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 * \param msg The message to be sent.
821 * \param do_send True if lnet_ni_send() should be called in this function.
822 * lnet_send() is going to lnet_net_unlock immediately after this, so
823 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
825 * \retval 0 If \a msg sent or OK to send.
826 * \retval EAGAIN If \a msg blocked for credit.
827 * \retval EHOSTUNREACH If the next hop of the message appears dead.
828 * \retval ECANCELED If the MD of the message has been unlinked.
831 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
833 lnet_peer_t *lp = msg->msg_txpeer;
834 lnet_ni_t *ni = lp->lp_ni;
835 int cpt = msg->msg_tx_cpt;
836 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
838 /* non-lnet_send() callers have checked before */
839 LASSERT(!do_send || msg->msg_tx_delayed);
840 LASSERT(!msg->msg_receiving);
841 LASSERT(msg->msg_tx_committed);
843 /* NB 'lp' is always the next hop */
844 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
845 lnet_peer_alive_locked(lp) == 0) {
846 the_lnet.ln_counters[cpt]->drop_count++;
847 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
848 lnet_net_unlock(cpt);
850 CNETERR("Dropping message for %s: peer not alive\n",
851 libcfs_id2str(msg->msg_target));
853 lnet_finalize(ni, msg, -EHOSTUNREACH);
859 if (msg->msg_md != NULL &&
860 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
861 lnet_net_unlock(cpt);
863 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
864 "called on the MD/ME.\n",
865 libcfs_id2str(msg->msg_target));
867 lnet_finalize(ni, msg, -ECANCELED);
873 if (!msg->msg_peertxcredit) {
874 LASSERT ((lp->lp_txcredits < 0) ==
875 !cfs_list_empty(&lp->lp_txq));
877 msg->msg_peertxcredit = 1;
878 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
881 if (lp->lp_txcredits < lp->lp_mintxcredits)
882 lp->lp_mintxcredits = lp->lp_txcredits;
884 if (lp->lp_txcredits < 0) {
885 msg->msg_tx_delayed = 1;
886 cfs_list_add_tail(&msg->msg_list, &lp->lp_txq);
891 if (!msg->msg_txcredit) {
892 LASSERT((tq->tq_credits < 0) ==
893 !cfs_list_empty(&tq->tq_delayed));
895 msg->msg_txcredit = 1;
898 if (tq->tq_credits < tq->tq_credits_min)
899 tq->tq_credits_min = tq->tq_credits;
901 if (tq->tq_credits < 0) {
902 msg->msg_tx_delayed = 1;
903 cfs_list_add_tail(&msg->msg_list, &tq->tq_delayed);
909 lnet_net_unlock(cpt);
910 lnet_ni_send(ni, msg);
919 lnet_msg2bufpool(lnet_msg_t *msg)
921 lnet_rtrbufpool_t *rbp;
924 LASSERT(msg->msg_rx_committed);
926 cpt = msg->msg_rx_cpt;
927 rbp = &the_lnet.ln_rtrpools[cpt][0];
929 LASSERT(msg->msg_len <= LNET_MTU);
930 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_CACHE_SIZE) {
932 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
939 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
941 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
942 * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
943 * return EAGAIN if msg blocked and 0 if received or OK to receive */
944 lnet_peer_t *lp = msg->msg_rxpeer;
945 lnet_rtrbufpool_t *rbp;
948 LASSERT (msg->msg_iov == NULL);
949 LASSERT (msg->msg_kiov == NULL);
950 LASSERT (msg->msg_niov == 0);
951 LASSERT (msg->msg_routing);
952 LASSERT (msg->msg_receiving);
953 LASSERT (!msg->msg_sending);
955 /* non-lnet_parse callers only receive delayed messages */
956 LASSERT(!do_recv || msg->msg_rx_delayed);
958 if (!msg->msg_peerrtrcredit) {
959 LASSERT ((lp->lp_rtrcredits < 0) ==
960 !cfs_list_empty(&lp->lp_rtrq));
962 msg->msg_peerrtrcredit = 1;
964 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
965 lp->lp_minrtrcredits = lp->lp_rtrcredits;
967 if (lp->lp_rtrcredits < 0) {
968 /* must have checked eager_recv before here */
969 LASSERT(msg->msg_rx_ready_delay);
970 msg->msg_rx_delayed = 1;
971 cfs_list_add_tail(&msg->msg_list, &lp->lp_rtrq);
976 rbp = lnet_msg2bufpool(msg);
978 if (!msg->msg_rtrcredit) {
979 LASSERT ((rbp->rbp_credits < 0) ==
980 !cfs_list_empty(&rbp->rbp_msgs));
982 msg->msg_rtrcredit = 1;
984 if (rbp->rbp_credits < rbp->rbp_mincredits)
985 rbp->rbp_mincredits = rbp->rbp_credits;
987 if (rbp->rbp_credits < 0) {
988 /* must have checked eager_recv before here */
989 LASSERT(msg->msg_rx_ready_delay);
990 msg->msg_rx_delayed = 1;
991 cfs_list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
996 LASSERT (!cfs_list_empty(&rbp->rbp_bufs));
997 rb = cfs_list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
998 cfs_list_del(&rb->rb_list);
1000 msg->msg_niov = rbp->rbp_npages;
1001 msg->msg_kiov = &rb->rb_kiov[0];
1004 int cpt = msg->msg_rx_cpt;
1006 lnet_net_unlock(cpt);
1007 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
1008 0, msg->msg_len, msg->msg_len);
1016 lnet_return_tx_credits_locked(lnet_msg_t *msg)
1018 lnet_peer_t *txpeer = msg->msg_txpeer;
1021 if (msg->msg_txcredit) {
1022 struct lnet_ni *ni = txpeer->lp_ni;
1023 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1025 /* give back NI txcredits */
1026 msg->msg_txcredit = 0;
1028 LASSERT((tq->tq_credits < 0) ==
1029 !cfs_list_empty(&tq->tq_delayed));
1032 if (tq->tq_credits <= 0) {
1033 msg2 = cfs_list_entry(tq->tq_delayed.next,
1034 lnet_msg_t, msg_list);
1035 cfs_list_del(&msg2->msg_list);
1037 LASSERT(msg2->msg_txpeer->lp_ni == ni);
1038 LASSERT(msg2->msg_tx_delayed);
1040 (void) lnet_post_send_locked(msg2, 1);
1044 if (msg->msg_peertxcredit) {
1045 /* give back peer txcredits */
1046 msg->msg_peertxcredit = 0;
1048 LASSERT((txpeer->lp_txcredits < 0) ==
1049 !cfs_list_empty(&txpeer->lp_txq));
1051 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1052 LASSERT (txpeer->lp_txqnob >= 0);
1054 txpeer->lp_txcredits++;
1055 if (txpeer->lp_txcredits <= 0) {
1056 msg2 = cfs_list_entry(txpeer->lp_txq.next,
1057 lnet_msg_t, msg_list);
1058 cfs_list_del(&msg2->msg_list);
1060 LASSERT(msg2->msg_txpeer == txpeer);
1061 LASSERT(msg2->msg_tx_delayed);
1063 (void) lnet_post_send_locked(msg2, 1);
1067 if (txpeer != NULL) {
1068 msg->msg_txpeer = NULL;
1069 lnet_peer_decref_locked(txpeer);
1074 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1076 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1080 if (msg->msg_rtrcredit) {
1081 /* give back global router credits */
1083 lnet_rtrbufpool_t *rbp;
1085 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1086 * there until it gets one allocated, or aborts the wait
1088 LASSERT (msg->msg_kiov != NULL);
1090 rb = cfs_list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1092 LASSERT (rbp == lnet_msg2bufpool(msg));
1094 msg->msg_kiov = NULL;
1095 msg->msg_rtrcredit = 0;
1097 LASSERT((rbp->rbp_credits < 0) ==
1098 !cfs_list_empty(&rbp->rbp_msgs));
1099 LASSERT((rbp->rbp_credits > 0) ==
1100 !cfs_list_empty(&rbp->rbp_bufs));
1102 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1104 if (rbp->rbp_credits <= 0) {
1105 msg2 = cfs_list_entry(rbp->rbp_msgs.next,
1106 lnet_msg_t, msg_list);
1107 cfs_list_del(&msg2->msg_list);
1109 (void) lnet_post_routed_recv_locked(msg2, 1);
1113 if (msg->msg_peerrtrcredit) {
1114 /* give back peer router credits */
1115 msg->msg_peerrtrcredit = 0;
1117 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1118 !cfs_list_empty(&rxpeer->lp_rtrq));
1120 rxpeer->lp_rtrcredits++;
1121 if (rxpeer->lp_rtrcredits <= 0) {
1122 msg2 = cfs_list_entry(rxpeer->lp_rtrq.next,
1123 lnet_msg_t, msg_list);
1124 cfs_list_del(&msg2->msg_list);
1126 (void) lnet_post_routed_recv_locked(msg2, 1);
1130 LASSERT (!msg->msg_rtrcredit);
1131 LASSERT (!msg->msg_peerrtrcredit);
1133 if (rxpeer != NULL) {
1134 msg->msg_rxpeer = NULL;
1135 lnet_peer_decref_locked(rxpeer);
1140 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1142 lnet_peer_t *p1 = r1->lr_gateway;
1143 lnet_peer_t *p2 = r2->lr_gateway;
1145 if (r1->lr_priority < r2->lr_priority)
1148 if (r1->lr_priority > r2->lr_priority)
1151 if (r1->lr_hops < r2->lr_hops)
1154 if (r1->lr_hops > r2->lr_hops)
1157 if (p1->lp_txqnob < p2->lp_txqnob)
1160 if (p1->lp_txqnob > p2->lp_txqnob)
1163 if (p1->lp_txcredits > p2->lp_txcredits)
1166 if (p1->lp_txcredits < p2->lp_txcredits)
1169 if (r1->lr_seq - r2->lr_seq <= 0)
1175 static lnet_peer_t *
1176 lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target, lnet_nid_t rtr_nid)
1178 lnet_remotenet_t *rnet;
1179 lnet_route_t *route;
1180 lnet_route_t *best_route;
1181 lnet_route_t *last_route;
1182 struct lnet_peer *lp_best;
1183 struct lnet_peer *lp;
1186 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1187 * rtr_nid nid, otherwise find the best gateway I can use */
1189 rnet = lnet_find_net_locked(LNET_NIDNET(target));
1194 best_route = last_route = NULL;
1195 cfs_list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1196 lp = route->lr_gateway;
1198 if (!lnet_is_route_alive(route))
1201 if (ni != NULL && lp->lp_ni != ni)
1204 if (lp->lp_nid == rtr_nid) /* it's pre-determined router */
1207 if (lp_best == NULL) {
1208 best_route = last_route = route;
1213 /* no protection on below fields, but it's harmless */
1214 if (last_route->lr_seq - route->lr_seq < 0)
1217 rc = lnet_compare_routes(route, best_route);
1225 /* set sequence number on the best router to the latest sequence + 1
1226 * so we can round-robin all routers, it's race and inaccurate but
1227 * harmless and functional */
1228 if (best_route != NULL)
1229 best_route->lr_seq = last_route->lr_seq + 1;
1234 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1236 lnet_nid_t dst_nid = msg->msg_target.nid;
1237 struct lnet_ni *src_ni;
1238 struct lnet_ni *local_ni;
1239 struct lnet_peer *lp;
1244 /* NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1245 * but we might want to use pre-determined router for ACK/REPLY
1247 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1248 LASSERT (msg->msg_txpeer == NULL);
1249 LASSERT (!msg->msg_sending);
1250 LASSERT (!msg->msg_target_is_router);
1251 LASSERT (!msg->msg_receiving);
1253 msg->msg_sending = 1;
1255 LASSERT(!msg->msg_tx_committed);
1256 cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid);
1260 if (the_lnet.ln_shutdown) {
1261 lnet_net_unlock(cpt);
1265 if (src_nid == LNET_NID_ANY) {
1268 src_ni = lnet_nid2ni_locked(src_nid, cpt);
1269 if (src_ni == NULL) {
1270 lnet_net_unlock(cpt);
1271 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1272 "local nid\n", libcfs_nid2str(dst_nid),
1273 libcfs_nid2str(src_nid));
1276 LASSERT (!msg->msg_routing);
1279 /* Is this for someone on a local network? */
1280 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt);
1282 if (local_ni != NULL) {
1283 if (src_ni == NULL) {
1285 src_nid = src_ni->ni_nid;
1286 } else if (src_ni == local_ni) {
1287 lnet_ni_decref_locked(local_ni, cpt);
1289 lnet_ni_decref_locked(local_ni, cpt);
1290 lnet_ni_decref_locked(src_ni, cpt);
1291 lnet_net_unlock(cpt);
1292 LCONSOLE_WARN("No route to %s via from %s\n",
1293 libcfs_nid2str(dst_nid),
1294 libcfs_nid2str(src_nid));
1298 LASSERT(src_nid != LNET_NID_ANY);
1299 lnet_msg_commit(msg, cpt);
1301 if (!msg->msg_routing)
1302 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1304 if (src_ni == the_lnet.ln_loni) {
1305 /* No send credit hassles with LOLND */
1306 lnet_net_unlock(cpt);
1307 lnet_ni_send(src_ni, msg);
1310 lnet_ni_decref_locked(src_ni, cpt);
1311 lnet_net_unlock(cpt);
1315 rc = lnet_nid2peer_locked(&lp, dst_nid, cpt);
1316 /* lp has ref on src_ni; lose mine */
1317 lnet_ni_decref_locked(src_ni, cpt);
1319 lnet_net_unlock(cpt);
1320 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1321 libcfs_nid2str(dst_nid));
1322 /* ENOMEM or shutting down */
1325 LASSERT (lp->lp_ni == src_ni);
1328 lnet_net_unlock(cpt);
1331 * - once application finishes computation, check here to update
1332 * router states before it waits for pending IO in LNetEQPoll
1333 * - recursion breaker: router checker sends no message
1334 * to remote networks */
1335 if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING)
1336 lnet_router_checker();
1340 /* sending to a remote network */
1341 lp = lnet_find_route_locked(src_ni, dst_nid, rtr_nid);
1344 lnet_ni_decref_locked(src_ni, cpt);
1345 lnet_net_unlock(cpt);
1347 LCONSOLE_WARN("No route to %s via %s "
1348 "(all routers down)\n",
1349 libcfs_id2str(msg->msg_target),
1350 libcfs_nid2str(src_nid));
1351 return -EHOSTUNREACH;
1354 /* rtr_nid is LNET_NID_ANY or NID of pre-determined router,
1355 * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't
1356 * pre-determined router, this can happen if router table
1357 * was changed when we release the lock */
1358 if (rtr_nid != lp->lp_nid) {
1359 cpt2 = lnet_cpt_of_nid_locked(lp->lp_nid);
1362 lnet_ni_decref_locked(src_ni, cpt);
1363 lnet_net_unlock(cpt);
1365 rtr_nid = lp->lp_nid;
1371 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1372 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1373 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1375 if (src_ni == NULL) {
1377 src_nid = src_ni->ni_nid;
1379 LASSERT (src_ni == lp->lp_ni);
1380 lnet_ni_decref_locked(src_ni, cpt);
1383 lnet_peer_addref_locked(lp);
1385 LASSERT(src_nid != LNET_NID_ANY);
1386 lnet_msg_commit(msg, cpt);
1388 if (!msg->msg_routing) {
1389 /* I'm the source and now I know which NI to send on */
1390 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1393 msg->msg_target_is_router = 1;
1394 msg->msg_target.nid = lp->lp_nid;
1395 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1398 /* 'lp' is our best choice of peer */
1400 LASSERT (!msg->msg_peertxcredit);
1401 LASSERT (!msg->msg_txcredit);
1402 LASSERT (msg->msg_txpeer == NULL);
1404 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1406 rc = lnet_post_send_locked(msg, 0);
1407 lnet_net_unlock(cpt);
1409 if (rc == EHOSTUNREACH || rc == ECANCELED)
1413 lnet_ni_send(src_ni, msg);
1415 return 0; /* rc == 0 or EAGAIN */
1419 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1422 the_lnet.ln_counters[cpt]->drop_count++;
1423 the_lnet.ln_counters[cpt]->drop_length += nob;
1424 lnet_net_unlock(cpt);
1426 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1430 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1432 lnet_hdr_t *hdr = &msg->msg_hdr;
1434 if (msg->msg_wanted != 0)
1435 lnet_setpayloadbuffer(msg);
1437 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1439 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1440 * it back into the ACK during lnet_finalize() */
1441 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1442 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1444 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1445 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1449 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1451 lnet_hdr_t *hdr = &msg->msg_hdr;
1452 struct lnet_match_info info;
1455 /* Convert put fields to host byte order */
1456 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1457 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1458 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1460 info.mi_id.nid = hdr->src_nid;
1461 info.mi_id.pid = hdr->src_pid;
1462 info.mi_opc = LNET_MD_OP_PUT;
1463 info.mi_portal = hdr->msg.put.ptl_index;
1464 info.mi_rlength = hdr->payload_length;
1465 info.mi_roffset = hdr->msg.put.offset;
1466 info.mi_mbits = hdr->msg.put.match_bits;
1468 msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL;
1471 rc = lnet_ptl_match_md(&info, msg);
1476 case LNET_MATCHMD_OK:
1477 lnet_recv_put(ni, msg);
1480 case LNET_MATCHMD_NONE:
1481 if (msg->msg_rx_delayed) /* attached on delayed list */
1484 rc = lnet_ni_eager_recv(ni, msg);
1489 case LNET_MATCHMD_DROP:
1490 CNETERR("Dropping PUT from %s portal %d match "LPU64
1491 " offset %d length %d: %d\n",
1492 libcfs_id2str(info.mi_id), info.mi_portal,
1493 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1495 return ENOENT; /* +ve: OK but no match */
1500 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1502 struct lnet_match_info info;
1503 lnet_hdr_t *hdr = &msg->msg_hdr;
1504 lnet_handle_wire_t reply_wmd;
1507 /* Convert get fields to host byte order */
1508 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1509 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1510 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1511 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1513 info.mi_id.nid = hdr->src_nid;
1514 info.mi_id.pid = hdr->src_pid;
1515 info.mi_opc = LNET_MD_OP_GET;
1516 info.mi_portal = hdr->msg.get.ptl_index;
1517 info.mi_rlength = hdr->msg.get.sink_length;
1518 info.mi_roffset = hdr->msg.get.src_offset;
1519 info.mi_mbits = hdr->msg.get.match_bits;
1521 rc = lnet_ptl_match_md(&info, msg);
1522 if (rc == LNET_MATCHMD_DROP) {
1523 CNETERR("Dropping GET from %s portal %d match "LPU64
1524 " offset %d length %d\n",
1525 libcfs_id2str(info.mi_id), info.mi_portal,
1526 info.mi_mbits, info.mi_roffset, info.mi_rlength);
1527 return ENOENT; /* +ve: OK but no match */
1530 LASSERT(rc == LNET_MATCHMD_OK);
1532 lnet_build_msg_event(msg, LNET_EVENT_GET);
1534 reply_wmd = hdr->msg.get.return_wmd;
1536 lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
1537 msg->msg_offset, msg->msg_wanted);
1539 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1542 /* The LND completes the REPLY from her recv procedure */
1543 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1544 msg->msg_offset, msg->msg_len, msg->msg_len);
1548 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1549 msg->msg_receiving = 0;
1551 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
1553 /* didn't get as far as lnet_ni_send() */
1554 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1555 libcfs_nid2str(ni->ni_nid),
1556 libcfs_id2str(info.mi_id), rc);
1558 lnet_finalize(ni, msg, rc);
1565 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1567 void *private = msg->msg_private;
1568 lnet_hdr_t *hdr = &msg->msg_hdr;
1569 lnet_process_id_t src = {0};
1575 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
1578 src.nid = hdr->src_nid;
1579 src.pid = hdr->src_pid;
1581 /* NB handles only looked up by creator (no flips) */
1582 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1583 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1584 CNETERR("%s: Dropping REPLY from %s for %s "
1585 "MD "LPX64"."LPX64"\n",
1586 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1587 (md == NULL) ? "invalid" : "inactive",
1588 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1589 hdr->msg.reply.dst_wmd.wh_object_cookie);
1590 if (md != NULL && md->md_me != NULL)
1591 CERROR("REPLY MD also attached to portal %d\n",
1592 md->md_me->me_portal);
1594 lnet_res_unlock(cpt);
1595 return ENOENT; /* +ve: OK but no match */
1598 LASSERT (md->md_offset == 0);
1600 rlength = hdr->payload_length;
1601 mlength = MIN(rlength, (int)md->md_length);
1603 if (mlength < rlength &&
1604 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1605 CNETERR("%s: Dropping REPLY from %s length %d "
1606 "for MD "LPX64" would overflow (%d)\n",
1607 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1608 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1610 lnet_res_unlock(cpt);
1611 return ENOENT; /* +ve: OK but no match */
1614 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1615 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1616 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1618 lnet_msg_attach_md(msg, md, 0, mlength);
1621 lnet_setpayloadbuffer(msg);
1623 lnet_res_unlock(cpt);
1625 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
1627 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1632 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1634 lnet_hdr_t *hdr = &msg->msg_hdr;
1635 lnet_process_id_t src = {0};
1639 src.nid = hdr->src_nid;
1640 src.pid = hdr->src_pid;
1642 /* Convert ack fields to host byte order */
1643 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1644 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1646 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
1649 /* NB handles only looked up by creator (no flips) */
1650 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1651 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1652 /* Don't moan; this is expected */
1654 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1655 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1656 (md == NULL) ? "invalid" : "inactive",
1657 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1658 hdr->msg.ack.dst_wmd.wh_object_cookie);
1659 if (md != NULL && md->md_me != NULL)
1660 CERROR("Source MD also attached to portal %d\n",
1661 md->md_me->me_portal);
1663 lnet_res_unlock(cpt);
1664 return ENOENT; /* +ve! */
1667 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1668 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1669 hdr->msg.ack.dst_wmd.wh_object_cookie);
1671 lnet_msg_attach_md(msg, md, 0, 0);
1673 lnet_res_unlock(cpt);
1675 lnet_build_msg_event(msg, LNET_EVENT_ACK);
1677 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
1682 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
1687 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
1688 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
1689 if (ni->ni_lnd->lnd_eager_recv == NULL) {
1690 msg->msg_rx_ready_delay = 1;
1692 lnet_net_unlock(msg->msg_rx_cpt);
1693 rc = lnet_ni_eager_recv(ni, msg);
1694 lnet_net_lock(msg->msg_rx_cpt);
1699 rc = lnet_post_routed_recv_locked(msg, 0);
1707 lnet_msgtyp2str (int type)
1716 case LNET_MSG_REPLY:
1718 case LNET_MSG_HELLO:
1721 return ("<UNKNOWN>");
1724 EXPORT_SYMBOL(lnet_msgtyp2str);
1727 lnet_print_hdr(lnet_hdr_t * hdr)
1729 lnet_process_id_t src = {0};
1730 lnet_process_id_t dst = {0};
1731 char *type_str = lnet_msgtyp2str (hdr->type);
1733 src.nid = hdr->src_nid;
1734 src.pid = hdr->src_pid;
1736 dst.nid = hdr->dest_nid;
1737 dst.pid = hdr->dest_pid;
1739 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
1740 CWARN(" From %s\n", libcfs_id2str(src));
1741 CWARN(" To %s\n", libcfs_id2str(dst));
1743 switch (hdr->type) {
1748 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
1749 "match bits "LPU64"\n",
1750 hdr->msg.put.ptl_index,
1751 hdr->msg.put.ack_wmd.wh_interface_cookie,
1752 hdr->msg.put.ack_wmd.wh_object_cookie,
1753 hdr->msg.put.match_bits);
1754 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
1755 hdr->payload_length, hdr->msg.put.offset,
1756 hdr->msg.put.hdr_data);
1760 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
1761 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
1762 hdr->msg.get.return_wmd.wh_interface_cookie,
1763 hdr->msg.get.return_wmd.wh_object_cookie,
1764 hdr->msg.get.match_bits);
1765 CWARN(" Length %d, src offset %d\n",
1766 hdr->msg.get.sink_length,
1767 hdr->msg.get.src_offset);
1771 CWARN(" dst md "LPX64"."LPX64", "
1772 "manipulated length %d\n",
1773 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1774 hdr->msg.ack.dst_wmd.wh_object_cookie,
1775 hdr->msg.ack.mlength);
1778 case LNET_MSG_REPLY:
1779 CWARN(" dst md "LPX64"."LPX64", "
1781 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1782 hdr->msg.reply.dst_wmd.wh_object_cookie,
1783 hdr->payload_length);
1789 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
1790 void *private, int rdma_req)
1795 struct lnet_msg *msg;
1796 lnet_pid_t dest_pid;
1797 lnet_nid_t dest_nid;
1799 __u32 payload_length;
1802 LASSERT (!in_interrupt ());
1804 type = le32_to_cpu(hdr->type);
1805 src_nid = le64_to_cpu(hdr->src_nid);
1806 dest_nid = le64_to_cpu(hdr->dest_nid);
1807 dest_pid = le32_to_cpu(hdr->dest_pid);
1808 payload_length = le32_to_cpu(hdr->payload_length);
1810 for_me = (ni->ni_nid == dest_nid);
1811 cpt = lnet_cpt_of_nid(from_nid);
1816 if (payload_length > 0) {
1817 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
1818 libcfs_nid2str(from_nid),
1819 libcfs_nid2str(src_nid),
1820 lnet_msgtyp2str(type), payload_length);
1826 case LNET_MSG_REPLY:
1827 if (payload_length > (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
1828 CERROR("%s, src %s: bad %s payload %d "
1829 "(%d max expected)\n",
1830 libcfs_nid2str(from_nid),
1831 libcfs_nid2str(src_nid),
1832 lnet_msgtyp2str(type),
1834 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
1840 CERROR("%s, src %s: Bad message type 0x%x\n",
1841 libcfs_nid2str(from_nid),
1842 libcfs_nid2str(src_nid), type);
1846 if (the_lnet.ln_routing &&
1847 ni->ni_last_alive != cfs_time_current_sec()) {
1850 /* NB: so far here is the only place to set NI status to "up */
1851 ni->ni_last_alive = cfs_time_current_sec();
1852 if (ni->ni_status != NULL &&
1853 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
1854 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
1858 /* Regard a bad destination NID as a protocol error. Senders should
1859 * know what they're doing; if they don't they're misconfigured, buggy
1860 * or malicious so we chop them off at the knees :) */
1863 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
1864 /* should have gone direct */
1865 CERROR ("%s, src %s: Bad dest nid %s "
1866 "(should have been sent direct)\n",
1867 libcfs_nid2str(from_nid),
1868 libcfs_nid2str(src_nid),
1869 libcfs_nid2str(dest_nid));
1873 if (lnet_islocalnid(dest_nid)) {
1874 /* dest is another local NI; sender should have used
1875 * this node's NID on its own network */
1876 CERROR ("%s, src %s: Bad dest nid %s "
1877 "(it's my nid but on a different network)\n",
1878 libcfs_nid2str(from_nid),
1879 libcfs_nid2str(src_nid),
1880 libcfs_nid2str(dest_nid));
1884 if (rdma_req && type == LNET_MSG_GET) {
1885 CERROR ("%s, src %s: Bad optimized GET for %s "
1886 "(final destination must be me)\n",
1887 libcfs_nid2str(from_nid),
1888 libcfs_nid2str(src_nid),
1889 libcfs_nid2str(dest_nid));
1893 if (!the_lnet.ln_routing) {
1894 CERROR ("%s, src %s: Dropping message for %s "
1895 "(routing not enabled)\n",
1896 libcfs_nid2str(from_nid),
1897 libcfs_nid2str(src_nid),
1898 libcfs_nid2str(dest_nid));
1903 /* Message looks OK; we're not going to return an error, so we MUST
1904 * call back lnd_recv() come what may... */
1906 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
1907 fail_peer (src_nid, 0)) /* shall we now? */
1909 CERROR("%s, src %s: Dropping %s to simulate failure\n",
1910 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1911 lnet_msgtyp2str(type));
1915 msg = lnet_msg_alloc();
1917 CERROR("%s, src %s: Dropping %s (out of memory)\n",
1918 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1919 lnet_msgtyp2str(type));
1923 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
1925 msg->msg_type = type;
1926 msg->msg_private = private;
1927 msg->msg_receiving = 1;
1928 msg->msg_len = msg->msg_wanted = payload_length;
1929 msg->msg_offset = 0;
1930 msg->msg_hdr = *hdr;
1931 /* for building message event */
1932 msg->msg_from = from_nid;
1934 msg->msg_target.pid = dest_pid;
1935 msg->msg_target.nid = dest_nid;
1936 msg->msg_routing = 1;
1939 /* convert common msg->hdr fields to host byteorder */
1940 msg->msg_hdr.type = type;
1941 msg->msg_hdr.src_nid = src_nid;
1942 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
1943 msg->msg_hdr.dest_nid = dest_nid;
1944 msg->msg_hdr.dest_pid = dest_pid;
1945 msg->msg_hdr.payload_length = payload_length;
1949 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid, cpt);
1951 lnet_net_unlock(cpt);
1952 CERROR("%s, src %s: Dropping %s "
1953 "(error %d looking up sender)\n",
1954 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1955 lnet_msgtyp2str(type), rc);
1960 lnet_msg_commit(msg, cpt);
1963 rc = lnet_parse_forward_locked(ni, msg);
1964 lnet_net_unlock(cpt);
1969 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1970 0, payload_length, payload_length);
1975 lnet_net_unlock(cpt);
1979 rc = lnet_parse_ack(ni, msg);
1982 rc = lnet_parse_put(ni, msg);
1985 rc = lnet_parse_get(ni, msg, rdma_req);
1987 case LNET_MSG_REPLY:
1988 rc = lnet_parse_reply(ni, msg);
1993 goto free_drop; /* prevent an unused label if !kernel */
1999 LASSERT (rc == ENOENT);
2002 LASSERT(msg->msg_md == NULL);
2003 lnet_finalize(ni, msg, rc);
2006 lnet_drop_message(ni, cpt, private, payload_length);
2009 EXPORT_SYMBOL(lnet_parse);
2012 lnet_drop_delayed_msg_list(cfs_list_t *head, char *reason)
2014 while (!cfs_list_empty(head)) {
2015 lnet_process_id_t id = {0};
2018 msg = cfs_list_entry(head->next, lnet_msg_t, msg_list);
2019 cfs_list_del(&msg->msg_list);
2021 id.nid = msg->msg_hdr.src_nid;
2022 id.pid = msg->msg_hdr.src_pid;
2024 LASSERT(msg->msg_md == NULL);
2025 LASSERT(msg->msg_rx_delayed);
2026 LASSERT(msg->msg_rxpeer != NULL);
2027 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2029 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
2030 " offset %d length %d: %s\n",
2032 msg->msg_hdr.msg.put.ptl_index,
2033 msg->msg_hdr.msg.put.match_bits,
2034 msg->msg_hdr.msg.put.offset,
2035 msg->msg_hdr.payload_length, reason);
2037 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2038 * called lnet_drop_message(), so I just hang onto msg as well
2039 * until that's done */
2041 lnet_drop_message(msg->msg_rxpeer->lp_ni,
2042 msg->msg_rxpeer->lp_cpt,
2043 msg->msg_private, msg->msg_len);
2045 * NB: message will not generate event because w/o attached MD,
2046 * but we still should give error code so lnet_msg_decommit()
2047 * can skip counters operations and other checks.
2049 lnet_finalize(msg->msg_rxpeer->lp_ni, msg, -ENOENT);
2054 lnet_recv_delayed_msg_list(cfs_list_t *head)
2056 while (!cfs_list_empty(head)) {
2058 lnet_process_id_t id;
2060 msg = cfs_list_entry(head->next, lnet_msg_t, msg_list);
2061 cfs_list_del(&msg->msg_list);
2063 /* md won't disappear under me, since each msg
2064 * holds a ref on it */
2066 id.nid = msg->msg_hdr.src_nid;
2067 id.pid = msg->msg_hdr.src_pid;
2069 LASSERT(msg->msg_rx_delayed);
2070 LASSERT(msg->msg_md != NULL);
2071 LASSERT(msg->msg_rxpeer != NULL);
2072 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2074 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2075 "match "LPU64" offset %d length %d.\n",
2076 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2077 msg->msg_hdr.msg.put.match_bits,
2078 msg->msg_hdr.msg.put.offset,
2079 msg->msg_hdr.payload_length);
2081 lnet_recv_put(msg->msg_rxpeer->lp_ni, msg);
2086 * Initiate an asynchronous PUT operation.
2088 * There are several events associated with a PUT: completion of the send on
2089 * the initiator node (LNET_EVENT_SEND), and when the send completes
2090 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2091 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2092 * used at the target node to indicate the completion of incoming data
2095 * The local events will be logged in the EQ associated with the MD pointed to
2096 * by \a mdh handle. Using a MD without an associated EQ results in these
2097 * events being discarded. In this case, the caller must have another
2098 * mechanism (e.g., a higher level protocol) for determining when it is safe
2099 * to modify the memory region associated with the MD.
2101 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2102 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2104 * \param self Indicates the NID of a local interface through which to send
2105 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2106 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2107 * must be "free floating" (See LNetMDBind()).
2108 * \param ack Controls whether an acknowledgment is requested.
2109 * Acknowledgments are only sent when they are requested by the initiating
2110 * process and the target MD enables them.
2111 * \param target A process identifier for the target process.
2112 * \param portal The index in the \a target's portal table.
2113 * \param match_bits The match bits to use for MD selection at the target
2115 * \param offset The offset into the target MD (only used when the target
2116 * MD has the LNET_MD_MANAGE_REMOTE option set).
2117 * \param hdr_data 64 bits of user data that can be included in the message
2118 * header. This data is written to an event queue entry at the target if an
2119 * EQ is present on the matching MD.
2121 * \retval 0 Success, and only in this case events will be generated
2122 * and logged to EQ (if it exists).
2123 * \retval -EIO Simulated failure.
2124 * \retval -ENOMEM Memory allocation failure.
2125 * \retval -ENOENT Invalid MD object.
2127 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2130 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2131 lnet_process_id_t target, unsigned int portal,
2132 __u64 match_bits, unsigned int offset,
2135 struct lnet_msg *msg;
2136 struct lnet_libmd *md;
2140 LASSERT (the_lnet.ln_init);
2141 LASSERT (the_lnet.ln_refcount > 0);
2143 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2144 fail_peer (target.nid, 1)) /* shall we now? */
2146 CERROR("Dropping PUT to %s: simulated failure\n",
2147 libcfs_id2str(target));
2151 msg = lnet_msg_alloc();
2153 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2154 libcfs_id2str(target));
2157 msg->msg_vmflush = !!memory_pressure_get();
2159 cpt = lnet_cpt_of_cookie(mdh.cookie);
2162 md = lnet_handle2md(&mdh);
2163 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2164 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2165 match_bits, portal, libcfs_id2str(target),
2166 md == NULL ? -1 : md->md_threshold);
2167 if (md != NULL && md->md_me != NULL)
2168 CERROR("Source MD also attached to portal %d\n",
2169 md->md_me->me_portal);
2170 lnet_res_unlock(cpt);
2176 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2178 lnet_msg_attach_md(msg, md, 0, 0);
2180 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2182 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2183 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2184 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2185 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2187 /* NB handles only looked up by creator (no flips) */
2188 if (ack == LNET_ACK_REQ) {
2189 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2190 the_lnet.ln_interface_cookie;
2191 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2192 md->md_lh.lh_cookie;
2194 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2195 LNET_WIRE_HANDLE_COOKIE_NONE;
2196 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2197 LNET_WIRE_HANDLE_COOKIE_NONE;
2200 lnet_res_unlock(cpt);
2202 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2204 rc = lnet_send(self, msg, LNET_NID_ANY);
2206 CNETERR( "Error sending PUT to %s: %d\n",
2207 libcfs_id2str(target), rc);
2208 lnet_finalize (NULL, msg, rc);
2211 /* completion will be signalled by an event */
2214 EXPORT_SYMBOL(LNetPut);
2217 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2219 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2220 * returns a msg for the LND to pass to lnet_finalize() when the sink
2221 * data has been received.
2223 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2224 * lnet_finalize() is called on it, so the LND must call this first */
2226 struct lnet_msg *msg = lnet_msg_alloc();
2227 struct lnet_libmd *getmd = getmsg->msg_md;
2228 lnet_process_id_t peer_id = getmsg->msg_target;
2231 LASSERT(!getmsg->msg_target_is_router);
2232 LASSERT(!getmsg->msg_routing);
2235 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2236 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2240 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2243 LASSERT(getmd->md_refcount > 0);
2245 if (getmd->md_threshold == 0) {
2246 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2247 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2249 lnet_res_unlock(cpt);
2253 LASSERT(getmd->md_offset == 0);
2255 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2256 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2258 /* setup information for lnet_build_msg_event */
2259 msg->msg_from = peer_id.nid;
2260 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2261 msg->msg_hdr.src_nid = peer_id.nid;
2262 msg->msg_hdr.payload_length = getmd->md_length;
2263 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2265 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2266 lnet_res_unlock(cpt);
2268 cpt = lnet_cpt_of_nid(peer_id.nid);
2271 lnet_msg_commit(msg, cpt);
2272 lnet_net_unlock(cpt);
2274 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2279 cpt = lnet_cpt_of_nid(peer_id.nid);
2282 the_lnet.ln_counters[cpt]->drop_count++;
2283 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2284 lnet_net_unlock(cpt);
2291 EXPORT_SYMBOL(lnet_create_reply_msg);
2294 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2296 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2297 * completed and I know it. */
2298 LASSERT (reply != NULL);
2299 LASSERT (reply->msg_type == LNET_MSG_GET);
2300 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2302 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2303 * the end of my buffer, I might as well be dead. */
2304 LASSERT (len <= reply->msg_ev.mlength);
2306 reply->msg_ev.mlength = len;
2308 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2311 * Initiate an asynchronous GET operation.
2313 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2314 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2315 * the target node in the REPLY has been written to local MD.
2317 * On the target node, an LNET_EVENT_GET is logged when the GET request
2318 * arrives and is accepted into a MD.
2320 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2321 * \param mdh A handle for the MD that describes the memory into which the
2322 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2324 * \retval 0 Success, and only in this case events will be generated
2325 * and logged to EQ (if it exists) of the MD.
2326 * \retval -EIO Simulated failure.
2327 * \retval -ENOMEM Memory allocation failure.
2328 * \retval -ENOENT Invalid MD object.
2331 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2332 lnet_process_id_t target, unsigned int portal,
2333 __u64 match_bits, unsigned int offset)
2335 struct lnet_msg *msg;
2336 struct lnet_libmd *md;
2340 LASSERT (the_lnet.ln_init);
2341 LASSERT (the_lnet.ln_refcount > 0);
2343 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2344 fail_peer (target.nid, 1)) /* shall we now? */
2346 CERROR("Dropping GET to %s: simulated failure\n",
2347 libcfs_id2str(target));
2351 msg = lnet_msg_alloc();
2353 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2354 libcfs_id2str(target));
2358 cpt = lnet_cpt_of_cookie(mdh.cookie);
2361 md = lnet_handle2md(&mdh);
2362 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2363 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2364 match_bits, portal, libcfs_id2str(target),
2365 md == NULL ? -1 : md->md_threshold);
2366 if (md != NULL && md->md_me != NULL)
2367 CERROR("REPLY MD also attached to portal %d\n",
2368 md->md_me->me_portal);
2370 lnet_res_unlock(cpt);
2376 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2378 lnet_msg_attach_md(msg, md, 0, 0);
2380 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2382 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2383 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2384 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2385 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2387 /* NB handles only looked up by creator (no flips) */
2388 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2389 the_lnet.ln_interface_cookie;
2390 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2391 md->md_lh.lh_cookie;
2393 lnet_res_unlock(cpt);
2395 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2397 rc = lnet_send(self, msg, LNET_NID_ANY);
2399 CNETERR("Error sending GET to %s: %d\n",
2400 libcfs_id2str(target), rc);
2401 lnet_finalize(NULL, msg, rc);
2404 /* completion will be signalled by an event */
2407 EXPORT_SYMBOL(LNetGet);
2410 * Calculate distance to node at \a dstnid.
2412 * \param dstnid Target NID.
2413 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2415 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2418 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2419 * local_nid_dist_zero is set, which is the default.
2420 * \retval positives Distance to target NID, i.e. number of hops plus one.
2421 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2424 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2428 lnet_remotenet_t *rnet;
2429 __u32 dstnet = LNET_NIDNET(dstnid);
2433 cfs_list_t *rn_list;
2435 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2436 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2437 * keep order 0 free for 0@lo and order 1 free for a local NID
2440 LASSERT (the_lnet.ln_init);
2441 LASSERT (the_lnet.ln_refcount > 0);
2443 cpt = lnet_net_lock_current();
2445 cfs_list_for_each (e, &the_lnet.ln_nis) {
2446 ni = cfs_list_entry(e, lnet_ni_t, ni_list);
2448 if (ni->ni_nid == dstnid) {
2449 if (srcnidp != NULL)
2451 if (orderp != NULL) {
2452 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2457 lnet_net_unlock(cpt);
2459 return local_nid_dist_zero ? 0 : 1;
2462 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2463 if (srcnidp != NULL)
2464 *srcnidp = ni->ni_nid;
2467 lnet_net_unlock(cpt);
2474 rn_list = lnet_net2rnethash(dstnet);
2475 cfs_list_for_each(e, rn_list) {
2476 rnet = cfs_list_entry(e, lnet_remotenet_t, lrn_list);
2478 if (rnet->lrn_net == dstnet) {
2479 lnet_route_t *route;
2480 lnet_route_t *shortest = NULL;
2482 LASSERT (!cfs_list_empty(&rnet->lrn_routes));
2484 cfs_list_for_each_entry(route, &rnet->lrn_routes,
2486 if (shortest == NULL ||
2487 route->lr_hops < shortest->lr_hops)
2491 LASSERT (shortest != NULL);
2492 hops = shortest->lr_hops;
2493 if (srcnidp != NULL)
2494 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2497 lnet_net_unlock(cpt);
2503 lnet_net_unlock(cpt);
2504 return -EHOSTUNREACH;
2506 EXPORT_SYMBOL(LNetDist);
2509 * Set the number of asynchronous messages expected from a target process.
2511 * This function is only meaningful for userspace callers. It's a no-op when
2512 * called from kernel.
2514 * Asynchronous messages are those that can come from a target when the
2515 * userspace process is not waiting for IO to complete; e.g., AST callbacks
2516 * from Lustre servers. Specifying the expected number of such messages
2517 * allows them to be eagerly received when user process is not running in
2518 * LNet; otherwise network errors may occur.
2520 * \param id Process ID of the target process.
2521 * \param nasync Number of asynchronous messages expected from the target.
2523 * \return 0 on success, and an error code otherwise.
2526 LNetSetAsync(lnet_process_id_t id, int nasync)
2532 lnet_remotenet_t *rnet;
2534 lnet_route_t *route;
2542 /* Target on a local network? */
2543 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2545 if (ni->ni_lnd->lnd_setasync != NULL)
2546 rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2551 /* Target on a remote network: apply to routers */
2553 LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2558 /* Snapshot all the router NIDs */
2559 cpt = lnet_net_lock_current();
2560 rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2562 cfs_list_for_each(tmp, &rnet->lrn_routes) {
2563 if (nnids == maxnids) {
2564 lnet_net_unlock(cpt);
2565 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2570 route = cfs_list_entry(tmp, lnet_route_t, lr_list);
2571 nids[nnids++] = route->lr_gateway->lp_nid;
2574 lnet_net_unlock(cpt);
2576 /* set async on all the routers */
2577 while (nnids-- > 0) {
2578 id.pid = LUSTRE_SRV_LNET_PID;
2579 id.nid = nids[nnids];
2581 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2585 if (ni->ni_lnd->lnd_setasync != NULL) {
2586 rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2593 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2597 EXPORT_SYMBOL(LNetSetAsync);