3009 * it's not on any list and if not then add it
3010 * to the processed list, which will eventually be
3011 * spliced back on to the recovery queue.
3013 ni->ni_ping_mdh = mdh;
3014 if (list_empty(&ni->ni_recovery)) {
3015 list_add_tail(&ni->ni_recovery, &processed_list);
3016 lnet_ni_addref_locked(ni, 0);
3022 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
3028 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3029 * reexamined in the next iteration.
3031 list_splice_init(&processed_list, &local_queue);
3033 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3037 static struct list_head **
3038 lnet_create_array_of_queues(void)
3040 struct list_head **qs;
3041 struct list_head *q;
3044 qs = cfs_percpt_alloc(lnet_cpt_table(),
3045 sizeof(struct list_head));
3047 CERROR("Failed to allocate queues\n");
3051 cfs_percpt_for_each(q, i, qs)
3058 lnet_resendqs_create(void)
3060 struct list_head **resendqs;
3061 resendqs = lnet_create_array_of_queues();
3066 lnet_net_lock(LNET_LOCK_EX);
3067 the_lnet.ln_mt_resendqs = resendqs;
3068 lnet_net_unlock(LNET_LOCK_EX);
3074 lnet_clean_local_ni_recoveryq(void)
3078 /* This is only called when the monitor thread has stopped */
3081 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3082 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3083 struct lnet_ni, ni_recovery);
3084 list_del_init(&ni->ni_recovery);
3086 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3088 lnet_ni_decref_locked(ni, 0);
3095 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3098 struct lnet_handle_md recovery_mdh;
3100 LNetInvalidateMDHandle(&recovery_mdh);
3102 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3103 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3104 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3106 spin_unlock(&lpni->lpni_lock);
3107 lnet_net_unlock(cpt);
3108 if (!LNetMDHandleIsInvalid(recovery_mdh))
3109 LNetMDUnlink(recovery_mdh);
3111 spin_lock(&lpni->lpni_lock);
3115 lnet_clean_peer_ni_recoveryq(void)
3117 struct lnet_peer_ni *lpni, *tmp;
3119 lnet_net_lock(LNET_LOCK_EX);
3121 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3123 list_del_init(&lpni->lpni_recovery);
3124 spin_lock(&lpni->lpni_lock);
3125 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3126 spin_unlock(&lpni->lpni_lock);
3127 lnet_peer_ni_decref_locked(lpni);
3130 lnet_net_unlock(LNET_LOCK_EX);
3134 lnet_clean_resendqs(void)
3136 struct lnet_msg *msg, *tmp;
3137 struct list_head msgs;
3140 INIT_LIST_HEAD(&msgs);
3142 cfs_cpt_for_each(i, lnet_cpt_table()) {
3144 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3146 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3147 list_del_init(&msg->msg_list);
3148 msg->msg_no_resend = true;
3149 lnet_finalize(msg, -ESHUTDOWN);
3153 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3157 lnet_recover_peer_nis(void)
3159 struct lnet_mt_event_info *ev_info;
3160 struct list_head processed_list;
3161 struct list_head local_queue;
3162 struct lnet_handle_md mdh;
3163 struct lnet_peer_ni *lpni;
3164 struct lnet_peer_ni *tmp;
3169 INIT_LIST_HEAD(&local_queue);
3170 INIT_LIST_HEAD(&processed_list);
3173 * Always use cpt 0 for locking across all interactions with
3174 * ln_mt_peerNIRecovq
3177 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3181 list_for_each_entry_safe(lpni, tmp, &local_queue,
3184 * The same protection strategy is used here as is in the
3185 * local recovery case.
3188 healthv = atomic_read(&lpni->lpni_healthv);
3189 spin_lock(&lpni->lpni_lock);
3190 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3191 healthv == LNET_MAX_HEALTH_VALUE) {
3192 list_del_init(&lpni->lpni_recovery);
3193 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3194 spin_unlock(&lpni->lpni_lock);
3195 lnet_peer_ni_decref_locked(lpni);
3201 * If the peer NI has failed recovery we must unlink the
3202 * md. But we want to keep the peer ni on the recovery
3203 * queue so we can try to continue recovering it
3205 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3206 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3207 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3210 spin_unlock(&lpni->lpni_lock);
3214 * NOTE: we're racing with peer deletion from user space.
3215 * It's possible that a peer is deleted after we check its
3216 * state. In this case the recovery can create a new peer
3218 spin_lock(&lpni->lpni_lock);
3219 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3220 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3221 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3222 spin_unlock(&lpni->lpni_lock);
3224 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3226 CERROR("out of memory. Can't recover %s\n",
3227 libcfs_nid2str(lpni->lpni_nid));
3228 spin_lock(&lpni->lpni_lock);
3229 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3230 spin_unlock(&lpni->lpni_lock);
3234 /* look at the comments in lnet_recover_local_nis() */
3235 mdh = lpni->lpni_recovery_ping_mdh;
3236 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3237 nid = lpni->lpni_nid;
3239 list_del_init(&lpni->lpni_recovery);
3240 lnet_peer_ni_decref_locked(lpni);
3243 ev_info->mt_type = MT_TYPE_PEER_NI;
3244 ev_info->mt_nid = nid;
3245 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3246 ev_info, the_lnet.ln_mt_eqh, true);
3249 * lnet_find_peer_ni_locked() grabs a refcount for
3250 * us. No need to take it explicitly.
3252 lpni = lnet_find_peer_ni_locked(nid);
3259 lpni->lpni_recovery_ping_mdh = mdh;
3261 * While we're unlocked the lpni could've been
3262 * readded on the recovery queue. In this case we
3263 * don't need to add it to the local queue, since
3264 * it's already on there and the thread that added
3265 * it would've incremented the refcount on the
3266 * peer, which means we need to decref the refcount
3267 * that was implicitly grabbed by find_peer_ni_locked.
3268 * Otherwise, if the lpni is still not on
3269 * the recovery queue, then we'll add it to the
3272 if (list_empty(&lpni->lpni_recovery))
3273 list_add_tail(&lpni->lpni_recovery, &processed_list);
3275 lnet_peer_ni_decref_locked(lpni);
3278 spin_lock(&lpni->lpni_lock);
3280 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3282 spin_unlock(&lpni->lpni_lock);
3285 list_splice_init(&processed_list, &local_queue);
3287 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3292 lnet_monitor_thread(void *arg)
3294 int wakeup_counter = 0;
3297 * The monitor thread takes care of the following:
3298 * 1. Checks the aliveness of routers
3299 * 2. Checks if there are messages on the resend queue to resend
3301 * 3. Check if there are any NIs on the local recovery queue and
3303 * 4. Checks if there are any NIs on the remote recovery queue
3306 cfs_block_allsigs();
3308 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3309 if (lnet_router_checker_active())
3310 lnet_check_routers();
3312 lnet_resend_pending_msgs();
3315 if (wakeup_counter >= lnet_transaction_timeout / 2) {
3316 lnet_finalize_expired_responses(false);
3320 lnet_recover_local_nis();
3322 lnet_recover_peer_nis();
3325 * TODO do we need to check if we should sleep without
3326 * timeout? Technically, an active system will always
3327 * have messages in flight so this check will always
3328 * evaluate to false. And on an idle system do we care
3329 * if we wake up every 1 second? Although, we've seen
3330 * cases where we get a complaint that an idle thread
3331 * is waking up unnecessarily.
3333 wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
3335 cfs_time_seconds(1));
3338 /* clean up the router checker */
3339 lnet_prune_rc_data(1);
3342 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3344 /* signal that the monitor thread is exiting */
3345 up(&the_lnet.ln_mt_signal);
3353 * Returns == 0 if success
3354 * Returns > 0 if LNetMDBind or prior fails
3355 * Returns < 0 if LNetGet fails
3358 lnet_send_ping(lnet_nid_t dest_nid,
3359 struct lnet_handle_md *mdh, int nnis,
3360 void *user_data, struct lnet_handle_eq eqh, bool recovery)
3362 struct lnet_md md = { NULL };
3363 struct lnet_process_id id;
3364 struct lnet_ping_buffer *pbuf;
3367 if (dest_nid == LNET_NID_ANY) {
3372 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3378 /* initialize md content */
3379 md.start = &pbuf->pb_info;
3380 md.length = LNET_PING_INFO_SIZE(nnis);
3381 md.threshold = 2; /* GET/REPLY */
3383 md.options = LNET_MD_TRUNCATE;
3384 md.user_ptr = user_data;
3387 rc = LNetMDBind(md, LNET_UNLINK, mdh);
3389 lnet_ping_buffer_decref(pbuf);
3390 CERROR("Can't bind MD: %d\n", rc);
3391 rc = -rc; /* change the rc to positive */
3394 id.pid = LNET_PID_LUSTRE;
3397 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3398 LNET_RESERVED_PORTAL,
3399 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3402 goto fail_unlink_md;
3408 LNetInvalidateMDHandle(mdh);
3414 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3417 lnet_nid_t nid = ev_info->mt_nid;
3419 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3423 ni = lnet_nid2ni_locked(nid, 0);
3429 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
3431 ni->ni_state |= LNET_NI_STATE_RECOVERY_FAILED;
3436 CERROR("local NI (%s) recovery failed with %d\n",
3437 libcfs_nid2str(nid), status);
3441 * need to increment healthv for the ni here, because in
3442 * the lnet_finalize() path we don't have access to this
3443 * NI. And in order to get access to it, we'll need to
3444 * carry forward too much information.
3445 * In the peer case, it'll naturally be incremented
3447 lnet_inc_healthv(&ni->ni_healthv);
3449 struct lnet_peer_ni *lpni;
3452 cpt = lnet_net_lock_current();
3453 lpni = lnet_find_peer_ni_locked(nid);
3455 lnet_net_unlock(cpt);
3458 spin_lock(&lpni->lpni_lock);
3459 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3461 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3462 spin_unlock(&lpni->lpni_lock);
3463 lnet_peer_ni_decref_locked(lpni);
3464 lnet_net_unlock(cpt);
3467 CERROR("peer NI (%s) recovery failed with %d\n",
3468 libcfs_nid2str(nid), status);
3473 lnet_mt_event_handler(struct lnet_event *event)
3475 struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3476 struct lnet_ping_buffer *pbuf;
3478 /* TODO: remove assert */
3479 LASSERT(event->type == LNET_EVENT_REPLY ||
3480 event->type == LNET_EVENT_SEND ||
3481 event->type == LNET_EVENT_UNLINK);
3483 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3486 switch (event->type) {
3487 case LNET_EVENT_UNLINK:
3488 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3489 libcfs_nid2str(ev_info->mt_nid));
3490 case LNET_EVENT_REPLY:
3491 lnet_handle_recovery_reply(ev_info, event->status);
3493 case LNET_EVENT_SEND:
3494 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3495 libcfs_nid2str(ev_info->mt_nid),
3496 (event->status) ? "unsuccessfully" :
3497 "successfully", event->status);
3498 lnet_handle_recovery_reply(ev_info, event->status);
3501 CERROR("Unexpected event: %d\n", event->type);
3504 if (event->unlinked) {
3505 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3506 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3507 lnet_ping_buffer_decref(pbuf);
3512 lnet_rsp_tracker_create(void)
3514 struct list_head **rstqs;
3515 rstqs = lnet_create_array_of_queues();
3520 the_lnet.ln_mt_rstq = rstqs;
3526 lnet_rsp_tracker_clean(void)
3528 lnet_finalize_expired_responses(true);
3530 cfs_percpt_free(the_lnet.ln_mt_rstq);
3531 the_lnet.ln_mt_rstq = NULL;
3534 int lnet_monitor_thr_start(void)
3537 struct task_struct *task;
3539 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3542 rc = lnet_resendqs_create();
3546 rc = lnet_rsp_tracker_create();
3550 rc = LNetEQAlloc(0, lnet_mt_event_handler, &the_lnet.ln_mt_eqh);
3552 CERROR("Can't allocate monitor thread EQ: %d\n", rc);
3556 /* Pre monitor thread start processing */
3557 rc = lnet_router_pre_mt_start();
3561 sema_init(&the_lnet.ln_mt_signal, 0);
3563 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3564 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3567 CERROR("Can't start monitor thread: %d\n", rc);
3571 /* post monitor thread start processing */
3572 lnet_router_post_mt_start();
3577 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3578 /* block until event callback signals exit */
3579 down(&the_lnet.ln_mt_signal);
3581 lnet_router_cleanup();
3583 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3584 lnet_rsp_tracker_clean();
3585 lnet_clean_local_ni_recoveryq();
3586 lnet_clean_peer_ni_recoveryq();
3587 lnet_clean_resendqs();
3588 LNetEQFree(the_lnet.ln_mt_eqh);
3589 LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3592 lnet_rsp_tracker_clean();
3593 lnet_clean_local_ni_recoveryq();
3594 lnet_clean_peer_ni_recoveryq();
3595 lnet_clean_resendqs();
3599 void lnet_monitor_thr_stop(void)
3603 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3606 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3607 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3609 /* tell the monitor thread that we're shutting down */
3610 wake_up(&the_lnet.ln_mt_waitq);
3612 /* block until monitor thread signals that it's done */
3613 down(&the_lnet.ln_mt_signal);
3614 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3616 /* perform cleanup tasks */
3617 lnet_router_cleanup();
3618 lnet_rsp_tracker_clean();
3619 lnet_clean_local_ni_recoveryq();
3620 lnet_clean_peer_ni_recoveryq();
3621 lnet_clean_resendqs();
3622 rc = LNetEQFree(the_lnet.ln_mt_eqh);
3628 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3632 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3633 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3634 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3635 lnet_net_unlock(cpt);
3637 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3641 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3643 struct lnet_hdr *hdr = &msg->msg_hdr;
3645 if (msg->msg_wanted != 0)
3646 lnet_setpayloadbuffer(msg);
3648 lnet_build_msg_event(msg, LNET_EVENT_PUT);
3650 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
3651 * it back into the ACK during lnet_finalize() */
3652 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3653 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3655 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3656 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3660 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3662 struct lnet_hdr *hdr = &msg->msg_hdr;
3663 struct lnet_match_info info;
3667 /* Convert put fields to host byte order */
3668 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3669 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
3670 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
3672 /* Primary peer NID. */
3673 info.mi_id.nid = msg->msg_initiator;
3674 info.mi_id.pid = hdr->src_pid;
3675 info.mi_opc = LNET_MD_OP_PUT;
3676 info.mi_portal = hdr->msg.put.ptl_index;
3677 info.mi_rlength = hdr->payload_length;
3678 info.mi_roffset = hdr->msg.put.offset;
3679 info.mi_mbits = hdr->msg.put.match_bits;
3680 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
3682 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3683 ready_delay = msg->msg_rx_ready_delay;
3686 rc = lnet_ptl_match_md(&info, msg);
3691 case LNET_MATCHMD_OK:
3692 lnet_recv_put(ni, msg);
3695 case LNET_MATCHMD_NONE:
3697 /* no eager_recv or has already called it, should
3698 * have been attached on delayed list */
3701 rc = lnet_ni_eager_recv(ni, msg);
3708 case LNET_MATCHMD_DROP:
3709 CNETERR("Dropping PUT from %s portal %d match %llu"
3710 " offset %d length %d: %d\n",
3711 libcfs_id2str(info.mi_id), info.mi_portal,
3712 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3714 return -ENOENT; /* -ve: OK but no match */
3719 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3721 struct lnet_match_info info;
3722 struct lnet_hdr *hdr = &msg->msg_hdr;
3723 struct lnet_process_id source_id;
3724 struct lnet_handle_wire reply_wmd;
3727 /* Convert get fields to host byte order */
3728 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
3729 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
3730 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
3731 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
3733 source_id.nid = hdr->src_nid;
3734 source_id.pid = hdr->src_pid;
3735 /* Primary peer NID */
3736 info.mi_id.nid = msg->msg_initiator;
3737 info.mi_id.pid = hdr->src_pid;
3738 info.mi_opc = LNET_MD_OP_GET;
3739 info.mi_portal = hdr->msg.get.ptl_index;
3740 info.mi_rlength = hdr->msg.get.sink_length;
3741 info.mi_roffset = hdr->msg.get.src_offset;
3742 info.mi_mbits = hdr->msg.get.match_bits;
3743 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
3745 rc = lnet_ptl_match_md(&info, msg);
3746 if (rc == LNET_MATCHMD_DROP) {
3747 CNETERR("Dropping GET from %s portal %d match %llu"
3748 " offset %d length %d\n",
3749 libcfs_id2str(info.mi_id), info.mi_portal,
3750 info.mi_mbits, info.mi_roffset, info.mi_rlength);
3751 return -ENOENT; /* -ve: OK but no match */
3754 LASSERT(rc == LNET_MATCHMD_OK);
3756 lnet_build_msg_event(msg, LNET_EVENT_GET);
3758 reply_wmd = hdr->msg.get.return_wmd;
3760 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3761 msg->msg_offset, msg->msg_wanted);
3763 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3766 /* The LND completes the REPLY from her recv procedure */
3767 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3768 msg->msg_offset, msg->msg_len, msg->msg_len);
3772 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3773 msg->msg_receiving = 0;
3775 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
3777 /* didn't get as far as lnet_ni_send() */
3778 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3779 libcfs_nid2str(ni->ni_nid),
3780 libcfs_id2str(info.mi_id), rc);
3782 lnet_finalize(msg, rc);
3789 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3791 void *private = msg->msg_private;
3792 struct lnet_hdr *hdr = &msg->msg_hdr;
3793 struct lnet_process_id src = {0};
3794 struct lnet_libmd *md;
3799 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3802 src.nid = hdr->src_nid;
3803 src.pid = hdr->src_pid;
3805 /* NB handles only looked up by creator (no flips) */
3806 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
3807 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3808 CNETERR("%s: Dropping REPLY from %s for %s "
3810 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3811 (md == NULL) ? "invalid" : "inactive",
3812 hdr->msg.reply.dst_wmd.wh_interface_cookie,
3813 hdr->msg.reply.dst_wmd.wh_object_cookie);
3814 if (md != NULL && md->md_me != NULL)
3815 CERROR("REPLY MD also attached to portal %d\n",
3816 md->md_me->me_portal);
3818 lnet_res_unlock(cpt);
3819 return -ENOENT; /* -ve: OK but no match */
3822 LASSERT(md->md_offset == 0);
3824 rlength = hdr->payload_length;
3825 mlength = MIN(rlength, (int)md->md_length);
3827 if (mlength < rlength &&
3828 (md->md_options & LNET_MD_TRUNCATE) == 0) {
3829 CNETERR("%s: Dropping REPLY from %s length %d "
3830 "for MD %#llx would overflow (%d)\n",
3831 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3832 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
3834 lnet_res_unlock(cpt);
3835 return -ENOENT; /* -ve: OK but no match */
3838 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
3839 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3840 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
3842 lnet_msg_attach_md(msg, md, 0, mlength);
3845 lnet_setpayloadbuffer(msg);
3847 lnet_res_unlock(cpt);
3849 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3851 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
3856 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
3858 struct lnet_hdr *hdr = &msg->msg_hdr;
3859 struct lnet_process_id src = {0};
3860 struct lnet_libmd *md;
3863 src.nid = hdr->src_nid;
3864 src.pid = hdr->src_pid;
3866 /* Convert ack fields to host byte order */
3867 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
3868 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
3870 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
3873 /* NB handles only looked up by creator (no flips) */
3874 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
3875 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3876 /* Don't moan; this is expected */
3878 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
3879 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3880 (md == NULL) ? "invalid" : "inactive",
3881 hdr->msg.ack.dst_wmd.wh_interface_cookie,
3882 hdr->msg.ack.dst_wmd.wh_object_cookie);
3883 if (md != NULL && md->md_me != NULL)
3884 CERROR("Source MD also attached to portal %d\n",
3885 md->md_me->me_portal);
3887 lnet_res_unlock(cpt);
3888 return -ENOENT; /* -ve! */
3891 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
3892 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3893 hdr->msg.ack.dst_wmd.wh_object_cookie);
3895 lnet_msg_attach_md(msg, md, 0, 0);
3897 lnet_res_unlock(cpt);
3899 lnet_build_msg_event(msg, LNET_EVENT_ACK);
3901 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
3906 * \retval LNET_CREDIT_OK If \a msg is forwarded
3907 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
3908 * \retval -ve error code
3911 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
3915 if (!the_lnet.ln_routing)
3918 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
3919 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
3920 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
3921 msg->msg_rx_ready_delay = 1;
3923 lnet_net_unlock(msg->msg_rx_cpt);
3924 rc = lnet_ni_eager_recv(ni, msg);
3925 lnet_net_lock(msg->msg_rx_cpt);
3930 rc = lnet_post_routed_recv_locked(msg, 0);
3935 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
3939 switch (msg->msg_type) {
3941 rc = lnet_parse_ack(ni, msg);
3944 rc = lnet_parse_put(ni, msg);
3947 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
3949 case LNET_MSG_REPLY:
3950 rc = lnet_parse_reply(ni, msg);
3952 default: /* prevent an unused label if !kernel */
3957 LASSERT(rc == 0 || rc == -ENOENT);
3962 lnet_msgtyp2str (int type)
3971 case LNET_MSG_REPLY:
3973 case LNET_MSG_HELLO:
3976 return ("<UNKNOWN>");
3981 lnet_print_hdr(struct lnet_hdr *hdr)
3983 struct lnet_process_id src = {
3984 .nid = hdr->src_nid,
3985 .pid = hdr->src_pid,
3987 struct lnet_process_id dst = {
3988 .nid = hdr->dest_nid,
3989 .pid = hdr->dest_pid,
3991 char *type_str = lnet_msgtyp2str(hdr->type);
3993 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
3994 CWARN(" From %s\n", libcfs_id2str(src));
3995 CWARN(" To %s\n", libcfs_id2str(dst));
3997 switch (hdr->type) {
4002 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
4003 "match bits %llu\n",
4004 hdr->msg.put.ptl_index,
4005 hdr->msg.put.ack_wmd.wh_interface_cookie,
4006 hdr->msg.put.ack_wmd.wh_object_cookie,
4007 hdr->msg.put.match_bits);
4008 CWARN(" Length %d, offset %d, hdr data %#llx\n",
4009 hdr->payload_length, hdr->msg.put.offset,
4010 hdr->msg.put.hdr_data);
4014 CWARN(" Ptl index %d, return md %#llx.%#llx, "
4015 "match bits %llu\n", hdr->msg.get.ptl_index,
4016 hdr->msg.get.return_wmd.wh_interface_cookie,
4017 hdr->msg.get.return_wmd.wh_object_cookie,
4018 hdr->msg.get.match_bits);
4019 CWARN(" Length %d, src offset %d\n",
4020 hdr->msg.get.sink_length,
4021 hdr->msg.get.src_offset);
4025 CWARN(" dst md %#llx.%#llx, "
4026 "manipulated length %d\n",
4027 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4028 hdr->msg.ack.dst_wmd.wh_object_cookie,
4029 hdr->msg.ack.mlength);
4032 case LNET_MSG_REPLY:
4033 CWARN(" dst md %#llx.%#llx, "
4035 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4036 hdr->msg.reply.dst_wmd.wh_object_cookie,
4037 hdr->payload_length);
4043 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4044 void *private, int rdma_req)
4049 struct lnet_msg *msg;
4050 lnet_pid_t dest_pid;
4051 lnet_nid_t dest_nid;
4053 struct lnet_peer_ni *lpni;
4054 __u32 payload_length;
4057 LASSERT (!in_interrupt ());
4059 type = le32_to_cpu(hdr->type);
4060 src_nid = le64_to_cpu(hdr->src_nid);
4061 dest_nid = le64_to_cpu(hdr->dest_nid);
4062 dest_pid = le32_to_cpu(hdr->dest_pid);
4063 payload_length = le32_to_cpu(hdr->payload_length);
4065 for_me = (ni->ni_nid == dest_nid);
4066 cpt = lnet_cpt_of_nid(from_nid, ni);
4068 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4069 libcfs_nid2str(dest_nid),
4070 libcfs_nid2str(ni->ni_nid),
4071 libcfs_nid2str(src_nid),
4072 lnet_msgtyp2str(type),
4073 (for_me) ? "for me" : "routed");
4078 if (payload_length > 0) {
4079 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4080 libcfs_nid2str(from_nid),
4081 libcfs_nid2str(src_nid),
4082 lnet_msgtyp2str(type), payload_length);
4088 case LNET_MSG_REPLY:
4089 if (payload_length >
4090 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4091 CERROR("%s, src %s: bad %s payload %d "
4092 "(%d max expected)\n",
4093 libcfs_nid2str(from_nid),
4094 libcfs_nid2str(src_nid),
4095 lnet_msgtyp2str(type),
4097 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4103 CERROR("%s, src %s: Bad message type 0x%x\n",
4104 libcfs_nid2str(from_nid),
4105 libcfs_nid2str(src_nid), type);
4109 if (the_lnet.ln_routing &&
4110 ni->ni_last_alive != ktime_get_real_seconds()) {
4111 /* NB: so far here is the only place to set NI status to "up */
4113 ni->ni_last_alive = ktime_get_real_seconds();
4114 if (ni->ni_status != NULL &&
4115 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
4116 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4120 /* Regard a bad destination NID as a protocol error. Senders should
4121 * know what they're doing; if they don't they're misconfigured, buggy
4122 * or malicious so we chop them off at the knees :) */
4125 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4126 /* should have gone direct */
4127 CERROR("%s, src %s: Bad dest nid %s "
4128 "(should have been sent direct)\n",
4129 libcfs_nid2str(from_nid),
4130 libcfs_nid2str(src_nid),
4131 libcfs_nid2str(dest_nid));
4135 if (lnet_islocalnid(dest_nid)) {
4136 /* dest is another local NI; sender should have used
4137 * this node's NID on its own network */
4138 CERROR("%s, src %s: Bad dest nid %s "
4139 "(it's my nid but on a different network)\n",
4140 libcfs_nid2str(from_nid),
4141 libcfs_nid2str(src_nid),
4142 libcfs_nid2str(dest_nid));
4146 if (rdma_req && type == LNET_MSG_GET) {
4147 CERROR("%s, src %s: Bad optimized GET for %s "
4148 "(final destination must be me)\n",
4149 libcfs_nid2str(from_nid),
4150 libcfs_nid2str(src_nid),
4151 libcfs_nid2str(dest_nid));
4155 if (!the_lnet.ln_routing) {
4156 CERROR("%s, src %s: Dropping message for %s "
4157 "(routing not enabled)\n",
4158 libcfs_nid2str(from_nid),
4159 libcfs_nid2str(src_nid),
4160 libcfs_nid2str(dest_nid));
4165 /* Message looks OK; we're not going to return an error, so we MUST
4166 * call back lnd_recv() come what may... */
4168 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4169 fail_peer(src_nid, 0)) { /* shall we now? */
4170 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4171 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4172 lnet_msgtyp2str(type));
4176 if (!list_empty(&the_lnet.ln_drop_rules) &&
4177 lnet_drop_rule_match(hdr, NULL)) {
4178 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4179 "silent message loss\n",
4180 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4181 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4186 msg = lnet_msg_alloc();
4188 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4189 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4190 lnet_msgtyp2str(type));
4194 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4195 * pointers NULL etc */
4197 msg->msg_type = type;
4198 msg->msg_private = private;
4199 msg->msg_receiving = 1;
4200 msg->msg_rdma_get = rdma_req;
4201 msg->msg_len = msg->msg_wanted = payload_length;
4202 msg->msg_offset = 0;
4203 msg->msg_hdr = *hdr;
4204 /* for building message event */
4205 msg->msg_from = from_nid;
4207 msg->msg_target.pid = dest_pid;
4208 msg->msg_target.nid = dest_nid;
4209 msg->msg_routing = 1;
4212 /* convert common msg->hdr fields to host byteorder */
4213 msg->msg_hdr.type = type;
4214 msg->msg_hdr.src_nid = src_nid;
4215 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4216 msg->msg_hdr.dest_nid = dest_nid;
4217 msg->msg_hdr.dest_pid = dest_pid;
4218 msg->msg_hdr.payload_length = payload_length;
4222 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4224 lnet_net_unlock(cpt);
4225 CERROR("%s, src %s: Dropping %s "
4226 "(error %ld looking up sender)\n",
4227 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4228 lnet_msgtyp2str(type), PTR_ERR(lpni));
4230 if (rc == -ESHUTDOWN)
4231 /* We are shutting down. Don't do anything more */
4235 msg->msg_rxpeer = lpni;
4237 lnet_ni_addref_locked(ni, cpt);
4238 /* Multi-Rail: Primary NID of source. */
4239 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4241 if (lnet_isrouter(msg->msg_rxpeer)) {
4242 lnet_peer_set_alive(msg->msg_rxpeer);
4243 if (avoid_asym_router_failure &&
4244 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4245 /* received a remote message from router, update
4246 * remote NI status on this router.
4247 * NB: multi-hop routed message will be ignored.
4249 lnet_router_ni_update_locked(msg->msg_rxpeer,
4250 LNET_NIDNET(src_nid));
4254 lnet_msg_commit(msg, cpt);
4256 /* message delay simulation */
4257 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4258 lnet_delay_rule_match_locked(hdr, msg))) {
4259 lnet_net_unlock(cpt);
4264 rc = lnet_parse_forward_locked(ni, msg);
4265 lnet_net_unlock(cpt);
4270 if (rc == LNET_CREDIT_OK) {
4271 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4272 0, payload_length, payload_length);
4277 lnet_net_unlock(cpt);
4279 rc = lnet_parse_local(ni, msg);
4285 LASSERT(msg->msg_md == NULL);
4286 lnet_finalize(msg, rc);
4289 lnet_drop_message(ni, cpt, private, payload_length, type);
4292 EXPORT_SYMBOL(lnet_parse);
4295 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4297 while (!list_empty(head)) {
4298 struct lnet_process_id id = {0};
4299 struct lnet_msg *msg;
4301 msg = list_entry(head->next, struct lnet_msg, msg_list);
4302 list_del(&msg->msg_list);
4304 id.nid = msg->msg_hdr.src_nid;
4305 id.pid = msg->msg_hdr.src_pid;
4307 LASSERT(msg->msg_md == NULL);
4308 LASSERT(msg->msg_rx_delayed);
4309 LASSERT(msg->msg_rxpeer != NULL);
4310 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4312 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4313 " offset %d length %d: %s\n",
4315 msg->msg_hdr.msg.put.ptl_index,
4316 msg->msg_hdr.msg.put.match_bits,
4317 msg->msg_hdr.msg.put.offset,
4318 msg->msg_hdr.payload_length, reason);
4320 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4321 * called lnet_drop_message(), so I just hang onto msg as well
4322 * until that's done */
4324 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4325 msg->msg_private, msg->msg_len,
4328 msg->msg_no_resend = true;
4330 * NB: message will not generate event because w/o attached MD,
4331 * but we still should give error code so lnet_msg_decommit()
4332 * can skip counters operations and other checks.
4334 lnet_finalize(msg, -ENOENT);
4339 lnet_recv_delayed_msg_list(struct list_head *head)
4341 while (!list_empty(head)) {
4342 struct lnet_msg *msg;
4343 struct lnet_process_id id;
4345 msg = list_entry(head->next, struct lnet_msg, msg_list);
4346 list_del(&msg->msg_list);
4348 /* md won't disappear under me, since each msg
4349 * holds a ref on it */
4351 id.nid = msg->msg_hdr.src_nid;
4352 id.pid = msg->msg_hdr.src_pid;
4354 LASSERT(msg->msg_rx_delayed);
4355 LASSERT(msg->msg_md != NULL);
4356 LASSERT(msg->msg_rxpeer != NULL);
4357 LASSERT(msg->msg_rxni != NULL);
4358 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4360 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4361 "match %llu offset %d length %d.\n",
4362 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4363 msg->msg_hdr.msg.put.match_bits,
4364 msg->msg_hdr.msg.put.offset,
4365 msg->msg_hdr.payload_length);
4367 lnet_recv_put(msg->msg_rxni, msg);
4372 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4373 struct lnet_libmd *md, struct lnet_handle_md mdh)
4378 * MD has a refcount taken by message so it's not going away.
4379 * The MD however can be looked up. We need to secure the access
4380 * to the md_rspt_ptr by taking the res_lock.
4381 * The rspt can be accessed without protection up to when it gets
4382 * added to the list.
4386 LASSERT(md->md_rspt_ptr == NULL);
4388 /* we'll use that same event in case we never get a response */
4389 rspt->rspt_mdh = mdh;
4390 rspt->rspt_cpt = cpt;
4391 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4392 rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4395 /* store the rspt so we can access it when we get the REPLY */
4396 md->md_rspt_ptr = rspt;
4397 lnet_res_unlock(cpt);
4400 * add to the list of tracked responses. It's added to tail of the
4401 * list in order to expire all the older entries first.
4404 list_add_tail(&rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4405 lnet_net_unlock(cpt);
4409 * Initiate an asynchronous PUT operation.
4411 * There are several events associated with a PUT: completion of the send on
4412 * the initiator node (LNET_EVENT_SEND), and when the send completes
4413 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4414 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4415 * used at the target node to indicate the completion of incoming data
4418 * The local events will be logged in the EQ associated with the MD pointed to
4419 * by \a mdh handle. Using a MD without an associated EQ results in these
4420 * events being discarded. In this case, the caller must have another
4421 * mechanism (e.g., a higher level protocol) for determining when it is safe
4422 * to modify the memory region associated with the MD.
4424 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4425 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4427 * \param self Indicates the NID of a local interface through which to send
4428 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4429 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4430 * must be "free floating" (See LNetMDBind()).
4431 * \param ack Controls whether an acknowledgment is requested.
4432 * Acknowledgments are only sent when they are requested by the initiating
4433 * process and the target MD enables them.
4434 * \param target A process identifier for the target process.
4435 * \param portal The index in the \a target's portal table.
4436 * \param match_bits The match bits to use for MD selection at the target
4438 * \param offset The offset into the target MD (only used when the target
4439 * MD has the LNET_MD_MANAGE_REMOTE option set).
4440 * \param hdr_data 64 bits of user data that can be included in the message
4441 * header. This data is written to an event queue entry at the target if an
4442 * EQ is present on the matching MD.
4444 * \retval 0 Success, and only in this case events will be generated
4445 * and logged to EQ (if it exists).
4446 * \retval -EIO Simulated failure.
4447 * \retval -ENOMEM Memory allocation failure.
4448 * \retval -ENOENT Invalid MD object.
4450 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4453 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4454 struct lnet_process_id target, unsigned int portal,
4455 __u64 match_bits, unsigned int offset,
4458 struct lnet_msg *msg;
4459 struct lnet_libmd *md;
4462 struct lnet_rsp_tracker *rspt = NULL;
4464 LASSERT(the_lnet.ln_refcount > 0);
4466 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4467 fail_peer(target.nid, 1)) { /* shall we now? */
4468 CERROR("Dropping PUT to %s: simulated failure\n",
4469 libcfs_id2str(target));
4473 msg = lnet_msg_alloc();
4475 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4476 libcfs_id2str(target));
4479 msg->msg_vmflush = !!memory_pressure_get();
4481 cpt = lnet_cpt_of_cookie(mdh.cookie);
4483 if (ack == LNET_ACK_REQ) {
4484 rspt = lnet_rspt_alloc(cpt);
4486 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4487 libcfs_id2str(target));
4490 INIT_LIST_HEAD(&rspt->rspt_on_list);
4495 md = lnet_handle2md(&mdh);
4496 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4497 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4498 match_bits, portal, libcfs_id2str(target),
4499 md == NULL ? -1 : md->md_threshold);
4500 if (md != NULL && md->md_me != NULL)
4501 CERROR("Source MD also attached to portal %d\n",
4502 md->md_me->me_portal);
4503 lnet_res_unlock(cpt);
4505 LIBCFS_FREE(rspt, sizeof(*rspt));
4510 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4512 lnet_msg_attach_md(msg, md, 0, 0);
4514 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4516 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4517 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4518 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4519 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4521 /* NB handles only looked up by creator (no flips) */
4522 if (ack == LNET_ACK_REQ) {
4523 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4524 the_lnet.ln_interface_cookie;
4525 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4526 md->md_lh.lh_cookie;
4528 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4529 LNET_WIRE_HANDLE_COOKIE_NONE;
4530 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4531 LNET_WIRE_HANDLE_COOKIE_NONE;
4534 lnet_res_unlock(cpt);
4536 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4538 if (ack == LNET_ACK_REQ)
4539 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4541 rc = lnet_send(self, msg, LNET_NID_ANY);
4543 CNETERR("Error sending PUT to %s: %d\n",
4544 libcfs_id2str(target), rc);
4545 msg->msg_no_resend = true;
4546 lnet_detach_rsp_tracker(msg->msg_md, cpt);
4547 lnet_finalize(msg, rc);
4550 /* completion will be signalled by an event */
4553 EXPORT_SYMBOL(LNetPut);
4556 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
4557 * returns a msg for the LND to pass to lnet_finalize() when the sink
4558 * data has been received.
4560 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4561 * lnet_finalize() is called on it, so the LND must call this first
4564 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4566 struct lnet_msg *msg = lnet_msg_alloc();
4567 struct lnet_libmd *getmd = getmsg->msg_md;
4568 struct lnet_process_id peer_id = getmsg->msg_target;
4571 LASSERT(!getmsg->msg_target_is_router);
4572 LASSERT(!getmsg->msg_routing);
4575 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4576 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4580 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4583 LASSERT(getmd->md_refcount > 0);
4585 if (getmd->md_threshold == 0) {
4586 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4587 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4589 lnet_res_unlock(cpt);
4593 LASSERT(getmd->md_offset == 0);
4595 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4596 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4598 /* setup information for lnet_build_msg_event */
4599 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4600 msg->msg_from = peer_id.nid;
4601 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4602 msg->msg_hdr.src_nid = peer_id.nid;
4603 msg->msg_hdr.payload_length = getmd->md_length;
4604 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4606 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4607 lnet_res_unlock(cpt);
4609 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4612 lnet_msg_commit(msg, cpt);
4613 lnet_net_unlock(cpt);
4615 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4620 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4623 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4624 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4625 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4627 lnet_net_unlock(cpt);
4634 EXPORT_SYMBOL(lnet_create_reply_msg);
4637 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4640 /* Set the REPLY length, now the RDMA that elides the REPLY message has
4641 * completed and I know it. */
4642 LASSERT(reply != NULL);
4643 LASSERT(reply->msg_type == LNET_MSG_GET);
4644 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4646 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
4647 * the end of my buffer, I might as well be dead. */
4648 LASSERT(len <= reply->msg_ev.mlength);
4650 reply->msg_ev.mlength = len;
4652 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4655 * Initiate an asynchronous GET operation.
4657 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4658 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4659 * the target node in the REPLY has been written to local MD.
4661 * On the target node, an LNET_EVENT_GET is logged when the GET request
4662 * arrives and is accepted into a MD.
4664 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4665 * \param mdh A handle for the MD that describes the memory into which the
4666 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4668 * \retval 0 Success, and only in this case events will be generated
4669 * and logged to EQ (if it exists) of the MD.
4670 * \retval -EIO Simulated failure.
4671 * \retval -ENOMEM Memory allocation failure.
4672 * \retval -ENOENT Invalid MD object.
4675 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4676 struct lnet_process_id target, unsigned int portal,
4677 __u64 match_bits, unsigned int offset, bool recovery)
4679 struct lnet_msg *msg;
4680 struct lnet_libmd *md;
4681 struct lnet_rsp_tracker *rspt;
4685 LASSERT(the_lnet.ln_refcount > 0);
4687 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4688 fail_peer(target.nid, 1)) /* shall we now? */
4690 CERROR("Dropping GET to %s: simulated failure\n",
4691 libcfs_id2str(target));
4695 msg = lnet_msg_alloc();
4697 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4698 libcfs_id2str(target));
4702 cpt = lnet_cpt_of_cookie(mdh.cookie);
4704 rspt = lnet_rspt_alloc(cpt);
4706 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4707 libcfs_id2str(target));
4710 INIT_LIST_HEAD(&rspt->rspt_on_list);
4712 msg->msg_recovery = recovery;
4716 md = lnet_handle2md(&mdh);
4717 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4718 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4719 match_bits, portal, libcfs_id2str(target),
4720 md == NULL ? -1 : md->md_threshold);
4721 if (md != NULL && md->md_me != NULL)
4722 CERROR("REPLY MD also attached to portal %d\n",
4723 md->md_me->me_portal);
4725 lnet_res_unlock(cpt);
4728 LIBCFS_FREE(rspt, sizeof(*rspt));
4732 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4734 lnet_msg_attach_md(msg, md, 0, 0);
4736 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
4738 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
4739 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
4740 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
4741 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
4743 /* NB handles only looked up by creator (no flips) */
4744 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
4745 the_lnet.ln_interface_cookie;
4746 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
4747 md->md_lh.lh_cookie;
4749 lnet_res_unlock(cpt);
4751 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4753 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4755 rc = lnet_send(self, msg, LNET_NID_ANY);
4757 CNETERR("Error sending GET to %s: %d\n",
4758 libcfs_id2str(target), rc);
4759 msg->msg_no_resend = true;
4760 lnet_detach_rsp_tracker(msg->msg_md, cpt);
4761 lnet_finalize(msg, rc);
4764 /* completion will be signalled by an event */
4767 EXPORT_SYMBOL(LNetGet);
4770 * Calculate distance to node at \a dstnid.
4772 * \param dstnid Target NID.
4773 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
4775 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
4778 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
4779 * local_nid_dist_zero is set, which is the default.
4780 * \retval positives Distance to target NID, i.e. number of hops plus one.
4781 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
4784 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
4786 struct list_head *e;
4787 struct lnet_ni *ni = NULL;
4788 struct lnet_remotenet *rnet;
4789 __u32 dstnet = LNET_NIDNET(dstnid);
4793 struct list_head *rn_list;
4795 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
4796 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
4797 * keep order 0 free for 0@lo and order 1 free for a local NID
4800 LASSERT(the_lnet.ln_refcount > 0);
4802 cpt = lnet_net_lock_current();
4804 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
4805 if (ni->ni_nid == dstnid) {
4806 if (srcnidp != NULL)
4808 if (orderp != NULL) {
4809 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
4814 lnet_net_unlock(cpt);
4816 return local_nid_dist_zero ? 0 : 1;
4819 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
4820 /* Check if ni was originally created in
4821 * current net namespace.
4822 * If not, assign order above 0xffff0000,
4823 * to make this ni not a priority. */
4824 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
4825 order += 0xffff0000;
4827 if (srcnidp != NULL)
4828 *srcnidp = ni->ni_nid;
4831 lnet_net_unlock(cpt);
4838 rn_list = lnet_net2rnethash(dstnet);
4839 list_for_each(e, rn_list) {
4840 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
4842 if (rnet->lrn_net == dstnet) {
4843 struct lnet_route *route;
4844 struct lnet_route *shortest = NULL;
4845 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
4848 LASSERT(!list_empty(&rnet->lrn_routes));
4850 list_for_each_entry(route, &rnet->lrn_routes,
4852 route_hops = route->lr_hops;
4853 if (route_hops == LNET_UNDEFINED_HOPS)
4855 if (shortest == NULL ||
4856 route_hops < shortest_hops) {
4858 shortest_hops = route_hops;
4862 LASSERT(shortest != NULL);
4863 hops = shortest_hops;
4864 if (srcnidp != NULL) {
4865 ni = lnet_get_next_ni_locked(
4866 shortest->lr_gateway->lpni_net,
4868 *srcnidp = ni->ni_nid;
4872 lnet_net_unlock(cpt);
4878 lnet_net_unlock(cpt);
4879 return -EHOSTUNREACH;
4881 EXPORT_SYMBOL(LNetDist);