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));
961 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_DROPPED;
962 lnet_finalize(msg, -EHOSTUNREACH);
966 return -EHOSTUNREACH;
969 if (msg->msg_md != NULL &&
970 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
971 lnet_net_unlock(cpt);
973 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
974 "called on the MD/ME.\n",
975 libcfs_id2str(msg->msg_target));
977 msg->msg_no_resend = true;
978 lnet_finalize(msg, -ECANCELED);
985 if (!msg->msg_peertxcredit) {
986 spin_lock(&lp->lpni_lock);
987 LASSERT((lp->lpni_txcredits < 0) ==
988 !list_empty(&lp->lpni_txq));
990 msg->msg_peertxcredit = 1;
991 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
992 lp->lpni_txcredits--;
994 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
995 lp->lpni_mintxcredits = lp->lpni_txcredits;
997 if (lp->lpni_txcredits < 0) {
998 msg->msg_tx_delayed = 1;
999 list_add_tail(&msg->msg_list, &lp->lpni_txq);
1000 spin_unlock(&lp->lpni_lock);
1001 return LNET_CREDIT_WAIT;
1003 spin_unlock(&lp->lpni_lock);
1006 if (!msg->msg_txcredit) {
1007 LASSERT((tq->tq_credits < 0) ==
1008 !list_empty(&tq->tq_delayed));
1010 msg->msg_txcredit = 1;
1012 atomic_dec(&ni->ni_tx_credits);
1014 if (tq->tq_credits < tq->tq_credits_min)
1015 tq->tq_credits_min = tq->tq_credits;
1017 if (tq->tq_credits < 0) {
1018 msg->msg_tx_delayed = 1;
1019 list_add_tail(&msg->msg_list, &tq->tq_delayed);
1020 return LNET_CREDIT_WAIT;
1024 /* unset the tx_delay flag as we're going to send it now */
1025 msg->msg_tx_delayed = 0;
1028 lnet_net_unlock(cpt);
1029 lnet_ni_send(ni, msg);
1032 return LNET_CREDIT_OK;
1036 static struct lnet_rtrbufpool *
1037 lnet_msg2bufpool(struct lnet_msg *msg)
1039 struct lnet_rtrbufpool *rbp;
1042 LASSERT(msg->msg_rx_committed);
1044 cpt = msg->msg_rx_cpt;
1045 rbp = &the_lnet.ln_rtrpools[cpt][0];
1047 LASSERT(msg->msg_len <= LNET_MTU);
1048 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
1050 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
1057 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
1059 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
1060 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
1061 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
1062 * received or OK to receive */
1063 struct lnet_peer_ni *lp = msg->msg_rxpeer;
1064 struct lnet_rtrbufpool *rbp;
1065 struct lnet_rtrbuf *rb;
1067 LASSERT (msg->msg_iov == NULL);
1068 LASSERT (msg->msg_kiov == NULL);
1069 LASSERT (msg->msg_niov == 0);
1070 LASSERT (msg->msg_routing);
1071 LASSERT (msg->msg_receiving);
1072 LASSERT (!msg->msg_sending);
1074 /* non-lnet_parse callers only receive delayed messages */
1075 LASSERT(!do_recv || msg->msg_rx_delayed);
1077 if (!msg->msg_peerrtrcredit) {
1078 spin_lock(&lp->lpni_lock);
1079 LASSERT((lp->lpni_rtrcredits < 0) ==
1080 !list_empty(&lp->lpni_rtrq));
1082 msg->msg_peerrtrcredit = 1;
1083 lp->lpni_rtrcredits--;
1084 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
1085 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
1087 if (lp->lpni_rtrcredits < 0) {
1088 /* must have checked eager_recv before here */
1089 LASSERT(msg->msg_rx_ready_delay);
1090 msg->msg_rx_delayed = 1;
1091 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
1092 spin_unlock(&lp->lpni_lock);
1093 return LNET_CREDIT_WAIT;
1095 spin_unlock(&lp->lpni_lock);
1098 rbp = lnet_msg2bufpool(msg);
1100 if (!msg->msg_rtrcredit) {
1101 msg->msg_rtrcredit = 1;
1103 if (rbp->rbp_credits < rbp->rbp_mincredits)
1104 rbp->rbp_mincredits = rbp->rbp_credits;
1106 if (rbp->rbp_credits < 0) {
1107 /* must have checked eager_recv before here */
1108 LASSERT(msg->msg_rx_ready_delay);
1109 msg->msg_rx_delayed = 1;
1110 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1111 return LNET_CREDIT_WAIT;
1115 LASSERT(!list_empty(&rbp->rbp_bufs));
1116 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1117 list_del(&rb->rb_list);
1119 msg->msg_niov = rbp->rbp_npages;
1120 msg->msg_kiov = &rb->rb_kiov[0];
1122 /* unset the msg-rx_delayed flag since we're receiving the message */
1123 msg->msg_rx_delayed = 0;
1126 int cpt = msg->msg_rx_cpt;
1128 lnet_net_unlock(cpt);
1129 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1130 0, msg->msg_len, msg->msg_len);
1133 return LNET_CREDIT_OK;
1137 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1139 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1140 struct lnet_ni *txni = msg->msg_txni;
1141 struct lnet_msg *msg2;
1143 if (msg->msg_txcredit) {
1144 struct lnet_ni *ni = msg->msg_txni;
1145 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1147 /* give back NI txcredits */
1148 msg->msg_txcredit = 0;
1150 LASSERT((tq->tq_credits < 0) ==
1151 !list_empty(&tq->tq_delayed));
1154 atomic_inc(&ni->ni_tx_credits);
1155 if (tq->tq_credits <= 0) {
1156 msg2 = list_entry(tq->tq_delayed.next,
1157 struct lnet_msg, msg_list);
1158 list_del(&msg2->msg_list);
1160 LASSERT(msg2->msg_txni == ni);
1161 LASSERT(msg2->msg_tx_delayed);
1162 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1164 (void) lnet_post_send_locked(msg2, 1);
1168 if (msg->msg_peertxcredit) {
1169 /* give back peer txcredits */
1170 msg->msg_peertxcredit = 0;
1172 spin_lock(&txpeer->lpni_lock);
1173 LASSERT((txpeer->lpni_txcredits < 0) ==
1174 !list_empty(&txpeer->lpni_txq));
1176 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1177 LASSERT(txpeer->lpni_txqnob >= 0);
1179 txpeer->lpni_txcredits++;
1180 if (txpeer->lpni_txcredits <= 0) {
1183 msg2 = list_entry(txpeer->lpni_txq.next,
1184 struct lnet_msg, msg_list);
1185 list_del(&msg2->msg_list);
1186 spin_unlock(&txpeer->lpni_lock);
1188 LASSERT(msg2->msg_txpeer == txpeer);
1189 LASSERT(msg2->msg_tx_delayed);
1191 msg2_cpt = msg2->msg_tx_cpt;
1194 * The msg_cpt can be different from the msg2_cpt
1195 * so we need to make sure we lock the correct cpt
1197 * Once we call lnet_post_send_locked() it is no
1198 * longer safe to access msg2, since it could've
1199 * been freed by lnet_finalize(), but we still
1200 * need to relock the correct cpt, so we cache the
1201 * msg2_cpt for the purpose of the check that
1202 * follows the call to lnet_pose_send_locked().
1204 if (msg2_cpt != msg->msg_tx_cpt) {
1205 lnet_net_unlock(msg->msg_tx_cpt);
1206 lnet_net_lock(msg2_cpt);
1208 (void) lnet_post_send_locked(msg2, 1);
1209 if (msg2_cpt != msg->msg_tx_cpt) {
1210 lnet_net_unlock(msg2_cpt);
1211 lnet_net_lock(msg->msg_tx_cpt);
1214 spin_unlock(&txpeer->lpni_lock);
1219 msg->msg_txni = NULL;
1220 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1223 if (txpeer != NULL) {
1224 msg->msg_txpeer = NULL;
1225 lnet_peer_ni_decref_locked(txpeer);
1230 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1232 struct lnet_msg *msg;
1234 if (list_empty(&rbp->rbp_msgs))
1236 msg = list_entry(rbp->rbp_msgs.next,
1237 struct lnet_msg, msg_list);
1238 list_del(&msg->msg_list);
1240 (void)lnet_post_routed_recv_locked(msg, 1);
1244 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1246 struct lnet_msg *msg;
1247 struct lnet_msg *tmp;
1249 lnet_net_unlock(cpt);
1251 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1252 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1253 0, 0, 0, msg->msg_hdr.payload_length);
1254 list_del_init(&msg->msg_list);
1255 msg->msg_no_resend = true;
1256 lnet_finalize(msg, -ECANCELED);
1263 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1265 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1266 struct lnet_ni *rxni = msg->msg_rxni;
1267 struct lnet_msg *msg2;
1269 if (msg->msg_rtrcredit) {
1270 /* give back global router credits */
1271 struct lnet_rtrbuf *rb;
1272 struct lnet_rtrbufpool *rbp;
1274 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1275 * there until it gets one allocated, or aborts the wait
1277 LASSERT(msg->msg_kiov != NULL);
1279 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1282 msg->msg_kiov = NULL;
1283 msg->msg_rtrcredit = 0;
1285 LASSERT(rbp == lnet_msg2bufpool(msg));
1287 LASSERT((rbp->rbp_credits > 0) ==
1288 !list_empty(&rbp->rbp_bufs));
1290 /* If routing is now turned off, we just drop this buffer and
1291 * don't bother trying to return credits. */
1292 if (!the_lnet.ln_routing) {
1293 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1297 /* It is possible that a user has lowered the desired number of
1298 * buffers in this pool. Make sure we never put back
1299 * more buffers than the stated number. */
1300 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1301 /* Discard this buffer so we don't have too
1303 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1304 rbp->rbp_nbuffers--;
1306 list_add(&rb->rb_list, &rbp->rbp_bufs);
1308 if (rbp->rbp_credits <= 0)
1309 lnet_schedule_blocked_locked(rbp);
1314 if (msg->msg_peerrtrcredit) {
1315 /* give back peer router credits */
1316 msg->msg_peerrtrcredit = 0;
1318 spin_lock(&rxpeer->lpni_lock);
1319 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1320 !list_empty(&rxpeer->lpni_rtrq));
1322 rxpeer->lpni_rtrcredits++;
1324 /* drop all messages which are queued to be routed on that
1326 if (!the_lnet.ln_routing) {
1327 struct list_head drop;
1328 INIT_LIST_HEAD(&drop);
1329 list_splice_init(&rxpeer->lpni_rtrq, &drop);
1330 spin_unlock(&rxpeer->lpni_lock);
1331 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1332 } else if (rxpeer->lpni_rtrcredits <= 0) {
1333 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1334 struct lnet_msg, msg_list);
1335 list_del(&msg2->msg_list);
1336 spin_unlock(&rxpeer->lpni_lock);
1337 (void) lnet_post_routed_recv_locked(msg2, 1);
1339 spin_unlock(&rxpeer->lpni_lock);
1343 msg->msg_rxni = NULL;
1344 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1346 if (rxpeer != NULL) {
1347 msg->msg_rxpeer = NULL;
1348 lnet_peer_ni_decref_locked(rxpeer);
1353 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1355 if (p1->lpni_txqnob < p2->lpni_txqnob)
1358 if (p1->lpni_txqnob > p2->lpni_txqnob)
1361 if (p1->lpni_txcredits > p2->lpni_txcredits)
1364 if (p1->lpni_txcredits < p2->lpni_txcredits)
1371 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1373 struct lnet_peer_ni *p1 = r1->lr_gateway;
1374 struct lnet_peer_ni *p2 = r2->lr_gateway;
1375 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1376 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1379 if (r1->lr_priority < r2->lr_priority)
1382 if (r1->lr_priority > r2->lr_priority)
1385 if (r1_hops < r2_hops)
1388 if (r1_hops > r2_hops)
1391 rc = lnet_compare_peers(p1, p2);
1395 if (r1->lr_seq - r2->lr_seq <= 0)
1401 static struct lnet_peer_ni *
1402 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1405 struct lnet_remotenet *rnet;
1406 struct lnet_route *route;
1407 struct lnet_route *best_route;
1408 struct lnet_route *last_route;
1409 struct lnet_peer_ni *lpni_best;
1410 struct lnet_peer_ni *lp;
1413 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1414 * rtr_nid nid, otherwise find the best gateway I can use */
1416 rnet = lnet_find_rnet_locked(remote_net);
1421 best_route = last_route = NULL;
1422 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1423 lp = route->lr_gateway;
1425 if (!lnet_is_route_alive(route))
1428 if (net != NULL && lp->lpni_net != net)
1431 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1434 if (lpni_best == NULL) {
1435 best_route = last_route = route;
1440 /* no protection on below fields, but it's harmless */
1441 if (last_route->lr_seq - route->lr_seq < 0)
1444 rc = lnet_compare_routes(route, best_route);
1452 /* set sequence number on the best router to the latest sequence + 1
1453 * so we can round-robin all routers, it's race and inaccurate but
1454 * harmless and functional */
1455 if (best_route != NULL)
1456 best_route->lr_seq = last_route->lr_seq + 1;
1460 static struct lnet_ni *
1461 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1462 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1465 struct lnet_ni *ni = NULL;
1466 unsigned int shortest_distance;
1471 * If there is no peer_ni that we can send to on this network,
1472 * then there is no point in looking for a new best_ni here.
1474 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1477 if (best_ni == NULL) {
1478 shortest_distance = UINT_MAX;
1479 best_credits = INT_MIN;
1482 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1483 best_ni->ni_dev_cpt);
1484 best_credits = atomic_read(&best_ni->ni_tx_credits);
1485 best_healthv = atomic_read(&best_ni->ni_healthv);
1488 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1489 unsigned int distance;
1494 ni_credits = atomic_read(&ni->ni_tx_credits);
1495 ni_healthv = atomic_read(&ni->ni_healthv);
1496 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1499 * calculate the distance from the CPT on which
1500 * the message memory is allocated to the CPT of
1501 * the NI's physical device
1503 distance = cfs_cpt_distance(lnet_cpt_table(),
1507 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1508 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1509 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1510 : "not seleced", best_credits, shortest_distance,
1511 (best_ni) ? best_ni->ni_seq : 0);
1514 * All distances smaller than the NUMA range
1515 * are treated equally.
1517 if (distance < lnet_numa_range)
1518 distance = lnet_numa_range;
1521 * Select on health, shorter distance, available
1522 * credits, then round-robin.
1526 } else if (ni_healthv < best_healthv) {
1528 } else if (ni_healthv > best_healthv) {
1529 best_healthv = ni_healthv;
1531 * If we're going to prefer this ni because it's
1532 * the healthiest, then we should set the
1533 * shortest_distance in the algorithm in case
1534 * there are multiple NIs with the same health but
1535 * different distances.
1537 if (distance < shortest_distance)
1538 shortest_distance = distance;
1539 } else if (distance > shortest_distance) {
1541 } else if (distance < shortest_distance) {
1542 shortest_distance = distance;
1543 } else if (ni_credits < best_credits) {
1545 } else if (ni_credits == best_credits) {
1546 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1550 best_credits = ni_credits;
1553 CDEBUG(D_NET, "selected best_ni %s\n",
1554 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1560 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1561 * because such traffic is required to perform discovery. We therefore
1562 * exclude all GET and PUT on that portal. We also exclude all ACK and
1563 * REPLY traffic, but that is because the portal is not tracked in the
1564 * message structure for these message types. We could restrict this
1565 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1568 lnet_msg_discovery(struct lnet_msg *msg)
1570 if (msg->msg_type == LNET_MSG_PUT) {
1571 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1573 } else if (msg->msg_type == LNET_MSG_GET) {
1574 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1580 #define SRC_SPEC 0x0001
1581 #define SRC_ANY 0x0002
1582 #define LOCAL_DST 0x0004
1583 #define REMOTE_DST 0x0008
1584 #define MR_DST 0x0010
1585 #define NMR_DST 0x0020
1586 #define SND_RESP 0x0040
1588 /* The following to defines are used for return codes */
1589 #define REPEAT_SEND 0x1000
1590 #define PASS_THROUGH 0x2000
1592 /* The different cases lnet_select pathway needs to handle */
1593 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1594 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1595 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1596 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1597 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1598 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1599 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1600 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1603 lnet_handle_send(struct lnet_send_data *sd)
1605 struct lnet_ni *best_ni = sd->sd_best_ni;
1606 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1607 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1608 struct lnet_msg *msg = sd->sd_msg;
1610 __u32 send_case = sd->sd_send_case;
1612 __u32 routing = send_case & REMOTE_DST;
1615 * Increment sequence number of the selected peer so that we
1616 * pick the next one in Round Robin.
1618 best_lpni->lpni_seq++;
1621 * grab a reference on the peer_ni so it sticks around even if
1622 * we need to drop and relock the lnet_net_lock below.
1624 lnet_peer_ni_addref_locked(best_lpni);
1627 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1628 * message. This ensures that we get a CPT that is correct for
1629 * the NI when the NI has been restricted to a subset of all CPTs.
1630 * If the selected CPT differs from the one currently locked, we
1631 * must unlock and relock the lnet_net_lock(), and then check whether
1632 * the configuration has changed. We don't have a hold on the best_ni
1633 * yet, and it may have vanished.
1635 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1636 if (sd->sd_cpt != cpt2) {
1637 __u32 seq = lnet_get_dlc_seq_locked();
1638 lnet_net_unlock(sd->sd_cpt);
1640 lnet_net_lock(sd->sd_cpt);
1641 if (seq != lnet_get_dlc_seq_locked()) {
1642 lnet_peer_ni_decref_locked(best_lpni);
1648 * store the best_lpni in the message right away to avoid having
1649 * to do the same operation under different conditions
1651 msg->msg_txpeer = best_lpni;
1652 msg->msg_txni = best_ni;
1655 * grab a reference for the best_ni since now it's in use in this
1656 * send. The reference will be dropped in lnet_finalize()
1658 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1661 * Always set the target.nid to the best peer picked. Either the
1662 * NID will be one of the peer NIDs selected, or the same NID as
1663 * what was originally set in the target or it will be the NID of
1664 * a router if this message should be routed
1666 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1669 * lnet_msg_commit assigns the correct cpt to the message, which
1670 * is used to decrement the correct refcount on the ni when it's
1671 * time to return the credits
1673 lnet_msg_commit(msg, sd->sd_cpt);
1676 * If we are routing the message then we keep the src_nid that was
1677 * set by the originator. If we are not routing then we are the
1678 * originator and set it here.
1680 if (!msg->msg_routing)
1681 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1684 msg->msg_target_is_router = 1;
1685 msg->msg_target.pid = LNET_PID_LUSTRE;
1687 * since we're routing we want to ensure that the
1688 * msg_hdr.dest_nid is set to the final destination. When
1689 * the router receives this message it knows how to route
1692 * final_dst_lpni is set at the beginning of the
1693 * lnet_select_pathway() function and is never changed.
1694 * It's safe to use it here.
1696 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1699 * if we're not routing set the dest_nid to the best peer
1700 * ni NID that we picked earlier in the algorithm.
1702 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1705 rc = lnet_post_send_locked(msg, 0);
1708 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1709 libcfs_nid2str(msg->msg_hdr.src_nid),
1710 libcfs_nid2str(msg->msg_txni->ni_nid),
1711 libcfs_nid2str(sd->sd_src_nid),
1712 libcfs_nid2str(msg->msg_hdr.dest_nid),
1713 libcfs_nid2str(sd->sd_dst_nid),
1714 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1715 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1720 static struct lnet_peer_ni *
1721 lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer,
1722 struct lnet_peer_net *peer_net)
1725 * Look at the peer NIs for the destination peer that connect
1726 * to the chosen net. If a peer_ni is preferred when using the
1727 * best_ni to communicate, we use that one. If there is no
1728 * preferred peer_ni, or there are multiple preferred peer_ni,
1729 * the available transmit credits are used. If the transmit
1730 * credits are equal, we round-robin over the peer_ni.
1732 struct lnet_peer_ni *lpni = NULL;
1733 struct lnet_peer_ni *best_lpni = NULL;
1734 struct lnet_ni *best_ni = sd->sd_best_ni;
1735 lnet_nid_t dst_nid = sd->sd_dst_nid;
1736 int best_lpni_credits = INT_MIN;
1737 bool preferred = false;
1739 int best_lpni_healthv = 0;
1742 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1744 * if the best_ni we've chosen aleady has this lpni
1745 * preferred, then let's use it
1747 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1750 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1752 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1753 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1756 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1757 libcfs_nid2str(lpni->lpni_nid),
1758 lpni->lpni_txcredits, best_lpni_credits,
1759 lpni->lpni_seq, best_lpni->lpni_seq);
1761 /* pick the healthiest peer ni */
1762 if (lpni_healthv < best_lpni_healthv) {
1764 } else if (lpni_healthv > best_lpni_healthv) {
1765 best_lpni_healthv = lpni_healthv;
1766 /* if this is a preferred peer use it */
1767 } else if (!preferred && ni_is_pref) {
1769 } else if (preferred && !ni_is_pref) {
1771 * this is not the preferred peer so let's ignore
1775 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1777 * We already have a peer that has more credits
1778 * available than this one. No need to consider
1779 * this peer further.
1782 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1784 * The best peer found so far and the current peer
1785 * have the same number of available credits let's
1786 * make sure to select between them using Round
1790 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1796 best_lpni_credits = lpni->lpni_txcredits;
1799 /* if we still can't find a peer ni then we can't reach it */
1801 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1802 LNET_NIDNET(dst_nid);
1803 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1804 libcfs_net2str(net_id));
1808 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1809 libcfs_nid2str(best_lpni->lpni_nid));
1815 * Prerequisite: the best_ni should already be set in the sd
1817 static inline struct lnet_peer_ni *
1818 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1821 struct lnet_peer_net *peer_net;
1824 * The gateway is Multi-Rail capable so now we must select the
1827 peer_net = lnet_peer_get_net_locked(peer, net_id);
1830 CERROR("gateway peer %s has no NI on net %s\n",
1831 libcfs_nid2str(peer->lp_primary_nid),
1832 libcfs_net2str(net_id));
1836 return lnet_select_peer_ni(sd, peer, peer_net);
1840 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1842 if (sd->sd_send_case & NMR_DST &&
1843 sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1844 sd->sd_msg->msg_type != LNET_MSG_ACK &&
1845 sd->sd_best_lpni->lpni_pref_nnids == 0) {
1846 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1847 libcfs_nid2str(sd->sd_best_ni->ni_nid),
1848 libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1849 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1850 sd->sd_best_ni->ni_nid);
1859 * use the source and destination NIDs as the pathway
1862 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1864 /* the destination lpni is set before we get here. */
1867 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1868 if (!sd->sd_best_ni) {
1869 CERROR("Can't send to %s: src %s is not a "
1870 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1871 libcfs_nid2str(sd->sd_src_nid));
1876 * the preferred NID will only be set for NMR peers
1878 lnet_set_non_mr_pref_nid(sd);
1880 return lnet_handle_send(sd);
1888 * Run the selection algorithm on the peer NIs unless we're sending
1889 * a response, in this case just send to the destination
1892 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1894 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1895 if (!sd->sd_best_ni) {
1896 CERROR("Can't send to %s: src %s is not a "
1897 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1898 libcfs_nid2str(sd->sd_src_nid));
1903 * only run the selection algorithm to pick the peer_ni if we're
1904 * sending a GET or a PUT. Responses are sent to the same
1905 * destination NID provided.
1907 if (!(sd->sd_send_case & SND_RESP)) {
1909 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1910 sd->sd_best_ni->ni_net->net_id);
1913 if (sd->sd_best_lpni)
1914 return lnet_handle_send(sd);
1916 CERROR("can't send to %s. no NI on %s\n",
1917 libcfs_nid2str(sd->sd_dst_nid),
1918 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1920 return -EHOSTUNREACH;
1924 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1925 struct lnet_peer *peer,
1926 struct lnet_peer_net *peer_net,
1930 struct lnet_net *local_net;
1931 struct lnet_ni *best_ni;
1933 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1938 * Iterate through the NIs in this local Net and select
1939 * the NI to send from. The selection is determined by
1940 * these 3 criterion in the following priority:
1942 * 2. NI available credits
1945 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1946 peer, peer_net, cpt);
1948 if (incr_seq && best_ni)
1955 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1957 struct lnet_peer_ni **gw_lpni,
1958 struct lnet_peer **gw_peer)
1960 struct lnet_peer_ni *gw;
1961 lnet_nid_t src_nid = sd->sd_src_nid;
1963 gw = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
1966 CERROR("no route to %s from %s\n",
1967 libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1968 return -EHOSTUNREACH;
1971 /* get the peer of the gw_ni */
1972 LASSERT(gw->lpni_peer_net);
1973 LASSERT(gw->lpni_peer_net->lpn_peer);
1975 *gw_peer = gw->lpni_peer_net->lpn_peer;
1977 if (!sd->sd_best_ni)
1978 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, *gw_peer,
1983 if (!sd->sd_best_ni) {
1984 CERROR("Internal Error. Expected local ni on %s "
1985 "but non found :%s\n",
1986 libcfs_net2str(gw->lpni_peer_net->lpn_net_id),
1987 libcfs_nid2str(sd->sd_src_nid));
1992 * if gw is MR let's find its best peer_ni
1994 if (lnet_peer_is_multi_rail(*gw_peer)) {
1995 gw = lnet_find_best_lpni_on_net(sd, *gw_peer,
1996 sd->sd_best_ni->ni_net->net_id);
1998 * We've already verified that the gw has an NI on that
1999 * desired net, but we're not finding it. Something is
2003 CERROR("Internal Error. Route expected to %s from %s\n",
2004 libcfs_nid2str(dst_nid),
2005 libcfs_nid2str(src_nid));
2020 * Remote destination
2021 * Non-MR destination
2025 * Remote destination
2028 * The handling of these two cases is similar. Even though the destination
2029 * can be MR or non-MR, we'll deal directly with the router.
2032 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2035 struct lnet_peer_ni *gw_lpni = NULL;
2036 struct lnet_peer *gw_peer = NULL;
2039 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2040 if (!sd->sd_best_ni) {
2041 CERROR("Can't send to %s: src %s is not a "
2042 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2043 libcfs_nid2str(sd->sd_src_nid));
2047 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2052 if (sd->sd_send_case & NMR_DST)
2054 * since the final destination is non-MR let's set its preferred
2055 * NID before we send
2057 lnet_set_non_mr_pref_nid(sd);
2060 * We're going to send to the gw found so let's set its
2063 sd->sd_peer = gw_peer;
2064 sd->sd_best_lpni = gw_lpni;
2066 return lnet_handle_send(sd);
2070 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
2072 struct lnet_peer_net *peer_net = NULL;
2073 struct lnet_ni *best_ni = NULL;
2076 * The peer can have multiple interfaces, some of them can be on
2077 * the local network and others on a routed network. We should
2078 * prefer the local network. However if the local network is not
2079 * available then we need to try the routed network
2082 /* go through all the peer nets and find the best_ni */
2083 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2085 * The peer's list of nets can contain non-local nets. We
2086 * want to only examine the local ones.
2088 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2090 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2091 peer_net, md_cpt, false);
2095 /* increment sequence number so we can round robin */
2101 static struct lnet_ni *
2102 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2104 struct lnet_ni *best_ni = NULL;
2105 struct lnet_peer_net *peer_net;
2106 struct lnet_peer *peer = sd->sd_peer;
2107 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2108 struct lnet_peer_ni *lpni;
2109 int cpt = sd->sd_cpt;
2112 * We must use a consistent source address when sending to a
2113 * non-MR peer. However, a non-MR peer can have multiple NIDs
2114 * on multiple networks, and we may even need to talk to this
2115 * peer on multiple networks -- certain types of
2116 * load-balancing configuration do this.
2118 * So we need to pick the NI the peer prefers for this
2119 * particular network.
2122 /* Get the target peer_ni */
2123 peer_net = lnet_peer_get_net_locked(peer,
2124 LNET_NIDNET(best_lpni->lpni_nid));
2125 LASSERT(peer_net != NULL);
2126 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2128 if (lpni->lpni_pref_nnids == 0)
2130 LASSERT(lpni->lpni_pref_nnids == 1);
2131 best_ni = lnet_nid2ni_locked(
2132 lpni->lpni_pref.nid, cpt);
2139 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2141 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2143 struct lnet_ni *best_ni = NULL;
2144 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2147 * We must use a consistent source address when sending to a
2148 * non-MR peer. However, a non-MR peer can have multiple NIDs
2149 * on multiple networks, and we may even need to talk to this
2150 * peer on multiple networks -- certain types of
2151 * load-balancing configuration do this.
2153 * So we need to pick the NI the peer prefers for this
2154 * particular network.
2157 best_ni = lnet_find_existing_preferred_best_ni(sd);
2159 /* if best_ni is still not set just pick one */
2162 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2163 sd->sd_best_lpni->lpni_peer_net,
2164 sd->sd_md_cpt, true);
2165 /* If there is no best_ni we don't have a route */
2167 CERROR("no path to %s from net %s\n",
2168 libcfs_nid2str(best_lpni->lpni_nid),
2169 libcfs_net2str(best_lpni->lpni_net->net_id));
2170 return -EHOSTUNREACH;
2174 sd->sd_best_ni = best_ni;
2176 /* Set preferred NI if necessary. */
2177 lnet_set_non_mr_pref_nid(sd);
2184 * Source not specified
2188 * always use the same source NID for NMR peers
2189 * If we've talked to that peer before then we already have a preferred
2190 * source NI associated with it. Otherwise, we select a preferred local NI
2191 * and store it in the peer
2194 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2198 /* sd->sd_best_lpni is already set to the final destination */
2201 * At this point we should've created the peer ni and peer. If we
2202 * can't find it, then something went wrong. Instead of assert
2203 * output a relevant message and fail the send
2205 if (!sd->sd_best_lpni) {
2206 CERROR("Internal fault. Unable to send msg %s to %s. "
2208 lnet_msgtyp2str(sd->sd_msg->msg_type),
2209 libcfs_nid2str(sd->sd_dst_nid));
2213 rc = lnet_select_preferred_best_ni(sd);
2215 rc = lnet_handle_send(sd);
2221 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2224 * NOTE we've already handled the remote peer case. So we only
2225 * need to worry about the local case here.
2227 * if we're sending a response, ACK or reply, we need to send it
2228 * to the destination NID given to us. At this point we already
2229 * have the peer_ni we're suppose to send to, so just find the
2230 * best_ni on the peer net and use that. Since we're sending to an
2231 * MR peer then we can just run the selection algorithm on our
2232 * local NIs and pick the best one.
2234 if (sd->sd_send_case & SND_RESP) {
2236 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2237 sd->sd_best_lpni->lpni_peer_net,
2238 sd->sd_md_cpt, true);
2240 if (!sd->sd_best_ni) {
2242 * We're not going to deal with not able to send
2243 * a response to the provided final destination
2245 CERROR("Can't send response to %s. "
2246 "No local NI available\n",
2247 libcfs_nid2str(sd->sd_dst_nid));
2248 return -EHOSTUNREACH;
2251 return lnet_handle_send(sd);
2255 * If we get here that means we're sending a fresh request, PUT or
2256 * GET, so we need to run our standard selection algorithm.
2257 * First find the best local interface that's on any of the peer's
2260 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2262 if (sd->sd_best_ni) {
2264 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2265 sd->sd_best_ni->ni_net->net_id);
2268 * if we're successful in selecting a peer_ni on the local
2269 * network, then send to it. Otherwise fall through and
2270 * try and see if we can reach it over another routed
2273 if (sd->sd_best_lpni) {
2275 * in case we initially started with a routed
2276 * destination, let's reset to local
2278 sd->sd_send_case &= ~REMOTE_DST;
2279 sd->sd_send_case |= LOCAL_DST;
2280 return lnet_handle_send(sd);
2283 CERROR("Internal Error. Expected to have a best_lpni: "
2285 libcfs_nid2str(sd->sd_src_nid),
2286 libcfs_nid2str(sd->sd_dst_nid));
2292 * Peer doesn't have a local network. Let's see if there is
2293 * a remote network we can reach it on.
2295 return PASS_THROUGH;
2300 * Source NID not specified
2305 * Source NID not speified
2306 * Remote destination
2309 * In both of these cases if we're sending a response, ACK or REPLY, then
2310 * we need to send to the destination NID provided.
2312 * In the remote case let's deal with MR routers.
2317 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2320 struct lnet_peer *gw_peer = NULL;
2321 struct lnet_peer_ni *gw_lpni = NULL;
2324 * handle sending a response to a remote peer here so we don't
2325 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2327 if (sd->sd_send_case & REMOTE_DST &&
2328 sd->sd_send_case & SND_RESP) {
2329 struct lnet_peer_ni *gw;
2330 struct lnet_peer *gw_peer;
2332 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2335 CERROR("Can't send response to %s. "
2336 "No route available\n",
2337 libcfs_nid2str(sd->sd_dst_nid));
2338 return -EHOSTUNREACH;
2341 sd->sd_best_lpni = gw;
2342 sd->sd_peer = gw_peer;
2344 return lnet_handle_send(sd);
2348 * Even though the NID for the peer might not be on a local network,
2349 * since the peer is MR there could be other interfaces on the
2350 * local network. In that case we'd still like to prefer the local
2351 * network over the routed network. If we're unable to do that
2352 * then we select the best router among the different routed networks,
2353 * and if the router is MR then we can deal with it as such.
2355 rc = lnet_handle_any_mr_dsta(sd);
2356 if (rc != PASS_THROUGH)
2360 * TODO; One possible enhancement is to run the selection
2361 * algorithm on the peer. However for remote peers the credits are
2362 * not decremented, so we'll be basically going over the peer NIs
2363 * in round robin. An MR router will run the selection algorithm
2364 * on the next-hop interfaces.
2366 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2371 sd->sd_send_case &= ~LOCAL_DST;
2372 sd->sd_send_case |= REMOTE_DST;
2374 sd->sd_peer = gw_peer;
2375 sd->sd_best_lpni = gw_lpni;
2377 return lnet_handle_send(sd);
2381 * Source not specified
2382 * Remote destination
2385 * Must send to the specified peer NID using the same source NID that
2386 * we've used before. If it's the first time to talk to that peer then
2387 * find the source NI and assign it as preferred to that peer
2390 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2393 struct lnet_peer_ni *gw_lpni = NULL;
2394 struct lnet_peer *gw_peer = NULL;
2397 * Let's set if we have a preferred NI to talk to this NMR peer
2399 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2402 * find the router and that'll find the best NI if we didn't find
2405 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2411 * set the best_ni we've chosen as the preferred one for
2414 lnet_set_non_mr_pref_nid(sd);
2416 /* we'll be sending to the gw */
2417 sd->sd_best_lpni = gw_lpni;
2418 sd->sd_peer = gw_peer;
2420 return lnet_handle_send(sd);
2424 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2427 * turn off the SND_RESP bit.
2428 * It will be checked in the case handling
2430 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2432 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2433 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2434 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2435 (send_case & MR_DST) ? "MR: " : "NMR: ",
2436 libcfs_nid2str(sd->sd_dst_nid),
2437 (send_case & LOCAL_DST) ? "local" : "routed");
2439 switch (send_case) {
2441 * For all cases where the source is specified, we should always
2442 * use the destination NID, whether it's an MR destination or not,
2443 * since we're continuing a series of related messages for the
2446 case SRC_SPEC_LOCAL_NMR_DST:
2447 return lnet_handle_spec_local_nmr_dst(sd);
2448 case SRC_SPEC_LOCAL_MR_DST:
2449 return lnet_handle_spec_local_mr_dst(sd);
2450 case SRC_SPEC_ROUTER_NMR_DST:
2451 case SRC_SPEC_ROUTER_MR_DST:
2452 return lnet_handle_spec_router_dst(sd);
2453 case SRC_ANY_LOCAL_NMR_DST:
2454 return lnet_handle_any_local_nmr_dst(sd);
2455 case SRC_ANY_LOCAL_MR_DST:
2456 case SRC_ANY_ROUTER_MR_DST:
2457 return lnet_handle_any_mr_dst(sd);
2458 case SRC_ANY_ROUTER_NMR_DST:
2459 return lnet_handle_any_router_nmr_dst(sd);
2461 CERROR("Unknown send case\n");
2467 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2468 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2470 struct lnet_peer_ni *lpni;
2471 struct lnet_peer *peer;
2472 struct lnet_send_data send_data;
2475 __u32 send_case = 0;
2477 memset(&send_data, 0, sizeof(send_data));
2480 * get an initial CPT to use for locking. The idea here is not to
2481 * serialize the calls to select_pathway, so that as many
2482 * operations can run concurrently as possible. To do that we use
2483 * the CPT where this call is being executed. Later on when we
2484 * determine the CPT to use in lnet_message_commit, we switch the
2485 * lock and check if there was any configuration change. If none,
2486 * then we proceed, if there is, then we restart the operation.
2488 cpt = lnet_net_lock_current();
2490 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2491 if (md_cpt == CFS_CPT_ANY)
2497 * If we're being asked to send to the loopback interface, there
2498 * is no need to go through any selection. We can just shortcut
2499 * the entire process and send over lolnd
2501 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2502 /* No send credit hassles with LOLND */
2503 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
2504 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
2505 if (!msg->msg_routing)
2506 msg->msg_hdr.src_nid =
2507 cpu_to_le64(the_lnet.ln_loni->ni_nid);
2508 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
2509 lnet_msg_commit(msg, cpt);
2510 msg->msg_txni = the_lnet.ln_loni;
2511 lnet_net_unlock(cpt);
2513 return LNET_CREDIT_OK;
2517 * find an existing peer_ni, or create one and mark it as having been
2518 * created due to network traffic. This call will create the
2519 * peer->peer_net->peer_ni tree.
2521 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2523 lnet_net_unlock(cpt);
2524 return PTR_ERR(lpni);
2528 * Cache the original src_nid. If we need to resend the message
2529 * then we'll need to know whether the src_nid was originally
2530 * specified for this message. If it was originally specified,
2531 * then we need to keep using the same src_nid since it's
2532 * continuing the same sequence of messages.
2534 msg->msg_src_nid_param = src_nid;
2537 * Now that we have a peer_ni, check if we want to discover
2538 * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2539 * trigger discovery.
2541 peer = lpni->lpni_peer_net->lpn_peer;
2542 if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
2543 lnet_nid_t primary_nid;
2544 rc = lnet_discover_peer_locked(lpni, cpt, false);
2546 lnet_peer_ni_decref_locked(lpni);
2547 lnet_net_unlock(cpt);
2550 /* The peer may have changed. */
2551 peer = lpni->lpni_peer_net->lpn_peer;
2552 /* queue message and return */
2553 msg->msg_rtr_nid_param = rtr_nid;
2554 msg->msg_sending = 0;
2555 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2556 lnet_peer_ni_decref_locked(lpni);
2557 primary_nid = peer->lp_primary_nid;
2558 lnet_net_unlock(cpt);
2560 CDEBUG(D_NET, "%s pending discovery\n",
2561 libcfs_nid2str(primary_nid));
2563 return LNET_DC_WAIT;
2565 lnet_peer_ni_decref_locked(lpni);
2568 * Identify the different send cases
2570 if (src_nid == LNET_NID_ANY)
2571 send_case |= SRC_ANY;
2573 send_case |= SRC_SPEC;
2575 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2576 send_case |= LOCAL_DST;
2578 send_case |= REMOTE_DST;
2581 * if this is a non-MR peer or if we're recovering a peer ni then
2582 * let's consider this an NMR case so we can hit the destination
2585 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2586 send_case |= NMR_DST;
2588 send_case |= MR_DST;
2590 if (msg->msg_type == LNET_MSG_REPLY ||
2591 msg->msg_type == LNET_MSG_ACK)
2592 send_case |= SND_RESP;
2594 /* assign parameters to the send_data */
2595 send_data.sd_msg = msg;
2596 send_data.sd_rtr_nid = rtr_nid;
2597 send_data.sd_src_nid = src_nid;
2598 send_data.sd_dst_nid = dst_nid;
2599 send_data.sd_best_lpni = lpni;
2601 * keep a pointer to the final destination in case we're going to
2602 * route, so we'll need to access it later
2604 send_data.sd_final_dst_lpni = lpni;
2605 send_data.sd_peer = peer;
2606 send_data.sd_md_cpt = md_cpt;
2607 send_data.sd_cpt = cpt;
2608 send_data.sd_send_case = send_case;
2610 rc = lnet_handle_send_case_locked(&send_data);
2612 if (rc == REPEAT_SEND)
2615 lnet_net_unlock(send_data.sd_cpt);
2621 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2623 lnet_nid_t dst_nid = msg->msg_target.nid;
2627 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2628 * but we might want to use pre-determined router for ACK/REPLY
2631 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2632 LASSERT(msg->msg_txpeer == NULL);
2633 LASSERT(msg->msg_txni == NULL);
2634 LASSERT(!msg->msg_sending);
2635 LASSERT(!msg->msg_target_is_router);
2636 LASSERT(!msg->msg_receiving);
2638 msg->msg_sending = 1;
2640 LASSERT(!msg->msg_tx_committed);
2642 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2646 if (rc == LNET_CREDIT_OK)
2647 lnet_ni_send(msg->msg_txni, msg);
2649 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2653 enum lnet_mt_event_type {
2654 MT_TYPE_LOCAL_NI = 0,
2658 struct lnet_mt_event_info {
2659 enum lnet_mt_event_type mt_type;
2664 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2666 struct lnet_rsp_tracker *rspt;
2669 * msg has a refcount on the MD so the MD is not going away.
2670 * The rspt queue for the cpt is protected by
2671 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2674 if (!md->md_rspt_ptr) {
2675 lnet_res_unlock(cpt);
2678 rspt = md->md_rspt_ptr;
2679 md->md_rspt_ptr = NULL;
2682 LASSERT(rspt->rspt_cpt == cpt);
2685 * invalidate the handle to indicate that a response has been
2686 * received, which will then lead the monitor thread to clean up
2689 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2690 lnet_res_unlock(cpt);
2694 lnet_finalize_expired_responses(bool force)
2696 struct lnet_libmd *md;
2697 struct list_head local_queue;
2698 struct lnet_rsp_tracker *rspt, *tmp;
2701 if (the_lnet.ln_mt_rstq == NULL)
2704 cfs_cpt_for_each(i, lnet_cpt_table()) {
2705 INIT_LIST_HEAD(&local_queue);
2708 if (!the_lnet.ln_mt_rstq[i]) {
2712 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2715 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2717 * The rspt mdh will be invalidated when a response
2718 * is received or whenever we want to discard the
2719 * block the monitor thread will walk the queue
2720 * and clean up any rsts with an invalid mdh.
2721 * The monitor thread will walk the queue until
2722 * the first unexpired rspt block. This means that
2723 * some rspt blocks which received their
2724 * corresponding responses will linger in the
2725 * queue until they are cleaned up eventually.
2728 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2730 list_del_init(&rspt->rspt_on_list);
2731 lnet_rspt_free(rspt, i);
2735 if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
2737 md = lnet_handle2md(&rspt->rspt_mdh);
2739 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2741 list_del_init(&rspt->rspt_on_list);
2742 lnet_rspt_free(rspt, i);
2745 LASSERT(md->md_rspt_ptr == rspt);
2746 md->md_rspt_ptr = NULL;
2750 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2753 list_del_init(&rspt->rspt_on_list);
2755 CNETERR("Response timed out: md = %p\n", md);
2756 LNetMDUnlink(rspt->rspt_mdh);
2757 lnet_rspt_free(rspt, i);
2765 if (!list_empty(&local_queue))
2766 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2772 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2774 struct lnet_msg *msg;
2776 while (!list_empty(resendq)) {
2777 struct lnet_peer_ni *lpni;
2779 msg = list_entry(resendq->next, struct lnet_msg,
2782 list_del_init(&msg->msg_list);
2784 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2786 lnet_net_unlock(cpt);
2787 CERROR("Expected that a peer is already created for %s\n",
2788 libcfs_nid2str(msg->msg_hdr.dest_nid));
2789 msg->msg_no_resend = true;
2790 lnet_finalize(msg, -EFAULT);
2793 struct lnet_peer *peer;
2795 lnet_nid_t src_nid = LNET_NID_ANY;
2798 * if this message is not being routed and the
2799 * peer is non-MR then we must use the same
2800 * src_nid that was used in the original send.
2801 * Otherwise if we're routing the message (IE
2802 * we're a router) then we can use any of our
2803 * local interfaces. It doesn't matter to the
2804 * final destination.
2806 peer = lpni->lpni_peer_net->lpn_peer;
2807 if (!msg->msg_routing &&
2808 !lnet_peer_is_multi_rail(peer))
2809 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
2812 * If we originally specified a src NID, then we
2813 * must attempt to reuse it in the resend as well.
2815 if (msg->msg_src_nid_param != LNET_NID_ANY)
2816 src_nid = msg->msg_src_nid_param;
2817 lnet_peer_ni_decref_locked(lpni);
2819 lnet_net_unlock(cpt);
2820 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
2821 libcfs_nid2str(src_nid),
2822 libcfs_id2str(msg->msg_target),
2823 lnet_msgtyp2str(msg->msg_type),
2825 msg->msg_retry_count);
2826 rc = lnet_send(src_nid, msg, LNET_NID_ANY);
2828 CERROR("Error sending %s to %s: %d\n",
2829 lnet_msgtyp2str(msg->msg_type),
2830 libcfs_id2str(msg->msg_target), rc);
2831 msg->msg_no_resend = true;
2832 lnet_finalize(msg, rc);
2836 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
2842 lnet_resend_pending_msgs(void)
2846 cfs_cpt_for_each(i, lnet_cpt_table()) {
2848 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
2853 /* called with cpt and ni_lock held */
2855 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
2857 struct lnet_handle_md recovery_mdh;
2859 LNetInvalidateMDHandle(&recovery_mdh);
2861 if (ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING || force) {
2862 recovery_mdh = ni->ni_ping_mdh;
2863 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2866 lnet_net_unlock(cpt);
2867 if (!LNetMDHandleIsInvalid(recovery_mdh))
2868 LNetMDUnlink(recovery_mdh);
2874 lnet_recover_local_nis(void)
2876 struct lnet_mt_event_info *ev_info;
2877 struct list_head processed_list;
2878 struct list_head local_queue;
2879 struct lnet_handle_md mdh;
2880 struct lnet_ni *tmp;
2886 INIT_LIST_HEAD(&local_queue);
2887 INIT_LIST_HEAD(&processed_list);
2890 * splice the recovery queue on a local queue. We will iterate
2891 * through the local queue and update it as needed. Once we're
2892 * done with the traversal, we'll splice the local queue back on
2893 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
2894 * will be traversed in the next iteration.
2897 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
2901 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
2903 * if an NI is being deleted or it is now healthy, there
2904 * is no need to keep it around in the recovery queue.
2905 * The monitor thread is the only thread responsible for
2906 * removing the NI from the recovery queue.
2907 * Multiple threads can be adding NIs to the recovery
2910 healthv = atomic_read(&ni->ni_healthv);
2914 if (!(ni->ni_state & LNET_NI_STATE_ACTIVE) ||
2915 healthv == LNET_MAX_HEALTH_VALUE) {
2916 list_del_init(&ni->ni_recovery);
2917 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
2919 lnet_ni_decref_locked(ni, 0);
2925 * if the local NI failed recovery we must unlink the md.
2926 * But we want to keep the local_ni on the recovery queue
2927 * so we can continue the attempts to recover it.
2929 if (ni->ni_state & LNET_NI_STATE_RECOVERY_FAILED) {
2930 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
2931 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_FAILED;
2938 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
2939 libcfs_nid2str(ni->ni_nid));
2942 if (!(ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING)) {
2943 ni->ni_state |= LNET_NI_STATE_RECOVERY_PENDING;
2946 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
2948 CERROR("out of memory. Can't recover %s\n",
2949 libcfs_nid2str(ni->ni_nid));
2951 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
2956 mdh = ni->ni_ping_mdh;
2958 * Invalidate the ni mdh in case it's deleted.
2959 * We'll unlink the mdh in this case below.
2961 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2965 * remove the NI from the local queue and drop the
2966 * reference count to it while we're recovering
2967 * it. The reason for that, is that the NI could
2968 * be deleted, and the way the code is structured
2969 * is if we don't drop the NI, then the deletion
2970 * code will enter a loop waiting for the
2971 * reference count to be removed while holding the
2972 * ln_mutex_lock(). When we look up the peer to
2973 * send to in lnet_select_pathway() we will try to
2974 * lock the ln_mutex_lock() as well, leading to
2975 * a deadlock. By dropping the refcount and
2976 * removing it from the list, we allow for the NI
2977 * to be removed, then we use the cached NID to
2978 * look it up again. If it's gone, then we just
2979 * continue examining the rest of the queue.
2982 list_del_init(&ni->ni_recovery);
2983 lnet_ni_decref_locked(ni, 0);
2986 ev_info->mt_type = MT_TYPE_LOCAL_NI;
2987 ev_info->mt_nid = nid;
2988 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
2989 ev_info, the_lnet.ln_mt_eqh, true);
2990 /* lookup the nid again */
2992 ni = lnet_nid2ni_locked(nid, 0);
2995 * the NI has been deleted when we dropped
3003 * Same note as in lnet_recover_peer_nis(). When
3004 * we're sending the ping, the NI is free to be
3005 * deleted or manipulated. By this point it
3006 * could've been added back on the recovery queue,
3007 * and a refcount taken on it.
3008 * So we can't just add it blindly again or we'll
3009 * corrupt the queue. We must check under lock if
3010 * it's not on any list and if not then add it
3011 * to the processed list, which will eventually be
3012 * spliced back on to the recovery queue.
3014 ni->ni_ping_mdh = mdh;
3015 if (list_empty(&ni->ni_recovery)) {
3016 list_add_tail(&ni->ni_recovery, &processed_list);
3017 lnet_ni_addref_locked(ni, 0);
3023 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
3029 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3030 * reexamined in the next iteration.
3032 list_splice_init(&processed_list, &local_queue);
3034 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3038 static struct list_head **
3039 lnet_create_array_of_queues(void)
3041 struct list_head **qs;
3042 struct list_head *q;
3045 qs = cfs_percpt_alloc(lnet_cpt_table(),
3046 sizeof(struct list_head));
3048 CERROR("Failed to allocate queues\n");
3052 cfs_percpt_for_each(q, i, qs)
3059 lnet_resendqs_create(void)
3061 struct list_head **resendqs;
3062 resendqs = lnet_create_array_of_queues();
3067 lnet_net_lock(LNET_LOCK_EX);
3068 the_lnet.ln_mt_resendqs = resendqs;
3069 lnet_net_unlock(LNET_LOCK_EX);
3075 lnet_clean_local_ni_recoveryq(void)
3079 /* This is only called when the monitor thread has stopped */
3082 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3083 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3084 struct lnet_ni, ni_recovery);
3085 list_del_init(&ni->ni_recovery);
3087 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3089 lnet_ni_decref_locked(ni, 0);
3096 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3099 struct lnet_handle_md recovery_mdh;
3101 LNetInvalidateMDHandle(&recovery_mdh);
3103 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3104 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3105 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3107 spin_unlock(&lpni->lpni_lock);
3108 lnet_net_unlock(cpt);
3109 if (!LNetMDHandleIsInvalid(recovery_mdh))
3110 LNetMDUnlink(recovery_mdh);
3112 spin_lock(&lpni->lpni_lock);
3116 lnet_clean_peer_ni_recoveryq(void)
3118 struct lnet_peer_ni *lpni, *tmp;
3120 lnet_net_lock(LNET_LOCK_EX);
3122 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3124 list_del_init(&lpni->lpni_recovery);
3125 spin_lock(&lpni->lpni_lock);
3126 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3127 spin_unlock(&lpni->lpni_lock);
3128 lnet_peer_ni_decref_locked(lpni);
3131 lnet_net_unlock(LNET_LOCK_EX);
3135 lnet_clean_resendqs(void)
3137 struct lnet_msg *msg, *tmp;
3138 struct list_head msgs;
3141 INIT_LIST_HEAD(&msgs);
3143 cfs_cpt_for_each(i, lnet_cpt_table()) {
3145 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3147 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3148 list_del_init(&msg->msg_list);
3149 msg->msg_no_resend = true;
3150 lnet_finalize(msg, -ESHUTDOWN);
3154 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3158 lnet_recover_peer_nis(void)
3160 struct lnet_mt_event_info *ev_info;
3161 struct list_head processed_list;
3162 struct list_head local_queue;
3163 struct lnet_handle_md mdh;
3164 struct lnet_peer_ni *lpni;
3165 struct lnet_peer_ni *tmp;
3170 INIT_LIST_HEAD(&local_queue);
3171 INIT_LIST_HEAD(&processed_list);
3174 * Always use cpt 0 for locking across all interactions with
3175 * ln_mt_peerNIRecovq
3178 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3182 list_for_each_entry_safe(lpni, tmp, &local_queue,
3185 * The same protection strategy is used here as is in the
3186 * local recovery case.
3189 healthv = atomic_read(&lpni->lpni_healthv);
3190 spin_lock(&lpni->lpni_lock);
3191 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3192 healthv == LNET_MAX_HEALTH_VALUE) {
3193 list_del_init(&lpni->lpni_recovery);
3194 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3195 spin_unlock(&lpni->lpni_lock);
3196 lnet_peer_ni_decref_locked(lpni);
3202 * If the peer NI has failed recovery we must unlink the
3203 * md. But we want to keep the peer ni on the recovery
3204 * queue so we can try to continue recovering it
3206 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3207 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3208 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3211 spin_unlock(&lpni->lpni_lock);
3215 * NOTE: we're racing with peer deletion from user space.
3216 * It's possible that a peer is deleted after we check its
3217 * state. In this case the recovery can create a new peer
3219 spin_lock(&lpni->lpni_lock);
3220 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3221 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3222 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3223 spin_unlock(&lpni->lpni_lock);
3225 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3227 CERROR("out of memory. Can't recover %s\n",
3228 libcfs_nid2str(lpni->lpni_nid));
3229 spin_lock(&lpni->lpni_lock);
3230 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3231 spin_unlock(&lpni->lpni_lock);
3235 /* look at the comments in lnet_recover_local_nis() */
3236 mdh = lpni->lpni_recovery_ping_mdh;
3237 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3238 nid = lpni->lpni_nid;
3240 list_del_init(&lpni->lpni_recovery);
3241 lnet_peer_ni_decref_locked(lpni);
3244 ev_info->mt_type = MT_TYPE_PEER_NI;
3245 ev_info->mt_nid = nid;
3246 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3247 ev_info, the_lnet.ln_mt_eqh, true);
3250 * lnet_find_peer_ni_locked() grabs a refcount for
3251 * us. No need to take it explicitly.
3253 lpni = lnet_find_peer_ni_locked(nid);
3260 lpni->lpni_recovery_ping_mdh = mdh;
3262 * While we're unlocked the lpni could've been
3263 * readded on the recovery queue. In this case we
3264 * don't need to add it to the local queue, since
3265 * it's already on there and the thread that added
3266 * it would've incremented the refcount on the
3267 * peer, which means we need to decref the refcount
3268 * that was implicitly grabbed by find_peer_ni_locked.
3269 * Otherwise, if the lpni is still not on
3270 * the recovery queue, then we'll add it to the
3273 if (list_empty(&lpni->lpni_recovery))
3274 list_add_tail(&lpni->lpni_recovery, &processed_list);
3276 lnet_peer_ni_decref_locked(lpni);
3279 spin_lock(&lpni->lpni_lock);
3281 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3283 spin_unlock(&lpni->lpni_lock);
3286 list_splice_init(&processed_list, &local_queue);
3288 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3293 lnet_monitor_thread(void *arg)
3295 int wakeup_counter = 0;
3298 * The monitor thread takes care of the following:
3299 * 1. Checks the aliveness of routers
3300 * 2. Checks if there are messages on the resend queue to resend
3302 * 3. Check if there are any NIs on the local recovery queue and
3304 * 4. Checks if there are any NIs on the remote recovery queue
3307 cfs_block_allsigs();
3309 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3310 if (lnet_router_checker_active())
3311 lnet_check_routers();
3313 lnet_resend_pending_msgs();
3316 if (wakeup_counter >= lnet_transaction_timeout / 2) {
3317 lnet_finalize_expired_responses(false);
3321 lnet_recover_local_nis();
3323 lnet_recover_peer_nis();
3326 * TODO do we need to check if we should sleep without
3327 * timeout? Technically, an active system will always
3328 * have messages in flight so this check will always
3329 * evaluate to false. And on an idle system do we care
3330 * if we wake up every 1 second? Although, we've seen
3331 * cases where we get a complaint that an idle thread
3332 * is waking up unnecessarily.
3334 wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
3336 cfs_time_seconds(1));
3339 /* clean up the router checker */
3340 lnet_prune_rc_data(1);
3343 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3345 /* signal that the monitor thread is exiting */
3346 up(&the_lnet.ln_mt_signal);
3354 * Returns == 0 if success
3355 * Returns > 0 if LNetMDBind or prior fails
3356 * Returns < 0 if LNetGet fails
3359 lnet_send_ping(lnet_nid_t dest_nid,
3360 struct lnet_handle_md *mdh, int nnis,
3361 void *user_data, struct lnet_handle_eq eqh, bool recovery)
3363 struct lnet_md md = { NULL };
3364 struct lnet_process_id id;
3365 struct lnet_ping_buffer *pbuf;
3368 if (dest_nid == LNET_NID_ANY) {
3373 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3379 /* initialize md content */
3380 md.start = &pbuf->pb_info;
3381 md.length = LNET_PING_INFO_SIZE(nnis);
3382 md.threshold = 2; /* GET/REPLY */
3384 md.options = LNET_MD_TRUNCATE;
3385 md.user_ptr = user_data;
3388 rc = LNetMDBind(md, LNET_UNLINK, mdh);
3390 lnet_ping_buffer_decref(pbuf);
3391 CERROR("Can't bind MD: %d\n", rc);
3392 rc = -rc; /* change the rc to positive */
3395 id.pid = LNET_PID_LUSTRE;
3398 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3399 LNET_RESERVED_PORTAL,
3400 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3403 goto fail_unlink_md;
3409 LNetInvalidateMDHandle(mdh);
3415 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3418 lnet_nid_t nid = ev_info->mt_nid;
3420 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3424 ni = lnet_nid2ni_locked(nid, 0);
3430 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
3432 ni->ni_state |= LNET_NI_STATE_RECOVERY_FAILED;
3437 CERROR("local NI (%s) recovery failed with %d\n",
3438 libcfs_nid2str(nid), status);
3442 * need to increment healthv for the ni here, because in
3443 * the lnet_finalize() path we don't have access to this
3444 * NI. And in order to get access to it, we'll need to
3445 * carry forward too much information.
3446 * In the peer case, it'll naturally be incremented
3448 lnet_inc_healthv(&ni->ni_healthv);
3450 struct lnet_peer_ni *lpni;
3453 cpt = lnet_net_lock_current();
3454 lpni = lnet_find_peer_ni_locked(nid);
3456 lnet_net_unlock(cpt);
3459 spin_lock(&lpni->lpni_lock);
3460 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3462 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3463 spin_unlock(&lpni->lpni_lock);
3464 lnet_peer_ni_decref_locked(lpni);
3465 lnet_net_unlock(cpt);
3468 CERROR("peer NI (%s) recovery failed with %d\n",
3469 libcfs_nid2str(nid), status);
3474 lnet_mt_event_handler(struct lnet_event *event)
3476 struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3477 struct lnet_ping_buffer *pbuf;
3479 /* TODO: remove assert */
3480 LASSERT(event->type == LNET_EVENT_REPLY ||
3481 event->type == LNET_EVENT_SEND ||
3482 event->type == LNET_EVENT_UNLINK);
3484 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3487 switch (event->type) {
3488 case LNET_EVENT_UNLINK:
3489 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3490 libcfs_nid2str(ev_info->mt_nid));
3491 case LNET_EVENT_REPLY:
3492 lnet_handle_recovery_reply(ev_info, event->status);
3494 case LNET_EVENT_SEND:
3495 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3496 libcfs_nid2str(ev_info->mt_nid),
3497 (event->status) ? "unsuccessfully" :
3498 "successfully", event->status);
3499 lnet_handle_recovery_reply(ev_info, event->status);
3502 CERROR("Unexpected event: %d\n", event->type);
3505 if (event->unlinked) {
3506 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3507 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3508 lnet_ping_buffer_decref(pbuf);
3513 lnet_rsp_tracker_create(void)
3515 struct list_head **rstqs;
3516 rstqs = lnet_create_array_of_queues();
3521 the_lnet.ln_mt_rstq = rstqs;
3527 lnet_rsp_tracker_clean(void)
3529 lnet_finalize_expired_responses(true);
3531 cfs_percpt_free(the_lnet.ln_mt_rstq);
3532 the_lnet.ln_mt_rstq = NULL;
3535 int lnet_monitor_thr_start(void)
3538 struct task_struct *task;
3540 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3543 rc = lnet_resendqs_create();
3547 rc = lnet_rsp_tracker_create();
3551 rc = LNetEQAlloc(0, lnet_mt_event_handler, &the_lnet.ln_mt_eqh);
3553 CERROR("Can't allocate monitor thread EQ: %d\n", rc);
3557 /* Pre monitor thread start processing */
3558 rc = lnet_router_pre_mt_start();
3562 sema_init(&the_lnet.ln_mt_signal, 0);
3564 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3565 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3568 CERROR("Can't start monitor thread: %d\n", rc);
3572 /* post monitor thread start processing */
3573 lnet_router_post_mt_start();
3578 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3579 /* block until event callback signals exit */
3580 down(&the_lnet.ln_mt_signal);
3582 lnet_router_cleanup();
3584 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3585 lnet_rsp_tracker_clean();
3586 lnet_clean_local_ni_recoveryq();
3587 lnet_clean_peer_ni_recoveryq();
3588 lnet_clean_resendqs();
3589 LNetEQFree(the_lnet.ln_mt_eqh);
3590 LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3593 lnet_rsp_tracker_clean();
3594 lnet_clean_local_ni_recoveryq();
3595 lnet_clean_peer_ni_recoveryq();
3596 lnet_clean_resendqs();
3600 void lnet_monitor_thr_stop(void)
3604 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3607 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3608 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3610 /* tell the monitor thread that we're shutting down */
3611 wake_up(&the_lnet.ln_mt_waitq);
3613 /* block until monitor thread signals that it's done */
3614 down(&the_lnet.ln_mt_signal);
3615 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3617 /* perform cleanup tasks */
3618 lnet_router_cleanup();
3619 lnet_rsp_tracker_clean();
3620 lnet_clean_local_ni_recoveryq();
3621 lnet_clean_peer_ni_recoveryq();
3622 lnet_clean_resendqs();
3623 rc = LNetEQFree(the_lnet.ln_mt_eqh);
3629 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3633 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3634 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3635 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3636 lnet_net_unlock(cpt);
3638 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3642 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3644 struct lnet_hdr *hdr = &msg->msg_hdr;
3646 if (msg->msg_wanted != 0)
3647 lnet_setpayloadbuffer(msg);
3649 lnet_build_msg_event(msg, LNET_EVENT_PUT);
3651 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
3652 * it back into the ACK during lnet_finalize() */
3653 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3654 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3656 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3657 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3661 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3663 struct lnet_hdr *hdr = &msg->msg_hdr;
3664 struct lnet_match_info info;
3668 /* Convert put fields to host byte order */
3669 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3670 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
3671 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
3673 /* Primary peer NID. */
3674 info.mi_id.nid = msg->msg_initiator;
3675 info.mi_id.pid = hdr->src_pid;
3676 info.mi_opc = LNET_MD_OP_PUT;
3677 info.mi_portal = hdr->msg.put.ptl_index;
3678 info.mi_rlength = hdr->payload_length;
3679 info.mi_roffset = hdr->msg.put.offset;
3680 info.mi_mbits = hdr->msg.put.match_bits;
3681 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
3683 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3684 ready_delay = msg->msg_rx_ready_delay;
3687 rc = lnet_ptl_match_md(&info, msg);
3692 case LNET_MATCHMD_OK:
3693 lnet_recv_put(ni, msg);
3696 case LNET_MATCHMD_NONE:
3698 /* no eager_recv or has already called it, should
3699 * have been attached on delayed list */
3702 rc = lnet_ni_eager_recv(ni, msg);
3709 case LNET_MATCHMD_DROP:
3710 CNETERR("Dropping PUT from %s portal %d match %llu"
3711 " offset %d length %d: %d\n",
3712 libcfs_id2str(info.mi_id), info.mi_portal,
3713 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3715 return -ENOENT; /* -ve: OK but no match */
3720 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3722 struct lnet_match_info info;
3723 struct lnet_hdr *hdr = &msg->msg_hdr;
3724 struct lnet_process_id source_id;
3725 struct lnet_handle_wire reply_wmd;
3728 /* Convert get fields to host byte order */
3729 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
3730 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
3731 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
3732 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
3734 source_id.nid = hdr->src_nid;
3735 source_id.pid = hdr->src_pid;
3736 /* Primary peer NID */
3737 info.mi_id.nid = msg->msg_initiator;
3738 info.mi_id.pid = hdr->src_pid;
3739 info.mi_opc = LNET_MD_OP_GET;
3740 info.mi_portal = hdr->msg.get.ptl_index;
3741 info.mi_rlength = hdr->msg.get.sink_length;
3742 info.mi_roffset = hdr->msg.get.src_offset;
3743 info.mi_mbits = hdr->msg.get.match_bits;
3744 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
3746 rc = lnet_ptl_match_md(&info, msg);
3747 if (rc == LNET_MATCHMD_DROP) {
3748 CNETERR("Dropping GET from %s portal %d match %llu"
3749 " offset %d length %d\n",
3750 libcfs_id2str(info.mi_id), info.mi_portal,
3751 info.mi_mbits, info.mi_roffset, info.mi_rlength);
3752 return -ENOENT; /* -ve: OK but no match */
3755 LASSERT(rc == LNET_MATCHMD_OK);
3757 lnet_build_msg_event(msg, LNET_EVENT_GET);
3759 reply_wmd = hdr->msg.get.return_wmd;
3761 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3762 msg->msg_offset, msg->msg_wanted);
3764 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3767 /* The LND completes the REPLY from her recv procedure */
3768 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3769 msg->msg_offset, msg->msg_len, msg->msg_len);
3773 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3774 msg->msg_receiving = 0;
3776 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
3778 /* didn't get as far as lnet_ni_send() */
3779 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3780 libcfs_nid2str(ni->ni_nid),
3781 libcfs_id2str(info.mi_id), rc);
3783 lnet_finalize(msg, rc);
3790 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3792 void *private = msg->msg_private;
3793 struct lnet_hdr *hdr = &msg->msg_hdr;
3794 struct lnet_process_id src = {0};
3795 struct lnet_libmd *md;
3800 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3803 src.nid = hdr->src_nid;
3804 src.pid = hdr->src_pid;
3806 /* NB handles only looked up by creator (no flips) */
3807 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
3808 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3809 CNETERR("%s: Dropping REPLY from %s for %s "
3811 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3812 (md == NULL) ? "invalid" : "inactive",
3813 hdr->msg.reply.dst_wmd.wh_interface_cookie,
3814 hdr->msg.reply.dst_wmd.wh_object_cookie);
3815 if (md != NULL && md->md_me != NULL)
3816 CERROR("REPLY MD also attached to portal %d\n",
3817 md->md_me->me_portal);
3819 lnet_res_unlock(cpt);
3820 return -ENOENT; /* -ve: OK but no match */
3823 LASSERT(md->md_offset == 0);
3825 rlength = hdr->payload_length;
3826 mlength = MIN(rlength, (int)md->md_length);
3828 if (mlength < rlength &&
3829 (md->md_options & LNET_MD_TRUNCATE) == 0) {
3830 CNETERR("%s: Dropping REPLY from %s length %d "
3831 "for MD %#llx would overflow (%d)\n",
3832 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3833 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
3835 lnet_res_unlock(cpt);
3836 return -ENOENT; /* -ve: OK but no match */
3839 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
3840 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3841 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
3843 lnet_msg_attach_md(msg, md, 0, mlength);
3846 lnet_setpayloadbuffer(msg);
3848 lnet_res_unlock(cpt);
3850 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3852 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
3857 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
3859 struct lnet_hdr *hdr = &msg->msg_hdr;
3860 struct lnet_process_id src = {0};
3861 struct lnet_libmd *md;
3864 src.nid = hdr->src_nid;
3865 src.pid = hdr->src_pid;
3867 /* Convert ack fields to host byte order */
3868 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
3869 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
3871 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
3874 /* NB handles only looked up by creator (no flips) */
3875 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
3876 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3877 /* Don't moan; this is expected */
3879 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
3880 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3881 (md == NULL) ? "invalid" : "inactive",
3882 hdr->msg.ack.dst_wmd.wh_interface_cookie,
3883 hdr->msg.ack.dst_wmd.wh_object_cookie);
3884 if (md != NULL && md->md_me != NULL)
3885 CERROR("Source MD also attached to portal %d\n",
3886 md->md_me->me_portal);
3888 lnet_res_unlock(cpt);
3889 return -ENOENT; /* -ve! */
3892 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
3893 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3894 hdr->msg.ack.dst_wmd.wh_object_cookie);
3896 lnet_msg_attach_md(msg, md, 0, 0);
3898 lnet_res_unlock(cpt);
3900 lnet_build_msg_event(msg, LNET_EVENT_ACK);
3902 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
3907 * \retval LNET_CREDIT_OK If \a msg is forwarded
3908 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
3909 * \retval -ve error code
3912 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
3916 if (!the_lnet.ln_routing)
3919 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
3920 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
3921 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
3922 msg->msg_rx_ready_delay = 1;
3924 lnet_net_unlock(msg->msg_rx_cpt);
3925 rc = lnet_ni_eager_recv(ni, msg);
3926 lnet_net_lock(msg->msg_rx_cpt);
3931 rc = lnet_post_routed_recv_locked(msg, 0);
3936 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
3940 switch (msg->msg_type) {
3942 rc = lnet_parse_ack(ni, msg);
3945 rc = lnet_parse_put(ni, msg);
3948 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
3950 case LNET_MSG_REPLY:
3951 rc = lnet_parse_reply(ni, msg);
3953 default: /* prevent an unused label if !kernel */
3958 LASSERT(rc == 0 || rc == -ENOENT);
3963 lnet_msgtyp2str (int type)
3972 case LNET_MSG_REPLY:
3974 case LNET_MSG_HELLO:
3977 return ("<UNKNOWN>");
3982 lnet_print_hdr(struct lnet_hdr *hdr)
3984 struct lnet_process_id src = {
3985 .nid = hdr->src_nid,
3986 .pid = hdr->src_pid,
3988 struct lnet_process_id dst = {
3989 .nid = hdr->dest_nid,
3990 .pid = hdr->dest_pid,
3992 char *type_str = lnet_msgtyp2str(hdr->type);
3994 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
3995 CWARN(" From %s\n", libcfs_id2str(src));
3996 CWARN(" To %s\n", libcfs_id2str(dst));
3998 switch (hdr->type) {
4003 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
4004 "match bits %llu\n",
4005 hdr->msg.put.ptl_index,
4006 hdr->msg.put.ack_wmd.wh_interface_cookie,
4007 hdr->msg.put.ack_wmd.wh_object_cookie,
4008 hdr->msg.put.match_bits);
4009 CWARN(" Length %d, offset %d, hdr data %#llx\n",
4010 hdr->payload_length, hdr->msg.put.offset,
4011 hdr->msg.put.hdr_data);
4015 CWARN(" Ptl index %d, return md %#llx.%#llx, "
4016 "match bits %llu\n", hdr->msg.get.ptl_index,
4017 hdr->msg.get.return_wmd.wh_interface_cookie,
4018 hdr->msg.get.return_wmd.wh_object_cookie,
4019 hdr->msg.get.match_bits);
4020 CWARN(" Length %d, src offset %d\n",
4021 hdr->msg.get.sink_length,
4022 hdr->msg.get.src_offset);
4026 CWARN(" dst md %#llx.%#llx, "
4027 "manipulated length %d\n",
4028 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4029 hdr->msg.ack.dst_wmd.wh_object_cookie,
4030 hdr->msg.ack.mlength);
4033 case LNET_MSG_REPLY:
4034 CWARN(" dst md %#llx.%#llx, "
4036 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4037 hdr->msg.reply.dst_wmd.wh_object_cookie,
4038 hdr->payload_length);
4044 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4045 void *private, int rdma_req)
4050 struct lnet_msg *msg;
4051 lnet_pid_t dest_pid;
4052 lnet_nid_t dest_nid;
4054 struct lnet_peer_ni *lpni;
4055 __u32 payload_length;
4058 LASSERT (!in_interrupt ());
4060 type = le32_to_cpu(hdr->type);
4061 src_nid = le64_to_cpu(hdr->src_nid);
4062 dest_nid = le64_to_cpu(hdr->dest_nid);
4063 dest_pid = le32_to_cpu(hdr->dest_pid);
4064 payload_length = le32_to_cpu(hdr->payload_length);
4066 for_me = (ni->ni_nid == dest_nid);
4067 cpt = lnet_cpt_of_nid(from_nid, ni);
4069 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4070 libcfs_nid2str(dest_nid),
4071 libcfs_nid2str(ni->ni_nid),
4072 libcfs_nid2str(src_nid),
4073 lnet_msgtyp2str(type),
4074 (for_me) ? "for me" : "routed");
4079 if (payload_length > 0) {
4080 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4081 libcfs_nid2str(from_nid),
4082 libcfs_nid2str(src_nid),
4083 lnet_msgtyp2str(type), payload_length);
4089 case LNET_MSG_REPLY:
4090 if (payload_length >
4091 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4092 CERROR("%s, src %s: bad %s payload %d "
4093 "(%d max expected)\n",
4094 libcfs_nid2str(from_nid),
4095 libcfs_nid2str(src_nid),
4096 lnet_msgtyp2str(type),
4098 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4104 CERROR("%s, src %s: Bad message type 0x%x\n",
4105 libcfs_nid2str(from_nid),
4106 libcfs_nid2str(src_nid), type);
4110 if (the_lnet.ln_routing &&
4111 ni->ni_last_alive != ktime_get_real_seconds()) {
4112 /* NB: so far here is the only place to set NI status to "up */
4114 ni->ni_last_alive = ktime_get_real_seconds();
4115 if (ni->ni_status != NULL &&
4116 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
4117 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4121 /* Regard a bad destination NID as a protocol error. Senders should
4122 * know what they're doing; if they don't they're misconfigured, buggy
4123 * or malicious so we chop them off at the knees :) */
4126 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4127 /* should have gone direct */
4128 CERROR("%s, src %s: Bad dest nid %s "
4129 "(should have been sent direct)\n",
4130 libcfs_nid2str(from_nid),
4131 libcfs_nid2str(src_nid),
4132 libcfs_nid2str(dest_nid));
4136 if (lnet_islocalnid(dest_nid)) {
4137 /* dest is another local NI; sender should have used
4138 * this node's NID on its own network */
4139 CERROR("%s, src %s: Bad dest nid %s "
4140 "(it's my nid but on a different network)\n",
4141 libcfs_nid2str(from_nid),
4142 libcfs_nid2str(src_nid),
4143 libcfs_nid2str(dest_nid));
4147 if (rdma_req && type == LNET_MSG_GET) {
4148 CERROR("%s, src %s: Bad optimized GET for %s "
4149 "(final destination must be me)\n",
4150 libcfs_nid2str(from_nid),
4151 libcfs_nid2str(src_nid),
4152 libcfs_nid2str(dest_nid));
4156 if (!the_lnet.ln_routing) {
4157 CERROR("%s, src %s: Dropping message for %s "
4158 "(routing not enabled)\n",
4159 libcfs_nid2str(from_nid),
4160 libcfs_nid2str(src_nid),
4161 libcfs_nid2str(dest_nid));
4166 /* Message looks OK; we're not going to return an error, so we MUST
4167 * call back lnd_recv() come what may... */
4169 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4170 fail_peer(src_nid, 0)) { /* shall we now? */
4171 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4172 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4173 lnet_msgtyp2str(type));
4177 if (!list_empty(&the_lnet.ln_drop_rules) &&
4178 lnet_drop_rule_match(hdr, NULL)) {
4179 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4180 "silent message loss\n",
4181 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4182 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4187 msg = lnet_msg_alloc();
4189 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4190 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4191 lnet_msgtyp2str(type));
4195 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4196 * pointers NULL etc */
4198 msg->msg_type = type;
4199 msg->msg_private = private;
4200 msg->msg_receiving = 1;
4201 msg->msg_rdma_get = rdma_req;
4202 msg->msg_len = msg->msg_wanted = payload_length;
4203 msg->msg_offset = 0;
4204 msg->msg_hdr = *hdr;
4205 /* for building message event */
4206 msg->msg_from = from_nid;
4208 msg->msg_target.pid = dest_pid;
4209 msg->msg_target.nid = dest_nid;
4210 msg->msg_routing = 1;
4213 /* convert common msg->hdr fields to host byteorder */
4214 msg->msg_hdr.type = type;
4215 msg->msg_hdr.src_nid = src_nid;
4216 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4217 msg->msg_hdr.dest_nid = dest_nid;
4218 msg->msg_hdr.dest_pid = dest_pid;
4219 msg->msg_hdr.payload_length = payload_length;
4223 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4225 lnet_net_unlock(cpt);
4226 CERROR("%s, src %s: Dropping %s "
4227 "(error %ld looking up sender)\n",
4228 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4229 lnet_msgtyp2str(type), PTR_ERR(lpni));
4231 if (rc == -ESHUTDOWN)
4232 /* We are shutting down. Don't do anything more */
4236 msg->msg_rxpeer = lpni;
4238 lnet_ni_addref_locked(ni, cpt);
4239 /* Multi-Rail: Primary NID of source. */
4240 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4242 if (lnet_isrouter(msg->msg_rxpeer)) {
4243 lnet_peer_set_alive(msg->msg_rxpeer);
4244 if (avoid_asym_router_failure &&
4245 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4246 /* received a remote message from router, update
4247 * remote NI status on this router.
4248 * NB: multi-hop routed message will be ignored.
4250 lnet_router_ni_update_locked(msg->msg_rxpeer,
4251 LNET_NIDNET(src_nid));
4255 lnet_msg_commit(msg, cpt);
4257 /* message delay simulation */
4258 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4259 lnet_delay_rule_match_locked(hdr, msg))) {
4260 lnet_net_unlock(cpt);
4265 rc = lnet_parse_forward_locked(ni, msg);
4266 lnet_net_unlock(cpt);
4271 if (rc == LNET_CREDIT_OK) {
4272 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4273 0, payload_length, payload_length);
4278 lnet_net_unlock(cpt);
4280 rc = lnet_parse_local(ni, msg);
4286 LASSERT(msg->msg_md == NULL);
4287 lnet_finalize(msg, rc);
4290 lnet_drop_message(ni, cpt, private, payload_length, type);
4293 EXPORT_SYMBOL(lnet_parse);
4296 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4298 while (!list_empty(head)) {
4299 struct lnet_process_id id = {0};
4300 struct lnet_msg *msg;
4302 msg = list_entry(head->next, struct lnet_msg, msg_list);
4303 list_del(&msg->msg_list);
4305 id.nid = msg->msg_hdr.src_nid;
4306 id.pid = msg->msg_hdr.src_pid;
4308 LASSERT(msg->msg_md == NULL);
4309 LASSERT(msg->msg_rx_delayed);
4310 LASSERT(msg->msg_rxpeer != NULL);
4311 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4313 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4314 " offset %d length %d: %s\n",
4316 msg->msg_hdr.msg.put.ptl_index,
4317 msg->msg_hdr.msg.put.match_bits,
4318 msg->msg_hdr.msg.put.offset,
4319 msg->msg_hdr.payload_length, reason);
4321 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4322 * called lnet_drop_message(), so I just hang onto msg as well
4323 * until that's done */
4325 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4326 msg->msg_private, msg->msg_len,
4329 msg->msg_no_resend = true;
4331 * NB: message will not generate event because w/o attached MD,
4332 * but we still should give error code so lnet_msg_decommit()
4333 * can skip counters operations and other checks.
4335 lnet_finalize(msg, -ENOENT);
4340 lnet_recv_delayed_msg_list(struct list_head *head)
4342 while (!list_empty(head)) {
4343 struct lnet_msg *msg;
4344 struct lnet_process_id id;
4346 msg = list_entry(head->next, struct lnet_msg, msg_list);
4347 list_del(&msg->msg_list);
4349 /* md won't disappear under me, since each msg
4350 * holds a ref on it */
4352 id.nid = msg->msg_hdr.src_nid;
4353 id.pid = msg->msg_hdr.src_pid;
4355 LASSERT(msg->msg_rx_delayed);
4356 LASSERT(msg->msg_md != NULL);
4357 LASSERT(msg->msg_rxpeer != NULL);
4358 LASSERT(msg->msg_rxni != NULL);
4359 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4361 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4362 "match %llu offset %d length %d.\n",
4363 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4364 msg->msg_hdr.msg.put.match_bits,
4365 msg->msg_hdr.msg.put.offset,
4366 msg->msg_hdr.payload_length);
4368 lnet_recv_put(msg->msg_rxni, msg);
4373 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4374 struct lnet_libmd *md, struct lnet_handle_md mdh)
4379 * MD has a refcount taken by message so it's not going away.
4380 * The MD however can be looked up. We need to secure the access
4381 * to the md_rspt_ptr by taking the res_lock.
4382 * The rspt can be accessed without protection up to when it gets
4383 * added to the list.
4387 LASSERT(md->md_rspt_ptr == NULL);
4389 /* we'll use that same event in case we never get a response */
4390 rspt->rspt_mdh = mdh;
4391 rspt->rspt_cpt = cpt;
4392 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4393 rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4396 /* store the rspt so we can access it when we get the REPLY */
4397 md->md_rspt_ptr = rspt;
4398 lnet_res_unlock(cpt);
4401 * add to the list of tracked responses. It's added to tail of the
4402 * list in order to expire all the older entries first.
4405 list_add_tail(&rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4406 lnet_net_unlock(cpt);
4410 * Initiate an asynchronous PUT operation.
4412 * There are several events associated with a PUT: completion of the send on
4413 * the initiator node (LNET_EVENT_SEND), and when the send completes
4414 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4415 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4416 * used at the target node to indicate the completion of incoming data
4419 * The local events will be logged in the EQ associated with the MD pointed to
4420 * by \a mdh handle. Using a MD without an associated EQ results in these
4421 * events being discarded. In this case, the caller must have another
4422 * mechanism (e.g., a higher level protocol) for determining when it is safe
4423 * to modify the memory region associated with the MD.
4425 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4426 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4428 * \param self Indicates the NID of a local interface through which to send
4429 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4430 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4431 * must be "free floating" (See LNetMDBind()).
4432 * \param ack Controls whether an acknowledgment is requested.
4433 * Acknowledgments are only sent when they are requested by the initiating
4434 * process and the target MD enables them.
4435 * \param target A process identifier for the target process.
4436 * \param portal The index in the \a target's portal table.
4437 * \param match_bits The match bits to use for MD selection at the target
4439 * \param offset The offset into the target MD (only used when the target
4440 * MD has the LNET_MD_MANAGE_REMOTE option set).
4441 * \param hdr_data 64 bits of user data that can be included in the message
4442 * header. This data is written to an event queue entry at the target if an
4443 * EQ is present on the matching MD.
4445 * \retval 0 Success, and only in this case events will be generated
4446 * and logged to EQ (if it exists).
4447 * \retval -EIO Simulated failure.
4448 * \retval -ENOMEM Memory allocation failure.
4449 * \retval -ENOENT Invalid MD object.
4451 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4454 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4455 struct lnet_process_id target, unsigned int portal,
4456 __u64 match_bits, unsigned int offset,
4459 struct lnet_msg *msg;
4460 struct lnet_libmd *md;
4463 struct lnet_rsp_tracker *rspt = NULL;
4465 LASSERT(the_lnet.ln_refcount > 0);
4467 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4468 fail_peer(target.nid, 1)) { /* shall we now? */
4469 CERROR("Dropping PUT to %s: simulated failure\n",
4470 libcfs_id2str(target));
4474 msg = lnet_msg_alloc();
4476 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4477 libcfs_id2str(target));
4480 msg->msg_vmflush = !!memory_pressure_get();
4482 cpt = lnet_cpt_of_cookie(mdh.cookie);
4484 if (ack == LNET_ACK_REQ) {
4485 rspt = lnet_rspt_alloc(cpt);
4487 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4488 libcfs_id2str(target));
4491 INIT_LIST_HEAD(&rspt->rspt_on_list);
4496 md = lnet_handle2md(&mdh);
4497 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4498 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4499 match_bits, portal, libcfs_id2str(target),
4500 md == NULL ? -1 : md->md_threshold);
4501 if (md != NULL && md->md_me != NULL)
4502 CERROR("Source MD also attached to portal %d\n",
4503 md->md_me->me_portal);
4504 lnet_res_unlock(cpt);
4506 LIBCFS_FREE(rspt, sizeof(*rspt));
4511 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4513 lnet_msg_attach_md(msg, md, 0, 0);
4515 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4517 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4518 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4519 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4520 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4522 /* NB handles only looked up by creator (no flips) */
4523 if (ack == LNET_ACK_REQ) {
4524 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4525 the_lnet.ln_interface_cookie;
4526 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4527 md->md_lh.lh_cookie;
4529 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4530 LNET_WIRE_HANDLE_COOKIE_NONE;
4531 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4532 LNET_WIRE_HANDLE_COOKIE_NONE;
4535 lnet_res_unlock(cpt);
4537 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4539 if (ack == LNET_ACK_REQ)
4540 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4542 rc = lnet_send(self, msg, LNET_NID_ANY);
4544 CNETERR("Error sending PUT to %s: %d\n",
4545 libcfs_id2str(target), rc);
4546 msg->msg_no_resend = true;
4547 lnet_detach_rsp_tracker(msg->msg_md, cpt);
4548 lnet_finalize(msg, rc);
4551 /* completion will be signalled by an event */
4554 EXPORT_SYMBOL(LNetPut);
4557 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
4558 * returns a msg for the LND to pass to lnet_finalize() when the sink
4559 * data has been received.
4561 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4562 * lnet_finalize() is called on it, so the LND must call this first
4565 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4567 struct lnet_msg *msg = lnet_msg_alloc();
4568 struct lnet_libmd *getmd = getmsg->msg_md;
4569 struct lnet_process_id peer_id = getmsg->msg_target;
4572 LASSERT(!getmsg->msg_target_is_router);
4573 LASSERT(!getmsg->msg_routing);
4576 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4577 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4581 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4584 LASSERT(getmd->md_refcount > 0);
4586 if (getmd->md_threshold == 0) {
4587 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4588 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4590 lnet_res_unlock(cpt);
4594 LASSERT(getmd->md_offset == 0);
4596 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4597 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4599 /* setup information for lnet_build_msg_event */
4600 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4601 msg->msg_from = peer_id.nid;
4602 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4603 msg->msg_hdr.src_nid = peer_id.nid;
4604 msg->msg_hdr.payload_length = getmd->md_length;
4605 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4607 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4608 lnet_res_unlock(cpt);
4610 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4613 lnet_msg_commit(msg, cpt);
4614 lnet_net_unlock(cpt);
4616 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4621 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4624 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4625 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4626 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4628 lnet_net_unlock(cpt);
4635 EXPORT_SYMBOL(lnet_create_reply_msg);
4638 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4641 /* Set the REPLY length, now the RDMA that elides the REPLY message has
4642 * completed and I know it. */
4643 LASSERT(reply != NULL);
4644 LASSERT(reply->msg_type == LNET_MSG_GET);
4645 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4647 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
4648 * the end of my buffer, I might as well be dead. */
4649 LASSERT(len <= reply->msg_ev.mlength);
4651 reply->msg_ev.mlength = len;
4653 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4656 * Initiate an asynchronous GET operation.
4658 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4659 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4660 * the target node in the REPLY has been written to local MD.
4662 * On the target node, an LNET_EVENT_GET is logged when the GET request
4663 * arrives and is accepted into a MD.
4665 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4666 * \param mdh A handle for the MD that describes the memory into which the
4667 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4669 * \retval 0 Success, and only in this case events will be generated
4670 * and logged to EQ (if it exists) of the MD.
4671 * \retval -EIO Simulated failure.
4672 * \retval -ENOMEM Memory allocation failure.
4673 * \retval -ENOENT Invalid MD object.
4676 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4677 struct lnet_process_id target, unsigned int portal,
4678 __u64 match_bits, unsigned int offset, bool recovery)
4680 struct lnet_msg *msg;
4681 struct lnet_libmd *md;
4682 struct lnet_rsp_tracker *rspt;
4686 LASSERT(the_lnet.ln_refcount > 0);
4688 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4689 fail_peer(target.nid, 1)) /* shall we now? */
4691 CERROR("Dropping GET to %s: simulated failure\n",
4692 libcfs_id2str(target));
4696 msg = lnet_msg_alloc();
4698 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4699 libcfs_id2str(target));
4703 cpt = lnet_cpt_of_cookie(mdh.cookie);
4705 rspt = lnet_rspt_alloc(cpt);
4707 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4708 libcfs_id2str(target));
4711 INIT_LIST_HEAD(&rspt->rspt_on_list);
4713 msg->msg_recovery = recovery;
4717 md = lnet_handle2md(&mdh);
4718 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4719 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4720 match_bits, portal, libcfs_id2str(target),
4721 md == NULL ? -1 : md->md_threshold);
4722 if (md != NULL && md->md_me != NULL)
4723 CERROR("REPLY MD also attached to portal %d\n",
4724 md->md_me->me_portal);
4726 lnet_res_unlock(cpt);
4729 LIBCFS_FREE(rspt, sizeof(*rspt));
4733 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4735 lnet_msg_attach_md(msg, md, 0, 0);
4737 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
4739 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
4740 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
4741 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
4742 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
4744 /* NB handles only looked up by creator (no flips) */
4745 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
4746 the_lnet.ln_interface_cookie;
4747 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
4748 md->md_lh.lh_cookie;
4750 lnet_res_unlock(cpt);
4752 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4754 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4756 rc = lnet_send(self, msg, LNET_NID_ANY);
4758 CNETERR("Error sending GET to %s: %d\n",
4759 libcfs_id2str(target), rc);
4760 msg->msg_no_resend = true;
4761 lnet_detach_rsp_tracker(msg->msg_md, cpt);
4762 lnet_finalize(msg, rc);
4765 /* completion will be signalled by an event */
4768 EXPORT_SYMBOL(LNetGet);
4771 * Calculate distance to node at \a dstnid.
4773 * \param dstnid Target NID.
4774 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
4776 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
4779 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
4780 * local_nid_dist_zero is set, which is the default.
4781 * \retval positives Distance to target NID, i.e. number of hops plus one.
4782 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
4785 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
4787 struct list_head *e;
4788 struct lnet_ni *ni = NULL;
4789 struct lnet_remotenet *rnet;
4790 __u32 dstnet = LNET_NIDNET(dstnid);
4794 struct list_head *rn_list;
4796 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
4797 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
4798 * keep order 0 free for 0@lo and order 1 free for a local NID
4801 LASSERT(the_lnet.ln_refcount > 0);
4803 cpt = lnet_net_lock_current();
4805 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
4806 if (ni->ni_nid == dstnid) {
4807 if (srcnidp != NULL)
4809 if (orderp != NULL) {
4810 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
4815 lnet_net_unlock(cpt);
4817 return local_nid_dist_zero ? 0 : 1;
4820 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
4821 /* Check if ni was originally created in
4822 * current net namespace.
4823 * If not, assign order above 0xffff0000,
4824 * to make this ni not a priority. */
4825 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
4826 order += 0xffff0000;
4828 if (srcnidp != NULL)
4829 *srcnidp = ni->ni_nid;
4832 lnet_net_unlock(cpt);
4839 rn_list = lnet_net2rnethash(dstnet);
4840 list_for_each(e, rn_list) {
4841 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
4843 if (rnet->lrn_net == dstnet) {
4844 struct lnet_route *route;
4845 struct lnet_route *shortest = NULL;
4846 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
4849 LASSERT(!list_empty(&rnet->lrn_routes));
4851 list_for_each_entry(route, &rnet->lrn_routes,
4853 route_hops = route->lr_hops;
4854 if (route_hops == LNET_UNDEFINED_HOPS)
4856 if (shortest == NULL ||
4857 route_hops < shortest_hops) {
4859 shortest_hops = route_hops;
4863 LASSERT(shortest != NULL);
4864 hops = shortest_hops;
4865 if (srcnidp != NULL) {
4866 ni = lnet_get_next_ni_locked(
4867 shortest->lr_gateway->lpni_net,
4869 *srcnidp = ni->ni_nid;
4873 lnet_net_unlock(cpt);
4879 lnet_net_unlock(cpt);
4880 return -EHOSTUNREACH;
4882 EXPORT_SYMBOL(LNetDist);