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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/lnet/lib-move.c
34 * Data movement routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <lnet/lib-lnet.h>
40 #include <linux/nsproxy.h>
41 #include <net/net_namespace.h>
43 static int local_nid_dist_zero = 1;
44 module_param(local_nid_dist_zero, int, 0444);
45 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
48 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
50 struct lnet_test_peer *tp;
52 struct list_head *next;
53 struct list_head cull;
55 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
57 /* Adding a new entry */
58 LIBCFS_ALLOC(tp, sizeof(*tp));
63 tp->tp_threshold = threshold;
66 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
71 /* removing entries */
72 INIT_LIST_HEAD(&cull);
76 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
77 tp = list_entry(el, struct lnet_test_peer, tp_list);
79 if (tp->tp_threshold == 0 || /* needs culling anyway */
80 nid == LNET_NID_ANY || /* removing all entries */
81 tp->tp_nid == nid) { /* matched this one */
82 list_del(&tp->tp_list);
83 list_add(&tp->tp_list, &cull);
89 while (!list_empty(&cull)) {
90 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
92 list_del(&tp->tp_list);
93 LIBCFS_FREE(tp, sizeof(*tp));
99 fail_peer (lnet_nid_t nid, int outgoing)
101 struct lnet_test_peer *tp;
102 struct list_head *el;
103 struct list_head *next;
104 struct list_head cull;
107 INIT_LIST_HEAD(&cull);
109 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
112 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
113 tp = list_entry(el, struct lnet_test_peer, tp_list);
115 if (tp->tp_threshold == 0) {
118 /* only cull zombies on outgoing tests,
119 * since we may be at interrupt priority on
120 * incoming messages. */
121 list_del(&tp->tp_list);
122 list_add(&tp->tp_list, &cull);
127 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
128 nid == tp->tp_nid) { /* fail this peer */
131 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
134 tp->tp_threshold == 0) {
136 list_del(&tp->tp_list);
137 list_add(&tp->tp_list, &cull);
146 while (!list_empty(&cull)) {
147 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
148 list_del(&tp->tp_list);
150 LIBCFS_FREE(tp, sizeof(*tp));
157 lnet_iov_nob(unsigned int niov, struct kvec *iov)
159 unsigned int nob = 0;
161 LASSERT(niov == 0 || iov != NULL);
163 nob += (iov++)->iov_len;
167 EXPORT_SYMBOL(lnet_iov_nob);
170 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
171 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
174 /* NB diov, siov are READ-ONLY */
175 unsigned int this_nob;
180 /* skip complete frags before 'doffset' */
182 while (doffset >= diov->iov_len) {
183 doffset -= diov->iov_len;
189 /* skip complete frags before 'soffset' */
191 while (soffset >= siov->iov_len) {
192 soffset -= siov->iov_len;
201 this_nob = MIN(diov->iov_len - doffset,
202 siov->iov_len - soffset);
203 this_nob = MIN(this_nob, nob);
205 memcpy((char *)diov->iov_base + doffset,
206 (char *)siov->iov_base + soffset, this_nob);
209 if (diov->iov_len > doffset + this_nob) {
217 if (siov->iov_len > soffset + this_nob) {
226 EXPORT_SYMBOL(lnet_copy_iov2iov);
229 lnet_extract_iov(int dst_niov, struct kvec *dst,
230 int src_niov, struct kvec *src,
231 unsigned int offset, unsigned int len)
233 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
234 * for exactly 'len' bytes, and return the number of entries.
235 * NB not destructive to 'src' */
236 unsigned int frag_len;
239 if (len == 0) /* no data => */
240 return (0); /* no frags */
242 LASSERT(src_niov > 0);
243 while (offset >= src->iov_len) { /* skip initial frags */
244 offset -= src->iov_len;
247 LASSERT(src_niov > 0);
252 LASSERT(src_niov > 0);
253 LASSERT((int)niov <= dst_niov);
255 frag_len = src->iov_len - offset;
256 dst->iov_base = ((char *)src->iov_base) + offset;
258 if (len <= frag_len) {
263 dst->iov_len = frag_len;
273 EXPORT_SYMBOL(lnet_extract_iov);
277 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
279 unsigned int nob = 0;
281 LASSERT(niov == 0 || kiov != NULL);
283 nob += (kiov++)->kiov_len;
287 EXPORT_SYMBOL(lnet_kiov_nob);
290 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
291 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
294 /* NB diov, siov are READ-ONLY */
295 unsigned int this_nob;
302 LASSERT (!in_interrupt ());
305 while (doffset >= diov->kiov_len) {
306 doffset -= diov->kiov_len;
313 while (soffset >= siov->kiov_len) {
314 soffset -= siov->kiov_len;
323 this_nob = MIN(diov->kiov_len - doffset,
324 siov->kiov_len - soffset);
325 this_nob = MIN(this_nob, nob);
328 daddr = ((char *)kmap(diov->kiov_page)) +
329 diov->kiov_offset + doffset;
331 saddr = ((char *)kmap(siov->kiov_page)) +
332 siov->kiov_offset + soffset;
334 /* Vanishing risk of kmap deadlock when mapping 2 pages.
335 * However in practice at least one of the kiovs will be mapped
336 * kernel pages and the map/unmap will be NOOPs */
338 memcpy (daddr, saddr, this_nob);
341 if (diov->kiov_len > doffset + this_nob) {
345 kunmap(diov->kiov_page);
352 if (siov->kiov_len > soffset + this_nob) {
356 kunmap(siov->kiov_page);
365 kunmap(diov->kiov_page);
367 kunmap(siov->kiov_page);
369 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
372 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
373 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
376 /* NB iov, kiov are READ-ONLY */
377 unsigned int this_nob;
383 LASSERT (!in_interrupt ());
386 while (iovoffset >= iov->iov_len) {
387 iovoffset -= iov->iov_len;
394 while (kiovoffset >= kiov->kiov_len) {
395 kiovoffset -= kiov->kiov_len;
404 this_nob = MIN(iov->iov_len - iovoffset,
405 kiov->kiov_len - kiovoffset);
406 this_nob = MIN(this_nob, nob);
409 addr = ((char *)kmap(kiov->kiov_page)) +
410 kiov->kiov_offset + kiovoffset;
412 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
415 if (iov->iov_len > iovoffset + this_nob) {
416 iovoffset += this_nob;
423 if (kiov->kiov_len > kiovoffset + this_nob) {
425 kiovoffset += this_nob;
427 kunmap(kiov->kiov_page);
437 kunmap(kiov->kiov_page);
439 EXPORT_SYMBOL(lnet_copy_kiov2iov);
442 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
443 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
446 /* NB kiov, iov are READ-ONLY */
447 unsigned int this_nob;
453 LASSERT (!in_interrupt ());
456 while (kiovoffset >= kiov->kiov_len) {
457 kiovoffset -= kiov->kiov_len;
464 while (iovoffset >= iov->iov_len) {
465 iovoffset -= iov->iov_len;
474 this_nob = MIN(kiov->kiov_len - kiovoffset,
475 iov->iov_len - iovoffset);
476 this_nob = MIN(this_nob, nob);
479 addr = ((char *)kmap(kiov->kiov_page)) +
480 kiov->kiov_offset + kiovoffset;
482 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
485 if (kiov->kiov_len > kiovoffset + this_nob) {
487 kiovoffset += this_nob;
489 kunmap(kiov->kiov_page);
496 if (iov->iov_len > iovoffset + this_nob) {
497 iovoffset += this_nob;
506 kunmap(kiov->kiov_page);
508 EXPORT_SYMBOL(lnet_copy_iov2kiov);
511 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
512 int src_niov, lnet_kiov_t *src,
513 unsigned int offset, unsigned int len)
515 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
516 * for exactly 'len' bytes, and return the number of entries.
517 * NB not destructive to 'src' */
518 unsigned int frag_len;
521 if (len == 0) /* no data => */
522 return (0); /* no frags */
524 LASSERT(src_niov > 0);
525 while (offset >= src->kiov_len) { /* skip initial frags */
526 offset -= src->kiov_len;
529 LASSERT(src_niov > 0);
534 LASSERT(src_niov > 0);
535 LASSERT((int)niov <= dst_niov);
537 frag_len = src->kiov_len - offset;
538 dst->kiov_page = src->kiov_page;
539 dst->kiov_offset = src->kiov_offset + offset;
541 if (len <= frag_len) {
543 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
547 dst->kiov_len = frag_len;
548 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
558 EXPORT_SYMBOL(lnet_extract_kiov);
561 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
562 int delayed, unsigned int offset, unsigned int mlen,
565 unsigned int niov = 0;
566 struct kvec *iov = NULL;
567 lnet_kiov_t *kiov = NULL;
570 LASSERT (!in_interrupt ());
571 LASSERT (mlen == 0 || msg != NULL);
574 LASSERT(msg->msg_receiving);
575 LASSERT(!msg->msg_sending);
576 LASSERT(rlen == msg->msg_len);
577 LASSERT(mlen <= msg->msg_len);
578 LASSERT(msg->msg_offset == offset);
579 LASSERT(msg->msg_wanted == mlen);
581 msg->msg_receiving = 0;
584 niov = msg->msg_niov;
586 kiov = msg->msg_kiov;
589 LASSERT ((iov == NULL) != (kiov == NULL));
593 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
594 niov, iov, kiov, offset, mlen,
597 lnet_finalize(msg, rc);
601 lnet_setpayloadbuffer(struct lnet_msg *msg)
603 struct lnet_libmd *md = msg->msg_md;
605 LASSERT(msg->msg_len > 0);
606 LASSERT(!msg->msg_routing);
608 LASSERT(msg->msg_niov == 0);
609 LASSERT(msg->msg_iov == NULL);
610 LASSERT(msg->msg_kiov == NULL);
612 msg->msg_niov = md->md_niov;
613 if ((md->md_options & LNET_MD_KIOV) != 0)
614 msg->msg_kiov = md->md_iov.kiov;
616 msg->msg_iov = md->md_iov.iov;
620 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
621 unsigned int offset, unsigned int len)
623 msg->msg_type = type;
624 msg->msg_target = target;
626 msg->msg_offset = offset;
629 lnet_setpayloadbuffer(msg);
631 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
632 msg->msg_hdr.type = cpu_to_le32(type);
633 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
634 /* src_nid will be set later */
635 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
636 msg->msg_hdr.payload_length = cpu_to_le32(len);
640 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
642 void *priv = msg->msg_private;
645 LASSERT (!in_interrupt ());
646 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
647 (msg->msg_txcredit && msg->msg_peertxcredit));
649 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
651 lnet_finalize(msg, rc);
655 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
659 LASSERT(!msg->msg_sending);
660 LASSERT(msg->msg_receiving);
661 LASSERT(!msg->msg_rx_ready_delay);
662 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
664 msg->msg_rx_ready_delay = 1;
665 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
668 CERROR("recv from %s / send to %s aborted: "
669 "eager_recv failed %d\n",
670 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
671 libcfs_id2str(msg->msg_target), rc);
672 LASSERT(rc < 0); /* required by my callers */
679 * This function can be called from two paths:
680 * 1. when sending a message
681 * 2. when decommiting a message (lnet_msg_decommit_tx())
682 * In both these cases the peer_ni should have it's reference count
683 * acquired by the caller and therefore it is safe to drop the spin
684 * lock before calling lnd_query()
687 lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
689 cfs_time_t last_alive = 0;
690 int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
692 LASSERT(lnet_peer_aliveness_enabled(lp));
693 LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
695 lnet_net_unlock(cpt);
696 (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
699 lp->lpni_last_query = cfs_time_current();
701 if (last_alive != 0) /* NI has updated timestamp */
702 lp->lpni_last_alive = last_alive;
705 /* NB: always called with lnet_net_lock held */
707 lnet_peer_is_alive (struct lnet_peer_ni *lp, cfs_time_t now)
712 LASSERT (lnet_peer_aliveness_enabled(lp));
715 * Trust lnet_notify() if it has more recent aliveness news, but
716 * ignore the initial assumed death (see lnet_peers_start_down()).
718 spin_lock(&lp->lpni_lock);
719 if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
720 cfs_time_aftereq(lp->lpni_timestamp, lp->lpni_last_alive)) {
721 spin_unlock(&lp->lpni_lock);
726 cfs_time_add(lp->lpni_last_alive,
727 cfs_time_seconds(lp->lpni_net->net_tunables.
729 alive = cfs_time_after(deadline, now);
732 * Update obsolete lp_alive except for routers assumed to be dead
733 * initially, because router checker would update aliveness in this
734 * case, and moreover lpni_last_alive at peer creation is assumed.
736 if (alive && !lp->lpni_alive &&
737 !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
738 spin_unlock(&lp->lpni_lock);
739 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
741 spin_unlock(&lp->lpni_lock);
748 /* NB: returns 1 when alive, 0 when dead, negative when error;
749 * may drop the lnet_net_lock */
751 lnet_peer_alive_locked (struct lnet_ni *ni, struct lnet_peer_ni *lp)
753 cfs_time_t now = cfs_time_current();
755 if (!lnet_peer_aliveness_enabled(lp))
758 if (lnet_peer_is_alive(lp, now))
762 * Peer appears dead, but we should avoid frequent NI queries (at
763 * most once per lnet_queryinterval seconds).
765 if (lp->lpni_last_query != 0) {
766 static const int lnet_queryinterval = 1;
768 cfs_time_t next_query =
769 cfs_time_add(lp->lpni_last_query,
770 cfs_time_seconds(lnet_queryinterval));
772 if (cfs_time_before(now, next_query)) {
774 CWARN("Unexpected aliveness of peer %s: "
776 libcfs_nid2str(lp->lpni_nid),
777 (int)now, (int)next_query,
779 lp->lpni_net->net_tunables.lct_peer_timeout);
784 /* query NI for latest aliveness news */
785 lnet_ni_query_locked(ni, lp);
787 if (lnet_peer_is_alive(lp, now))
790 lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
795 * \param msg The message to be sent.
796 * \param do_send True if lnet_ni_send() should be called in this function.
797 * lnet_send() is going to lnet_net_unlock immediately after this, so
798 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
800 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
801 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
802 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
803 * \retval -ECANCELED If the MD of the message has been unlinked.
806 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
808 struct lnet_peer_ni *lp = msg->msg_txpeer;
809 struct lnet_ni *ni = msg->msg_txni;
810 int cpt = msg->msg_tx_cpt;
811 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
813 /* non-lnet_send() callers have checked before */
814 LASSERT(!do_send || msg->msg_tx_delayed);
815 LASSERT(!msg->msg_receiving);
816 LASSERT(msg->msg_tx_committed);
818 /* NB 'lp' is always the next hop */
819 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
820 lnet_peer_alive_locked(ni, lp) == 0) {
821 the_lnet.ln_counters[cpt]->drop_count++;
822 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
823 lnet_net_unlock(cpt);
825 atomic_inc(&msg->msg_txpeer->lpni_stats.drop_count);
827 atomic_inc(&msg->msg_txni->ni_stats.drop_count);
829 CNETERR("Dropping message for %s: peer not alive\n",
830 libcfs_id2str(msg->msg_target));
832 lnet_finalize(msg, -EHOSTUNREACH);
835 return -EHOSTUNREACH;
838 if (msg->msg_md != NULL &&
839 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
840 lnet_net_unlock(cpt);
842 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
843 "called on the MD/ME.\n",
844 libcfs_id2str(msg->msg_target));
846 lnet_finalize(msg, -ECANCELED);
852 if (!msg->msg_peertxcredit) {
853 spin_lock(&lp->lpni_lock);
854 LASSERT((lp->lpni_txcredits < 0) ==
855 !list_empty(&lp->lpni_txq));
857 msg->msg_peertxcredit = 1;
858 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
859 lp->lpni_txcredits--;
861 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
862 lp->lpni_mintxcredits = lp->lpni_txcredits;
864 if (lp->lpni_txcredits < 0) {
865 msg->msg_tx_delayed = 1;
866 list_add_tail(&msg->msg_list, &lp->lpni_txq);
867 spin_unlock(&lp->lpni_lock);
868 return LNET_CREDIT_WAIT;
870 spin_unlock(&lp->lpni_lock);
873 if (!msg->msg_txcredit) {
874 LASSERT((tq->tq_credits < 0) ==
875 !list_empty(&tq->tq_delayed));
877 msg->msg_txcredit = 1;
879 atomic_dec(&ni->ni_tx_credits);
881 if (tq->tq_credits < tq->tq_credits_min)
882 tq->tq_credits_min = tq->tq_credits;
884 if (tq->tq_credits < 0) {
885 msg->msg_tx_delayed = 1;
886 list_add_tail(&msg->msg_list, &tq->tq_delayed);
887 return LNET_CREDIT_WAIT;
892 lnet_net_unlock(cpt);
893 lnet_ni_send(ni, msg);
896 return LNET_CREDIT_OK;
900 static struct lnet_rtrbufpool *
901 lnet_msg2bufpool(struct lnet_msg *msg)
903 struct lnet_rtrbufpool *rbp;
906 LASSERT(msg->msg_rx_committed);
908 cpt = msg->msg_rx_cpt;
909 rbp = &the_lnet.ln_rtrpools[cpt][0];
911 LASSERT(msg->msg_len <= LNET_MTU);
912 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
914 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
921 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
923 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
924 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
925 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
926 * received or OK to receive */
927 struct lnet_peer_ni *lp = msg->msg_rxpeer;
928 struct lnet_rtrbufpool *rbp;
929 struct lnet_rtrbuf *rb;
931 LASSERT (msg->msg_iov == NULL);
932 LASSERT (msg->msg_kiov == NULL);
933 LASSERT (msg->msg_niov == 0);
934 LASSERT (msg->msg_routing);
935 LASSERT (msg->msg_receiving);
936 LASSERT (!msg->msg_sending);
938 /* non-lnet_parse callers only receive delayed messages */
939 LASSERT(!do_recv || msg->msg_rx_delayed);
941 if (!msg->msg_peerrtrcredit) {
942 spin_lock(&lp->lpni_lock);
943 LASSERT((lp->lpni_rtrcredits < 0) ==
944 !list_empty(&lp->lpni_rtrq));
946 msg->msg_peerrtrcredit = 1;
947 lp->lpni_rtrcredits--;
948 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
949 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
951 if (lp->lpni_rtrcredits < 0) {
952 /* must have checked eager_recv before here */
953 LASSERT(msg->msg_rx_ready_delay);
954 msg->msg_rx_delayed = 1;
955 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
956 spin_unlock(&lp->lpni_lock);
957 return LNET_CREDIT_WAIT;
959 spin_unlock(&lp->lpni_lock);
962 rbp = lnet_msg2bufpool(msg);
964 if (!msg->msg_rtrcredit) {
965 msg->msg_rtrcredit = 1;
967 if (rbp->rbp_credits < rbp->rbp_mincredits)
968 rbp->rbp_mincredits = rbp->rbp_credits;
970 if (rbp->rbp_credits < 0) {
971 /* must have checked eager_recv before here */
972 LASSERT(msg->msg_rx_ready_delay);
973 msg->msg_rx_delayed = 1;
974 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
975 return LNET_CREDIT_WAIT;
979 LASSERT(!list_empty(&rbp->rbp_bufs));
980 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
981 list_del(&rb->rb_list);
983 msg->msg_niov = rbp->rbp_npages;
984 msg->msg_kiov = &rb->rb_kiov[0];
987 int cpt = msg->msg_rx_cpt;
989 lnet_net_unlock(cpt);
990 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
991 0, msg->msg_len, msg->msg_len);
994 return LNET_CREDIT_OK;
998 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1000 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1001 struct lnet_ni *txni = msg->msg_txni;
1002 struct lnet_msg *msg2;
1004 if (msg->msg_txcredit) {
1005 struct lnet_ni *ni = msg->msg_txni;
1006 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1008 /* give back NI txcredits */
1009 msg->msg_txcredit = 0;
1011 LASSERT((tq->tq_credits < 0) ==
1012 !list_empty(&tq->tq_delayed));
1015 atomic_inc(&ni->ni_tx_credits);
1016 if (tq->tq_credits <= 0) {
1017 msg2 = list_entry(tq->tq_delayed.next,
1018 struct lnet_msg, msg_list);
1019 list_del(&msg2->msg_list);
1021 LASSERT(msg2->msg_txni == ni);
1022 LASSERT(msg2->msg_tx_delayed);
1023 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1025 (void) lnet_post_send_locked(msg2, 1);
1029 if (msg->msg_peertxcredit) {
1030 /* give back peer txcredits */
1031 msg->msg_peertxcredit = 0;
1033 spin_lock(&txpeer->lpni_lock);
1034 LASSERT((txpeer->lpni_txcredits < 0) ==
1035 !list_empty(&txpeer->lpni_txq));
1037 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1038 LASSERT(txpeer->lpni_txqnob >= 0);
1040 txpeer->lpni_txcredits++;
1041 if (txpeer->lpni_txcredits <= 0) {
1044 msg2 = list_entry(txpeer->lpni_txq.next,
1045 struct lnet_msg, msg_list);
1046 list_del(&msg2->msg_list);
1047 spin_unlock(&txpeer->lpni_lock);
1049 LASSERT(msg2->msg_txpeer == txpeer);
1050 LASSERT(msg2->msg_tx_delayed);
1052 msg2_cpt = msg2->msg_tx_cpt;
1055 * The msg_cpt can be different from the msg2_cpt
1056 * so we need to make sure we lock the correct cpt
1058 * Once we call lnet_post_send_locked() it is no
1059 * longer safe to access msg2, since it could've
1060 * been freed by lnet_finalize(), but we still
1061 * need to relock the correct cpt, so we cache the
1062 * msg2_cpt for the purpose of the check that
1063 * follows the call to lnet_pose_send_locked().
1065 if (msg2_cpt != msg->msg_tx_cpt) {
1066 lnet_net_unlock(msg->msg_tx_cpt);
1067 lnet_net_lock(msg2_cpt);
1069 (void) lnet_post_send_locked(msg2, 1);
1070 if (msg2_cpt != msg->msg_tx_cpt) {
1071 lnet_net_unlock(msg2_cpt);
1072 lnet_net_lock(msg->msg_tx_cpt);
1075 spin_unlock(&txpeer->lpni_lock);
1080 msg->msg_txni = NULL;
1081 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1084 if (txpeer != NULL) {
1087 * Once the patch for the health comes in we need to set
1088 * the health of the peer ni to bad when we fail to send
1090 * int status = msg->msg_ev.status;
1092 * lnet_set_peer_ni_health_locked(txpeer, false)
1094 msg->msg_txpeer = NULL;
1095 lnet_peer_ni_decref_locked(txpeer);
1100 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1102 struct lnet_msg *msg;
1104 if (list_empty(&rbp->rbp_msgs))
1106 msg = list_entry(rbp->rbp_msgs.next,
1107 struct lnet_msg, msg_list);
1108 list_del(&msg->msg_list);
1110 (void)lnet_post_routed_recv_locked(msg, 1);
1114 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1116 struct lnet_msg *msg;
1117 struct lnet_msg *tmp;
1119 lnet_net_unlock(cpt);
1121 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1122 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1123 0, 0, 0, msg->msg_hdr.payload_length);
1124 list_del_init(&msg->msg_list);
1125 lnet_finalize(msg, -ECANCELED);
1132 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1134 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1135 struct lnet_ni *rxni = msg->msg_rxni;
1136 struct lnet_msg *msg2;
1138 if (msg->msg_rtrcredit) {
1139 /* give back global router credits */
1140 struct lnet_rtrbuf *rb;
1141 struct lnet_rtrbufpool *rbp;
1143 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1144 * there until it gets one allocated, or aborts the wait
1146 LASSERT(msg->msg_kiov != NULL);
1148 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1151 msg->msg_kiov = NULL;
1152 msg->msg_rtrcredit = 0;
1154 LASSERT(rbp == lnet_msg2bufpool(msg));
1156 LASSERT((rbp->rbp_credits > 0) ==
1157 !list_empty(&rbp->rbp_bufs));
1159 /* If routing is now turned off, we just drop this buffer and
1160 * don't bother trying to return credits. */
1161 if (!the_lnet.ln_routing) {
1162 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1166 /* It is possible that a user has lowered the desired number of
1167 * buffers in this pool. Make sure we never put back
1168 * more buffers than the stated number. */
1169 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1170 /* Discard this buffer so we don't have too
1172 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1173 rbp->rbp_nbuffers--;
1175 list_add(&rb->rb_list, &rbp->rbp_bufs);
1177 if (rbp->rbp_credits <= 0)
1178 lnet_schedule_blocked_locked(rbp);
1183 if (msg->msg_peerrtrcredit) {
1184 /* give back peer router credits */
1185 msg->msg_peerrtrcredit = 0;
1187 spin_lock(&rxpeer->lpni_lock);
1188 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1189 !list_empty(&rxpeer->lpni_rtrq));
1191 rxpeer->lpni_rtrcredits++;
1193 /* drop all messages which are queued to be routed on that
1195 if (!the_lnet.ln_routing) {
1196 struct list_head drop;
1197 INIT_LIST_HEAD(&drop);
1198 list_splice_init(&rxpeer->lpni_rtrq, &drop);
1199 spin_unlock(&rxpeer->lpni_lock);
1200 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1201 } else if (rxpeer->lpni_rtrcredits <= 0) {
1202 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1203 struct lnet_msg, msg_list);
1204 list_del(&msg2->msg_list);
1205 spin_unlock(&rxpeer->lpni_lock);
1206 (void) lnet_post_routed_recv_locked(msg2, 1);
1208 spin_unlock(&rxpeer->lpni_lock);
1212 msg->msg_rxni = NULL;
1213 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1215 if (rxpeer != NULL) {
1216 msg->msg_rxpeer = NULL;
1217 lnet_peer_ni_decref_locked(rxpeer);
1222 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1224 if (p1->lpni_txqnob < p2->lpni_txqnob)
1227 if (p1->lpni_txqnob > p2->lpni_txqnob)
1230 if (p1->lpni_txcredits > p2->lpni_txcredits)
1233 if (p1->lpni_txcredits < p2->lpni_txcredits)
1240 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1242 struct lnet_peer_ni *p1 = r1->lr_gateway;
1243 struct lnet_peer_ni *p2 = r2->lr_gateway;
1244 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1245 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1248 if (r1->lr_priority < r2->lr_priority)
1251 if (r1->lr_priority > r2->lr_priority)
1254 if (r1_hops < r2_hops)
1257 if (r1_hops > r2_hops)
1260 rc = lnet_compare_peers(p1, p2);
1264 if (r1->lr_seq - r2->lr_seq <= 0)
1270 static struct lnet_peer_ni *
1271 lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
1274 struct lnet_remotenet *rnet;
1275 struct lnet_route *route;
1276 struct lnet_route *best_route;
1277 struct lnet_route *last_route;
1278 struct lnet_peer_ni *lpni_best;
1279 struct lnet_peer_ni *lp;
1282 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1283 * rtr_nid nid, otherwise find the best gateway I can use */
1285 rnet = lnet_find_rnet_locked(LNET_NIDNET(target));
1290 best_route = last_route = NULL;
1291 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1292 lp = route->lr_gateway;
1294 if (!lnet_is_route_alive(route))
1297 if (net != NULL && lp->lpni_net != net)
1300 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1303 if (lpni_best == NULL) {
1304 best_route = last_route = route;
1309 /* no protection on below fields, but it's harmless */
1310 if (last_route->lr_seq - route->lr_seq < 0)
1313 rc = lnet_compare_routes(route, best_route);
1321 /* set sequence number on the best router to the latest sequence + 1
1322 * so we can round-robin all routers, it's race and inaccurate but
1323 * harmless and functional */
1324 if (best_route != NULL)
1325 best_route->lr_seq = last_route->lr_seq + 1;
1329 static struct lnet_ni *
1330 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *cur_ni,
1333 struct lnet_ni *ni = NULL, *best_ni = cur_ni;
1334 unsigned int shortest_distance;
1337 if (best_ni == NULL) {
1338 shortest_distance = UINT_MAX;
1339 best_credits = INT_MIN;
1341 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1342 best_ni->ni_dev_cpt);
1343 best_credits = atomic_read(&best_ni->ni_tx_credits);
1346 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1347 unsigned int distance;
1350 if (!lnet_is_ni_healthy_locked(ni))
1353 ni_credits = atomic_read(&ni->ni_tx_credits);
1356 * calculate the distance from the CPT on which
1357 * the message memory is allocated to the CPT of
1358 * the NI's physical device
1360 distance = cfs_cpt_distance(lnet_cpt_table(),
1365 * All distances smaller than the NUMA range
1366 * are treated equally.
1368 if (distance < lnet_numa_range)
1369 distance = lnet_numa_range;
1372 * Select on shorter distance, then available
1373 * credits, then round-robin.
1375 if (distance > shortest_distance) {
1377 } else if (distance < shortest_distance) {
1378 shortest_distance = distance;
1379 } else if (ni_credits < best_credits) {
1381 } else if (ni_credits == best_credits) {
1382 if (best_ni && (best_ni)->ni_seq <= ni->ni_seq)
1386 best_credits = ni_credits;
1393 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
1394 struct lnet_msg *msg, lnet_nid_t rtr_nid)
1396 struct lnet_ni *best_ni;
1397 struct lnet_peer_ni *best_lpni;
1398 struct lnet_peer_ni *best_gw;
1399 struct lnet_peer_ni *lpni;
1400 struct lnet_peer_ni *final_dst;
1401 struct lnet_peer *peer;
1402 struct lnet_peer_net *peer_net;
1403 struct lnet_net *local_net;
1411 int best_lpni_credits;
1415 * get an initial CPT to use for locking. The idea here is not to
1416 * serialize the calls to select_pathway, so that as many
1417 * operations can run concurrently as possible. To do that we use
1418 * the CPT where this call is being executed. Later on when we
1419 * determine the CPT to use in lnet_message_commit, we switch the
1420 * lock and check if there was any configuration change. If none,
1421 * then we proceed, if there is, then we restart the operation.
1423 cpt = lnet_net_lock_current();
1425 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
1426 if (md_cpt == CFS_CPT_ANY)
1437 local_found = false;
1439 seq = lnet_get_dlc_seq_locked();
1441 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1442 lnet_net_unlock(cpt);
1447 * lnet_nid2peerni_locked() is the path that will find an
1448 * existing peer_ni, or create one and mark it as having been
1449 * created due to network traffic.
1451 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
1453 lnet_net_unlock(cpt);
1454 return PTR_ERR(lpni);
1456 peer = lpni->lpni_peer_net->lpn_peer;
1457 lnet_peer_ni_decref_locked(lpni);
1459 /* If peer is not healthy then can not send anything to it */
1460 if (!lnet_is_peer_healthy_locked(peer)) {
1461 lnet_net_unlock(cpt);
1462 return -EHOSTUNREACH;
1466 * STEP 1: first jab at determining best_ni
1467 * if src_nid is explicitly specified, then best_ni is already
1468 * pre-determiend for us. Otherwise we need to select the best
1469 * one to use later on
1471 if (src_nid != LNET_NID_ANY) {
1472 best_ni = lnet_nid2ni_locked(src_nid, cpt);
1474 lnet_net_unlock(cpt);
1475 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1476 "local nid\n", libcfs_nid2str(dst_nid),
1477 libcfs_nid2str(src_nid));
1482 if (msg->msg_type == LNET_MSG_REPLY ||
1483 msg->msg_type == LNET_MSG_ACK ||
1484 !lnet_peer_is_multi_rail(peer) ||
1487 * for replies we want to respond on the same peer_ni we
1488 * received the message on if possible. If not, then pick
1489 * a peer_ni to send to
1491 * if the peer is non-multi-rail then you want to send to
1492 * the dst_nid provided as well.
1494 * If the best_ni has already been determined, IE the
1495 * src_nid has been specified, then use the
1496 * destination_nid provided as well, since we're
1497 * continuing a series of related messages for the same
1500 * It is expected to find the lpni using dst_nid, since we
1501 * created it earlier.
1503 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1505 lnet_peer_ni_decref_locked(best_lpni);
1507 if (best_lpni && !lnet_get_net_locked(LNET_NIDNET(dst_nid))) {
1509 * this lpni is not on a local network so we need
1510 * to route this reply.
1512 best_gw = lnet_find_route_locked(NULL,
1513 best_lpni->lpni_nid,
1517 * RULE: Each node considers only the next-hop
1519 * We're going to route the message, so change the peer to
1522 LASSERT(best_gw->lpni_peer_net);
1523 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1524 peer = best_gw->lpni_peer_net->lpn_peer;
1527 * if the router is not multi-rail then use the best_gw
1528 * found to send the message to
1530 if (!lnet_peer_is_multi_rail(peer))
1531 best_lpni = best_gw;
1539 } else if (!best_lpni) {
1540 lnet_net_unlock(cpt);
1541 CERROR("unable to send msg_type %d to "
1542 "originating %s. Destination NID not in DB\n",
1543 msg->msg_type, libcfs_nid2str(dst_nid));
1549 * We must use a consistent source address when sending to a
1550 * non-MR peer. However, a non-MR peer can have multiple NIDs
1551 * on multiple networks, and we may even need to talk to this
1552 * peer on multiple networks -- certain types of
1553 * load-balancing configuration do this.
1555 * So we need to pick the NI the peer prefers for this
1556 * particular network.
1558 if (!lnet_peer_is_multi_rail(peer)) {
1560 lnet_net_unlock(cpt);
1561 CERROR("no route to %s\n",
1562 libcfs_nid2str(dst_nid));
1563 return -EHOSTUNREACH;
1566 /* best ni is already set if src_nid was provided */
1568 /* Get the target peer_ni */
1569 peer_net = lnet_peer_get_net_locked(peer,
1570 LNET_NIDNET(best_lpni->lpni_nid));
1571 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
1572 lpni_on_peer_net_list) {
1573 if (lpni->lpni_pref_nnids == 0)
1575 LASSERT(lpni->lpni_pref_nnids == 1);
1576 best_ni = lnet_nid2ni_locked(
1577 lpni->lpni_pref.nid, cpt);
1581 /* if best_ni is still not set just pick one */
1583 best_ni = lnet_net2ni_locked(
1584 best_lpni->lpni_net->net_id, cpt);
1585 /* If there is no best_ni we don't have a route */
1587 lnet_net_unlock(cpt);
1588 CERROR("no path to %s from net %s\n",
1589 libcfs_nid2str(best_lpni->lpni_nid),
1590 libcfs_net2str(best_lpni->lpni_net->net_id));
1591 return -EHOSTUNREACH;
1593 lpni = list_entry(peer_net->lpn_peer_nis.next,
1594 struct lnet_peer_ni,
1595 lpni_on_peer_net_list);
1597 /* Set preferred NI if necessary. */
1598 if (lpni->lpni_pref_nnids == 0)
1599 lnet_peer_ni_set_non_mr_pref_nid(lpni, best_ni->ni_nid);
1603 * if we already found a best_ni because src_nid is specified and
1604 * best_lpni because we are replying to a message then just send
1607 if (best_ni && best_lpni)
1611 * If we already found a best_ni because src_nid is specified then
1612 * pick the peer then send the message
1618 * pick the best_ni by going through all the possible networks of
1619 * that peer and see which local NI is best suited to talk to that
1622 * Locally connected networks will always be preferred over
1623 * a routed network. If there are only routed paths to the peer,
1624 * then the best route is chosen. If all routes are equal then
1625 * they are used in round robin.
1627 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_on_peer_list) {
1628 if (!lnet_is_peer_net_healthy_locked(peer_net))
1631 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1632 if (!local_net && !routing && !local_found) {
1633 struct lnet_peer_ni *net_gw;
1635 lpni = list_entry(peer_net->lpn_peer_nis.next,
1636 struct lnet_peer_ni,
1637 lpni_on_peer_net_list);
1639 net_gw = lnet_find_route_locked(NULL,
1647 * lnet_find_route_locked() call
1648 * will return the best_Gw on the
1649 * lpni->lpni_nid network.
1650 * However, best_gw and net_gw can
1651 * be on different networks.
1652 * Therefore need to compare them
1653 * to pick the better of either.
1655 if (lnet_compare_peers(best_gw, net_gw) > 0)
1657 if (best_gw->lpni_gw_seq <= net_gw->lpni_gw_seq)
1672 * a gw on this network is found, but there could be
1673 * other better gateways on other networks. So don't pick
1674 * the best_ni until we determine the best_gw.
1679 /* if no local_net found continue */
1684 * Iterate through the NIs in this local Net and select
1685 * the NI to send from. The selection is determined by
1686 * these 3 criterion in the following priority:
1688 * 2. NI available credits
1691 best_ni = lnet_get_best_ni(local_net, best_ni, md_cpt);
1694 if (!best_ni && !best_gw) {
1695 lnet_net_unlock(cpt);
1696 LCONSOLE_WARN("No local ni found to send from to %s\n",
1697 libcfs_nid2str(dst_nid));
1702 best_ni = lnet_get_best_ni(best_gw->lpni_net, best_ni, md_cpt);
1703 LASSERT(best_gw && best_ni);
1706 * We're going to route the message, so change the peer to
1709 LASSERT(best_gw->lpni_peer_net);
1710 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1711 best_gw->lpni_gw_seq++;
1712 peer = best_gw->lpni_peer_net->lpn_peer;
1716 * Now that we selected the NI to use increment its sequence
1717 * number so the Round Robin algorithm will detect that it has
1718 * been used and pick the next NI.
1724 * At this point the best_ni is on a local network on which
1725 * the peer has a peer_ni as well
1727 peer_net = lnet_peer_get_net_locked(peer,
1728 best_ni->ni_net->net_id);
1730 * peer_net is not available or the src_nid is explicitly defined
1731 * and the peer_net for that src_nid is unhealthy. find a route to
1732 * the destination nid.
1735 (src_nid != LNET_NID_ANY &&
1736 !lnet_is_peer_net_healthy_locked(peer_net))) {
1737 best_gw = lnet_find_route_locked(best_ni->ni_net,
1741 * if no route is found for that network then
1742 * move onto the next peer_ni in the peer
1745 lnet_net_unlock(cpt);
1746 LCONSOLE_WARN("No route to peer from %s\n",
1747 libcfs_nid2str(best_ni->ni_nid));
1748 return -EHOSTUNREACH;
1751 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1752 libcfs_nid2str(dst_nid),
1753 libcfs_nid2str(best_gw->lpni_nid),
1754 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1758 * RULE: Each node considers only the next-hop
1760 * We're going to route the message, so change the peer to
1763 LASSERT(best_gw->lpni_peer_net);
1764 LASSERT(best_gw->lpni_peer_net->lpn_peer);
1765 peer = best_gw->lpni_peer_net->lpn_peer;
1766 } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
1768 * this peer_net is unhealthy but we still have an opportunity
1769 * to find another peer_net that we can use
1771 __u32 net_id = peer_net->lpn_net_id;
1772 LCONSOLE_WARN("peer net %s unhealthy\n",
1773 libcfs_net2str(net_id));
1778 * Look at the peer NIs for the destination peer that connect
1779 * to the chosen net. If a peer_ni is preferred when using the
1780 * best_ni to communicate, we use that one. If there is no
1781 * preferred peer_ni, or there are multiple preferred peer_ni,
1782 * the available transmit credits are used. If the transmit
1783 * credits are equal, we round-robin over the peer_ni.
1786 best_lpni_credits = INT_MIN;
1789 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1791 * if this peer ni is not healthy just skip it, no point in
1792 * examining it further
1794 if (!lnet_is_peer_ni_healthy_locked(lpni))
1796 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1799 /* if this is a preferred peer use it */
1800 if (!preferred && ni_is_pref) {
1802 } else if (preferred && !ni_is_pref) {
1804 * this is not the preferred peer so let's ignore
1808 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1810 * We already have a peer that has more credits
1811 * available than this one. No need to consider
1812 * this peer further.
1815 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1817 * The best peer found so far and the current peer
1818 * have the same number of available credits let's
1819 * make sure to select between them using Round
1823 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1829 best_lpni_credits = lpni->lpni_txcredits;
1832 /* if we still can't find a peer ni then we can't reach it */
1834 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1835 LNET_NIDNET(dst_nid);
1836 lnet_net_unlock(cpt);
1837 LCONSOLE_WARN("no peer_ni found on peer net %s\n",
1838 libcfs_net2str(net_id));
1839 return -EHOSTUNREACH;
1844 /* Shortcut for loopback. */
1845 if (best_ni == the_lnet.ln_loni) {
1846 /* No send credit hassles with LOLND */
1847 lnet_ni_addref_locked(best_ni, cpt);
1848 msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
1849 if (!msg->msg_routing)
1850 msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
1851 msg->msg_target.nid = best_ni->ni_nid;
1852 lnet_msg_commit(msg, cpt);
1853 msg->msg_txni = best_ni;
1854 lnet_net_unlock(cpt);
1856 return LNET_CREDIT_OK;
1859 routing = routing || routing2;
1862 * Increment sequence number of the peer selected so that we
1863 * pick the next one in Round Robin.
1865 best_lpni->lpni_seq++;
1868 * grab a reference on the peer_ni so it sticks around even if
1869 * we need to drop and relock the lnet_net_lock below.
1871 lnet_peer_ni_addref_locked(best_lpni);
1874 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1875 * message. This ensures that we get a CPT that is correct for
1876 * the NI when the NI has been restricted to a subset of all CPTs.
1877 * If the selected CPT differs from the one currently locked, we
1878 * must unlock and relock the lnet_net_lock(), and then check whether
1879 * the configuration has changed. We don't have a hold on the best_ni
1880 * yet, and it may have vanished.
1882 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1884 lnet_net_unlock(cpt);
1887 if (seq != lnet_get_dlc_seq_locked()) {
1888 lnet_peer_ni_decref_locked(best_lpni);
1894 * store the best_lpni in the message right away to avoid having
1895 * to do the same operation under different conditions
1897 msg->msg_txpeer = best_lpni;
1898 msg->msg_txni = best_ni;
1901 * grab a reference for the best_ni since now it's in use in this
1902 * send. the reference will need to be dropped when the message is
1903 * finished in lnet_finalize()
1905 lnet_ni_addref_locked(msg->msg_txni, cpt);
1908 * Always set the target.nid to the best peer picked. Either the
1909 * nid will be one of the preconfigured NIDs, or the same NID as
1910 * what was originally set in the target or it will be the NID of
1911 * a router if this message should be routed
1913 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1916 * lnet_msg_commit assigns the correct cpt to the message, which
1917 * is used to decrement the correct refcount on the ni when it's
1918 * time to return the credits
1920 lnet_msg_commit(msg, cpt);
1923 * If we are routing the message then we don't need to overwrite
1924 * the src_nid since it would've been set at the origin. Otherwise
1925 * we are the originator so we need to set it.
1927 if (!msg->msg_routing)
1928 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1931 msg->msg_target_is_router = 1;
1932 msg->msg_target.pid = LNET_PID_LUSTRE;
1934 * since we're routing we want to ensure that the
1935 * msg_hdr.dest_nid is set to the final destination. When
1936 * the router receives this message it knows how to route
1939 msg->msg_hdr.dest_nid =
1940 cpu_to_le64(final_dst ? final_dst->lpni_nid : dst_nid);
1943 * if we're not routing set the dest_nid to the best peer
1944 * ni that we picked earlier in the algorithm.
1946 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1949 rc = lnet_post_send_locked(msg, 0);
1952 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s\n",
1953 libcfs_nid2str(msg->msg_hdr.src_nid),
1954 libcfs_nid2str(msg->msg_txni->ni_nid),
1955 libcfs_nid2str(src_nid),
1956 libcfs_nid2str(msg->msg_hdr.dest_nid),
1957 libcfs_nid2str(dst_nid),
1958 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1959 lnet_msgtyp2str(msg->msg_type));
1961 lnet_net_unlock(cpt);
1967 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
1969 lnet_nid_t dst_nid = msg->msg_target.nid;
1973 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1974 * but we might want to use pre-determined router for ACK/REPLY
1977 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1978 LASSERT (msg->msg_txpeer == NULL);
1979 LASSERT (!msg->msg_sending);
1980 LASSERT (!msg->msg_target_is_router);
1981 LASSERT (!msg->msg_receiving);
1983 msg->msg_sending = 1;
1985 LASSERT(!msg->msg_tx_committed);
1987 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
1991 if (rc == LNET_CREDIT_OK)
1992 lnet_ni_send(msg->msg_txni, msg);
1994 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
1999 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob)
2002 the_lnet.ln_counters[cpt]->drop_count++;
2003 the_lnet.ln_counters[cpt]->drop_length += nob;
2004 lnet_net_unlock(cpt);
2006 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
2010 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
2012 struct lnet_hdr *hdr = &msg->msg_hdr;
2014 if (msg->msg_wanted != 0)
2015 lnet_setpayloadbuffer(msg);
2017 lnet_build_msg_event(msg, LNET_EVENT_PUT);
2019 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
2020 * it back into the ACK during lnet_finalize() */
2021 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
2022 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
2024 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
2025 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
2029 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
2031 struct lnet_hdr *hdr = &msg->msg_hdr;
2032 struct lnet_match_info info;
2036 /* Convert put fields to host byte order */
2037 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
2038 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
2039 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
2041 /* Primary peer NID. */
2042 info.mi_id.nid = msg->msg_initiator;
2043 info.mi_id.pid = hdr->src_pid;
2044 info.mi_opc = LNET_MD_OP_PUT;
2045 info.mi_portal = hdr->msg.put.ptl_index;
2046 info.mi_rlength = hdr->payload_length;
2047 info.mi_roffset = hdr->msg.put.offset;
2048 info.mi_mbits = hdr->msg.put.match_bits;
2049 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
2051 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
2052 ready_delay = msg->msg_rx_ready_delay;
2055 rc = lnet_ptl_match_md(&info, msg);
2060 case LNET_MATCHMD_OK:
2061 lnet_recv_put(ni, msg);
2064 case LNET_MATCHMD_NONE:
2066 /* no eager_recv or has already called it, should
2067 * have been attached on delayed list */
2070 rc = lnet_ni_eager_recv(ni, msg);
2077 case LNET_MATCHMD_DROP:
2078 CNETERR("Dropping PUT from %s portal %d match %llu"
2079 " offset %d length %d: %d\n",
2080 libcfs_id2str(info.mi_id), info.mi_portal,
2081 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
2083 return -ENOENT; /* -ve: OK but no match */
2088 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
2090 struct lnet_match_info info;
2091 struct lnet_hdr *hdr = &msg->msg_hdr;
2092 struct lnet_process_id source_id;
2093 struct lnet_handle_wire reply_wmd;
2096 /* Convert get fields to host byte order */
2097 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
2098 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
2099 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
2100 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
2102 source_id.nid = hdr->src_nid;
2103 source_id.pid = hdr->src_pid;
2104 /* Primary peer NID */
2105 info.mi_id.nid = msg->msg_initiator;
2106 info.mi_id.pid = hdr->src_pid;
2107 info.mi_opc = LNET_MD_OP_GET;
2108 info.mi_portal = hdr->msg.get.ptl_index;
2109 info.mi_rlength = hdr->msg.get.sink_length;
2110 info.mi_roffset = hdr->msg.get.src_offset;
2111 info.mi_mbits = hdr->msg.get.match_bits;
2112 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
2114 rc = lnet_ptl_match_md(&info, msg);
2115 if (rc == LNET_MATCHMD_DROP) {
2116 CNETERR("Dropping GET from %s portal %d match %llu"
2117 " offset %d length %d\n",
2118 libcfs_id2str(info.mi_id), info.mi_portal,
2119 info.mi_mbits, info.mi_roffset, info.mi_rlength);
2120 return -ENOENT; /* -ve: OK but no match */
2123 LASSERT(rc == LNET_MATCHMD_OK);
2125 lnet_build_msg_event(msg, LNET_EVENT_GET);
2127 reply_wmd = hdr->msg.get.return_wmd;
2129 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
2130 msg->msg_offset, msg->msg_wanted);
2132 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
2135 /* The LND completes the REPLY from her recv procedure */
2136 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2137 msg->msg_offset, msg->msg_len, msg->msg_len);
2141 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
2142 msg->msg_receiving = 0;
2144 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
2146 /* didn't get as far as lnet_ni_send() */
2147 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
2148 libcfs_nid2str(ni->ni_nid),
2149 libcfs_id2str(info.mi_id), rc);
2151 lnet_finalize(msg, rc);
2158 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
2160 void *private = msg->msg_private;
2161 struct lnet_hdr *hdr = &msg->msg_hdr;
2162 struct lnet_process_id src = {0};
2163 struct lnet_libmd *md;
2168 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
2171 src.nid = hdr->src_nid;
2172 src.pid = hdr->src_pid;
2174 /* NB handles only looked up by creator (no flips) */
2175 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
2176 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2177 CNETERR("%s: Dropping REPLY from %s for %s "
2179 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2180 (md == NULL) ? "invalid" : "inactive",
2181 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2182 hdr->msg.reply.dst_wmd.wh_object_cookie);
2183 if (md != NULL && md->md_me != NULL)
2184 CERROR("REPLY MD also attached to portal %d\n",
2185 md->md_me->me_portal);
2187 lnet_res_unlock(cpt);
2188 return -ENOENT; /* -ve: OK but no match */
2191 LASSERT(md->md_offset == 0);
2193 rlength = hdr->payload_length;
2194 mlength = MIN(rlength, (int)md->md_length);
2196 if (mlength < rlength &&
2197 (md->md_options & LNET_MD_TRUNCATE) == 0) {
2198 CNETERR("%s: Dropping REPLY from %s length %d "
2199 "for MD %#llx would overflow (%d)\n",
2200 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2201 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2203 lnet_res_unlock(cpt);
2204 return -ENOENT; /* -ve: OK but no match */
2207 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
2208 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2209 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2211 lnet_msg_attach_md(msg, md, 0, mlength);
2214 lnet_setpayloadbuffer(msg);
2216 lnet_res_unlock(cpt);
2218 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2220 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2225 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
2227 struct lnet_hdr *hdr = &msg->msg_hdr;
2228 struct lnet_process_id src = {0};
2229 struct lnet_libmd *md;
2232 src.nid = hdr->src_nid;
2233 src.pid = hdr->src_pid;
2235 /* Convert ack fields to host byte order */
2236 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2237 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2239 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
2242 /* NB handles only looked up by creator (no flips) */
2243 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2244 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2245 /* Don't moan; this is expected */
2247 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
2248 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2249 (md == NULL) ? "invalid" : "inactive",
2250 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2251 hdr->msg.ack.dst_wmd.wh_object_cookie);
2252 if (md != NULL && md->md_me != NULL)
2253 CERROR("Source MD also attached to portal %d\n",
2254 md->md_me->me_portal);
2256 lnet_res_unlock(cpt);
2257 return -ENOENT; /* -ve! */
2260 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
2261 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2262 hdr->msg.ack.dst_wmd.wh_object_cookie);
2264 lnet_msg_attach_md(msg, md, 0, 0);
2266 lnet_res_unlock(cpt);
2268 lnet_build_msg_event(msg, LNET_EVENT_ACK);
2270 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2275 * \retval LNET_CREDIT_OK If \a msg is forwarded
2276 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
2277 * \retval -ve error code
2280 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
2284 if (!the_lnet.ln_routing)
2287 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
2288 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2289 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
2290 msg->msg_rx_ready_delay = 1;
2292 lnet_net_unlock(msg->msg_rx_cpt);
2293 rc = lnet_ni_eager_recv(ni, msg);
2294 lnet_net_lock(msg->msg_rx_cpt);
2299 rc = lnet_post_routed_recv_locked(msg, 0);
2304 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
2308 switch (msg->msg_type) {
2310 rc = lnet_parse_ack(ni, msg);
2313 rc = lnet_parse_put(ni, msg);
2316 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
2318 case LNET_MSG_REPLY:
2319 rc = lnet_parse_reply(ni, msg);
2321 default: /* prevent an unused label if !kernel */
2326 LASSERT(rc == 0 || rc == -ENOENT);
2331 lnet_msgtyp2str (int type)
2340 case LNET_MSG_REPLY:
2342 case LNET_MSG_HELLO:
2345 return ("<UNKNOWN>");
2350 lnet_print_hdr(struct lnet_hdr *hdr)
2352 struct lnet_process_id src = {
2353 .nid = hdr->src_nid,
2354 .pid = hdr->src_pid,
2356 struct lnet_process_id dst = {
2357 .nid = hdr->dest_nid,
2358 .pid = hdr->dest_pid,
2360 char *type_str = lnet_msgtyp2str(hdr->type);
2362 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2363 CWARN(" From %s\n", libcfs_id2str(src));
2364 CWARN(" To %s\n", libcfs_id2str(dst));
2366 switch (hdr->type) {
2371 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
2372 "match bits %llu\n",
2373 hdr->msg.put.ptl_index,
2374 hdr->msg.put.ack_wmd.wh_interface_cookie,
2375 hdr->msg.put.ack_wmd.wh_object_cookie,
2376 hdr->msg.put.match_bits);
2377 CWARN(" Length %d, offset %d, hdr data %#llx\n",
2378 hdr->payload_length, hdr->msg.put.offset,
2379 hdr->msg.put.hdr_data);
2383 CWARN(" Ptl index %d, return md %#llx.%#llx, "
2384 "match bits %llu\n", hdr->msg.get.ptl_index,
2385 hdr->msg.get.return_wmd.wh_interface_cookie,
2386 hdr->msg.get.return_wmd.wh_object_cookie,
2387 hdr->msg.get.match_bits);
2388 CWARN(" Length %d, src offset %d\n",
2389 hdr->msg.get.sink_length,
2390 hdr->msg.get.src_offset);
2394 CWARN(" dst md %#llx.%#llx, "
2395 "manipulated length %d\n",
2396 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2397 hdr->msg.ack.dst_wmd.wh_object_cookie,
2398 hdr->msg.ack.mlength);
2401 case LNET_MSG_REPLY:
2402 CWARN(" dst md %#llx.%#llx, "
2404 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2405 hdr->msg.reply.dst_wmd.wh_object_cookie,
2406 hdr->payload_length);
2412 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
2413 void *private, int rdma_req)
2418 struct lnet_msg *msg;
2419 lnet_pid_t dest_pid;
2420 lnet_nid_t dest_nid;
2422 struct lnet_peer_ni *lpni;
2423 __u32 payload_length;
2426 LASSERT (!in_interrupt ());
2428 type = le32_to_cpu(hdr->type);
2429 src_nid = le64_to_cpu(hdr->src_nid);
2430 dest_nid = le64_to_cpu(hdr->dest_nid);
2431 dest_pid = le32_to_cpu(hdr->dest_pid);
2432 payload_length = le32_to_cpu(hdr->payload_length);
2434 for_me = (ni->ni_nid == dest_nid);
2435 cpt = lnet_cpt_of_nid(from_nid, ni);
2437 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s\n",
2438 libcfs_nid2str(dest_nid),
2439 libcfs_nid2str(ni->ni_nid),
2440 libcfs_nid2str(src_nid),
2441 lnet_msgtyp2str(type));
2446 if (payload_length > 0) {
2447 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2448 libcfs_nid2str(from_nid),
2449 libcfs_nid2str(src_nid),
2450 lnet_msgtyp2str(type), payload_length);
2456 case LNET_MSG_REPLY:
2457 if (payload_length >
2458 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2459 CERROR("%s, src %s: bad %s payload %d "
2460 "(%d max expected)\n",
2461 libcfs_nid2str(from_nid),
2462 libcfs_nid2str(src_nid),
2463 lnet_msgtyp2str(type),
2465 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2471 CERROR("%s, src %s: Bad message type 0x%x\n",
2472 libcfs_nid2str(from_nid),
2473 libcfs_nid2str(src_nid), type);
2477 if (the_lnet.ln_routing &&
2478 ni->ni_last_alive != ktime_get_real_seconds()) {
2479 /* NB: so far here is the only place to set NI status to "up */
2481 ni->ni_last_alive = ktime_get_real_seconds();
2482 if (ni->ni_status != NULL &&
2483 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
2484 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
2488 /* Regard a bad destination NID as a protocol error. Senders should
2489 * know what they're doing; if they don't they're misconfigured, buggy
2490 * or malicious so we chop them off at the knees :) */
2493 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2494 /* should have gone direct */
2495 CERROR("%s, src %s: Bad dest nid %s "
2496 "(should have been sent direct)\n",
2497 libcfs_nid2str(from_nid),
2498 libcfs_nid2str(src_nid),
2499 libcfs_nid2str(dest_nid));
2503 if (lnet_islocalnid(dest_nid)) {
2504 /* dest is another local NI; sender should have used
2505 * this node's NID on its own network */
2506 CERROR("%s, src %s: Bad dest nid %s "
2507 "(it's my nid but on a different network)\n",
2508 libcfs_nid2str(from_nid),
2509 libcfs_nid2str(src_nid),
2510 libcfs_nid2str(dest_nid));
2514 if (rdma_req && type == LNET_MSG_GET) {
2515 CERROR("%s, src %s: Bad optimized GET for %s "
2516 "(final destination must be me)\n",
2517 libcfs_nid2str(from_nid),
2518 libcfs_nid2str(src_nid),
2519 libcfs_nid2str(dest_nid));
2523 if (!the_lnet.ln_routing) {
2524 CERROR("%s, src %s: Dropping message for %s "
2525 "(routing not enabled)\n",
2526 libcfs_nid2str(from_nid),
2527 libcfs_nid2str(src_nid),
2528 libcfs_nid2str(dest_nid));
2533 /* Message looks OK; we're not going to return an error, so we MUST
2534 * call back lnd_recv() come what may... */
2536 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2537 fail_peer(src_nid, 0)) { /* shall we now? */
2538 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2539 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2540 lnet_msgtyp2str(type));
2544 if (!list_empty(&the_lnet.ln_drop_rules) &&
2545 lnet_drop_rule_match(hdr)) {
2546 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
2547 "silent message loss\n",
2548 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2549 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
2554 msg = lnet_msg_alloc();
2556 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2557 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2558 lnet_msgtyp2str(type));
2562 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
2563 * pointers NULL etc */
2565 msg->msg_type = type;
2566 msg->msg_private = private;
2567 msg->msg_receiving = 1;
2568 msg->msg_rdma_get = rdma_req;
2569 msg->msg_len = msg->msg_wanted = payload_length;
2570 msg->msg_offset = 0;
2571 msg->msg_hdr = *hdr;
2572 /* for building message event */
2573 msg->msg_from = from_nid;
2575 msg->msg_target.pid = dest_pid;
2576 msg->msg_target.nid = dest_nid;
2577 msg->msg_routing = 1;
2580 /* convert common msg->hdr fields to host byteorder */
2581 msg->msg_hdr.type = type;
2582 msg->msg_hdr.src_nid = src_nid;
2583 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2584 msg->msg_hdr.dest_nid = dest_nid;
2585 msg->msg_hdr.dest_pid = dest_pid;
2586 msg->msg_hdr.payload_length = payload_length;
2590 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
2592 lnet_net_unlock(cpt);
2593 CERROR("%s, src %s: Dropping %s "
2594 "(error %ld looking up sender)\n",
2595 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2596 lnet_msgtyp2str(type), PTR_ERR(lpni));
2598 if (rc == -ESHUTDOWN)
2599 /* We are shutting down. Don't do anything more */
2603 msg->msg_rxpeer = lpni;
2605 lnet_ni_addref_locked(ni, cpt);
2606 /* Multi-Rail: Primary NID of source. */
2607 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
2609 if (lnet_isrouter(msg->msg_rxpeer)) {
2610 lnet_peer_set_alive(msg->msg_rxpeer);
2611 if (avoid_asym_router_failure &&
2612 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
2613 /* received a remote message from router, update
2614 * remote NI status on this router.
2615 * NB: multi-hop routed message will be ignored.
2617 lnet_router_ni_update_locked(msg->msg_rxpeer,
2618 LNET_NIDNET(src_nid));
2622 lnet_msg_commit(msg, cpt);
2624 /* message delay simulation */
2625 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
2626 lnet_delay_rule_match_locked(hdr, msg))) {
2627 lnet_net_unlock(cpt);
2632 rc = lnet_parse_forward_locked(ni, msg);
2633 lnet_net_unlock(cpt);
2638 if (rc == LNET_CREDIT_OK) {
2639 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2640 0, payload_length, payload_length);
2645 lnet_net_unlock(cpt);
2647 rc = lnet_parse_local(ni, msg);
2653 LASSERT(msg->msg_md == NULL);
2654 lnet_finalize(msg, rc);
2657 lnet_drop_message(ni, cpt, private, payload_length);
2660 EXPORT_SYMBOL(lnet_parse);
2663 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
2665 while (!list_empty(head)) {
2666 struct lnet_process_id id = {0};
2667 struct lnet_msg *msg;
2669 msg = list_entry(head->next, struct lnet_msg, msg_list);
2670 list_del(&msg->msg_list);
2672 id.nid = msg->msg_hdr.src_nid;
2673 id.pid = msg->msg_hdr.src_pid;
2675 LASSERT(msg->msg_md == NULL);
2676 LASSERT(msg->msg_rx_delayed);
2677 LASSERT(msg->msg_rxpeer != NULL);
2678 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2680 CWARN("Dropping delayed PUT from %s portal %d match %llu"
2681 " offset %d length %d: %s\n",
2683 msg->msg_hdr.msg.put.ptl_index,
2684 msg->msg_hdr.msg.put.match_bits,
2685 msg->msg_hdr.msg.put.offset,
2686 msg->msg_hdr.payload_length, reason);
2688 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2689 * called lnet_drop_message(), so I just hang onto msg as well
2690 * until that's done */
2692 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
2693 msg->msg_private, msg->msg_len);
2695 * NB: message will not generate event because w/o attached MD,
2696 * but we still should give error code so lnet_msg_decommit()
2697 * can skip counters operations and other checks.
2699 lnet_finalize(msg, -ENOENT);
2704 lnet_recv_delayed_msg_list(struct list_head *head)
2706 while (!list_empty(head)) {
2707 struct lnet_msg *msg;
2708 struct lnet_process_id id;
2710 msg = list_entry(head->next, struct lnet_msg, msg_list);
2711 list_del(&msg->msg_list);
2713 /* md won't disappear under me, since each msg
2714 * holds a ref on it */
2716 id.nid = msg->msg_hdr.src_nid;
2717 id.pid = msg->msg_hdr.src_pid;
2719 LASSERT(msg->msg_rx_delayed);
2720 LASSERT(msg->msg_md != NULL);
2721 LASSERT(msg->msg_rxpeer != NULL);
2722 LASSERT(msg->msg_rxni != NULL);
2723 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2725 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2726 "match %llu offset %d length %d.\n",
2727 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2728 msg->msg_hdr.msg.put.match_bits,
2729 msg->msg_hdr.msg.put.offset,
2730 msg->msg_hdr.payload_length);
2732 lnet_recv_put(msg->msg_rxni, msg);
2737 * Initiate an asynchronous PUT operation.
2739 * There are several events associated with a PUT: completion of the send on
2740 * the initiator node (LNET_EVENT_SEND), and when the send completes
2741 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2742 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2743 * used at the target node to indicate the completion of incoming data
2746 * The local events will be logged in the EQ associated with the MD pointed to
2747 * by \a mdh handle. Using a MD without an associated EQ results in these
2748 * events being discarded. In this case, the caller must have another
2749 * mechanism (e.g., a higher level protocol) for determining when it is safe
2750 * to modify the memory region associated with the MD.
2752 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2753 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2755 * \param self Indicates the NID of a local interface through which to send
2756 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2757 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2758 * must be "free floating" (See LNetMDBind()).
2759 * \param ack Controls whether an acknowledgment is requested.
2760 * Acknowledgments are only sent when they are requested by the initiating
2761 * process and the target MD enables them.
2762 * \param target A process identifier for the target process.
2763 * \param portal The index in the \a target's portal table.
2764 * \param match_bits The match bits to use for MD selection at the target
2766 * \param offset The offset into the target MD (only used when the target
2767 * MD has the LNET_MD_MANAGE_REMOTE option set).
2768 * \param hdr_data 64 bits of user data that can be included in the message
2769 * header. This data is written to an event queue entry at the target if an
2770 * EQ is present on the matching MD.
2772 * \retval 0 Success, and only in this case events will be generated
2773 * and logged to EQ (if it exists).
2774 * \retval -EIO Simulated failure.
2775 * \retval -ENOMEM Memory allocation failure.
2776 * \retval -ENOENT Invalid MD object.
2778 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
2781 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
2782 struct lnet_process_id target, unsigned int portal,
2783 __u64 match_bits, unsigned int offset,
2786 struct lnet_msg *msg;
2787 struct lnet_libmd *md;
2791 LASSERT(the_lnet.ln_refcount > 0);
2793 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2794 fail_peer(target.nid, 1)) { /* shall we now? */
2795 CERROR("Dropping PUT to %s: simulated failure\n",
2796 libcfs_id2str(target));
2800 msg = lnet_msg_alloc();
2802 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
2803 libcfs_id2str(target));
2806 msg->msg_vmflush = !!memory_pressure_get();
2808 cpt = lnet_cpt_of_cookie(mdh.cookie);
2811 md = lnet_handle2md(&mdh);
2812 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2813 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
2814 match_bits, portal, libcfs_id2str(target),
2815 md == NULL ? -1 : md->md_threshold);
2816 if (md != NULL && md->md_me != NULL)
2817 CERROR("Source MD also attached to portal %d\n",
2818 md->md_me->me_portal);
2819 lnet_res_unlock(cpt);
2825 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2827 lnet_msg_attach_md(msg, md, 0, 0);
2829 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2831 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2832 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2833 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2834 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2836 /* NB handles only looked up by creator (no flips) */
2837 if (ack == LNET_ACK_REQ) {
2838 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2839 the_lnet.ln_interface_cookie;
2840 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2841 md->md_lh.lh_cookie;
2843 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2844 LNET_WIRE_HANDLE_COOKIE_NONE;
2845 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2846 LNET_WIRE_HANDLE_COOKIE_NONE;
2849 lnet_res_unlock(cpt);
2851 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2853 rc = lnet_send(self, msg, LNET_NID_ANY);
2855 CNETERR("Error sending PUT to %s: %d\n",
2856 libcfs_id2str(target), rc);
2857 lnet_finalize(msg, rc);
2860 /* completion will be signalled by an event */
2863 EXPORT_SYMBOL(LNetPut);
2866 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2867 * returns a msg for the LND to pass to lnet_finalize() when the sink
2868 * data has been received.
2870 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2871 * lnet_finalize() is called on it, so the LND must call this first
2874 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
2876 struct lnet_msg *msg = lnet_msg_alloc();
2877 struct lnet_libmd *getmd = getmsg->msg_md;
2878 struct lnet_process_id peer_id = getmsg->msg_target;
2881 LASSERT(!getmsg->msg_target_is_router);
2882 LASSERT(!getmsg->msg_routing);
2885 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2886 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2890 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2893 LASSERT(getmd->md_refcount > 0);
2895 if (getmd->md_threshold == 0) {
2896 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
2897 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2899 lnet_res_unlock(cpt);
2903 LASSERT(getmd->md_offset == 0);
2905 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2906 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2908 /* setup information for lnet_build_msg_event */
2909 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
2910 msg->msg_from = peer_id.nid;
2911 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2912 msg->msg_hdr.src_nid = peer_id.nid;
2913 msg->msg_hdr.payload_length = getmd->md_length;
2914 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2916 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2917 lnet_res_unlock(cpt);
2919 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
2922 lnet_msg_commit(msg, cpt);
2923 lnet_net_unlock(cpt);
2925 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2930 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
2933 the_lnet.ln_counters[cpt]->drop_count++;
2934 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2935 lnet_net_unlock(cpt);
2942 EXPORT_SYMBOL(lnet_create_reply_msg);
2945 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
2948 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2949 * completed and I know it. */
2950 LASSERT(reply != NULL);
2951 LASSERT(reply->msg_type == LNET_MSG_GET);
2952 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
2954 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2955 * the end of my buffer, I might as well be dead. */
2956 LASSERT(len <= reply->msg_ev.mlength);
2958 reply->msg_ev.mlength = len;
2960 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2963 * Initiate an asynchronous GET operation.
2965 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2966 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2967 * the target node in the REPLY has been written to local MD.
2969 * On the target node, an LNET_EVENT_GET is logged when the GET request
2970 * arrives and is accepted into a MD.
2972 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2973 * \param mdh A handle for the MD that describes the memory into which the
2974 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2976 * \retval 0 Success, and only in this case events will be generated
2977 * and logged to EQ (if it exists) of the MD.
2978 * \retval -EIO Simulated failure.
2979 * \retval -ENOMEM Memory allocation failure.
2980 * \retval -ENOENT Invalid MD object.
2983 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
2984 struct lnet_process_id target, unsigned int portal,
2985 __u64 match_bits, unsigned int offset)
2987 struct lnet_msg *msg;
2988 struct lnet_libmd *md;
2992 LASSERT(the_lnet.ln_refcount > 0);
2994 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2995 fail_peer(target.nid, 1)) /* shall we now? */
2997 CERROR("Dropping GET to %s: simulated failure\n",
2998 libcfs_id2str(target));
3002 msg = lnet_msg_alloc();
3004 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
3005 libcfs_id2str(target));
3009 cpt = lnet_cpt_of_cookie(mdh.cookie);
3012 md = lnet_handle2md(&mdh);
3013 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3014 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
3015 match_bits, portal, libcfs_id2str(target),
3016 md == NULL ? -1 : md->md_threshold);
3017 if (md != NULL && md->md_me != NULL)
3018 CERROR("REPLY MD also attached to portal %d\n",
3019 md->md_me->me_portal);
3021 lnet_res_unlock(cpt);
3027 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
3029 lnet_msg_attach_md(msg, md, 0, 0);
3031 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
3033 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
3034 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
3035 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
3036 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
3038 /* NB handles only looked up by creator (no flips) */
3039 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
3040 the_lnet.ln_interface_cookie;
3041 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
3042 md->md_lh.lh_cookie;
3044 lnet_res_unlock(cpt);
3046 lnet_build_msg_event(msg, LNET_EVENT_SEND);
3048 rc = lnet_send(self, msg, LNET_NID_ANY);
3050 CNETERR("Error sending GET to %s: %d\n",
3051 libcfs_id2str(target), rc);
3052 lnet_finalize(msg, rc);
3055 /* completion will be signalled by an event */
3058 EXPORT_SYMBOL(LNetGet);
3061 * Calculate distance to node at \a dstnid.
3063 * \param dstnid Target NID.
3064 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
3066 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
3069 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
3070 * local_nid_dist_zero is set, which is the default.
3071 * \retval positives Distance to target NID, i.e. number of hops plus one.
3072 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
3075 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
3077 struct list_head *e;
3078 struct lnet_ni *ni = NULL;
3079 struct lnet_remotenet *rnet;
3080 __u32 dstnet = LNET_NIDNET(dstnid);
3084 struct list_head *rn_list;
3086 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
3087 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
3088 * keep order 0 free for 0@lo and order 1 free for a local NID
3091 LASSERT(the_lnet.ln_refcount > 0);
3093 cpt = lnet_net_lock_current();
3095 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
3096 if (ni->ni_nid == dstnid) {
3097 if (srcnidp != NULL)
3099 if (orderp != NULL) {
3100 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
3105 lnet_net_unlock(cpt);
3107 return local_nid_dist_zero ? 0 : 1;
3110 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
3111 /* Check if ni was originally created in
3112 * current net namespace.
3113 * If not, assign order above 0xffff0000,
3114 * to make this ni not a priority. */
3115 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
3116 order += 0xffff0000;
3118 if (srcnidp != NULL)
3119 *srcnidp = ni->ni_nid;
3122 lnet_net_unlock(cpt);
3129 rn_list = lnet_net2rnethash(dstnet);
3130 list_for_each(e, rn_list) {
3131 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
3133 if (rnet->lrn_net == dstnet) {
3134 struct lnet_route *route;
3135 struct lnet_route *shortest = NULL;
3136 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
3139 LASSERT(!list_empty(&rnet->lrn_routes));
3141 list_for_each_entry(route, &rnet->lrn_routes,
3143 route_hops = route->lr_hops;
3144 if (route_hops == LNET_UNDEFINED_HOPS)
3146 if (shortest == NULL ||
3147 route_hops < shortest_hops) {
3149 shortest_hops = route_hops;
3153 LASSERT(shortest != NULL);
3154 hops = shortest_hops;
3155 if (srcnidp != NULL) {
3156 ni = lnet_get_next_ni_locked(
3157 shortest->lr_gateway->lpni_net,
3159 *srcnidp = ni->ni_nid;
3163 lnet_net_unlock(cpt);
3169 lnet_net_unlock(cpt);
3170 return -EHOSTUNREACH;
3172 EXPORT_SYMBOL(LNetDist);