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 */
772 /* returns true if this message should be dropped */
774 lnet_check_message_drop(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
775 struct lnet_msg *msg)
777 if (msg->msg_target.pid & LNET_PID_USERFLAG)
780 if (!lnet_peer_aliveness_enabled(lpni))
783 /* If we're resending a message, let's attempt to send it even if
784 * the peer is down to fulfill our resend quota on the message
786 if (msg->msg_retry_count > 0)
789 /* try and send recovery messages irregardless */
790 if (msg->msg_recovery)
793 /* always send any responses */
794 if (lnet_msg_is_response(msg))
797 /* always send non-routed messages */
798 if (!msg->msg_routing)
801 /* assume peer_ni is alive as long as we're within the configured
804 return ktime_get_seconds() >=
805 (lpni->lpni_last_alive +
806 lpni->lpni_net->net_tunables.lct_peer_timeout);
810 * \param msg The message to be sent.
811 * \param do_send True if lnet_ni_send() should be called in this function.
812 * lnet_send() is going to lnet_net_unlock immediately after this, so
813 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
815 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
816 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
817 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
818 * \retval -ECANCELED If the MD of the message has been unlinked.
821 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
823 struct lnet_peer_ni *lp = msg->msg_txpeer;
824 struct lnet_ni *ni = msg->msg_txni;
825 int cpt = msg->msg_tx_cpt;
826 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
828 /* non-lnet_send() callers have checked before */
829 LASSERT(!do_send || msg->msg_tx_delayed);
830 LASSERT(!msg->msg_receiving);
831 LASSERT(msg->msg_tx_committed);
833 /* can't get here if we're sending to the loopback interface */
834 if (the_lnet.ln_loni)
835 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
837 /* NB 'lp' is always the next hop */
838 if (lnet_check_message_drop(ni, lp, msg)) {
839 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
840 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
842 lnet_net_unlock(cpt);
844 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
846 LNET_STATS_TYPE_DROP);
848 lnet_incr_stats(&msg->msg_txni->ni_stats,
850 LNET_STATS_TYPE_DROP);
852 CNETERR("Dropping message for %s: peer not alive\n",
853 libcfs_idstr(&msg->msg_target));
854 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
856 lnet_finalize(msg, -EHOSTUNREACH);
859 return -EHOSTUNREACH;
862 if (msg->msg_md != NULL &&
863 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
864 lnet_net_unlock(cpt);
866 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
867 "called on the MD/ME.\n",
868 libcfs_idstr(&msg->msg_target));
870 msg->msg_no_resend = true;
871 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
872 msg, libcfs_idstr(&msg->msg_target));
873 lnet_finalize(msg, -ECANCELED);
880 if (!msg->msg_peertxcredit) {
881 spin_lock(&lp->lpni_lock);
882 LASSERT((lp->lpni_txcredits < 0) ==
883 !list_empty(&lp->lpni_txq));
885 msg->msg_peertxcredit = 1;
886 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr_nid4);
887 lp->lpni_txcredits--;
889 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
890 lp->lpni_mintxcredits = lp->lpni_txcredits;
892 if (lp->lpni_txcredits < 0) {
893 msg->msg_tx_delayed = 1;
894 list_add_tail(&msg->msg_list, &lp->lpni_txq);
895 spin_unlock(&lp->lpni_lock);
896 return LNET_CREDIT_WAIT;
898 spin_unlock(&lp->lpni_lock);
901 if (!msg->msg_txcredit) {
902 LASSERT((tq->tq_credits < 0) ==
903 !list_empty(&tq->tq_delayed));
905 msg->msg_txcredit = 1;
907 atomic_dec(&ni->ni_tx_credits);
909 if (tq->tq_credits < tq->tq_credits_min)
910 tq->tq_credits_min = tq->tq_credits;
912 if (tq->tq_credits < 0) {
913 msg->msg_tx_delayed = 1;
914 list_add_tail(&msg->msg_list, &tq->tq_delayed);
915 return LNET_CREDIT_WAIT;
919 if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
920 lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
921 msg->msg_tx_delayed = 1;
922 return LNET_CREDIT_WAIT;
925 /* unset the tx_delay flag as we're going to send it now */
926 msg->msg_tx_delayed = 0;
929 lnet_net_unlock(cpt);
930 lnet_ni_send(ni, msg);
933 return LNET_CREDIT_OK;
937 static struct lnet_rtrbufpool *
938 lnet_msg2bufpool(struct lnet_msg *msg)
940 struct lnet_rtrbufpool *rbp;
943 LASSERT(msg->msg_rx_committed);
945 cpt = msg->msg_rx_cpt;
946 rbp = &the_lnet.ln_rtrpools[cpt][0];
948 LASSERT(msg->msg_len <= LNET_MTU);
949 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
951 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
958 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
960 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
961 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
962 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
963 * received or OK to receive */
964 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
965 struct lnet_peer *lp;
966 struct lnet_rtrbufpool *rbp;
967 struct lnet_rtrbuf *rb;
969 LASSERT(msg->msg_kiov == NULL);
970 LASSERT(msg->msg_niov == 0);
971 LASSERT(msg->msg_routing);
972 LASSERT(msg->msg_receiving);
973 LASSERT(!msg->msg_sending);
974 LASSERT(lpni->lpni_peer_net);
975 LASSERT(lpni->lpni_peer_net->lpn_peer);
977 lp = lpni->lpni_peer_net->lpn_peer;
979 /* non-lnet_parse callers only receive delayed messages */
980 LASSERT(!do_recv || msg->msg_rx_delayed);
982 if (!msg->msg_peerrtrcredit) {
983 /* lpni_lock protects the credit manipulation */
984 spin_lock(&lpni->lpni_lock);
986 msg->msg_peerrtrcredit = 1;
987 lpni->lpni_rtrcredits--;
988 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
989 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
991 if (lpni->lpni_rtrcredits < 0) {
992 spin_unlock(&lpni->lpni_lock);
993 /* must have checked eager_recv before here */
994 LASSERT(msg->msg_rx_ready_delay);
995 msg->msg_rx_delayed = 1;
996 /* lp_lock protects the lp_rtrq */
997 spin_lock(&lp->lp_lock);
998 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
999 spin_unlock(&lp->lp_lock);
1000 return LNET_CREDIT_WAIT;
1002 spin_unlock(&lpni->lpni_lock);
1005 rbp = lnet_msg2bufpool(msg);
1007 if (!msg->msg_rtrcredit) {
1008 msg->msg_rtrcredit = 1;
1010 if (rbp->rbp_credits < rbp->rbp_mincredits)
1011 rbp->rbp_mincredits = rbp->rbp_credits;
1013 if (rbp->rbp_credits < 0) {
1014 /* must have checked eager_recv before here */
1015 LASSERT(msg->msg_rx_ready_delay);
1016 msg->msg_rx_delayed = 1;
1017 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1018 return LNET_CREDIT_WAIT;
1022 LASSERT(!list_empty(&rbp->rbp_bufs));
1023 rb = list_first_entry(&rbp->rbp_bufs, struct lnet_rtrbuf, rb_list);
1024 list_del(&rb->rb_list);
1026 msg->msg_niov = rbp->rbp_npages;
1027 msg->msg_kiov = &rb->rb_kiov[0];
1029 /* unset the msg-rx_delayed flag since we're receiving the message */
1030 msg->msg_rx_delayed = 0;
1033 int cpt = msg->msg_rx_cpt;
1035 lnet_net_unlock(cpt);
1036 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1037 0, msg->msg_len, msg->msg_len);
1040 return LNET_CREDIT_OK;
1044 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1046 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1047 struct lnet_ni *txni = msg->msg_txni;
1048 struct lnet_msg *msg2;
1050 if (msg->msg_txcredit) {
1051 struct lnet_ni *ni = msg->msg_txni;
1052 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1054 /* give back NI txcredits */
1055 msg->msg_txcredit = 0;
1057 LASSERT((tq->tq_credits < 0) ==
1058 !list_empty(&tq->tq_delayed));
1061 atomic_inc(&ni->ni_tx_credits);
1062 if (tq->tq_credits <= 0) {
1063 msg2 = list_first_entry(&tq->tq_delayed,
1064 struct lnet_msg, msg_list);
1065 list_del(&msg2->msg_list);
1067 LASSERT(msg2->msg_txni == ni);
1068 LASSERT(msg2->msg_tx_delayed);
1069 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1071 (void) lnet_post_send_locked(msg2, 1);
1075 if (msg->msg_peertxcredit) {
1076 /* give back peer txcredits */
1077 msg->msg_peertxcredit = 0;
1079 spin_lock(&txpeer->lpni_lock);
1080 LASSERT((txpeer->lpni_txcredits < 0) ==
1081 !list_empty(&txpeer->lpni_txq));
1083 txpeer->lpni_txqnob -= msg->msg_len +
1084 sizeof(struct lnet_hdr_nid4);
1085 LASSERT(txpeer->lpni_txqnob >= 0);
1087 txpeer->lpni_txcredits++;
1088 if (txpeer->lpni_txcredits <= 0) {
1091 msg2 = list_first_entry(&txpeer->lpni_txq,
1092 struct lnet_msg, msg_list);
1093 list_del(&msg2->msg_list);
1094 spin_unlock(&txpeer->lpni_lock);
1096 LASSERT(msg2->msg_txpeer == txpeer);
1097 LASSERT(msg2->msg_tx_delayed);
1099 msg2_cpt = msg2->msg_tx_cpt;
1102 * The msg_cpt can be different from the msg2_cpt
1103 * so we need to make sure we lock the correct cpt
1105 * Once we call lnet_post_send_locked() it is no
1106 * longer safe to access msg2, since it could've
1107 * been freed by lnet_finalize(), but we still
1108 * need to relock the correct cpt, so we cache the
1109 * msg2_cpt for the purpose of the check that
1110 * follows the call to lnet_pose_send_locked().
1112 if (msg2_cpt != msg->msg_tx_cpt) {
1113 lnet_net_unlock(msg->msg_tx_cpt);
1114 lnet_net_lock(msg2_cpt);
1116 (void) lnet_post_send_locked(msg2, 1);
1117 if (msg2_cpt != msg->msg_tx_cpt) {
1118 lnet_net_unlock(msg2_cpt);
1119 lnet_net_lock(msg->msg_tx_cpt);
1122 spin_unlock(&txpeer->lpni_lock);
1127 msg->msg_txni = NULL;
1128 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1131 if (txpeer != NULL) {
1132 msg->msg_txpeer = NULL;
1133 lnet_peer_ni_decref_locked(txpeer);
1138 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1140 struct lnet_msg *msg;
1142 if (list_empty(&rbp->rbp_msgs))
1144 msg = list_first_entry(&rbp->rbp_msgs,
1145 struct lnet_msg, msg_list);
1146 list_del(&msg->msg_list);
1148 (void)lnet_post_routed_recv_locked(msg, 1);
1152 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1154 struct lnet_msg *msg;
1155 struct lnet_msg *tmp;
1157 lnet_net_unlock(cpt);
1159 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1160 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1161 0, 0, 0, msg->msg_hdr.payload_length);
1162 list_del_init(&msg->msg_list);
1163 msg->msg_no_resend = true;
1164 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1165 lnet_finalize(msg, -ECANCELED);
1172 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1174 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1175 struct lnet_peer *lp;
1176 struct lnet_ni *rxni = msg->msg_rxni;
1177 struct lnet_msg *msg2;
1179 if (msg->msg_rtrcredit) {
1180 /* give back global router credits */
1181 struct lnet_rtrbuf *rb;
1182 struct lnet_rtrbufpool *rbp;
1184 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1185 * there until it gets one allocated, or aborts the wait
1187 LASSERT(msg->msg_kiov != NULL);
1189 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1192 msg->msg_kiov = NULL;
1193 msg->msg_rtrcredit = 0;
1195 LASSERT(rbp == lnet_msg2bufpool(msg));
1197 LASSERT((rbp->rbp_credits > 0) ==
1198 !list_empty(&rbp->rbp_bufs));
1200 /* If routing is now turned off, we just drop this buffer and
1201 * don't bother trying to return credits. */
1202 if (!the_lnet.ln_routing) {
1203 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1207 /* It is possible that a user has lowered the desired number of
1208 * buffers in this pool. Make sure we never put back
1209 * more buffers than the stated number. */
1210 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1211 /* Discard this buffer so we don't have too
1213 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1214 rbp->rbp_nbuffers--;
1216 list_add(&rb->rb_list, &rbp->rbp_bufs);
1218 if (rbp->rbp_credits <= 0)
1219 lnet_schedule_blocked_locked(rbp);
1224 if (msg->msg_peerrtrcredit) {
1226 LASSERT(rxpeerni->lpni_peer_net);
1227 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1229 /* give back peer router credits */
1230 msg->msg_peerrtrcredit = 0;
1232 spin_lock(&rxpeerni->lpni_lock);
1233 rxpeerni->lpni_rtrcredits++;
1234 spin_unlock(&rxpeerni->lpni_lock);
1236 lp = rxpeerni->lpni_peer_net->lpn_peer;
1237 spin_lock(&lp->lp_lock);
1239 /* drop all messages which are queued to be routed on that
1241 if (!the_lnet.ln_routing) {
1243 list_splice_init(&lp->lp_rtrq, &drop);
1244 spin_unlock(&lp->lp_lock);
1245 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1246 } else if (!list_empty(&lp->lp_rtrq)) {
1249 msg2 = list_first_entry(&lp->lp_rtrq,
1250 struct lnet_msg, msg_list);
1251 list_del(&msg2->msg_list);
1252 msg2_cpt = msg2->msg_rx_cpt;
1253 spin_unlock(&lp->lp_lock);
1255 * messages on the lp_rtrq can be from any NID in
1256 * the peer, which means they might have different
1257 * cpts. We need to make sure we lock the right
1260 if (msg2_cpt != msg->msg_rx_cpt) {
1261 lnet_net_unlock(msg->msg_rx_cpt);
1262 lnet_net_lock(msg2_cpt);
1264 (void) lnet_post_routed_recv_locked(msg2, 1);
1265 if (msg2_cpt != msg->msg_rx_cpt) {
1266 lnet_net_unlock(msg2_cpt);
1267 lnet_net_lock(msg->msg_rx_cpt);
1270 spin_unlock(&lp->lp_lock);
1274 msg->msg_rxni = NULL;
1275 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1277 if (rxpeerni != NULL) {
1278 msg->msg_rxpeer = NULL;
1279 lnet_peer_ni_decref_locked(rxpeerni);
1283 static struct lnet_peer_ni *
1284 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1285 struct lnet_peer *peer,
1286 struct lnet_peer_ni *best_lpni,
1287 struct lnet_peer_net *peer_net)
1290 * Look at the peer NIs for the destination peer that connect
1291 * to the chosen net. If a peer_ni is preferred when using the
1292 * best_ni to communicate, we use that one. If there is no
1293 * preferred peer_ni, or there are multiple preferred peer_ni,
1294 * the available transmit credits are used. If the transmit
1295 * credits are equal, we round-robin over the peer_ni.
1297 struct lnet_peer_ni *lpni = NULL;
1298 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1300 int best_lpni_healthv = (best_lpni) ?
1301 atomic_read(&best_lpni->lpni_healthv) : 0;
1302 bool best_lpni_is_preferred = false;
1303 bool lpni_is_preferred;
1305 __u32 lpni_sel_prio;
1306 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1308 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1310 * if the best_ni we've chosen aleady has this lpni
1311 * preferred, then let's use it
1314 lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1315 lpni, &best_ni->ni_nid);
1316 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1317 libcfs_nidstr(&best_ni->ni_nid),
1320 lpni_is_preferred = false;
1323 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1324 lpni_sel_prio = lpni->lpni_sel_priority;
1327 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1328 libcfs_nidstr(&lpni->lpni_nid),
1329 libcfs_nidstr(&best_lpni->lpni_nid),
1330 lpni_healthv, best_lpni_healthv,
1331 lpni_sel_prio, best_sel_prio,
1332 lpni->lpni_txcredits, best_lpni_credits,
1333 lpni->lpni_seq, best_lpni->lpni_seq);
1337 /* pick the healthiest peer ni */
1338 if (lpni_healthv < best_lpni_healthv)
1340 else if (lpni_healthv > best_lpni_healthv) {
1341 if (best_lpni_is_preferred)
1342 best_lpni_is_preferred = false;
1346 if (lpni_sel_prio > best_sel_prio)
1348 else if (lpni_sel_prio < best_sel_prio) {
1349 if (best_lpni_is_preferred)
1350 best_lpni_is_preferred = false;
1354 /* if this is a preferred peer use it */
1355 if (!best_lpni_is_preferred && lpni_is_preferred) {
1356 best_lpni_is_preferred = true;
1358 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1359 /* this is not the preferred peer so let's ignore
1365 if (lpni->lpni_txcredits < best_lpni_credits)
1366 /* We already have a peer that has more credits
1367 * available than this one. No need to consider
1368 * this peer further.
1371 else if (lpni->lpni_txcredits > best_lpni_credits)
1374 /* The best peer found so far and the current peer
1375 * have the same number of available credits let's
1376 * make sure to select between them using Round Robin
1378 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1381 best_lpni_is_preferred = lpni_is_preferred;
1382 best_lpni_healthv = lpni_healthv;
1383 best_sel_prio = lpni_sel_prio;
1385 best_lpni_credits = lpni->lpni_txcredits;
1388 /* if we still can't find a peer ni then we can't reach it */
1390 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1391 LNET_NIDNET(dst_nid);
1392 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1393 libcfs_net2str(net_id));
1397 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1398 libcfs_nidstr(&best_lpni->lpni_nid));
1404 * Prerequisite: the best_ni should already be set in the sd
1405 * Find the best lpni.
1406 * If the net id is provided then restrict lpni selection on
1407 * that particular net.
1408 * Otherwise find any reachable lpni. When dealing with an MR
1409 * gateway and it has multiple lpnis which we can use
1410 * we want to select the best one from the list of reachable
1413 static inline struct lnet_peer_ni *
1414 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1415 struct lnet_peer *peer, __u32 net_id)
1417 struct lnet_peer_net *peer_net;
1419 /* find the best_lpni on any local network */
1420 if (net_id == LNET_NET_ANY) {
1421 struct lnet_peer_ni *best_lpni = NULL;
1422 struct lnet_peer_net *lpn;
1423 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1424 /* no net specified find any reachable peer ni */
1425 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1427 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1433 /* restrict on the specified net */
1434 peer_net = lnet_peer_get_net_locked(peer, net_id);
1436 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1442 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1444 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1447 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1450 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1453 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1459 /* Compare route priorities and hop counts */
1461 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1463 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1464 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1466 if (r1->lr_priority < r2->lr_priority)
1469 if (r1->lr_priority > r2->lr_priority)
1472 if (r1_hops < r2_hops)
1475 if (r1_hops > r2_hops)
1481 static struct lnet_route *
1482 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1483 struct lnet_peer_ni *remote_lpni,
1484 struct lnet_route **prev_route,
1485 struct lnet_peer_ni **gwni)
1487 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1488 struct lnet_route *best_route;
1489 struct lnet_route *last_route;
1490 struct lnet_route *route;
1492 bool best_rte_is_preferred = false;
1493 struct lnet_nid *gw_pnid;
1495 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1496 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1498 best_route = last_route = NULL;
1499 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1500 if (!lnet_is_route_alive(route))
1502 gw_pnid = &route->lr_gateway->lp_primary_nid;
1504 /* no protection on below fields, but it's harmless */
1505 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1508 /* if the best route found is in the preferred list then
1509 * tag it as preferred and use it later on. But if we
1510 * didn't find any routes which are on the preferred list
1511 * then just use the best route possible.
1513 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1515 if (!best_route || (rc && !best_rte_is_preferred)) {
1516 /* Restrict the selection of the router NI on the
1517 * src_net provided. If the src_net is LNET_NID_ANY,
1518 * then select the best interface available.
1520 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1525 "Gateway %s does not have a peer NI on net %s\n",
1526 libcfs_nidstr(gw_pnid),
1527 libcfs_net2str(src_net));
1532 if (rc && !best_rte_is_preferred) {
1533 /* This is the first preferred route we found,
1534 * so it beats any route found previously
1540 best_rte_is_preferred = true;
1541 CDEBUG(D_NET, "preferred gw = %s\n",
1542 libcfs_nidstr(gw_pnid));
1544 } else if ((!rc) && best_rte_is_preferred)
1545 /* The best route we found so far is in the preferred
1546 * list, so it beats any non-preferred route
1551 best_route = last_route = route;
1556 rc = lnet_compare_routes(route, best_route);
1560 /* Restrict the selection of the router NI on the
1561 * src_net provided. If the src_net is LNET_NID_ANY,
1562 * then select the best interface available.
1564 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1569 "Gateway %s does not have a peer NI on net %s\n",
1570 libcfs_nidstr(gw_pnid),
1571 libcfs_net2str(src_net));
1581 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1585 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1592 *prev_route = last_route;
1598 static inline unsigned int
1599 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1601 if (dev_idx == UINT_MAX)
1604 if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1605 !ni->ni_net->net_lnd->lnd_get_dev_prio)
1608 return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1611 static struct lnet_ni *
1612 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1613 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1614 struct lnet_msg *msg, int md_cpt)
1616 struct lnet_libmd *md = msg->msg_md;
1617 unsigned int offset = msg->msg_offset;
1618 unsigned int shortest_distance;
1619 struct lnet_ni *ni = NULL;
1622 __u32 best_sel_prio;
1623 unsigned int best_dev_prio;
1625 unsigned int dev_idx = UINT_MAX;
1626 bool gpu = md ? (md->md_flags & LNET_MD_FLAG_GPU) : false;
1629 struct page *page = lnet_get_first_page(md, offset);
1631 dev_idx = lnet_get_dev_idx(page);
1635 * If there is no peer_ni that we can send to on this network,
1636 * then there is no point in looking for a new best_ni here.
1638 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1641 if (best_ni == NULL) {
1642 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1643 shortest_distance = UINT_MAX;
1644 best_dev_prio = UINT_MAX;
1645 best_credits = INT_MIN;
1647 best_ni_fatal = true;
1649 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1650 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1651 best_ni->ni_dev_cpt);
1652 best_credits = atomic_read(&best_ni->ni_tx_credits);
1653 best_healthv = atomic_read(&best_ni->ni_healthv);
1654 best_sel_prio = best_ni->ni_sel_priority;
1655 best_ni_fatal = atomic_read(&best_ni->ni_fatal_error_on);
1658 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1659 unsigned int distance;
1664 unsigned int ni_dev_prio;
1666 ni_credits = atomic_read(&ni->ni_tx_credits);
1667 ni_healthv = atomic_read(&ni->ni_healthv);
1668 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1669 ni_sel_prio = ni->ni_sel_priority;
1672 * calculate the distance from the CPT on which
1673 * the message memory is allocated to the CPT of
1674 * the NI's physical device
1676 distance = cfs_cpt_distance(lnet_cpt_table(),
1680 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1683 * All distances smaller than the NUMA range
1684 * are treated equally.
1686 if (!gpu && distance < lnet_numa_range)
1687 distance = lnet_numa_range;
1690 * Select on health, selection policy, direct dma prio,
1691 * shorter distance, available credits, then round-robin.
1694 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",
1695 libcfs_nidstr(&ni->ni_nid),
1696 ni_fatal ? "y" : "n", ni_credits, distance,
1697 ni->ni_seq, ni_sel_prio, ni_dev_prio, ni_healthv,
1698 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1700 best_ni_fatal ? "y" : "n", best_credits,
1702 (best_ni) ? best_ni->ni_seq : 0,
1703 best_sel_prio, best_dev_prio, best_healthv);
1707 if (ni_fatal && !best_ni_fatal)
1709 else if (!ni_fatal && best_ni_fatal)
1712 if (ni_healthv < best_healthv)
1714 else if (ni_healthv > best_healthv)
1717 if (ni_sel_prio > best_sel_prio)
1719 else if (ni_sel_prio < best_sel_prio)
1722 if (ni_dev_prio > best_dev_prio)
1724 else if (ni_dev_prio < best_dev_prio)
1727 if (distance > shortest_distance)
1729 else if (distance < shortest_distance)
1732 if (ni_credits < best_credits)
1734 else if (ni_credits > best_credits)
1737 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1741 best_sel_prio = ni_sel_prio;
1742 best_dev_prio = ni_dev_prio;
1743 shortest_distance = distance;
1744 best_healthv = ni_healthv;
1746 best_credits = ni_credits;
1747 best_ni_fatal = ni_fatal;
1750 CDEBUG(D_NET, "selected best_ni %s\n",
1751 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1757 lnet_reserved_msg(struct lnet_msg *msg)
1759 if (msg->msg_type == LNET_MSG_PUT) {
1760 if (msg->msg_hdr.msg.put.ptl_index == LNET_RESERVED_PORTAL)
1762 } else if (msg->msg_type == LNET_MSG_GET) {
1763 if (msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1769 /* Can the specified message trigger peer discovery?
1771 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1772 * because such traffic is required to perform discovery. We therefore
1773 * exclude all GET and PUT on that portal. We also exclude all ACK and
1774 * REPLY traffic, but that is because the portal is not tracked in the
1775 * message structure for these message types. We could restrict this
1776 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1779 lnet_msg_discovery(struct lnet_msg *msg)
1781 return !(lnet_reserved_msg(msg) || lnet_msg_is_response(msg));
1784 /* Is the specified message an LNet ping?
1787 lnet_msg_is_ping(struct lnet_msg *msg)
1789 if (msg->msg_type == LNET_MSG_GET &&
1790 msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1796 #define SRC_SPEC 0x0001
1797 #define SRC_ANY 0x0002
1798 #define LOCAL_DST 0x0004
1799 #define REMOTE_DST 0x0008
1800 #define MR_DST 0x0010
1801 #define NMR_DST 0x0020
1802 #define SND_RESP 0x0040
1804 /* The following to defines are used for return codes */
1805 #define REPEAT_SEND 0x1000
1806 #define PASS_THROUGH 0x2000
1808 /* The different cases lnet_select pathway needs to handle */
1809 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1810 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1811 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1812 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1813 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1814 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1815 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1816 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1819 lnet_handle_lo_send(struct lnet_send_data *sd)
1821 struct lnet_msg *msg = sd->sd_msg;
1822 int cpt = sd->sd_cpt;
1824 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1827 /* No send credit hassles with LOLND */
1828 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1829 msg->msg_hdr.dest_nid = the_lnet.ln_loni->ni_nid;
1830 if (!msg->msg_routing)
1831 msg->msg_hdr.src_nid = the_lnet.ln_loni->ni_nid;
1832 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1833 lnet_msg_commit(msg, cpt);
1834 msg->msg_txni = the_lnet.ln_loni;
1836 return LNET_CREDIT_OK;
1840 lnet_handle_send(struct lnet_send_data *sd)
1842 struct lnet_ni *best_ni = sd->sd_best_ni;
1843 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1844 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1845 struct lnet_msg *msg = sd->sd_msg;
1847 __u32 send_case = sd->sd_send_case;
1849 __u32 routing = send_case & REMOTE_DST;
1850 struct lnet_rsp_tracker *rspt;
1852 /* Increment sequence number of the selected peer, peer net,
1853 * local ni and local net so that we pick the next ones
1856 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1857 best_lpni->lpni_peer_net->lpn_seq =
1858 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1859 best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1860 the_lnet.ln_net_seq++;
1861 best_ni->ni_net->net_seq = the_lnet.ln_net_seq;
1862 best_ni->ni_seq = best_ni->ni_net->net_seq;
1864 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1865 libcfs_nidstr(&best_ni->ni_nid),
1866 best_ni->ni_seq, best_ni->ni_net->net_seq,
1867 atomic_read(&best_ni->ni_tx_credits),
1868 best_ni->ni_sel_priority,
1869 libcfs_nidstr(&best_lpni->lpni_nid),
1870 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1871 best_lpni->lpni_txcredits,
1872 best_lpni->lpni_sel_priority);
1875 * grab a reference on the peer_ni so it sticks around even if
1876 * we need to drop and relock the lnet_net_lock below.
1878 lnet_peer_ni_addref_locked(best_lpni);
1881 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1882 * message. This ensures that we get a CPT that is correct for
1883 * the NI when the NI has been restricted to a subset of all CPTs.
1884 * If the selected CPT differs from the one currently locked, we
1885 * must unlock and relock the lnet_net_lock(), and then check whether
1886 * the configuration has changed. We don't have a hold on the best_ni
1887 * yet, and it may have vanished.
1889 cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1890 if (sd->sd_cpt != cpt2) {
1891 __u32 seq = lnet_get_dlc_seq_locked();
1892 lnet_net_unlock(sd->sd_cpt);
1894 lnet_net_lock(sd->sd_cpt);
1895 if (seq != lnet_get_dlc_seq_locked()) {
1896 lnet_peer_ni_decref_locked(best_lpni);
1902 * store the best_lpni in the message right away to avoid having
1903 * to do the same operation under different conditions
1905 msg->msg_txpeer = best_lpni;
1906 msg->msg_txni = best_ni;
1909 * grab a reference for the best_ni since now it's in use in this
1910 * send. The reference will be dropped in lnet_finalize()
1912 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1915 * Always set the target.nid to the best peer picked. Either the
1916 * NID will be one of the peer NIDs selected, or the same NID as
1917 * what was originally set in the target or it will be the NID of
1918 * a router if this message should be routed
1920 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1923 * lnet_msg_commit assigns the correct cpt to the message, which
1924 * is used to decrement the correct refcount on the ni when it's
1925 * time to return the credits
1927 lnet_msg_commit(msg, sd->sd_cpt);
1930 * If we are routing the message then we keep the src_nid that was
1931 * set by the originator. If we are not routing then we are the
1932 * originator and set it here.
1934 if (!msg->msg_routing)
1935 msg->msg_hdr.src_nid = msg->msg_txni->ni_nid;
1938 msg->msg_target_is_router = 1;
1939 msg->msg_target.pid = LNET_PID_LUSTRE;
1941 * since we're routing we want to ensure that the
1942 * msg_hdr.dest_nid is set to the final destination. When
1943 * the router receives this message it knows how to route
1946 * final_dst_lpni is set at the beginning of the
1947 * lnet_select_pathway() function and is never changed.
1948 * It's safe to use it here.
1950 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1951 final_dst_lpni->lpni_peer_net->lpn_seq =
1952 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1953 final_dst_lpni->lpni_seq =
1954 final_dst_lpni->lpni_peer_net->lpn_seq;
1955 msg->msg_hdr.dest_nid = final_dst_lpni->lpni_nid;
1958 * if we're not routing set the dest_nid to the best peer
1959 * ni NID that we picked earlier in the algorithm.
1961 msg->msg_hdr.dest_nid = msg->msg_txpeer->lpni_nid;
1965 * if we have response tracker block update it with the next hop
1969 rspt = msg->msg_md->md_rspt_ptr;
1971 rspt->rspt_next_hop_nid =
1972 msg->msg_txpeer->lpni_nid;
1973 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1974 libcfs_nidstr(&rspt->rspt_next_hop_nid));
1978 rc = lnet_post_send_locked(msg, 0);
1981 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1982 libcfs_nidstr(&msg->msg_hdr.src_nid),
1983 libcfs_nidstr(&msg->msg_txni->ni_nid),
1984 libcfs_nidstr(&sd->sd_src_nid),
1985 libcfs_nidstr(&msg->msg_hdr.dest_nid),
1986 libcfs_nidstr(&sd->sd_dst_nid),
1987 libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
1988 libcfs_nidstr(&sd->sd_rtr_nid),
1989 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1995 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1996 struct lnet_msg *msg)
1998 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1999 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
2000 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
2001 libcfs_nidstr(&lni->ni_nid),
2002 libcfs_nidstr(&lpni->lpni_nid));
2003 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
2012 * use the source and destination NIDs as the pathway
2015 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2017 /* the destination lpni is set before we get here. */
2020 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2021 if (!sd->sd_best_ni) {
2022 CERROR("Can't send to %s: src %s is not a local nid\n",
2023 libcfs_nidstr(&sd->sd_dst_nid),
2024 libcfs_nidstr(&sd->sd_src_nid));
2028 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2030 return lnet_handle_send(sd);
2038 * Don't run the selection algorithm on the peer NIs. By specifying the
2039 * local NID, we're also saying that we should always use the destination NID
2040 * provided. This handles the case where we should be using the same
2041 * destination NID for the all the messages which belong to the same RPC
2045 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2047 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2048 if (!sd->sd_best_ni) {
2049 CERROR("Can't send to %s: src %s is not a local nid\n",
2050 libcfs_nidstr(&sd->sd_dst_nid),
2051 libcfs_nidstr(&sd->sd_src_nid));
2055 if (sd->sd_best_lpni &&
2056 nid_same(&sd->sd_best_lpni->lpni_nid,
2057 &the_lnet.ln_loni->ni_nid))
2058 return lnet_handle_lo_send(sd);
2059 else if (sd->sd_best_lpni)
2060 return lnet_handle_send(sd);
2062 CERROR("can't send to %s. no NI on %s\n",
2063 libcfs_nidstr(&sd->sd_dst_nid),
2064 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2066 return -EHOSTUNREACH;
2070 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2071 struct lnet_peer *peer,
2072 struct lnet_peer_net *peer_net,
2073 struct lnet_msg *msg,
2076 struct lnet_net *local_net;
2077 struct lnet_ni *best_ni;
2079 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2084 * Iterate through the NIs in this local Net and select
2085 * the NI to send from. The selection is determined by
2086 * these 3 criterion in the following priority:
2088 * 2. NI available credits
2091 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2092 peer, peer_net, msg, cpt);
2098 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2101 struct lnet_peer *peer;
2102 struct lnet_peer_ni *new_lpni;
2105 lnet_peer_ni_addref_locked(lpni);
2107 peer = lpni->lpni_peer_net->lpn_peer;
2109 if (lnet_peer_gw_discovery(peer)) {
2110 lnet_peer_ni_decref_locked(lpni);
2114 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2115 lnet_peer_ni_decref_locked(lpni);
2119 rc = lnet_discover_peer_locked(lpni, cpt, false);
2121 lnet_peer_ni_decref_locked(lpni);
2125 new_lpni = lnet_peer_ni_find_locked(&lpni->lpni_nid);
2127 lnet_peer_ni_decref_locked(lpni);
2131 peer = new_lpni->lpni_peer_net->lpn_peer;
2132 spin_lock(&peer->lp_lock);
2133 if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2134 /* The peer NI did not change and the peer is up to date.
2135 * Nothing more to do.
2137 spin_unlock(&peer->lp_lock);
2138 lnet_peer_ni_decref_locked(lpni);
2139 lnet_peer_ni_decref_locked(new_lpni);
2142 spin_unlock(&peer->lp_lock);
2144 /* Either the peer NI changed during discovery, or the peer isn't up
2145 * to date. In both cases we want to queue the message on the
2146 * (possibly new) peer's pending queue and queue the peer for discovery
2148 msg->msg_sending = 0;
2149 msg->msg_txpeer = NULL;
2150 lnet_net_unlock(cpt);
2151 lnet_peer_queue_message(peer, msg);
2154 lnet_peer_ni_decref_locked(lpni);
2155 lnet_peer_ni_decref_locked(new_lpni);
2157 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2158 msg, libcfs_nidstr(&peer->lp_primary_nid));
2160 return LNET_DC_WAIT;
2164 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2165 struct lnet_nid *dst_nid,
2166 struct lnet_peer_ni **gw_lpni,
2167 struct lnet_peer **gw_peer)
2170 struct lnet_peer *gw;
2171 struct lnet_peer *lp;
2172 struct lnet_peer_net *lpn;
2173 struct lnet_peer_net *best_lpn = NULL;
2174 struct lnet_remotenet *rnet, *best_rnet = NULL;
2175 struct lnet_route *best_route = NULL;
2176 struct lnet_route *last_route = NULL;
2177 struct lnet_peer_ni *lpni = NULL;
2178 struct lnet_peer_ni *gwni = NULL;
2179 bool route_found = false;
2180 struct lnet_nid *src_nid =
2181 !LNET_NID_IS_ANY(&sd->sd_src_nid) || !sd->sd_best_ni
2183 : &sd->sd_best_ni->ni_nid;
2184 int best_lpn_healthv = 0;
2185 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2187 CDEBUG(D_NET, "%s route (%s) from local NI %s to destination %s\n",
2188 LNET_NID_IS_ANY(&sd->sd_rtr_nid) ? "Lookup" : "Specified",
2189 libcfs_nidstr(&sd->sd_rtr_nid), libcfs_nidstr(src_nid),
2190 libcfs_nidstr(&sd->sd_dst_nid));
2192 /* If a router nid was specified then we are replying to a GET or
2193 * sending an ACK. In this case we use the gateway associated with the
2194 * specified router nid.
2196 if (!LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2197 gwni = lnet_peer_ni_find_locked(&sd->sd_rtr_nid);
2199 gw = gwni->lpni_peer_net->lpn_peer;
2200 lnet_peer_ni_decref_locked(gwni);
2201 if (gw->lp_rtr_refcount)
2204 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2205 libcfs_nidstr(&sd->sd_rtr_nid));
2210 if (sd->sd_msg->msg_routing || !LNET_NID_IS_ANY(src_nid)) {
2211 /* If I'm routing this message then I need to find the
2212 * next hop based on the destination NID
2214 * We also find next hop based on the destination NID
2215 * if the source NI was specified
2217 best_rnet = lnet_find_rnet_locked(LNET_NID_NET(&sd->sd_dst_nid));
2219 CERROR("Unable to send message from %s to %s - Route table may be misconfigured\n",
2220 (src_nid && LNET_NID_IS_ANY(src_nid)) ?
2222 libcfs_nidstr(src_nid),
2223 libcfs_nidstr(&sd->sd_dst_nid));
2224 return -EHOSTUNREACH;
2226 CDEBUG(D_NET, "best_rnet %s\n",
2227 libcfs_net2str(best_rnet->lrn_net));
2229 /* we've already looked up the initial lpni using
2232 lpni = sd->sd_best_lpni;
2233 /* the peer tree must be in existence */
2234 LASSERT(lpni && lpni->lpni_peer_net &&
2235 lpni->lpni_peer_net->lpn_peer);
2236 lp = lpni->lpni_peer_net->lpn_peer;
2238 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2239 /* is this remote network reachable? */
2240 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2247 CDEBUG(D_NET, "n[%s, %s] h[%d, %d], p[%u, %u], s[%d, %d]\n",
2248 libcfs_net2str(lpn->lpn_net_id),
2249 libcfs_net2str(best_lpn->lpn_net_id),
2251 best_lpn->lpn_healthv,
2252 lpn->lpn_sel_priority,
2253 best_lpn->lpn_sel_priority,
2257 /* select the preferred peer net */
2258 if (best_lpn_healthv > lpn->lpn_healthv)
2260 else if (best_lpn_healthv < lpn->lpn_healthv)
2263 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2265 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2268 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2271 best_lpn_healthv = lpn->lpn_healthv;
2272 best_lpn_sel_prio = lpn->lpn_sel_priority;
2278 CERROR("peer %s has no available nets\n",
2279 libcfs_nidstr(&sd->sd_dst_nid));
2280 return -EHOSTUNREACH;
2283 CDEBUG(D_NET, "selected best_lpn %s\n",
2284 libcfs_net2str(best_lpn->lpn_net_id));
2286 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2287 lnet_nid_to_nid4(&sd->sd_dst_nid),
2289 best_lpn->lpn_net_id);
2290 if (!sd->sd_best_lpni) {
2291 CERROR("peer %s is unreachable\n",
2292 libcfs_nidstr(&sd->sd_dst_nid));
2293 return -EHOSTUNREACH;
2296 /* We're attempting to round robin over the remote peer
2297 * NI's so update the final destination we selected
2299 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2303 * find the best route. Restrict the selection on the net of the
2304 * local NI if we've already picked the local NI to send from.
2305 * Otherwise, let's pick any route we can find and then find
2306 * a local NI we can reach the route's gateway on. Any route we
2307 * select will be reachable by virtue of the restriction we have
2308 * when adding a route.
2310 best_route = lnet_find_route_locked(best_rnet,
2311 LNET_NID_NET(src_nid),
2313 &last_route, &gwni);
2316 CERROR("no route to %s from %s\n",
2317 libcfs_nidstr(dst_nid),
2318 libcfs_nidstr(src_nid));
2319 return -EHOSTUNREACH;
2323 CERROR("Internal Error. Route expected to %s from %s\n",
2324 libcfs_nidstr(dst_nid),
2325 libcfs_nidstr(src_nid));
2329 gw = best_route->lr_gateway;
2330 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2334 * If the router checker is not active then discover the gateway here.
2335 * This ensures we are able to take advantage of multi-rail routing, but
2336 * if the router checker is active then we do not unecessarily delay
2337 * messages while the gateway is being checked by the dedicated monitor
2340 * NB: We're only checking the alive_router_check_interval here, rather
2341 * than calling lnet_router_checker_active(), because the other
2342 * conditions that are checked by that function are either
2343 * irrelevant (the_lnet.ln_routing) or must be true (list of routers
2346 if (alive_router_check_interval <= 0) {
2347 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2352 if (!sd->sd_best_ni) {
2353 lpn = gwni->lpni_peer_net;
2354 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2357 if (!sd->sd_best_ni) {
2358 CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2359 libcfs_net2str(lpn->lpn_net_id),
2360 libcfs_nidstr(&sd->sd_src_nid));
2369 * increment the sequence number since now we're sure we're
2370 * going to use this route
2372 if (LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2373 LASSERT(best_route && last_route);
2374 best_route->lr_seq = last_route->lr_seq + 1;
2385 * Remote destination
2386 * Non-MR destination
2390 * Remote destination
2393 * The handling of these two cases is similar. Even though the destination
2394 * can be MR or non-MR, we'll deal directly with the router.
2397 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2400 struct lnet_peer_ni *gw_lpni = NULL;
2401 struct lnet_peer *gw_peer = NULL;
2404 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2405 if (!sd->sd_best_ni) {
2406 CERROR("Can't send to %s: src %s is not a local nid\n",
2407 libcfs_nidstr(&sd->sd_dst_nid),
2408 libcfs_nidstr(&sd->sd_src_nid));
2412 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2413 &gw_lpni, &gw_peer);
2417 if (sd->sd_send_case & NMR_DST)
2419 * since the final destination is non-MR let's set its preferred
2420 * NID before we send
2422 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2426 * We're going to send to the gw found so let's set its
2429 sd->sd_peer = gw_peer;
2430 sd->sd_best_lpni = gw_lpni;
2432 return lnet_handle_send(sd);
2436 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2437 struct lnet_msg *msg, bool discovery)
2439 struct lnet_peer_net *lpn = NULL;
2440 struct lnet_peer_net *best_lpn = NULL;
2441 struct lnet_net *net = NULL;
2442 struct lnet_net *best_net = NULL;
2443 struct lnet_ni *best_ni = NULL;
2444 int best_lpn_healthv = 0;
2445 int best_net_healthv = 0;
2447 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2449 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2452 /* If lp_disc_net_id is set, this peer is a router undergoing
2453 * discovery, and this message is an LNet ping, then this may be a
2454 * discovery message and we need to select an NI on the peer net
2455 * specified by lp_disc_net_id
2457 if (peer->lp_disc_net_id &&
2458 (peer->lp_state & LNET_PEER_RTR_DISCOVERY) &&
2459 lnet_msg_is_ping(msg)) {
2460 best_lpn = lnet_peer_get_net_locked(peer, peer->lp_disc_net_id);
2461 if (best_lpn && lnet_get_net_locked(best_lpn->lpn_net_id))
2462 goto select_best_ni;
2466 * The peer can have multiple interfaces, some of them can be on
2467 * the local network and others on a routed network. We should
2468 * prefer the local network. However if the local network is not
2469 * available then we need to try the routed network
2472 /* go through all the peer nets and find the best_ni */
2473 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2475 * The peer's list of nets can contain non-local nets. We
2476 * want to only examine the local ones.
2478 net = lnet_get_net_locked(lpn->lpn_net_id);
2482 lpn_sel_prio = lpn->lpn_sel_priority;
2483 net_healthv = lnet_get_net_healthv_locked(net);
2484 net_sel_prio = net->net_sel_priority;
2490 "n[%s, %s] ph[%d, %d], pp[%u, %u], nh[%d, %d], np[%u, %u], ps[%u, %u], ns[%u, %u]\n",
2491 libcfs_net2str(lpn->lpn_net_id),
2492 libcfs_net2str(best_lpn->lpn_net_id),
2506 /* always select the lpn with the best health */
2507 if (best_lpn_healthv > lpn->lpn_healthv)
2509 else if (best_lpn_healthv < lpn->lpn_healthv)
2512 /* select the preferred peer and local nets */
2513 if (best_lpn_sel_prio < lpn_sel_prio)
2515 else if (best_lpn_sel_prio > lpn_sel_prio)
2518 if (best_net_healthv > net_healthv)
2520 else if (best_net_healthv < net_healthv)
2523 if (best_net_sel_prio < net_sel_prio)
2525 else if (best_net_sel_prio > net_sel_prio)
2528 if (best_lpn->lpn_seq < lpn->lpn_seq)
2530 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2533 /* round robin over the local networks */
2534 if (best_net->net_seq <= net->net_seq)
2538 best_net_healthv = net_healthv;
2539 best_net_sel_prio = net_sel_prio;
2540 best_lpn_healthv = lpn->lpn_healthv;
2541 best_lpn_sel_prio = lpn_sel_prio;
2547 /* Select the best NI on the same net as best_lpn chosen
2551 CDEBUG(D_NET, "selected best_lpn %s\n",
2552 libcfs_net2str(best_lpn->lpn_net_id));
2553 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2560 static struct lnet_ni *
2561 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2563 struct lnet_ni *best_ni = NULL;
2564 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2565 struct lnet_peer_ni *lpni_entry;
2568 * We must use a consistent source address when sending to a
2569 * non-MR peer. However, a non-MR peer can have multiple NIDs
2570 * on multiple networks, and we may even need to talk to this
2571 * peer on multiple networks -- certain types of
2572 * load-balancing configuration do this.
2574 * So we need to pick the NI the peer prefers for this
2575 * particular network.
2578 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2580 if (lpni_entry->lpni_pref_nnids == 0)
2582 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2583 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2591 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2593 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2595 struct lnet_ni *best_ni = NULL;
2598 * We must use a consistent source address when sending to a
2599 * non-MR peer. However, a non-MR peer can have multiple NIDs
2600 * on multiple networks, and we may even need to talk to this
2601 * peer on multiple networks -- certain types of
2602 * load-balancing configuration do this.
2604 * So we need to pick the NI the peer prefers for this
2605 * particular network.
2607 * An exception is traffic on LNET_RESERVED_PORTAL. Internal LNet
2608 * traffic doesn't care which source NI is used, and we don't actually
2609 * want to restrict local recovery pings to a single source NI.
2611 if (!lnet_reserved_msg(sd->sd_msg))
2612 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2616 best_ni = lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2617 sd->sd_best_lpni->lpni_peer_net,
2621 /* If there is no best_ni we don't have a route */
2623 CERROR("no path to %s from net %s\n",
2624 libcfs_nidstr(&sd->sd_best_lpni->lpni_nid),
2625 libcfs_net2str(sd->sd_best_lpni->lpni_net->net_id));
2626 return -EHOSTUNREACH;
2629 sd->sd_best_ni = best_ni;
2631 /* Set preferred NI if necessary. */
2632 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2639 * Source not specified
2643 * always use the same source NID for NMR peers
2644 * If we've talked to that peer before then we already have a preferred
2645 * source NI associated with it. Otherwise, we select a preferred local NI
2646 * and store it in the peer
2649 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2653 /* sd->sd_best_lpni is already set to the final destination */
2656 * At this point we should've created the peer ni and peer. If we
2657 * can't find it, then something went wrong. Instead of assert
2658 * output a relevant message and fail the send
2660 if (!sd->sd_best_lpni) {
2661 CERROR("Internal fault. Unable to send msg %s to %s. NID not known\n",
2662 lnet_msgtyp2str(sd->sd_msg->msg_type),
2663 libcfs_nidstr(&sd->sd_dst_nid));
2667 if (sd->sd_msg->msg_routing) {
2668 /* If I'm forwarding this message then I can choose any NI
2669 * on the destination peer net
2671 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2673 sd->sd_best_lpni->lpni_peer_net,
2676 if (!sd->sd_best_ni) {
2677 CERROR("Unable to forward message to %s. No local NI available\n",
2678 libcfs_nidstr(&sd->sd_dst_nid));
2682 rc = lnet_select_preferred_best_ni(sd);
2685 rc = lnet_handle_send(sd);
2691 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2694 * NOTE we've already handled the remote peer case. So we only
2695 * need to worry about the local case here.
2697 * if we're sending a response, ACK or reply, we need to send it
2698 * to the destination NID given to us. At this point we already
2699 * have the peer_ni we're suppose to send to, so just find the
2700 * best_ni on the peer net and use that. Since we're sending to an
2701 * MR peer then we can just run the selection algorithm on our
2702 * local NIs and pick the best one.
2704 if (sd->sd_send_case & SND_RESP) {
2706 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2707 sd->sd_best_lpni->lpni_peer_net,
2711 if (!sd->sd_best_ni) {
2713 * We're not going to deal with not able to send
2714 * a response to the provided final destination
2716 CERROR("Can't send response to %s. No local NI available\n",
2717 libcfs_nidstr(&sd->sd_dst_nid));
2718 return -EHOSTUNREACH;
2721 return lnet_handle_send(sd);
2725 * If we get here that means we're sending a fresh request, PUT or
2726 * GET, so we need to run our standard selection algorithm.
2727 * First find the best local interface that's on any of the peer's
2730 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2733 lnet_msg_discovery(sd->sd_msg));
2734 if (sd->sd_best_ni) {
2736 lnet_find_best_lpni(sd->sd_best_ni,
2737 lnet_nid_to_nid4(&sd->sd_dst_nid),
2739 sd->sd_best_ni->ni_net->net_id);
2742 * if we're successful in selecting a peer_ni on the local
2743 * network, then send to it. Otherwise fall through and
2744 * try and see if we can reach it over another routed
2747 if (sd->sd_best_lpni &&
2748 nid_same(&sd->sd_best_lpni->lpni_nid,
2749 &the_lnet.ln_loni->ni_nid)) {
2751 * in case we initially started with a routed
2752 * destination, let's reset to local
2754 sd->sd_send_case &= ~REMOTE_DST;
2755 sd->sd_send_case |= LOCAL_DST;
2756 return lnet_handle_lo_send(sd);
2757 } else if (sd->sd_best_lpni) {
2759 * in case we initially started with a routed
2760 * destination, let's reset to local
2762 sd->sd_send_case &= ~REMOTE_DST;
2763 sd->sd_send_case |= LOCAL_DST;
2764 return lnet_handle_send(sd);
2767 CERROR("Internal Error. Expected to have a best_lpni: "
2769 libcfs_nidstr(&sd->sd_src_nid),
2770 libcfs_nidstr(&sd->sd_dst_nid));
2776 * Peer doesn't have a local network. Let's see if there is
2777 * a remote network we can reach it on.
2779 return PASS_THROUGH;
2784 * Source NID not specified
2789 * Source NID not speified
2790 * Remote destination
2793 * In both of these cases if we're sending a response, ACK or REPLY, then
2794 * we need to send to the destination NID provided.
2796 * In the remote case let's deal with MR routers.
2801 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2804 struct lnet_peer *gw_peer = NULL;
2805 struct lnet_peer_ni *gw_lpni = NULL;
2808 * handle sending a response to a remote peer here so we don't
2809 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2811 if (sd->sd_send_case & REMOTE_DST &&
2812 sd->sd_send_case & SND_RESP) {
2813 struct lnet_peer_ni *gw;
2814 struct lnet_peer *gw_peer;
2816 rc = lnet_handle_find_routed_path(
2817 sd, &sd->sd_dst_nid, &gw, &gw_peer);
2819 CERROR("Can't send response to %s. No route available\n",
2820 libcfs_nidstr(&sd->sd_dst_nid));
2821 return -EHOSTUNREACH;
2822 } else if (rc > 0) {
2826 sd->sd_best_lpni = gw;
2827 sd->sd_peer = gw_peer;
2829 return lnet_handle_send(sd);
2833 * Even though the NID for the peer might not be on a local network,
2834 * since the peer is MR there could be other interfaces on the
2835 * local network. In that case we'd still like to prefer the local
2836 * network over the routed network. If we're unable to do that
2837 * then we select the best router among the different routed networks,
2838 * and if the router is MR then we can deal with it as such.
2840 rc = lnet_handle_any_mr_dsta(sd);
2841 if (rc != PASS_THROUGH)
2845 * Now that we must route to the destination, we must consider the
2846 * MR case, where the destination has multiple interfaces, some of
2847 * which we can route to and others we do not. For this reason we
2848 * need to select the destination which we can route to and if
2849 * there are multiple, we need to round robin.
2851 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2852 &gw_lpni, &gw_peer);
2856 sd->sd_send_case &= ~LOCAL_DST;
2857 sd->sd_send_case |= REMOTE_DST;
2859 sd->sd_peer = gw_peer;
2860 sd->sd_best_lpni = gw_lpni;
2862 return lnet_handle_send(sd);
2866 * Source not specified
2867 * Remote destination
2870 * Must send to the specified peer NID using the same source NID that
2871 * we've used before. If it's the first time to talk to that peer then
2872 * find the source NI and assign it as preferred to that peer
2875 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2878 struct lnet_peer_ni *gw_lpni = NULL;
2879 struct lnet_peer *gw_peer = NULL;
2882 * Let's see if we have a preferred NI to talk to this NMR peer
2884 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2888 * find the router and that'll find the best NI if we didn't find
2891 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid, &gw_lpni,
2897 * set the best_ni we've chosen as the preferred one for
2900 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2902 /* we'll be sending to the gw */
2903 sd->sd_best_lpni = gw_lpni;
2904 sd->sd_peer = gw_peer;
2906 return lnet_handle_send(sd);
2910 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2913 * turn off the SND_RESP bit.
2914 * It will be checked in the case handling
2916 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2918 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2919 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2920 (send_case & SRC_SPEC) ? libcfs_nidstr(&sd->sd_src_nid) : "",
2921 (send_case & MR_DST) ? "MR: " : "NMR: ",
2922 libcfs_nidstr(&sd->sd_dst_nid),
2923 (send_case & LOCAL_DST) ? "local" : "routed");
2925 switch (send_case) {
2927 * For all cases where the source is specified, we should always
2928 * use the destination NID, whether it's an MR destination or not,
2929 * since we're continuing a series of related messages for the
2932 case SRC_SPEC_LOCAL_NMR_DST:
2933 return lnet_handle_spec_local_nmr_dst(sd);
2934 case SRC_SPEC_LOCAL_MR_DST:
2935 return lnet_handle_spec_local_mr_dst(sd);
2936 case SRC_SPEC_ROUTER_NMR_DST:
2937 case SRC_SPEC_ROUTER_MR_DST:
2938 return lnet_handle_spec_router_dst(sd);
2939 case SRC_ANY_LOCAL_NMR_DST:
2940 return lnet_handle_any_local_nmr_dst(sd);
2941 case SRC_ANY_LOCAL_MR_DST:
2942 case SRC_ANY_ROUTER_MR_DST:
2943 return lnet_handle_any_mr_dst(sd);
2944 case SRC_ANY_ROUTER_NMR_DST:
2945 return lnet_handle_any_router_nmr_dst(sd);
2947 CERROR("Unknown send case\n");
2953 lnet_select_pathway(struct lnet_nid *src_nid,
2954 struct lnet_nid *dst_nid,
2955 struct lnet_msg *msg,
2956 struct lnet_nid *rtr_nid)
2958 struct lnet_peer_ni *lpni;
2959 struct lnet_peer *peer;
2960 struct lnet_send_data send_data;
2963 __u32 send_case = 0;
2965 bool mr_forwarding_allowed;
2967 memset(&send_data, 0, sizeof(send_data));
2970 * get an initial CPT to use for locking. The idea here is not to
2971 * serialize the calls to select_pathway, so that as many
2972 * operations can run concurrently as possible. To do that we use
2973 * the CPT where this call is being executed. Later on when we
2974 * determine the CPT to use in lnet_message_commit, we switch the
2975 * lock and check if there was any configuration change. If none,
2976 * then we proceed, if there is, then we restart the operation.
2978 cpt = lnet_net_lock_current();
2980 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2981 if (md_cpt == CFS_CPT_ANY)
2987 * If we're being asked to send to the loopback interface, there
2988 * is no need to go through any selection. We can just shortcut
2989 * the entire process and send over lolnd
2991 send_data.sd_msg = msg;
2992 send_data.sd_cpt = cpt;
2993 if (nid_is_lo0(dst_nid)) {
2994 rc = lnet_handle_lo_send(&send_data);
2995 lnet_net_unlock(cpt);
3000 * find an existing peer_ni, or create one and mark it as having been
3001 * created due to network traffic. This call will create the
3002 * peer->peer_net->peer_ni tree.
3004 lpni = lnet_peerni_by_nid_locked(dst_nid, NULL, cpt);
3006 lnet_net_unlock(cpt);
3007 return PTR_ERR(lpni);
3011 * Cache the original src_nid and rtr_nid. If we need to resend the
3012 * message then we'll need to know whether the src_nid was originally
3013 * specified for this message. If it was originally specified,
3014 * then we need to keep using the same src_nid since it's
3015 * continuing the same sequence of messages. Similarly, rtr_nid will
3016 * affect our choice of next hop.
3019 msg->msg_src_nid_param = *src_nid;
3021 msg->msg_src_nid_param = LNET_ANY_NID;
3023 msg->msg_rtr_nid_param = *rtr_nid;
3025 msg->msg_rtr_nid_param = LNET_ANY_NID;
3028 * If necessary, perform discovery on the peer that owns this peer_ni.
3029 * Note, this can result in the ownership of this peer_ni changing
3030 * to another peer object.
3032 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
3034 lnet_peer_ni_decref_locked(lpni);
3035 lnet_net_unlock(cpt);
3038 lnet_peer_ni_decref_locked(lpni);
3040 peer = lpni->lpni_peer_net->lpn_peer;
3043 * Identify the different send cases
3045 if (!src_nid || LNET_NID_IS_ANY(src_nid)) {
3046 send_case |= SRC_ANY;
3047 if (lnet_get_net_locked(LNET_NID_NET(dst_nid)))
3048 send_case |= LOCAL_DST;
3050 send_case |= REMOTE_DST;
3052 send_case |= SRC_SPEC;
3053 if (LNET_NID_NET(src_nid) == LNET_NID_NET(dst_nid))
3054 send_case |= LOCAL_DST;
3056 send_case |= REMOTE_DST;
3060 if (msg->msg_routing && (send_case & LOCAL_DST))
3063 /* Determine whether to allow MR forwarding for this message.
3064 * NB: MR forwarding is allowed if the message originator and the
3065 * destination are both MR capable, and the destination lpni that was
3066 * originally chosen by the originator is unhealthy or down.
3067 * We check the MR capability of the destination further below
3069 mr_forwarding_allowed = false;
3071 struct lnet_peer *src_lp;
3072 struct lnet_peer_ni *src_lpni;
3074 src_lpni = lnet_peerni_by_nid_locked(&msg->msg_hdr.src_nid,
3076 /* We don't fail the send if we hit any errors here. We'll just
3077 * try to send it via non-multi-rail criteria
3079 if (!IS_ERR(src_lpni)) {
3080 /* Drop ref taken by lnet_peerni_by_nid_locked() */
3081 lnet_peer_ni_decref_locked(src_lpni);
3082 src_lp = lpni->lpni_peer_net->lpn_peer;
3083 if (lnet_peer_is_multi_rail(src_lp) &&
3084 !lnet_is_peer_ni_alive(lpni))
3085 mr_forwarding_allowed = true;
3088 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3089 mr_forwarding_allowed ? "allowed" : "not allowed");
3093 * Deal with the peer as NMR in the following cases:
3094 * 1. the peer is NMR
3095 * 2. We're trying to recover a specific peer NI
3096 * 3. I'm a router sending to the final destination and MR forwarding is
3097 * not allowed for this message (as determined above).
3098 * In this case the source of the message would've
3099 * already selected the final destination so my job
3100 * is to honor the selection.
3102 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3103 (final_hop && !mr_forwarding_allowed))
3104 send_case |= NMR_DST;
3106 send_case |= MR_DST;
3108 if (lnet_msg_is_response(msg))
3109 send_case |= SND_RESP;
3111 /* assign parameters to the send_data */
3113 send_data.sd_rtr_nid = *rtr_nid;
3115 send_data.sd_rtr_nid = LNET_ANY_NID;
3117 send_data.sd_src_nid = *src_nid;
3119 send_data.sd_src_nid = LNET_ANY_NID;
3120 send_data.sd_dst_nid = *dst_nid;
3121 send_data.sd_best_lpni = lpni;
3123 * keep a pointer to the final destination in case we're going to
3124 * route, so we'll need to access it later
3126 send_data.sd_final_dst_lpni = lpni;
3127 send_data.sd_peer = peer;
3128 send_data.sd_md_cpt = md_cpt;
3129 send_data.sd_send_case = send_case;
3131 rc = lnet_handle_send_case_locked(&send_data);
3134 * Update the local cpt since send_data.sd_cpt might've been
3135 * updated as a result of calling lnet_handle_send_case_locked().
3137 cpt = send_data.sd_cpt;
3139 if (rc == REPEAT_SEND)
3142 lnet_net_unlock(cpt);
3148 lnet_send(struct lnet_nid *src_nid, struct lnet_msg *msg,
3149 struct lnet_nid *rtr_nid)
3151 struct lnet_nid *dst_nid = &msg->msg_target.nid;
3154 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3155 LASSERT(msg->msg_txpeer == NULL);
3156 LASSERT(msg->msg_txni == NULL);
3157 LASSERT(!msg->msg_sending);
3158 LASSERT(!msg->msg_target_is_router);
3159 LASSERT(!msg->msg_receiving);
3161 msg->msg_sending = 1;
3163 LASSERT(!msg->msg_tx_committed);
3165 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3167 if (rc == -EHOSTUNREACH)
3168 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3170 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3174 if (rc == LNET_CREDIT_OK)
3175 lnet_ni_send(msg->msg_txni, msg);
3177 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3181 enum lnet_mt_event_type {
3182 MT_TYPE_LOCAL_NI = 0,
3186 struct lnet_mt_event_info {
3187 enum lnet_mt_event_type mt_type;
3188 struct lnet_nid mt_nid;
3191 /* called with res_lock held */
3193 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3195 struct lnet_rsp_tracker *rspt;
3198 * msg has a refcount on the MD so the MD is not going away.
3199 * The rspt queue for the cpt is protected by
3200 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3202 if (!md->md_rspt_ptr)
3205 rspt = md->md_rspt_ptr;
3208 LASSERT(rspt->rspt_cpt == cpt);
3210 md->md_rspt_ptr = NULL;
3212 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3214 * The monitor thread has invalidated this handle because the
3215 * response timed out, but it failed to lookup the MD. That
3216 * means this response tracker is on the zombie list. We can
3217 * safely remove it under the resource lock (held by caller) and
3218 * free the response tracker block.
3220 list_del(&rspt->rspt_on_list);
3221 lnet_rspt_free(rspt, cpt);
3224 * invalidate the handle to indicate that a response has been
3225 * received, which will then lead the monitor thread to clean up
3228 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3233 lnet_clean_zombie_rstqs(void)
3235 struct lnet_rsp_tracker *rspt, *tmp;
3238 cfs_cpt_for_each(i, lnet_cpt_table()) {
3239 list_for_each_entry_safe(rspt, tmp,
3240 the_lnet.ln_mt_zombie_rstqs[i],
3242 list_del(&rspt->rspt_on_list);
3243 lnet_rspt_free(rspt, i);
3247 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3251 lnet_finalize_expired_responses(void)
3253 struct lnet_libmd *md;
3254 struct lnet_rsp_tracker *rspt, *tmp;
3258 if (the_lnet.ln_mt_rstq == NULL)
3261 cfs_cpt_for_each(i, lnet_cpt_table()) {
3262 LIST_HEAD(local_queue);
3265 if (!the_lnet.ln_mt_rstq[i]) {
3269 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3274 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3276 * The rspt mdh will be invalidated when a response
3277 * is received or whenever we want to discard the
3278 * block the monitor thread will walk the queue
3279 * and clean up any rsts with an invalid mdh.
3280 * The monitor thread will walk the queue until
3281 * the first unexpired rspt block. This means that
3282 * some rspt blocks which received their
3283 * corresponding responses will linger in the
3284 * queue until they are cleaned up eventually.
3287 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3289 list_del(&rspt->rspt_on_list);
3290 lnet_rspt_free(rspt, i);
3294 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3295 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3296 struct lnet_peer_ni *lpni;
3297 struct lnet_nid nid;
3299 md = lnet_handle2md(&rspt->rspt_mdh);
3301 /* MD has been queued for unlink, but
3302 * rspt hasn't been detached (Note we've
3303 * checked above that the rspt_mdh is
3304 * valid). Since we cannot lookup the MD
3305 * we're unable to detach the rspt
3306 * ourselves. Thus, move the rspt to the
3307 * zombie list where we'll wait for
3309 * 1. The remaining operations on the
3310 * MD to complete. In this case the
3311 * final operation will result in
3312 * lnet_msg_detach_md()->
3313 * lnet_detach_rsp_tracker() where
3314 * we will clean up this response
3316 * 2. LNet to shutdown. In this case
3317 * we'll wait until after all LND Nets
3318 * have shutdown and then we can
3319 * safely free any remaining response
3320 * tracker blocks on the zombie list.
3321 * Note: We need to hold the resource
3322 * lock when adding to the zombie list
3323 * because we may have concurrent access
3324 * with lnet_detach_rsp_tracker().
3326 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3327 list_move(&rspt->rspt_on_list,
3328 the_lnet.ln_mt_zombie_rstqs[i]);
3332 LASSERT(md->md_rspt_ptr == rspt);
3333 md->md_rspt_ptr = NULL;
3336 LNetMDUnlink(rspt->rspt_mdh);
3338 nid = rspt->rspt_next_hop_nid;
3340 list_del(&rspt->rspt_on_list);
3341 lnet_rspt_free(rspt, i);
3343 /* If we're shutting down we just want to clean
3344 * up the rspt blocks
3346 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3350 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3354 "Response timeout: md = %p: nid = %s\n",
3355 md, libcfs_nidstr(&nid));
3358 * If there is a timeout on the response
3359 * from the next hop decrement its health
3360 * value so that we don't use it
3363 lpni = lnet_peer_ni_find_locked(&nid);
3365 lnet_handle_remote_failure_locked(lpni);
3366 lnet_peer_ni_decref_locked(lpni);
3375 if (!list_empty(&local_queue)) {
3377 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3384 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3386 struct lnet_msg *msg;
3388 while (!list_empty(resendq)) {
3389 struct lnet_peer_ni *lpni;
3391 msg = list_entry(resendq->next, struct lnet_msg,
3394 list_del_init(&msg->msg_list);
3396 lpni = lnet_peer_ni_find_locked(&msg->msg_hdr.dest_nid);
3398 lnet_net_unlock(cpt);
3399 CERROR("Expected that a peer is already created for %s\n",
3400 libcfs_nidstr(&msg->msg_hdr.dest_nid));
3401 msg->msg_no_resend = true;
3402 lnet_finalize(msg, -EFAULT);
3407 lnet_peer_ni_decref_locked(lpni);
3409 lnet_net_unlock(cpt);
3410 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3411 libcfs_nidstr(&msg->msg_src_nid_param),
3412 libcfs_idstr(&msg->msg_target),
3413 lnet_msgtyp2str(msg->msg_type),
3415 msg->msg_retry_count);
3416 rc = lnet_send(&msg->msg_src_nid_param, msg,
3417 &msg->msg_rtr_nid_param);
3419 CERROR("Error sending %s to %s: %d\n",
3420 lnet_msgtyp2str(msg->msg_type),
3421 libcfs_idstr(&msg->msg_target), rc);
3422 msg->msg_no_resend = true;
3423 lnet_finalize(msg, rc);
3427 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3433 lnet_resend_pending_msgs(void)
3437 cfs_cpt_for_each(i, lnet_cpt_table()) {
3439 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3444 /* called with cpt and ni_lock held */
3446 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3448 struct lnet_handle_md recovery_mdh;
3450 LNetInvalidateMDHandle(&recovery_mdh);
3452 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3454 recovery_mdh = ni->ni_ping_mdh;
3455 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3458 lnet_net_unlock(cpt);
3459 if (!LNetMDHandleIsInvalid(recovery_mdh))
3460 LNetMDUnlink(recovery_mdh);
3466 lnet_recover_local_nis(void)
3468 struct lnet_mt_event_info *ev_info;
3469 LIST_HEAD(processed_list);
3470 LIST_HEAD(local_queue);
3471 struct lnet_handle_md mdh;
3472 struct lnet_ni *tmp;
3474 struct lnet_nid nid;
3480 * splice the recovery queue on a local queue. We will iterate
3481 * through the local queue and update it as needed. Once we're
3482 * done with the traversal, we'll splice the local queue back on
3483 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3484 * will be traversed in the next iteration.
3487 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3491 now = ktime_get_seconds();
3493 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3495 * if an NI is being deleted or it is now healthy, there
3496 * is no need to keep it around in the recovery queue.
3497 * The monitor thread is the only thread responsible for
3498 * removing the NI from the recovery queue.
3499 * Multiple threads can be adding NIs to the recovery
3502 healthv = atomic_read(&ni->ni_healthv);
3506 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3507 healthv == LNET_MAX_HEALTH_VALUE) {
3508 list_del_init(&ni->ni_recovery);
3509 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3511 lnet_ni_decref_locked(ni, 0);
3517 * if the local NI failed recovery we must unlink the md.
3518 * But we want to keep the local_ni on the recovery queue
3519 * so we can continue the attempts to recover it.
3521 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3522 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3523 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3529 if (now < ni->ni_next_ping) {
3536 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3537 libcfs_nidstr(&ni->ni_nid));
3540 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3541 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3544 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3546 CERROR("out of memory. Can't recover %s\n",
3547 libcfs_nidstr(&ni->ni_nid));
3549 ni->ni_recovery_state &=
3550 ~LNET_NI_RECOVERY_PENDING;
3555 mdh = ni->ni_ping_mdh;
3557 * Invalidate the ni mdh in case it's deleted.
3558 * We'll unlink the mdh in this case below.
3560 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3564 * remove the NI from the local queue and drop the
3565 * reference count to it while we're recovering
3566 * it. The reason for that, is that the NI could
3567 * be deleted, and the way the code is structured
3568 * is if we don't drop the NI, then the deletion
3569 * code will enter a loop waiting for the
3570 * reference count to be removed while holding the
3571 * ln_mutex_lock(). When we look up the peer to
3572 * send to in lnet_select_pathway() we will try to
3573 * lock the ln_mutex_lock() as well, leading to
3574 * a deadlock. By dropping the refcount and
3575 * removing it from the list, we allow for the NI
3576 * to be removed, then we use the cached NID to
3577 * look it up again. If it's gone, then we just
3578 * continue examining the rest of the queue.
3581 list_del_init(&ni->ni_recovery);
3582 lnet_ni_decref_locked(ni, 0);
3585 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3586 ev_info->mt_nid = nid;
3587 rc = lnet_send_ping(&nid, &mdh, LNET_INTERFACES_MIN,
3588 ev_info, the_lnet.ln_mt_handler,
3590 /* lookup the nid again */
3592 ni = lnet_nid_to_ni_locked(&nid, 0);
3595 * the NI has been deleted when we dropped
3602 ni->ni_ping_count++;
3604 ni->ni_ping_mdh = mdh;
3605 lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3610 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3619 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3620 * reexamined in the next iteration.
3622 list_splice_init(&processed_list, &local_queue);
3624 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3629 lnet_resendqs_create(void)
3631 struct list_head **resendqs;
3632 resendqs = lnet_create_array_of_queues();
3637 lnet_net_lock(LNET_LOCK_EX);
3638 the_lnet.ln_mt_resendqs = resendqs;
3639 lnet_net_unlock(LNET_LOCK_EX);
3645 lnet_clean_local_ni_recoveryq(void)
3649 /* This is only called when the monitor thread has stopped */
3652 while ((ni = list_first_entry_or_null(&the_lnet.ln_mt_localNIRecovq,
3654 ni_recovery)) != NULL) {
3655 list_del_init(&ni->ni_recovery);
3657 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3659 lnet_ni_decref_locked(ni, 0);
3666 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3669 struct lnet_handle_md recovery_mdh;
3671 LNetInvalidateMDHandle(&recovery_mdh);
3673 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3674 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3675 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3677 spin_unlock(&lpni->lpni_lock);
3678 lnet_net_unlock(cpt);
3679 if (!LNetMDHandleIsInvalid(recovery_mdh))
3680 LNetMDUnlink(recovery_mdh);
3682 spin_lock(&lpni->lpni_lock);
3686 lnet_clean_peer_ni_recoveryq(void)
3688 struct lnet_peer_ni *lpni, *tmp;
3690 lnet_net_lock(LNET_LOCK_EX);
3692 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3694 list_del_init(&lpni->lpni_recovery);
3695 spin_lock(&lpni->lpni_lock);
3696 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3697 spin_unlock(&lpni->lpni_lock);
3698 lnet_peer_ni_decref_locked(lpni);
3701 lnet_net_unlock(LNET_LOCK_EX);
3705 lnet_clean_resendqs(void)
3707 struct lnet_msg *msg, *tmp;
3711 cfs_cpt_for_each(i, lnet_cpt_table()) {
3713 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3715 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3716 list_del_init(&msg->msg_list);
3717 msg->msg_no_resend = true;
3718 lnet_finalize(msg, -ESHUTDOWN);
3722 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3726 lnet_recover_peer_nis(void)
3728 struct lnet_mt_event_info *ev_info;
3729 LIST_HEAD(processed_list);
3730 LIST_HEAD(local_queue);
3731 struct lnet_handle_md mdh;
3732 struct lnet_peer_ni *lpni;
3733 struct lnet_peer_ni *tmp;
3734 struct lnet_nid nid;
3740 * Always use cpt 0 for locking across all interactions with
3741 * ln_mt_peerNIRecovq
3744 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3748 now = ktime_get_seconds();
3750 list_for_each_entry_safe(lpni, tmp, &local_queue,
3753 * The same protection strategy is used here as is in the
3754 * local recovery case.
3757 healthv = atomic_read(&lpni->lpni_healthv);
3758 spin_lock(&lpni->lpni_lock);
3759 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3760 healthv == LNET_MAX_HEALTH_VALUE) {
3761 list_del_init(&lpni->lpni_recovery);
3762 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3763 spin_unlock(&lpni->lpni_lock);
3764 lnet_peer_ni_decref_locked(lpni);
3770 * If the peer NI has failed recovery we must unlink the
3771 * md. But we want to keep the peer ni on the recovery
3772 * queue so we can try to continue recovering it
3774 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3775 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3776 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3779 spin_unlock(&lpni->lpni_lock);
3781 if (now < lpni->lpni_next_ping) {
3789 * NOTE: we're racing with peer deletion from user space.
3790 * It's possible that a peer is deleted after we check its
3791 * state. In this case the recovery can create a new peer
3793 spin_lock(&lpni->lpni_lock);
3794 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3795 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3796 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3797 spin_unlock(&lpni->lpni_lock);
3799 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3801 CERROR("out of memory. Can't recover %s\n",
3802 libcfs_nidstr(&lpni->lpni_nid));
3803 spin_lock(&lpni->lpni_lock);
3804 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3805 spin_unlock(&lpni->lpni_lock);
3809 /* look at the comments in lnet_recover_local_nis() */
3810 mdh = lpni->lpni_recovery_ping_mdh;
3811 nid = lpni->lpni_nid;
3812 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3814 list_del_init(&lpni->lpni_recovery);
3815 lnet_peer_ni_decref_locked(lpni);
3818 ev_info->mt_type = MT_TYPE_PEER_NI;
3819 ev_info->mt_nid = nid;
3820 rc = lnet_send_ping(&nid, &mdh, LNET_INTERFACES_MIN,
3821 ev_info, the_lnet.ln_mt_handler,
3825 * lnet_peer_ni_find_locked() grabs a refcount for
3826 * us. No need to take it explicitly.
3828 lpni = lnet_peer_ni_find_locked(&nid);
3835 lpni->lpni_ping_count++;
3837 lpni->lpni_recovery_ping_mdh = mdh;
3839 lnet_peer_ni_add_to_recoveryq_locked(lpni,
3843 spin_lock(&lpni->lpni_lock);
3844 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3845 spin_unlock(&lpni->lpni_lock);
3848 /* Drop the ref taken by lnet_peer_ni_find_locked() */
3849 lnet_peer_ni_decref_locked(lpni);
3852 spin_unlock(&lpni->lpni_lock);
3855 list_splice_init(&processed_list, &local_queue);
3857 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3862 lnet_monitor_thread(void *arg)
3864 time64_t rsp_timeout = 0;
3867 wait_for_completion(&the_lnet.ln_started);
3869 * The monitor thread takes care of the following:
3870 * 1. Checks the aliveness of routers
3871 * 2. Checks if there are messages on the resend queue to resend
3873 * 3. Check if there are any NIs on the local recovery queue and
3875 * 4. Checks if there are any NIs on the remote recovery queue
3878 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3879 now = ktime_get_real_seconds();
3881 if (lnet_router_checker_active())
3882 lnet_check_routers();
3884 lnet_resend_pending_msgs();
3886 if (now >= rsp_timeout) {
3887 lnet_finalize_expired_responses();
3888 rsp_timeout = now + (lnet_transaction_timeout / 2);
3891 lnet_recover_local_nis();
3892 lnet_recover_peer_nis();
3895 * TODO do we need to check if we should sleep without
3896 * timeout? Technically, an active system will always
3897 * have messages in flight so this check will always
3898 * evaluate to false. And on an idle system do we care
3899 * if we wake up every 1 second? Although, we've seen
3900 * cases where we get a complaint that an idle thread
3901 * is waking up unnecessarily.
3903 wait_for_completion_interruptible_timeout(
3904 &the_lnet.ln_mt_wait_complete,
3905 cfs_time_seconds(1));
3906 /* Must re-init the completion before testing anything,
3907 * including ln_mt_state.
3909 reinit_completion(&the_lnet.ln_mt_wait_complete);
3913 lnet_net_lock(LNET_LOCK_EX);
3914 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3915 lnet_net_unlock(LNET_LOCK_EX);
3917 /* signal that the monitor thread is exiting */
3918 up(&the_lnet.ln_mt_signal);
3926 * Returns == 0 if success
3927 * Returns > 0 if LNetMDBind or prior fails
3928 * Returns < 0 if LNetGet fails
3931 lnet_send_ping(struct lnet_nid *dest_nid,
3932 struct lnet_handle_md *mdh, int nnis,
3933 void *user_data, lnet_handler_t handler, bool recovery)
3935 struct lnet_md md = { NULL };
3936 struct lnet_processid id;
3937 struct lnet_ping_buffer *pbuf;
3940 if (LNET_NID_IS_ANY(dest_nid)) {
3945 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3951 /* initialize md content */
3952 md.start = &pbuf->pb_info;
3953 md.length = LNET_PING_INFO_SIZE(nnis);
3954 md.threshold = 2; /* GET/REPLY */
3956 md.options = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3957 md.user_ptr = user_data;
3958 md.handler = handler;
3960 rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3962 lnet_ping_buffer_decref(pbuf);
3963 CERROR("Can't bind MD: %d\n", rc);
3964 rc = -rc; /* change the rc to positive */
3967 id.pid = LNET_PID_LUSTRE;
3970 rc = LNetGet(NULL, *mdh, &id,
3971 LNET_RESERVED_PORTAL,
3972 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3975 goto fail_unlink_md;
3981 LNetInvalidateMDHandle(mdh);
3987 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3988 int status, bool send, bool unlink_event)
3990 struct lnet_nid *nid = &ev_info->mt_nid;
3992 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3996 ni = lnet_nid_to_ni_locked(nid, 0);
4002 if (!send || (send && status != 0))
4003 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
4005 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
4010 CERROR("local NI (%s) recovery failed with %d\n",
4011 libcfs_nidstr(nid), status);
4015 * need to increment healthv for the ni here, because in
4016 * the lnet_finalize() path we don't have access to this
4017 * NI. And in order to get access to it, we'll need to
4018 * carry forward too much information.
4019 * In the peer case, it'll naturally be incremented
4022 lnet_inc_healthv(&ni->ni_healthv,
4023 lnet_health_sensitivity);
4025 struct lnet_peer_ni *lpni;
4028 cpt = lnet_net_lock_current();
4029 lpni = lnet_peer_ni_find_locked(nid);
4031 lnet_net_unlock(cpt);
4034 spin_lock(&lpni->lpni_lock);
4035 if (!send || (send && status != 0))
4036 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
4038 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
4039 spin_unlock(&lpni->lpni_lock);
4040 lnet_peer_ni_decref_locked(lpni);
4041 lnet_net_unlock(cpt);
4044 CERROR("peer NI (%s) recovery failed with %d\n",
4045 libcfs_nidstr(nid), status);
4050 lnet_mt_event_handler(struct lnet_event *event)
4052 struct lnet_mt_event_info *ev_info = event->md_user_ptr;
4053 struct lnet_ping_buffer *pbuf;
4055 /* TODO: remove assert */
4056 LASSERT(event->type == LNET_EVENT_REPLY ||
4057 event->type == LNET_EVENT_SEND ||
4058 event->type == LNET_EVENT_UNLINK);
4060 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
4063 switch (event->type) {
4064 case LNET_EVENT_UNLINK:
4065 CDEBUG(D_NET, "%s recovery ping unlinked\n",
4066 libcfs_nidstr(&ev_info->mt_nid));
4068 case LNET_EVENT_REPLY:
4069 lnet_handle_recovery_reply(ev_info, event->status, false,
4070 event->type == LNET_EVENT_UNLINK);
4072 case LNET_EVENT_SEND:
4073 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4074 libcfs_nidstr(&ev_info->mt_nid),
4075 (event->status) ? "unsuccessfully" :
4076 "successfully", event->status);
4077 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4080 CERROR("Unexpected event: %d\n", event->type);
4083 if (event->unlinked) {
4084 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4085 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4086 lnet_ping_buffer_decref(pbuf);
4091 lnet_rsp_tracker_create(void)
4093 struct list_head **rstqs;
4094 rstqs = lnet_create_array_of_queues();
4099 the_lnet.ln_mt_rstq = rstqs;
4105 lnet_rsp_tracker_clean(void)
4107 lnet_finalize_expired_responses();
4109 cfs_percpt_free(the_lnet.ln_mt_rstq);
4110 the_lnet.ln_mt_rstq = NULL;
4113 int lnet_monitor_thr_start(void)
4116 struct task_struct *task;
4118 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4121 rc = lnet_resendqs_create();
4125 rc = lnet_rsp_tracker_create();
4129 sema_init(&the_lnet.ln_mt_signal, 0);
4131 lnet_net_lock(LNET_LOCK_EX);
4132 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4133 lnet_net_unlock(LNET_LOCK_EX);
4134 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4137 CERROR("Can't start monitor thread: %d\n", rc);
4144 lnet_net_lock(LNET_LOCK_EX);
4145 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4146 lnet_net_unlock(LNET_LOCK_EX);
4147 /* block until event callback signals exit */
4148 down(&the_lnet.ln_mt_signal);
4150 lnet_net_lock(LNET_LOCK_EX);
4151 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4152 lnet_net_unlock(LNET_LOCK_EX);
4153 lnet_rsp_tracker_clean();
4154 lnet_clean_local_ni_recoveryq();
4155 lnet_clean_peer_ni_recoveryq();
4156 lnet_clean_resendqs();
4157 the_lnet.ln_mt_handler = NULL;
4160 lnet_rsp_tracker_clean();
4161 lnet_clean_local_ni_recoveryq();
4162 lnet_clean_peer_ni_recoveryq();
4163 lnet_clean_resendqs();
4167 void lnet_monitor_thr_stop(void)
4169 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4172 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4173 lnet_net_lock(LNET_LOCK_EX);
4174 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4175 lnet_net_unlock(LNET_LOCK_EX);
4177 /* tell the monitor thread that we're shutting down */
4178 complete(&the_lnet.ln_mt_wait_complete);
4180 /* block until monitor thread signals that it's done */
4181 mutex_unlock(&the_lnet.ln_api_mutex);
4182 down(&the_lnet.ln_mt_signal);
4183 mutex_lock(&the_lnet.ln_api_mutex);
4184 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4186 /* perform cleanup tasks */
4187 lnet_rsp_tracker_clean();
4188 lnet_clean_local_ni_recoveryq();
4189 lnet_clean_peer_ni_recoveryq();
4190 lnet_clean_resendqs();
4194 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4198 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4199 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4200 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4201 lnet_net_unlock(cpt);
4203 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4207 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4209 struct lnet_hdr *hdr = &msg->msg_hdr;
4211 if (msg->msg_wanted != 0)
4212 lnet_setpayloadbuffer(msg);
4214 lnet_build_msg_event(msg, LNET_EVENT_PUT);
4216 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
4217 * it back into the ACK during lnet_finalize() */
4218 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4219 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4221 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4222 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4226 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4228 struct lnet_hdr *hdr = &msg->msg_hdr;
4229 struct lnet_match_info info;
4233 /* Convert put fields to host byte order */
4234 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4235 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
4236 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
4238 /* Primary peer NID. */
4239 info.mi_id.nid = msg->msg_initiator;
4240 info.mi_id.pid = hdr->src_pid;
4241 info.mi_opc = LNET_MD_OP_PUT;
4242 info.mi_portal = hdr->msg.put.ptl_index;
4243 info.mi_rlength = hdr->payload_length;
4244 info.mi_roffset = hdr->msg.put.offset;
4245 info.mi_mbits = hdr->msg.put.match_bits;
4246 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4248 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4249 ready_delay = msg->msg_rx_ready_delay;
4252 rc = lnet_ptl_match_md(&info, msg);
4257 case LNET_MATCHMD_OK:
4258 lnet_recv_put(ni, msg);
4261 case LNET_MATCHMD_NONE:
4263 /* no eager_recv or has already called it, should
4264 * have been attached on delayed list */
4267 rc = lnet_ni_eager_recv(ni, msg);
4274 case LNET_MATCHMD_DROP:
4275 CNETERR("Dropping PUT from %s portal %d match %llu"
4276 " offset %d length %d: %d\n",
4277 libcfs_idstr(&info.mi_id), info.mi_portal,
4278 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4280 return -ENOENT; /* -ve: OK but no match */
4285 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4287 struct lnet_match_info info;
4288 struct lnet_hdr *hdr = &msg->msg_hdr;
4289 struct lnet_processid source_id;
4290 struct lnet_handle_wire reply_wmd;
4293 /* Convert get fields to host byte order */
4294 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
4295 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
4296 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
4297 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
4299 source_id.nid = hdr->src_nid;
4300 source_id.pid = hdr->src_pid;
4301 /* Primary peer NID */
4302 info.mi_id.nid = msg->msg_initiator;
4303 info.mi_id.pid = hdr->src_pid;
4304 info.mi_opc = LNET_MD_OP_GET;
4305 info.mi_portal = hdr->msg.get.ptl_index;
4306 info.mi_rlength = hdr->msg.get.sink_length;
4307 info.mi_roffset = hdr->msg.get.src_offset;
4308 info.mi_mbits = hdr->msg.get.match_bits;
4309 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4311 rc = lnet_ptl_match_md(&info, msg);
4312 if (rc == LNET_MATCHMD_DROP) {
4313 CNETERR("Dropping GET from %s portal %d match %llu"
4314 " offset %d length %d\n",
4315 libcfs_idstr(&info.mi_id), info.mi_portal,
4316 info.mi_mbits, info.mi_roffset, info.mi_rlength);
4317 return -ENOENT; /* -ve: OK but no match */
4320 LASSERT(rc == LNET_MATCHMD_OK);
4322 lnet_build_msg_event(msg, LNET_EVENT_GET);
4324 reply_wmd = hdr->msg.get.return_wmd;
4326 lnet_prep_send(msg, LNET_MSG_REPLY, &source_id,
4327 msg->msg_offset, msg->msg_wanted);
4329 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4332 /* The LND completes the REPLY from her recv procedure */
4333 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4334 msg->msg_offset, msg->msg_len, msg->msg_len);
4338 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4339 msg->msg_receiving = 0;
4341 rc = lnet_send(&ni->ni_nid, msg, &msg->msg_from);
4343 /* didn't get as far as lnet_ni_send() */
4344 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4345 libcfs_nidstr(&ni->ni_nid),
4346 libcfs_idstr(&info.mi_id), rc);
4348 lnet_finalize(msg, rc);
4355 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4357 void *private = msg->msg_private;
4358 struct lnet_hdr *hdr = &msg->msg_hdr;
4359 struct lnet_processid src = {};
4360 struct lnet_libmd *md;
4361 unsigned int rlength;
4362 unsigned int mlength;
4365 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4368 src.nid = hdr->src_nid;
4369 src.pid = hdr->src_pid;
4371 /* NB handles only looked up by creator (no flips) */
4372 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4373 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4374 CNETERR("%s: Dropping REPLY from %s for %s "
4376 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4377 (md == NULL) ? "invalid" : "inactive",
4378 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4379 hdr->msg.reply.dst_wmd.wh_object_cookie);
4380 if (md != NULL && md->md_me != NULL)
4381 CERROR("REPLY MD also attached to portal %d\n",
4382 md->md_me->me_portal);
4384 lnet_res_unlock(cpt);
4385 return -ENOENT; /* -ve: OK but no match */
4388 LASSERT(md->md_offset == 0);
4390 rlength = hdr->payload_length;
4391 mlength = min(rlength, md->md_length);
4393 if (mlength < rlength &&
4394 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4395 CNETERR("%s: Dropping REPLY from %s length %d "
4396 "for MD %#llx would overflow (%d)\n",
4397 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4398 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4400 lnet_res_unlock(cpt);
4401 return -ENOENT; /* -ve: OK but no match */
4404 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4405 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4406 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4408 lnet_msg_attach_md(msg, md, 0, mlength);
4411 lnet_setpayloadbuffer(msg);
4413 lnet_res_unlock(cpt);
4415 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4417 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4422 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4424 struct lnet_hdr *hdr = &msg->msg_hdr;
4425 struct lnet_processid src = {};
4426 struct lnet_libmd *md;
4429 src.nid = hdr->src_nid;
4430 src.pid = hdr->src_pid;
4432 /* Convert ack fields to host byte order */
4433 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4434 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4436 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4439 /* NB handles only looked up by creator (no flips) */
4440 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4441 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4442 /* Don't moan; this is expected */
4444 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4445 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4446 (md == NULL) ? "invalid" : "inactive",
4447 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4448 hdr->msg.ack.dst_wmd.wh_object_cookie);
4449 if (md != NULL && md->md_me != NULL)
4450 CERROR("Source MD also attached to portal %d\n",
4451 md->md_me->me_portal);
4453 lnet_res_unlock(cpt);
4454 return -ENOENT; /* -ve! */
4457 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4458 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4459 hdr->msg.ack.dst_wmd.wh_object_cookie);
4461 lnet_msg_attach_md(msg, md, 0, 0);
4463 lnet_res_unlock(cpt);
4465 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4467 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4472 * \retval LNET_CREDIT_OK If \a msg is forwarded
4473 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4474 * \retval -ve error code
4477 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4481 if (!the_lnet.ln_routing)
4484 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4485 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4486 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4487 msg->msg_rx_ready_delay = 1;
4489 lnet_net_unlock(msg->msg_rx_cpt);
4490 rc = lnet_ni_eager_recv(ni, msg);
4491 lnet_net_lock(msg->msg_rx_cpt);
4496 rc = lnet_post_routed_recv_locked(msg, 0);
4501 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4505 switch (msg->msg_type) {
4507 rc = lnet_parse_ack(ni, msg);
4510 rc = lnet_parse_put(ni, msg);
4513 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4515 case LNET_MSG_REPLY:
4516 rc = lnet_parse_reply(ni, msg);
4518 default: /* prevent an unused label if !kernel */
4523 LASSERT(rc == 0 || rc == -ENOENT);
4528 lnet_msgtyp2str (int type)
4537 case LNET_MSG_REPLY:
4539 case LNET_MSG_HELLO:
4542 return ("<UNKNOWN>");
4545 EXPORT_SYMBOL(lnet_msgtyp2str);
4548 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr,
4549 struct lnet_nid *from_nid, void *private, int rdma_req)
4551 struct lnet_peer_ni *lpni;
4552 struct lnet_msg *msg;
4553 __u32 payload_length;
4554 lnet_pid_t dest_pid;
4555 struct lnet_nid dest_nid;
4556 struct lnet_nid src_nid;
4562 time64_t now = ktime_get_seconds();
4564 LASSERT (!in_interrupt ());
4567 src_nid = hdr->src_nid;
4568 dest_nid = hdr->dest_nid;
4569 dest_pid = hdr->dest_pid;
4570 payload_length = hdr->payload_length;
4572 for_me = nid_same(&ni->ni_nid, &dest_nid);
4573 cpt = lnet_nid2cpt(from_nid, ni);
4575 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4576 libcfs_nidstr(&dest_nid),
4577 libcfs_nidstr(&ni->ni_nid),
4578 libcfs_nidstr(&src_nid),
4579 lnet_msgtyp2str(type),
4580 (for_me) ? "for me" : "routed");
4585 if (payload_length > 0) {
4586 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4587 libcfs_nidstr(from_nid),
4588 libcfs_nidstr(&src_nid),
4589 lnet_msgtyp2str(type), payload_length);
4595 case LNET_MSG_REPLY:
4596 if (payload_length >
4597 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4598 CERROR("%s, src %s: bad %s payload %d "
4599 "(%d max expected)\n",
4600 libcfs_nidstr(from_nid),
4601 libcfs_nidstr(&src_nid),
4602 lnet_msgtyp2str(type),
4604 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4610 CERROR("%s, src %s: Bad message type 0x%x\n",
4611 libcfs_nidstr(from_nid),
4612 libcfs_nidstr(&src_nid), type);
4616 /* Only update net_last_alive for incoming GETs on the reserved portal
4617 * (i.e. incoming lnet/discovery pings).
4618 * This avoids situations where the router's own traffic results in NI
4621 if (the_lnet.ln_routing && type == LNET_MSG_GET &&
4622 hdr->msg.get.ptl_index == LNET_RESERVED_PORTAL &&
4623 !lnet_islocalnid(&src_nid) &&
4624 ni->ni_net->net_last_alive != now) {
4626 spin_lock(&ni->ni_net->net_lock);
4627 ni->ni_net->net_last_alive = now;
4628 spin_unlock(&ni->ni_net->net_lock);
4629 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4634 lnet_push_update_to_peers(1);
4636 /* Regard a bad destination NID as a protocol error. Senders should
4637 * know what they're doing; if they don't they're misconfigured, buggy
4638 * or malicious so we chop them off at the knees :) */
4641 if (LNET_NID_NET(&dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4642 /* should have gone direct */
4643 CERROR("%s, src %s: Bad dest nid %s "
4644 "(should have been sent direct)\n",
4645 libcfs_nidstr(from_nid),
4646 libcfs_nidstr(&src_nid),
4647 libcfs_nidstr(&dest_nid));
4651 if (lnet_islocalnid(&dest_nid)) {
4652 /* dest is another local NI; sender should have used
4653 * this node's NID on its own network */
4654 CERROR("%s, src %s: Bad dest nid %s "
4655 "(it's my nid but on a different network)\n",
4656 libcfs_nidstr(from_nid),
4657 libcfs_nidstr(&src_nid),
4658 libcfs_nidstr(&dest_nid));
4662 if (rdma_req && type == LNET_MSG_GET) {
4663 CERROR("%s, src %s: Bad optimized GET for %s "
4664 "(final destination must be me)\n",
4665 libcfs_nidstr(from_nid),
4666 libcfs_nidstr(&src_nid),
4667 libcfs_nidstr(&dest_nid));
4671 if (!the_lnet.ln_routing) {
4672 CERROR("%s, src %s: Dropping message for %s "
4673 "(routing not enabled)\n",
4674 libcfs_nidstr(from_nid),
4675 libcfs_nidstr(&src_nid),
4676 libcfs_nidstr(&dest_nid));
4681 /* Message looks OK; we're not going to return an error, so we MUST
4682 * call back lnd_recv() come what may... */
4684 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4685 fail_peer(&src_nid, 0)) { /* shall we now? */
4686 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4687 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4688 lnet_msgtyp2str(type));
4692 if (!list_empty(&the_lnet.ln_drop_rules) &&
4693 lnet_drop_rule_match(hdr, &ni->ni_nid, NULL)) {
4695 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4696 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4697 libcfs_nidstr(&dest_nid), lnet_msgtyp2str(type));
4701 msg = lnet_msg_alloc();
4703 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4704 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4705 lnet_msgtyp2str(type));
4709 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4710 * pointers NULL etc */
4712 msg->msg_type = type;
4713 msg->msg_private = private;
4714 msg->msg_receiving = 1;
4715 msg->msg_rdma_get = rdma_req;
4716 msg->msg_len = msg->msg_wanted = payload_length;
4717 msg->msg_offset = 0;
4718 msg->msg_hdr = *hdr;
4719 /* for building message event */
4720 msg->msg_from = *from_nid;
4722 msg->msg_target.pid = dest_pid;
4723 msg->msg_target.nid = dest_nid;
4724 msg->msg_routing = 1;
4728 lpni = lnet_peerni_by_nid_locked(from_nid, &ni->ni_nid, cpt);
4730 lnet_net_unlock(cpt);
4732 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4733 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4734 lnet_msgtyp2str(type), rc);
4736 if (rc == -ESHUTDOWN)
4737 /* We are shutting down. Don't do anything more */
4742 /* If this message was forwarded to us from a router then we may need
4743 * to update router aliveness or check for an asymmetrical route
4746 if (((lnet_drop_asym_route && for_me) ||
4747 !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4748 LNET_NID_NET(&src_nid) != LNET_NID_NET(from_nid)) {
4749 __u32 src_net_id = LNET_NID_NET(&src_nid);
4750 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4751 struct lnet_route *route;
4754 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4755 if (route->lr_net == src_net_id) {
4757 /* If we're transitioning the gateway from
4758 * dead -> alive, and discovery is disabled
4759 * locally or on the gateway, then we need to
4760 * update the cached route aliveness for each
4761 * route to the src_nid's net.
4763 * Otherwise, we're only checking for
4764 * symmetrical route, and we can break the
4767 if (!gw->lp_alive &&
4768 lnet_is_discovery_disabled(gw))
4769 lnet_set_route_aliveness(route, true);
4774 if (lnet_drop_asym_route && for_me && !found) {
4775 /* Drop ref taken by lnet_nid2peerni_locked() */
4776 lnet_peer_ni_decref_locked(lpni);
4777 lnet_net_unlock(cpt);
4778 /* we would not use from_nid to route a message to
4780 * => asymmetric routing detected but forbidden
4782 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4783 libcfs_nidstr(from_nid),
4784 libcfs_nidstr(&src_nid), lnet_msgtyp2str(type));
4788 if (!gw->lp_alive) {
4789 struct lnet_peer_net *lpn;
4790 struct lnet_peer_ni *lpni2;
4792 gw->lp_alive = true;
4793 /* Mark all remote NIs on src_nid's net UP */
4794 lpn = lnet_peer_get_net_locked(gw, src_net_id);
4796 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4798 lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4802 lpni->lpni_last_alive = now;
4804 msg->msg_rxpeer = lpni;
4806 lnet_ni_addref_locked(ni, cpt);
4807 /* Multi-Rail: Primary NID of source. */
4808 lnet_peer_primary_nid_locked(&src_nid, &msg->msg_initiator);
4811 * mark the status of this lpni as UP since we received a message
4812 * from it. The ping response reports back the ns_status which is
4813 * marked on the remote as up or down and we cache it here.
4815 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4817 lnet_msg_commit(msg, cpt);
4819 /* message delay simulation */
4820 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4821 lnet_delay_rule_match_locked(hdr, msg))) {
4822 lnet_net_unlock(cpt);
4827 rc = lnet_parse_forward_locked(ni, msg);
4828 lnet_net_unlock(cpt);
4833 if (rc == LNET_CREDIT_OK) {
4834 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4835 0, payload_length, payload_length);
4840 lnet_net_unlock(cpt);
4842 rc = lnet_parse_local(ni, msg);
4848 LASSERT(msg->msg_md == NULL);
4849 lnet_finalize(msg, rc);
4852 lnet_drop_message(ni, cpt, private, payload_length, type);
4855 EXPORT_SYMBOL(lnet_parse);
4858 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4860 struct lnet_msg *msg;
4862 while ((msg = list_first_entry_or_null(head, struct lnet_msg,
4863 msg_list)) != NULL) {
4864 struct lnet_processid id = {};
4866 list_del(&msg->msg_list);
4868 id.nid = msg->msg_hdr.src_nid;
4869 id.pid = msg->msg_hdr.src_pid;
4871 LASSERT(msg->msg_md == NULL);
4872 LASSERT(msg->msg_rx_delayed);
4873 LASSERT(msg->msg_rxpeer != NULL);
4874 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4876 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4877 " offset %d length %d: %s\n",
4879 msg->msg_hdr.msg.put.ptl_index,
4880 msg->msg_hdr.msg.put.match_bits,
4881 msg->msg_hdr.msg.put.offset,
4882 msg->msg_hdr.payload_length, reason);
4884 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4885 * called lnet_drop_message(), so I just hang onto msg as well
4886 * until that's done */
4888 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4889 msg->msg_private, msg->msg_len,
4892 msg->msg_no_resend = true;
4894 * NB: message will not generate event because w/o attached MD,
4895 * but we still should give error code so lnet_msg_decommit()
4896 * can skip counters operations and other checks.
4898 lnet_finalize(msg, -ENOENT);
4903 lnet_recv_delayed_msg_list(struct list_head *head)
4905 struct lnet_msg *msg;
4907 while ((msg = list_first_entry_or_null(head, struct lnet_msg,
4908 msg_list)) != NULL) {
4909 struct lnet_processid id;
4911 list_del(&msg->msg_list);
4913 /* md won't disappear under me, since each msg
4914 * holds a ref on it */
4916 id.nid = msg->msg_hdr.src_nid;
4917 id.pid = msg->msg_hdr.src_pid;
4919 LASSERT(msg->msg_rx_delayed);
4920 LASSERT(msg->msg_md != NULL);
4921 LASSERT(msg->msg_rxpeer != NULL);
4922 LASSERT(msg->msg_rxni != NULL);
4923 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4925 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4926 "match %llu offset %d length %d.\n",
4927 libcfs_idstr(&id), msg->msg_hdr.msg.put.ptl_index,
4928 msg->msg_hdr.msg.put.match_bits,
4929 msg->msg_hdr.msg.put.offset,
4930 msg->msg_hdr.payload_length);
4932 lnet_recv_put(msg->msg_rxni, msg);
4937 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4938 struct lnet_libmd *md, struct lnet_handle_md mdh)
4941 struct lnet_rsp_tracker *local_rspt;
4944 * MD has a refcount taken by message so it's not going away.
4945 * The MD however can be looked up. We need to secure the access
4946 * to the md_rspt_ptr by taking the res_lock.
4947 * The rspt can be accessed without protection up to when it gets
4948 * added to the list.
4952 local_rspt = md->md_rspt_ptr;
4953 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4954 if (local_rspt != NULL) {
4956 * we already have an rspt attached to the md, so we'll
4957 * update the deadline on that one.
4959 lnet_rspt_free(rspt, cpt);
4962 rspt->rspt_mdh = mdh;
4963 rspt->rspt_cpt = cpt;
4964 /* store the rspt so we can access it when we get the REPLY */
4965 md->md_rspt_ptr = rspt;
4968 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4971 * add to the list of tracked responses. It's added to tail of the
4972 * list in order to expire all the older entries first.
4975 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4976 lnet_net_unlock(cpt);
4977 lnet_res_unlock(cpt);
4981 * Initiate an asynchronous PUT operation.
4983 * There are several events associated with a PUT: completion of the send on
4984 * the initiator node (LNET_EVENT_SEND), and when the send completes
4985 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4986 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4987 * used at the target node to indicate the completion of incoming data
4990 * The local events will be logged in the EQ associated with the MD pointed to
4991 * by \a mdh handle. Using a MD without an associated EQ results in these
4992 * events being discarded. In this case, the caller must have another
4993 * mechanism (e.g., a higher level protocol) for determining when it is safe
4994 * to modify the memory region associated with the MD.
4996 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4997 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4999 * \param self Indicates the NID of a local interface through which to send
5000 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
5001 * \param mdh A handle for the MD that describes the memory to be sent. The MD
5002 * must be "free floating" (See LNetMDBind()).
5003 * \param ack Controls whether an acknowledgment is requested.
5004 * Acknowledgments are only sent when they are requested by the initiating
5005 * process and the target MD enables them.
5006 * \param target A process identifier for the target process.
5007 * \param portal The index in the \a target's portal table.
5008 * \param match_bits The match bits to use for MD selection at the target
5010 * \param offset The offset into the target MD (only used when the target
5011 * MD has the LNET_MD_MANAGE_REMOTE option set).
5012 * \param hdr_data 64 bits of user data that can be included in the message
5013 * header. This data is written to an event queue entry at the target if an
5014 * EQ is present on the matching MD.
5016 * \retval 0 Success, and only in this case events will be generated
5017 * and logged to EQ (if it exists).
5018 * \retval -EIO Simulated failure.
5019 * \retval -ENOMEM Memory allocation failure.
5020 * \retval -ENOENT Invalid MD object.
5022 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
5025 LNetPut(struct lnet_nid *self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
5026 struct lnet_processid *target, unsigned int portal,
5027 __u64 match_bits, unsigned int offset,
5030 struct lnet_msg *msg;
5031 struct lnet_libmd *md;
5034 struct lnet_rsp_tracker *rspt = NULL;
5036 LASSERT(the_lnet.ln_refcount > 0);
5038 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5039 fail_peer(&target->nid, 1)) { /* shall we now? */
5040 CERROR("Dropping PUT to %s: simulated failure\n",
5041 libcfs_idstr(target));
5045 msg = lnet_msg_alloc();
5047 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
5048 libcfs_idstr(target));
5051 msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
5053 cpt = lnet_cpt_of_cookie(mdh.cookie);
5055 if (ack == LNET_ACK_REQ) {
5056 rspt = lnet_rspt_alloc(cpt);
5058 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
5059 libcfs_idstr(target));
5062 INIT_LIST_HEAD(&rspt->rspt_on_list);
5067 md = lnet_handle2md(&mdh);
5068 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5069 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5070 match_bits, portal, libcfs_idstr(target),
5071 md == NULL ? -1 : md->md_threshold);
5072 if (md != NULL && md->md_me != NULL)
5073 CERROR("Source MD also attached to portal %d\n",
5074 md->md_me->me_portal);
5075 lnet_res_unlock(cpt);
5078 lnet_rspt_free(rspt, cpt);
5084 CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5086 lnet_msg_attach_md(msg, md, 0, 0);
5088 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5090 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5091 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5092 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5093 msg->msg_hdr.msg.put.hdr_data = hdr_data;
5095 /* NB handles only looked up by creator (no flips) */
5096 if (ack == LNET_ACK_REQ) {
5097 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5098 the_lnet.ln_interface_cookie;
5099 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5100 md->md_lh.lh_cookie;
5102 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5103 LNET_WIRE_HANDLE_COOKIE_NONE;
5104 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5105 LNET_WIRE_HANDLE_COOKIE_NONE;
5108 lnet_res_unlock(cpt);
5110 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5112 if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5114 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5116 lnet_rspt_free(rspt, cpt);
5118 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5122 rc = lnet_send(self, msg, NULL);
5125 CNETERR("Error sending PUT to %s: %d\n",
5126 libcfs_idstr(target), rc);
5127 msg->msg_no_resend = true;
5128 lnet_finalize(msg, rc);
5131 /* completion will be signalled by an event */
5134 EXPORT_SYMBOL(LNetPut);
5137 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
5138 * returns a msg for the LND to pass to lnet_finalize() when the sink
5139 * data has been received.
5141 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5142 * lnet_finalize() is called on it, so the LND must call this first
5145 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5147 struct lnet_msg *msg = lnet_msg_alloc();
5148 struct lnet_libmd *getmd = getmsg->msg_md;
5149 struct lnet_processid *peer_id = &getmsg->msg_target;
5152 LASSERT(!getmsg->msg_target_is_router);
5153 LASSERT(!getmsg->msg_routing);
5156 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5157 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id));
5161 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5164 LASSERT(getmd->md_refcount > 0);
5166 if (getmd->md_threshold == 0) {
5167 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5168 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id),
5170 lnet_res_unlock(cpt);
5174 LASSERT(getmd->md_offset == 0);
5176 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5177 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id), getmd);
5179 /* setup information for lnet_build_msg_event */
5180 msg->msg_initiator =
5181 getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5182 msg->msg_from = peer_id->nid;
5183 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5184 msg->msg_hdr.src_nid = peer_id->nid;
5185 msg->msg_hdr.payload_length = getmd->md_length;
5186 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5188 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5189 lnet_res_unlock(cpt);
5191 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5194 lnet_msg_commit(msg, cpt);
5195 lnet_net_unlock(cpt);
5197 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5202 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5205 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5206 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5207 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5209 lnet_net_unlock(cpt);
5216 EXPORT_SYMBOL(lnet_create_reply_msg);
5219 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5222 /* Set the REPLY length, now the RDMA that elides the REPLY message has
5223 * completed and I know it. */
5224 LASSERT(reply != NULL);
5225 LASSERT(reply->msg_type == LNET_MSG_GET);
5226 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5228 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
5229 * the end of my buffer, I might as well be dead. */
5230 LASSERT(len <= reply->msg_ev.mlength);
5232 reply->msg_ev.mlength = len;
5234 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5237 * Initiate an asynchronous GET operation.
5239 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5240 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5241 * the target node in the REPLY has been written to local MD.
5243 * On the target node, an LNET_EVENT_GET is logged when the GET request
5244 * arrives and is accepted into a MD.
5246 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5247 * \param mdh A handle for the MD that describes the memory into which the
5248 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5250 * \retval 0 Success, and only in this case events will be generated
5251 * and logged to EQ (if it exists) of the MD.
5252 * \retval -EIO Simulated failure.
5253 * \retval -ENOMEM Memory allocation failure.
5254 * \retval -ENOENT Invalid MD object.
5257 LNetGet(struct lnet_nid *self, struct lnet_handle_md mdh,
5258 struct lnet_processid *target, unsigned int portal,
5259 __u64 match_bits, unsigned int offset, bool recovery)
5261 struct lnet_msg *msg;
5262 struct lnet_libmd *md;
5263 struct lnet_rsp_tracker *rspt;
5267 LASSERT(the_lnet.ln_refcount > 0);
5269 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5270 fail_peer(&target->nid, 1)) /* shall we now? */
5272 CERROR("Dropping GET to %s: simulated failure\n",
5273 libcfs_idstr(target));
5277 msg = lnet_msg_alloc();
5279 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5280 libcfs_idstr(target));
5284 cpt = lnet_cpt_of_cookie(mdh.cookie);
5286 rspt = lnet_rspt_alloc(cpt);
5288 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5289 libcfs_idstr(target));
5292 INIT_LIST_HEAD(&rspt->rspt_on_list);
5294 msg->msg_recovery = recovery;
5298 md = lnet_handle2md(&mdh);
5299 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5300 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5301 match_bits, portal, libcfs_idstr(target),
5302 md == NULL ? -1 : md->md_threshold);
5303 if (md != NULL && md->md_me != NULL)
5304 CERROR("REPLY MD also attached to portal %d\n",
5305 md->md_me->me_portal);
5307 lnet_res_unlock(cpt);
5310 lnet_rspt_free(rspt, cpt);
5314 CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5316 lnet_msg_attach_md(msg, md, 0, 0);
5318 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5320 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5321 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5322 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5323 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5325 /* NB handles only looked up by creator (no flips) */
5326 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5327 the_lnet.ln_interface_cookie;
5328 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5329 md->md_lh.lh_cookie;
5331 lnet_res_unlock(cpt);
5333 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5335 if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5336 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5338 lnet_rspt_free(rspt, cpt);
5340 rc = lnet_send(self, msg, NULL);
5342 CNETERR("Error sending GET to %s: %d\n",
5343 libcfs_idstr(target), rc);
5344 msg->msg_no_resend = true;
5345 lnet_finalize(msg, rc);
5348 /* completion will be signalled by an event */
5351 EXPORT_SYMBOL(LNetGet);
5354 * Calculate distance to node at \a dstnid.
5356 * \param dstnid Target NID.
5357 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5359 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5362 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5363 * local_nid_dist_zero is set, which is the default.
5364 * \retval positives Distance to target NID, i.e. number of hops plus one.
5365 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5368 LNetDist(struct lnet_nid *dstnid, struct lnet_nid *srcnid, __u32 *orderp)
5370 struct lnet_ni *ni = NULL;
5371 struct lnet_remotenet *rnet;
5372 __u32 dstnet = LNET_NID_NET(dstnid);
5376 struct list_head *rn_list;
5377 struct lnet_ni *matched_dstnet = NULL;
5379 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5380 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5381 * keep order 0 free for 0@lo and order 1 free for a local NID
5383 * WARNING: dstnid and srcnid might point to same place.
5384 * Don't set *srcnid until late.
5387 LASSERT(the_lnet.ln_refcount > 0);
5389 cpt = lnet_net_lock_current();
5391 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5392 if (nid_same(&ni->ni_nid, dstnid)) {
5393 if (orderp != NULL) {
5394 if (nid_is_lo0(dstnid))
5401 lnet_net_unlock(cpt);
5403 return local_nid_dist_zero ? 0 : 1;
5406 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5407 matched_dstnet = ni;
5408 /* We matched the destination net, but we may have
5409 * additional local NIs to inspect.
5411 * We record the order as appropriate, but
5412 * they may be overwritten if we match local NI above.
5416 /* Check if ni was originally created in
5417 * current net namespace.
5418 * If not, assign order above 0xffff0000,
5419 * to make this ni not a priority.
5421 if (current->nsproxy &&
5422 !net_eq(ni->ni_net_ns,
5423 current->nsproxy->net_ns))
5424 *orderp = order + 0xffff0000;
5433 if (matched_dstnet) {
5435 *srcnid = matched_dstnet->ni_nid;
5436 lnet_net_unlock(cpt);
5440 rn_list = lnet_net2rnethash(dstnet);
5441 list_for_each_entry(rnet, rn_list, lrn_list) {
5442 if (rnet->lrn_net == dstnet) {
5443 struct lnet_route *route;
5444 struct lnet_route *shortest = NULL;
5445 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5448 LASSERT(!list_empty(&rnet->lrn_routes));
5450 list_for_each_entry(route, &rnet->lrn_routes,
5452 route_hops = route->lr_hops;
5453 if (route_hops == LNET_UNDEFINED_HOPS)
5455 if (shortest == NULL ||
5456 route_hops < shortest_hops) {
5458 shortest_hops = route_hops;
5462 LASSERT(shortest != NULL);
5463 hops = shortest_hops;
5465 struct lnet_net *net;
5466 net = lnet_get_net_locked(shortest->lr_lnet);
5468 ni = lnet_get_next_ni_locked(net, NULL);
5469 *srcnid = ni->ni_nid;
5473 lnet_net_unlock(cpt);
5479 lnet_net_unlock(cpt);
5480 return -EHOSTUNREACH;
5482 EXPORT_SYMBOL(LNetDist);