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, 2014, 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)
54 struct list_head *next;
55 struct list_head cull;
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 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
75 /* removing entries */
76 INIT_LIST_HEAD(&cull);
80 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
81 tp = 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 */
86 list_del(&tp->tp_list);
87 list_add(&tp->tp_list, &cull);
93 while (!list_empty(&cull)) {
94 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
96 list_del(&tp->tp_list);
97 LIBCFS_FREE(tp, sizeof(*tp));
103 fail_peer (lnet_nid_t nid, int outgoing)
105 lnet_test_peer_t *tp;
106 struct list_head *el;
107 struct list_head *next;
108 struct list_head cull;
111 INIT_LIST_HEAD(&cull);
113 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
116 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
117 tp = list_entry(el, lnet_test_peer_t, tp_list);
119 if (tp->tp_threshold == 0) {
122 /* only cull zombies on outgoing tests,
123 * since we may be at interrupt priority on
124 * incoming messages. */
125 list_del(&tp->tp_list);
126 list_add(&tp->tp_list, &cull);
131 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
132 nid == tp->tp_nid) { /* fail this peer */
135 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
138 tp->tp_threshold == 0) {
140 list_del(&tp->tp_list);
141 list_add(&tp->tp_list, &cull);
150 while (!list_empty(&cull)) {
151 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
152 list_del(&tp->tp_list);
154 LIBCFS_FREE(tp, sizeof(*tp));
161 lnet_iov_nob (unsigned int niov, struct iovec *iov)
163 unsigned int nob = 0;
165 LASSERT(niov == 0 || iov != NULL);
167 nob += (iov++)->iov_len;
171 EXPORT_SYMBOL(lnet_iov_nob);
174 lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
175 unsigned int nsiov, struct iovec *siov, unsigned int soffset,
178 /* NB diov, siov are READ-ONLY */
179 unsigned int this_nob;
184 /* skip complete frags before 'doffset' */
186 while (doffset >= diov->iov_len) {
187 doffset -= diov->iov_len;
193 /* skip complete frags before 'soffset' */
195 while (soffset >= siov->iov_len) {
196 soffset -= siov->iov_len;
205 this_nob = MIN(diov->iov_len - doffset,
206 siov->iov_len - soffset);
207 this_nob = MIN(this_nob, nob);
209 memcpy ((char *)diov->iov_base + doffset,
210 (char *)siov->iov_base + soffset, this_nob);
213 if (diov->iov_len > doffset + this_nob) {
221 if (siov->iov_len > soffset + this_nob) {
230 EXPORT_SYMBOL(lnet_copy_iov2iov);
233 lnet_extract_iov (int dst_niov, struct iovec *dst,
234 int src_niov, struct iovec *src,
235 unsigned int offset, unsigned int len)
237 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
238 * for exactly 'len' bytes, and return the number of entries.
239 * NB not destructive to 'src' */
240 unsigned int frag_len;
243 if (len == 0) /* no data => */
244 return (0); /* no frags */
246 LASSERT (src_niov > 0);
247 while (offset >= src->iov_len) { /* skip initial frags */
248 offset -= src->iov_len;
251 LASSERT (src_niov > 0);
256 LASSERT (src_niov > 0);
257 LASSERT ((int)niov <= dst_niov);
259 frag_len = src->iov_len - offset;
260 dst->iov_base = ((char *)src->iov_base) + offset;
262 if (len <= frag_len) {
267 dst->iov_len = frag_len;
277 EXPORT_SYMBOL(lnet_extract_iov);
281 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
288 lnet_copy_kiov2kiov (unsigned int ndkiov, lnet_kiov_t *dkiov, unsigned int doffset,
289 unsigned int nskiov, lnet_kiov_t *skiov, unsigned int soffset,
296 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
297 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
304 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
305 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
312 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
313 int src_niov, lnet_kiov_t *src,
314 unsigned int offset, unsigned int len)
319 #else /* __KERNEL__ */
322 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
324 unsigned int nob = 0;
326 LASSERT(niov == 0 || kiov != NULL);
328 nob += (kiov++)->kiov_len;
332 EXPORT_SYMBOL(lnet_kiov_nob);
335 lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
336 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
339 /* NB diov, siov are READ-ONLY */
340 unsigned int this_nob;
347 LASSERT (!in_interrupt ());
350 while (doffset >= diov->kiov_len) {
351 doffset -= diov->kiov_len;
358 while (soffset >= siov->kiov_len) {
359 soffset -= siov->kiov_len;
368 this_nob = MIN(diov->kiov_len - doffset,
369 siov->kiov_len - soffset);
370 this_nob = MIN(this_nob, nob);
373 daddr = ((char *)kmap(diov->kiov_page)) +
374 diov->kiov_offset + doffset;
376 saddr = ((char *)kmap(siov->kiov_page)) +
377 siov->kiov_offset + soffset;
379 /* Vanishing risk of kmap deadlock when mapping 2 pages.
380 * However in practice at least one of the kiovs will be mapped
381 * kernel pages and the map/unmap will be NOOPs */
383 memcpy (daddr, saddr, this_nob);
386 if (diov->kiov_len > doffset + this_nob) {
390 kunmap(diov->kiov_page);
397 if (siov->kiov_len > soffset + this_nob) {
401 kunmap(siov->kiov_page);
410 kunmap(diov->kiov_page);
412 kunmap(siov->kiov_page);
414 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
417 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
418 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
421 /* NB iov, kiov are READ-ONLY */
422 unsigned int this_nob;
428 LASSERT (!in_interrupt ());
431 while (iovoffset >= iov->iov_len) {
432 iovoffset -= iov->iov_len;
439 while (kiovoffset >= kiov->kiov_len) {
440 kiovoffset -= kiov->kiov_len;
449 this_nob = MIN(iov->iov_len - iovoffset,
450 kiov->kiov_len - kiovoffset);
451 this_nob = MIN(this_nob, nob);
454 addr = ((char *)kmap(kiov->kiov_page)) +
455 kiov->kiov_offset + kiovoffset;
457 memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
460 if (iov->iov_len > iovoffset + this_nob) {
461 iovoffset += this_nob;
468 if (kiov->kiov_len > kiovoffset + this_nob) {
470 kiovoffset += this_nob;
472 kunmap(kiov->kiov_page);
482 kunmap(kiov->kiov_page);
484 EXPORT_SYMBOL(lnet_copy_kiov2iov);
487 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
488 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
491 /* NB kiov, iov are READ-ONLY */
492 unsigned int this_nob;
498 LASSERT (!in_interrupt ());
501 while (kiovoffset >= kiov->kiov_len) {
502 kiovoffset -= kiov->kiov_len;
509 while (iovoffset >= iov->iov_len) {
510 iovoffset -= iov->iov_len;
519 this_nob = MIN(kiov->kiov_len - kiovoffset,
520 iov->iov_len - iovoffset);
521 this_nob = MIN(this_nob, nob);
524 addr = ((char *)kmap(kiov->kiov_page)) +
525 kiov->kiov_offset + kiovoffset;
527 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
530 if (kiov->kiov_len > kiovoffset + this_nob) {
532 kiovoffset += this_nob;
534 kunmap(kiov->kiov_page);
541 if (iov->iov_len > iovoffset + this_nob) {
542 iovoffset += this_nob;
551 kunmap(kiov->kiov_page);
553 EXPORT_SYMBOL(lnet_copy_iov2kiov);
556 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
557 int src_niov, lnet_kiov_t *src,
558 unsigned int offset, unsigned int len)
560 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
561 * for exactly 'len' bytes, and return the number of entries.
562 * NB not destructive to 'src' */
563 unsigned int frag_len;
566 if (len == 0) /* no data => */
567 return (0); /* no frags */
569 LASSERT (src_niov > 0);
570 while (offset >= src->kiov_len) { /* skip initial frags */
571 offset -= src->kiov_len;
574 LASSERT (src_niov > 0);
579 LASSERT (src_niov > 0);
580 LASSERT ((int)niov <= dst_niov);
582 frag_len = src->kiov_len - offset;
583 dst->kiov_page = src->kiov_page;
584 dst->kiov_offset = src->kiov_offset + offset;
586 if (len <= frag_len) {
588 LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE);
592 dst->kiov_len = frag_len;
593 LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE);
603 EXPORT_SYMBOL(lnet_extract_kiov);
607 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
608 unsigned int offset, unsigned int mlen, unsigned int rlen)
610 unsigned int niov = 0;
611 struct iovec *iov = NULL;
612 lnet_kiov_t *kiov = NULL;
615 LASSERT (!in_interrupt ());
616 LASSERT (mlen == 0 || msg != NULL);
619 LASSERT(msg->msg_receiving);
620 LASSERT(!msg->msg_sending);
621 LASSERT(rlen == msg->msg_len);
622 LASSERT(mlen <= msg->msg_len);
623 LASSERT(msg->msg_offset == offset);
624 LASSERT(msg->msg_wanted == mlen);
626 msg->msg_receiving = 0;
629 niov = msg->msg_niov;
631 kiov = msg->msg_kiov;
634 LASSERT ((iov == NULL) != (kiov == NULL));
638 rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
639 niov, iov, kiov, offset, mlen, rlen);
641 lnet_finalize(ni, msg, rc);
645 lnet_setpayloadbuffer(lnet_msg_t *msg)
647 lnet_libmd_t *md = msg->msg_md;
649 LASSERT (msg->msg_len > 0);
650 LASSERT (!msg->msg_routing);
651 LASSERT (md != NULL);
652 LASSERT (msg->msg_niov == 0);
653 LASSERT (msg->msg_iov == NULL);
654 LASSERT (msg->msg_kiov == NULL);
656 msg->msg_niov = md->md_niov;
657 if ((md->md_options & LNET_MD_KIOV) != 0)
658 msg->msg_kiov = md->md_iov.kiov;
660 msg->msg_iov = md->md_iov.iov;
664 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
665 unsigned int offset, unsigned int len)
667 msg->msg_type = type;
668 msg->msg_target = target;
670 msg->msg_offset = offset;
673 lnet_setpayloadbuffer(msg);
675 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
676 msg->msg_hdr.type = cpu_to_le32(type);
677 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
678 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
679 /* src_nid will be set later */
680 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
681 msg->msg_hdr.payload_length = cpu_to_le32(len);
685 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
687 void *priv = msg->msg_private;
690 LASSERT (!in_interrupt ());
691 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
692 (msg->msg_txcredit && msg->msg_peertxcredit));
694 rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
696 lnet_finalize(ni, msg, rc);
700 lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
704 LASSERT(!msg->msg_sending);
705 LASSERT(msg->msg_receiving);
706 LASSERT(!msg->msg_rx_ready_delay);
707 LASSERT(ni->ni_lnd->lnd_eager_recv != NULL);
709 msg->msg_rx_ready_delay = 1;
710 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
713 CERROR("recv from %s / send to %s aborted: "
714 "eager_recv failed %d\n",
715 libcfs_nid2str(msg->msg_rxpeer->lp_nid),
716 libcfs_id2str(msg->msg_target), rc);
717 LASSERT(rc < 0); /* required by my callers */
723 /* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */
725 lnet_ni_query_locked(lnet_ni_t *ni, lnet_peer_t *lp)
727 cfs_time_t last_alive = 0;
729 LASSERT(lnet_peer_aliveness_enabled(lp));
730 LASSERT(ni->ni_lnd->lnd_query != NULL);
732 lnet_net_unlock(lp->lp_cpt);
733 (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
734 lnet_net_lock(lp->lp_cpt);
736 lp->lp_last_query = cfs_time_current();
738 if (last_alive != 0) /* NI has updated timestamp */
739 lp->lp_last_alive = last_alive;
742 /* NB: always called with lnet_net_lock held */
744 lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
749 LASSERT (lnet_peer_aliveness_enabled(lp));
751 /* Trust lnet_notify() if it has more recent aliveness news, but
752 * ignore the initial assumed death (see lnet_peers_start_down()).
754 if (!lp->lp_alive && lp->lp_alive_count > 0 &&
755 cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive))
758 deadline = cfs_time_add(lp->lp_last_alive,
759 cfs_time_seconds(lp->lp_ni->ni_peertimeout));
760 alive = cfs_time_after(deadline, now);
762 /* Update obsolete lp_alive except for routers assumed to be dead
763 * initially, because router checker would update aliveness in this
764 * case, and moreover lp_last_alive at peer creation is assumed.
766 if (alive && !lp->lp_alive &&
767 !(lnet_isrouter(lp) && lp->lp_alive_count == 0))
768 lnet_notify_locked(lp, 0, 1, lp->lp_last_alive);
774 /* NB: returns 1 when alive, 0 when dead, negative when error;
775 * may drop the lnet_net_lock */
777 lnet_peer_alive_locked (lnet_peer_t *lp)
779 cfs_time_t now = cfs_time_current();
781 if (!lnet_peer_aliveness_enabled(lp))
784 if (lnet_peer_is_alive(lp, now))
787 /* Peer appears dead, but we should avoid frequent NI queries (at
788 * most once per lnet_queryinterval seconds). */
789 if (lp->lp_last_query != 0) {
790 static const int lnet_queryinterval = 1;
792 cfs_time_t next_query =
793 cfs_time_add(lp->lp_last_query,
794 cfs_time_seconds(lnet_queryinterval));
796 if (cfs_time_before(now, next_query)) {
798 CWARN("Unexpected aliveness of peer %s: "
800 libcfs_nid2str(lp->lp_nid),
801 (int)now, (int)next_query,
803 lp->lp_ni->ni_peertimeout);
808 /* query NI for latest aliveness news */
809 lnet_ni_query_locked(lp->lp_ni, lp);
811 if (lnet_peer_is_alive(lp, now))
814 lnet_notify_locked(lp, 0, 0, lp->lp_last_alive);
819 * \param msg The message to be sent.
820 * \param do_send True if lnet_ni_send() should be called in this function.
821 * lnet_send() is going to lnet_net_unlock immediately after this, so
822 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
824 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
825 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
826 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
827 * \retval -ECANCELED If the MD of the message has been unlinked.
830 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
832 lnet_peer_t *lp = msg->msg_txpeer;
833 lnet_ni_t *ni = lp->lp_ni;
834 int cpt = msg->msg_tx_cpt;
835 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
837 /* non-lnet_send() callers have checked before */
838 LASSERT(!do_send || msg->msg_tx_delayed);
839 LASSERT(!msg->msg_receiving);
840 LASSERT(msg->msg_tx_committed);
842 /* NB 'lp' is always the next hop */
843 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
844 lnet_peer_alive_locked(lp) == 0) {
845 the_lnet.ln_counters[cpt]->drop_count++;
846 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
847 lnet_net_unlock(cpt);
849 CNETERR("Dropping message for %s: peer not alive\n",
850 libcfs_id2str(msg->msg_target));
852 lnet_finalize(ni, msg, -EHOSTUNREACH);
855 return -EHOSTUNREACH;
858 if (msg->msg_md != NULL &&
859 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
860 lnet_net_unlock(cpt);
862 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
863 "called on the MD/ME.\n",
864 libcfs_id2str(msg->msg_target));
866 lnet_finalize(ni, msg, -ECANCELED);
872 if (!msg->msg_peertxcredit) {
873 LASSERT((lp->lp_txcredits < 0) ==
874 !list_empty(&lp->lp_txq));
876 msg->msg_peertxcredit = 1;
877 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
880 if (lp->lp_txcredits < lp->lp_mintxcredits)
881 lp->lp_mintxcredits = lp->lp_txcredits;
883 if (lp->lp_txcredits < 0) {
884 msg->msg_tx_delayed = 1;
885 list_add_tail(&msg->msg_list, &lp->lp_txq);
886 return LNET_CREDIT_WAIT;
890 if (!msg->msg_txcredit) {
891 LASSERT((tq->tq_credits < 0) ==
892 !list_empty(&tq->tq_delayed));
894 msg->msg_txcredit = 1;
897 if (tq->tq_credits < tq->tq_credits_min)
898 tq->tq_credits_min = tq->tq_credits;
900 if (tq->tq_credits < 0) {
901 msg->msg_tx_delayed = 1;
902 list_add_tail(&msg->msg_list, &tq->tq_delayed);
903 return LNET_CREDIT_WAIT;
908 lnet_net_unlock(cpt);
909 lnet_ni_send(ni, msg);
912 return LNET_CREDIT_OK;
917 static lnet_rtrbufpool_t *
918 lnet_msg2bufpool(lnet_msg_t *msg)
920 lnet_rtrbufpool_t *rbp;
923 LASSERT(msg->msg_rx_committed);
925 cpt = msg->msg_rx_cpt;
926 rbp = &the_lnet.ln_rtrpools[cpt][0];
928 LASSERT(msg->msg_len <= LNET_MTU);
929 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_CACHE_SIZE) {
931 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
938 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
940 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
941 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
942 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
943 * 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 !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 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
972 return LNET_CREDIT_WAIT;
976 rbp = lnet_msg2bufpool(msg);
978 if (!msg->msg_rtrcredit) {
979 msg->msg_rtrcredit = 1;
981 if (rbp->rbp_credits < rbp->rbp_mincredits)
982 rbp->rbp_mincredits = rbp->rbp_credits;
984 if (rbp->rbp_credits < 0) {
985 /* must have checked eager_recv before here */
986 LASSERT(msg->msg_rx_ready_delay);
987 msg->msg_rx_delayed = 1;
988 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
989 return LNET_CREDIT_WAIT;
993 LASSERT(!list_empty(&rbp->rbp_bufs));
994 rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
995 list_del(&rb->rb_list);
997 msg->msg_niov = rbp->rbp_npages;
998 msg->msg_kiov = &rb->rb_kiov[0];
1001 int cpt = msg->msg_rx_cpt;
1003 lnet_net_unlock(cpt);
1004 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
1005 0, msg->msg_len, msg->msg_len);
1008 return LNET_CREDIT_OK;
1013 lnet_return_tx_credits_locked(lnet_msg_t *msg)
1015 lnet_peer_t *txpeer = msg->msg_txpeer;
1018 if (msg->msg_txcredit) {
1019 struct lnet_ni *ni = txpeer->lp_ni;
1020 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1022 /* give back NI txcredits */
1023 msg->msg_txcredit = 0;
1025 LASSERT((tq->tq_credits < 0) ==
1026 !list_empty(&tq->tq_delayed));
1029 if (tq->tq_credits <= 0) {
1030 msg2 = list_entry(tq->tq_delayed.next,
1031 lnet_msg_t, msg_list);
1032 list_del(&msg2->msg_list);
1034 LASSERT(msg2->msg_txpeer->lp_ni == ni);
1035 LASSERT(msg2->msg_tx_delayed);
1037 (void) lnet_post_send_locked(msg2, 1);
1041 if (msg->msg_peertxcredit) {
1042 /* give back peer txcredits */
1043 msg->msg_peertxcredit = 0;
1045 LASSERT((txpeer->lp_txcredits < 0) ==
1046 !list_empty(&txpeer->lp_txq));
1048 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1049 LASSERT (txpeer->lp_txqnob >= 0);
1051 txpeer->lp_txcredits++;
1052 if (txpeer->lp_txcredits <= 0) {
1053 msg2 = list_entry(txpeer->lp_txq.next,
1054 lnet_msg_t, msg_list);
1055 list_del(&msg2->msg_list);
1057 LASSERT(msg2->msg_txpeer == txpeer);
1058 LASSERT(msg2->msg_tx_delayed);
1060 (void) lnet_post_send_locked(msg2, 1);
1064 if (txpeer != NULL) {
1065 msg->msg_txpeer = NULL;
1066 lnet_peer_decref_locked(txpeer);
1072 lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp)
1076 if (list_empty(&rbp->rbp_msgs))
1078 msg = list_entry(rbp->rbp_msgs.next,
1079 lnet_msg_t, msg_list);
1080 list_del(&msg->msg_list);
1082 (void)lnet_post_routed_recv_locked(msg, 1);
1087 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1091 struct list_head drop;
1093 INIT_LIST_HEAD(&drop);
1095 list_splice_init(list, &drop);
1097 lnet_net_unlock(cpt);
1099 list_for_each_entry_safe(msg, tmp, &drop, msg_list) {
1100 lnet_ni_recv(msg->msg_rxpeer->lp_ni, msg->msg_private, NULL,
1101 0, 0, 0, msg->msg_hdr.payload_length);
1102 list_del_init(&msg->msg_list);
1103 lnet_finalize(NULL, msg, -ECANCELED);
1110 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1112 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1116 if (msg->msg_rtrcredit) {
1117 /* give back global router credits */
1119 lnet_rtrbufpool_t *rbp;
1121 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1122 * there until it gets one allocated, or aborts the wait
1124 LASSERT(msg->msg_kiov != NULL);
1126 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1129 msg->msg_kiov = NULL;
1130 msg->msg_rtrcredit = 0;
1132 LASSERT(rbp == lnet_msg2bufpool(msg));
1134 LASSERT((rbp->rbp_credits > 0) ==
1135 !list_empty(&rbp->rbp_bufs));
1137 /* If routing is now turned off, we just drop this buffer and
1138 * don't bother trying to return credits. */
1139 if (!the_lnet.ln_routing) {
1140 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1144 /* It is possible that a user has lowered the desired number of
1145 * buffers in this pool. Make sure we never put back
1146 * more buffers than the stated number. */
1147 if (rbp->rbp_credits >= rbp->rbp_nbuffers) {
1148 /* Discard this buffer so we don't have too many. */
1149 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1151 list_add(&rb->rb_list, &rbp->rbp_bufs);
1153 if (rbp->rbp_credits <= 0)
1154 lnet_schedule_blocked_locked(rbp);
1159 if (msg->msg_peerrtrcredit) {
1160 /* give back peer router credits */
1161 msg->msg_peerrtrcredit = 0;
1163 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1164 !list_empty(&rxpeer->lp_rtrq));
1166 rxpeer->lp_rtrcredits++;
1168 /* drop all messages which are queued to be routed on that
1170 if (!the_lnet.ln_routing) {
1171 lnet_drop_routed_msgs_locked(&rxpeer->lp_rtrq,
1173 } else if (rxpeer->lp_rtrcredits <= 0) {
1174 msg2 = list_entry(rxpeer->lp_rtrq.next,
1175 lnet_msg_t, msg_list);
1176 list_del(&msg2->msg_list);
1178 (void) lnet_post_routed_recv_locked(msg2, 1);
1182 LASSERT(!msg->msg_rtrcredit);
1183 LASSERT(!msg->msg_peerrtrcredit);
1185 if (rxpeer != NULL) {
1186 msg->msg_rxpeer = NULL;
1187 lnet_peer_decref_locked(rxpeer);
1192 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1194 lnet_peer_t *p1 = r1->lr_gateway;
1195 lnet_peer_t *p2 = r2->lr_gateway;
1197 if (r1->lr_priority < r2->lr_priority)
1200 if (r1->lr_priority > r2->lr_priority)
1203 if (r1->lr_hops < r2->lr_hops)
1206 if (r1->lr_hops > r2->lr_hops)
1209 if (p1->lp_txqnob < p2->lp_txqnob)
1212 if (p1->lp_txqnob > p2->lp_txqnob)
1215 if (p1->lp_txcredits > p2->lp_txcredits)
1218 if (p1->lp_txcredits < p2->lp_txcredits)
1221 if (r1->lr_seq - r2->lr_seq <= 0)
1227 static lnet_peer_t *
1228 lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target, lnet_nid_t rtr_nid)
1230 lnet_remotenet_t *rnet;
1231 lnet_route_t *route;
1232 lnet_route_t *best_route;
1233 lnet_route_t *last_route;
1234 struct lnet_peer *lp_best;
1235 struct lnet_peer *lp;
1238 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1239 * rtr_nid nid, otherwise find the best gateway I can use */
1241 rnet = lnet_find_net_locked(LNET_NIDNET(target));
1246 best_route = last_route = NULL;
1247 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1248 lp = route->lr_gateway;
1250 if (!lnet_is_route_alive(route))
1253 if (ni != NULL && lp->lp_ni != ni)
1256 if (lp->lp_nid == rtr_nid) /* it's pre-determined router */
1259 if (lp_best == NULL) {
1260 best_route = last_route = route;
1265 /* no protection on below fields, but it's harmless */
1266 if (last_route->lr_seq - route->lr_seq < 0)
1269 rc = lnet_compare_routes(route, best_route);
1277 /* set sequence number on the best router to the latest sequence + 1
1278 * so we can round-robin all routers, it's race and inaccurate but
1279 * harmless and functional */
1280 if (best_route != NULL)
1281 best_route->lr_seq = last_route->lr_seq + 1;
1286 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1288 lnet_nid_t dst_nid = msg->msg_target.nid;
1289 struct lnet_ni *src_ni;
1290 struct lnet_ni *local_ni;
1291 struct lnet_peer *lp;
1296 /* NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1297 * but we might want to use pre-determined router for ACK/REPLY
1299 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1300 LASSERT (msg->msg_txpeer == NULL);
1301 LASSERT (!msg->msg_sending);
1302 LASSERT (!msg->msg_target_is_router);
1303 LASSERT (!msg->msg_receiving);
1305 msg->msg_sending = 1;
1307 LASSERT(!msg->msg_tx_committed);
1308 cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid);
1312 if (the_lnet.ln_shutdown) {
1313 lnet_net_unlock(cpt);
1317 if (src_nid == LNET_NID_ANY) {
1320 src_ni = lnet_nid2ni_locked(src_nid, cpt);
1321 if (src_ni == NULL) {
1322 lnet_net_unlock(cpt);
1323 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1324 "local nid\n", libcfs_nid2str(dst_nid),
1325 libcfs_nid2str(src_nid));
1328 LASSERT (!msg->msg_routing);
1331 /* Is this for someone on a local network? */
1332 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt);
1334 if (local_ni != NULL) {
1335 if (src_ni == NULL) {
1337 src_nid = src_ni->ni_nid;
1338 } else if (src_ni == local_ni) {
1339 lnet_ni_decref_locked(local_ni, cpt);
1341 lnet_ni_decref_locked(local_ni, cpt);
1342 lnet_ni_decref_locked(src_ni, cpt);
1343 lnet_net_unlock(cpt);
1344 LCONSOLE_WARN("No route to %s via from %s\n",
1345 libcfs_nid2str(dst_nid),
1346 libcfs_nid2str(src_nid));
1350 LASSERT(src_nid != LNET_NID_ANY);
1351 lnet_msg_commit(msg, cpt);
1353 if (!msg->msg_routing)
1354 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1356 if (src_ni == the_lnet.ln_loni) {
1357 /* No send credit hassles with LOLND */
1358 lnet_net_unlock(cpt);
1359 lnet_ni_send(src_ni, msg);
1362 lnet_ni_decref_locked(src_ni, cpt);
1363 lnet_net_unlock(cpt);
1367 rc = lnet_nid2peer_locked(&lp, dst_nid, cpt);
1368 /* lp has ref on src_ni; lose mine */
1369 lnet_ni_decref_locked(src_ni, cpt);
1371 lnet_net_unlock(cpt);
1372 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1373 libcfs_nid2str(dst_nid));
1374 /* ENOMEM or shutting down */
1377 LASSERT (lp->lp_ni == src_ni);
1380 lnet_net_unlock(cpt);
1383 * - once application finishes computation, check here to update
1384 * router states before it waits for pending IO in LNetEQPoll
1385 * - recursion breaker: router checker sends no message
1386 * to remote networks */
1387 if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING)
1388 lnet_router_checker();
1392 /* sending to a remote network */
1393 lp = lnet_find_route_locked(src_ni, dst_nid, rtr_nid);
1396 lnet_ni_decref_locked(src_ni, cpt);
1397 lnet_net_unlock(cpt);
1399 LCONSOLE_WARN("No route to %s via %s "
1400 "(all routers down)\n",
1401 libcfs_id2str(msg->msg_target),
1402 libcfs_nid2str(src_nid));
1403 return -EHOSTUNREACH;
1406 /* rtr_nid is LNET_NID_ANY or NID of pre-determined router,
1407 * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't
1408 * pre-determined router, this can happen if router table
1409 * was changed when we release the lock */
1410 if (rtr_nid != lp->lp_nid) {
1411 cpt2 = lnet_cpt_of_nid_locked(lp->lp_nid);
1414 lnet_ni_decref_locked(src_ni, cpt);
1415 lnet_net_unlock(cpt);
1417 rtr_nid = lp->lp_nid;
1423 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1424 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1425 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1427 if (src_ni == NULL) {
1429 src_nid = src_ni->ni_nid;
1431 LASSERT (src_ni == lp->lp_ni);
1432 lnet_ni_decref_locked(src_ni, cpt);
1435 lnet_peer_addref_locked(lp);
1437 LASSERT(src_nid != LNET_NID_ANY);
1438 lnet_msg_commit(msg, cpt);
1440 if (!msg->msg_routing) {
1441 /* I'm the source and now I know which NI to send on */
1442 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1445 msg->msg_target_is_router = 1;
1446 msg->msg_target.nid = lp->lp_nid;
1447 msg->msg_target.pid = LNET_PID_LUSTRE;
1450 /* 'lp' is our best choice of peer */
1452 LASSERT (!msg->msg_peertxcredit);
1453 LASSERT (!msg->msg_txcredit);
1454 LASSERT (msg->msg_txpeer == NULL);
1456 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1458 rc = lnet_post_send_locked(msg, 0);
1459 lnet_net_unlock(cpt);
1464 if (rc == LNET_CREDIT_OK)
1465 lnet_ni_send(src_ni, msg);
1467 return 0; /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
1471 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1474 the_lnet.ln_counters[cpt]->drop_count++;
1475 the_lnet.ln_counters[cpt]->drop_length += nob;
1476 lnet_net_unlock(cpt);
1478 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1482 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1484 lnet_hdr_t *hdr = &msg->msg_hdr;
1486 if (msg->msg_wanted != 0)
1487 lnet_setpayloadbuffer(msg);
1489 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1491 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1492 * it back into the ACK during lnet_finalize() */
1493 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1494 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1496 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1497 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1501 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1503 lnet_hdr_t *hdr = &msg->msg_hdr;
1504 struct lnet_match_info info;
1507 /* Convert put fields to host byte order */
1508 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1509 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1510 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1512 info.mi_id.nid = hdr->src_nid;
1513 info.mi_id.pid = hdr->src_pid;
1514 info.mi_opc = LNET_MD_OP_PUT;
1515 info.mi_portal = hdr->msg.put.ptl_index;
1516 info.mi_rlength = hdr->payload_length;
1517 info.mi_roffset = hdr->msg.put.offset;
1518 info.mi_mbits = hdr->msg.put.match_bits;
1520 msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL;
1523 rc = lnet_ptl_match_md(&info, msg);
1528 case LNET_MATCHMD_OK:
1529 lnet_recv_put(ni, msg);
1532 case LNET_MATCHMD_NONE:
1533 if (msg->msg_rx_delayed) /* attached on delayed list */
1536 rc = lnet_ni_eager_recv(ni, msg);
1541 case LNET_MATCHMD_DROP:
1542 CNETERR("Dropping PUT from %s portal %d match "LPU64
1543 " offset %d length %d: %d\n",
1544 libcfs_id2str(info.mi_id), info.mi_portal,
1545 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1547 return ENOENT; /* +ve: OK but no match */
1552 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1554 struct lnet_match_info info;
1555 lnet_hdr_t *hdr = &msg->msg_hdr;
1556 lnet_handle_wire_t reply_wmd;
1559 /* Convert get fields to host byte order */
1560 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1561 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1562 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1563 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1565 info.mi_id.nid = hdr->src_nid;
1566 info.mi_id.pid = hdr->src_pid;
1567 info.mi_opc = LNET_MD_OP_GET;
1568 info.mi_portal = hdr->msg.get.ptl_index;
1569 info.mi_rlength = hdr->msg.get.sink_length;
1570 info.mi_roffset = hdr->msg.get.src_offset;
1571 info.mi_mbits = hdr->msg.get.match_bits;
1573 rc = lnet_ptl_match_md(&info, msg);
1574 if (rc == LNET_MATCHMD_DROP) {
1575 CNETERR("Dropping GET from %s portal %d match "LPU64
1576 " offset %d length %d\n",
1577 libcfs_id2str(info.mi_id), info.mi_portal,
1578 info.mi_mbits, info.mi_roffset, info.mi_rlength);
1579 return ENOENT; /* +ve: OK but no match */
1582 LASSERT(rc == LNET_MATCHMD_OK);
1584 lnet_build_msg_event(msg, LNET_EVENT_GET);
1586 reply_wmd = hdr->msg.get.return_wmd;
1588 lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
1589 msg->msg_offset, msg->msg_wanted);
1591 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1594 /* The LND completes the REPLY from her recv procedure */
1595 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1596 msg->msg_offset, msg->msg_len, msg->msg_len);
1600 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1601 msg->msg_receiving = 0;
1603 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
1605 /* didn't get as far as lnet_ni_send() */
1606 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1607 libcfs_nid2str(ni->ni_nid),
1608 libcfs_id2str(info.mi_id), rc);
1610 lnet_finalize(ni, msg, rc);
1617 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1619 void *private = msg->msg_private;
1620 lnet_hdr_t *hdr = &msg->msg_hdr;
1621 lnet_process_id_t src = {0};
1627 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
1630 src.nid = hdr->src_nid;
1631 src.pid = hdr->src_pid;
1633 /* NB handles only looked up by creator (no flips) */
1634 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1635 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1636 CNETERR("%s: Dropping REPLY from %s for %s "
1637 "MD "LPX64"."LPX64"\n",
1638 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1639 (md == NULL) ? "invalid" : "inactive",
1640 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1641 hdr->msg.reply.dst_wmd.wh_object_cookie);
1642 if (md != NULL && md->md_me != NULL)
1643 CERROR("REPLY MD also attached to portal %d\n",
1644 md->md_me->me_portal);
1646 lnet_res_unlock(cpt);
1647 return ENOENT; /* +ve: OK but no match */
1650 LASSERT (md->md_offset == 0);
1652 rlength = hdr->payload_length;
1653 mlength = MIN(rlength, (int)md->md_length);
1655 if (mlength < rlength &&
1656 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1657 CNETERR("%s: Dropping REPLY from %s length %d "
1658 "for MD "LPX64" would overflow (%d)\n",
1659 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1660 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1662 lnet_res_unlock(cpt);
1663 return ENOENT; /* +ve: OK but no match */
1666 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1667 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1668 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1670 lnet_msg_attach_md(msg, md, 0, mlength);
1673 lnet_setpayloadbuffer(msg);
1675 lnet_res_unlock(cpt);
1677 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
1679 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1684 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1686 lnet_hdr_t *hdr = &msg->msg_hdr;
1687 lnet_process_id_t src = {0};
1691 src.nid = hdr->src_nid;
1692 src.pid = hdr->src_pid;
1694 /* Convert ack fields to host byte order */
1695 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1696 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1698 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
1701 /* NB handles only looked up by creator (no flips) */
1702 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1703 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1704 /* Don't moan; this is expected */
1706 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1707 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1708 (md == NULL) ? "invalid" : "inactive",
1709 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1710 hdr->msg.ack.dst_wmd.wh_object_cookie);
1711 if (md != NULL && md->md_me != NULL)
1712 CERROR("Source MD also attached to portal %d\n",
1713 md->md_me->me_portal);
1715 lnet_res_unlock(cpt);
1716 return ENOENT; /* +ve! */
1719 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1720 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1721 hdr->msg.ack.dst_wmd.wh_object_cookie);
1723 lnet_msg_attach_md(msg, md, 0, 0);
1725 lnet_res_unlock(cpt);
1727 lnet_build_msg_event(msg, LNET_EVENT_ACK);
1729 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
1734 * \retval LNET_CREDIT_OK If \a msg is forwarded
1735 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
1736 * \retval -ve error code
1739 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
1744 if (!the_lnet.ln_routing)
1747 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
1748 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
1749 if (ni->ni_lnd->lnd_eager_recv == NULL) {
1750 msg->msg_rx_ready_delay = 1;
1752 lnet_net_unlock(msg->msg_rx_cpt);
1753 rc = lnet_ni_eager_recv(ni, msg);
1754 lnet_net_lock(msg->msg_rx_cpt);
1759 rc = lnet_post_routed_recv_locked(msg, 0);
1767 lnet_parse_local(lnet_ni_t *ni, lnet_msg_t *msg)
1771 switch (msg->msg_type) {
1773 rc = lnet_parse_ack(ni, msg);
1776 rc = lnet_parse_put(ni, msg);
1779 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
1781 case LNET_MSG_REPLY:
1782 rc = lnet_parse_reply(ni, msg);
1784 default: /* prevent an unused label if !kernel */
1789 LASSERT(rc == 0 || rc == ENOENT);
1794 lnet_msgtyp2str (int type)
1803 case LNET_MSG_REPLY:
1805 case LNET_MSG_HELLO:
1808 return ("<UNKNOWN>");
1811 EXPORT_SYMBOL(lnet_msgtyp2str);
1814 lnet_print_hdr(lnet_hdr_t * hdr)
1816 lnet_process_id_t src = {0};
1817 lnet_process_id_t dst = {0};
1818 char *type_str = lnet_msgtyp2str (hdr->type);
1820 src.nid = hdr->src_nid;
1821 src.pid = hdr->src_pid;
1823 dst.nid = hdr->dest_nid;
1824 dst.pid = hdr->dest_pid;
1826 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
1827 CWARN(" From %s\n", libcfs_id2str(src));
1828 CWARN(" To %s\n", libcfs_id2str(dst));
1830 switch (hdr->type) {
1835 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
1836 "match bits "LPU64"\n",
1837 hdr->msg.put.ptl_index,
1838 hdr->msg.put.ack_wmd.wh_interface_cookie,
1839 hdr->msg.put.ack_wmd.wh_object_cookie,
1840 hdr->msg.put.match_bits);
1841 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
1842 hdr->payload_length, hdr->msg.put.offset,
1843 hdr->msg.put.hdr_data);
1847 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
1848 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
1849 hdr->msg.get.return_wmd.wh_interface_cookie,
1850 hdr->msg.get.return_wmd.wh_object_cookie,
1851 hdr->msg.get.match_bits);
1852 CWARN(" Length %d, src offset %d\n",
1853 hdr->msg.get.sink_length,
1854 hdr->msg.get.src_offset);
1858 CWARN(" dst md "LPX64"."LPX64", "
1859 "manipulated length %d\n",
1860 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1861 hdr->msg.ack.dst_wmd.wh_object_cookie,
1862 hdr->msg.ack.mlength);
1865 case LNET_MSG_REPLY:
1866 CWARN(" dst md "LPX64"."LPX64", "
1868 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1869 hdr->msg.reply.dst_wmd.wh_object_cookie,
1870 hdr->payload_length);
1876 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
1877 void *private, int rdma_req)
1882 struct lnet_msg *msg;
1883 lnet_pid_t dest_pid;
1884 lnet_nid_t dest_nid;
1886 __u32 payload_length;
1889 LASSERT (!in_interrupt ());
1891 type = le32_to_cpu(hdr->type);
1892 src_nid = le64_to_cpu(hdr->src_nid);
1893 dest_nid = le64_to_cpu(hdr->dest_nid);
1894 dest_pid = le32_to_cpu(hdr->dest_pid);
1895 payload_length = le32_to_cpu(hdr->payload_length);
1897 for_me = (ni->ni_nid == dest_nid);
1898 cpt = lnet_cpt_of_nid(from_nid);
1903 if (payload_length > 0) {
1904 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
1905 libcfs_nid2str(from_nid),
1906 libcfs_nid2str(src_nid),
1907 lnet_msgtyp2str(type), payload_length);
1913 case LNET_MSG_REPLY:
1914 if (payload_length >
1915 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
1916 CERROR("%s, src %s: bad %s payload %d "
1917 "(%d max expected)\n",
1918 libcfs_nid2str(from_nid),
1919 libcfs_nid2str(src_nid),
1920 lnet_msgtyp2str(type),
1922 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
1928 CERROR("%s, src %s: Bad message type 0x%x\n",
1929 libcfs_nid2str(from_nid),
1930 libcfs_nid2str(src_nid), type);
1934 if (the_lnet.ln_routing &&
1935 ni->ni_last_alive != cfs_time_current_sec()) {
1936 /* NB: so far here is the only place to set NI status to "up */
1938 ni->ni_last_alive = cfs_time_current_sec();
1939 if (ni->ni_status != NULL &&
1940 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
1941 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
1945 /* Regard a bad destination NID as a protocol error. Senders should
1946 * know what they're doing; if they don't they're misconfigured, buggy
1947 * or malicious so we chop them off at the knees :) */
1950 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
1951 /* should have gone direct */
1952 CERROR("%s, src %s: Bad dest nid %s "
1953 "(should have been sent direct)\n",
1954 libcfs_nid2str(from_nid),
1955 libcfs_nid2str(src_nid),
1956 libcfs_nid2str(dest_nid));
1960 if (lnet_islocalnid(dest_nid)) {
1961 /* dest is another local NI; sender should have used
1962 * this node's NID on its own network */
1963 CERROR("%s, src %s: Bad dest nid %s "
1964 "(it's my nid but on a different network)\n",
1965 libcfs_nid2str(from_nid),
1966 libcfs_nid2str(src_nid),
1967 libcfs_nid2str(dest_nid));
1971 if (rdma_req && type == LNET_MSG_GET) {
1972 CERROR("%s, src %s: Bad optimized GET for %s "
1973 "(final destination must be me)\n",
1974 libcfs_nid2str(from_nid),
1975 libcfs_nid2str(src_nid),
1976 libcfs_nid2str(dest_nid));
1980 if (!the_lnet.ln_routing) {
1981 CERROR("%s, src %s: Dropping message for %s "
1982 "(routing not enabled)\n",
1983 libcfs_nid2str(from_nid),
1984 libcfs_nid2str(src_nid),
1985 libcfs_nid2str(dest_nid));
1990 /* Message looks OK; we're not going to return an error, so we MUST
1991 * call back lnd_recv() come what may... */
1993 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
1994 fail_peer(src_nid, 0)) { /* shall we now? */
1995 CERROR("%s, src %s: Dropping %s to simulate failure\n",
1996 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1997 lnet_msgtyp2str(type));
2001 if (!list_empty(&the_lnet.ln_drop_rules) &&
2002 lnet_drop_rule_match(hdr)) {
2003 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
2004 "silent message loss\n",
2005 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2006 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
2011 msg = lnet_msg_alloc();
2013 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2014 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2015 lnet_msgtyp2str(type));
2019 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
2020 * pointers NULL etc */
2022 msg->msg_type = type;
2023 msg->msg_private = private;
2024 msg->msg_receiving = 1;
2025 msg->msg_rdma_get = rdma_req;
2026 msg->msg_len = msg->msg_wanted = payload_length;
2027 msg->msg_offset = 0;
2028 msg->msg_hdr = *hdr;
2029 /* for building message event */
2030 msg->msg_from = from_nid;
2032 msg->msg_target.pid = dest_pid;
2033 msg->msg_target.nid = dest_nid;
2034 msg->msg_routing = 1;
2037 /* convert common msg->hdr fields to host byteorder */
2038 msg->msg_hdr.type = type;
2039 msg->msg_hdr.src_nid = src_nid;
2040 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2041 msg->msg_hdr.dest_nid = dest_nid;
2042 msg->msg_hdr.dest_pid = dest_pid;
2043 msg->msg_hdr.payload_length = payload_length;
2047 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid, cpt);
2049 lnet_net_unlock(cpt);
2050 CERROR("%s, src %s: Dropping %s "
2051 "(error %d looking up sender)\n",
2052 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2053 lnet_msgtyp2str(type), rc);
2058 if (lnet_isrouter(msg->msg_rxpeer)) {
2059 lnet_peer_set_alive(msg->msg_rxpeer);
2060 if (avoid_asym_router_failure &&
2061 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
2062 /* received a remote message from router, update
2063 * remote NI status on this router.
2064 * NB: multi-hop routed message will be ignored.
2066 lnet_router_ni_update_locked(msg->msg_rxpeer,
2067 LNET_NIDNET(src_nid));
2071 lnet_msg_commit(msg, cpt);
2073 /* message delay simulation */
2074 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
2075 lnet_delay_rule_match_locked(hdr, msg))) {
2076 lnet_net_unlock(cpt);
2081 rc = lnet_parse_forward_locked(ni, msg);
2082 lnet_net_unlock(cpt);
2087 if (rc == LNET_CREDIT_OK) {
2088 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2089 0, payload_length, payload_length);
2094 lnet_net_unlock(cpt);
2096 rc = lnet_parse_local(ni, msg);
2102 LASSERT(msg->msg_md == NULL);
2103 lnet_finalize(ni, msg, rc);
2106 lnet_drop_message(ni, cpt, private, payload_length);
2109 EXPORT_SYMBOL(lnet_parse);
2112 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
2114 while (!list_empty(head)) {
2115 lnet_process_id_t id = {0};
2118 msg = list_entry(head->next, lnet_msg_t, msg_list);
2119 list_del(&msg->msg_list);
2121 id.nid = msg->msg_hdr.src_nid;
2122 id.pid = msg->msg_hdr.src_pid;
2124 LASSERT(msg->msg_md == NULL);
2125 LASSERT(msg->msg_rx_delayed);
2126 LASSERT(msg->msg_rxpeer != NULL);
2127 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2129 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
2130 " offset %d length %d: %s\n",
2132 msg->msg_hdr.msg.put.ptl_index,
2133 msg->msg_hdr.msg.put.match_bits,
2134 msg->msg_hdr.msg.put.offset,
2135 msg->msg_hdr.payload_length, reason);
2137 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2138 * called lnet_drop_message(), so I just hang onto msg as well
2139 * until that's done */
2141 lnet_drop_message(msg->msg_rxpeer->lp_ni,
2142 msg->msg_rxpeer->lp_cpt,
2143 msg->msg_private, msg->msg_len);
2145 * NB: message will not generate event because w/o attached MD,
2146 * but we still should give error code so lnet_msg_decommit()
2147 * can skip counters operations and other checks.
2149 lnet_finalize(msg->msg_rxpeer->lp_ni, msg, -ENOENT);
2154 lnet_recv_delayed_msg_list(struct list_head *head)
2156 while (!list_empty(head)) {
2158 lnet_process_id_t id;
2160 msg = list_entry(head->next, lnet_msg_t, msg_list);
2161 list_del(&msg->msg_list);
2163 /* md won't disappear under me, since each msg
2164 * holds a ref on it */
2166 id.nid = msg->msg_hdr.src_nid;
2167 id.pid = msg->msg_hdr.src_pid;
2169 LASSERT(msg->msg_rx_delayed);
2170 LASSERT(msg->msg_md != NULL);
2171 LASSERT(msg->msg_rxpeer != NULL);
2172 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2174 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2175 "match "LPU64" offset %d length %d.\n",
2176 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2177 msg->msg_hdr.msg.put.match_bits,
2178 msg->msg_hdr.msg.put.offset,
2179 msg->msg_hdr.payload_length);
2181 lnet_recv_put(msg->msg_rxpeer->lp_ni, msg);
2186 * Initiate an asynchronous PUT operation.
2188 * There are several events associated with a PUT: completion of the send on
2189 * the initiator node (LNET_EVENT_SEND), and when the send completes
2190 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2191 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2192 * used at the target node to indicate the completion of incoming data
2195 * The local events will be logged in the EQ associated with the MD pointed to
2196 * by \a mdh handle. Using a MD without an associated EQ results in these
2197 * events being discarded. In this case, the caller must have another
2198 * mechanism (e.g., a higher level protocol) for determining when it is safe
2199 * to modify the memory region associated with the MD.
2201 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2202 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2204 * \param self Indicates the NID of a local interface through which to send
2205 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2206 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2207 * must be "free floating" (See LNetMDBind()).
2208 * \param ack Controls whether an acknowledgment is requested.
2209 * Acknowledgments are only sent when they are requested by the initiating
2210 * process and the target MD enables them.
2211 * \param target A process identifier for the target process.
2212 * \param portal The index in the \a target's portal table.
2213 * \param match_bits The match bits to use for MD selection at the target
2215 * \param offset The offset into the target MD (only used when the target
2216 * MD has the LNET_MD_MANAGE_REMOTE option set).
2217 * \param hdr_data 64 bits of user data that can be included in the message
2218 * header. This data is written to an event queue entry at the target if an
2219 * EQ is present on the matching MD.
2221 * \retval 0 Success, and only in this case events will be generated
2222 * and logged to EQ (if it exists).
2223 * \retval -EIO Simulated failure.
2224 * \retval -ENOMEM Memory allocation failure.
2225 * \retval -ENOENT Invalid MD object.
2227 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2230 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2231 lnet_process_id_t target, unsigned int portal,
2232 __u64 match_bits, unsigned int offset,
2235 struct lnet_msg *msg;
2236 struct lnet_libmd *md;
2240 LASSERT(the_lnet.ln_init);
2241 LASSERT(the_lnet.ln_refcount > 0);
2243 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2244 fail_peer(target.nid, 1)) { /* shall we now? */
2245 CERROR("Dropping PUT to %s: simulated failure\n",
2246 libcfs_id2str(target));
2250 msg = lnet_msg_alloc();
2252 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2253 libcfs_id2str(target));
2256 msg->msg_vmflush = !!memory_pressure_get();
2258 cpt = lnet_cpt_of_cookie(mdh.cookie);
2261 md = lnet_handle2md(&mdh);
2262 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2263 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2264 match_bits, portal, libcfs_id2str(target),
2265 md == NULL ? -1 : md->md_threshold);
2266 if (md != NULL && md->md_me != NULL)
2267 CERROR("Source MD also attached to portal %d\n",
2268 md->md_me->me_portal);
2269 lnet_res_unlock(cpt);
2275 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2277 lnet_msg_attach_md(msg, md, 0, 0);
2279 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2281 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2282 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2283 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2284 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2286 /* NB handles only looked up by creator (no flips) */
2287 if (ack == LNET_ACK_REQ) {
2288 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2289 the_lnet.ln_interface_cookie;
2290 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2291 md->md_lh.lh_cookie;
2293 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2294 LNET_WIRE_HANDLE_COOKIE_NONE;
2295 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2296 LNET_WIRE_HANDLE_COOKIE_NONE;
2299 lnet_res_unlock(cpt);
2301 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2303 rc = lnet_send(self, msg, LNET_NID_ANY);
2305 CNETERR( "Error sending PUT to %s: %d\n",
2306 libcfs_id2str(target), rc);
2307 lnet_finalize (NULL, msg, rc);
2310 /* completion will be signalled by an event */
2313 EXPORT_SYMBOL(LNetPut);
2316 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2318 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2319 * returns a msg for the LND to pass to lnet_finalize() when the sink
2320 * data has been received.
2322 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2323 * lnet_finalize() is called on it, so the LND must call this first */
2325 struct lnet_msg *msg = lnet_msg_alloc();
2326 struct lnet_libmd *getmd = getmsg->msg_md;
2327 lnet_process_id_t peer_id = getmsg->msg_target;
2330 LASSERT(!getmsg->msg_target_is_router);
2331 LASSERT(!getmsg->msg_routing);
2334 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2335 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2339 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2342 LASSERT(getmd->md_refcount > 0);
2344 if (getmd->md_threshold == 0) {
2345 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2346 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2348 lnet_res_unlock(cpt);
2352 LASSERT(getmd->md_offset == 0);
2354 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2355 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2357 /* setup information for lnet_build_msg_event */
2358 msg->msg_from = peer_id.nid;
2359 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2360 msg->msg_hdr.src_nid = peer_id.nid;
2361 msg->msg_hdr.payload_length = getmd->md_length;
2362 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2364 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2365 lnet_res_unlock(cpt);
2367 cpt = lnet_cpt_of_nid(peer_id.nid);
2370 lnet_msg_commit(msg, cpt);
2371 lnet_net_unlock(cpt);
2373 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2378 cpt = lnet_cpt_of_nid(peer_id.nid);
2381 the_lnet.ln_counters[cpt]->drop_count++;
2382 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2383 lnet_net_unlock(cpt);
2390 EXPORT_SYMBOL(lnet_create_reply_msg);
2393 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2395 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2396 * completed and I know it. */
2397 LASSERT (reply != NULL);
2398 LASSERT (reply->msg_type == LNET_MSG_GET);
2399 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2401 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2402 * the end of my buffer, I might as well be dead. */
2403 LASSERT (len <= reply->msg_ev.mlength);
2405 reply->msg_ev.mlength = len;
2407 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2410 * Initiate an asynchronous GET operation.
2412 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2413 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2414 * the target node in the REPLY has been written to local MD.
2416 * On the target node, an LNET_EVENT_GET is logged when the GET request
2417 * arrives and is accepted into a MD.
2419 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2420 * \param mdh A handle for the MD that describes the memory into which the
2421 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2423 * \retval 0 Success, and only in this case events will be generated
2424 * and logged to EQ (if it exists) of the MD.
2425 * \retval -EIO Simulated failure.
2426 * \retval -ENOMEM Memory allocation failure.
2427 * \retval -ENOENT Invalid MD object.
2430 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2431 lnet_process_id_t target, unsigned int portal,
2432 __u64 match_bits, unsigned int offset)
2434 struct lnet_msg *msg;
2435 struct lnet_libmd *md;
2439 LASSERT (the_lnet.ln_init);
2440 LASSERT (the_lnet.ln_refcount > 0);
2442 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2443 fail_peer(target.nid, 1)) /* shall we now? */
2445 CERROR("Dropping GET to %s: simulated failure\n",
2446 libcfs_id2str(target));
2450 msg = lnet_msg_alloc();
2452 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2453 libcfs_id2str(target));
2457 cpt = lnet_cpt_of_cookie(mdh.cookie);
2460 md = lnet_handle2md(&mdh);
2461 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2462 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2463 match_bits, portal, libcfs_id2str(target),
2464 md == NULL ? -1 : md->md_threshold);
2465 if (md != NULL && md->md_me != NULL)
2466 CERROR("REPLY MD also attached to portal %d\n",
2467 md->md_me->me_portal);
2469 lnet_res_unlock(cpt);
2475 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2477 lnet_msg_attach_md(msg, md, 0, 0);
2479 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2481 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2482 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2483 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2484 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2486 /* NB handles only looked up by creator (no flips) */
2487 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2488 the_lnet.ln_interface_cookie;
2489 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2490 md->md_lh.lh_cookie;
2492 lnet_res_unlock(cpt);
2494 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2496 rc = lnet_send(self, msg, LNET_NID_ANY);
2498 CNETERR("Error sending GET to %s: %d\n",
2499 libcfs_id2str(target), rc);
2500 lnet_finalize(NULL, msg, rc);
2503 /* completion will be signalled by an event */
2506 EXPORT_SYMBOL(LNetGet);
2509 * Calculate distance to node at \a dstnid.
2511 * \param dstnid Target NID.
2512 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2514 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2517 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2518 * local_nid_dist_zero is set, which is the default.
2519 * \retval positives Distance to target NID, i.e. number of hops plus one.
2520 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2523 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2525 struct list_head *e;
2527 lnet_remotenet_t *rnet;
2528 __u32 dstnet = LNET_NIDNET(dstnid);
2532 struct list_head *rn_list;
2534 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2535 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2536 * keep order 0 free for 0@lo and order 1 free for a local NID
2539 LASSERT (the_lnet.ln_init);
2540 LASSERT (the_lnet.ln_refcount > 0);
2542 cpt = lnet_net_lock_current();
2544 list_for_each(e, &the_lnet.ln_nis) {
2545 ni = list_entry(e, lnet_ni_t, ni_list);
2547 if (ni->ni_nid == dstnid) {
2548 if (srcnidp != NULL)
2550 if (orderp != NULL) {
2551 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2556 lnet_net_unlock(cpt);
2558 return local_nid_dist_zero ? 0 : 1;
2561 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2562 if (srcnidp != NULL)
2563 *srcnidp = ni->ni_nid;
2566 lnet_net_unlock(cpt);
2573 rn_list = lnet_net2rnethash(dstnet);
2574 list_for_each(e, rn_list) {
2575 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2577 if (rnet->lrn_net == dstnet) {
2578 lnet_route_t *route;
2579 lnet_route_t *shortest = NULL;
2581 LASSERT(!list_empty(&rnet->lrn_routes));
2583 list_for_each_entry(route, &rnet->lrn_routes,
2585 if (shortest == NULL ||
2586 route->lr_hops < shortest->lr_hops)
2590 LASSERT (shortest != NULL);
2591 hops = shortest->lr_hops;
2592 if (srcnidp != NULL)
2593 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2596 lnet_net_unlock(cpt);
2602 lnet_net_unlock(cpt);
2603 return -EHOSTUNREACH;
2605 EXPORT_SYMBOL(LNetDist);
2608 * Set the number of asynchronous messages expected from a target process.
2610 * This function is only meaningful for userspace callers. It's a no-op when
2611 * called from kernel.
2613 * Asynchronous messages are those that can come from a target when the
2614 * userspace process is not waiting for IO to complete; e.g., AST callbacks
2615 * from Lustre servers. Specifying the expected number of such messages
2616 * allows them to be eagerly received when user process is not running in
2617 * LNet; otherwise network errors may occur.
2619 * \param id Process ID of the target process.
2620 * \param nasync Number of asynchronous messages expected from the target.
2622 * \return 0 on success, and an error code otherwise.
2625 LNetSetAsync(lnet_process_id_t id, int nasync)
2631 lnet_remotenet_t *rnet;
2632 struct list_head *tmp;
2633 lnet_route_t *route;
2641 /* Target on a local network? */
2642 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2644 if (ni->ni_lnd->lnd_setasync != NULL)
2645 rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2650 /* Target on a remote network: apply to routers */
2652 LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2657 /* Snapshot all the router NIDs */
2658 cpt = lnet_net_lock_current();
2659 rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2661 list_for_each(tmp, &rnet->lrn_routes) {
2662 if (nnids == maxnids) {
2663 lnet_net_unlock(cpt);
2664 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2669 route = list_entry(tmp, lnet_route_t, lr_list);
2670 nids[nnids++] = route->lr_gateway->lp_nid;
2673 lnet_net_unlock(cpt);
2675 /* set async on all the routers */
2676 while (nnids-- > 0) {
2677 id.pid = LNET_PID_LUSTRE;
2678 id.nid = nids[nnids];
2680 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2684 if (ni->ni_lnd->lnd_setasync != NULL) {
2685 rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2692 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2696 EXPORT_SYMBOL(LNetSetAsync);