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, 2017, 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 <linux/pagemap.h>
41 #include <lnet/lib-lnet.h>
42 #include <linux/nsproxy.h>
43 #include <net/net_namespace.h>
45 static int local_nid_dist_zero = 1;
46 module_param(local_nid_dist_zero, int, 0444);
47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
49 struct lnet_send_data {
50 struct lnet_ni *sd_best_ni;
51 struct lnet_peer_ni *sd_best_lpni;
52 struct lnet_peer_ni *sd_final_dst_lpni;
53 struct lnet_peer *sd_peer;
54 struct lnet_peer *sd_gw_peer;
55 struct lnet_peer_ni *sd_gw_lpni;
56 struct lnet_peer_net *sd_peer_net;
57 struct lnet_msg *sd_msg;
58 lnet_nid_t sd_dst_nid;
59 lnet_nid_t sd_src_nid;
60 lnet_nid_t sd_rtr_nid;
66 static inline struct lnet_comm_count *
67 get_stats_counts(struct lnet_element_stats *stats,
68 enum lnet_stats_type stats_type)
71 case LNET_STATS_TYPE_SEND:
72 return &stats->el_send_stats;
73 case LNET_STATS_TYPE_RECV:
74 return &stats->el_recv_stats;
75 case LNET_STATS_TYPE_DROP:
76 return &stats->el_drop_stats;
78 CERROR("Unknown stats type\n");
84 void lnet_incr_stats(struct lnet_element_stats *stats,
85 enum lnet_msg_type msg_type,
86 enum lnet_stats_type stats_type)
88 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
94 atomic_inc(&counts->co_ack_count);
97 atomic_inc(&counts->co_put_count);
100 atomic_inc(&counts->co_get_count);
103 atomic_inc(&counts->co_reply_count);
106 atomic_inc(&counts->co_hello_count);
109 CERROR("There is a BUG in the code. Unknown message type\n");
114 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
115 enum lnet_stats_type stats_type)
117 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
121 return (atomic_read(&counts->co_ack_count) +
122 atomic_read(&counts->co_put_count) +
123 atomic_read(&counts->co_get_count) +
124 atomic_read(&counts->co_reply_count) +
125 atomic_read(&counts->co_hello_count));
128 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
129 struct lnet_comm_count *counts)
131 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
132 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
133 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
134 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
135 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
138 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
139 struct lnet_element_stats *stats)
141 struct lnet_comm_count *counts;
146 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
149 assign_stats(&msg_stats->im_send_stats, counts);
151 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
154 assign_stats(&msg_stats->im_recv_stats, counts);
156 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
159 assign_stats(&msg_stats->im_drop_stats, counts);
163 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
165 struct lnet_test_peer *tp;
166 struct list_head *el;
167 struct list_head *next;
168 struct list_head cull;
170 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
171 if (threshold != 0) {
172 /* Adding a new entry */
173 LIBCFS_ALLOC(tp, sizeof(*tp));
178 tp->tp_threshold = threshold;
181 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
186 /* removing entries */
187 INIT_LIST_HEAD(&cull);
191 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
192 tp = list_entry(el, struct lnet_test_peer, tp_list);
194 if (tp->tp_threshold == 0 || /* needs culling anyway */
195 nid == LNET_NID_ANY || /* removing all entries */
196 tp->tp_nid == nid) { /* matched this one */
197 list_del(&tp->tp_list);
198 list_add(&tp->tp_list, &cull);
204 while (!list_empty(&cull)) {
205 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
207 list_del(&tp->tp_list);
208 LIBCFS_FREE(tp, sizeof(*tp));
214 fail_peer (lnet_nid_t nid, int outgoing)
216 struct lnet_test_peer *tp;
217 struct list_head *el;
218 struct list_head *next;
219 struct list_head cull;
222 INIT_LIST_HEAD(&cull);
224 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
227 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
228 tp = list_entry(el, struct lnet_test_peer, tp_list);
230 if (tp->tp_threshold == 0) {
233 /* only cull zombies on outgoing tests,
234 * since we may be at interrupt priority on
235 * incoming messages. */
236 list_del(&tp->tp_list);
237 list_add(&tp->tp_list, &cull);
242 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
243 nid == tp->tp_nid) { /* fail this peer */
246 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
249 tp->tp_threshold == 0) {
251 list_del(&tp->tp_list);
252 list_add(&tp->tp_list, &cull);
261 while (!list_empty(&cull)) {
262 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
263 list_del(&tp->tp_list);
265 LIBCFS_FREE(tp, sizeof(*tp));
272 lnet_iov_nob(unsigned int niov, struct kvec *iov)
274 unsigned int nob = 0;
276 LASSERT(niov == 0 || iov != NULL);
278 nob += (iov++)->iov_len;
282 EXPORT_SYMBOL(lnet_iov_nob);
285 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
286 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
289 /* NB diov, siov are READ-ONLY */
290 unsigned int this_nob;
295 /* skip complete frags before 'doffset' */
297 while (doffset >= diov->iov_len) {
298 doffset -= diov->iov_len;
304 /* skip complete frags before 'soffset' */
306 while (soffset >= siov->iov_len) {
307 soffset -= siov->iov_len;
316 this_nob = MIN(diov->iov_len - doffset,
317 siov->iov_len - soffset);
318 this_nob = MIN(this_nob, nob);
320 memcpy((char *)diov->iov_base + doffset,
321 (char *)siov->iov_base + soffset, this_nob);
324 if (diov->iov_len > doffset + this_nob) {
332 if (siov->iov_len > soffset + this_nob) {
341 EXPORT_SYMBOL(lnet_copy_iov2iov);
344 lnet_extract_iov(int dst_niov, struct kvec *dst,
345 int src_niov, struct kvec *src,
346 unsigned int offset, unsigned int len)
348 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
349 * for exactly 'len' bytes, and return the number of entries.
350 * NB not destructive to 'src' */
351 unsigned int frag_len;
354 if (len == 0) /* no data => */
355 return (0); /* no frags */
357 LASSERT(src_niov > 0);
358 while (offset >= src->iov_len) { /* skip initial frags */
359 offset -= src->iov_len;
362 LASSERT(src_niov > 0);
367 LASSERT(src_niov > 0);
368 LASSERT((int)niov <= dst_niov);
370 frag_len = src->iov_len - offset;
371 dst->iov_base = ((char *)src->iov_base) + offset;
373 if (len <= frag_len) {
378 dst->iov_len = frag_len;
388 EXPORT_SYMBOL(lnet_extract_iov);
392 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
394 unsigned int nob = 0;
396 LASSERT(niov == 0 || kiov != NULL);
398 nob += (kiov++)->kiov_len;
402 EXPORT_SYMBOL(lnet_kiov_nob);
405 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
406 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
409 /* NB diov, siov are READ-ONLY */
410 unsigned int this_nob;
417 LASSERT (!in_interrupt ());
420 while (doffset >= diov->kiov_len) {
421 doffset -= diov->kiov_len;
428 while (soffset >= siov->kiov_len) {
429 soffset -= siov->kiov_len;
438 this_nob = MIN(diov->kiov_len - doffset,
439 siov->kiov_len - soffset);
440 this_nob = MIN(this_nob, nob);
443 daddr = ((char *)kmap(diov->kiov_page)) +
444 diov->kiov_offset + doffset;
446 saddr = ((char *)kmap(siov->kiov_page)) +
447 siov->kiov_offset + soffset;
449 /* Vanishing risk of kmap deadlock when mapping 2 pages.
450 * However in practice at least one of the kiovs will be mapped
451 * kernel pages and the map/unmap will be NOOPs */
453 memcpy (daddr, saddr, this_nob);
456 if (diov->kiov_len > doffset + this_nob) {
460 kunmap(diov->kiov_page);
467 if (siov->kiov_len > soffset + this_nob) {
471 kunmap(siov->kiov_page);
480 kunmap(diov->kiov_page);
482 kunmap(siov->kiov_page);
484 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
487 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
488 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
491 /* NB iov, kiov are READ-ONLY */
492 unsigned int this_nob;
498 LASSERT (!in_interrupt ());
501 while (iovoffset >= iov->iov_len) {
502 iovoffset -= iov->iov_len;
509 while (kiovoffset >= kiov->kiov_len) {
510 kiovoffset -= kiov->kiov_len;
519 this_nob = MIN(iov->iov_len - iovoffset,
520 kiov->kiov_len - kiovoffset);
521 this_nob = MIN(this_nob, nob);
524 addr = ((char *)kmap(kiov->kiov_page)) +
525 kiov->kiov_offset + kiovoffset;
527 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
530 if (iov->iov_len > iovoffset + this_nob) {
531 iovoffset += this_nob;
538 if (kiov->kiov_len > kiovoffset + this_nob) {
540 kiovoffset += this_nob;
542 kunmap(kiov->kiov_page);
552 kunmap(kiov->kiov_page);
554 EXPORT_SYMBOL(lnet_copy_kiov2iov);
557 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
558 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
561 /* NB kiov, iov are READ-ONLY */
562 unsigned int this_nob;
568 LASSERT (!in_interrupt ());
571 while (kiovoffset >= kiov->kiov_len) {
572 kiovoffset -= kiov->kiov_len;
579 while (iovoffset >= iov->iov_len) {
580 iovoffset -= iov->iov_len;
589 this_nob = MIN(kiov->kiov_len - kiovoffset,
590 iov->iov_len - iovoffset);
591 this_nob = MIN(this_nob, nob);
594 addr = ((char *)kmap(kiov->kiov_page)) +
595 kiov->kiov_offset + kiovoffset;
597 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
600 if (kiov->kiov_len > kiovoffset + this_nob) {
602 kiovoffset += this_nob;
604 kunmap(kiov->kiov_page);
611 if (iov->iov_len > iovoffset + this_nob) {
612 iovoffset += this_nob;
621 kunmap(kiov->kiov_page);
623 EXPORT_SYMBOL(lnet_copy_iov2kiov);
626 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
627 int src_niov, lnet_kiov_t *src,
628 unsigned int offset, unsigned int len)
630 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
631 * for exactly 'len' bytes, and return the number of entries.
632 * NB not destructive to 'src' */
633 unsigned int frag_len;
636 if (len == 0) /* no data => */
637 return (0); /* no frags */
639 LASSERT(src_niov > 0);
640 while (offset >= src->kiov_len) { /* skip initial frags */
641 offset -= src->kiov_len;
644 LASSERT(src_niov > 0);
649 LASSERT(src_niov > 0);
650 LASSERT((int)niov <= dst_niov);
652 frag_len = src->kiov_len - offset;
653 dst->kiov_page = src->kiov_page;
654 dst->kiov_offset = src->kiov_offset + offset;
656 if (len <= frag_len) {
658 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
662 dst->kiov_len = frag_len;
663 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
673 EXPORT_SYMBOL(lnet_extract_kiov);
676 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
677 int delayed, unsigned int offset, unsigned int mlen,
680 unsigned int niov = 0;
681 struct kvec *iov = NULL;
682 lnet_kiov_t *kiov = NULL;
685 LASSERT (!in_interrupt ());
686 LASSERT (mlen == 0 || msg != NULL);
689 LASSERT(msg->msg_receiving);
690 LASSERT(!msg->msg_sending);
691 LASSERT(rlen == msg->msg_len);
692 LASSERT(mlen <= msg->msg_len);
693 LASSERT(msg->msg_offset == offset);
694 LASSERT(msg->msg_wanted == mlen);
696 msg->msg_receiving = 0;
699 niov = msg->msg_niov;
701 kiov = msg->msg_kiov;
704 LASSERT ((iov == NULL) != (kiov == NULL));
708 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
709 niov, iov, kiov, offset, mlen,
712 lnet_finalize(msg, rc);
716 lnet_setpayloadbuffer(struct lnet_msg *msg)
718 struct lnet_libmd *md = msg->msg_md;
720 LASSERT(msg->msg_len > 0);
721 LASSERT(!msg->msg_routing);
723 LASSERT(msg->msg_niov == 0);
724 LASSERT(msg->msg_iov == NULL);
725 LASSERT(msg->msg_kiov == NULL);
727 msg->msg_niov = md->md_niov;
728 if ((md->md_options & LNET_MD_KIOV) != 0)
729 msg->msg_kiov = md->md_iov.kiov;
731 msg->msg_iov = md->md_iov.iov;
735 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
736 unsigned int offset, unsigned int len)
738 msg->msg_type = type;
739 msg->msg_target = target;
741 msg->msg_offset = offset;
744 lnet_setpayloadbuffer(msg);
746 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
747 msg->msg_hdr.type = cpu_to_le32(type);
748 /* dest_nid will be overwritten by lnet_select_pathway() */
749 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
750 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
751 /* src_nid will be set later */
752 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
753 msg->msg_hdr.payload_length = cpu_to_le32(len);
757 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
759 void *priv = msg->msg_private;
762 LASSERT (!in_interrupt ());
763 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
764 (msg->msg_txcredit && msg->msg_peertxcredit));
766 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
768 lnet_finalize(msg, rc);
772 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
776 LASSERT(!msg->msg_sending);
777 LASSERT(msg->msg_receiving);
778 LASSERT(!msg->msg_rx_ready_delay);
779 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
781 msg->msg_rx_ready_delay = 1;
782 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
785 CERROR("recv from %s / send to %s aborted: "
786 "eager_recv failed %d\n",
787 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
788 libcfs_id2str(msg->msg_target), rc);
789 LASSERT(rc < 0); /* required by my callers */
796 * This function can be called from two paths:
797 * 1. when sending a message
798 * 2. when decommiting a message (lnet_msg_decommit_tx())
799 * In both these cases the peer_ni should have it's reference count
800 * acquired by the caller and therefore it is safe to drop the spin
801 * lock before calling lnd_query()
804 lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
806 time64_t last_alive = 0;
807 int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
809 LASSERT(lnet_peer_aliveness_enabled(lp));
810 LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
812 lnet_net_unlock(cpt);
813 (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
816 lp->lpni_last_query = ktime_get_seconds();
818 if (last_alive != 0) /* NI has updated timestamp */
819 lp->lpni_last_alive = last_alive;
822 /* NB: always called with lnet_net_lock held */
824 lnet_peer_is_alive(struct lnet_peer_ni *lp, time64_t now)
829 LASSERT (lnet_peer_aliveness_enabled(lp));
832 * Trust lnet_notify() if it has more recent aliveness news, but
833 * ignore the initial assumed death (see lnet_peers_start_down()).
835 spin_lock(&lp->lpni_lock);
836 if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
837 lp->lpni_timestamp >= lp->lpni_last_alive) {
838 spin_unlock(&lp->lpni_lock);
842 deadline = lp->lpni_last_alive +
843 lp->lpni_net->net_tunables.lct_peer_timeout;
844 alive = deadline > now;
847 * Update obsolete lp_alive except for routers assumed to be dead
848 * initially, because router checker would update aliveness in this
849 * case, and moreover lpni_last_alive at peer creation is assumed.
851 if (alive && !lp->lpni_alive &&
852 !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
853 spin_unlock(&lp->lpni_lock);
854 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
856 spin_unlock(&lp->lpni_lock);
863 /* NB: returns 1 when alive, 0 when dead, negative when error;
864 * may drop the lnet_net_lock */
866 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
868 time64_t now = ktime_get_seconds();
870 if (!lnet_peer_aliveness_enabled(lp))
873 if (lnet_peer_is_alive(lp, now))
877 * Peer appears dead, but we should avoid frequent NI queries (at
878 * most once per lnet_queryinterval seconds).
880 if (lp->lpni_last_query != 0) {
881 static const int lnet_queryinterval = 1;
884 next_query = lp->lpni_last_query + lnet_queryinterval;
886 if (now < next_query) {
888 CWARN("Unexpected aliveness of peer %s: "
889 "%lld < %lld (%d/%d)\n",
890 libcfs_nid2str(lp->lpni_nid),
893 lp->lpni_net->net_tunables.lct_peer_timeout);
898 /* query NI for latest aliveness news */
899 lnet_ni_query_locked(ni, lp);
901 if (lnet_peer_is_alive(lp, now))
904 lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
909 * \param msg The message to be sent.
910 * \param do_send True if lnet_ni_send() should be called in this function.
911 * lnet_send() is going to lnet_net_unlock immediately after this, so
912 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
914 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
915 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
916 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
917 * \retval -ECANCELED If the MD of the message has been unlinked.
920 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
922 struct lnet_peer_ni *lp = msg->msg_txpeer;
923 struct lnet_ni *ni = msg->msg_txni;
924 int cpt = msg->msg_tx_cpt;
925 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
927 /* non-lnet_send() callers have checked before */
928 LASSERT(!do_send || msg->msg_tx_delayed);
929 LASSERT(!msg->msg_receiving);
930 LASSERT(msg->msg_tx_committed);
932 /* NB 'lp' is always the next hop */
933 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
934 lnet_peer_alive_locked(ni, lp) == 0) {
935 the_lnet.ln_counters[cpt]->drop_count++;
936 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
937 lnet_net_unlock(cpt);
939 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
941 LNET_STATS_TYPE_DROP);
943 lnet_incr_stats(&msg->msg_txni->ni_stats,
945 LNET_STATS_TYPE_DROP);
947 CNETERR("Dropping message for %s: peer not alive\n",
948 libcfs_id2str(msg->msg_target));
950 lnet_finalize(msg, -EHOSTUNREACH);
953 return -EHOSTUNREACH;
956 if (msg->msg_md != NULL &&
957 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
958 lnet_net_unlock(cpt);
960 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
961 "called on the MD/ME.\n",
962 libcfs_id2str(msg->msg_target));
964 lnet_finalize(msg, -ECANCELED);
970 if (!msg->msg_peertxcredit) {
971 spin_lock(&lp->lpni_lock);
972 LASSERT((lp->lpni_txcredits < 0) ==
973 !list_empty(&lp->lpni_txq));
975 msg->msg_peertxcredit = 1;
976 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
977 lp->lpni_txcredits--;
979 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
980 lp->lpni_mintxcredits = lp->lpni_txcredits;
982 if (lp->lpni_txcredits < 0) {
983 msg->msg_tx_delayed = 1;
984 list_add_tail(&msg->msg_list, &lp->lpni_txq);
985 spin_unlock(&lp->lpni_lock);
986 return LNET_CREDIT_WAIT;
988 spin_unlock(&lp->lpni_lock);
991 if (!msg->msg_txcredit) {
992 LASSERT((tq->tq_credits < 0) ==
993 !list_empty(&tq->tq_delayed));
995 msg->msg_txcredit = 1;
997 atomic_dec(&ni->ni_tx_credits);
999 if (tq->tq_credits < tq->tq_credits_min)
1000 tq->tq_credits_min = tq->tq_credits;
1002 if (tq->tq_credits < 0) {
1003 msg->msg_tx_delayed = 1;
1004 list_add_tail(&msg->msg_list, &tq->tq_delayed);
1005 return LNET_CREDIT_WAIT;
1009 /* unset the tx_delay flag as we're going to send it now */
1010 msg->msg_tx_delayed = 0;
1013 lnet_net_unlock(cpt);
1014 lnet_ni_send(ni, msg);
1017 return LNET_CREDIT_OK;
1021 static struct lnet_rtrbufpool *
1022 lnet_msg2bufpool(struct lnet_msg *msg)
1024 struct lnet_rtrbufpool *rbp;
1027 LASSERT(msg->msg_rx_committed);
1029 cpt = msg->msg_rx_cpt;
1030 rbp = &the_lnet.ln_rtrpools[cpt][0];
1032 LASSERT(msg->msg_len <= LNET_MTU);
1033 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
1035 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
1042 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
1044 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
1045 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
1046 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
1047 * received or OK to receive */
1048 struct lnet_peer_ni *lp = msg->msg_rxpeer;
1049 struct lnet_rtrbufpool *rbp;
1050 struct lnet_rtrbuf *rb;
1052 LASSERT (msg->msg_iov == NULL);
1053 LASSERT (msg->msg_kiov == NULL);
1054 LASSERT (msg->msg_niov == 0);
1055 LASSERT (msg->msg_routing);
1056 LASSERT (msg->msg_receiving);
1057 LASSERT (!msg->msg_sending);
1059 /* non-lnet_parse callers only receive delayed messages */
1060 LASSERT(!do_recv || msg->msg_rx_delayed);
1062 if (!msg->msg_peerrtrcredit) {
1063 spin_lock(&lp->lpni_lock);
1064 LASSERT((lp->lpni_rtrcredits < 0) ==
1065 !list_empty(&lp->lpni_rtrq));
1067 msg->msg_peerrtrcredit = 1;
1068 lp->lpni_rtrcredits--;
1069 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
1070 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
1072 if (lp->lpni_rtrcredits < 0) {
1073 /* must have checked eager_recv before here */
1074 LASSERT(msg->msg_rx_ready_delay);
1075 msg->msg_rx_delayed = 1;
1076 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
1077 spin_unlock(&lp->lpni_lock);
1078 return LNET_CREDIT_WAIT;
1080 spin_unlock(&lp->lpni_lock);
1083 rbp = lnet_msg2bufpool(msg);
1085 if (!msg->msg_rtrcredit) {
1086 msg->msg_rtrcredit = 1;
1088 if (rbp->rbp_credits < rbp->rbp_mincredits)
1089 rbp->rbp_mincredits = rbp->rbp_credits;
1091 if (rbp->rbp_credits < 0) {
1092 /* must have checked eager_recv before here */
1093 LASSERT(msg->msg_rx_ready_delay);
1094 msg->msg_rx_delayed = 1;
1095 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1096 return LNET_CREDIT_WAIT;
1100 LASSERT(!list_empty(&rbp->rbp_bufs));
1101 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1102 list_del(&rb->rb_list);
1104 msg->msg_niov = rbp->rbp_npages;
1105 msg->msg_kiov = &rb->rb_kiov[0];
1107 /* unset the msg-rx_delayed flag since we're receiving the message */
1108 msg->msg_rx_delayed = 0;
1111 int cpt = msg->msg_rx_cpt;
1113 lnet_net_unlock(cpt);
1114 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1115 0, msg->msg_len, msg->msg_len);
1118 return LNET_CREDIT_OK;
1122 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1124 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1125 struct lnet_ni *txni = msg->msg_txni;
1126 struct lnet_msg *msg2;
1128 if (msg->msg_txcredit) {
1129 struct lnet_ni *ni = msg->msg_txni;
1130 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1132 /* give back NI txcredits */
1133 msg->msg_txcredit = 0;
1135 LASSERT((tq->tq_credits < 0) ==
1136 !list_empty(&tq->tq_delayed));
1139 atomic_inc(&ni->ni_tx_credits);
1140 if (tq->tq_credits <= 0) {
1141 msg2 = list_entry(tq->tq_delayed.next,
1142 struct lnet_msg, msg_list);
1143 list_del(&msg2->msg_list);
1145 LASSERT(msg2->msg_txni == ni);
1146 LASSERT(msg2->msg_tx_delayed);
1147 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1149 (void) lnet_post_send_locked(msg2, 1);
1153 if (msg->msg_peertxcredit) {
1154 /* give back peer txcredits */
1155 msg->msg_peertxcredit = 0;
1157 spin_lock(&txpeer->lpni_lock);
1158 LASSERT((txpeer->lpni_txcredits < 0) ==
1159 !list_empty(&txpeer->lpni_txq));
1161 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1162 LASSERT(txpeer->lpni_txqnob >= 0);
1164 txpeer->lpni_txcredits++;
1165 if (txpeer->lpni_txcredits <= 0) {
1168 msg2 = list_entry(txpeer->lpni_txq.next,
1169 struct lnet_msg, msg_list);
1170 list_del(&msg2->msg_list);
1171 spin_unlock(&txpeer->lpni_lock);
1173 LASSERT(msg2->msg_txpeer == txpeer);
1174 LASSERT(msg2->msg_tx_delayed);
1176 msg2_cpt = msg2->msg_tx_cpt;
1179 * The msg_cpt can be different from the msg2_cpt
1180 * so we need to make sure we lock the correct cpt
1182 * Once we call lnet_post_send_locked() it is no
1183 * longer safe to access msg2, since it could've
1184 * been freed by lnet_finalize(), but we still
1185 * need to relock the correct cpt, so we cache the
1186 * msg2_cpt for the purpose of the check that
1187 * follows the call to lnet_pose_send_locked().
1189 if (msg2_cpt != msg->msg_tx_cpt) {
1190 lnet_net_unlock(msg->msg_tx_cpt);
1191 lnet_net_lock(msg2_cpt);
1193 (void) lnet_post_send_locked(msg2, 1);
1194 if (msg2_cpt != msg->msg_tx_cpt) {
1195 lnet_net_unlock(msg2_cpt);
1196 lnet_net_lock(msg->msg_tx_cpt);
1199 spin_unlock(&txpeer->lpni_lock);
1204 msg->msg_txni = NULL;
1205 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1208 if (txpeer != NULL) {
1211 * Once the patch for the health comes in we need to set
1212 * the health of the peer ni to bad when we fail to send
1214 * int status = msg->msg_ev.status;
1216 * lnet_set_peer_ni_health_locked(txpeer, false)
1218 msg->msg_txpeer = NULL;
1219 lnet_peer_ni_decref_locked(txpeer);
1224 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1226 struct lnet_msg *msg;
1228 if (list_empty(&rbp->rbp_msgs))
1230 msg = list_entry(rbp->rbp_msgs.next,
1231 struct lnet_msg, msg_list);
1232 list_del(&msg->msg_list);
1234 (void)lnet_post_routed_recv_locked(msg, 1);
1238 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1240 struct lnet_msg *msg;
1241 struct lnet_msg *tmp;
1243 lnet_net_unlock(cpt);
1245 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1246 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1247 0, 0, 0, msg->msg_hdr.payload_length);
1248 list_del_init(&msg->msg_list);
1249 lnet_finalize(msg, -ECANCELED);
1256 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1258 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1259 struct lnet_ni *rxni = msg->msg_rxni;
1260 struct lnet_msg *msg2;
1262 if (msg->msg_rtrcredit) {
1263 /* give back global router credits */
1264 struct lnet_rtrbuf *rb;
1265 struct lnet_rtrbufpool *rbp;
1267 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1268 * there until it gets one allocated, or aborts the wait
1270 LASSERT(msg->msg_kiov != NULL);
1272 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1275 msg->msg_kiov = NULL;
1276 msg->msg_rtrcredit = 0;
1278 LASSERT(rbp == lnet_msg2bufpool(msg));
1280 LASSERT((rbp->rbp_credits > 0) ==
1281 !list_empty(&rbp->rbp_bufs));
1283 /* If routing is now turned off, we just drop this buffer and
1284 * don't bother trying to return credits. */
1285 if (!the_lnet.ln_routing) {
1286 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1290 /* It is possible that a user has lowered the desired number of
1291 * buffers in this pool. Make sure we never put back
1292 * more buffers than the stated number. */
1293 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1294 /* Discard this buffer so we don't have too
1296 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1297 rbp->rbp_nbuffers--;
1299 list_add(&rb->rb_list, &rbp->rbp_bufs);
1301 if (rbp->rbp_credits <= 0)
1302 lnet_schedule_blocked_locked(rbp);
1307 if (msg->msg_peerrtrcredit) {
1308 /* give back peer router credits */
1309 msg->msg_peerrtrcredit = 0;
1311 spin_lock(&rxpeer->lpni_lock);
1312 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1313 !list_empty(&rxpeer->lpni_rtrq));
1315 rxpeer->lpni_rtrcredits++;
1317 /* drop all messages which are queued to be routed on that
1319 if (!the_lnet.ln_routing) {
1320 struct list_head drop;
1321 INIT_LIST_HEAD(&drop);
1322 list_splice_init(&rxpeer->lpni_rtrq, &drop);
1323 spin_unlock(&rxpeer->lpni_lock);
1324 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1325 } else if (rxpeer->lpni_rtrcredits <= 0) {
1326 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1327 struct lnet_msg, msg_list);
1328 list_del(&msg2->msg_list);
1329 spin_unlock(&rxpeer->lpni_lock);
1330 (void) lnet_post_routed_recv_locked(msg2, 1);
1332 spin_unlock(&rxpeer->lpni_lock);
1336 msg->msg_rxni = NULL;
1337 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1339 if (rxpeer != NULL) {
1340 msg->msg_rxpeer = NULL;
1341 lnet_peer_ni_decref_locked(rxpeer);
1346 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1348 if (p1->lpni_txqnob < p2->lpni_txqnob)
1351 if (p1->lpni_txqnob > p2->lpni_txqnob)
1354 if (p1->lpni_txcredits > p2->lpni_txcredits)
1357 if (p1->lpni_txcredits < p2->lpni_txcredits)
1364 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1366 struct lnet_peer_ni *p1 = r1->lr_gateway;
1367 struct lnet_peer_ni *p2 = r2->lr_gateway;
1368 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1369 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1372 if (r1->lr_priority < r2->lr_priority)
1375 if (r1->lr_priority > r2->lr_priority)
1378 if (r1_hops < r2_hops)
1381 if (r1_hops > r2_hops)
1384 rc = lnet_compare_peers(p1, p2);
1388 if (r1->lr_seq - r2->lr_seq <= 0)
1394 static struct lnet_peer_ni *
1395 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1398 struct lnet_remotenet *rnet;
1399 struct lnet_route *route;
1400 struct lnet_route *best_route;
1401 struct lnet_route *last_route;
1402 struct lnet_peer_ni *lpni_best;
1403 struct lnet_peer_ni *lp;
1406 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1407 * rtr_nid nid, otherwise find the best gateway I can use */
1409 rnet = lnet_find_rnet_locked(remote_net);
1414 best_route = last_route = NULL;
1415 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1416 lp = route->lr_gateway;
1418 if (!lnet_is_route_alive(route))
1421 if (net != NULL && lp->lpni_net != net)
1424 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1427 if (lpni_best == NULL) {
1428 best_route = last_route = route;
1433 /* no protection on below fields, but it's harmless */
1434 if (last_route->lr_seq - route->lr_seq < 0)
1437 rc = lnet_compare_routes(route, best_route);
1445 /* set sequence number on the best router to the latest sequence + 1
1446 * so we can round-robin all routers, it's race and inaccurate but
1447 * harmless and functional */
1448 if (best_route != NULL)
1449 best_route->lr_seq = last_route->lr_seq + 1;
1453 static struct lnet_ni *
1454 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1455 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1458 struct lnet_ni *ni = NULL;
1459 unsigned int shortest_distance;
1464 * If there is no peer_ni that we can send to on this network,
1465 * then there is no point in looking for a new best_ni here.
1467 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1470 if (best_ni == NULL) {
1471 shortest_distance = UINT_MAX;
1472 best_credits = INT_MIN;
1475 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1476 best_ni->ni_dev_cpt);
1477 best_credits = atomic_read(&best_ni->ni_tx_credits);
1478 best_healthv = atomic_read(&best_ni->ni_healthv);
1481 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1482 unsigned int distance;
1486 ni_credits = atomic_read(&ni->ni_tx_credits);
1487 ni_healthv = atomic_read(&ni->ni_healthv);
1490 * calculate the distance from the CPT on which
1491 * the message memory is allocated to the CPT of
1492 * the NI's physical device
1494 distance = cfs_cpt_distance(lnet_cpt_table(),
1498 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1499 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1500 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1501 : "not seleced", best_credits, shortest_distance,
1502 (best_ni) ? best_ni->ni_seq : 0);
1505 * All distances smaller than the NUMA range
1506 * are treated equally.
1508 if (distance < lnet_numa_range)
1509 distance = lnet_numa_range;
1512 * Select on health, shorter distance, available
1513 * credits, then round-robin.
1515 if (ni_healthv < best_healthv) {
1517 } else if (ni_healthv > best_healthv) {
1518 best_healthv = ni_healthv;
1520 * If we're going to prefer this ni because it's
1521 * the healthiest, then we should set the
1522 * shortest_distance in the algorithm in case
1523 * there are multiple NIs with the same health but
1524 * different distances.
1526 if (distance < shortest_distance)
1527 shortest_distance = distance;
1528 } else if (distance > shortest_distance) {
1530 } else if (distance < shortest_distance) {
1531 shortest_distance = distance;
1532 } else if (ni_credits < best_credits) {
1534 } else if (ni_credits == best_credits) {
1535 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1539 best_credits = ni_credits;
1542 CDEBUG(D_NET, "selected best_ni %s\n",
1543 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1549 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1550 * because such traffic is required to perform discovery. We therefore
1551 * exclude all GET and PUT on that portal. We also exclude all ACK and
1552 * REPLY traffic, but that is because the portal is not tracked in the
1553 * message structure for these message types. We could restrict this
1554 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1557 lnet_msg_discovery(struct lnet_msg *msg)
1559 if (msg->msg_type == LNET_MSG_PUT) {
1560 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1562 } else if (msg->msg_type == LNET_MSG_GET) {
1563 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1569 #define SRC_SPEC 0x0001
1570 #define SRC_ANY 0x0002
1571 #define LOCAL_DST 0x0004
1572 #define REMOTE_DST 0x0008
1573 #define MR_DST 0x0010
1574 #define NMR_DST 0x0020
1575 #define SND_RESP 0x0040
1577 /* The following to defines are used for return codes */
1578 #define REPEAT_SEND 0x1000
1579 #define PASS_THROUGH 0x2000
1581 /* The different cases lnet_select pathway needs to handle */
1582 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1583 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1584 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1585 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1586 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1587 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1588 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1589 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1592 lnet_handle_send(struct lnet_send_data *sd)
1594 struct lnet_ni *best_ni = sd->sd_best_ni;
1595 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1596 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1597 struct lnet_msg *msg = sd->sd_msg;
1599 __u32 send_case = sd->sd_send_case;
1601 __u32 routing = send_case & REMOTE_DST;
1604 * Increment sequence number of the selected peer so that we
1605 * pick the next one in Round Robin.
1607 best_lpni->lpni_seq++;
1610 * grab a reference on the peer_ni so it sticks around even if
1611 * we need to drop and relock the lnet_net_lock below.
1613 lnet_peer_ni_addref_locked(best_lpni);
1616 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1617 * message. This ensures that we get a CPT that is correct for
1618 * the NI when the NI has been restricted to a subset of all CPTs.
1619 * If the selected CPT differs from the one currently locked, we
1620 * must unlock and relock the lnet_net_lock(), and then check whether
1621 * the configuration has changed. We don't have a hold on the best_ni
1622 * yet, and it may have vanished.
1624 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1625 if (sd->sd_cpt != cpt2) {
1626 __u32 seq = lnet_get_dlc_seq_locked();
1627 lnet_net_unlock(sd->sd_cpt);
1629 lnet_net_lock(sd->sd_cpt);
1630 if (seq != lnet_get_dlc_seq_locked()) {
1631 lnet_peer_ni_decref_locked(best_lpni);
1637 * store the best_lpni in the message right away to avoid having
1638 * to do the same operation under different conditions
1640 msg->msg_txpeer = best_lpni;
1641 msg->msg_txni = best_ni;
1644 * grab a reference for the best_ni since now it's in use in this
1645 * send. The reference will be dropped in lnet_finalize()
1647 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1650 * Always set the target.nid to the best peer picked. Either the
1651 * NID will be one of the peer NIDs selected, or the same NID as
1652 * what was originally set in the target or it will be the NID of
1653 * a router if this message should be routed
1655 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1658 * lnet_msg_commit assigns the correct cpt to the message, which
1659 * is used to decrement the correct refcount on the ni when it's
1660 * time to return the credits
1662 lnet_msg_commit(msg, sd->sd_cpt);
1665 * If we are routing the message then we keep the src_nid that was
1666 * set by the originator. If we are not routing then we are the
1667 * originator and set it here.
1669 if (!msg->msg_routing)
1670 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1673 msg->msg_target_is_router = 1;
1674 msg->msg_target.pid = LNET_PID_LUSTRE;
1676 * since we're routing we want to ensure that the
1677 * msg_hdr.dest_nid is set to the final destination. When
1678 * the router receives this message it knows how to route
1681 * final_dst_lpni is set at the beginning of the
1682 * lnet_select_pathway() function and is never changed.
1683 * It's safe to use it here.
1685 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1688 * if we're not routing set the dest_nid to the best peer
1689 * ni NID that we picked earlier in the algorithm.
1691 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1694 rc = lnet_post_send_locked(msg, 0);
1697 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s\n",
1698 libcfs_nid2str(msg->msg_hdr.src_nid),
1699 libcfs_nid2str(msg->msg_txni->ni_nid),
1700 libcfs_nid2str(sd->sd_src_nid),
1701 libcfs_nid2str(msg->msg_hdr.dest_nid),
1702 libcfs_nid2str(sd->sd_dst_nid),
1703 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1704 lnet_msgtyp2str(msg->msg_type));
1709 static struct lnet_peer_ni *
1710 lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer,
1711 struct lnet_peer_net *peer_net)
1714 * Look at the peer NIs for the destination peer that connect
1715 * to the chosen net. If a peer_ni is preferred when using the
1716 * best_ni to communicate, we use that one. If there is no
1717 * preferred peer_ni, or there are multiple preferred peer_ni,
1718 * the available transmit credits are used. If the transmit
1719 * credits are equal, we round-robin over the peer_ni.
1721 struct lnet_peer_ni *lpni = NULL;
1722 struct lnet_peer_ni *best_lpni = NULL;
1723 struct lnet_ni *best_ni = sd->sd_best_ni;
1724 lnet_nid_t dst_nid = sd->sd_dst_nid;
1725 int best_lpni_credits = INT_MIN;
1726 bool preferred = false;
1729 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1731 * if the best_ni we've chosen aleady has this lpni
1732 * preferred, then let's use it
1734 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1737 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1738 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1741 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1742 libcfs_nid2str(lpni->lpni_nid),
1743 lpni->lpni_txcredits, best_lpni_credits,
1744 lpni->lpni_seq, best_lpni->lpni_seq);
1746 /* if this is a preferred peer use it */
1747 if (!preferred && ni_is_pref) {
1749 } else if (preferred && !ni_is_pref) {
1751 * this is not the preferred peer so let's ignore
1755 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1757 * We already have a peer that has more credits
1758 * available than this one. No need to consider
1759 * this peer further.
1762 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1764 * The best peer found so far and the current peer
1765 * have the same number of available credits let's
1766 * make sure to select between them using Round
1770 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1776 best_lpni_credits = lpni->lpni_txcredits;
1779 /* if we still can't find a peer ni then we can't reach it */
1781 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1782 LNET_NIDNET(dst_nid);
1783 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1784 libcfs_net2str(net_id));
1788 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1789 libcfs_nid2str(best_lpni->lpni_nid));
1795 * Prerequisite: the best_ni should already be set in the sd
1797 static inline struct lnet_peer_ni *
1798 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1801 struct lnet_peer_net *peer_net;
1804 * The gateway is Multi-Rail capable so now we must select the
1807 peer_net = lnet_peer_get_net_locked(peer, net_id);
1810 CERROR("gateway peer %s has no NI on net %s\n",
1811 libcfs_nid2str(peer->lp_primary_nid),
1812 libcfs_net2str(net_id));
1816 return lnet_select_peer_ni(sd, peer, peer_net);
1820 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1822 if (sd->sd_send_case & NMR_DST &&
1823 sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1824 sd->sd_msg->msg_type != LNET_MSG_ACK &&
1825 sd->sd_best_lpni->lpni_pref_nnids == 0) {
1826 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1827 libcfs_nid2str(sd->sd_best_ni->ni_nid),
1828 libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1829 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1830 sd->sd_best_ni->ni_nid);
1839 * use the source and destination NIDs as the pathway
1842 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1844 /* the destination lpni is set before we get here. */
1847 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1848 if (!sd->sd_best_ni) {
1849 CERROR("Can't send to %s: src %s is not a "
1850 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1851 libcfs_nid2str(sd->sd_src_nid));
1856 * the preferred NID will only be set for NMR peers
1858 lnet_set_non_mr_pref_nid(sd);
1860 return lnet_handle_send(sd);
1868 * Run the selection algorithm on the peer NIs unless we're sending
1869 * a response, in this case just send to the destination
1872 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1874 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1875 if (!sd->sd_best_ni) {
1876 CERROR("Can't send to %s: src %s is not a "
1877 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1878 libcfs_nid2str(sd->sd_src_nid));
1883 * only run the selection algorithm to pick the peer_ni if we're
1884 * sending a GET or a PUT. Responses are sent to the same
1885 * destination NID provided.
1887 if (!(sd->sd_send_case & SND_RESP)) {
1889 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1890 sd->sd_best_ni->ni_net->net_id);
1893 if (sd->sd_best_lpni)
1894 return lnet_handle_send(sd);
1896 CERROR("can't send to %s. no NI on %s\n",
1897 libcfs_nid2str(sd->sd_dst_nid),
1898 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1900 return -EHOSTUNREACH;
1904 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1905 struct lnet_peer *peer,
1906 struct lnet_peer_net *peer_net,
1910 struct lnet_net *local_net;
1911 struct lnet_ni *best_ni;
1913 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1918 * Iterate through the NIs in this local Net and select
1919 * the NI to send from. The selection is determined by
1920 * these 3 criterion in the following priority:
1922 * 2. NI available credits
1925 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1926 peer, peer_net, cpt);
1928 if (incr_seq && best_ni)
1935 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1937 struct lnet_peer_ni **gw_lpni,
1938 struct lnet_peer **gw_peer)
1940 struct lnet_peer_ni *gw;
1941 lnet_nid_t src_nid = sd->sd_src_nid;
1943 gw = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
1946 CERROR("no route to %s from %s\n",
1947 libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1948 return -EHOSTUNREACH;
1951 /* get the peer of the gw_ni */
1952 LASSERT(gw->lpni_peer_net);
1953 LASSERT(gw->lpni_peer_net->lpn_peer);
1955 *gw_peer = gw->lpni_peer_net->lpn_peer;
1957 if (!sd->sd_best_ni)
1958 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, *gw_peer,
1963 if (!sd->sd_best_ni) {
1964 CERROR("Internal Error. Expected local ni on %s "
1965 "but non found :%s\n",
1966 libcfs_net2str(gw->lpni_peer_net->lpn_net_id),
1967 libcfs_nid2str(sd->sd_src_nid));
1972 * if gw is MR let's find its best peer_ni
1974 if (lnet_peer_is_multi_rail(*gw_peer)) {
1975 gw = lnet_find_best_lpni_on_net(sd, *gw_peer,
1976 sd->sd_best_ni->ni_net->net_id);
1978 * We've already verified that the gw has an NI on that
1979 * desired net, but we're not finding it. Something is
1983 CERROR("Internal Error. Route expected to %s from %s\n",
1984 libcfs_nid2str(dst_nid),
1985 libcfs_nid2str(src_nid));
2000 * Remote destination
2001 * Non-MR destination
2005 * Remote destination
2008 * The handling of these two cases is similar. Even though the destination
2009 * can be MR or non-MR, we'll deal directly with the router.
2012 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2015 struct lnet_peer_ni *gw_lpni = NULL;
2016 struct lnet_peer *gw_peer = NULL;
2019 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2020 if (!sd->sd_best_ni) {
2021 CERROR("Can't send to %s: src %s is not a "
2022 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2023 libcfs_nid2str(sd->sd_src_nid));
2027 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2032 if (sd->sd_send_case & NMR_DST)
2034 * since the final destination is non-MR let's set its preferred
2035 * NID before we send
2037 lnet_set_non_mr_pref_nid(sd);
2040 * We're going to send to the gw found so let's set its
2043 sd->sd_peer = gw_peer;
2044 sd->sd_best_lpni = gw_lpni;
2046 return lnet_handle_send(sd);
2050 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
2052 struct lnet_peer_net *peer_net = NULL;
2053 struct lnet_ni *best_ni = NULL;
2056 * The peer can have multiple interfaces, some of them can be on
2057 * the local network and others on a routed network. We should
2058 * prefer the local network. However if the local network is not
2059 * available then we need to try the routed network
2062 /* go through all the peer nets and find the best_ni */
2063 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2065 * The peer's list of nets can contain non-local nets. We
2066 * want to only examine the local ones.
2068 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2070 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2071 peer_net, md_cpt, false);
2075 /* increment sequence number so we can round robin */
2081 static struct lnet_ni *
2082 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2084 struct lnet_ni *best_ni = NULL;
2085 struct lnet_peer_net *peer_net;
2086 struct lnet_peer *peer = sd->sd_peer;
2087 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2088 struct lnet_peer_ni *lpni;
2089 int cpt = sd->sd_cpt;
2092 * We must use a consistent source address when sending to a
2093 * non-MR peer. However, a non-MR peer can have multiple NIDs
2094 * on multiple networks, and we may even need to talk to this
2095 * peer on multiple networks -- certain types of
2096 * load-balancing configuration do this.
2098 * So we need to pick the NI the peer prefers for this
2099 * particular network.
2102 /* Get the target peer_ni */
2103 peer_net = lnet_peer_get_net_locked(peer,
2104 LNET_NIDNET(best_lpni->lpni_nid));
2105 LASSERT(peer_net != NULL);
2106 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2108 if (lpni->lpni_pref_nnids == 0)
2110 LASSERT(lpni->lpni_pref_nnids == 1);
2111 best_ni = lnet_nid2ni_locked(
2112 lpni->lpni_pref.nid, cpt);
2119 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2121 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2123 struct lnet_ni *best_ni = NULL;
2124 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2127 * We must use a consistent source address when sending to a
2128 * non-MR peer. However, a non-MR peer can have multiple NIDs
2129 * on multiple networks, and we may even need to talk to this
2130 * peer on multiple networks -- certain types of
2131 * load-balancing configuration do this.
2133 * So we need to pick the NI the peer prefers for this
2134 * particular network.
2137 best_ni = lnet_find_existing_preferred_best_ni(sd);
2139 /* if best_ni is still not set just pick one */
2142 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2143 sd->sd_best_lpni->lpni_peer_net,
2144 sd->sd_md_cpt, true);
2145 /* If there is no best_ni we don't have a route */
2147 CERROR("no path to %s from net %s\n",
2148 libcfs_nid2str(best_lpni->lpni_nid),
2149 libcfs_net2str(best_lpni->lpni_net->net_id));
2150 return -EHOSTUNREACH;
2154 sd->sd_best_ni = best_ni;
2156 /* Set preferred NI if necessary. */
2157 lnet_set_non_mr_pref_nid(sd);
2164 * Source not specified
2168 * always use the same source NID for NMR peers
2169 * If we've talked to that peer before then we already have a preferred
2170 * source NI associated with it. Otherwise, we select a preferred local NI
2171 * and store it in the peer
2174 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2178 /* sd->sd_best_lpni is already set to the final destination */
2181 * At this point we should've created the peer ni and peer. If we
2182 * can't find it, then something went wrong. Instead of assert
2183 * output a relevant message and fail the send
2185 if (!sd->sd_best_lpni) {
2186 CERROR("Internal fault. Unable to send msg %s to %s. "
2188 lnet_msgtyp2str(sd->sd_msg->msg_type),
2189 libcfs_nid2str(sd->sd_dst_nid));
2193 rc = lnet_select_preferred_best_ni(sd);
2195 rc = lnet_handle_send(sd);
2201 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2204 * NOTE we've already handled the remote peer case. So we only
2205 * need to worry about the local case here.
2207 * if we're sending a response, ACK or reply, we need to send it
2208 * to the destination NID given to us. At this point we already
2209 * have the peer_ni we're suppose to send to, so just find the
2210 * best_ni on the peer net and use that. Since we're sending to an
2211 * MR peer then we can just run the selection algorithm on our
2212 * local NIs and pick the best one.
2214 if (sd->sd_send_case & SND_RESP) {
2216 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2217 sd->sd_best_lpni->lpni_peer_net,
2218 sd->sd_md_cpt, true);
2220 if (!sd->sd_best_ni) {
2222 * We're not going to deal with not able to send
2223 * a response to the provided final destination
2225 CERROR("Can't send response to %s. "
2226 "No local NI available\n",
2227 libcfs_nid2str(sd->sd_dst_nid));
2228 return -EHOSTUNREACH;
2231 return lnet_handle_send(sd);
2235 * If we get here that means we're sending a fresh request, PUT or
2236 * GET, so we need to run our standard selection algorithm.
2237 * First find the best local interface that's on any of the peer's
2240 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2242 if (sd->sd_best_ni) {
2244 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2245 sd->sd_best_ni->ni_net->net_id);
2248 * if we're successful in selecting a peer_ni on the local
2249 * network, then send to it. Otherwise fall through and
2250 * try and see if we can reach it over another routed
2253 if (sd->sd_best_lpni) {
2255 * in case we initially started with a routed
2256 * destination, let's reset to local
2258 sd->sd_send_case &= ~REMOTE_DST;
2259 sd->sd_send_case |= LOCAL_DST;
2260 return lnet_handle_send(sd);
2263 CERROR("Internal Error. Expected to have a best_lpni: "
2265 libcfs_nid2str(sd->sd_src_nid),
2266 libcfs_nid2str(sd->sd_dst_nid));
2272 * Peer doesn't have a local network. Let's see if there is
2273 * a remote network we can reach it on.
2275 return PASS_THROUGH;
2280 * Source NID not specified
2285 * Source NID not speified
2286 * Remote destination
2289 * In both of these cases if we're sending a response, ACK or REPLY, then
2290 * we need to send to the destination NID provided.
2292 * In the remote case let's deal with MR routers.
2297 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2300 struct lnet_peer *gw_peer = NULL;
2301 struct lnet_peer_ni *gw_lpni = NULL;
2304 * handle sending a response to a remote peer here so we don't
2305 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2307 if (sd->sd_send_case & REMOTE_DST &&
2308 sd->sd_send_case & SND_RESP) {
2309 struct lnet_peer_ni *gw;
2310 struct lnet_peer *gw_peer;
2312 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2315 CERROR("Can't send response to %s. "
2316 "No route available\n",
2317 libcfs_nid2str(sd->sd_dst_nid));
2318 return -EHOSTUNREACH;
2321 sd->sd_best_lpni = gw;
2322 sd->sd_peer = gw_peer;
2324 return lnet_handle_send(sd);
2328 * Even though the NID for the peer might not be on a local network,
2329 * since the peer is MR there could be other interfaces on the
2330 * local network. In that case we'd still like to prefer the local
2331 * network over the routed network. If we're unable to do that
2332 * then we select the best router among the different routed networks,
2333 * and if the router is MR then we can deal with it as such.
2335 rc = lnet_handle_any_mr_dsta(sd);
2336 if (rc != PASS_THROUGH)
2340 * TODO; One possible enhancement is to run the selection
2341 * algorithm on the peer. However for remote peers the credits are
2342 * not decremented, so we'll be basically going over the peer NIs
2343 * in round robin. An MR router will run the selection algorithm
2344 * on the next-hop interfaces.
2346 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2351 sd->sd_send_case &= ~LOCAL_DST;
2352 sd->sd_send_case |= REMOTE_DST;
2354 sd->sd_peer = gw_peer;
2355 sd->sd_best_lpni = gw_lpni;
2357 return lnet_handle_send(sd);
2361 * Source not specified
2362 * Remote destination
2365 * Must send to the specified peer NID using the same source NID that
2366 * we've used before. If it's the first time to talk to that peer then
2367 * find the source NI and assign it as preferred to that peer
2370 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2373 struct lnet_peer_ni *gw_lpni = NULL;
2374 struct lnet_peer *gw_peer = NULL;
2377 * Let's set if we have a preferred NI to talk to this NMR peer
2379 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2382 * find the router and that'll find the best NI if we didn't find
2385 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2391 * set the best_ni we've chosen as the preferred one for
2394 lnet_set_non_mr_pref_nid(sd);
2396 /* we'll be sending to the gw */
2397 sd->sd_best_lpni = gw_lpni;
2398 sd->sd_peer = gw_peer;
2400 return lnet_handle_send(sd);
2404 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2407 * turn off the SND_RESP bit.
2408 * It will be checked in the case handling
2410 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2412 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2413 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2414 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2415 (send_case & MR_DST) ? "MR: " : "NMR: ",
2416 libcfs_nid2str(sd->sd_dst_nid),
2417 (send_case & LOCAL_DST) ? "local" : "routed");
2419 switch (send_case) {
2421 * For all cases where the source is specified, we should always
2422 * use the destination NID, whether it's an MR destination or not,
2423 * since we're continuing a series of related messages for the
2426 case SRC_SPEC_LOCAL_NMR_DST:
2427 return lnet_handle_spec_local_nmr_dst(sd);
2428 case SRC_SPEC_LOCAL_MR_DST:
2429 return lnet_handle_spec_local_mr_dst(sd);
2430 case SRC_SPEC_ROUTER_NMR_DST:
2431 case SRC_SPEC_ROUTER_MR_DST:
2432 return lnet_handle_spec_router_dst(sd);
2433 case SRC_ANY_LOCAL_NMR_DST:
2434 return lnet_handle_any_local_nmr_dst(sd);
2435 case SRC_ANY_LOCAL_MR_DST:
2436 case SRC_ANY_ROUTER_MR_DST:
2437 return lnet_handle_any_mr_dst(sd);
2438 case SRC_ANY_ROUTER_NMR_DST:
2439 return lnet_handle_any_router_nmr_dst(sd);
2441 CERROR("Unknown send case\n");
2447 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2448 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2450 struct lnet_peer_ni *lpni;
2451 struct lnet_peer *peer;
2452 struct lnet_send_data send_data;
2455 __u32 send_case = 0;
2457 memset(&send_data, 0, sizeof(send_data));
2460 * get an initial CPT to use for locking. The idea here is not to
2461 * serialize the calls to select_pathway, so that as many
2462 * operations can run concurrently as possible. To do that we use
2463 * the CPT where this call is being executed. Later on when we
2464 * determine the CPT to use in lnet_message_commit, we switch the
2465 * lock and check if there was any configuration change. If none,
2466 * then we proceed, if there is, then we restart the operation.
2468 cpt = lnet_net_lock_current();
2470 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2471 if (md_cpt == CFS_CPT_ANY)
2477 * If we're being asked to send to the loopback interface, there
2478 * is no need to go through any selection. We can just shortcut
2479 * the entire process and send over lolnd
2481 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2482 /* No send credit hassles with LOLND */
2483 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
2484 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
2485 if (!msg->msg_routing)
2486 msg->msg_hdr.src_nid =
2487 cpu_to_le64(the_lnet.ln_loni->ni_nid);
2488 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
2489 lnet_msg_commit(msg, cpt);
2490 msg->msg_txni = the_lnet.ln_loni;
2491 lnet_net_unlock(cpt);
2493 return LNET_CREDIT_OK;
2497 * find an existing peer_ni, or create one and mark it as having been
2498 * created due to network traffic. This call will create the
2499 * peer->peer_net->peer_ni tree.
2501 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2503 lnet_net_unlock(cpt);
2504 return PTR_ERR(lpni);
2508 * Now that we have a peer_ni, check if we want to discover
2509 * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2510 * trigger discovery.
2512 peer = lpni->lpni_peer_net->lpn_peer;
2513 if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
2514 lnet_nid_t primary_nid;
2515 rc = lnet_discover_peer_locked(lpni, cpt, false);
2517 lnet_peer_ni_decref_locked(lpni);
2518 lnet_net_unlock(cpt);
2521 /* The peer may have changed. */
2522 peer = lpni->lpni_peer_net->lpn_peer;
2523 /* queue message and return */
2524 msg->msg_src_nid_param = src_nid;
2525 msg->msg_rtr_nid_param = rtr_nid;
2526 msg->msg_sending = 0;
2527 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2528 lnet_peer_ni_decref_locked(lpni);
2529 primary_nid = peer->lp_primary_nid;
2530 lnet_net_unlock(cpt);
2532 CDEBUG(D_NET, "%s pending discovery\n",
2533 libcfs_nid2str(primary_nid));
2535 return LNET_DC_WAIT;
2537 lnet_peer_ni_decref_locked(lpni);
2539 /* If peer is not healthy then can not send anything to it */
2540 if (!lnet_is_peer_healthy_locked(peer)) {
2541 lnet_net_unlock(cpt);
2542 return -EHOSTUNREACH;
2546 * Identify the different send cases
2548 if (src_nid == LNET_NID_ANY)
2549 send_case |= SRC_ANY;
2551 send_case |= SRC_SPEC;
2553 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2554 send_case |= LOCAL_DST;
2556 send_case |= REMOTE_DST;
2558 if (!lnet_peer_is_multi_rail(peer))
2559 send_case |= NMR_DST;
2561 send_case |= MR_DST;
2563 if (msg->msg_type == LNET_MSG_REPLY ||
2564 msg->msg_type == LNET_MSG_ACK)
2565 send_case |= SND_RESP;
2567 /* assign parameters to the send_data */
2568 send_data.sd_msg = msg;
2569 send_data.sd_rtr_nid = rtr_nid;
2570 send_data.sd_src_nid = src_nid;
2571 send_data.sd_dst_nid = dst_nid;
2572 send_data.sd_best_lpni = lpni;
2574 * keep a pointer to the final destination in case we're going to
2575 * route, so we'll need to access it later
2577 send_data.sd_final_dst_lpni = lpni;
2578 send_data.sd_peer = peer;
2579 send_data.sd_md_cpt = md_cpt;
2580 send_data.sd_cpt = cpt;
2581 send_data.sd_send_case = send_case;
2583 rc = lnet_handle_send_case_locked(&send_data);
2585 if (rc == REPEAT_SEND)
2588 lnet_net_unlock(send_data.sd_cpt);
2594 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2596 lnet_nid_t dst_nid = msg->msg_target.nid;
2600 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2601 * but we might want to use pre-determined router for ACK/REPLY
2604 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2605 LASSERT (msg->msg_txpeer == NULL);
2606 LASSERT (!msg->msg_sending);
2607 LASSERT (!msg->msg_target_is_router);
2608 LASSERT (!msg->msg_receiving);
2610 msg->msg_sending = 1;
2612 LASSERT(!msg->msg_tx_committed);
2614 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2618 if (rc == LNET_CREDIT_OK)
2619 lnet_ni_send(msg->msg_txni, msg);
2621 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2626 lnet_monitor_thread(void *arg)
2629 * The monitor thread takes care of the following:
2630 * 1. Checks the aliveness of routers
2631 * 2. Checks if there are messages on the resend queue to resend
2633 * 3. Check if there are any NIs on the local recovery queue and
2635 * 4. Checks if there are any NIs on the remote recovery queue
2638 cfs_block_allsigs();
2640 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
2641 if (lnet_router_checker_active())
2642 lnet_check_routers();
2645 * TODO do we need to check if we should sleep without
2646 * timeout? Technically, an active system will always
2647 * have messages in flight so this check will always
2648 * evaluate to false. And on an idle system do we care
2649 * if we wake up every 1 second? Although, we've seen
2650 * cases where we get a complaint that an idle thread
2651 * is waking up unnecessarily.
2653 wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
2655 cfs_time_seconds(1));
2658 /* clean up the router checker */
2659 lnet_prune_rc_data(1);
2662 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
2664 /* signal that the monitor thread is exiting */
2665 up(&the_lnet.ln_mt_signal);
2670 int lnet_monitor_thr_start(void)
2673 struct task_struct *task;
2675 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
2677 sema_init(&the_lnet.ln_mt_signal, 0);
2679 /* Pre monitor thread start processing */
2680 rc = lnet_router_pre_mt_start();
2684 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
2685 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
2688 CERROR("Can't start monitor thread: %d\n", rc);
2689 /* block until event callback signals exit */
2690 down(&the_lnet.ln_mt_signal);
2693 lnet_router_cleanup();
2694 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
2698 /* post monitor thread start processing */
2699 lnet_router_post_mt_start();
2704 void lnet_monitor_thr_stop(void)
2706 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2709 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
2710 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
2712 /* tell the monitor thread that we're shutting down */
2713 wake_up(&the_lnet.ln_mt_waitq);
2715 /* block until monitor thread signals that it's done */
2716 down(&the_lnet.ln_mt_signal);
2717 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
2719 lnet_router_cleanup();
2724 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
2728 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
2729 the_lnet.ln_counters[cpt]->drop_count++;
2730 the_lnet.ln_counters[cpt]->drop_length += nob;
2731 lnet_net_unlock(cpt);
2733 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
2737 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
2739 struct lnet_hdr *hdr = &msg->msg_hdr;
2741 if (msg->msg_wanted != 0)
2742 lnet_setpayloadbuffer(msg);
2744 lnet_build_msg_event(msg, LNET_EVENT_PUT);
2746 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
2747 * it back into the ACK during lnet_finalize() */
2748 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
2749 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
2751 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
2752 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
2756 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
2758 struct lnet_hdr *hdr = &msg->msg_hdr;
2759 struct lnet_match_info info;
2763 /* Convert put fields to host byte order */
2764 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
2765 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
2766 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
2768 /* Primary peer NID. */
2769 info.mi_id.nid = msg->msg_initiator;
2770 info.mi_id.pid = hdr->src_pid;
2771 info.mi_opc = LNET_MD_OP_PUT;
2772 info.mi_portal = hdr->msg.put.ptl_index;
2773 info.mi_rlength = hdr->payload_length;
2774 info.mi_roffset = hdr->msg.put.offset;
2775 info.mi_mbits = hdr->msg.put.match_bits;
2776 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
2778 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
2779 ready_delay = msg->msg_rx_ready_delay;
2782 rc = lnet_ptl_match_md(&info, msg);
2787 case LNET_MATCHMD_OK:
2788 lnet_recv_put(ni, msg);
2791 case LNET_MATCHMD_NONE:
2793 /* no eager_recv or has already called it, should
2794 * have been attached on delayed list */
2797 rc = lnet_ni_eager_recv(ni, msg);
2804 case LNET_MATCHMD_DROP:
2805 CNETERR("Dropping PUT from %s portal %d match %llu"
2806 " offset %d length %d: %d\n",
2807 libcfs_id2str(info.mi_id), info.mi_portal,
2808 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
2810 return -ENOENT; /* -ve: OK but no match */
2815 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
2817 struct lnet_match_info info;
2818 struct lnet_hdr *hdr = &msg->msg_hdr;
2819 struct lnet_process_id source_id;
2820 struct lnet_handle_wire reply_wmd;
2823 /* Convert get fields to host byte order */
2824 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
2825 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
2826 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
2827 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
2829 source_id.nid = hdr->src_nid;
2830 source_id.pid = hdr->src_pid;
2831 /* Primary peer NID */
2832 info.mi_id.nid = msg->msg_initiator;
2833 info.mi_id.pid = hdr->src_pid;
2834 info.mi_opc = LNET_MD_OP_GET;
2835 info.mi_portal = hdr->msg.get.ptl_index;
2836 info.mi_rlength = hdr->msg.get.sink_length;
2837 info.mi_roffset = hdr->msg.get.src_offset;
2838 info.mi_mbits = hdr->msg.get.match_bits;
2839 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
2841 rc = lnet_ptl_match_md(&info, msg);
2842 if (rc == LNET_MATCHMD_DROP) {
2843 CNETERR("Dropping GET from %s portal %d match %llu"
2844 " offset %d length %d\n",
2845 libcfs_id2str(info.mi_id), info.mi_portal,
2846 info.mi_mbits, info.mi_roffset, info.mi_rlength);
2847 return -ENOENT; /* -ve: OK but no match */
2850 LASSERT(rc == LNET_MATCHMD_OK);
2852 lnet_build_msg_event(msg, LNET_EVENT_GET);
2854 reply_wmd = hdr->msg.get.return_wmd;
2856 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
2857 msg->msg_offset, msg->msg_wanted);
2859 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
2862 /* The LND completes the REPLY from her recv procedure */
2863 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2864 msg->msg_offset, msg->msg_len, msg->msg_len);
2868 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
2869 msg->msg_receiving = 0;
2871 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
2873 /* didn't get as far as lnet_ni_send() */
2874 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
2875 libcfs_nid2str(ni->ni_nid),
2876 libcfs_id2str(info.mi_id), rc);
2878 lnet_finalize(msg, rc);
2885 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
2887 void *private = msg->msg_private;
2888 struct lnet_hdr *hdr = &msg->msg_hdr;
2889 struct lnet_process_id src = {0};
2890 struct lnet_libmd *md;
2895 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
2898 src.nid = hdr->src_nid;
2899 src.pid = hdr->src_pid;
2901 /* NB handles only looked up by creator (no flips) */
2902 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
2903 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2904 CNETERR("%s: Dropping REPLY from %s for %s "
2906 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2907 (md == NULL) ? "invalid" : "inactive",
2908 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2909 hdr->msg.reply.dst_wmd.wh_object_cookie);
2910 if (md != NULL && md->md_me != NULL)
2911 CERROR("REPLY MD also attached to portal %d\n",
2912 md->md_me->me_portal);
2914 lnet_res_unlock(cpt);
2915 return -ENOENT; /* -ve: OK but no match */
2918 LASSERT(md->md_offset == 0);
2920 rlength = hdr->payload_length;
2921 mlength = MIN(rlength, (int)md->md_length);
2923 if (mlength < rlength &&
2924 (md->md_options & LNET_MD_TRUNCATE) == 0) {
2925 CNETERR("%s: Dropping REPLY from %s length %d "
2926 "for MD %#llx would overflow (%d)\n",
2927 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2928 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2930 lnet_res_unlock(cpt);
2931 return -ENOENT; /* -ve: OK but no match */
2934 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
2935 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2936 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2938 lnet_msg_attach_md(msg, md, 0, mlength);
2941 lnet_setpayloadbuffer(msg);
2943 lnet_res_unlock(cpt);
2945 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2947 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2952 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
2954 struct lnet_hdr *hdr = &msg->msg_hdr;
2955 struct lnet_process_id src = {0};
2956 struct lnet_libmd *md;
2959 src.nid = hdr->src_nid;
2960 src.pid = hdr->src_pid;
2962 /* Convert ack fields to host byte order */
2963 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2964 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2966 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
2969 /* NB handles only looked up by creator (no flips) */
2970 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2971 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2972 /* Don't moan; this is expected */
2974 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
2975 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2976 (md == NULL) ? "invalid" : "inactive",
2977 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2978 hdr->msg.ack.dst_wmd.wh_object_cookie);
2979 if (md != NULL && md->md_me != NULL)
2980 CERROR("Source MD also attached to portal %d\n",
2981 md->md_me->me_portal);
2983 lnet_res_unlock(cpt);
2984 return -ENOENT; /* -ve! */
2987 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
2988 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2989 hdr->msg.ack.dst_wmd.wh_object_cookie);
2991 lnet_msg_attach_md(msg, md, 0, 0);
2993 lnet_res_unlock(cpt);
2995 lnet_build_msg_event(msg, LNET_EVENT_ACK);
2997 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
3002 * \retval LNET_CREDIT_OK If \a msg is forwarded
3003 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
3004 * \retval -ve error code
3007 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
3011 if (!the_lnet.ln_routing)
3014 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
3015 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
3016 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
3017 msg->msg_rx_ready_delay = 1;
3019 lnet_net_unlock(msg->msg_rx_cpt);
3020 rc = lnet_ni_eager_recv(ni, msg);
3021 lnet_net_lock(msg->msg_rx_cpt);
3026 rc = lnet_post_routed_recv_locked(msg, 0);
3031 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
3035 switch (msg->msg_type) {
3037 rc = lnet_parse_ack(ni, msg);
3040 rc = lnet_parse_put(ni, msg);
3043 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
3045 case LNET_MSG_REPLY:
3046 rc = lnet_parse_reply(ni, msg);
3048 default: /* prevent an unused label if !kernel */
3053 LASSERT(rc == 0 || rc == -ENOENT);
3058 lnet_msgtyp2str (int type)
3067 case LNET_MSG_REPLY:
3069 case LNET_MSG_HELLO:
3072 return ("<UNKNOWN>");
3077 lnet_print_hdr(struct lnet_hdr *hdr)
3079 struct lnet_process_id src = {
3080 .nid = hdr->src_nid,
3081 .pid = hdr->src_pid,
3083 struct lnet_process_id dst = {
3084 .nid = hdr->dest_nid,
3085 .pid = hdr->dest_pid,
3087 char *type_str = lnet_msgtyp2str(hdr->type);
3089 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
3090 CWARN(" From %s\n", libcfs_id2str(src));
3091 CWARN(" To %s\n", libcfs_id2str(dst));
3093 switch (hdr->type) {
3098 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
3099 "match bits %llu\n",
3100 hdr->msg.put.ptl_index,
3101 hdr->msg.put.ack_wmd.wh_interface_cookie,
3102 hdr->msg.put.ack_wmd.wh_object_cookie,
3103 hdr->msg.put.match_bits);
3104 CWARN(" Length %d, offset %d, hdr data %#llx\n",
3105 hdr->payload_length, hdr->msg.put.offset,
3106 hdr->msg.put.hdr_data);
3110 CWARN(" Ptl index %d, return md %#llx.%#llx, "
3111 "match bits %llu\n", hdr->msg.get.ptl_index,
3112 hdr->msg.get.return_wmd.wh_interface_cookie,
3113 hdr->msg.get.return_wmd.wh_object_cookie,
3114 hdr->msg.get.match_bits);
3115 CWARN(" Length %d, src offset %d\n",
3116 hdr->msg.get.sink_length,
3117 hdr->msg.get.src_offset);
3121 CWARN(" dst md %#llx.%#llx, "
3122 "manipulated length %d\n",
3123 hdr->msg.ack.dst_wmd.wh_interface_cookie,
3124 hdr->msg.ack.dst_wmd.wh_object_cookie,
3125 hdr->msg.ack.mlength);
3128 case LNET_MSG_REPLY:
3129 CWARN(" dst md %#llx.%#llx, "
3131 hdr->msg.reply.dst_wmd.wh_interface_cookie,
3132 hdr->msg.reply.dst_wmd.wh_object_cookie,
3133 hdr->payload_length);
3139 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
3140 void *private, int rdma_req)
3145 struct lnet_msg *msg;
3146 lnet_pid_t dest_pid;
3147 lnet_nid_t dest_nid;
3149 struct lnet_peer_ni *lpni;
3150 __u32 payload_length;
3153 LASSERT (!in_interrupt ());
3155 type = le32_to_cpu(hdr->type);
3156 src_nid = le64_to_cpu(hdr->src_nid);
3157 dest_nid = le64_to_cpu(hdr->dest_nid);
3158 dest_pid = le32_to_cpu(hdr->dest_pid);
3159 payload_length = le32_to_cpu(hdr->payload_length);
3161 for_me = (ni->ni_nid == dest_nid);
3162 cpt = lnet_cpt_of_nid(from_nid, ni);
3164 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
3165 libcfs_nid2str(dest_nid),
3166 libcfs_nid2str(ni->ni_nid),
3167 libcfs_nid2str(src_nid),
3168 lnet_msgtyp2str(type),
3169 (for_me) ? "for me" : "routed");
3174 if (payload_length > 0) {
3175 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
3176 libcfs_nid2str(from_nid),
3177 libcfs_nid2str(src_nid),
3178 lnet_msgtyp2str(type), payload_length);
3184 case LNET_MSG_REPLY:
3185 if (payload_length >
3186 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
3187 CERROR("%s, src %s: bad %s payload %d "
3188 "(%d max expected)\n",
3189 libcfs_nid2str(from_nid),
3190 libcfs_nid2str(src_nid),
3191 lnet_msgtyp2str(type),
3193 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
3199 CERROR("%s, src %s: Bad message type 0x%x\n",
3200 libcfs_nid2str(from_nid),
3201 libcfs_nid2str(src_nid), type);
3205 if (the_lnet.ln_routing &&
3206 ni->ni_last_alive != ktime_get_real_seconds()) {
3207 /* NB: so far here is the only place to set NI status to "up */
3209 ni->ni_last_alive = ktime_get_real_seconds();
3210 if (ni->ni_status != NULL &&
3211 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
3212 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
3216 /* Regard a bad destination NID as a protocol error. Senders should
3217 * know what they're doing; if they don't they're misconfigured, buggy
3218 * or malicious so we chop them off at the knees :) */
3221 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
3222 /* should have gone direct */
3223 CERROR("%s, src %s: Bad dest nid %s "
3224 "(should have been sent direct)\n",
3225 libcfs_nid2str(from_nid),
3226 libcfs_nid2str(src_nid),
3227 libcfs_nid2str(dest_nid));
3231 if (lnet_islocalnid(dest_nid)) {
3232 /* dest is another local NI; sender should have used
3233 * this node's NID on its own network */
3234 CERROR("%s, src %s: Bad dest nid %s "
3235 "(it's my nid but on a different network)\n",
3236 libcfs_nid2str(from_nid),
3237 libcfs_nid2str(src_nid),
3238 libcfs_nid2str(dest_nid));
3242 if (rdma_req && type == LNET_MSG_GET) {
3243 CERROR("%s, src %s: Bad optimized GET for %s "
3244 "(final destination must be me)\n",
3245 libcfs_nid2str(from_nid),
3246 libcfs_nid2str(src_nid),
3247 libcfs_nid2str(dest_nid));
3251 if (!the_lnet.ln_routing) {
3252 CERROR("%s, src %s: Dropping message for %s "
3253 "(routing not enabled)\n",
3254 libcfs_nid2str(from_nid),
3255 libcfs_nid2str(src_nid),
3256 libcfs_nid2str(dest_nid));
3261 /* Message looks OK; we're not going to return an error, so we MUST
3262 * call back lnd_recv() come what may... */
3264 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
3265 fail_peer(src_nid, 0)) { /* shall we now? */
3266 CERROR("%s, src %s: Dropping %s to simulate failure\n",
3267 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
3268 lnet_msgtyp2str(type));
3272 if (!list_empty(&the_lnet.ln_drop_rules) &&
3273 lnet_drop_rule_match(hdr)) {
3274 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
3275 "silent message loss\n",
3276 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
3277 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
3282 msg = lnet_msg_alloc();
3284 CERROR("%s, src %s: Dropping %s (out of memory)\n",
3285 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
3286 lnet_msgtyp2str(type));
3290 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
3291 * pointers NULL etc */
3293 msg->msg_type = type;
3294 msg->msg_private = private;
3295 msg->msg_receiving = 1;
3296 msg->msg_rdma_get = rdma_req;
3297 msg->msg_len = msg->msg_wanted = payload_length;
3298 msg->msg_offset = 0;
3299 msg->msg_hdr = *hdr;
3300 /* for building message event */
3301 msg->msg_from = from_nid;
3303 msg->msg_target.pid = dest_pid;
3304 msg->msg_target.nid = dest_nid;
3305 msg->msg_routing = 1;
3308 /* convert common msg->hdr fields to host byteorder */
3309 msg->msg_hdr.type = type;
3310 msg->msg_hdr.src_nid = src_nid;
3311 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
3312 msg->msg_hdr.dest_nid = dest_nid;
3313 msg->msg_hdr.dest_pid = dest_pid;
3314 msg->msg_hdr.payload_length = payload_length;
3318 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
3320 lnet_net_unlock(cpt);
3321 CERROR("%s, src %s: Dropping %s "
3322 "(error %ld looking up sender)\n",
3323 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
3324 lnet_msgtyp2str(type), PTR_ERR(lpni));
3326 if (rc == -ESHUTDOWN)
3327 /* We are shutting down. Don't do anything more */
3331 msg->msg_rxpeer = lpni;
3333 lnet_ni_addref_locked(ni, cpt);
3334 /* Multi-Rail: Primary NID of source. */
3335 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
3337 if (lnet_isrouter(msg->msg_rxpeer)) {
3338 lnet_peer_set_alive(msg->msg_rxpeer);
3339 if (avoid_asym_router_failure &&
3340 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
3341 /* received a remote message from router, update
3342 * remote NI status on this router.
3343 * NB: multi-hop routed message will be ignored.
3345 lnet_router_ni_update_locked(msg->msg_rxpeer,
3346 LNET_NIDNET(src_nid));
3350 lnet_msg_commit(msg, cpt);
3352 /* message delay simulation */
3353 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
3354 lnet_delay_rule_match_locked(hdr, msg))) {
3355 lnet_net_unlock(cpt);
3360 rc = lnet_parse_forward_locked(ni, msg);
3361 lnet_net_unlock(cpt);
3366 if (rc == LNET_CREDIT_OK) {
3367 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3368 0, payload_length, payload_length);
3373 lnet_net_unlock(cpt);
3375 rc = lnet_parse_local(ni, msg);
3381 LASSERT(msg->msg_md == NULL);
3382 lnet_finalize(msg, rc);
3385 lnet_drop_message(ni, cpt, private, payload_length, type);
3388 EXPORT_SYMBOL(lnet_parse);
3391 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
3393 while (!list_empty(head)) {
3394 struct lnet_process_id id = {0};
3395 struct lnet_msg *msg;
3397 msg = list_entry(head->next, struct lnet_msg, msg_list);
3398 list_del(&msg->msg_list);
3400 id.nid = msg->msg_hdr.src_nid;
3401 id.pid = msg->msg_hdr.src_pid;
3403 LASSERT(msg->msg_md == NULL);
3404 LASSERT(msg->msg_rx_delayed);
3405 LASSERT(msg->msg_rxpeer != NULL);
3406 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
3408 CWARN("Dropping delayed PUT from %s portal %d match %llu"
3409 " offset %d length %d: %s\n",
3411 msg->msg_hdr.msg.put.ptl_index,
3412 msg->msg_hdr.msg.put.match_bits,
3413 msg->msg_hdr.msg.put.offset,
3414 msg->msg_hdr.payload_length, reason);
3416 /* NB I can't drop msg's ref on msg_rxpeer until after I've
3417 * called lnet_drop_message(), so I just hang onto msg as well
3418 * until that's done */
3420 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
3421 msg->msg_private, msg->msg_len,
3424 * NB: message will not generate event because w/o attached MD,
3425 * but we still should give error code so lnet_msg_decommit()
3426 * can skip counters operations and other checks.
3428 lnet_finalize(msg, -ENOENT);
3433 lnet_recv_delayed_msg_list(struct list_head *head)
3435 while (!list_empty(head)) {
3436 struct lnet_msg *msg;
3437 struct lnet_process_id id;
3439 msg = list_entry(head->next, struct lnet_msg, msg_list);
3440 list_del(&msg->msg_list);
3442 /* md won't disappear under me, since each msg
3443 * holds a ref on it */
3445 id.nid = msg->msg_hdr.src_nid;
3446 id.pid = msg->msg_hdr.src_pid;
3448 LASSERT(msg->msg_rx_delayed);
3449 LASSERT(msg->msg_md != NULL);
3450 LASSERT(msg->msg_rxpeer != NULL);
3451 LASSERT(msg->msg_rxni != NULL);
3452 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
3454 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
3455 "match %llu offset %d length %d.\n",
3456 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
3457 msg->msg_hdr.msg.put.match_bits,
3458 msg->msg_hdr.msg.put.offset,
3459 msg->msg_hdr.payload_length);
3461 lnet_recv_put(msg->msg_rxni, msg);
3466 * Initiate an asynchronous PUT operation.
3468 * There are several events associated with a PUT: completion of the send on
3469 * the initiator node (LNET_EVENT_SEND), and when the send completes
3470 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
3471 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
3472 * used at the target node to indicate the completion of incoming data
3475 * The local events will be logged in the EQ associated with the MD pointed to
3476 * by \a mdh handle. Using a MD without an associated EQ results in these
3477 * events being discarded. In this case, the caller must have another
3478 * mechanism (e.g., a higher level protocol) for determining when it is safe
3479 * to modify the memory region associated with the MD.
3481 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
3482 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
3484 * \param self Indicates the NID of a local interface through which to send
3485 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
3486 * \param mdh A handle for the MD that describes the memory to be sent. The MD
3487 * must be "free floating" (See LNetMDBind()).
3488 * \param ack Controls whether an acknowledgment is requested.
3489 * Acknowledgments are only sent when they are requested by the initiating
3490 * process and the target MD enables them.
3491 * \param target A process identifier for the target process.
3492 * \param portal The index in the \a target's portal table.
3493 * \param match_bits The match bits to use for MD selection at the target
3495 * \param offset The offset into the target MD (only used when the target
3496 * MD has the LNET_MD_MANAGE_REMOTE option set).
3497 * \param hdr_data 64 bits of user data that can be included in the message
3498 * header. This data is written to an event queue entry at the target if an
3499 * EQ is present on the matching MD.
3501 * \retval 0 Success, and only in this case events will be generated
3502 * and logged to EQ (if it exists).
3503 * \retval -EIO Simulated failure.
3504 * \retval -ENOMEM Memory allocation failure.
3505 * \retval -ENOENT Invalid MD object.
3507 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
3510 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
3511 struct lnet_process_id target, unsigned int portal,
3512 __u64 match_bits, unsigned int offset,
3515 struct lnet_msg *msg;
3516 struct lnet_libmd *md;
3520 LASSERT(the_lnet.ln_refcount > 0);
3522 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
3523 fail_peer(target.nid, 1)) { /* shall we now? */
3524 CERROR("Dropping PUT to %s: simulated failure\n",
3525 libcfs_id2str(target));
3529 msg = lnet_msg_alloc();
3531 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
3532 libcfs_id2str(target));
3535 msg->msg_vmflush = !!memory_pressure_get();
3537 cpt = lnet_cpt_of_cookie(mdh.cookie);
3540 md = lnet_handle2md(&mdh);
3541 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3542 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
3543 match_bits, portal, libcfs_id2str(target),
3544 md == NULL ? -1 : md->md_threshold);
3545 if (md != NULL && md->md_me != NULL)
3546 CERROR("Source MD also attached to portal %d\n",
3547 md->md_me->me_portal);
3548 lnet_res_unlock(cpt);
3554 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
3556 lnet_msg_attach_md(msg, md, 0, 0);
3558 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
3560 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
3561 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
3562 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
3563 msg->msg_hdr.msg.put.hdr_data = hdr_data;
3565 /* NB handles only looked up by creator (no flips) */
3566 if (ack == LNET_ACK_REQ) {
3567 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
3568 the_lnet.ln_interface_cookie;
3569 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
3570 md->md_lh.lh_cookie;
3572 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
3573 LNET_WIRE_HANDLE_COOKIE_NONE;
3574 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
3575 LNET_WIRE_HANDLE_COOKIE_NONE;
3578 lnet_res_unlock(cpt);
3580 lnet_build_msg_event(msg, LNET_EVENT_SEND);
3582 rc = lnet_send(self, msg, LNET_NID_ANY);
3584 CNETERR("Error sending PUT to %s: %d\n",
3585 libcfs_id2str(target), rc);
3586 lnet_finalize(msg, rc);
3589 /* completion will be signalled by an event */
3592 EXPORT_SYMBOL(LNetPut);
3595 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
3596 * returns a msg for the LND to pass to lnet_finalize() when the sink
3597 * data has been received.
3599 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
3600 * lnet_finalize() is called on it, so the LND must call this first
3603 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
3605 struct lnet_msg *msg = lnet_msg_alloc();
3606 struct lnet_libmd *getmd = getmsg->msg_md;
3607 struct lnet_process_id peer_id = getmsg->msg_target;
3610 LASSERT(!getmsg->msg_target_is_router);
3611 LASSERT(!getmsg->msg_routing);
3614 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
3615 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
3619 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
3622 LASSERT(getmd->md_refcount > 0);
3624 if (getmd->md_threshold == 0) {
3625 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
3626 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
3628 lnet_res_unlock(cpt);
3632 LASSERT(getmd->md_offset == 0);
3634 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
3635 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
3637 /* setup information for lnet_build_msg_event */
3638 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
3639 msg->msg_from = peer_id.nid;
3640 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
3641 msg->msg_hdr.src_nid = peer_id.nid;
3642 msg->msg_hdr.payload_length = getmd->md_length;
3643 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
3645 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
3646 lnet_res_unlock(cpt);
3648 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
3651 lnet_msg_commit(msg, cpt);
3652 lnet_net_unlock(cpt);
3654 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3659 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
3662 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
3663 the_lnet.ln_counters[cpt]->drop_count++;
3664 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
3665 lnet_net_unlock(cpt);
3672 EXPORT_SYMBOL(lnet_create_reply_msg);
3675 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
3678 /* Set the REPLY length, now the RDMA that elides the REPLY message has
3679 * completed and I know it. */
3680 LASSERT(reply != NULL);
3681 LASSERT(reply->msg_type == LNET_MSG_GET);
3682 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
3684 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
3685 * the end of my buffer, I might as well be dead. */
3686 LASSERT(len <= reply->msg_ev.mlength);
3688 reply->msg_ev.mlength = len;
3690 EXPORT_SYMBOL(lnet_set_reply_msg_len);
3693 * Initiate an asynchronous GET operation.
3695 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
3696 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
3697 * the target node in the REPLY has been written to local MD.
3699 * On the target node, an LNET_EVENT_GET is logged when the GET request
3700 * arrives and is accepted into a MD.
3702 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
3703 * \param mdh A handle for the MD that describes the memory into which the
3704 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
3706 * \retval 0 Success, and only in this case events will be generated
3707 * and logged to EQ (if it exists) of the MD.
3708 * \retval -EIO Simulated failure.
3709 * \retval -ENOMEM Memory allocation failure.
3710 * \retval -ENOENT Invalid MD object.
3713 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
3714 struct lnet_process_id target, unsigned int portal,
3715 __u64 match_bits, unsigned int offset)
3717 struct lnet_msg *msg;
3718 struct lnet_libmd *md;
3722 LASSERT(the_lnet.ln_refcount > 0);
3724 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
3725 fail_peer(target.nid, 1)) /* shall we now? */
3727 CERROR("Dropping GET to %s: simulated failure\n",
3728 libcfs_id2str(target));
3732 msg = lnet_msg_alloc();
3734 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
3735 libcfs_id2str(target));
3739 cpt = lnet_cpt_of_cookie(mdh.cookie);
3742 md = lnet_handle2md(&mdh);
3743 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3744 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
3745 match_bits, portal, libcfs_id2str(target),
3746 md == NULL ? -1 : md->md_threshold);
3747 if (md != NULL && md->md_me != NULL)
3748 CERROR("REPLY MD also attached to portal %d\n",
3749 md->md_me->me_portal);
3751 lnet_res_unlock(cpt);
3757 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
3759 lnet_msg_attach_md(msg, md, 0, 0);
3761 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
3763 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
3764 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
3765 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
3766 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
3768 /* NB handles only looked up by creator (no flips) */
3769 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
3770 the_lnet.ln_interface_cookie;
3771 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
3772 md->md_lh.lh_cookie;
3774 lnet_res_unlock(cpt);
3776 lnet_build_msg_event(msg, LNET_EVENT_SEND);
3778 rc = lnet_send(self, msg, LNET_NID_ANY);
3780 CNETERR("Error sending GET to %s: %d\n",
3781 libcfs_id2str(target), rc);
3782 lnet_finalize(msg, rc);
3785 /* completion will be signalled by an event */
3788 EXPORT_SYMBOL(LNetGet);
3791 * Calculate distance to node at \a dstnid.
3793 * \param dstnid Target NID.
3794 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
3796 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
3799 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
3800 * local_nid_dist_zero is set, which is the default.
3801 * \retval positives Distance to target NID, i.e. number of hops plus one.
3802 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
3805 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
3807 struct list_head *e;
3808 struct lnet_ni *ni = NULL;
3809 struct lnet_remotenet *rnet;
3810 __u32 dstnet = LNET_NIDNET(dstnid);
3814 struct list_head *rn_list;
3816 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
3817 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
3818 * keep order 0 free for 0@lo and order 1 free for a local NID
3821 LASSERT(the_lnet.ln_refcount > 0);
3823 cpt = lnet_net_lock_current();
3825 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
3826 if (ni->ni_nid == dstnid) {
3827 if (srcnidp != NULL)
3829 if (orderp != NULL) {
3830 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
3835 lnet_net_unlock(cpt);
3837 return local_nid_dist_zero ? 0 : 1;
3840 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
3841 /* Check if ni was originally created in
3842 * current net namespace.
3843 * If not, assign order above 0xffff0000,
3844 * to make this ni not a priority. */
3845 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
3846 order += 0xffff0000;
3848 if (srcnidp != NULL)
3849 *srcnidp = ni->ni_nid;
3852 lnet_net_unlock(cpt);
3859 rn_list = lnet_net2rnethash(dstnet);
3860 list_for_each(e, rn_list) {
3861 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
3863 if (rnet->lrn_net == dstnet) {
3864 struct lnet_route *route;
3865 struct lnet_route *shortest = NULL;
3866 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
3869 LASSERT(!list_empty(&rnet->lrn_routes));
3871 list_for_each_entry(route, &rnet->lrn_routes,
3873 route_hops = route->lr_hops;
3874 if (route_hops == LNET_UNDEFINED_HOPS)
3876 if (shortest == NULL ||
3877 route_hops < shortest_hops) {
3879 shortest_hops = route_hops;
3883 LASSERT(shortest != NULL);
3884 hops = shortest_hops;
3885 if (srcnidp != NULL) {
3886 ni = lnet_get_next_ni_locked(
3887 shortest->lr_gateway->lpni_net,
3889 *srcnidp = ni->ni_nid;
3893 lnet_net_unlock(cpt);
3899 lnet_net_unlock(cpt);
3900 return -EHOSTUNREACH;
3902 EXPORT_SYMBOL(LNetDist);