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 lnet_nid_t sd_dst_nid;
59 lnet_nid_t sd_src_nid;
60 lnet_nid_t 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 (!list_empty(&cull)) {
233 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
235 list_del(&tp->tp_list);
236 LIBCFS_FREE(tp, sizeof(*tp));
242 fail_peer(lnet_nid_t nid4, int outgoing)
244 struct lnet_test_peer *tp;
245 struct list_head *el;
246 struct list_head *next;
251 lnet_nid4_to_nid(nid4, &nid);
252 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
255 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
256 tp = list_entry(el, struct lnet_test_peer, tp_list);
258 if (tp->tp_threshold == 0) {
261 /* only cull zombies on outgoing tests,
262 * since we may be at interrupt priority on
263 * incoming messages. */
264 list_move(&tp->tp_list, &cull);
269 if (LNET_NID_IS_ANY(&tp->tp_nid) || /* fail every peer */
270 nid_same(&nid, &tp->tp_nid)) { /* fail this peer */
273 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
276 tp->tp_threshold == 0) {
278 list_move(&tp->tp_list, &cull);
287 while (!list_empty(&cull)) {
288 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
289 list_del(&tp->tp_list);
291 LIBCFS_FREE(tp, sizeof(*tp));
298 lnet_iov_nob(unsigned int niov, struct kvec *iov)
300 unsigned int nob = 0;
302 LASSERT(niov == 0 || iov != NULL);
304 nob += (iov++)->iov_len;
308 EXPORT_SYMBOL(lnet_iov_nob);
311 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
312 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
315 /* NB diov, siov are READ-ONLY */
316 unsigned int this_nob;
321 /* skip complete frags before 'doffset' */
323 while (doffset >= diov->iov_len) {
324 doffset -= diov->iov_len;
330 /* skip complete frags before 'soffset' */
332 while (soffset >= siov->iov_len) {
333 soffset -= siov->iov_len;
342 this_nob = min3((unsigned int)diov->iov_len - doffset,
343 (unsigned int)siov->iov_len - soffset,
346 memcpy((char *)diov->iov_base + doffset,
347 (char *)siov->iov_base + soffset, this_nob);
350 if (diov->iov_len > doffset + this_nob) {
358 if (siov->iov_len > soffset + this_nob) {
367 EXPORT_SYMBOL(lnet_copy_iov2iov);
370 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
372 unsigned int nob = 0;
374 LASSERT(niov == 0 || kiov != NULL);
376 nob += (kiov++)->bv_len;
380 EXPORT_SYMBOL(lnet_kiov_nob);
383 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
384 unsigned int doffset,
385 unsigned int nsiov, struct bio_vec *siov,
386 unsigned int soffset,
389 /* NB diov, siov are READ-ONLY */
390 unsigned int this_nob;
397 LASSERT (!in_interrupt ());
400 while (doffset >= diov->bv_len) {
401 doffset -= diov->bv_len;
408 while (soffset >= siov->bv_len) {
409 soffset -= siov->bv_len;
418 this_nob = min3(diov->bv_len - doffset,
419 siov->bv_len - soffset,
423 daddr = ((char *)kmap(diov->bv_page)) +
424 diov->bv_offset + doffset;
426 saddr = ((char *)kmap(siov->bv_page)) +
427 siov->bv_offset + soffset;
429 /* Vanishing risk of kmap deadlock when mapping 2 pages.
430 * However in practice at least one of the kiovs will be mapped
431 * kernel pages and the map/unmap will be NOOPs */
433 memcpy (daddr, saddr, this_nob);
436 if (diov->bv_len > doffset + this_nob) {
440 kunmap(diov->bv_page);
447 if (siov->bv_len > soffset + this_nob) {
451 kunmap(siov->bv_page);
460 kunmap(diov->bv_page);
462 kunmap(siov->bv_page);
464 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
467 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
468 unsigned int nkiov, struct bio_vec *kiov,
469 unsigned int kiovoffset,
472 /* NB iov, kiov are READ-ONLY */
473 unsigned int this_nob;
479 LASSERT (!in_interrupt ());
482 while (iovoffset >= iov->iov_len) {
483 iovoffset -= iov->iov_len;
490 while (kiovoffset >= kiov->bv_len) {
491 kiovoffset -= kiov->bv_len;
500 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
501 (unsigned int)kiov->bv_len - kiovoffset,
505 addr = ((char *)kmap(kiov->bv_page)) +
506 kiov->bv_offset + kiovoffset;
508 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
511 if (iov->iov_len > iovoffset + this_nob) {
512 iovoffset += this_nob;
519 if (kiov->bv_len > kiovoffset + this_nob) {
521 kiovoffset += this_nob;
523 kunmap(kiov->bv_page);
533 kunmap(kiov->bv_page);
535 EXPORT_SYMBOL(lnet_copy_kiov2iov);
538 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
539 unsigned int kiovoffset,
540 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
543 /* NB kiov, iov are READ-ONLY */
544 unsigned int this_nob;
550 LASSERT (!in_interrupt ());
553 while (kiovoffset >= kiov->bv_len) {
554 kiovoffset -= kiov->bv_len;
561 while (iovoffset >= iov->iov_len) {
562 iovoffset -= iov->iov_len;
571 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
572 (unsigned int)iov->iov_len - iovoffset,
576 addr = ((char *)kmap(kiov->bv_page)) +
577 kiov->bv_offset + kiovoffset;
579 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
582 if (kiov->bv_len > kiovoffset + this_nob) {
584 kiovoffset += this_nob;
586 kunmap(kiov->bv_page);
593 if (iov->iov_len > iovoffset + this_nob) {
594 iovoffset += this_nob;
603 kunmap(kiov->bv_page);
605 EXPORT_SYMBOL(lnet_copy_iov2kiov);
608 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
609 int src_niov, struct bio_vec *src,
610 unsigned int offset, unsigned int len)
612 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
613 * for exactly 'len' bytes, and return the number of entries.
614 * NB not destructive to 'src' */
615 unsigned int frag_len;
618 if (len == 0) /* no data => */
619 return (0); /* no frags */
621 LASSERT(src_niov > 0);
622 while (offset >= src->bv_len) { /* skip initial frags */
623 offset -= src->bv_len;
626 LASSERT(src_niov > 0);
631 LASSERT(src_niov > 0);
632 LASSERT((int)niov <= dst_niov);
634 frag_len = src->bv_len - offset;
635 dst->bv_page = src->bv_page;
636 dst->bv_offset = src->bv_offset + offset;
638 if (len <= frag_len) {
640 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
644 dst->bv_len = frag_len;
645 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
655 EXPORT_SYMBOL(lnet_extract_kiov);
658 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
659 int delayed, unsigned int offset, unsigned int mlen,
662 unsigned int niov = 0;
663 struct kvec *iov = NULL;
664 struct bio_vec *kiov = NULL;
667 LASSERT (!in_interrupt ());
668 LASSERT (mlen == 0 || msg != NULL);
671 LASSERT(msg->msg_receiving);
672 LASSERT(!msg->msg_sending);
673 LASSERT(rlen == msg->msg_len);
674 LASSERT(mlen <= msg->msg_len);
675 LASSERT(msg->msg_offset == offset);
676 LASSERT(msg->msg_wanted == mlen);
678 msg->msg_receiving = 0;
681 niov = msg->msg_niov;
682 kiov = msg->msg_kiov;
685 LASSERT ((iov == NULL) != (kiov == NULL));
689 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
690 niov, kiov, offset, mlen,
693 lnet_finalize(msg, rc);
697 lnet_setpayloadbuffer(struct lnet_msg *msg)
699 struct lnet_libmd *md = msg->msg_md;
701 LASSERT(msg->msg_len > 0);
702 LASSERT(!msg->msg_routing);
704 LASSERT(msg->msg_niov == 0);
705 LASSERT(msg->msg_kiov == NULL);
707 msg->msg_niov = md->md_niov;
708 msg->msg_kiov = md->md_kiov;
712 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
713 unsigned int offset, unsigned int len)
715 msg->msg_type = type;
716 msg->msg_target = target;
718 msg->msg_offset = offset;
721 lnet_setpayloadbuffer(msg);
723 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
724 msg->msg_hdr.type = cpu_to_le32(type);
725 /* dest_nid will be overwritten by lnet_select_pathway() */
726 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
727 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
728 /* src_nid will be set later */
729 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
730 msg->msg_hdr.payload_length = cpu_to_le32(len);
734 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
736 void *priv = msg->msg_private;
739 LASSERT(!in_interrupt());
740 LASSERT(nid_is_lo0(&ni->ni_nid) ||
741 (msg->msg_txcredit && msg->msg_peertxcredit));
743 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
745 msg->msg_no_resend = true;
746 lnet_finalize(msg, rc);
751 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
755 LASSERT(!msg->msg_sending);
756 LASSERT(msg->msg_receiving);
757 LASSERT(!msg->msg_rx_ready_delay);
758 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
760 msg->msg_rx_ready_delay = 1;
761 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
764 CERROR("recv from %s / send to %s aborted: "
765 "eager_recv failed %d\n",
766 libcfs_nidstr(&msg->msg_rxpeer->lpni_nid),
767 libcfs_id2str(msg->msg_target), rc);
768 LASSERT(rc < 0); /* required by my callers */
775 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
779 deadline = lpni->lpni_last_alive +
780 lpni->lpni_net->net_tunables.lct_peer_timeout;
783 * assume peer_ni is alive as long as we're within the configured
792 /* NB: returns 1 when alive, 0 when dead, negative when error;
793 * may drop the lnet_net_lock */
795 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
796 struct lnet_msg *msg)
798 time64_t now = ktime_get_seconds();
800 if (!lnet_peer_aliveness_enabled(lpni))
804 * If we're resending a message, let's attempt to send it even if
805 * the peer is down to fulfill our resend quota on the message
807 if (msg->msg_retry_count > 0)
810 /* try and send recovery messages irregardless */
811 if (msg->msg_recovery)
814 /* always send any responses */
815 if (lnet_msg_is_response(msg))
818 if (!lnet_is_peer_deadline_passed(lpni, now))
821 return lnet_is_peer_ni_alive(lpni);
825 * \param msg The message to be sent.
826 * \param do_send True if lnet_ni_send() should be called in this function.
827 * lnet_send() is going to lnet_net_unlock immediately after this, so
828 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
830 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
831 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
832 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
833 * \retval -ECANCELED If the MD of the message has been unlinked.
836 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
838 struct lnet_peer_ni *lp = msg->msg_txpeer;
839 struct lnet_ni *ni = msg->msg_txni;
840 int cpt = msg->msg_tx_cpt;
841 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
843 /* non-lnet_send() callers have checked before */
844 LASSERT(!do_send || msg->msg_tx_delayed);
845 LASSERT(!msg->msg_receiving);
846 LASSERT(msg->msg_tx_committed);
848 /* can't get here if we're sending to the loopback interface */
849 if (the_lnet.ln_loni)
850 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
852 /* NB 'lp' is always the next hop */
853 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
854 lnet_peer_alive_locked(ni, lp, msg) == 0) {
855 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
856 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
858 lnet_net_unlock(cpt);
860 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
862 LNET_STATS_TYPE_DROP);
864 lnet_incr_stats(&msg->msg_txni->ni_stats,
866 LNET_STATS_TYPE_DROP);
868 CNETERR("Dropping message for %s: peer not alive\n",
869 libcfs_id2str(msg->msg_target));
870 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
872 lnet_finalize(msg, -EHOSTUNREACH);
875 return -EHOSTUNREACH;
878 if (msg->msg_md != NULL &&
879 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
880 lnet_net_unlock(cpt);
882 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
883 "called on the MD/ME.\n",
884 libcfs_id2str(msg->msg_target));
886 msg->msg_no_resend = true;
887 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
888 msg, libcfs_id2str(msg->msg_target));
889 lnet_finalize(msg, -ECANCELED);
896 if (!msg->msg_peertxcredit) {
897 spin_lock(&lp->lpni_lock);
898 LASSERT((lp->lpni_txcredits < 0) ==
899 !list_empty(&lp->lpni_txq));
901 msg->msg_peertxcredit = 1;
902 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
903 lp->lpni_txcredits--;
905 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
906 lp->lpni_mintxcredits = lp->lpni_txcredits;
908 if (lp->lpni_txcredits < 0) {
909 msg->msg_tx_delayed = 1;
910 list_add_tail(&msg->msg_list, &lp->lpni_txq);
911 spin_unlock(&lp->lpni_lock);
912 return LNET_CREDIT_WAIT;
914 spin_unlock(&lp->lpni_lock);
917 if (!msg->msg_txcredit) {
918 LASSERT((tq->tq_credits < 0) ==
919 !list_empty(&tq->tq_delayed));
921 msg->msg_txcredit = 1;
923 atomic_dec(&ni->ni_tx_credits);
925 if (tq->tq_credits < tq->tq_credits_min)
926 tq->tq_credits_min = tq->tq_credits;
928 if (tq->tq_credits < 0) {
929 msg->msg_tx_delayed = 1;
930 list_add_tail(&msg->msg_list, &tq->tq_delayed);
931 return LNET_CREDIT_WAIT;
935 if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
936 lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
937 msg->msg_tx_delayed = 1;
938 return LNET_CREDIT_WAIT;
941 /* unset the tx_delay flag as we're going to send it now */
942 msg->msg_tx_delayed = 0;
945 lnet_net_unlock(cpt);
946 lnet_ni_send(ni, msg);
949 return LNET_CREDIT_OK;
953 static struct lnet_rtrbufpool *
954 lnet_msg2bufpool(struct lnet_msg *msg)
956 struct lnet_rtrbufpool *rbp;
959 LASSERT(msg->msg_rx_committed);
961 cpt = msg->msg_rx_cpt;
962 rbp = &the_lnet.ln_rtrpools[cpt][0];
964 LASSERT(msg->msg_len <= LNET_MTU);
965 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
967 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
974 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
976 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
977 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
978 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
979 * received or OK to receive */
980 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
981 struct lnet_peer *lp;
982 struct lnet_rtrbufpool *rbp;
983 struct lnet_rtrbuf *rb;
985 LASSERT(msg->msg_kiov == NULL);
986 LASSERT(msg->msg_niov == 0);
987 LASSERT(msg->msg_routing);
988 LASSERT(msg->msg_receiving);
989 LASSERT(!msg->msg_sending);
990 LASSERT(lpni->lpni_peer_net);
991 LASSERT(lpni->lpni_peer_net->lpn_peer);
993 lp = lpni->lpni_peer_net->lpn_peer;
995 /* non-lnet_parse callers only receive delayed messages */
996 LASSERT(!do_recv || msg->msg_rx_delayed);
998 if (!msg->msg_peerrtrcredit) {
999 /* lpni_lock protects the credit manipulation */
1000 spin_lock(&lpni->lpni_lock);
1002 msg->msg_peerrtrcredit = 1;
1003 lpni->lpni_rtrcredits--;
1004 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1005 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1007 if (lpni->lpni_rtrcredits < 0) {
1008 spin_unlock(&lpni->lpni_lock);
1009 /* must have checked eager_recv before here */
1010 LASSERT(msg->msg_rx_ready_delay);
1011 msg->msg_rx_delayed = 1;
1012 /* lp_lock protects the lp_rtrq */
1013 spin_lock(&lp->lp_lock);
1014 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1015 spin_unlock(&lp->lp_lock);
1016 return LNET_CREDIT_WAIT;
1018 spin_unlock(&lpni->lpni_lock);
1021 rbp = lnet_msg2bufpool(msg);
1023 if (!msg->msg_rtrcredit) {
1024 msg->msg_rtrcredit = 1;
1026 if (rbp->rbp_credits < rbp->rbp_mincredits)
1027 rbp->rbp_mincredits = rbp->rbp_credits;
1029 if (rbp->rbp_credits < 0) {
1030 /* must have checked eager_recv before here */
1031 LASSERT(msg->msg_rx_ready_delay);
1032 msg->msg_rx_delayed = 1;
1033 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1034 return LNET_CREDIT_WAIT;
1038 LASSERT(!list_empty(&rbp->rbp_bufs));
1039 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1040 list_del(&rb->rb_list);
1042 msg->msg_niov = rbp->rbp_npages;
1043 msg->msg_kiov = &rb->rb_kiov[0];
1045 /* unset the msg-rx_delayed flag since we're receiving the message */
1046 msg->msg_rx_delayed = 0;
1049 int cpt = msg->msg_rx_cpt;
1051 lnet_net_unlock(cpt);
1052 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1053 0, msg->msg_len, msg->msg_len);
1056 return LNET_CREDIT_OK;
1060 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1062 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1063 struct lnet_ni *txni = msg->msg_txni;
1064 struct lnet_msg *msg2;
1066 if (msg->msg_txcredit) {
1067 struct lnet_ni *ni = msg->msg_txni;
1068 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1070 /* give back NI txcredits */
1071 msg->msg_txcredit = 0;
1073 LASSERT((tq->tq_credits < 0) ==
1074 !list_empty(&tq->tq_delayed));
1077 atomic_inc(&ni->ni_tx_credits);
1078 if (tq->tq_credits <= 0) {
1079 msg2 = list_entry(tq->tq_delayed.next,
1080 struct lnet_msg, msg_list);
1081 list_del(&msg2->msg_list);
1083 LASSERT(msg2->msg_txni == ni);
1084 LASSERT(msg2->msg_tx_delayed);
1085 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1087 (void) lnet_post_send_locked(msg2, 1);
1091 if (msg->msg_peertxcredit) {
1092 /* give back peer txcredits */
1093 msg->msg_peertxcredit = 0;
1095 spin_lock(&txpeer->lpni_lock);
1096 LASSERT((txpeer->lpni_txcredits < 0) ==
1097 !list_empty(&txpeer->lpni_txq));
1099 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1100 LASSERT(txpeer->lpni_txqnob >= 0);
1102 txpeer->lpni_txcredits++;
1103 if (txpeer->lpni_txcredits <= 0) {
1106 msg2 = list_entry(txpeer->lpni_txq.next,
1107 struct lnet_msg, msg_list);
1108 list_del(&msg2->msg_list);
1109 spin_unlock(&txpeer->lpni_lock);
1111 LASSERT(msg2->msg_txpeer == txpeer);
1112 LASSERT(msg2->msg_tx_delayed);
1114 msg2_cpt = msg2->msg_tx_cpt;
1117 * The msg_cpt can be different from the msg2_cpt
1118 * so we need to make sure we lock the correct cpt
1120 * Once we call lnet_post_send_locked() it is no
1121 * longer safe to access msg2, since it could've
1122 * been freed by lnet_finalize(), but we still
1123 * need to relock the correct cpt, so we cache the
1124 * msg2_cpt for the purpose of the check that
1125 * follows the call to lnet_pose_send_locked().
1127 if (msg2_cpt != msg->msg_tx_cpt) {
1128 lnet_net_unlock(msg->msg_tx_cpt);
1129 lnet_net_lock(msg2_cpt);
1131 (void) lnet_post_send_locked(msg2, 1);
1132 if (msg2_cpt != msg->msg_tx_cpt) {
1133 lnet_net_unlock(msg2_cpt);
1134 lnet_net_lock(msg->msg_tx_cpt);
1137 spin_unlock(&txpeer->lpni_lock);
1142 msg->msg_txni = NULL;
1143 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1146 if (txpeer != NULL) {
1147 msg->msg_txpeer = NULL;
1148 lnet_peer_ni_decref_locked(txpeer);
1153 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1155 struct lnet_msg *msg;
1157 if (list_empty(&rbp->rbp_msgs))
1159 msg = list_entry(rbp->rbp_msgs.next,
1160 struct lnet_msg, msg_list);
1161 list_del(&msg->msg_list);
1163 (void)lnet_post_routed_recv_locked(msg, 1);
1167 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1169 struct lnet_msg *msg;
1170 struct lnet_msg *tmp;
1172 lnet_net_unlock(cpt);
1174 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1175 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1176 0, 0, 0, msg->msg_hdr.payload_length);
1177 list_del_init(&msg->msg_list);
1178 msg->msg_no_resend = true;
1179 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1180 lnet_finalize(msg, -ECANCELED);
1187 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1189 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1190 struct lnet_peer *lp;
1191 struct lnet_ni *rxni = msg->msg_rxni;
1192 struct lnet_msg *msg2;
1194 if (msg->msg_rtrcredit) {
1195 /* give back global router credits */
1196 struct lnet_rtrbuf *rb;
1197 struct lnet_rtrbufpool *rbp;
1199 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1200 * there until it gets one allocated, or aborts the wait
1202 LASSERT(msg->msg_kiov != NULL);
1204 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1207 msg->msg_kiov = NULL;
1208 msg->msg_rtrcredit = 0;
1210 LASSERT(rbp == lnet_msg2bufpool(msg));
1212 LASSERT((rbp->rbp_credits > 0) ==
1213 !list_empty(&rbp->rbp_bufs));
1215 /* If routing is now turned off, we just drop this buffer and
1216 * don't bother trying to return credits. */
1217 if (!the_lnet.ln_routing) {
1218 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1222 /* It is possible that a user has lowered the desired number of
1223 * buffers in this pool. Make sure we never put back
1224 * more buffers than the stated number. */
1225 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1226 /* Discard this buffer so we don't have too
1228 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1229 rbp->rbp_nbuffers--;
1231 list_add(&rb->rb_list, &rbp->rbp_bufs);
1233 if (rbp->rbp_credits <= 0)
1234 lnet_schedule_blocked_locked(rbp);
1239 if (msg->msg_peerrtrcredit) {
1241 LASSERT(rxpeerni->lpni_peer_net);
1242 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1244 /* give back peer router credits */
1245 msg->msg_peerrtrcredit = 0;
1247 spin_lock(&rxpeerni->lpni_lock);
1248 rxpeerni->lpni_rtrcredits++;
1249 spin_unlock(&rxpeerni->lpni_lock);
1251 lp = rxpeerni->lpni_peer_net->lpn_peer;
1252 spin_lock(&lp->lp_lock);
1254 /* drop all messages which are queued to be routed on that
1256 if (!the_lnet.ln_routing) {
1258 list_splice_init(&lp->lp_rtrq, &drop);
1259 spin_unlock(&lp->lp_lock);
1260 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1261 } else if (!list_empty(&lp->lp_rtrq)) {
1264 msg2 = list_entry(lp->lp_rtrq.next,
1265 struct lnet_msg, msg_list);
1266 list_del(&msg2->msg_list);
1267 msg2_cpt = msg2->msg_rx_cpt;
1268 spin_unlock(&lp->lp_lock);
1270 * messages on the lp_rtrq can be from any NID in
1271 * the peer, which means they might have different
1272 * cpts. We need to make sure we lock the right
1275 if (msg2_cpt != msg->msg_rx_cpt) {
1276 lnet_net_unlock(msg->msg_rx_cpt);
1277 lnet_net_lock(msg2_cpt);
1279 (void) lnet_post_routed_recv_locked(msg2, 1);
1280 if (msg2_cpt != msg->msg_rx_cpt) {
1281 lnet_net_unlock(msg2_cpt);
1282 lnet_net_lock(msg->msg_rx_cpt);
1285 spin_unlock(&lp->lp_lock);
1289 msg->msg_rxni = NULL;
1290 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1292 if (rxpeerni != NULL) {
1293 msg->msg_rxpeer = NULL;
1294 lnet_peer_ni_decref_locked(rxpeerni);
1298 static struct lnet_peer_ni *
1299 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1300 struct lnet_peer *peer,
1301 struct lnet_peer_ni *best_lpni,
1302 struct lnet_peer_net *peer_net)
1305 * Look at the peer NIs for the destination peer that connect
1306 * to the chosen net. If a peer_ni is preferred when using the
1307 * best_ni to communicate, we use that one. If there is no
1308 * preferred peer_ni, or there are multiple preferred peer_ni,
1309 * the available transmit credits are used. If the transmit
1310 * credits are equal, we round-robin over the peer_ni.
1312 struct lnet_peer_ni *lpni = NULL;
1313 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1315 int best_lpni_healthv = (best_lpni) ?
1316 atomic_read(&best_lpni->lpni_healthv) : 0;
1317 bool best_lpni_is_preferred = false;
1318 bool lpni_is_preferred;
1320 __u32 lpni_sel_prio;
1321 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1323 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1325 * if the best_ni we've chosen aleady has this lpni
1326 * preferred, then let's use it
1329 lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1330 lpni, &best_ni->ni_nid);
1331 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1332 libcfs_nidstr(&best_ni->ni_nid),
1335 lpni_is_preferred = false;
1338 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1339 lpni_sel_prio = lpni->lpni_sel_priority;
1342 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1343 libcfs_nidstr(&lpni->lpni_nid),
1344 libcfs_nidstr(&best_lpni->lpni_nid),
1345 lpni_healthv, best_lpni_healthv,
1346 lpni_sel_prio, best_sel_prio,
1347 lpni->lpni_txcredits, best_lpni_credits,
1348 lpni->lpni_seq, best_lpni->lpni_seq);
1352 /* pick the healthiest peer ni */
1353 if (lpni_healthv < best_lpni_healthv)
1355 else if (lpni_healthv > best_lpni_healthv) {
1356 if (best_lpni_is_preferred)
1357 best_lpni_is_preferred = false;
1361 if (lpni_sel_prio > best_sel_prio)
1363 else if (lpni_sel_prio < best_sel_prio) {
1364 if (best_lpni_is_preferred)
1365 best_lpni_is_preferred = false;
1369 /* if this is a preferred peer use it */
1370 if (!best_lpni_is_preferred && lpni_is_preferred) {
1371 best_lpni_is_preferred = true;
1373 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1374 /* this is not the preferred peer so let's ignore
1380 if (lpni->lpni_txcredits < best_lpni_credits)
1381 /* We already have a peer that has more credits
1382 * available than this one. No need to consider
1383 * this peer further.
1386 else if (lpni->lpni_txcredits > best_lpni_credits)
1389 /* The best peer found so far and the current peer
1390 * have the same number of available credits let's
1391 * make sure to select between them using Round Robin
1393 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1396 best_lpni_is_preferred = lpni_is_preferred;
1397 best_lpni_healthv = lpni_healthv;
1398 best_sel_prio = lpni_sel_prio;
1400 best_lpni_credits = lpni->lpni_txcredits;
1403 /* if we still can't find a peer ni then we can't reach it */
1405 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1406 LNET_NIDNET(dst_nid);
1407 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1408 libcfs_net2str(net_id));
1412 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1413 libcfs_nidstr(&best_lpni->lpni_nid));
1419 * Prerequisite: the best_ni should already be set in the sd
1420 * Find the best lpni.
1421 * If the net id is provided then restrict lpni selection on
1422 * that particular net.
1423 * Otherwise find any reachable lpni. When dealing with an MR
1424 * gateway and it has multiple lpnis which we can use
1425 * we want to select the best one from the list of reachable
1428 static inline struct lnet_peer_ni *
1429 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1430 struct lnet_peer *peer, __u32 net_id)
1432 struct lnet_peer_net *peer_net;
1434 /* find the best_lpni on any local network */
1435 if (net_id == LNET_NET_ANY) {
1436 struct lnet_peer_ni *best_lpni = NULL;
1437 struct lnet_peer_net *lpn;
1438 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1439 /* no net specified find any reachable peer ni */
1440 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1442 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1448 /* restrict on the specified net */
1449 peer_net = lnet_peer_get_net_locked(peer, net_id);
1451 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1457 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1459 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1462 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1465 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1468 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1474 /* Compare route priorities and hop counts */
1476 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1478 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1479 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1481 if (r1->lr_priority < r2->lr_priority)
1484 if (r1->lr_priority > r2->lr_priority)
1487 if (r1_hops < r2_hops)
1490 if (r1_hops > r2_hops)
1496 static struct lnet_route *
1497 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1498 struct lnet_peer_ni *remote_lpni,
1499 struct lnet_route **prev_route,
1500 struct lnet_peer_ni **gwni)
1502 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1503 struct lnet_route *best_route;
1504 struct lnet_route *last_route;
1505 struct lnet_route *route;
1507 bool best_rte_is_preferred = false;
1508 struct lnet_nid *gw_pnid;
1510 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1511 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1513 best_route = last_route = NULL;
1514 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1515 if (!lnet_is_route_alive(route))
1517 gw_pnid = &route->lr_gateway->lp_primary_nid;
1519 /* no protection on below fields, but it's harmless */
1520 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1523 /* if the best route found is in the preferred list then
1524 * tag it as preferred and use it later on. But if we
1525 * didn't find any routes which are on the preferred list
1526 * then just use the best route possible.
1528 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1530 if (!best_route || (rc && !best_rte_is_preferred)) {
1531 /* Restrict the selection of the router NI on the
1532 * src_net provided. If the src_net is LNET_NID_ANY,
1533 * then select the best interface available.
1535 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1540 "Gateway %s does not have a peer NI on net %s\n",
1541 libcfs_nidstr(gw_pnid),
1542 libcfs_net2str(src_net));
1547 if (rc && !best_rte_is_preferred) {
1548 /* This is the first preferred route we found,
1549 * so it beats any route found previously
1555 best_rte_is_preferred = true;
1556 CDEBUG(D_NET, "preferred gw = %s\n",
1557 libcfs_nidstr(gw_pnid));
1559 } else if ((!rc) && best_rte_is_preferred)
1560 /* The best route we found so far is in the preferred
1561 * list, so it beats any non-preferred route
1566 best_route = last_route = route;
1571 rc = lnet_compare_routes(route, best_route);
1575 /* Restrict the selection of the router NI on the
1576 * src_net provided. If the src_net is LNET_NID_ANY,
1577 * then select the best interface available.
1579 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1584 "Gateway %s does not have a peer NI on net %s\n",
1585 libcfs_nidstr(gw_pnid),
1586 libcfs_net2str(src_net));
1596 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1600 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1607 *prev_route = last_route;
1613 static inline unsigned int
1614 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1616 if (dev_idx == UINT_MAX)
1619 if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1620 !ni->ni_net->net_lnd->lnd_get_dev_prio)
1623 return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1626 static struct lnet_ni *
1627 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1628 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1629 struct lnet_msg *msg, int md_cpt)
1631 struct lnet_libmd *md = msg->msg_md;
1632 unsigned int offset = msg->msg_offset;
1633 unsigned int shortest_distance;
1634 struct lnet_ni *ni = NULL;
1637 __u32 best_sel_prio;
1638 unsigned int best_dev_prio;
1639 unsigned int dev_idx = UINT_MAX;
1640 struct page *page = lnet_get_first_page(md, offset);
1641 msg->msg_rdma_force = lnet_is_rdma_only_page(page);
1643 if (msg->msg_rdma_force)
1644 dev_idx = lnet_get_dev_idx(page);
1647 * If there is no peer_ni that we can send to on this network,
1648 * then there is no point in looking for a new best_ni here.
1650 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1653 if (best_ni == NULL) {
1654 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1655 shortest_distance = UINT_MAX;
1656 best_dev_prio = UINT_MAX;
1657 best_credits = INT_MIN;
1660 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1661 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1662 best_ni->ni_dev_cpt);
1663 best_credits = atomic_read(&best_ni->ni_tx_credits);
1664 best_healthv = atomic_read(&best_ni->ni_healthv);
1665 best_sel_prio = best_ni->ni_sel_priority;
1668 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1669 unsigned int distance;
1674 unsigned int ni_dev_prio;
1676 ni_credits = atomic_read(&ni->ni_tx_credits);
1677 ni_healthv = atomic_read(&ni->ni_healthv);
1678 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1679 ni_sel_prio = ni->ni_sel_priority;
1682 * calculate the distance from the CPT on which
1683 * the message memory is allocated to the CPT of
1684 * the NI's physical device
1686 distance = cfs_cpt_distance(lnet_cpt_table(),
1690 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1693 * All distances smaller than the NUMA range
1694 * are treated equally.
1696 if (distance < lnet_numa_range)
1697 distance = lnet_numa_range;
1700 * Select on health, selection policy, direct dma prio,
1701 * shorter distance, available credits, then round-robin.
1707 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d, p:%u, g:%u] with best_ni %s [c:%d, d:%d, s:%d, p:%u, g:%u]\n",
1708 libcfs_nidstr(&ni->ni_nid), ni_credits, distance,
1709 ni->ni_seq, ni_sel_prio, ni_dev_prio,
1710 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1711 : "not selected", best_credits, shortest_distance,
1712 (best_ni) ? best_ni->ni_seq : 0,
1713 best_sel_prio, best_dev_prio);
1717 if (ni_healthv < best_healthv)
1719 else if (ni_healthv > best_healthv)
1722 if (ni_sel_prio > best_sel_prio)
1724 else if (ni_sel_prio < best_sel_prio)
1727 if (ni_dev_prio > best_dev_prio)
1729 else if (ni_dev_prio < best_dev_prio)
1732 if (distance > shortest_distance)
1734 else if (distance < shortest_distance)
1737 if (ni_credits < best_credits)
1739 else if (ni_credits > best_credits)
1742 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1746 best_sel_prio = ni_sel_prio;
1747 best_dev_prio = ni_dev_prio;
1748 shortest_distance = distance;
1749 best_healthv = ni_healthv;
1751 best_credits = ni_credits;
1754 CDEBUG(D_NET, "selected best_ni %s\n",
1755 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1761 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1762 * because such traffic is required to perform discovery. We therefore
1763 * exclude all GET and PUT on that portal. We also exclude all ACK and
1764 * REPLY traffic, but that is because the portal is not tracked in the
1765 * message structure for these message types. We could restrict this
1766 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1769 lnet_msg_discovery(struct lnet_msg *msg)
1771 if (msg->msg_type == LNET_MSG_PUT) {
1772 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1774 } else if (msg->msg_type == LNET_MSG_GET) {
1775 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1781 #define SRC_SPEC 0x0001
1782 #define SRC_ANY 0x0002
1783 #define LOCAL_DST 0x0004
1784 #define REMOTE_DST 0x0008
1785 #define MR_DST 0x0010
1786 #define NMR_DST 0x0020
1787 #define SND_RESP 0x0040
1789 /* The following to defines are used for return codes */
1790 #define REPEAT_SEND 0x1000
1791 #define PASS_THROUGH 0x2000
1793 /* The different cases lnet_select pathway needs to handle */
1794 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1795 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1796 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1797 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1798 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1799 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1800 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1801 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1804 lnet_handle_lo_send(struct lnet_send_data *sd)
1806 struct lnet_msg *msg = sd->sd_msg;
1807 int cpt = sd->sd_cpt;
1809 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1812 /* No send credit hassles with LOLND */
1813 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1814 msg->msg_hdr.dest_nid =
1815 cpu_to_le64(lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid));
1816 if (!msg->msg_routing)
1817 msg->msg_hdr.src_nid =
1818 cpu_to_le64(lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid));
1819 msg->msg_target.nid = lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid);
1820 lnet_msg_commit(msg, cpt);
1821 msg->msg_txni = the_lnet.ln_loni;
1823 return LNET_CREDIT_OK;
1827 lnet_handle_send(struct lnet_send_data *sd)
1829 struct lnet_ni *best_ni = sd->sd_best_ni;
1830 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1831 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1832 struct lnet_msg *msg = sd->sd_msg;
1834 __u32 send_case = sd->sd_send_case;
1836 __u32 routing = send_case & REMOTE_DST;
1837 struct lnet_rsp_tracker *rspt;
1839 /* Increment sequence number of the selected peer, peer net,
1840 * local ni and local net so that we pick the next ones
1843 best_lpni->lpni_peer_net->lpn_seq++;
1844 best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1845 best_ni->ni_net->net_seq++;
1846 best_ni->ni_seq = best_ni->ni_net->net_seq;
1848 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1849 libcfs_nidstr(&best_ni->ni_nid),
1850 best_ni->ni_seq, best_ni->ni_net->net_seq,
1851 atomic_read(&best_ni->ni_tx_credits),
1852 best_ni->ni_sel_priority,
1853 libcfs_nidstr(&best_lpni->lpni_nid),
1854 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1855 best_lpni->lpni_txcredits,
1856 best_lpni->lpni_sel_priority);
1859 * grab a reference on the peer_ni so it sticks around even if
1860 * we need to drop and relock the lnet_net_lock below.
1862 lnet_peer_ni_addref_locked(best_lpni);
1865 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1866 * message. This ensures that we get a CPT that is correct for
1867 * the NI when the NI has been restricted to a subset of all CPTs.
1868 * If the selected CPT differs from the one currently locked, we
1869 * must unlock and relock the lnet_net_lock(), and then check whether
1870 * the configuration has changed. We don't have a hold on the best_ni
1871 * yet, and it may have vanished.
1873 cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1874 if (sd->sd_cpt != cpt2) {
1875 __u32 seq = lnet_get_dlc_seq_locked();
1876 lnet_net_unlock(sd->sd_cpt);
1878 lnet_net_lock(sd->sd_cpt);
1879 if (seq != lnet_get_dlc_seq_locked()) {
1880 lnet_peer_ni_decref_locked(best_lpni);
1886 * store the best_lpni in the message right away to avoid having
1887 * to do the same operation under different conditions
1889 msg->msg_txpeer = best_lpni;
1890 msg->msg_txni = best_ni;
1893 * grab a reference for the best_ni since now it's in use in this
1894 * send. The reference will be dropped in lnet_finalize()
1896 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1899 * Always set the target.nid to the best peer picked. Either the
1900 * NID will be one of the peer NIDs selected, or the same NID as
1901 * what was originally set in the target or it will be the NID of
1902 * a router if this message should be routed
1904 /* FIXME handle large-addr nids */
1905 msg->msg_target.nid = lnet_nid_to_nid4(&msg->msg_txpeer->lpni_nid);
1908 * lnet_msg_commit assigns the correct cpt to the message, which
1909 * is used to decrement the correct refcount on the ni when it's
1910 * time to return the credits
1912 lnet_msg_commit(msg, sd->sd_cpt);
1915 * If we are routing the message then we keep the src_nid that was
1916 * set by the originator. If we are not routing then we are the
1917 * originator and set it here.
1919 if (!msg->msg_routing)
1920 msg->msg_hdr.src_nid =
1921 cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txni->ni_nid));
1924 msg->msg_target_is_router = 1;
1925 msg->msg_target.pid = LNET_PID_LUSTRE;
1927 * since we're routing we want to ensure that the
1928 * msg_hdr.dest_nid is set to the final destination. When
1929 * the router receives this message it knows how to route
1932 * final_dst_lpni is set at the beginning of the
1933 * lnet_select_pathway() function and is never changed.
1934 * It's safe to use it here.
1936 /* FIXME handle large-addr nid */
1937 msg->msg_hdr.dest_nid =
1938 cpu_to_le64(lnet_nid_to_nid4(&final_dst_lpni->lpni_nid));
1941 * if we're not routing set the dest_nid to the best peer
1942 * ni NID that we picked earlier in the algorithm.
1944 msg->msg_hdr.dest_nid =
1945 cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txpeer->lpni_nid));
1949 * if we have response tracker block update it with the next hop
1953 rspt = msg->msg_md->md_rspt_ptr;
1955 rspt->rspt_next_hop_nid =
1956 msg->msg_txpeer->lpni_nid;
1957 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1958 libcfs_nidstr(&rspt->rspt_next_hop_nid));
1962 rc = lnet_post_send_locked(msg, 0);
1965 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1966 libcfs_nid2str(msg->msg_hdr.src_nid),
1967 libcfs_nidstr(&msg->msg_txni->ni_nid),
1968 libcfs_nid2str(sd->sd_src_nid),
1969 libcfs_nid2str(msg->msg_hdr.dest_nid),
1970 libcfs_nid2str(sd->sd_dst_nid),
1971 libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
1972 libcfs_nid2str(sd->sd_rtr_nid),
1973 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1979 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1980 struct lnet_msg *msg)
1982 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1983 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1984 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1985 libcfs_nidstr(&lni->ni_nid),
1986 libcfs_nidstr(&lpni->lpni_nid));
1987 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
1996 * use the source and destination NIDs as the pathway
1999 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2001 /* the destination lpni is set before we get here. */
2004 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2005 if (!sd->sd_best_ni) {
2006 CERROR("Can't send to %s: src %s is not a "
2007 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2008 libcfs_nid2str(sd->sd_src_nid));
2012 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2014 return lnet_handle_send(sd);
2022 * Don't run the selection algorithm on the peer NIs. By specifying the
2023 * local NID, we're also saying that we should always use the destination NID
2024 * provided. This handles the case where we should be using the same
2025 * destination NID for the all the messages which belong to the same RPC
2029 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2031 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2032 if (!sd->sd_best_ni) {
2033 CERROR("Can't send to %s: src %s is not a "
2034 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2035 libcfs_nid2str(sd->sd_src_nid));
2039 if (sd->sd_best_lpni &&
2040 nid_same(&sd->sd_best_lpni->lpni_nid,
2041 &the_lnet.ln_loni->ni_nid))
2042 return lnet_handle_lo_send(sd);
2043 else if (sd->sd_best_lpni)
2044 return lnet_handle_send(sd);
2046 CERROR("can't send to %s. no NI on %s\n",
2047 libcfs_nid2str(sd->sd_dst_nid),
2048 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2050 return -EHOSTUNREACH;
2054 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2055 struct lnet_peer *peer,
2056 struct lnet_peer_net *peer_net,
2057 struct lnet_msg *msg,
2060 struct lnet_net *local_net;
2061 struct lnet_ni *best_ni;
2063 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2068 * Iterate through the NIs in this local Net and select
2069 * the NI to send from. The selection is determined by
2070 * these 3 criterion in the following priority:
2072 * 2. NI available credits
2075 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2076 peer, peer_net, msg, cpt);
2082 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2085 struct lnet_peer *peer;
2086 struct lnet_peer_ni *new_lpni;
2089 lnet_peer_ni_addref_locked(lpni);
2091 peer = lpni->lpni_peer_net->lpn_peer;
2093 if (lnet_peer_gw_discovery(peer)) {
2094 lnet_peer_ni_decref_locked(lpni);
2098 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2099 lnet_peer_ni_decref_locked(lpni);
2103 rc = lnet_discover_peer_locked(lpni, cpt, false);
2105 lnet_peer_ni_decref_locked(lpni);
2109 new_lpni = lnet_find_peer_ni_locked(lnet_nid_to_nid4(&lpni->lpni_nid));
2111 lnet_peer_ni_decref_locked(lpni);
2115 peer = new_lpni->lpni_peer_net->lpn_peer;
2116 spin_lock(&peer->lp_lock);
2117 if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2118 /* The peer NI did not change and the peer is up to date.
2119 * Nothing more to do.
2121 spin_unlock(&peer->lp_lock);
2122 lnet_peer_ni_decref_locked(lpni);
2123 lnet_peer_ni_decref_locked(new_lpni);
2126 spin_unlock(&peer->lp_lock);
2128 /* Either the peer NI changed during discovery, or the peer isn't up
2129 * to date. In both cases we want to queue the message on the
2130 * (possibly new) peer's pending queue and queue the peer for discovery
2132 msg->msg_sending = 0;
2133 msg->msg_txpeer = NULL;
2134 lnet_net_unlock(cpt);
2135 lnet_peer_queue_message(peer, msg);
2138 lnet_peer_ni_decref_locked(lpni);
2139 lnet_peer_ni_decref_locked(new_lpni);
2141 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2142 msg, libcfs_nidstr(&peer->lp_primary_nid));
2144 return LNET_DC_WAIT;
2148 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2150 struct lnet_peer_ni **gw_lpni,
2151 struct lnet_peer **gw_peer)
2154 struct lnet_peer *gw;
2155 struct lnet_peer *lp;
2156 struct lnet_peer_net *lpn;
2157 struct lnet_peer_net *best_lpn = NULL;
2158 struct lnet_remotenet *rnet, *best_rnet = NULL;
2159 struct lnet_route *best_route = NULL;
2160 struct lnet_route *last_route = NULL;
2161 struct lnet_peer_ni *lpni = NULL;
2162 struct lnet_peer_ni *gwni = NULL;
2163 bool route_found = false;
2164 lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
2165 (sd->sd_best_ni != NULL)
2166 ? lnet_nid_to_nid4(&sd->sd_best_ni->ni_nid)
2168 int best_lpn_healthv = 0;
2169 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2171 CDEBUG(D_NET, "using src nid %s for route restriction\n",
2172 libcfs_nid2str(src_nid));
2174 /* If a router nid was specified then we are replying to a GET or
2175 * sending an ACK. In this case we use the gateway associated with the
2176 * specified router nid.
2178 if (sd->sd_rtr_nid != LNET_NID_ANY) {
2179 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2181 gw = gwni->lpni_peer_net->lpn_peer;
2182 lnet_peer_ni_decref_locked(gwni);
2183 if (gw->lp_rtr_refcount)
2186 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2187 libcfs_nid2str(sd->sd_rtr_nid));
2192 if (sd->sd_msg->msg_routing) {
2193 /* If I'm routing this message then I need to find the
2194 * next hop based on the destination NID
2196 best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
2198 CERROR("Unable to route message to %s - Route table may be misconfigured\n",
2199 libcfs_nid2str(sd->sd_dst_nid));
2200 return -EHOSTUNREACH;
2203 /* we've already looked up the initial lpni using
2206 lpni = sd->sd_best_lpni;
2207 /* the peer tree must be in existence */
2208 LASSERT(lpni && lpni->lpni_peer_net &&
2209 lpni->lpni_peer_net->lpn_peer);
2210 lp = lpni->lpni_peer_net->lpn_peer;
2212 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2213 /* is this remote network reachable? */
2214 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2223 /* select the preferred peer net */
2224 if (best_lpn_healthv > lpn->lpn_healthv)
2226 else if (best_lpn_healthv < lpn->lpn_healthv)
2229 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2231 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2234 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2237 best_lpn_healthv = lpn->lpn_healthv;
2238 best_lpn_sel_prio = lpn->lpn_sel_priority;
2244 CERROR("peer %s has no available nets\n",
2245 libcfs_nid2str(sd->sd_dst_nid));
2246 return -EHOSTUNREACH;
2249 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2252 best_lpn->lpn_net_id);
2253 if (!sd->sd_best_lpni) {
2254 CERROR("peer %s is unreachable\n",
2255 libcfs_nid2str(sd->sd_dst_nid));
2256 return -EHOSTUNREACH;
2259 /* We're attempting to round robin over the remote peer
2260 * NI's so update the final destination we selected
2262 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2264 /* Increment the sequence number of the remote lpni so
2265 * we can round robin over the different interfaces of
2268 sd->sd_best_lpni->lpni_seq++;
2272 * find the best route. Restrict the selection on the net of the
2273 * local NI if we've already picked the local NI to send from.
2274 * Otherwise, let's pick any route we can find and then find
2275 * a local NI we can reach the route's gateway on. Any route we
2276 * select will be reachable by virtue of the restriction we have
2277 * when adding a route.
2279 best_route = lnet_find_route_locked(best_rnet,
2280 LNET_NIDNET(src_nid),
2282 &last_route, &gwni);
2285 CERROR("no route to %s from %s\n",
2286 libcfs_nid2str(dst_nid),
2287 libcfs_nid2str(src_nid));
2288 return -EHOSTUNREACH;
2292 CERROR("Internal Error. Route expected to %s from %s\n",
2293 libcfs_nid2str(dst_nid),
2294 libcfs_nid2str(src_nid));
2298 gw = best_route->lr_gateway;
2299 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2303 * Discover this gateway if it hasn't already been discovered.
2304 * This means we might delay the message until discovery has
2307 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2311 if (!sd->sd_best_ni) {
2312 lpn = gwni->lpni_peer_net;
2313 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2316 if (!sd->sd_best_ni) {
2317 CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2318 libcfs_net2str(lpn->lpn_net_id),
2319 libcfs_nid2str(sd->sd_src_nid));
2328 * increment the sequence numbers since now we're sure we're
2329 * going to use this path
2331 if (sd->sd_rtr_nid == LNET_NID_ANY) {
2332 LASSERT(best_route && last_route);
2333 best_route->lr_seq = last_route->lr_seq + 1;
2335 best_lpn->lpn_seq++;
2346 * Remote destination
2347 * Non-MR destination
2351 * Remote destination
2354 * The handling of these two cases is similar. Even though the destination
2355 * can be MR or non-MR, we'll deal directly with the router.
2358 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2361 struct lnet_peer_ni *gw_lpni = NULL;
2362 struct lnet_peer *gw_peer = NULL;
2365 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2366 if (!sd->sd_best_ni) {
2367 CERROR("Can't send to %s: src %s is not a "
2368 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2369 libcfs_nid2str(sd->sd_src_nid));
2373 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2378 if (sd->sd_send_case & NMR_DST)
2380 * since the final destination is non-MR let's set its preferred
2381 * NID before we send
2383 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2387 * We're going to send to the gw found so let's set its
2390 sd->sd_peer = gw_peer;
2391 sd->sd_best_lpni = gw_lpni;
2393 return lnet_handle_send(sd);
2397 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2398 struct lnet_msg *msg, bool discovery)
2400 struct lnet_peer_net *lpn = NULL;
2401 struct lnet_peer_net *best_lpn = NULL;
2402 struct lnet_net *net = NULL;
2403 struct lnet_net *best_net = NULL;
2404 struct lnet_ni *best_ni = NULL;
2405 int best_lpn_healthv = 0;
2406 int best_net_healthv = 0;
2408 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2410 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2415 * The peer can have multiple interfaces, some of them can be on
2416 * the local network and others on a routed network. We should
2417 * prefer the local network. However if the local network is not
2418 * available then we need to try the routed network
2421 /* go through all the peer nets and find the best_ni */
2422 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2424 * The peer's list of nets can contain non-local nets. We
2425 * want to only examine the local ones.
2427 net = lnet_get_net_locked(lpn->lpn_net_id);
2431 lpn_sel_prio = lpn->lpn_sel_priority;
2432 net_healthv = lnet_get_net_healthv_locked(net);
2433 net_sel_prio = net->net_sel_priority;
2436 * if this is a discovery message and lp_disc_net_id is
2437 * specified then use that net to send the discovery on.
2439 if (peer->lp_disc_net_id == lpn->lpn_net_id &&
2448 /* always select the lpn with the best health */
2449 if (best_lpn_healthv > lpn->lpn_healthv)
2451 else if (best_lpn_healthv < lpn->lpn_healthv)
2454 /* select the preferred peer and local nets */
2455 if (best_lpn_sel_prio < lpn_sel_prio)
2457 else if (best_lpn_sel_prio > lpn_sel_prio)
2460 if (best_net_healthv > net_healthv)
2462 else if (best_net_healthv < net_healthv)
2465 if (best_net_sel_prio < net_sel_prio)
2467 else if (best_net_sel_prio > net_sel_prio)
2470 if (best_lpn->lpn_seq < lpn->lpn_seq)
2472 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2475 /* round robin over the local networks */
2476 if (best_net->net_seq <= net->net_seq)
2480 best_net_healthv = net_healthv;
2481 best_net_sel_prio = net_sel_prio;
2482 best_lpn_healthv = lpn->lpn_healthv;
2483 best_lpn_sel_prio = lpn_sel_prio;
2492 /* Select the best NI on the same net as best_lpn chosen
2495 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2502 static struct lnet_ni *
2503 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2505 struct lnet_ni *best_ni = NULL;
2506 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2507 struct lnet_peer_ni *lpni_entry;
2510 * We must use a consistent source address when sending to a
2511 * non-MR peer. However, a non-MR peer can have multiple NIDs
2512 * on multiple networks, and we may even need to talk to this
2513 * peer on multiple networks -- certain types of
2514 * load-balancing configuration do this.
2516 * So we need to pick the NI the peer prefers for this
2517 * particular network.
2520 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2522 if (lpni_entry->lpni_pref_nnids == 0)
2524 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2525 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2533 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2535 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2537 struct lnet_ni *best_ni = NULL;
2538 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2541 * We must use a consistent source address when sending to a
2542 * non-MR peer. However, a non-MR peer can have multiple NIDs
2543 * on multiple networks, and we may even need to talk to this
2544 * peer on multiple networks -- certain types of
2545 * load-balancing configuration do this.
2547 * So we need to pick the NI the peer prefers for this
2548 * particular network.
2551 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2554 /* if best_ni is still not set just pick one */
2557 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2558 sd->sd_best_lpni->lpni_peer_net,
2561 /* If there is no best_ni we don't have a route */
2563 CERROR("no path to %s from net %s\n",
2564 libcfs_nidstr(&best_lpni->lpni_nid),
2565 libcfs_net2str(best_lpni->lpni_net->net_id));
2566 return -EHOSTUNREACH;
2570 sd->sd_best_ni = best_ni;
2572 /* Set preferred NI if necessary. */
2573 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2580 * Source not specified
2584 * always use the same source NID for NMR peers
2585 * If we've talked to that peer before then we already have a preferred
2586 * source NI associated with it. Otherwise, we select a preferred local NI
2587 * and store it in the peer
2590 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2594 /* sd->sd_best_lpni is already set to the final destination */
2597 * At this point we should've created the peer ni and peer. If we
2598 * can't find it, then something went wrong. Instead of assert
2599 * output a relevant message and fail the send
2601 if (!sd->sd_best_lpni) {
2602 CERROR("Internal fault. Unable to send msg %s to %s. "
2604 lnet_msgtyp2str(sd->sd_msg->msg_type),
2605 libcfs_nid2str(sd->sd_dst_nid));
2609 if (sd->sd_msg->msg_routing) {
2610 /* If I'm forwarding this message then I can choose any NI
2611 * on the destination peer net
2613 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2615 sd->sd_best_lpni->lpni_peer_net,
2618 if (!sd->sd_best_ni) {
2619 CERROR("Unable to forward message to %s. No local NI available\n",
2620 libcfs_nid2str(sd->sd_dst_nid));
2624 rc = lnet_select_preferred_best_ni(sd);
2627 rc = lnet_handle_send(sd);
2633 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2636 * NOTE we've already handled the remote peer case. So we only
2637 * need to worry about the local case here.
2639 * if we're sending a response, ACK or reply, we need to send it
2640 * to the destination NID given to us. At this point we already
2641 * have the peer_ni we're suppose to send to, so just find the
2642 * best_ni on the peer net and use that. Since we're sending to an
2643 * MR peer then we can just run the selection algorithm on our
2644 * local NIs and pick the best one.
2646 if (sd->sd_send_case & SND_RESP) {
2648 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2649 sd->sd_best_lpni->lpni_peer_net,
2653 if (!sd->sd_best_ni) {
2655 * We're not going to deal with not able to send
2656 * a response to the provided final destination
2658 CERROR("Can't send response to %s. "
2659 "No local NI available\n",
2660 libcfs_nid2str(sd->sd_dst_nid));
2661 return -EHOSTUNREACH;
2664 return lnet_handle_send(sd);
2668 * If we get here that means we're sending a fresh request, PUT or
2669 * GET, so we need to run our standard selection algorithm.
2670 * First find the best local interface that's on any of the peer's
2673 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2676 lnet_msg_discovery(sd->sd_msg));
2677 if (sd->sd_best_ni) {
2679 lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2681 sd->sd_best_ni->ni_net->net_id);
2684 * if we're successful in selecting a peer_ni on the local
2685 * network, then send to it. Otherwise fall through and
2686 * try and see if we can reach it over another routed
2689 if (sd->sd_best_lpni &&
2690 nid_same(&sd->sd_best_lpni->lpni_nid,
2691 &the_lnet.ln_loni->ni_nid)) {
2693 * in case we initially started with a routed
2694 * destination, let's reset to local
2696 sd->sd_send_case &= ~REMOTE_DST;
2697 sd->sd_send_case |= LOCAL_DST;
2698 return lnet_handle_lo_send(sd);
2699 } else if (sd->sd_best_lpni) {
2701 * in case we initially started with a routed
2702 * destination, let's reset to local
2704 sd->sd_send_case &= ~REMOTE_DST;
2705 sd->sd_send_case |= LOCAL_DST;
2706 return lnet_handle_send(sd);
2709 CERROR("Internal Error. Expected to have a best_lpni: "
2711 libcfs_nid2str(sd->sd_src_nid),
2712 libcfs_nid2str(sd->sd_dst_nid));
2718 * Peer doesn't have a local network. Let's see if there is
2719 * a remote network we can reach it on.
2721 return PASS_THROUGH;
2726 * Source NID not specified
2731 * Source NID not speified
2732 * Remote destination
2735 * In both of these cases if we're sending a response, ACK or REPLY, then
2736 * we need to send to the destination NID provided.
2738 * In the remote case let's deal with MR routers.
2743 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2746 struct lnet_peer *gw_peer = NULL;
2747 struct lnet_peer_ni *gw_lpni = NULL;
2750 * handle sending a response to a remote peer here so we don't
2751 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2753 if (sd->sd_send_case & REMOTE_DST &&
2754 sd->sd_send_case & SND_RESP) {
2755 struct lnet_peer_ni *gw;
2756 struct lnet_peer *gw_peer;
2758 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2761 CERROR("Can't send response to %s. "
2762 "No route available\n",
2763 libcfs_nid2str(sd->sd_dst_nid));
2764 return -EHOSTUNREACH;
2765 } else if (rc > 0) {
2769 sd->sd_best_lpni = gw;
2770 sd->sd_peer = gw_peer;
2772 return lnet_handle_send(sd);
2776 * Even though the NID for the peer might not be on a local network,
2777 * since the peer is MR there could be other interfaces on the
2778 * local network. In that case we'd still like to prefer the local
2779 * network over the routed network. If we're unable to do that
2780 * then we select the best router among the different routed networks,
2781 * and if the router is MR then we can deal with it as such.
2783 rc = lnet_handle_any_mr_dsta(sd);
2784 if (rc != PASS_THROUGH)
2788 * Now that we must route to the destination, we must consider the
2789 * MR case, where the destination has multiple interfaces, some of
2790 * which we can route to and others we do not. For this reason we
2791 * need to select the destination which we can route to and if
2792 * there are multiple, we need to round robin.
2794 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2799 sd->sd_send_case &= ~LOCAL_DST;
2800 sd->sd_send_case |= REMOTE_DST;
2802 sd->sd_peer = gw_peer;
2803 sd->sd_best_lpni = gw_lpni;
2805 return lnet_handle_send(sd);
2809 * Source not specified
2810 * Remote destination
2813 * Must send to the specified peer NID using the same source NID that
2814 * we've used before. If it's the first time to talk to that peer then
2815 * find the source NI and assign it as preferred to that peer
2818 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2821 struct lnet_peer_ni *gw_lpni = NULL;
2822 struct lnet_peer *gw_peer = NULL;
2825 * Let's see if we have a preferred NI to talk to this NMR peer
2827 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2831 * find the router and that'll find the best NI if we didn't find
2834 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2840 * set the best_ni we've chosen as the preferred one for
2843 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2845 /* we'll be sending to the gw */
2846 sd->sd_best_lpni = gw_lpni;
2847 sd->sd_peer = gw_peer;
2849 return lnet_handle_send(sd);
2853 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2856 * turn off the SND_RESP bit.
2857 * It will be checked in the case handling
2859 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2861 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2862 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2863 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2864 (send_case & MR_DST) ? "MR: " : "NMR: ",
2865 libcfs_nid2str(sd->sd_dst_nid),
2866 (send_case & LOCAL_DST) ? "local" : "routed");
2868 switch (send_case) {
2870 * For all cases where the source is specified, we should always
2871 * use the destination NID, whether it's an MR destination or not,
2872 * since we're continuing a series of related messages for the
2875 case SRC_SPEC_LOCAL_NMR_DST:
2876 return lnet_handle_spec_local_nmr_dst(sd);
2877 case SRC_SPEC_LOCAL_MR_DST:
2878 return lnet_handle_spec_local_mr_dst(sd);
2879 case SRC_SPEC_ROUTER_NMR_DST:
2880 case SRC_SPEC_ROUTER_MR_DST:
2881 return lnet_handle_spec_router_dst(sd);
2882 case SRC_ANY_LOCAL_NMR_DST:
2883 return lnet_handle_any_local_nmr_dst(sd);
2884 case SRC_ANY_LOCAL_MR_DST:
2885 case SRC_ANY_ROUTER_MR_DST:
2886 return lnet_handle_any_mr_dst(sd);
2887 case SRC_ANY_ROUTER_NMR_DST:
2888 return lnet_handle_any_router_nmr_dst(sd);
2890 CERROR("Unknown send case\n");
2896 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2897 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2899 struct lnet_peer_ni *lpni;
2900 struct lnet_peer *peer;
2901 struct lnet_send_data send_data;
2904 __u32 send_case = 0;
2906 bool mr_forwarding_allowed;
2908 memset(&send_data, 0, sizeof(send_data));
2911 * get an initial CPT to use for locking. The idea here is not to
2912 * serialize the calls to select_pathway, so that as many
2913 * operations can run concurrently as possible. To do that we use
2914 * the CPT where this call is being executed. Later on when we
2915 * determine the CPT to use in lnet_message_commit, we switch the
2916 * lock and check if there was any configuration change. If none,
2917 * then we proceed, if there is, then we restart the operation.
2919 cpt = lnet_net_lock_current();
2921 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2922 if (md_cpt == CFS_CPT_ANY)
2928 * If we're being asked to send to the loopback interface, there
2929 * is no need to go through any selection. We can just shortcut
2930 * the entire process and send over lolnd
2932 send_data.sd_msg = msg;
2933 send_data.sd_cpt = cpt;
2934 if (dst_nid == LNET_NID_LO_0) {
2935 rc = lnet_handle_lo_send(&send_data);
2936 lnet_net_unlock(cpt);
2941 * find an existing peer_ni, or create one and mark it as having been
2942 * created due to network traffic. This call will create the
2943 * peer->peer_net->peer_ni tree.
2945 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2947 lnet_net_unlock(cpt);
2948 return PTR_ERR(lpni);
2952 * Cache the original src_nid and rtr_nid. If we need to resend the
2953 * message then we'll need to know whether the src_nid was originally
2954 * specified for this message. If it was originally specified,
2955 * then we need to keep using the same src_nid since it's
2956 * continuing the same sequence of messages. Similarly, rtr_nid will
2957 * affect our choice of next hop.
2959 msg->msg_src_nid_param = src_nid;
2960 msg->msg_rtr_nid_param = rtr_nid;
2963 * If necessary, perform discovery on the peer that owns this peer_ni.
2964 * Note, this can result in the ownership of this peer_ni changing
2965 * to another peer object.
2967 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2969 lnet_peer_ni_decref_locked(lpni);
2970 lnet_net_unlock(cpt);
2973 lnet_peer_ni_decref_locked(lpni);
2975 peer = lpni->lpni_peer_net->lpn_peer;
2978 * Identify the different send cases
2980 if (src_nid == LNET_NID_ANY)
2981 send_case |= SRC_ANY;
2983 send_case |= SRC_SPEC;
2985 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2986 send_case |= LOCAL_DST;
2988 send_case |= REMOTE_DST;
2991 if (msg->msg_routing && (send_case & LOCAL_DST))
2994 /* Determine whether to allow MR forwarding for this message.
2995 * NB: MR forwarding is allowed if the message originator and the
2996 * destination are both MR capable, and the destination lpni that was
2997 * originally chosen by the originator is unhealthy or down.
2998 * We check the MR capability of the destination further below
3000 mr_forwarding_allowed = false;
3002 struct lnet_peer *src_lp;
3003 struct lnet_peer_ni *src_lpni;
3005 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
3007 /* We don't fail the send if we hit any errors here. We'll just
3008 * try to send it via non-multi-rail criteria
3010 if (!IS_ERR(src_lpni)) {
3011 /* Drop ref taken by lnet_nid2peerni_locked() */
3012 lnet_peer_ni_decref_locked(src_lpni);
3013 src_lp = lpni->lpni_peer_net->lpn_peer;
3014 if (lnet_peer_is_multi_rail(src_lp) &&
3015 !lnet_is_peer_ni_alive(lpni))
3016 mr_forwarding_allowed = true;
3019 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3020 mr_forwarding_allowed ? "allowed" : "not allowed");
3024 * Deal with the peer as NMR in the following cases:
3025 * 1. the peer is NMR
3026 * 2. We're trying to recover a specific peer NI
3027 * 3. I'm a router sending to the final destination and MR forwarding is
3028 * not allowed for this message (as determined above).
3029 * In this case the source of the message would've
3030 * already selected the final destination so my job
3031 * is to honor the selection.
3033 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3034 (final_hop && !mr_forwarding_allowed))
3035 send_case |= NMR_DST;
3037 send_case |= MR_DST;
3039 if (lnet_msg_is_response(msg))
3040 send_case |= SND_RESP;
3042 /* assign parameters to the send_data */
3043 send_data.sd_rtr_nid = rtr_nid;
3044 send_data.sd_src_nid = src_nid;
3045 send_data.sd_dst_nid = dst_nid;
3046 send_data.sd_best_lpni = lpni;
3048 * keep a pointer to the final destination in case we're going to
3049 * route, so we'll need to access it later
3051 send_data.sd_final_dst_lpni = lpni;
3052 send_data.sd_peer = peer;
3053 send_data.sd_md_cpt = md_cpt;
3054 send_data.sd_send_case = send_case;
3056 rc = lnet_handle_send_case_locked(&send_data);
3059 * Update the local cpt since send_data.sd_cpt might've been
3060 * updated as a result of calling lnet_handle_send_case_locked().
3062 cpt = send_data.sd_cpt;
3064 if (rc == REPEAT_SEND)
3067 lnet_net_unlock(cpt);
3073 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
3075 lnet_nid_t dst_nid = msg->msg_target.nid;
3079 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
3080 * but we might want to use pre-determined router for ACK/REPLY
3083 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3084 LASSERT(msg->msg_txpeer == NULL);
3085 LASSERT(msg->msg_txni == NULL);
3086 LASSERT(!msg->msg_sending);
3087 LASSERT(!msg->msg_target_is_router);
3088 LASSERT(!msg->msg_receiving);
3090 msg->msg_sending = 1;
3092 LASSERT(!msg->msg_tx_committed);
3094 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3096 if (rc == -EHOSTUNREACH)
3097 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3099 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3103 if (rc == LNET_CREDIT_OK)
3104 lnet_ni_send(msg->msg_txni, msg);
3106 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3110 enum lnet_mt_event_type {
3111 MT_TYPE_LOCAL_NI = 0,
3115 struct lnet_mt_event_info {
3116 enum lnet_mt_event_type mt_type;
3120 /* called with res_lock held */
3122 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3124 struct lnet_rsp_tracker *rspt;
3127 * msg has a refcount on the MD so the MD is not going away.
3128 * The rspt queue for the cpt is protected by
3129 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3131 if (!md->md_rspt_ptr)
3134 rspt = md->md_rspt_ptr;
3137 LASSERT(rspt->rspt_cpt == cpt);
3139 md->md_rspt_ptr = NULL;
3141 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3143 * The monitor thread has invalidated this handle because the
3144 * response timed out, but it failed to lookup the MD. That
3145 * means this response tracker is on the zombie list. We can
3146 * safely remove it under the resource lock (held by caller) and
3147 * free the response tracker block.
3149 list_del(&rspt->rspt_on_list);
3150 lnet_rspt_free(rspt, cpt);
3153 * invalidate the handle to indicate that a response has been
3154 * received, which will then lead the monitor thread to clean up
3157 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3162 lnet_clean_zombie_rstqs(void)
3164 struct lnet_rsp_tracker *rspt, *tmp;
3167 cfs_cpt_for_each(i, lnet_cpt_table()) {
3168 list_for_each_entry_safe(rspt, tmp,
3169 the_lnet.ln_mt_zombie_rstqs[i],
3171 list_del(&rspt->rspt_on_list);
3172 lnet_rspt_free(rspt, i);
3176 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3180 lnet_finalize_expired_responses(void)
3182 struct lnet_libmd *md;
3183 struct lnet_rsp_tracker *rspt, *tmp;
3187 if (the_lnet.ln_mt_rstq == NULL)
3190 cfs_cpt_for_each(i, lnet_cpt_table()) {
3191 LIST_HEAD(local_queue);
3194 if (!the_lnet.ln_mt_rstq[i]) {
3198 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3203 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3205 * The rspt mdh will be invalidated when a response
3206 * is received or whenever we want to discard the
3207 * block the monitor thread will walk the queue
3208 * and clean up any rsts with an invalid mdh.
3209 * The monitor thread will walk the queue until
3210 * the first unexpired rspt block. This means that
3211 * some rspt blocks which received their
3212 * corresponding responses will linger in the
3213 * queue until they are cleaned up eventually.
3216 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3218 list_del(&rspt->rspt_on_list);
3219 lnet_rspt_free(rspt, i);
3223 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3224 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {