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, 2013, 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 CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444,
50 lnet_fail_nid (lnet_nid_t nid, unsigned int threshold)
57 LASSERT (the_lnet.ln_init);
59 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
61 /* Adding a new entry */
62 LIBCFS_ALLOC(tp, sizeof(*tp));
67 tp->tp_threshold = threshold;
70 cfs_list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
75 /* removing entries */
76 CFS_INIT_LIST_HEAD(&cull);
80 cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
81 tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
83 if (tp->tp_threshold == 0 || /* needs culling anyway */
84 nid == LNET_NID_ANY || /* removing all entries */
85 tp->tp_nid == nid) /* matched this one */
87 cfs_list_del (&tp->tp_list);
88 cfs_list_add (&tp->tp_list, &cull);
94 while (!cfs_list_empty (&cull)) {
95 tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
97 cfs_list_del (&tp->tp_list);
98 LIBCFS_FREE(tp, sizeof (*tp));
104 fail_peer (lnet_nid_t nid, int outgoing)
106 lnet_test_peer_t *tp;
112 CFS_INIT_LIST_HEAD (&cull);
114 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
117 cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
118 tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
120 if (tp->tp_threshold == 0) {
123 /* only cull zombies on outgoing tests,
124 * since we may be at interrupt priority on
125 * incoming messages. */
126 cfs_list_del (&tp->tp_list);
127 cfs_list_add (&tp->tp_list, &cull);
132 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
133 nid == tp->tp_nid) { /* fail this peer */
136 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
139 tp->tp_threshold == 0) {
141 cfs_list_del (&tp->tp_list);
142 cfs_list_add (&tp->tp_list, &cull);
151 while (!cfs_list_empty (&cull)) {
152 tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
153 cfs_list_del (&tp->tp_list);
155 LIBCFS_FREE(tp, sizeof (*tp));
162 lnet_iov_nob (unsigned int niov, struct iovec *iov)
164 unsigned int nob = 0;
166 LASSERT(niov == 0 || iov != NULL);
168 nob += (iov++)->iov_len;
172 EXPORT_SYMBOL(lnet_iov_nob);
175 lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
176 unsigned int nsiov, struct iovec *siov, unsigned int soffset,
179 /* NB diov, siov are READ-ONLY */
180 unsigned int this_nob;
185 /* skip complete frags before 'doffset' */
187 while (doffset >= diov->iov_len) {
188 doffset -= diov->iov_len;
194 /* skip complete frags before 'soffset' */
196 while (soffset >= siov->iov_len) {
197 soffset -= siov->iov_len;
206 this_nob = MIN(diov->iov_len - doffset,
207 siov->iov_len - soffset);
208 this_nob = MIN(this_nob, nob);
210 memcpy ((char *)diov->iov_base + doffset,
211 (char *)siov->iov_base + soffset, this_nob);
214 if (diov->iov_len > doffset + this_nob) {
222 if (siov->iov_len > soffset + this_nob) {
231 EXPORT_SYMBOL(lnet_copy_iov2iov);
234 lnet_extract_iov (int dst_niov, struct iovec *dst,
235 int src_niov, struct iovec *src,
236 unsigned int offset, unsigned int len)
238 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
239 * for exactly 'len' bytes, and return the number of entries.
240 * NB not destructive to 'src' */
241 unsigned int frag_len;
244 if (len == 0) /* no data => */
245 return (0); /* no frags */
247 LASSERT (src_niov > 0);
248 while (offset >= src->iov_len) { /* skip initial frags */
249 offset -= src->iov_len;
252 LASSERT (src_niov > 0);
257 LASSERT (src_niov > 0);
258 LASSERT ((int)niov <= dst_niov);
260 frag_len = src->iov_len - offset;
261 dst->iov_base = ((char *)src->iov_base) + offset;
263 if (len <= frag_len) {
268 dst->iov_len = frag_len;
278 EXPORT_SYMBOL(lnet_extract_iov);
282 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
289 lnet_copy_kiov2kiov (unsigned int ndkiov, lnet_kiov_t *dkiov, unsigned int doffset,
290 unsigned int nskiov, lnet_kiov_t *skiov, unsigned int soffset,
297 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
298 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
305 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
306 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
313 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
314 int src_niov, lnet_kiov_t *src,
315 unsigned int offset, unsigned int len)
320 #else /* __KERNEL__ */
323 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
325 unsigned int nob = 0;
327 LASSERT(niov == 0 || kiov != NULL);
329 nob += (kiov++)->kiov_len;
333 EXPORT_SYMBOL(lnet_kiov_nob);
336 lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
337 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
340 /* NB diov, siov are READ-ONLY */
341 unsigned int this_nob;
348 LASSERT (!in_interrupt ());
351 while (doffset >= diov->kiov_len) {
352 doffset -= diov->kiov_len;
359 while (soffset >= siov->kiov_len) {
360 soffset -= siov->kiov_len;
369 this_nob = MIN(diov->kiov_len - doffset,
370 siov->kiov_len - soffset);
371 this_nob = MIN(this_nob, nob);
374 daddr = ((char *)kmap(diov->kiov_page)) +
375 diov->kiov_offset + doffset;
377 saddr = ((char *)kmap(siov->kiov_page)) +
378 siov->kiov_offset + soffset;
380 /* Vanishing risk of kmap deadlock when mapping 2 pages.
381 * However in practice at least one of the kiovs will be mapped
382 * kernel pages and the map/unmap will be NOOPs */
384 memcpy (daddr, saddr, this_nob);
387 if (diov->kiov_len > doffset + this_nob) {
391 kunmap(diov->kiov_page);
398 if (siov->kiov_len > soffset + this_nob) {
402 kunmap(siov->kiov_page);
411 kunmap(diov->kiov_page);
413 kunmap(siov->kiov_page);
415 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
418 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
419 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
422 /* NB iov, kiov are READ-ONLY */
423 unsigned int this_nob;
429 LASSERT (!in_interrupt ());
432 while (iovoffset >= iov->iov_len) {
433 iovoffset -= iov->iov_len;
440 while (kiovoffset >= kiov->kiov_len) {
441 kiovoffset -= kiov->kiov_len;
450 this_nob = MIN(iov->iov_len - iovoffset,
451 kiov->kiov_len - kiovoffset);
452 this_nob = MIN(this_nob, nob);
455 addr = ((char *)kmap(kiov->kiov_page)) +
456 kiov->kiov_offset + kiovoffset;
458 memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
461 if (iov->iov_len > iovoffset + this_nob) {
462 iovoffset += this_nob;
469 if (kiov->kiov_len > kiovoffset + this_nob) {
471 kiovoffset += this_nob;
473 kunmap(kiov->kiov_page);
483 kunmap(kiov->kiov_page);
485 EXPORT_SYMBOL(lnet_copy_kiov2iov);
488 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
489 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
492 /* NB kiov, iov are READ-ONLY */
493 unsigned int this_nob;
499 LASSERT (!in_interrupt ());
502 while (kiovoffset >= kiov->kiov_len) {
503 kiovoffset -= kiov->kiov_len;
510 while (iovoffset >= iov->iov_len) {
511 iovoffset -= iov->iov_len;
520 this_nob = MIN(kiov->kiov_len - kiovoffset,
521 iov->iov_len - iovoffset);
522 this_nob = MIN(this_nob, nob);
525 addr = ((char *)kmap(kiov->kiov_page)) +
526 kiov->kiov_offset + kiovoffset;
528 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
531 if (kiov->kiov_len > kiovoffset + this_nob) {
533 kiovoffset += this_nob;
535 kunmap(kiov->kiov_page);
542 if (iov->iov_len > iovoffset + this_nob) {
543 iovoffset += this_nob;
552 kunmap(kiov->kiov_page);
554 EXPORT_SYMBOL(lnet_copy_iov2kiov);
557 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
558 int src_niov, lnet_kiov_t *src,
559 unsigned int offset, unsigned int len)
561 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
562 * for exactly 'len' bytes, and return the number of entries.
563 * NB not destructive to 'src' */
564 unsigned int frag_len;
567 if (len == 0) /* no data => */
568 return (0); /* no frags */
570 LASSERT (src_niov > 0);
571 while (offset >= src->kiov_len) { /* skip initial frags */
572 offset -= src->kiov_len;
575 LASSERT (src_niov > 0);
580 LASSERT (src_niov > 0);
581 LASSERT ((int)niov <= dst_niov);
583 frag_len = src->kiov_len - offset;
584 dst->kiov_page = src->kiov_page;
585 dst->kiov_offset = src->kiov_offset + offset;
587 if (len <= frag_len) {
589 LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE);
593 dst->kiov_len = frag_len;
594 LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE);
604 EXPORT_SYMBOL(lnet_extract_kiov);
608 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
609 unsigned int offset, unsigned int mlen, unsigned int rlen)
611 unsigned int niov = 0;
612 struct iovec *iov = NULL;
613 lnet_kiov_t *kiov = NULL;
616 LASSERT (!in_interrupt ());
617 LASSERT (mlen == 0 || msg != NULL);
620 LASSERT(msg->msg_receiving);
621 LASSERT(!msg->msg_sending);
622 LASSERT(rlen == msg->msg_len);
623 LASSERT(mlen <= msg->msg_len);
624 LASSERT(msg->msg_offset == offset);
625 LASSERT(msg->msg_wanted == mlen);
627 msg->msg_receiving = 0;
630 niov = msg->msg_niov;
632 kiov = msg->msg_kiov;
635 LASSERT ((iov == NULL) != (kiov == NULL));
639 rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
640 niov, iov, kiov, offset, mlen, rlen);
642 lnet_finalize(ni, msg, rc);
646 lnet_setpayloadbuffer(lnet_msg_t *msg)
648 lnet_libmd_t *md = msg->msg_md;
650 LASSERT (msg->msg_len > 0);
651 LASSERT (!msg->msg_routing);
652 LASSERT (md != NULL);
653 LASSERT (msg->msg_niov == 0);
654 LASSERT (msg->msg_iov == NULL);
655 LASSERT (msg->msg_kiov == NULL);
657 msg->msg_niov = md->md_niov;
658 if ((md->md_options & LNET_MD_KIOV) != 0)
659 msg->msg_kiov = md->md_iov.kiov;
661 msg->msg_iov = md->md_iov.iov;
665 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
666 unsigned int offset, unsigned int len)
668 msg->msg_type = type;
669 msg->msg_target = target;
671 msg->msg_offset = offset;
674 lnet_setpayloadbuffer(msg);
676 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
677 msg->msg_hdr.type = cpu_to_le32(type);
678 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
679 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
680 /* src_nid will be set later */
681 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
682 msg->msg_hdr.payload_length = cpu_to_le32(len);
686 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
688 void *priv = msg->msg_private;
691 LASSERT (!in_interrupt ());
692 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
693 (msg->msg_txcredit && msg->msg_peertxcredit));
695 rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
697 lnet_finalize(ni, msg, rc);
701 lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
705 LASSERT(!msg->msg_sending);
706 LASSERT(msg->msg_receiving);
707 LASSERT(!msg->msg_rx_ready_delay);
708 LASSERT(ni->ni_lnd->lnd_eager_recv != NULL);
710 msg->msg_rx_ready_delay = 1;
711 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
714 CERROR("recv from %s / send to %s aborted: "
715 "eager_recv failed %d\n",
716 libcfs_nid2str(msg->msg_rxpeer->lp_nid),
717 libcfs_id2str(msg->msg_target), rc);
718 LASSERT(rc < 0); /* required by my callers */
724 /* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */
726 lnet_ni_query_locked(lnet_ni_t *ni, lnet_peer_t *lp)
728 cfs_time_t last_alive = 0;
730 LASSERT(lnet_peer_aliveness_enabled(lp));
731 LASSERT(ni->ni_lnd->lnd_query != NULL);
733 lnet_net_unlock(lp->lp_cpt);
734 (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
735 lnet_net_lock(lp->lp_cpt);
737 lp->lp_last_query = cfs_time_current();
739 if (last_alive != 0) /* NI has updated timestamp */
740 lp->lp_last_alive = last_alive;
743 /* NB: always called with lnet_net_lock held */
745 lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
750 LASSERT (lnet_peer_aliveness_enabled(lp));
752 /* Trust lnet_notify() if it has more recent aliveness news, but
753 * ignore the initial assumed death (see lnet_peers_start_down()).
755 if (!lp->lp_alive && lp->lp_alive_count > 0 &&
756 cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive))
759 deadline = cfs_time_add(lp->lp_last_alive,
760 cfs_time_seconds(lp->lp_ni->ni_peertimeout));
761 alive = cfs_time_after(deadline, now);
763 /* Update obsolete lp_alive except for routers assumed to be dead
764 * initially, because router checker would update aliveness in this
765 * case, and moreover lp_last_alive at peer creation is assumed.
767 if (alive && !lp->lp_alive &&
768 !(lnet_isrouter(lp) && lp->lp_alive_count == 0))
769 lnet_notify_locked(lp, 0, 1, lp->lp_last_alive);
775 /* NB: returns 1 when alive, 0 when dead, negative when error;
776 * may drop the lnet_net_lock */
778 lnet_peer_alive_locked (lnet_peer_t *lp)
780 cfs_time_t now = cfs_time_current();
782 if (!lnet_peer_aliveness_enabled(lp))
785 if (lnet_peer_is_alive(lp, now))
788 /* Peer appears dead, but we should avoid frequent NI queries (at
789 * most once per lnet_queryinterval seconds). */
790 if (lp->lp_last_query != 0) {
791 static const int lnet_queryinterval = 1;
793 cfs_time_t next_query =
794 cfs_time_add(lp->lp_last_query,
795 cfs_time_seconds(lnet_queryinterval));
797 if (cfs_time_before(now, next_query)) {
799 CWARN("Unexpected aliveness of peer %s: "
801 libcfs_nid2str(lp->lp_nid),
802 (int)now, (int)next_query,
804 lp->lp_ni->ni_peertimeout);
809 /* query NI for latest aliveness news */
810 lnet_ni_query_locked(lp->lp_ni, lp);
812 if (lnet_peer_is_alive(lp, now))
815 lnet_notify_locked(lp, 0, 0, lp->lp_last_alive);
820 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
822 /* lnet_send is going to lnet_net_unlock immediately after this,
823 * so it sets do_send FALSE and I don't do the unlock/send/lock bit.
824 * I return EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer
825 * appears dead, and 0 if sent or OK to send */
826 struct lnet_peer *lp = msg->msg_txpeer;
827 struct lnet_ni *ni = lp->lp_ni;
828 struct lnet_tx_queue *tq;
831 /* non-lnet_send() callers have checked before */
832 LASSERT(!do_send || msg->msg_tx_delayed);
833 LASSERT(!msg->msg_receiving);
834 LASSERT(msg->msg_tx_committed);
836 cpt = msg->msg_tx_cpt;
837 tq = ni->ni_tx_queues[cpt];
839 /* NB 'lp' is always the next hop */
840 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
841 lnet_peer_alive_locked(lp) == 0) {
842 the_lnet.ln_counters[cpt]->drop_count++;
843 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
844 lnet_net_unlock(cpt);
846 CNETERR("Dropping message for %s: peer not alive\n",
847 libcfs_id2str(msg->msg_target));
849 lnet_finalize(ni, msg, -EHOSTUNREACH);
855 if (!msg->msg_peertxcredit) {
856 LASSERT ((lp->lp_txcredits < 0) ==
857 !cfs_list_empty(&lp->lp_txq));
859 msg->msg_peertxcredit = 1;
860 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
863 if (lp->lp_txcredits < lp->lp_mintxcredits)
864 lp->lp_mintxcredits = lp->lp_txcredits;
866 if (lp->lp_txcredits < 0) {
867 msg->msg_tx_delayed = 1;
868 cfs_list_add_tail(&msg->msg_list, &lp->lp_txq);
873 if (!msg->msg_txcredit) {
874 LASSERT((tq->tq_credits < 0) ==
875 !cfs_list_empty(&tq->tq_delayed));
877 msg->msg_txcredit = 1;
880 if (tq->tq_credits < tq->tq_credits_min)
881 tq->tq_credits_min = tq->tq_credits;
883 if (tq->tq_credits < 0) {
884 msg->msg_tx_delayed = 1;
885 cfs_list_add_tail(&msg->msg_list, &tq->tq_delayed);
891 lnet_net_unlock(cpt);
892 lnet_ni_send(ni, msg);
901 lnet_msg2bufpool(lnet_msg_t *msg)
903 lnet_rtrbufpool_t *rbp;
906 LASSERT(msg->msg_rx_committed);
908 cpt = msg->msg_rx_cpt;
909 rbp = &the_lnet.ln_rtrpools[cpt][0];
911 LASSERT(msg->msg_len <= LNET_MTU);
912 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_CACHE_SIZE) {
914 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
921 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
923 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
924 * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
925 * return EAGAIN if msg blocked and 0 if received or OK to receive */
926 lnet_peer_t *lp = msg->msg_rxpeer;
927 lnet_rtrbufpool_t *rbp;
930 LASSERT (msg->msg_iov == NULL);
931 LASSERT (msg->msg_kiov == NULL);
932 LASSERT (msg->msg_niov == 0);
933 LASSERT (msg->msg_routing);
934 LASSERT (msg->msg_receiving);
935 LASSERT (!msg->msg_sending);
937 /* non-lnet_parse callers only receive delayed messages */
938 LASSERT(!do_recv || msg->msg_rx_delayed);
940 if (!msg->msg_peerrtrcredit) {
941 LASSERT ((lp->lp_rtrcredits < 0) ==
942 !cfs_list_empty(&lp->lp_rtrq));
944 msg->msg_peerrtrcredit = 1;
946 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
947 lp->lp_minrtrcredits = lp->lp_rtrcredits;
949 if (lp->lp_rtrcredits < 0) {
950 /* must have checked eager_recv before here */
951 LASSERT(msg->msg_rx_ready_delay);
952 msg->msg_rx_delayed = 1;
953 cfs_list_add_tail(&msg->msg_list, &lp->lp_rtrq);
958 rbp = lnet_msg2bufpool(msg);
960 if (!msg->msg_rtrcredit) {
961 LASSERT ((rbp->rbp_credits < 0) ==
962 !cfs_list_empty(&rbp->rbp_msgs));
964 msg->msg_rtrcredit = 1;
966 if (rbp->rbp_credits < rbp->rbp_mincredits)
967 rbp->rbp_mincredits = rbp->rbp_credits;
969 if (rbp->rbp_credits < 0) {
970 /* must have checked eager_recv before here */
971 LASSERT(msg->msg_rx_ready_delay);
972 msg->msg_rx_delayed = 1;
973 cfs_list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
978 LASSERT (!cfs_list_empty(&rbp->rbp_bufs));
979 rb = cfs_list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
980 cfs_list_del(&rb->rb_list);
982 msg->msg_niov = rbp->rbp_npages;
983 msg->msg_kiov = &rb->rb_kiov[0];
986 int cpt = msg->msg_rx_cpt;
988 lnet_net_unlock(cpt);
989 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
990 0, msg->msg_len, msg->msg_len);
998 lnet_return_tx_credits_locked(lnet_msg_t *msg)
1000 lnet_peer_t *txpeer = msg->msg_txpeer;
1003 if (msg->msg_txcredit) {
1004 struct lnet_ni *ni = txpeer->lp_ni;
1005 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1007 /* give back NI txcredits */
1008 msg->msg_txcredit = 0;
1010 LASSERT((tq->tq_credits < 0) ==
1011 !cfs_list_empty(&tq->tq_delayed));
1014 if (tq->tq_credits <= 0) {
1015 msg2 = cfs_list_entry(tq->tq_delayed.next,
1016 lnet_msg_t, msg_list);
1017 cfs_list_del(&msg2->msg_list);
1019 LASSERT(msg2->msg_txpeer->lp_ni == ni);
1020 LASSERT(msg2->msg_tx_delayed);
1022 (void) lnet_post_send_locked(msg2, 1);
1026 if (msg->msg_peertxcredit) {
1027 /* give back peer txcredits */
1028 msg->msg_peertxcredit = 0;
1030 LASSERT((txpeer->lp_txcredits < 0) ==
1031 !cfs_list_empty(&txpeer->lp_txq));
1033 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1034 LASSERT (txpeer->lp_txqnob >= 0);
1036 txpeer->lp_txcredits++;
1037 if (txpeer->lp_txcredits <= 0) {
1038 msg2 = cfs_list_entry(txpeer->lp_txq.next,
1039 lnet_msg_t, msg_list);
1040 cfs_list_del(&msg2->msg_list);
1042 LASSERT(msg2->msg_txpeer == txpeer);
1043 LASSERT(msg2->msg_tx_delayed);
1045 (void) lnet_post_send_locked(msg2, 1);
1049 if (txpeer != NULL) {
1050 msg->msg_txpeer = NULL;
1051 lnet_peer_decref_locked(txpeer);
1056 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1058 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1062 if (msg->msg_rtrcredit) {
1063 /* give back global router credits */
1065 lnet_rtrbufpool_t *rbp;
1067 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1068 * there until it gets one allocated, or aborts the wait
1070 LASSERT (msg->msg_kiov != NULL);
1072 rb = cfs_list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1074 LASSERT (rbp == lnet_msg2bufpool(msg));
1076 msg->msg_kiov = NULL;
1077 msg->msg_rtrcredit = 0;
1079 LASSERT((rbp->rbp_credits < 0) ==
1080 !cfs_list_empty(&rbp->rbp_msgs));
1081 LASSERT((rbp->rbp_credits > 0) ==
1082 !cfs_list_empty(&rbp->rbp_bufs));
1084 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1086 if (rbp->rbp_credits <= 0) {
1087 msg2 = cfs_list_entry(rbp->rbp_msgs.next,
1088 lnet_msg_t, msg_list);
1089 cfs_list_del(&msg2->msg_list);
1091 (void) lnet_post_routed_recv_locked(msg2, 1);
1095 if (msg->msg_peerrtrcredit) {
1096 /* give back peer router credits */
1097 msg->msg_peerrtrcredit = 0;
1099 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1100 !cfs_list_empty(&rxpeer->lp_rtrq));
1102 rxpeer->lp_rtrcredits++;
1103 if (rxpeer->lp_rtrcredits <= 0) {
1104 msg2 = cfs_list_entry(rxpeer->lp_rtrq.next,
1105 lnet_msg_t, msg_list);
1106 cfs_list_del(&msg2->msg_list);
1108 (void) lnet_post_routed_recv_locked(msg2, 1);
1112 LASSERT (!msg->msg_rtrcredit);
1113 LASSERT (!msg->msg_peerrtrcredit);
1115 if (rxpeer != NULL) {
1116 msg->msg_rxpeer = NULL;
1117 lnet_peer_decref_locked(rxpeer);
1122 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1124 lnet_peer_t *p1 = r1->lr_gateway;
1125 lnet_peer_t *p2 = r2->lr_gateway;
1127 if (r1->lr_priority < r2->lr_priority)
1130 if (r1->lr_priority > r2->lr_priority)
1133 if (r1->lr_hops < r2->lr_hops)
1136 if (r1->lr_hops > r2->lr_hops)
1139 if (p1->lp_txqnob < p2->lp_txqnob)
1142 if (p1->lp_txqnob > p2->lp_txqnob)
1145 if (p1->lp_txcredits > p2->lp_txcredits)
1148 if (p1->lp_txcredits < p2->lp_txcredits)
1151 if (r1->lr_seq - r2->lr_seq <= 0)
1157 static lnet_peer_t *
1158 lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target, lnet_nid_t rtr_nid)
1160 lnet_remotenet_t *rnet;
1162 lnet_route_t *rtr_best;
1163 lnet_route_t *rtr_last;
1164 struct lnet_peer *lp_best;
1165 struct lnet_peer *lp;
1168 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1169 * rtr_nid nid, otherwise find the best gateway I can use */
1171 rnet = lnet_find_net_locked(LNET_NIDNET(target));
1176 rtr_best = rtr_last = NULL;
1177 cfs_list_for_each_entry(rtr, &rnet->lrn_routes, lr_list) {
1178 lp = rtr->lr_gateway;
1180 if (!lp->lp_alive || /* gateway is down */
1181 ((lp->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0 &&
1182 rtr->lr_downis != 0)) /* NI to target is down */
1185 if (ni != NULL && lp->lp_ni != ni)
1188 if (lp->lp_nid == rtr_nid) /* it's pre-determined router */
1191 if (lp_best == NULL) {
1192 rtr_best = rtr_last = rtr;
1197 /* no protection on below fields, but it's harmless */
1198 if (rtr_last->lr_seq - rtr->lr_seq < 0)
1201 rc = lnet_compare_routes(rtr, rtr_best);
1209 /* set sequence number on the best router to the latest sequence + 1
1210 * so we can round-robin all routers, it's race and inaccurate but
1211 * harmless and functional */
1212 if (rtr_best != NULL)
1213 rtr_best->lr_seq = rtr_last->lr_seq + 1;
1218 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1220 lnet_nid_t dst_nid = msg->msg_target.nid;
1221 struct lnet_ni *src_ni;
1222 struct lnet_ni *local_ni;
1223 struct lnet_peer *lp;
1228 /* NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1229 * but we might want to use pre-determined router for ACK/REPLY
1231 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1232 LASSERT (msg->msg_txpeer == NULL);
1233 LASSERT (!msg->msg_sending);
1234 LASSERT (!msg->msg_target_is_router);
1235 LASSERT (!msg->msg_receiving);
1237 msg->msg_sending = 1;
1239 LASSERT(!msg->msg_tx_committed);
1240 cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid);
1244 if (the_lnet.ln_shutdown) {
1245 lnet_net_unlock(cpt);
1249 if (src_nid == LNET_NID_ANY) {
1252 src_ni = lnet_nid2ni_locked(src_nid, cpt);
1253 if (src_ni == NULL) {
1254 lnet_net_unlock(cpt);
1255 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1256 "local nid\n", libcfs_nid2str(dst_nid),
1257 libcfs_nid2str(src_nid));
1260 LASSERT (!msg->msg_routing);
1263 /* Is this for someone on a local network? */
1264 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt);
1266 if (local_ni != NULL) {
1267 if (src_ni == NULL) {
1269 src_nid = src_ni->ni_nid;
1270 } else if (src_ni == local_ni) {
1271 lnet_ni_decref_locked(local_ni, cpt);
1273 lnet_ni_decref_locked(local_ni, cpt);
1274 lnet_ni_decref_locked(src_ni, cpt);
1275 lnet_net_unlock(cpt);
1276 LCONSOLE_WARN("No route to %s via from %s\n",
1277 libcfs_nid2str(dst_nid),
1278 libcfs_nid2str(src_nid));
1282 LASSERT(src_nid != LNET_NID_ANY);
1283 lnet_msg_commit(msg, cpt);
1285 if (!msg->msg_routing)
1286 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1288 if (src_ni == the_lnet.ln_loni) {
1289 /* No send credit hassles with LOLND */
1290 lnet_net_unlock(cpt);
1291 lnet_ni_send(src_ni, msg);
1294 lnet_ni_decref_locked(src_ni, cpt);
1295 lnet_net_unlock(cpt);
1299 rc = lnet_nid2peer_locked(&lp, dst_nid, cpt);
1300 /* lp has ref on src_ni; lose mine */
1301 lnet_ni_decref_locked(src_ni, cpt);
1303 lnet_net_unlock(cpt);
1304 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1305 libcfs_nid2str(dst_nid));
1306 /* ENOMEM or shutting down */
1309 LASSERT (lp->lp_ni == src_ni);
1312 lnet_net_unlock(cpt);
1315 * - once application finishes computation, check here to update
1316 * router states before it waits for pending IO in LNetEQPoll
1317 * - recursion breaker: router checker sends no message
1318 * to remote networks */
1319 if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING)
1320 lnet_router_checker();
1324 /* sending to a remote network */
1325 lp = lnet_find_route_locked(src_ni, dst_nid, rtr_nid);
1328 lnet_ni_decref_locked(src_ni, cpt);
1329 lnet_net_unlock(cpt);
1331 LCONSOLE_WARN("No route to %s via %s "
1332 "(all routers down)\n",
1333 libcfs_id2str(msg->msg_target),
1334 libcfs_nid2str(src_nid));
1335 return -EHOSTUNREACH;
1338 /* rtr_nid is LNET_NID_ANY or NID of pre-determined router,
1339 * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't
1340 * pre-determined router, this can happen if router table
1341 * was changed when we release the lock */
1342 if (rtr_nid != lp->lp_nid) {
1343 cpt2 = lnet_cpt_of_nid_locked(lp->lp_nid);
1346 lnet_ni_decref_locked(src_ni, cpt);
1347 lnet_net_unlock(cpt);
1349 rtr_nid = lp->lp_nid;
1355 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1356 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1357 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1359 if (src_ni == NULL) {
1361 src_nid = src_ni->ni_nid;
1363 LASSERT (src_ni == lp->lp_ni);
1364 lnet_ni_decref_locked(src_ni, cpt);
1367 lnet_peer_addref_locked(lp);
1369 LASSERT(src_nid != LNET_NID_ANY);
1370 lnet_msg_commit(msg, cpt);
1372 if (!msg->msg_routing) {
1373 /* I'm the source and now I know which NI to send on */
1374 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1377 msg->msg_target_is_router = 1;
1378 msg->msg_target.nid = lp->lp_nid;
1379 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1382 /* 'lp' is our best choice of peer */
1384 LASSERT (!msg->msg_peertxcredit);
1385 LASSERT (!msg->msg_txcredit);
1386 LASSERT (msg->msg_txpeer == NULL);
1388 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1390 rc = lnet_post_send_locked(msg, 0);
1391 lnet_net_unlock(cpt);
1393 if (rc == EHOSTUNREACH)
1394 return -EHOSTUNREACH;
1397 lnet_ni_send(src_ni, msg);
1403 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1406 the_lnet.ln_counters[cpt]->drop_count++;
1407 the_lnet.ln_counters[cpt]->drop_length += nob;
1408 lnet_net_unlock(cpt);
1410 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1414 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1416 lnet_hdr_t *hdr = &msg->msg_hdr;
1418 if (msg->msg_wanted != 0)
1419 lnet_setpayloadbuffer(msg);
1421 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1423 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1424 * it back into the ACK during lnet_finalize() */
1425 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1426 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1428 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1429 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1433 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1435 lnet_hdr_t *hdr = &msg->msg_hdr;
1436 struct lnet_match_info info;
1439 /* Convert put fields to host byte order */
1440 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1441 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1442 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1444 info.mi_id.nid = hdr->src_nid;
1445 info.mi_id.pid = hdr->src_pid;
1446 info.mi_opc = LNET_MD_OP_PUT;
1447 info.mi_portal = hdr->msg.put.ptl_index;
1448 info.mi_rlength = hdr->payload_length;
1449 info.mi_roffset = hdr->msg.put.offset;
1450 info.mi_mbits = hdr->msg.put.match_bits;
1452 msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL;
1455 rc = lnet_ptl_match_md(&info, msg);
1460 case LNET_MATCHMD_OK:
1461 lnet_recv_put(ni, msg);
1464 case LNET_MATCHMD_NONE:
1465 if (msg->msg_rx_delayed) /* attached on delayed list */
1468 rc = lnet_ni_eager_recv(ni, msg);
1473 case LNET_MATCHMD_DROP:
1474 CNETERR("Dropping PUT from %s portal %d match "LPU64
1475 " offset %d length %d: %d\n",
1476 libcfs_id2str(info.mi_id), info.mi_portal,
1477 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1479 return ENOENT; /* +ve: OK but no match */
1484 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1486 struct lnet_match_info info;
1487 lnet_hdr_t *hdr = &msg->msg_hdr;
1488 lnet_handle_wire_t reply_wmd;
1491 /* Convert get fields to host byte order */
1492 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1493 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1494 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1495 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1497 info.mi_id.nid = hdr->src_nid;
1498 info.mi_id.pid = hdr->src_pid;
1499 info.mi_opc = LNET_MD_OP_GET;
1500 info.mi_portal = hdr->msg.get.ptl_index;
1501 info.mi_rlength = hdr->msg.get.sink_length;
1502 info.mi_roffset = hdr->msg.get.src_offset;
1503 info.mi_mbits = hdr->msg.get.match_bits;
1505 rc = lnet_ptl_match_md(&info, msg);
1506 if (rc == LNET_MATCHMD_DROP) {
1507 CNETERR("Dropping GET from %s portal %d match "LPU64
1508 " offset %d length %d\n",
1509 libcfs_id2str(info.mi_id), info.mi_portal,
1510 info.mi_mbits, info.mi_roffset, info.mi_rlength);
1511 return ENOENT; /* +ve: OK but no match */
1514 LASSERT(rc == LNET_MATCHMD_OK);
1516 lnet_build_msg_event(msg, LNET_EVENT_GET);
1518 reply_wmd = hdr->msg.get.return_wmd;
1520 lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
1521 msg->msg_offset, msg->msg_wanted);
1523 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1526 /* The LND completes the REPLY from her recv procedure */
1527 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1528 msg->msg_offset, msg->msg_len, msg->msg_len);
1532 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1533 msg->msg_receiving = 0;
1535 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
1537 /* didn't get as far as lnet_ni_send() */
1538 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1539 libcfs_nid2str(ni->ni_nid),
1540 libcfs_id2str(info.mi_id), rc);
1542 lnet_finalize(ni, msg, rc);
1549 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1551 void *private = msg->msg_private;
1552 lnet_hdr_t *hdr = &msg->msg_hdr;
1553 lnet_process_id_t src = {0};
1559 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
1562 src.nid = hdr->src_nid;
1563 src.pid = hdr->src_pid;
1565 /* NB handles only looked up by creator (no flips) */
1566 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1567 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1568 CNETERR("%s: Dropping REPLY from %s for %s "
1569 "MD "LPX64"."LPX64"\n",
1570 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1571 (md == NULL) ? "invalid" : "inactive",
1572 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1573 hdr->msg.reply.dst_wmd.wh_object_cookie);
1574 if (md != NULL && md->md_me != NULL)
1575 CERROR("REPLY MD also attached to portal %d\n",
1576 md->md_me->me_portal);
1578 lnet_res_unlock(cpt);
1579 return ENOENT; /* +ve: OK but no match */
1582 LASSERT (md->md_offset == 0);
1584 rlength = hdr->payload_length;
1585 mlength = MIN(rlength, (int)md->md_length);
1587 if (mlength < rlength &&
1588 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1589 CNETERR("%s: Dropping REPLY from %s length %d "
1590 "for MD "LPX64" would overflow (%d)\n",
1591 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1592 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1594 lnet_res_unlock(cpt);
1595 return ENOENT; /* +ve: OK but no match */
1598 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1599 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1600 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1602 lnet_msg_attach_md(msg, md, 0, mlength);
1605 lnet_setpayloadbuffer(msg);
1607 lnet_res_unlock(cpt);
1609 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
1611 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1616 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1618 lnet_hdr_t *hdr = &msg->msg_hdr;
1619 lnet_process_id_t src = {0};
1623 src.nid = hdr->src_nid;
1624 src.pid = hdr->src_pid;
1626 /* Convert ack fields to host byte order */
1627 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1628 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1630 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
1633 /* NB handles only looked up by creator (no flips) */
1634 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1635 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1636 /* Don't moan; this is expected */
1638 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1639 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1640 (md == NULL) ? "invalid" : "inactive",
1641 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1642 hdr->msg.ack.dst_wmd.wh_object_cookie);
1643 if (md != NULL && md->md_me != NULL)
1644 CERROR("Source MD also attached to portal %d\n",
1645 md->md_me->me_portal);
1647 lnet_res_unlock(cpt);
1648 return ENOENT; /* +ve! */
1651 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1652 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1653 hdr->msg.ack.dst_wmd.wh_object_cookie);
1655 lnet_msg_attach_md(msg, md, 0, 0);
1657 lnet_res_unlock(cpt);
1659 lnet_build_msg_event(msg, LNET_EVENT_ACK);
1661 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
1666 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
1671 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
1672 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
1673 if (ni->ni_lnd->lnd_eager_recv == NULL) {
1674 msg->msg_rx_ready_delay = 1;
1676 lnet_net_unlock(msg->msg_rx_cpt);
1677 rc = lnet_ni_eager_recv(ni, msg);
1678 lnet_net_lock(msg->msg_rx_cpt);
1683 rc = lnet_post_routed_recv_locked(msg, 0);
1691 lnet_msgtyp2str (int type)
1700 case LNET_MSG_REPLY:
1702 case LNET_MSG_HELLO:
1705 return ("<UNKNOWN>");
1708 EXPORT_SYMBOL(lnet_msgtyp2str);
1711 lnet_print_hdr(lnet_hdr_t * hdr)
1713 lnet_process_id_t src = {0};
1714 lnet_process_id_t dst = {0};
1715 char *type_str = lnet_msgtyp2str (hdr->type);
1717 src.nid = hdr->src_nid;
1718 src.pid = hdr->src_pid;
1720 dst.nid = hdr->dest_nid;
1721 dst.pid = hdr->dest_pid;
1723 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
1724 CWARN(" From %s\n", libcfs_id2str(src));
1725 CWARN(" To %s\n", libcfs_id2str(dst));
1727 switch (hdr->type) {
1732 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
1733 "match bits "LPU64"\n",
1734 hdr->msg.put.ptl_index,
1735 hdr->msg.put.ack_wmd.wh_interface_cookie,
1736 hdr->msg.put.ack_wmd.wh_object_cookie,
1737 hdr->msg.put.match_bits);
1738 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
1739 hdr->payload_length, hdr->msg.put.offset,
1740 hdr->msg.put.hdr_data);
1744 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
1745 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
1746 hdr->msg.get.return_wmd.wh_interface_cookie,
1747 hdr->msg.get.return_wmd.wh_object_cookie,
1748 hdr->msg.get.match_bits);
1749 CWARN(" Length %d, src offset %d\n",
1750 hdr->msg.get.sink_length,
1751 hdr->msg.get.src_offset);
1755 CWARN(" dst md "LPX64"."LPX64", "
1756 "manipulated length %d\n",
1757 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1758 hdr->msg.ack.dst_wmd.wh_object_cookie,
1759 hdr->msg.ack.mlength);
1762 case LNET_MSG_REPLY:
1763 CWARN(" dst md "LPX64"."LPX64", "
1765 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1766 hdr->msg.reply.dst_wmd.wh_object_cookie,
1767 hdr->payload_length);
1773 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
1774 void *private, int rdma_req)
1779 struct lnet_msg *msg;
1780 lnet_pid_t dest_pid;
1781 lnet_nid_t dest_nid;
1783 __u32 payload_length;
1786 LASSERT (!in_interrupt ());
1788 type = le32_to_cpu(hdr->type);
1789 src_nid = le64_to_cpu(hdr->src_nid);
1790 dest_nid = le64_to_cpu(hdr->dest_nid);
1791 dest_pid = le32_to_cpu(hdr->dest_pid);
1792 payload_length = le32_to_cpu(hdr->payload_length);
1794 for_me = (ni->ni_nid == dest_nid);
1795 cpt = lnet_cpt_of_nid(from_nid);
1800 if (payload_length > 0) {
1801 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
1802 libcfs_nid2str(from_nid),
1803 libcfs_nid2str(src_nid),
1804 lnet_msgtyp2str(type), payload_length);
1810 case LNET_MSG_REPLY:
1811 if (payload_length > (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
1812 CERROR("%s, src %s: bad %s payload %d "
1813 "(%d max expected)\n",
1814 libcfs_nid2str(from_nid),
1815 libcfs_nid2str(src_nid),
1816 lnet_msgtyp2str(type),
1818 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
1824 CERROR("%s, src %s: Bad message type 0x%x\n",
1825 libcfs_nid2str(from_nid),
1826 libcfs_nid2str(src_nid), type);
1830 if (the_lnet.ln_routing &&
1831 ni->ni_last_alive != cfs_time_current_sec()) {
1834 /* NB: so far here is the only place to set NI status to "up */
1835 ni->ni_last_alive = cfs_time_current_sec();
1836 if (ni->ni_status != NULL &&
1837 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
1838 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
1842 /* Regard a bad destination NID as a protocol error. Senders should
1843 * know what they're doing; if they don't they're misconfigured, buggy
1844 * or malicious so we chop them off at the knees :) */
1847 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
1848 /* should have gone direct */
1849 CERROR ("%s, src %s: Bad dest nid %s "
1850 "(should have been sent direct)\n",
1851 libcfs_nid2str(from_nid),
1852 libcfs_nid2str(src_nid),
1853 libcfs_nid2str(dest_nid));
1857 if (lnet_islocalnid(dest_nid)) {
1858 /* dest is another local NI; sender should have used
1859 * this node's NID on its own network */
1860 CERROR ("%s, src %s: Bad dest nid %s "
1861 "(it's my nid but on a different network)\n",
1862 libcfs_nid2str(from_nid),
1863 libcfs_nid2str(src_nid),
1864 libcfs_nid2str(dest_nid));
1868 if (rdma_req && type == LNET_MSG_GET) {
1869 CERROR ("%s, src %s: Bad optimized GET for %s "
1870 "(final destination must be me)\n",
1871 libcfs_nid2str(from_nid),
1872 libcfs_nid2str(src_nid),
1873 libcfs_nid2str(dest_nid));
1877 if (!the_lnet.ln_routing) {
1878 CERROR ("%s, src %s: Dropping message for %s "
1879 "(routing not enabled)\n",
1880 libcfs_nid2str(from_nid),
1881 libcfs_nid2str(src_nid),
1882 libcfs_nid2str(dest_nid));
1887 /* Message looks OK; we're not going to return an error, so we MUST
1888 * call back lnd_recv() come what may... */
1890 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
1891 fail_peer (src_nid, 0)) /* shall we now? */
1893 CERROR("%s, src %s: Dropping %s to simulate failure\n",
1894 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1895 lnet_msgtyp2str(type));
1899 msg = lnet_msg_alloc();
1901 CERROR("%s, src %s: Dropping %s (out of memory)\n",
1902 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1903 lnet_msgtyp2str(type));
1907 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
1909 msg->msg_type = type;
1910 msg->msg_private = private;
1911 msg->msg_receiving = 1;
1912 msg->msg_len = msg->msg_wanted = payload_length;
1913 msg->msg_offset = 0;
1914 msg->msg_hdr = *hdr;
1915 /* for building message event */
1916 msg->msg_from = from_nid;
1918 msg->msg_target.pid = dest_pid;
1919 msg->msg_target.nid = dest_nid;
1920 msg->msg_routing = 1;
1923 /* convert common msg->hdr fields to host byteorder */
1924 msg->msg_hdr.type = type;
1925 msg->msg_hdr.src_nid = src_nid;
1926 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
1927 msg->msg_hdr.dest_nid = dest_nid;
1928 msg->msg_hdr.dest_pid = dest_pid;
1929 msg->msg_hdr.payload_length = payload_length;
1933 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid, cpt);
1935 lnet_net_unlock(cpt);
1936 CERROR("%s, src %s: Dropping %s "
1937 "(error %d looking up sender)\n",
1938 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1939 lnet_msgtyp2str(type), rc);
1944 lnet_msg_commit(msg, cpt);
1947 rc = lnet_parse_forward_locked(ni, msg);
1948 lnet_net_unlock(cpt);
1953 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1954 0, payload_length, payload_length);
1959 lnet_net_unlock(cpt);
1963 rc = lnet_parse_ack(ni, msg);
1966 rc = lnet_parse_put(ni, msg);
1969 rc = lnet_parse_get(ni, msg, rdma_req);
1971 case LNET_MSG_REPLY:
1972 rc = lnet_parse_reply(ni, msg);
1977 goto free_drop; /* prevent an unused label if !kernel */
1983 LASSERT (rc == ENOENT);
1986 LASSERT(msg->msg_md == NULL);
1987 lnet_finalize(ni, msg, rc);
1990 lnet_drop_message(ni, cpt, private, payload_length);
1993 EXPORT_SYMBOL(lnet_parse);
1996 lnet_drop_delayed_msg_list(cfs_list_t *head, char *reason)
1998 while (!cfs_list_empty(head)) {
1999 lnet_process_id_t id = {0};
2002 msg = cfs_list_entry(head->next, lnet_msg_t, msg_list);
2003 cfs_list_del(&msg->msg_list);
2005 id.nid = msg->msg_hdr.src_nid;
2006 id.pid = msg->msg_hdr.src_pid;
2008 LASSERT(msg->msg_md == NULL);
2009 LASSERT(msg->msg_rx_delayed);
2010 LASSERT(msg->msg_rxpeer != NULL);
2011 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2013 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
2014 " offset %d length %d: %s\n",
2016 msg->msg_hdr.msg.put.ptl_index,
2017 msg->msg_hdr.msg.put.match_bits,
2018 msg->msg_hdr.msg.put.offset,
2019 msg->msg_hdr.payload_length, reason);
2021 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2022 * called lnet_drop_message(), so I just hang onto msg as well
2023 * until that's done */
2025 lnet_drop_message(msg->msg_rxpeer->lp_ni,
2026 msg->msg_rxpeer->lp_cpt,
2027 msg->msg_private, msg->msg_len);
2029 * NB: message will not generate event because w/o attached MD,
2030 * but we still should give error code so lnet_msg_decommit()
2031 * can skip counters operations and other checks.
2033 lnet_finalize(msg->msg_rxpeer->lp_ni, msg, -ENOENT);
2038 lnet_recv_delayed_msg_list(cfs_list_t *head)
2040 while (!cfs_list_empty(head)) {
2042 lnet_process_id_t id;
2044 msg = cfs_list_entry(head->next, lnet_msg_t, msg_list);
2045 cfs_list_del(&msg->msg_list);
2047 /* md won't disappear under me, since each msg
2048 * holds a ref on it */
2050 id.nid = msg->msg_hdr.src_nid;
2051 id.pid = msg->msg_hdr.src_pid;
2053 LASSERT(msg->msg_rx_delayed);
2054 LASSERT(msg->msg_md != NULL);
2055 LASSERT(msg->msg_rxpeer != NULL);
2056 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2058 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2059 "match "LPU64" offset %d length %d.\n",
2060 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2061 msg->msg_hdr.msg.put.match_bits,
2062 msg->msg_hdr.msg.put.offset,
2063 msg->msg_hdr.payload_length);
2065 lnet_recv_put(msg->msg_rxpeer->lp_ni, msg);
2070 * Initiate an asynchronous PUT operation.
2072 * There are several events associated with a PUT: completion of the send on
2073 * the initiator node (LNET_EVENT_SEND), and when the send completes
2074 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2075 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2076 * used at the target node to indicate the completion of incoming data
2079 * The local events will be logged in the EQ associated with the MD pointed to
2080 * by \a mdh handle. Using a MD without an associated EQ results in these
2081 * events being discarded. In this case, the caller must have another
2082 * mechanism (e.g., a higher level protocol) for determining when it is safe
2083 * to modify the memory region associated with the MD.
2085 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2086 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2088 * \param self Indicates the NID of a local interface through which to send
2089 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2090 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2091 * must be "free floating" (See LNetMDBind()).
2092 * \param ack Controls whether an acknowledgment is requested.
2093 * Acknowledgments are only sent when they are requested by the initiating
2094 * process and the target MD enables them.
2095 * \param target A process identifier for the target process.
2096 * \param portal The index in the \a target's portal table.
2097 * \param match_bits The match bits to use for MD selection at the target
2099 * \param offset The offset into the target MD (only used when the target
2100 * MD has the LNET_MD_MANAGE_REMOTE option set).
2101 * \param hdr_data 64 bits of user data that can be included in the message
2102 * header. This data is written to an event queue entry at the target if an
2103 * EQ is present on the matching MD.
2105 * \retval 0 Success, and only in this case events will be generated
2106 * and logged to EQ (if it exists).
2107 * \retval -EIO Simulated failure.
2108 * \retval -ENOMEM Memory allocation failure.
2109 * \retval -ENOENT Invalid MD object.
2111 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2114 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2115 lnet_process_id_t target, unsigned int portal,
2116 __u64 match_bits, unsigned int offset,
2119 struct lnet_msg *msg;
2120 struct lnet_libmd *md;
2124 LASSERT (the_lnet.ln_init);
2125 LASSERT (the_lnet.ln_refcount > 0);
2127 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2128 fail_peer (target.nid, 1)) /* shall we now? */
2130 CERROR("Dropping PUT to %s: simulated failure\n",
2131 libcfs_id2str(target));
2135 msg = lnet_msg_alloc();
2137 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2138 libcfs_id2str(target));
2141 msg->msg_vmflush = !!memory_pressure_get();
2143 cpt = lnet_cpt_of_cookie(mdh.cookie);
2146 md = lnet_handle2md(&mdh);
2147 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2148 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2149 match_bits, portal, libcfs_id2str(target),
2150 md == NULL ? -1 : md->md_threshold);
2151 if (md != NULL && md->md_me != NULL)
2152 CERROR("Source MD also attached to portal %d\n",
2153 md->md_me->me_portal);
2154 lnet_res_unlock(cpt);
2160 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2162 lnet_msg_attach_md(msg, md, 0, 0);
2164 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2166 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2167 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2168 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2169 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2171 /* NB handles only looked up by creator (no flips) */
2172 if (ack == LNET_ACK_REQ) {
2173 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2174 the_lnet.ln_interface_cookie;
2175 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2176 md->md_lh.lh_cookie;
2178 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2179 LNET_WIRE_HANDLE_COOKIE_NONE;
2180 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2181 LNET_WIRE_HANDLE_COOKIE_NONE;
2184 lnet_res_unlock(cpt);
2186 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2188 rc = lnet_send(self, msg, LNET_NID_ANY);
2190 CNETERR( "Error sending PUT to %s: %d\n",
2191 libcfs_id2str(target), rc);
2192 lnet_finalize (NULL, msg, rc);
2195 /* completion will be signalled by an event */
2198 EXPORT_SYMBOL(LNetPut);
2201 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2203 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2204 * returns a msg for the LND to pass to lnet_finalize() when the sink
2205 * data has been received.
2207 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2208 * lnet_finalize() is called on it, so the LND must call this first */
2210 struct lnet_msg *msg = lnet_msg_alloc();
2211 struct lnet_libmd *getmd = getmsg->msg_md;
2212 lnet_process_id_t peer_id = getmsg->msg_target;
2215 LASSERT(!getmsg->msg_target_is_router);
2216 LASSERT(!getmsg->msg_routing);
2218 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2221 LASSERT (getmd->md_refcount > 0);
2224 CERROR ("%s: Dropping REPLY from %s: can't allocate msg\n",
2225 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2229 if (getmd->md_threshold == 0) {
2230 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2231 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2233 lnet_res_unlock(cpt);
2237 LASSERT(getmd->md_offset == 0);
2239 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2240 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2242 /* setup information for lnet_build_msg_event */
2243 msg->msg_from = peer_id.nid;
2244 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2245 msg->msg_hdr.src_nid = peer_id.nid;
2246 msg->msg_hdr.payload_length = getmd->md_length;
2247 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2249 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2250 lnet_res_unlock(cpt);
2252 cpt = lnet_cpt_of_nid(peer_id.nid);
2255 lnet_msg_commit(msg, cpt);
2256 lnet_net_unlock(cpt);
2258 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2263 cpt = lnet_cpt_of_nid(peer_id.nid);
2266 the_lnet.ln_counters[cpt]->drop_count++;
2267 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2268 lnet_net_unlock(cpt);
2275 EXPORT_SYMBOL(lnet_create_reply_msg);
2278 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2280 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2281 * completed and I know it. */
2282 LASSERT (reply != NULL);
2283 LASSERT (reply->msg_type == LNET_MSG_GET);
2284 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2286 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2287 * the end of my buffer, I might as well be dead. */
2288 LASSERT (len <= reply->msg_ev.mlength);
2290 reply->msg_ev.mlength = len;
2292 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2295 * Initiate an asynchronous GET operation.
2297 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2298 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2299 * the target node in the REPLY has been written to local MD.
2301 * On the target node, an LNET_EVENT_GET is logged when the GET request
2302 * arrives and is accepted into a MD.
2304 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2305 * \param mdh A handle for the MD that describes the memory into which the
2306 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2308 * \retval 0 Success, and only in this case events will be generated
2309 * and logged to EQ (if it exists) of the MD.
2310 * \retval -EIO Simulated failure.
2311 * \retval -ENOMEM Memory allocation failure.
2312 * \retval -ENOENT Invalid MD object.
2315 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2316 lnet_process_id_t target, unsigned int portal,
2317 __u64 match_bits, unsigned int offset)
2319 struct lnet_msg *msg;
2320 struct lnet_libmd *md;
2324 LASSERT (the_lnet.ln_init);
2325 LASSERT (the_lnet.ln_refcount > 0);
2327 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2328 fail_peer (target.nid, 1)) /* shall we now? */
2330 CERROR("Dropping GET to %s: simulated failure\n",
2331 libcfs_id2str(target));
2335 msg = lnet_msg_alloc();
2337 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2338 libcfs_id2str(target));
2342 cpt = lnet_cpt_of_cookie(mdh.cookie);
2345 md = lnet_handle2md(&mdh);
2346 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2347 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2348 match_bits, portal, libcfs_id2str(target),
2349 md == NULL ? -1 : md->md_threshold);
2350 if (md != NULL && md->md_me != NULL)
2351 CERROR("REPLY MD also attached to portal %d\n",
2352 md->md_me->me_portal);
2354 lnet_res_unlock(cpt);
2361 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2363 lnet_msg_attach_md(msg, md, 0, 0);
2365 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2367 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2368 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2369 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2370 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2372 /* NB handles only looked up by creator (no flips) */
2373 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2374 the_lnet.ln_interface_cookie;
2375 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2376 md->md_lh.lh_cookie;
2378 lnet_res_unlock(cpt);
2380 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2382 rc = lnet_send(self, msg, LNET_NID_ANY);
2384 CNETERR( "Error sending GET to %s: %d\n",
2385 libcfs_id2str(target), rc);
2386 lnet_finalize (NULL, msg, rc);
2389 /* completion will be signalled by an event */
2392 EXPORT_SYMBOL(LNetGet);
2395 * Calculate distance to node at \a dstnid.
2397 * \param dstnid Target NID.
2398 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2400 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2403 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2404 * local_nid_dist_zero is set, which is the default.
2405 * \retval positives Distance to target NID, i.e. number of hops plus one.
2406 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2409 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2413 lnet_remotenet_t *rnet;
2414 __u32 dstnet = LNET_NIDNET(dstnid);
2418 cfs_list_t *rn_list;
2420 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2421 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2422 * keep order 0 free for 0@lo and order 1 free for a local NID
2425 LASSERT (the_lnet.ln_init);
2426 LASSERT (the_lnet.ln_refcount > 0);
2428 cpt = lnet_net_lock_current();
2430 cfs_list_for_each (e, &the_lnet.ln_nis) {
2431 ni = cfs_list_entry(e, lnet_ni_t, ni_list);
2433 if (ni->ni_nid == dstnid) {
2434 if (srcnidp != NULL)
2436 if (orderp != NULL) {
2437 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2442 lnet_net_unlock(cpt);
2444 return local_nid_dist_zero ? 0 : 1;
2447 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2448 if (srcnidp != NULL)
2449 *srcnidp = ni->ni_nid;
2452 lnet_net_unlock(cpt);
2459 rn_list = lnet_net2rnethash(dstnet);
2460 cfs_list_for_each(e, rn_list) {
2461 rnet = cfs_list_entry(e, lnet_remotenet_t, lrn_list);
2463 if (rnet->lrn_net == dstnet) {
2464 lnet_route_t *route;
2465 lnet_route_t *shortest = NULL;
2467 LASSERT (!cfs_list_empty(&rnet->lrn_routes));
2469 cfs_list_for_each_entry(route, &rnet->lrn_routes,
2471 if (shortest == NULL ||
2472 route->lr_hops < shortest->lr_hops)
2476 LASSERT (shortest != NULL);
2477 hops = shortest->lr_hops;
2478 if (srcnidp != NULL)
2479 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2482 lnet_net_unlock(cpt);
2488 lnet_net_unlock(cpt);
2489 return -EHOSTUNREACH;
2491 EXPORT_SYMBOL(LNetDist);
2494 * Set the number of asynchronous messages expected from a target process.
2496 * This function is only meaningful for userspace callers. It's a no-op when
2497 * called from kernel.
2499 * Asynchronous messages are those that can come from a target when the
2500 * userspace process is not waiting for IO to complete; e.g., AST callbacks
2501 * from Lustre servers. Specifying the expected number of such messages
2502 * allows them to be eagerly received when user process is not running in
2503 * LNet; otherwise network errors may occur.
2505 * \param id Process ID of the target process.
2506 * \param nasync Number of asynchronous messages expected from the target.
2508 * \return 0 on success, and an error code otherwise.
2511 LNetSetAsync(lnet_process_id_t id, int nasync)
2517 lnet_remotenet_t *rnet;
2519 lnet_route_t *route;
2527 /* Target on a local network? */
2528 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2530 if (ni->ni_lnd->lnd_setasync != NULL)
2531 rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2536 /* Target on a remote network: apply to routers */
2538 LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2543 /* Snapshot all the router NIDs */
2544 cpt = lnet_net_lock_current();
2545 rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2547 cfs_list_for_each(tmp, &rnet->lrn_routes) {
2548 if (nnids == maxnids) {
2549 lnet_net_unlock(cpt);
2550 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2555 route = cfs_list_entry(tmp, lnet_route_t, lr_list);
2556 nids[nnids++] = route->lr_gateway->lp_nid;
2559 lnet_net_unlock(cpt);
2561 /* set async on all the routers */
2562 while (nnids-- > 0) {
2563 id.pid = LUSTRE_SRV_LNET_PID;
2564 id.nid = nids[nnids];
2566 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2570 if (ni->ni_lnd->lnd_setasync != NULL) {
2571 rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2578 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2582 EXPORT_SYMBOL(LNetSetAsync);