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 nid, unsigned int threshold)
195 struct lnet_test_peer *tp;
196 struct list_head *el;
197 struct list_head *next;
200 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
201 if (threshold != 0) {
202 /* Adding a new entry */
203 LIBCFS_ALLOC(tp, sizeof(*tp));
208 tp->tp_threshold = threshold;
211 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
218 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
219 tp = list_entry(el, struct lnet_test_peer, tp_list);
221 if (tp->tp_threshold == 0 || /* needs culling anyway */
222 nid == LNET_NID_ANY || /* removing all entries */
223 tp->tp_nid == nid) { /* matched this one */
224 list_move(&tp->tp_list, &cull);
230 while (!list_empty(&cull)) {
231 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
233 list_del(&tp->tp_list);
234 LIBCFS_FREE(tp, sizeof(*tp));
240 fail_peer (lnet_nid_t nid, int outgoing)
242 struct lnet_test_peer *tp;
243 struct list_head *el;
244 struct list_head *next;
248 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
251 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
252 tp = list_entry(el, struct lnet_test_peer, tp_list);
254 if (tp->tp_threshold == 0) {
257 /* only cull zombies on outgoing tests,
258 * since we may be at interrupt priority on
259 * incoming messages. */
260 list_move(&tp->tp_list, &cull);
265 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
266 nid == tp->tp_nid) { /* fail this peer */
269 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
272 tp->tp_threshold == 0) {
274 list_move(&tp->tp_list, &cull);
283 while (!list_empty(&cull)) {
284 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
285 list_del(&tp->tp_list);
287 LIBCFS_FREE(tp, sizeof(*tp));
294 lnet_iov_nob(unsigned int niov, struct kvec *iov)
296 unsigned int nob = 0;
298 LASSERT(niov == 0 || iov != NULL);
300 nob += (iov++)->iov_len;
304 EXPORT_SYMBOL(lnet_iov_nob);
307 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
308 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
311 /* NB diov, siov are READ-ONLY */
312 unsigned int this_nob;
317 /* skip complete frags before 'doffset' */
319 while (doffset >= diov->iov_len) {
320 doffset -= diov->iov_len;
326 /* skip complete frags before 'soffset' */
328 while (soffset >= siov->iov_len) {
329 soffset -= siov->iov_len;
338 this_nob = min3((unsigned int)diov->iov_len - doffset,
339 (unsigned int)siov->iov_len - soffset,
342 memcpy((char *)diov->iov_base + doffset,
343 (char *)siov->iov_base + soffset, this_nob);
346 if (diov->iov_len > doffset + this_nob) {
354 if (siov->iov_len > soffset + this_nob) {
363 EXPORT_SYMBOL(lnet_copy_iov2iov);
366 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
368 unsigned int nob = 0;
370 LASSERT(niov == 0 || kiov != NULL);
372 nob += (kiov++)->bv_len;
376 EXPORT_SYMBOL(lnet_kiov_nob);
379 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
380 unsigned int doffset,
381 unsigned int nsiov, struct bio_vec *siov,
382 unsigned int soffset,
385 /* NB diov, siov are READ-ONLY */
386 unsigned int this_nob;
393 LASSERT (!in_interrupt ());
396 while (doffset >= diov->bv_len) {
397 doffset -= diov->bv_len;
404 while (soffset >= siov->bv_len) {
405 soffset -= siov->bv_len;
414 this_nob = min3(diov->bv_len - doffset,
415 siov->bv_len - soffset,
419 daddr = ((char *)kmap(diov->bv_page)) +
420 diov->bv_offset + doffset;
422 saddr = ((char *)kmap(siov->bv_page)) +
423 siov->bv_offset + soffset;
425 /* Vanishing risk of kmap deadlock when mapping 2 pages.
426 * However in practice at least one of the kiovs will be mapped
427 * kernel pages and the map/unmap will be NOOPs */
429 memcpy (daddr, saddr, this_nob);
432 if (diov->bv_len > doffset + this_nob) {
436 kunmap(diov->bv_page);
443 if (siov->bv_len > soffset + this_nob) {
447 kunmap(siov->bv_page);
456 kunmap(diov->bv_page);
458 kunmap(siov->bv_page);
460 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
463 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
464 unsigned int nkiov, struct bio_vec *kiov,
465 unsigned int kiovoffset,
468 /* NB iov, kiov are READ-ONLY */
469 unsigned int this_nob;
475 LASSERT (!in_interrupt ());
478 while (iovoffset >= iov->iov_len) {
479 iovoffset -= iov->iov_len;
486 while (kiovoffset >= kiov->bv_len) {
487 kiovoffset -= kiov->bv_len;
496 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
497 (unsigned int)kiov->bv_len - kiovoffset,
501 addr = ((char *)kmap(kiov->bv_page)) +
502 kiov->bv_offset + kiovoffset;
504 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
507 if (iov->iov_len > iovoffset + this_nob) {
508 iovoffset += this_nob;
515 if (kiov->bv_len > kiovoffset + this_nob) {
517 kiovoffset += this_nob;
519 kunmap(kiov->bv_page);
529 kunmap(kiov->bv_page);
531 EXPORT_SYMBOL(lnet_copy_kiov2iov);
534 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
535 unsigned int kiovoffset,
536 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
539 /* NB kiov, iov are READ-ONLY */
540 unsigned int this_nob;
546 LASSERT (!in_interrupt ());
549 while (kiovoffset >= kiov->bv_len) {
550 kiovoffset -= kiov->bv_len;
557 while (iovoffset >= iov->iov_len) {
558 iovoffset -= iov->iov_len;
567 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
568 (unsigned int)iov->iov_len - iovoffset,
572 addr = ((char *)kmap(kiov->bv_page)) +
573 kiov->bv_offset + kiovoffset;
575 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
578 if (kiov->bv_len > kiovoffset + this_nob) {
580 kiovoffset += this_nob;
582 kunmap(kiov->bv_page);
589 if (iov->iov_len > iovoffset + this_nob) {
590 iovoffset += this_nob;
599 kunmap(kiov->bv_page);
601 EXPORT_SYMBOL(lnet_copy_iov2kiov);
604 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
605 int src_niov, struct bio_vec *src,
606 unsigned int offset, unsigned int len)
608 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
609 * for exactly 'len' bytes, and return the number of entries.
610 * NB not destructive to 'src' */
611 unsigned int frag_len;
614 if (len == 0) /* no data => */
615 return (0); /* no frags */
617 LASSERT(src_niov > 0);
618 while (offset >= src->bv_len) { /* skip initial frags */
619 offset -= src->bv_len;
622 LASSERT(src_niov > 0);
627 LASSERT(src_niov > 0);
628 LASSERT((int)niov <= dst_niov);
630 frag_len = src->bv_len - offset;
631 dst->bv_page = src->bv_page;
632 dst->bv_offset = src->bv_offset + offset;
634 if (len <= frag_len) {
636 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
640 dst->bv_len = frag_len;
641 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
651 EXPORT_SYMBOL(lnet_extract_kiov);
654 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
655 int delayed, unsigned int offset, unsigned int mlen,
658 unsigned int niov = 0;
659 struct kvec *iov = NULL;
660 struct bio_vec *kiov = NULL;
663 LASSERT (!in_interrupt ());
664 LASSERT (mlen == 0 || msg != NULL);
667 LASSERT(msg->msg_receiving);
668 LASSERT(!msg->msg_sending);
669 LASSERT(rlen == msg->msg_len);
670 LASSERT(mlen <= msg->msg_len);
671 LASSERT(msg->msg_offset == offset);
672 LASSERT(msg->msg_wanted == mlen);
674 msg->msg_receiving = 0;
677 niov = msg->msg_niov;
678 kiov = msg->msg_kiov;
681 LASSERT ((iov == NULL) != (kiov == NULL));
685 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
686 niov, kiov, offset, mlen,
689 lnet_finalize(msg, rc);
693 lnet_setpayloadbuffer(struct lnet_msg *msg)
695 struct lnet_libmd *md = msg->msg_md;
697 LASSERT(msg->msg_len > 0);
698 LASSERT(!msg->msg_routing);
700 LASSERT(msg->msg_niov == 0);
701 LASSERT(msg->msg_kiov == NULL);
703 msg->msg_niov = md->md_niov;
704 msg->msg_kiov = md->md_kiov;
708 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
709 unsigned int offset, unsigned int len)
711 msg->msg_type = type;
712 msg->msg_target = target;
714 msg->msg_offset = offset;
717 lnet_setpayloadbuffer(msg);
719 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
720 msg->msg_hdr.type = cpu_to_le32(type);
721 /* dest_nid will be overwritten by lnet_select_pathway() */
722 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
723 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
724 /* src_nid will be set later */
725 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
726 msg->msg_hdr.payload_length = cpu_to_le32(len);
730 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
732 void *priv = msg->msg_private;
735 LASSERT(!in_interrupt());
736 LASSERT(ni->ni_nid == LNET_NID_LO_0 ||
737 (msg->msg_txcredit && msg->msg_peertxcredit));
739 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
741 msg->msg_no_resend = true;
742 lnet_finalize(msg, rc);
747 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
751 LASSERT(!msg->msg_sending);
752 LASSERT(msg->msg_receiving);
753 LASSERT(!msg->msg_rx_ready_delay);
754 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
756 msg->msg_rx_ready_delay = 1;
757 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
760 CERROR("recv from %s / send to %s aborted: "
761 "eager_recv failed %d\n",
762 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
763 libcfs_id2str(msg->msg_target), rc);
764 LASSERT(rc < 0); /* required by my callers */
771 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
775 deadline = lpni->lpni_last_alive +
776 lpni->lpni_net->net_tunables.lct_peer_timeout;
779 * assume peer_ni is alive as long as we're within the configured
788 /* NB: returns 1 when alive, 0 when dead, negative when error;
789 * may drop the lnet_net_lock */
791 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
792 struct lnet_msg *msg)
794 time64_t now = ktime_get_seconds();
796 if (!lnet_peer_aliveness_enabled(lpni))
800 * If we're resending a message, let's attempt to send it even if
801 * the peer is down to fulfill our resend quota on the message
803 if (msg->msg_retry_count > 0)
806 /* try and send recovery messages irregardless */
807 if (msg->msg_recovery)
810 /* always send any responses */
811 if (lnet_msg_is_response(msg))
814 if (!lnet_is_peer_deadline_passed(lpni, now))
817 return lnet_is_peer_ni_alive(lpni);
821 * \param msg The message to be sent.
822 * \param do_send True if lnet_ni_send() should be called in this function.
823 * lnet_send() is going to lnet_net_unlock immediately after this, so
824 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
826 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
827 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
828 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
829 * \retval -ECANCELED If the MD of the message has been unlinked.
832 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
834 struct lnet_peer_ni *lp = msg->msg_txpeer;
835 struct lnet_ni *ni = msg->msg_txni;
836 int cpt = msg->msg_tx_cpt;
837 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
839 /* non-lnet_send() callers have checked before */
840 LASSERT(!do_send || msg->msg_tx_delayed);
841 LASSERT(!msg->msg_receiving);
842 LASSERT(msg->msg_tx_committed);
844 /* can't get here if we're sending to the loopback interface */
845 if (the_lnet.ln_loni)
846 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
848 /* NB 'lp' is always the next hop */
849 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
850 lnet_peer_alive_locked(ni, lp, msg) == 0) {
851 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
852 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
854 lnet_net_unlock(cpt);
856 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
858 LNET_STATS_TYPE_DROP);
860 lnet_incr_stats(&msg->msg_txni->ni_stats,
862 LNET_STATS_TYPE_DROP);
864 CNETERR("Dropping message for %s: peer not alive\n",
865 libcfs_id2str(msg->msg_target));
866 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
868 lnet_finalize(msg, -EHOSTUNREACH);
871 return -EHOSTUNREACH;
874 if (msg->msg_md != NULL &&
875 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
876 lnet_net_unlock(cpt);
878 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
879 "called on the MD/ME.\n",
880 libcfs_id2str(msg->msg_target));
882 msg->msg_no_resend = true;
883 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
884 msg, libcfs_id2str(msg->msg_target));
885 lnet_finalize(msg, -ECANCELED);
892 if (!msg->msg_peertxcredit) {
893 spin_lock(&lp->lpni_lock);
894 LASSERT((lp->lpni_txcredits < 0) ==
895 !list_empty(&lp->lpni_txq));
897 msg->msg_peertxcredit = 1;
898 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
899 lp->lpni_txcredits--;
901 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
902 lp->lpni_mintxcredits = lp->lpni_txcredits;
904 if (lp->lpni_txcredits < 0) {
905 msg->msg_tx_delayed = 1;
906 list_add_tail(&msg->msg_list, &lp->lpni_txq);
907 spin_unlock(&lp->lpni_lock);
908 return LNET_CREDIT_WAIT;
910 spin_unlock(&lp->lpni_lock);
913 if (!msg->msg_txcredit) {
914 LASSERT((tq->tq_credits < 0) ==
915 !list_empty(&tq->tq_delayed));
917 msg->msg_txcredit = 1;
919 atomic_dec(&ni->ni_tx_credits);
921 if (tq->tq_credits < tq->tq_credits_min)
922 tq->tq_credits_min = tq->tq_credits;
924 if (tq->tq_credits < 0) {
925 msg->msg_tx_delayed = 1;
926 list_add_tail(&msg->msg_list, &tq->tq_delayed);
927 return LNET_CREDIT_WAIT;
931 if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
932 lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
933 msg->msg_tx_delayed = 1;
934 return LNET_CREDIT_WAIT;
937 /* unset the tx_delay flag as we're going to send it now */
938 msg->msg_tx_delayed = 0;
941 lnet_net_unlock(cpt);
942 lnet_ni_send(ni, msg);
945 return LNET_CREDIT_OK;
949 static struct lnet_rtrbufpool *
950 lnet_msg2bufpool(struct lnet_msg *msg)
952 struct lnet_rtrbufpool *rbp;
955 LASSERT(msg->msg_rx_committed);
957 cpt = msg->msg_rx_cpt;
958 rbp = &the_lnet.ln_rtrpools[cpt][0];
960 LASSERT(msg->msg_len <= LNET_MTU);
961 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
963 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
970 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
972 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
973 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
974 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
975 * received or OK to receive */
976 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
977 struct lnet_peer *lp;
978 struct lnet_rtrbufpool *rbp;
979 struct lnet_rtrbuf *rb;
981 LASSERT(msg->msg_kiov == NULL);
982 LASSERT(msg->msg_niov == 0);
983 LASSERT(msg->msg_routing);
984 LASSERT(msg->msg_receiving);
985 LASSERT(!msg->msg_sending);
986 LASSERT(lpni->lpni_peer_net);
987 LASSERT(lpni->lpni_peer_net->lpn_peer);
989 lp = lpni->lpni_peer_net->lpn_peer;
991 /* non-lnet_parse callers only receive delayed messages */
992 LASSERT(!do_recv || msg->msg_rx_delayed);
994 if (!msg->msg_peerrtrcredit) {
995 /* lpni_lock protects the credit manipulation */
996 spin_lock(&lpni->lpni_lock);
998 msg->msg_peerrtrcredit = 1;
999 lpni->lpni_rtrcredits--;
1000 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1001 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1003 if (lpni->lpni_rtrcredits < 0) {
1004 spin_unlock(&lpni->lpni_lock);
1005 /* must have checked eager_recv before here */
1006 LASSERT(msg->msg_rx_ready_delay);
1007 msg->msg_rx_delayed = 1;
1008 /* lp_lock protects the lp_rtrq */
1009 spin_lock(&lp->lp_lock);
1010 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1011 spin_unlock(&lp->lp_lock);
1012 return LNET_CREDIT_WAIT;
1014 spin_unlock(&lpni->lpni_lock);
1017 rbp = lnet_msg2bufpool(msg);
1019 if (!msg->msg_rtrcredit) {
1020 msg->msg_rtrcredit = 1;
1022 if (rbp->rbp_credits < rbp->rbp_mincredits)
1023 rbp->rbp_mincredits = rbp->rbp_credits;
1025 if (rbp->rbp_credits < 0) {
1026 /* must have checked eager_recv before here */
1027 LASSERT(msg->msg_rx_ready_delay);
1028 msg->msg_rx_delayed = 1;
1029 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1030 return LNET_CREDIT_WAIT;
1034 LASSERT(!list_empty(&rbp->rbp_bufs));
1035 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1036 list_del(&rb->rb_list);
1038 msg->msg_niov = rbp->rbp_npages;
1039 msg->msg_kiov = &rb->rb_kiov[0];
1041 /* unset the msg-rx_delayed flag since we're receiving the message */
1042 msg->msg_rx_delayed = 0;
1045 int cpt = msg->msg_rx_cpt;
1047 lnet_net_unlock(cpt);
1048 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1049 0, msg->msg_len, msg->msg_len);
1052 return LNET_CREDIT_OK;
1056 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1058 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1059 struct lnet_ni *txni = msg->msg_txni;
1060 struct lnet_msg *msg2;
1062 if (msg->msg_txcredit) {
1063 struct lnet_ni *ni = msg->msg_txni;
1064 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1066 /* give back NI txcredits */
1067 msg->msg_txcredit = 0;
1069 LASSERT((tq->tq_credits < 0) ==
1070 !list_empty(&tq->tq_delayed));
1073 atomic_inc(&ni->ni_tx_credits);
1074 if (tq->tq_credits <= 0) {
1075 msg2 = list_entry(tq->tq_delayed.next,
1076 struct lnet_msg, msg_list);
1077 list_del(&msg2->msg_list);
1079 LASSERT(msg2->msg_txni == ni);
1080 LASSERT(msg2->msg_tx_delayed);
1081 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1083 (void) lnet_post_send_locked(msg2, 1);
1087 if (msg->msg_peertxcredit) {
1088 /* give back peer txcredits */
1089 msg->msg_peertxcredit = 0;
1091 spin_lock(&txpeer->lpni_lock);
1092 LASSERT((txpeer->lpni_txcredits < 0) ==
1093 !list_empty(&txpeer->lpni_txq));
1095 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1096 LASSERT(txpeer->lpni_txqnob >= 0);
1098 txpeer->lpni_txcredits++;
1099 if (txpeer->lpni_txcredits <= 0) {
1102 msg2 = list_entry(txpeer->lpni_txq.next,
1103 struct lnet_msg, msg_list);
1104 list_del(&msg2->msg_list);
1105 spin_unlock(&txpeer->lpni_lock);
1107 LASSERT(msg2->msg_txpeer == txpeer);
1108 LASSERT(msg2->msg_tx_delayed);
1110 msg2_cpt = msg2->msg_tx_cpt;
1113 * The msg_cpt can be different from the msg2_cpt
1114 * so we need to make sure we lock the correct cpt
1116 * Once we call lnet_post_send_locked() it is no
1117 * longer safe to access msg2, since it could've
1118 * been freed by lnet_finalize(), but we still
1119 * need to relock the correct cpt, so we cache the
1120 * msg2_cpt for the purpose of the check that
1121 * follows the call to lnet_pose_send_locked().
1123 if (msg2_cpt != msg->msg_tx_cpt) {
1124 lnet_net_unlock(msg->msg_tx_cpt);
1125 lnet_net_lock(msg2_cpt);
1127 (void) lnet_post_send_locked(msg2, 1);
1128 if (msg2_cpt != msg->msg_tx_cpt) {
1129 lnet_net_unlock(msg2_cpt);
1130 lnet_net_lock(msg->msg_tx_cpt);
1133 spin_unlock(&txpeer->lpni_lock);
1138 msg->msg_txni = NULL;
1139 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1142 if (txpeer != NULL) {
1143 msg->msg_txpeer = NULL;
1144 lnet_peer_ni_decref_locked(txpeer);
1149 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1151 struct lnet_msg *msg;
1153 if (list_empty(&rbp->rbp_msgs))
1155 msg = list_entry(rbp->rbp_msgs.next,
1156 struct lnet_msg, msg_list);
1157 list_del(&msg->msg_list);
1159 (void)lnet_post_routed_recv_locked(msg, 1);
1163 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1165 struct lnet_msg *msg;
1166 struct lnet_msg *tmp;
1168 lnet_net_unlock(cpt);
1170 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1171 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1172 0, 0, 0, msg->msg_hdr.payload_length);
1173 list_del_init(&msg->msg_list);
1174 msg->msg_no_resend = true;
1175 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1176 lnet_finalize(msg, -ECANCELED);
1183 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1185 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1186 struct lnet_peer *lp;
1187 struct lnet_ni *rxni = msg->msg_rxni;
1188 struct lnet_msg *msg2;
1190 if (msg->msg_rtrcredit) {
1191 /* give back global router credits */
1192 struct lnet_rtrbuf *rb;
1193 struct lnet_rtrbufpool *rbp;
1195 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1196 * there until it gets one allocated, or aborts the wait
1198 LASSERT(msg->msg_kiov != NULL);
1200 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1203 msg->msg_kiov = NULL;
1204 msg->msg_rtrcredit = 0;
1206 LASSERT(rbp == lnet_msg2bufpool(msg));
1208 LASSERT((rbp->rbp_credits > 0) ==
1209 !list_empty(&rbp->rbp_bufs));
1211 /* If routing is now turned off, we just drop this buffer and
1212 * don't bother trying to return credits. */
1213 if (!the_lnet.ln_routing) {
1214 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1218 /* It is possible that a user has lowered the desired number of
1219 * buffers in this pool. Make sure we never put back
1220 * more buffers than the stated number. */
1221 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1222 /* Discard this buffer so we don't have too
1224 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1225 rbp->rbp_nbuffers--;
1227 list_add(&rb->rb_list, &rbp->rbp_bufs);
1229 if (rbp->rbp_credits <= 0)
1230 lnet_schedule_blocked_locked(rbp);
1235 if (msg->msg_peerrtrcredit) {
1237 LASSERT(rxpeerni->lpni_peer_net);
1238 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1240 /* give back peer router credits */
1241 msg->msg_peerrtrcredit = 0;
1243 spin_lock(&rxpeerni->lpni_lock);
1244 rxpeerni->lpni_rtrcredits++;
1245 spin_unlock(&rxpeerni->lpni_lock);
1247 lp = rxpeerni->lpni_peer_net->lpn_peer;
1248 spin_lock(&lp->lp_lock);
1250 /* drop all messages which are queued to be routed on that
1252 if (!the_lnet.ln_routing) {
1254 list_splice_init(&lp->lp_rtrq, &drop);
1255 spin_unlock(&lp->lp_lock);
1256 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1257 } else if (!list_empty(&lp->lp_rtrq)) {
1260 msg2 = list_entry(lp->lp_rtrq.next,
1261 struct lnet_msg, msg_list);
1262 list_del(&msg2->msg_list);
1263 msg2_cpt = msg2->msg_rx_cpt;
1264 spin_unlock(&lp->lp_lock);
1266 * messages on the lp_rtrq can be from any NID in
1267 * the peer, which means they might have different
1268 * cpts. We need to make sure we lock the right
1271 if (msg2_cpt != msg->msg_rx_cpt) {
1272 lnet_net_unlock(msg->msg_rx_cpt);
1273 lnet_net_lock(msg2_cpt);
1275 (void) lnet_post_routed_recv_locked(msg2, 1);
1276 if (msg2_cpt != msg->msg_rx_cpt) {
1277 lnet_net_unlock(msg2_cpt);
1278 lnet_net_lock(msg->msg_rx_cpt);
1281 spin_unlock(&lp->lp_lock);
1285 msg->msg_rxni = NULL;
1286 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1288 if (rxpeerni != NULL) {
1289 msg->msg_rxpeer = NULL;
1290 lnet_peer_ni_decref_locked(rxpeerni);
1294 static struct lnet_peer_ni *
1295 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1296 struct lnet_peer *peer,
1297 struct lnet_peer_ni *best_lpni,
1298 struct lnet_peer_net *peer_net)
1301 * Look at the peer NIs for the destination peer that connect
1302 * to the chosen net. If a peer_ni is preferred when using the
1303 * best_ni to communicate, we use that one. If there is no
1304 * preferred peer_ni, or there are multiple preferred peer_ni,
1305 * the available transmit credits are used. If the transmit
1306 * credits are equal, we round-robin over the peer_ni.
1308 struct lnet_peer_ni *lpni = NULL;
1309 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1311 int best_lpni_healthv = (best_lpni) ?
1312 atomic_read(&best_lpni->lpni_healthv) : 0;
1313 bool best_lpni_is_preferred = false;
1314 bool lpni_is_preferred;
1316 __u32 lpni_sel_prio;
1317 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1319 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1321 * if the best_ni we've chosen aleady has this lpni
1322 * preferred, then let's use it
1325 lpni_is_preferred = lnet_peer_is_pref_nid_locked(lpni,
1327 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1328 libcfs_nid2str(best_ni->ni_nid),
1331 lpni_is_preferred = false;
1334 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1335 lpni_sel_prio = lpni->lpni_sel_priority;
1338 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1339 libcfs_nid2str(lpni->lpni_nid),
1340 libcfs_nid2str(best_lpni->lpni_nid),
1341 lpni_healthv, best_lpni_healthv,
1342 lpni_sel_prio, best_sel_prio,
1343 lpni->lpni_txcredits, best_lpni_credits,
1344 lpni->lpni_seq, best_lpni->lpni_seq);
1348 /* pick the healthiest peer ni */
1349 if (lpni_healthv < best_lpni_healthv)
1351 else if (lpni_healthv > best_lpni_healthv) {
1352 if (best_lpni_is_preferred)
1353 best_lpni_is_preferred = false;
1357 if (lpni_sel_prio > best_sel_prio)
1359 else if (lpni_sel_prio < best_sel_prio) {
1360 if (best_lpni_is_preferred)
1361 best_lpni_is_preferred = false;
1365 /* if this is a preferred peer use it */
1366 if (!best_lpni_is_preferred && lpni_is_preferred) {
1367 best_lpni_is_preferred = true;
1369 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1370 /* this is not the preferred peer so let's ignore
1376 if (lpni->lpni_txcredits < best_lpni_credits)
1377 /* We already have a peer that has more credits
1378 * available than this one. No need to consider
1379 * this peer further.
1382 else if (lpni->lpni_txcredits > best_lpni_credits)
1385 /* The best peer found so far and the current peer
1386 * have the same number of available credits let's
1387 * make sure to select between them using Round Robin
1389 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1392 best_lpni_is_preferred = lpni_is_preferred;
1393 best_lpni_healthv = lpni_healthv;
1394 best_sel_prio = lpni_sel_prio;
1396 best_lpni_credits = lpni->lpni_txcredits;
1399 /* if we still can't find a peer ni then we can't reach it */
1401 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1402 LNET_NIDNET(dst_nid);
1403 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1404 libcfs_net2str(net_id));
1408 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1409 libcfs_nid2str(best_lpni->lpni_nid));
1415 * Prerequisite: the best_ni should already be set in the sd
1416 * Find the best lpni.
1417 * If the net id is provided then restrict lpni selection on
1418 * that particular net.
1419 * Otherwise find any reachable lpni. When dealing with an MR
1420 * gateway and it has multiple lpnis which we can use
1421 * we want to select the best one from the list of reachable
1424 static inline struct lnet_peer_ni *
1425 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1426 struct lnet_peer *peer, __u32 net_id)
1428 struct lnet_peer_net *peer_net;
1430 /* find the best_lpni on any local network */
1431 if (net_id == LNET_NET_ANY) {
1432 struct lnet_peer_ni *best_lpni = NULL;
1433 struct lnet_peer_net *lpn;
1434 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1435 /* no net specified find any reachable peer ni */
1436 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1438 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1444 /* restrict on the specified net */
1445 peer_net = lnet_peer_get_net_locked(peer, net_id);
1447 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1453 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1455 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1458 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1461 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1464 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1470 /* Compare route priorities and hop counts */
1472 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1474 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1475 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1477 if (r1->lr_priority < r2->lr_priority)
1480 if (r1->lr_priority > r2->lr_priority)
1483 if (r1_hops < r2_hops)
1486 if (r1_hops > r2_hops)
1492 static struct lnet_route *
1493 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1494 struct lnet_peer_ni *remote_lpni,
1495 struct lnet_route **prev_route,
1496 struct lnet_peer_ni **gwni)
1498 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1499 struct lnet_route *best_route;
1500 struct lnet_route *last_route;
1501 struct lnet_route *route;
1503 bool best_rte_is_preferred = false;
1506 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1507 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1509 best_route = last_route = NULL;
1510 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1511 if (!lnet_is_route_alive(route))
1513 gw_pnid = route->lr_gateway->lp_primary_nid;
1515 /* no protection on below fields, but it's harmless */
1516 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1519 /* if the best route found is in the preferred list then
1520 * tag it as preferred and use it later on. But if we
1521 * didn't find any routes which are on the preferred list
1522 * then just use the best route possible.
1524 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1526 if (!best_route || (rc && !best_rte_is_preferred)) {
1527 /* Restrict the selection of the router NI on the
1528 * src_net provided. If the src_net is LNET_NID_ANY,
1529 * then select the best interface available.
1531 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1536 "Gateway %s does not have a peer NI on net %s\n",
1537 libcfs_nid2str(gw_pnid),
1538 libcfs_net2str(src_net));
1543 if (rc && !best_rte_is_preferred) {
1544 /* This is the first preferred route we found,
1545 * so it beats any route found previously
1551 best_rte_is_preferred = true;
1552 CDEBUG(D_NET, "preferred gw = %s\n",
1553 libcfs_nid2str(gw_pnid));
1555 } else if ((!rc) && best_rte_is_preferred)
1556 /* The best route we found so far is in the preferred
1557 * list, so it beats any non-preferred route
1562 best_route = last_route = route;
1567 rc = lnet_compare_routes(route, best_route);
1571 /* Restrict the selection of the router NI on the
1572 * src_net provided. If the src_net is LNET_NID_ANY,
1573 * then select the best interface available.
1575 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1580 "Gateway %s does not have a peer NI on net %s\n",
1581 libcfs_nid2str(gw_pnid),
1582 libcfs_net2str(src_net));
1592 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1596 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1603 *prev_route = last_route;
1609 static inline unsigned int
1610 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1612 if (dev_idx == UINT_MAX)
1615 if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1616 !ni->ni_net->net_lnd->lnd_get_dev_prio)
1619 return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1622 static struct lnet_ni *
1623 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1624 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1625 struct lnet_msg *msg, int md_cpt)
1627 struct lnet_libmd *md = msg->msg_md;
1628 unsigned int offset = msg->msg_offset;
1629 unsigned int shortest_distance;
1630 struct lnet_ni *ni = NULL;
1633 __u32 best_sel_prio;
1634 unsigned int best_dev_prio;
1635 unsigned int dev_idx = UINT_MAX;
1636 struct page *page = lnet_get_first_page(md, offset);
1637 msg->msg_rdma_force = lnet_is_rdma_only_page(page);
1639 if (msg->msg_rdma_force)
1640 dev_idx = lnet_get_dev_idx(page);
1643 * If there is no peer_ni that we can send to on this network,
1644 * then there is no point in looking for a new best_ni here.
1646 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1649 if (best_ni == NULL) {
1650 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1651 shortest_distance = UINT_MAX;
1652 best_dev_prio = UINT_MAX;
1653 best_credits = INT_MIN;
1656 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1657 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1658 best_ni->ni_dev_cpt);
1659 best_credits = atomic_read(&best_ni->ni_tx_credits);
1660 best_healthv = atomic_read(&best_ni->ni_healthv);
1661 best_sel_prio = best_ni->ni_sel_priority;
1664 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1665 unsigned int distance;
1670 unsigned int ni_dev_prio;
1672 ni_credits = atomic_read(&ni->ni_tx_credits);
1673 ni_healthv = atomic_read(&ni->ni_healthv);
1674 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1675 ni_sel_prio = ni->ni_sel_priority;
1678 * calculate the distance from the CPT on which
1679 * the message memory is allocated to the CPT of
1680 * the NI's physical device
1682 distance = cfs_cpt_distance(lnet_cpt_table(),
1686 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1689 * All distances smaller than the NUMA range
1690 * are treated equally.
1692 if (distance < lnet_numa_range)
1693 distance = lnet_numa_range;
1696 * Select on health, selection policy, direct dma prio,
1697 * shorter distance, available credits, then round-robin.
1703 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",
1704 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1705 ni->ni_seq, ni_sel_prio, ni_dev_prio,
1706 (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1707 : "not selected", best_credits, shortest_distance,
1708 (best_ni) ? best_ni->ni_seq : 0,
1709 best_sel_prio, best_dev_prio);
1713 if (ni_healthv < best_healthv)
1715 else if (ni_healthv > best_healthv)
1718 if (ni_sel_prio > best_sel_prio)
1720 else if (ni_sel_prio < best_sel_prio)
1723 if (ni_dev_prio > best_dev_prio)
1725 else if (ni_dev_prio < best_dev_prio)
1728 if (distance > shortest_distance)
1730 else if (distance < shortest_distance)
1733 if (ni_credits < best_credits)
1735 else if (ni_credits > best_credits)
1738 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1742 best_sel_prio = ni_sel_prio;
1743 best_dev_prio = ni_dev_prio;
1744 shortest_distance = distance;
1745 best_healthv = ni_healthv;
1747 best_credits = ni_credits;
1750 CDEBUG(D_NET, "selected best_ni %s\n",
1751 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1757 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1758 * because such traffic is required to perform discovery. We therefore
1759 * exclude all GET and PUT on that portal. We also exclude all ACK and
1760 * REPLY traffic, but that is because the portal is not tracked in the
1761 * message structure for these message types. We could restrict this
1762 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1765 lnet_msg_discovery(struct lnet_msg *msg)
1767 if (msg->msg_type == LNET_MSG_PUT) {
1768 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1770 } else if (msg->msg_type == LNET_MSG_GET) {
1771 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1777 #define SRC_SPEC 0x0001
1778 #define SRC_ANY 0x0002
1779 #define LOCAL_DST 0x0004
1780 #define REMOTE_DST 0x0008
1781 #define MR_DST 0x0010
1782 #define NMR_DST 0x0020
1783 #define SND_RESP 0x0040
1785 /* The following to defines are used for return codes */
1786 #define REPEAT_SEND 0x1000
1787 #define PASS_THROUGH 0x2000
1789 /* The different cases lnet_select pathway needs to handle */
1790 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1791 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1792 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1793 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1794 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1795 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1796 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1797 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1800 lnet_handle_lo_send(struct lnet_send_data *sd)
1802 struct lnet_msg *msg = sd->sd_msg;
1803 int cpt = sd->sd_cpt;
1805 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1808 /* No send credit hassles with LOLND */
1809 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1810 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1811 if (!msg->msg_routing)
1812 msg->msg_hdr.src_nid =
1813 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1814 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1815 lnet_msg_commit(msg, cpt);
1816 msg->msg_txni = the_lnet.ln_loni;
1818 return LNET_CREDIT_OK;
1822 lnet_handle_send(struct lnet_send_data *sd)
1824 struct lnet_ni *best_ni = sd->sd_best_ni;
1825 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1826 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1827 struct lnet_msg *msg = sd->sd_msg;
1829 __u32 send_case = sd->sd_send_case;
1831 __u32 routing = send_case & REMOTE_DST;
1832 struct lnet_rsp_tracker *rspt;
1834 /* Increment sequence number of the selected peer, peer net,
1835 * local ni and local net so that we pick the next ones
1838 best_lpni->lpni_seq++;
1839 best_lpni->lpni_peer_net->lpn_seq++;
1841 best_ni->ni_net->net_seq++;
1843 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1844 libcfs_nid2str(best_ni->ni_nid),
1845 best_ni->ni_seq, best_ni->ni_net->net_seq,
1846 atomic_read(&best_ni->ni_tx_credits),
1847 best_ni->ni_sel_priority,
1848 libcfs_nid2str(best_lpni->lpni_nid),
1849 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1850 best_lpni->lpni_txcredits,
1851 best_lpni->lpni_sel_priority);
1854 * grab a reference on the peer_ni so it sticks around even if
1855 * we need to drop and relock the lnet_net_lock below.
1857 lnet_peer_ni_addref_locked(best_lpni);
1860 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1861 * message. This ensures that we get a CPT that is correct for
1862 * the NI when the NI has been restricted to a subset of all CPTs.
1863 * If the selected CPT differs from the one currently locked, we
1864 * must unlock and relock the lnet_net_lock(), and then check whether
1865 * the configuration has changed. We don't have a hold on the best_ni
1866 * yet, and it may have vanished.
1868 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1869 if (sd->sd_cpt != cpt2) {
1870 __u32 seq = lnet_get_dlc_seq_locked();
1871 lnet_net_unlock(sd->sd_cpt);
1873 lnet_net_lock(sd->sd_cpt);
1874 if (seq != lnet_get_dlc_seq_locked()) {
1875 lnet_peer_ni_decref_locked(best_lpni);
1881 * store the best_lpni in the message right away to avoid having
1882 * to do the same operation under different conditions
1884 msg->msg_txpeer = best_lpni;
1885 msg->msg_txni = best_ni;
1888 * grab a reference for the best_ni since now it's in use in this
1889 * send. The reference will be dropped in lnet_finalize()
1891 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1894 * Always set the target.nid to the best peer picked. Either the
1895 * NID will be one of the peer NIDs selected, or the same NID as
1896 * what was originally set in the target or it will be the NID of
1897 * a router if this message should be routed
1899 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1902 * lnet_msg_commit assigns the correct cpt to the message, which
1903 * is used to decrement the correct refcount on the ni when it's
1904 * time to return the credits
1906 lnet_msg_commit(msg, sd->sd_cpt);
1909 * If we are routing the message then we keep the src_nid that was
1910 * set by the originator. If we are not routing then we are the
1911 * originator and set it here.
1913 if (!msg->msg_routing)
1914 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1917 msg->msg_target_is_router = 1;
1918 msg->msg_target.pid = LNET_PID_LUSTRE;
1920 * since we're routing we want to ensure that the
1921 * msg_hdr.dest_nid is set to the final destination. When
1922 * the router receives this message it knows how to route
1925 * final_dst_lpni is set at the beginning of the
1926 * lnet_select_pathway() function and is never changed.
1927 * It's safe to use it here.
1929 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1932 * if we're not routing set the dest_nid to the best peer
1933 * ni NID that we picked earlier in the algorithm.
1935 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1939 * if we have response tracker block update it with the next hop
1943 rspt = msg->msg_md->md_rspt_ptr;
1945 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1946 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1947 libcfs_nid2str(rspt->rspt_next_hop_nid));
1951 rc = lnet_post_send_locked(msg, 0);
1954 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1955 libcfs_nid2str(msg->msg_hdr.src_nid),
1956 libcfs_nid2str(msg->msg_txni->ni_nid),
1957 libcfs_nid2str(sd->sd_src_nid),
1958 libcfs_nid2str(msg->msg_hdr.dest_nid),
1959 libcfs_nid2str(sd->sd_dst_nid),
1960 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1961 libcfs_nid2str(sd->sd_rtr_nid),
1962 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1968 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1969 struct lnet_msg *msg)
1971 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1972 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1973 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1974 libcfs_nid2str(lni->ni_nid),
1975 libcfs_nid2str(lpni->lpni_nid));
1976 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1985 * use the source and destination NIDs as the pathway
1988 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1990 /* the destination lpni is set before we get here. */
1993 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1994 if (!sd->sd_best_ni) {
1995 CERROR("Can't send to %s: src %s is not a "
1996 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1997 libcfs_nid2str(sd->sd_src_nid));
2001 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2003 return lnet_handle_send(sd);
2011 * Don't run the selection algorithm on the peer NIs. By specifying the
2012 * local NID, we're also saying that we should always use the destination NID
2013 * provided. This handles the case where we should be using the same
2014 * destination NID for the all the messages which belong to the same RPC
2018 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2020 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2021 if (!sd->sd_best_ni) {
2022 CERROR("Can't send to %s: src %s is not a "
2023 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2024 libcfs_nid2str(sd->sd_src_nid));
2028 if (sd->sd_best_lpni &&
2029 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
2030 return lnet_handle_lo_send(sd);
2031 else if (sd->sd_best_lpni)
2032 return lnet_handle_send(sd);
2034 CERROR("can't send to %s. no NI on %s\n",
2035 libcfs_nid2str(sd->sd_dst_nid),
2036 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2038 return -EHOSTUNREACH;
2042 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2043 struct lnet_peer *peer,
2044 struct lnet_peer_net *peer_net,
2045 struct lnet_msg *msg,
2048 struct lnet_net *local_net;
2049 struct lnet_ni *best_ni;
2051 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2056 * Iterate through the NIs in this local Net and select
2057 * the NI to send from. The selection is determined by
2058 * these 3 criterion in the following priority:
2060 * 2. NI available credits
2063 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2064 peer, peer_net, msg, cpt);
2070 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2073 struct lnet_peer *peer;
2074 struct lnet_peer_ni *new_lpni;
2077 lnet_peer_ni_addref_locked(lpni);
2079 peer = lpni->lpni_peer_net->lpn_peer;
2081 if (lnet_peer_gw_discovery(peer)) {
2082 lnet_peer_ni_decref_locked(lpni);
2086 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2087 lnet_peer_ni_decref_locked(lpni);
2091 rc = lnet_discover_peer_locked(lpni, cpt, false);
2093 lnet_peer_ni_decref_locked(lpni);
2097 new_lpni = lnet_find_peer_ni_locked(lpni->lpni_nid);
2099 lnet_peer_ni_decref_locked(lpni);
2103 peer = new_lpni->lpni_peer_net->lpn_peer;
2104 spin_lock(&peer->lp_lock);
2105 if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2106 /* The peer NI did not change and the peer is up to date.
2107 * Nothing more to do.
2109 spin_unlock(&peer->lp_lock);
2110 lnet_peer_ni_decref_locked(lpni);
2111 lnet_peer_ni_decref_locked(new_lpni);
2114 spin_unlock(&peer->lp_lock);
2116 /* Either the peer NI changed during discovery, or the peer isn't up
2117 * to date. In both cases we want to queue the message on the
2118 * (possibly new) peer's pending queue and queue the peer for discovery
2120 msg->msg_sending = 0;
2121 msg->msg_txpeer = NULL;
2122 lnet_net_unlock(cpt);
2123 lnet_peer_queue_message(peer, msg);
2126 lnet_peer_ni_decref_locked(lpni);
2127 lnet_peer_ni_decref_locked(new_lpni);
2129 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2130 msg, libcfs_nid2str(peer->lp_primary_nid));
2132 return LNET_DC_WAIT;
2136 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2138 struct lnet_peer_ni **gw_lpni,
2139 struct lnet_peer **gw_peer)
2142 struct lnet_peer *gw;
2143 struct lnet_peer *lp;
2144 struct lnet_peer_net *lpn;
2145 struct lnet_peer_net *best_lpn = NULL;
2146 struct lnet_remotenet *rnet, *best_rnet = NULL;
2147 struct lnet_route *best_route = NULL;
2148 struct lnet_route *last_route = NULL;
2149 struct lnet_peer_ni *lpni = NULL;
2150 struct lnet_peer_ni *gwni = NULL;
2151 bool route_found = false;
2152 lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
2153 (sd->sd_best_ni != NULL) ? sd->sd_best_ni->ni_nid :
2155 int best_lpn_healthv = 0;
2156 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2158 CDEBUG(D_NET, "using src nid %s for route restriction\n",
2159 libcfs_nid2str(src_nid));
2161 /* If a router nid was specified then we are replying to a GET or
2162 * sending an ACK. In this case we use the gateway associated with the
2163 * specified router nid.
2165 if (sd->sd_rtr_nid != LNET_NID_ANY) {
2166 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2168 gw = gwni->lpni_peer_net->lpn_peer;
2169 lnet_peer_ni_decref_locked(gwni);
2170 if (gw->lp_rtr_refcount)
2173 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2174 libcfs_nid2str(sd->sd_rtr_nid));
2179 if (sd->sd_msg->msg_routing) {
2180 /* If I'm routing this message then I need to find the
2181 * next hop based on the destination NID
2183 best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
2185 CERROR("Unable to route message to %s - Route table may be misconfigured\n",
2186 libcfs_nid2str(sd->sd_dst_nid));
2187 return -EHOSTUNREACH;
2190 /* we've already looked up the initial lpni using
2193 lpni = sd->sd_best_lpni;
2194 /* the peer tree must be in existence */
2195 LASSERT(lpni && lpni->lpni_peer_net &&
2196 lpni->lpni_peer_net->lpn_peer);
2197 lp = lpni->lpni_peer_net->lpn_peer;
2199 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2200 /* is this remote network reachable? */
2201 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2210 /* select the preferred peer net */
2211 if (best_lpn_healthv > lpn->lpn_healthv)
2213 else if (best_lpn_healthv < lpn->lpn_healthv)
2216 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2218 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2221 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2224 best_lpn_healthv = lpn->lpn_healthv;
2225 best_lpn_sel_prio = lpn->lpn_sel_priority;
2231 CERROR("peer %s has no available nets\n",
2232 libcfs_nid2str(sd->sd_dst_nid));
2233 return -EHOSTUNREACH;
2236 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2239 best_lpn->lpn_net_id);
2240 if (!sd->sd_best_lpni) {
2241 CERROR("peer %s is unreachable\n",
2242 libcfs_nid2str(sd->sd_dst_nid));
2243 return -EHOSTUNREACH;
2246 /* We're attempting to round robin over the remote peer
2247 * NI's so update the final destination we selected
2249 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2251 /* Increment the sequence number of the remote lpni so
2252 * we can round robin over the different interfaces of
2255 sd->sd_best_lpni->lpni_seq++;
2259 * find the best route. Restrict the selection on the net of the
2260 * local NI if we've already picked the local NI to send from.
2261 * Otherwise, let's pick any route we can find and then find
2262 * a local NI we can reach the route's gateway on. Any route we
2263 * select will be reachable by virtue of the restriction we have
2264 * when adding a route.
2266 best_route = lnet_find_route_locked(best_rnet,
2267 LNET_NIDNET(src_nid),
2269 &last_route, &gwni);
2272 CERROR("no route to %s from %s\n",
2273 libcfs_nid2str(dst_nid),
2274 libcfs_nid2str(src_nid));
2275 return -EHOSTUNREACH;
2279 CERROR("Internal Error. Route expected to %s from %s\n",
2280 libcfs_nid2str(dst_nid),
2281 libcfs_nid2str(src_nid));
2285 gw = best_route->lr_gateway;
2286 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2290 * Discover this gateway if it hasn't already been discovered.
2291 * This means we might delay the message until discovery has
2294 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2298 if (!sd->sd_best_ni) {
2299 lpn = gwni->lpni_peer_net;
2300 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2303 if (!sd->sd_best_ni) {
2304 CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2305 libcfs_net2str(lpn->lpn_net_id),
2306 libcfs_nid2str(sd->sd_src_nid));
2315 * increment the sequence numbers since now we're sure we're
2316 * going to use this path
2318 if (sd->sd_rtr_nid == LNET_NID_ANY) {
2319 LASSERT(best_route && last_route);
2320 best_route->lr_seq = last_route->lr_seq + 1;
2322 best_lpn->lpn_seq++;
2333 * Remote destination
2334 * Non-MR destination
2338 * Remote destination
2341 * The handling of these two cases is similar. Even though the destination
2342 * can be MR or non-MR, we'll deal directly with the router.
2345 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2348 struct lnet_peer_ni *gw_lpni = NULL;
2349 struct lnet_peer *gw_peer = NULL;
2352 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2353 if (!sd->sd_best_ni) {
2354 CERROR("Can't send to %s: src %s is not a "
2355 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2356 libcfs_nid2str(sd->sd_src_nid));
2360 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2365 if (sd->sd_send_case & NMR_DST)
2367 * since the final destination is non-MR let's set its preferred
2368 * NID before we send
2370 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2374 * We're going to send to the gw found so let's set its
2377 sd->sd_peer = gw_peer;
2378 sd->sd_best_lpni = gw_lpni;
2380 return lnet_handle_send(sd);
2384 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2385 struct lnet_msg *msg, bool discovery)
2387 struct lnet_peer_net *lpn = NULL;
2388 struct lnet_peer_net *best_lpn = NULL;
2389 struct lnet_net *net = NULL;
2390 struct lnet_net *best_net = NULL;
2391 struct lnet_ni *best_ni = NULL;
2392 int best_lpn_healthv = 0;
2393 int best_net_healthv = 0;
2395 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2397 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2402 * The peer can have multiple interfaces, some of them can be on
2403 * the local network and others on a routed network. We should
2404 * prefer the local network. However if the local network is not
2405 * available then we need to try the routed network
2408 /* go through all the peer nets and find the best_ni */
2409 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2411 * The peer's list of nets can contain non-local nets. We
2412 * want to only examine the local ones.
2414 net = lnet_get_net_locked(lpn->lpn_net_id);
2418 lpn_sel_prio = lpn->lpn_sel_priority;
2419 net_healthv = lnet_get_net_healthv_locked(net);
2420 net_sel_prio = net->net_sel_priority;
2423 * if this is a discovery message and lp_disc_net_id is
2424 * specified then use that net to send the discovery on.
2426 if (peer->lp_disc_net_id == lpn->lpn_net_id &&
2435 /* always select the lpn with the best health */
2436 if (best_lpn_healthv > lpn->lpn_healthv)
2438 else if (best_lpn_healthv < lpn->lpn_healthv)
2441 /* select the preferred peer and local nets */
2442 if (best_lpn_sel_prio < lpn_sel_prio)
2444 else if (best_lpn_sel_prio > lpn_sel_prio)
2447 if (best_net_healthv > net_healthv)
2449 else if (best_net_healthv < net_healthv)
2452 if (best_net_sel_prio < net_sel_prio)
2454 else if (best_net_sel_prio > net_sel_prio)
2457 if (best_lpn->lpn_seq < lpn->lpn_seq)
2459 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2462 /* round robin over the local networks */
2463 if (best_net->net_seq <= net->net_seq)
2467 best_net_healthv = net_healthv;
2468 best_net_sel_prio = net_sel_prio;
2469 best_lpn_healthv = lpn->lpn_healthv;
2470 best_lpn_sel_prio = lpn_sel_prio;
2479 /* Select the best NI on the same net as best_lpn chosen
2482 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2489 static struct lnet_ni *
2490 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2492 struct lnet_ni *best_ni = NULL;
2493 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2494 struct lnet_peer_ni *lpni_entry;
2497 * We must use a consistent source address when sending to a
2498 * non-MR peer. However, a non-MR peer can have multiple NIDs
2499 * on multiple networks, and we may even need to talk to this
2500 * peer on multiple networks -- certain types of
2501 * load-balancing configuration do this.
2503 * So we need to pick the NI the peer prefers for this
2504 * particular network.
2507 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2509 if (lpni_entry->lpni_pref_nnids == 0)
2511 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2512 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2519 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2521 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2523 struct lnet_ni *best_ni = NULL;
2524 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2527 * We must use a consistent source address when sending to a
2528 * non-MR peer. However, a non-MR peer can have multiple NIDs
2529 * on multiple networks, and we may even need to talk to this
2530 * peer on multiple networks -- certain types of
2531 * load-balancing configuration do this.
2533 * So we need to pick the NI the peer prefers for this
2534 * particular network.
2537 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2540 /* if best_ni is still not set just pick one */
2543 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2544 sd->sd_best_lpni->lpni_peer_net,
2547 /* If there is no best_ni we don't have a route */
2549 CERROR("no path to %s from net %s\n",
2550 libcfs_nid2str(best_lpni->lpni_nid),
2551 libcfs_net2str(best_lpni->lpni_net->net_id));
2552 return -EHOSTUNREACH;
2556 sd->sd_best_ni = best_ni;
2558 /* Set preferred NI if necessary. */
2559 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2566 * Source not specified
2570 * always use the same source NID for NMR peers
2571 * If we've talked to that peer before then we already have a preferred
2572 * source NI associated with it. Otherwise, we select a preferred local NI
2573 * and store it in the peer
2576 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2580 /* sd->sd_best_lpni is already set to the final destination */
2583 * At this point we should've created the peer ni and peer. If we
2584 * can't find it, then something went wrong. Instead of assert
2585 * output a relevant message and fail the send
2587 if (!sd->sd_best_lpni) {
2588 CERROR("Internal fault. Unable to send msg %s to %s. "
2590 lnet_msgtyp2str(sd->sd_msg->msg_type),
2591 libcfs_nid2str(sd->sd_dst_nid));
2595 if (sd->sd_msg->msg_routing) {
2596 /* If I'm forwarding this message then I can choose any NI
2597 * on the destination peer net
2599 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2601 sd->sd_best_lpni->lpni_peer_net,
2604 if (!sd->sd_best_ni) {
2605 CERROR("Unable to forward message to %s. No local NI available\n",
2606 libcfs_nid2str(sd->sd_dst_nid));
2610 rc = lnet_select_preferred_best_ni(sd);
2613 rc = lnet_handle_send(sd);
2619 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2622 * NOTE we've already handled the remote peer case. So we only
2623 * need to worry about the local case here.
2625 * if we're sending a response, ACK or reply, we need to send it
2626 * to the destination NID given to us. At this point we already
2627 * have the peer_ni we're suppose to send to, so just find the
2628 * best_ni on the peer net and use that. Since we're sending to an
2629 * MR peer then we can just run the selection algorithm on our
2630 * local NIs and pick the best one.
2632 if (sd->sd_send_case & SND_RESP) {
2634 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2635 sd->sd_best_lpni->lpni_peer_net,
2639 if (!sd->sd_best_ni) {
2641 * We're not going to deal with not able to send
2642 * a response to the provided final destination
2644 CERROR("Can't send response to %s. "
2645 "No local NI available\n",
2646 libcfs_nid2str(sd->sd_dst_nid));
2647 return -EHOSTUNREACH;
2650 return lnet_handle_send(sd);
2654 * If we get here that means we're sending a fresh request, PUT or
2655 * GET, so we need to run our standard selection algorithm.
2656 * First find the best local interface that's on any of the peer's
2659 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2662 lnet_msg_discovery(sd->sd_msg));
2663 if (sd->sd_best_ni) {
2665 lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2667 sd->sd_best_ni->ni_net->net_id);
2670 * if we're successful in selecting a peer_ni on the local
2671 * network, then send to it. Otherwise fall through and
2672 * try and see if we can reach it over another routed
2675 if (sd->sd_best_lpni &&
2676 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2678 * in case we initially started with a routed
2679 * destination, let's reset to local
2681 sd->sd_send_case &= ~REMOTE_DST;
2682 sd->sd_send_case |= LOCAL_DST;
2683 return lnet_handle_lo_send(sd);
2684 } else if (sd->sd_best_lpni) {
2686 * in case we initially started with a routed
2687 * destination, let's reset to local
2689 sd->sd_send_case &= ~REMOTE_DST;
2690 sd->sd_send_case |= LOCAL_DST;
2691 return lnet_handle_send(sd);
2694 CERROR("Internal Error. Expected to have a best_lpni: "
2696 libcfs_nid2str(sd->sd_src_nid),
2697 libcfs_nid2str(sd->sd_dst_nid));
2703 * Peer doesn't have a local network. Let's see if there is
2704 * a remote network we can reach it on.
2706 return PASS_THROUGH;
2711 * Source NID not specified
2716 * Source NID not speified
2717 * Remote destination
2720 * In both of these cases if we're sending a response, ACK or REPLY, then
2721 * we need to send to the destination NID provided.
2723 * In the remote case let's deal with MR routers.
2728 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2731 struct lnet_peer *gw_peer = NULL;
2732 struct lnet_peer_ni *gw_lpni = NULL;
2735 * handle sending a response to a remote peer here so we don't
2736 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2738 if (sd->sd_send_case & REMOTE_DST &&
2739 sd->sd_send_case & SND_RESP) {
2740 struct lnet_peer_ni *gw;
2741 struct lnet_peer *gw_peer;
2743 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2746 CERROR("Can't send response to %s. "
2747 "No route available\n",
2748 libcfs_nid2str(sd->sd_dst_nid));
2749 return -EHOSTUNREACH;
2750 } else if (rc > 0) {
2754 sd->sd_best_lpni = gw;
2755 sd->sd_peer = gw_peer;
2757 return lnet_handle_send(sd);
2761 * Even though the NID for the peer might not be on a local network,
2762 * since the peer is MR there could be other interfaces on the
2763 * local network. In that case we'd still like to prefer the local
2764 * network over the routed network. If we're unable to do that
2765 * then we select the best router among the different routed networks,
2766 * and if the router is MR then we can deal with it as such.
2768 rc = lnet_handle_any_mr_dsta(sd);
2769 if (rc != PASS_THROUGH)
2773 * Now that we must route to the destination, we must consider the
2774 * MR case, where the destination has multiple interfaces, some of
2775 * which we can route to and others we do not. For this reason we
2776 * need to select the destination which we can route to and if
2777 * there are multiple, we need to round robin.
2779 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2784 sd->sd_send_case &= ~LOCAL_DST;
2785 sd->sd_send_case |= REMOTE_DST;
2787 sd->sd_peer = gw_peer;
2788 sd->sd_best_lpni = gw_lpni;
2790 return lnet_handle_send(sd);
2794 * Source not specified
2795 * Remote destination
2798 * Must send to the specified peer NID using the same source NID that
2799 * we've used before. If it's the first time to talk to that peer then
2800 * find the source NI and assign it as preferred to that peer
2803 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2806 struct lnet_peer_ni *gw_lpni = NULL;
2807 struct lnet_peer *gw_peer = NULL;
2810 * Let's see if we have a preferred NI to talk to this NMR peer
2812 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2816 * find the router and that'll find the best NI if we didn't find
2819 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2825 * set the best_ni we've chosen as the preferred one for
2828 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2830 /* we'll be sending to the gw */
2831 sd->sd_best_lpni = gw_lpni;
2832 sd->sd_peer = gw_peer;
2834 return lnet_handle_send(sd);
2838 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2841 * turn off the SND_RESP bit.
2842 * It will be checked in the case handling
2844 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2846 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2847 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2848 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2849 (send_case & MR_DST) ? "MR: " : "NMR: ",
2850 libcfs_nid2str(sd->sd_dst_nid),
2851 (send_case & LOCAL_DST) ? "local" : "routed");
2853 switch (send_case) {
2855 * For all cases where the source is specified, we should always
2856 * use the destination NID, whether it's an MR destination or not,
2857 * since we're continuing a series of related messages for the
2860 case SRC_SPEC_LOCAL_NMR_DST:
2861 return lnet_handle_spec_local_nmr_dst(sd);
2862 case SRC_SPEC_LOCAL_MR_DST:
2863 return lnet_handle_spec_local_mr_dst(sd);
2864 case SRC_SPEC_ROUTER_NMR_DST:
2865 case SRC_SPEC_ROUTER_MR_DST:
2866 return lnet_handle_spec_router_dst(sd);
2867 case SRC_ANY_LOCAL_NMR_DST:
2868 return lnet_handle_any_local_nmr_dst(sd);
2869 case SRC_ANY_LOCAL_MR_DST:
2870 case SRC_ANY_ROUTER_MR_DST:
2871 return lnet_handle_any_mr_dst(sd);
2872 case SRC_ANY_ROUTER_NMR_DST:
2873 return lnet_handle_any_router_nmr_dst(sd);
2875 CERROR("Unknown send case\n");
2881 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2882 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2884 struct lnet_peer_ni *lpni;
2885 struct lnet_peer *peer;
2886 struct lnet_send_data send_data;
2889 __u32 send_case = 0;
2891 bool mr_forwarding_allowed;
2893 memset(&send_data, 0, sizeof(send_data));
2896 * get an initial CPT to use for locking. The idea here is not to
2897 * serialize the calls to select_pathway, so that as many
2898 * operations can run concurrently as possible. To do that we use
2899 * the CPT where this call is being executed. Later on when we
2900 * determine the CPT to use in lnet_message_commit, we switch the
2901 * lock and check if there was any configuration change. If none,
2902 * then we proceed, if there is, then we restart the operation.
2904 cpt = lnet_net_lock_current();
2906 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2907 if (md_cpt == CFS_CPT_ANY)
2913 * If we're being asked to send to the loopback interface, there
2914 * is no need to go through any selection. We can just shortcut
2915 * the entire process and send over lolnd
2917 send_data.sd_msg = msg;
2918 send_data.sd_cpt = cpt;
2919 if (dst_nid == LNET_NID_LO_0) {
2920 rc = lnet_handle_lo_send(&send_data);
2921 lnet_net_unlock(cpt);
2926 * find an existing peer_ni, or create one and mark it as having been
2927 * created due to network traffic. This call will create the
2928 * peer->peer_net->peer_ni tree.
2930 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2932 lnet_net_unlock(cpt);
2933 return PTR_ERR(lpni);
2937 * Cache the original src_nid and rtr_nid. If we need to resend the
2938 * message then we'll need to know whether the src_nid was originally
2939 * specified for this message. If it was originally specified,
2940 * then we need to keep using the same src_nid since it's
2941 * continuing the same sequence of messages. Similarly, rtr_nid will
2942 * affect our choice of next hop.
2944 msg->msg_src_nid_param = src_nid;
2945 msg->msg_rtr_nid_param = rtr_nid;
2948 * If necessary, perform discovery on the peer that owns this peer_ni.
2949 * Note, this can result in the ownership of this peer_ni changing
2950 * to another peer object.
2952 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2954 lnet_peer_ni_decref_locked(lpni);
2955 lnet_net_unlock(cpt);
2958 lnet_peer_ni_decref_locked(lpni);
2960 peer = lpni->lpni_peer_net->lpn_peer;
2963 * Identify the different send cases
2965 if (src_nid == LNET_NID_ANY)
2966 send_case |= SRC_ANY;
2968 send_case |= SRC_SPEC;
2970 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2971 send_case |= LOCAL_DST;
2973 send_case |= REMOTE_DST;
2976 if (msg->msg_routing && (send_case & LOCAL_DST))
2979 /* Determine whether to allow MR forwarding for this message.
2980 * NB: MR forwarding is allowed if the message originator and the
2981 * destination are both MR capable, and the destination lpni that was
2982 * originally chosen by the originator is unhealthy or down.
2983 * We check the MR capability of the destination further below
2985 mr_forwarding_allowed = false;
2987 struct lnet_peer *src_lp;
2988 struct lnet_peer_ni *src_lpni;
2990 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
2992 /* We don't fail the send if we hit any errors here. We'll just
2993 * try to send it via non-multi-rail criteria
2995 if (!IS_ERR(src_lpni)) {
2996 /* Drop ref taken by lnet_nid2peerni_locked() */
2997 lnet_peer_ni_decref_locked(src_lpni);
2998 src_lp = lpni->lpni_peer_net->lpn_peer;
2999 if (lnet_peer_is_multi_rail(src_lp) &&
3000 !lnet_is_peer_ni_alive(lpni))
3001 mr_forwarding_allowed = true;
3004 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3005 mr_forwarding_allowed ? "allowed" : "not allowed");
3009 * Deal with the peer as NMR in the following cases:
3010 * 1. the peer is NMR
3011 * 2. We're trying to recover a specific peer NI
3012 * 3. I'm a router sending to the final destination and MR forwarding is
3013 * not allowed for this message (as determined above).
3014 * In this case the source of the message would've
3015 * already selected the final destination so my job
3016 * is to honor the selection.
3018 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3019 (final_hop && !mr_forwarding_allowed))
3020 send_case |= NMR_DST;
3022 send_case |= MR_DST;
3024 if (lnet_msg_is_response(msg))
3025 send_case |= SND_RESP;
3027 /* assign parameters to the send_data */
3028 send_data.sd_rtr_nid = rtr_nid;
3029 send_data.sd_src_nid = src_nid;
3030 send_data.sd_dst_nid = dst_nid;
3031 send_data.sd_best_lpni = lpni;
3033 * keep a pointer to the final destination in case we're going to
3034 * route, so we'll need to access it later
3036 send_data.sd_final_dst_lpni = lpni;
3037 send_data.sd_peer = peer;
3038 send_data.sd_md_cpt = md_cpt;
3039 send_data.sd_send_case = send_case;
3041 rc = lnet_handle_send_case_locked(&send_data);
3044 * Update the local cpt since send_data.sd_cpt might've been
3045 * updated as a result of calling lnet_handle_send_case_locked().
3047 cpt = send_data.sd_cpt;
3049 if (rc == REPEAT_SEND)
3052 lnet_net_unlock(cpt);
3058 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
3060 lnet_nid_t dst_nid = msg->msg_target.nid;
3064 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
3065 * but we might want to use pre-determined router for ACK/REPLY
3068 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3069 LASSERT(msg->msg_txpeer == NULL);
3070 LASSERT(msg->msg_txni == NULL);
3071 LASSERT(!msg->msg_sending);
3072 LASSERT(!msg->msg_target_is_router);
3073 LASSERT(!msg->msg_receiving);
3075 msg->msg_sending = 1;
3077 LASSERT(!msg->msg_tx_committed);
3079 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3081 if (rc == -EHOSTUNREACH)
3082 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3084 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3088 if (rc == LNET_CREDIT_OK)
3089 lnet_ni_send(msg->msg_txni, msg);
3091 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3095 enum lnet_mt_event_type {
3096 MT_TYPE_LOCAL_NI = 0,
3100 struct lnet_mt_event_info {
3101 enum lnet_mt_event_type mt_type;
3105 /* called with res_lock held */
3107 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3109 struct lnet_rsp_tracker *rspt;
3112 * msg has a refcount on the MD so the MD is not going away.
3113 * The rspt queue for the cpt is protected by
3114 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3116 if (!md->md_rspt_ptr)
3119 rspt = md->md_rspt_ptr;
3122 LASSERT(rspt->rspt_cpt == cpt);
3124 md->md_rspt_ptr = NULL;
3126 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3128 * The monitor thread has invalidated this handle because the
3129 * response timed out, but it failed to lookup the MD. That
3130 * means this response tracker is on the zombie list. We can
3131 * safely remove it under the resource lock (held by caller) and
3132 * free the response tracker block.
3134 list_del(&rspt->rspt_on_list);
3135 lnet_rspt_free(rspt, cpt);
3138 * invalidate the handle to indicate that a response has been
3139 * received, which will then lead the monitor thread to clean up
3142 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3147 lnet_clean_zombie_rstqs(void)
3149 struct lnet_rsp_tracker *rspt, *tmp;
3152 cfs_cpt_for_each(i, lnet_cpt_table()) {
3153 list_for_each_entry_safe(rspt, tmp,
3154 the_lnet.ln_mt_zombie_rstqs[i],
3156 list_del(&rspt->rspt_on_list);
3157 lnet_rspt_free(rspt, i);
3161 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3165 lnet_finalize_expired_responses(void)
3167 struct lnet_libmd *md;
3168 struct lnet_rsp_tracker *rspt, *tmp;
3172 if (the_lnet.ln_mt_rstq == NULL)
3175 cfs_cpt_for_each(i, lnet_cpt_table()) {
3176 LIST_HEAD(local_queue);
3179 if (!the_lnet.ln_mt_rstq[i]) {
3183 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3188 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3190 * The rspt mdh will be invalidated when a response
3191 * is received or whenever we want to discard the
3192 * block the monitor thread will walk the queue
3193 * and clean up any rsts with an invalid mdh.
3194 * The monitor thread will walk the queue until
3195 * the first unexpired rspt block. This means that
3196 * some rspt blocks which received their
3197 * corresponding responses will linger in the
3198 * queue until they are cleaned up eventually.
3201 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3203 list_del(&rspt->rspt_on_list);
3204 lnet_rspt_free(rspt, i);
3208 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3209 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3210 struct lnet_peer_ni *lpni;
3213 md = lnet_handle2md(&rspt->rspt_mdh);
3215 /* MD has been queued for unlink, but
3216 * rspt hasn't been detached (Note we've
3217 * checked above that the rspt_mdh is
3218 * valid). Since we cannot lookup the MD
3219 * we're unable to detach the rspt
3220 * ourselves. Thus, move the rspt to the
3221 * zombie list where we'll wait for
3223 * 1. The remaining operations on the
3224 * MD to complete. In this case the
3225 * final operation will result in
3226 * lnet_msg_detach_md()->
3227 * lnet_detach_rsp_tracker() where
3228 * we will clean up this response
3230 * 2. LNet to shutdown. In this case
3231 * we'll wait until after all LND Nets
3232 * have shutdown and then we can
3233 * safely free any remaining response
3234 * tracker blocks on the zombie list.
3235 * Note: We need to hold the resource
3236 * lock when adding to the zombie list
3237 * because we may have concurrent access
3238 * with lnet_detach_rsp_tracker().
3240 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3241 list_move(&rspt->rspt_on_list,
3242 the_lnet.ln_mt_zombie_rstqs[i]);
3246 LASSERT(md->md_rspt_ptr == rspt);
3247 md->md_rspt_ptr = NULL;
3250 LNetMDUnlink(rspt->rspt_mdh);
3252 nid = rspt->rspt_next_hop_nid;
3254 list_del(&rspt->rspt_on_list);
3255 lnet_rspt_free(rspt, i);
3257 /* If we're shutting down we just want to clean
3258 * up the rspt blocks
3260 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3264 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3268 "Response timeout: md = %p: nid = %s\n",
3269 md, libcfs_nid2str(nid));
3272 * If there is a timeout on the response
3273 * from the next hop decrement its health
3274 * value so that we don't use it
3277 lpni = lnet_find_peer_ni_locked(nid);
3279 lnet_handle_remote_failure_locked(lpni);
3280 lnet_peer_ni_decref_locked(lpni);
3289 if (!list_empty(&local_queue)) {
3291 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3298 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3300 struct lnet_msg *msg;
3302 while (!list_empty(resendq)) {
3303 struct lnet_peer_ni *lpni;
3305 msg = list_entry(resendq->next, struct lnet_msg,
3308 list_del_init(&msg->msg_list);
3310 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3312 lnet_net_unlock(cpt);
3313 CERROR("Expected that a peer is already created for %s\n",
3314 libcfs_nid2str(msg->msg_hdr.dest_nid));
3315 msg->msg_no_resend = true;
3316 lnet_finalize(msg, -EFAULT);
3321 lnet_peer_ni_decref_locked(lpni);
3323 lnet_net_unlock(cpt);
3324 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3325 libcfs_nid2str(msg->msg_src_nid_param),
3326 libcfs_id2str(msg->msg_target),
3327 lnet_msgtyp2str(msg->msg_type),
3329 msg->msg_retry_count);
3330 rc = lnet_send(msg->msg_src_nid_param, msg,
3331 msg->msg_rtr_nid_param);
3333 CERROR("Error sending %s to %s: %d\n",
3334 lnet_msgtyp2str(msg->msg_type),
3335 libcfs_id2str(msg->msg_target), rc);
3336 msg->msg_no_resend = true;
3337 lnet_finalize(msg, rc);
3341 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3347 lnet_resend_pending_msgs(void)
3351 cfs_cpt_for_each(i, lnet_cpt_table()) {
3353 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3358 /* called with cpt and ni_lock held */
3360 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3362 struct lnet_handle_md recovery_mdh;
3364 LNetInvalidateMDHandle(&recovery_mdh);
3366 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3368 recovery_mdh = ni->ni_ping_mdh;
3369 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3372 lnet_net_unlock(cpt);
3373 if (!LNetMDHandleIsInvalid(recovery_mdh))
3374 LNetMDUnlink(recovery_mdh);
3380 lnet_recover_local_nis(void)
3382 struct lnet_mt_event_info *ev_info;
3383 LIST_HEAD(processed_list);
3384 LIST_HEAD(local_queue);
3385 struct lnet_handle_md mdh;
3386 struct lnet_ni *tmp;
3394 * splice the recovery queue on a local queue. We will iterate
3395 * through the local queue and update it as needed. Once we're
3396 * done with the traversal, we'll splice the local queue back on
3397 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3398 * will be traversed in the next iteration.
3401 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3405 now = ktime_get_seconds();
3407 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3409 * if an NI is being deleted or it is now healthy, there
3410 * is no need to keep it around in the recovery queue.
3411 * The monitor thread is the only thread responsible for
3412 * removing the NI from the recovery queue.
3413 * Multiple threads can be adding NIs to the recovery
3416 healthv = atomic_read(&ni->ni_healthv);
3420 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3421 healthv == LNET_MAX_HEALTH_VALUE) {
3422 list_del_init(&ni->ni_recovery);
3423 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3425 lnet_ni_decref_locked(ni, 0);
3431 * if the local NI failed recovery we must unlink the md.
3432 * But we want to keep the local_ni on the recovery queue
3433 * so we can continue the attempts to recover it.
3435 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3436 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3437 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3443 if (now < ni->ni_next_ping) {
3450 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3451 libcfs_nid2str(ni->ni_nid));
3454 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3455 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3458 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3460 CERROR("out of memory. Can't recover %s\n",
3461 libcfs_nid2str(ni->ni_nid));
3463 ni->ni_recovery_state &=
3464 ~LNET_NI_RECOVERY_PENDING;
3469 mdh = ni->ni_ping_mdh;
3471 * Invalidate the ni mdh in case it's deleted.
3472 * We'll unlink the mdh in this case below.
3474 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3478 * remove the NI from the local queue and drop the
3479 * reference count to it while we're recovering
3480 * it. The reason for that, is that the NI could
3481 * be deleted, and the way the code is structured
3482 * is if we don't drop the NI, then the deletion
3483 * code will enter a loop waiting for the
3484 * reference count to be removed while holding the
3485 * ln_mutex_lock(). When we look up the peer to
3486 * send to in lnet_select_pathway() we will try to
3487 * lock the ln_mutex_lock() as well, leading to
3488 * a deadlock. By dropping the refcount and
3489 * removing it from the list, we allow for the NI
3490 * to be removed, then we use the cached NID to
3491 * look it up again. If it's gone, then we just
3492 * continue examining the rest of the queue.
3495 list_del_init(&ni->ni_recovery);
3496 lnet_ni_decref_locked(ni, 0);
3499 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3500 ev_info->mt_nid = nid;
3501 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3502 ev_info, the_lnet.ln_mt_handler,
3504 /* lookup the nid again */
3506 ni = lnet_nid2ni_locked(nid, 0);
3509 * the NI has been deleted when we dropped
3516 ni->ni_ping_count++;
3518 ni->ni_ping_mdh = mdh;
3519 lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3524 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3533 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3534 * reexamined in the next iteration.
3536 list_splice_init(&processed_list, &local_queue);
3538 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3543 lnet_resendqs_create(void)
3545 struct list_head **resendqs;
3546 resendqs = lnet_create_array_of_queues();
3551 lnet_net_lock(LNET_LOCK_EX);
3552 the_lnet.ln_mt_resendqs = resendqs;
3553 lnet_net_unlock(LNET_LOCK_EX);
3559 lnet_clean_local_ni_recoveryq(void)
3563 /* This is only called when the monitor thread has stopped */
3566 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3567 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3568 struct lnet_ni, ni_recovery);
3569 list_del_init(&ni->ni_recovery);
3571 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3573 lnet_ni_decref_locked(ni, 0);
3580 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3583 struct lnet_handle_md recovery_mdh;
3585 LNetInvalidateMDHandle(&recovery_mdh);
3587 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3588 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3589 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3591 spin_unlock(&lpni->lpni_lock);
3592 lnet_net_unlock(cpt);
3593 if (!LNetMDHandleIsInvalid(recovery_mdh))
3594 LNetMDUnlink(recovery_mdh);
3596 spin_lock(&lpni->lpni_lock);
3600 lnet_clean_peer_ni_recoveryq(void)
3602 struct lnet_peer_ni *lpni, *tmp;
3604 lnet_net_lock(LNET_LOCK_EX);
3606 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3608 list_del_init(&lpni->lpni_recovery);
3609 spin_lock(&lpni->lpni_lock);
3610 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3611 spin_unlock(&lpni->lpni_lock);
3612 lnet_peer_ni_decref_locked(lpni);
3615 lnet_net_unlock(LNET_LOCK_EX);
3619 lnet_clean_resendqs(void)
3621 struct lnet_msg *msg, *tmp;
3625 cfs_cpt_for_each(i, lnet_cpt_table()) {
3627 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3629 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3630 list_del_init(&msg->msg_list);
3631 msg->msg_no_resend = true;
3632 lnet_finalize(msg, -ESHUTDOWN);
3636 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3640 lnet_recover_peer_nis(void)
3642 struct lnet_mt_event_info *ev_info;
3643 LIST_HEAD(processed_list);
3644 LIST_HEAD(local_queue);
3645 struct lnet_handle_md mdh;
3646 struct lnet_peer_ni *lpni;
3647 struct lnet_peer_ni *tmp;
3654 * Always use cpt 0 for locking across all interactions with
3655 * ln_mt_peerNIRecovq
3658 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3662 now = ktime_get_seconds();
3664 list_for_each_entry_safe(lpni, tmp, &local_queue,
3667 * The same protection strategy is used here as is in the
3668 * local recovery case.
3671 healthv = atomic_read(&lpni->lpni_healthv);
3672 spin_lock(&lpni->lpni_lock);
3673 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3674 healthv == LNET_MAX_HEALTH_VALUE) {
3675 list_del_init(&lpni->lpni_recovery);
3676 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3677 spin_unlock(&lpni->lpni_lock);
3678 lnet_peer_ni_decref_locked(lpni);
3684 * If the peer NI has failed recovery we must unlink the
3685 * md. But we want to keep the peer ni on the recovery
3686 * queue so we can try to continue recovering it
3688 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3689 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3690 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3693 spin_unlock(&lpni->lpni_lock);
3695 if (now < lpni->lpni_next_ping) {
3703 * NOTE: we're racing with peer deletion from user space.
3704 * It's possible that a peer is deleted after we check its
3705 * state. In this case the recovery can create a new peer
3707 spin_lock(&lpni->lpni_lock);
3708 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3709 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3710 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3711 spin_unlock(&lpni->lpni_lock);
3713 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3715 CERROR("out of memory. Can't recover %s\n",
3716 libcfs_nid2str(lpni->lpni_nid));
3717 spin_lock(&lpni->lpni_lock);
3718 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3719 spin_unlock(&lpni->lpni_lock);
3723 /* look at the comments in lnet_recover_local_nis() */
3724 mdh = lpni->lpni_recovery_ping_mdh;
3725 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3726 nid = lpni->lpni_nid;
3728 list_del_init(&lpni->lpni_recovery);
3729 lnet_peer_ni_decref_locked(lpni);
3732 ev_info->mt_type = MT_TYPE_PEER_NI;
3733 ev_info->mt_nid = nid;
3734 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3735 ev_info, the_lnet.ln_mt_handler,
3739 * lnet_find_peer_ni_locked() grabs a refcount for
3740 * us. No need to take it explicitly.
3742 lpni = lnet_find_peer_ni_locked(nid);
3749 lpni->lpni_ping_count++;
3751 lpni->lpni_recovery_ping_mdh = mdh;
3753 lnet_peer_ni_add_to_recoveryq_locked(lpni,
3757 spin_lock(&lpni->lpni_lock);
3758 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3759 spin_unlock(&lpni->lpni_lock);
3762 /* Drop the ref taken by lnet_find_peer_ni_locked() */
3763 lnet_peer_ni_decref_locked(lpni);
3766 spin_unlock(&lpni->lpni_lock);
3769 list_splice_init(&processed_list, &local_queue);
3771 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3776 lnet_monitor_thread(void *arg)
3778 time64_t rsp_timeout = 0;
3781 wait_for_completion(&the_lnet.ln_started);
3783 * The monitor thread takes care of the following:
3784 * 1. Checks the aliveness of routers
3785 * 2. Checks if there are messages on the resend queue to resend
3787 * 3. Check if there are any NIs on the local recovery queue and
3789 * 4. Checks if there are any NIs on the remote recovery queue
3792 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3793 now = ktime_get_real_seconds();
3795 if (lnet_router_checker_active())
3796 lnet_check_routers();
3798 lnet_resend_pending_msgs();
3800 if (now >= rsp_timeout) {
3801 lnet_finalize_expired_responses();
3802 rsp_timeout = now + (lnet_transaction_timeout / 2);
3805 lnet_recover_local_nis();
3806 lnet_recover_peer_nis();
3809 * TODO do we need to check if we should sleep without
3810 * timeout? Technically, an active system will always
3811 * have messages in flight so this check will always
3812 * evaluate to false. And on an idle system do we care
3813 * if we wake up every 1 second? Although, we've seen
3814 * cases where we get a complaint that an idle thread
3815 * is waking up unnecessarily.
3817 wait_for_completion_interruptible_timeout(
3818 &the_lnet.ln_mt_wait_complete,
3819 cfs_time_seconds(1));
3820 /* Must re-init the completion before testing anything,
3821 * including ln_mt_state.
3823 reinit_completion(&the_lnet.ln_mt_wait_complete);
3827 lnet_net_lock(LNET_LOCK_EX);
3828 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3829 lnet_net_unlock(LNET_LOCK_EX);
3831 /* signal that the monitor thread is exiting */
3832 up(&the_lnet.ln_mt_signal);
3840 * Returns == 0 if success
3841 * Returns > 0 if LNetMDBind or prior fails
3842 * Returns < 0 if LNetGet fails
3845 lnet_send_ping(lnet_nid_t dest_nid,
3846 struct lnet_handle_md *mdh, int nnis,
3847 void *user_data, lnet_handler_t handler, bool recovery)
3849 struct lnet_md md = { NULL };
3850 struct lnet_process_id id;
3851 struct lnet_ping_buffer *pbuf;
3854 if (dest_nid == LNET_NID_ANY) {
3859 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3865 /* initialize md content */
3866 md.start = &pbuf->pb_info;
3867 md.length = LNET_PING_INFO_SIZE(nnis);
3868 md.threshold = 2; /* GET/REPLY */
3870 md.options = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3871 md.user_ptr = user_data;
3872 md.handler = handler;
3874 rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3876 lnet_ping_buffer_decref(pbuf);
3877 CERROR("Can't bind MD: %d\n", rc);
3878 rc = -rc; /* change the rc to positive */
3881 id.pid = LNET_PID_LUSTRE;
3884 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3885 LNET_RESERVED_PORTAL,
3886 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3889 goto fail_unlink_md;
3895 LNetInvalidateMDHandle(mdh);
3901 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3902 int status, bool send, bool unlink_event)
3904 lnet_nid_t nid = ev_info->mt_nid;
3906 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3910 ni = lnet_nid2ni_locked(nid, 0);
3916 if (!send || (send && status != 0))
3917 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3919 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3924 CERROR("local NI (%s) recovery failed with %d\n",
3925 libcfs_nid2str(nid), status);
3929 * need to increment healthv for the ni here, because in
3930 * the lnet_finalize() path we don't have access to this
3931 * NI. And in order to get access to it, we'll need to
3932 * carry forward too much information.
3933 * In the peer case, it'll naturally be incremented
3936 lnet_inc_healthv(&ni->ni_healthv,
3937 lnet_health_sensitivity);
3939 struct lnet_peer_ni *lpni;
3942 cpt = lnet_net_lock_current();
3943 lpni = lnet_find_peer_ni_locked(nid);
3945 lnet_net_unlock(cpt);
3948 spin_lock(&lpni->lpni_lock);
3949 if (!send || (send && status != 0))
3950 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3952 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3953 spin_unlock(&lpni->lpni_lock);
3954 lnet_peer_ni_decref_locked(lpni);
3955 lnet_net_unlock(cpt);
3958 CERROR("peer NI (%s) recovery failed with %d\n",
3959 libcfs_nid2str(nid), status);
3964 lnet_mt_event_handler(struct lnet_event *event)
3966 struct lnet_mt_event_info *ev_info = event->md_user_ptr;
3967 struct lnet_ping_buffer *pbuf;
3969 /* TODO: remove assert */
3970 LASSERT(event->type == LNET_EVENT_REPLY ||
3971 event->type == LNET_EVENT_SEND ||
3972 event->type == LNET_EVENT_UNLINK);
3974 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3977 switch (event->type) {
3978 case LNET_EVENT_UNLINK:
3979 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3980 libcfs_nid2str(ev_info->mt_nid));
3982 case LNET_EVENT_REPLY:
3983 lnet_handle_recovery_reply(ev_info, event->status, false,
3984 event->type == LNET_EVENT_UNLINK);
3986 case LNET_EVENT_SEND:
3987 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3988 libcfs_nid2str(ev_info->mt_nid),
3989 (event->status) ? "unsuccessfully" :
3990 "successfully", event->status);
3991 lnet_handle_recovery_reply(ev_info, event->status, true, false);
3994 CERROR("Unexpected event: %d\n", event->type);
3997 if (event->unlinked) {
3998 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3999 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4000 lnet_ping_buffer_decref(pbuf);
4005 lnet_rsp_tracker_create(void)
4007 struct list_head **rstqs;
4008 rstqs = lnet_create_array_of_queues();
4013 the_lnet.ln_mt_rstq = rstqs;
4019 lnet_rsp_tracker_clean(void)
4021 lnet_finalize_expired_responses();
4023 cfs_percpt_free(the_lnet.ln_mt_rstq);
4024 the_lnet.ln_mt_rstq = NULL;
4027 int lnet_monitor_thr_start(void)
4030 struct task_struct *task;
4032 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4035 rc = lnet_resendqs_create();
4039 rc = lnet_rsp_tracker_create();
4043 sema_init(&the_lnet.ln_mt_signal, 0);
4045 lnet_net_lock(LNET_LOCK_EX);
4046 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4047 lnet_net_unlock(LNET_LOCK_EX);
4048 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4051 CERROR("Can't start monitor thread: %d\n", rc);
4058 lnet_net_lock(LNET_LOCK_EX);
4059 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4060 lnet_net_unlock(LNET_LOCK_EX);
4061 /* block until event callback signals exit */
4062 down(&the_lnet.ln_mt_signal);
4064 lnet_net_lock(LNET_LOCK_EX);
4065 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4066 lnet_net_unlock(LNET_LOCK_EX);
4067 lnet_rsp_tracker_clean();
4068 lnet_clean_local_ni_recoveryq();
4069 lnet_clean_peer_ni_recoveryq();
4070 lnet_clean_resendqs();
4071 the_lnet.ln_mt_handler = NULL;
4074 lnet_rsp_tracker_clean();
4075 lnet_clean_local_ni_recoveryq();
4076 lnet_clean_peer_ni_recoveryq();
4077 lnet_clean_resendqs();
4081 void lnet_monitor_thr_stop(void)
4083 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4086 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4087 lnet_net_lock(LNET_LOCK_EX);
4088 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4089 lnet_net_unlock(LNET_LOCK_EX);
4091 /* tell the monitor thread that we're shutting down */
4092 complete(&the_lnet.ln_mt_wait_complete);
4094 /* block until monitor thread signals that it's done */
4095 down(&the_lnet.ln_mt_signal);
4096 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4098 /* perform cleanup tasks */
4099 lnet_rsp_tracker_clean();
4100 lnet_clean_local_ni_recoveryq();
4101 lnet_clean_peer_ni_recoveryq();
4102 lnet_clean_resendqs();
4106 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4110 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4111 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4112 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4113 lnet_net_unlock(cpt);
4115 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4119 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4121 struct lnet_hdr *hdr = &msg->msg_hdr;
4123 if (msg->msg_wanted != 0)
4124 lnet_setpayloadbuffer(msg);
4126 lnet_build_msg_event(msg, LNET_EVENT_PUT);
4128 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
4129 * it back into the ACK during lnet_finalize() */
4130 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4131 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4133 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4134 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4138 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4140 struct lnet_hdr *hdr = &msg->msg_hdr;
4141 struct lnet_match_info info;
4145 /* Convert put fields to host byte order */
4146 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4147 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
4148 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
4150 /* Primary peer NID. */
4151 info.mi_id.nid = msg->msg_initiator;
4152 info.mi_id.pid = hdr->src_pid;
4153 info.mi_opc = LNET_MD_OP_PUT;
4154 info.mi_portal = hdr->msg.put.ptl_index;
4155 info.mi_rlength = hdr->payload_length;
4156 info.mi_roffset = hdr->msg.put.offset;
4157 info.mi_mbits = hdr->msg.put.match_bits;
4158 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
4160 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4161 ready_delay = msg->msg_rx_ready_delay;
4164 rc = lnet_ptl_match_md(&info, msg);
4169 case LNET_MATCHMD_OK:
4170 lnet_recv_put(ni, msg);
4173 case LNET_MATCHMD_NONE:
4175 /* no eager_recv or has already called it, should
4176 * have been attached on delayed list */
4179 rc = lnet_ni_eager_recv(ni, msg);
4186 case LNET_MATCHMD_DROP:
4187 CNETERR("Dropping PUT from %s portal %d match %llu"
4188 " offset %d length %d: %d\n",
4189 libcfs_id2str(info.mi_id), info.mi_portal,
4190 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4192 return -ENOENT; /* -ve: OK but no match */
4197 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4199 struct lnet_match_info info;
4200 struct lnet_hdr *hdr = &msg->msg_hdr;
4201 struct lnet_process_id source_id;
4202 struct lnet_handle_wire reply_wmd;
4205 /* Convert get fields to host byte order */
4206 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
4207 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
4208 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
4209 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
4211 source_id.nid = hdr->src_nid;
4212 source_id.pid = hdr->src_pid;
4213 /* Primary peer NID */
4214 info.mi_id.nid = msg->msg_initiator;
4215 info.mi_id.pid = hdr->src_pid;
4216 info.mi_opc = LNET_MD_OP_GET;
4217 info.mi_portal = hdr->msg.get.ptl_index;
4218 info.mi_rlength = hdr->msg.get.sink_length;
4219 info.mi_roffset = hdr->msg.get.src_offset;
4220 info.mi_mbits = hdr->msg.get.match_bits;
4221 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
4223 rc = lnet_ptl_match_md(&info, msg);
4224 if (rc == LNET_MATCHMD_DROP) {
4225 CNETERR("Dropping GET from %s portal %d match %llu"
4226 " offset %d length %d\n",
4227 libcfs_id2str(info.mi_id), info.mi_portal,
4228 info.mi_mbits, info.mi_roffset, info.mi_rlength);
4229 return -ENOENT; /* -ve: OK but no match */
4232 LASSERT(rc == LNET_MATCHMD_OK);
4234 lnet_build_msg_event(msg, LNET_EVENT_GET);
4236 reply_wmd = hdr->msg.get.return_wmd;
4238 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
4239 msg->msg_offset, msg->msg_wanted);
4241 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4244 /* The LND completes the REPLY from her recv procedure */
4245 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4246 msg->msg_offset, msg->msg_len, msg->msg_len);
4250 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4251 msg->msg_receiving = 0;
4253 rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
4255 /* didn't get as far as lnet_ni_send() */
4256 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4257 libcfs_nid2str(ni->ni_nid),
4258 libcfs_id2str(info.mi_id), rc);
4260 lnet_finalize(msg, rc);
4267 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4269 void *private = msg->msg_private;
4270 struct lnet_hdr *hdr = &msg->msg_hdr;
4271 struct lnet_process_id src = {0};
4272 struct lnet_libmd *md;
4273 unsigned int rlength;
4274 unsigned int mlength;
4277 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4280 src.nid = hdr->src_nid;
4281 src.pid = hdr->src_pid;
4283 /* NB handles only looked up by creator (no flips) */
4284 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4285 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4286 CNETERR("%s: Dropping REPLY from %s for %s "
4288 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4289 (md == NULL) ? "invalid" : "inactive",
4290 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4291 hdr->msg.reply.dst_wmd.wh_object_cookie);
4292 if (md != NULL && md->md_me != NULL)
4293 CERROR("REPLY MD also attached to portal %d\n",
4294 md->md_me->me_portal);
4296 lnet_res_unlock(cpt);
4297 return -ENOENT; /* -ve: OK but no match */
4300 LASSERT(md->md_offset == 0);
4302 rlength = hdr->payload_length;
4303 mlength = min(rlength, md->md_length);
4305 if (mlength < rlength &&
4306 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4307 CNETERR("%s: Dropping REPLY from %s length %d "
4308 "for MD %#llx would overflow (%d)\n",
4309 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4310 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4312 lnet_res_unlock(cpt);
4313 return -ENOENT; /* -ve: OK but no match */
4316 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4317 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4318 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4320 lnet_msg_attach_md(msg, md, 0, mlength);
4323 lnet_setpayloadbuffer(msg);
4325 lnet_res_unlock(cpt);
4327 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4329 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4334 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4336 struct lnet_hdr *hdr = &msg->msg_hdr;
4337 struct lnet_process_id src = {0};
4338 struct lnet_libmd *md;
4341 src.nid = hdr->src_nid;
4342 src.pid = hdr->src_pid;
4344 /* Convert ack fields to host byte order */
4345 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4346 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4348 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4351 /* NB handles only looked up by creator (no flips) */
4352 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4353 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4354 /* Don't moan; this is expected */
4356 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4357 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4358 (md == NULL) ? "invalid" : "inactive",
4359 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4360 hdr->msg.ack.dst_wmd.wh_object_cookie);
4361 if (md != NULL && md->md_me != NULL)
4362 CERROR("Source MD also attached to portal %d\n",
4363 md->md_me->me_portal);
4365 lnet_res_unlock(cpt);
4366 return -ENOENT; /* -ve! */
4369 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4370 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4371 hdr->msg.ack.dst_wmd.wh_object_cookie);
4373 lnet_msg_attach_md(msg, md, 0, 0);
4375 lnet_res_unlock(cpt);
4377 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4379 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4384 * \retval LNET_CREDIT_OK If \a msg is forwarded
4385 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4386 * \retval -ve error code
4389 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4393 if (!the_lnet.ln_routing)
4396 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4397 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4398 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4399 msg->msg_rx_ready_delay = 1;
4401 lnet_net_unlock(msg->msg_rx_cpt);
4402 rc = lnet_ni_eager_recv(ni, msg);
4403 lnet_net_lock(msg->msg_rx_cpt);
4408 rc = lnet_post_routed_recv_locked(msg, 0);
4413 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4417 switch (msg->msg_type) {
4419 rc = lnet_parse_ack(ni, msg);
4422 rc = lnet_parse_put(ni, msg);
4425 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4427 case LNET_MSG_REPLY:
4428 rc = lnet_parse_reply(ni, msg);
4430 default: /* prevent an unused label if !kernel */
4435 LASSERT(rc == 0 || rc == -ENOENT);
4440 lnet_msgtyp2str (int type)
4449 case LNET_MSG_REPLY:
4451 case LNET_MSG_HELLO:
4454 return ("<UNKNOWN>");
4459 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4460 void *private, int rdma_req)
4462 struct lnet_peer_ni *lpni;
4463 struct lnet_msg *msg;
4464 __u32 payload_length;
4465 lnet_pid_t dest_pid;
4466 lnet_nid_t dest_nid;
4474 LASSERT (!in_interrupt ());
4476 type = le32_to_cpu(hdr->type);
4477 src_nid = le64_to_cpu(hdr->src_nid);
4478 dest_nid = le64_to_cpu(hdr->dest_nid);
4479 dest_pid = le32_to_cpu(hdr->dest_pid);
4480 payload_length = le32_to_cpu(hdr->payload_length);
4482 for_me = (ni->ni_nid == dest_nid);
4483 cpt = lnet_cpt_of_nid(from_nid, ni);
4485 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4486 libcfs_nid2str(dest_nid),
4487 libcfs_nid2str(ni->ni_nid),
4488 libcfs_nid2str(src_nid),
4489 lnet_msgtyp2str(type),
4490 (for_me) ? "for me" : "routed");
4495 if (payload_length > 0) {
4496 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4497 libcfs_nid2str(from_nid),
4498 libcfs_nid2str(src_nid),
4499 lnet_msgtyp2str(type), payload_length);
4505 case LNET_MSG_REPLY:
4506 if (payload_length >
4507 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4508 CERROR("%s, src %s: bad %s payload %d "
4509 "(%d max expected)\n",
4510 libcfs_nid2str(from_nid),
4511 libcfs_nid2str(src_nid),
4512 lnet_msgtyp2str(type),
4514 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4520 CERROR("%s, src %s: Bad message type 0x%x\n",
4521 libcfs_nid2str(from_nid),
4522 libcfs_nid2str(src_nid), type);
4526 if (the_lnet.ln_routing &&
4527 ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4529 spin_lock(&ni->ni_net->net_lock);
4530 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4531 spin_unlock(&ni->ni_net->net_lock);
4532 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4537 lnet_push_update_to_peers(1);
4539 /* Regard a bad destination NID as a protocol error. Senders should
4540 * know what they're doing; if they don't they're misconfigured, buggy
4541 * or malicious so we chop them off at the knees :) */
4544 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4545 /* should have gone direct */
4546 CERROR("%s, src %s: Bad dest nid %s "
4547 "(should have been sent direct)\n",
4548 libcfs_nid2str(from_nid),
4549 libcfs_nid2str(src_nid),
4550 libcfs_nid2str(dest_nid));
4554 if (lnet_islocalnid(dest_nid)) {
4555 /* dest is another local NI; sender should have used
4556 * this node's NID on its own network */
4557 CERROR("%s, src %s: Bad dest nid %s "
4558 "(it's my nid but on a different network)\n",
4559 libcfs_nid2str(from_nid),
4560 libcfs_nid2str(src_nid),
4561 libcfs_nid2str(dest_nid));
4565 if (rdma_req && type == LNET_MSG_GET) {
4566 CERROR("%s, src %s: Bad optimized GET for %s "
4567 "(final destination must be me)\n",
4568 libcfs_nid2str(from_nid),
4569 libcfs_nid2str(src_nid),
4570 libcfs_nid2str(dest_nid));
4574 if (!the_lnet.ln_routing) {
4575 CERROR("%s, src %s: Dropping message for %s "
4576 "(routing not enabled)\n",
4577 libcfs_nid2str(from_nid),
4578 libcfs_nid2str(src_nid),
4579 libcfs_nid2str(dest_nid));
4584 /* Message looks OK; we're not going to return an error, so we MUST
4585 * call back lnd_recv() come what may... */
4587 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4588 fail_peer(src_nid, 0)) { /* shall we now? */
4589 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4590 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4591 lnet_msgtyp2str(type));
4595 if (!list_empty(&the_lnet.ln_drop_rules) &&
4596 lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4598 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4599 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4600 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4604 msg = lnet_msg_alloc();
4606 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4607 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4608 lnet_msgtyp2str(type));
4612 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4613 * pointers NULL etc */
4615 msg->msg_type = type;
4616 msg->msg_private = private;
4617 msg->msg_receiving = 1;
4618 msg->msg_rdma_get = rdma_req;
4619 msg->msg_len = msg->msg_wanted = payload_length;
4620 msg->msg_offset = 0;
4621 msg->msg_hdr = *hdr;
4622 /* for building message event */
4623 msg->msg_from = from_nid;
4625 msg->msg_target.pid = dest_pid;
4626 msg->msg_target.nid = dest_nid;
4627 msg->msg_routing = 1;
4630 /* convert common msg->hdr fields to host byteorder */
4631 msg->msg_hdr.type = type;
4632 msg->msg_hdr.src_nid = src_nid;
4633 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4634 msg->msg_hdr.dest_nid = dest_nid;
4635 msg->msg_hdr.dest_pid = dest_pid;
4636 msg->msg_hdr.payload_length = payload_length;
4640 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4642 lnet_net_unlock(cpt);
4643 CERROR("%s, src %s: Dropping %s "
4644 "(error %ld looking up sender)\n",
4645 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4646 lnet_msgtyp2str(type), PTR_ERR(lpni));
4648 if (rc == -ESHUTDOWN)
4649 /* We are shutting down. Don't do anything more */
4654 /* If this message was forwarded to us from a router then we may need
4655 * to update router aliveness or check for an asymmetrical route
4658 if (((lnet_drop_asym_route && for_me) ||
4659 !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4660 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4661 __u32 src_net_id = LNET_NIDNET(src_nid);
4662 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4663 struct lnet_route *route;
4666 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4667 if (route->lr_net == src_net_id) {
4669 /* If we're transitioning the gateway from
4670 * dead -> alive, and discovery is disabled
4671 * locally or on the gateway, then we need to
4672 * update the cached route aliveness for each
4673 * route to the src_nid's net.
4675 * Otherwise, we're only checking for
4676 * symmetrical route, and we can break the
4679 if (!gw->lp_alive &&
4680 lnet_is_discovery_disabled(gw))
4681 lnet_set_route_aliveness(route, true);
4686 if (lnet_drop_asym_route && for_me && !found) {
4687 lnet_net_unlock(cpt);
4688 /* we would not use from_nid to route a message to
4690 * => asymmetric routing detected but forbidden
4692 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4693 libcfs_nid2str(from_nid),
4694 libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4698 if (!gw->lp_alive) {
4699 struct lnet_peer_net *lpn;
4700 struct lnet_peer_ni *lpni2;
4702 gw->lp_alive = true;
4703 /* Mark all remote NIs on src_nid's net UP */
4704 lpn = lnet_peer_get_net_locked(gw, src_net_id);
4706 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4708 lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4712 lpni->lpni_last_alive = ktime_get_seconds();
4714 msg->msg_rxpeer = lpni;
4716 lnet_ni_addref_locked(ni, cpt);
4717 /* Multi-Rail: Primary NID of source. */
4718 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4721 * mark the status of this lpni as UP since we received a message
4722 * from it. The ping response reports back the ns_status which is
4723 * marked on the remote as up or down and we cache it here.
4725 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4727 lnet_msg_commit(msg, cpt);
4729 /* message delay simulation */
4730 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4731 lnet_delay_rule_match_locked(hdr, msg))) {
4732 lnet_net_unlock(cpt);
4737 rc = lnet_parse_forward_locked(ni, msg);
4738 lnet_net_unlock(cpt);
4743 if (rc == LNET_CREDIT_OK) {
4744 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4745 0, payload_length, payload_length);
4750 lnet_net_unlock(cpt);
4752 rc = lnet_parse_local(ni, msg);
4758 LASSERT(msg->msg_md == NULL);
4759 lnet_finalize(msg, rc);
4762 lnet_drop_message(ni, cpt, private, payload_length, type);
4765 EXPORT_SYMBOL(lnet_parse);
4768 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4770 while (!list_empty(head)) {
4771 struct lnet_process_id id = {0};
4772 struct lnet_msg *msg;
4774 msg = list_entry(head->next, struct lnet_msg, msg_list);
4775 list_del(&msg->msg_list);
4777 id.nid = msg->msg_hdr.src_nid;
4778 id.pid = msg->msg_hdr.src_pid;
4780 LASSERT(msg->msg_md == NULL);
4781 LASSERT(msg->msg_rx_delayed);
4782 LASSERT(msg->msg_rxpeer != NULL);
4783 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4785 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4786 " offset %d length %d: %s\n",
4788 msg->msg_hdr.msg.put.ptl_index,
4789 msg->msg_hdr.msg.put.match_bits,
4790 msg->msg_hdr.msg.put.offset,
4791 msg->msg_hdr.payload_length, reason);
4793 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4794 * called lnet_drop_message(), so I just hang onto msg as well
4795 * until that's done */
4797 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4798 msg->msg_private, msg->msg_len,
4801 msg->msg_no_resend = true;
4803 * NB: message will not generate event because w/o attached MD,
4804 * but we still should give error code so lnet_msg_decommit()
4805 * can skip counters operations and other checks.
4807 lnet_finalize(msg, -ENOENT);
4812 lnet_recv_delayed_msg_list(struct list_head *head)
4814 while (!list_empty(head)) {
4815 struct lnet_msg *msg;
4816 struct lnet_process_id id;
4818 msg = list_entry(head->next, struct lnet_msg, msg_list);
4819 list_del(&msg->msg_list);
4821 /* md won't disappear under me, since each msg
4822 * holds a ref on it */
4824 id.nid = msg->msg_hdr.src_nid;
4825 id.pid = msg->msg_hdr.src_pid;
4827 LASSERT(msg->msg_rx_delayed);
4828 LASSERT(msg->msg_md != NULL);
4829 LASSERT(msg->msg_rxpeer != NULL);
4830 LASSERT(msg->msg_rxni != NULL);
4831 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4833 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4834 "match %llu offset %d length %d.\n",
4835 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4836 msg->msg_hdr.msg.put.match_bits,
4837 msg->msg_hdr.msg.put.offset,
4838 msg->msg_hdr.payload_length);
4840 lnet_recv_put(msg->msg_rxni, msg);
4845 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4846 struct lnet_libmd *md, struct lnet_handle_md mdh)
4849 struct lnet_rsp_tracker *local_rspt;
4852 * MD has a refcount taken by message so it's not going away.
4853 * The MD however can be looked up. We need to secure the access
4854 * to the md_rspt_ptr by taking the res_lock.
4855 * The rspt can be accessed without protection up to when it gets
4856 * added to the list.
4860 local_rspt = md->md_rspt_ptr;
4861 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4862 if (local_rspt != NULL) {
4864 * we already have an rspt attached to the md, so we'll
4865 * update the deadline on that one.
4867 lnet_rspt_free(rspt, cpt);
4870 rspt->rspt_mdh = mdh;
4871 rspt->rspt_cpt = cpt;
4872 /* store the rspt so we can access it when we get the REPLY */
4873 md->md_rspt_ptr = rspt;
4876 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4879 * add to the list of tracked responses. It's added to tail of the
4880 * list in order to expire all the older entries first.
4883 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4884 lnet_net_unlock(cpt);
4885 lnet_res_unlock(cpt);
4889 * Initiate an asynchronous PUT operation.
4891 * There are several events associated with a PUT: completion of the send on
4892 * the initiator node (LNET_EVENT_SEND), and when the send completes
4893 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4894 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4895 * used at the target node to indicate the completion of incoming data
4898 * The local events will be logged in the EQ associated with the MD pointed to
4899 * by \a mdh handle. Using a MD without an associated EQ results in these
4900 * events being discarded. In this case, the caller must have another
4901 * mechanism (e.g., a higher level protocol) for determining when it is safe
4902 * to modify the memory region associated with the MD.
4904 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4905 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4907 * \param self Indicates the NID of a local interface through which to send
4908 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4909 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4910 * must be "free floating" (See LNetMDBind()).
4911 * \param ack Controls whether an acknowledgment is requested.
4912 * Acknowledgments are only sent when they are requested by the initiating
4913 * process and the target MD enables them.
4914 * \param target A process identifier for the target process.
4915 * \param portal The index in the \a target's portal table.
4916 * \param match_bits The match bits to use for MD selection at the target
4918 * \param offset The offset into the target MD (only used when the target
4919 * MD has the LNET_MD_MANAGE_REMOTE option set).
4920 * \param hdr_data 64 bits of user data that can be included in the message
4921 * header. This data is written to an event queue entry at the target if an
4922 * EQ is present on the matching MD.
4924 * \retval 0 Success, and only in this case events will be generated
4925 * and logged to EQ (if it exists).
4926 * \retval -EIO Simulated failure.
4927 * \retval -ENOMEM Memory allocation failure.
4928 * \retval -ENOENT Invalid MD object.
4930 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4933 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4934 struct lnet_process_id target, unsigned int portal,
4935 __u64 match_bits, unsigned int offset,
4938 struct lnet_msg *msg;
4939 struct lnet_libmd *md;
4942 struct lnet_rsp_tracker *rspt = NULL;
4944 LASSERT(the_lnet.ln_refcount > 0);
4946 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4947 fail_peer(target.nid, 1)) { /* shall we now? */
4948 CERROR("Dropping PUT to %s: simulated failure\n",
4949 libcfs_id2str(target));
4953 msg = lnet_msg_alloc();
4955 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4956 libcfs_id2str(target));
4959 msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
4961 cpt = lnet_cpt_of_cookie(mdh.cookie);
4963 if (ack == LNET_ACK_REQ) {
4964 rspt = lnet_rspt_alloc(cpt);
4966 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4967 libcfs_id2str(target));
4970 INIT_LIST_HEAD(&rspt->rspt_on_list);
4975 md = lnet_handle2md(&mdh);
4976 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4977 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4978 match_bits, portal, libcfs_id2str(target),
4979 md == NULL ? -1 : md->md_threshold);
4980 if (md != NULL && md->md_me != NULL)
4981 CERROR("Source MD also attached to portal %d\n",
4982 md->md_me->me_portal);
4983 lnet_res_unlock(cpt);
4986 lnet_rspt_free(rspt, cpt);
4992 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4994 lnet_msg_attach_md(msg, md, 0, 0);
4996 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4998 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4999 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5000 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5001 msg->msg_hdr.msg.put.hdr_data = hdr_data;
5003 /* NB handles only looked up by creator (no flips) */
5004 if (ack == LNET_ACK_REQ) {
5005 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5006 the_lnet.ln_interface_cookie;
5007 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5008 md->md_lh.lh_cookie;
5010 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5011 LNET_WIRE_HANDLE_COOKIE_NONE;
5012 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5013 LNET_WIRE_HANDLE_COOKIE_NONE;
5016 lnet_res_unlock(cpt);
5018 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5020 if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5022 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5024 lnet_rspt_free(rspt, cpt);
5026 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5030 rc = lnet_send(self, msg, LNET_NID_ANY);
5033 CNETERR("Error sending PUT to %s: %d\n",
5034 libcfs_id2str(target), rc);
5035 msg->msg_no_resend = true;
5036 lnet_finalize(msg, rc);
5039 /* completion will be signalled by an event */
5042 EXPORT_SYMBOL(LNetPut);
5045 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
5046 * returns a msg for the LND to pass to lnet_finalize() when the sink
5047 * data has been received.
5049 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5050 * lnet_finalize() is called on it, so the LND must call this first
5053 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5055 struct lnet_msg *msg = lnet_msg_alloc();
5056 struct lnet_libmd *getmd = getmsg->msg_md;
5057 struct lnet_process_id peer_id = getmsg->msg_target;
5060 LASSERT(!getmsg->msg_target_is_router);
5061 LASSERT(!getmsg->msg_routing);
5064 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5065 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
5069 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5072 LASSERT(getmd->md_refcount > 0);
5074 if (getmd->md_threshold == 0) {
5075 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5076 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
5078 lnet_res_unlock(cpt);
5082 LASSERT(getmd->md_offset == 0);
5084 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5085 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
5087 /* setup information for lnet_build_msg_event */
5088 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5089 msg->msg_from = peer_id.nid;
5090 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5091 msg->msg_hdr.src_nid = peer_id.nid;
5092 msg->msg_hdr.payload_length = getmd->md_length;
5093 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5095 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5096 lnet_res_unlock(cpt);
5098 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
5101 lnet_msg_commit(msg, cpt);
5102 lnet_net_unlock(cpt);
5104 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5109 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
5112 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5113 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5114 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5116 lnet_net_unlock(cpt);
5123 EXPORT_SYMBOL(lnet_create_reply_msg);
5126 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5129 /* Set the REPLY length, now the RDMA that elides the REPLY message has
5130 * completed and I know it. */
5131 LASSERT(reply != NULL);
5132 LASSERT(reply->msg_type == LNET_MSG_GET);
5133 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5135 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
5136 * the end of my buffer, I might as well be dead. */
5137 LASSERT(len <= reply->msg_ev.mlength);
5139 reply->msg_ev.mlength = len;
5141 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5144 * Initiate an asynchronous GET operation.
5146 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5147 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5148 * the target node in the REPLY has been written to local MD.
5150 * On the target node, an LNET_EVENT_GET is logged when the GET request
5151 * arrives and is accepted into a MD.
5153 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5154 * \param mdh A handle for the MD that describes the memory into which the
5155 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5157 * \retval 0 Success, and only in this case events will be generated
5158 * and logged to EQ (if it exists) of the MD.
5159 * \retval -EIO Simulated failure.
5160 * \retval -ENOMEM Memory allocation failure.
5161 * \retval -ENOENT Invalid MD object.
5164 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
5165 struct lnet_process_id target, unsigned int portal,
5166 __u64 match_bits, unsigned int offset, bool recovery)
5168 struct lnet_msg *msg;
5169 struct lnet_libmd *md;
5170 struct lnet_rsp_tracker *rspt;
5174 LASSERT(the_lnet.ln_refcount > 0);
5176 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5177 fail_peer(target.nid, 1)) /* shall we now? */
5179 CERROR("Dropping GET to %s: simulated failure\n",
5180 libcfs_id2str(target));
5184 msg = lnet_msg_alloc();
5186 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5187 libcfs_id2str(target));
5191 cpt = lnet_cpt_of_cookie(mdh.cookie);
5193 rspt = lnet_rspt_alloc(cpt);
5195 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5196 libcfs_id2str(target));
5199 INIT_LIST_HEAD(&rspt->rspt_on_list);
5201 msg->msg_recovery = recovery;
5205 md = lnet_handle2md(&mdh);
5206 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5207 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5208 match_bits, portal, libcfs_id2str(target),
5209 md == NULL ? -1 : md->md_threshold);
5210 if (md != NULL && md->md_me != NULL)
5211 CERROR("REPLY MD also attached to portal %d\n",
5212 md->md_me->me_portal);
5214 lnet_res_unlock(cpt);
5217 lnet_rspt_free(rspt, cpt);
5221 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5223 lnet_msg_attach_md(msg, md, 0, 0);
5225 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5227 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5228 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5229 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5230 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5232 /* NB handles only looked up by creator (no flips) */
5233 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5234 the_lnet.ln_interface_cookie;
5235 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5236 md->md_lh.lh_cookie;
5238 lnet_res_unlock(cpt);
5240 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5242 if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5243 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5245 lnet_rspt_free(rspt, cpt);
5247 rc = lnet_send(self, msg, LNET_NID_ANY);
5249 CNETERR("Error sending GET to %s: %d\n",
5250 libcfs_id2str(target), rc);
5251 msg->msg_no_resend = true;
5252 lnet_finalize(msg, rc);
5255 /* completion will be signalled by an event */
5258 EXPORT_SYMBOL(LNetGet);
5261 * Calculate distance to node at \a dstnid.
5263 * \param dstnid Target NID.
5264 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5266 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5269 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5270 * local_nid_dist_zero is set, which is the default.
5271 * \retval positives Distance to target NID, i.e. number of hops plus one.
5272 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5275 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5277 struct list_head *e;
5278 struct lnet_ni *ni = NULL;
5279 struct lnet_remotenet *rnet;
5280 __u32 dstnet = LNET_NIDNET(dstnid);
5284 struct list_head *rn_list;
5285 bool matched_dstnet = false;
5287 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5288 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5289 * keep order 0 free for 0@lo and order 1 free for a local NID
5292 LASSERT(the_lnet.ln_refcount > 0);
5294 cpt = lnet_net_lock_current();
5296 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5297 if (ni->ni_nid == dstnid) {
5298 if (srcnidp != NULL)
5300 if (orderp != NULL) {
5301 if (dstnid == LNET_NID_LO_0)
5306 lnet_net_unlock(cpt);
5308 return local_nid_dist_zero ? 0 : 1;
5311 if (!matched_dstnet && LNET_NIDNET(ni->ni_nid) == dstnet) {
5312 matched_dstnet = true;
5313 /* We matched the destination net, but we may have
5314 * additional local NIs to inspect.
5316 * We record the nid and order as appropriate, but
5317 * they may be overwritten if we match local NI above.
5320 *srcnidp = ni->ni_nid;
5323 /* Check if ni was originally created in
5324 * current net namespace.
5325 * If not, assign order above 0xffff0000,
5326 * to make this ni not a priority.
5328 if (current->nsproxy &&
5329 !net_eq(ni->ni_net_ns,
5330 current->nsproxy->net_ns))
5331 *orderp = order + 0xffff0000;
5340 if (matched_dstnet) {
5341 lnet_net_unlock(cpt);
5345 rn_list = lnet_net2rnethash(dstnet);
5346 list_for_each(e, rn_list) {
5347 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5349 if (rnet->lrn_net == dstnet) {
5350 struct lnet_route *route;
5351 struct lnet_route *shortest = NULL;
5352 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5355 LASSERT(!list_empty(&rnet->lrn_routes));
5357 list_for_each_entry(route, &rnet->lrn_routes,
5359 route_hops = route->lr_hops;
5360 if (route_hops == LNET_UNDEFINED_HOPS)
5362 if (shortest == NULL ||
5363 route_hops < shortest_hops) {
5365 shortest_hops = route_hops;
5369 LASSERT(shortest != NULL);
5370 hops = shortest_hops;
5371 if (srcnidp != NULL) {
5372 struct lnet_net *net;
5373 net = lnet_get_net_locked(shortest->lr_lnet);
5375 ni = lnet_get_next_ni_locked(net, NULL);
5376 *srcnidp = ni->ni_nid;
5380 lnet_net_unlock(cpt);
5386 lnet_net_unlock(cpt);
5387 return -EHOSTUNREACH;
5389 EXPORT_SYMBOL(LNetDist);