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 /* assume peer_ni is alive as long as we're within the configured
800 return ktime_get_seconds() >=
801 (lpni->lpni_last_alive +
802 lpni->lpni_net->net_tunables.lct_peer_timeout);
806 * \param msg The message to be sent.
807 * \param do_send True if lnet_ni_send() should be called in this function.
808 * lnet_send() is going to lnet_net_unlock immediately after this, so
809 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
811 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
812 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
813 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
814 * \retval -ECANCELED If the MD of the message has been unlinked.
817 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
819 struct lnet_peer_ni *lp = msg->msg_txpeer;
820 struct lnet_ni *ni = msg->msg_txni;
821 int cpt = msg->msg_tx_cpt;
822 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
824 /* non-lnet_send() callers have checked before */
825 LASSERT(!do_send || msg->msg_tx_delayed);
826 LASSERT(!msg->msg_receiving);
827 LASSERT(msg->msg_tx_committed);
829 /* can't get here if we're sending to the loopback interface */
830 if (the_lnet.ln_loni)
831 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
833 /* NB 'lp' is always the next hop */
834 if (lnet_check_message_drop(ni, lp, msg)) {
835 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
836 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
838 lnet_net_unlock(cpt);
840 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
842 LNET_STATS_TYPE_DROP);
844 lnet_incr_stats(&msg->msg_txni->ni_stats,
846 LNET_STATS_TYPE_DROP);
848 CNETERR("Dropping message for %s: peer not alive\n",
849 libcfs_idstr(&msg->msg_target));
850 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
852 lnet_finalize(msg, -EHOSTUNREACH);
855 return -EHOSTUNREACH;
858 if (msg->msg_md != NULL &&
859 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
860 lnet_net_unlock(cpt);
862 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
863 "called on the MD/ME.\n",
864 libcfs_idstr(&msg->msg_target));
866 msg->msg_no_resend = true;
867 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
868 msg, libcfs_idstr(&msg->msg_target));
869 lnet_finalize(msg, -ECANCELED);
876 if (!msg->msg_peertxcredit) {
877 spin_lock(&lp->lpni_lock);
878 LASSERT((lp->lpni_txcredits < 0) ==
879 !list_empty(&lp->lpni_txq));
881 msg->msg_peertxcredit = 1;
882 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr_nid4);
883 lp->lpni_txcredits--;
885 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
886 lp->lpni_mintxcredits = lp->lpni_txcredits;
888 if (lp->lpni_txcredits < 0) {
889 msg->msg_tx_delayed = 1;
890 list_add_tail(&msg->msg_list, &lp->lpni_txq);
891 spin_unlock(&lp->lpni_lock);
892 return LNET_CREDIT_WAIT;
894 spin_unlock(&lp->lpni_lock);
897 if (!msg->msg_txcredit) {
898 LASSERT((tq->tq_credits < 0) ==
899 !list_empty(&tq->tq_delayed));
901 msg->msg_txcredit = 1;
903 atomic_dec(&ni->ni_tx_credits);
905 if (tq->tq_credits < tq->tq_credits_min)
906 tq->tq_credits_min = tq->tq_credits;
908 if (tq->tq_credits < 0) {
909 msg->msg_tx_delayed = 1;
910 list_add_tail(&msg->msg_list, &tq->tq_delayed);
911 return LNET_CREDIT_WAIT;
915 if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
916 lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
917 msg->msg_tx_delayed = 1;
918 return LNET_CREDIT_WAIT;
921 /* unset the tx_delay flag as we're going to send it now */
922 msg->msg_tx_delayed = 0;
925 lnet_net_unlock(cpt);
926 lnet_ni_send(ni, msg);
929 return LNET_CREDIT_OK;
933 static struct lnet_rtrbufpool *
934 lnet_msg2bufpool(struct lnet_msg *msg)
936 struct lnet_rtrbufpool *rbp;
939 LASSERT(msg->msg_rx_committed);
941 cpt = msg->msg_rx_cpt;
942 rbp = &the_lnet.ln_rtrpools[cpt][0];
944 LASSERT(msg->msg_len <= LNET_MTU);
945 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
947 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
954 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
956 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
957 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
958 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
959 * received or OK to receive */
960 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
961 struct lnet_peer *lp;
962 struct lnet_rtrbufpool *rbp;
963 struct lnet_rtrbuf *rb;
965 LASSERT(msg->msg_kiov == NULL);
966 LASSERT(msg->msg_niov == 0);
967 LASSERT(msg->msg_routing);
968 LASSERT(msg->msg_receiving);
969 LASSERT(!msg->msg_sending);
970 LASSERT(lpni->lpni_peer_net);
971 LASSERT(lpni->lpni_peer_net->lpn_peer);
973 lp = lpni->lpni_peer_net->lpn_peer;
975 /* non-lnet_parse callers only receive delayed messages */
976 LASSERT(!do_recv || msg->msg_rx_delayed);
978 if (!msg->msg_peerrtrcredit) {
979 /* lpni_lock protects the credit manipulation */
980 spin_lock(&lpni->lpni_lock);
982 msg->msg_peerrtrcredit = 1;
983 lpni->lpni_rtrcredits--;
984 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
985 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
987 if (lpni->lpni_rtrcredits < 0) {
988 spin_unlock(&lpni->lpni_lock);
989 /* must have checked eager_recv before here */
990 LASSERT(msg->msg_rx_ready_delay);
991 msg->msg_rx_delayed = 1;
992 /* lp_lock protects the lp_rtrq */
993 spin_lock(&lp->lp_lock);
994 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
995 spin_unlock(&lp->lp_lock);
996 return LNET_CREDIT_WAIT;
998 spin_unlock(&lpni->lpni_lock);
1001 rbp = lnet_msg2bufpool(msg);
1003 if (!msg->msg_rtrcredit) {
1004 msg->msg_rtrcredit = 1;
1006 if (rbp->rbp_credits < rbp->rbp_mincredits)
1007 rbp->rbp_mincredits = rbp->rbp_credits;
1009 if (rbp->rbp_credits < 0) {
1010 /* must have checked eager_recv before here */
1011 LASSERT(msg->msg_rx_ready_delay);
1012 msg->msg_rx_delayed = 1;
1013 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1014 return LNET_CREDIT_WAIT;
1018 LASSERT(!list_empty(&rbp->rbp_bufs));
1019 rb = list_first_entry(&rbp->rbp_bufs, struct lnet_rtrbuf, rb_list);
1020 list_del(&rb->rb_list);
1022 msg->msg_niov = rbp->rbp_npages;
1023 msg->msg_kiov = &rb->rb_kiov[0];
1025 /* unset the msg-rx_delayed flag since we're receiving the message */
1026 msg->msg_rx_delayed = 0;
1029 int cpt = msg->msg_rx_cpt;
1031 lnet_net_unlock(cpt);
1032 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1033 0, msg->msg_len, msg->msg_len);
1036 return LNET_CREDIT_OK;
1040 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1042 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1043 struct lnet_ni *txni = msg->msg_txni;
1044 struct lnet_msg *msg2;
1046 if (msg->msg_txcredit) {
1047 struct lnet_ni *ni = msg->msg_txni;
1048 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1050 /* give back NI txcredits */
1051 msg->msg_txcredit = 0;
1053 LASSERT((tq->tq_credits < 0) ==
1054 !list_empty(&tq->tq_delayed));
1057 atomic_inc(&ni->ni_tx_credits);
1058 if (tq->tq_credits <= 0) {
1059 msg2 = list_first_entry(&tq->tq_delayed,
1060 struct lnet_msg, msg_list);
1061 list_del(&msg2->msg_list);
1063 LASSERT(msg2->msg_txni == ni);
1064 LASSERT(msg2->msg_tx_delayed);
1065 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1067 (void) lnet_post_send_locked(msg2, 1);
1071 if (msg->msg_peertxcredit) {
1072 /* give back peer txcredits */
1073 msg->msg_peertxcredit = 0;
1075 spin_lock(&txpeer->lpni_lock);
1076 LASSERT((txpeer->lpni_txcredits < 0) ==
1077 !list_empty(&txpeer->lpni_txq));
1079 txpeer->lpni_txqnob -= msg->msg_len +
1080 sizeof(struct lnet_hdr_nid4);
1081 LASSERT(txpeer->lpni_txqnob >= 0);
1083 txpeer->lpni_txcredits++;
1084 if (txpeer->lpni_txcredits <= 0) {
1087 msg2 = list_first_entry(&txpeer->lpni_txq,
1088 struct lnet_msg, msg_list);
1089 list_del(&msg2->msg_list);
1090 spin_unlock(&txpeer->lpni_lock);
1092 LASSERT(msg2->msg_txpeer == txpeer);
1093 LASSERT(msg2->msg_tx_delayed);
1095 msg2_cpt = msg2->msg_tx_cpt;
1098 * The msg_cpt can be different from the msg2_cpt
1099 * so we need to make sure we lock the correct cpt
1101 * Once we call lnet_post_send_locked() it is no
1102 * longer safe to access msg2, since it could've
1103 * been freed by lnet_finalize(), but we still
1104 * need to relock the correct cpt, so we cache the
1105 * msg2_cpt for the purpose of the check that
1106 * follows the call to lnet_pose_send_locked().
1108 if (msg2_cpt != msg->msg_tx_cpt) {
1109 lnet_net_unlock(msg->msg_tx_cpt);
1110 lnet_net_lock(msg2_cpt);
1112 (void) lnet_post_send_locked(msg2, 1);
1113 if (msg2_cpt != msg->msg_tx_cpt) {
1114 lnet_net_unlock(msg2_cpt);
1115 lnet_net_lock(msg->msg_tx_cpt);
1118 spin_unlock(&txpeer->lpni_lock);
1123 msg->msg_txni = NULL;
1124 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1127 if (txpeer != NULL) {
1128 msg->msg_txpeer = NULL;
1129 lnet_peer_ni_decref_locked(txpeer);
1134 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1136 struct lnet_msg *msg;
1138 if (list_empty(&rbp->rbp_msgs))
1140 msg = list_first_entry(&rbp->rbp_msgs,
1141 struct lnet_msg, msg_list);
1142 list_del(&msg->msg_list);
1144 (void)lnet_post_routed_recv_locked(msg, 1);
1148 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1150 struct lnet_msg *msg;
1151 struct lnet_msg *tmp;
1153 lnet_net_unlock(cpt);
1155 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1156 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1157 0, 0, 0, msg->msg_hdr.payload_length);
1158 list_del_init(&msg->msg_list);
1159 msg->msg_no_resend = true;
1160 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1161 lnet_finalize(msg, -ECANCELED);
1168 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1170 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1171 struct lnet_peer *lp;
1172 struct lnet_ni *rxni = msg->msg_rxni;
1173 struct lnet_msg *msg2;
1175 if (msg->msg_rtrcredit) {
1176 /* give back global router credits */
1177 struct lnet_rtrbuf *rb;
1178 struct lnet_rtrbufpool *rbp;
1180 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1181 * there until it gets one allocated, or aborts the wait
1183 LASSERT(msg->msg_kiov != NULL);
1185 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1188 msg->msg_kiov = NULL;
1189 msg->msg_rtrcredit = 0;
1191 LASSERT(rbp == lnet_msg2bufpool(msg));
1193 LASSERT((rbp->rbp_credits > 0) ==
1194 !list_empty(&rbp->rbp_bufs));
1196 /* If routing is now turned off, we just drop this buffer and
1197 * don't bother trying to return credits. */
1198 if (!the_lnet.ln_routing) {
1199 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1203 /* It is possible that a user has lowered the desired number of
1204 * buffers in this pool. Make sure we never put back
1205 * more buffers than the stated number. */
1206 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1207 /* Discard this buffer so we don't have too
1209 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1210 rbp->rbp_nbuffers--;
1212 list_add(&rb->rb_list, &rbp->rbp_bufs);
1214 if (rbp->rbp_credits <= 0)
1215 lnet_schedule_blocked_locked(rbp);
1220 if (msg->msg_peerrtrcredit) {
1222 LASSERT(rxpeerni->lpni_peer_net);
1223 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1225 /* give back peer router credits */
1226 msg->msg_peerrtrcredit = 0;
1228 spin_lock(&rxpeerni->lpni_lock);
1229 rxpeerni->lpni_rtrcredits++;
1230 spin_unlock(&rxpeerni->lpni_lock);
1232 lp = rxpeerni->lpni_peer_net->lpn_peer;
1233 spin_lock(&lp->lp_lock);
1235 /* drop all messages which are queued to be routed on that
1237 if (!the_lnet.ln_routing) {
1239 list_splice_init(&lp->lp_rtrq, &drop);
1240 spin_unlock(&lp->lp_lock);
1241 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1242 } else if (!list_empty(&lp->lp_rtrq)) {
1245 msg2 = list_first_entry(&lp->lp_rtrq,
1246 struct lnet_msg, msg_list);
1247 list_del(&msg2->msg_list);
1248 msg2_cpt = msg2->msg_rx_cpt;
1249 spin_unlock(&lp->lp_lock);
1251 * messages on the lp_rtrq can be from any NID in
1252 * the peer, which means they might have different
1253 * cpts. We need to make sure we lock the right
1256 if (msg2_cpt != msg->msg_rx_cpt) {
1257 lnet_net_unlock(msg->msg_rx_cpt);
1258 lnet_net_lock(msg2_cpt);
1260 (void) lnet_post_routed_recv_locked(msg2, 1);
1261 if (msg2_cpt != msg->msg_rx_cpt) {
1262 lnet_net_unlock(msg2_cpt);
1263 lnet_net_lock(msg->msg_rx_cpt);
1266 spin_unlock(&lp->lp_lock);
1270 msg->msg_rxni = NULL;
1271 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1273 if (rxpeerni != NULL) {
1274 msg->msg_rxpeer = NULL;
1275 lnet_peer_ni_decref_locked(rxpeerni);
1279 static struct lnet_peer_ni *
1280 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1281 struct lnet_peer *peer,
1282 struct lnet_peer_ni *best_lpni,
1283 struct lnet_peer_net *peer_net)
1286 * Look at the peer NIs for the destination peer that connect
1287 * to the chosen net. If a peer_ni is preferred when using the
1288 * best_ni to communicate, we use that one. If there is no
1289 * preferred peer_ni, or there are multiple preferred peer_ni,
1290 * the available transmit credits are used. If the transmit
1291 * credits are equal, we round-robin over the peer_ni.
1293 struct lnet_peer_ni *lpni = NULL;
1294 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1296 int best_lpni_healthv = (best_lpni) ?
1297 atomic_read(&best_lpni->lpni_healthv) : 0;
1298 bool best_lpni_is_preferred = false;
1299 bool lpni_is_preferred;
1301 __u32 lpni_sel_prio;
1302 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1304 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1306 * if the best_ni we've chosen aleady has this lpni
1307 * preferred, then let's use it
1310 lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1311 lpni, &best_ni->ni_nid);
1312 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1313 libcfs_nidstr(&best_ni->ni_nid),
1316 lpni_is_preferred = false;
1319 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1320 lpni_sel_prio = lpni->lpni_sel_priority;
1323 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1324 libcfs_nidstr(&lpni->lpni_nid),
1325 libcfs_nidstr(&best_lpni->lpni_nid),
1326 lpni_healthv, best_lpni_healthv,
1327 lpni_sel_prio, best_sel_prio,
1328 lpni->lpni_txcredits, best_lpni_credits,
1329 lpni->lpni_seq, best_lpni->lpni_seq);
1333 /* pick the healthiest peer ni */
1334 if (lpni_healthv < best_lpni_healthv)
1336 else if (lpni_healthv > best_lpni_healthv) {
1337 if (best_lpni_is_preferred)
1338 best_lpni_is_preferred = false;
1342 if (lpni_sel_prio > best_sel_prio)
1344 else if (lpni_sel_prio < best_sel_prio) {
1345 if (best_lpni_is_preferred)
1346 best_lpni_is_preferred = false;
1350 /* if this is a preferred peer use it */
1351 if (!best_lpni_is_preferred && lpni_is_preferred) {
1352 best_lpni_is_preferred = true;
1354 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1355 /* this is not the preferred peer so let's ignore
1361 if (lpni->lpni_txcredits < best_lpni_credits)
1362 /* We already have a peer that has more credits
1363 * available than this one. No need to consider
1364 * this peer further.
1367 else if (lpni->lpni_txcredits > best_lpni_credits)
1370 /* The best peer found so far and the current peer
1371 * have the same number of available credits let's
1372 * make sure to select between them using Round Robin
1374 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1377 best_lpni_is_preferred = lpni_is_preferred;
1378 best_lpni_healthv = lpni_healthv;
1379 best_sel_prio = lpni_sel_prio;
1381 best_lpni_credits = lpni->lpni_txcredits;
1384 /* if we still can't find a peer ni then we can't reach it */
1386 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1387 LNET_NIDNET(dst_nid);
1388 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1389 libcfs_net2str(net_id));
1393 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1394 libcfs_nidstr(&best_lpni->lpni_nid));
1400 * Prerequisite: the best_ni should already be set in the sd
1401 * Find the best lpni.
1402 * If the net id is provided then restrict lpni selection on
1403 * that particular net.
1404 * Otherwise find any reachable lpni. When dealing with an MR
1405 * gateway and it has multiple lpnis which we can use
1406 * we want to select the best one from the list of reachable
1409 static inline struct lnet_peer_ni *
1410 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1411 struct lnet_peer *peer, __u32 net_id)
1413 struct lnet_peer_net *peer_net;
1415 /* find the best_lpni on any local network */
1416 if (net_id == LNET_NET_ANY) {
1417 struct lnet_peer_ni *best_lpni = NULL;
1418 struct lnet_peer_net *lpn;
1419 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1420 /* no net specified find any reachable peer ni */
1421 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1423 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1429 /* restrict on the specified net */
1430 peer_net = lnet_peer_get_net_locked(peer, net_id);
1432 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1438 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1440 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1443 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1446 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1449 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1455 /* Compare route priorities and hop counts */
1457 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1459 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1460 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1462 if (r1->lr_priority < r2->lr_priority)
1465 if (r1->lr_priority > r2->lr_priority)
1468 if (r1_hops < r2_hops)
1471 if (r1_hops > r2_hops)
1477 static struct lnet_route *
1478 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1479 struct lnet_peer_ni *remote_lpni,
1480 struct lnet_route **prev_route,
1481 struct lnet_peer_ni **gwni)
1483 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1484 struct lnet_route *best_route;
1485 struct lnet_route *last_route;
1486 struct lnet_route *route;
1488 bool best_rte_is_preferred = false;
1489 struct lnet_nid *gw_pnid;
1491 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1492 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1494 best_route = last_route = NULL;
1495 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1496 if (!lnet_is_route_alive(route))
1498 gw_pnid = &route->lr_gateway->lp_primary_nid;
1500 /* no protection on below fields, but it's harmless */
1501 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1504 /* if the best route found is in the preferred list then
1505 * tag it as preferred and use it later on. But if we
1506 * didn't find any routes which are on the preferred list
1507 * then just use the best route possible.
1509 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1511 if (!best_route || (rc && !best_rte_is_preferred)) {
1512 /* Restrict the selection of the router NI on the
1513 * src_net provided. If the src_net is LNET_NID_ANY,
1514 * then select the best interface available.
1516 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1521 "Gateway %s does not have a peer NI on net %s\n",
1522 libcfs_nidstr(gw_pnid),
1523 libcfs_net2str(src_net));
1528 if (rc && !best_rte_is_preferred) {
1529 /* This is the first preferred route we found,
1530 * so it beats any route found previously
1536 best_rte_is_preferred = true;
1537 CDEBUG(D_NET, "preferred gw = %s\n",
1538 libcfs_nidstr(gw_pnid));
1540 } else if ((!rc) && best_rte_is_preferred)
1541 /* The best route we found so far is in the preferred
1542 * list, so it beats any non-preferred route
1547 best_route = last_route = route;
1552 rc = lnet_compare_routes(route, best_route);
1556 /* Restrict the selection of the router NI on the
1557 * src_net provided. If the src_net is LNET_NID_ANY,
1558 * then select the best interface available.
1560 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1565 "Gateway %s does not have a peer NI on net %s\n",
1566 libcfs_nidstr(gw_pnid),
1567 libcfs_net2str(src_net));
1577 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1581 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1588 *prev_route = last_route;
1594 static inline unsigned int
1595 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1597 if (dev_idx == UINT_MAX)
1600 if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1601 !ni->ni_net->net_lnd->lnd_get_dev_prio)
1604 return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1607 static struct lnet_ni *
1608 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1609 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1610 struct lnet_msg *msg, int md_cpt)
1612 struct lnet_libmd *md = msg->msg_md;
1613 unsigned int offset = msg->msg_offset;
1614 unsigned int shortest_distance;
1615 struct lnet_ni *ni = NULL;
1618 __u32 best_sel_prio;
1619 unsigned int best_dev_prio;
1621 unsigned int dev_idx = UINT_MAX;
1622 bool gpu = md ? (md->md_flags & LNET_MD_FLAG_GPU) : false;
1625 struct page *page = lnet_get_first_page(md, offset);
1627 dev_idx = lnet_get_dev_idx(page);
1631 * If there is no peer_ni that we can send to on this network,
1632 * then there is no point in looking for a new best_ni here.
1634 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1637 if (best_ni == NULL) {
1638 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1639 shortest_distance = UINT_MAX;
1640 best_dev_prio = UINT_MAX;
1641 best_credits = INT_MIN;
1643 best_ni_fatal = true;
1645 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1646 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1647 best_ni->ni_dev_cpt);
1648 best_credits = atomic_read(&best_ni->ni_tx_credits);
1649 best_healthv = atomic_read(&best_ni->ni_healthv);
1650 best_sel_prio = best_ni->ni_sel_priority;
1651 best_ni_fatal = atomic_read(&best_ni->ni_fatal_error_on);
1654 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1655 unsigned int distance;
1660 unsigned int ni_dev_prio;
1662 ni_credits = atomic_read(&ni->ni_tx_credits);
1663 ni_healthv = atomic_read(&ni->ni_healthv);
1664 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1665 ni_sel_prio = ni->ni_sel_priority;
1668 * calculate the distance from the CPT on which
1669 * the message memory is allocated to the CPT of
1670 * the NI's physical device
1672 distance = cfs_cpt_distance(lnet_cpt_table(),
1676 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1679 * All distances smaller than the NUMA range
1680 * are treated equally.
1682 if (!gpu && distance < lnet_numa_range)
1683 distance = lnet_numa_range;
1686 * Select on health, selection policy, direct dma prio,
1687 * shorter distance, available credits, then round-robin.
1690 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",
1691 libcfs_nidstr(&ni->ni_nid),
1692 ni_fatal ? "y" : "n", ni_credits, distance,
1693 ni->ni_seq, ni_sel_prio, ni_dev_prio, ni_healthv,
1694 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1696 best_ni_fatal ? "y" : "n", best_credits,
1698 (best_ni) ? best_ni->ni_seq : 0,
1699 best_sel_prio, best_dev_prio, best_healthv);
1703 if (ni_fatal && !best_ni_fatal)
1705 else if (!ni_fatal && best_ni_fatal)
1708 if (ni_healthv < best_healthv)
1710 else if (ni_healthv > best_healthv)
1713 if (ni_sel_prio > best_sel_prio)
1715 else if (ni_sel_prio < best_sel_prio)
1718 if (ni_dev_prio > best_dev_prio)
1720 else if (ni_dev_prio < best_dev_prio)
1723 if (distance > shortest_distance)
1725 else if (distance < shortest_distance)
1728 if (ni_credits < best_credits)
1730 else if (ni_credits > best_credits)
1733 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1737 best_sel_prio = ni_sel_prio;
1738 best_dev_prio = ni_dev_prio;
1739 shortest_distance = distance;
1740 best_healthv = ni_healthv;
1742 best_credits = ni_credits;
1743 best_ni_fatal = ni_fatal;
1746 CDEBUG(D_NET, "selected best_ni %s\n",
1747 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1753 lnet_reserved_msg(struct lnet_msg *msg)
1755 if (msg->msg_type == LNET_MSG_PUT) {
1756 if (msg->msg_hdr.msg.put.ptl_index == LNET_RESERVED_PORTAL)
1758 } else if (msg->msg_type == LNET_MSG_GET) {
1759 if (msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1765 /* Can the specified message trigger peer discovery?
1767 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1768 * because such traffic is required to perform discovery. We therefore
1769 * exclude all GET and PUT on that portal. We also exclude all ACK and
1770 * REPLY traffic, but that is because the portal is not tracked in the
1771 * message structure for these message types. We could restrict this
1772 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1775 lnet_msg_discovery(struct lnet_msg *msg)
1777 return !(lnet_reserved_msg(msg) || lnet_msg_is_response(msg));
1780 /* Is the specified message an LNet ping?
1783 lnet_msg_is_ping(struct lnet_msg *msg)
1785 if (msg->msg_type == LNET_MSG_GET &&
1786 msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1792 #define SRC_SPEC 0x0001
1793 #define SRC_ANY 0x0002
1794 #define LOCAL_DST 0x0004
1795 #define REMOTE_DST 0x0008
1796 #define MR_DST 0x0010
1797 #define NMR_DST 0x0020
1798 #define SND_RESP 0x0040
1800 /* The following to defines are used for return codes */
1801 #define REPEAT_SEND 0x1000
1802 #define PASS_THROUGH 0x2000
1804 /* The different cases lnet_select pathway needs to handle */
1805 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1806 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1807 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1808 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1809 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1810 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1811 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1812 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1815 lnet_handle_lo_send(struct lnet_send_data *sd)
1817 struct lnet_msg *msg = sd->sd_msg;
1818 int cpt = sd->sd_cpt;
1820 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1823 /* No send credit hassles with LOLND */
1824 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1825 msg->msg_hdr.dest_nid = the_lnet.ln_loni->ni_nid;
1826 if (!msg->msg_routing)
1827 msg->msg_hdr.src_nid = the_lnet.ln_loni->ni_nid;
1828 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1829 lnet_msg_commit(msg, cpt);
1830 msg->msg_txni = the_lnet.ln_loni;
1832 return LNET_CREDIT_OK;
1836 lnet_handle_send(struct lnet_send_data *sd)
1838 struct lnet_ni *best_ni = sd->sd_best_ni;
1839 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1840 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1841 struct lnet_msg *msg = sd->sd_msg;
1843 __u32 send_case = sd->sd_send_case;
1845 __u32 routing = send_case & REMOTE_DST;
1846 struct lnet_rsp_tracker *rspt;
1848 /* Increment sequence number of the selected peer, peer net,
1849 * local ni and local net so that we pick the next ones
1852 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1853 best_lpni->lpni_peer_net->lpn_seq =
1854 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1855 best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1856 the_lnet.ln_net_seq++;
1857 best_ni->ni_net->net_seq = the_lnet.ln_net_seq;
1858 best_ni->ni_seq = best_ni->ni_net->net_seq;
1860 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1861 libcfs_nidstr(&best_ni->ni_nid),
1862 best_ni->ni_seq, best_ni->ni_net->net_seq,
1863 atomic_read(&best_ni->ni_tx_credits),
1864 best_ni->ni_sel_priority,
1865 libcfs_nidstr(&best_lpni->lpni_nid),
1866 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1867 best_lpni->lpni_txcredits,
1868 best_lpni->lpni_sel_priority);
1871 * grab a reference on the peer_ni so it sticks around even if
1872 * we need to drop and relock the lnet_net_lock below.
1874 lnet_peer_ni_addref_locked(best_lpni);
1877 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1878 * message. This ensures that we get a CPT that is correct for
1879 * the NI when the NI has been restricted to a subset of all CPTs.
1880 * If the selected CPT differs from the one currently locked, we
1881 * must unlock and relock the lnet_net_lock(), and then check whether
1882 * the configuration has changed. We don't have a hold on the best_ni
1883 * yet, and it may have vanished.
1885 cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1886 if (sd->sd_cpt != cpt2) {
1887 __u32 seq = lnet_get_dlc_seq_locked();
1888 lnet_net_unlock(sd->sd_cpt);
1890 lnet_net_lock(sd->sd_cpt);
1891 if (seq != lnet_get_dlc_seq_locked()) {
1892 lnet_peer_ni_decref_locked(best_lpni);
1898 * store the best_lpni in the message right away to avoid having
1899 * to do the same operation under different conditions
1901 msg->msg_txpeer = best_lpni;
1902 msg->msg_txni = best_ni;
1905 * grab a reference for the best_ni since now it's in use in this
1906 * send. The reference will be dropped in lnet_finalize()
1908 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1911 * Always set the target.nid to the best peer picked. Either the
1912 * NID will be one of the peer NIDs selected, or the same NID as
1913 * what was originally set in the target or it will be the NID of
1914 * a router if this message should be routed
1916 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1919 * lnet_msg_commit assigns the correct cpt to the message, which
1920 * is used to decrement the correct refcount on the ni when it's
1921 * time to return the credits
1923 lnet_msg_commit(msg, sd->sd_cpt);
1926 * If we are routing the message then we keep the src_nid that was
1927 * set by the originator. If we are not routing then we are the
1928 * originator and set it here.
1930 if (!msg->msg_routing)
1931 msg->msg_hdr.src_nid = msg->msg_txni->ni_nid;
1934 msg->msg_target_is_router = 1;
1935 msg->msg_target.pid = LNET_PID_LUSTRE;
1937 * since we're routing we want to ensure that the
1938 * msg_hdr.dest_nid is set to the final destination. When
1939 * the router receives this message it knows how to route
1942 * final_dst_lpni is set at the beginning of the
1943 * lnet_select_pathway() function and is never changed.
1944 * It's safe to use it here.
1946 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1947 final_dst_lpni->lpni_peer_net->lpn_seq =
1948 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1949 final_dst_lpni->lpni_seq =
1950 final_dst_lpni->lpni_peer_net->lpn_seq;
1951 msg->msg_hdr.dest_nid = final_dst_lpni->lpni_nid;
1954 * if we're not routing set the dest_nid to the best peer
1955 * ni NID that we picked earlier in the algorithm.
1957 msg->msg_hdr.dest_nid = msg->msg_txpeer->lpni_nid;
1961 * if we have response tracker block update it with the next hop
1965 rspt = msg->msg_md->md_rspt_ptr;
1967 rspt->rspt_next_hop_nid =
1968 msg->msg_txpeer->lpni_nid;
1969 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1970 libcfs_nidstr(&rspt->rspt_next_hop_nid));
1974 rc = lnet_post_send_locked(msg, 0);
1977 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1978 libcfs_nidstr(&msg->msg_hdr.src_nid),
1979 libcfs_nidstr(&msg->msg_txni->ni_nid),
1980 libcfs_nidstr(&sd->sd_src_nid),
1981 libcfs_nidstr(&msg->msg_hdr.dest_nid),
1982 libcfs_nidstr(&sd->sd_dst_nid),
1983 libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
1984 libcfs_nidstr(&sd->sd_rtr_nid),
1985 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1991 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1992 struct lnet_msg *msg)
1994 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1995 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1996 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1997 libcfs_nidstr(&lni->ni_nid),
1998 libcfs_nidstr(&lpni->lpni_nid));
1999 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
2008 * use the source and destination NIDs as the pathway
2011 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2013 /* the destination lpni is set before we get here. */
2016 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2017 if (!sd->sd_best_ni) {
2018 CERROR("Can't send to %s: src %s is not a local nid\n",
2019 libcfs_nidstr(&sd->sd_dst_nid),
2020 libcfs_nidstr(&sd->sd_src_nid));
2024 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2026 return lnet_handle_send(sd);
2034 * Don't run the selection algorithm on the peer NIs. By specifying the
2035 * local NID, we're also saying that we should always use the destination NID
2036 * provided. This handles the case where we should be using the same
2037 * destination NID for the all the messages which belong to the same RPC
2041 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2043 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2044 if (!sd->sd_best_ni) {
2045 CERROR("Can't send to %s: src %s is not a local nid\n",
2046 libcfs_nidstr(&sd->sd_dst_nid),
2047 libcfs_nidstr(&sd->sd_src_nid));
2051 if (sd->sd_best_lpni &&
2052 nid_same(&sd->sd_best_lpni->lpni_nid,
2053 &the_lnet.ln_loni->ni_nid))
2054 return lnet_handle_lo_send(sd);
2055 else if (sd->sd_best_lpni)
2056 return lnet_handle_send(sd);
2058 CERROR("can't send to %s. no NI on %s\n",
2059 libcfs_nidstr(&sd->sd_dst_nid),
2060 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2062 return -EHOSTUNREACH;
2066 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2067 struct lnet_peer *peer,
2068 struct lnet_peer_net *peer_net,
2069 struct lnet_msg *msg,
2072 struct lnet_net *local_net;
2073 struct lnet_ni *best_ni;
2075 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2080 * Iterate through the NIs in this local Net and select
2081 * the NI to send from. The selection is determined by
2082 * these 3 criterion in the following priority:
2084 * 2. NI available credits
2087 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2088 peer, peer_net, msg, cpt);
2094 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2097 struct lnet_peer *peer;
2098 struct lnet_peer_ni *new_lpni;
2101 lnet_peer_ni_addref_locked(lpni);
2103 peer = lpni->lpni_peer_net->lpn_peer;
2105 if (lnet_peer_gw_discovery(peer)) {
2106 lnet_peer_ni_decref_locked(lpni);
2110 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2111 lnet_peer_ni_decref_locked(lpni);
2115 rc = lnet_discover_peer_locked(lpni, cpt, false);
2117 lnet_peer_ni_decref_locked(lpni);
2121 new_lpni = lnet_peer_ni_find_locked(&lpni->lpni_nid);
2123 lnet_peer_ni_decref_locked(lpni);
2127 peer = new_lpni->lpni_peer_net->lpn_peer;
2128 spin_lock(&peer->lp_lock);
2129 if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2130 /* The peer NI did not change and the peer is up to date.
2131 * Nothing more to do.
2133 spin_unlock(&peer->lp_lock);
2134 lnet_peer_ni_decref_locked(lpni);
2135 lnet_peer_ni_decref_locked(new_lpni);
2138 spin_unlock(&peer->lp_lock);
2140 /* Either the peer NI changed during discovery, or the peer isn't up
2141 * to date. In both cases we want to queue the message on the
2142 * (possibly new) peer's pending queue and queue the peer for discovery
2144 msg->msg_sending = 0;
2145 msg->msg_txpeer = NULL;
2146 lnet_net_unlock(cpt);
2147 lnet_peer_queue_message(peer, msg);
2150 lnet_peer_ni_decref_locked(lpni);
2151 lnet_peer_ni_decref_locked(new_lpni);
2153 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2154 msg, libcfs_nidstr(&peer->lp_primary_nid));
2156 return LNET_DC_WAIT;
2160 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2161 struct lnet_nid *dst_nid,
2162 struct lnet_peer_ni **gw_lpni,
2163 struct lnet_peer **gw_peer)
2166 struct lnet_peer *gw;
2167 struct lnet_peer *lp;
2168 struct lnet_peer_net *lpn;
2169 struct lnet_peer_net *best_lpn = NULL;
2170 struct lnet_remotenet *rnet, *best_rnet = NULL;
2171 struct lnet_route *best_route = NULL;
2172 struct lnet_route *last_route = NULL;
2173 struct lnet_peer_ni *lpni = NULL;
2174 struct lnet_peer_ni *gwni = NULL;
2175 bool route_found = false;
2176 struct lnet_nid *src_nid =
2177 !LNET_NID_IS_ANY(&sd->sd_src_nid) || !sd->sd_best_ni
2179 : &sd->sd_best_ni->ni_nid;
2180 int best_lpn_healthv = 0;
2181 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2183 CDEBUG(D_NET, "%s route (%s) from local NI %s to destination %s\n",
2184 LNET_NID_IS_ANY(&sd->sd_rtr_nid) ? "Lookup" : "Specified",
2185 libcfs_nidstr(&sd->sd_rtr_nid), libcfs_nidstr(src_nid),
2186 libcfs_nidstr(&sd->sd_dst_nid));
2188 /* If a router nid was specified then we are replying to a GET or
2189 * sending an ACK. In this case we use the gateway associated with the
2190 * specified router nid.
2192 if (!LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2193 gwni = lnet_peer_ni_find_locked(&sd->sd_rtr_nid);
2195 gw = gwni->lpni_peer_net->lpn_peer;
2196 lnet_peer_ni_decref_locked(gwni);
2197 if (gw->lp_rtr_refcount)
2200 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2201 libcfs_nidstr(&sd->sd_rtr_nid));
2206 if (sd->sd_msg->msg_routing || !LNET_NID_IS_ANY(src_nid)) {
2207 /* If I'm routing this message then I need to find the
2208 * next hop based on the destination NID
2210 * We also find next hop based on the destination NID
2211 * if the source NI was specified
2213 best_rnet = lnet_find_rnet_locked(LNET_NID_NET(&sd->sd_dst_nid));
2215 CERROR("Unable to send message from %s to %s - Route table may be misconfigured\n",
2216 (src_nid && LNET_NID_IS_ANY(src_nid)) ?
2218 libcfs_nidstr(src_nid),
2219 libcfs_nidstr(&sd->sd_dst_nid));
2220 return -EHOSTUNREACH;
2222 CDEBUG(D_NET, "best_rnet %s\n",
2223 libcfs_net2str(best_rnet->lrn_net));
2225 /* we've already looked up the initial lpni using
2228 lpni = sd->sd_best_lpni;
2229 /* the peer tree must be in existence */
2230 LASSERT(lpni && lpni->lpni_peer_net &&
2231 lpni->lpni_peer_net->lpn_peer);
2232 lp = lpni->lpni_peer_net->lpn_peer;
2234 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2235 /* is this remote network reachable? */
2236 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2243 CDEBUG(D_NET, "n[%s, %s] h[%d, %d], p[%u, %u], s[%d, %d]\n",
2244 libcfs_net2str(lpn->lpn_net_id),
2245 libcfs_net2str(best_lpn->lpn_net_id),
2247 best_lpn->lpn_healthv,
2248 lpn->lpn_sel_priority,
2249 best_lpn->lpn_sel_priority,
2253 /* select the preferred peer net */
2254 if (best_lpn_healthv > lpn->lpn_healthv)
2256 else if (best_lpn_healthv < lpn->lpn_healthv)
2259 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2261 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2264 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2267 best_lpn_healthv = lpn->lpn_healthv;
2268 best_lpn_sel_prio = lpn->lpn_sel_priority;
2274 CERROR("peer %s has no available nets\n",
2275 libcfs_nidstr(&sd->sd_dst_nid));
2276 return -EHOSTUNREACH;
2279 CDEBUG(D_NET, "selected best_lpn %s\n",
2280 libcfs_net2str(best_lpn->lpn_net_id));
2282 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2283 lnet_nid_to_nid4(&sd->sd_dst_nid),
2285 best_lpn->lpn_net_id);
2286 if (!sd->sd_best_lpni) {
2287 CERROR("peer %s is unreachable\n",
2288 libcfs_nidstr(&sd->sd_dst_nid));
2289 return -EHOSTUNREACH;
2292 /* We're attempting to round robin over the remote peer
2293 * NI's so update the final destination we selected
2295 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2299 * find the best route. Restrict the selection on the net of the
2300 * local NI if we've already picked the local NI to send from.
2301 * Otherwise, let's pick any route we can find and then find
2302 * a local NI we can reach the route's gateway on. Any route we
2303 * select will be reachable by virtue of the restriction we have
2304 * when adding a route.
2306 best_route = lnet_find_route_locked(best_rnet,
2307 LNET_NID_NET(src_nid),
2309 &last_route, &gwni);
2312 CERROR("no route to %s from %s\n",
2313 libcfs_nidstr(dst_nid),
2314 libcfs_nidstr(src_nid));
2315 return -EHOSTUNREACH;
2319 CERROR("Internal Error. Route expected to %s from %s\n",
2320 libcfs_nidstr(dst_nid),
2321 libcfs_nidstr(src_nid));
2325 gw = best_route->lr_gateway;
2326 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2330 * If the router checker is not active then discover the gateway here.
2331 * This ensures we are able to take advantage of multi-rail routing, but
2332 * if the router checker is active then we do not unecessarily delay
2333 * messages while the gateway is being checked by the dedicated monitor
2336 * NB: We're only checking the alive_router_check_interval here, rather
2337 * than calling lnet_router_checker_active(), because the other
2338 * conditions that are checked by that function are either
2339 * irrelevant (the_lnet.ln_routing) or must be true (list of routers
2342 if (alive_router_check_interval <= 0) {
2343 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2348 if (!sd->sd_best_ni) {
2349 lpn = gwni->lpni_peer_net;
2350 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2353 if (!sd->sd_best_ni) {
2354 CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2355 libcfs_net2str(lpn->lpn_net_id),
2356 libcfs_nidstr(&sd->sd_src_nid));
2365 * increment the sequence number since now we're sure we're
2366 * going to use this route
2368 if (LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2369 LASSERT(best_route && last_route);
2370 best_route->lr_seq = last_route->lr_seq + 1;
2381 * Remote destination
2382 * Non-MR destination
2386 * Remote destination
2389 * The handling of these two cases is similar. Even though the destination
2390 * can be MR or non-MR, we'll deal directly with the router.
2393 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2396 struct lnet_peer_ni *gw_lpni = NULL;
2397 struct lnet_peer *gw_peer = NULL;
2400 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2401 if (!sd->sd_best_ni) {
2402 CERROR("Can't send to %s: src %s is not a local nid\n",
2403 libcfs_nidstr(&sd->sd_dst_nid),
2404 libcfs_nidstr(&sd->sd_src_nid));
2408 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2409 &gw_lpni, &gw_peer);
2413 if (sd->sd_send_case & NMR_DST)
2415 * since the final destination is non-MR let's set its preferred
2416 * NID before we send
2418 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2422 * We're going to send to the gw found so let's set its
2425 sd->sd_peer = gw_peer;
2426 sd->sd_best_lpni = gw_lpni;
2428 return lnet_handle_send(sd);
2432 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2433 struct lnet_msg *msg, bool discovery)
2435 struct lnet_peer_net *lpn = NULL;
2436 struct lnet_peer_net *best_lpn = NULL;
2437 struct lnet_net *net = NULL;
2438 struct lnet_net *best_net = NULL;
2439 struct lnet_ni *best_ni = NULL;
2440 int best_lpn_healthv = 0;
2441 int best_net_healthv = 0;
2443 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2445 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2448 /* If lp_disc_net_id is set, this peer is a router undergoing
2449 * discovery, and this message is an LNet ping, then this may be a
2450 * discovery message and we need to select an NI on the peer net
2451 * specified by lp_disc_net_id
2453 if (peer->lp_disc_net_id &&
2454 (peer->lp_state & LNET_PEER_RTR_DISCOVERY) &&
2455 lnet_msg_is_ping(msg)) {
2456 best_lpn = lnet_peer_get_net_locked(peer, peer->lp_disc_net_id);
2457 if (best_lpn && lnet_get_net_locked(best_lpn->lpn_net_id))
2458 goto select_best_ni;
2462 * The peer can have multiple interfaces, some of them can be on
2463 * the local network and others on a routed network. We should
2464 * prefer the local network. However if the local network is not
2465 * available then we need to try the routed network
2468 /* go through all the peer nets and find the best_ni */
2469 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2471 * The peer's list of nets can contain non-local nets. We
2472 * want to only examine the local ones.
2474 net = lnet_get_net_locked(lpn->lpn_net_id);
2478 lpn_sel_prio = lpn->lpn_sel_priority;
2479 net_healthv = lnet_get_net_healthv_locked(net);
2480 net_sel_prio = net->net_sel_priority;
2486 "n[%s, %s] ph[%d, %d], pp[%u, %u], nh[%d, %d], np[%u, %u], ps[%u, %u], ns[%u, %u]\n",
2487 libcfs_net2str(lpn->lpn_net_id),
2488 libcfs_net2str(best_lpn->lpn_net_id),
2502 /* always select the lpn with the best health */
2503 if (best_lpn_healthv > lpn->lpn_healthv)
2505 else if (best_lpn_healthv < lpn->lpn_healthv)
2508 /* select the preferred peer and local nets */
2509 if (best_lpn_sel_prio < lpn_sel_prio)
2511 else if (best_lpn_sel_prio > lpn_sel_prio)
2514 if (best_net_healthv > net_healthv)
2516 else if (best_net_healthv < net_healthv)
2519 if (best_net_sel_prio < net_sel_prio)
2521 else if (best_net_sel_prio > net_sel_prio)
2524 if (best_lpn->lpn_seq < lpn->lpn_seq)
2526 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2529 /* round robin over the local networks */
2530 if (best_net->net_seq <= net->net_seq)
2534 best_net_healthv = net_healthv;
2535 best_net_sel_prio = net_sel_prio;
2536 best_lpn_healthv = lpn->lpn_healthv;
2537 best_lpn_sel_prio = lpn_sel_prio;
2543 /* Select the best NI on the same net as best_lpn chosen
2547 CDEBUG(D_NET, "selected best_lpn %s\n",
2548 libcfs_net2str(best_lpn->lpn_net_id));
2549 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2556 static struct lnet_ni *
2557 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2559 struct lnet_ni *best_ni = NULL;
2560 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2561 struct lnet_peer_ni *lpni_entry;
2564 * We must use a consistent source address when sending to a
2565 * non-MR peer. However, a non-MR peer can have multiple NIDs
2566 * on multiple networks, and we may even need to talk to this
2567 * peer on multiple networks -- certain types of
2568 * load-balancing configuration do this.
2570 * So we need to pick the NI the peer prefers for this
2571 * particular network.
2574 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2576 if (lpni_entry->lpni_pref_nnids == 0)
2578 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2579 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2587 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2589 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2591 struct lnet_ni *best_ni = NULL;
2594 * We must use a consistent source address when sending to a
2595 * non-MR peer. However, a non-MR peer can have multiple NIDs
2596 * on multiple networks, and we may even need to talk to this
2597 * peer on multiple networks -- certain types of
2598 * load-balancing configuration do this.
2600 * So we need to pick the NI the peer prefers for this
2601 * particular network.
2603 * An exception is traffic on LNET_RESERVED_PORTAL. Internal LNet
2604 * traffic doesn't care which source NI is used, and we don't actually
2605 * want to restrict local recovery pings to a single source NI.
2607 if (!lnet_reserved_msg(sd->sd_msg))
2608 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2612 best_ni = lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2613 sd->sd_best_lpni->lpni_peer_net,
2617 /* If there is no best_ni we don't have a route */
2619 CERROR("no path to %s from net %s\n",
2620 libcfs_nidstr(&sd->sd_best_lpni->lpni_nid),
2621 libcfs_net2str(sd->sd_best_lpni->lpni_net->net_id));
2622 return -EHOSTUNREACH;
2625 sd->sd_best_ni = best_ni;
2627 /* Set preferred NI if necessary. */
2628 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2635 * Source not specified
2639 * always use the same source NID for NMR peers
2640 * If we've talked to that peer before then we already have a preferred
2641 * source NI associated with it. Otherwise, we select a preferred local NI
2642 * and store it in the peer
2645 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2649 /* sd->sd_best_lpni is already set to the final destination */
2652 * At this point we should've created the peer ni and peer. If we
2653 * can't find it, then something went wrong. Instead of assert
2654 * output a relevant message and fail the send
2656 if (!sd->sd_best_lpni) {
2657 CERROR("Internal fault. Unable to send msg %s to %s. NID not known\n",
2658 lnet_msgtyp2str(sd->sd_msg->msg_type),
2659 libcfs_nidstr(&sd->sd_dst_nid));
2663 if (sd->sd_msg->msg_routing) {
2664 /* If I'm forwarding this message then I can choose any NI
2665 * on the destination peer net
2667 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2669 sd->sd_best_lpni->lpni_peer_net,
2672 if (!sd->sd_best_ni) {
2673 CERROR("Unable to forward message to %s. No local NI available\n",
2674 libcfs_nidstr(&sd->sd_dst_nid));
2678 rc = lnet_select_preferred_best_ni(sd);
2681 rc = lnet_handle_send(sd);
2687 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2690 * NOTE we've already handled the remote peer case. So we only
2691 * need to worry about the local case here.
2693 * if we're sending a response, ACK or reply, we need to send it
2694 * to the destination NID given to us. At this point we already
2695 * have the peer_ni we're suppose to send to, so just find the
2696 * best_ni on the peer net and use that. Since we're sending to an
2697 * MR peer then we can just run the selection algorithm on our
2698 * local NIs and pick the best one.
2700 if (sd->sd_send_case & SND_RESP) {
2702 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2703 sd->sd_best_lpni->lpni_peer_net,
2707 if (!sd->sd_best_ni) {
2709 * We're not going to deal with not able to send
2710 * a response to the provided final destination
2712 CERROR("Can't send response to %s. No local NI available\n",
2713 libcfs_nidstr(&sd->sd_dst_nid));
2714 return -EHOSTUNREACH;
2717 return lnet_handle_send(sd);
2721 * If we get here that means we're sending a fresh request, PUT or
2722 * GET, so we need to run our standard selection algorithm.
2723 * First find the best local interface that's on any of the peer's
2726 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2729 lnet_msg_discovery(sd->sd_msg));
2730 if (sd->sd_best_ni) {
2732 lnet_find_best_lpni(sd->sd_best_ni,
2733 lnet_nid_to_nid4(&sd->sd_dst_nid),
2735 sd->sd_best_ni->ni_net->net_id);
2738 * if we're successful in selecting a peer_ni on the local
2739 * network, then send to it. Otherwise fall through and
2740 * try and see if we can reach it over another routed
2743 if (sd->sd_best_lpni &&
2744 nid_same(&sd->sd_best_lpni->lpni_nid,
2745 &the_lnet.ln_loni->ni_nid)) {
2747 * in case we initially started with a routed
2748 * destination, let's reset to local
2750 sd->sd_send_case &= ~REMOTE_DST;
2751 sd->sd_send_case |= LOCAL_DST;
2752 return lnet_handle_lo_send(sd);
2753 } else if (sd->sd_best_lpni) {
2755 * in case we initially started with a routed
2756 * destination, let's reset to local
2758 sd->sd_send_case &= ~REMOTE_DST;
2759 sd->sd_send_case |= LOCAL_DST;
2760 return lnet_handle_send(sd);
2763 CERROR("Internal Error. Expected to have a best_lpni: "
2765 libcfs_nidstr(&sd->sd_src_nid),
2766 libcfs_nidstr(&sd->sd_dst_nid));
2772 * Peer doesn't have a local network. Let's see if there is
2773 * a remote network we can reach it on.
2775 return PASS_THROUGH;
2780 * Source NID not specified
2785 * Source NID not speified
2786 * Remote destination
2789 * In both of these cases if we're sending a response, ACK or REPLY, then
2790 * we need to send to the destination NID provided.
2792 * In the remote case let's deal with MR routers.
2797 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2800 struct lnet_peer *gw_peer = NULL;
2801 struct lnet_peer_ni *gw_lpni = NULL;
2804 * handle sending a response to a remote peer here so we don't
2805 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2807 if (sd->sd_send_case & REMOTE_DST &&
2808 sd->sd_send_case & SND_RESP) {
2809 struct lnet_peer_ni *gw;
2810 struct lnet_peer *gw_peer;
2812 rc = lnet_handle_find_routed_path(
2813 sd, &sd->sd_dst_nid, &gw, &gw_peer);
2815 CERROR("Can't send response to %s. No route available\n",
2816 libcfs_nidstr(&sd->sd_dst_nid));
2817 return -EHOSTUNREACH;
2818 } else if (rc > 0) {
2822 sd->sd_best_lpni = gw;
2823 sd->sd_peer = gw_peer;
2825 return lnet_handle_send(sd);
2829 * Even though the NID for the peer might not be on a local network,
2830 * since the peer is MR there could be other interfaces on the
2831 * local network. In that case we'd still like to prefer the local
2832 * network over the routed network. If we're unable to do that
2833 * then we select the best router among the different routed networks,
2834 * and if the router is MR then we can deal with it as such.
2836 rc = lnet_handle_any_mr_dsta(sd);
2837 if (rc != PASS_THROUGH)
2841 * Now that we must route to the destination, we must consider the
2842 * MR case, where the destination has multiple interfaces, some of
2843 * which we can route to and others we do not. For this reason we
2844 * need to select the destination which we can route to and if
2845 * there are multiple, we need to round robin.
2847 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2848 &gw_lpni, &gw_peer);
2852 sd->sd_send_case &= ~LOCAL_DST;
2853 sd->sd_send_case |= REMOTE_DST;
2855 sd->sd_peer = gw_peer;
2856 sd->sd_best_lpni = gw_lpni;
2858 return lnet_handle_send(sd);
2862 * Source not specified
2863 * Remote destination
2866 * Must send to the specified peer NID using the same source NID that
2867 * we've used before. If it's the first time to talk to that peer then
2868 * find the source NI and assign it as preferred to that peer
2871 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2874 struct lnet_peer_ni *gw_lpni = NULL;
2875 struct lnet_peer *gw_peer = NULL;
2878 * Let's see if we have a preferred NI to talk to this NMR peer
2880 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2884 * find the router and that'll find the best NI if we didn't find
2887 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid, &gw_lpni,
2893 * set the best_ni we've chosen as the preferred one for
2896 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2898 /* we'll be sending to the gw */
2899 sd->sd_best_lpni = gw_lpni;
2900 sd->sd_peer = gw_peer;
2902 return lnet_handle_send(sd);
2906 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2909 * turn off the SND_RESP bit.
2910 * It will be checked in the case handling
2912 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2914 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2915 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2916 (send_case & SRC_SPEC) ? libcfs_nidstr(&sd->sd_src_nid) : "",
2917 (send_case & MR_DST) ? "MR: " : "NMR: ",
2918 libcfs_nidstr(&sd->sd_dst_nid),
2919 (send_case & LOCAL_DST) ? "local" : "routed");
2921 switch (send_case) {
2923 * For all cases where the source is specified, we should always
2924 * use the destination NID, whether it's an MR destination or not,
2925 * since we're continuing a series of related messages for the
2928 case SRC_SPEC_LOCAL_NMR_DST:
2929 return lnet_handle_spec_local_nmr_dst(sd);
2930 case SRC_SPEC_LOCAL_MR_DST:
2931 return lnet_handle_spec_local_mr_dst(sd);
2932 case SRC_SPEC_ROUTER_NMR_DST:
2933 case SRC_SPEC_ROUTER_MR_DST:
2934 return lnet_handle_spec_router_dst(sd);
2935 case SRC_ANY_LOCAL_NMR_DST:
2936 return lnet_handle_any_local_nmr_dst(sd);
2937 case SRC_ANY_LOCAL_MR_DST:
2938 case SRC_ANY_ROUTER_MR_DST:
2939 return lnet_handle_any_mr_dst(sd);
2940 case SRC_ANY_ROUTER_NMR_DST:
2941 return lnet_handle_any_router_nmr_dst(sd);
2943 CERROR("Unknown send case\n");
2949 lnet_select_pathway(struct lnet_nid *src_nid,
2950 struct lnet_nid *dst_nid,
2951 struct lnet_msg *msg,
2952 struct lnet_nid *rtr_nid)
2954 struct lnet_peer_ni *lpni;
2955 struct lnet_peer *peer;
2956 struct lnet_send_data send_data;
2959 __u32 send_case = 0;
2961 bool mr_forwarding_allowed;
2963 memset(&send_data, 0, sizeof(send_data));
2966 * get an initial CPT to use for locking. The idea here is not to
2967 * serialize the calls to select_pathway, so that as many
2968 * operations can run concurrently as possible. To do that we use
2969 * the CPT where this call is being executed. Later on when we
2970 * determine the CPT to use in lnet_message_commit, we switch the
2971 * lock and check if there was any configuration change. If none,
2972 * then we proceed, if there is, then we restart the operation.
2974 cpt = lnet_net_lock_current();
2976 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2977 if (md_cpt == CFS_CPT_ANY)
2983 * If we're being asked to send to the loopback interface, there
2984 * is no need to go through any selection. We can just shortcut
2985 * the entire process and send over lolnd
2987 send_data.sd_msg = msg;
2988 send_data.sd_cpt = cpt;
2989 if (nid_is_lo0(dst_nid)) {
2990 rc = lnet_handle_lo_send(&send_data);
2991 lnet_net_unlock(cpt);
2996 * find an existing peer_ni, or create one and mark it as having been
2997 * created due to network traffic. This call will create the
2998 * peer->peer_net->peer_ni tree.
3000 lpni = lnet_peerni_by_nid_locked(dst_nid, NULL, cpt);
3002 lnet_net_unlock(cpt);
3003 return PTR_ERR(lpni);
3007 * Cache the original src_nid and rtr_nid. If we need to resend the
3008 * message then we'll need to know whether the src_nid was originally
3009 * specified for this message. If it was originally specified,
3010 * then we need to keep using the same src_nid since it's
3011 * continuing the same sequence of messages. Similarly, rtr_nid will
3012 * affect our choice of next hop.
3015 msg->msg_src_nid_param = *src_nid;
3017 msg->msg_src_nid_param = LNET_ANY_NID;
3019 msg->msg_rtr_nid_param = *rtr_nid;
3021 msg->msg_rtr_nid_param = LNET_ANY_NID;
3024 * If necessary, perform discovery on the peer that owns this peer_ni.
3025 * Note, this can result in the ownership of this peer_ni changing
3026 * to another peer object.
3028 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
3030 lnet_peer_ni_decref_locked(lpni);
3031 lnet_net_unlock(cpt);
3034 lnet_peer_ni_decref_locked(lpni);
3036 peer = lpni->lpni_peer_net->lpn_peer;
3039 * Identify the different send cases
3041 if (!src_nid || LNET_NID_IS_ANY(src_nid)) {
3042 send_case |= SRC_ANY;
3043 if (lnet_get_net_locked(LNET_NID_NET(dst_nid)))
3044 send_case |= LOCAL_DST;
3046 send_case |= REMOTE_DST;
3048 send_case |= SRC_SPEC;
3049 if (LNET_NID_NET(src_nid) == LNET_NID_NET(dst_nid))
3050 send_case |= LOCAL_DST;
3052 send_case |= REMOTE_DST;
3056 if (msg->msg_routing && (send_case & LOCAL_DST))
3059 /* Determine whether to allow MR forwarding for this message.
3060 * NB: MR forwarding is allowed if the message originator and the
3061 * destination are both MR capable, and the destination lpni that was
3062 * originally chosen by the originator is unhealthy or down.
3063 * We check the MR capability of the destination further below
3065 mr_forwarding_allowed = false;
3067 struct lnet_peer *src_lp;
3068 struct lnet_peer_ni *src_lpni;
3070 src_lpni = lnet_peerni_by_nid_locked(&msg->msg_hdr.src_nid,
3072 /* We don't fail the send if we hit any errors here. We'll just
3073 * try to send it via non-multi-rail criteria
3075 if (!IS_ERR(src_lpni)) {
3076 /* Drop ref taken by lnet_peerni_by_nid_locked() */
3077 lnet_peer_ni_decref_locked(src_lpni);
3078 src_lp = lpni->lpni_peer_net->lpn_peer;
3079 if (lnet_peer_is_multi_rail(src_lp) &&
3080 !lnet_is_peer_ni_alive(lpni))
3081 mr_forwarding_allowed = true;
3084 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3085 mr_forwarding_allowed ? "allowed" : "not allowed");
3089 * Deal with the peer as NMR in the following cases:
3090 * 1. the peer is NMR
3091 * 2. We're trying to recover a specific peer NI
3092 * 3. I'm a router sending to the final destination and MR forwarding is
3093 * not allowed for this message (as determined above).
3094 * In this case the source of the message would've
3095 * already selected the final destination so my job
3096 * is to honor the selection.
3098 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3099 (final_hop && !mr_forwarding_allowed))
3100 send_case |= NMR_DST;
3102 send_case |= MR_DST;
3104 if (lnet_msg_is_response(msg))
3105 send_case |= SND_RESP;
3107 /* assign parameters to the send_data */
3109 send_data.sd_rtr_nid = *rtr_nid;
3111 send_data.sd_rtr_nid = LNET_ANY_NID;
3113 send_data.sd_src_nid = *src_nid;
3115 send_data.sd_src_nid = LNET_ANY_NID;
3116 send_data.sd_dst_nid = *dst_nid;
3117 send_data.sd_best_lpni = lpni;
3119 * keep a pointer to the final destination in case we're going to
3120 * route, so we'll need to access it later
3122 send_data.sd_final_dst_lpni = lpni;
3123 send_data.sd_peer = peer;
3124 send_data.sd_md_cpt = md_cpt;
3125 send_data.sd_send_case = send_case;
3127 rc = lnet_handle_send_case_locked(&send_data);
3130 * Update the local cpt since send_data.sd_cpt might've been
3131 * updated as a result of calling lnet_handle_send_case_locked().
3133 cpt = send_data.sd_cpt;
3135 if (rc == REPEAT_SEND)
3138 lnet_net_unlock(cpt);
3144 lnet_send(struct lnet_nid *src_nid, struct lnet_msg *msg,
3145 struct lnet_nid *rtr_nid)
3147 struct lnet_nid *dst_nid = &msg->msg_target.nid;
3150 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3151 LASSERT(msg->msg_txpeer == NULL);
3152 LASSERT(msg->msg_txni == NULL);
3153 LASSERT(!msg->msg_sending);
3154 LASSERT(!msg->msg_target_is_router);
3155 LASSERT(!msg->msg_receiving);
3157 msg->msg_sending = 1;
3159 LASSERT(!msg->msg_tx_committed);
3161 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3163 if (rc == -EHOSTUNREACH)
3164 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3166 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3170 if (rc == LNET_CREDIT_OK)
3171 lnet_ni_send(msg->msg_txni, msg);
3173 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3177 enum lnet_mt_event_type {
3178 MT_TYPE_LOCAL_NI = 0,
3182 struct lnet_mt_event_info {
3183 enum lnet_mt_event_type mt_type;
3184 struct lnet_nid mt_nid;
3187 /* called with res_lock held */
3189 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3191 struct lnet_rsp_tracker *rspt;
3194 * msg has a refcount on the MD so the MD is not going away.
3195 * The rspt queue for the cpt is protected by
3196 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3198 if (!md->md_rspt_ptr)
3201 rspt = md->md_rspt_ptr;
3204 LASSERT(rspt->rspt_cpt == cpt);
3206 md->md_rspt_ptr = NULL;
3208 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3210 * The monitor thread has invalidated this handle because the
3211 * response timed out, but it failed to lookup the MD. That
3212 * means this response tracker is on the zombie list. We can
3213 * safely remove it under the resource lock (held by caller) and
3214 * free the response tracker block.
3216 list_del(&rspt->rspt_on_list);
3217 lnet_rspt_free(rspt, cpt);
3220 * invalidate the handle to indicate that a response has been
3221 * received, which will then lead the monitor thread to clean up
3224 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3229 lnet_clean_zombie_rstqs(void)
3231 struct lnet_rsp_tracker *rspt, *tmp;
3234 cfs_cpt_for_each(i, lnet_cpt_table()) {
3235 list_for_each_entry_safe(rspt, tmp,
3236 the_lnet.ln_mt_zombie_rstqs[i],
3238 list_del(&rspt->rspt_on_list);
3239 lnet_rspt_free(rspt, i);
3243 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3247 lnet_finalize_expired_responses(void)
3249 struct lnet_libmd *md;
3250 struct lnet_rsp_tracker *rspt, *tmp;
3254 if (the_lnet.ln_mt_rstq == NULL)
3257 cfs_cpt_for_each(i, lnet_cpt_table()) {
3258 LIST_HEAD(local_queue);
3261 if (!the_lnet.ln_mt_rstq[i]) {
3265 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3270 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3272 * The rspt mdh will be invalidated when a response
3273 * is received or whenever we want to discard the
3274 * block the monitor thread will walk the queue
3275 * and clean up any rsts with an invalid mdh.
3276 * The monitor thread will walk the queue until
3277 * the first unexpired rspt block. This means that
3278 * some rspt blocks which received their
3279 * corresponding responses will linger in the
3280 * queue until they are cleaned up eventually.
3283 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3285 list_del(&rspt->rspt_on_list);
3286 lnet_rspt_free(rspt, i);
3290 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3291 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3292 struct lnet_peer_ni *lpni;
3293 struct lnet_nid nid;
3295 md = lnet_handle2md(&rspt->rspt_mdh);
3297 /* MD has been queued for unlink, but
3298 * rspt hasn't been detached (Note we've
3299 * checked above that the rspt_mdh is
3300 * valid). Since we cannot lookup the MD
3301 * we're unable to detach the rspt
3302 * ourselves. Thus, move the rspt to the
3303 * zombie list where we'll wait for
3305 * 1. The remaining operations on the
3306 * MD to complete. In this case the
3307 * final operation will result in
3308 * lnet_msg_detach_md()->
3309 * lnet_detach_rsp_tracker() where
3310 * we will clean up this response
3312 * 2. LNet to shutdown. In this case
3313 * we'll wait until after all LND Nets
3314 * have shutdown and then we can
3315 * safely free any remaining response
3316 * tracker blocks on the zombie list.
3317 * Note: We need to hold the resource
3318 * lock when adding to the zombie list
3319 * because we may have concurrent access
3320 * with lnet_detach_rsp_tracker().
3322 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3323 list_move(&rspt->rspt_on_list,
3324 the_lnet.ln_mt_zombie_rstqs[i]);
3328 LASSERT(md->md_rspt_ptr == rspt);
3329 md->md_rspt_ptr = NULL;
3332 LNetMDUnlink(rspt->rspt_mdh);
3334 nid = rspt->rspt_next_hop_nid;
3336 list_del(&rspt->rspt_on_list);
3337 lnet_rspt_free(rspt, i);
3339 /* If we're shutting down we just want to clean
3340 * up the rspt blocks
3342 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3346 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3350 "Response timeout: md = %p: nid = %s\n",
3351 md, libcfs_nidstr(&nid));
3354 * If there is a timeout on the response
3355 * from the next hop decrement its health
3356 * value so that we don't use it
3359 lpni = lnet_peer_ni_find_locked(&nid);
3361 lnet_handle_remote_failure_locked(lpni);
3362 lnet_peer_ni_decref_locked(lpni);
3371 if (!list_empty(&local_queue)) {
3373 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3380 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3382 struct lnet_msg *msg;
3384 while (!list_empty(resendq)) {
3385 struct lnet_peer_ni *lpni;
3387 msg = list_entry(resendq->next, struct lnet_msg,
3390 list_del_init(&msg->msg_list);
3392 lpni = lnet_peer_ni_find_locked(&msg->msg_hdr.dest_nid);
3394 lnet_net_unlock(cpt);
3395 CERROR("Expected that a peer is already created for %s\n",
3396 libcfs_nidstr(&msg->msg_hdr.dest_nid));
3397 msg->msg_no_resend = true;
3398 lnet_finalize(msg, -EFAULT);
3403 lnet_peer_ni_decref_locked(lpni);
3405 lnet_net_unlock(cpt);
3406 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3407 libcfs_nidstr(&msg->msg_src_nid_param),
3408 libcfs_idstr(&msg->msg_target),
3409 lnet_msgtyp2str(msg->msg_type),
3411 msg->msg_retry_count);
3412 rc = lnet_send(&msg->msg_src_nid_param, msg,
3413 &msg->msg_rtr_nid_param);
3415 CERROR("Error sending %s to %s: %d\n",
3416 lnet_msgtyp2str(msg->msg_type),
3417 libcfs_idstr(&msg->msg_target), rc);
3418 msg->msg_no_resend = true;
3419 lnet_finalize(msg, rc);
3423 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3429 lnet_resend_pending_msgs(void)
3433 cfs_cpt_for_each(i, lnet_cpt_table()) {
3435 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3440 /* called with cpt and ni_lock held */
3442 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3444 struct lnet_handle_md recovery_mdh;
3446 LNetInvalidateMDHandle(&recovery_mdh);
3448 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3450 recovery_mdh = ni->ni_ping_mdh;
3451 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3454 lnet_net_unlock(cpt);
3455 if (!LNetMDHandleIsInvalid(recovery_mdh))
3456 LNetMDUnlink(recovery_mdh);
3462 lnet_recover_local_nis(void)
3464 struct lnet_mt_event_info *ev_info;
3465 LIST_HEAD(processed_list);
3466 LIST_HEAD(local_queue);
3467 struct lnet_handle_md mdh;
3468 struct lnet_ni *tmp;
3470 struct lnet_nid nid;
3476 * splice the recovery queue on a local queue. We will iterate
3477 * through the local queue and update it as needed. Once we're
3478 * done with the traversal, we'll splice the local queue back on
3479 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3480 * will be traversed in the next iteration.
3483 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3487 now = ktime_get_seconds();
3489 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3491 * if an NI is being deleted or it is now healthy, there
3492 * is no need to keep it around in the recovery queue.
3493 * The monitor thread is the only thread responsible for
3494 * removing the NI from the recovery queue.
3495 * Multiple threads can be adding NIs to the recovery
3498 healthv = atomic_read(&ni->ni_healthv);
3502 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3503 healthv == LNET_MAX_HEALTH_VALUE) {
3504 list_del_init(&ni->ni_recovery);
3505 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3507 lnet_ni_decref_locked(ni, 0);
3513 * if the local NI failed recovery we must unlink the md.
3514 * But we want to keep the local_ni on the recovery queue
3515 * so we can continue the attempts to recover it.
3517 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3518 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3519 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3525 if (now < ni->ni_next_ping) {
3532 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3533 libcfs_nidstr(&ni->ni_nid));
3536 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3537 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3540 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3542 CERROR("out of memory. Can't recover %s\n",
3543 libcfs_nidstr(&ni->ni_nid));
3545 ni->ni_recovery_state &=
3546 ~LNET_NI_RECOVERY_PENDING;
3551 mdh = ni->ni_ping_mdh;
3553 * Invalidate the ni mdh in case it's deleted.
3554 * We'll unlink the mdh in this case below.
3556 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3560 * remove the NI from the local queue and drop the
3561 * reference count to it while we're recovering
3562 * it. The reason for that, is that the NI could
3563 * be deleted, and the way the code is structured
3564 * is if we don't drop the NI, then the deletion
3565 * code will enter a loop waiting for the
3566 * reference count to be removed while holding the
3567 * ln_mutex_lock(). When we look up the peer to
3568 * send to in lnet_select_pathway() we will try to
3569 * lock the ln_mutex_lock() as well, leading to
3570 * a deadlock. By dropping the refcount and
3571 * removing it from the list, we allow for the NI
3572 * to be removed, then we use the cached NID to
3573 * look it up again. If it's gone, then we just
3574 * continue examining the rest of the queue.
3577 list_del_init(&ni->ni_recovery);
3578 lnet_ni_decref_locked(ni, 0);
3581 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3582 ev_info->mt_nid = nid;
3583 rc = lnet_send_ping(&nid, &mdh, LNET_INTERFACES_MIN,
3584 ev_info, the_lnet.ln_mt_handler,
3586 /* lookup the nid again */
3588 ni = lnet_nid_to_ni_locked(&nid, 0);
3591 * the NI has been deleted when we dropped
3598 ni->ni_ping_count++;
3600 ni->ni_ping_mdh = mdh;
3601 lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3606 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3615 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3616 * reexamined in the next iteration.
3618 list_splice_init(&processed_list, &local_queue);
3620 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3625 lnet_resendqs_create(void)
3627 struct list_head **resendqs;
3628 resendqs = lnet_create_array_of_queues();
3633 lnet_net_lock(LNET_LOCK_EX);
3634 the_lnet.ln_mt_resendqs = resendqs;
3635 lnet_net_unlock(LNET_LOCK_EX);
3641 lnet_clean_local_ni_recoveryq(void)
3645 /* This is only called when the monitor thread has stopped */
3648 while ((ni = list_first_entry_or_null(&the_lnet.ln_mt_localNIRecovq,
3650 ni_recovery)) != NULL) {
3651 list_del_init(&ni->ni_recovery);
3653 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3655 lnet_ni_decref_locked(ni, 0);
3662 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3665 struct lnet_handle_md recovery_mdh;
3667 LNetInvalidateMDHandle(&recovery_mdh);
3669 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3670 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3671 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3673 spin_unlock(&lpni->lpni_lock);
3674 lnet_net_unlock(cpt);
3675 if (!LNetMDHandleIsInvalid(recovery_mdh))
3676 LNetMDUnlink(recovery_mdh);
3678 spin_lock(&lpni->lpni_lock);
3682 lnet_clean_peer_ni_recoveryq(void)
3684 struct lnet_peer_ni *lpni, *tmp;
3686 lnet_net_lock(LNET_LOCK_EX);
3688 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3690 list_del_init(&lpni->lpni_recovery);
3691 spin_lock(&lpni->lpni_lock);
3692 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3693 spin_unlock(&lpni->lpni_lock);
3694 lnet_peer_ni_decref_locked(lpni);
3697 lnet_net_unlock(LNET_LOCK_EX);
3701 lnet_clean_resendqs(void)
3703 struct lnet_msg *msg, *tmp;
3707 cfs_cpt_for_each(i, lnet_cpt_table()) {
3709 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3711 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3712 list_del_init(&msg->msg_list);
3713 msg->msg_no_resend = true;
3714 lnet_finalize(msg, -ESHUTDOWN);
3718 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3722 lnet_recover_peer_nis(void)
3724 struct lnet_mt_event_info *ev_info;
3725 LIST_HEAD(processed_list);
3726 LIST_HEAD(local_queue);
3727 struct lnet_handle_md mdh;
3728 struct lnet_peer_ni *lpni;
3729 struct lnet_peer_ni *tmp;
3730 struct lnet_nid nid;
3736 * Always use cpt 0 for locking across all interactions with
3737 * ln_mt_peerNIRecovq
3740 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3744 now = ktime_get_seconds();
3746 list_for_each_entry_safe(lpni, tmp, &local_queue,
3749 * The same protection strategy is used here as is in the
3750 * local recovery case.
3753 healthv = atomic_read(&lpni->lpni_healthv);
3754 spin_lock(&lpni->lpni_lock);
3755 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3756 healthv == LNET_MAX_HEALTH_VALUE) {
3757 list_del_init(&lpni->lpni_recovery);
3758 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3759 spin_unlock(&lpni->lpni_lock);
3760 lnet_peer_ni_decref_locked(lpni);
3766 * If the peer NI has failed recovery we must unlink the
3767 * md. But we want to keep the peer ni on the recovery
3768 * queue so we can try to continue recovering it
3770 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3771 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3772 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3775 spin_unlock(&lpni->lpni_lock);
3777 if (now < lpni->lpni_next_ping) {
3785 * NOTE: we're racing with peer deletion from user space.
3786 * It's possible that a peer is deleted after we check its
3787 * state. In this case the recovery can create a new peer
3789 spin_lock(&lpni->lpni_lock);
3790 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3791 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3792 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3793 spin_unlock(&lpni->lpni_lock);
3795 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3797 CERROR("out of memory. Can't recover %s\n",
3798 libcfs_nidstr(&lpni->lpni_nid));
3799 spin_lock(&lpni->lpni_lock);
3800 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3801 spin_unlock(&lpni->lpni_lock);
3805 /* look at the comments in lnet_recover_local_nis() */
3806 mdh = lpni->lpni_recovery_ping_mdh;
3807 nid = lpni->lpni_nid;
3808 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3810 list_del_init(&lpni->lpni_recovery);
3811 lnet_peer_ni_decref_locked(lpni);
3814 ev_info->mt_type = MT_TYPE_PEER_NI;
3815 ev_info->mt_nid = nid;
3816 rc = lnet_send_ping(&nid, &mdh, LNET_INTERFACES_MIN,
3817 ev_info, the_lnet.ln_mt_handler,
3821 * lnet_peer_ni_find_locked() grabs a refcount for
3822 * us. No need to take it explicitly.
3824 lpni = lnet_peer_ni_find_locked(&nid);
3831 lpni->lpni_ping_count++;
3833 lpni->lpni_recovery_ping_mdh = mdh;
3835 lnet_peer_ni_add_to_recoveryq_locked(lpni,
3839 spin_lock(&lpni->lpni_lock);
3840 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3841 spin_unlock(&lpni->lpni_lock);
3844 /* Drop the ref taken by lnet_peer_ni_find_locked() */
3845 lnet_peer_ni_decref_locked(lpni);
3848 spin_unlock(&lpni->lpni_lock);
3851 list_splice_init(&processed_list, &local_queue);
3853 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3858 lnet_monitor_thread(void *arg)
3860 time64_t rsp_timeout = 0;
3863 wait_for_completion(&the_lnet.ln_started);
3865 * The monitor thread takes care of the following:
3866 * 1. Checks the aliveness of routers
3867 * 2. Checks if there are messages on the resend queue to resend
3869 * 3. Check if there are any NIs on the local recovery queue and
3871 * 4. Checks if there are any NIs on the remote recovery queue
3874 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3875 now = ktime_get_real_seconds();
3877 if (lnet_router_checker_active())
3878 lnet_check_routers();
3880 lnet_resend_pending_msgs();
3882 if (now >= rsp_timeout) {
3883 lnet_finalize_expired_responses();
3884 rsp_timeout = now + (lnet_transaction_timeout / 2);
3887 lnet_recover_local_nis();
3888 lnet_recover_peer_nis();
3891 * TODO do we need to check if we should sleep without
3892 * timeout? Technically, an active system will always
3893 * have messages in flight so this check will always
3894 * evaluate to false. And on an idle system do we care
3895 * if we wake up every 1 second? Although, we've seen
3896 * cases where we get a complaint that an idle thread
3897 * is waking up unnecessarily.
3899 wait_for_completion_interruptible_timeout(
3900 &the_lnet.ln_mt_wait_complete,
3901 cfs_time_seconds(1));
3902 /* Must re-init the completion before testing anything,
3903 * including ln_mt_state.
3905 reinit_completion(&the_lnet.ln_mt_wait_complete);
3909 lnet_net_lock(LNET_LOCK_EX);
3910 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3911 lnet_net_unlock(LNET_LOCK_EX);
3913 /* signal that the monitor thread is exiting */
3914 up(&the_lnet.ln_mt_signal);
3922 * Returns == 0 if success
3923 * Returns > 0 if LNetMDBind or prior fails
3924 * Returns < 0 if LNetGet fails
3927 lnet_send_ping(struct lnet_nid *dest_nid,
3928 struct lnet_handle_md *mdh, int nnis,
3929 void *user_data, lnet_handler_t handler, bool recovery)
3931 struct lnet_md md = { NULL };
3932 struct lnet_processid id;
3933 struct lnet_ping_buffer *pbuf;
3936 if (LNET_NID_IS_ANY(dest_nid)) {
3941 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3947 /* initialize md content */
3948 md.start = &pbuf->pb_info;
3949 md.length = LNET_PING_INFO_SIZE(nnis);
3950 md.threshold = 2; /* GET/REPLY */
3952 md.options = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3953 md.user_ptr = user_data;
3954 md.handler = handler;
3956 rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3958 lnet_ping_buffer_decref(pbuf);
3959 CERROR("Can't bind MD: %d\n", rc);
3960 rc = -rc; /* change the rc to positive */
3963 id.pid = LNET_PID_LUSTRE;
3966 rc = LNetGet(NULL, *mdh, &id,
3967 LNET_RESERVED_PORTAL,
3968 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3971 goto fail_unlink_md;
3977 LNetInvalidateMDHandle(mdh);
3983 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3984 int status, bool send, bool unlink_event)
3986 struct lnet_nid *nid = &ev_info->mt_nid;
3988 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3992 ni = lnet_nid_to_ni_locked(nid, 0);
3998 if (!send || (send && status != 0))
3999 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
4001 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
4006 CERROR("local NI (%s) recovery failed with %d\n",
4007 libcfs_nidstr(nid), status);
4011 * need to increment healthv for the ni here, because in
4012 * the lnet_finalize() path we don't have access to this
4013 * NI. And in order to get access to it, we'll need to
4014 * carry forward too much information.
4015 * In the peer case, it'll naturally be incremented
4018 lnet_inc_healthv(&ni->ni_healthv,
4019 lnet_health_sensitivity);
4021 struct lnet_peer_ni *lpni;
4024 cpt = lnet_net_lock_current();
4025 lpni = lnet_peer_ni_find_locked(nid);
4027 lnet_net_unlock(cpt);
4030 spin_lock(&lpni->lpni_lock);
4031 if (!send || (send && status != 0))
4032 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
4034 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
4035 spin_unlock(&lpni->lpni_lock);
4036 lnet_peer_ni_decref_locked(lpni);
4037 lnet_net_unlock(cpt);
4040 CERROR("peer NI (%s) recovery failed with %d\n",
4041 libcfs_nidstr(nid), status);
4046 lnet_mt_event_handler(struct lnet_event *event)
4048 struct lnet_mt_event_info *ev_info = event->md_user_ptr;
4049 struct lnet_ping_buffer *pbuf;
4051 /* TODO: remove assert */
4052 LASSERT(event->type == LNET_EVENT_REPLY ||
4053 event->type == LNET_EVENT_SEND ||
4054 event->type == LNET_EVENT_UNLINK);
4056 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
4059 switch (event->type) {
4060 case LNET_EVENT_UNLINK:
4061 CDEBUG(D_NET, "%s recovery ping unlinked\n",
4062 libcfs_nidstr(&ev_info->mt_nid));
4064 case LNET_EVENT_REPLY:
4065 lnet_handle_recovery_reply(ev_info, event->status, false,
4066 event->type == LNET_EVENT_UNLINK);
4068 case LNET_EVENT_SEND:
4069 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4070 libcfs_nidstr(&ev_info->mt_nid),
4071 (event->status) ? "unsuccessfully" :
4072 "successfully", event->status);
4073 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4076 CERROR("Unexpected event: %d\n", event->type);
4079 if (event->unlinked) {
4080 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4081 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4082 lnet_ping_buffer_decref(pbuf);
4087 lnet_rsp_tracker_create(void)
4089 struct list_head **rstqs;
4090 rstqs = lnet_create_array_of_queues();
4095 the_lnet.ln_mt_rstq = rstqs;
4101 lnet_rsp_tracker_clean(void)
4103 lnet_finalize_expired_responses();
4105 cfs_percpt_free(the_lnet.ln_mt_rstq);
4106 the_lnet.ln_mt_rstq = NULL;
4109 int lnet_monitor_thr_start(void)
4112 struct task_struct *task;
4114 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4117 rc = lnet_resendqs_create();
4121 rc = lnet_rsp_tracker_create();
4125 sema_init(&the_lnet.ln_mt_signal, 0);
4127 lnet_net_lock(LNET_LOCK_EX);
4128 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4129 lnet_net_unlock(LNET_LOCK_EX);
4130 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4133 CERROR("Can't start monitor thread: %d\n", rc);
4140 lnet_net_lock(LNET_LOCK_EX);
4141 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4142 lnet_net_unlock(LNET_LOCK_EX);
4143 /* block until event callback signals exit */
4144 down(&the_lnet.ln_mt_signal);
4146 lnet_net_lock(LNET_LOCK_EX);
4147 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4148 lnet_net_unlock(LNET_LOCK_EX);
4149 lnet_rsp_tracker_clean();
4150 lnet_clean_local_ni_recoveryq();
4151 lnet_clean_peer_ni_recoveryq();
4152 lnet_clean_resendqs();
4153 the_lnet.ln_mt_handler = NULL;
4156 lnet_rsp_tracker_clean();
4157 lnet_clean_local_ni_recoveryq();
4158 lnet_clean_peer_ni_recoveryq();
4159 lnet_clean_resendqs();
4163 void lnet_monitor_thr_stop(void)
4165 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4168 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4169 lnet_net_lock(LNET_LOCK_EX);
4170 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4171 lnet_net_unlock(LNET_LOCK_EX);
4173 /* tell the monitor thread that we're shutting down */
4174 complete(&the_lnet.ln_mt_wait_complete);
4176 /* block until monitor thread signals that it's done */
4177 mutex_unlock(&the_lnet.ln_api_mutex);
4178 down(&the_lnet.ln_mt_signal);
4179 mutex_lock(&the_lnet.ln_api_mutex);
4180 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4182 /* perform cleanup tasks */
4183 lnet_rsp_tracker_clean();
4184 lnet_clean_local_ni_recoveryq();
4185 lnet_clean_peer_ni_recoveryq();
4186 lnet_clean_resendqs();
4190 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4194 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4195 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4196 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4197 lnet_net_unlock(cpt);
4199 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4203 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4205 struct lnet_hdr *hdr = &msg->msg_hdr;
4207 if (msg->msg_wanted != 0)
4208 lnet_setpayloadbuffer(msg);
4210 lnet_build_msg_event(msg, LNET_EVENT_PUT);
4212 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
4213 * it back into the ACK during lnet_finalize() */
4214 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4215 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4217 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4218 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4222 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4224 struct lnet_hdr *hdr = &msg->msg_hdr;
4225 struct lnet_match_info info;
4229 /* Convert put fields to host byte order */
4230 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4231 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
4232 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
4234 /* Primary peer NID. */
4235 info.mi_id.nid = msg->msg_initiator;
4236 info.mi_id.pid = hdr->src_pid;
4237 info.mi_opc = LNET_MD_OP_PUT;
4238 info.mi_portal = hdr->msg.put.ptl_index;
4239 info.mi_rlength = hdr->payload_length;
4240 info.mi_roffset = hdr->msg.put.offset;
4241 info.mi_mbits = hdr->msg.put.match_bits;
4242 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4244 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4245 ready_delay = msg->msg_rx_ready_delay;
4248 rc = lnet_ptl_match_md(&info, msg);
4253 case LNET_MATCHMD_OK:
4254 lnet_recv_put(ni, msg);
4257 case LNET_MATCHMD_NONE:
4259 /* no eager_recv or has already called it, should
4260 * have been attached on delayed list */
4263 rc = lnet_ni_eager_recv(ni, msg);
4270 case LNET_MATCHMD_DROP:
4271 CNETERR("Dropping PUT from %s portal %d match %llu"
4272 " offset %d length %d: %d\n",
4273 libcfs_idstr(&info.mi_id), info.mi_portal,
4274 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4276 return -ENOENT; /* -ve: OK but no match */
4281 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4283 struct lnet_match_info info;
4284 struct lnet_hdr *hdr = &msg->msg_hdr;
4285 struct lnet_processid source_id;
4286 struct lnet_handle_wire reply_wmd;
4289 /* Convert get fields to host byte order */
4290 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
4291 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
4292 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
4293 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
4295 source_id.nid = hdr->src_nid;
4296 source_id.pid = hdr->src_pid;
4297 /* Primary peer NID */
4298 info.mi_id.nid = msg->msg_initiator;
4299 info.mi_id.pid = hdr->src_pid;
4300 info.mi_opc = LNET_MD_OP_GET;
4301 info.mi_portal = hdr->msg.get.ptl_index;
4302 info.mi_rlength = hdr->msg.get.sink_length;
4303 info.mi_roffset = hdr->msg.get.src_offset;
4304 info.mi_mbits = hdr->msg.get.match_bits;
4305 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4307 rc = lnet_ptl_match_md(&info, msg);
4308 if (rc == LNET_MATCHMD_DROP) {
4309 CNETERR("Dropping GET from %s portal %d match %llu"
4310 " offset %d length %d\n",
4311 libcfs_idstr(&info.mi_id), info.mi_portal,
4312 info.mi_mbits, info.mi_roffset, info.mi_rlength);
4313 return -ENOENT; /* -ve: OK but no match */
4316 LASSERT(rc == LNET_MATCHMD_OK);
4318 lnet_build_msg_event(msg, LNET_EVENT_GET);
4320 reply_wmd = hdr->msg.get.return_wmd;
4322 lnet_prep_send(msg, LNET_MSG_REPLY, &source_id,
4323 msg->msg_offset, msg->msg_wanted);
4325 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4328 /* The LND completes the REPLY from her recv procedure */
4329 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4330 msg->msg_offset, msg->msg_len, msg->msg_len);
4334 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4335 msg->msg_receiving = 0;
4337 rc = lnet_send(&ni->ni_nid, msg, &msg->msg_from);
4339 /* didn't get as far as lnet_ni_send() */
4340 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4341 libcfs_nidstr(&ni->ni_nid),
4342 libcfs_idstr(&info.mi_id), rc);
4344 lnet_finalize(msg, rc);
4351 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4353 void *private = msg->msg_private;
4354 struct lnet_hdr *hdr = &msg->msg_hdr;
4355 struct lnet_processid src = {};
4356 struct lnet_libmd *md;
4357 unsigned int rlength;
4358 unsigned int mlength;
4361 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4364 src.nid = hdr->src_nid;
4365 src.pid = hdr->src_pid;
4367 /* NB handles only looked up by creator (no flips) */
4368 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4369 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4370 CNETERR("%s: Dropping REPLY from %s for %s "
4372 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4373 (md == NULL) ? "invalid" : "inactive",
4374 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4375 hdr->msg.reply.dst_wmd.wh_object_cookie);
4376 if (md != NULL && md->md_me != NULL)
4377 CERROR("REPLY MD also attached to portal %d\n",
4378 md->md_me->me_portal);
4380 lnet_res_unlock(cpt);
4381 return -ENOENT; /* -ve: OK but no match */
4384 LASSERT(md->md_offset == 0);
4386 rlength = hdr->payload_length;
4387 mlength = min(rlength, md->md_length);
4389 if (mlength < rlength &&
4390 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4391 CNETERR("%s: Dropping REPLY from %s length %d "
4392 "for MD %#llx would overflow (%d)\n",
4393 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4394 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4396 lnet_res_unlock(cpt);
4397 return -ENOENT; /* -ve: OK but no match */
4400 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4401 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4402 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4404 lnet_msg_attach_md(msg, md, 0, mlength);
4407 lnet_setpayloadbuffer(msg);
4409 lnet_res_unlock(cpt);
4411 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4413 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4418 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4420 struct lnet_hdr *hdr = &msg->msg_hdr;
4421 struct lnet_processid src = {};
4422 struct lnet_libmd *md;
4425 src.nid = hdr->src_nid;
4426 src.pid = hdr->src_pid;
4428 /* Convert ack fields to host byte order */
4429 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4430 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4432 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4435 /* NB handles only looked up by creator (no flips) */
4436 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4437 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4438 /* Don't moan; this is expected */
4440 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4441 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4442 (md == NULL) ? "invalid" : "inactive",
4443 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4444 hdr->msg.ack.dst_wmd.wh_object_cookie);
4445 if (md != NULL && md->md_me != NULL)
4446 CERROR("Source MD also attached to portal %d\n",
4447 md->md_me->me_portal);
4449 lnet_res_unlock(cpt);
4450 return -ENOENT; /* -ve! */
4453 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4454 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4455 hdr->msg.ack.dst_wmd.wh_object_cookie);
4457 lnet_msg_attach_md(msg, md, 0, 0);
4459 lnet_res_unlock(cpt);
4461 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4463 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4468 * \retval LNET_CREDIT_OK If \a msg is forwarded
4469 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4470 * \retval -ve error code
4473 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4477 if (!the_lnet.ln_routing)
4480 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4481 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4482 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4483 msg->msg_rx_ready_delay = 1;
4485 lnet_net_unlock(msg->msg_rx_cpt);
4486 rc = lnet_ni_eager_recv(ni, msg);
4487 lnet_net_lock(msg->msg_rx_cpt);
4492 rc = lnet_post_routed_recv_locked(msg, 0);
4497 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4501 switch (msg->msg_type) {
4503 rc = lnet_parse_ack(ni, msg);
4506 rc = lnet_parse_put(ni, msg);
4509 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4511 case LNET_MSG_REPLY:
4512 rc = lnet_parse_reply(ni, msg);
4514 default: /* prevent an unused label if !kernel */
4519 LASSERT(rc == 0 || rc == -ENOENT);
4524 lnet_msgtyp2str (int type)
4533 case LNET_MSG_REPLY:
4535 case LNET_MSG_HELLO:
4538 return ("<UNKNOWN>");
4541 EXPORT_SYMBOL(lnet_msgtyp2str);
4544 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr,
4545 struct lnet_nid *from_nid, void *private, int rdma_req)
4547 struct lnet_peer_ni *lpni;
4548 struct lnet_msg *msg;
4549 __u32 payload_length;
4550 lnet_pid_t dest_pid;
4551 struct lnet_nid dest_nid;
4552 struct lnet_nid src_nid;
4558 time64_t now = ktime_get_seconds();
4560 LASSERT (!in_interrupt ());
4563 src_nid = hdr->src_nid;
4564 dest_nid = hdr->dest_nid;
4565 dest_pid = hdr->dest_pid;
4566 payload_length = hdr->payload_length;
4568 for_me = nid_same(&ni->ni_nid, &dest_nid);
4569 cpt = lnet_nid2cpt(from_nid, ni);
4571 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4572 libcfs_nidstr(&dest_nid),
4573 libcfs_nidstr(&ni->ni_nid),
4574 libcfs_nidstr(&src_nid),
4575 lnet_msgtyp2str(type),
4576 (for_me) ? "for me" : "routed");
4581 if (payload_length > 0) {
4582 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4583 libcfs_nidstr(from_nid),
4584 libcfs_nidstr(&src_nid),
4585 lnet_msgtyp2str(type), payload_length);
4591 case LNET_MSG_REPLY:
4592 if (payload_length >
4593 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4594 CERROR("%s, src %s: bad %s payload %d "
4595 "(%d max expected)\n",
4596 libcfs_nidstr(from_nid),
4597 libcfs_nidstr(&src_nid),
4598 lnet_msgtyp2str(type),
4600 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4606 CERROR("%s, src %s: Bad message type 0x%x\n",
4607 libcfs_nidstr(from_nid),
4608 libcfs_nidstr(&src_nid), type);
4612 /* Only update net_last_alive for incoming GETs on the reserved portal
4613 * (i.e. incoming lnet/discovery pings).
4614 * This avoids situations where the router's own traffic results in NI
4617 if (the_lnet.ln_routing && type == LNET_MSG_GET &&
4618 hdr->msg.get.ptl_index == LNET_RESERVED_PORTAL &&
4619 !lnet_islocalnid(&src_nid) &&
4620 ni->ni_net->net_last_alive != now) {
4622 spin_lock(&ni->ni_net->net_lock);
4623 ni->ni_net->net_last_alive = now;
4624 spin_unlock(&ni->ni_net->net_lock);
4625 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4630 lnet_push_update_to_peers(1);
4632 /* Regard a bad destination NID as a protocol error. Senders should
4633 * know what they're doing; if they don't they're misconfigured, buggy
4634 * or malicious so we chop them off at the knees :) */
4637 if (LNET_NID_NET(&dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4638 /* should have gone direct */
4639 CERROR("%s, src %s: Bad dest nid %s "
4640 "(should have been sent direct)\n",
4641 libcfs_nidstr(from_nid),
4642 libcfs_nidstr(&src_nid),
4643 libcfs_nidstr(&dest_nid));
4647 if (lnet_islocalnid(&dest_nid)) {
4648 /* dest is another local NI; sender should have used
4649 * this node's NID on its own network */
4650 CERROR("%s, src %s: Bad dest nid %s "
4651 "(it's my nid but on a different network)\n",
4652 libcfs_nidstr(from_nid),
4653 libcfs_nidstr(&src_nid),
4654 libcfs_nidstr(&dest_nid));
4658 if (rdma_req && type == LNET_MSG_GET) {
4659 CERROR("%s, src %s: Bad optimized GET for %s "
4660 "(final destination must be me)\n",
4661 libcfs_nidstr(from_nid),
4662 libcfs_nidstr(&src_nid),
4663 libcfs_nidstr(&dest_nid));
4667 if (!the_lnet.ln_routing) {
4668 CERROR("%s, src %s: Dropping message for %s "
4669 "(routing not enabled)\n",
4670 libcfs_nidstr(from_nid),
4671 libcfs_nidstr(&src_nid),
4672 libcfs_nidstr(&dest_nid));
4677 /* Message looks OK; we're not going to return an error, so we MUST
4678 * call back lnd_recv() come what may... */
4680 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4681 fail_peer(&src_nid, 0)) { /* shall we now? */
4682 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4683 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4684 lnet_msgtyp2str(type));
4688 if (!list_empty(&the_lnet.ln_drop_rules) &&
4689 lnet_drop_rule_match(hdr, &ni->ni_nid, NULL)) {
4691 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4692 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4693 libcfs_nidstr(&dest_nid), lnet_msgtyp2str(type));
4697 msg = lnet_msg_alloc();
4699 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4700 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4701 lnet_msgtyp2str(type));
4705 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4706 * pointers NULL etc */
4708 msg->msg_type = type;
4709 msg->msg_private = private;
4710 msg->msg_receiving = 1;
4711 msg->msg_rdma_get = rdma_req;
4712 msg->msg_len = msg->msg_wanted = payload_length;
4713 msg->msg_offset = 0;
4714 msg->msg_hdr = *hdr;
4715 /* for building message event */
4716 msg->msg_from = *from_nid;
4718 msg->msg_target.pid = dest_pid;
4719 msg->msg_target.nid = dest_nid;
4720 msg->msg_routing = 1;
4724 lpni = lnet_peerni_by_nid_locked(from_nid, &ni->ni_nid, cpt);
4726 lnet_net_unlock(cpt);
4728 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4729 libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4730 lnet_msgtyp2str(type), rc);
4732 if (rc == -ESHUTDOWN)
4733 /* We are shutting down. Don't do anything more */
4738 /* If this message was forwarded to us from a router then we may need
4739 * to update router aliveness or check for an asymmetrical route
4742 if (((lnet_drop_asym_route && for_me) ||
4743 !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4744 LNET_NID_NET(&src_nid) != LNET_NID_NET(from_nid)) {
4745 __u32 src_net_id = LNET_NID_NET(&src_nid);
4746 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4747 struct lnet_route *route;
4750 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4751 if (route->lr_net == src_net_id) {
4753 /* If we're transitioning the gateway from
4754 * dead -> alive, and discovery is disabled
4755 * locally or on the gateway, then we need to
4756 * update the cached route aliveness for each
4757 * route to the src_nid's net.
4759 * Otherwise, we're only checking for
4760 * symmetrical route, and we can break the
4763 if (!gw->lp_alive &&
4764 lnet_is_discovery_disabled(gw))
4765 lnet_set_route_aliveness(route, true);
4770 if (lnet_drop_asym_route && for_me && !found) {
4771 /* Drop ref taken by lnet_nid2peerni_locked() */
4772 lnet_peer_ni_decref_locked(lpni);
4773 lnet_net_unlock(cpt);
4774 /* we would not use from_nid to route a message to
4776 * => asymmetric routing detected but forbidden
4778 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4779 libcfs_nidstr(from_nid),
4780 libcfs_nidstr(&src_nid), lnet_msgtyp2str(type));
4784 if (!gw->lp_alive) {
4785 struct lnet_peer_net *lpn;
4786 struct lnet_peer_ni *lpni2;
4788 gw->lp_alive = true;
4789 /* Mark all remote NIs on src_nid's net UP */
4790 lpn = lnet_peer_get_net_locked(gw, src_net_id);
4792 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4794 lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4798 lpni->lpni_last_alive = now;
4800 msg->msg_rxpeer = lpni;
4802 lnet_ni_addref_locked(ni, cpt);
4803 /* Multi-Rail: Primary NID of source. */
4804 lnet_peer_primary_nid_locked(&src_nid, &msg->msg_initiator);
4807 * mark the status of this lpni as UP since we received a message
4808 * from it. The ping response reports back the ns_status which is
4809 * marked on the remote as up or down and we cache it here.
4811 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4813 lnet_msg_commit(msg, cpt);
4815 /* message delay simulation */
4816 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4817 lnet_delay_rule_match_locked(hdr, msg))) {
4818 lnet_net_unlock(cpt);
4823 rc = lnet_parse_forward_locked(ni, msg);
4824 lnet_net_unlock(cpt);
4829 if (rc == LNET_CREDIT_OK) {
4830 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4831 0, payload_length, payload_length);
4836 lnet_net_unlock(cpt);
4838 rc = lnet_parse_local(ni, msg);
4844 LASSERT(msg->msg_md == NULL);
4845 lnet_finalize(msg, rc);
4848 lnet_drop_message(ni, cpt, private, payload_length, type);
4851 EXPORT_SYMBOL(lnet_parse);
4854 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4856 struct lnet_msg *msg;
4858 while ((msg = list_first_entry_or_null(head, struct lnet_msg,
4859 msg_list)) != NULL) {
4860 struct lnet_processid id = {};
4862 list_del(&msg->msg_list);
4864 id.nid = msg->msg_hdr.src_nid;
4865 id.pid = msg->msg_hdr.src_pid;
4867 LASSERT(msg->msg_md == NULL);
4868 LASSERT(msg->msg_rx_delayed);
4869 LASSERT(msg->msg_rxpeer != NULL);
4870 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4872 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4873 " offset %d length %d: %s\n",
4875 msg->msg_hdr.msg.put.ptl_index,
4876 msg->msg_hdr.msg.put.match_bits,
4877 msg->msg_hdr.msg.put.offset,
4878 msg->msg_hdr.payload_length, reason);
4880 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4881 * called lnet_drop_message(), so I just hang onto msg as well
4882 * until that's done */
4884 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4885 msg->msg_private, msg->msg_len,
4888 msg->msg_no_resend = true;
4890 * NB: message will not generate event because w/o attached MD,
4891 * but we still should give error code so lnet_msg_decommit()
4892 * can skip counters operations and other checks.
4894 lnet_finalize(msg, -ENOENT);
4899 lnet_recv_delayed_msg_list(struct list_head *head)
4901 struct lnet_msg *msg;
4903 while ((msg = list_first_entry_or_null(head, struct lnet_msg,
4904 msg_list)) != NULL) {
4905 struct lnet_processid id;
4907 list_del(&msg->msg_list);
4909 /* md won't disappear under me, since each msg
4910 * holds a ref on it */
4912 id.nid = msg->msg_hdr.src_nid;
4913 id.pid = msg->msg_hdr.src_pid;
4915 LASSERT(msg->msg_rx_delayed);
4916 LASSERT(msg->msg_md != NULL);
4917 LASSERT(msg->msg_rxpeer != NULL);
4918 LASSERT(msg->msg_rxni != NULL);
4919 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4921 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4922 "match %llu offset %d length %d.\n",
4923 libcfs_idstr(&id), msg->msg_hdr.msg.put.ptl_index,
4924 msg->msg_hdr.msg.put.match_bits,
4925 msg->msg_hdr.msg.put.offset,
4926 msg->msg_hdr.payload_length);
4928 lnet_recv_put(msg->msg_rxni, msg);
4933 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4934 struct lnet_libmd *md, struct lnet_handle_md mdh)
4937 struct lnet_rsp_tracker *local_rspt;
4940 * MD has a refcount taken by message so it's not going away.
4941 * The MD however can be looked up. We need to secure the access
4942 * to the md_rspt_ptr by taking the res_lock.
4943 * The rspt can be accessed without protection up to when it gets
4944 * added to the list.
4948 local_rspt = md->md_rspt_ptr;
4949 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4950 if (local_rspt != NULL) {
4952 * we already have an rspt attached to the md, so we'll
4953 * update the deadline on that one.
4955 lnet_rspt_free(rspt, cpt);
4958 rspt->rspt_mdh = mdh;
4959 rspt->rspt_cpt = cpt;
4960 /* store the rspt so we can access it when we get the REPLY */
4961 md->md_rspt_ptr = rspt;
4964 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4967 * add to the list of tracked responses. It's added to tail of the
4968 * list in order to expire all the older entries first.
4971 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4972 lnet_net_unlock(cpt);
4973 lnet_res_unlock(cpt);
4977 * Initiate an asynchronous PUT operation.
4979 * There are several events associated with a PUT: completion of the send on
4980 * the initiator node (LNET_EVENT_SEND), and when the send completes
4981 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4982 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4983 * used at the target node to indicate the completion of incoming data
4986 * The local events will be logged in the EQ associated with the MD pointed to
4987 * by \a mdh handle. Using a MD without an associated EQ results in these
4988 * events being discarded. In this case, the caller must have another
4989 * mechanism (e.g., a higher level protocol) for determining when it is safe
4990 * to modify the memory region associated with the MD.
4992 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4993 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4995 * \param self Indicates the NID of a local interface through which to send
4996 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4997 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4998 * must be "free floating" (See LNetMDBind()).
4999 * \param ack Controls whether an acknowledgment is requested.
5000 * Acknowledgments are only sent when they are requested by the initiating
5001 * process and the target MD enables them.
5002 * \param target A process identifier for the target process.
5003 * \param portal The index in the \a target's portal table.
5004 * \param match_bits The match bits to use for MD selection at the target
5006 * \param offset The offset into the target MD (only used when the target
5007 * MD has the LNET_MD_MANAGE_REMOTE option set).
5008 * \param hdr_data 64 bits of user data that can be included in the message
5009 * header. This data is written to an event queue entry at the target if an
5010 * EQ is present on the matching MD.
5012 * \retval 0 Success, and only in this case events will be generated
5013 * and logged to EQ (if it exists).
5014 * \retval -EIO Simulated failure.
5015 * \retval -ENOMEM Memory allocation failure.
5016 * \retval -ENOENT Invalid MD object.
5018 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
5021 LNetPut(struct lnet_nid *self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
5022 struct lnet_processid *target, unsigned int portal,
5023 __u64 match_bits, unsigned int offset,
5026 struct lnet_msg *msg;
5027 struct lnet_libmd *md;
5030 struct lnet_rsp_tracker *rspt = NULL;
5032 LASSERT(the_lnet.ln_refcount > 0);
5034 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5035 fail_peer(&target->nid, 1)) { /* shall we now? */
5036 CERROR("Dropping PUT to %s: simulated failure\n",
5037 libcfs_idstr(target));
5041 msg = lnet_msg_alloc();
5043 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
5044 libcfs_idstr(target));
5047 msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
5049 cpt = lnet_cpt_of_cookie(mdh.cookie);
5051 if (ack == LNET_ACK_REQ) {
5052 rspt = lnet_rspt_alloc(cpt);
5054 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
5055 libcfs_idstr(target));
5058 INIT_LIST_HEAD(&rspt->rspt_on_list);
5063 md = lnet_handle2md(&mdh);
5064 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5065 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5066 match_bits, portal, libcfs_idstr(target),
5067 md == NULL ? -1 : md->md_threshold);
5068 if (md != NULL && md->md_me != NULL)
5069 CERROR("Source MD also attached to portal %d\n",
5070 md->md_me->me_portal);
5071 lnet_res_unlock(cpt);
5074 lnet_rspt_free(rspt, cpt);
5080 CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5082 lnet_msg_attach_md(msg, md, 0, 0);
5084 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5086 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5087 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5088 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5089 msg->msg_hdr.msg.put.hdr_data = hdr_data;
5091 /* NB handles only looked up by creator (no flips) */
5092 if (ack == LNET_ACK_REQ) {
5093 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5094 the_lnet.ln_interface_cookie;
5095 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5096 md->md_lh.lh_cookie;
5098 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5099 LNET_WIRE_HANDLE_COOKIE_NONE;
5100 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5101 LNET_WIRE_HANDLE_COOKIE_NONE;
5104 lnet_res_unlock(cpt);
5106 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5108 if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5110 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5112 lnet_rspt_free(rspt, cpt);
5114 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5118 rc = lnet_send(self, msg, NULL);
5121 CNETERR("Error sending PUT to %s: %d\n",
5122 libcfs_idstr(target), rc);
5123 msg->msg_no_resend = true;
5124 lnet_finalize(msg, rc);
5127 /* completion will be signalled by an event */
5130 EXPORT_SYMBOL(LNetPut);
5133 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
5134 * returns a msg for the LND to pass to lnet_finalize() when the sink
5135 * data has been received.
5137 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5138 * lnet_finalize() is called on it, so the LND must call this first
5141 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5143 struct lnet_msg *msg = lnet_msg_alloc();
5144 struct lnet_libmd *getmd = getmsg->msg_md;
5145 struct lnet_processid *peer_id = &getmsg->msg_target;
5148 LASSERT(!getmsg->msg_target_is_router);
5149 LASSERT(!getmsg->msg_routing);
5152 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5153 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id));
5157 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5160 LASSERT(getmd->md_refcount > 0);
5162 if (getmd->md_threshold == 0) {
5163 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5164 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id),
5166 lnet_res_unlock(cpt);
5170 LASSERT(getmd->md_offset == 0);
5172 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5173 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id), getmd);
5175 /* setup information for lnet_build_msg_event */
5176 msg->msg_initiator =
5177 getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5178 msg->msg_from = peer_id->nid;
5179 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5180 msg->msg_hdr.src_nid = peer_id->nid;
5181 msg->msg_hdr.payload_length = getmd->md_length;
5182 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5184 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5185 lnet_res_unlock(cpt);
5187 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5190 lnet_msg_commit(msg, cpt);
5191 lnet_net_unlock(cpt);
5193 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5198 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5201 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5202 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5203 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5205 lnet_net_unlock(cpt);
5212 EXPORT_SYMBOL(lnet_create_reply_msg);
5215 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5218 /* Set the REPLY length, now the RDMA that elides the REPLY message has
5219 * completed and I know it. */
5220 LASSERT(reply != NULL);
5221 LASSERT(reply->msg_type == LNET_MSG_GET);
5222 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5224 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
5225 * the end of my buffer, I might as well be dead. */
5226 LASSERT(len <= reply->msg_ev.mlength);
5228 reply->msg_ev.mlength = len;
5230 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5233 * Initiate an asynchronous GET operation.
5235 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5236 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5237 * the target node in the REPLY has been written to local MD.
5239 * On the target node, an LNET_EVENT_GET is logged when the GET request
5240 * arrives and is accepted into a MD.
5242 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5243 * \param mdh A handle for the MD that describes the memory into which the
5244 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5246 * \retval 0 Success, and only in this case events will be generated
5247 * and logged to EQ (if it exists) of the MD.
5248 * \retval -EIO Simulated failure.
5249 * \retval -ENOMEM Memory allocation failure.
5250 * \retval -ENOENT Invalid MD object.
5253 LNetGet(struct lnet_nid *self, struct lnet_handle_md mdh,
5254 struct lnet_processid *target, unsigned int portal,
5255 __u64 match_bits, unsigned int offset, bool recovery)
5257 struct lnet_msg *msg;
5258 struct lnet_libmd *md;
5259 struct lnet_rsp_tracker *rspt;
5263 LASSERT(the_lnet.ln_refcount > 0);
5265 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5266 fail_peer(&target->nid, 1)) /* shall we now? */
5268 CERROR("Dropping GET to %s: simulated failure\n",
5269 libcfs_idstr(target));
5273 msg = lnet_msg_alloc();
5275 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5276 libcfs_idstr(target));
5280 cpt = lnet_cpt_of_cookie(mdh.cookie);
5282 rspt = lnet_rspt_alloc(cpt);
5284 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5285 libcfs_idstr(target));
5288 INIT_LIST_HEAD(&rspt->rspt_on_list);
5290 msg->msg_recovery = recovery;
5294 md = lnet_handle2md(&mdh);
5295 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5296 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5297 match_bits, portal, libcfs_idstr(target),
5298 md == NULL ? -1 : md->md_threshold);
5299 if (md != NULL && md->md_me != NULL)
5300 CERROR("REPLY MD also attached to portal %d\n",
5301 md->md_me->me_portal);
5303 lnet_res_unlock(cpt);
5306 lnet_rspt_free(rspt, cpt);
5310 CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5312 lnet_msg_attach_md(msg, md, 0, 0);
5314 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5316 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5317 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5318 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5319 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5321 /* NB handles only looked up by creator (no flips) */
5322 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5323 the_lnet.ln_interface_cookie;
5324 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5325 md->md_lh.lh_cookie;
5327 lnet_res_unlock(cpt);
5329 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5331 if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5332 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5334 lnet_rspt_free(rspt, cpt);
5336 rc = lnet_send(self, msg, NULL);
5338 CNETERR("Error sending GET to %s: %d\n",
5339 libcfs_idstr(target), rc);
5340 msg->msg_no_resend = true;
5341 lnet_finalize(msg, rc);
5344 /* completion will be signalled by an event */
5347 EXPORT_SYMBOL(LNetGet);
5350 * Calculate distance to node at \a dstnid.
5352 * \param dstnid Target NID.
5353 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5355 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5358 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5359 * local_nid_dist_zero is set, which is the default.
5360 * \retval positives Distance to target NID, i.e. number of hops plus one.
5361 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5364 LNetDist(struct lnet_nid *dstnid, struct lnet_nid *srcnid, __u32 *orderp)
5366 struct lnet_ni *ni = NULL;
5367 struct lnet_remotenet *rnet;
5368 __u32 dstnet = LNET_NID_NET(dstnid);
5372 struct list_head *rn_list;
5373 struct lnet_ni *matched_dstnet = NULL;
5375 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5376 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5377 * keep order 0 free for 0@lo and order 1 free for a local NID
5379 * WARNING: dstnid and srcnid might point to same place.
5380 * Don't set *srcnid until late.
5383 LASSERT(the_lnet.ln_refcount > 0);
5385 cpt = lnet_net_lock_current();
5387 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5388 if (nid_same(&ni->ni_nid, dstnid)) {
5389 if (orderp != NULL) {
5390 if (nid_is_lo0(dstnid))
5397 lnet_net_unlock(cpt);
5399 return local_nid_dist_zero ? 0 : 1;
5402 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5403 matched_dstnet = ni;
5404 /* We matched the destination net, but we may have
5405 * additional local NIs to inspect.
5407 * We record the order as appropriate, but
5408 * they may be overwritten if we match local NI above.
5412 /* Check if ni was originally created in
5413 * current net namespace.
5414 * If not, assign order above 0xffff0000,
5415 * to make this ni not a priority.
5417 if (current->nsproxy &&
5418 !net_eq(ni->ni_net_ns,
5419 current->nsproxy->net_ns))
5420 *orderp = order + 0xffff0000;
5429 if (matched_dstnet) {
5431 *srcnid = matched_dstnet->ni_nid;
5432 lnet_net_unlock(cpt);
5436 rn_list = lnet_net2rnethash(dstnet);
5437 list_for_each_entry(rnet, rn_list, lrn_list) {
5438 if (rnet->lrn_net == dstnet) {
5439 struct lnet_route *route;
5440 struct lnet_route *shortest = NULL;
5441 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5444 LASSERT(!list_empty(&rnet->lrn_routes));
5446 list_for_each_entry(route, &rnet->lrn_routes,
5448 route_hops = route->lr_hops;
5449 if (route_hops == LNET_UNDEFINED_HOPS)
5451 if (shortest == NULL ||
5452 route_hops < shortest_hops) {
5454 shortest_hops = route_hops;
5458 LASSERT(shortest != NULL);
5459 hops = shortest_hops;
5461 struct lnet_net *net;
5462 net = lnet_get_net_locked(shortest->lr_lnet);
5464 ni = lnet_get_next_ni_locked(net, NULL);
5465 *srcnid = ni->ni_nid;
5469 lnet_net_unlock(cpt);
5475 lnet_net_unlock(cpt);
5476 return -EHOSTUNREACH;
5478 EXPORT_SYMBOL(LNetDist);