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))
815 return -EHOSTUNREACH;
821 * \param msg The message to be sent.
822 * \param do_send True if lnet_ni_send() should be called in this function.
823 * lnet_send() is going to lnet_net_unlock immediately after this, so
824 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
826 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
827 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
828 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
829 * \retval -ECANCELED If the MD of the message has been unlinked.
832 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
834 struct lnet_peer_ni *lp = msg->msg_txpeer;
835 struct lnet_ni *ni = msg->msg_txni;
836 int cpt = msg->msg_tx_cpt;
837 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
840 /* non-lnet_send() callers have checked before */
841 LASSERT(!do_send || msg->msg_tx_delayed);
842 LASSERT(!msg->msg_receiving);
843 LASSERT(msg->msg_tx_committed);
845 /* can't get here if we're sending to the loopback interface */
846 if (the_lnet.ln_loni)
847 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
849 /* NB 'lp' is always the next hop */
850 rc = lnet_check_message_drop(ni, lp, msg);
852 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
853 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
855 lnet_net_unlock(cpt);
857 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
859 LNET_STATS_TYPE_DROP);
861 lnet_incr_stats(&msg->msg_txni->ni_stats,
863 LNET_STATS_TYPE_DROP);
865 if (rc == -EHOSTUNREACH) {
866 CNETERR("Dropping message for %s: peer not alive\n",
867 libcfs_idstr(&msg->msg_target));
868 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
870 CNETERR("Dropping message for %s: exceeded message deadline\n",
871 libcfs_idstr(&msg->msg_target));
872 msg->msg_health_status =
873 LNET_MSG_STATUS_NETWORK_TIMEOUT;
877 lnet_finalize(msg, rc);
883 if (msg->msg_md != NULL &&
884 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
885 lnet_net_unlock(cpt);
887 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
888 "called on the MD/ME.\n",
889 libcfs_idstr(&msg->msg_target));
891 msg->msg_no_resend = true;
892 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
893 msg, libcfs_idstr(&msg->msg_target));
894 lnet_finalize(msg, -ECANCELED);
901 if (!msg->msg_peertxcredit) {
902 spin_lock(&lp->lpni_lock);
903 LASSERT((lp->lpni_txcredits < 0) ==
904 !list_empty(&lp->lpni_txq));
906 msg->msg_peertxcredit = 1;
907 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr_nid4);
908 lp->lpni_txcredits--;
910 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
911 lp->lpni_mintxcredits = lp->lpni_txcredits;
913 if (lp->lpni_txcredits < 0) {
914 msg->msg_tx_delayed = 1;
915 list_add_tail(&msg->msg_list, &lp->lpni_txq);
916 spin_unlock(&lp->lpni_lock);
917 return LNET_CREDIT_WAIT;
919 spin_unlock(&lp->lpni_lock);
922 if (!msg->msg_txcredit) {
923 LASSERT((tq->tq_credits < 0) ==
924 !list_empty(&tq->tq_delayed));
926 msg->msg_txcredit = 1;
928 atomic_dec(&ni->ni_tx_credits);
930 if (tq->tq_credits < tq->tq_credits_min)
931 tq->tq_credits_min = tq->tq_credits;
933 if (tq->tq_credits < 0) {
934 msg->msg_tx_delayed = 1;
935 list_add_tail(&msg->msg_list, &tq->tq_delayed);
936 return LNET_CREDIT_WAIT;
940 if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
941 lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
942 msg->msg_tx_delayed = 1;
943 return LNET_CREDIT_WAIT;
946 /* unset the tx_delay flag as we're going to send it now */
947 msg->msg_tx_delayed = 0;
950 lnet_net_unlock(cpt);
951 lnet_ni_send(ni, msg);
954 return LNET_CREDIT_OK;
958 static struct lnet_rtrbufpool *
959 lnet_msg2bufpool(struct lnet_msg *msg)
961 struct lnet_rtrbufpool *rbp;
964 LASSERT(msg->msg_rx_committed);
966 cpt = msg->msg_rx_cpt;
967 rbp = &the_lnet.ln_rtrpools[cpt][0];
969 LASSERT(msg->msg_len <= LNET_MTU);
970 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
972 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
979 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
981 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
982 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
983 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
984 * received or OK to receive */
985 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
986 struct lnet_peer *lp;
987 struct lnet_rtrbufpool *rbp;
988 struct lnet_rtrbuf *rb;
990 LASSERT(msg->msg_kiov == NULL);
991 LASSERT(msg->msg_niov == 0);
992 LASSERT(msg->msg_routing);
993 LASSERT(msg->msg_receiving);
994 LASSERT(!msg->msg_sending);
995 LASSERT(lpni->lpni_peer_net);
996 LASSERT(lpni->lpni_peer_net->lpn_peer);
998 lp = lpni->lpni_peer_net->lpn_peer;
1000 /* non-lnet_parse callers only receive delayed messages */
1001 LASSERT(!do_recv || msg->msg_rx_delayed);
1003 if (!msg->msg_peerrtrcredit) {
1004 /* lpni_lock protects the credit manipulation */
1005 spin_lock(&lpni->lpni_lock);
1007 msg->msg_peerrtrcredit = 1;
1008 lpni->lpni_rtrcredits--;
1009 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1010 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1012 if (lpni->lpni_rtrcredits < 0) {
1013 spin_unlock(&lpni->lpni_lock);
1014 /* must have checked eager_recv before here */
1015 LASSERT(msg->msg_rx_ready_delay);
1016 msg->msg_rx_delayed = 1;
1017 /* lp_lock protects the lp_rtrq */
1018 spin_lock(&lp->lp_lock);
1019 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1020 spin_unlock(&lp->lp_lock);
1021 return LNET_CREDIT_WAIT;
1023 spin_unlock(&lpni->lpni_lock);
1026 rbp = lnet_msg2bufpool(msg);
1028 if (!msg->msg_rtrcredit) {
1029 msg->msg_rtrcredit = 1;
1031 if (rbp->rbp_credits < rbp->rbp_mincredits)
1032 rbp->rbp_mincredits = rbp->rbp_credits;
1034 if (rbp->rbp_credits < 0) {
1035 /* must have checked eager_recv before here */
1036 LASSERT(msg->msg_rx_ready_delay);
1037 msg->msg_rx_delayed = 1;
1038 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1039 return LNET_CREDIT_WAIT;
1043 LASSERT(!list_empty(&rbp->rbp_bufs));
1044 rb = list_first_entry(&rbp->rbp_bufs, struct lnet_rtrbuf, rb_list);
1045 list_del(&rb->rb_list);
1047 msg->msg_niov = rbp->rbp_npages;
1048 msg->msg_kiov = &rb->rb_kiov[0];
1050 /* unset the msg-rx_delayed flag since we're receiving the message */
1051 msg->msg_rx_delayed = 0;
1054 int cpt = msg->msg_rx_cpt;
1056 lnet_net_unlock(cpt);
1057 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1058 0, msg->msg_len, msg->msg_len);
1061 return LNET_CREDIT_OK;
1065 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1067 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1068 struct lnet_ni *txni = msg->msg_txni;
1069 struct lnet_msg *msg2;
1071 if (msg->msg_txcredit) {
1072 struct lnet_ni *ni = msg->msg_txni;
1073 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1075 /* give back NI txcredits */
1076 msg->msg_txcredit = 0;
1078 LASSERT((tq->tq_credits < 0) ==
1079 !list_empty(&tq->tq_delayed));
1082 atomic_inc(&ni->ni_tx_credits);
1083 if (tq->tq_credits <= 0) {
1084 msg2 = list_first_entry(&tq->tq_delayed,
1085 struct lnet_msg, msg_list);
1086 list_del(&msg2->msg_list);
1088 LASSERT(msg2->msg_txni == ni);
1089 LASSERT(msg2->msg_tx_delayed);
1090 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1092 (void) lnet_post_send_locked(msg2, 1);
1096 if (msg->msg_peertxcredit) {
1097 /* give back peer txcredits */
1098 msg->msg_peertxcredit = 0;
1100 spin_lock(&txpeer->lpni_lock);
1101 LASSERT((txpeer->lpni_txcredits < 0) ==
1102 !list_empty(&txpeer->lpni_txq));
1104 txpeer->lpni_txqnob -= msg->msg_len +
1105 sizeof(struct lnet_hdr_nid4);
1106 LASSERT(txpeer->lpni_txqnob >= 0);
1108 txpeer->lpni_txcredits++;
1109 if (txpeer->lpni_txcredits <= 0) {
1112 msg2 = list_first_entry(&txpeer->lpni_txq,
1113 struct lnet_msg, msg_list);
1114 list_del(&msg2->msg_list);
1115 spin_unlock(&txpeer->lpni_lock);
1117 LASSERT(msg2->msg_txpeer == txpeer);
1118 LASSERT(msg2->msg_tx_delayed);
1120 msg2_cpt = msg2->msg_tx_cpt;
1123 * The msg_cpt can be different from the msg2_cpt
1124 * so we need to make sure we lock the correct cpt
1126 * Once we call lnet_post_send_locked() it is no
1127 * longer safe to access msg2, since it could've
1128 * been freed by lnet_finalize(), but we still
1129 * need to relock the correct cpt, so we cache the
1130 * msg2_cpt for the purpose of the check that
1131 * follows the call to lnet_pose_send_locked().
1133 if (msg2_cpt != msg->msg_tx_cpt) {
1134 lnet_net_unlock(msg->msg_tx_cpt);
1135 lnet_net_lock(msg2_cpt);
1137 (void) lnet_post_send_locked(msg2, 1);
1138 if (msg2_cpt != msg->msg_tx_cpt) {
1139 lnet_net_unlock(msg2_cpt);
1140 lnet_net_lock(msg->msg_tx_cpt);
1143 spin_unlock(&txpeer->lpni_lock);
1148 msg->msg_txni = NULL;
1149 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1152 if (txpeer != NULL) {
1153 msg->msg_txpeer = NULL;
1154 lnet_peer_ni_decref_locked(txpeer);
1159 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1161 struct lnet_msg *msg;
1163 if (list_empty(&rbp->rbp_msgs))
1165 msg = list_first_entry(&rbp->rbp_msgs,
1166 struct lnet_msg, msg_list);
1167 list_del(&msg->msg_list);
1169 (void)lnet_post_routed_recv_locked(msg, 1);
1173 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1175 struct lnet_msg *msg;
1176 struct lnet_msg *tmp;
1178 lnet_net_unlock(cpt);
1180 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1181 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1182 0, 0, 0, msg->msg_hdr.payload_length);
1183 list_del_init(&msg->msg_list);
1184 msg->msg_no_resend = true;
1185 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1186 lnet_finalize(msg, -ECANCELED);
1193 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1195 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1196 struct lnet_peer *lp;
1197 struct lnet_ni *rxni = msg->msg_rxni;
1198 struct lnet_msg *msg2;
1200 if (msg->msg_rtrcredit) {
1201 /* give back global router credits */
1202 struct lnet_rtrbuf *rb;
1203 struct lnet_rtrbufpool *rbp;
1205 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1206 * there until it gets one allocated, or aborts the wait
1208 LASSERT(msg->msg_kiov != NULL);
1210 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1213 msg->msg_kiov = NULL;
1214 msg->msg_rtrcredit = 0;
1216 LASSERT(rbp == lnet_msg2bufpool(msg));
1218 LASSERT((rbp->rbp_credits > 0) ==
1219 !list_empty(&rbp->rbp_bufs));
1221 /* If routing is now turned off, we just drop this buffer and
1222 * don't bother trying to return credits. */
1223 if (!the_lnet.ln_routing) {
1224 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1228 /* It is possible that a user has lowered the desired number of
1229 * buffers in this pool. Make sure we never put back
1230 * more buffers than the stated number. */
1231 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1232 /* Discard this buffer so we don't have too
1234 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1235 rbp->rbp_nbuffers--;
1237 list_add(&rb->rb_list, &rbp->rbp_bufs);
1239 if (rbp->rbp_credits <= 0)
1240 lnet_schedule_blocked_locked(rbp);
1245 if (msg->msg_peerrtrcredit) {
1247 LASSERT(rxpeerni->lpni_peer_net);
1248 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1250 /* give back peer router credits */
1251 msg->msg_peerrtrcredit = 0;
1253 spin_lock(&rxpeerni->lpni_lock);
1254 rxpeerni->lpni_rtrcredits++;
1255 spin_unlock(&rxpeerni->lpni_lock);
1257 lp = rxpeerni->lpni_peer_net->lpn_peer;
1258 spin_lock(&lp->lp_lock);
1260 /* drop all messages which are queued to be routed on that
1262 if (!the_lnet.ln_routing) {
1264 list_splice_init(&lp->lp_rtrq, &drop);
1265 spin_unlock(&lp->lp_lock);
1266 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1267 } else if (!list_empty(&lp->lp_rtrq)) {
1270 msg2 = list_first_entry(&lp->lp_rtrq,
1271 struct lnet_msg, msg_list);
1272 list_del(&msg2->msg_list);
1273 msg2_cpt = msg2->msg_rx_cpt;
1274 spin_unlock(&lp->lp_lock);
1276 * messages on the lp_rtrq can be from any NID in
1277 * the peer, which means they might have different
1278 * cpts. We need to make sure we lock the right
1281 if (msg2_cpt != msg->msg_rx_cpt) {
1282 lnet_net_unlock(msg->msg_rx_cpt);
1283 lnet_net_lock(msg2_cpt);
1285 (void) lnet_post_routed_recv_locked(msg2, 1);
1286 if (msg2_cpt != msg->msg_rx_cpt) {
1287 lnet_net_unlock(msg2_cpt);
1288 lnet_net_lock(msg->msg_rx_cpt);
1291 spin_unlock(&lp->lp_lock);
1295 msg->msg_rxni = NULL;
1296 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1298 if (rxpeerni != NULL) {
1299 msg->msg_rxpeer = NULL;
1300 lnet_peer_ni_decref_locked(rxpeerni);
1304 static struct lnet_peer_ni *
1305 lnet_select_peer_ni(struct lnet_ni *best_ni, struct lnet_nid *dst_nid,
1306 struct lnet_peer *peer,
1307 struct lnet_peer_ni *best_lpni,
1308 struct lnet_peer_net *peer_net)
1311 * Look at the peer NIs for the destination peer that connect
1312 * to the chosen net. If a peer_ni is preferred when using the
1313 * best_ni to communicate, we use that one. If there is no
1314 * preferred peer_ni, or there are multiple preferred peer_ni,
1315 * the available transmit credits are used. If the transmit
1316 * credits are equal, we round-robin over the peer_ni.
1318 struct lnet_peer_ni *lpni = NULL;
1319 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1321 int best_lpni_healthv = (best_lpni) ?
1322 atomic_read(&best_lpni->lpni_healthv) : 0;
1323 bool best_lpni_is_preferred = false;
1324 bool lpni_is_preferred;
1326 __u32 lpni_sel_prio;
1327 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1329 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1331 * if the best_ni we've chosen aleady has this lpni
1332 * preferred, then let's use it
1335 lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1336 lpni, &best_ni->ni_nid);
1337 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1338 libcfs_nidstr(&best_ni->ni_nid),
1341 lpni_is_preferred = false;
1344 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1345 lpni_sel_prio = lpni->lpni_sel_priority;
1348 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1349 libcfs_nidstr(&lpni->lpni_nid),
1350 libcfs_nidstr(&best_lpni->lpni_nid),
1351 lpni_healthv, best_lpni_healthv,
1352 lpni_sel_prio, best_sel_prio,
1353 lpni->lpni_txcredits, best_lpni_credits,
1354 lpni->lpni_seq, best_lpni->lpni_seq);
1358 /* pick the healthiest peer ni */
1359 if (lpni_healthv < best_lpni_healthv)
1361 else if (lpni_healthv > best_lpni_healthv) {
1362 if (best_lpni_is_preferred)
1363 best_lpni_is_preferred = false;
1367 if (lpni_sel_prio > best_sel_prio)
1369 else if (lpni_sel_prio < best_sel_prio) {
1370 if (best_lpni_is_preferred)
1371 best_lpni_is_preferred = false;
1375 /* if this is a preferred peer use it */
1376 if (!best_lpni_is_preferred && lpni_is_preferred) {
1377 best_lpni_is_preferred = true;
1379 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1380 /* this is not the preferred peer so let's ignore
1386 if (lpni->lpni_txcredits < best_lpni_credits)
1387 /* We already have a peer that has more credits
1388 * available than this one. No need to consider
1389 * this peer further.
1392 else if (lpni->lpni_txcredits > best_lpni_credits)
1395 /* The best peer found so far and the current peer
1396 * have the same number of available credits let's
1397 * make sure to select between them using Round Robin
1399 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1402 best_lpni_is_preferred = lpni_is_preferred;
1403 best_lpni_healthv = lpni_healthv;
1404 best_sel_prio = lpni_sel_prio;
1406 best_lpni_credits = lpni->lpni_txcredits;
1409 /* if we still can't find a peer ni then we can't reach it */
1411 u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1412 LNET_NID_NET(dst_nid);
1413 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1414 libcfs_net2str(net_id));
1418 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1419 libcfs_nidstr(&best_lpni->lpni_nid));
1425 * Prerequisite: the best_ni should already be set in the sd
1426 * Find the best lpni.
1427 * If the net id is provided then restrict lpni selection on
1428 * that particular net.
1429 * Otherwise find any reachable lpni. When dealing with an MR
1430 * gateway and it has multiple lpnis which we can use
1431 * we want to select the best one from the list of reachable
1434 static inline struct lnet_peer_ni *
1435 lnet_find_best_lpni(struct lnet_ni *lni, struct lnet_nid *dst_nid,
1436 struct lnet_peer *peer, u32 net_id)
1438 struct lnet_peer_net *peer_net;
1440 /* find the best_lpni on any local network */
1441 if (net_id == LNET_NET_ANY) {
1442 struct lnet_peer_ni *best_lpni = NULL;
1443 struct lnet_peer_net *lpn;
1444 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1445 /* no net specified find any reachable peer ni */
1446 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1448 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1454 /* restrict on the specified net */
1455 peer_net = lnet_peer_get_net_locked(peer, net_id);
1457 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1463 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1465 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1468 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1471 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1474 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1480 /* Compare route priorities and hop counts */
1482 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1484 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1485 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1487 if (r1->lr_priority < r2->lr_priority)
1490 if (r1->lr_priority > r2->lr_priority)
1493 if (r1_hops < r2_hops)
1496 if (r1_hops > r2_hops)
1502 static struct lnet_route *
1503 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1504 struct lnet_peer_ni *remote_lpni,
1505 struct lnet_route **prev_route,
1506 struct lnet_peer_ni **gwni)
1508 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1509 struct lnet_route *best_route;
1510 struct lnet_route *last_route;
1511 struct lnet_route *route;
1513 bool best_rte_is_preferred = false;
1514 struct lnet_nid *gw_pnid;
1516 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1517 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1519 best_route = last_route = NULL;
1520 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1521 if (!lnet_is_route_alive(route))
1523 gw_pnid = &route->lr_gateway->lp_primary_nid;
1525 /* no protection on below fields, but it's harmless */
1526 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1529 /* if the best route found is in the preferred list then
1530 * tag it as preferred and use it later on. But if we
1531 * didn't find any routes which are on the preferred list
1532 * then just use the best route possible.
1534 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1536 if (!best_route || (rc && !best_rte_is_preferred)) {
1537 /* Restrict the selection of the router NI on the
1538 * src_net provided. If the src_net is LNET_NID_ANY,
1539 * then select the best interface available.
1541 lpni = lnet_find_best_lpni(NULL, NULL,
1546 "Gateway %s does not have a peer NI on net %s\n",
1547 libcfs_nidstr(gw_pnid),
1548 libcfs_net2str(src_net));
1553 if (rc && !best_rte_is_preferred) {
1554 /* This is the first preferred route we found,
1555 * so it beats any route found previously
1561 best_rte_is_preferred = true;
1562 CDEBUG(D_NET, "preferred gw = %s\n",
1563 libcfs_nidstr(gw_pnid));
1565 } else if ((!rc) && best_rte_is_preferred)
1566 /* The best route we found so far is in the preferred
1567 * list, so it beats any non-preferred route
1572 best_route = last_route = route;
1577 rc = lnet_compare_routes(route, best_route);
1581 /* Restrict the selection of the router NI on the
1582 * src_net provided. If the src_net is LNET_NID_ANY,
1583 * then select the best interface available.
1585 lpni = lnet_find_best_lpni(NULL, NULL, route->lr_gateway,
1589 "Gateway %s does not have a peer NI on net %s\n",
1590 libcfs_nidstr(gw_pnid),
1591 libcfs_net2str(src_net));
1601 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1605 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1612 *prev_route = last_route;
1618 static inline unsigned int
1619 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1621 if (dev_idx == UINT_MAX)
1624 if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1625 !ni->ni_net->net_lnd->lnd_get_dev_prio)
1628 return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1631 static struct lnet_ni *
1632 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1633 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1634 struct lnet_msg *msg, int md_cpt)
1636 struct lnet_libmd *md = msg->msg_md;
1637 unsigned int offset = msg->msg_offset;
1638 unsigned int shortest_distance;
1639 struct lnet_ni *ni = NULL;
1642 __u32 best_sel_prio;
1643 unsigned int best_dev_prio;
1645 unsigned int dev_idx = UINT_MAX;
1646 bool gpu = md ? (md->md_flags & LNET_MD_FLAG_GPU) : false;
1649 struct page *page = lnet_get_first_page(md, offset);
1651 dev_idx = lnet_get_dev_idx(page);
1655 * If there is no peer_ni that we can send to on this network,
1656 * then there is no point in looking for a new best_ni here.
1658 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1661 if (best_ni == NULL) {
1662 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1663 shortest_distance = UINT_MAX;
1664 best_dev_prio = UINT_MAX;
1665 best_credits = INT_MIN;
1667 best_ni_fatal = true;
1669 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1670 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1671 best_ni->ni_dev_cpt);
1672 best_credits = atomic_read(&best_ni->ni_tx_credits);
1673 best_healthv = atomic_read(&best_ni->ni_healthv);
1674 best_sel_prio = best_ni->ni_sel_priority;
1675 best_ni_fatal = atomic_read(&best_ni->ni_fatal_error_on);
1678 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1679 unsigned int distance;
1684 unsigned int ni_dev_prio;
1686 ni_credits = atomic_read(&ni->ni_tx_credits);
1687 ni_healthv = atomic_read(&ni->ni_healthv);
1688 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1689 ni_sel_prio = ni->ni_sel_priority;
1692 * calculate the distance from the CPT on which
1693 * the message memory is allocated to the CPT of
1694 * the NI's physical device
1696 distance = cfs_cpt_distance(lnet_cpt_table(),
1700 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1703 * All distances smaller than the NUMA range
1704 * are treated equally.
1706 if (!gpu && distance < lnet_numa_range)
1707 distance = lnet_numa_range;
1710 * Select on health, selection policy, direct dma prio,
1711 * shorter distance, available credits, then round-robin.
1714 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",
1715 libcfs_nidstr(&ni->ni_nid),
1716 ni_fatal ? "y" : "n", ni_credits, distance,
1717 ni->ni_seq, ni_sel_prio, ni_dev_prio, ni_healthv,
1718 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1720 best_ni_fatal ? "y" : "n", best_credits,
1722 (best_ni) ? best_ni->ni_seq : 0,
1723 best_sel_prio, best_dev_prio, best_healthv);
1727 if (ni_fatal && !best_ni_fatal)
1729 else if (!ni_fatal && best_ni_fatal)
1732 if (ni_healthv < best_healthv)
1734 else if (ni_healthv > best_healthv)
1737 if (ni_sel_prio > best_sel_prio)
1739 else if (ni_sel_prio < best_sel_prio)
1742 if (ni_dev_prio > best_dev_prio)
1744 else if (ni_dev_prio < best_dev_prio)
1747 if (distance > shortest_distance)
1749 else if (distance < shortest_distance)
1752 if (ni_credits < best_credits)
1754 else if (ni_credits > best_credits)
1757 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1761 best_sel_prio = ni_sel_prio;
1762 best_dev_prio = ni_dev_prio;
1763 shortest_distance = distance;
1764 best_healthv = ni_healthv;
1766 best_credits = ni_credits;
1767 best_ni_fatal = ni_fatal;
1770 CDEBUG(D_NET, "selected best_ni %s\n",
1771 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1777 lnet_reserved_msg(struct lnet_msg *msg)
1779 if (msg->msg_type == LNET_MSG_PUT) {
1780 if (msg->msg_hdr.msg.put.ptl_index == LNET_RESERVED_PORTAL)
1782 } else if (msg->msg_type == LNET_MSG_GET) {
1783 if (msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1789 /* Can the specified message trigger peer discovery?
1791 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1792 * because such traffic is required to perform discovery. We therefore
1793 * exclude all GET and PUT on that portal. We also exclude all ACK and
1794 * REPLY traffic, but that is because the portal is not tracked in the
1795 * message structure for these message types. We could restrict this
1796 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1799 lnet_msg_discovery(struct lnet_msg *msg)
1801 return !(lnet_reserved_msg(msg) || lnet_msg_is_response(msg));
1804 /* Is the specified message an LNet ping?
1807 lnet_msg_is_ping(struct lnet_msg *msg)
1809 if (msg->msg_type == LNET_MSG_GET &&
1810 msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1816 #define SRC_SPEC 0x0001
1817 #define SRC_ANY 0x0002
1818 #define LOCAL_DST 0x0004
1819 #define REMOTE_DST 0x0008
1820 #define MR_DST 0x0010
1821 #define NMR_DST 0x0020
1822 #define SND_RESP 0x0040
1824 /* The following to defines are used for return codes */
1825 #define REPEAT_SEND 0x1000
1826 #define PASS_THROUGH 0x2000
1828 /* The different cases lnet_select pathway needs to handle */
1829 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1830 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1831 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1832 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1833 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1834 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1835 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1836 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1839 lnet_handle_lo_send(struct lnet_send_data *sd)
1841 struct lnet_msg *msg = sd->sd_msg;
1842 int cpt = sd->sd_cpt;
1844 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1847 /* No send credit hassles with LOLND */
1848 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1849 msg->msg_hdr.dest_nid = the_lnet.ln_loni->ni_nid;
1850 if (!msg->msg_routing)
1851 msg->msg_hdr.src_nid = the_lnet.ln_loni->ni_nid;
1852 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1853 lnet_msg_commit(msg, cpt);
1854 msg->msg_txni = the_lnet.ln_loni;
1856 return LNET_CREDIT_OK;
1860 lnet_handle_send(struct lnet_send_data *sd)
1862 struct lnet_ni *best_ni = sd->sd_best_ni;
1863 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1864 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1865 struct lnet_msg *msg = sd->sd_msg;
1867 __u32 send_case = sd->sd_send_case;
1869 __u32 routing = send_case & REMOTE_DST;
1870 struct lnet_rsp_tracker *rspt;
1872 /* Increment sequence number of the selected peer, peer net,
1873 * local ni and local net so that we pick the next ones
1876 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1877 best_lpni->lpni_peer_net->lpn_seq =
1878 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1879 best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1880 the_lnet.ln_net_seq++;
1881 best_ni->ni_net->net_seq = the_lnet.ln_net_seq;
1882 best_ni->ni_seq = best_ni->ni_net->net_seq;
1884 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1885 libcfs_nidstr(&best_ni->ni_nid),
1886 best_ni->ni_seq, best_ni->ni_net->net_seq,
1887 atomic_read(&best_ni->ni_tx_credits),
1888 best_ni->ni_sel_priority,
1889 libcfs_nidstr(&best_lpni->lpni_nid),
1890 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1891 best_lpni->lpni_txcredits,
1892 best_lpni->lpni_sel_priority);
1895 * grab a reference on the peer_ni so it sticks around even if
1896 * we need to drop and relock the lnet_net_lock below.
1898 lnet_peer_ni_addref_locked(best_lpni);
1901 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1902 * message. This ensures that we get a CPT that is correct for
1903 * the NI when the NI has been restricted to a subset of all CPTs.
1904 * If the selected CPT differs from the one currently locked, we
1905 * must unlock and relock the lnet_net_lock(), and then check whether
1906 * the configuration has changed. We don't have a hold on the best_ni
1907 * yet, and it may have vanished.
1909 cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1910 if (sd->sd_cpt != cpt2) {
1911 __u32 seq = lnet_get_dlc_seq_locked();
1912 lnet_net_unlock(sd->sd_cpt);
1914 lnet_net_lock(sd->sd_cpt);
1915 if (seq != lnet_get_dlc_seq_locked()) {
1916 lnet_peer_ni_decref_locked(best_lpni);
1922 * store the best_lpni in the message right away to avoid having
1923 * to do the same operation under different conditions
1925 msg->msg_txpeer = best_lpni;
1926 msg->msg_txni = best_ni;
1929 * grab a reference for the best_ni since now it's in use in this
1930 * send. The reference will be dropped in lnet_finalize()
1932 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1935 * Always set the target.nid to the best peer picked. Either the
1936 * NID will be one of the peer NIDs selected, or the same NID as
1937 * what was originally set in the target or it will be the NID of
1938 * a router if this message should be routed
1940 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1943 * lnet_msg_commit assigns the correct cpt to the message, which
1944 * is used to decrement the correct refcount on the ni when it's
1945 * time to return the credits
1947 lnet_msg_commit(msg, sd->sd_cpt);
1950 * If we are routing the message then we keep the src_nid that was
1951 * set by the originator. If we are not routing then we are the
1952 * originator and set it here.
1954 if (!msg->msg_routing)
1955 msg->msg_hdr.src_nid = msg->msg_txni->ni_nid;
1958 msg->msg_target_is_router = 1;
1959 msg->msg_target.pid = LNET_PID_LUSTRE;
1961 * since we're routing we want to ensure that the
1962 * msg_hdr.dest_nid is set to the final destination. When
1963 * the router receives this message it knows how to route
1966 * final_dst_lpni is set at the beginning of the
1967 * lnet_select_pathway() function and is never changed.
1968 * It's safe to use it here.
1970 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1971 final_dst_lpni->lpni_peer_net->lpn_seq =
1972 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1973 final_dst_lpni->lpni_seq =
1974 final_dst_lpni->lpni_peer_net->lpn_seq;
1975 msg->msg_hdr.dest_nid = final_dst_lpni->lpni_nid;
1978 * if we're not routing set the dest_nid to the best peer
1979 * ni NID that we picked earlier in the algorithm.
1981 msg->msg_hdr.dest_nid = msg->msg_txpeer->lpni_nid;
1985 * if we have response tracker block update it with the next hop
1989 rspt = msg->msg_md->md_rspt_ptr;
1991 rspt->rspt_next_hop_nid =
1992 msg->msg_txpeer->lpni_nid;
1993 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1994 libcfs_nidstr(&rspt->rspt_next_hop_nid));
1998 rc = lnet_post_send_locked(msg, 0);
2001 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
2002 libcfs_nidstr(&msg->msg_hdr.src_nid),
2003 libcfs_nidstr(&msg->msg_txni->ni_nid),
2004 libcfs_nidstr(&sd->sd_src_nid),
2005 libcfs_nidstr(&msg->msg_hdr.dest_nid),
2006 libcfs_nidstr(&sd->sd_dst_nid),
2007 libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
2008 libcfs_nidstr(&sd->sd_rtr_nid),
2009 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
2015 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
2016 struct lnet_msg *msg)
2018 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
2019 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
2020 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
2021 libcfs_nidstr(&lni->ni_nid),
2022 libcfs_nidstr(&lpni->lpni_nid));
2023 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
2032 * use the source and destination NIDs as the pathway
2035 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2037 /* the destination lpni is set before we get here. */
2040 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2041 if (!sd->sd_best_ni) {
2042 CERROR("Can't send to %s: src %s is not a local nid\n",
2043 libcfs_nidstr(&sd->sd_dst_nid),
2044 libcfs_nidstr(&sd->sd_src_nid));
2048 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2050 return lnet_handle_send(sd);
2058 * Don't run the selection algorithm on the peer NIs. By specifying the
2059 * local NID, we're also saying that we should always use the destination NID
2060 * provided. This handles the case where we should be using the same
2061 * destination NID for the all the messages which belong to the same RPC
2065 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2067 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2068 if (!sd->sd_best_ni) {
2069 CERROR("Can't send to %s: src %s is not a local nid\n",
2070 libcfs_nidstr(&sd->sd_dst_nid),
2071 libcfs_nidstr(&sd->sd_src_nid));
2075 if (sd->sd_best_lpni &&
2076 nid_same(&sd->sd_best_lpni->lpni_nid,
2077 &the_lnet.ln_loni->ni_nid))
2078 return lnet_handle_lo_send(sd);
2079 else if (sd->sd_best_lpni)
2080 return lnet_handle_send(sd);
2082 CERROR("can't send to %s. no NI on %s\n",
2083 libcfs_nidstr(&sd->sd_dst_nid),
2084 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2086 return -EHOSTUNREACH;
2090 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2091 struct lnet_peer *peer,
2092 struct lnet_peer_net *peer_net,
2093 struct lnet_msg *msg,
2096 struct lnet_net *local_net;
2097 struct lnet_ni *best_ni;
2099 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2104 * Iterate through the NIs in this local Net and select
2105 * the NI to send from. The selection is determined by
2106 * these 3 criterion in the following priority:
2108 * 2. NI available credits
2111 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2112 peer, peer_net, msg, cpt);
2118 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2121 struct lnet_peer *peer;
2122 struct lnet_peer_ni *new_lpni;
2125 lnet_peer_ni_addref_locked(lpni);
2127 peer = lpni->lpni_peer_net->lpn_peer;
2129 if (lnet_peer_gw_discovery(peer)) {
2130 lnet_peer_ni_decref_locked(lpni);
2134 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2135 lnet_peer_ni_decref_locked(lpni);
2139 rc = lnet_discover_peer_locked(lpni, cpt, false);
2141 lnet_peer_ni_decref_locked(lpni);
2145 new_lpni = lnet_peer_ni_find_locked(&lpni->lpni_nid);
2147 lnet_peer_ni_decref_locked(lpni);
2151 peer = new_lpni->lpni_peer_net->lpn_peer;
2152 spin_lock(&peer->lp_lock);
2153 if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2154 /* The peer NI did not change and the peer is up to date.
2155 * Nothing more to do.
2157 spin_unlock(&peer->lp_lock);
2158 lnet_peer_ni_decref_locked(lpni);
2159 lnet_peer_ni_decref_locked(new_lpni);
2162 spin_unlock(&peer->lp_lock);
2164 /* Either the peer NI changed during discovery, or the peer isn't up
2165 * to date. In both cases we want to queue the message on the
2166 * (possibly new) peer's pending queue and queue the peer for discovery
2168 msg->msg_sending = 0;
2169 msg->msg_txpeer = NULL;
2170 lnet_net_unlock(cpt);
2171 lnet_peer_queue_message(peer, msg);
2174 lnet_peer_ni_decref_locked(lpni);
2175 lnet_peer_ni_decref_locked(new_lpni);
2177 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2178 msg, libcfs_nidstr(&peer->lp_primary_nid));
2180 return LNET_DC_WAIT;
2184 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2185 struct lnet_nid *dst_nid,
2186 struct lnet_peer_ni **gw_lpni,
2187 struct lnet_peer **gw_peer)
2190 struct lnet_peer *gw;
2191 struct lnet_peer *lp;
2192 struct lnet_peer_net *lpn;
2193 struct lnet_peer_net *best_lpn = NULL;
2194 struct lnet_remotenet *rnet, *best_rnet = NULL;
2195 struct lnet_route *best_route = NULL;
2196 struct lnet_route *last_route = NULL;
2197 struct lnet_peer_ni *lpni = NULL;
2198 struct lnet_peer_ni *gwni = NULL;
2199 bool route_found = false;
2200 struct lnet_nid *src_nid =
2201 !LNET_NID_IS_ANY(&sd->sd_src_nid) || !sd->sd_best_ni
2203 : &sd->sd_best_ni->ni_nid;
2204 int best_lpn_healthv = 0;
2205 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2207 CDEBUG(D_NET, "%s route (%s) from local NI %s to destination %s\n",
2208 LNET_NID_IS_ANY(&sd->sd_rtr_nid) ? "Lookup" : "Specified",
2209 libcfs_nidstr(&sd->sd_rtr_nid), libcfs_nidstr(src_nid),
2210 libcfs_nidstr(&sd->sd_dst_nid));
2212 /* If a router nid was specified then we are replying to a GET or
2213 * sending an ACK. In this case we use the gateway associated with the
2214 * specified router nid.
2216 if (!LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2217 gwni = lnet_peer_ni_find_locked(&sd->sd_rtr_nid);
2219 gw = gwni->lpni_peer_net->lpn_peer;
2220 lnet_peer_ni_decref_locked(gwni);
2221 if (gw->lp_rtr_refcount)
2224 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2225 libcfs_nidstr(&sd->sd_rtr_nid));
2230 if (sd->sd_msg->msg_routing || !LNET_NID_IS_ANY(src_nid)) {
2231 /* If I'm routing this message then I need to find the
2232 * next hop based on the destination NID
2234 * We also find next hop based on the destination NID
2235 * if the source NI was specified
2237 best_rnet = lnet_find_rnet_locked(LNET_NID_NET(&sd->sd_dst_nid));
2239 CERROR("Unable to send message from %s to %s - Route table may be misconfigured\n",
2240 (src_nid && LNET_NID_IS_ANY(src_nid)) ?
2242 libcfs_nidstr(src_nid),
2243 libcfs_nidstr(&sd->sd_dst_nid));
2244 return -EHOSTUNREACH;
2246 CDEBUG(D_NET, "best_rnet %s\n",
2247 libcfs_net2str(best_rnet->lrn_net));
2249 /* we've already looked up the initial lpni using
2252 lpni = sd->sd_best_lpni;
2253 /* the peer tree must be in existence */
2254 LASSERT(lpni && lpni->lpni_peer_net &&
2255 lpni->lpni_peer_net->lpn_peer);
2256 lp = lpni->lpni_peer_net->lpn_peer;
2258 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2259 /* is this remote network reachable? */
2260 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2267 CDEBUG(D_NET, "n[%s, %s] h[%d, %d], p[%u, %u], s[%d, %d]\n",
2268 libcfs_net2str(lpn->lpn_net_id),
2269 libcfs_net2str(best_lpn->lpn_net_id),
2271 best_lpn->lpn_healthv,
2272 lpn->lpn_sel_priority,
2273 best_lpn->lpn_sel_priority,
2277 /* select the preferred peer net */
2278 if (best_lpn_healthv > lpn->lpn_healthv)
2280 else if (best_lpn_healthv < lpn->lpn_healthv)
2283 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2285 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2288 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2291 best_lpn_healthv = lpn->lpn_healthv;
2292 best_lpn_sel_prio = lpn->lpn_sel_priority;
2298 CERROR("peer %s has no available nets\n",
2299 libcfs_nidstr(&sd->sd_dst_nid));
2300 return -EHOSTUNREACH;
2303 CDEBUG(D_NET, "selected best_lpn %s\n",
2304 libcfs_net2str(best_lpn->lpn_net_id));
2306 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2309 best_lpn->lpn_net_id);
2310 if (!sd->sd_best_lpni) {
2311 CERROR("peer %s is unreachable\n",
2312 libcfs_nidstr(&sd->sd_dst_nid));
2313 return -EHOSTUNREACH;
2316 /* We're attempting to round robin over the remote peer
2317 * NI's so update the final destination we selected
2319 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2323 * find the best route. Restrict the selection on the net of the
2324 * local NI if we've already picked the local NI to send from.
2325 * Otherwise, let's pick any route we can find and then find
2326 * a local NI we can reach the route's gateway on. Any route we
2327 * select will be reachable by virtue of the restriction we have
2328 * when adding a route.
2330 best_route = lnet_find_route_locked(best_rnet,
2331 LNET_NID_NET(src_nid),
2333 &last_route, &gwni);
2336 CERROR("no route to %s from %s\n",
2337 libcfs_nidstr(dst_nid),
2338 libcfs_nidstr(src_nid));
2339 return -EHOSTUNREACH;
2343 CERROR("Internal Error. Route expected to %s from %s\n",
2344 libcfs_nidstr(dst_nid),
2345 libcfs_nidstr(src_nid));
2349 gw = best_route->lr_gateway;
2350 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2354 * If the router checker is not active then discover the gateway here.
2355 * This ensures we are able to take advantage of multi-rail routing, but
2356 * if the router checker is active then we do not unecessarily delay
2357 * messages while the gateway is being checked by the dedicated monitor
2360 * NB: We're only checking the alive_router_check_interval here, rather
2361 * than calling lnet_router_checker_active(), because the other
2362 * conditions that are checked by that function are either
2363 * irrelevant (the_lnet.ln_routing) or must be true (list of routers
2366 if (alive_router_check_interval <= 0) {
2367 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2372 if (!sd->sd_best_ni) {
2373 lpn = gwni->lpni_peer_net;
2374 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2377 if (!sd->sd_best_ni) {
2378 CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2379 libcfs_net2str(lpn->lpn_net_id),
2380 libcfs_nidstr(&sd->sd_src_nid));
2389 * increment the sequence number since now we're sure we're
2390 * going to use this route
2392 if (LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2393 LASSERT(best_route && last_route);
2394 best_route->lr_seq = last_route->lr_seq + 1;
2405 * Remote destination
2406 * Non-MR destination
2410 * Remote destination
2413 * The handling of these two cases is similar. Even though the destination
2414 * can be MR or non-MR, we'll deal directly with the router.
2417 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2420 struct lnet_peer_ni *gw_lpni = NULL;
2421 struct lnet_peer *gw_peer = NULL;
2424 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2425 if (!sd->sd_best_ni) {
2426 CERROR("Can't send to %s: src %s is not a local nid\n",
2427 libcfs_nidstr(&sd->sd_dst_nid),
2428 libcfs_nidstr(&sd->sd_src_nid));
2432 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2433 &gw_lpni, &gw_peer);
2437 if (sd->sd_send_case & NMR_DST)
2439 * since the final destination is non-MR let's set its preferred
2440 * NID before we send
2442 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2446 * We're going to send to the gw found so let's set its
2449 sd->sd_peer = gw_peer;
2450 sd->sd_best_lpni = gw_lpni;
2452 return lnet_handle_send(sd);
2456 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2457 struct lnet_msg *msg, bool discovery)
2459 struct lnet_peer_net *lpn = NULL;
2460 struct lnet_peer_net *best_lpn = NULL;
2461 struct lnet_net *net = NULL;
2462 struct lnet_net *best_net = NULL;
2463 struct lnet_ni *best_ni = NULL;
2464 int best_lpn_healthv = 0;
2465 int best_net_healthv = 0;
2467 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2469 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2472 /* If lp_disc_net_id is set, this peer is a router undergoing
2473 * discovery, and this message is an LNet ping, then this may be a
2474 * discovery message and we need to select an NI on the peer net
2475 * specified by lp_disc_net_id
2477 if (peer->lp_disc_net_id &&
2478 (peer->lp_state & LNET_PEER_RTR_DISCOVERY) &&
2479 lnet_msg_is_ping(msg)) {
2480 best_lpn = lnet_peer_get_net_locked(peer, peer->lp_disc_net_id);
2481 if (best_lpn && lnet_get_net_locked(best_lpn->lpn_net_id))
2482 goto select_best_ni;
2486 * The peer can have multiple interfaces, some of them can be on
2487 * the local network and others on a routed network. We should
2488 * prefer the local network. However if the local network is not
2489 * available then we need to try the routed network
2492 /* go through all the peer nets and find the best_ni */
2493 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2495 * The peer's list of nets can contain non-local nets. We
2496 * want to only examine the local ones.
2498 net = lnet_get_net_locked(lpn->lpn_net_id);
2502 lpn_sel_prio = lpn->lpn_sel_priority;
2503 net_healthv = lnet_get_net_healthv_locked(net);
2504 net_sel_prio = net->net_sel_priority;
2510 "n[%s, %s] ph[%d, %d], pp[%u, %u], nh[%d, %d], np[%u, %u], ps[%u, %u], ns[%u, %u]\n",
2511 libcfs_net2str(lpn->lpn_net_id),
2512 libcfs_net2str(best_lpn->lpn_net_id),
2526 /* always select the lpn with the best health */
2527 if (best_lpn_healthv > lpn->lpn_healthv)
2529 else if (best_lpn_healthv < lpn->lpn_healthv)
2532 /* select the preferred peer and local nets */
2533 if (best_lpn_sel_prio < lpn_sel_prio)
2535 else if (best_lpn_sel_prio > lpn_sel_prio)
2538 if (best_net_healthv > net_healthv)
2540 else if (best_net_healthv < net_healthv)
2543 if (best_net_sel_prio < net_sel_prio)
2545 else if (best_net_sel_prio > net_sel_prio)
2548 if (best_lpn->lpn_seq < lpn->lpn_seq)
2550 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2553 /* round robin over the local networks */
2554 if (best_net->net_seq <= net->net_seq)
2558 best_net_healthv = net_healthv;
2559 best_net_sel_prio = net_sel_prio;
2560 best_lpn_healthv = lpn->lpn_healthv;
2561 best_lpn_sel_prio = lpn_sel_prio;
2567 /* Select the best NI on the same net as best_lpn chosen
2571 CDEBUG(D_NET, "selected best_lpn %s\n",
2572 libcfs_net2str(best_lpn->lpn_net_id));
2573 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2580 static struct lnet_ni *
2581 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2583 struct lnet_ni *best_ni = NULL;
2584 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2585 struct lnet_peer_ni *lpni_entry;
2588 * We must use a consistent source address when sending to a
2589 * non-MR peer. However, a non-MR peer can have multiple NIDs
2590 * on multiple networks, and we may even need to talk to this
2591 * peer on multiple networks -- certain types of
2592 * load-balancing configuration do this.
2594 * So we need to pick the NI the peer prefers for this
2595 * particular network.
2598 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2600 if (lpni_entry->lpni_pref_nnids == 0)
2602 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2603 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2611 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2613 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2615 struct lnet_ni *best_ni = NULL;
2618 * We must use a consistent source address when sending to a
2619 * non-MR peer. However, a non-MR peer can have multiple NIDs
2620 * on multiple networks, and we may even need to talk to this
2621 * peer on multiple networks -- certain types of
2622 * load-balancing configuration do this.
2624 * So we need to pick the NI the peer prefers for this
2625 * particular network.
2627 * An exception is traffic on LNET_RESERVED_PORTAL. Internal LNet
2628 * traffic doesn't care which source NI is used, and we don't actually
2629 * want to restrict local recovery pings to a single source NI.
2631 if (!lnet_reserved_msg(sd->sd_msg))
2632 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2636 best_ni = lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2637 sd->sd_best_lpni->lpni_peer_net,
2641 /* If there is no best_ni we don't have a route */
2643 CERROR("no path to %s from net %s\n",
2644 libcfs_nidstr(&sd->sd_best_lpni->lpni_nid),
2645 libcfs_net2str(sd->sd_best_lpni->lpni_net->net_id));
2646 return -EHOSTUNREACH;
2649 sd->sd_best_ni = best_ni;
2651 /* Set preferred NI if necessary. */
2652 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2659 * Source not specified
2663 * always use the same source NID for NMR peers
2664 * If we've talked to that peer before then we already have a preferred
2665 * source NI associated with it. Otherwise, we select a preferred local NI
2666 * and store it in the peer
2669 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2673 /* sd->sd_best_lpni is already set to the final destination */
2676 * At this point we should've created the peer ni and peer. If we
2677 * can't find it, then something went wrong. Instead of assert
2678 * output a relevant message and fail the send
2680 if (!sd->sd_best_lpni) {
2681 CERROR("Internal fault. Unable to send msg %s to %s. NID not known\n",
2682 lnet_msgtyp2str(sd->sd_msg->msg_type),
2683 libcfs_nidstr(&sd->sd_dst_nid));
2687 if (sd->sd_msg->msg_routing) {
2688 /* If I'm forwarding this message then I can choose any NI
2689 * on the destination peer net
2691 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2693 sd->sd_best_lpni->lpni_peer_net,
2696 if (!sd->sd_best_ni) {
2697 CERROR("Unable to forward message to %s. No local NI available\n",
2698 libcfs_nidstr(&sd->sd_dst_nid));
2702 rc = lnet_select_preferred_best_ni(sd);
2705 rc = lnet_handle_send(sd);
2711 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2714 * NOTE we've already handled the remote peer case. So we only
2715 * need to worry about the local case here.
2717 * if we're sending a response, ACK or reply, we need to send it
2718 * to the destination NID given to us. At this point we already
2719 * have the peer_ni we're suppose to send to, so just find the
2720 * best_ni on the peer net and use that. Since we're sending to an
2721 * MR peer then we can just run the selection algorithm on our
2722 * local NIs and pick the best one.
2724 if (sd->sd_send_case & SND_RESP) {
2726 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2727 sd->sd_best_lpni->lpni_peer_net,
2731 if (!sd->sd_best_ni) {
2733 * We're not going to deal with not able to send
2734 * a response to the provided final destination
2736 CERROR("Can't send response to %s. No local NI available\n",
2737 libcfs_nidstr(&sd->sd_dst_nid));
2738 return -EHOSTUNREACH;
2741 return lnet_handle_send(sd);
2745 * If we get here that means we're sending a fresh request, PUT or
2746 * GET, so we need to run our standard selection algorithm.
2747 * First find the best local interface that's on any of the peer's
2750 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2753 lnet_msg_discovery(sd->sd_msg));
2754 if (sd->sd_best_ni) {
2756 lnet_find_best_lpni(sd->sd_best_ni, &sd->sd_dst_nid,
2758 sd->sd_best_ni->ni_net->net_id);
2761 * if we're successful in selecting a peer_ni on the local
2762 * network, then send to it. Otherwise fall through and
2763 * try and see if we can reach it over another routed
2766 if (sd->sd_best_lpni &&
2767 nid_same(&sd->sd_best_lpni->lpni_nid,
2768 &the_lnet.ln_loni->ni_nid)) {
2770 * in case we initially started with a routed
2771 * destination, let's reset to local
2773 sd->sd_send_case &= ~REMOTE_DST;
2774 sd->sd_send_case |= LOCAL_DST;
2775 return lnet_handle_lo_send(sd);
2776 } else if (sd->sd_best_lpni) {
2778 * in case we initially started with a routed
2779 * destination, let's reset to local
2781 sd->sd_send_case &= ~REMOTE_DST;
2782 sd->sd_send_case |= LOCAL_DST;
2783 return lnet_handle_send(sd);
2786 CERROR("Internal Error. Expected to have a best_lpni: "
2788 libcfs_nidstr(&sd->sd_src_nid),
2789 libcfs_nidstr(&sd->sd_dst_nid));
2795 * Peer doesn't have a local network. Let's see if there is
2796 * a remote network we can reach it on.
2798 return PASS_THROUGH;
2803 * Source NID not specified
2808 * Source NID not speified
2809 * Remote destination
2812 * In both of these cases if we're sending a response, ACK or REPLY, then
2813 * we need to send to the destination NID provided.
2815 * In the remote case let's deal with MR routers.
2820 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2823 struct lnet_peer *gw_peer = NULL;
2824 struct lnet_peer_ni *gw_lpni = NULL;
2827 * handle sending a response to a remote peer here so we don't
2828 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2830 if (sd->sd_send_case & REMOTE_DST &&
2831 sd->sd_send_case & SND_RESP) {
2832 struct lnet_peer_ni *gw;
2833 struct lnet_peer *gw_peer;
2835 rc = lnet_handle_find_routed_path(
2836 sd, &sd->sd_dst_nid, &gw, &gw_peer);
2838 CERROR("Can't send response to %s. No route available\n",
2839 libcfs_nidstr(&sd->sd_dst_nid));
2840 return -EHOSTUNREACH;
2841 } else if (rc > 0) {
2845 sd->sd_best_lpni = gw;
2846 sd->sd_peer = gw_peer;
2848 return lnet_handle_send(sd);
2852 * Even though the NID for the peer might not be on a local network,
2853 * since the peer is MR there could be other interfaces on the
2854 * local network. In that case we'd still like to prefer the local
2855 * network over the routed network. If we're unable to do that
2856 * then we select the best router among the different routed networks,
2857 * and if the router is MR then we can deal with it as such.
2859 rc = lnet_handle_any_mr_dsta(sd);
2860 if (rc != PASS_THROUGH)
2864 * Now that we must route to the destination, we must consider the
2865 * MR case, where the destination has multiple interfaces, some of
2866 * which we can route to and others we do not. For this reason we
2867 * need to select the destination which we can route to and if
2868 * there are multiple, we need to round robin.
2870 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2871 &gw_lpni, &gw_peer);
2875 sd->sd_send_case &= ~LOCAL_DST;
2876 sd->sd_send_case |= REMOTE_DST;
2878 sd->sd_peer = gw_peer;
2879 sd->sd_best_lpni = gw_lpni;
2881 return lnet_handle_send(sd);
2885 * Source not specified
2886 * Remote destination
2889 * Must send to the specified peer NID using the same source NID that
2890 * we've used before. If it's the first time to talk to that peer then
2891 * find the source NI and assign it as preferred to that peer
2894 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2897 struct lnet_peer_ni *gw_lpni = NULL;
2898 struct lnet_peer *gw_peer = NULL;
2901 * Let's see if we have a preferred NI to talk to this NMR peer
2903 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2907 * find the router and that'll find the best NI if we didn't find
2910 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid, &gw_lpni,
2916 * set the best_ni we've chosen as the preferred one for
2919 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2921 /* we'll be sending to the gw */
2922 sd->sd_best_lpni = gw_lpni;
2923 sd->sd_peer = gw_peer;
2925 return lnet_handle_send(sd);
2929 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2932 * turn off the SND_RESP bit.
2933 * It will be checked in the case handling
2935 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2937 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2938 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2939 (send_case & SRC_SPEC) ? libcfs_nidstr(&sd->sd_src_nid) : "",
2940 (send_case & MR_DST) ? "MR: " : "NMR: ",
2941 libcfs_nidstr(&sd->sd_dst_nid),
2942 (send_case & LOCAL_DST) ? "local" : "routed");
2944 switch (send_case) {
2946 * For all cases where the source is specified, we should always
2947 * use the destination NID, whether it's an MR destination or not,
2948 * since we're continuing a series of related messages for the
2951 case SRC_SPEC_LOCAL_NMR_DST:
2952 return lnet_handle_spec_local_nmr_dst(sd);
2953 case SRC_SPEC_LOCAL_MR_DST:
2954 return lnet_handle_spec_local_mr_dst(sd);
2955 case SRC_SPEC_ROUTER_NMR_DST:
2956 case SRC_SPEC_ROUTER_MR_DST:
2957 return lnet_handle_spec_router_dst(sd);
2958 case SRC_ANY_LOCAL_NMR_DST:
2959 return lnet_handle_any_local_nmr_dst(sd);
2960 case SRC_ANY_LOCAL_MR_DST:
2961 case SRC_ANY_ROUTER_MR_DST:
2962 return lnet_handle_any_mr_dst(sd);
2963 case SRC_ANY_ROUTER_NMR_DST:
2964 return lnet_handle_any_router_nmr_dst(sd);
2966 CERROR("Unknown send case\n");
2972 lnet_select_pathway(struct lnet_nid *src_nid,
2973 struct lnet_nid *dst_nid,
2974 struct lnet_msg *msg,
2975 struct lnet_nid *rtr_nid)
2977 struct lnet_peer_ni *lpni;
2978 struct lnet_peer *peer;
2979 struct lnet_send_data send_data;
2982 __u32 send_case = 0;
2984 bool mr_forwarding_allowed;
2986 memset(&send_data, 0, sizeof(send_data));
2989 * get an initial CPT to use for locking. The idea here is not to
2990 * serialize the calls to select_pathway, so that as many
2991 * operations can run concurrently as possible. To do that we use
2992 * the CPT where this call is being executed. Later on when we
2993 * determine the CPT to use in lnet_message_commit, we switch the
2994 * lock and check if there was any configuration change. If none,
2995 * then we proceed, if there is, then we restart the operation.
2997 cpt = lnet_net_lock_current();
2999 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
3000 if (md_cpt == CFS_CPT_ANY)
3006 * If we're being asked to send to the loopback interface, there
3007 * is no need to go through any selection. We can just shortcut
3008 * the entire process and send over lolnd
3010 send_data.sd_msg = msg;
3011 send_data.sd_cpt = cpt;
3012 if (nid_is_lo0(dst_nid)) {
3013 rc = lnet_handle_lo_send(&send_data);
3014 lnet_net_unlock(cpt);
3019 * find an existing peer_ni, or create one and mark it as having been
3020 * created due to network traffic. This call will create the
3021 * peer->peer_net->peer_ni tree.
3023 lpni = lnet_peerni_by_nid_locked(dst_nid, NULL, cpt);
3025 lnet_net_unlock(cpt);
3026 return PTR_ERR(lpni);
3030 * Cache the original src_nid and rtr_nid. If we need to resend the
3031 * message then we'll need to know whether the src_nid was originally
3032 * specified for this message. If it was originally specified,
3033 * then we need to keep using the same src_nid since it's
3034 * continuing the same sequence of messages. Similarly, rtr_nid will
3035 * affect our choice of next hop.
3038 msg->msg_src_nid_param = *src_nid;
3040 msg->msg_src_nid_param = LNET_ANY_NID;
3042 msg->msg_rtr_nid_param = *rtr_nid;
3044 msg->msg_rtr_nid_param = LNET_ANY_NID;
3047 * If necessary, perform discovery on the peer that owns this peer_ni.
3048 * Note, this can result in the ownership of this peer_ni changing
3049 * to another peer object.
3051 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
3053 lnet_peer_ni_decref_locked(lpni);
3054 lnet_net_unlock(cpt);
3057 lnet_peer_ni_decref_locked(lpni);
3059 peer = lpni->lpni_peer_net->lpn_peer;
3062 * Identify the different send cases
3064 if (!src_nid || LNET_NID_IS_ANY(src_nid)) {
3065 send_case |= SRC_ANY;
3066 if (lnet_get_net_locked(LNET_NID_NET(dst_nid)))
3067 send_case |= LOCAL_DST;
3069 send_case |= REMOTE_DST;
3071 send_case |= SRC_SPEC;
3072 if (LNET_NID_NET(src_nid) == LNET_NID_NET(dst_nid))
3073 send_case |= LOCAL_DST;
3075 send_case |= REMOTE_DST;
3079 if (msg->msg_routing && (send_case & LOCAL_DST))
3082 /* Determine whether to allow MR forwarding for this message.
3083 * NB: MR forwarding is allowed if the message originator and the
3084 * destination are both MR capable, and the destination lpni that was
3085 * originally chosen by the originator is unhealthy or down.
3086 * We check the MR capability of the destination further below
3088 mr_forwarding_allowed = false;
3090 struct lnet_peer *src_lp;
3091 struct lnet_peer_ni *src_lpni;
3093 src_lpni = lnet_peerni_by_nid_locked(&msg->msg_hdr.src_nid,
3095 /* We don't fail the send if we hit any errors here. We'll just
3096 * try to send it via non-multi-rail criteria
3098 if (!IS_ERR(src_lpni)) {
3099 /* Drop ref taken by lnet_peerni_by_nid_locked() */
3100 lnet_peer_ni_decref_locked(src_lpni);
3101 src_lp = lpni->lpni_peer_net->lpn_peer;
3102 if (lnet_peer_is_multi_rail(src_lp) &&
3103 !lnet_is_peer_ni_alive(lpni))
3104 mr_forwarding_allowed = true;
3107 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3108 mr_forwarding_allowed ? "allowed" : "not allowed");
3112 * Deal with the peer as NMR in the following cases:
3113 * 1. the peer is NMR
3114 * 2. We're trying to recover a specific peer NI
3115 * 3. I'm a router sending to the final destination and MR forwarding is
3116 * not allowed for this message (as determined above).
3117 * In this case the source of the message would've
3118 * already selected the final destination so my job
3119 * is to honor the selection.
3121 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3122 (final_hop && !mr_forwarding_allowed))
3123 send_case |= NMR_DST;
3125 send_case |= MR_DST;
3127 if (lnet_msg_is_response(msg))
3128 send_case |= SND_RESP;
3130 /* assign parameters to the send_data */
3132 send_data.sd_rtr_nid = *rtr_nid;
3134 send_data.sd_rtr_nid = LNET_ANY_NID;
3136 send_data.sd_src_nid = *src_nid;
3138 send_data.sd_src_nid = LNET_ANY_NID;
3139 send_data.sd_dst_nid = *dst_nid;
3140 send_data.sd_best_lpni = lpni;
3142 * keep a pointer to the final destination in case we're going to
3143 * route, so we'll need to access it later
3145 send_data.sd_final_dst_lpni = lpni;
3146 send_data.sd_peer = peer;
3147 send_data.sd_md_cpt = md_cpt;
3148 send_data.sd_send_case = send_case;
3150 rc = lnet_handle_send_case_locked(&send_data);
3153 * Update the local cpt since send_data.sd_cpt might've been
3154 * updated as a result of calling lnet_handle_send_case_locked().
3156 cpt = send_data.sd_cpt;
3158 if (rc == REPEAT_SEND)
3161 lnet_net_unlock(cpt);
3167 lnet_send(struct lnet_nid *src_nid, struct lnet_msg *msg,
3168 struct lnet_nid *rtr_nid)
3170 struct lnet_nid *dst_nid = &msg->msg_target.nid;
3173 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3174 LASSERT(msg->msg_txpeer == NULL);
3175 LASSERT(msg->msg_txni == NULL);
3176 LASSERT(!msg->msg_sending);
3177 LASSERT(!msg->msg_target_is_router);
3178 LASSERT(!msg->msg_receiving);
3180 msg->msg_sending = 1;
3182 LASSERT(!msg->msg_tx_committed);
3184 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3186 if (rc == -EHOSTUNREACH)
3187 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3189 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3193 if (rc == LNET_CREDIT_OK)
3194 lnet_ni_send(msg->msg_txni, msg);
3196 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3200 enum lnet_mt_event_type {
3201 MT_TYPE_LOCAL_NI = 0,
3205 struct lnet_mt_event_info {
3206 enum lnet_mt_event_type mt_type;
3207 struct lnet_nid mt_nid;
3210 /* called with res_lock held */
3212 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3214 struct lnet_rsp_tracker *rspt;
3217 * msg has a refcount on the MD so the MD is not going away.
3218 * The rspt queue for the cpt is protected by
3219 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3221 if (!md->md_rspt_ptr)
3224 rspt = md->md_rspt_ptr;
3227 LASSERT(rspt->rspt_cpt == cpt);
3229 md->md_rspt_ptr = NULL;
3231 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3233 * The monitor thread has invalidated this handle because the
3234 * response timed out, but it failed to lookup the MD. That
3235 * means this response tracker is on the zombie list. We can
3236 * safely remove it under the resource lock (held by caller) and
3237 * free the response tracker block.
3239 list_del(&rspt->rspt_on_list);
3240 lnet_rspt_free(rspt, cpt);
3243 * invalidate the handle to indicate that a response has been
3244 * received, which will then lead the monitor thread to clean up
3247 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3252 lnet_clean_zombie_rstqs(void)
3254 struct lnet_rsp_tracker *rspt, *tmp;
3257 cfs_cpt_for_each(i, lnet_cpt_table()) {
3258 list_for_each_entry_safe(rspt, tmp,
3259 the_lnet.ln_mt_zombie_rstqs[i],
3261 list_del(&rspt->rspt_on_list);
3262 lnet_rspt_free(rspt, i);
3266 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3270 lnet_finalize_expired_responses(void)
3272 struct lnet_libmd *md;
3273 struct lnet_rsp_tracker *rspt, *tmp;
3277 if (the_lnet.ln_mt_rstq == NULL)
3280 cfs_cpt_for_each(i, lnet_cpt_table()) {
3281 LIST_HEAD(local_queue);
3284 if (!the_lnet.ln_mt_rstq[i]) {
3288 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3293 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3295 * The rspt mdh will be invalidated when a response
3296 * is received or whenever we want to discard the
3297 * block the monitor thread will walk the queue
3298 * and clean up any rsts with an invalid mdh.
3299 * The monitor thread will walk the queue until
3300 * the first unexpired rspt block. This means that
3301 * some rspt blocks which received their
3302 * corresponding responses will linger in the
3303 * queue until they are cleaned up eventually.
3306 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3308 list_del(&rspt->rspt_on_list);
3309 lnet_rspt_free(rspt, i);
3313 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3314 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3315 struct lnet_peer_ni *lpni;
3316 struct lnet_nid nid;
3318 md = lnet_handle2md(&rspt->rspt_mdh);
3320 /* MD has been queued for unlink, but
3321 * rspt hasn't been detached (Note we've
3322 * checked above that the rspt_mdh is
3323 * valid). Since we cannot lookup the MD
3324 * we're unable to detach the rspt
3325 * ourselves. Thus, move the rspt to the
3326 * zombie list where we'll wait for
3328 * 1. The remaining operations on the
3329 * MD to complete. In this case the
3330 * final operation will result in
3331 * lnet_msg_detach_md()->
3332 * lnet_detach_rsp_tracker() where
3333 * we will clean up this response
3335 * 2. LNet to shutdown. In this case
3336 * we'll wait until after all LND Nets
3337 * have shutdown and then we can
3338 * safely free any remaining response
3339 * tracker blocks on the zombie list.
3340 * Note: We need to hold the resource
3341 * lock when adding to the zombie list
3342 * because we may have concurrent access
3343 * with lnet_detach_rsp_tracker().
3345 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3346 list_move(&rspt->rspt_on_list,
3347 the_lnet.ln_mt_zombie_rstqs[i]);
3351 LASSERT(md->md_rspt_ptr == rspt);
3352 md->md_rspt_ptr = NULL;
3355 LNetMDUnlink(rspt->rspt_mdh);
3357 nid = rspt->rspt_next_hop_nid;
3359 list_del(&rspt->rspt_on_list);
3360 lnet_rspt_free(rspt, i);
3362 /* If we're shutting down we just want to clean
3363 * up the rspt blocks
3365 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3369 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3373 "Response timeout: md = %p: nid = %s\n",
3374 md, libcfs_nidstr(&nid));
3377 * If there is a timeout on the response
3378 * from the next hop decrement its health
3379 * value so that we don't use it
3382 lpni = lnet_peer_ni_find_locked(&nid);
3384 lnet_handle_remote_failure_locked(lpni);
3385 lnet_peer_ni_decref_locked(lpni);
3394 if (!list_empty(&local_queue)) {
3396 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3403 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3405 struct lnet_msg *msg;
3407 while (!list_empty(resendq)) {
3408 struct lnet_peer_ni *lpni;
3410 msg = list_entry(resendq->next, struct lnet_msg,
3413 list_del_init(&msg->msg_list);
3415 lpni = lnet_peer_ni_find_locked(&msg->msg_hdr.dest_nid);
3417 lnet_net_unlock(cpt);
3418 CERROR("Expected that a peer is already created for %s\n",
3419 libcfs_nidstr(&msg->msg_hdr.dest_nid));
3420 msg->msg_no_resend = true;
3421 lnet_finalize(msg, -EFAULT);
3426 lnet_peer_ni_decref_locked(lpni);
3428 lnet_net_unlock(cpt);
3429 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3430 libcfs_nidstr(&msg->msg_src_nid_param),
3431 libcfs_idstr(&msg->msg_target),
3432 lnet_msgtyp2str(msg->msg_type),
3434 msg->msg_retry_count);
3435 rc = lnet_send(&msg->msg_src_nid_param, msg,
3436 &msg->msg_rtr_nid_param);
3438 CERROR("Error sending %s to %s: %d\n",
3439 lnet_msgtyp2str(msg->msg_type),
3440 libcfs_idstr(&msg->msg_target), rc);
3441 msg->msg_no_resend = true;
3442 lnet_finalize(msg, rc);
3446 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3452 lnet_resend_pending_msgs(void)
3456 cfs_cpt_for_each(i, lnet_cpt_table()) {
3458 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3463 /* called with cpt and ni_lock held */
3465 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3467 struct lnet_handle_md recovery_mdh;
3469 LNetInvalidateMDHandle(&recovery_mdh);
3471 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3473 recovery_mdh = ni->ni_ping_mdh;
3474 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3477 lnet_net_unlock(cpt);
3478 if (!LNetMDHandleIsInvalid(recovery_mdh))
3479 LNetMDUnlink(recovery_mdh);
3485 lnet_recover_local_nis(void)
3487 struct lnet_mt_event_info *ev_info;
3488 LIST_HEAD(processed_list);
3489 LIST_HEAD(local_queue);
3490 struct lnet_handle_md mdh;
3491 struct lnet_ni *tmp;
3493 struct lnet_nid nid;
3499 * splice the recovery queue on a local queue. We will iterate
3500 * through the local queue and update it as needed. Once we're
3501 * done with the traversal, we'll splice the local queue back on
3502 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3503 * will be traversed in the next iteration.
3506 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3510 now = ktime_get_seconds();
3512 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3514 * if an NI is being deleted or it is now healthy, there
3515 * is no need to keep it around in the recovery queue.
3516 * The monitor thread is the only thread responsible for
3517 * removing the NI from the recovery queue.
3518 * Multiple threads can be adding NIs to the recovery
3521 healthv = atomic_read(&ni->ni_healthv);
3525 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3526 healthv == LNET_MAX_HEALTH_VALUE) {
3527 list_del_init(&ni->ni_recovery);
3528 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3530 lnet_ni_decref_locked(ni, 0);
3536 * if the local NI failed recovery we must unlink the md.
3537 * But we want to keep the local_ni on the recovery queue
3538 * so we can continue the attempts to recover it.
3540 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3541 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3542 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3548 if (now < ni->ni_next_ping) {
3555 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3556 libcfs_nidstr(&ni->ni_nid));
3559 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3560 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3563 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3565 CERROR("out of memory. Can't recover %s\n",
3566 libcfs_nidstr(&ni->ni_nid));
3568 ni->ni_recovery_state &=
3569 ~LNET_NI_RECOVERY_PENDING;
3574 mdh = ni->ni_ping_mdh;
3576 * Invalidate the ni mdh in case it's deleted.
3577 * We'll unlink the mdh in this case below.
3579 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3583 * remove the NI from the local queue and drop the
3584 * reference count to it while we're recovering
3585 * it. The reason for that, is that the NI could
3586 * be deleted, and the way the code is structured
3587 * is if we don't drop the NI, then the deletion
3588 * code will enter a loop waiting for the
3589 * reference count to be removed while holding the
3590 * ln_mutex_lock(). When we look up the peer to
3591 * send to in lnet_select_pathway() we will try to
3592 * lock the ln_mutex_lock() as well, leading to
3593 * a deadlock. By dropping the refcount and
3594 * removing it from the list, we allow for the NI
3595 * to be removed, then we use the cached NID to
3596 * look it up again. If it's gone, then we just
3597 * continue examining the rest of the queue.
3600 list_del_init(&ni->ni_recovery);
3601 lnet_ni_decref_locked(ni, 0);
3604 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3605 ev_info->mt_nid = nid;
3606 rc = lnet_send_ping(&nid, &mdh, LNET_PING_INFO_MIN_SIZE,
3607 ev_info, the_lnet.ln_mt_handler,
3609 /* lookup the nid again */
3611 ni = lnet_nid_to_ni_locked(&nid, 0);
3614 * the NI has been deleted when we dropped
3621 ni->ni_ping_count++;
3623 ni->ni_ping_mdh = mdh;
3624 lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3629 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3638 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3639 * reexamined in the next iteration.
3641 list_splice_init(&processed_list, &local_queue);
3643 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3648 lnet_resendqs_create(void)
3650 struct list_head **resendqs;
3651 resendqs = lnet_create_array_of_queues();
3656 lnet_net_lock(LNET_LOCK_EX);
3657 the_lnet.ln_mt_resendqs = resendqs;
3658 lnet_net_unlock(LNET_LOCK_EX);
3664 lnet_clean_local_ni_recoveryq(void)
3668 /* This is only called when the monitor thread has stopped */
3671 while ((ni = list_first_entry_or_null(&the_lnet.ln_mt_localNIRecovq,
3673 ni_recovery)) != NULL) {
3674 list_del_init(&ni->ni_recovery);
3676 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3678 lnet_ni_decref_locked(ni, 0);
3685 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3688 struct lnet_handle_md recovery_mdh;
3690 LNetInvalidateMDHandle(&recovery_mdh);
3692 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3693 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3694 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3696 spin_unlock(&lpni->lpni_lock);
3697 lnet_net_unlock(cpt);
3698 if (!LNetMDHandleIsInvalid(recovery_mdh))
3699 LNetMDUnlink(recovery_mdh);
3701 spin_lock(&lpni->lpni_lock);
3705 lnet_clean_peer_ni_recoveryq(void)
3707 struct lnet_peer_ni *lpni, *tmp;
3709 lnet_net_lock(LNET_LOCK_EX);
3711 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3713 list_del_init(&lpni->lpni_recovery);
3714 spin_lock(&lpni->lpni_lock);
3715 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3716 spin_unlock(&lpni->lpni_lock);
3717 lnet_peer_ni_decref_locked(lpni);
3720 lnet_net_unlock(LNET_LOCK_EX);
3724 lnet_clean_resendqs(void)
3726 struct lnet_msg *msg, *tmp;
3730 cfs_cpt_for_each(i, lnet_cpt_table()) {
3732 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3734 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3735 list_del_init(&msg->msg_list);
3736 msg->msg_no_resend = true;
3737 lnet_finalize(msg, -ESHUTDOWN);
3741 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3745 lnet_recover_peer_nis(void)
3747 struct lnet_mt_event_info *ev_info;
3748 LIST_HEAD(processed_list);
3749 LIST_HEAD(local_queue);
3750 struct lnet_handle_md mdh;
3751 struct lnet_peer_ni *lpni;
3752 struct lnet_peer_ni *tmp;
3753 struct lnet_nid nid;
3759 * Always use cpt 0 for locking across all interactions with
3760 * ln_mt_peerNIRecovq
3763 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3767 now = ktime_get_seconds();
3769 list_for_each_entry_safe(lpni, tmp, &local_queue,
3772 * The same protection strategy is used here as is in the
3773 * local recovery case.
3776 healthv = atomic_read(&lpni->lpni_healthv);
3777 spin_lock(&lpni->lpni_lock);
3778 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3779 healthv == LNET_MAX_HEALTH_VALUE) {
3780 list_del_init(&lpni->lpni_recovery);
3781 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3782 spin_unlock(&lpni->lpni_lock);
3783 lnet_peer_ni_decref_locked(lpni);
3789 * If the peer NI has failed recovery we must unlink the
3790 * md. But we want to keep the peer ni on the recovery
3791 * queue so we can try to continue recovering it
3793 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3794 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3795 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3798 spin_unlock(&lpni->lpni_lock);
3800 if (now < lpni->lpni_next_ping) {
3808 * NOTE: we're racing with peer deletion from user space.
3809 * It's possible that a peer is deleted after we check its
3810 * state. In this case the recovery can create a new peer
3812 spin_lock(&lpni->lpni_lock);
3813 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3814 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3815 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3816 spin_unlock(&lpni->lpni_lock);
3818 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3820 CERROR("out of memory. Can't recover %s\n",
3821 libcfs_nidstr(&lpni->lpni_nid));
3822 spin_lock(&lpni->lpni_lock);
3823 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3824 spin_unlock(&lpni->lpni_lock);
3828 /* look at the comments in lnet_recover_local_nis() */
3829 mdh = lpni->lpni_recovery_ping_mdh;
3830 nid = lpni->lpni_nid;
3831 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3833 list_del_init(&lpni->lpni_recovery);
3834 lnet_peer_ni_decref_locked(lpni);
3837 ev_info->mt_type = MT_TYPE_PEER_NI;
3838 ev_info->mt_nid = nid;
3839 rc = lnet_send_ping(&nid, &mdh, LNET_PING_INFO_MIN_SIZE,
3840 ev_info, the_lnet.ln_mt_handler,
3844 * lnet_peer_ni_find_locked() grabs a refcount for
3845 * us. No need to take it explicitly.
3847 lpni = lnet_peer_ni_find_locked(&nid);
3854 lpni->lpni_ping_count++;
3856 lpni->lpni_recovery_ping_mdh = mdh;
3858 lnet_peer_ni_add_to_recoveryq_locked(lpni,
3862 spin_lock(&lpni->lpni_lock);
3863 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3864 spin_unlock(&lpni->lpni_lock);
3867 /* Drop the ref taken by lnet_peer_ni_find_locked() */
3868 lnet_peer_ni_decref_locked(lpni);
3871 spin_unlock(&lpni->lpni_lock);
3874 list_splice_init(&processed_list, &local_queue);
3876 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3881 lnet_monitor_thread(void *arg)
3883 time64_t rsp_timeout = 0;
3886 wait_for_completion(&the_lnet.ln_started);
3888 * The monitor thread takes care of the following:
3889 * 1. Checks the aliveness of routers
3890 * 2. Checks if there are messages on the resend queue to resend
3892 * 3. Check if there are any NIs on the local recovery queue and
3894 * 4. Checks if there are any NIs on the remote recovery queue
3897 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3898 now = ktime_get_real_seconds();
3900 if (lnet_router_checker_active())
3901 lnet_check_routers();
3903 lnet_resend_pending_msgs();
3905 if (now >= rsp_timeout) {
3906 lnet_finalize_expired_responses();
3907 rsp_timeout = now + (lnet_transaction_timeout / 2);
3910 lnet_recover_local_nis();
3911 lnet_recover_peer_nis();
3914 * TODO do we need to check if we should sleep without
3915 * timeout? Technically, an active system will always
3916 * have messages in flight so this check will always
3917 * evaluate to false. And on an idle system do we care
3918 * if we wake up every 1 second? Although, we've seen
3919 * cases where we get a complaint that an idle thread
3920 * is waking up unnecessarily.
3922 wait_for_completion_interruptible_timeout(
3923 &the_lnet.ln_mt_wait_complete,
3924 cfs_time_seconds(1));
3925 /* Must re-init the completion before testing anything,
3926 * including ln_mt_state.
3928 reinit_completion(&the_lnet.ln_mt_wait_complete);
3932 lnet_net_lock(LNET_LOCK_EX);
3933 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3934 lnet_net_unlock(LNET_LOCK_EX);
3936 /* signal that the monitor thread is exiting */
3937 up(&the_lnet.ln_mt_signal);
3945 * Returns == 0 if success
3946 * Returns > 0 if LNetMDBind or prior fails
3947 * Returns < 0 if LNetGet fails
3950 lnet_send_ping(struct lnet_nid *dest_nid,
3951 struct lnet_handle_md *mdh, int bytes,
3952 void *user_data, lnet_handler_t handler, bool recovery)
3954 struct lnet_md md = { NULL };
3955 struct lnet_processid id;
3956 struct lnet_ping_buffer *pbuf;
3959 if (LNET_NID_IS_ANY(dest_nid)) {
3964 pbuf = lnet_ping_buffer_alloc(bytes, GFP_NOFS);
3970 /* initialize md content */
3971 md.start = &pbuf->pb_info;
3973 md.threshold = 2; /* GET/REPLY */
3975 md.options = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3976 md.user_ptr = user_data;
3977 md.handler = handler;
3979 rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3981 lnet_ping_buffer_decref(pbuf);
3982 CERROR("Can't bind MD: %d\n", rc);
3983 rc = -rc; /* change the rc to positive */
3986 id.pid = LNET_PID_LUSTRE;
3989 rc = LNetGet(NULL, *mdh, &id,
3990 LNET_RESERVED_PORTAL,
3991 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3994 goto fail_unlink_md;
4000 LNetInvalidateMDHandle(mdh);
4006 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
4007 int status, bool send, bool unlink_event)
4009 struct lnet_nid *nid = &ev_info->mt_nid;
4011 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
4015 ni = lnet_nid_to_ni_locked(nid, 0);
4021 if (!send || (send && status != 0))
4022 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
4024 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
4029 CERROR("local NI (%s) recovery failed with %d\n",
4030 libcfs_nidstr(nid), status);
4034 * need to increment healthv for the ni here, because in
4035 * the lnet_finalize() path we don't have access to this
4036 * NI. And in order to get access to it, we'll need to
4037 * carry forward too much information.
4038 * In the peer case, it'll naturally be incremented
4041 lnet_inc_healthv(&ni->ni_healthv,
4042 lnet_health_sensitivity);
4044 struct lnet_peer_ni *lpni;
4047 cpt = lnet_net_lock_current();
4048 lpni = lnet_peer_ni_find_locked(nid);
4050 lnet_net_unlock(cpt);
4053 spin_lock(&lpni->lpni_lock);
4054 if (!send || (send && status != 0))
4055 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
4057 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
4058 spin_unlock(&lpni->lpni_lock);
4059 lnet_peer_ni_decref_locked(lpni);
4060 lnet_net_unlock(cpt);
4063 CERROR("peer NI (%s) recovery failed with %d\n",
4064 libcfs_nidstr(nid), status);
4069 lnet_mt_event_handler(struct lnet_event *event)
4071 struct lnet_mt_event_info *ev_info = event->md_user_ptr;
4072 struct lnet_ping_buffer *pbuf;
4074 /* TODO: remove assert */
4075 LASSERT(event->type == LNET_EVENT_REPLY ||
4076 event->type == LNET_EVENT_SEND ||
4077 event->type == LNET_EVENT_UNLINK);
4079 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
4082 switch (event->type) {
4083 case LNET_EVENT_UNLINK:
4084 CDEBUG(D_NET, "%s recovery ping unlinked\n",
4085 libcfs_nidstr(&ev_info->mt_nid));
4087 case LNET_EVENT_REPLY:
4088 lnet_handle_recovery_reply(ev_info, event->status, false,
4089 event->type == LNET_EVENT_UNLINK);
4091 case LNET_EVENT_SEND:
4092 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4093 libcfs_nidstr(&ev_info->mt_nid),
4094 (event->status) ? "unsuccessfully" :
4095 "successfully", event->status);
4096 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4099 CERROR("Unexpected event: %d\n", event->type);
4102 if (event->unlinked) {
4103 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4104 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4105 lnet_ping_buffer_decref(pbuf);
4110 lnet_rsp_tracker_create(void)
4112 struct list_head **rstqs;
4113 rstqs = lnet_create_array_of_queues();
4118 the_lnet.ln_mt_rstq = rstqs;
4124 lnet_rsp_tracker_clean(void)
4126 lnet_finalize_expired_responses();
4128 cfs_percpt_free(the_lnet.ln_mt_rstq);
4129 the_lnet.ln_mt_rstq = NULL;
4132 int lnet_monitor_thr_start(void)
4135 struct task_struct *task;
4137 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4140 rc = lnet_resendqs_create();
4144 rc = lnet_rsp_tracker_create();
4148 sema_init(&the_lnet.ln_mt_signal, 0);
4150 lnet_net_lock(LNET_LOCK_EX);
4151 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4152 lnet_net_unlock(LNET_LOCK_EX);
4153 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4156 CERROR("Can't start monitor thread: %d\n", rc);
4163 lnet_net_lock(LNET_LOCK_EX);
4164 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4165 lnet_net_unlock(LNET_LOCK_EX);
4166 /* block until event callback signals exit */
4167 down(&the_lnet.ln_mt_signal);
4169 lnet_net_lock(LNET_LOCK_EX);
4170 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4171 lnet_net_unlock(LNET_LOCK_EX);
4172 lnet_rsp_tracker_clean();
4173 lnet_clean_local_ni_recoveryq();
4174 lnet_clean_peer_ni_recoveryq();
4175 lnet_clean_resendqs();
4176 the_lnet.ln_mt_handler = NULL;
4179 lnet_rsp_tracker_clean();
4180 lnet_clean_local_ni_recoveryq();
4181 lnet_clean_peer_ni_recoveryq();
4182 lnet_clean_resendqs();
4186 void lnet_monitor_thr_stop(void)
4188 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4191 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4192 lnet_net_lock(LNET_LOCK_EX);
4193 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4194 lnet_net_unlock(LNET_LOCK_EX);
4196 /* tell the monitor thread that we're shutting down */
4197 complete(&the_lnet.ln_mt_wait_complete);
4199 /* block until monitor thread signals that it's done */
4200 mutex_unlock(&the_lnet.ln_api_mutex);
4201 down(&the_lnet.ln_mt_signal);
4202 mutex_lock(&the_lnet.ln_api_mutex);
4203 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4205 /* perform cleanup tasks */
4206 lnet_rsp_tracker_clean();
4207 lnet_clean_local_ni_recoveryq();
4208 lnet_clean_peer_ni_recoveryq();
4209 lnet_clean_resendqs();
4213 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4217 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4218 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4219 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4220 lnet_net_unlock(cpt);
4222 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4226 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4228 struct lnet_hdr *hdr = &msg->msg_hdr;
4230 if (msg->msg_wanted != 0)
4231 lnet_setpayloadbuffer(msg);
4233 lnet_build_msg_event(msg, LNET_EVENT_PUT);
4235 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
4236 * it back into the ACK during lnet_finalize() */
4237 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4238 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4240 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4241 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4245 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4247 struct lnet_hdr *hdr = &msg->msg_hdr;
4248 struct lnet_match_info info;
4252 /* Convert put fields to host byte order */
4253 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4254 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
4255 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
4257 /* Primary peer NID. */
4258 info.mi_id.nid = msg->msg_initiator;
4259 info.mi_id.pid = hdr->src_pid;
4260 info.mi_opc = LNET_MD_OP_PUT;
4261 info.mi_portal = hdr->msg.put.ptl_index;
4262 info.mi_rlength = hdr->payload_length;
4263 info.mi_roffset = hdr->msg.put.offset;
4264 info.mi_mbits = hdr->msg.put.match_bits;
4265 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4267 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4268 ready_delay = msg->msg_rx_ready_delay;
4271 rc = lnet_ptl_match_md(&info, msg);
4276 case LNET_MATCHMD_OK:
4277 lnet_recv_put(ni, msg);
4280 case LNET_MATCHMD_NONE:
4282 /* no eager_recv or has already called it, should
4283 * have been attached on delayed list */
4286 rc = lnet_ni_eager_recv(ni, msg);
4293 case LNET_MATCHMD_DROP:
4294 CNETERR("Dropping PUT from %s portal %d match %llu"
4295 " offset %d length %d: %d\n",
4296 libcfs_idstr(&info.mi_id), info.mi_portal,
4297 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4299 return -ENOENT; /* -ve: OK but no match */
4304 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4306 struct lnet_match_info info;
4307 struct lnet_hdr *hdr = &msg->msg_hdr;
4308 struct lnet_processid source_id;
4309 struct lnet_handle_wire reply_wmd;
4312 /* Convert get fields to host byte order */
4313 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
4314 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
4315 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
4316 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
4318 source_id.nid = hdr->src_nid;
4319 source_id.pid = hdr->src_pid;
4320 /* Primary peer NID */
4321 info.mi_id.nid = msg->msg_initiator;
4322 info.mi_id.pid = hdr->src_pid;
4323 info.mi_opc = LNET_MD_OP_GET;
4324 info.mi_portal = hdr->msg.get.ptl_index;
4325 info.mi_rlength = hdr->msg.get.sink_length;
4326 info.mi_roffset = hdr->msg.get.src_offset;
4327 info.mi_mbits = hdr->msg.get.match_bits;
4328 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4330 rc = lnet_ptl_match_md(&info, msg);
4331 if (rc == LNET_MATCHMD_DROP) {
4332 CNETERR("Dropping GET from %s portal %d match %llu"
4333 " offset %d length %d\n",
4334 libcfs_idstr(&info.mi_id), info.mi_portal,
4335 info.mi_mbits, info.mi_roffset, info.mi_rlength);
4336 return -ENOENT; /* -ve: OK but no match */
4339 LASSERT(rc == LNET_MATCHMD_OK);
4341 lnet_build_msg_event(msg, LNET_EVENT_GET);
4343 reply_wmd = hdr->msg.get.return_wmd;
4345 lnet_prep_send(msg, LNET_MSG_REPLY, &source_id,
4346 msg->msg_offset, msg->msg_wanted);
4348 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4351 /* The LND completes the REPLY from her recv procedure */
4352 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4353 msg->msg_offset, msg->msg_len, msg->msg_len);
4357 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4358 msg->msg_receiving = 0;
4360 rc = lnet_send(&ni->ni_nid, msg, &msg->msg_from);
4362 /* didn't get as far as lnet_ni_send() */
4363 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4364 libcfs_nidstr(&ni->ni_nid),
4365 libcfs_idstr(&info.mi_id), rc);
4367 lnet_finalize(msg, rc);
4374 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4376 void *private = msg->msg_private;
4377 struct lnet_hdr *hdr = &msg->msg_hdr;
4378 struct lnet_processid src = {};
4379 struct lnet_libmd *md;
4380 unsigned int rlength;
4381 unsigned int mlength;
4384 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4387 src.nid = hdr->src_nid;
4388 src.pid = hdr->src_pid;
4390 /* NB handles only looked up by creator (no flips) */
4391 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4392 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4393 CNETERR("%s: Dropping REPLY from %s for %s "
4395 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4396 (md == NULL) ? "invalid" : "inactive",
4397 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4398 hdr->msg.reply.dst_wmd.wh_object_cookie);
4399 if (md != NULL && md->md_me != NULL)
4400 CERROR("REPLY MD also attached to portal %d\n",
4401 md->md_me->me_portal);
4403 lnet_res_unlock(cpt);
4404 return -ENOENT; /* -ve: OK but no match */
4407 LASSERT(md->md_offset == 0);
4409 rlength = hdr->payload_length;
4410 mlength = min(rlength, md->md_length);
4412 if (mlength < rlength &&
4413 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4414 CNETERR("%s: Dropping REPLY from %s length %d "
4415 "for MD %#llx would overflow (%d)\n",
4416 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4417 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4419 lnet_res_unlock(cpt);
4420 return -ENOENT; /* -ve: OK but no match */
4423 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4424 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4425 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4427 lnet_msg_attach_md(msg, md, 0, mlength);
4430 lnet_setpayloadbuffer(msg);
4432 lnet_res_unlock(cpt);
4434 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4436 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4441 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4443 struct lnet_hdr *hdr = &msg->msg_hdr;
4444 struct lnet_processid src = {};
4445 struct lnet_libmd *md;
4448 src.nid = hdr->src_nid;
4449 src.pid = hdr->src_pid;
4451 /* Convert ack fields to host byte order */
4452 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4453 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4455 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4458 /* NB handles only looked up by creator (no flips) */
4459 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4460 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4461 /* Don't moan; this is expected */
4463 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4464 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4465 (md == NULL) ? "invalid" : "inactive",
4466 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4467 hdr->msg.ack.dst_wmd.wh_object_cookie);
4468 if (md != NULL && md->md_me != NULL)
4469 CERROR("Source MD also attached to portal %d\n",
4470 md->md_me->me_portal);
4472 lnet_res_unlock(cpt);
4473 return -ENOENT; /* -ve! */
4476 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4477 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4478 hdr->msg.ack.dst_wmd.wh_object_cookie);
4480 lnet_msg_attach_md(msg, md, 0, 0);
4482 lnet_res_unlock(cpt);
4484 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4486 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4491 * \retval LNET_CREDIT_OK If \a msg is forwarded
4492 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4493 * \retval -ve error code
4496 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4500 if (!the_lnet.ln_routing)
4503 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4504 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4505 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4506 msg->msg_rx_ready_delay = 1;
4508 lnet_net_unlock(msg->msg_rx_cpt);
4509 rc = lnet_ni_eager_recv(ni, msg);
4510 lnet_net_lock(msg->msg_rx_cpt);
4515 rc = lnet_post_routed_recv_locked(msg, 0);
4520 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4524 switch (msg->msg_type) {
4526 rc = lnet_parse_ack(ni, msg);
4529 rc = lnet_parse_put(ni, msg);
4532 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4534 case LNET_MSG_REPLY:
4535 rc = lnet_parse_reply(ni, msg);
4537 default: /* prevent an unused label if !kernel */
4542 LASSERT(rc == 0 || rc == -ENOENT);
4547 lnet_msgtyp2str (int type)
4556 case LNET_MSG_REPLY:
4558 case LNET_MSG_HELLO:
4561 return ("<UNKNOWN>");
4564 EXPORT_SYMBOL(lnet_msgtyp2str);
4567 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr,
4568 struct lnet_nid *from_nid, void *private, int rdma_req)
4570 struct lnet_peer_ni *lpni;
4571 struct lnet_msg *msg;
4572 __u32 payload_length;
4573 lnet_pid_t dest_pid;
4574 struct lnet_nid dest_nid;
4575 struct lnet_nid src_nid;
4581 time64_t now = ktime_get_seconds();
4583 LASSERT (!in_interrupt ());
4586 src_nid = hdr->src_nid;
4587 dest_nid = hdr->dest_nid;
4588 dest_pid = hdr->dest_pid;
4589 payload_length = hdr->payload_length;
4591 for_me = nid_same(&ni->ni_nid, &dest_nid);
4592 cpt = lnet_nid2cpt(from_nid, ni);
4594 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4595 libcfs_nidstr(&dest_nid),
4596 libcfs_nidstr(&ni->ni_nid),
4597 libcfs_nidstr(&src_nid),
4598 lnet_msgtyp2str(type),
4599 (for_me) ? "for me" : "routed");
4604 if (payload_length > 0) {
4605 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4606 libcfs_nidstr(from_nid),
4607 libcfs_nidstr(&src_nid),
4608 lnet_msgtyp2str(type), payload_length);
4614 case LNET_MSG_REPLY:
4615 if (payload_length >
4616 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4617 CERROR("%s, src %s: bad %s payload %d "
4618 "(%d max expected)\n",
4619 libcfs_nidstr(from_nid),
4620 libcfs_nidstr(&src_nid),
4621 lnet_msgtyp2str(type),
4623 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4629 CERROR("%s, src %s: Bad message type 0x%x\n",
4630 libcfs_nidstr(from_nid),
4631 libcfs_nidstr(&src_nid), type);
4635 /* Only update net_last_alive for incoming GETs on the reserved portal
4636 * (i.e. incoming lnet/discovery pings).
4637 * This avoids situations where the router's own traffic results in NI
4640 if (the_lnet.ln_routing && type == LNET_MSG_GET &&
4641 hdr->msg.get.ptl_index == LNET_RESERVED_PORTAL &&
4642 !lnet_islocalnid(&src_nid) &&
4643 ni->ni_net->net_last_alive != now) {
4645 spin_lock(&ni->ni_net->net_lock);
4646 ni->ni_net->net_last_alive = now;
4647 spin_unlock(&ni->ni_net->net_lock);
4648 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4653 lnet_push_update_to_peers(1);
4655 /* Regard a bad destination NID as a protocol error. Senders should
4656 * know what they're doing; if they don't they're misconfigured, buggy
4657 * or malicious so we chop them off at the knees :) */
4660 if (LNET_NID_NET(&dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4661 /* should have gone direct */
4662 CERROR("%s, src %s: Bad dest nid %s "
4663 "(should have been sent direct)\n",
4664 libcfs_nidstr(from_nid),
4665 libcfs_nidstr(&src_nid),
4666 libcfs_nidstr(&dest_nid));
4670 if (lnet_islocalnid(&dest_nid)) {
4671 /* dest is another local NI; sender should have used
4672 * this node's NID on its own network */
4673 CERROR("%s, src %s: Bad dest nid %s "
4674 "(it's my nid but on a different network)\n",
4675 libcfs_nidstr(from_nid),
4676 libcfs_nidstr(&src_nid),
4677 libcfs_nidstr(&dest_nid));
4681 if (rdma_req && type == LNET_MSG_GET) {
4682 CERROR("%s, src %s: Bad optimized GET for %s "
4683 "(final destination must be me)\n",
4684 libcfs_nidstr(from_nid),
4685 libcfs_nidstr(&src_nid),
4686 libcfs_nidstr(&dest_nid));
4690 if (!the_lnet.ln_routing) {
4691 CERROR("%s, src %s: Dropping message for %s "
4692 "(routing not enabled)\n",
4693 libcfs_nidstr(from_nid),
4694 libcfs_nidstr(&src_nid),
4695 libcfs_nidstr(&dest_nid));
4700 /* Message looks OK; we're not going to return an error, so we MUST
4701 * call back lnd_recv() come what may... */
4703 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4704 fail_peer(&src_nid, 0)) { /* shall we now? */
4705 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4706 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4707 lnet_msgtyp2str(type));
4711 if (!list_empty(&the_lnet.ln_drop_rules) &&
4712 lnet_drop_rule_match(hdr, &ni->ni_nid, NULL)) {
4714 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4715 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4716 libcfs_nidstr(&dest_nid), lnet_msgtyp2str(type));
4720 msg = lnet_msg_alloc();
4722 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4723 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4724 lnet_msgtyp2str(type));
4728 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4729 * pointers NULL etc */
4731 msg->msg_type = type;
4732 msg->msg_private = private;
4733 msg->msg_receiving = 1;
4734 msg->msg_rdma_get = rdma_req;
4735 msg->msg_len = msg->msg_wanted = payload_length;
4736 msg->msg_offset = 0;
4737 msg->msg_hdr = *hdr;
4738 /* for building message event */
4739 msg->msg_from = *from_nid;
4741 msg->msg_target.pid = dest_pid;
4742 msg->msg_target.nid = dest_nid;
4743 msg->msg_routing = 1;
4747 lpni = lnet_peerni_by_nid_locked(from_nid, &ni->ni_nid, cpt);
4749 lnet_net_unlock(cpt);
4751 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4752 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4753 lnet_msgtyp2str(type), rc);
4755 if (rc == -ESHUTDOWN)
4756 /* We are shutting down. Don't do anything more */
4761 /* If this message was forwarded to us from a router then we may need
4762 * to update router aliveness or check for an asymmetrical route
4765 if (((lnet_drop_asym_route && for_me) ||
4766 !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4767 LNET_NID_NET(&src_nid) != LNET_NID_NET(from_nid)) {
4768 __u32 src_net_id = LNET_NID_NET(&src_nid);
4769 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4770 struct lnet_route *route;
4773 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4774 if (route->lr_net == src_net_id) {
4776 /* If we're transitioning the gateway from
4777 * dead -> alive, and discovery is disabled
4778 * locally or on the gateway, then we need to
4779 * update the cached route aliveness for each
4780 * route to the src_nid's net.
4782 * Otherwise, we're only checking for
4783 * symmetrical route, and we can break the
4786 if (!gw->lp_alive &&
4787 lnet_is_discovery_disabled(gw))
4788 lnet_set_route_aliveness(route, true);
4793 if (lnet_drop_asym_route && for_me && !found) {
4794 /* Drop ref taken by lnet_nid2peerni_locked() */
4795 lnet_peer_ni_decref_locked(lpni);
4796 lnet_net_unlock(cpt);
4797 /* we would not use from_nid to route a message to
4799 * => asymmetric routing detected but forbidden
4801 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4802 libcfs_nidstr(from_nid),
4803 libcfs_nidstr(&src_nid), lnet_msgtyp2str(type));
4807 if (!gw->lp_alive) {
4808 struct lnet_peer_net *lpn;
4809 struct lnet_peer_ni *lpni2;
4811 gw->lp_alive = true;
4812 /* Mark all remote NIs on src_nid's net UP */
4813 lpn = lnet_peer_get_net_locked(gw, src_net_id);
4815 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4817 lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4821 lpni->lpni_last_alive = now;
4823 msg->msg_rxpeer = lpni;
4825 lnet_ni_addref_locked(ni, cpt);
4826 /* Multi-Rail: Primary NID of source. */
4827 lnet_peer_primary_nid_locked(&src_nid, &msg->msg_initiator);
4830 * mark the status of this lpni as UP since we received a message
4831 * from it. The ping response reports back the ns_status which is
4832 * marked on the remote as up or down and we cache it here.
4834 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4836 lnet_msg_commit(msg, cpt);
4838 /* message delay simulation */
4839 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4840 lnet_delay_rule_match_locked(hdr, msg))) {
4841 lnet_net_unlock(cpt);
4846 rc = lnet_parse_forward_locked(ni, msg);
4847 lnet_net_unlock(cpt);
4852 if (rc == LNET_CREDIT_OK) {
4853 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4854 0, payload_length, payload_length);
4859 lnet_net_unlock(cpt);
4861 rc = lnet_parse_local(ni, msg);
4867 LASSERT(msg->msg_md == NULL);
4868 lnet_finalize(msg, rc);
4871 lnet_drop_message(ni, cpt, private, payload_length, type);
4874 EXPORT_SYMBOL(lnet_parse);
4877 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4879 struct lnet_msg *msg;
4881 while ((msg = list_first_entry_or_null(head, struct lnet_msg,
4882 msg_list)) != NULL) {
4883 struct lnet_processid id = {};
4885 list_del(&msg->msg_list);
4887 id.nid = msg->msg_hdr.src_nid;
4888 id.pid = msg->msg_hdr.src_pid;
4890 LASSERT(msg->msg_md == NULL);
4891 LASSERT(msg->msg_rx_delayed);
4892 LASSERT(msg->msg_rxpeer != NULL);
4893 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4895 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4896 " offset %d length %d: %s\n",
4898 msg->msg_hdr.msg.put.ptl_index,
4899 msg->msg_hdr.msg.put.match_bits,
4900 msg->msg_hdr.msg.put.offset,
4901 msg->msg_hdr.payload_length, reason);
4903 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4904 * called lnet_drop_message(), so I just hang onto msg as well
4905 * until that's done */
4907 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4908 msg->msg_private, msg->msg_len,
4911 msg->msg_no_resend = true;
4913 * NB: message will not generate event because w/o attached MD,
4914 * but we still should give error code so lnet_msg_decommit()
4915 * can skip counters operations and other checks.
4917 lnet_finalize(msg, -ENOENT);
4922 lnet_recv_delayed_msg_list(struct list_head *head)
4924 struct lnet_msg *msg;
4926 while ((msg = list_first_entry_or_null(head, struct lnet_msg,
4927 msg_list)) != NULL) {
4928 struct lnet_processid id;
4930 list_del(&msg->msg_list);
4932 /* md won't disappear under me, since each msg
4933 * holds a ref on it */
4935 id.nid = msg->msg_hdr.src_nid;
4936 id.pid = msg->msg_hdr.src_pid;
4938 LASSERT(msg->msg_rx_delayed);
4939 LASSERT(msg->msg_md != NULL);
4940 LASSERT(msg->msg_rxpeer != NULL);
4941 LASSERT(msg->msg_rxni != NULL);
4942 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4944 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4945 "match %llu offset %d length %d.\n",
4946 libcfs_idstr(&id), msg->msg_hdr.msg.put.ptl_index,
4947 msg->msg_hdr.msg.put.match_bits,
4948 msg->msg_hdr.msg.put.offset,
4949 msg->msg_hdr.payload_length);
4951 lnet_recv_put(msg->msg_rxni, msg);
4956 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4957 struct lnet_libmd *md, struct lnet_handle_md mdh)
4960 struct lnet_rsp_tracker *local_rspt;
4963 * MD has a refcount taken by message so it's not going away.
4964 * The MD however can be looked up. We need to secure the access
4965 * to the md_rspt_ptr by taking the res_lock.
4966 * The rspt can be accessed without protection up to when it gets
4967 * added to the list.
4971 local_rspt = md->md_rspt_ptr;
4972 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4973 if (local_rspt != NULL) {
4975 * we already have an rspt attached to the md, so we'll
4976 * update the deadline on that one.
4978 lnet_rspt_free(rspt, cpt);
4981 rspt->rspt_mdh = mdh;
4982 rspt->rspt_cpt = cpt;
4983 /* store the rspt so we can access it when we get the REPLY */
4984 md->md_rspt_ptr = rspt;
4987 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4990 * add to the list of tracked responses. It's added to tail of the
4991 * list in order to expire all the older entries first.
4994 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4995 lnet_net_unlock(cpt);
4996 lnet_res_unlock(cpt);
5000 * Initiate an asynchronous PUT operation.
5002 * There are several events associated with a PUT: completion of the send on
5003 * the initiator node (LNET_EVENT_SEND), and when the send completes
5004 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
5005 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
5006 * used at the target node to indicate the completion of incoming data
5009 * The local events will be logged in the EQ associated with the MD pointed to
5010 * by \a mdh handle. Using a MD without an associated EQ results in these
5011 * events being discarded. In this case, the caller must have another
5012 * mechanism (e.g., a higher level protocol) for determining when it is safe
5013 * to modify the memory region associated with the MD.
5015 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
5016 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
5018 * \param self Indicates the NID of a local interface through which to send
5019 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
5020 * \param mdh A handle for the MD that describes the memory to be sent. The MD
5021 * must be "free floating" (See LNetMDBind()).
5022 * \param ack Controls whether an acknowledgment is requested.
5023 * Acknowledgments are only sent when they are requested by the initiating
5024 * process and the target MD enables them.
5025 * \param target A process identifier for the target process.
5026 * \param portal The index in the \a target's portal table.
5027 * \param match_bits The match bits to use for MD selection at the target
5029 * \param offset The offset into the target MD (only used when the target
5030 * MD has the LNET_MD_MANAGE_REMOTE option set).
5031 * \param hdr_data 64 bits of user data that can be included in the message
5032 * header. This data is written to an event queue entry at the target if an
5033 * EQ is present on the matching MD.
5035 * \retval 0 Success, and only in this case events will be generated
5036 * and logged to EQ (if it exists).
5037 * \retval -EIO Simulated failure.
5038 * \retval -ENOMEM Memory allocation failure.
5039 * \retval -ENOENT Invalid MD object.
5041 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
5044 LNetPut(struct lnet_nid *self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
5045 struct lnet_processid *target, unsigned int portal,
5046 __u64 match_bits, unsigned int offset,
5049 struct lnet_msg *msg;
5050 struct lnet_libmd *md;
5053 struct lnet_rsp_tracker *rspt = NULL;
5055 LASSERT(the_lnet.ln_refcount > 0);
5057 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5058 fail_peer(&target->nid, 1)) { /* shall we now? */
5059 CERROR("Dropping PUT to %s: simulated failure\n",
5060 libcfs_idstr(target));
5064 msg = lnet_msg_alloc();
5066 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
5067 libcfs_idstr(target));
5070 msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
5072 cpt = lnet_cpt_of_cookie(mdh.cookie);
5074 if (ack == LNET_ACK_REQ) {
5075 rspt = lnet_rspt_alloc(cpt);
5077 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
5078 libcfs_idstr(target));
5081 INIT_LIST_HEAD(&rspt->rspt_on_list);
5086 md = lnet_handle2md(&mdh);
5087 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5088 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5089 match_bits, portal, libcfs_idstr(target),
5090 md == NULL ? -1 : md->md_threshold);
5091 if (md != NULL && md->md_me != NULL)
5092 CERROR("Source MD also attached to portal %d\n",
5093 md->md_me->me_portal);
5094 lnet_res_unlock(cpt);
5097 lnet_rspt_free(rspt, cpt);
5103 CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5105 lnet_msg_attach_md(msg, md, 0, 0);
5107 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5109 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5110 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5111 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5112 msg->msg_hdr.msg.put.hdr_data = hdr_data;
5114 /* NB handles only looked up by creator (no flips) */
5115 if (ack == LNET_ACK_REQ) {
5116 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5117 the_lnet.ln_interface_cookie;
5118 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5119 md->md_lh.lh_cookie;
5121 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5122 LNET_WIRE_HANDLE_COOKIE_NONE;
5123 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5124 LNET_WIRE_HANDLE_COOKIE_NONE;
5127 lnet_res_unlock(cpt);
5129 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5131 if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5133 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5135 lnet_rspt_free(rspt, cpt);
5137 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5141 rc = lnet_send(self, msg, NULL);
5144 CNETERR("Error sending PUT to %s: %d\n",
5145 libcfs_idstr(target), rc);
5146 msg->msg_no_resend = true;
5147 lnet_finalize(msg, rc);
5150 /* completion will be signalled by an event */
5153 EXPORT_SYMBOL(LNetPut);
5156 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
5157 * returns a msg for the LND to pass to lnet_finalize() when the sink
5158 * data has been received.
5160 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5161 * lnet_finalize() is called on it, so the LND must call this first
5164 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5166 struct lnet_msg *msg = lnet_msg_alloc();
5167 struct lnet_libmd *getmd = getmsg->msg_md;
5168 struct lnet_processid *peer_id = &getmsg->msg_target;
5171 LASSERT(!getmsg->msg_target_is_router);
5172 LASSERT(!getmsg->msg_routing);
5175 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5176 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id));
5180 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5183 LASSERT(getmd->md_refcount > 0);
5185 if (getmd->md_threshold == 0) {
5186 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5187 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id),
5189 lnet_res_unlock(cpt);
5193 LASSERT(getmd->md_offset == 0);
5195 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5196 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id), getmd);
5198 /* setup information for lnet_build_msg_event */
5199 msg->msg_initiator =
5200 getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5201 msg->msg_from = peer_id->nid;
5202 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5203 msg->msg_hdr.src_nid = peer_id->nid;
5204 msg->msg_hdr.payload_length = getmd->md_length;
5205 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5207 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5208 lnet_res_unlock(cpt);
5210 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5213 lnet_msg_commit(msg, cpt);
5214 lnet_net_unlock(cpt);
5216 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5221 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5224 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5225 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5226 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5228 lnet_net_unlock(cpt);
5235 EXPORT_SYMBOL(lnet_create_reply_msg);
5238 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5241 /* Set the REPLY length, now the RDMA that elides the REPLY message has
5242 * completed and I know it. */
5243 LASSERT(reply != NULL);
5244 LASSERT(reply->msg_type == LNET_MSG_GET);
5245 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5247 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
5248 * the end of my buffer, I might as well be dead. */
5249 LASSERT(len <= reply->msg_ev.mlength);
5251 reply->msg_ev.mlength = len;
5253 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5256 * Initiate an asynchronous GET operation.
5258 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5259 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5260 * the target node in the REPLY has been written to local MD.
5262 * On the target node, an LNET_EVENT_GET is logged when the GET request
5263 * arrives and is accepted into a MD.
5265 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5266 * \param mdh A handle for the MD that describes the memory into which the
5267 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5269 * \retval 0 Success, and only in this case events will be generated
5270 * and logged to EQ (if it exists) of the MD.
5271 * \retval -EIO Simulated failure.
5272 * \retval -ENOMEM Memory allocation failure.
5273 * \retval -ENOENT Invalid MD object.
5276 LNetGet(struct lnet_nid *self, struct lnet_handle_md mdh,
5277 struct lnet_processid *target, unsigned int portal,
5278 __u64 match_bits, unsigned int offset, bool recovery)
5280 struct lnet_msg *msg;
5281 struct lnet_libmd *md;
5282 struct lnet_rsp_tracker *rspt;
5286 LASSERT(the_lnet.ln_refcount > 0);
5288 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5289 fail_peer(&target->nid, 1)) /* shall we now? */
5291 CERROR("Dropping GET to %s: simulated failure\n",
5292 libcfs_idstr(target));
5296 msg = lnet_msg_alloc();
5298 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5299 libcfs_idstr(target));
5303 cpt = lnet_cpt_of_cookie(mdh.cookie);
5305 rspt = lnet_rspt_alloc(cpt);
5307 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5308 libcfs_idstr(target));
5311 INIT_LIST_HEAD(&rspt->rspt_on_list);
5313 msg->msg_recovery = recovery;
5317 md = lnet_handle2md(&mdh);
5318 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5319 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5320 match_bits, portal, libcfs_idstr(target),
5321 md == NULL ? -1 : md->md_threshold);
5322 if (md != NULL && md->md_me != NULL)
5323 CERROR("REPLY MD also attached to portal %d\n",
5324 md->md_me->me_portal);
5326 lnet_res_unlock(cpt);
5329 lnet_rspt_free(rspt, cpt);
5333 CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5335 lnet_msg_attach_md(msg, md, 0, 0);
5337 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5339 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5340 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5341 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5342 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5344 /* NB handles only looked up by creator (no flips) */
5345 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5346 the_lnet.ln_interface_cookie;
5347 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5348 md->md_lh.lh_cookie;
5350 lnet_res_unlock(cpt);
5352 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5354 if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5355 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5357 lnet_rspt_free(rspt, cpt);
5359 rc = lnet_send(self, msg, NULL);
5361 CNETERR("Error sending GET to %s: %d\n",
5362 libcfs_idstr(target), rc);
5363 msg->msg_no_resend = true;
5364 lnet_finalize(msg, rc);
5367 /* completion will be signalled by an event */
5370 EXPORT_SYMBOL(LNetGet);
5373 * Calculate distance to node at \a dstnid.
5375 * \param dstnid Target NID.
5376 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5378 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5381 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5382 * local_nid_dist_zero is set, which is the default.
5383 * \retval positives Distance to target NID, i.e. number of hops plus one.
5384 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5387 LNetDist(struct lnet_nid *dstnid, struct lnet_nid *srcnid, __u32 *orderp)
5389 struct lnet_ni *ni = NULL;
5390 struct lnet_remotenet *rnet;
5391 __u32 dstnet = LNET_NID_NET(dstnid);
5395 struct list_head *rn_list;
5396 struct lnet_ni *matched_dstnet = NULL;
5398 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5399 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5400 * keep order 0 free for 0@lo and order 1 free for a local NID
5402 * WARNING: dstnid and srcnid might point to same place.
5403 * Don't set *srcnid until late.
5406 LASSERT(the_lnet.ln_refcount > 0);
5408 cpt = lnet_net_lock_current();
5410 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5411 if (nid_same(&ni->ni_nid, dstnid)) {
5412 if (orderp != NULL) {
5413 if (nid_is_lo0(dstnid))
5420 lnet_net_unlock(cpt);
5422 return local_nid_dist_zero ? 0 : 1;
5425 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5426 matched_dstnet = ni;
5427 /* We matched the destination net, but we may have
5428 * additional local NIs to inspect.
5430 * We record the order as appropriate, but
5431 * they may be overwritten if we match local NI above.
5435 /* Check if ni was originally created in
5436 * current net namespace.
5437 * If not, assign order above 0xffff0000,
5438 * to make this ni not a priority.
5440 if (current->nsproxy &&
5441 !net_eq(ni->ni_net_ns,
5442 current->nsproxy->net_ns))
5443 *orderp = order + 0xffff0000;
5452 if (matched_dstnet) {
5454 *srcnid = matched_dstnet->ni_nid;
5455 lnet_net_unlock(cpt);
5459 rn_list = lnet_net2rnethash(dstnet);
5460 list_for_each_entry(rnet, rn_list, lrn_list) {
5461 if (rnet->lrn_net == dstnet) {
5462 struct lnet_route *route;
5463 struct lnet_route *shortest = NULL;
5464 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5467 LASSERT(!list_empty(&rnet->lrn_routes));
5469 list_for_each_entry(route, &rnet->lrn_routes,
5471 route_hops = route->lr_hops;
5472 if (route_hops == LNET_UNDEFINED_HOPS)
5474 if (shortest == NULL ||
5475 route_hops < shortest_hops) {
5477 shortest_hops = route_hops;
5481 LASSERT(shortest != NULL);
5482 hops = shortest_hops;
5484 struct lnet_net *net;
5485 net = lnet_get_net_locked(shortest->lr_lnet);
5487 ni = lnet_get_next_ni_locked(net, NULL);
5488 *srcnid = ni->ni_nid;
5492 lnet_net_unlock(cpt);
5498 lnet_net_unlock(cpt);
5499 return -EHOSTUNREACH;
5501 EXPORT_SYMBOL(LNetDist);