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/
31 * lnet/lnet/lib-move.c
33 * Data movement routines
36 #define DEBUG_SUBSYSTEM S_LNET
38 #include <linux/pagemap.h>
40 #include <lnet/lib-lnet.h>
41 #include <linux/nsproxy.h>
42 #include <lnet/lnet_rdma.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 struct lnet_nid sd_dst_nid;
59 struct lnet_nid sd_src_nid;
60 struct lnet_nid sd_rtr_nid;
67 lnet_msg_is_response(struct lnet_msg *msg)
69 return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
73 lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
75 if (md_options & LNET_MD_NO_TRACK_RESPONSE)
76 /* Explicitly disabled in MD options */
79 if (md_options & LNET_MD_TRACK_RESPONSE)
80 /* Explicity enabled in MD options */
83 if (lnet_response_tracking == 3)
84 /* Enabled for all message types */
87 if (msg_type == LNET_MSG_PUT)
88 return lnet_response_tracking == 2;
90 if (msg_type == LNET_MSG_GET)
91 return lnet_response_tracking == 1;
96 static inline struct lnet_comm_count *
97 get_stats_counts(struct lnet_element_stats *stats,
98 enum lnet_stats_type stats_type)
100 switch (stats_type) {
101 case LNET_STATS_TYPE_SEND:
102 return &stats->el_send_stats;
103 case LNET_STATS_TYPE_RECV:
104 return &stats->el_recv_stats;
105 case LNET_STATS_TYPE_DROP:
106 return &stats->el_drop_stats;
108 CERROR("Unknown stats type\n");
114 void lnet_incr_stats(struct lnet_element_stats *stats,
115 enum lnet_msg_type msg_type,
116 enum lnet_stats_type stats_type)
118 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
124 atomic_inc(&counts->co_ack_count);
127 atomic_inc(&counts->co_put_count);
130 atomic_inc(&counts->co_get_count);
133 atomic_inc(&counts->co_reply_count);
136 atomic_inc(&counts->co_hello_count);
139 CERROR("There is a BUG in the code. Unknown message type\n");
144 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
145 enum lnet_stats_type stats_type)
147 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
151 return (atomic_read(&counts->co_ack_count) +
152 atomic_read(&counts->co_put_count) +
153 atomic_read(&counts->co_get_count) +
154 atomic_read(&counts->co_reply_count) +
155 atomic_read(&counts->co_hello_count));
158 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
159 struct lnet_comm_count *counts)
161 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
162 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
163 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
164 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
165 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
168 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
169 struct lnet_element_stats *stats)
171 struct lnet_comm_count *counts;
176 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
179 assign_stats(&msg_stats->im_send_stats, counts);
181 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
184 assign_stats(&msg_stats->im_recv_stats, counts);
186 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
189 assign_stats(&msg_stats->im_drop_stats, counts);
193 lnet_fail_nid(lnet_nid_t nid4, unsigned int threshold)
195 struct lnet_test_peer *tp;
196 struct list_head *el;
197 struct list_head *next;
201 lnet_nid4_to_nid(nid4, &nid);
202 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
203 if (threshold != 0) {
204 /* Adding a new entry */
205 LIBCFS_ALLOC(tp, sizeof(*tp));
210 tp->tp_threshold = threshold;
213 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
220 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
221 tp = list_entry(el, struct lnet_test_peer, tp_list);
223 if (tp->tp_threshold == 0 || /* needs culling anyway */
224 LNET_NID_IS_ANY(&nid) || /* removing all entries */
225 nid_same(&tp->tp_nid, &nid)) { /* matched this one */
226 list_move(&tp->tp_list, &cull);
232 while ((tp = list_first_entry_or_null(&cull,
233 struct lnet_test_peer,
235 list_del(&tp->tp_list);
236 LIBCFS_FREE(tp, sizeof(*tp));
242 fail_peer(struct lnet_nid *nid, int outgoing)
244 struct lnet_test_peer *tp;
245 struct list_head *el;
246 struct list_head *next;
250 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
253 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
254 tp = list_entry(el, struct lnet_test_peer, tp_list);
256 if (tp->tp_threshold == 0) {
259 /* only cull zombies on outgoing tests,
260 * since we may be at interrupt priority on
261 * incoming messages. */
262 list_move(&tp->tp_list, &cull);
267 if (LNET_NID_IS_ANY(&tp->tp_nid) || /* fail every peer */
268 nid_same(nid, &tp->tp_nid)) { /* fail this peer */
271 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
274 tp->tp_threshold == 0) {
276 list_move(&tp->tp_list, &cull);
285 while ((tp = list_first_entry_or_null(&cull,
286 struct lnet_test_peer,
288 list_del(&tp->tp_list);
289 LIBCFS_FREE(tp, sizeof(*tp));
296 lnet_iov_nob(unsigned int niov, struct kvec *iov)
298 unsigned int nob = 0;
300 LASSERT(niov == 0 || iov != NULL);
302 nob += (iov++)->iov_len;
306 EXPORT_SYMBOL(lnet_iov_nob);
309 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
310 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
313 /* NB diov, siov are READ-ONLY */
314 unsigned int this_nob;
319 /* skip complete frags before 'doffset' */
321 while (doffset >= diov->iov_len) {
322 doffset -= diov->iov_len;
328 /* skip complete frags before 'soffset' */
330 while (soffset >= siov->iov_len) {
331 soffset -= siov->iov_len;
340 this_nob = min3((unsigned int)diov->iov_len - doffset,
341 (unsigned int)siov->iov_len - soffset,
344 memcpy((char *)diov->iov_base + doffset,
345 (char *)siov->iov_base + soffset, this_nob);
348 if (diov->iov_len > doffset + this_nob) {
356 if (siov->iov_len > soffset + this_nob) {
365 EXPORT_SYMBOL(lnet_copy_iov2iov);
368 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
370 unsigned int nob = 0;
372 LASSERT(niov == 0 || kiov != NULL);
374 nob += (kiov++)->bv_len;
378 EXPORT_SYMBOL(lnet_kiov_nob);
381 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
382 unsigned int doffset,
383 unsigned int nsiov, struct bio_vec *siov,
384 unsigned int soffset,
387 /* NB diov, siov are READ-ONLY */
388 unsigned int this_nob;
395 LASSERT (!in_interrupt ());
398 while (doffset >= diov->bv_len) {
399 doffset -= diov->bv_len;
406 while (soffset >= siov->bv_len) {
407 soffset -= siov->bv_len;
416 this_nob = min3(diov->bv_len - doffset,
417 siov->bv_len - soffset,
421 daddr = ((char *)kmap(diov->bv_page)) +
422 diov->bv_offset + doffset;
424 saddr = ((char *)kmap(siov->bv_page)) +
425 siov->bv_offset + soffset;
427 /* Vanishing risk of kmap deadlock when mapping 2 pages.
428 * However in practice at least one of the kiovs will be mapped
429 * kernel pages and the map/unmap will be NOOPs */
431 memcpy (daddr, saddr, this_nob);
434 if (diov->bv_len > doffset + this_nob) {
438 kunmap(diov->bv_page);
445 if (siov->bv_len > soffset + this_nob) {
449 kunmap(siov->bv_page);
458 kunmap(diov->bv_page);
460 kunmap(siov->bv_page);
462 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
465 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
466 unsigned int nkiov, struct bio_vec *kiov,
467 unsigned int kiovoffset,
470 /* NB iov, kiov are READ-ONLY */
471 unsigned int this_nob;
477 LASSERT (!in_interrupt ());
480 while (iovoffset >= iov->iov_len) {
481 iovoffset -= iov->iov_len;
488 while (kiovoffset >= kiov->bv_len) {
489 kiovoffset -= kiov->bv_len;
498 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
499 (unsigned int)kiov->bv_len - kiovoffset,
503 addr = ((char *)kmap(kiov->bv_page)) +
504 kiov->bv_offset + kiovoffset;
506 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
509 if (iov->iov_len > iovoffset + this_nob) {
510 iovoffset += this_nob;
517 if (kiov->bv_len > kiovoffset + this_nob) {
519 kiovoffset += this_nob;
521 kunmap(kiov->bv_page);
531 kunmap(kiov->bv_page);
533 EXPORT_SYMBOL(lnet_copy_kiov2iov);
536 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
537 unsigned int kiovoffset,
538 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
541 /* NB kiov, iov are READ-ONLY */
542 unsigned int this_nob;
548 LASSERT (!in_interrupt ());
551 while (kiovoffset >= kiov->bv_len) {
552 kiovoffset -= kiov->bv_len;
559 while (iovoffset >= iov->iov_len) {
560 iovoffset -= iov->iov_len;
569 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
570 (unsigned int)iov->iov_len - iovoffset,
574 addr = ((char *)kmap(kiov->bv_page)) +
575 kiov->bv_offset + kiovoffset;
577 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
580 if (kiov->bv_len > kiovoffset + this_nob) {
582 kiovoffset += this_nob;
584 kunmap(kiov->bv_page);
591 if (iov->iov_len > iovoffset + this_nob) {
592 iovoffset += this_nob;
601 kunmap(kiov->bv_page);
603 EXPORT_SYMBOL(lnet_copy_iov2kiov);
606 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
607 int src_niov, struct bio_vec *src,
608 unsigned int offset, unsigned int len)
610 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
611 * for exactly 'len' bytes, and return the number of entries.
612 * NB not destructive to 'src' */
613 unsigned int frag_len;
616 if (len == 0) /* no data => */
617 return (0); /* no frags */
619 LASSERT(src_niov > 0);
620 while (offset >= src->bv_len) { /* skip initial frags */
621 offset -= src->bv_len;
624 LASSERT(src_niov > 0);
629 LASSERT(src_niov > 0);
630 LASSERT((int)niov <= dst_niov);
632 frag_len = src->bv_len - offset;
633 dst->bv_page = src->bv_page;
634 dst->bv_offset = src->bv_offset + offset;
636 if (len <= frag_len) {
638 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
642 dst->bv_len = frag_len;
643 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
653 EXPORT_SYMBOL(lnet_extract_kiov);
656 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
657 int delayed, unsigned int offset, unsigned int mlen,
660 unsigned int niov = 0;
661 struct kvec *iov = NULL;
662 struct bio_vec *kiov = NULL;
665 LASSERT (!in_interrupt ());
666 LASSERT (mlen == 0 || msg != NULL);
669 LASSERT(msg->msg_receiving);
670 LASSERT(!msg->msg_sending);
671 LASSERT(rlen == msg->msg_len);
672 LASSERT(mlen <= msg->msg_len);
673 LASSERT(msg->msg_offset == offset);
674 LASSERT(msg->msg_wanted == mlen);
676 msg->msg_receiving = 0;
679 niov = msg->msg_niov;
680 kiov = msg->msg_kiov;
683 LASSERT ((iov == NULL) != (kiov == NULL));
687 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
688 niov, kiov, offset, mlen,
691 lnet_finalize(msg, rc);
695 lnet_setpayloadbuffer(struct lnet_msg *msg)
697 struct lnet_libmd *md = msg->msg_md;
699 LASSERT(msg->msg_len > 0);
700 LASSERT(!msg->msg_routing);
702 LASSERT(msg->msg_niov == 0);
703 LASSERT(msg->msg_kiov == NULL);
705 msg->msg_niov = md->md_niov;
706 msg->msg_kiov = md->md_kiov;
710 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_processid *target,
711 unsigned int offset, unsigned int len)
713 msg->msg_type = type;
714 msg->msg_target = *target;
716 msg->msg_offset = offset;
719 lnet_setpayloadbuffer(msg);
721 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
722 msg->msg_hdr.type = type;
723 /* dest_nid will be overwritten by lnet_select_pathway() */
724 msg->msg_hdr.dest_nid = target->nid;
725 msg->msg_hdr.dest_pid = target->pid;
726 /* src_nid will be set later */
727 msg->msg_hdr.src_pid = the_lnet.ln_pid;
728 msg->msg_hdr.payload_length = len;
732 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
734 void *priv = msg->msg_private;
737 LASSERT(!in_interrupt());
738 LASSERT(nid_is_lo0(&ni->ni_nid) ||
739 (msg->msg_txcredit && msg->msg_peertxcredit));
741 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
743 msg->msg_no_resend = true;
744 lnet_finalize(msg, rc);
749 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
753 LASSERT(!msg->msg_sending);
754 LASSERT(msg->msg_receiving);
755 LASSERT(!msg->msg_rx_ready_delay);
756 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
758 msg->msg_rx_ready_delay = 1;
759 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
762 CERROR("recv from %s / send to %s aborted: "
763 "eager_recv failed %d\n",
764 libcfs_nidstr(&msg->msg_rxpeer->lpni_nid),
765 libcfs_idstr(&msg->msg_target), rc);
766 LASSERT(rc < 0); /* required by my callers */
773 * -ETIMEDOUT if the message deadline has been exceeded
774 * -EHOSTUNREACH if the peer is down
775 * 0 if this message should not be dropped
778 lnet_check_message_drop(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
779 struct lnet_msg *msg)
781 /* Drop message if we've exceeded the message deadline */
782 if (ktime_after(ktime_get(), msg->msg_deadline))
785 if (msg->msg_target.pid & LNET_PID_USERFLAG)
788 if (!lnet_peer_aliveness_enabled(lpni))
791 /* If we're resending a message, let's attempt to send it even if
792 * the peer is down to fulfill our resend quota on the message
794 if (msg->msg_retry_count > 0)
797 /* try and send recovery messages regardless */
798 if (msg->msg_recovery)
801 /* always send any responses */
802 if (lnet_msg_is_response(msg))
805 /* always send non-routed messages */
806 if (!msg->msg_routing)
809 /* assume peer_ni is alive as long as we're within the configured
812 if (ktime_get_seconds() <
813 (lpni->lpni_last_alive +
814 lpni->lpni_net->net_tunables.lct_peer_timeout))
817 if (!lnet_is_peer_ni_alive(lpni))
818 return -EHOSTUNREACH;
824 * \param msg The message to be sent.
825 * \param do_send True if lnet_ni_send() should be called in this function.
826 * lnet_send() is going to lnet_net_unlock immediately after this, so
827 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
829 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
830 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
831 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
832 * \retval -ECANCELED If the MD of the message has been unlinked.
835 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
837 struct lnet_peer_ni *lp = msg->msg_txpeer;
838 struct lnet_ni *ni = msg->msg_txni;
839 int cpt = msg->msg_tx_cpt;
840 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
843 /* non-lnet_send() callers have checked before */
844 LASSERT(!do_send || msg->msg_tx_delayed);
845 LASSERT(!msg->msg_receiving);
846 LASSERT(msg->msg_tx_committed);
848 /* can't get here if we're sending to the loopback interface */
849 if (the_lnet.ln_loni)
850 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
852 /* NB 'lp' is always the next hop */
853 rc = lnet_check_message_drop(ni, lp, msg);
855 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
856 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
858 lnet_net_unlock(cpt);
860 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
862 LNET_STATS_TYPE_DROP);
864 lnet_incr_stats(&msg->msg_txni->ni_stats,
866 LNET_STATS_TYPE_DROP);
868 if (rc == -EHOSTUNREACH) {
869 CNETERR("Dropping message for %s: peer not alive\n",
870 libcfs_idstr(&msg->msg_target));
871 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
873 CNETERR("Dropping message for %s: exceeded message deadline\n",
874 libcfs_idstr(&msg->msg_target));
875 msg->msg_health_status =
876 LNET_MSG_STATUS_NETWORK_TIMEOUT;
880 lnet_finalize(msg, rc);
886 if (msg->msg_md != NULL &&
887 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
888 lnet_net_unlock(cpt);
890 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
891 "called on the MD/ME.\n",
892 libcfs_idstr(&msg->msg_target));
894 msg->msg_no_resend = true;
895 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
896 msg, libcfs_idstr(&msg->msg_target));
897 lnet_finalize(msg, -ECANCELED);
904 if (!msg->msg_peertxcredit) {
905 spin_lock(&lp->lpni_lock);
906 LASSERT((lp->lpni_txcredits < 0) ==
907 !list_empty(&lp->lpni_txq));
909 msg->msg_peertxcredit = 1;
910 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr_nid4);
911 lp->lpni_txcredits--;
913 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
914 lp->lpni_mintxcredits = lp->lpni_txcredits;
916 if (lp->lpni_txcredits < 0) {
917 msg->msg_tx_delayed = 1;
918 list_add_tail(&msg->msg_list, &lp->lpni_txq);
919 spin_unlock(&lp->lpni_lock);
920 return LNET_CREDIT_WAIT;
922 spin_unlock(&lp->lpni_lock);
925 if (!msg->msg_txcredit) {
926 LASSERT((tq->tq_credits < 0) ==
927 !list_empty(&tq->tq_delayed));
929 msg->msg_txcredit = 1;
931 atomic_dec(&ni->ni_tx_credits);
933 if (tq->tq_credits < tq->tq_credits_min)
934 tq->tq_credits_min = tq->tq_credits;
936 if (tq->tq_credits < 0) {
937 msg->msg_tx_delayed = 1;
938 list_add_tail(&msg->msg_list, &tq->tq_delayed);
939 return LNET_CREDIT_WAIT;
943 if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
944 lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
945 msg->msg_tx_delayed = 1;
946 return LNET_CREDIT_WAIT;
949 /* unset the tx_delay flag as we're going to send it now */
950 msg->msg_tx_delayed = 0;
953 lnet_net_unlock(cpt);
954 lnet_ni_send(ni, msg);
957 return LNET_CREDIT_OK;
961 static struct lnet_rtrbufpool *
962 lnet_msg2bufpool(struct lnet_msg *msg)
964 struct lnet_rtrbufpool *rbp;
967 LASSERT(msg->msg_rx_committed);
969 cpt = msg->msg_rx_cpt;
970 rbp = &the_lnet.ln_rtrpools[cpt][0];
972 LASSERT(msg->msg_len <= LNET_MTU);
973 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
975 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
982 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
984 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
985 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
986 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
987 * received or OK to receive */
988 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
989 struct lnet_peer *lp;
990 struct lnet_rtrbufpool *rbp;
991 struct lnet_rtrbuf *rb;
993 LASSERT(msg->msg_kiov == NULL);
994 LASSERT(msg->msg_niov == 0);
995 LASSERT(msg->msg_routing);
996 LASSERT(msg->msg_receiving);
997 LASSERT(!msg->msg_sending);
998 LASSERT(lpni->lpni_peer_net);
999 LASSERT(lpni->lpni_peer_net->lpn_peer);
1001 lp = lpni->lpni_peer_net->lpn_peer;
1003 /* non-lnet_parse callers only receive delayed messages */
1004 LASSERT(!do_recv || msg->msg_rx_delayed);
1006 if (!msg->msg_peerrtrcredit) {
1007 /* lpni_lock protects the credit manipulation */
1008 spin_lock(&lpni->lpni_lock);
1010 msg->msg_peerrtrcredit = 1;
1011 lpni->lpni_rtrcredits--;
1012 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1013 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1015 if (lpni->lpni_rtrcredits < 0) {
1016 spin_unlock(&lpni->lpni_lock);
1017 /* must have checked eager_recv before here */
1018 LASSERT(msg->msg_rx_ready_delay);
1019 msg->msg_rx_delayed = 1;
1020 /* lp_lock protects the lp_rtrq */
1021 spin_lock(&lp->lp_lock);
1022 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1023 spin_unlock(&lp->lp_lock);
1024 return LNET_CREDIT_WAIT;
1026 spin_unlock(&lpni->lpni_lock);
1029 rbp = lnet_msg2bufpool(msg);
1031 if (!msg->msg_rtrcredit) {
1032 msg->msg_rtrcredit = 1;
1034 if (rbp->rbp_credits < rbp->rbp_mincredits)
1035 rbp->rbp_mincredits = rbp->rbp_credits;
1037 if (rbp->rbp_credits < 0) {
1038 /* must have checked eager_recv before here */
1039 LASSERT(msg->msg_rx_ready_delay);
1040 msg->msg_rx_delayed = 1;
1041 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1042 return LNET_CREDIT_WAIT;
1046 LASSERT(!list_empty(&rbp->rbp_bufs));
1047 rb = list_first_entry(&rbp->rbp_bufs, struct lnet_rtrbuf, rb_list);
1048 list_del(&rb->rb_list);
1050 msg->msg_niov = rbp->rbp_npages;
1051 msg->msg_kiov = &rb->rb_kiov[0];
1053 /* unset the msg-rx_delayed flag since we're receiving the message */
1054 msg->msg_rx_delayed = 0;
1057 int cpt = msg->msg_rx_cpt;
1059 lnet_net_unlock(cpt);
1060 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1061 0, msg->msg_len, msg->msg_len);
1064 return LNET_CREDIT_OK;
1068 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1070 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1071 struct lnet_ni *txni = msg->msg_txni;
1072 struct lnet_msg *msg2;
1074 if (msg->msg_txcredit) {
1075 struct lnet_ni *ni = msg->msg_txni;
1076 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1078 /* give back NI txcredits */
1079 msg->msg_txcredit = 0;
1081 LASSERT((tq->tq_credits < 0) ==
1082 !list_empty(&tq->tq_delayed));
1085 atomic_inc(&ni->ni_tx_credits);
1086 if (tq->tq_credits <= 0) {
1087 msg2 = list_first_entry(&tq->tq_delayed,
1088 struct lnet_msg, msg_list);
1089 list_del(&msg2->msg_list);
1091 LASSERT(msg2->msg_txni == ni);
1092 LASSERT(msg2->msg_tx_delayed);
1093 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1095 (void) lnet_post_send_locked(msg2, 1);
1099 if (msg->msg_peertxcredit) {
1100 /* give back peer txcredits */
1101 msg->msg_peertxcredit = 0;
1103 spin_lock(&txpeer->lpni_lock);
1104 LASSERT((txpeer->lpni_txcredits < 0) ==
1105 !list_empty(&txpeer->lpni_txq));
1107 txpeer->lpni_txqnob -= msg->msg_len +
1108 sizeof(struct lnet_hdr_nid4);
1109 LASSERT(txpeer->lpni_txqnob >= 0);
1111 txpeer->lpni_txcredits++;
1112 if (txpeer->lpni_txcredits <= 0) {
1115 msg2 = list_first_entry(&txpeer->lpni_txq,
1116 struct lnet_msg, msg_list);
1117 list_del(&msg2->msg_list);
1118 spin_unlock(&txpeer->lpni_lock);
1120 LASSERT(msg2->msg_txpeer == txpeer);
1121 LASSERT(msg2->msg_tx_delayed);
1123 msg2_cpt = msg2->msg_tx_cpt;
1126 * The msg_cpt can be different from the msg2_cpt
1127 * so we need to make sure we lock the correct cpt
1129 * Once we call lnet_post_send_locked() it is no
1130 * longer safe to access msg2, since it could've
1131 * been freed by lnet_finalize(), but we still
1132 * need to relock the correct cpt, so we cache the
1133 * msg2_cpt for the purpose of the check that
1134 * follows the call to lnet_pose_send_locked().
1136 if (msg2_cpt != msg->msg_tx_cpt) {
1137 lnet_net_unlock(msg->msg_tx_cpt);
1138 lnet_net_lock(msg2_cpt);
1140 (void) lnet_post_send_locked(msg2, 1);
1141 if (msg2_cpt != msg->msg_tx_cpt) {
1142 lnet_net_unlock(msg2_cpt);
1143 lnet_net_lock(msg->msg_tx_cpt);
1146 spin_unlock(&txpeer->lpni_lock);
1151 msg->msg_txni = NULL;
1152 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1155 if (txpeer != NULL) {
1156 msg->msg_txpeer = NULL;
1157 lnet_peer_ni_decref_locked(txpeer);
1162 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1164 struct lnet_msg *msg;
1166 if (list_empty(&rbp->rbp_msgs))
1168 msg = list_first_entry(&rbp->rbp_msgs,
1169 struct lnet_msg, msg_list);
1170 list_del(&msg->msg_list);
1172 (void)lnet_post_routed_recv_locked(msg, 1);
1176 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1178 struct lnet_msg *msg;
1179 struct lnet_msg *tmp;
1181 lnet_net_unlock(cpt);
1183 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1184 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1185 0, 0, 0, msg->msg_hdr.payload_length);
1186 list_del_init(&msg->msg_list);
1187 msg->msg_no_resend = true;
1188 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1189 lnet_finalize(msg, -ECANCELED);
1196 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1198 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1199 struct lnet_peer *lp;
1200 struct lnet_ni *rxni = msg->msg_rxni;
1201 struct lnet_msg *msg2;
1203 if (msg->msg_rtrcredit) {
1204 /* give back global router credits */
1205 struct lnet_rtrbuf *rb;
1206 struct lnet_rtrbufpool *rbp;
1208 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1209 * there until it gets one allocated, or aborts the wait
1211 LASSERT(msg->msg_kiov != NULL);
1213 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1216 msg->msg_kiov = NULL;
1217 msg->msg_rtrcredit = 0;
1219 LASSERT(rbp == lnet_msg2bufpool(msg));
1221 LASSERT((rbp->rbp_credits > 0) ==
1222 !list_empty(&rbp->rbp_bufs));
1224 /* If routing is now turned off, we just drop this buffer and
1225 * don't bother trying to return credits. */
1226 if (!the_lnet.ln_routing) {
1227 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1231 /* It is possible that a user has lowered the desired number of
1232 * buffers in this pool. Make sure we never put back
1233 * more buffers than the stated number. */
1234 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1235 /* Discard this buffer so we don't have too
1237 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1238 rbp->rbp_nbuffers--;
1240 list_add(&rb->rb_list, &rbp->rbp_bufs);
1242 if (rbp->rbp_credits <= 0)
1243 lnet_schedule_blocked_locked(rbp);
1248 if (msg->msg_peerrtrcredit) {
1250 LASSERT(rxpeerni->lpni_peer_net);
1251 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1253 /* give back peer router credits */
1254 msg->msg_peerrtrcredit = 0;
1256 spin_lock(&rxpeerni->lpni_lock);
1257 rxpeerni->lpni_rtrcredits++;
1258 spin_unlock(&rxpeerni->lpni_lock);
1260 lp = rxpeerni->lpni_peer_net->lpn_peer;
1261 spin_lock(&lp->lp_lock);
1263 /* drop all messages which are queued to be routed on that
1265 if (!the_lnet.ln_routing) {
1267 list_splice_init(&lp->lp_rtrq, &drop);
1268 spin_unlock(&lp->lp_lock);
1269 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1270 } else if (!list_empty(&lp->lp_rtrq)) {
1273 msg2 = list_first_entry(&lp->lp_rtrq,
1274 struct lnet_msg, msg_list);
1275 list_del(&msg2->msg_list);
1276 msg2_cpt = msg2->msg_rx_cpt;
1277 spin_unlock(&lp->lp_lock);
1279 * messages on the lp_rtrq can be from any NID in
1280 * the peer, which means they might have different
1281 * cpts. We need to make sure we lock the right
1284 if (msg2_cpt != msg->msg_rx_cpt) {
1285 lnet_net_unlock(msg->msg_rx_cpt);
1286 lnet_net_lock(msg2_cpt);
1288 (void) lnet_post_routed_recv_locked(msg2, 1);
1289 if (msg2_cpt != msg->msg_rx_cpt) {
1290 lnet_net_unlock(msg2_cpt);
1291 lnet_net_lock(msg->msg_rx_cpt);
1294 spin_unlock(&lp->lp_lock);
1298 msg->msg_rxni = NULL;
1299 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1301 if (rxpeerni != NULL) {
1302 msg->msg_rxpeer = NULL;
1303 lnet_peer_ni_decref_locked(rxpeerni);
1307 static struct lnet_peer_ni *
1308 lnet_select_peer_ni(struct lnet_ni *best_ni, struct lnet_nid *dst_nid,
1309 struct lnet_peer *peer,
1310 struct lnet_peer_ni *best_lpni,
1311 struct lnet_peer_net *peer_net)
1314 * Look at the peer NIs for the destination peer that connect
1315 * to the chosen net. If a peer_ni is preferred when using the
1316 * best_ni to communicate, we use that one. If there is no
1317 * preferred peer_ni, or there are multiple preferred peer_ni,
1318 * the available transmit credits are used. If the transmit
1319 * credits are equal, we round-robin over the peer_ni.
1321 struct lnet_peer_ni *lpni = NULL;
1322 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1324 int best_lpni_healthv = (best_lpni) ?
1325 atomic_read(&best_lpni->lpni_healthv) : 0;
1326 bool best_lpni_is_preferred = false;
1327 bool lpni_is_preferred;
1329 __u32 lpni_sel_prio;
1330 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1332 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1334 * if the best_ni we've chosen aleady has this lpni
1335 * preferred, then let's use it
1338 lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1339 lpni, &best_ni->ni_nid);
1340 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1341 libcfs_nidstr(&best_ni->ni_nid),
1344 lpni_is_preferred = false;
1347 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1348 lpni_sel_prio = lpni->lpni_sel_priority;
1351 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1352 libcfs_nidstr(&lpni->lpni_nid),
1353 libcfs_nidstr(&best_lpni->lpni_nid),
1354 lpni_healthv, best_lpni_healthv,
1355 lpni_sel_prio, best_sel_prio,
1356 lpni->lpni_txcredits, best_lpni_credits,
1357 lpni->lpni_seq, best_lpni->lpni_seq);
1361 /* pick the healthiest peer ni */
1362 if (lpni_healthv < best_lpni_healthv)
1364 else if (lpni_healthv > best_lpni_healthv) {
1365 if (best_lpni_is_preferred)
1366 best_lpni_is_preferred = false;
1370 if (lpni_sel_prio > best_sel_prio)
1372 else if (lpni_sel_prio < best_sel_prio) {
1373 if (best_lpni_is_preferred)
1374 best_lpni_is_preferred = false;
1378 /* if this is a preferred peer use it */
1379 if (!best_lpni_is_preferred && lpni_is_preferred) {
1380 best_lpni_is_preferred = true;
1382 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1383 /* this is not the preferred peer so let's ignore
1389 if (lpni->lpni_txcredits < best_lpni_credits)
1390 /* We already have a peer that has more credits
1391 * available than this one. No need to consider
1392 * this peer further.
1395 else if (lpni->lpni_txcredits > best_lpni_credits)
1398 /* The best peer found so far and the current peer
1399 * have the same number of available credits let's
1400 * make sure to select between them using Round Robin
1402 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1405 best_lpni_is_preferred = lpni_is_preferred;
1406 best_lpni_healthv = lpni_healthv;
1407 best_sel_prio = lpni_sel_prio;
1409 best_lpni_credits = lpni->lpni_txcredits;
1412 /* if we still can't find a peer ni then we can't reach it */
1414 u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1415 LNET_NID_NET(dst_nid);
1416 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1417 libcfs_net2str(net_id));
1421 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1422 libcfs_nidstr(&best_lpni->lpni_nid));
1428 * Prerequisite: the best_ni should already be set in the sd
1429 * Find the best lpni.
1430 * If the net id is provided then restrict lpni selection on
1431 * that particular net.
1432 * Otherwise find any reachable lpni. When dealing with an MR
1433 * gateway and it has multiple lpnis which we can use
1434 * we want to select the best one from the list of reachable
1437 static inline struct lnet_peer_ni *
1438 lnet_find_best_lpni(struct lnet_ni *lni, struct lnet_nid *dst_nid,
1439 struct lnet_peer *peer, u32 net_id)
1441 struct lnet_peer_net *peer_net;
1443 /* find the best_lpni on any local network */
1444 if (net_id == LNET_NET_ANY) {
1445 struct lnet_peer_ni *best_lpni = NULL;
1446 struct lnet_peer_net *lpn;
1447 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1448 /* no net specified find any reachable peer ni */
1449 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1451 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1457 /* restrict on the specified net */
1458 peer_net = lnet_peer_get_net_locked(peer, net_id);
1460 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1466 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1468 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1471 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1474 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1477 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1483 /* Compare route priorities and hop counts */
1485 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1487 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1488 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1490 if (r1->lr_priority < r2->lr_priority)
1493 if (r1->lr_priority > r2->lr_priority)
1496 if (r1_hops < r2_hops)
1499 if (r1_hops > r2_hops)
1505 static struct lnet_route *
1506 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1507 struct lnet_peer_ni *remote_lpni,
1508 struct lnet_route **prev_route,
1509 struct lnet_peer_ni **gwni)
1511 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1512 struct lnet_route *best_route;
1513 struct lnet_route *last_route;
1514 struct lnet_route *route;
1516 bool best_rte_is_preferred = false;
1517 struct lnet_nid *gw_pnid;
1519 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1520 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1522 best_route = last_route = NULL;
1523 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1524 if (!lnet_is_route_alive(route))
1526 gw_pnid = &route->lr_gateway->lp_primary_nid;
1528 /* no protection on below fields, but it's harmless */
1529 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1532 /* if the best route found is in the preferred list then
1533 * tag it as preferred and use it later on. But if we
1534 * didn't find any routes which are on the preferred list
1535 * then just use the best route possible.
1537 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1539 if (!best_route || (rc && !best_rte_is_preferred)) {
1540 /* Restrict the selection of the router NI on the
1541 * src_net provided. If the src_net is LNET_NID_ANY,
1542 * then select the best interface available.
1544 lpni = lnet_find_best_lpni(NULL, NULL,
1549 "Gateway %s does not have a peer NI on net %s\n",
1550 libcfs_nidstr(gw_pnid),
1551 libcfs_net2str(src_net));
1556 if (rc && !best_rte_is_preferred) {
1557 /* This is the first preferred route we found,
1558 * so it beats any route found previously
1564 best_rte_is_preferred = true;
1565 CDEBUG(D_NET, "preferred gw = %s\n",
1566 libcfs_nidstr(gw_pnid));
1568 } else if ((!rc) && best_rte_is_preferred)
1569 /* The best route we found so far is in the preferred
1570 * list, so it beats any non-preferred route
1575 best_route = last_route = route;
1580 rc = lnet_compare_routes(route, best_route);
1584 /* Restrict the selection of the router NI on the
1585 * src_net provided. If the src_net is LNET_NID_ANY,
1586 * then select the best interface available.
1588 lpni = lnet_find_best_lpni(NULL, NULL, route->lr_gateway,
1592 "Gateway %s does not have a peer NI on net %s\n",
1593 libcfs_nidstr(gw_pnid),
1594 libcfs_net2str(src_net));
1604 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1608 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1615 *prev_route = last_route;
1621 static inline unsigned int
1622 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1624 if (dev_idx == UINT_MAX)
1627 if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1628 !ni->ni_net->net_lnd->lnd_get_dev_prio)
1631 return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1634 static struct lnet_ni *
1635 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1636 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1637 struct lnet_msg *msg, int md_cpt)
1639 struct lnet_libmd *md = msg->msg_md;
1640 unsigned int offset = msg->msg_offset;
1641 unsigned int shortest_distance;
1642 struct lnet_ni *ni = NULL;
1645 __u32 best_sel_prio;
1646 unsigned int best_dev_prio;
1648 unsigned int dev_idx = UINT_MAX;
1649 bool gpu = md ? (md->md_flags & LNET_MD_FLAG_GPU) : false;
1652 struct page *page = lnet_get_first_page(md, offset);
1654 dev_idx = lnet_get_dev_idx(page);
1658 * If there is no peer_ni that we can send to on this network,
1659 * then there is no point in looking for a new best_ni here.
1661 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1664 if (best_ni == NULL) {
1665 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1666 shortest_distance = UINT_MAX;
1667 best_dev_prio = UINT_MAX;
1668 best_credits = INT_MIN;
1670 best_ni_fatal = true;
1672 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1673 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1674 best_ni->ni_dev_cpt);
1675 best_credits = atomic_read(&best_ni->ni_tx_credits);
1676 best_healthv = atomic_read(&best_ni->ni_healthv);
1677 best_sel_prio = best_ni->ni_sel_priority;
1678 best_ni_fatal = atomic_read(&best_ni->ni_fatal_error_on);
1681 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1682 unsigned int distance;
1687 unsigned int ni_dev_prio;
1689 ni_credits = atomic_read(&ni->ni_tx_credits);
1690 ni_healthv = atomic_read(&ni->ni_healthv);
1691 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1692 ni_sel_prio = ni->ni_sel_priority;
1695 * calculate the distance from the CPT on which
1696 * the message memory is allocated to the CPT of
1697 * the NI's physical device
1699 distance = cfs_cpt_distance(lnet_cpt_table(),
1703 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1706 * All distances smaller than the NUMA range
1707 * are treated equally.
1709 if (!gpu && distance < lnet_numa_range)
1710 distance = lnet_numa_range;
1713 * Select on health, selection policy, direct dma prio,
1714 * shorter distance, available credits, then round-robin.
1717 CDEBUG(D_NET, "compare ni %s [f:%s, c:%d, d:%d, s:%d, p:%u, g:%u, h:%d] with best_ni %s [f:%s, c:%d, d:%d, s:%d, p:%u, g:%u, h:%d]\n",
1718 libcfs_nidstr(&ni->ni_nid),
1719 ni_fatal ? "y" : "n", ni_credits, distance,
1720 ni->ni_seq, ni_sel_prio, ni_dev_prio, ni_healthv,
1721 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1723 best_ni_fatal ? "y" : "n", best_credits,
1725 (best_ni) ? best_ni->ni_seq : 0,
1726 best_sel_prio, best_dev_prio, best_healthv);
1730 if (ni_fatal && !best_ni_fatal)
1732 else if (!ni_fatal && best_ni_fatal)
1735 if (ni_healthv < best_healthv)
1737 else if (ni_healthv > best_healthv)
1740 if (ni_sel_prio > best_sel_prio)
1742 else if (ni_sel_prio < best_sel_prio)
1745 if (ni_dev_prio > best_dev_prio)
1747 else if (ni_dev_prio < best_dev_prio)
1750 if (distance > shortest_distance)
1752 else if (distance < shortest_distance)
1755 if (ni_credits < best_credits)
1757 else if (ni_credits > best_credits)
1760 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1764 best_sel_prio = ni_sel_prio;
1765 best_dev_prio = ni_dev_prio;
1766 shortest_distance = distance;
1767 best_healthv = ni_healthv;
1769 best_credits = ni_credits;
1770 best_ni_fatal = ni_fatal;
1773 CDEBUG(D_NET, "selected best_ni %s\n",
1774 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1780 lnet_reserved_msg(struct lnet_msg *msg)
1782 if (msg->msg_type == LNET_MSG_PUT) {
1783 if (msg->msg_hdr.msg.put.ptl_index == LNET_RESERVED_PORTAL)
1785 } else if (msg->msg_type == LNET_MSG_GET) {
1786 if (msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1792 /* Can the specified message trigger peer discovery?
1794 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1795 * because such traffic is required to perform discovery. We therefore
1796 * exclude all GET and PUT on that portal. We also exclude all ACK and
1797 * REPLY traffic, but that is because the portal is not tracked in the
1798 * message structure for these message types. We could restrict this
1799 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1802 lnet_msg_discovery(struct lnet_msg *msg)
1804 return !(lnet_reserved_msg(msg) || lnet_msg_is_response(msg));
1807 /* Is the specified message an LNet ping?
1810 lnet_msg_is_ping(struct lnet_msg *msg)
1812 if (msg->msg_type == LNET_MSG_GET &&
1813 msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1819 #define SRC_SPEC 0x0001
1820 #define SRC_ANY 0x0002
1821 #define LOCAL_DST 0x0004
1822 #define REMOTE_DST 0x0008
1823 #define MR_DST 0x0010
1824 #define NMR_DST 0x0020
1825 #define SND_RESP 0x0040
1827 /* The following to defines are used for return codes */
1828 #define REPEAT_SEND 0x1000
1829 #define PASS_THROUGH 0x2000
1831 /* The different cases lnet_select pathway needs to handle */
1832 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1833 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1834 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1835 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1836 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1837 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1838 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1839 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1842 lnet_handle_lo_send(struct lnet_send_data *sd)
1844 struct lnet_msg *msg = sd->sd_msg;
1845 int cpt = sd->sd_cpt;
1847 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1850 /* No send credit hassles with LOLND */
1851 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1852 msg->msg_hdr.dest_nid = the_lnet.ln_loni->ni_nid;
1853 if (!msg->msg_routing)
1854 msg->msg_hdr.src_nid = the_lnet.ln_loni->ni_nid;
1855 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1856 lnet_msg_commit(msg, cpt);
1857 msg->msg_txni = the_lnet.ln_loni;
1859 return LNET_CREDIT_OK;
1863 lnet_handle_send(struct lnet_send_data *sd)
1865 struct lnet_ni *best_ni = sd->sd_best_ni;
1866 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1867 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1868 struct lnet_msg *msg = sd->sd_msg;
1870 __u32 send_case = sd->sd_send_case;
1872 __u32 routing = send_case & REMOTE_DST;
1873 struct lnet_rsp_tracker *rspt;
1875 /* Increment sequence number of the selected peer, peer net,
1876 * local ni and local net so that we pick the next ones
1879 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1880 best_lpni->lpni_peer_net->lpn_seq =
1881 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1882 best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1883 the_lnet.ln_net_seq++;
1884 best_ni->ni_net->net_seq = the_lnet.ln_net_seq;
1885 best_ni->ni_seq = best_ni->ni_net->net_seq;
1887 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1888 libcfs_nidstr(&best_ni->ni_nid),
1889 best_ni->ni_seq, best_ni->ni_net->net_seq,
1890 atomic_read(&best_ni->ni_tx_credits),
1891 best_ni->ni_sel_priority,
1892 libcfs_nidstr(&best_lpni->lpni_nid),
1893 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1894 best_lpni->lpni_txcredits,
1895 best_lpni->lpni_sel_priority);
1898 * grab a reference on the peer_ni so it sticks around even if
1899 * we need to drop and relock the lnet_net_lock below.
1901 lnet_peer_ni_addref_locked(best_lpni);
1904 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1905 * message. This ensures that we get a CPT that is correct for
1906 * the NI when the NI has been restricted to a subset of all CPTs.
1907 * If the selected CPT differs from the one currently locked, we
1908 * must unlock and relock the lnet_net_lock(), and then check whether
1909 * the configuration has changed. We don't have a hold on the best_ni
1910 * yet, and it may have vanished.
1912 cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1913 if (sd->sd_cpt != cpt2) {
1914 __u32 seq = lnet_get_dlc_seq_locked();
1915 lnet_net_unlock(sd->sd_cpt);
1917 lnet_net_lock(sd->sd_cpt);
1918 if (seq != lnet_get_dlc_seq_locked()) {
1919 lnet_peer_ni_decref_locked(best_lpni);
1925 * store the best_lpni in the message right away to avoid having
1926 * to do the same operation under different conditions
1928 msg->msg_txpeer = best_lpni;
1929 msg->msg_txni = best_ni;
1932 * grab a reference for the best_ni since now it's in use in this
1933 * send. The reference will be dropped in lnet_finalize()
1935 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1938 * Always set the target.nid to the best peer picked. Either the
1939 * NID will be one of the peer NIDs selected, or the same NID as
1940 * what was originally set in the target or it will be the NID of
1941 * a router if this message should be routed
1943 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1946 * lnet_msg_commit assigns the correct cpt to the message, which
1947 * is used to decrement the correct refcount on the ni when it's
1948 * time to return the credits
1950 lnet_msg_commit(msg, sd->sd_cpt);
1953 * If we are routing the message then we keep the src_nid that was
1954 * set by the originator. If we are not routing then we are the
1955 * originator and set it here.
1957 if (!msg->msg_routing)
1958 msg->msg_hdr.src_nid = msg->msg_txni->ni_nid;
1961 msg->msg_target_is_router = 1;
1962 msg->msg_target.pid = LNET_PID_LUSTRE;
1964 * since we're routing we want to ensure that the
1965 * msg_hdr.dest_nid is set to the final destination. When
1966 * the router receives this message it knows how to route
1969 * final_dst_lpni is set at the beginning of the
1970 * lnet_select_pathway() function and is never changed.
1971 * It's safe to use it here.
1973 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1974 final_dst_lpni->lpni_peer_net->lpn_seq =
1975 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1976 final_dst_lpni->lpni_seq =
1977 final_dst_lpni->lpni_peer_net->lpn_seq;
1978 msg->msg_hdr.dest_nid = final_dst_lpni->lpni_nid;
1981 * if we're not routing set the dest_nid to the best peer
1982 * ni NID that we picked earlier in the algorithm.
1984 msg->msg_hdr.dest_nid = msg->msg_txpeer->lpni_nid;
1988 * if we have response tracker block update it with the next hop
1992 rspt = msg->msg_md->md_rspt_ptr;
1994 rspt->rspt_next_hop_nid =
1995 msg->msg_txpeer->lpni_nid;
1996 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1997 libcfs_nidstr(&rspt->rspt_next_hop_nid));
2001 rc = lnet_post_send_locked(msg, 0);
2004 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
2005 libcfs_nidstr(&msg->msg_hdr.src_nid),
2006 libcfs_nidstr(&msg->msg_txni->ni_nid),
2007 libcfs_nidstr(&sd->sd_src_nid),
2008 libcfs_nidstr(&msg->msg_hdr.dest_nid),
2009 libcfs_nidstr(&sd->sd_dst_nid),
2010 libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
2011 libcfs_nidstr(&sd->sd_rtr_nid),
2012 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
2018 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
2019 struct lnet_msg *msg)
2021 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
2022 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
2023 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
2024 libcfs_nidstr(&lni->ni_nid),
2025 libcfs_nidstr(&lpni->lpni_nid));
2026 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
2035 * use the source and destination NIDs as the pathway
2038 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2040 /* the destination lpni is set before we get here. */
2043 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2044 if (!sd->sd_best_ni) {
2045 CERROR("Can't send to %s: src %s is not a local nid\n",
2046 libcfs_nidstr(&sd->sd_dst_nid),
2047 libcfs_nidstr(&sd->sd_src_nid));
2051 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2053 return lnet_handle_send(sd);
2061 * Don't run the selection algorithm on the peer NIs. By specifying the
2062 * local NID, we're also saying that we should always use the destination NID
2063 * provided. This handles the case where we should be using the same
2064 * destination NID for the all the messages which belong to the same RPC
2068 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2070 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2071 if (!sd->sd_best_ni) {
2072 CERROR("Can't send to %s: src %s is not a local nid\n",
2073 libcfs_nidstr(&sd->sd_dst_nid),
2074 libcfs_nidstr(&sd->sd_src_nid));
2078 if (sd->sd_best_lpni &&
2079 nid_same(&sd->sd_best_lpni->lpni_nid,
2080 &the_lnet.ln_loni->ni_nid))
2081 return lnet_handle_lo_send(sd);
2082 else if (sd->sd_best_lpni)
2083 return lnet_handle_send(sd);
2085 CERROR("can't send to %s. no NI on %s\n",
2086 libcfs_nidstr(&sd->sd_dst_nid),
2087 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2089 return -EHOSTUNREACH;
2092 static struct lnet_ni *
2093 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2094 struct lnet_peer *peer,
2095 struct lnet_peer_net *peer_net,
2096 struct lnet_msg *msg,
2099 struct lnet_net *local_net;
2100 struct lnet_ni *best_ni;
2102 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2107 * Iterate through the NIs in this local Net and select
2108 * the NI to send from. The selection is determined by
2109 * these 3 criterion in the following priority:
2111 * 2. NI available credits
2114 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2115 peer, peer_net, msg, cpt);
2121 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2124 struct lnet_peer *peer;
2125 struct lnet_peer_ni *new_lpni;
2128 lnet_peer_ni_addref_locked(lpni);
2130 peer = lpni->lpni_peer_net->lpn_peer;
2132 if (lnet_peer_gw_discovery(peer)) {
2133 lnet_peer_ni_decref_locked(lpni);
2137 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2138 lnet_peer_ni_decref_locked(lpni);
2142 rc = lnet_discover_peer_locked(lpni, cpt, false);
2144 lnet_peer_ni_decref_locked(lpni);
2148 new_lpni = lnet_peer_ni_find_locked(&lpni->lpni_nid);
2150 lnet_peer_ni_decref_locked(lpni);
2154 peer = new_lpni->lpni_peer_net->lpn_peer;
2155 spin_lock(&peer->lp_lock);
2156 if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2157 /* The peer NI did not change and the peer is up to date.
2158 * Nothing more to do.
2160 spin_unlock(&peer->lp_lock);
2161 lnet_peer_ni_decref_locked(lpni);
2162 lnet_peer_ni_decref_locked(new_lpni);
2165 spin_unlock(&peer->lp_lock);
2167 /* Either the peer NI changed during discovery, or the peer isn't up
2168 * to date. In both cases we want to queue the message on the
2169 * (possibly new) peer's pending queue and queue the peer for discovery
2171 msg->msg_sending = 0;
2172 msg->msg_txpeer = NULL;
2173 lnet_net_unlock(cpt);
2174 lnet_peer_queue_message(peer, msg);
2177 lnet_peer_ni_decref_locked(lpni);
2178 lnet_peer_ni_decref_locked(new_lpni);
2180 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2181 msg, libcfs_nidstr(&peer->lp_primary_nid));
2183 return LNET_DC_WAIT;
2187 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2188 struct lnet_nid *dst_nid,
2189 struct lnet_peer_ni **gw_lpni,
2190 struct lnet_peer **gw_peer)
2193 struct lnet_peer *gw;
2194 struct lnet_peer *lp;
2195 struct lnet_peer_net *lpn;
2196 struct lnet_peer_net *best_lpn = NULL;
2197 struct lnet_remotenet *rnet, *best_rnet = NULL;
2198 struct lnet_route *best_route = NULL;
2199 struct lnet_route *last_route = NULL;
2200 struct lnet_peer_ni *lpni = NULL;
2201 struct lnet_peer_ni *gwni = NULL;
2202 bool route_found = false;
2203 struct lnet_nid *src_nid =
2204 !LNET_NID_IS_ANY(&sd->sd_src_nid) || !sd->sd_best_ni
2206 : &sd->sd_best_ni->ni_nid;
2207 int best_lpn_healthv = 0;
2208 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2210 CDEBUG(D_NET, "%s route (%s) from local NI %s to destination %s\n",
2211 LNET_NID_IS_ANY(&sd->sd_rtr_nid) ? "Lookup" : "Specified",
2212 libcfs_nidstr(&sd->sd_rtr_nid), libcfs_nidstr(src_nid),
2213 libcfs_nidstr(&sd->sd_dst_nid));
2215 /* If a router nid was specified then we are replying to a GET or
2216 * sending an ACK. In this case we use the gateway associated with the
2217 * specified router nid.
2219 if (!LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2220 gwni = lnet_peer_ni_find_locked(&sd->sd_rtr_nid);
2222 gw = gwni->lpni_peer_net->lpn_peer;
2223 lnet_peer_ni_decref_locked(gwni);
2224 if (gw->lp_rtr_refcount)
2227 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2228 libcfs_nidstr(&sd->sd_rtr_nid));
2233 if (sd->sd_msg->msg_routing || !LNET_NID_IS_ANY(src_nid)) {
2234 /* If I'm routing this message then I need to find the
2235 * next hop based on the destination NID
2237 * We also find next hop based on the destination NID
2238 * if the source NI was specified
2240 best_rnet = lnet_find_rnet_locked(LNET_NID_NET(&sd->sd_dst_nid));
2242 CERROR("Unable to send message from %s to %s - Route table may be misconfigured\n",
2243 (src_nid && LNET_NID_IS_ANY(src_nid)) ?
2245 libcfs_nidstr(src_nid),
2246 libcfs_nidstr(&sd->sd_dst_nid));
2247 return -EHOSTUNREACH;
2249 CDEBUG(D_NET, "best_rnet %s\n",
2250 libcfs_net2str(best_rnet->lrn_net));
2252 /* we've already looked up the initial lpni using
2255 lpni = sd->sd_best_lpni;
2256 /* the peer tree must be in existence */
2257 LASSERT(lpni && lpni->lpni_peer_net &&
2258 lpni->lpni_peer_net->lpn_peer);
2259 lp = lpni->lpni_peer_net->lpn_peer;
2261 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2262 /* is this remote network reachable? */
2263 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2270 CDEBUG(D_NET, "n[%s, %s] h[%d, %d], p[%u, %u], s[%d, %d]\n",
2271 libcfs_net2str(lpn->lpn_net_id),
2272 libcfs_net2str(best_lpn->lpn_net_id),
2274 best_lpn->lpn_healthv,
2275 lpn->lpn_sel_priority,
2276 best_lpn->lpn_sel_priority,
2280 /* select the preferred peer net */
2281 if (best_lpn_healthv > lpn->lpn_healthv)
2283 else if (best_lpn_healthv < lpn->lpn_healthv)
2286 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2288 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2291 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2294 best_lpn_healthv = lpn->lpn_healthv;
2295 best_lpn_sel_prio = lpn->lpn_sel_priority;
2301 CERROR("peer %s has no available nets\n",
2302 libcfs_nidstr(&sd->sd_dst_nid));
2303 return -EHOSTUNREACH;
2306 CDEBUG(D_NET, "selected best_lpn %s\n",
2307 libcfs_net2str(best_lpn->lpn_net_id));
2309 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2312 best_lpn->lpn_net_id);
2313 if (!sd->sd_best_lpni) {
2314 CERROR("peer %s is unreachable\n",
2315 libcfs_nidstr(&sd->sd_dst_nid));
2316 return -EHOSTUNREACH;
2319 /* We're attempting to round robin over the remote peer
2320 * NI's so update the final destination we selected
2322 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2326 * find the best route. Restrict the selection on the net of the
2327 * local NI if we've already picked the local NI to send from.
2328 * Otherwise, let's pick any route we can find and then find
2329 * a local NI we can reach the route's gateway on. Any route we
2330 * select will be reachable by virtue of the restriction we have
2331 * when adding a route.
2333 best_route = lnet_find_route_locked(best_rnet,
2334 LNET_NID_NET(src_nid),
2336 &last_route, &gwni);
2339 CERROR("no route to %s from %s\n",
2340 libcfs_nidstr(dst_nid),
2341 libcfs_nidstr(src_nid));
2342 return -EHOSTUNREACH;
2346 CERROR("Internal Error. Route expected to %s from %s\n",
2347 libcfs_nidstr(dst_nid),
2348 libcfs_nidstr(src_nid));
2352 gw = best_route->lr_gateway;
2353 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2357 * If the router checker is not active then discover the gateway here.
2358 * This ensures we are able to take advantage of multi-rail routing, but
2359 * if the router checker is active then we do not unecessarily delay
2360 * messages while the gateway is being checked by the dedicated monitor
2363 * NB: We're only checking the alive_router_check_interval here, rather
2364 * than calling lnet_router_checker_active(), because the other
2365 * conditions that are checked by that function are either
2366 * irrelevant (the_lnet.ln_routing) or must be true (list of routers
2369 if (alive_router_check_interval <= 0) {
2370 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2375 if (!sd->sd_best_ni) {
2376 lpn = gwni->lpni_peer_net;
2377 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2380 if (!sd->sd_best_ni) {
2381 CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2382 libcfs_net2str(lpn->lpn_net_id),
2383 libcfs_nidstr(&sd->sd_src_nid));
2392 * increment the sequence number since now we're sure we're
2393 * going to use this route
2395 if (LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2396 LASSERT(best_route && last_route);
2397 best_route->lr_seq = last_route->lr_seq + 1;
2408 * Remote destination
2409 * Non-MR destination
2413 * Remote destination
2416 * The handling of these two cases is similar. Even though the destination
2417 * can be MR or non-MR, we'll deal directly with the router.
2420 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2423 struct lnet_peer_ni *gw_lpni = NULL;
2424 struct lnet_peer *gw_peer = NULL;
2427 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2428 if (!sd->sd_best_ni) {
2429 CERROR("Can't send to %s: src %s is not a local nid\n",
2430 libcfs_nidstr(&sd->sd_dst_nid),
2431 libcfs_nidstr(&sd->sd_src_nid));
2435 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2436 &gw_lpni, &gw_peer);
2440 if (sd->sd_send_case & NMR_DST)
2442 * since the final destination is non-MR let's set its preferred
2443 * NID before we send
2445 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2449 * We're going to send to the gw found so let's set its
2452 sd->sd_peer = gw_peer;
2453 sd->sd_best_lpni = gw_lpni;
2455 return lnet_handle_send(sd);
2458 static struct lnet_ni *
2459 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2460 struct lnet_msg *msg, bool discovery)
2462 struct lnet_peer_net *lpn = NULL;
2463 struct lnet_peer_net *best_lpn = NULL;
2464 struct lnet_net *net = NULL;
2465 struct lnet_net *best_net = NULL;
2466 struct lnet_ni *best_ni = NULL;
2467 int best_lpn_healthv = 0;
2468 int best_net_healthv = 0;
2470 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2472 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2475 /* If lp_disc_net_id is set, this peer is a router undergoing
2476 * discovery, and this message is an LNet ping, then this may be a
2477 * discovery message and we need to select an NI on the peer net
2478 * specified by lp_disc_net_id
2480 if (peer->lp_disc_net_id &&
2481 (peer->lp_state & LNET_PEER_RTR_DISCOVERY) &&
2482 lnet_msg_is_ping(msg)) {
2483 best_lpn = lnet_peer_get_net_locked(peer, peer->lp_disc_net_id);
2484 if (best_lpn && lnet_get_net_locked(best_lpn->lpn_net_id))
2485 goto select_best_ni;
2489 * The peer can have multiple interfaces, some of them can be on
2490 * the local network and others on a routed network. We should
2491 * prefer the local network. However if the local network is not
2492 * available then we need to try the routed network
2495 /* go through all the peer nets and find the best_ni */
2496 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2498 * The peer's list of nets can contain non-local nets. We
2499 * want to only examine the local ones.
2501 net = lnet_get_net_locked(lpn->lpn_net_id);
2505 lpn_sel_prio = lpn->lpn_sel_priority;
2506 net_healthv = lnet_get_net_healthv_locked(net);
2507 net_sel_prio = net->net_sel_priority;
2513 "n[%s, %s] ph[%d, %d], pp[%u, %u], nh[%d, %d], np[%u, %u], ps[%u, %u], ns[%u, %u]\n",
2514 libcfs_net2str(lpn->lpn_net_id),
2515 libcfs_net2str(best_lpn->lpn_net_id),
2529 /* always select the lpn with the best health */
2530 if (best_lpn_healthv > lpn->lpn_healthv)
2532 else if (best_lpn_healthv < lpn->lpn_healthv)
2535 /* select the preferred peer and local nets */
2536 if (best_lpn_sel_prio < lpn_sel_prio)
2538 else if (best_lpn_sel_prio > lpn_sel_prio)
2541 if (best_net_healthv > net_healthv)
2543 else if (best_net_healthv < net_healthv)
2546 if (best_net_sel_prio < net_sel_prio)
2548 else if (best_net_sel_prio > net_sel_prio)
2551 if (best_lpn->lpn_seq < lpn->lpn_seq)
2553 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2556 /* round robin over the local networks */
2557 if (best_net->net_seq <= net->net_seq)
2561 best_net_healthv = net_healthv;
2562 best_net_sel_prio = net_sel_prio;
2563 best_lpn_healthv = lpn->lpn_healthv;
2564 best_lpn_sel_prio = lpn_sel_prio;
2570 /* Select the best NI on the same net as best_lpn chosen
2574 CDEBUG(D_NET, "selected best_lpn %s\n",
2575 libcfs_net2str(best_lpn->lpn_net_id));
2576 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2583 static struct lnet_ni *
2584 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2586 struct lnet_ni *best_ni = NULL;
2587 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2588 struct lnet_peer_ni *lpni_entry;
2591 * We must use a consistent source address when sending to a
2592 * non-MR peer. However, a non-MR peer can have multiple NIDs
2593 * on multiple networks, and we may even need to talk to this
2594 * peer on multiple networks -- certain types of
2595 * load-balancing configuration do this.
2597 * So we need to pick the NI the peer prefers for this
2598 * particular network.
2601 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2603 if (lpni_entry->lpni_pref_nnids == 0)
2605 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2606 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2614 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2616 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2618 struct lnet_ni *best_ni = NULL;
2621 * We must use a consistent source address when sending to a
2622 * non-MR peer. However, a non-MR peer can have multiple NIDs
2623 * on multiple networks, and we may even need to talk to this
2624 * peer on multiple networks -- certain types of
2625 * load-balancing configuration do this.
2627 * So we need to pick the NI the peer prefers for this
2628 * particular network.
2630 * An exception is traffic on LNET_RESERVED_PORTAL. Internal LNet
2631 * traffic doesn't care which source NI is used, and we don't actually
2632 * want to restrict local recovery pings to a single source NI.
2634 if (!lnet_reserved_msg(sd->sd_msg))
2635 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2639 best_ni = lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2640 sd->sd_best_lpni->lpni_peer_net,
2644 /* If there is no best_ni we don't have a route */
2646 CERROR("no path to %s from net %s\n",
2647 libcfs_nidstr(&sd->sd_best_lpni->lpni_nid),
2648 libcfs_net2str(sd->sd_best_lpni->lpni_net->net_id));
2649 return -EHOSTUNREACH;
2652 sd->sd_best_ni = best_ni;
2654 /* Set preferred NI if necessary. */
2655 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2662 * Source not specified
2666 * always use the same source NID for NMR peers
2667 * If we've talked to that peer before then we already have a preferred
2668 * source NI associated with it. Otherwise, we select a preferred local NI
2669 * and store it in the peer
2672 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2676 /* sd->sd_best_lpni is already set to the final destination */
2679 * At this point we should've created the peer ni and peer. If we
2680 * can't find it, then something went wrong. Instead of assert
2681 * output a relevant message and fail the send
2683 if (!sd->sd_best_lpni) {
2684 CERROR("Internal fault. Unable to send msg %s to %s. NID not known\n",
2685 lnet_msgtyp2str(sd->sd_msg->msg_type),
2686 libcfs_nidstr(&sd->sd_dst_nid));
2690 if (sd->sd_msg->msg_routing) {
2691 /* If I'm forwarding this message then I can choose any NI
2692 * on the destination peer net
2694 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2696 sd->sd_best_lpni->lpni_peer_net,
2699 if (!sd->sd_best_ni) {
2700 CERROR("Unable to forward message to %s. No local NI available\n",
2701 libcfs_nidstr(&sd->sd_dst_nid));
2705 rc = lnet_select_preferred_best_ni(sd);
2708 rc = lnet_handle_send(sd);
2714 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2717 * NOTE we've already handled the remote peer case. So we only
2718 * need to worry about the local case here.
2720 * if we're sending a response, ACK or reply, we need to send it
2721 * to the destination NID given to us. At this point we already
2722 * have the peer_ni we're suppose to send to, so just find the
2723 * best_ni on the peer net and use that. Since we're sending to an
2724 * MR peer then we can just run the selection algorithm on our
2725 * local NIs and pick the best one.
2727 if (sd->sd_send_case & SND_RESP) {
2729 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2730 sd->sd_best_lpni->lpni_peer_net,
2734 if (!sd->sd_best_ni) {
2736 * We're not going to deal with not able to send
2737 * a response to the provided final destination
2739 CERROR("Can't send response to %s. No local NI available\n",
2740 libcfs_nidstr(&sd->sd_dst_nid));
2741 return -EHOSTUNREACH;
2744 return lnet_handle_send(sd);
2748 * If we get here that means we're sending a fresh request, PUT or
2749 * GET, so we need to run our standard selection algorithm.
2750 * First find the best local interface that's on any of the peer's
2753 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2756 lnet_msg_discovery(sd->sd_msg));
2757 if (sd->sd_best_ni) {
2759 lnet_find_best_lpni(sd->sd_best_ni, &sd->sd_dst_nid,
2761 sd->sd_best_ni->ni_net->net_id);
2764 * if we're successful in selecting a peer_ni on the local
2765 * network, then send to it. Otherwise fall through and
2766 * try and see if we can reach it over another routed
2769 if (sd->sd_best_lpni &&
2770 nid_same(&sd->sd_best_lpni->lpni_nid,
2771 &the_lnet.ln_loni->ni_nid)) {
2773 * in case we initially started with a routed
2774 * destination, let's reset to local
2776 sd->sd_send_case &= ~REMOTE_DST;
2777 sd->sd_send_case |= LOCAL_DST;
2778 return lnet_handle_lo_send(sd);
2779 } else if (sd->sd_best_lpni) {
2781 * in case we initially started with a routed
2782 * destination, let's reset to local
2784 sd->sd_send_case &= ~REMOTE_DST;
2785 sd->sd_send_case |= LOCAL_DST;
2786 return lnet_handle_send(sd);
2789 CERROR("Internal Error. Expected to have a best_lpni: "
2791 libcfs_nidstr(&sd->sd_src_nid),
2792 libcfs_nidstr(&sd->sd_dst_nid));
2798 * Peer doesn't have a local network. Let's see if there is
2799 * a remote network we can reach it on.
2801 return PASS_THROUGH;
2806 * Source NID not specified
2811 * Source NID not speified
2812 * Remote destination
2815 * In both of these cases if we're sending a response, ACK or REPLY, then
2816 * we need to send to the destination NID provided.
2818 * In the remote case let's deal with MR routers.
2823 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2826 struct lnet_peer *gw_peer = NULL;
2827 struct lnet_peer_ni *gw_lpni = NULL;
2830 * handle sending a response to a remote peer here so we don't
2831 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2833 if (sd->sd_send_case & REMOTE_DST &&
2834 sd->sd_send_case & SND_RESP) {
2835 struct lnet_peer_ni *gw;
2836 struct lnet_peer *gw_peer;
2838 rc = lnet_handle_find_routed_path(
2839 sd, &sd->sd_dst_nid, &gw, &gw_peer);
2841 CERROR("Can't send response to %s. No route available\n",
2842 libcfs_nidstr(&sd->sd_dst_nid));
2843 return -EHOSTUNREACH;
2844 } else if (rc > 0) {
2848 sd->sd_best_lpni = gw;
2849 sd->sd_peer = gw_peer;
2851 return lnet_handle_send(sd);
2855 * Even though the NID for the peer might not be on a local network,
2856 * since the peer is MR there could be other interfaces on the
2857 * local network. In that case we'd still like to prefer the local
2858 * network over the routed network. If we're unable to do that
2859 * then we select the best router among the different routed networks,
2860 * and if the router is MR then we can deal with it as such.
2862 rc = lnet_handle_any_mr_dsta(sd);
2863 if (rc != PASS_THROUGH)
2867 * Now that we must route to the destination, we must consider the
2868 * MR case, where the destination has multiple interfaces, some of
2869 * which we can route to and others we do not. For this reason we
2870 * need to select the destination which we can route to and if
2871 * there are multiple, we need to round robin.
2873 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2874 &gw_lpni, &gw_peer);
2878 sd->sd_send_case &= ~LOCAL_DST;
2879 sd->sd_send_case |= REMOTE_DST;
2881 sd->sd_peer = gw_peer;
2882 sd->sd_best_lpni = gw_lpni;
2884 return lnet_handle_send(sd);
2888 * Source not specified
2889 * Remote destination
2892 * Must send to the specified peer NID using the same source NID that
2893 * we've used before. If it's the first time to talk to that peer then
2894 * find the source NI and assign it as preferred to that peer
2897 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2900 struct lnet_peer_ni *gw_lpni = NULL;
2901 struct lnet_peer *gw_peer = NULL;
2904 * Let's see if we have a preferred NI to talk to this NMR peer
2906 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2910 * find the router and that'll find the best NI if we didn't find
2913 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid, &gw_lpni,
2919 * set the best_ni we've chosen as the preferred one for
2922 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2924 /* we'll be sending to the gw */
2925 sd->sd_best_lpni = gw_lpni;
2926 sd->sd_peer = gw_peer;
2928 return lnet_handle_send(sd);
2932 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2935 * turn off the SND_RESP bit.
2936 * It will be checked in the case handling
2938 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2940 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2941 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2942 (send_case & SRC_SPEC) ? libcfs_nidstr(&sd->sd_src_nid) : "",
2943 (send_case & MR_DST) ? "MR: " : "NMR: ",
2944 libcfs_nidstr(&sd->sd_dst_nid),
2945 (send_case & LOCAL_DST) ? "local" : "routed");
2947 switch (send_case) {
2949 * For all cases where the source is specified, we should always
2950 * use the destination NID, whether it's an MR destination or not,
2951 * since we're continuing a series of related messages for the
2954 case SRC_SPEC_LOCAL_NMR_DST:
2955 return lnet_handle_spec_local_nmr_dst(sd);
2956 case SRC_SPEC_LOCAL_MR_DST:
2957 return lnet_handle_spec_local_mr_dst(sd);
2958 case SRC_SPEC_ROUTER_NMR_DST:
2959 case SRC_SPEC_ROUTER_MR_DST:
2960 return lnet_handle_spec_router_dst(sd);
2961 case SRC_ANY_LOCAL_NMR_DST:
2962 return lnet_handle_any_local_nmr_dst(sd);
2963 case SRC_ANY_LOCAL_MR_DST:
2964 case SRC_ANY_ROUTER_MR_DST:
2965 return lnet_handle_any_mr_dst(sd);
2966 case SRC_ANY_ROUTER_NMR_DST:
2967 return lnet_handle_any_router_nmr_dst(sd);
2969 CERROR("Unknown send case\n");
2975 lnet_select_pathway(struct lnet_nid *src_nid,
2976 struct lnet_nid *dst_nid,
2977 struct lnet_msg *msg,
2978 struct lnet_nid *rtr_nid)
2980 struct lnet_peer_ni *lpni;
2981 struct lnet_peer *peer;
2982 struct lnet_send_data send_data;
2985 __u32 send_case = 0;
2987 bool mr_forwarding_allowed;
2989 memset(&send_data, 0, sizeof(send_data));
2992 * get an initial CPT to use for locking. The idea here is not to
2993 * serialize the calls to select_pathway, so that as many
2994 * operations can run concurrently as possible. To do that we use
2995 * the CPT where this call is being executed. Later on when we
2996 * determine the CPT to use in lnet_message_commit, we switch the
2997 * lock and check if there was any configuration change. If none,
2998 * then we proceed, if there is, then we restart the operation.
3000 cpt = lnet_net_lock_current();
3002 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
3003 if (md_cpt == CFS_CPT_ANY)
3009 * If we're being asked to send to the loopback interface, there
3010 * is no need to go through any selection. We can just shortcut
3011 * the entire process and send over lolnd
3013 send_data.sd_msg = msg;
3014 send_data.sd_cpt = cpt;
3015 if (nid_is_lo0(dst_nid)) {
3016 rc = lnet_handle_lo_send(&send_data);
3017 lnet_net_unlock(cpt);
3022 * find an existing peer_ni, or create one and mark it as having been
3023 * created due to network traffic. This call will create the
3024 * peer->peer_net->peer_ni tree.
3026 lpni = lnet_peerni_by_nid_locked(dst_nid, NULL, cpt);
3028 lnet_net_unlock(cpt);
3029 return PTR_ERR(lpni);
3033 * Cache the original src_nid and rtr_nid. If we need to resend the
3034 * message then we'll need to know whether the src_nid was originally
3035 * specified for this message. If it was originally specified,
3036 * then we need to keep using the same src_nid since it's
3037 * continuing the same sequence of messages. Similarly, rtr_nid will
3038 * affect our choice of next hop.
3041 msg->msg_src_nid_param = *src_nid;
3043 msg->msg_src_nid_param = LNET_ANY_NID;
3045 msg->msg_rtr_nid_param = *rtr_nid;
3047 msg->msg_rtr_nid_param = LNET_ANY_NID;
3050 * If necessary, perform discovery on the peer that owns this peer_ni.
3051 * Note, this can result in the ownership of this peer_ni changing
3052 * to another peer object.
3054 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
3056 lnet_peer_ni_decref_locked(lpni);
3057 lnet_net_unlock(cpt);
3060 lnet_peer_ni_decref_locked(lpni);
3062 peer = lpni->lpni_peer_net->lpn_peer;
3065 * Identify the different send cases
3067 if (!src_nid || LNET_NID_IS_ANY(src_nid)) {
3068 send_case |= SRC_ANY;
3069 if (lnet_get_net_locked(LNET_NID_NET(dst_nid)))
3070 send_case |= LOCAL_DST;
3072 send_case |= REMOTE_DST;
3074 send_case |= SRC_SPEC;
3075 if (LNET_NID_NET(src_nid) == LNET_NID_NET(dst_nid))
3076 send_case |= LOCAL_DST;
3078 send_case |= REMOTE_DST;
3082 if (msg->msg_routing && (send_case & LOCAL_DST))
3085 /* Determine whether to allow MR forwarding for this message.
3086 * NB: MR forwarding is allowed if the message originator and the
3087 * destination are both MR capable, and the destination lpni that was
3088 * originally chosen by the originator is unhealthy or down.
3089 * We check the MR capability of the destination further below
3091 mr_forwarding_allowed = false;
3093 struct lnet_peer *src_lp;
3094 struct lnet_peer_ni *src_lpni;
3096 src_lpni = lnet_peerni_by_nid_locked(&msg->msg_hdr.src_nid,
3098 /* We don't fail the send if we hit any errors here. We'll just
3099 * try to send it via non-multi-rail criteria
3101 if (!IS_ERR(src_lpni)) {
3102 /* Drop ref taken by lnet_peerni_by_nid_locked() */
3103 lnet_peer_ni_decref_locked(src_lpni);
3104 src_lp = lpni->lpni_peer_net->lpn_peer;
3105 if (lnet_peer_is_multi_rail(src_lp) &&
3106 !lnet_is_peer_ni_alive(lpni))
3107 mr_forwarding_allowed = true;
3110 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3111 mr_forwarding_allowed ? "allowed" : "not allowed");
3115 * Deal with the peer as NMR in the following cases:
3116 * 1. the peer is NMR
3117 * 2. We're trying to recover a specific peer NI
3118 * 3. I'm a router sending to the final destination and MR forwarding is
3119 * not allowed for this message (as determined above).
3120 * In this case the source of the message would've
3121 * already selected the final destination so my job
3122 * is to honor the selection.
3124 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3125 (final_hop && !mr_forwarding_allowed))
3126 send_case |= NMR_DST;
3128 send_case |= MR_DST;
3130 if (lnet_msg_is_response(msg))
3131 send_case |= SND_RESP;
3133 /* assign parameters to the send_data */
3135 send_data.sd_rtr_nid = *rtr_nid;
3137 send_data.sd_rtr_nid = LNET_ANY_NID;
3139 send_data.sd_src_nid = *src_nid;
3141 send_data.sd_src_nid = LNET_ANY_NID;
3142 send_data.sd_dst_nid = *dst_nid;
3143 send_data.sd_best_lpni = lpni;
3145 * keep a pointer to the final destination in case we're going to
3146 * route, so we'll need to access it later
3148 send_data.sd_final_dst_lpni = lpni;
3149 send_data.sd_peer = peer;
3150 send_data.sd_md_cpt = md_cpt;
3151 send_data.sd_send_case = send_case;
3153 rc = lnet_handle_send_case_locked(&send_data);
3156 * Update the local cpt since send_data.sd_cpt might've been
3157 * updated as a result of calling lnet_handle_send_case_locked().
3159 cpt = send_data.sd_cpt;
3161 if (rc == REPEAT_SEND)
3164 lnet_net_unlock(cpt);
3170 lnet_send(struct lnet_nid *src_nid, struct lnet_msg *msg,
3171 struct lnet_nid *rtr_nid)
3173 struct lnet_nid *dst_nid = &msg->msg_target.nid;
3176 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3177 LASSERT(msg->msg_txpeer == NULL);
3178 LASSERT(msg->msg_txni == NULL);
3179 LASSERT(!msg->msg_sending);
3180 LASSERT(!msg->msg_target_is_router);
3181 LASSERT(!msg->msg_receiving);
3183 msg->msg_sending = 1;
3185 LASSERT(!msg->msg_tx_committed);
3187 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3189 if (rc == -EHOSTUNREACH)
3190 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3192 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3196 if (rc == LNET_CREDIT_OK)
3197 lnet_ni_send(msg->msg_txni, msg);
3199 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3203 enum lnet_mt_event_type {
3204 MT_TYPE_LOCAL_NI = 0,
3208 struct lnet_mt_event_info {
3209 enum lnet_mt_event_type mt_type;
3210 struct lnet_nid mt_nid;
3213 /* called with res_lock held */
3215 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3217 struct lnet_rsp_tracker *rspt;
3220 * msg has a refcount on the MD so the MD is not going away.
3221 * The rspt queue for the cpt is protected by
3222 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3224 if (!md->md_rspt_ptr)
3227 rspt = md->md_rspt_ptr;
3230 LASSERT(rspt->rspt_cpt == cpt);
3232 md->md_rspt_ptr = NULL;
3234 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3236 * The monitor thread has invalidated this handle because the
3237 * response timed out, but it failed to lookup the MD. That
3238 * means this response tracker is on the zombie list. We can
3239 * safely remove it under the resource lock (held by caller) and
3240 * free the response tracker block.
3242 list_del(&rspt->rspt_on_list);
3243 lnet_rspt_free(rspt, cpt);
3246 * invalidate the handle to indicate that a response has been
3247 * received, which will then lead the monitor thread to clean up
3250 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3255 lnet_clean_zombie_rstqs(void)
3257 struct lnet_rsp_tracker *rspt, *tmp;
3260 cfs_cpt_for_each(i, lnet_cpt_table()) {
3261 list_for_each_entry_safe(rspt, tmp,
3262 the_lnet.ln_mt_zombie_rstqs[i],
3264 list_del(&rspt->rspt_on_list);
3265 lnet_rspt_free(rspt, i);
3269 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3273 lnet_finalize_expired_responses(void)
3275 struct lnet_libmd *md;
3276 struct lnet_rsp_tracker *rspt, *tmp;
3280 if (the_lnet.ln_mt_rstq == NULL)
3283 cfs_cpt_for_each(i, lnet_cpt_table()) {
3284 LIST_HEAD(local_queue);
3287 if (!the_lnet.ln_mt_rstq[i]) {
3291 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3296 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3298 * The rspt mdh will be invalidated when a response
3299 * is received or whenever we want to discard the
3300 * block the monitor thread will walk the queue
3301 * and clean up any rsts with an invalid mdh.
3302 * The monitor thread will walk the queue until
3303 * the first unexpired rspt block. This means that
3304 * some rspt blocks which received their
3305 * corresponding responses will linger in the
3306 * queue until they are cleaned up eventually.
3309 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3311 list_del(&rspt->rspt_on_list);
3312 lnet_rspt_free(rspt, i);
3316 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3317 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3318 struct lnet_peer_ni *lpni;
3319 struct lnet_nid nid;
3321 md = lnet_handle2md(&rspt->rspt_mdh);
3323 /* MD has been queued for unlink, but
3324 * rspt hasn't been detached (Note we've
3325 * checked above that the rspt_mdh is
3326 * valid). Since we cannot lookup the MD
3327 * we're unable to detach the rspt
3328 * ourselves. Thus, move the rspt to the
3329 * zombie list where we'll wait for
3331 * 1. The remaining operations on the
3332 * MD to complete. In this case the
3333 * final operation will result in
3334 * lnet_msg_detach_md()->
3335 * lnet_detach_rsp_tracker() where
3336 * we will clean up this response
3338 * 2. LNet to shutdown. In this case
3339 * we'll wait until after all LND Nets
3340 * have shutdown and then we can
3341 * safely free any remaining response
3342 * tracker blocks on the zombie list.
3343 * Note: We need to hold the resource
3344 * lock when adding to the zombie list
3345 * because we may have concurrent access
3346 * with lnet_detach_rsp_tracker().
3348 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3349 list_move(&rspt->rspt_on_list,
3350 the_lnet.ln_mt_zombie_rstqs[i]);
3354 LASSERT(md->md_rspt_ptr == rspt);
3355 md->md_rspt_ptr = NULL;
3358 LNetMDUnlink(rspt->rspt_mdh);
3360 nid = rspt->rspt_next_hop_nid;
3362 list_del(&rspt->rspt_on_list);
3363 lnet_rspt_free(rspt, i);
3365 /* If we're shutting down we just want to clean
3366 * up the rspt blocks
3368 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3372 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3376 "Response timeout: md = %p: nid = %s\n",
3377 md, libcfs_nidstr(&nid));
3380 * If there is a timeout on the response
3381 * from the next hop decrement its health
3382 * value so that we don't use it
3385 lpni = lnet_peer_ni_find_locked(&nid);
3387 lnet_handle_remote_failure_locked(lpni);
3388 lnet_peer_ni_decref_locked(lpni);
3397 if (!list_empty(&local_queue)) {
3399 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3406 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3408 struct lnet_msg *msg;
3410 while (!list_empty(resendq)) {
3411 struct lnet_peer_ni *lpni;
3413 msg = list_first_entry(resendq, struct lnet_msg,
3416 list_del_init(&msg->msg_list);
3418 lpni = lnet_peer_ni_find_locked(&msg->msg_hdr.dest_nid);
3420 lnet_net_unlock(cpt);
3421 CERROR("Expected that a peer is already created for %s\n",
3422 libcfs_nidstr(&msg->msg_hdr.dest_nid));
3423 msg->msg_no_resend = true;
3424 lnet_finalize(msg, -EFAULT);
3429 lnet_peer_ni_decref_locked(lpni);
3431 lnet_net_unlock(cpt);
3432 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3433 libcfs_nidstr(&msg->msg_src_nid_param),
3434 libcfs_idstr(&msg->msg_target),
3435 lnet_msgtyp2str(msg->msg_type),
3437 msg->msg_retry_count);
3438 rc = lnet_send(&msg->msg_src_nid_param, msg,
3439 &msg->msg_rtr_nid_param);
3441 CERROR("Error sending %s to %s: %d\n",
3442 lnet_msgtyp2str(msg->msg_type),
3443 libcfs_idstr(&msg->msg_target), rc);
3444 msg->msg_no_resend = true;
3445 lnet_finalize(msg, rc);
3449 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3455 lnet_resend_pending_msgs(void)
3459 cfs_cpt_for_each(i, lnet_cpt_table()) {
3461 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3466 /* called with cpt and ni_lock held */
3468 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3470 struct lnet_handle_md recovery_mdh;
3472 LNetInvalidateMDHandle(&recovery_mdh);
3474 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3476 recovery_mdh = ni->ni_ping_mdh;
3477 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3480 lnet_net_unlock(cpt);
3481 if (!LNetMDHandleIsInvalid(recovery_mdh))
3482 LNetMDUnlink(recovery_mdh);
3487 /* Returns the total number of local NIs in recovery.
3488 * Records up to @arrsz of the associated NIDs in the @nidarr array
3491 lnet_recover_local_nis(struct lnet_nid *nidarr, unsigned int arrsz)
3493 struct lnet_mt_event_info *ev_info;
3494 LIST_HEAD(processed_list);
3495 LIST_HEAD(local_queue);
3496 struct lnet_handle_md mdh;
3497 struct lnet_ni *tmp;
3499 struct lnet_nid nid;
3503 unsigned int nnis = 0;
3506 * splice the recovery queue on a local queue. We will iterate
3507 * through the local queue and update it as needed. Once we're
3508 * done with the traversal, we'll splice the local queue back on
3509 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3510 * will be traversed in the next iteration.
3513 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3517 now = ktime_get_seconds();
3519 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3521 * if an NI is being deleted or it is now healthy, there
3522 * is no need to keep it around in the recovery queue.
3523 * The monitor thread is the only thread responsible for
3524 * removing the NI from the recovery queue.
3525 * Multiple threads can be adding NIs to the recovery
3528 healthv = atomic_read(&ni->ni_healthv);
3532 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3533 healthv == LNET_MAX_HEALTH_VALUE) {
3534 list_del_init(&ni->ni_recovery);
3535 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3537 lnet_ni_decref_locked(ni, 0);
3543 nidarr[nnis] = ni->ni_nid;
3547 * if the local NI failed recovery we must unlink the md.
3548 * But we want to keep the local_ni on the recovery queue
3549 * so we can continue the attempts to recover it.
3551 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3552 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3553 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3559 if (now < ni->ni_next_ping) {
3566 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3567 libcfs_nidstr(&ni->ni_nid));
3570 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3571 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3574 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3576 CERROR("out of memory. Can't recover %s\n",
3577 libcfs_nidstr(&ni->ni_nid));
3579 ni->ni_recovery_state &=
3580 ~LNET_NI_RECOVERY_PENDING;
3585 mdh = ni->ni_ping_mdh;
3587 * Invalidate the ni mdh in case it's deleted.
3588 * We'll unlink the mdh in this case below.
3590 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3594 * remove the NI from the local queue and drop the
3595 * reference count to it while we're recovering
3596 * it. The reason for that, is that the NI could
3597 * be deleted, and the way the code is structured
3598 * is if we don't drop the NI, then the deletion
3599 * code will enter a loop waiting for the
3600 * reference count to be removed while holding the
3601 * ln_mutex_lock(). When we look up the peer to
3602 * send to in lnet_select_pathway() we will try to
3603 * lock the ln_mutex_lock() as well, leading to
3604 * a deadlock. By dropping the refcount and
3605 * removing it from the list, we allow for the NI
3606 * to be removed, then we use the cached NID to
3607 * look it up again. If it's gone, then we just
3608 * continue examining the rest of the queue.
3611 list_del_init(&ni->ni_recovery);
3612 lnet_ni_decref_locked(ni, 0);
3615 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3616 ev_info->mt_nid = nid;
3617 rc = lnet_send_ping(&nid, &mdh, LNET_PING_INFO_MIN_SIZE,
3618 ev_info, the_lnet.ln_mt_handler,
3620 /* lookup the nid again */
3622 ni = lnet_nid_to_ni_locked(&nid, 0);
3625 * the NI has been deleted when we dropped
3632 ni->ni_ping_count++;
3634 ni->ni_ping_mdh = mdh;
3635 lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3640 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3649 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3650 * reexamined in the next iteration.
3652 list_splice_init(&processed_list, &local_queue);
3654 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3661 lnet_resendqs_create(void)
3663 struct list_head **resendqs;
3664 resendqs = lnet_create_array_of_queues();
3669 lnet_net_lock(LNET_LOCK_EX);
3670 the_lnet.ln_mt_resendqs = resendqs;
3671 lnet_net_unlock(LNET_LOCK_EX);
3677 lnet_clean_local_ni_recoveryq(void)
3681 /* This is only called when the monitor thread has stopped */
3684 while ((ni = list_first_entry_or_null(&the_lnet.ln_mt_localNIRecovq,
3686 ni_recovery)) != NULL) {
3687 list_del_init(&ni->ni_recovery);
3689 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3691 lnet_ni_decref_locked(ni, 0);
3698 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3701 struct lnet_handle_md recovery_mdh;
3703 LNetInvalidateMDHandle(&recovery_mdh);
3705 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3706 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3707 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3709 spin_unlock(&lpni->lpni_lock);
3710 lnet_net_unlock(cpt);
3711 if (!LNetMDHandleIsInvalid(recovery_mdh))
3712 LNetMDUnlink(recovery_mdh);
3714 spin_lock(&lpni->lpni_lock);
3718 lnet_clean_peer_ni_recoveryq(void)
3720 struct lnet_peer_ni *lpni, *tmp;
3722 lnet_net_lock(LNET_LOCK_EX);
3724 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3726 list_del_init(&lpni->lpni_recovery);
3727 spin_lock(&lpni->lpni_lock);
3728 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3729 spin_unlock(&lpni->lpni_lock);
3730 lnet_peer_ni_decref_locked(lpni);
3733 lnet_net_unlock(LNET_LOCK_EX);
3737 lnet_clean_resendqs(void)
3739 struct lnet_msg *msg, *tmp;
3743 cfs_cpt_for_each(i, lnet_cpt_table()) {
3745 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3747 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3748 list_del_init(&msg->msg_list);
3749 msg->msg_no_resend = true;
3750 lnet_finalize(msg, -ESHUTDOWN);
3754 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3757 /* Returns the total number of peer NIs in recovery.
3758 * Records up to @arrsz of the associated NIDs in the @nidarr array
3761 lnet_recover_peer_nis(struct lnet_nid *nidarr, unsigned int arrsz)
3763 struct lnet_mt_event_info *ev_info;
3764 LIST_HEAD(processed_list);
3765 LIST_HEAD(local_queue);
3766 struct lnet_handle_md mdh;
3767 struct lnet_peer_ni *lpni;
3768 struct lnet_peer_ni *tmp;
3769 struct lnet_nid nid;
3773 unsigned int nlpnis = 0;
3776 * Always use cpt 0 for locking across all interactions with
3777 * ln_mt_peerNIRecovq
3780 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3784 now = ktime_get_seconds();
3786 list_for_each_entry_safe(lpni, tmp, &local_queue,
3789 * The same protection strategy is used here as is in the
3790 * local recovery case.
3793 healthv = atomic_read(&lpni->lpni_healthv);
3794 spin_lock(&lpni->lpni_lock);
3795 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3796 healthv == LNET_MAX_HEALTH_VALUE) {
3797 list_del_init(&lpni->lpni_recovery);
3798 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3799 spin_unlock(&lpni->lpni_lock);
3800 lnet_peer_ni_decref_locked(lpni);
3806 nidarr[nlpnis] = lpni->lpni_nid;
3810 * If the peer NI has failed recovery we must unlink the
3811 * md. But we want to keep the peer ni on the recovery
3812 * queue so we can try to continue recovering it
3814 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3815 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3816 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3819 spin_unlock(&lpni->lpni_lock);
3821 if (now < lpni->lpni_next_ping) {
3829 * NOTE: we're racing with peer deletion from user space.
3830 * It's possible that a peer is deleted after we check its
3831 * state. In this case the recovery can create a new peer
3833 spin_lock(&lpni->lpni_lock);
3834 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3835 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3836 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3837 spin_unlock(&lpni->lpni_lock);
3839 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3841 CERROR("out of memory. Can't recover %s\n",
3842 libcfs_nidstr(&lpni->lpni_nid));
3843 spin_lock(&lpni->lpni_lock);
3844 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3845 spin_unlock(&lpni->lpni_lock);
3849 /* look at the comments in lnet_recover_local_nis() */
3850 mdh = lpni->lpni_recovery_ping_mdh;
3851 nid = lpni->lpni_nid;
3852 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3854 list_del_init(&lpni->lpni_recovery);
3855 lnet_peer_ni_decref_locked(lpni);
3858 ev_info->mt_type = MT_TYPE_PEER_NI;
3859 ev_info->mt_nid = nid;
3860 rc = lnet_send_ping(&nid, &mdh, LNET_PING_INFO_MIN_SIZE,
3861 ev_info, the_lnet.ln_mt_handler,
3865 * lnet_peer_ni_find_locked() grabs a refcount for
3866 * us. No need to take it explicitly.
3868 lpni = lnet_peer_ni_find_locked(&nid);
3875 lpni->lpni_ping_count++;
3877 lpni->lpni_recovery_ping_mdh = mdh;
3879 lnet_peer_ni_add_to_recoveryq_locked(lpni,
3883 spin_lock(&lpni->lpni_lock);
3884 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3885 spin_unlock(&lpni->lpni_lock);
3888 /* Drop the ref taken by lnet_peer_ni_find_locked() */
3889 lnet_peer_ni_decref_locked(lpni);
3892 spin_unlock(&lpni->lpni_lock);
3896 list_splice_init(&processed_list, &local_queue);
3898 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3904 #define LNET_MAX_NNIDS 20
3905 /* @nids is array of nids that are in recovery. It has max size of
3907 * @nnids is the total number of nids that are in recovery. It can be
3908 * larger than LNET_MAX_NNIDS.
3909 * @local tells us whether these are local or peer NIs in recovery.
3912 lnet_print_recovery_list(struct lnet_nid *nids, unsigned int nnids,
3915 static bool printed;
3925 arrsz = nnids < LNET_MAX_NNIDS ? nnids : LNET_MAX_NNIDS;
3927 /* Printing arrsz NIDs, each has max size LNET_NIDSTR_SIZE, a comma
3928 * and space for each nid after the first (2 * (arrsz - 1)),
3929 * + 1 for terminating null byte
3931 bufsz = (arrsz * LNET_NIDSTR_SIZE) + (2 * (arrsz - 1)) + 1;
3932 LIBCFS_ALLOC(buf, bufsz);
3934 LCONSOLE(D_INFO, "%u %s NIs in recovery\n",
3935 nnids, local ? "local" : "peer");
3940 tmp += sprintf(tmp, "%s", libcfs_nidstr(&nids[0]));
3941 for (i = 1; i < arrsz; i++)
3942 tmp += sprintf(tmp, ", %s", libcfs_nidstr(&nids[i]));
3944 /* LCONSOLE() used to avoid rate limiting when we have both local
3945 * and peer NIs in recovery
3947 LCONSOLE(D_INFO, "%u %s NIs in recovery (showing %u): %s\n",
3948 nnids, local ? "local" : "peer", arrsz, buf);
3950 LIBCFS_FREE(buf, bufsz);
3952 if (!printed && nnids > LNET_MAX_NNIDS) {
3953 LCONSOLE(D_INFO, "See full list with 'lnetctl debug recovery -(p|l)'\n");
3959 lnet_health_update_console(struct lnet_nid *lnids, unsigned int nnis,
3960 struct lnet_nid *rnids, unsigned int nlpnis,
3963 static time64_t next_ni_update;
3964 static time64_t next_lpni_update;
3965 static time64_t next_msg_update;
3966 static unsigned int num_ni_updates;
3967 static unsigned int num_lpni_updates;
3968 static unsigned int num_msg_updates = 1;
3971 if (now >= next_ni_update) {
3973 lnet_print_recovery_list(lnids, nnis, true);
3974 if (num_ni_updates < 5)
3976 next_ni_update = now + (60 * num_ni_updates);
3984 if (now >= next_lpni_update) {
3986 lnet_print_recovery_list(rnids, nlpnis, false);
3987 if (num_lpni_updates < 5)
3989 next_lpni_update = now + (60 * num_lpni_updates);
3991 next_lpni_update = 0;
3992 num_lpni_updates = 0;
3996 /* Let late_count accumulate for 60 seconds */
3997 if (unlikely(!next_msg_update))
3998 next_msg_update = now + 60;
4000 if (now >= next_msg_update) {
4001 late_count = atomic_read(&the_lnet.ln_late_msg_count);
4004 s64 avg = atomic64_xchg(&the_lnet.ln_late_msg_nsecs, 0) /
4005 atomic_xchg(&the_lnet.ln_late_msg_count, 0);
4007 if (avg > NSEC_PER_SEC) {
4008 unsigned int avg_msec;
4010 avg_msec = do_div(avg, NSEC_PER_SEC) /
4012 LCONSOLE_INFO("%u messages in past %us over their deadline by avg %lld.%03us\n",
4014 (60 * num_msg_updates), avg,
4017 if (num_msg_updates < 5)
4019 next_msg_update = now + (60 * num_msg_updates);
4022 next_msg_update = now + 60;
4023 num_msg_updates = 1;
4029 lnet_monitor_thread(void *arg)
4031 time64_t rsp_timeout = 0;
4034 unsigned int nlpnis;
4035 struct lnet_nid local_nids[LNET_MAX_NNIDS];
4036 struct lnet_nid peer_nids[LNET_MAX_NNIDS];
4038 wait_for_completion(&the_lnet.ln_started);
4041 * The monitor thread takes care of the following:
4042 * 1. Checks the aliveness of routers
4043 * 2. Checks if there are messages on the resend queue to resend
4045 * 3. Checks if there are any NIs on the local recovery queue and
4047 * 4. Checks if there are any NIs on the remote recovery queue
4049 * 5. Updates the ping buffer if requested by LNDs upon interface
4052 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
4053 now = ktime_get_real_seconds();
4055 if (lnet_router_checker_active())
4056 lnet_check_routers();
4058 lnet_resend_pending_msgs();
4060 if (now >= rsp_timeout) {
4061 lnet_finalize_expired_responses();
4062 rsp_timeout = now + (lnet_transaction_timeout / 2);
4065 nnis = lnet_recover_local_nis(local_nids, LNET_MAX_NNIDS);
4066 nlpnis = lnet_recover_peer_nis(peer_nids, LNET_MAX_NNIDS);
4067 lnet_health_update_console(local_nids, nnis, peer_nids, nlpnis,
4069 lnet_update_ping_buffer();
4072 * TODO do we need to check if we should sleep without
4073 * timeout? Technically, an active system will always
4074 * have messages in flight so this check will always
4075 * evaluate to false. And on an idle system do we care
4076 * if we wake up every 1 second? Although, we've seen
4077 * cases where we get a complaint that an idle thread
4078 * is waking up unnecessarily.
4080 wait_for_completion_interruptible_timeout(
4081 &the_lnet.ln_mt_wait_complete,
4082 cfs_time_seconds(1));
4083 /* Must re-init the completion before testing anything,
4084 * including ln_mt_state.
4086 reinit_completion(&the_lnet.ln_mt_wait_complete);
4090 lnet_net_lock(LNET_LOCK_EX);
4091 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4092 lnet_net_unlock(LNET_LOCK_EX);
4094 /* signal that the monitor thread is exiting */
4095 up(&the_lnet.ln_mt_signal);
4103 * Returns == 0 if success
4104 * Returns > 0 if LNetMDBind or prior fails
4105 * Returns < 0 if LNetGet fails
4108 lnet_send_ping(struct lnet_nid *dest_nid,
4109 struct lnet_handle_md *mdh, int bytes,
4110 void *user_data, lnet_handler_t handler, bool recovery)
4112 struct lnet_md md = { NULL };
4113 struct lnet_processid id;
4114 struct lnet_ping_buffer *pbuf;
4117 if (LNET_NID_IS_ANY(dest_nid)) {
4122 pbuf = lnet_ping_buffer_alloc(bytes, GFP_NOFS);
4128 /* initialize md content */
4129 md.start = &pbuf->pb_info;
4131 md.threshold = 2; /* GET/REPLY */
4133 md.options = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
4134 md.user_ptr = user_data;
4135 md.handler = handler;
4137 rc = LNetMDBind(&md, LNET_UNLINK, mdh);
4139 lnet_ping_buffer_decref(pbuf);
4140 CERROR("Can't bind MD: %d\n", rc);
4141 rc = -rc; /* change the rc to positive */
4144 id.pid = LNET_PID_LUSTRE;
4147 rc = LNetGet(NULL, *mdh, &id,
4148 LNET_RESERVED_PORTAL,
4149 LNET_PROTO_PING_MATCHBITS, 0, recovery);
4152 goto fail_unlink_md;
4158 LNetInvalidateMDHandle(mdh);
4164 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
4165 int status, bool send, bool unlink_event)
4167 struct lnet_nid *nid = &ev_info->mt_nid;
4169 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
4173 ni = lnet_nid_to_ni_locked(nid, 0);
4179 if (!send || (send && status != 0))
4180 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
4182 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
4187 CDEBUG(D_NET, "local NI (%s) recovery failed with %d\n",
4188 libcfs_nidstr(nid), status);
4192 * need to increment healthv for the ni here, because in
4193 * the lnet_finalize() path we don't have access to this
4194 * NI. And in order to get access to it, we'll need to
4195 * carry forward too much information.
4196 * In the peer case, it'll naturally be incremented
4199 lnet_inc_healthv(&ni->ni_healthv,
4200 lnet_health_sensitivity);
4202 struct lnet_peer_ni *lpni;
4205 cpt = lnet_net_lock_current();
4206 lpni = lnet_peer_ni_find_locked(nid);
4208 lnet_net_unlock(cpt);
4211 spin_lock(&lpni->lpni_lock);
4212 if (!send || (send && status != 0))
4213 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
4215 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
4216 spin_unlock(&lpni->lpni_lock);
4217 lnet_peer_ni_decref_locked(lpni);
4218 lnet_net_unlock(cpt);
4221 CDEBUG(D_NET, "peer NI (%s) recovery failed with %d\n",
4222 libcfs_nidstr(nid), status);
4227 lnet_mt_event_handler(struct lnet_event *event)
4229 struct lnet_mt_event_info *ev_info = event->md_user_ptr;
4230 struct lnet_ping_buffer *pbuf;
4232 /* TODO: remove assert */
4233 LASSERT(event->type == LNET_EVENT_REPLY ||
4234 event->type == LNET_EVENT_SEND ||
4235 event->type == LNET_EVENT_UNLINK);
4237 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
4240 switch (event->type) {
4241 case LNET_EVENT_UNLINK:
4242 CDEBUG(D_NET, "%s recovery ping unlinked\n",
4243 libcfs_nidstr(&ev_info->mt_nid));
4245 case LNET_EVENT_REPLY:
4246 lnet_handle_recovery_reply(ev_info, event->status, false,
4247 event->type == LNET_EVENT_UNLINK);
4249 case LNET_EVENT_SEND:
4250 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4251 libcfs_nidstr(&ev_info->mt_nid),
4252 (event->status) ? "unsuccessfully" :
4253 "successfully", event->status);
4254 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4257 CERROR("Unexpected event: %d\n", event->type);
4260 if (event->unlinked) {
4261 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4262 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4263 lnet_ping_buffer_decref(pbuf);
4268 lnet_rsp_tracker_create(void)
4270 struct list_head **rstqs;
4271 rstqs = lnet_create_array_of_queues();
4276 the_lnet.ln_mt_rstq = rstqs;
4282 lnet_rsp_tracker_clean(void)
4284 lnet_finalize_expired_responses();
4286 cfs_percpt_free(the_lnet.ln_mt_rstq);
4287 the_lnet.ln_mt_rstq = NULL;
4290 int lnet_monitor_thr_start(void)
4293 struct task_struct *task;
4295 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4298 rc = lnet_resendqs_create();
4302 rc = lnet_rsp_tracker_create();
4306 sema_init(&the_lnet.ln_mt_signal, 0);
4308 lnet_net_lock(LNET_LOCK_EX);
4309 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4310 lnet_net_unlock(LNET_LOCK_EX);
4311 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4314 CERROR("Can't start monitor thread: %d\n", rc);
4321 lnet_net_lock(LNET_LOCK_EX);
4322 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4323 lnet_net_unlock(LNET_LOCK_EX);
4324 /* block until event callback signals exit */
4325 down(&the_lnet.ln_mt_signal);
4327 lnet_net_lock(LNET_LOCK_EX);
4328 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4329 lnet_net_unlock(LNET_LOCK_EX);
4330 lnet_rsp_tracker_clean();
4331 lnet_clean_local_ni_recoveryq();
4332 lnet_clean_peer_ni_recoveryq();
4333 lnet_clean_resendqs();
4334 the_lnet.ln_mt_handler = NULL;
4337 lnet_rsp_tracker_clean();
4338 lnet_clean_local_ni_recoveryq();
4339 lnet_clean_peer_ni_recoveryq();
4340 lnet_clean_resendqs();
4344 void lnet_monitor_thr_stop(void)
4346 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4349 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4350 lnet_net_lock(LNET_LOCK_EX);
4351 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4352 lnet_net_unlock(LNET_LOCK_EX);
4354 /* tell the monitor thread that we're shutting down */
4355 complete(&the_lnet.ln_mt_wait_complete);
4357 /* block until monitor thread signals that it's done */
4358 mutex_unlock(&the_lnet.ln_api_mutex);
4359 down(&the_lnet.ln_mt_signal);
4360 mutex_lock(&the_lnet.ln_api_mutex);
4361 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4363 /* perform cleanup tasks */
4364 lnet_rsp_tracker_clean();
4365 lnet_clean_local_ni_recoveryq();
4366 lnet_clean_peer_ni_recoveryq();
4367 lnet_clean_resendqs();
4371 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4375 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4376 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4377 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4378 lnet_net_unlock(cpt);
4380 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4384 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4386 struct lnet_hdr *hdr = &msg->msg_hdr;
4388 if (msg->msg_wanted != 0)
4389 lnet_setpayloadbuffer(msg);
4391 lnet_build_msg_event(msg, LNET_EVENT_PUT);
4393 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
4394 * it back into the ACK during lnet_finalize() */
4395 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4396 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4398 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4399 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4403 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4405 struct lnet_hdr *hdr = &msg->msg_hdr;
4406 struct lnet_match_info info;
4410 /* Convert put fields to host byte order */
4411 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4412 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
4413 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
4415 /* Primary peer NID. */
4416 info.mi_id.nid = msg->msg_initiator;
4417 info.mi_id.pid = hdr->src_pid;
4418 info.mi_opc = LNET_MD_OP_PUT;
4419 info.mi_portal = hdr->msg.put.ptl_index;
4420 info.mi_rlength = hdr->payload_length;
4421 info.mi_roffset = hdr->msg.put.offset;
4422 info.mi_mbits = hdr->msg.put.match_bits;
4423 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4425 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4426 ready_delay = msg->msg_rx_ready_delay;
4429 rc = lnet_ptl_match_md(&info, msg);
4434 case LNET_MATCHMD_OK:
4435 lnet_recv_put(ni, msg);
4438 case LNET_MATCHMD_NONE:
4440 /* no eager_recv or has already called it, should
4441 * have been attached on delayed list */
4444 rc = lnet_ni_eager_recv(ni, msg);
4451 case LNET_MATCHMD_DROP:
4452 CNETERR("Dropping PUT from %s portal %d match %llu"
4453 " offset %d length %d: %d\n",
4454 libcfs_idstr(&info.mi_id), info.mi_portal,
4455 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4457 return -ENOENT; /* -ve: OK but no match */
4462 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4464 struct lnet_match_info info;
4465 struct lnet_hdr *hdr = &msg->msg_hdr;
4466 struct lnet_processid source_id;
4467 struct lnet_handle_wire reply_wmd;
4470 /* Convert get fields to host byte order */
4471 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
4472 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
4473 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
4474 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
4476 source_id.nid = hdr->src_nid;
4477 source_id.pid = hdr->src_pid;
4478 /* Primary peer NID */
4479 info.mi_id.nid = msg->msg_initiator;
4480 info.mi_id.pid = hdr->src_pid;
4481 info.mi_opc = LNET_MD_OP_GET;
4482 info.mi_portal = hdr->msg.get.ptl_index;
4483 info.mi_rlength = hdr->msg.get.sink_length;
4484 info.mi_roffset = hdr->msg.get.src_offset;
4485 info.mi_mbits = hdr->msg.get.match_bits;
4486 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4488 rc = lnet_ptl_match_md(&info, msg);
4489 if (rc == LNET_MATCHMD_DROP) {
4490 CNETERR("Dropping GET from %s portal %d match %llu"
4491 " offset %d length %d\n",
4492 libcfs_idstr(&info.mi_id), info.mi_portal,
4493 info.mi_mbits, info.mi_roffset, info.mi_rlength);
4494 return -ENOENT; /* -ve: OK but no match */
4497 LASSERT(rc == LNET_MATCHMD_OK);
4499 lnet_build_msg_event(msg, LNET_EVENT_GET);
4501 reply_wmd = hdr->msg.get.return_wmd;
4503 lnet_prep_send(msg, LNET_MSG_REPLY, &source_id,
4504 msg->msg_offset, msg->msg_wanted);
4506 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4509 /* The LND completes the REPLY from her recv procedure */
4510 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4511 msg->msg_offset, msg->msg_len, msg->msg_len);
4515 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4516 msg->msg_receiving = 0;
4518 rc = lnet_send(&ni->ni_nid, msg, &msg->msg_from);
4520 /* didn't get as far as lnet_ni_send() */
4521 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4522 libcfs_nidstr(&ni->ni_nid),
4523 libcfs_idstr(&info.mi_id), rc);
4525 lnet_finalize(msg, rc);
4532 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4534 void *private = msg->msg_private;
4535 struct lnet_hdr *hdr = &msg->msg_hdr;
4536 struct lnet_processid src = {};
4537 struct lnet_libmd *md;
4538 unsigned int rlength;
4539 unsigned int mlength;
4542 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4545 src.nid = hdr->src_nid;
4546 src.pid = hdr->src_pid;
4548 /* NB handles only looked up by creator (no flips) */
4549 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4550 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4551 CNETERR("%s: Dropping REPLY from %s for %s "
4553 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4554 (md == NULL) ? "invalid" : "inactive",
4555 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4556 hdr->msg.reply.dst_wmd.wh_object_cookie);
4557 if (md != NULL && md->md_me != NULL)
4558 CERROR("REPLY MD also attached to portal %d\n",
4559 md->md_me->me_portal);
4561 lnet_res_unlock(cpt);
4562 return -ENOENT; /* -ve: OK but no match */
4565 LASSERT(md->md_offset == 0);
4567 rlength = hdr->payload_length;
4568 mlength = min(rlength, md->md_length);
4570 if (mlength < rlength &&
4571 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4572 CNETERR("%s: Dropping REPLY from %s length %d "
4573 "for MD %#llx would overflow (%d)\n",
4574 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4575 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4577 lnet_res_unlock(cpt);
4578 return -ENOENT; /* -ve: OK but no match */
4581 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4582 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4583 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4585 lnet_msg_attach_md(msg, md, 0, mlength);
4588 lnet_setpayloadbuffer(msg);
4590 lnet_res_unlock(cpt);
4592 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4594 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4599 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4601 struct lnet_hdr *hdr = &msg->msg_hdr;
4602 struct lnet_processid src = {};
4603 struct lnet_libmd *md;
4606 src.nid = hdr->src_nid;
4607 src.pid = hdr->src_pid;
4609 /* Convert ack fields to host byte order */
4610 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4611 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4613 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4616 /* NB handles only looked up by creator (no flips) */
4617 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4618 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4619 /* Don't moan; this is expected */
4621 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4622 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4623 (md == NULL) ? "invalid" : "inactive",
4624 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4625 hdr->msg.ack.dst_wmd.wh_object_cookie);
4626 if (md != NULL && md->md_me != NULL)
4627 CERROR("Source MD also attached to portal %d\n",
4628 md->md_me->me_portal);
4630 lnet_res_unlock(cpt);
4631 return -ENOENT; /* -ve! */
4634 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4635 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4636 hdr->msg.ack.dst_wmd.wh_object_cookie);
4638 lnet_msg_attach_md(msg, md, 0, 0);
4640 lnet_res_unlock(cpt);
4642 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4644 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4649 * \retval LNET_CREDIT_OK If \a msg is forwarded
4650 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4651 * \retval -ve error code
4654 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4658 if (!the_lnet.ln_routing)
4661 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4662 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4663 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4664 msg->msg_rx_ready_delay = 1;
4666 lnet_net_unlock(msg->msg_rx_cpt);
4667 rc = lnet_ni_eager_recv(ni, msg);
4668 lnet_net_lock(msg->msg_rx_cpt);
4673 rc = lnet_post_routed_recv_locked(msg, 0);
4678 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4682 switch (msg->msg_type) {
4684 rc = lnet_parse_ack(ni, msg);
4687 rc = lnet_parse_put(ni, msg);
4690 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4692 case LNET_MSG_REPLY:
4693 rc = lnet_parse_reply(ni, msg);
4695 default: /* prevent an unused label if !kernel */
4700 LASSERT(rc == 0 || rc == -ENOENT);
4705 lnet_msgtyp2str (int type)
4714 case LNET_MSG_REPLY:
4716 case LNET_MSG_HELLO:
4719 return ("<UNKNOWN>");
4722 EXPORT_SYMBOL(lnet_msgtyp2str);
4725 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr,
4726 struct lnet_nid *from_nid, void *private, int rdma_req)
4728 struct lnet_peer_ni *lpni;
4729 struct lnet_msg *msg;
4730 __u32 payload_length;
4731 lnet_pid_t dest_pid;
4732 struct lnet_nid dest_nid;
4733 struct lnet_nid src_nid;
4739 time64_t now = ktime_get_seconds();
4741 LASSERT (!in_interrupt ());
4744 src_nid = hdr->src_nid;
4745 dest_nid = hdr->dest_nid;
4746 dest_pid = hdr->dest_pid;
4747 payload_length = hdr->payload_length;
4749 for_me = nid_same(&ni->ni_nid, &dest_nid);
4750 cpt = lnet_nid2cpt(from_nid, ni);
4752 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4753 libcfs_nidstr(&dest_nid),
4754 libcfs_nidstr(&ni->ni_nid),
4755 libcfs_nidstr(&src_nid),
4756 lnet_msgtyp2str(type),
4757 (for_me) ? "for me" : "routed");
4762 if (payload_length > 0) {
4763 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4764 libcfs_nidstr(from_nid),
4765 libcfs_nidstr(&src_nid),
4766 lnet_msgtyp2str(type), payload_length);
4772 case LNET_MSG_REPLY:
4773 if (payload_length >
4774 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4775 CERROR("%s, src %s: bad %s payload %d "
4776 "(%d max expected)\n",
4777 libcfs_nidstr(from_nid),
4778 libcfs_nidstr(&src_nid),
4779 lnet_msgtyp2str(type),
4781 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4787 CERROR("%s, src %s: Bad message type 0x%x\n",
4788 libcfs_nidstr(from_nid),
4789 libcfs_nidstr(&src_nid), type);
4793 /* Only update net_last_alive for incoming GETs on the reserved portal
4794 * (i.e. incoming lnet/discovery pings).
4795 * This avoids situations where the router's own traffic results in NI
4798 if (the_lnet.ln_routing && type == LNET_MSG_GET &&
4799 hdr->msg.get.ptl_index == LNET_RESERVED_PORTAL &&
4800 !lnet_islocalnid(&src_nid) &&
4801 ni->ni_net->net_last_alive != now) {
4803 spin_lock(&ni->ni_net->net_lock);
4804 ni->ni_net->net_last_alive = now;
4805 spin_unlock(&ni->ni_net->net_lock);
4806 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4811 lnet_push_update_to_peers(1);
4813 /* Regard a bad destination NID as a protocol error. Senders should
4814 * know what they're doing; if they don't they're misconfigured, buggy
4815 * or malicious so we chop them off at the knees :) */
4818 if (LNET_NID_NET(&dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4819 /* should have gone direct */
4820 CERROR("%s, src %s: Bad dest nid %s "
4821 "(should have been sent direct)\n",
4822 libcfs_nidstr(from_nid),
4823 libcfs_nidstr(&src_nid),
4824 libcfs_nidstr(&dest_nid));
4828 if (lnet_islocalnid(&dest_nid)) {
4829 /* dest is another local NI; sender should have used
4830 * this node's NID on its own network */
4831 CERROR("%s, src %s: Bad dest nid %s "
4832 "(it's my nid but on a different network)\n",
4833 libcfs_nidstr(from_nid),
4834 libcfs_nidstr(&src_nid),
4835 libcfs_nidstr(&dest_nid));
4839 if (rdma_req && type == LNET_MSG_GET) {
4840 CERROR("%s, src %s: Bad optimized GET for %s "
4841 "(final destination must be me)\n",
4842 libcfs_nidstr(from_nid),
4843 libcfs_nidstr(&src_nid),
4844 libcfs_nidstr(&dest_nid));
4848 if (!the_lnet.ln_routing) {
4849 CERROR("%s, src %s: Dropping message for %s "
4850 "(routing not enabled)\n",
4851 libcfs_nidstr(from_nid),
4852 libcfs_nidstr(&src_nid),
4853 libcfs_nidstr(&dest_nid));
4858 /* Message looks OK; we're not going to return an error, so we MUST
4859 * call back lnd_recv() come what may... */
4861 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4862 fail_peer(&src_nid, 0)) { /* shall we now? */
4863 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4864 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4865 lnet_msgtyp2str(type));
4869 if (!list_empty(&the_lnet.ln_drop_rules) &&
4870 lnet_drop_rule_match(hdr, &ni->ni_nid, NULL)) {
4872 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4873 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4874 libcfs_nidstr(&dest_nid), lnet_msgtyp2str(type));
4878 msg = lnet_msg_alloc();
4880 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4881 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4882 lnet_msgtyp2str(type));
4886 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4887 * pointers NULL etc */
4889 msg->msg_type = type;
4890 msg->msg_private = private;
4891 msg->msg_receiving = 1;
4892 msg->msg_rdma_get = rdma_req;
4893 msg->msg_len = msg->msg_wanted = payload_length;
4894 msg->msg_offset = 0;
4895 msg->msg_hdr = *hdr;
4896 /* for building message event */
4897 msg->msg_from = *from_nid;
4899 msg->msg_target.pid = dest_pid;
4900 msg->msg_target.nid = dest_nid;
4901 msg->msg_routing = 1;
4905 lpni = lnet_peerni_by_nid_locked(from_nid, &ni->ni_nid, cpt);
4907 lnet_net_unlock(cpt);
4909 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4910 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4911 lnet_msgtyp2str(type), rc);
4913 if (rc == -ESHUTDOWN)
4914 /* We are shutting down. Don't do anything more */
4919 /* If this message was forwarded to us from a router then we may need
4920 * to update router aliveness or check for an asymmetrical route
4923 if (((lnet_drop_asym_route && for_me) ||
4924 !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4925 LNET_NID_NET(&src_nid) != LNET_NID_NET(from_nid)) {
4926 __u32 src_net_id = LNET_NID_NET(&src_nid);
4927 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4928 struct lnet_route *route;
4931 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4932 if (route->lr_net == src_net_id) {
4934 /* If we're transitioning the gateway from
4935 * dead -> alive, and discovery is disabled
4936 * locally or on the gateway, then we need to
4937 * update the cached route aliveness for each
4938 * route to the src_nid's net.
4940 * Otherwise, we're only checking for
4941 * symmetrical route, and we can break the
4944 if (!gw->lp_alive &&
4945 lnet_is_discovery_disabled(gw))
4946 lnet_set_route_aliveness(route, true);
4951 if (lnet_drop_asym_route && for_me && !found) {
4952 /* Drop ref taken by lnet_nid2peerni_locked() */
4953 lnet_peer_ni_decref_locked(lpni);
4954 lnet_net_unlock(cpt);
4955 /* we would not use from_nid to route a message to
4957 * => asymmetric routing detected but forbidden
4959 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4960 libcfs_nidstr(from_nid),
4961 libcfs_nidstr(&src_nid), lnet_msgtyp2str(type));
4965 if (!gw->lp_alive) {
4966 struct lnet_peer_net *lpn;
4967 struct lnet_peer_ni *lpni2;
4969 gw->lp_alive = true;
4970 /* Mark all remote NIs on src_nid's net UP */
4971 lpn = lnet_peer_get_net_locked(gw, src_net_id);
4973 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4975 lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4979 lpni->lpni_last_alive = now;
4981 msg->msg_rxpeer = lpni;
4983 lnet_ni_addref_locked(ni, cpt);
4984 /* Multi-Rail: Primary NID of source. */
4985 lnet_peer_primary_nid_locked(&src_nid, &msg->msg_initiator);
4988 * mark the status of this lpni as UP since we received a message
4989 * from it. The ping response reports back the ns_status which is
4990 * marked on the remote as up or down and we cache it here.
4992 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4994 lnet_msg_commit(msg, cpt);
4996 /* message delay simulation */
4997 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4998 lnet_delay_rule_match_locked(hdr, msg))) {
4999 lnet_net_unlock(cpt);
5004 rc = lnet_parse_forward_locked(ni, msg);
5005 lnet_net_unlock(cpt);
5010 if (rc == LNET_CREDIT_OK) {
5011 lnet_ni_recv(ni, msg->msg_private, msg, 0,
5012 0, payload_length, payload_length);
5017 lnet_net_unlock(cpt);
5019 rc = lnet_parse_local(ni, msg);
5025 LASSERT(msg->msg_md == NULL);
5026 lnet_finalize(msg, rc);
5029 lnet_drop_message(ni, cpt, private, payload_length, type);
5032 EXPORT_SYMBOL(lnet_parse);
5035 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
5037 struct lnet_msg *msg;
5039 while ((msg = list_first_entry_or_null(head, struct lnet_msg,
5040 msg_list)) != NULL) {
5041 struct lnet_processid id = {};
5043 list_del(&msg->msg_list);
5045 id.nid = msg->msg_hdr.src_nid;
5046 id.pid = msg->msg_hdr.src_pid;
5048 LASSERT(msg->msg_md == NULL);
5049 LASSERT(msg->msg_rx_delayed);
5050 LASSERT(msg->msg_rxpeer != NULL);
5051 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
5053 CWARN("Dropping delayed PUT from %s portal %d match %llu"
5054 " offset %d length %d: %s\n",
5056 msg->msg_hdr.msg.put.ptl_index,
5057 msg->msg_hdr.msg.put.match_bits,
5058 msg->msg_hdr.msg.put.offset,
5059 msg->msg_hdr.payload_length, reason);
5061 /* NB I can't drop msg's ref on msg_rxpeer until after I've
5062 * called lnet_drop_message(), so I just hang onto msg as well
5063 * until that's done */
5065 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
5066 msg->msg_private, msg->msg_len,
5069 msg->msg_no_resend = true;
5071 * NB: message will not generate event because w/o attached MD,
5072 * but we still should give error code so lnet_msg_decommit()
5073 * can skip counters operations and other checks.
5075 lnet_finalize(msg, -ENOENT);
5080 lnet_recv_delayed_msg_list(struct list_head *head)
5082 struct lnet_msg *msg;
5084 while ((msg = list_first_entry_or_null(head, struct lnet_msg,
5085 msg_list)) != NULL) {
5086 struct lnet_processid id;
5088 list_del(&msg->msg_list);
5090 /* md won't disappear under me, since each msg
5091 * holds a ref on it */
5093 id.nid = msg->msg_hdr.src_nid;
5094 id.pid = msg->msg_hdr.src_pid;
5096 LASSERT(msg->msg_rx_delayed);
5097 LASSERT(msg->msg_md != NULL);
5098 LASSERT(msg->msg_rxpeer != NULL);
5099 LASSERT(msg->msg_rxni != NULL);
5100 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
5102 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
5103 "match %llu offset %d length %d.\n",
5104 libcfs_idstr(&id), msg->msg_hdr.msg.put.ptl_index,
5105 msg->msg_hdr.msg.put.match_bits,
5106 msg->msg_hdr.msg.put.offset,
5107 msg->msg_hdr.payload_length);
5109 lnet_recv_put(msg->msg_rxni, msg);
5114 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
5115 struct lnet_libmd *md, struct lnet_handle_md mdh)
5118 struct lnet_rsp_tracker *local_rspt;
5121 * MD has a refcount taken by message so it's not going away.
5122 * The MD however can be looked up. We need to secure the access
5123 * to the md_rspt_ptr by taking the res_lock.
5124 * The rspt can be accessed without protection up to when it gets
5125 * added to the list.
5129 local_rspt = md->md_rspt_ptr;
5130 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
5131 if (local_rspt != NULL) {
5133 * we already have an rspt attached to the md, so we'll
5134 * update the deadline on that one.
5136 lnet_rspt_free(rspt, cpt);
5139 rspt->rspt_mdh = mdh;
5140 rspt->rspt_cpt = cpt;
5141 /* store the rspt so we can access it when we get the REPLY */
5142 md->md_rspt_ptr = rspt;
5145 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
5148 * add to the list of tracked responses. It's added to tail of the
5149 * list in order to expire all the older entries first.
5152 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
5153 lnet_net_unlock(cpt);
5154 lnet_res_unlock(cpt);
5158 * Initiate an asynchronous PUT operation.
5160 * There are several events associated with a PUT: completion of the send on
5161 * the initiator node (LNET_EVENT_SEND), and when the send completes
5162 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
5163 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
5164 * used at the target node to indicate the completion of incoming data
5167 * The local events will be logged in the EQ associated with the MD pointed to
5168 * by \a mdh handle. Using a MD without an associated EQ results in these
5169 * events being discarded. In this case, the caller must have another
5170 * mechanism (e.g., a higher level protocol) for determining when it is safe
5171 * to modify the memory region associated with the MD.
5173 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
5174 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
5176 * \param self Indicates the NID of a local interface through which to send
5177 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
5178 * \param mdh A handle for the MD that describes the memory to be sent. The MD
5179 * must be "free floating" (See LNetMDBind()).
5180 * \param ack Controls whether an acknowledgment is requested.
5181 * Acknowledgments are only sent when they are requested by the initiating
5182 * process and the target MD enables them.
5183 * \param target A process identifier for the target process.
5184 * \param portal The index in the \a target's portal table.
5185 * \param match_bits The match bits to use for MD selection at the target
5187 * \param offset The offset into the target MD (only used when the target
5188 * MD has the LNET_MD_MANAGE_REMOTE option set).
5189 * \param hdr_data 64 bits of user data that can be included in the message
5190 * header. This data is written to an event queue entry at the target if an
5191 * EQ is present on the matching MD.
5193 * \retval 0 Success, and only in this case events will be generated
5194 * and logged to EQ (if it exists).
5195 * \retval -EIO Simulated failure.
5196 * \retval -ENOMEM Memory allocation failure.
5197 * \retval -ENOENT Invalid MD object.
5199 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
5202 LNetPut(struct lnet_nid *self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
5203 struct lnet_processid *target, unsigned int portal,
5204 __u64 match_bits, unsigned int offset,
5207 struct lnet_msg *msg;
5208 struct lnet_libmd *md;
5211 struct lnet_rsp_tracker *rspt = NULL;
5213 LASSERT(the_lnet.ln_refcount > 0);
5215 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5216 fail_peer(&target->nid, 1)) { /* shall we now? */
5217 CERROR("Dropping PUT to %s: simulated failure\n",
5218 libcfs_idstr(target));
5222 msg = lnet_msg_alloc();
5224 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
5225 libcfs_idstr(target));
5228 msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
5230 cpt = lnet_cpt_of_cookie(mdh.cookie);
5232 if (ack == LNET_ACK_REQ) {
5233 rspt = lnet_rspt_alloc(cpt);
5235 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
5236 libcfs_idstr(target));
5239 INIT_LIST_HEAD(&rspt->rspt_on_list);
5244 md = lnet_handle2md(&mdh);
5245 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5246 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5247 match_bits, portal, libcfs_idstr(target),
5248 md == NULL ? -1 : md->md_threshold);
5249 if (md != NULL && md->md_me != NULL)
5250 CERROR("Source MD also attached to portal %d\n",
5251 md->md_me->me_portal);
5252 lnet_res_unlock(cpt);
5255 lnet_rspt_free(rspt, cpt);
5261 CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5263 lnet_msg_attach_md(msg, md, 0, 0);
5265 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5267 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5268 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5269 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5270 msg->msg_hdr.msg.put.hdr_data = hdr_data;
5272 /* NB handles only looked up by creator (no flips) */
5273 if (ack == LNET_ACK_REQ) {
5274 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5275 the_lnet.ln_interface_cookie;
5276 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5277 md->md_lh.lh_cookie;
5279 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5280 LNET_WIRE_HANDLE_COOKIE_NONE;
5281 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5282 LNET_WIRE_HANDLE_COOKIE_NONE;
5285 lnet_res_unlock(cpt);
5287 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5289 if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5291 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5293 lnet_rspt_free(rspt, cpt);
5295 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5299 rc = lnet_send(self, msg, NULL);
5302 CNETERR("Error sending PUT to %s: %d\n",
5303 libcfs_idstr(target), rc);
5304 msg->msg_no_resend = true;
5305 lnet_finalize(msg, rc);
5308 /* completion will be signalled by an event */
5311 EXPORT_SYMBOL(LNetPut);
5314 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
5315 * returns a msg for the LND to pass to lnet_finalize() when the sink
5316 * data has been received.
5318 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5319 * lnet_finalize() is called on it, so the LND must call this first
5322 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5324 struct lnet_msg *msg = lnet_msg_alloc();
5325 struct lnet_libmd *getmd = getmsg->msg_md;
5326 struct lnet_processid *peer_id = &getmsg->msg_target;
5329 LASSERT(!getmsg->msg_target_is_router);
5330 LASSERT(!getmsg->msg_routing);
5333 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5334 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id));
5338 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5341 LASSERT(getmd->md_refcount > 0);
5343 if (getmd->md_threshold == 0) {
5344 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5345 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id),
5347 lnet_res_unlock(cpt);
5351 LASSERT(getmd->md_offset == 0);
5353 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5354 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id), getmd);
5356 /* setup information for lnet_build_msg_event */
5357 msg->msg_initiator =
5358 getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5359 msg->msg_from = peer_id->nid;
5360 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5361 msg->msg_hdr.src_nid = peer_id->nid;
5362 msg->msg_hdr.payload_length = getmd->md_length;
5363 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5365 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5366 lnet_res_unlock(cpt);
5368 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5371 lnet_msg_commit(msg, cpt);
5372 lnet_net_unlock(cpt);
5374 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5379 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5382 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5383 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5384 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5386 lnet_net_unlock(cpt);
5393 EXPORT_SYMBOL(lnet_create_reply_msg);
5396 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5399 /* Set the REPLY length, now the RDMA that elides the REPLY message has
5400 * completed and I know it. */
5401 LASSERT(reply != NULL);
5402 LASSERT(reply->msg_type == LNET_MSG_GET);
5403 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5405 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
5406 * the end of my buffer, I might as well be dead. */
5407 LASSERT(len <= reply->msg_ev.mlength);
5409 reply->msg_ev.mlength = len;
5411 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5414 * Initiate an asynchronous GET operation.
5416 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5417 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5418 * the target node in the REPLY has been written to local MD.
5420 * On the target node, an LNET_EVENT_GET is logged when the GET request
5421 * arrives and is accepted into a MD.
5423 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5424 * \param mdh A handle for the MD that describes the memory into which the
5425 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5427 * \retval 0 Success, and only in this case events will be generated
5428 * and logged to EQ (if it exists) of the MD.
5429 * \retval -EIO Simulated failure.
5430 * \retval -ENOMEM Memory allocation failure.
5431 * \retval -ENOENT Invalid MD object.
5434 LNetGet(struct lnet_nid *self, struct lnet_handle_md mdh,
5435 struct lnet_processid *target, unsigned int portal,
5436 __u64 match_bits, unsigned int offset, bool recovery)
5438 struct lnet_msg *msg;
5439 struct lnet_libmd *md;
5440 struct lnet_rsp_tracker *rspt;
5444 LASSERT(the_lnet.ln_refcount > 0);
5446 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5447 fail_peer(&target->nid, 1)) /* shall we now? */
5449 CERROR("Dropping GET to %s: simulated failure\n",
5450 libcfs_idstr(target));
5454 msg = lnet_msg_alloc();
5456 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5457 libcfs_idstr(target));
5461 cpt = lnet_cpt_of_cookie(mdh.cookie);
5463 rspt = lnet_rspt_alloc(cpt);
5465 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5466 libcfs_idstr(target));
5469 INIT_LIST_HEAD(&rspt->rspt_on_list);
5471 msg->msg_recovery = recovery;
5475 md = lnet_handle2md(&mdh);
5476 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5477 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5478 match_bits, portal, libcfs_idstr(target),
5479 md == NULL ? -1 : md->md_threshold);
5480 if (md != NULL && md->md_me != NULL)
5481 CERROR("REPLY MD also attached to portal %d\n",
5482 md->md_me->me_portal);
5484 lnet_res_unlock(cpt);
5487 lnet_rspt_free(rspt, cpt);
5491 CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5493 lnet_msg_attach_md(msg, md, 0, 0);
5495 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5497 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5498 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5499 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5500 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5502 /* NB handles only looked up by creator (no flips) */
5503 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5504 the_lnet.ln_interface_cookie;
5505 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5506 md->md_lh.lh_cookie;
5508 lnet_res_unlock(cpt);
5510 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5512 if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5513 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5515 lnet_rspt_free(rspt, cpt);
5517 rc = lnet_send(self, msg, NULL);
5519 CNETERR("Error sending GET to %s: %d\n",
5520 libcfs_idstr(target), rc);
5521 msg->msg_no_resend = true;
5522 lnet_finalize(msg, rc);
5525 /* completion will be signalled by an event */
5528 EXPORT_SYMBOL(LNetGet);
5531 * Calculate distance to node at \a dstnid.
5533 * \param dstnid Target NID.
5534 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5536 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5539 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5540 * local_nid_dist_zero is set, which is the default.
5541 * \retval positives Distance to target NID, i.e. number of hops plus one.
5542 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5545 LNetDist(struct lnet_nid *dstnid, struct lnet_nid *srcnid, __u32 *orderp)
5547 struct lnet_ni *ni = NULL;
5548 struct lnet_remotenet *rnet;
5549 __u32 dstnet = LNET_NID_NET(dstnid);
5553 struct list_head *rn_list;
5554 struct lnet_ni *matched_dstnet = NULL;
5556 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5557 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5558 * keep order 0 free for 0@lo and order 1 free for a local NID
5560 * WARNING: dstnid and srcnid might point to same place.
5561 * Don't set *srcnid until late.
5564 LASSERT(the_lnet.ln_refcount > 0);
5566 cpt = lnet_net_lock_current();
5568 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5569 if (nid_same(&ni->ni_nid, dstnid)) {
5570 if (orderp != NULL) {
5571 if (nid_is_lo0(dstnid))
5578 lnet_net_unlock(cpt);
5580 return local_nid_dist_zero ? 0 : 1;
5583 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5584 matched_dstnet = ni;
5585 /* We matched the destination net, but we may have
5586 * additional local NIs to inspect.
5588 * We record the order as appropriate, but
5589 * they may be overwritten if we match local NI above.
5593 /* Check if ni was originally created in
5594 * current net namespace.
5595 * If not, assign order above 0xffff0000,
5596 * to make this ni not a priority.
5598 if (current->nsproxy &&
5599 !net_eq(ni->ni_net_ns,
5600 current->nsproxy->net_ns))
5601 *orderp = order + 0xffff0000;
5610 if (matched_dstnet) {
5612 *srcnid = matched_dstnet->ni_nid;
5613 lnet_net_unlock(cpt);
5617 rn_list = lnet_net2rnethash(dstnet);
5618 list_for_each_entry(rnet, rn_list, lrn_list) {
5619 if (rnet->lrn_net == dstnet) {
5620 struct lnet_route *route;
5621 struct lnet_route *shortest = NULL;
5622 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5625 LASSERT(!list_empty(&rnet->lrn_routes));
5627 list_for_each_entry(route, &rnet->lrn_routes,
5629 route_hops = route->lr_hops;
5630 if (route_hops == LNET_UNDEFINED_HOPS)
5632 if (shortest == NULL ||
5633 route_hops < shortest_hops) {
5635 shortest_hops = route_hops;
5639 LASSERT(shortest != NULL);
5640 hops = shortest_hops;
5642 struct lnet_net *net;
5643 net = lnet_get_net_locked(shortest->lr_lnet);
5645 ni = lnet_get_next_ni_locked(net, NULL);
5646 *srcnid = ni->ni_nid;
5650 lnet_net_unlock(cpt);
5656 lnet_net_unlock(cpt);
5657 return -EHOSTUNREACH;
5659 EXPORT_SYMBOL(LNetDist);