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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/lnet/lib-move.c
38 * Data movement routines
41 #define DEBUG_SUBSYSTEM S_LNET
43 #include <lnet/lib-lnet.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");
50 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
54 struct list_head *next;
55 struct list_head cull;
57 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
59 /* Adding a new entry */
60 LIBCFS_ALLOC(tp, sizeof(*tp));
65 tp->tp_threshold = threshold;
68 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
73 /* removing entries */
74 INIT_LIST_HEAD(&cull);
78 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
79 tp = list_entry(el, lnet_test_peer_t, tp_list);
81 if (tp->tp_threshold == 0 || /* needs culling anyway */
82 nid == LNET_NID_ANY || /* removing all entries */
83 tp->tp_nid == nid) { /* matched this one */
84 list_del(&tp->tp_list);
85 list_add(&tp->tp_list, &cull);
91 while (!list_empty(&cull)) {
92 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
94 list_del(&tp->tp_list);
95 LIBCFS_FREE(tp, sizeof(*tp));
101 fail_peer (lnet_nid_t nid, int outgoing)
103 lnet_test_peer_t *tp;
104 struct list_head *el;
105 struct list_head *next;
106 struct list_head cull;
109 INIT_LIST_HEAD(&cull);
111 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
114 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
115 tp = list_entry(el, lnet_test_peer_t, tp_list);
117 if (tp->tp_threshold == 0) {
120 /* only cull zombies on outgoing tests,
121 * since we may be at interrupt priority on
122 * incoming messages. */
123 list_del(&tp->tp_list);
124 list_add(&tp->tp_list, &cull);
129 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
130 nid == tp->tp_nid) { /* fail this peer */
133 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
136 tp->tp_threshold == 0) {
138 list_del(&tp->tp_list);
139 list_add(&tp->tp_list, &cull);
148 while (!list_empty(&cull)) {
149 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
150 list_del(&tp->tp_list);
152 LIBCFS_FREE(tp, sizeof(*tp));
159 lnet_iov_nob(unsigned int niov, struct kvec *iov)
161 unsigned int nob = 0;
163 LASSERT(niov == 0 || iov != NULL);
165 nob += (iov++)->iov_len;
169 EXPORT_SYMBOL(lnet_iov_nob);
172 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
173 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
176 /* NB diov, siov are READ-ONLY */
177 unsigned int this_nob;
182 /* skip complete frags before 'doffset' */
184 while (doffset >= diov->iov_len) {
185 doffset -= diov->iov_len;
191 /* skip complete frags before 'soffset' */
193 while (soffset >= siov->iov_len) {
194 soffset -= siov->iov_len;
203 this_nob = MIN(diov->iov_len - doffset,
204 siov->iov_len - soffset);
205 this_nob = MIN(this_nob, nob);
207 memcpy ((char *)diov->iov_base + doffset,
208 (char *)siov->iov_base + soffset, this_nob);
211 if (diov->iov_len > doffset + this_nob) {
219 if (siov->iov_len > soffset + this_nob) {
228 EXPORT_SYMBOL(lnet_copy_iov2iov);
231 lnet_extract_iov(int dst_niov, struct kvec *dst,
232 int src_niov, struct kvec *src,
233 unsigned int offset, unsigned int len)
235 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
236 * for exactly 'len' bytes, and return the number of entries.
237 * NB not destructive to 'src' */
238 unsigned int frag_len;
241 if (len == 0) /* no data => */
242 return (0); /* no frags */
244 LASSERT (src_niov > 0);
245 while (offset >= src->iov_len) { /* skip initial frags */
246 offset -= src->iov_len;
249 LASSERT (src_niov > 0);
254 LASSERT (src_niov > 0);
255 LASSERT ((int)niov <= dst_niov);
257 frag_len = src->iov_len - offset;
258 dst->iov_base = ((char *)src->iov_base) + offset;
260 if (len <= frag_len) {
265 dst->iov_len = frag_len;
275 EXPORT_SYMBOL(lnet_extract_iov);
279 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
281 unsigned int nob = 0;
283 LASSERT(niov == 0 || kiov != NULL);
285 nob += (kiov++)->kiov_len;
289 EXPORT_SYMBOL(lnet_kiov_nob);
292 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
293 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
296 /* NB diov, siov are READ-ONLY */
297 unsigned int this_nob;
304 LASSERT (!in_interrupt ());
307 while (doffset >= diov->kiov_len) {
308 doffset -= diov->kiov_len;
315 while (soffset >= siov->kiov_len) {
316 soffset -= siov->kiov_len;
325 this_nob = MIN(diov->kiov_len - doffset,
326 siov->kiov_len - soffset);
327 this_nob = MIN(this_nob, nob);
330 daddr = ((char *)kmap(diov->kiov_page)) +
331 diov->kiov_offset + doffset;
333 saddr = ((char *)kmap(siov->kiov_page)) +
334 siov->kiov_offset + soffset;
336 /* Vanishing risk of kmap deadlock when mapping 2 pages.
337 * However in practice at least one of the kiovs will be mapped
338 * kernel pages and the map/unmap will be NOOPs */
340 memcpy (daddr, saddr, this_nob);
343 if (diov->kiov_len > doffset + this_nob) {
347 kunmap(diov->kiov_page);
354 if (siov->kiov_len > soffset + this_nob) {
358 kunmap(siov->kiov_page);
367 kunmap(diov->kiov_page);
369 kunmap(siov->kiov_page);
371 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
374 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
375 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
378 /* NB iov, kiov are READ-ONLY */
379 unsigned int this_nob;
385 LASSERT (!in_interrupt ());
388 while (iovoffset >= iov->iov_len) {
389 iovoffset -= iov->iov_len;
396 while (kiovoffset >= kiov->kiov_len) {
397 kiovoffset -= kiov->kiov_len;
406 this_nob = MIN(iov->iov_len - iovoffset,
407 kiov->kiov_len - kiovoffset);
408 this_nob = MIN(this_nob, nob);
411 addr = ((char *)kmap(kiov->kiov_page)) +
412 kiov->kiov_offset + kiovoffset;
414 memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
417 if (iov->iov_len > iovoffset + this_nob) {
418 iovoffset += this_nob;
425 if (kiov->kiov_len > kiovoffset + this_nob) {
427 kiovoffset += this_nob;
429 kunmap(kiov->kiov_page);
439 kunmap(kiov->kiov_page);
441 EXPORT_SYMBOL(lnet_copy_kiov2iov);
444 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
445 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
448 /* NB kiov, iov are READ-ONLY */
449 unsigned int this_nob;
455 LASSERT (!in_interrupt ());
458 while (kiovoffset >= kiov->kiov_len) {
459 kiovoffset -= kiov->kiov_len;
466 while (iovoffset >= iov->iov_len) {
467 iovoffset -= iov->iov_len;
476 this_nob = MIN(kiov->kiov_len - kiovoffset,
477 iov->iov_len - iovoffset);
478 this_nob = MIN(this_nob, nob);
481 addr = ((char *)kmap(kiov->kiov_page)) +
482 kiov->kiov_offset + kiovoffset;
484 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
487 if (kiov->kiov_len > kiovoffset + this_nob) {
489 kiovoffset += this_nob;
491 kunmap(kiov->kiov_page);
498 if (iov->iov_len > iovoffset + this_nob) {
499 iovoffset += this_nob;
508 kunmap(kiov->kiov_page);
510 EXPORT_SYMBOL(lnet_copy_iov2kiov);
513 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
514 int src_niov, lnet_kiov_t *src,
515 unsigned int offset, unsigned int len)
517 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
518 * for exactly 'len' bytes, and return the number of entries.
519 * NB not destructive to 'src' */
520 unsigned int frag_len;
523 if (len == 0) /* no data => */
524 return (0); /* no frags */
526 LASSERT (src_niov > 0);
527 while (offset >= src->kiov_len) { /* skip initial frags */
528 offset -= src->kiov_len;
531 LASSERT (src_niov > 0);
536 LASSERT (src_niov > 0);
537 LASSERT ((int)niov <= dst_niov);
539 frag_len = src->kiov_len - offset;
540 dst->kiov_page = src->kiov_page;
541 dst->kiov_offset = src->kiov_offset + offset;
543 if (len <= frag_len) {
545 LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE);
549 dst->kiov_len = frag_len;
550 LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE);
560 EXPORT_SYMBOL(lnet_extract_kiov);
563 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
564 unsigned int offset, unsigned int mlen, unsigned int rlen)
566 unsigned int niov = 0;
567 struct kvec *iov = NULL;
568 lnet_kiov_t *kiov = NULL;
571 LASSERT (!in_interrupt ());
572 LASSERT (mlen == 0 || msg != NULL);
575 LASSERT(msg->msg_receiving);
576 LASSERT(!msg->msg_sending);
577 LASSERT(rlen == msg->msg_len);
578 LASSERT(mlen <= msg->msg_len);
579 LASSERT(msg->msg_offset == offset);
580 LASSERT(msg->msg_wanted == mlen);
582 msg->msg_receiving = 0;
585 niov = msg->msg_niov;
587 kiov = msg->msg_kiov;
590 LASSERT ((iov == NULL) != (kiov == NULL));
594 rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
595 niov, iov, kiov, offset, mlen, rlen);
597 lnet_finalize(ni, msg, rc);
601 lnet_setpayloadbuffer(lnet_msg_t *msg)
603 lnet_libmd_t *md = msg->msg_md;
605 LASSERT (msg->msg_len > 0);
606 LASSERT (!msg->msg_routing);
607 LASSERT (md != NULL);
608 LASSERT (msg->msg_niov == 0);
609 LASSERT (msg->msg_iov == NULL);
610 LASSERT (msg->msg_kiov == NULL);
612 msg->msg_niov = md->md_niov;
613 if ((md->md_options & LNET_MD_KIOV) != 0)
614 msg->msg_kiov = md->md_iov.kiov;
616 msg->msg_iov = md->md_iov.iov;
620 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
621 unsigned int offset, unsigned int len)
623 msg->msg_type = type;
624 msg->msg_target = target;
626 msg->msg_offset = offset;
629 lnet_setpayloadbuffer(msg);
631 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
632 msg->msg_hdr.type = cpu_to_le32(type);
633 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
634 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
635 /* src_nid will be set later */
636 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
637 msg->msg_hdr.payload_length = cpu_to_le32(len);
641 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
643 void *priv = msg->msg_private;
646 LASSERT (!in_interrupt ());
647 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
648 (msg->msg_txcredit && msg->msg_peertxcredit));
650 rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
652 lnet_finalize(ni, msg, rc);
656 lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
660 LASSERT(!msg->msg_sending);
661 LASSERT(msg->msg_receiving);
662 LASSERT(!msg->msg_rx_ready_delay);
663 LASSERT(ni->ni_lnd->lnd_eager_recv != NULL);
665 msg->msg_rx_ready_delay = 1;
666 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
669 CERROR("recv from %s / send to %s aborted: "
670 "eager_recv failed %d\n",
671 libcfs_nid2str(msg->msg_rxpeer->lp_nid),
672 libcfs_id2str(msg->msg_target), rc);
673 LASSERT(rc < 0); /* required by my callers */
679 /* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */
681 lnet_ni_query_locked(lnet_ni_t *ni, lnet_peer_t *lp)
683 cfs_time_t last_alive = 0;
685 LASSERT(lnet_peer_aliveness_enabled(lp));
686 LASSERT(ni->ni_lnd->lnd_query != NULL);
688 lnet_net_unlock(lp->lp_cpt);
689 (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
690 lnet_net_lock(lp->lp_cpt);
692 lp->lp_last_query = cfs_time_current();
694 if (last_alive != 0) /* NI has updated timestamp */
695 lp->lp_last_alive = last_alive;
698 /* NB: always called with lnet_net_lock held */
700 lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
705 LASSERT (lnet_peer_aliveness_enabled(lp));
707 /* Trust lnet_notify() if it has more recent aliveness news, but
708 * ignore the initial assumed death (see lnet_peers_start_down()).
710 if (!lp->lp_alive && lp->lp_alive_count > 0 &&
711 cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive))
714 deadline = cfs_time_add(lp->lp_last_alive,
715 cfs_time_seconds(lp->lp_ni->ni_peertimeout));
716 alive = cfs_time_after(deadline, now);
718 /* Update obsolete lp_alive except for routers assumed to be dead
719 * initially, because router checker would update aliveness in this
720 * case, and moreover lp_last_alive at peer creation is assumed.
722 if (alive && !lp->lp_alive &&
723 !(lnet_isrouter(lp) && lp->lp_alive_count == 0))
724 lnet_notify_locked(lp, 0, 1, lp->lp_last_alive);
730 /* NB: returns 1 when alive, 0 when dead, negative when error;
731 * may drop the lnet_net_lock */
733 lnet_peer_alive_locked (lnet_peer_t *lp)
735 cfs_time_t now = cfs_time_current();
737 if (!lnet_peer_aliveness_enabled(lp))
740 if (lnet_peer_is_alive(lp, now))
743 /* Peer appears dead, but we should avoid frequent NI queries (at
744 * most once per lnet_queryinterval seconds). */
745 if (lp->lp_last_query != 0) {
746 static const int lnet_queryinterval = 1;
748 cfs_time_t next_query =
749 cfs_time_add(lp->lp_last_query,
750 cfs_time_seconds(lnet_queryinterval));
752 if (cfs_time_before(now, next_query)) {
754 CWARN("Unexpected aliveness of peer %s: "
756 libcfs_nid2str(lp->lp_nid),
757 (int)now, (int)next_query,
759 lp->lp_ni->ni_peertimeout);
764 /* query NI for latest aliveness news */
765 lnet_ni_query_locked(lp->lp_ni, lp);
767 if (lnet_peer_is_alive(lp, now))
770 lnet_notify_locked(lp, 0, 0, lp->lp_last_alive);
775 * \param msg The message to be sent.
776 * \param do_send True if lnet_ni_send() should be called in this function.
777 * lnet_send() is going to lnet_net_unlock immediately after this, so
778 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
780 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
781 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
782 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
783 * \retval -ECANCELED If the MD of the message has been unlinked.
786 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
788 lnet_peer_t *lp = msg->msg_txpeer;
789 lnet_ni_t *ni = lp->lp_ni;
790 int cpt = msg->msg_tx_cpt;
791 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
793 /* non-lnet_send() callers have checked before */
794 LASSERT(!do_send || msg->msg_tx_delayed);
795 LASSERT(!msg->msg_receiving);
796 LASSERT(msg->msg_tx_committed);
798 /* NB 'lp' is always the next hop */
799 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
800 lnet_peer_alive_locked(lp) == 0) {
801 the_lnet.ln_counters[cpt]->drop_count++;
802 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
803 lnet_net_unlock(cpt);
805 CNETERR("Dropping message for %s: peer not alive\n",
806 libcfs_id2str(msg->msg_target));
808 lnet_finalize(ni, msg, -EHOSTUNREACH);
811 return -EHOSTUNREACH;
814 if (msg->msg_md != NULL &&
815 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
816 lnet_net_unlock(cpt);
818 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
819 "called on the MD/ME.\n",
820 libcfs_id2str(msg->msg_target));
822 lnet_finalize(ni, msg, -ECANCELED);
828 if (!msg->msg_peertxcredit) {
829 LASSERT((lp->lp_txcredits < 0) ==
830 !list_empty(&lp->lp_txq));
832 msg->msg_peertxcredit = 1;
833 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
836 if (lp->lp_txcredits < lp->lp_mintxcredits)
837 lp->lp_mintxcredits = lp->lp_txcredits;
839 if (lp->lp_txcredits < 0) {
840 msg->msg_tx_delayed = 1;
841 list_add_tail(&msg->msg_list, &lp->lp_txq);
842 return LNET_CREDIT_WAIT;
846 if (!msg->msg_txcredit) {
847 LASSERT((tq->tq_credits < 0) ==
848 !list_empty(&tq->tq_delayed));
850 msg->msg_txcredit = 1;
853 if (tq->tq_credits < tq->tq_credits_min)
854 tq->tq_credits_min = tq->tq_credits;
856 if (tq->tq_credits < 0) {
857 msg->msg_tx_delayed = 1;
858 list_add_tail(&msg->msg_list, &tq->tq_delayed);
859 return LNET_CREDIT_WAIT;
864 lnet_net_unlock(cpt);
865 lnet_ni_send(ni, msg);
868 return LNET_CREDIT_OK;
872 static lnet_rtrbufpool_t *
873 lnet_msg2bufpool(lnet_msg_t *msg)
875 lnet_rtrbufpool_t *rbp;
878 LASSERT(msg->msg_rx_committed);
880 cpt = msg->msg_rx_cpt;
881 rbp = &the_lnet.ln_rtrpools[cpt][0];
883 LASSERT(msg->msg_len <= LNET_MTU);
884 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_CACHE_SIZE) {
886 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
893 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
895 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
896 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
897 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
898 * received or OK to receive */
899 lnet_peer_t *lp = msg->msg_rxpeer;
900 lnet_rtrbufpool_t *rbp;
903 LASSERT (msg->msg_iov == NULL);
904 LASSERT (msg->msg_kiov == NULL);
905 LASSERT (msg->msg_niov == 0);
906 LASSERT (msg->msg_routing);
907 LASSERT (msg->msg_receiving);
908 LASSERT (!msg->msg_sending);
910 /* non-lnet_parse callers only receive delayed messages */
911 LASSERT(!do_recv || msg->msg_rx_delayed);
913 if (!msg->msg_peerrtrcredit) {
914 LASSERT((lp->lp_rtrcredits < 0) ==
915 !list_empty(&lp->lp_rtrq));
917 msg->msg_peerrtrcredit = 1;
919 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
920 lp->lp_minrtrcredits = lp->lp_rtrcredits;
922 if (lp->lp_rtrcredits < 0) {
923 /* must have checked eager_recv before here */
924 LASSERT(msg->msg_rx_ready_delay);
925 msg->msg_rx_delayed = 1;
926 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
927 return LNET_CREDIT_WAIT;
931 rbp = lnet_msg2bufpool(msg);
933 if (!msg->msg_rtrcredit) {
934 msg->msg_rtrcredit = 1;
936 if (rbp->rbp_credits < rbp->rbp_mincredits)
937 rbp->rbp_mincredits = rbp->rbp_credits;
939 if (rbp->rbp_credits < 0) {
940 /* must have checked eager_recv before here */
941 LASSERT(msg->msg_rx_ready_delay);
942 msg->msg_rx_delayed = 1;
943 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
944 return LNET_CREDIT_WAIT;
948 LASSERT(!list_empty(&rbp->rbp_bufs));
949 rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
950 list_del(&rb->rb_list);
952 msg->msg_niov = rbp->rbp_npages;
953 msg->msg_kiov = &rb->rb_kiov[0];
956 int cpt = msg->msg_rx_cpt;
958 lnet_net_unlock(cpt);
959 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
960 0, msg->msg_len, msg->msg_len);
963 return LNET_CREDIT_OK;
967 lnet_return_tx_credits_locked(lnet_msg_t *msg)
969 lnet_peer_t *txpeer = msg->msg_txpeer;
972 if (msg->msg_txcredit) {
973 struct lnet_ni *ni = txpeer->lp_ni;
974 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
976 /* give back NI txcredits */
977 msg->msg_txcredit = 0;
979 LASSERT((tq->tq_credits < 0) ==
980 !list_empty(&tq->tq_delayed));
983 if (tq->tq_credits <= 0) {
984 msg2 = list_entry(tq->tq_delayed.next,
985 lnet_msg_t, msg_list);
986 list_del(&msg2->msg_list);
988 LASSERT(msg2->msg_txpeer->lp_ni == ni);
989 LASSERT(msg2->msg_tx_delayed);
991 (void) lnet_post_send_locked(msg2, 1);
995 if (msg->msg_peertxcredit) {
996 /* give back peer txcredits */
997 msg->msg_peertxcredit = 0;
999 LASSERT((txpeer->lp_txcredits < 0) ==
1000 !list_empty(&txpeer->lp_txq));
1002 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1003 LASSERT (txpeer->lp_txqnob >= 0);
1005 txpeer->lp_txcredits++;
1006 if (txpeer->lp_txcredits <= 0) {
1007 msg2 = list_entry(txpeer->lp_txq.next,
1008 lnet_msg_t, msg_list);
1009 list_del(&msg2->msg_list);
1011 LASSERT(msg2->msg_txpeer == txpeer);
1012 LASSERT(msg2->msg_tx_delayed);
1014 (void) lnet_post_send_locked(msg2, 1);
1018 if (txpeer != NULL) {
1019 msg->msg_txpeer = NULL;
1020 lnet_peer_decref_locked(txpeer);
1025 lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp)
1029 if (list_empty(&rbp->rbp_msgs))
1031 msg = list_entry(rbp->rbp_msgs.next,
1032 lnet_msg_t, msg_list);
1033 list_del(&msg->msg_list);
1035 (void)lnet_post_routed_recv_locked(msg, 1);
1039 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1043 struct list_head drop;
1045 INIT_LIST_HEAD(&drop);
1047 list_splice_init(list, &drop);
1049 lnet_net_unlock(cpt);
1051 list_for_each_entry_safe(msg, tmp, &drop, msg_list) {
1052 lnet_ni_recv(msg->msg_rxpeer->lp_ni, msg->msg_private, NULL,
1053 0, 0, 0, msg->msg_hdr.payload_length);
1054 list_del_init(&msg->msg_list);
1055 lnet_finalize(NULL, msg, -ECANCELED);
1062 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1064 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1067 if (msg->msg_rtrcredit) {
1068 /* give back global router credits */
1070 lnet_rtrbufpool_t *rbp;
1072 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1073 * there until it gets one allocated, or aborts the wait
1075 LASSERT(msg->msg_kiov != NULL);
1077 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1080 msg->msg_kiov = NULL;
1081 msg->msg_rtrcredit = 0;
1083 LASSERT(rbp == lnet_msg2bufpool(msg));
1085 LASSERT((rbp->rbp_credits > 0) ==
1086 !list_empty(&rbp->rbp_bufs));
1088 /* If routing is now turned off, we just drop this buffer and
1089 * don't bother trying to return credits. */
1090 if (!the_lnet.ln_routing) {
1091 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1095 /* It is possible that a user has lowered the desired number of
1096 * buffers in this pool. Make sure we never put back
1097 * more buffers than the stated number. */
1098 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1099 /* Discard this buffer so we don't have too
1101 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1102 rbp->rbp_nbuffers--;
1104 list_add(&rb->rb_list, &rbp->rbp_bufs);
1106 if (rbp->rbp_credits <= 0)
1107 lnet_schedule_blocked_locked(rbp);
1112 if (msg->msg_peerrtrcredit) {
1113 /* give back peer router credits */
1114 msg->msg_peerrtrcredit = 0;
1116 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1117 !list_empty(&rxpeer->lp_rtrq));
1119 rxpeer->lp_rtrcredits++;
1121 /* drop all messages which are queued to be routed on that
1123 if (!the_lnet.ln_routing) {
1124 lnet_drop_routed_msgs_locked(&rxpeer->lp_rtrq,
1126 } else if (rxpeer->lp_rtrcredits <= 0) {
1127 msg2 = list_entry(rxpeer->lp_rtrq.next,
1128 lnet_msg_t, msg_list);
1129 list_del(&msg2->msg_list);
1131 (void) lnet_post_routed_recv_locked(msg2, 1);
1134 if (rxpeer != NULL) {
1135 msg->msg_rxpeer = NULL;
1136 lnet_peer_decref_locked(rxpeer);
1141 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1143 lnet_peer_t *p1 = r1->lr_gateway;
1144 lnet_peer_t *p2 = r2->lr_gateway;
1145 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1146 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1148 if (r1->lr_priority < r2->lr_priority)
1151 if (r1->lr_priority > r2->lr_priority)
1154 if (r1_hops < r2_hops)
1157 if (r1_hops > r2_hops)
1160 if (p1->lp_txqnob < p2->lp_txqnob)
1163 if (p1->lp_txqnob > p2->lp_txqnob)
1166 if (p1->lp_txcredits > p2->lp_txcredits)
1169 if (p1->lp_txcredits < p2->lp_txcredits)
1172 if (r1->lr_seq - r2->lr_seq <= 0)
1178 static lnet_peer_t *
1179 lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target, lnet_nid_t rtr_nid)
1181 lnet_remotenet_t *rnet;
1182 lnet_route_t *route;
1183 lnet_route_t *best_route;
1184 lnet_route_t *last_route;
1185 struct lnet_peer *lp_best;
1186 struct lnet_peer *lp;
1189 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1190 * rtr_nid nid, otherwise find the best gateway I can use */
1192 rnet = lnet_find_net_locked(LNET_NIDNET(target));
1197 best_route = last_route = NULL;
1198 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1199 lp = route->lr_gateway;
1201 if (!lnet_is_route_alive(route))
1204 if (ni != NULL && lp->lp_ni != ni)
1207 if (lp->lp_nid == rtr_nid) /* it's pre-determined router */
1210 if (lp_best == NULL) {
1211 best_route = last_route = route;
1216 /* no protection on below fields, but it's harmless */
1217 if (last_route->lr_seq - route->lr_seq < 0)
1220 rc = lnet_compare_routes(route, best_route);
1228 /* set sequence number on the best router to the latest sequence + 1
1229 * so we can round-robin all routers, it's race and inaccurate but
1230 * harmless and functional */
1231 if (best_route != NULL)
1232 best_route->lr_seq = last_route->lr_seq + 1;
1237 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1239 lnet_nid_t dst_nid = msg->msg_target.nid;
1240 struct lnet_ni *src_ni;
1241 struct lnet_ni *local_ni;
1242 struct lnet_peer *lp;
1247 /* NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1248 * but we might want to use pre-determined router for ACK/REPLY
1250 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1251 LASSERT (msg->msg_txpeer == NULL);
1252 LASSERT (!msg->msg_sending);
1253 LASSERT (!msg->msg_target_is_router);
1254 LASSERT (!msg->msg_receiving);
1256 msg->msg_sending = 1;
1258 LASSERT(!msg->msg_tx_committed);
1259 cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid);
1263 if (the_lnet.ln_shutdown) {
1264 lnet_net_unlock(cpt);
1268 if (src_nid == LNET_NID_ANY) {
1271 src_ni = lnet_nid2ni_locked(src_nid, cpt);
1272 if (src_ni == NULL) {
1273 lnet_net_unlock(cpt);
1274 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1275 "local nid\n", libcfs_nid2str(dst_nid),
1276 libcfs_nid2str(src_nid));
1279 LASSERT (!msg->msg_routing);
1282 /* Is this for someone on a local network? */
1283 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt);
1285 if (local_ni != NULL) {
1286 if (src_ni == NULL) {
1288 src_nid = src_ni->ni_nid;
1289 } else if (src_ni == local_ni) {
1290 lnet_ni_decref_locked(local_ni, cpt);
1292 lnet_ni_decref_locked(local_ni, cpt);
1293 lnet_ni_decref_locked(src_ni, cpt);
1294 lnet_net_unlock(cpt);
1295 LCONSOLE_WARN("No route to %s via from %s\n",
1296 libcfs_nid2str(dst_nid),
1297 libcfs_nid2str(src_nid));
1301 LASSERT(src_nid != LNET_NID_ANY);
1302 lnet_msg_commit(msg, cpt);
1304 if (!msg->msg_routing)
1305 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1307 if (src_ni == the_lnet.ln_loni) {
1308 /* No send credit hassles with LOLND */
1309 lnet_net_unlock(cpt);
1310 lnet_ni_send(src_ni, msg);
1313 lnet_ni_decref_locked(src_ni, cpt);
1314 lnet_net_unlock(cpt);
1318 rc = lnet_nid2peer_locked(&lp, dst_nid, cpt);
1319 /* lp has ref on src_ni; lose mine */
1320 lnet_ni_decref_locked(src_ni, cpt);
1322 lnet_net_unlock(cpt);
1323 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1324 libcfs_nid2str(dst_nid));
1325 /* ENOMEM or shutting down */
1328 LASSERT (lp->lp_ni == src_ni);
1330 /* sending to a remote network */
1331 lp = lnet_find_route_locked(src_ni, dst_nid, rtr_nid);
1334 lnet_ni_decref_locked(src_ni, cpt);
1335 lnet_net_unlock(cpt);
1337 LCONSOLE_WARN("No route to %s via %s "
1338 "(all routers down)\n",
1339 libcfs_id2str(msg->msg_target),
1340 libcfs_nid2str(src_nid));
1341 return -EHOSTUNREACH;
1344 /* rtr_nid is LNET_NID_ANY or NID of pre-determined router,
1345 * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't
1346 * pre-determined router, this can happen if router table
1347 * was changed when we release the lock */
1348 if (rtr_nid != lp->lp_nid) {
1349 cpt2 = lnet_cpt_of_nid_locked(lp->lp_nid);
1352 lnet_ni_decref_locked(src_ni, cpt);
1353 lnet_net_unlock(cpt);
1355 rtr_nid = lp->lp_nid;
1361 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1362 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1363 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1365 if (src_ni == NULL) {
1367 src_nid = src_ni->ni_nid;
1369 LASSERT (src_ni == lp->lp_ni);
1370 lnet_ni_decref_locked(src_ni, cpt);
1373 lnet_peer_addref_locked(lp);
1375 LASSERT(src_nid != LNET_NID_ANY);
1376 lnet_msg_commit(msg, cpt);
1378 if (!msg->msg_routing) {
1379 /* I'm the source and now I know which NI to send on */
1380 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1383 msg->msg_target_is_router = 1;
1384 msg->msg_target.nid = lp->lp_nid;
1385 msg->msg_target.pid = LNET_PID_LUSTRE;
1388 /* 'lp' is our best choice of peer */
1390 LASSERT (!msg->msg_peertxcredit);
1391 LASSERT (!msg->msg_txcredit);
1392 LASSERT (msg->msg_txpeer == NULL);
1394 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1396 rc = lnet_post_send_locked(msg, 0);
1397 lnet_net_unlock(cpt);
1402 if (rc == LNET_CREDIT_OK)
1403 lnet_ni_send(src_ni, msg);
1405 return 0; /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
1409 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1412 the_lnet.ln_counters[cpt]->drop_count++;
1413 the_lnet.ln_counters[cpt]->drop_length += nob;
1414 lnet_net_unlock(cpt);
1416 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1420 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1422 lnet_hdr_t *hdr = &msg->msg_hdr;
1424 if (msg->msg_wanted != 0)
1425 lnet_setpayloadbuffer(msg);
1427 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1429 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1430 * it back into the ACK during lnet_finalize() */
1431 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1432 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1434 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1435 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1439 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1441 lnet_hdr_t *hdr = &msg->msg_hdr;
1442 struct lnet_match_info info;
1446 /* Convert put fields to host byte order */
1447 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1448 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1449 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1451 info.mi_id.nid = hdr->src_nid;
1452 info.mi_id.pid = hdr->src_pid;
1453 info.mi_opc = LNET_MD_OP_PUT;
1454 info.mi_portal = hdr->msg.put.ptl_index;
1455 info.mi_rlength = hdr->payload_length;
1456 info.mi_roffset = hdr->msg.put.offset;
1457 info.mi_mbits = hdr->msg.put.match_bits;
1459 msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL;
1460 ready_delay = msg->msg_rx_ready_delay;
1463 rc = lnet_ptl_match_md(&info, msg);
1468 case LNET_MATCHMD_OK:
1469 lnet_recv_put(ni, msg);
1472 case LNET_MATCHMD_NONE:
1474 /* no eager_recv or has already called it, should
1475 * have been attached on delayed list */
1478 rc = lnet_ni_eager_recv(ni, msg);
1485 case LNET_MATCHMD_DROP:
1486 CNETERR("Dropping PUT from %s portal %d match "LPU64
1487 " offset %d length %d: %d\n",
1488 libcfs_id2str(info.mi_id), info.mi_portal,
1489 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1491 return -ENOENT; /* -ve: OK but no match */
1496 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1498 struct lnet_match_info info;
1499 lnet_hdr_t *hdr = &msg->msg_hdr;
1500 lnet_handle_wire_t reply_wmd;
1503 /* Convert get fields to host byte order */
1504 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1505 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1506 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1507 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1509 info.mi_id.nid = hdr->src_nid;
1510 info.mi_id.pid = hdr->src_pid;
1511 info.mi_opc = LNET_MD_OP_GET;
1512 info.mi_portal = hdr->msg.get.ptl_index;
1513 info.mi_rlength = hdr->msg.get.sink_length;
1514 info.mi_roffset = hdr->msg.get.src_offset;
1515 info.mi_mbits = hdr->msg.get.match_bits;
1517 rc = lnet_ptl_match_md(&info, msg);
1518 if (rc == LNET_MATCHMD_DROP) {
1519 CNETERR("Dropping GET from %s portal %d match "LPU64
1520 " offset %d length %d\n",
1521 libcfs_id2str(info.mi_id), info.mi_portal,
1522 info.mi_mbits, info.mi_roffset, info.mi_rlength);
1523 return -ENOENT; /* -ve: OK but no match */
1526 LASSERT(rc == LNET_MATCHMD_OK);
1528 lnet_build_msg_event(msg, LNET_EVENT_GET);
1530 reply_wmd = hdr->msg.get.return_wmd;
1532 lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
1533 msg->msg_offset, msg->msg_wanted);
1535 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1538 /* The LND completes the REPLY from her recv procedure */
1539 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1540 msg->msg_offset, msg->msg_len, msg->msg_len);
1544 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1545 msg->msg_receiving = 0;
1547 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
1549 /* didn't get as far as lnet_ni_send() */
1550 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1551 libcfs_nid2str(ni->ni_nid),
1552 libcfs_id2str(info.mi_id), rc);
1554 lnet_finalize(ni, msg, rc);
1561 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1563 void *private = msg->msg_private;
1564 lnet_hdr_t *hdr = &msg->msg_hdr;
1565 lnet_process_id_t src = {0};
1571 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
1574 src.nid = hdr->src_nid;
1575 src.pid = hdr->src_pid;
1577 /* NB handles only looked up by creator (no flips) */
1578 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1579 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1580 CNETERR("%s: Dropping REPLY from %s for %s "
1581 "MD "LPX64"."LPX64"\n",
1582 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1583 (md == NULL) ? "invalid" : "inactive",
1584 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1585 hdr->msg.reply.dst_wmd.wh_object_cookie);
1586 if (md != NULL && md->md_me != NULL)
1587 CERROR("REPLY MD also attached to portal %d\n",
1588 md->md_me->me_portal);
1590 lnet_res_unlock(cpt);
1591 return -ENOENT; /* -ve: OK but no match */
1594 LASSERT (md->md_offset == 0);
1596 rlength = hdr->payload_length;
1597 mlength = MIN(rlength, (int)md->md_length);
1599 if (mlength < rlength &&
1600 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1601 CNETERR("%s: Dropping REPLY from %s length %d "
1602 "for MD "LPX64" would overflow (%d)\n",
1603 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1604 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1606 lnet_res_unlock(cpt);
1607 return -ENOENT; /* -ve: OK but no match */
1610 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1611 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1612 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1614 lnet_msg_attach_md(msg, md, 0, mlength);
1617 lnet_setpayloadbuffer(msg);
1619 lnet_res_unlock(cpt);
1621 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
1623 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1628 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1630 lnet_hdr_t *hdr = &msg->msg_hdr;
1631 lnet_process_id_t src = {0};
1635 src.nid = hdr->src_nid;
1636 src.pid = hdr->src_pid;
1638 /* Convert ack fields to host byte order */
1639 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1640 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1642 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
1645 /* NB handles only looked up by creator (no flips) */
1646 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1647 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1648 /* Don't moan; this is expected */
1650 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1651 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1652 (md == NULL) ? "invalid" : "inactive",
1653 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1654 hdr->msg.ack.dst_wmd.wh_object_cookie);
1655 if (md != NULL && md->md_me != NULL)
1656 CERROR("Source MD also attached to portal %d\n",
1657 md->md_me->me_portal);
1659 lnet_res_unlock(cpt);
1660 return -ENOENT; /* -ve! */
1663 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1664 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1665 hdr->msg.ack.dst_wmd.wh_object_cookie);
1667 lnet_msg_attach_md(msg, md, 0, 0);
1669 lnet_res_unlock(cpt);
1671 lnet_build_msg_event(msg, LNET_EVENT_ACK);
1673 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
1678 * \retval LNET_CREDIT_OK If \a msg is forwarded
1679 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
1680 * \retval -ve error code
1683 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
1687 if (!the_lnet.ln_routing)
1690 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
1691 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
1692 if (ni->ni_lnd->lnd_eager_recv == NULL) {
1693 msg->msg_rx_ready_delay = 1;
1695 lnet_net_unlock(msg->msg_rx_cpt);
1696 rc = lnet_ni_eager_recv(ni, msg);
1697 lnet_net_lock(msg->msg_rx_cpt);
1702 rc = lnet_post_routed_recv_locked(msg, 0);
1707 lnet_parse_local(lnet_ni_t *ni, lnet_msg_t *msg)
1711 switch (msg->msg_type) {
1713 rc = lnet_parse_ack(ni, msg);
1716 rc = lnet_parse_put(ni, msg);
1719 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
1721 case LNET_MSG_REPLY:
1722 rc = lnet_parse_reply(ni, msg);
1724 default: /* prevent an unused label if !kernel */
1729 LASSERT(rc == 0 || rc == -ENOENT);
1734 lnet_msgtyp2str (int type)
1743 case LNET_MSG_REPLY:
1745 case LNET_MSG_HELLO:
1748 return ("<UNKNOWN>");
1753 lnet_print_hdr(lnet_hdr_t * hdr)
1755 lnet_process_id_t src = {0};
1756 lnet_process_id_t dst = {0};
1757 char *type_str = lnet_msgtyp2str (hdr->type);
1759 src.nid = hdr->src_nid;
1760 src.pid = hdr->src_pid;
1762 dst.nid = hdr->dest_nid;
1763 dst.pid = hdr->dest_pid;
1765 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
1766 CWARN(" From %s\n", libcfs_id2str(src));
1767 CWARN(" To %s\n", libcfs_id2str(dst));
1769 switch (hdr->type) {
1774 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
1775 "match bits "LPU64"\n",
1776 hdr->msg.put.ptl_index,
1777 hdr->msg.put.ack_wmd.wh_interface_cookie,
1778 hdr->msg.put.ack_wmd.wh_object_cookie,
1779 hdr->msg.put.match_bits);
1780 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
1781 hdr->payload_length, hdr->msg.put.offset,
1782 hdr->msg.put.hdr_data);
1786 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
1787 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
1788 hdr->msg.get.return_wmd.wh_interface_cookie,
1789 hdr->msg.get.return_wmd.wh_object_cookie,
1790 hdr->msg.get.match_bits);
1791 CWARN(" Length %d, src offset %d\n",
1792 hdr->msg.get.sink_length,
1793 hdr->msg.get.src_offset);
1797 CWARN(" dst md "LPX64"."LPX64", "
1798 "manipulated length %d\n",
1799 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1800 hdr->msg.ack.dst_wmd.wh_object_cookie,
1801 hdr->msg.ack.mlength);
1804 case LNET_MSG_REPLY:
1805 CWARN(" dst md "LPX64"."LPX64", "
1807 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1808 hdr->msg.reply.dst_wmd.wh_object_cookie,
1809 hdr->payload_length);
1815 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
1816 void *private, int rdma_req)
1821 struct lnet_msg *msg;
1822 lnet_pid_t dest_pid;
1823 lnet_nid_t dest_nid;
1825 __u32 payload_length;
1828 LASSERT (!in_interrupt ());
1830 type = le32_to_cpu(hdr->type);
1831 src_nid = le64_to_cpu(hdr->src_nid);
1832 dest_nid = le64_to_cpu(hdr->dest_nid);
1833 dest_pid = le32_to_cpu(hdr->dest_pid);
1834 payload_length = le32_to_cpu(hdr->payload_length);
1836 for_me = (ni->ni_nid == dest_nid);
1837 cpt = lnet_cpt_of_nid(from_nid);
1842 if (payload_length > 0) {
1843 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
1844 libcfs_nid2str(from_nid),
1845 libcfs_nid2str(src_nid),
1846 lnet_msgtyp2str(type), payload_length);
1852 case LNET_MSG_REPLY:
1853 if (payload_length >
1854 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
1855 CERROR("%s, src %s: bad %s payload %d "
1856 "(%d max expected)\n",
1857 libcfs_nid2str(from_nid),
1858 libcfs_nid2str(src_nid),
1859 lnet_msgtyp2str(type),
1861 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
1867 CERROR("%s, src %s: Bad message type 0x%x\n",
1868 libcfs_nid2str(from_nid),
1869 libcfs_nid2str(src_nid), type);
1873 if (the_lnet.ln_routing &&
1874 ni->ni_last_alive != cfs_time_current_sec()) {
1875 /* NB: so far here is the only place to set NI status to "up */
1877 ni->ni_last_alive = cfs_time_current_sec();
1878 if (ni->ni_status != NULL &&
1879 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
1880 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
1884 /* Regard a bad destination NID as a protocol error. Senders should
1885 * know what they're doing; if they don't they're misconfigured, buggy
1886 * or malicious so we chop them off at the knees :) */
1889 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
1890 /* should have gone direct */
1891 CERROR("%s, src %s: Bad dest nid %s "
1892 "(should have been sent direct)\n",
1893 libcfs_nid2str(from_nid),
1894 libcfs_nid2str(src_nid),
1895 libcfs_nid2str(dest_nid));
1899 if (lnet_islocalnid(dest_nid)) {
1900 /* dest is another local NI; sender should have used
1901 * this node's NID on its own network */
1902 CERROR("%s, src %s: Bad dest nid %s "
1903 "(it's my nid but on a different network)\n",
1904 libcfs_nid2str(from_nid),
1905 libcfs_nid2str(src_nid),
1906 libcfs_nid2str(dest_nid));
1910 if (rdma_req && type == LNET_MSG_GET) {
1911 CERROR("%s, src %s: Bad optimized GET for %s "
1912 "(final destination must be me)\n",
1913 libcfs_nid2str(from_nid),
1914 libcfs_nid2str(src_nid),
1915 libcfs_nid2str(dest_nid));
1919 if (!the_lnet.ln_routing) {
1920 CERROR("%s, src %s: Dropping message for %s "
1921 "(routing not enabled)\n",
1922 libcfs_nid2str(from_nid),
1923 libcfs_nid2str(src_nid),
1924 libcfs_nid2str(dest_nid));
1929 /* Message looks OK; we're not going to return an error, so we MUST
1930 * call back lnd_recv() come what may... */
1932 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
1933 fail_peer(src_nid, 0)) { /* shall we now? */
1934 CERROR("%s, src %s: Dropping %s to simulate failure\n",
1935 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1936 lnet_msgtyp2str(type));
1940 if (!list_empty(&the_lnet.ln_drop_rules) &&
1941 lnet_drop_rule_match(hdr)) {
1942 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
1943 "silent message loss\n",
1944 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1945 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
1950 msg = lnet_msg_alloc();
1952 CERROR("%s, src %s: Dropping %s (out of memory)\n",
1953 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1954 lnet_msgtyp2str(type));
1958 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
1959 * pointers NULL etc */
1961 msg->msg_type = type;
1962 msg->msg_private = private;
1963 msg->msg_receiving = 1;
1964 msg->msg_rdma_get = rdma_req;
1965 msg->msg_len = msg->msg_wanted = payload_length;
1966 msg->msg_offset = 0;
1967 msg->msg_hdr = *hdr;
1968 /* for building message event */
1969 msg->msg_from = from_nid;
1971 msg->msg_target.pid = dest_pid;
1972 msg->msg_target.nid = dest_nid;
1973 msg->msg_routing = 1;
1976 /* convert common msg->hdr fields to host byteorder */
1977 msg->msg_hdr.type = type;
1978 msg->msg_hdr.src_nid = src_nid;
1979 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
1980 msg->msg_hdr.dest_nid = dest_nid;
1981 msg->msg_hdr.dest_pid = dest_pid;
1982 msg->msg_hdr.payload_length = payload_length;
1986 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid, cpt);
1988 lnet_net_unlock(cpt);
1989 CERROR("%s, src %s: Dropping %s "
1990 "(error %d looking up sender)\n",
1991 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1992 lnet_msgtyp2str(type), rc);
1997 if (lnet_isrouter(msg->msg_rxpeer)) {
1998 lnet_peer_set_alive(msg->msg_rxpeer);
1999 if (avoid_asym_router_failure &&
2000 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
2001 /* received a remote message from router, update
2002 * remote NI status on this router.
2003 * NB: multi-hop routed message will be ignored.
2005 lnet_router_ni_update_locked(msg->msg_rxpeer,
2006 LNET_NIDNET(src_nid));
2010 lnet_msg_commit(msg, cpt);
2012 /* message delay simulation */
2013 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
2014 lnet_delay_rule_match_locked(hdr, msg))) {
2015 lnet_net_unlock(cpt);
2020 rc = lnet_parse_forward_locked(ni, msg);
2021 lnet_net_unlock(cpt);
2026 if (rc == LNET_CREDIT_OK) {
2027 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2028 0, payload_length, payload_length);
2033 lnet_net_unlock(cpt);
2035 rc = lnet_parse_local(ni, msg);
2041 LASSERT(msg->msg_md == NULL);
2042 lnet_finalize(ni, msg, rc);
2045 lnet_drop_message(ni, cpt, private, payload_length);
2048 EXPORT_SYMBOL(lnet_parse);
2051 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
2053 while (!list_empty(head)) {
2054 lnet_process_id_t id = {0};
2057 msg = list_entry(head->next, lnet_msg_t, msg_list);
2058 list_del(&msg->msg_list);
2060 id.nid = msg->msg_hdr.src_nid;
2061 id.pid = msg->msg_hdr.src_pid;
2063 LASSERT(msg->msg_md == NULL);
2064 LASSERT(msg->msg_rx_delayed);
2065 LASSERT(msg->msg_rxpeer != NULL);
2066 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2068 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
2069 " offset %d length %d: %s\n",
2071 msg->msg_hdr.msg.put.ptl_index,
2072 msg->msg_hdr.msg.put.match_bits,
2073 msg->msg_hdr.msg.put.offset,
2074 msg->msg_hdr.payload_length, reason);
2076 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2077 * called lnet_drop_message(), so I just hang onto msg as well
2078 * until that's done */
2080 lnet_drop_message(msg->msg_rxpeer->lp_ni,
2081 msg->msg_rxpeer->lp_cpt,
2082 msg->msg_private, msg->msg_len);
2084 * NB: message will not generate event because w/o attached MD,
2085 * but we still should give error code so lnet_msg_decommit()
2086 * can skip counters operations and other checks.
2088 lnet_finalize(msg->msg_rxpeer->lp_ni, msg, -ENOENT);
2093 lnet_recv_delayed_msg_list(struct list_head *head)
2095 while (!list_empty(head)) {
2097 lnet_process_id_t id;
2099 msg = list_entry(head->next, lnet_msg_t, msg_list);
2100 list_del(&msg->msg_list);
2102 /* md won't disappear under me, since each msg
2103 * holds a ref on it */
2105 id.nid = msg->msg_hdr.src_nid;
2106 id.pid = msg->msg_hdr.src_pid;
2108 LASSERT(msg->msg_rx_delayed);
2109 LASSERT(msg->msg_md != NULL);
2110 LASSERT(msg->msg_rxpeer != NULL);
2111 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2113 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2114 "match "LPU64" offset %d length %d.\n",
2115 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2116 msg->msg_hdr.msg.put.match_bits,
2117 msg->msg_hdr.msg.put.offset,
2118 msg->msg_hdr.payload_length);
2120 lnet_recv_put(msg->msg_rxpeer->lp_ni, msg);
2125 * Initiate an asynchronous PUT operation.
2127 * There are several events associated with a PUT: completion of the send on
2128 * the initiator node (LNET_EVENT_SEND), and when the send completes
2129 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2130 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2131 * used at the target node to indicate the completion of incoming data
2134 * The local events will be logged in the EQ associated with the MD pointed to
2135 * by \a mdh handle. Using a MD without an associated EQ results in these
2136 * events being discarded. In this case, the caller must have another
2137 * mechanism (e.g., a higher level protocol) for determining when it is safe
2138 * to modify the memory region associated with the MD.
2140 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2141 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2143 * \param self Indicates the NID of a local interface through which to send
2144 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2145 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2146 * must be "free floating" (See LNetMDBind()).
2147 * \param ack Controls whether an acknowledgment is requested.
2148 * Acknowledgments are only sent when they are requested by the initiating
2149 * process and the target MD enables them.
2150 * \param target A process identifier for the target process.
2151 * \param portal The index in the \a target's portal table.
2152 * \param match_bits The match bits to use for MD selection at the target
2154 * \param offset The offset into the target MD (only used when the target
2155 * MD has the LNET_MD_MANAGE_REMOTE option set).
2156 * \param hdr_data 64 bits of user data that can be included in the message
2157 * header. This data is written to an event queue entry at the target if an
2158 * EQ is present on the matching MD.
2160 * \retval 0 Success, and only in this case events will be generated
2161 * and logged to EQ (if it exists).
2162 * \retval -EIO Simulated failure.
2163 * \retval -ENOMEM Memory allocation failure.
2164 * \retval -ENOENT Invalid MD object.
2166 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2169 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2170 lnet_process_id_t target, unsigned int portal,
2171 __u64 match_bits, unsigned int offset,
2174 struct lnet_msg *msg;
2175 struct lnet_libmd *md;
2179 LASSERT(the_lnet.ln_refcount > 0);
2181 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2182 fail_peer(target.nid, 1)) { /* shall we now? */
2183 CERROR("Dropping PUT to %s: simulated failure\n",
2184 libcfs_id2str(target));
2188 msg = lnet_msg_alloc();
2190 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2191 libcfs_id2str(target));
2194 msg->msg_vmflush = !!memory_pressure_get();
2196 cpt = lnet_cpt_of_cookie(mdh.cookie);
2199 md = lnet_handle2md(&mdh);
2200 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2201 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2202 match_bits, portal, libcfs_id2str(target),
2203 md == NULL ? -1 : md->md_threshold);
2204 if (md != NULL && md->md_me != NULL)
2205 CERROR("Source MD also attached to portal %d\n",
2206 md->md_me->me_portal);
2207 lnet_res_unlock(cpt);
2213 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2215 lnet_msg_attach_md(msg, md, 0, 0);
2217 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2219 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2220 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2221 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2222 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2224 /* NB handles only looked up by creator (no flips) */
2225 if (ack == LNET_ACK_REQ) {
2226 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2227 the_lnet.ln_interface_cookie;
2228 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2229 md->md_lh.lh_cookie;
2231 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2232 LNET_WIRE_HANDLE_COOKIE_NONE;
2233 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2234 LNET_WIRE_HANDLE_COOKIE_NONE;
2237 lnet_res_unlock(cpt);
2239 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2241 rc = lnet_send(self, msg, LNET_NID_ANY);
2243 CNETERR( "Error sending PUT to %s: %d\n",
2244 libcfs_id2str(target), rc);
2245 lnet_finalize (NULL, msg, rc);
2248 /* completion will be signalled by an event */
2251 EXPORT_SYMBOL(LNetPut);
2254 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2256 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2257 * returns a msg for the LND to pass to lnet_finalize() when the sink
2258 * data has been received.
2260 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2261 * lnet_finalize() is called on it, so the LND must call this first */
2263 struct lnet_msg *msg = lnet_msg_alloc();
2264 struct lnet_libmd *getmd = getmsg->msg_md;
2265 lnet_process_id_t peer_id = getmsg->msg_target;
2268 LASSERT(!getmsg->msg_target_is_router);
2269 LASSERT(!getmsg->msg_routing);
2272 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2273 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2277 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2280 LASSERT(getmd->md_refcount > 0);
2282 if (getmd->md_threshold == 0) {
2283 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2284 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2286 lnet_res_unlock(cpt);
2290 LASSERT(getmd->md_offset == 0);
2292 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2293 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2295 /* setup information for lnet_build_msg_event */
2296 msg->msg_from = peer_id.nid;
2297 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2298 msg->msg_hdr.src_nid = peer_id.nid;
2299 msg->msg_hdr.payload_length = getmd->md_length;
2300 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2302 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2303 lnet_res_unlock(cpt);
2305 cpt = lnet_cpt_of_nid(peer_id.nid);
2308 lnet_msg_commit(msg, cpt);
2309 lnet_net_unlock(cpt);
2311 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2316 cpt = lnet_cpt_of_nid(peer_id.nid);
2319 the_lnet.ln_counters[cpt]->drop_count++;
2320 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2321 lnet_net_unlock(cpt);
2328 EXPORT_SYMBOL(lnet_create_reply_msg);
2331 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2333 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2334 * completed and I know it. */
2335 LASSERT (reply != NULL);
2336 LASSERT (reply->msg_type == LNET_MSG_GET);
2337 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2339 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2340 * the end of my buffer, I might as well be dead. */
2341 LASSERT (len <= reply->msg_ev.mlength);
2343 reply->msg_ev.mlength = len;
2345 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2348 * Initiate an asynchronous GET operation.
2350 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2351 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2352 * the target node in the REPLY has been written to local MD.
2354 * On the target node, an LNET_EVENT_GET is logged when the GET request
2355 * arrives and is accepted into a MD.
2357 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2358 * \param mdh A handle for the MD that describes the memory into which the
2359 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2361 * \retval 0 Success, and only in this case events will be generated
2362 * and logged to EQ (if it exists) of the MD.
2363 * \retval -EIO Simulated failure.
2364 * \retval -ENOMEM Memory allocation failure.
2365 * \retval -ENOENT Invalid MD object.
2368 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2369 lnet_process_id_t target, unsigned int portal,
2370 __u64 match_bits, unsigned int offset)
2372 struct lnet_msg *msg;
2373 struct lnet_libmd *md;
2377 LASSERT (the_lnet.ln_refcount > 0);
2379 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2380 fail_peer(target.nid, 1)) /* shall we now? */
2382 CERROR("Dropping GET to %s: simulated failure\n",
2383 libcfs_id2str(target));
2387 msg = lnet_msg_alloc();
2389 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2390 libcfs_id2str(target));
2394 cpt = lnet_cpt_of_cookie(mdh.cookie);
2397 md = lnet_handle2md(&mdh);
2398 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2399 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2400 match_bits, portal, libcfs_id2str(target),
2401 md == NULL ? -1 : md->md_threshold);
2402 if (md != NULL && md->md_me != NULL)
2403 CERROR("REPLY MD also attached to portal %d\n",
2404 md->md_me->me_portal);
2406 lnet_res_unlock(cpt);
2412 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2414 lnet_msg_attach_md(msg, md, 0, 0);
2416 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2418 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2419 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2420 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2421 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2423 /* NB handles only looked up by creator (no flips) */
2424 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2425 the_lnet.ln_interface_cookie;
2426 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2427 md->md_lh.lh_cookie;
2429 lnet_res_unlock(cpt);
2431 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2433 rc = lnet_send(self, msg, LNET_NID_ANY);
2435 CNETERR("Error sending GET to %s: %d\n",
2436 libcfs_id2str(target), rc);
2437 lnet_finalize(NULL, msg, rc);
2440 /* completion will be signalled by an event */
2443 EXPORT_SYMBOL(LNetGet);
2446 * Calculate distance to node at \a dstnid.
2448 * \param dstnid Target NID.
2449 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2451 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2454 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2455 * local_nid_dist_zero is set, which is the default.
2456 * \retval positives Distance to target NID, i.e. number of hops plus one.
2457 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2460 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2462 struct list_head *e;
2464 lnet_remotenet_t *rnet;
2465 __u32 dstnet = LNET_NIDNET(dstnid);
2469 struct list_head *rn_list;
2471 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2472 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2473 * keep order 0 free for 0@lo and order 1 free for a local NID
2476 LASSERT (the_lnet.ln_refcount > 0);
2478 cpt = lnet_net_lock_current();
2480 list_for_each(e, &the_lnet.ln_nis) {
2481 ni = list_entry(e, lnet_ni_t, ni_list);
2483 if (ni->ni_nid == dstnid) {
2484 if (srcnidp != NULL)
2486 if (orderp != NULL) {
2487 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2492 lnet_net_unlock(cpt);
2494 return local_nid_dist_zero ? 0 : 1;
2497 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2498 if (srcnidp != NULL)
2499 *srcnidp = ni->ni_nid;
2502 lnet_net_unlock(cpt);
2509 rn_list = lnet_net2rnethash(dstnet);
2510 list_for_each(e, rn_list) {
2511 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2513 if (rnet->lrn_net == dstnet) {
2514 lnet_route_t *route;
2515 lnet_route_t *shortest = NULL;
2516 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
2519 LASSERT(!list_empty(&rnet->lrn_routes));
2521 list_for_each_entry(route, &rnet->lrn_routes,
2523 route_hops = route->lr_hops;
2524 if (route_hops == LNET_UNDEFINED_HOPS)
2526 if (shortest == NULL ||
2527 route_hops < shortest_hops) {
2529 shortest_hops = route_hops;
2533 LASSERT(shortest != NULL);
2534 hops = shortest_hops;
2535 if (srcnidp != NULL)
2536 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2539 lnet_net_unlock(cpt);
2545 lnet_net_unlock(cpt);
2546 return -EHOSTUNREACH;
2548 EXPORT_SYMBOL(LNetDist);