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, 2016, 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 <lnet/lib-lnet.h>
40 #include <linux/nsproxy.h>
41 #include <net/net_namespace.h>
43 static int local_nid_dist_zero = 1;
44 module_param(local_nid_dist_zero, int, 0444);
45 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
47 static inline struct lnet_comm_count *
48 get_stats_counts(struct lnet_element_stats *stats,
49 enum lnet_stats_type stats_type)
52 case LNET_STATS_TYPE_SEND:
53 return &stats->el_send_stats;
54 case LNET_STATS_TYPE_RECV:
55 return &stats->el_recv_stats;
56 case LNET_STATS_TYPE_DROP:
57 return &stats->el_drop_stats;
59 CERROR("Unknown stats type\n");
65 void lnet_incr_stats(struct lnet_element_stats *stats, lnet_msg_type_t msg_type,
66 enum lnet_stats_type stats_type)
68 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
74 atomic_inc(&counts->co_ack_count);
77 atomic_inc(&counts->co_put_count);
80 atomic_inc(&counts->co_get_count);
83 atomic_inc(&counts->co_reply_count);
86 atomic_inc(&counts->co_hello_count);
89 CERROR("There is a BUG in the code. Unknown message type\n");
94 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
95 enum lnet_stats_type stats_type)
97 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
101 return (atomic_read(&counts->co_ack_count) +
102 atomic_read(&counts->co_put_count) +
103 atomic_read(&counts->co_get_count) +
104 atomic_read(&counts->co_reply_count) +
105 atomic_read(&counts->co_hello_count));
108 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
109 struct lnet_comm_count *counts)
111 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
112 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
113 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
114 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
115 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
118 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
119 struct lnet_element_stats *stats)
121 struct lnet_comm_count *counts;
126 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
129 assign_stats(&msg_stats->im_send_stats, counts);
131 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
134 assign_stats(&msg_stats->im_recv_stats, counts);
136 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
139 assign_stats(&msg_stats->im_drop_stats, counts);
143 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
145 struct lnet_test_peer *tp;
146 struct list_head *el;
147 struct list_head *next;
148 struct list_head cull;
150 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
151 if (threshold != 0) {
152 /* Adding a new entry */
153 LIBCFS_ALLOC(tp, sizeof(*tp));
158 tp->tp_threshold = threshold;
161 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
166 /* removing entries */
167 INIT_LIST_HEAD(&cull);
171 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
172 tp = list_entry(el, struct lnet_test_peer, tp_list);
174 if (tp->tp_threshold == 0 || /* needs culling anyway */
175 nid == LNET_NID_ANY || /* removing all entries */
176 tp->tp_nid == nid) { /* matched this one */
177 list_del(&tp->tp_list);
178 list_add(&tp->tp_list, &cull);
184 while (!list_empty(&cull)) {
185 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
187 list_del(&tp->tp_list);
188 LIBCFS_FREE(tp, sizeof(*tp));
194 fail_peer (lnet_nid_t nid, int outgoing)
196 struct lnet_test_peer *tp;
197 struct list_head *el;
198 struct list_head *next;
199 struct list_head cull;
202 INIT_LIST_HEAD(&cull);
204 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
207 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
208 tp = list_entry(el, struct lnet_test_peer, tp_list);
210 if (tp->tp_threshold == 0) {
213 /* only cull zombies on outgoing tests,
214 * since we may be at interrupt priority on
215 * incoming messages. */
216 list_del(&tp->tp_list);
217 list_add(&tp->tp_list, &cull);
222 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
223 nid == tp->tp_nid) { /* fail this peer */
226 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
229 tp->tp_threshold == 0) {
231 list_del(&tp->tp_list);
232 list_add(&tp->tp_list, &cull);
241 while (!list_empty(&cull)) {
242 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
243 list_del(&tp->tp_list);
245 LIBCFS_FREE(tp, sizeof(*tp));
252 lnet_iov_nob(unsigned int niov, struct kvec *iov)
254 unsigned int nob = 0;
256 LASSERT(niov == 0 || iov != NULL);
258 nob += (iov++)->iov_len;
262 EXPORT_SYMBOL(lnet_iov_nob);
265 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
266 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
269 /* NB diov, siov are READ-ONLY */
270 unsigned int this_nob;
275 /* skip complete frags before 'doffset' */
277 while (doffset >= diov->iov_len) {
278 doffset -= diov->iov_len;
284 /* skip complete frags before 'soffset' */
286 while (soffset >= siov->iov_len) {
287 soffset -= siov->iov_len;
296 this_nob = MIN(diov->iov_len - doffset,
297 siov->iov_len - soffset);
298 this_nob = MIN(this_nob, nob);
300 memcpy((char *)diov->iov_base + doffset,
301 (char *)siov->iov_base + soffset, this_nob);
304 if (diov->iov_len > doffset + this_nob) {
312 if (siov->iov_len > soffset + this_nob) {
321 EXPORT_SYMBOL(lnet_copy_iov2iov);
324 lnet_extract_iov(int dst_niov, struct kvec *dst,
325 int src_niov, struct kvec *src,
326 unsigned int offset, unsigned int len)
328 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
329 * for exactly 'len' bytes, and return the number of entries.
330 * NB not destructive to 'src' */
331 unsigned int frag_len;
334 if (len == 0) /* no data => */
335 return (0); /* no frags */
337 LASSERT(src_niov > 0);
338 while (offset >= src->iov_len) { /* skip initial frags */
339 offset -= src->iov_len;
342 LASSERT(src_niov > 0);
347 LASSERT(src_niov > 0);
348 LASSERT((int)niov <= dst_niov);
350 frag_len = src->iov_len - offset;
351 dst->iov_base = ((char *)src->iov_base) + offset;
353 if (len <= frag_len) {
358 dst->iov_len = frag_len;
368 EXPORT_SYMBOL(lnet_extract_iov);
372 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
374 unsigned int nob = 0;
376 LASSERT(niov == 0 || kiov != NULL);
378 nob += (kiov++)->kiov_len;
382 EXPORT_SYMBOL(lnet_kiov_nob);
385 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
386 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
389 /* NB diov, siov are READ-ONLY */
390 unsigned int this_nob;
397 LASSERT (!in_interrupt ());
400 while (doffset >= diov->kiov_len) {
401 doffset -= diov->kiov_len;
408 while (soffset >= siov->kiov_len) {
409 soffset -= siov->kiov_len;
418 this_nob = MIN(diov->kiov_len - doffset,
419 siov->kiov_len - soffset);
420 this_nob = MIN(this_nob, nob);
423 daddr = ((char *)kmap(diov->kiov_page)) +
424 diov->kiov_offset + doffset;
426 saddr = ((char *)kmap(siov->kiov_page)) +
427 siov->kiov_offset + soffset;
429 /* Vanishing risk of kmap deadlock when mapping 2 pages.
430 * However in practice at least one of the kiovs will be mapped
431 * kernel pages and the map/unmap will be NOOPs */
433 memcpy (daddr, saddr, this_nob);
436 if (diov->kiov_len > doffset + this_nob) {
440 kunmap(diov->kiov_page);
447 if (siov->kiov_len > soffset + this_nob) {
451 kunmap(siov->kiov_page);
460 kunmap(diov->kiov_page);
462 kunmap(siov->kiov_page);
464 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
467 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
468 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
471 /* NB iov, kiov are READ-ONLY */
472 unsigned int this_nob;
478 LASSERT (!in_interrupt ());
481 while (iovoffset >= iov->iov_len) {
482 iovoffset -= iov->iov_len;
489 while (kiovoffset >= kiov->kiov_len) {
490 kiovoffset -= kiov->kiov_len;
499 this_nob = MIN(iov->iov_len - iovoffset,
500 kiov->kiov_len - kiovoffset);
501 this_nob = MIN(this_nob, nob);
504 addr = ((char *)kmap(kiov->kiov_page)) +
505 kiov->kiov_offset + kiovoffset;
507 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
510 if (iov->iov_len > iovoffset + this_nob) {
511 iovoffset += this_nob;
518 if (kiov->kiov_len > kiovoffset + this_nob) {
520 kiovoffset += this_nob;
522 kunmap(kiov->kiov_page);
532 kunmap(kiov->kiov_page);
534 EXPORT_SYMBOL(lnet_copy_kiov2iov);
537 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
538 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
541 /* NB kiov, iov are READ-ONLY */
542 unsigned int this_nob;
548 LASSERT (!in_interrupt ());
551 while (kiovoffset >= kiov->kiov_len) {
552 kiovoffset -= kiov->kiov_len;
559 while (iovoffset >= iov->iov_len) {
560 iovoffset -= iov->iov_len;
569 this_nob = MIN(kiov->kiov_len - kiovoffset,
570 iov->iov_len - iovoffset);
571 this_nob = MIN(this_nob, nob);
574 addr = ((char *)kmap(kiov->kiov_page)) +
575 kiov->kiov_offset + kiovoffset;
577 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
580 if (kiov->kiov_len > kiovoffset + this_nob) {
582 kiovoffset += this_nob;
584 kunmap(kiov->kiov_page);
591 if (iov->iov_len > iovoffset + this_nob) {
592 iovoffset += this_nob;
601 kunmap(kiov->kiov_page);
603 EXPORT_SYMBOL(lnet_copy_iov2kiov);
606 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
607 int src_niov, lnet_kiov_t *src,
608 unsigned int offset, unsigned int len)
610 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
611 * for exactly 'len' bytes, and return the number of entries.
612 * NB not destructive to 'src' */
613 unsigned int frag_len;
616 if (len == 0) /* no data => */
617 return (0); /* no frags */
619 LASSERT(src_niov > 0);
620 while (offset >= src->kiov_len) { /* skip initial frags */
621 offset -= src->kiov_len;
624 LASSERT(src_niov > 0);
629 LASSERT(src_niov > 0);
630 LASSERT((int)niov <= dst_niov);
632 frag_len = src->kiov_len - offset;
633 dst->kiov_page = src->kiov_page;
634 dst->kiov_offset = src->kiov_offset + offset;
636 if (len <= frag_len) {
638 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
642 dst->kiov_len = frag_len;
643 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
653 EXPORT_SYMBOL(lnet_extract_kiov);
656 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
657 int delayed, unsigned int offset, unsigned int mlen,
660 unsigned int niov = 0;
661 struct kvec *iov = NULL;
662 lnet_kiov_t *kiov = NULL;
665 LASSERT (!in_interrupt ());
666 LASSERT (mlen == 0 || msg != NULL);
669 LASSERT(msg->msg_receiving);
670 LASSERT(!msg->msg_sending);
671 LASSERT(rlen == msg->msg_len);
672 LASSERT(mlen <= msg->msg_len);
673 LASSERT(msg->msg_offset == offset);
674 LASSERT(msg->msg_wanted == mlen);
676 msg->msg_receiving = 0;
679 niov = msg->msg_niov;
681 kiov = msg->msg_kiov;
684 LASSERT ((iov == NULL) != (kiov == NULL));
688 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
689 niov, iov, kiov, offset, mlen,
692 lnet_finalize(msg, rc);
696 lnet_setpayloadbuffer(struct lnet_msg *msg)
698 struct lnet_libmd *md = msg->msg_md;
700 LASSERT(msg->msg_len > 0);
701 LASSERT(!msg->msg_routing);
703 LASSERT(msg->msg_niov == 0);
704 LASSERT(msg->msg_iov == NULL);
705 LASSERT(msg->msg_kiov == NULL);
707 msg->msg_niov = md->md_niov;
708 if ((md->md_options & LNET_MD_KIOV) != 0)
709 msg->msg_kiov = md->md_iov.kiov;
711 msg->msg_iov = md->md_iov.iov;
715 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
716 unsigned int offset, unsigned int len)
718 msg->msg_type = type;
719 msg->msg_target = target;
721 msg->msg_offset = offset;
724 lnet_setpayloadbuffer(msg);
726 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
727 msg->msg_hdr.type = cpu_to_le32(type);
728 /* dest_nid will be overwritten by lnet_select_pathway() */
729 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
730 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
731 /* src_nid will be set later */
732 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
733 msg->msg_hdr.payload_length = cpu_to_le32(len);
737 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
739 void *priv = msg->msg_private;
742 LASSERT (!in_interrupt ());
743 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
744 (msg->msg_txcredit && msg->msg_peertxcredit));
746 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
748 lnet_finalize(msg, rc);
752 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
756 LASSERT(!msg->msg_sending);
757 LASSERT(msg->msg_receiving);
758 LASSERT(!msg->msg_rx_ready_delay);
759 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
761 msg->msg_rx_ready_delay = 1;
762 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
765 CERROR("recv from %s / send to %s aborted: "
766 "eager_recv failed %d\n",
767 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
768 libcfs_id2str(msg->msg_target), rc);
769 LASSERT(rc < 0); /* required by my callers */
776 * This function can be called from two paths:
777 * 1. when sending a message
778 * 2. when decommiting a message (lnet_msg_decommit_tx())
779 * In both these cases the peer_ni should have it's reference count
780 * acquired by the caller and therefore it is safe to drop the spin
781 * lock before calling lnd_query()
784 lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
786 cfs_time_t last_alive = 0;
787 int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
789 LASSERT(lnet_peer_aliveness_enabled(lp));
790 LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
792 lnet_net_unlock(cpt);
793 (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
796 lp->lpni_last_query = cfs_time_current();
798 if (last_alive != 0) /* NI has updated timestamp */
799 lp->lpni_last_alive = last_alive;
802 /* NB: always called with lnet_net_lock held */
804 lnet_peer_is_alive (struct lnet_peer_ni *lp, cfs_time_t now)
809 LASSERT (lnet_peer_aliveness_enabled(lp));
812 * Trust lnet_notify() if it has more recent aliveness news, but
813 * ignore the initial assumed death (see lnet_peers_start_down()).
815 spin_lock(&lp->lpni_lock);
816 if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
817 cfs_time_aftereq(lp->lpni_timestamp, lp->lpni_last_alive)) {
818 spin_unlock(&lp->lpni_lock);
823 cfs_time_add(lp->lpni_last_alive,
824 cfs_time_seconds(lp->lpni_net->net_tunables.
826 alive = cfs_time_after(deadline, now);
829 * Update obsolete lp_alive except for routers assumed to be dead
830 * initially, because router checker would update aliveness in this
831 * case, and moreover lpni_last_alive at peer creation is assumed.
833 if (alive && !lp->lpni_alive &&
834 !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
835 spin_unlock(&lp->lpni_lock);
836 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
838 spin_unlock(&lp->lpni_lock);
845 /* NB: returns 1 when alive, 0 when dead, negative when error;
846 * may drop the lnet_net_lock */
848 lnet_peer_alive_locked (struct lnet_ni *ni, struct lnet_peer_ni *lp)
850 cfs_time_t now = cfs_time_current();
852 if (!lnet_peer_aliveness_enabled(lp))
855 if (lnet_peer_is_alive(lp, now))
859 * Peer appears dead, but we should avoid frequent NI queries (at
860 * most once per lnet_queryinterval seconds).
862 if (lp->lpni_last_query != 0) {
863 static const int lnet_queryinterval = 1;
865 cfs_time_t next_query =
866 cfs_time_add(lp->lpni_last_query,
867 cfs_time_seconds(lnet_queryinterval));
869 if (cfs_time_before(now, next_query)) {
871 CWARN("Unexpected aliveness of peer %s: "
873 libcfs_nid2str(lp->lpni_nid),
874 (int)now, (int)next_query,
876 lp->lpni_net->net_tunables.lct_peer_timeout);
881 /* query NI for latest aliveness news */
882 lnet_ni_query_locked(ni, lp);
884 if (lnet_peer_is_alive(lp, now))
887 lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
892 * \param msg The message to be sent.
893 * \param do_send True if lnet_ni_send() should be called in this function.
894 * lnet_send() is going to lnet_net_unlock immediately after this, so
895 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
897 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
898 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
899 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
900 * \retval -ECANCELED If the MD of the message has been unlinked.
903 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
905 struct lnet_peer_ni *lp = msg->msg_txpeer;
906 struct lnet_ni *ni = msg->msg_txni;
907 int cpt = msg->msg_tx_cpt;
908 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
910 /* non-lnet_send() callers have checked before */
911 LASSERT(!do_send || msg->msg_tx_delayed);
912 LASSERT(!msg->msg_receiving);
913 LASSERT(msg->msg_tx_committed);
915 /* NB 'lp' is always the next hop */
916 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
917 lnet_peer_alive_locked(ni, lp) == 0) {
918 the_lnet.ln_counters[cpt]->drop_count++;
919 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
920 lnet_net_unlock(cpt);
922 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
924 LNET_STATS_TYPE_DROP);
926 lnet_incr_stats(&msg->msg_txni->ni_stats,
928 LNET_STATS_TYPE_DROP);
930 CNETERR("Dropping message for %s: peer not alive\n",
931 libcfs_id2str(msg->msg_target));
933 lnet_finalize(msg, -EHOSTUNREACH);
936 return -EHOSTUNREACH;
939 if (msg->msg_md != NULL &&
940 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
941 lnet_net_unlock(cpt);
943 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
944 "called on the MD/ME.\n",
945 libcfs_id2str(msg->msg_target));
947 lnet_finalize(msg, -ECANCELED);
953 if (!msg->msg_peertxcredit) {
954 spin_lock(&lp->lpni_lock);
955 LASSERT((lp->lpni_txcredits < 0) ==
956 !list_empty(&lp->lpni_txq));
958 msg->msg_peertxcredit = 1;
959 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
960 lp->lpni_txcredits--;
962 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
963 lp->lpni_mintxcredits = lp->lpni_txcredits;
965 if (lp->lpni_txcredits < 0) {
966 msg->msg_tx_delayed = 1;
967 list_add_tail(&msg->msg_list, &lp->lpni_txq);
968 spin_unlock(&lp->lpni_lock);
969 return LNET_CREDIT_WAIT;
971 spin_unlock(&lp->lpni_lock);
974 if (!msg->msg_txcredit) {
975 LASSERT((tq->tq_credits < 0) ==
976 !list_empty(&tq->tq_delayed));
978 msg->msg_txcredit = 1;
980 atomic_dec(&ni->ni_tx_credits);
982 if (tq->tq_credits < tq->tq_credits_min)
983 tq->tq_credits_min = tq->tq_credits;
985 if (tq->tq_credits < 0) {
986 msg->msg_tx_delayed = 1;
987 list_add_tail(&msg->msg_list, &tq->tq_delayed);
988 return LNET_CREDIT_WAIT;
993 lnet_net_unlock(cpt);
994 lnet_ni_send(ni, msg);
997 return LNET_CREDIT_OK;
1001 static struct lnet_rtrbufpool *
1002 lnet_msg2bufpool(struct lnet_msg *msg)
1004 struct lnet_rtrbufpool *rbp;
1007 LASSERT(msg->msg_rx_committed);
1009 cpt = msg->msg_rx_cpt;
1010 rbp = &the_lnet.ln_rtrpools[cpt][0];
1012 LASSERT(msg->msg_len <= LNET_MTU);
1013 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
1015 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
1022 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
1024 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
1025 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
1026 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
1027 * received or OK to receive */
1028 struct lnet_peer_ni *lp = msg->msg_rxpeer;
1029 struct lnet_rtrbufpool *rbp;
1030 struct lnet_rtrbuf *rb;
1032 LASSERT (msg->msg_iov == NULL);
1033 LASSERT (msg->msg_kiov == NULL);
1034 LASSERT (msg->msg_niov == 0);
1035 LASSERT (msg->msg_routing);
1036 LASSERT (msg->msg_receiving);
1037 LASSERT (!msg->msg_sending);
1039 /* non-lnet_parse callers only receive delayed messages */
1040 LASSERT(!do_recv || msg->msg_rx_delayed);
1042 if (!msg->msg_peerrtrcredit) {
1043 spin_lock(&lp->lpni_lock);
1044 LASSERT((lp->lpni_rtrcredits < 0) ==
1045 !list_empty(&lp->lpni_rtrq));
1047 msg->msg_peerrtrcredit = 1;
1048 lp->lpni_rtrcredits--;
1049 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
1050 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
1052 if (lp->lpni_rtrcredits < 0) {
1053 /* must have checked eager_recv before here */
1054 LASSERT(msg->msg_rx_ready_delay);
1055 msg->msg_rx_delayed = 1;
1056 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
1057 spin_unlock(&lp->lpni_lock);
1058 return LNET_CREDIT_WAIT;
1060 spin_unlock(&lp->lpni_lock);
1063 rbp = lnet_msg2bufpool(msg);
1065 if (!msg->msg_rtrcredit) {
1066 msg->msg_rtrcredit = 1;
1068 if (rbp->rbp_credits < rbp->rbp_mincredits)
1069 rbp->rbp_mincredits = rbp->rbp_credits;
1071 if (rbp->rbp_credits < 0) {
1072 /* must have checked eager_recv before here */
1073 LASSERT(msg->msg_rx_ready_delay);
1074 msg->msg_rx_delayed = 1;
1075 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1076 return LNET_CREDIT_WAIT;
1080 LASSERT(!list_empty(&rbp->rbp_bufs));
1081 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1082 list_del(&rb->rb_list);
1084 msg->msg_niov = rbp->rbp_npages;
1085 msg->msg_kiov = &rb->rb_kiov[0];
1088 int cpt = msg->msg_rx_cpt;
1090 lnet_net_unlock(cpt);
1091 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1092 0, msg->msg_len, msg->msg_len);
1095 return LNET_CREDIT_OK;
1099 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1101 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1102 struct lnet_ni *txni = msg->msg_txni;
1103 struct lnet_msg *msg2;
1105 if (msg->msg_txcredit) {
1106 struct lnet_ni *ni = msg->msg_txni;
1107 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1109 /* give back NI txcredits */
1110 msg->msg_txcredit = 0;
1112 LASSERT((tq->tq_credits < 0) ==
1113 !list_empty(&tq->tq_delayed));
1116 atomic_inc(&ni->ni_tx_credits);
1117 if (tq->tq_credits <= 0) {
1118 msg2 = list_entry(tq->tq_delayed.next,
1119 struct lnet_msg, msg_list);
1120 list_del(&msg2->msg_list);
1122 LASSERT(msg2->msg_txni == ni);
1123 LASSERT(msg2->msg_tx_delayed);
1124 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1126 (void) lnet_post_send_locked(msg2, 1);
1130 if (msg->msg_peertxcredit) {
1131 /* give back peer txcredits */
1132 msg->msg_peertxcredit = 0;
1134 spin_lock(&txpeer->lpni_lock);
1135 LASSERT((txpeer->lpni_txcredits < 0) ==
1136 !list_empty(&txpeer->lpni_txq));
1138 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1139 LASSERT(txpeer->lpni_txqnob >= 0);
1141 txpeer->lpni_txcredits++;
1142 if (txpeer->lpni_txcredits <= 0) {
1145 msg2 = list_entry(txpeer->lpni_txq.next,
1146 struct lnet_msg, msg_list);
1147 list_del(&msg2->msg_list);
1148 spin_unlock(&txpeer->lpni_lock);
1150 LASSERT(msg2->msg_txpeer == txpeer);
1151 LASSERT(msg2->msg_tx_delayed);
1153 msg2_cpt = msg2->msg_tx_cpt;
1156 * The msg_cpt can be different from the msg2_cpt
1157 * so we need to make sure we lock the correct cpt
1159 * Once we call lnet_post_send_locked() it is no
1160 * longer safe to access msg2, since it could've
1161 * been freed by lnet_finalize(), but we still
1162 * need to relock the correct cpt, so we cache the
1163 * msg2_cpt for the purpose of the check that
1164 * follows the call to lnet_pose_send_locked().
1166 if (msg2_cpt != msg->msg_tx_cpt) {
1167 lnet_net_unlock(msg->msg_tx_cpt);
1168 lnet_net_lock(msg2_cpt);
1170 (void) lnet_post_send_locked(msg2, 1);
1171 if (msg2_cpt != msg->msg_tx_cpt) {
1172 lnet_net_unlock(msg2_cpt);
1173 lnet_net_lock(msg->msg_tx_cpt);
1176 spin_unlock(&txpeer->lpni_lock);
1181 msg->msg_txni = NULL;
1182 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1185 if (txpeer != NULL) {
1188 * Once the patch for the health comes in we need to set
1189 * the health of the peer ni to bad when we fail to send
1191 * int status = msg->msg_ev.status;
1193 * lnet_set_peer_ni_health_locked(txpeer, false)
1195 msg->msg_txpeer = NULL;
1196 lnet_peer_ni_decref_locked(txpeer);
1201 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1203 struct lnet_msg *msg;
1205 if (list_empty(&rbp->rbp_msgs))
1207 msg = list_entry(rbp->rbp_msgs.next,
1208 struct lnet_msg, msg_list);
1209 list_del(&msg->msg_list);
1211 (void)lnet_post_routed_recv_locked(msg, 1);
1215 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1217 struct lnet_msg *msg;
1218 struct lnet_msg *tmp;
1220 lnet_net_unlock(cpt);
1222 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1223 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1224 0, 0, 0, msg->msg_hdr.payload_length);
1225 list_del_init(&msg->msg_list);
1226 lnet_finalize(msg, -ECANCELED);
1233 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1235 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1236 struct lnet_ni *rxni = msg->msg_rxni;
1237 struct lnet_msg *msg2;
1239 if (msg->msg_rtrcredit) {
1240 /* give back global router credits */
1241 struct lnet_rtrbuf *rb;
1242 struct lnet_rtrbufpool *rbp;
1244 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1245 * there until it gets one allocated, or aborts the wait
1247 LASSERT(msg->msg_kiov != NULL);
1249 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1252 msg->msg_kiov = NULL;
1253 msg->msg_rtrcredit = 0;
1255 LASSERT(rbp == lnet_msg2bufpool(msg));
1257 LASSERT((rbp->rbp_credits > 0) ==
1258 !list_empty(&rbp->rbp_bufs));
1260 /* If routing is now turned off, we just drop this buffer and
1261 * don't bother trying to return credits. */
1262 if (!the_lnet.ln_routing) {
1263 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1267 /* It is possible that a user has lowered the desired number of
1268 * buffers in this pool. Make sure we never put back
1269 * more buffers than the stated number. */
1270 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1271 /* Discard this buffer so we don't have too
1273 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1274 rbp->rbp_nbuffers--;
1276 list_add(&rb->rb_list, &rbp->rbp_bufs);
1278 if (rbp->rbp_credits <= 0)
1279 lnet_schedule_blocked_locked(rbp);
1284 if (msg->msg_peerrtrcredit) {
1285 /* give back peer router credits */
1286 msg->msg_peerrtrcredit = 0;
1288 spin_lock(&rxpeer->lpni_lock);
1289 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1290 !list_empty(&rxpeer->lpni_rtrq));
1292 rxpeer->lpni_rtrcredits++;
1294 /* drop all messages which are queued to be routed on that
1296 if (!the_lnet.ln_routing) {
1297 struct list_head drop;
1298 INIT_LIST_HEAD(&drop);
1299 list_splice_init(&rxpeer->lpni_rtrq, &drop);
1300 spin_unlock(&rxpeer->lpni_lock);
1301 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1302 } else if (rxpeer->lpni_rtrcredits <= 0) {
1303 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1304 struct lnet_msg, msg_list);
1305 list_del(&msg2->msg_list);
1306 spin_unlock(&rxpeer->lpni_lock);
1307 (void) lnet_post_routed_recv_locked(msg2, 1);
1309 spin_unlock(&rxpeer->lpni_lock);
1313 msg->msg_rxni = NULL;
1314 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1316 if (rxpeer != NULL) {
1317 msg->msg_rxpeer = NULL;
1318 lnet_peer_ni_decref_locked(rxpeer);
1323 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1325 if (p1->lpni_txqnob < p2->lpni_txqnob)
1328 if (p1->lpni_txqnob > p2->lpni_txqnob)
1331 if (p1->lpni_txcredits > p2->lpni_txcredits)
1334 if (p1->lpni_txcredits < p2->lpni_txcredits)
1341 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1343 struct lnet_peer_ni *p1 = r1->lr_gateway;
1344 struct lnet_peer_ni *p2 = r2->lr_gateway;
1345 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1346 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1349 if (r1->lr_priority < r2->lr_priority)
1352 if (r1->lr_priority > r2->lr_priority)
1355 if (r1_hops < r2_hops)
1358 if (r1_hops > r2_hops)
1361 rc = lnet_compare_peers(p1, p2);
1365 if (r1->lr_seq - r2->lr_seq <= 0)
1371 static struct lnet_peer_ni *
1372 lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
1375 struct lnet_remotenet *rnet;
1376 struct lnet_route *route;
1377 struct lnet_route *best_route;
1378 struct lnet_route *last_route;
1379 struct lnet_peer_ni *lpni_best;
1380 struct lnet_peer_ni *lp;
1383 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1384 * rtr_nid nid, otherwise find the best gateway I can use */
1386 rnet = lnet_find_rnet_locked(LNET_NIDNET(target));
1391 best_route = last_route = NULL;
1392 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1393 lp = route->lr_gateway;
1395 if (!lnet_is_route_alive(route))
1398 if (net != NULL && lp->lpni_net != net)
1401 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1404 if (lpni_best == NULL) {
1405 best_route = last_route = route;
1410 /* no protection on below fields, but it's harmless */
1411 if (last_route->lr_seq - route->lr_seq < 0)
1414 rc = lnet_compare_routes(route, best_route);
1422 /* set sequence number on the best router to the latest sequence + 1
1423 * so we can round-robin all routers, it's race and inaccurate but
1424 * harmless and functional */
1425 if (best_route != NULL)
1426 best_route->lr_seq = last_route->lr_seq + 1;
1430 static struct lnet_ni *
1431 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *cur_ni,
1434 struct lnet_ni *ni = NULL, *best_ni = cur_ni;
1435 unsigned int shortest_distance;
1438 if (best_ni == NULL) {
1439 shortest_distance = UINT_MAX;
1440 best_credits = INT_MIN;
1442 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1443 best_ni->ni_dev_cpt);
1444 best_credits = atomic_read(&best_ni->ni_tx_credits);
1447 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1448 unsigned int distance;
1451 if (!lnet_is_ni_healthy_locked(ni))
1454 ni_credits = atomic_read(&ni->ni_tx_credits);
1457 * calculate the distance from the CPT on which
1458 * the message memory is allocated to the CPT of
1459 * the NI's physical device
1461 distance = cfs_cpt_distance(lnet_cpt_table(),
1466 * All distances smaller than the NUMA range
1467 * are treated equally.
1469 if (distance < lnet_numa_range)
1470 distance = lnet_numa_range;
1473 * Select on shorter distance, then available
1474 * credits, then round-robin.
1476 if (distance > shortest_distance) {
1478 } else if (distance < shortest_distance) {
1479 shortest_distance = distance;
1480 } else if (ni_credits < best_credits) {
1482 } else if (ni_credits == best_credits) {
1483 if (best_ni && (best_ni)->ni_seq <= ni->ni_seq)
1487 best_credits = ni_credits;
1494 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1495 * because such traffic is required to perform discovery. We therefore
1496 * exclude all GET and PUT on that portal. We also exclude all ACK and
1497 * REPLY traffic, but that is because the portal is not tracked in the
1498 * message structure for these message types. We could restrict this
1499 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1502 lnet_msg_discovery(struct lnet_msg *msg)
1504 if (msg->msg_type == LNET_MSG_PUT) {
1505 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1507 } else if (msg->msg_type == LNET_MSG_GET) {
1508 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1515 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
1516 struct lnet_msg *msg, lnet_nid_t rtr_nid)
1518 struct lnet_ni *best_ni;
1519 struct lnet_peer_ni *best_lpni;
1520 struct lnet_peer_ni *best_gw;
1521 struct lnet_peer_ni *lpni;
1522 struct lnet_peer_ni *final_dst;
1523 struct lnet_peer *peer;
1524 struct lnet_peer_net *peer_net;
1525 struct lnet_net *local_net;
1532 int best_lpni_credits;
1536 * get an initial CPT to use for locking. The idea here is not to
1537 * serialize the calls to select_pathway, so that as many
1538 * operations can run concurrently as possible. To do that we use
1539 * the CPT where this call is being executed. Later on when we
1540 * determine the CPT to use in lnet_message_commit, we switch the
1541 * lock and check if there was any configuration change. If none,
1542 * then we proceed, if there is, then we restart the operation.
1544 cpt = lnet_net_lock_current();
1546 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
1547 if (md_cpt == CFS_CPT_ANY)
1558 local_found = false;
1561 * lnet_nid2peerni_locked() is the path that will find an
1562 * existing peer_ni, or create one and mark it as having been
1563 * created due to network traffic.
1565 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
1567 lnet_net_unlock(cpt);
1568 return PTR_ERR(lpni);
1571 * Now that we have a peer_ni, check if we want to discover
1572 * the peer. Traffic to the LNET_RESERVED_PORTAL should not
1573 * trigger discovery.
1575 peer = lpni->lpni_peer_net->lpn_peer;
1576 if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
1577 rc = lnet_discover_peer_locked(lpni, cpt, false);
1579 lnet_peer_ni_decref_locked(lpni);
1580 lnet_net_unlock(cpt);
1583 /* The peer may have changed. */
1584 peer = lpni->lpni_peer_net->lpn_peer;
1585 /* queue message and return */
1586 msg->msg_src_nid_param = src_nid;
1587 msg->msg_rtr_nid_param = rtr_nid;
1588 msg->msg_sending = 0;
1589 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1590 lnet_peer_ni_decref_locked(lpni);
1591 lnet_net_unlock(cpt);
1593 CDEBUG(D_NET, "%s pending discovery\n",
1594 libcfs_nid2str(peer->lp_primary_nid));
1596 return LNET_DC_WAIT;
1598 lnet_peer_ni_decref_locked(lpni);
1600 /* If peer is not healthy then can not send anything to it */
1601 if (!lnet_is_peer_healthy_locked(peer)) {
1602 lnet_net_unlock(cpt);
1603 return -EHOSTUNREACH;
1607 * STEP 1: first jab at determining best_ni
1608 * if src_nid is explicitly specified, then best_ni is already
1609 * pre-determiend for us. Otherwise we need to select the best
1610 * one to use later on
1612 if (src_nid != LNET_NID_ANY) {
1613 best_ni = lnet_nid2ni_locked(src_nid, cpt);
1615 lnet_net_unlock(cpt);
1616 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1617 "local nid\n", libcfs_nid2str(dst_nid),
1618 libcfs_nid2str(src_nid));
1623 if (msg->msg_type == LNET_MSG_REPLY ||
1624 msg->msg_type == LNET_MSG_ACK ||
1625 !lnet_peer_is_multi_rail(peer) ||
1628 * for replies we want to respond on the same peer_ni we
1629 * received the message on if possible. If not, then pick
1630 * a peer_ni to send to
1632 * if the peer is non-multi-rail then you want to send to
1633 * the dst_nid provided as well.
1635 * If the best_ni has already been determined, IE the
1636 * src_nid has been specified, then use the
1637 * destination_nid provided as well, since we're
1638 * continuing a series of related messages for the same
1641 * It is expected to find the lpni using dst_nid, since we
1642 * created it earlier.
1644 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1646 lnet_peer_ni_decref_locked(best_lpni);
1648 if (best_lpni && !lnet_get_net_locked(LNET_NIDNET(dst_nid))) {
1650 * this lpni is not on a local network so we need
1651 * to route this reply.
1653 best_gw = lnet_find_route_locked(NULL,
1654 best_lpni->lpni_nid,
1658 * RULE: Each node considers only the next-hop
1660 * We're going to route the message, so change the peer to
1663 LASSERT(best_gw->lpni_peer_net);
1664 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1665 peer = best_gw->lpni_peer_net->lpn_peer;
1668 * if the router is not multi-rail then use the best_gw
1669 * found to send the message to
1671 if (!lnet_peer_is_multi_rail(peer))
1672 best_lpni = best_gw;
1680 } else if (!best_lpni) {
1681 lnet_net_unlock(cpt);
1682 CERROR("unable to send msg_type %d to "
1683 "originating %s. Destination NID not in DB\n",
1684 msg->msg_type, libcfs_nid2str(dst_nid));
1690 * We must use a consistent source address when sending to a
1691 * non-MR peer. However, a non-MR peer can have multiple NIDs
1692 * on multiple networks, and we may even need to talk to this
1693 * peer on multiple networks -- certain types of
1694 * load-balancing configuration do this.
1696 * So we need to pick the NI the peer prefers for this
1697 * particular network.
1699 if (!lnet_peer_is_multi_rail(peer)) {
1701 lnet_net_unlock(cpt);
1702 CERROR("no route to %s\n",
1703 libcfs_nid2str(dst_nid));
1704 return -EHOSTUNREACH;
1707 /* best ni is already set if src_nid was provided */
1709 /* Get the target peer_ni */
1710 peer_net = lnet_peer_get_net_locked(peer,
1711 LNET_NIDNET(best_lpni->lpni_nid));
1712 LASSERT(peer_net != NULL);
1713 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
1715 if (lpni->lpni_pref_nnids == 0)
1717 LASSERT(lpni->lpni_pref_nnids == 1);
1718 best_ni = lnet_nid2ni_locked(
1719 lpni->lpni_pref.nid, cpt);
1723 /* if best_ni is still not set just pick one */
1725 best_ni = lnet_net2ni_locked(
1726 best_lpni->lpni_net->net_id, cpt);
1727 /* If there is no best_ni we don't have a route */
1729 lnet_net_unlock(cpt);
1730 CERROR("no path to %s from net %s\n",
1731 libcfs_nid2str(best_lpni->lpni_nid),
1732 libcfs_net2str(best_lpni->lpni_net->net_id));
1733 return -EHOSTUNREACH;
1735 lpni = list_entry(peer_net->lpn_peer_nis.next,
1736 struct lnet_peer_ni,
1739 /* Set preferred NI if necessary. */
1740 if (lpni->lpni_pref_nnids == 0)
1741 lnet_peer_ni_set_non_mr_pref_nid(lpni, best_ni->ni_nid);
1745 * if we already found a best_ni because src_nid is specified and
1746 * best_lpni because we are replying to a message then just send
1749 if (best_ni && best_lpni)
1753 * If we already found a best_ni because src_nid is specified then
1754 * pick the peer then send the message
1760 * pick the best_ni by going through all the possible networks of
1761 * that peer and see which local NI is best suited to talk to that
1764 * Locally connected networks will always be preferred over
1765 * a routed network. If there are only routed paths to the peer,
1766 * then the best route is chosen. If all routes are equal then
1767 * they are used in round robin.
1769 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
1770 if (!lnet_is_peer_net_healthy_locked(peer_net))
1773 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1774 if (!local_net && !routing && !local_found) {
1775 struct lnet_peer_ni *net_gw;
1777 lpni = list_entry(peer_net->lpn_peer_nis.next,
1778 struct lnet_peer_ni,
1781 net_gw = lnet_find_route_locked(NULL,
1789 * lnet_find_route_locked() call
1790 * will return the best_Gw on the
1791 * lpni->lpni_nid network.
1792 * However, best_gw and net_gw can
1793 * be on different networks.
1794 * Therefore need to compare them
1795 * to pick the better of either.
1797 if (lnet_compare_peers(best_gw, net_gw) > 0)
1799 if (best_gw->lpni_gw_seq <= net_gw->lpni_gw_seq)
1814 * a gw on this network is found, but there could be
1815 * other better gateways on other networks. So don't pick
1816 * the best_ni until we determine the best_gw.
1821 /* if no local_net found continue */
1826 * Iterate through the NIs in this local Net and select
1827 * the NI to send from. The selection is determined by
1828 * these 3 criterion in the following priority:
1830 * 2. NI available credits
1833 best_ni = lnet_get_best_ni(local_net, best_ni, md_cpt);
1836 if (!best_ni && !best_gw) {
1837 lnet_net_unlock(cpt);
1838 LCONSOLE_WARN("No local ni found to send from to %s\n",
1839 libcfs_nid2str(dst_nid));
1844 best_ni = lnet_get_best_ni(best_gw->lpni_net, best_ni, md_cpt);
1845 LASSERT(best_gw && best_ni);
1848 * We're going to route the message, so change the peer to
1851 LASSERT(best_gw->lpni_peer_net);
1852 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1853 best_gw->lpni_gw_seq++;
1854 peer = best_gw->lpni_peer_net->lpn_peer;
1858 * Now that we selected the NI to use increment its sequence
1859 * number so the Round Robin algorithm will detect that it has
1860 * been used and pick the next NI.
1866 * At this point the best_ni is on a local network on which
1867 * the peer has a peer_ni as well
1869 peer_net = lnet_peer_get_net_locked(peer,
1870 best_ni->ni_net->net_id);
1872 * peer_net is not available or the src_nid is explicitly defined
1873 * and the peer_net for that src_nid is unhealthy. find a route to
1874 * the destination nid.
1877 (src_nid != LNET_NID_ANY &&
1878 !lnet_is_peer_net_healthy_locked(peer_net))) {
1879 best_gw = lnet_find_route_locked(best_ni->ni_net,
1883 * if no route is found for that network then
1884 * move onto the next peer_ni in the peer
1887 lnet_net_unlock(cpt);
1888 LCONSOLE_WARN("No route to peer from %s\n",
1889 libcfs_nid2str(best_ni->ni_nid));
1890 return -EHOSTUNREACH;
1893 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1894 libcfs_nid2str(dst_nid),
1895 libcfs_nid2str(best_gw->lpni_nid),
1896 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1900 * RULE: Each node considers only the next-hop
1902 * We're going to route the message, so change the peer to
1905 LASSERT(best_gw->lpni_peer_net);
1906 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1907 peer = best_gw->lpni_peer_net->lpn_peer;
1908 } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
1910 * this peer_net is unhealthy but we still have an opportunity
1911 * to find another peer_net that we can use
1913 __u32 net_id = peer_net->lpn_net_id;
1914 LCONSOLE_WARN("peer net %s unhealthy\n",
1915 libcfs_net2str(net_id));
1920 * Look at the peer NIs for the destination peer that connect
1921 * to the chosen net. If a peer_ni is preferred when using the
1922 * best_ni to communicate, we use that one. If there is no
1923 * preferred peer_ni, or there are multiple preferred peer_ni,
1924 * the available transmit credits are used. If the transmit
1925 * credits are equal, we round-robin over the peer_ni.
1928 best_lpni_credits = INT_MIN;
1931 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1933 * if this peer ni is not healthy just skip it, no point in
1934 * examining it further
1936 if (!lnet_is_peer_ni_healthy_locked(lpni))
1938 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1941 /* if this is a preferred peer use it */
1942 if (!preferred && ni_is_pref) {
1944 } else if (preferred && !ni_is_pref) {
1946 * this is not the preferred peer so let's ignore
1950 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1952 * We already have a peer that has more credits
1953 * available than this one. No need to consider
1954 * this peer further.
1957 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1959 * The best peer found so far and the current peer
1960 * have the same number of available credits let's
1961 * make sure to select between them using Round
1965 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1971 best_lpni_credits = lpni->lpni_txcredits;
1974 /* if we still can't find a peer ni then we can't reach it */
1976 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1977 LNET_NIDNET(dst_nid);
1978 lnet_net_unlock(cpt);
1979 LCONSOLE_WARN("no peer_ni found on peer net %s\n",
1980 libcfs_net2str(net_id));
1981 return -EHOSTUNREACH;
1986 /* Shortcut for loopback. */
1987 if (best_ni == the_lnet.ln_loni) {
1988 /* No send credit hassles with LOLND */
1989 lnet_ni_addref_locked(best_ni, cpt);
1990 msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
1991 if (!msg->msg_routing)
1992 msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
1993 msg->msg_target.nid = best_ni->ni_nid;
1994 lnet_msg_commit(msg, cpt);
1995 msg->msg_txni = best_ni;
1996 lnet_net_unlock(cpt);
1998 return LNET_CREDIT_OK;
2001 routing = routing || routing2;
2004 * Increment sequence number of the peer selected so that we
2005 * pick the next one in Round Robin.
2007 best_lpni->lpni_seq++;
2010 * grab a reference on the peer_ni so it sticks around even if
2011 * we need to drop and relock the lnet_net_lock below.
2013 lnet_peer_ni_addref_locked(best_lpni);
2016 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
2017 * message. This ensures that we get a CPT that is correct for
2018 * the NI when the NI has been restricted to a subset of all CPTs.
2019 * If the selected CPT differs from the one currently locked, we
2020 * must unlock and relock the lnet_net_lock(), and then check whether
2021 * the configuration has changed. We don't have a hold on the best_ni
2022 * yet, and it may have vanished.
2024 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
2026 __u32 seq = lnet_get_dlc_seq_locked();
2027 lnet_net_unlock(cpt);
2030 if (seq != lnet_get_dlc_seq_locked()) {
2031 lnet_peer_ni_decref_locked(best_lpni);
2037 * store the best_lpni in the message right away to avoid having
2038 * to do the same operation under different conditions
2040 msg->msg_txpeer = best_lpni;
2041 msg->msg_txni = best_ni;
2044 * grab a reference for the best_ni since now it's in use in this
2045 * send. the reference will need to be dropped when the message is
2046 * finished in lnet_finalize()
2048 lnet_ni_addref_locked(msg->msg_txni, cpt);
2051 * Always set the target.nid to the best peer picked. Either the
2052 * nid will be one of the preconfigured NIDs, or the same NID as
2053 * what was originally set in the target or it will be the NID of
2054 * a router if this message should be routed
2056 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
2059 * lnet_msg_commit assigns the correct cpt to the message, which
2060 * is used to decrement the correct refcount on the ni when it's
2061 * time to return the credits
2063 lnet_msg_commit(msg, cpt);
2066 * If we are routing the message then we don't need to overwrite
2067 * the src_nid since it would've been set at the origin. Otherwise
2068 * we are the originator so we need to set it.
2070 if (!msg->msg_routing)
2071 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
2074 msg->msg_target_is_router = 1;
2075 msg->msg_target.pid = LNET_PID_LUSTRE;
2077 * since we're routing we want to ensure that the
2078 * msg_hdr.dest_nid is set to the final destination. When
2079 * the router receives this message it knows how to route
2082 msg->msg_hdr.dest_nid =
2083 cpu_to_le64(final_dst ? final_dst->lpni_nid : dst_nid);
2086 * if we're not routing set the dest_nid to the best peer
2087 * ni that we picked earlier in the algorithm.
2089 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
2092 rc = lnet_post_send_locked(msg, 0);
2095 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s\n",
2096 libcfs_nid2str(msg->msg_hdr.src_nid),
2097 libcfs_nid2str(msg->msg_txni->ni_nid),
2098 libcfs_nid2str(src_nid),
2099 libcfs_nid2str(msg->msg_hdr.dest_nid),
2100 libcfs_nid2str(dst_nid),
2101 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
2102 lnet_msgtyp2str(msg->msg_type));
2104 lnet_net_unlock(cpt);
2110 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2112 lnet_nid_t dst_nid = msg->msg_target.nid;
2116 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2117 * but we might want to use pre-determined router for ACK/REPLY
2120 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2121 LASSERT (msg->msg_txpeer == NULL);
2122 LASSERT (!msg->msg_sending);
2123 LASSERT (!msg->msg_target_is_router);
2124 LASSERT (!msg->msg_receiving);
2126 msg->msg_sending = 1;
2128 LASSERT(!msg->msg_tx_committed);
2130 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2134 if (rc == LNET_CREDIT_OK)
2135 lnet_ni_send(msg->msg_txni, msg);
2137 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2142 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
2146 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
2147 the_lnet.ln_counters[cpt]->drop_count++;
2148 the_lnet.ln_counters[cpt]->drop_length += nob;
2149 lnet_net_unlock(cpt);
2151 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
2155 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
2157 struct lnet_hdr *hdr = &msg->msg_hdr;
2159 if (msg->msg_wanted != 0)
2160 lnet_setpayloadbuffer(msg);
2162 lnet_build_msg_event(msg, LNET_EVENT_PUT);
2164 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
2165 * it back into the ACK during lnet_finalize() */
2166 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
2167 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
2169 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
2170 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
2174 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
2176 struct lnet_hdr *hdr = &msg->msg_hdr;
2177 struct lnet_match_info info;
2181 /* Convert put fields to host byte order */
2182 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
2183 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
2184 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
2186 /* Primary peer NID. */
2187 info.mi_id.nid = msg->msg_initiator;
2188 info.mi_id.pid = hdr->src_pid;
2189 info.mi_opc = LNET_MD_OP_PUT;
2190 info.mi_portal = hdr->msg.put.ptl_index;
2191 info.mi_rlength = hdr->payload_length;
2192 info.mi_roffset = hdr->msg.put.offset;
2193 info.mi_mbits = hdr->msg.put.match_bits;
2194 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
2196 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
2197 ready_delay = msg->msg_rx_ready_delay;
2200 rc = lnet_ptl_match_md(&info, msg);
2205 case LNET_MATCHMD_OK:
2206 lnet_recv_put(ni, msg);
2209 case LNET_MATCHMD_NONE:
2211 /* no eager_recv or has already called it, should
2212 * have been attached on delayed list */
2215 rc = lnet_ni_eager_recv(ni, msg);
2222 case LNET_MATCHMD_DROP:
2223 CNETERR("Dropping PUT from %s portal %d match %llu"
2224 " offset %d length %d: %d\n",
2225 libcfs_id2str(info.mi_id), info.mi_portal,
2226 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
2228 return -ENOENT; /* -ve: OK but no match */
2233 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
2235 struct lnet_match_info info;
2236 struct lnet_hdr *hdr = &msg->msg_hdr;
2237 struct lnet_process_id source_id;
2238 struct lnet_handle_wire reply_wmd;
2241 /* Convert get fields to host byte order */
2242 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
2243 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
2244 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
2245 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
2247 source_id.nid = hdr->src_nid;
2248 source_id.pid = hdr->src_pid;
2249 /* Primary peer NID */
2250 info.mi_id.nid = msg->msg_initiator;
2251 info.mi_id.pid = hdr->src_pid;
2252 info.mi_opc = LNET_MD_OP_GET;
2253 info.mi_portal = hdr->msg.get.ptl_index;
2254 info.mi_rlength = hdr->msg.get.sink_length;
2255 info.mi_roffset = hdr->msg.get.src_offset;
2256 info.mi_mbits = hdr->msg.get.match_bits;
2257 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
2259 rc = lnet_ptl_match_md(&info, msg);
2260 if (rc == LNET_MATCHMD_DROP) {
2261 CNETERR("Dropping GET from %s portal %d match %llu"
2262 " offset %d length %d\n",
2263 libcfs_id2str(info.mi_id), info.mi_portal,
2264 info.mi_mbits, info.mi_roffset, info.mi_rlength);
2265 return -ENOENT; /* -ve: OK but no match */
2268 LASSERT(rc == LNET_MATCHMD_OK);
2270 lnet_build_msg_event(msg, LNET_EVENT_GET);
2272 reply_wmd = hdr->msg.get.return_wmd;
2274 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
2275 msg->msg_offset, msg->msg_wanted);
2277 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
2280 /* The LND completes the REPLY from her recv procedure */
2281 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2282 msg->msg_offset, msg->msg_len, msg->msg_len);
2286 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
2287 msg->msg_receiving = 0;
2289 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
2291 /* didn't get as far as lnet_ni_send() */
2292 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
2293 libcfs_nid2str(ni->ni_nid),
2294 libcfs_id2str(info.mi_id), rc);
2296 lnet_finalize(msg, rc);
2303 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
2305 void *private = msg->msg_private;
2306 struct lnet_hdr *hdr = &msg->msg_hdr;
2307 struct lnet_process_id src = {0};
2308 struct lnet_libmd *md;
2313 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
2316 src.nid = hdr->src_nid;
2317 src.pid = hdr->src_pid;
2319 /* NB handles only looked up by creator (no flips) */
2320 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
2321 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2322 CNETERR("%s: Dropping REPLY from %s for %s "
2324 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2325 (md == NULL) ? "invalid" : "inactive",
2326 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2327 hdr->msg.reply.dst_wmd.wh_object_cookie);
2328 if (md != NULL && md->md_me != NULL)
2329 CERROR("REPLY MD also attached to portal %d\n",
2330 md->md_me->me_portal);
2332 lnet_res_unlock(cpt);
2333 return -ENOENT; /* -ve: OK but no match */
2336 LASSERT(md->md_offset == 0);
2338 rlength = hdr->payload_length;
2339 mlength = MIN(rlength, (int)md->md_length);
2341 if (mlength < rlength &&
2342 (md->md_options & LNET_MD_TRUNCATE) == 0) {
2343 CNETERR("%s: Dropping REPLY from %s length %d "
2344 "for MD %#llx would overflow (%d)\n",
2345 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2346 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2348 lnet_res_unlock(cpt);
2349 return -ENOENT; /* -ve: OK but no match */
2352 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
2353 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2354 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2356 lnet_msg_attach_md(msg, md, 0, mlength);
2359 lnet_setpayloadbuffer(msg);
2361 lnet_res_unlock(cpt);
2363 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2365 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2370 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
2372 struct lnet_hdr *hdr = &msg->msg_hdr;
2373 struct lnet_process_id src = {0};
2374 struct lnet_libmd *md;
2377 src.nid = hdr->src_nid;
2378 src.pid = hdr->src_pid;
2380 /* Convert ack fields to host byte order */
2381 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2382 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2384 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
2387 /* NB handles only looked up by creator (no flips) */
2388 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2389 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2390 /* Don't moan; this is expected */
2392 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
2393 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2394 (md == NULL) ? "invalid" : "inactive",
2395 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2396 hdr->msg.ack.dst_wmd.wh_object_cookie);
2397 if (md != NULL && md->md_me != NULL)
2398 CERROR("Source MD also attached to portal %d\n",
2399 md->md_me->me_portal);
2401 lnet_res_unlock(cpt);
2402 return -ENOENT; /* -ve! */
2405 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
2406 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2407 hdr->msg.ack.dst_wmd.wh_object_cookie);
2409 lnet_msg_attach_md(msg, md, 0, 0);
2411 lnet_res_unlock(cpt);
2413 lnet_build_msg_event(msg, LNET_EVENT_ACK);
2415 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2420 * \retval LNET_CREDIT_OK If \a msg is forwarded
2421 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
2422 * \retval -ve error code
2425 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
2429 if (!the_lnet.ln_routing)
2432 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
2433 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2434 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
2435 msg->msg_rx_ready_delay = 1;
2437 lnet_net_unlock(msg->msg_rx_cpt);
2438 rc = lnet_ni_eager_recv(ni, msg);
2439 lnet_net_lock(msg->msg_rx_cpt);
2444 rc = lnet_post_routed_recv_locked(msg, 0);
2449 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
2453 switch (msg->msg_type) {
2455 rc = lnet_parse_ack(ni, msg);
2458 rc = lnet_parse_put(ni, msg);
2461 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
2463 case LNET_MSG_REPLY:
2464 rc = lnet_parse_reply(ni, msg);
2466 default: /* prevent an unused label if !kernel */
2471 LASSERT(rc == 0 || rc == -ENOENT);
2476 lnet_msgtyp2str (int type)
2485 case LNET_MSG_REPLY:
2487 case LNET_MSG_HELLO:
2490 return ("<UNKNOWN>");
2495 lnet_print_hdr(struct lnet_hdr *hdr)
2497 struct lnet_process_id src = {
2498 .nid = hdr->src_nid,
2499 .pid = hdr->src_pid,
2501 struct lnet_process_id dst = {
2502 .nid = hdr->dest_nid,
2503 .pid = hdr->dest_pid,
2505 char *type_str = lnet_msgtyp2str(hdr->type);
2507 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2508 CWARN(" From %s\n", libcfs_id2str(src));
2509 CWARN(" To %s\n", libcfs_id2str(dst));
2511 switch (hdr->type) {
2516 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
2517 "match bits %llu\n",
2518 hdr->msg.put.ptl_index,
2519 hdr->msg.put.ack_wmd.wh_interface_cookie,
2520 hdr->msg.put.ack_wmd.wh_object_cookie,
2521 hdr->msg.put.match_bits);
2522 CWARN(" Length %d, offset %d, hdr data %#llx\n",
2523 hdr->payload_length, hdr->msg.put.offset,
2524 hdr->msg.put.hdr_data);
2528 CWARN(" Ptl index %d, return md %#llx.%#llx, "
2529 "match bits %llu\n", hdr->msg.get.ptl_index,
2530 hdr->msg.get.return_wmd.wh_interface_cookie,
2531 hdr->msg.get.return_wmd.wh_object_cookie,
2532 hdr->msg.get.match_bits);
2533 CWARN(" Length %d, src offset %d\n",
2534 hdr->msg.get.sink_length,
2535 hdr->msg.get.src_offset);
2539 CWARN(" dst md %#llx.%#llx, "
2540 "manipulated length %d\n",
2541 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2542 hdr->msg.ack.dst_wmd.wh_object_cookie,
2543 hdr->msg.ack.mlength);
2546 case LNET_MSG_REPLY:
2547 CWARN(" dst md %#llx.%#llx, "
2549 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2550 hdr->msg.reply.dst_wmd.wh_object_cookie,
2551 hdr->payload_length);
2557 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
2558 void *private, int rdma_req)
2563 struct lnet_msg *msg;
2564 lnet_pid_t dest_pid;
2565 lnet_nid_t dest_nid;
2567 struct lnet_peer_ni *lpni;
2568 __u32 payload_length;
2571 LASSERT (!in_interrupt ());
2573 type = le32_to_cpu(hdr->type);
2574 src_nid = le64_to_cpu(hdr->src_nid);
2575 dest_nid = le64_to_cpu(hdr->dest_nid);
2576 dest_pid = le32_to_cpu(hdr->dest_pid);
2577 payload_length = le32_to_cpu(hdr->payload_length);
2579 for_me = (ni->ni_nid == dest_nid);
2580 cpt = lnet_cpt_of_nid(from_nid, ni);
2582 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
2583 libcfs_nid2str(dest_nid),
2584 libcfs_nid2str(ni->ni_nid),
2585 libcfs_nid2str(src_nid),
2586 lnet_msgtyp2str(type),
2587 (for_me) ? "for me" : "routed");
2592 if (payload_length > 0) {
2593 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2594 libcfs_nid2str(from_nid),
2595 libcfs_nid2str(src_nid),
2596 lnet_msgtyp2str(type), payload_length);
2602 case LNET_MSG_REPLY:
2603 if (payload_length >
2604 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2605 CERROR("%s, src %s: bad %s payload %d "
2606 "(%d max expected)\n",
2607 libcfs_nid2str(from_nid),
2608 libcfs_nid2str(src_nid),
2609 lnet_msgtyp2str(type),
2611 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2617 CERROR("%s, src %s: Bad message type 0x%x\n",
2618 libcfs_nid2str(from_nid),
2619 libcfs_nid2str(src_nid), type);
2623 if (the_lnet.ln_routing &&
2624 ni->ni_last_alive != ktime_get_real_seconds()) {
2625 /* NB: so far here is the only place to set NI status to "up */
2627 ni->ni_last_alive = ktime_get_real_seconds();
2628 if (ni->ni_status != NULL &&
2629 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
2630 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
2634 /* Regard a bad destination NID as a protocol error. Senders should
2635 * know what they're doing; if they don't they're misconfigured, buggy
2636 * or malicious so we chop them off at the knees :) */
2639 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2640 /* should have gone direct */
2641 CERROR("%s, src %s: Bad dest nid %s "
2642 "(should have been sent direct)\n",
2643 libcfs_nid2str(from_nid),
2644 libcfs_nid2str(src_nid),
2645 libcfs_nid2str(dest_nid));
2649 if (lnet_islocalnid(dest_nid)) {
2650 /* dest is another local NI; sender should have used
2651 * this node's NID on its own network */
2652 CERROR("%s, src %s: Bad dest nid %s "
2653 "(it's my nid but on a different network)\n",
2654 libcfs_nid2str(from_nid),
2655 libcfs_nid2str(src_nid),
2656 libcfs_nid2str(dest_nid));
2660 if (rdma_req && type == LNET_MSG_GET) {
2661 CERROR("%s, src %s: Bad optimized GET for %s "
2662 "(final destination must be me)\n",
2663 libcfs_nid2str(from_nid),
2664 libcfs_nid2str(src_nid),
2665 libcfs_nid2str(dest_nid));
2669 if (!the_lnet.ln_routing) {
2670 CERROR("%s, src %s: Dropping message for %s "
2671 "(routing not enabled)\n",
2672 libcfs_nid2str(from_nid),
2673 libcfs_nid2str(src_nid),
2674 libcfs_nid2str(dest_nid));
2679 /* Message looks OK; we're not going to return an error, so we MUST
2680 * call back lnd_recv() come what may... */
2682 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2683 fail_peer(src_nid, 0)) { /* shall we now? */
2684 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2685 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2686 lnet_msgtyp2str(type));
2690 if (!list_empty(&the_lnet.ln_drop_rules) &&
2691 lnet_drop_rule_match(hdr)) {
2692 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
2693 "silent message loss\n",
2694 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2695 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
2700 msg = lnet_msg_alloc();
2702 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2703 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2704 lnet_msgtyp2str(type));
2708 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
2709 * pointers NULL etc */
2711 msg->msg_type = type;
2712 msg->msg_private = private;
2713 msg->msg_receiving = 1;
2714 msg->msg_rdma_get = rdma_req;
2715 msg->msg_len = msg->msg_wanted = payload_length;
2716 msg->msg_offset = 0;
2717 msg->msg_hdr = *hdr;
2718 /* for building message event */
2719 msg->msg_from = from_nid;
2721 msg->msg_target.pid = dest_pid;
2722 msg->msg_target.nid = dest_nid;
2723 msg->msg_routing = 1;
2726 /* convert common msg->hdr fields to host byteorder */
2727 msg->msg_hdr.type = type;
2728 msg->msg_hdr.src_nid = src_nid;
2729 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2730 msg->msg_hdr.dest_nid = dest_nid;
2731 msg->msg_hdr.dest_pid = dest_pid;
2732 msg->msg_hdr.payload_length = payload_length;
2736 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
2738 lnet_net_unlock(cpt);
2739 CERROR("%s, src %s: Dropping %s "
2740 "(error %ld looking up sender)\n",
2741 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2742 lnet_msgtyp2str(type), PTR_ERR(lpni));
2744 if (rc == -ESHUTDOWN)
2745 /* We are shutting down. Don't do anything more */
2749 msg->msg_rxpeer = lpni;
2751 lnet_ni_addref_locked(ni, cpt);
2752 /* Multi-Rail: Primary NID of source. */
2753 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
2755 if (lnet_isrouter(msg->msg_rxpeer)) {
2756 lnet_peer_set_alive(msg->msg_rxpeer);
2757 if (avoid_asym_router_failure &&
2758 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
2759 /* received a remote message from router, update
2760 * remote NI status on this router.
2761 * NB: multi-hop routed message will be ignored.
2763 lnet_router_ni_update_locked(msg->msg_rxpeer,
2764 LNET_NIDNET(src_nid));
2768 lnet_msg_commit(msg, cpt);
2770 /* message delay simulation */
2771 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
2772 lnet_delay_rule_match_locked(hdr, msg))) {
2773 lnet_net_unlock(cpt);
2778 rc = lnet_parse_forward_locked(ni, msg);
2779 lnet_net_unlock(cpt);
2784 if (rc == LNET_CREDIT_OK) {
2785 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2786 0, payload_length, payload_length);
2791 lnet_net_unlock(cpt);
2793 rc = lnet_parse_local(ni, msg);
2799 LASSERT(msg->msg_md == NULL);
2800 lnet_finalize(msg, rc);
2803 lnet_drop_message(ni, cpt, private, payload_length, type);
2806 EXPORT_SYMBOL(lnet_parse);
2809 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
2811 while (!list_empty(head)) {
2812 struct lnet_process_id id = {0};
2813 struct lnet_msg *msg;
2815 msg = list_entry(head->next, struct lnet_msg, msg_list);
2816 list_del(&msg->msg_list);
2818 id.nid = msg->msg_hdr.src_nid;
2819 id.pid = msg->msg_hdr.src_pid;
2821 LASSERT(msg->msg_md == NULL);
2822 LASSERT(msg->msg_rx_delayed);
2823 LASSERT(msg->msg_rxpeer != NULL);
2824 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2826 CWARN("Dropping delayed PUT from %s portal %d match %llu"
2827 " offset %d length %d: %s\n",
2829 msg->msg_hdr.msg.put.ptl_index,
2830 msg->msg_hdr.msg.put.match_bits,
2831 msg->msg_hdr.msg.put.offset,
2832 msg->msg_hdr.payload_length, reason);
2834 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2835 * called lnet_drop_message(), so I just hang onto msg as well
2836 * until that's done */
2838 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
2839 msg->msg_private, msg->msg_len,
2842 * NB: message will not generate event because w/o attached MD,
2843 * but we still should give error code so lnet_msg_decommit()
2844 * can skip counters operations and other checks.
2846 lnet_finalize(msg, -ENOENT);
2851 lnet_recv_delayed_msg_list(struct list_head *head)
2853 while (!list_empty(head)) {
2854 struct lnet_msg *msg;
2855 struct lnet_process_id id;
2857 msg = list_entry(head->next, struct lnet_msg, msg_list);
2858 list_del(&msg->msg_list);
2860 /* md won't disappear under me, since each msg
2861 * holds a ref on it */
2863 id.nid = msg->msg_hdr.src_nid;
2864 id.pid = msg->msg_hdr.src_pid;
2866 LASSERT(msg->msg_rx_delayed);
2867 LASSERT(msg->msg_md != NULL);
2868 LASSERT(msg->msg_rxpeer != NULL);
2869 LASSERT(msg->msg_rxni != NULL);
2870 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2872 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2873 "match %llu offset %d length %d.\n",
2874 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2875 msg->msg_hdr.msg.put.match_bits,
2876 msg->msg_hdr.msg.put.offset,
2877 msg->msg_hdr.payload_length);
2879 lnet_recv_put(msg->msg_rxni, msg);
2884 * Initiate an asynchronous PUT operation.
2886 * There are several events associated with a PUT: completion of the send on
2887 * the initiator node (LNET_EVENT_SEND), and when the send completes
2888 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2889 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2890 * used at the target node to indicate the completion of incoming data
2893 * The local events will be logged in the EQ associated with the MD pointed to
2894 * by \a mdh handle. Using a MD without an associated EQ results in these
2895 * events being discarded. In this case, the caller must have another
2896 * mechanism (e.g., a higher level protocol) for determining when it is safe
2897 * to modify the memory region associated with the MD.
2899 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2900 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2902 * \param self Indicates the NID of a local interface through which to send
2903 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2904 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2905 * must be "free floating" (See LNetMDBind()).
2906 * \param ack Controls whether an acknowledgment is requested.
2907 * Acknowledgments are only sent when they are requested by the initiating
2908 * process and the target MD enables them.
2909 * \param target A process identifier for the target process.
2910 * \param portal The index in the \a target's portal table.
2911 * \param match_bits The match bits to use for MD selection at the target
2913 * \param offset The offset into the target MD (only used when the target
2914 * MD has the LNET_MD_MANAGE_REMOTE option set).
2915 * \param hdr_data 64 bits of user data that can be included in the message
2916 * header. This data is written to an event queue entry at the target if an
2917 * EQ is present on the matching MD.
2919 * \retval 0 Success, and only in this case events will be generated
2920 * and logged to EQ (if it exists).
2921 * \retval -EIO Simulated failure.
2922 * \retval -ENOMEM Memory allocation failure.
2923 * \retval -ENOENT Invalid MD object.
2925 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
2928 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
2929 struct lnet_process_id target, unsigned int portal,
2930 __u64 match_bits, unsigned int offset,
2933 struct lnet_msg *msg;
2934 struct lnet_libmd *md;
2938 LASSERT(the_lnet.ln_refcount > 0);
2940 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2941 fail_peer(target.nid, 1)) { /* shall we now? */
2942 CERROR("Dropping PUT to %s: simulated failure\n",
2943 libcfs_id2str(target));
2947 msg = lnet_msg_alloc();
2949 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
2950 libcfs_id2str(target));
2953 msg->msg_vmflush = !!memory_pressure_get();
2955 cpt = lnet_cpt_of_cookie(mdh.cookie);
2958 md = lnet_handle2md(&mdh);
2959 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2960 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
2961 match_bits, portal, libcfs_id2str(target),
2962 md == NULL ? -1 : md->md_threshold);
2963 if (md != NULL && md->md_me != NULL)
2964 CERROR("Source MD also attached to portal %d\n",
2965 md->md_me->me_portal);
2966 lnet_res_unlock(cpt);
2972 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2974 lnet_msg_attach_md(msg, md, 0, 0);
2976 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2978 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2979 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2980 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2981 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2983 /* NB handles only looked up by creator (no flips) */
2984 if (ack == LNET_ACK_REQ) {
2985 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2986 the_lnet.ln_interface_cookie;
2987 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2988 md->md_lh.lh_cookie;
2990 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2991 LNET_WIRE_HANDLE_COOKIE_NONE;
2992 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2993 LNET_WIRE_HANDLE_COOKIE_NONE;
2996 lnet_res_unlock(cpt);
2998 lnet_build_msg_event(msg, LNET_EVENT_SEND);
3000 rc = lnet_send(self, msg, LNET_NID_ANY);
3002 CNETERR("Error sending PUT to %s: %d\n",
3003 libcfs_id2str(target), rc);
3004 lnet_finalize(msg, rc);
3007 /* completion will be signalled by an event */
3010 EXPORT_SYMBOL(LNetPut);
3013 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
3014 * returns a msg for the LND to pass to lnet_finalize() when the sink
3015 * data has been received.
3017 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
3018 * lnet_finalize() is called on it, so the LND must call this first
3021 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
3023 struct lnet_msg *msg = lnet_msg_alloc();
3024 struct lnet_libmd *getmd = getmsg->msg_md;
3025 struct lnet_process_id peer_id = getmsg->msg_target;
3028 LASSERT(!getmsg->msg_target_is_router);
3029 LASSERT(!getmsg->msg_routing);
3032 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
3033 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
3037 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
3040 LASSERT(getmd->md_refcount > 0);
3042 if (getmd->md_threshold == 0) {
3043 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
3044 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
3046 lnet_res_unlock(cpt);
3050 LASSERT(getmd->md_offset == 0);
3052 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
3053 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
3055 /* setup information for lnet_build_msg_event */
3056 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
3057 msg->msg_from = peer_id.nid;
3058 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
3059 msg->msg_hdr.src_nid = peer_id.nid;
3060 msg->msg_hdr.payload_length = getmd->md_length;
3061 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
3063 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
3064 lnet_res_unlock(cpt);
3066 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
3069 lnet_msg_commit(msg, cpt);
3070 lnet_net_unlock(cpt);
3072 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3077 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
3080 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
3081 the_lnet.ln_counters[cpt]->drop_count++;
3082 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
3083 lnet_net_unlock(cpt);
3090 EXPORT_SYMBOL(lnet_create_reply_msg);
3093 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
3096 /* Set the REPLY length, now the RDMA that elides the REPLY message has
3097 * completed and I know it. */
3098 LASSERT(reply != NULL);
3099 LASSERT(reply->msg_type == LNET_MSG_GET);
3100 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
3102 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
3103 * the end of my buffer, I might as well be dead. */
3104 LASSERT(len <= reply->msg_ev.mlength);
3106 reply->msg_ev.mlength = len;
3108 EXPORT_SYMBOL(lnet_set_reply_msg_len);
3111 * Initiate an asynchronous GET operation.
3113 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
3114 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
3115 * the target node in the REPLY has been written to local MD.
3117 * On the target node, an LNET_EVENT_GET is logged when the GET request
3118 * arrives and is accepted into a MD.
3120 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
3121 * \param mdh A handle for the MD that describes the memory into which the
3122 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
3124 * \retval 0 Success, and only in this case events will be generated
3125 * and logged to EQ (if it exists) of the MD.
3126 * \retval -EIO Simulated failure.
3127 * \retval -ENOMEM Memory allocation failure.
3128 * \retval -ENOENT Invalid MD object.
3131 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
3132 struct lnet_process_id target, unsigned int portal,
3133 __u64 match_bits, unsigned int offset)
3135 struct lnet_msg *msg;
3136 struct lnet_libmd *md;
3140 LASSERT(the_lnet.ln_refcount > 0);
3142 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
3143 fail_peer(target.nid, 1)) /* shall we now? */
3145 CERROR("Dropping GET to %s: simulated failure\n",
3146 libcfs_id2str(target));
3150 msg = lnet_msg_alloc();
3152 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
3153 libcfs_id2str(target));
3157 cpt = lnet_cpt_of_cookie(mdh.cookie);
3160 md = lnet_handle2md(&mdh);
3161 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3162 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
3163 match_bits, portal, libcfs_id2str(target),
3164 md == NULL ? -1 : md->md_threshold);
3165 if (md != NULL && md->md_me != NULL)
3166 CERROR("REPLY MD also attached to portal %d\n",
3167 md->md_me->me_portal);
3169 lnet_res_unlock(cpt);
3175 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
3177 lnet_msg_attach_md(msg, md, 0, 0);
3179 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
3181 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
3182 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
3183 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
3184 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
3186 /* NB handles only looked up by creator (no flips) */
3187 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
3188 the_lnet.ln_interface_cookie;
3189 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
3190 md->md_lh.lh_cookie;
3192 lnet_res_unlock(cpt);
3194 lnet_build_msg_event(msg, LNET_EVENT_SEND);
3196 rc = lnet_send(self, msg, LNET_NID_ANY);
3198 CNETERR("Error sending GET to %s: %d\n",
3199 libcfs_id2str(target), rc);
3200 lnet_finalize(msg, rc);
3203 /* completion will be signalled by an event */
3206 EXPORT_SYMBOL(LNetGet);
3209 * Calculate distance to node at \a dstnid.
3211 * \param dstnid Target NID.
3212 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
3214 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
3217 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
3218 * local_nid_dist_zero is set, which is the default.
3219 * \retval positives Distance to target NID, i.e. number of hops plus one.
3220 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
3223 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
3225 struct list_head *e;
3226 struct lnet_ni *ni = NULL;
3227 struct lnet_remotenet *rnet;
3228 __u32 dstnet = LNET_NIDNET(dstnid);
3232 struct list_head *rn_list;
3234 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
3235 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
3236 * keep order 0 free for 0@lo and order 1 free for a local NID
3239 LASSERT(the_lnet.ln_refcount > 0);
3241 cpt = lnet_net_lock_current();
3243 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
3244 if (ni->ni_nid == dstnid) {
3245 if (srcnidp != NULL)
3247 if (orderp != NULL) {
3248 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
3253 lnet_net_unlock(cpt);
3255 return local_nid_dist_zero ? 0 : 1;
3258 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
3259 /* Check if ni was originally created in
3260 * current net namespace.
3261 * If not, assign order above 0xffff0000,
3262 * to make this ni not a priority. */
3263 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
3264 order += 0xffff0000;
3266 if (srcnidp != NULL)
3267 *srcnidp = ni->ni_nid;
3270 lnet_net_unlock(cpt);
3277 rn_list = lnet_net2rnethash(dstnet);
3278 list_for_each(e, rn_list) {
3279 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
3281 if (rnet->lrn_net == dstnet) {
3282 struct lnet_route *route;
3283 struct lnet_route *shortest = NULL;
3284 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
3287 LASSERT(!list_empty(&rnet->lrn_routes));
3289 list_for_each_entry(route, &rnet->lrn_routes,
3291 route_hops = route->lr_hops;
3292 if (route_hops == LNET_UNDEFINED_HOPS)
3294 if (shortest == NULL ||
3295 route_hops < shortest_hops) {
3297 shortest_hops = route_hops;
3301 LASSERT(shortest != NULL);
3302 hops = shortest_hops;
3303 if (srcnidp != NULL) {
3304 ni = lnet_get_next_ni_locked(
3305 shortest->lr_gateway->lpni_net,
3307 *srcnidp = ni->ni_nid;
3311 lnet_net_unlock(cpt);
3317 lnet_net_unlock(cpt);
3318 return -EHOSTUNREACH;
3320 EXPORT_SYMBOL(LNetDist);