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.
31 * This file is part of Lustre, http://www.lustre.org/
32 * Lustre is a trademark of Sun Microsystems, Inc.
34 * lnet/lnet/lib-move.c
36 * Data movement routines
39 #define DEBUG_SUBSYSTEM S_LNET
41 #include <lnet/lib-lnet.h>
43 static int local_nid_dist_zero = 1;
44 CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444,
48 lnet_fail_nid (lnet_nid_t nid, unsigned int threshold)
55 LASSERT (the_lnet.ln_init);
58 /* Adding a new entry */
59 LIBCFS_ALLOC(tp, sizeof(*tp));
64 tp->tp_threshold = threshold;
67 cfs_list_add_tail (&tp->tp_list, &the_lnet.ln_test_peers);
72 /* removing entries */
73 CFS_INIT_LIST_HEAD (&cull);
77 cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
78 tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
80 if (tp->tp_threshold == 0 || /* needs culling anyway */
81 nid == LNET_NID_ANY || /* removing all entries */
82 tp->tp_nid == nid) /* matched this one */
84 cfs_list_del (&tp->tp_list);
85 cfs_list_add (&tp->tp_list, &cull);
91 while (!cfs_list_empty (&cull)) {
92 tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
94 cfs_list_del (&tp->tp_list);
95 LIBCFS_FREE(tp, sizeof (*tp));
101 fail_peer (lnet_nid_t nid, int outgoing)
103 lnet_test_peer_t *tp;
109 CFS_INIT_LIST_HEAD (&cull);
113 cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
114 tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
116 if (tp->tp_threshold == 0) {
119 /* only cull zombies on outgoing tests,
120 * since we may be at interrupt priority on
121 * incoming messages. */
122 cfs_list_del (&tp->tp_list);
123 cfs_list_add (&tp->tp_list, &cull);
128 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
129 nid == tp->tp_nid) { /* fail this peer */
132 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
135 tp->tp_threshold == 0) {
137 cfs_list_del (&tp->tp_list);
138 cfs_list_add (&tp->tp_list, &cull);
147 while (!cfs_list_empty (&cull)) {
148 tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
149 cfs_list_del (&tp->tp_list);
151 LIBCFS_FREE(tp, sizeof (*tp));
158 lnet_iov_nob (unsigned int niov, struct iovec *iov)
160 unsigned int nob = 0;
163 nob += (iov++)->iov_len;
169 lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
170 unsigned int nsiov, struct iovec *siov, unsigned int soffset,
173 /* NB diov, siov are READ-ONLY */
174 unsigned int this_nob;
179 /* skip complete frags before 'doffset' */
181 while (doffset >= diov->iov_len) {
182 doffset -= diov->iov_len;
188 /* skip complete frags before 'soffset' */
190 while (soffset >= siov->iov_len) {
191 soffset -= siov->iov_len;
200 this_nob = MIN(diov->iov_len - doffset,
201 siov->iov_len - soffset);
202 this_nob = MIN(this_nob, nob);
204 memcpy ((char *)diov->iov_base + doffset,
205 (char *)siov->iov_base + soffset, this_nob);
208 if (diov->iov_len > doffset + this_nob) {
216 if (siov->iov_len > soffset + this_nob) {
227 lnet_extract_iov (int dst_niov, struct iovec *dst,
228 int src_niov, struct iovec *src,
229 unsigned int offset, unsigned int len)
231 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
232 * for exactly 'len' bytes, and return the number of entries.
233 * NB not destructive to 'src' */
234 unsigned int frag_len;
237 if (len == 0) /* no data => */
238 return (0); /* no frags */
240 LASSERT (src_niov > 0);
241 while (offset >= src->iov_len) { /* skip initial frags */
242 offset -= src->iov_len;
245 LASSERT (src_niov > 0);
250 LASSERT (src_niov > 0);
251 LASSERT ((int)niov <= dst_niov);
253 frag_len = src->iov_len - offset;
254 dst->iov_base = ((char *)src->iov_base) + offset;
256 if (len <= frag_len) {
261 dst->iov_len = frag_len;
274 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
281 lnet_copy_kiov2kiov (unsigned int ndkiov, lnet_kiov_t *dkiov, unsigned int doffset,
282 unsigned int nskiov, lnet_kiov_t *skiov, unsigned int soffset,
289 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
290 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
297 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
298 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
305 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
306 int src_niov, lnet_kiov_t *src,
307 unsigned int offset, unsigned int len)
312 #else /* __KERNEL__ */
315 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
317 unsigned int nob = 0;
320 nob += (kiov++)->kiov_len;
326 lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
327 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
330 /* NB diov, siov are READ-ONLY */
331 unsigned int this_nob;
338 LASSERT (!cfs_in_interrupt ());
341 while (doffset >= diov->kiov_len) {
342 doffset -= diov->kiov_len;
349 while (soffset >= siov->kiov_len) {
350 soffset -= siov->kiov_len;
359 this_nob = MIN(diov->kiov_len - doffset,
360 siov->kiov_len - soffset);
361 this_nob = MIN(this_nob, nob);
364 daddr = ((char *)cfs_kmap(diov->kiov_page)) +
365 diov->kiov_offset + doffset;
367 saddr = ((char *)cfs_kmap(siov->kiov_page)) +
368 siov->kiov_offset + soffset;
370 /* Vanishing risk of kmap deadlock when mapping 2 pages.
371 * However in practice at least one of the kiovs will be mapped
372 * kernel pages and the map/unmap will be NOOPs */
374 memcpy (daddr, saddr, this_nob);
377 if (diov->kiov_len > doffset + this_nob) {
381 cfs_kunmap(diov->kiov_page);
388 if (siov->kiov_len > soffset + this_nob) {
392 cfs_kunmap(siov->kiov_page);
401 cfs_kunmap(diov->kiov_page);
403 cfs_kunmap(siov->kiov_page);
407 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
408 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
411 /* NB iov, kiov are READ-ONLY */
412 unsigned int this_nob;
418 LASSERT (!cfs_in_interrupt ());
421 while (iovoffset >= iov->iov_len) {
422 iovoffset -= iov->iov_len;
429 while (kiovoffset >= kiov->kiov_len) {
430 kiovoffset -= kiov->kiov_len;
439 this_nob = MIN(iov->iov_len - iovoffset,
440 kiov->kiov_len - kiovoffset);
441 this_nob = MIN(this_nob, nob);
444 addr = ((char *)cfs_kmap(kiov->kiov_page)) +
445 kiov->kiov_offset + kiovoffset;
447 memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
450 if (iov->iov_len > iovoffset + this_nob) {
451 iovoffset += this_nob;
458 if (kiov->kiov_len > kiovoffset + this_nob) {
460 kiovoffset += this_nob;
462 cfs_kunmap(kiov->kiov_page);
472 cfs_kunmap(kiov->kiov_page);
476 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
477 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
480 /* NB kiov, iov are READ-ONLY */
481 unsigned int this_nob;
487 LASSERT (!cfs_in_interrupt ());
490 while (kiovoffset >= kiov->kiov_len) {
491 kiovoffset -= kiov->kiov_len;
498 while (iovoffset >= iov->iov_len) {
499 iovoffset -= iov->iov_len;
508 this_nob = MIN(kiov->kiov_len - kiovoffset,
509 iov->iov_len - iovoffset);
510 this_nob = MIN(this_nob, nob);
513 addr = ((char *)cfs_kmap(kiov->kiov_page)) +
514 kiov->kiov_offset + kiovoffset;
516 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
519 if (kiov->kiov_len > kiovoffset + this_nob) {
521 kiovoffset += this_nob;
523 cfs_kunmap(kiov->kiov_page);
530 if (iov->iov_len > iovoffset + this_nob) {
531 iovoffset += this_nob;
540 cfs_kunmap(kiov->kiov_page);
544 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
545 int src_niov, lnet_kiov_t *src,
546 unsigned int offset, unsigned int len)
548 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
549 * for exactly 'len' bytes, and return the number of entries.
550 * NB not destructive to 'src' */
551 unsigned int frag_len;
554 if (len == 0) /* no data => */
555 return (0); /* no frags */
557 LASSERT (src_niov > 0);
558 while (offset >= src->kiov_len) { /* skip initial frags */
559 offset -= src->kiov_len;
562 LASSERT (src_niov > 0);
567 LASSERT (src_niov > 0);
568 LASSERT ((int)niov <= dst_niov);
570 frag_len = src->kiov_len - offset;
571 dst->kiov_page = src->kiov_page;
572 dst->kiov_offset = src->kiov_offset + offset;
574 if (len <= frag_len) {
576 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
580 dst->kiov_len = frag_len;
581 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
594 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
595 unsigned int offset, unsigned int mlen, unsigned int rlen)
597 unsigned int niov = 0;
598 struct iovec *iov = NULL;
599 lnet_kiov_t *kiov = NULL;
602 LASSERT (!cfs_in_interrupt ());
603 LASSERT (mlen == 0 || msg != NULL);
606 LASSERT(msg->msg_receiving);
607 LASSERT(!msg->msg_sending);
608 LASSERT(rlen == msg->msg_len);
609 LASSERT(mlen <= msg->msg_len);
610 LASSERT(msg->msg_offset == offset);
611 LASSERT(msg->msg_wanted == mlen);
613 msg->msg_receiving = 0;
616 niov = msg->msg_niov;
618 kiov = msg->msg_kiov;
621 LASSERT ((iov == NULL) != (kiov == NULL));
625 rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
626 niov, iov, kiov, offset, mlen, rlen);
628 lnet_finalize(ni, msg, rc);
632 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
634 lnet_peer_t *p1 = r1->lr_gateway;
635 lnet_peer_t *p2 = r2->lr_gateway;
637 if (r1->lr_hops < r2->lr_hops)
640 if (r1->lr_hops > r2->lr_hops)
643 if (p1->lp_txqnob < p2->lp_txqnob)
646 if (p1->lp_txqnob > p2->lp_txqnob)
649 if (p1->lp_txcredits > p2->lp_txcredits)
652 if (p1->lp_txcredits < p2->lp_txcredits)
660 lnet_setpayloadbuffer(lnet_msg_t *msg)
662 lnet_libmd_t *md = msg->msg_md;
664 LASSERT (msg->msg_len > 0);
665 LASSERT (!msg->msg_routing);
666 LASSERT (md != NULL);
667 LASSERT (msg->msg_niov == 0);
668 LASSERT (msg->msg_iov == NULL);
669 LASSERT (msg->msg_kiov == NULL);
671 msg->msg_niov = md->md_niov;
672 if ((md->md_options & LNET_MD_KIOV) != 0)
673 msg->msg_kiov = md->md_iov.kiov;
675 msg->msg_iov = md->md_iov.iov;
679 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
680 unsigned int offset, unsigned int len)
682 msg->msg_type = type;
683 msg->msg_target = target;
685 msg->msg_offset = offset;
688 lnet_setpayloadbuffer(msg);
690 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
691 msg->msg_hdr.type = cpu_to_le32(type);
692 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
693 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
694 /* src_nid will be set later */
695 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
696 msg->msg_hdr.payload_length = cpu_to_le32(len);
700 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
702 void *priv = msg->msg_private;
705 LASSERT (!cfs_in_interrupt ());
706 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
707 (msg->msg_txcredit && msg->msg_peertxcredit));
709 rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
711 lnet_finalize(ni, msg, rc);
715 lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
719 LASSERT(!msg->msg_sending);
720 LASSERT(msg->msg_receiving);
721 LASSERT(ni->ni_lnd->lnd_eager_recv != NULL);
723 msg->msg_rx_ready_delay = 1;
724 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
727 CERROR("recv from %s / send to %s aborted: "
728 "eager_recv failed %d\n",
729 libcfs_nid2str(msg->msg_rxpeer->lp_nid),
730 libcfs_id2str(msg->msg_target), rc);
731 LASSERT(rc < 0); /* required by my callers */
738 lnet_ni_eager_recv_locked(lnet_ni_t *ni, lnet_msg_t *msg)
742 if (ni->ni_lnd->lnd_eager_recv == NULL) {
743 msg->msg_rx_ready_delay = 1;
748 rc = lnet_ni_eager_recv(ni, msg);
754 /* NB: caller shall hold a ref on 'lp' as I'd drop LNET_LOCK */
756 lnet_ni_peer_alive(lnet_peer_t *lp)
758 cfs_time_t last_alive = 0;
759 lnet_ni_t *ni = lp->lp_ni;
761 LASSERT (lnet_peer_aliveness_enabled(lp));
762 LASSERT (ni->ni_lnd->lnd_query != NULL);
763 LASSERT (the_lnet.ln_routing == 1);
766 (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
769 lp->lp_last_query = cfs_time_current();
771 if (last_alive != 0) /* NI has updated timestamp */
772 lp->lp_last_alive = last_alive;
775 /* NB: always called with LNET_LOCK held */
777 lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
782 LASSERT (lnet_peer_aliveness_enabled(lp));
783 LASSERT (the_lnet.ln_routing == 1);
785 /* Trust lnet_notify() if it has more recent aliveness news, but
786 * ignore the initial assumed death (see lnet_peers_start_down()).
788 if (!lp->lp_alive && lp->lp_alive_count > 0 &&
789 cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive))
792 deadline = cfs_time_add(lp->lp_last_alive,
793 cfs_time_seconds(lp->lp_ni->ni_peertimeout));
794 alive = cfs_time_after(deadline, now);
796 /* Update obsolete lp_alive except for routers assumed to be dead
797 * initially, because router checker would update aliveness in this
798 * case, and moreover lp_last_alive at peer creation is assumed.
800 if (alive && !lp->lp_alive &&
801 !(lnet_isrouter(lp) && lp->lp_alive_count == 0))
802 lnet_notify_locked(lp, 0, 1, lp->lp_last_alive);
808 /* NB: returns 1 when alive, 0 when dead, negative when error;
809 * may drop the LNET_LOCK */
811 lnet_peer_alive_locked (lnet_peer_t *lp)
813 cfs_time_t now = cfs_time_current();
815 /* LU-630: only router checks peer health. */
816 if (the_lnet.ln_routing == 0)
819 if (!lnet_peer_aliveness_enabled(lp))
822 if (lnet_peer_is_alive(lp, now))
825 /* Peer appears dead, but we should avoid frequent NI queries (at
826 * most once per lnet_queryinterval seconds). */
827 if (lp->lp_last_query != 0) {
828 static const int lnet_queryinterval = 1;
830 cfs_time_t next_query =
831 cfs_time_add(lp->lp_last_query,
832 cfs_time_seconds(lnet_queryinterval));
834 if (cfs_time_before(now, next_query)) {
836 CWARN("Unexpected aliveness of peer %s: "
838 libcfs_nid2str(lp->lp_nid),
839 (int)now, (int)next_query,
841 lp->lp_ni->ni_peertimeout);
846 /* query NI for latest aliveness news */
847 lnet_ni_peer_alive(lp);
849 if (lnet_peer_is_alive(lp, now))
852 lnet_notify_locked(lp, 0, 0, lp->lp_last_alive);
857 lnet_post_send_locked (lnet_msg_t *msg, int do_send)
859 /* lnet_send is going to LNET_UNLOCK immediately after this, so it sets
860 * do_send FALSE and I don't do the unlock/send/lock bit. I return
861 * EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer appears dead, and
862 * 0 if sent or OK to send */
863 lnet_peer_t *lp = msg->msg_txpeer;
864 lnet_ni_t *ni = lp->lp_ni;
866 /* non-lnet_send() callers have checked before */
867 LASSERT(!do_send || msg->msg_tx_delayed);
868 LASSERT(!msg->msg_receiving);
870 /* NB 'lp' is always the next hop */
871 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
872 lnet_peer_alive_locked(lp) == 0) {
873 the_lnet.ln_counters.drop_count++;
874 the_lnet.ln_counters.drop_length += msg->msg_len;
877 CNETERR("Dropping message for %s: peer not alive\n",
878 libcfs_id2str(msg->msg_target));
880 lnet_finalize(ni, msg, -EHOSTUNREACH);
886 if (!msg->msg_peertxcredit) {
887 LASSERT ((lp->lp_txcredits < 0) ==
888 !cfs_list_empty(&lp->lp_txq));
890 msg->msg_peertxcredit = 1;
891 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
894 if (lp->lp_txcredits < lp->lp_mintxcredits)
895 lp->lp_mintxcredits = lp->lp_txcredits;
897 if (lp->lp_txcredits < 0) {
898 msg->msg_tx_delayed = 1;
899 cfs_list_add_tail(&msg->msg_list, &lp->lp_txq);
904 if (!msg->msg_txcredit) {
905 LASSERT ((ni->ni_txcredits < 0) ==
906 !cfs_list_empty(&ni->ni_txq));
908 msg->msg_txcredit = 1;
911 if (ni->ni_txcredits < ni->ni_mintxcredits)
912 ni->ni_mintxcredits = ni->ni_txcredits;
914 if (ni->ni_txcredits < 0) {
915 msg->msg_tx_delayed = 1;
916 cfs_list_add_tail(&msg->msg_list, &ni->ni_txq);
923 lnet_ni_send(ni, msg);
932 lnet_msg2bufpool(lnet_msg_t *msg)
934 lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[0];
936 LASSERT (msg->msg_len <= LNET_MTU);
937 while (msg->msg_len > (unsigned int)rbp->rbp_npages * CFS_PAGE_SIZE) {
939 LASSERT (rbp < &the_lnet.ln_rtrpools[LNET_NRBPOOLS]);
946 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
948 /* lnet_parse is going to LNET_UNLOCK immediately after this, so it
949 * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
950 * return EAGAIN if msg blocked and 0 if received or OK to receive */
951 lnet_peer_t *lp = msg->msg_rxpeer;
952 lnet_rtrbufpool_t *rbp;
955 LASSERT (msg->msg_iov == NULL);
956 LASSERT (msg->msg_kiov == NULL);
957 LASSERT (msg->msg_niov == 0);
958 LASSERT (msg->msg_routing);
959 LASSERT (msg->msg_receiving);
960 LASSERT (!msg->msg_sending);
962 /* non-lnet_parse callers only receive delayed messages */
963 LASSERT(!do_recv || msg->msg_rx_delayed);
965 if (!msg->msg_peerrtrcredit) {
966 LASSERT ((lp->lp_rtrcredits < 0) ==
967 !cfs_list_empty(&lp->lp_rtrq));
969 msg->msg_peerrtrcredit = 1;
971 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
972 lp->lp_minrtrcredits = lp->lp_rtrcredits;
974 if (lp->lp_rtrcredits < 0) {
975 /* must have checked eager_recv before here */
976 LASSERT(msg->msg_rx_ready_delay);
977 msg->msg_rx_delayed = 1;
978 cfs_list_add_tail(&msg->msg_list, &lp->lp_rtrq);
983 rbp = lnet_msg2bufpool(msg);
985 if (!msg->msg_rtrcredit) {
986 LASSERT ((rbp->rbp_credits < 0) ==
987 !cfs_list_empty(&rbp->rbp_msgs));
989 msg->msg_rtrcredit = 1;
991 if (rbp->rbp_credits < rbp->rbp_mincredits)
992 rbp->rbp_mincredits = rbp->rbp_credits;
994 if (rbp->rbp_credits < 0) {
995 /* must have checked eager_recv before here */
996 LASSERT(msg->msg_rx_ready_delay);
997 msg->msg_rx_delayed = 1;
998 cfs_list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1003 LASSERT (!cfs_list_empty(&rbp->rbp_bufs));
1004 rb = cfs_list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
1005 cfs_list_del(&rb->rb_list);
1007 msg->msg_niov = rbp->rbp_npages;
1008 msg->msg_kiov = &rb->rb_kiov[0];
1012 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
1013 0, msg->msg_len, msg->msg_len);
1021 lnet_return_tx_credits_locked(lnet_msg_t *msg)
1023 lnet_peer_t *txpeer = msg->msg_txpeer;
1027 if (msg->msg_txcredit) {
1028 /* give back NI txcredits */
1029 msg->msg_txcredit = 0;
1032 LASSERT((ni->ni_txcredits < 0) == !cfs_list_empty(&ni->ni_txq));
1035 if (ni->ni_txcredits <= 0) {
1036 msg2 = cfs_list_entry(ni->ni_txq.next, lnet_msg_t,
1038 cfs_list_del(&msg2->msg_list);
1040 LASSERT(msg2->msg_txpeer->lp_ni == ni);
1041 LASSERT(msg2->msg_tx_delayed);
1043 (void) lnet_post_send_locked(msg2, 1);
1047 if (msg->msg_peertxcredit) {
1048 /* give back peer txcredits */
1049 msg->msg_peertxcredit = 0;
1051 LASSERT((txpeer->lp_txcredits < 0) ==
1052 !cfs_list_empty(&txpeer->lp_txq));
1054 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1055 LASSERT (txpeer->lp_txqnob >= 0);
1057 txpeer->lp_txcredits++;
1058 if (txpeer->lp_txcredits <= 0) {
1059 msg2 = cfs_list_entry(txpeer->lp_txq.next,
1060 lnet_msg_t, msg_list);
1061 cfs_list_del(&msg2->msg_list);
1063 LASSERT(msg2->msg_txpeer == txpeer);
1064 LASSERT(msg2->msg_tx_delayed);
1066 (void) lnet_post_send_locked(msg2, 1);
1070 if (txpeer != NULL) {
1071 msg->msg_txpeer = NULL;
1072 lnet_peer_decref_locked(txpeer);
1077 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1079 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1083 if (msg->msg_rtrcredit) {
1084 /* give back global router credits */
1086 lnet_rtrbufpool_t *rbp;
1088 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1089 * there until it gets one allocated, or aborts the wait
1091 LASSERT (msg->msg_kiov != NULL);
1093 rb = cfs_list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1095 LASSERT (rbp == lnet_msg2bufpool(msg));
1097 msg->msg_kiov = NULL;
1098 msg->msg_rtrcredit = 0;
1100 LASSERT((rbp->rbp_credits < 0) ==
1101 !cfs_list_empty(&rbp->rbp_msgs));
1102 LASSERT((rbp->rbp_credits > 0) ==
1103 !cfs_list_empty(&rbp->rbp_bufs));
1105 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1107 if (rbp->rbp_credits <= 0) {
1108 msg2 = cfs_list_entry(rbp->rbp_msgs.next,
1109 lnet_msg_t, msg_list);
1110 cfs_list_del(&msg2->msg_list);
1112 (void) lnet_post_routed_recv_locked(msg2, 1);
1116 if (msg->msg_peerrtrcredit) {
1117 /* give back peer router credits */
1118 msg->msg_peerrtrcredit = 0;
1120 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1121 !cfs_list_empty(&rxpeer->lp_rtrq));
1123 rxpeer->lp_rtrcredits++;
1124 if (rxpeer->lp_rtrcredits <= 0) {
1125 msg2 = cfs_list_entry(rxpeer->lp_rtrq.next,
1126 lnet_msg_t, msg_list);
1127 cfs_list_del(&msg2->msg_list);
1129 (void) lnet_post_routed_recv_locked(msg2, 1);
1133 LASSERT (!msg->msg_rtrcredit);
1134 LASSERT (!msg->msg_peerrtrcredit);
1136 if (rxpeer != NULL) {
1137 msg->msg_rxpeer = NULL;
1138 lnet_peer_decref_locked(rxpeer);
1142 static lnet_peer_t *
1143 lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target)
1145 lnet_remotenet_t *rnet;
1147 lnet_route_t *rtr_best;
1148 lnet_route_t *rtr_last;
1149 struct lnet_peer *lp_best;
1150 struct lnet_peer *lp;
1153 rnet = lnet_find_net_locked(LNET_NIDNET(target));
1158 rtr_best = rtr_last = NULL;
1159 cfs_list_for_each_entry(rtr, &rnet->lrn_routes, lr_list) {
1160 lp = rtr->lr_gateway;
1162 if (!lp->lp_alive || /* gateway is down */
1163 (lp->lp_ping_version == LNET_PROTO_PING_VERSION &&
1164 rtr->lr_downis != 0)) /* NI to target is down */
1167 if (ni != NULL && lp->lp_ni != ni)
1170 if (lp_best == NULL) {
1171 rtr_best = rtr_last = rtr;
1176 rc = lnet_compare_routes(rtr, rtr_best);
1184 if (rtr_best != NULL) {
1185 /* Place selected route at the end of the route list to ensure
1186 * fairness; everything else being equal... */
1187 cfs_list_del(&rtr_best->lr_list);
1188 cfs_list_add_tail(&rtr_best->lr_list, &rnet->lrn_routes);
1195 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg)
1197 lnet_nid_t dst_nid = msg->msg_target.nid;
1199 lnet_ni_t *local_ni;
1203 LASSERT (msg->msg_txpeer == NULL);
1204 LASSERT (!msg->msg_sending);
1205 LASSERT (!msg->msg_target_is_router);
1206 LASSERT (!msg->msg_receiving);
1208 msg->msg_sending = 1;
1210 /* NB! ni != NULL == interface pre-determined (ACK/REPLY) */
1214 if (the_lnet.ln_shutdown) {
1219 if (src_nid == LNET_NID_ANY) {
1222 src_ni = lnet_nid2ni_locked(src_nid);
1223 if (src_ni == NULL) {
1225 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1226 "local nid\n", libcfs_nid2str(dst_nid),
1227 libcfs_nid2str(src_nid));
1230 LASSERT (!msg->msg_routing);
1233 lnet_msg_commit(msg, 1);
1234 /* Is this for someone on a local network? */
1235 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid));
1237 if (local_ni != NULL) {
1238 if (src_ni == NULL) {
1240 src_nid = src_ni->ni_nid;
1241 } else if (src_ni == local_ni) {
1242 lnet_ni_decref_locked(local_ni);
1244 lnet_ni_decref_locked(local_ni);
1245 lnet_ni_decref_locked(src_ni);
1247 LCONSOLE_WARN("No route to %s via from %s\n",
1248 libcfs_nid2str(dst_nid),
1249 libcfs_nid2str(src_nid));
1253 LASSERT (src_nid != LNET_NID_ANY);
1255 if (!msg->msg_routing)
1256 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1258 if (src_ni == the_lnet.ln_loni) {
1259 /* No send credit hassles with LOLND */
1261 lnet_ni_send(src_ni, msg);
1262 lnet_ni_decref(src_ni);
1266 rc = lnet_nid2peer_locked(&lp, dst_nid);
1267 lnet_ni_decref_locked(src_ni); /* lp has ref on src_ni; lose mine */
1270 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1271 libcfs_nid2str(dst_nid));
1272 /* ENOMEM or shutting down */
1275 LASSERT (lp->lp_ni == src_ni);
1281 * - once application finishes computation, check here to update
1282 * router states before it waits for pending IO in LNetEQPoll
1283 * - recursion breaker: router checker sends no message
1284 * to remote networks */
1285 if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING)
1286 lnet_router_checker();
1290 /* sending to a remote network */
1291 lp = lnet_find_route_locked(src_ni, dst_nid);
1294 lnet_ni_decref_locked(src_ni);
1297 LCONSOLE_WARN("No route to %s via %s "
1298 "(all routers down)\n",
1299 libcfs_id2str(msg->msg_target),
1300 libcfs_nid2str(src_nid));
1301 return -EHOSTUNREACH;
1304 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1305 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1306 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1308 if (src_ni == NULL) {
1310 src_nid = src_ni->ni_nid;
1312 LASSERT (src_ni == lp->lp_ni);
1313 lnet_ni_decref_locked(src_ni);
1316 lnet_peer_addref_locked(lp);
1318 LASSERT (src_nid != LNET_NID_ANY);
1320 if (!msg->msg_routing) {
1321 /* I'm the source and now I know which NI to send on */
1322 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1325 msg->msg_target_is_router = 1;
1326 msg->msg_target.nid = lp->lp_nid;
1327 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1330 /* 'lp' is our best choice of peer */
1332 LASSERT (!msg->msg_peertxcredit);
1333 LASSERT (!msg->msg_txcredit);
1334 LASSERT (msg->msg_txpeer == NULL);
1336 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1338 rc = lnet_post_send_locked(msg, 0);
1341 if (rc == EHOSTUNREACH)
1342 return -EHOSTUNREACH;
1345 lnet_ni_send(src_ni, msg);
1351 lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob)
1354 the_lnet.ln_counters.drop_count++;
1355 the_lnet.ln_counters.drop_length += nob;
1358 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1362 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1364 lnet_hdr_t *hdr = &msg->msg_hdr;
1366 if (msg->msg_wanted != 0)
1367 lnet_setpayloadbuffer(msg);
1369 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1371 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1372 * it back into the ACK during lnet_finalize() */
1373 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1374 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1376 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1377 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1381 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1385 lnet_hdr_t *hdr = &msg->msg_hdr;
1386 unsigned int rlength = hdr->payload_length;
1387 lnet_process_id_t src= {0};
1389 src.nid = hdr->src_nid;
1390 src.pid = hdr->src_pid;
1392 /* Convert put fields to host byte order */
1393 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1394 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1395 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1397 index = hdr->msg.put.ptl_index;
1399 msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL;
1402 rc = lnet_ptl_match_md(index, LNET_MD_OP_PUT, src,
1403 rlength, hdr->msg.put.offset,
1404 hdr->msg.put.match_bits, msg);
1409 case LNET_MATCHMD_OK:
1410 lnet_recv_put(ni, msg);
1413 case LNET_MATCHMD_NONE:
1414 if (msg->msg_rx_delayed) /* attached on delayed list */
1417 rc = lnet_ni_eager_recv(ni, msg);
1422 case LNET_MATCHMD_DROP:
1423 CNETERR("Dropping PUT from %s portal %d match "LPU64
1424 " offset %d length %d: %d\n",
1425 libcfs_id2str(src), index,
1426 hdr->msg.put.match_bits,
1427 hdr->msg.put.offset, rlength, rc);
1429 return ENOENT; /* +ve: OK but no match */
1434 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1436 lnet_hdr_t *hdr = &msg->msg_hdr;
1437 lnet_process_id_t src = {0};
1438 lnet_handle_wire_t reply_wmd;
1441 src.nid = hdr->src_nid;
1442 src.pid = hdr->src_pid;
1444 /* Convert get fields to host byte order */
1445 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1446 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1447 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1448 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1450 rc = lnet_ptl_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src,
1451 hdr->msg.get.sink_length,
1452 hdr->msg.get.src_offset,
1453 hdr->msg.get.match_bits, msg);
1454 if (rc == LNET_MATCHMD_DROP) {
1455 CNETERR("Dropping GET from %s portal %d match "LPU64
1456 " offset %d length %d\n",
1458 hdr->msg.get.ptl_index,
1459 hdr->msg.get.match_bits,
1460 hdr->msg.get.src_offset,
1461 hdr->msg.get.sink_length);
1462 return ENOENT; /* +ve: OK but no match */
1465 LASSERT (rc == LNET_MATCHMD_OK);
1467 lnet_build_msg_event(msg, LNET_EVENT_GET);
1469 reply_wmd = hdr->msg.get.return_wmd;
1471 lnet_prep_send(msg, LNET_MSG_REPLY, src,
1472 msg->msg_offset, msg->msg_wanted);
1474 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1477 /* The LND completes the REPLY from her recv procedure */
1478 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1479 msg->msg_offset, msg->msg_len, msg->msg_len);
1483 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1484 msg->msg_receiving = 0;
1486 rc = lnet_send(ni->ni_nid, msg);
1488 /* didn't get as far as lnet_ni_send() */
1489 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1490 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc);
1492 lnet_finalize(ni, msg, rc);
1499 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1501 void *private = msg->msg_private;
1502 lnet_hdr_t *hdr = &msg->msg_hdr;
1503 lnet_process_id_t src = {0};
1510 src.nid = hdr->src_nid;
1511 src.pid = hdr->src_pid;
1513 /* NB handles only looked up by creator (no flips) */
1514 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1515 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1516 CNETERR("%s: Dropping REPLY from %s for %s "
1517 "MD "LPX64"."LPX64"\n",
1518 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1519 (md == NULL) ? "invalid" : "inactive",
1520 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1521 hdr->msg.reply.dst_wmd.wh_object_cookie);
1522 if (md != NULL && md->md_me != NULL)
1523 CERROR("REPLY MD also attached to portal %d\n",
1524 md->md_me->me_portal);
1527 return ENOENT; /* +ve: OK but no match */
1530 LASSERT (md->md_offset == 0);
1532 rlength = hdr->payload_length;
1533 mlength = MIN(rlength, (int)md->md_length);
1535 if (mlength < rlength &&
1536 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1537 CNETERR("%s: Dropping REPLY from %s length %d "
1538 "for MD "LPX64" would overflow (%d)\n",
1539 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1540 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1543 return ENOENT; /* +ve: OK but no match */
1546 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1547 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1548 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1550 lnet_msg_attach_md(msg, md, 0, mlength);
1553 lnet_setpayloadbuffer(msg);
1557 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
1559 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1564 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1566 lnet_hdr_t *hdr = &msg->msg_hdr;
1567 lnet_process_id_t src = {0};
1570 src.nid = hdr->src_nid;
1571 src.pid = hdr->src_pid;
1573 /* Convert ack fields to host byte order */
1574 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1575 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1579 /* NB handles only looked up by creator (no flips) */
1580 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1581 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1582 /* Don't moan; this is expected */
1584 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1585 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1586 (md == NULL) ? "invalid" : "inactive",
1587 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1588 hdr->msg.ack.dst_wmd.wh_object_cookie);
1589 if (md != NULL && md->md_me != NULL)
1590 CERROR("Source MD also attached to portal %d\n",
1591 md->md_me->me_portal);
1594 return ENOENT; /* +ve! */
1597 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1598 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1599 hdr->msg.ack.dst_wmd.wh_object_cookie);
1601 lnet_msg_attach_md(msg, md, 0, 0);
1605 lnet_build_msg_event(msg, LNET_EVENT_ACK);
1607 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
1612 lnet_msgtyp2str (int type)
1621 case LNET_MSG_REPLY:
1623 case LNET_MSG_HELLO:
1626 return ("<UNKNOWN>");
1631 lnet_print_hdr(lnet_hdr_t * hdr)
1633 lnet_process_id_t src = {0};
1634 lnet_process_id_t dst = {0};
1635 char *type_str = lnet_msgtyp2str (hdr->type);
1637 src.nid = hdr->src_nid;
1638 src.pid = hdr->src_pid;
1640 dst.nid = hdr->dest_nid;
1641 dst.pid = hdr->dest_pid;
1643 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
1644 CWARN(" From %s\n", libcfs_id2str(src));
1645 CWARN(" To %s\n", libcfs_id2str(dst));
1647 switch (hdr->type) {
1652 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
1653 "match bits "LPU64"\n",
1654 hdr->msg.put.ptl_index,
1655 hdr->msg.put.ack_wmd.wh_interface_cookie,
1656 hdr->msg.put.ack_wmd.wh_object_cookie,
1657 hdr->msg.put.match_bits);
1658 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
1659 hdr->payload_length, hdr->msg.put.offset,
1660 hdr->msg.put.hdr_data);
1664 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
1665 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
1666 hdr->msg.get.return_wmd.wh_interface_cookie,
1667 hdr->msg.get.return_wmd.wh_object_cookie,
1668 hdr->msg.get.match_bits);
1669 CWARN(" Length %d, src offset %d\n",
1670 hdr->msg.get.sink_length,
1671 hdr->msg.get.src_offset);
1675 CWARN(" dst md "LPX64"."LPX64", "
1676 "manipulated length %d\n",
1677 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1678 hdr->msg.ack.dst_wmd.wh_object_cookie,
1679 hdr->msg.ack.mlength);
1682 case LNET_MSG_REPLY:
1683 CWARN(" dst md "LPX64"."LPX64", "
1685 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1686 hdr->msg.reply.dst_wmd.wh_object_cookie,
1687 hdr->payload_length);
1693 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
1694 void *private, int rdma_req)
1699 lnet_pid_t dest_pid;
1700 lnet_nid_t dest_nid;
1702 __u32 payload_length;
1705 LASSERT (!cfs_in_interrupt ());
1707 type = le32_to_cpu(hdr->type);
1708 src_nid = le64_to_cpu(hdr->src_nid);
1709 dest_nid = le64_to_cpu(hdr->dest_nid);
1710 dest_pid = le32_to_cpu(hdr->dest_pid);
1711 payload_length = le32_to_cpu(hdr->payload_length);
1713 for_me = (ni->ni_nid == dest_nid);
1718 if (payload_length > 0) {
1719 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
1720 libcfs_nid2str(from_nid),
1721 libcfs_nid2str(src_nid),
1722 lnet_msgtyp2str(type), payload_length);
1728 case LNET_MSG_REPLY:
1729 if (payload_length > (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
1730 CERROR("%s, src %s: bad %s payload %d "
1731 "(%d max expected)\n",
1732 libcfs_nid2str(from_nid),
1733 libcfs_nid2str(src_nid),
1734 lnet_msgtyp2str(type),
1736 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
1742 CERROR("%s, src %s: Bad message type 0x%x\n",
1743 libcfs_nid2str(from_nid),
1744 libcfs_nid2str(src_nid), type);
1748 if (the_lnet.ln_routing &&
1749 ni->ni_last_alive != cfs_time_current_sec()) {
1752 /* NB: so far here is the only place to set NI status to "up */
1753 ni->ni_last_alive = cfs_time_current_sec();
1754 if (ni->ni_status != NULL &&
1755 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
1756 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
1760 /* Regard a bad destination NID as a protocol error. Senders should
1761 * know what they're doing; if they don't they're misconfigured, buggy
1762 * or malicious so we chop them off at the knees :) */
1765 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
1766 /* should have gone direct */
1767 CERROR ("%s, src %s: Bad dest nid %s "
1768 "(should have been sent direct)\n",
1769 libcfs_nid2str(from_nid),
1770 libcfs_nid2str(src_nid),
1771 libcfs_nid2str(dest_nid));
1775 if (lnet_islocalnid(dest_nid)) {
1776 /* dest is another local NI; sender should have used
1777 * this node's NID on its own network */
1778 CERROR ("%s, src %s: Bad dest nid %s "
1779 "(it's my nid but on a different network)\n",
1780 libcfs_nid2str(from_nid),
1781 libcfs_nid2str(src_nid),
1782 libcfs_nid2str(dest_nid));
1786 if (rdma_req && type == LNET_MSG_GET) {
1787 CERROR ("%s, src %s: Bad optimized GET for %s "
1788 "(final destination must be me)\n",
1789 libcfs_nid2str(from_nid),
1790 libcfs_nid2str(src_nid),
1791 libcfs_nid2str(dest_nid));
1795 if (!the_lnet.ln_routing) {
1796 CERROR ("%s, src %s: Dropping message for %s "
1797 "(routing not enabled)\n",
1798 libcfs_nid2str(from_nid),
1799 libcfs_nid2str(src_nid),
1800 libcfs_nid2str(dest_nid));
1805 /* Message looks OK; we're not going to return an error, so we MUST
1806 * call back lnd_recv() come what may... */
1808 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
1809 fail_peer (src_nid, 0)) /* shall we now? */
1811 CERROR("%s, src %s: Dropping %s to simulate failure\n",
1812 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1813 lnet_msgtyp2str(type));
1817 msg = lnet_msg_alloc();
1819 CERROR("%s, src %s: Dropping %s (out of memory)\n",
1820 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1821 lnet_msgtyp2str(type));
1825 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
1827 msg->msg_type = type;
1828 msg->msg_private = private;
1829 msg->msg_receiving = 1;
1830 msg->msg_len = msg->msg_wanted = payload_length;
1831 msg->msg_offset = 0;
1832 msg->msg_hdr = *hdr;
1833 /* for building message event */
1834 msg->msg_from = from_nid;
1837 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid);
1840 CERROR("%s, src %s: Dropping %s "
1841 "(error %d looking up sender)\n",
1842 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1843 lnet_msgtyp2str(type), rc);
1848 lnet_msg_commit(msg, 0);
1855 msg->msg_target.pid = dest_pid;
1856 msg->msg_target.nid = dest_nid;
1857 msg->msg_routing = 1;
1858 msg->msg_offset = 0;
1861 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
1862 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
1863 rc = lnet_ni_eager_recv_locked(ni, msg);
1869 rc = lnet_post_routed_recv_locked(msg, 0);
1873 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1874 0, payload_length, payload_length);
1878 /* convert common msg->hdr fields to host byteorder */
1879 msg->msg_hdr.type = type;
1880 msg->msg_hdr.src_nid = src_nid;
1881 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
1882 msg->msg_hdr.dest_nid = dest_nid;
1883 msg->msg_hdr.dest_pid = dest_pid;
1884 msg->msg_hdr.payload_length = payload_length;
1888 rc = lnet_parse_ack(ni, msg);
1891 rc = lnet_parse_put(ni, msg);
1894 rc = lnet_parse_get(ni, msg, rdma_req);
1896 case LNET_MSG_REPLY:
1897 rc = lnet_parse_reply(ni, msg);
1902 goto free_drop; /* prevent an unused label if !kernel */
1908 LASSERT (rc == ENOENT);
1911 LASSERT(msg->msg_md == NULL);
1912 lnet_finalize(ni, msg, rc);
1915 lnet_drop_message(ni, private, payload_length);
1920 lnet_drop_delayed_msg_list(cfs_list_t *head, char *reason)
1922 while (!cfs_list_empty(head)) {
1923 lnet_process_id_t id = {0};
1926 msg = cfs_list_entry(head->next, lnet_msg_t, msg_list);
1927 cfs_list_del(&msg->msg_list);
1929 id.nid = msg->msg_hdr.src_nid;
1930 id.pid = msg->msg_hdr.src_pid;
1932 LASSERT(msg->msg_md == NULL);
1933 LASSERT(msg->msg_rx_delayed);
1934 LASSERT(msg->msg_rxpeer != NULL);
1935 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
1937 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
1938 " offset %d length %d: %s\n",
1940 msg->msg_hdr.msg.put.ptl_index,
1941 msg->msg_hdr.msg.put.match_bits,
1942 msg->msg_hdr.msg.put.offset,
1943 msg->msg_hdr.payload_length, reason);
1945 /* NB I can't drop msg's ref on msg_rxpeer until after I've
1946 * called lnet_drop_message(), so I just hang onto msg as well
1947 * until that's done */
1949 lnet_drop_message(msg->msg_rxpeer->lp_ni,
1950 msg->msg_private, msg->msg_len);
1953 lnet_peer_decref_locked(msg->msg_rxpeer);
1961 lnet_recv_delayed_msg_list(cfs_list_t *head)
1963 while (!cfs_list_empty(head)) {
1965 lnet_process_id_t id;
1967 msg = cfs_list_entry(head->next, lnet_msg_t, msg_list);
1968 cfs_list_del(&msg->msg_list);
1970 /* md won't disappear under me, since each msg
1971 * holds a ref on it */
1973 id.nid = msg->msg_hdr.src_nid;
1974 id.pid = msg->msg_hdr.src_pid;
1976 LASSERT(msg->msg_rx_delayed);
1977 LASSERT(msg->msg_md != NULL);
1978 LASSERT(msg->msg_rxpeer != NULL);
1979 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
1981 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
1982 "match "LPU64" offset %d length %d.\n",
1983 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
1984 msg->msg_hdr.msg.put.match_bits,
1985 msg->msg_hdr.msg.put.offset,
1986 msg->msg_hdr.payload_length);
1988 lnet_recv_put(msg->msg_rxpeer->lp_ni, msg);
1993 * Initiate an asynchronous PUT operation.
1995 * There are several events associated with a PUT: completion of the send on
1996 * the initiator node (LNET_EVENT_SEND), and when the send completes
1997 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
1998 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
1999 * used at the target node to indicate the completion of incoming data
2002 * The local events will be logged in the EQ associated with the MD pointed to
2003 * by \a mdh handle. Using a MD without an associated EQ results in these
2004 * events being discarded. In this case, the caller must have another
2005 * mechanism (e.g., a higher level protocol) for determining when it is safe
2006 * to modify the memory region associated with the MD.
2008 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2009 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2011 * \param self Indicates the NID of a local interface through which to send
2012 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2013 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2014 * must be "free floating" (See LNetMDBind()).
2015 * \param ack Controls whether an acknowledgment is requested.
2016 * Acknowledgments are only sent when they are requested by the initiating
2017 * process and the target MD enables them.
2018 * \param target A process identifier for the target process.
2019 * \param portal The index in the \a target's portal table.
2020 * \param match_bits The match bits to use for MD selection at the target
2022 * \param offset The offset into the target MD (only used when the target
2023 * MD has the LNET_MD_MANAGE_REMOTE option set).
2024 * \param hdr_data 64 bits of user data that can be included in the message
2025 * header. This data is written to an event queue entry at the target if an
2026 * EQ is present on the matching MD.
2028 * \retval 0 Success, and only in this case events will be generated
2029 * and logged to EQ (if it exists).
2030 * \retval -EIO Simulated failure.
2031 * \retval -ENOMEM Memory allocation failure.
2032 * \retval -ENOENT Invalid MD object.
2034 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2037 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2038 lnet_process_id_t target, unsigned int portal,
2039 __u64 match_bits, unsigned int offset,
2046 LASSERT (the_lnet.ln_init);
2047 LASSERT (the_lnet.ln_refcount > 0);
2049 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2050 fail_peer (target.nid, 1)) /* shall we now? */
2052 CERROR("Dropping PUT to %s: simulated failure\n",
2053 libcfs_id2str(target));
2057 msg = lnet_msg_alloc();
2059 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2060 libcfs_id2str(target));
2063 msg->msg_vmflush = !!cfs_memory_pressure_get();
2067 md = lnet_handle2md(&mdh);
2068 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2069 lnet_msg_free_locked(msg);
2071 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2072 match_bits, portal, libcfs_id2str(target),
2073 md == NULL ? -1 : md->md_threshold);
2074 if (md != NULL && md->md_me != NULL)
2075 CERROR("Source MD also attached to portal %d\n",
2076 md->md_me->me_portal);
2082 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2084 lnet_msg_attach_md(msg, md, 0, 0);
2086 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2088 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2089 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2090 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2091 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2093 /* NB handles only looked up by creator (no flips) */
2094 if (ack == LNET_ACK_REQ) {
2095 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2096 the_lnet.ln_interface_cookie;
2097 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2098 md->md_lh.lh_cookie;
2100 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2101 LNET_WIRE_HANDLE_COOKIE_NONE;
2102 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2103 LNET_WIRE_HANDLE_COOKIE_NONE;
2108 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2110 rc = lnet_send(self, msg);
2112 CNETERR( "Error sending PUT to %s: %d\n",
2113 libcfs_id2str(target), rc);
2114 lnet_finalize (NULL, msg, rc);
2117 /* completion will be signalled by an event */
2122 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2124 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2125 * returns a msg for the LND to pass to lnet_finalize() when the sink
2126 * data has been received.
2128 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2129 * lnet_finalize() is called on it, so the LND must call this first */
2131 lnet_msg_t *msg = lnet_msg_alloc();
2132 lnet_libmd_t *getmd = getmsg->msg_md;
2133 lnet_process_id_t peer_id = getmsg->msg_target;
2135 LASSERT (!getmsg->msg_target_is_router);
2136 LASSERT (!getmsg->msg_routing);
2140 LASSERT (getmd->md_refcount > 0);
2143 CERROR ("%s: Dropping REPLY from %s: can't allocate msg\n",
2144 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2148 if (getmd->md_threshold == 0) {
2149 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2150 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2155 LASSERT (getmd->md_offset == 0);
2157 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2158 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2160 /* setup information for lnet_build_msg_event */
2161 msg->msg_from = peer_id.nid;
2162 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2163 msg->msg_hdr.src_nid = peer_id.nid;
2164 msg->msg_hdr.payload_length = getmd->md_length;
2165 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2167 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2168 lnet_msg_commit(msg, 0);
2172 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2177 lnet_msg_free_locked(msg);
2179 the_lnet.ln_counters.drop_count++;
2180 the_lnet.ln_counters.drop_length += getmd->md_length;
2188 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2190 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2191 * completed and I know it. */
2192 LASSERT (reply != NULL);
2193 LASSERT (reply->msg_type == LNET_MSG_GET);
2194 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2196 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2197 * the end of my buffer, I might as well be dead. */
2198 LASSERT (len <= reply->msg_ev.mlength);
2200 reply->msg_ev.mlength = len;
2204 * Initiate an asynchronous GET operation.
2206 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2207 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2208 * the target node in the REPLY has been written to local MD.
2210 * On the target node, an LNET_EVENT_GET is logged when the GET request
2211 * arrives and is accepted into a MD.
2213 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2214 * \param mdh A handle for the MD that describes the memory into which the
2215 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2217 * \retval 0 Success, and only in this case events will be generated
2218 * and logged to EQ (if it exists) of the MD.
2219 * \retval -EIO Simulated failure.
2220 * \retval -ENOMEM Memory allocation failure.
2221 * \retval -ENOENT Invalid MD object.
2224 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2225 lnet_process_id_t target, unsigned int portal,
2226 __u64 match_bits, unsigned int offset)
2232 LASSERT (the_lnet.ln_init);
2233 LASSERT (the_lnet.ln_refcount > 0);
2235 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2236 fail_peer (target.nid, 1)) /* shall we now? */
2238 CERROR("Dropping GET to %s: simulated failure\n",
2239 libcfs_id2str(target));
2243 msg = lnet_msg_alloc();
2245 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2246 libcfs_id2str(target));
2252 md = lnet_handle2md(&mdh);
2253 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2254 lnet_msg_free_locked(msg);
2256 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2257 match_bits, portal, libcfs_id2str(target),
2258 md == NULL ? -1 : md->md_threshold);
2259 if (md != NULL && md->md_me != NULL)
2260 CERROR("REPLY MD also attached to portal %d\n",
2261 md->md_me->me_portal);
2267 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2269 lnet_msg_attach_md(msg, md, 0, 0);
2271 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2273 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2274 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2275 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2276 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2278 /* NB handles only looked up by creator (no flips) */
2279 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2280 the_lnet.ln_interface_cookie;
2281 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2282 md->md_lh.lh_cookie;
2286 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2288 rc = lnet_send(self, msg);
2290 CNETERR( "Error sending GET to %s: %d\n",
2291 libcfs_id2str(target), rc);
2292 lnet_finalize (NULL, msg, rc);
2295 /* completion will be signalled by an event */
2300 * Calculate distance to node at \a dstnid.
2302 * \param dstnid Target NID.
2303 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2305 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2308 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2309 * local_nid_dist_zero is set, which is the default.
2310 * \retval positives Distance to target NID, i.e. number of hops plus one.
2311 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2314 LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2318 lnet_remotenet_t *rnet;
2319 __u32 dstnet = LNET_NIDNET(dstnid);
2323 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2324 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2325 * keep order 0 free for 0@lo and order 1 free for a local NID
2328 LASSERT (the_lnet.ln_init);
2329 LASSERT (the_lnet.ln_refcount > 0);
2333 cfs_list_for_each (e, &the_lnet.ln_nis) {
2334 ni = cfs_list_entry(e, lnet_ni_t, ni_list);
2336 if (ni->ni_nid == dstnid) {
2337 if (srcnidp != NULL)
2339 if (orderp != NULL) {
2340 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2347 return local_nid_dist_zero ? 0 : 1;
2350 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2351 if (srcnidp != NULL)
2352 *srcnidp = ni->ni_nid;
2362 cfs_list_for_each (e, &the_lnet.ln_remote_nets) {
2363 rnet = cfs_list_entry(e, lnet_remotenet_t, lrn_list);
2365 if (rnet->lrn_net == dstnet) {
2366 lnet_route_t *route;
2367 lnet_route_t *shortest = NULL;
2369 LASSERT (!cfs_list_empty(&rnet->lrn_routes));
2371 cfs_list_for_each_entry(route, &rnet->lrn_routes,
2373 if (shortest == NULL ||
2374 route->lr_hops < shortest->lr_hops)
2378 LASSERT (shortest != NULL);
2379 hops = shortest->lr_hops;
2380 if (srcnidp != NULL)
2381 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2391 return -EHOSTUNREACH;
2395 * Set the number of asynchronous messages expected from a target process.
2397 * This function is only meaningful for userspace callers. It's a no-op when
2398 * called from kernel.
2400 * Asynchronous messages are those that can come from a target when the
2401 * userspace process is not waiting for IO to complete; e.g., AST callbacks
2402 * from Lustre servers. Specifying the expected number of such messages
2403 * allows them to be eagerly received when user process is not running in
2404 * LNet; otherwise network errors may occur.
2406 * \param id Process ID of the target process.
2407 * \param nasync Number of asynchronous messages expected from the target.
2409 * \return 0 on success, and an error code otherwise.
2412 LNetSetAsync(lnet_process_id_t id, int nasync)
2418 lnet_remotenet_t *rnet;
2420 lnet_route_t *route;
2427 /* Target on a local network? */
2428 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2430 if (ni->ni_lnd->lnd_setasync != NULL)
2431 rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2436 /* Target on a remote network: apply to routers */
2438 LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2443 /* Snapshot all the router NIDs */
2445 rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2447 cfs_list_for_each(tmp, &rnet->lrn_routes) {
2448 if (nnids == maxnids) {
2450 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2455 route = cfs_list_entry(tmp, lnet_route_t, lr_list);
2456 nids[nnids++] = route->lr_gateway->lp_nid;
2461 /* set async on all the routers */
2462 while (nnids-- > 0) {
2463 id.pid = LUSTRE_SRV_LNET_PID;
2464 id.nid = nids[nnids];
2466 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2470 if (ni->ni_lnd->lnd_setasync != NULL) {
2471 rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2478 LIBCFS_FREE(nids, maxnids * sizeof(*nids));