4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lnet/lnet/lib-move.c
33 * Data movement routines
36 #define DEBUG_SUBSYSTEM S_LNET
38 #include <linux/pagemap.h>
40 #include <lnet/lib-lnet.h>
41 #include <linux/nsproxy.h>
42 #include <lnet/lnet_rdma.h>
43 #include <net/net_namespace.h>
45 static int local_nid_dist_zero = 1;
46 module_param(local_nid_dist_zero, int, 0444);
47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
49 struct lnet_send_data {
50 struct lnet_ni *sd_best_ni;
51 struct lnet_peer_ni *sd_best_lpni;
52 struct lnet_peer_ni *sd_final_dst_lpni;
53 struct lnet_peer *sd_peer;
54 struct lnet_peer *sd_gw_peer;
55 struct lnet_peer_ni *sd_gw_lpni;
56 struct lnet_peer_net *sd_peer_net;
57 struct lnet_msg *sd_msg;
58 struct lnet_nid sd_dst_nid;
59 struct lnet_nid sd_src_nid;
60 struct lnet_nid sd_rtr_nid;
67 lnet_msg_is_response(struct lnet_msg *msg)
69 return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
73 lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
75 if (md_options & LNET_MD_NO_TRACK_RESPONSE)
76 /* Explicitly disabled in MD options */
79 if (md_options & LNET_MD_TRACK_RESPONSE)
80 /* Explicity enabled in MD options */
83 if (lnet_response_tracking == 3)
84 /* Enabled for all message types */
87 if (msg_type == LNET_MSG_PUT)
88 return lnet_response_tracking == 2;
90 if (msg_type == LNET_MSG_GET)
91 return lnet_response_tracking == 1;
96 static inline struct lnet_comm_count *
97 get_stats_counts(struct lnet_element_stats *stats,
98 enum lnet_stats_type stats_type)
100 switch (stats_type) {
101 case LNET_STATS_TYPE_SEND:
102 return &stats->el_send_stats;
103 case LNET_STATS_TYPE_RECV:
104 return &stats->el_recv_stats;
105 case LNET_STATS_TYPE_DROP:
106 return &stats->el_drop_stats;
108 CERROR("Unknown stats type\n");
114 void lnet_incr_stats(struct lnet_element_stats *stats,
115 enum lnet_msg_type msg_type,
116 enum lnet_stats_type stats_type)
118 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
124 atomic_inc(&counts->co_ack_count);
127 atomic_inc(&counts->co_put_count);
130 atomic_inc(&counts->co_get_count);
133 atomic_inc(&counts->co_reply_count);
136 atomic_inc(&counts->co_hello_count);
139 CERROR("There is a BUG in the code. Unknown message type\n");
144 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
145 enum lnet_stats_type stats_type)
147 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
151 return (atomic_read(&counts->co_ack_count) +
152 atomic_read(&counts->co_put_count) +
153 atomic_read(&counts->co_get_count) +
154 atomic_read(&counts->co_reply_count) +
155 atomic_read(&counts->co_hello_count));
158 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
159 struct lnet_comm_count *counts)
161 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
162 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
163 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
164 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
165 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
168 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
169 struct lnet_element_stats *stats)
171 struct lnet_comm_count *counts;
176 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
179 assign_stats(&msg_stats->im_send_stats, counts);
181 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
184 assign_stats(&msg_stats->im_recv_stats, counts);
186 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
189 assign_stats(&msg_stats->im_drop_stats, counts);
193 lnet_fail_nid(lnet_nid_t nid4, unsigned int threshold)
195 struct lnet_test_peer *tp;
196 struct list_head *el;
197 struct list_head *next;
201 lnet_nid4_to_nid(nid4, &nid);
202 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
203 if (threshold != 0) {
204 /* Adding a new entry */
205 LIBCFS_ALLOC(tp, sizeof(*tp));
210 tp->tp_threshold = threshold;
213 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
220 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
221 tp = list_entry(el, struct lnet_test_peer, tp_list);
223 if (tp->tp_threshold == 0 || /* needs culling anyway */
224 LNET_NID_IS_ANY(&nid) || /* removing all entries */
225 nid_same(&tp->tp_nid, &nid)) { /* matched this one */
226 list_move(&tp->tp_list, &cull);
232 while (!list_empty(&cull)) {
233 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
235 list_del(&tp->tp_list);
236 LIBCFS_FREE(tp, sizeof(*tp));
242 fail_peer(lnet_nid_t nid4, int outgoing)
244 struct lnet_test_peer *tp;
245 struct list_head *el;
246 struct list_head *next;
251 lnet_nid4_to_nid(nid4, &nid);
252 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
255 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
256 tp = list_entry(el, struct lnet_test_peer, tp_list);
258 if (tp->tp_threshold == 0) {
261 /* only cull zombies on outgoing tests,
262 * since we may be at interrupt priority on
263 * incoming messages. */
264 list_move(&tp->tp_list, &cull);
269 if (LNET_NID_IS_ANY(&tp->tp_nid) || /* fail every peer */
270 nid_same(&nid, &tp->tp_nid)) { /* fail this peer */
273 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
276 tp->tp_threshold == 0) {
278 list_move(&tp->tp_list, &cull);
287 while (!list_empty(&cull)) {
288 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
289 list_del(&tp->tp_list);
291 LIBCFS_FREE(tp, sizeof(*tp));
298 lnet_iov_nob(unsigned int niov, struct kvec *iov)
300 unsigned int nob = 0;
302 LASSERT(niov == 0 || iov != NULL);
304 nob += (iov++)->iov_len;
308 EXPORT_SYMBOL(lnet_iov_nob);
311 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
312 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
315 /* NB diov, siov are READ-ONLY */
316 unsigned int this_nob;
321 /* skip complete frags before 'doffset' */
323 while (doffset >= diov->iov_len) {
324 doffset -= diov->iov_len;
330 /* skip complete frags before 'soffset' */
332 while (soffset >= siov->iov_len) {
333 soffset -= siov->iov_len;
342 this_nob = min3((unsigned int)diov->iov_len - doffset,
343 (unsigned int)siov->iov_len - soffset,
346 memcpy((char *)diov->iov_base + doffset,
347 (char *)siov->iov_base + soffset, this_nob);
350 if (diov->iov_len > doffset + this_nob) {
358 if (siov->iov_len > soffset + this_nob) {
367 EXPORT_SYMBOL(lnet_copy_iov2iov);
370 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
372 unsigned int nob = 0;
374 LASSERT(niov == 0 || kiov != NULL);
376 nob += (kiov++)->bv_len;
380 EXPORT_SYMBOL(lnet_kiov_nob);
383 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
384 unsigned int doffset,
385 unsigned int nsiov, struct bio_vec *siov,
386 unsigned int soffset,
389 /* NB diov, siov are READ-ONLY */
390 unsigned int this_nob;
397 LASSERT (!in_interrupt ());
400 while (doffset >= diov->bv_len) {
401 doffset -= diov->bv_len;
408 while (soffset >= siov->bv_len) {
409 soffset -= siov->bv_len;
418 this_nob = min3(diov->bv_len - doffset,
419 siov->bv_len - soffset,
423 daddr = ((char *)kmap(diov->bv_page)) +
424 diov->bv_offset + doffset;
426 saddr = ((char *)kmap(siov->bv_page)) +
427 siov->bv_offset + soffset;
429 /* Vanishing risk of kmap deadlock when mapping 2 pages.
430 * However in practice at least one of the kiovs will be mapped
431 * kernel pages and the map/unmap will be NOOPs */
433 memcpy (daddr, saddr, this_nob);
436 if (diov->bv_len > doffset + this_nob) {
440 kunmap(diov->bv_page);
447 if (siov->bv_len > soffset + this_nob) {
451 kunmap(siov->bv_page);
460 kunmap(diov->bv_page);
462 kunmap(siov->bv_page);
464 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
467 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
468 unsigned int nkiov, struct bio_vec *kiov,
469 unsigned int kiovoffset,
472 /* NB iov, kiov are READ-ONLY */
473 unsigned int this_nob;
479 LASSERT (!in_interrupt ());
482 while (iovoffset >= iov->iov_len) {
483 iovoffset -= iov->iov_len;
490 while (kiovoffset >= kiov->bv_len) {
491 kiovoffset -= kiov->bv_len;
500 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
501 (unsigned int)kiov->bv_len - kiovoffset,
505 addr = ((char *)kmap(kiov->bv_page)) +
506 kiov->bv_offset + kiovoffset;
508 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
511 if (iov->iov_len > iovoffset + this_nob) {
512 iovoffset += this_nob;
519 if (kiov->bv_len > kiovoffset + this_nob) {
521 kiovoffset += this_nob;
523 kunmap(kiov->bv_page);
533 kunmap(kiov->bv_page);
535 EXPORT_SYMBOL(lnet_copy_kiov2iov);
538 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
539 unsigned int kiovoffset,
540 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
543 /* NB kiov, iov are READ-ONLY */
544 unsigned int this_nob;
550 LASSERT (!in_interrupt ());
553 while (kiovoffset >= kiov->bv_len) {
554 kiovoffset -= kiov->bv_len;
561 while (iovoffset >= iov->iov_len) {
562 iovoffset -= iov->iov_len;
571 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
572 (unsigned int)iov->iov_len - iovoffset,
576 addr = ((char *)kmap(kiov->bv_page)) +
577 kiov->bv_offset + kiovoffset;
579 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
582 if (kiov->bv_len > kiovoffset + this_nob) {
584 kiovoffset += this_nob;
586 kunmap(kiov->bv_page);
593 if (iov->iov_len > iovoffset + this_nob) {
594 iovoffset += this_nob;
603 kunmap(kiov->bv_page);
605 EXPORT_SYMBOL(lnet_copy_iov2kiov);
608 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
609 int src_niov, struct bio_vec *src,
610 unsigned int offset, unsigned int len)
612 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
613 * for exactly 'len' bytes, and return the number of entries.
614 * NB not destructive to 'src' */
615 unsigned int frag_len;
618 if (len == 0) /* no data => */
619 return (0); /* no frags */
621 LASSERT(src_niov > 0);
622 while (offset >= src->bv_len) { /* skip initial frags */
623 offset -= src->bv_len;
626 LASSERT(src_niov > 0);
631 LASSERT(src_niov > 0);
632 LASSERT((int)niov <= dst_niov);
634 frag_len = src->bv_len - offset;
635 dst->bv_page = src->bv_page;
636 dst->bv_offset = src->bv_offset + offset;
638 if (len <= frag_len) {
640 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
644 dst->bv_len = frag_len;
645 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
655 EXPORT_SYMBOL(lnet_extract_kiov);
658 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
659 int delayed, unsigned int offset, unsigned int mlen,
662 unsigned int niov = 0;
663 struct kvec *iov = NULL;
664 struct bio_vec *kiov = NULL;
667 LASSERT (!in_interrupt ());
668 LASSERT (mlen == 0 || msg != NULL);
671 LASSERT(msg->msg_receiving);
672 LASSERT(!msg->msg_sending);
673 LASSERT(rlen == msg->msg_len);
674 LASSERT(mlen <= msg->msg_len);
675 LASSERT(msg->msg_offset == offset);
676 LASSERT(msg->msg_wanted == mlen);
678 msg->msg_receiving = 0;
681 niov = msg->msg_niov;
682 kiov = msg->msg_kiov;
685 LASSERT ((iov == NULL) != (kiov == NULL));
689 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
690 niov, kiov, offset, mlen,
693 lnet_finalize(msg, rc);
697 lnet_setpayloadbuffer(struct lnet_msg *msg)
699 struct lnet_libmd *md = msg->msg_md;
701 LASSERT(msg->msg_len > 0);
702 LASSERT(!msg->msg_routing);
704 LASSERT(msg->msg_niov == 0);
705 LASSERT(msg->msg_kiov == NULL);
707 msg->msg_niov = md->md_niov;
708 msg->msg_kiov = md->md_kiov;
712 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
713 unsigned int offset, unsigned int len)
715 msg->msg_type = type;
716 msg->msg_target.pid = target.pid;
717 lnet_nid4_to_nid(target.nid, &msg->msg_target.nid);
719 msg->msg_offset = offset;
722 lnet_setpayloadbuffer(msg);
724 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
725 msg->msg_hdr.type = cpu_to_le32(type);
726 /* dest_nid will be overwritten by lnet_select_pathway() */
727 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
728 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
729 /* src_nid will be set later */
730 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
731 msg->msg_hdr.payload_length = cpu_to_le32(len);
735 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
737 void *priv = msg->msg_private;
740 LASSERT(!in_interrupt());
741 LASSERT(nid_is_lo0(&ni->ni_nid) ||
742 (msg->msg_txcredit && msg->msg_peertxcredit));
744 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
746 msg->msg_no_resend = true;
747 lnet_finalize(msg, rc);
752 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
756 LASSERT(!msg->msg_sending);
757 LASSERT(msg->msg_receiving);
758 LASSERT(!msg->msg_rx_ready_delay);
759 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
761 msg->msg_rx_ready_delay = 1;
762 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
765 CERROR("recv from %s / send to %s aborted: "
766 "eager_recv failed %d\n",
767 libcfs_nidstr(&msg->msg_rxpeer->lpni_nid),
768 libcfs_idstr(&msg->msg_target), rc);
769 LASSERT(rc < 0); /* required by my callers */
776 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
780 deadline = lpni->lpni_last_alive +
781 lpni->lpni_net->net_tunables.lct_peer_timeout;
784 * assume peer_ni is alive as long as we're within the configured
793 /* NB: returns 1 when alive, 0 when dead, negative when error;
794 * may drop the lnet_net_lock */
796 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
797 struct lnet_msg *msg)
799 time64_t now = ktime_get_seconds();
801 if (!lnet_peer_aliveness_enabled(lpni))
805 * If we're resending a message, let's attempt to send it even if
806 * the peer is down to fulfill our resend quota on the message
808 if (msg->msg_retry_count > 0)
811 /* try and send recovery messages irregardless */
812 if (msg->msg_recovery)
815 /* always send any responses */
816 if (lnet_msg_is_response(msg))
819 if (!lnet_is_peer_deadline_passed(lpni, now))
822 return lnet_is_peer_ni_alive(lpni);
826 * \param msg The message to be sent.
827 * \param do_send True if lnet_ni_send() should be called in this function.
828 * lnet_send() is going to lnet_net_unlock immediately after this, so
829 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
831 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
832 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
833 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
834 * \retval -ECANCELED If the MD of the message has been unlinked.
837 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
839 struct lnet_peer_ni *lp = msg->msg_txpeer;
840 struct lnet_ni *ni = msg->msg_txni;
841 int cpt = msg->msg_tx_cpt;
842 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
844 /* non-lnet_send() callers have checked before */
845 LASSERT(!do_send || msg->msg_tx_delayed);
846 LASSERT(!msg->msg_receiving);
847 LASSERT(msg->msg_tx_committed);
849 /* can't get here if we're sending to the loopback interface */
850 if (the_lnet.ln_loni)
851 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
853 /* NB 'lp' is always the next hop */
854 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
855 lnet_peer_alive_locked(ni, lp, msg) == 0) {
856 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
857 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
859 lnet_net_unlock(cpt);
861 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
863 LNET_STATS_TYPE_DROP);
865 lnet_incr_stats(&msg->msg_txni->ni_stats,
867 LNET_STATS_TYPE_DROP);
869 CNETERR("Dropping message for %s: peer not alive\n",
870 libcfs_idstr(&msg->msg_target));
871 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
873 lnet_finalize(msg, -EHOSTUNREACH);
876 return -EHOSTUNREACH;
879 if (msg->msg_md != NULL &&
880 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
881 lnet_net_unlock(cpt);
883 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
884 "called on the MD/ME.\n",
885 libcfs_idstr(&msg->msg_target));
887 msg->msg_no_resend = true;
888 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
889 msg, libcfs_idstr(&msg->msg_target));
890 lnet_finalize(msg, -ECANCELED);
897 if (!msg->msg_peertxcredit) {
898 spin_lock(&lp->lpni_lock);
899 LASSERT((lp->lpni_txcredits < 0) ==
900 !list_empty(&lp->lpni_txq));
902 msg->msg_peertxcredit = 1;
903 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
904 lp->lpni_txcredits--;
906 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
907 lp->lpni_mintxcredits = lp->lpni_txcredits;
909 if (lp->lpni_txcredits < 0) {
910 msg->msg_tx_delayed = 1;
911 list_add_tail(&msg->msg_list, &lp->lpni_txq);
912 spin_unlock(&lp->lpni_lock);
913 return LNET_CREDIT_WAIT;
915 spin_unlock(&lp->lpni_lock);
918 if (!msg->msg_txcredit) {
919 LASSERT((tq->tq_credits < 0) ==
920 !list_empty(&tq->tq_delayed));
922 msg->msg_txcredit = 1;
924 atomic_dec(&ni->ni_tx_credits);
926 if (tq->tq_credits < tq->tq_credits_min)
927 tq->tq_credits_min = tq->tq_credits;
929 if (tq->tq_credits < 0) {
930 msg->msg_tx_delayed = 1;
931 list_add_tail(&msg->msg_list, &tq->tq_delayed);
932 return LNET_CREDIT_WAIT;
936 if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
937 lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
938 msg->msg_tx_delayed = 1;
939 return LNET_CREDIT_WAIT;
942 /* unset the tx_delay flag as we're going to send it now */
943 msg->msg_tx_delayed = 0;
946 lnet_net_unlock(cpt);
947 lnet_ni_send(ni, msg);
950 return LNET_CREDIT_OK;
954 static struct lnet_rtrbufpool *
955 lnet_msg2bufpool(struct lnet_msg *msg)
957 struct lnet_rtrbufpool *rbp;
960 LASSERT(msg->msg_rx_committed);
962 cpt = msg->msg_rx_cpt;
963 rbp = &the_lnet.ln_rtrpools[cpt][0];
965 LASSERT(msg->msg_len <= LNET_MTU);
966 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
968 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
975 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
977 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
978 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
979 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
980 * received or OK to receive */
981 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
982 struct lnet_peer *lp;
983 struct lnet_rtrbufpool *rbp;
984 struct lnet_rtrbuf *rb;
986 LASSERT(msg->msg_kiov == NULL);
987 LASSERT(msg->msg_niov == 0);
988 LASSERT(msg->msg_routing);
989 LASSERT(msg->msg_receiving);
990 LASSERT(!msg->msg_sending);
991 LASSERT(lpni->lpni_peer_net);
992 LASSERT(lpni->lpni_peer_net->lpn_peer);
994 lp = lpni->lpni_peer_net->lpn_peer;
996 /* non-lnet_parse callers only receive delayed messages */
997 LASSERT(!do_recv || msg->msg_rx_delayed);
999 if (!msg->msg_peerrtrcredit) {
1000 /* lpni_lock protects the credit manipulation */
1001 spin_lock(&lpni->lpni_lock);
1003 msg->msg_peerrtrcredit = 1;
1004 lpni->lpni_rtrcredits--;
1005 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1006 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1008 if (lpni->lpni_rtrcredits < 0) {
1009 spin_unlock(&lpni->lpni_lock);
1010 /* must have checked eager_recv before here */
1011 LASSERT(msg->msg_rx_ready_delay);
1012 msg->msg_rx_delayed = 1;
1013 /* lp_lock protects the lp_rtrq */
1014 spin_lock(&lp->lp_lock);
1015 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1016 spin_unlock(&lp->lp_lock);
1017 return LNET_CREDIT_WAIT;
1019 spin_unlock(&lpni->lpni_lock);
1022 rbp = lnet_msg2bufpool(msg);
1024 if (!msg->msg_rtrcredit) {
1025 msg->msg_rtrcredit = 1;
1027 if (rbp->rbp_credits < rbp->rbp_mincredits)
1028 rbp->rbp_mincredits = rbp->rbp_credits;
1030 if (rbp->rbp_credits < 0) {
1031 /* must have checked eager_recv before here */
1032 LASSERT(msg->msg_rx_ready_delay);
1033 msg->msg_rx_delayed = 1;
1034 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1035 return LNET_CREDIT_WAIT;
1039 LASSERT(!list_empty(&rbp->rbp_bufs));
1040 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1041 list_del(&rb->rb_list);
1043 msg->msg_niov = rbp->rbp_npages;
1044 msg->msg_kiov = &rb->rb_kiov[0];
1046 /* unset the msg-rx_delayed flag since we're receiving the message */
1047 msg->msg_rx_delayed = 0;
1050 int cpt = msg->msg_rx_cpt;
1052 lnet_net_unlock(cpt);
1053 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1054 0, msg->msg_len, msg->msg_len);
1057 return LNET_CREDIT_OK;
1061 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1063 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1064 struct lnet_ni *txni = msg->msg_txni;
1065 struct lnet_msg *msg2;
1067 if (msg->msg_txcredit) {
1068 struct lnet_ni *ni = msg->msg_txni;
1069 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1071 /* give back NI txcredits */
1072 msg->msg_txcredit = 0;
1074 LASSERT((tq->tq_credits < 0) ==
1075 !list_empty(&tq->tq_delayed));
1078 atomic_inc(&ni->ni_tx_credits);
1079 if (tq->tq_credits <= 0) {
1080 msg2 = list_entry(tq->tq_delayed.next,
1081 struct lnet_msg, msg_list);
1082 list_del(&msg2->msg_list);
1084 LASSERT(msg2->msg_txni == ni);
1085 LASSERT(msg2->msg_tx_delayed);
1086 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1088 (void) lnet_post_send_locked(msg2, 1);
1092 if (msg->msg_peertxcredit) {
1093 /* give back peer txcredits */
1094 msg->msg_peertxcredit = 0;
1096 spin_lock(&txpeer->lpni_lock);
1097 LASSERT((txpeer->lpni_txcredits < 0) ==
1098 !list_empty(&txpeer->lpni_txq));
1100 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1101 LASSERT(txpeer->lpni_txqnob >= 0);
1103 txpeer->lpni_txcredits++;
1104 if (txpeer->lpni_txcredits <= 0) {
1107 msg2 = list_entry(txpeer->lpni_txq.next,
1108 struct lnet_msg, msg_list);
1109 list_del(&msg2->msg_list);
1110 spin_unlock(&txpeer->lpni_lock);
1112 LASSERT(msg2->msg_txpeer == txpeer);
1113 LASSERT(msg2->msg_tx_delayed);
1115 msg2_cpt = msg2->msg_tx_cpt;
1118 * The msg_cpt can be different from the msg2_cpt
1119 * so we need to make sure we lock the correct cpt
1121 * Once we call lnet_post_send_locked() it is no
1122 * longer safe to access msg2, since it could've
1123 * been freed by lnet_finalize(), but we still
1124 * need to relock the correct cpt, so we cache the
1125 * msg2_cpt for the purpose of the check that
1126 * follows the call to lnet_pose_send_locked().
1128 if (msg2_cpt != msg->msg_tx_cpt) {
1129 lnet_net_unlock(msg->msg_tx_cpt);
1130 lnet_net_lock(msg2_cpt);
1132 (void) lnet_post_send_locked(msg2, 1);
1133 if (msg2_cpt != msg->msg_tx_cpt) {
1134 lnet_net_unlock(msg2_cpt);
1135 lnet_net_lock(msg->msg_tx_cpt);
1138 spin_unlock(&txpeer->lpni_lock);
1143 msg->msg_txni = NULL;
1144 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1147 if (txpeer != NULL) {
1148 msg->msg_txpeer = NULL;
1149 lnet_peer_ni_decref_locked(txpeer);
1154 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1156 struct lnet_msg *msg;
1158 if (list_empty(&rbp->rbp_msgs))
1160 msg = list_entry(rbp->rbp_msgs.next,
1161 struct lnet_msg, msg_list);
1162 list_del(&msg->msg_list);
1164 (void)lnet_post_routed_recv_locked(msg, 1);
1168 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1170 struct lnet_msg *msg;
1171 struct lnet_msg *tmp;
1173 lnet_net_unlock(cpt);
1175 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1176 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1177 0, 0, 0, msg->msg_hdr.payload_length);
1178 list_del_init(&msg->msg_list);
1179 msg->msg_no_resend = true;
1180 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1181 lnet_finalize(msg, -ECANCELED);
1188 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1190 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1191 struct lnet_peer *lp;
1192 struct lnet_ni *rxni = msg->msg_rxni;
1193 struct lnet_msg *msg2;
1195 if (msg->msg_rtrcredit) {
1196 /* give back global router credits */
1197 struct lnet_rtrbuf *rb;
1198 struct lnet_rtrbufpool *rbp;
1200 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1201 * there until it gets one allocated, or aborts the wait
1203 LASSERT(msg->msg_kiov != NULL);
1205 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1208 msg->msg_kiov = NULL;
1209 msg->msg_rtrcredit = 0;
1211 LASSERT(rbp == lnet_msg2bufpool(msg));
1213 LASSERT((rbp->rbp_credits > 0) ==
1214 !list_empty(&rbp->rbp_bufs));
1216 /* If routing is now turned off, we just drop this buffer and
1217 * don't bother trying to return credits. */
1218 if (!the_lnet.ln_routing) {
1219 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1223 /* It is possible that a user has lowered the desired number of
1224 * buffers in this pool. Make sure we never put back
1225 * more buffers than the stated number. */
1226 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1227 /* Discard this buffer so we don't have too
1229 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1230 rbp->rbp_nbuffers--;
1232 list_add(&rb->rb_list, &rbp->rbp_bufs);
1234 if (rbp->rbp_credits <= 0)
1235 lnet_schedule_blocked_locked(rbp);
1240 if (msg->msg_peerrtrcredit) {
1242 LASSERT(rxpeerni->lpni_peer_net);
1243 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1245 /* give back peer router credits */
1246 msg->msg_peerrtrcredit = 0;
1248 spin_lock(&rxpeerni->lpni_lock);
1249 rxpeerni->lpni_rtrcredits++;
1250 spin_unlock(&rxpeerni->lpni_lock);
1252 lp = rxpeerni->lpni_peer_net->lpn_peer;
1253 spin_lock(&lp->lp_lock);
1255 /* drop all messages which are queued to be routed on that
1257 if (!the_lnet.ln_routing) {
1259 list_splice_init(&lp->lp_rtrq, &drop);
1260 spin_unlock(&lp->lp_lock);
1261 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1262 } else if (!list_empty(&lp->lp_rtrq)) {
1265 msg2 = list_entry(lp->lp_rtrq.next,
1266 struct lnet_msg, msg_list);
1267 list_del(&msg2->msg_list);
1268 msg2_cpt = msg2->msg_rx_cpt;
1269 spin_unlock(&lp->lp_lock);
1271 * messages on the lp_rtrq can be from any NID in
1272 * the peer, which means they might have different
1273 * cpts. We need to make sure we lock the right
1276 if (msg2_cpt != msg->msg_rx_cpt) {
1277 lnet_net_unlock(msg->msg_rx_cpt);
1278 lnet_net_lock(msg2_cpt);
1280 (void) lnet_post_routed_recv_locked(msg2, 1);
1281 if (msg2_cpt != msg->msg_rx_cpt) {
1282 lnet_net_unlock(msg2_cpt);
1283 lnet_net_lock(msg->msg_rx_cpt);
1286 spin_unlock(&lp->lp_lock);
1290 msg->msg_rxni = NULL;
1291 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1293 if (rxpeerni != NULL) {
1294 msg->msg_rxpeer = NULL;
1295 lnet_peer_ni_decref_locked(rxpeerni);
1299 static struct lnet_peer_ni *
1300 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1301 struct lnet_peer *peer,
1302 struct lnet_peer_ni *best_lpni,
1303 struct lnet_peer_net *peer_net)
1306 * Look at the peer NIs for the destination peer that connect
1307 * to the chosen net. If a peer_ni is preferred when using the
1308 * best_ni to communicate, we use that one. If there is no
1309 * preferred peer_ni, or there are multiple preferred peer_ni,
1310 * the available transmit credits are used. If the transmit
1311 * credits are equal, we round-robin over the peer_ni.
1313 struct lnet_peer_ni *lpni = NULL;
1314 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1316 int best_lpni_healthv = (best_lpni) ?
1317 atomic_read(&best_lpni->lpni_healthv) : 0;
1318 bool best_lpni_is_preferred = false;
1319 bool lpni_is_preferred;
1321 __u32 lpni_sel_prio;
1322 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1324 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1326 * if the best_ni we've chosen aleady has this lpni
1327 * preferred, then let's use it
1330 lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1331 lpni, &best_ni->ni_nid);
1332 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1333 libcfs_nidstr(&best_ni->ni_nid),
1336 lpni_is_preferred = false;
1339 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1340 lpni_sel_prio = lpni->lpni_sel_priority;
1343 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1344 libcfs_nidstr(&lpni->lpni_nid),
1345 libcfs_nidstr(&best_lpni->lpni_nid),
1346 lpni_healthv, best_lpni_healthv,
1347 lpni_sel_prio, best_sel_prio,
1348 lpni->lpni_txcredits, best_lpni_credits,
1349 lpni->lpni_seq, best_lpni->lpni_seq);
1353 /* pick the healthiest peer ni */
1354 if (lpni_healthv < best_lpni_healthv)
1356 else if (lpni_healthv > best_lpni_healthv) {
1357 if (best_lpni_is_preferred)
1358 best_lpni_is_preferred = false;
1362 if (lpni_sel_prio > best_sel_prio)
1364 else if (lpni_sel_prio < best_sel_prio) {
1365 if (best_lpni_is_preferred)
1366 best_lpni_is_preferred = false;
1370 /* if this is a preferred peer use it */
1371 if (!best_lpni_is_preferred && lpni_is_preferred) {
1372 best_lpni_is_preferred = true;
1374 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1375 /* this is not the preferred peer so let's ignore
1381 if (lpni->lpni_txcredits < best_lpni_credits)
1382 /* We already have a peer that has more credits
1383 * available than this one. No need to consider
1384 * this peer further.
1387 else if (lpni->lpni_txcredits > best_lpni_credits)
1390 /* The best peer found so far and the current peer
1391 * have the same number of available credits let's
1392 * make sure to select between them using Round Robin
1394 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1397 best_lpni_is_preferred = lpni_is_preferred;
1398 best_lpni_healthv = lpni_healthv;
1399 best_sel_prio = lpni_sel_prio;
1401 best_lpni_credits = lpni->lpni_txcredits;
1404 /* if we still can't find a peer ni then we can't reach it */
1406 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1407 LNET_NIDNET(dst_nid);
1408 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1409 libcfs_net2str(net_id));
1413 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1414 libcfs_nidstr(&best_lpni->lpni_nid));
1420 * Prerequisite: the best_ni should already be set in the sd
1421 * Find the best lpni.
1422 * If the net id is provided then restrict lpni selection on
1423 * that particular net.
1424 * Otherwise find any reachable lpni. When dealing with an MR
1425 * gateway and it has multiple lpnis which we can use
1426 * we want to select the best one from the list of reachable
1429 static inline struct lnet_peer_ni *
1430 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1431 struct lnet_peer *peer, __u32 net_id)
1433 struct lnet_peer_net *peer_net;
1435 /* find the best_lpni on any local network */
1436 if (net_id == LNET_NET_ANY) {
1437 struct lnet_peer_ni *best_lpni = NULL;
1438 struct lnet_peer_net *lpn;
1439 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1440 /* no net specified find any reachable peer ni */
1441 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1443 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1449 /* restrict on the specified net */
1450 peer_net = lnet_peer_get_net_locked(peer, net_id);
1452 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1458 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1460 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1463 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1466 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1469 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1475 /* Compare route priorities and hop counts */
1477 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1479 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1480 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1482 if (r1->lr_priority < r2->lr_priority)
1485 if (r1->lr_priority > r2->lr_priority)
1488 if (r1_hops < r2_hops)
1491 if (r1_hops > r2_hops)
1497 static struct lnet_route *
1498 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1499 struct lnet_peer_ni *remote_lpni,
1500 struct lnet_route **prev_route,
1501 struct lnet_peer_ni **gwni)
1503 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1504 struct lnet_route *best_route;
1505 struct lnet_route *last_route;
1506 struct lnet_route *route;
1508 bool best_rte_is_preferred = false;
1509 struct lnet_nid *gw_pnid;
1511 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1512 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1514 best_route = last_route = NULL;
1515 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1516 if (!lnet_is_route_alive(route))
1518 gw_pnid = &route->lr_gateway->lp_primary_nid;
1520 /* no protection on below fields, but it's harmless */
1521 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1524 /* if the best route found is in the preferred list then
1525 * tag it as preferred and use it later on. But if we
1526 * didn't find any routes which are on the preferred list
1527 * then just use the best route possible.
1529 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1531 if (!best_route || (rc && !best_rte_is_preferred)) {
1532 /* Restrict the selection of the router NI on the
1533 * src_net provided. If the src_net is LNET_NID_ANY,
1534 * then select the best interface available.
1536 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1541 "Gateway %s does not have a peer NI on net %s\n",
1542 libcfs_nidstr(gw_pnid),
1543 libcfs_net2str(src_net));
1548 if (rc && !best_rte_is_preferred) {
1549 /* This is the first preferred route we found,
1550 * so it beats any route found previously
1556 best_rte_is_preferred = true;
1557 CDEBUG(D_NET, "preferred gw = %s\n",
1558 libcfs_nidstr(gw_pnid));
1560 } else if ((!rc) && best_rte_is_preferred)
1561 /* The best route we found so far is in the preferred
1562 * list, so it beats any non-preferred route
1567 best_route = last_route = route;
1572 rc = lnet_compare_routes(route, best_route);
1576 /* Restrict the selection of the router NI on the
1577 * src_net provided. If the src_net is LNET_NID_ANY,
1578 * then select the best interface available.
1580 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1585 "Gateway %s does not have a peer NI on net %s\n",
1586 libcfs_nidstr(gw_pnid),
1587 libcfs_net2str(src_net));
1597 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1601 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1608 *prev_route = last_route;
1614 static inline unsigned int
1615 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1617 if (dev_idx == UINT_MAX)
1620 if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1621 !ni->ni_net->net_lnd->lnd_get_dev_prio)
1624 return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1627 static struct lnet_ni *
1628 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1629 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1630 struct lnet_msg *msg, int md_cpt)
1632 struct lnet_libmd *md = msg->msg_md;
1633 unsigned int offset = msg->msg_offset;
1634 unsigned int shortest_distance;
1635 struct lnet_ni *ni = NULL;
1638 __u32 best_sel_prio;
1639 unsigned int best_dev_prio;
1640 unsigned int dev_idx = UINT_MAX;
1641 struct page *page = lnet_get_first_page(md, offset);
1642 msg->msg_rdma_force = lnet_is_rdma_only_page(page);
1644 if (msg->msg_rdma_force)
1645 dev_idx = lnet_get_dev_idx(page);
1648 * If there is no peer_ni that we can send to on this network,
1649 * then there is no point in looking for a new best_ni here.
1651 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1654 if (best_ni == NULL) {
1655 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1656 shortest_distance = UINT_MAX;
1657 best_dev_prio = UINT_MAX;
1658 best_credits = INT_MIN;
1661 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1662 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1663 best_ni->ni_dev_cpt);
1664 best_credits = atomic_read(&best_ni->ni_tx_credits);
1665 best_healthv = atomic_read(&best_ni->ni_healthv);
1666 best_sel_prio = best_ni->ni_sel_priority;
1669 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1670 unsigned int distance;
1675 unsigned int ni_dev_prio;
1677 ni_credits = atomic_read(&ni->ni_tx_credits);
1678 ni_healthv = atomic_read(&ni->ni_healthv);
1679 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1680 ni_sel_prio = ni->ni_sel_priority;
1683 * calculate the distance from the CPT on which
1684 * the message memory is allocated to the CPT of
1685 * the NI's physical device
1687 distance = cfs_cpt_distance(lnet_cpt_table(),
1691 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1694 * All distances smaller than the NUMA range
1695 * are treated equally.
1697 if (distance < lnet_numa_range)
1698 distance = lnet_numa_range;
1701 * Select on health, selection policy, direct dma prio,
1702 * shorter distance, available credits, then round-robin.
1708 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",
1709 libcfs_nidstr(&ni->ni_nid), ni_credits, distance,
1710 ni->ni_seq, ni_sel_prio, ni_dev_prio,
1711 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1712 : "not selected", best_credits, shortest_distance,
1713 (best_ni) ? best_ni->ni_seq : 0,
1714 best_sel_prio, best_dev_prio);
1718 if (ni_healthv < best_healthv)
1720 else if (ni_healthv > best_healthv)
1723 if (ni_sel_prio > best_sel_prio)
1725 else if (ni_sel_prio < best_sel_prio)
1728 if (ni_dev_prio > best_dev_prio)
1730 else if (ni_dev_prio < best_dev_prio)
1733 if (distance > shortest_distance)
1735 else if (distance < shortest_distance)
1738 if (ni_credits < best_credits)
1740 else if (ni_credits > best_credits)
1743 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1747 best_sel_prio = ni_sel_prio;
1748 best_dev_prio = ni_dev_prio;
1749 shortest_distance = distance;
1750 best_healthv = ni_healthv;
1752 best_credits = ni_credits;
1755 CDEBUG(D_NET, "selected best_ni %s\n",
1756 (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1762 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1763 * because such traffic is required to perform discovery. We therefore
1764 * exclude all GET and PUT on that portal. We also exclude all ACK and
1765 * REPLY traffic, but that is because the portal is not tracked in the
1766 * message structure for these message types. We could restrict this
1767 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1770 lnet_msg_discovery(struct lnet_msg *msg)
1772 if (msg->msg_type == LNET_MSG_PUT) {
1773 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1775 } else if (msg->msg_type == LNET_MSG_GET) {
1776 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1782 #define SRC_SPEC 0x0001
1783 #define SRC_ANY 0x0002
1784 #define LOCAL_DST 0x0004
1785 #define REMOTE_DST 0x0008
1786 #define MR_DST 0x0010
1787 #define NMR_DST 0x0020
1788 #define SND_RESP 0x0040
1790 /* The following to defines are used for return codes */
1791 #define REPEAT_SEND 0x1000
1792 #define PASS_THROUGH 0x2000
1794 /* The different cases lnet_select pathway needs to handle */
1795 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1796 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1797 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1798 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1799 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1800 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1801 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1802 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1805 lnet_handle_lo_send(struct lnet_send_data *sd)
1807 struct lnet_msg *msg = sd->sd_msg;
1808 int cpt = sd->sd_cpt;
1810 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1813 /* No send credit hassles with LOLND */
1814 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1815 msg->msg_hdr.dest_nid =
1816 cpu_to_le64(lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid));
1817 if (!msg->msg_routing)
1818 msg->msg_hdr.src_nid =
1819 cpu_to_le64(lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid));
1820 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1821 lnet_msg_commit(msg, cpt);
1822 msg->msg_txni = the_lnet.ln_loni;
1824 return LNET_CREDIT_OK;
1828 lnet_handle_send(struct lnet_send_data *sd)
1830 struct lnet_ni *best_ni = sd->sd_best_ni;
1831 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1832 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1833 struct lnet_msg *msg = sd->sd_msg;
1835 __u32 send_case = sd->sd_send_case;
1837 __u32 routing = send_case & REMOTE_DST;
1838 struct lnet_rsp_tracker *rspt;
1840 /* Increment sequence number of the selected peer, peer net,
1841 * local ni and local net so that we pick the next ones
1844 best_lpni->lpni_peer_net->lpn_seq++;
1845 best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1846 best_ni->ni_net->net_seq++;
1847 best_ni->ni_seq = best_ni->ni_net->net_seq;
1849 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1850 libcfs_nidstr(&best_ni->ni_nid),
1851 best_ni->ni_seq, best_ni->ni_net->net_seq,
1852 atomic_read(&best_ni->ni_tx_credits),
1853 best_ni->ni_sel_priority,
1854 libcfs_nidstr(&best_lpni->lpni_nid),
1855 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1856 best_lpni->lpni_txcredits,
1857 best_lpni->lpni_sel_priority);
1860 * grab a reference on the peer_ni so it sticks around even if
1861 * we need to drop and relock the lnet_net_lock below.
1863 lnet_peer_ni_addref_locked(best_lpni);
1866 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1867 * message. This ensures that we get a CPT that is correct for
1868 * the NI when the NI has been restricted to a subset of all CPTs.
1869 * If the selected CPT differs from the one currently locked, we
1870 * must unlock and relock the lnet_net_lock(), and then check whether
1871 * the configuration has changed. We don't have a hold on the best_ni
1872 * yet, and it may have vanished.
1874 cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1875 if (sd->sd_cpt != cpt2) {
1876 __u32 seq = lnet_get_dlc_seq_locked();
1877 lnet_net_unlock(sd->sd_cpt);
1879 lnet_net_lock(sd->sd_cpt);
1880 if (seq != lnet_get_dlc_seq_locked()) {
1881 lnet_peer_ni_decref_locked(best_lpni);
1887 * store the best_lpni in the message right away to avoid having
1888 * to do the same operation under different conditions
1890 msg->msg_txpeer = best_lpni;
1891 msg->msg_txni = best_ni;
1894 * grab a reference for the best_ni since now it's in use in this
1895 * send. The reference will be dropped in lnet_finalize()
1897 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1900 * Always set the target.nid to the best peer picked. Either the
1901 * NID will be one of the peer NIDs selected, or the same NID as
1902 * what was originally set in the target or it will be the NID of
1903 * a router if this message should be routed
1905 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1908 * lnet_msg_commit assigns the correct cpt to the message, which
1909 * is used to decrement the correct refcount on the ni when it's
1910 * time to return the credits
1912 lnet_msg_commit(msg, sd->sd_cpt);
1915 * If we are routing the message then we keep the src_nid that was
1916 * set by the originator. If we are not routing then we are the
1917 * originator and set it here.
1919 if (!msg->msg_routing)
1920 msg->msg_hdr.src_nid =
1921 cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txni->ni_nid));
1924 msg->msg_target_is_router = 1;
1925 msg->msg_target.pid = LNET_PID_LUSTRE;
1927 * since we're routing we want to ensure that the
1928 * msg_hdr.dest_nid is set to the final destination. When
1929 * the router receives this message it knows how to route
1932 * final_dst_lpni is set at the beginning of the
1933 * lnet_select_pathway() function and is never changed.
1934 * It's safe to use it here.
1936 /* FIXME handle large-addr nid */
1937 msg->msg_hdr.dest_nid =
1938 cpu_to_le64(lnet_nid_to_nid4(&final_dst_lpni->lpni_nid));
1941 * if we're not routing set the dest_nid to the best peer
1942 * ni NID that we picked earlier in the algorithm.
1944 msg->msg_hdr.dest_nid =
1945 cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txpeer->lpni_nid));
1949 * if we have response tracker block update it with the next hop
1953 rspt = msg->msg_md->md_rspt_ptr;
1955 rspt->rspt_next_hop_nid =
1956 msg->msg_txpeer->lpni_nid;
1957 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1958 libcfs_nidstr(&rspt->rspt_next_hop_nid));
1962 rc = lnet_post_send_locked(msg, 0);
1965 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1966 libcfs_nid2str(msg->msg_hdr.src_nid),
1967 libcfs_nidstr(&msg->msg_txni->ni_nid),
1968 libcfs_nidstr(&sd->sd_src_nid),
1969 libcfs_nid2str(msg->msg_hdr.dest_nid),
1970 libcfs_nidstr(&sd->sd_dst_nid),
1971 libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
1972 libcfs_nidstr(&sd->sd_rtr_nid),
1973 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1979 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1980 struct lnet_msg *msg)
1982 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1983 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1984 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1985 libcfs_nidstr(&lni->ni_nid),
1986 libcfs_nidstr(&lpni->lpni_nid));
1987 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
1996 * use the source and destination NIDs as the pathway
1999 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2001 /* the destination lpni is set before we get here. */
2004 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2005 if (!sd->sd_best_ni) {
2006 CERROR("Can't send to %s: src %s is not a local nid\n",
2007 libcfs_nidstr(&sd->sd_dst_nid),
2008 libcfs_nidstr(&sd->sd_src_nid));
2012 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2014 return lnet_handle_send(sd);
2022 * Don't run the selection algorithm on the peer NIs. By specifying the
2023 * local NID, we're also saying that we should always use the destination NID
2024 * provided. This handles the case where we should be using the same
2025 * destination NID for the all the messages which belong to the same RPC
2029 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2031 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2032 if (!sd->sd_best_ni) {
2033 CERROR("Can't send to %s: src %s is not a local nid\n",
2034 libcfs_nidstr(&sd->sd_dst_nid),
2035 libcfs_nidstr(&sd->sd_src_nid));
2039 if (sd->sd_best_lpni &&
2040 nid_same(&sd->sd_best_lpni->lpni_nid,
2041 &the_lnet.ln_loni->ni_nid))
2042 return lnet_handle_lo_send(sd);
2043 else if (sd->sd_best_lpni)
2044 return lnet_handle_send(sd);
2046 CERROR("can't send to %s. no NI on %s\n",
2047 libcfs_nidstr(&sd->sd_dst_nid),
2048 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2050 return -EHOSTUNREACH;
2054 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2055 struct lnet_peer *peer,
2056 struct lnet_peer_net *peer_net,
2057 struct lnet_msg *msg,
2060 struct lnet_net *local_net;
2061 struct lnet_ni *best_ni;
2063 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2068 * Iterate through the NIs in this local Net and select
2069 * the NI to send from. The selection is determined by
2070 * these 3 criterion in the following priority:
2072 * 2. NI available credits
2075 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2076 peer, peer_net, msg, cpt);
2082 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2085 struct lnet_peer *peer;
2086 struct lnet_peer_ni *new_lpni;
2089 lnet_peer_ni_addref_locked(lpni);
2091 peer = lpni->lpni_peer_net->lpn_peer;
2093 if (lnet_peer_gw_discovery(peer)) {
2094 lnet_peer_ni_decref_locked(lpni);
2098 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2099 lnet_peer_ni_decref_locked(lpni);
2103 rc = lnet_discover_peer_locked(lpni, cpt, false);
2105 lnet_peer_ni_decref_locked(lpni);
2109 new_lpni = lnet_find_peer_ni_locked(lnet_nid_to_nid4(&lpni->lpni_nid));
2111 lnet_peer_ni_decref_locked(lpni);
2115 peer = new_lpni->lpni_peer_net->lpn_peer;
2116 spin_lock(&peer->lp_lock);
2117 if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2118 /* The peer NI did not change and the peer is up to date.
2119 * Nothing more to do.
2121 spin_unlock(&peer->lp_lock);
2122 lnet_peer_ni_decref_locked(lpni);
2123 lnet_peer_ni_decref_locked(new_lpni);
2126 spin_unlock(&peer->lp_lock);
2128 /* Either the peer NI changed during discovery, or the peer isn't up
2129 * to date. In both cases we want to queue the message on the
2130 * (possibly new) peer's pending queue and queue the peer for discovery
2132 msg->msg_sending = 0;
2133 msg->msg_txpeer = NULL;
2134 lnet_net_unlock(cpt);
2135 lnet_peer_queue_message(peer, msg);
2138 lnet_peer_ni_decref_locked(lpni);
2139 lnet_peer_ni_decref_locked(new_lpni);
2141 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2142 msg, libcfs_nidstr(&peer->lp_primary_nid));
2144 return LNET_DC_WAIT;
2148 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2149 struct lnet_nid *dst_nid,
2150 struct lnet_peer_ni **gw_lpni,
2151 struct lnet_peer **gw_peer)
2154 struct lnet_peer *gw;
2155 struct lnet_peer *lp;
2156 struct lnet_peer_net *lpn;
2157 struct lnet_peer_net *best_lpn = NULL;
2158 struct lnet_remotenet *rnet, *best_rnet = NULL;
2159 struct lnet_route *best_route = NULL;
2160 struct lnet_route *last_route = NULL;
2161 struct lnet_peer_ni *lpni = NULL;
2162 struct lnet_peer_ni *gwni = NULL;
2163 bool route_found = false;
2164 struct lnet_nid *src_nid =
2165 !LNET_NID_IS_ANY(&sd->sd_src_nid) || !sd->sd_best_ni
2167 : &sd->sd_best_ni->ni_nid;
2168 int best_lpn_healthv = 0;
2169 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2171 CDEBUG(D_NET, "using src nid %s for route restriction\n",
2172 src_nid ? libcfs_nidstr(src_nid) : "ANY");
2174 /* If a router nid was specified then we are replying to a GET or
2175 * sending an ACK. In this case we use the gateway associated with the
2176 * specified router nid.
2178 if (!LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2179 gwni = lnet_peer_ni_find_locked(&sd->sd_rtr_nid);
2181 gw = gwni->lpni_peer_net->lpn_peer;
2182 lnet_peer_ni_decref_locked(gwni);
2183 if (gw->lp_rtr_refcount)
2186 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2187 libcfs_nidstr(&sd->sd_rtr_nid));
2192 if (sd->sd_msg->msg_routing || (src_nid && !LNET_NID_IS_ANY(src_nid))) {
2193 /* If I'm routing this message then I need to find the
2194 * next hop based on the destination NID
2196 * We also find next hop based on the destination NID
2197 * if the source NI was specified
2199 best_rnet = lnet_find_rnet_locked(LNET_NID_NET(&sd->sd_dst_nid));
2201 CERROR("Unable to send message from %s to %s - Route table may be misconfigured\n",
2202 (src_nid && LNET_NID_IS_ANY(src_nid)) ?
2204 libcfs_nidstr(src_nid),
2205 libcfs_nidstr(&sd->sd_dst_nid));
2206 return -EHOSTUNREACH;
2209 /* we've already looked up the initial lpni using
2212 lpni = sd->sd_best_lpni;
2213 /* the peer tree must be in existence */
2214 LASSERT(lpni && lpni->lpni_peer_net &&
2215 lpni->lpni_peer_net->lpn_peer);
2216 lp = lpni->lpni_peer_net->lpn_peer;
2218 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2219 /* is this remote network reachable? */
2220 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2229 /* select the preferred peer net */
2230 if (best_lpn_healthv > lpn->lpn_healthv)
2232 else if (best_lpn_healthv < lpn->lpn_healthv)
2235 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2237 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2240 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2243 best_lpn_healthv = lpn->lpn_healthv;
2244 best_lpn_sel_prio = lpn->lpn_sel_priority;
2250 CERROR("peer %s has no available nets\n",
2251 libcfs_nidstr(&sd->sd_dst_nid));
2252 return -EHOSTUNREACH;
2255 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2256 lnet_nid_to_nid4(&sd->sd_dst_nid),
2258 best_lpn->lpn_net_id);
2259 if (!sd->sd_best_lpni) {
2260 CERROR("peer %s is unreachable\n",
2261 libcfs_nidstr(&sd->sd_dst_nid));
2262 return -EHOSTUNREACH;
2265 /* We're attempting to round robin over the remote peer
2266 * NI's so update the final destination we selected
2268 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2270 /* Increment the sequence number of the remote lpni so
2271 * we can round robin over the different interfaces of
2274 sd->sd_best_lpni->lpni_seq++;
2278 * find the best route. Restrict the selection on the net of the
2279 * local NI if we've already picked the local NI to send from.
2280 * Otherwise, let's pick any route we can find and then find
2281 * a local NI we can reach the route's gateway on. Any route we
2282 * select will be reachable by virtue of the restriction we have
2283 * when adding a route.
2285 best_route = lnet_find_route_locked(best_rnet,
2286 LNET_NID_NET(src_nid),
2288 &last_route, &gwni);
2291 CERROR("no route to %s from %s\n",
2292 libcfs_nidstr(dst_nid),
2293 libcfs_nidstr(src_nid));
2294 return -EHOSTUNREACH;
2298 CERROR("Internal Error. Route expected to %s from %s\n",
2299 libcfs_nidstr(dst_nid),
2300 libcfs_nidstr(src_nid));
2304 gw = best_route->lr_gateway;
2305 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2309 * If the router checker is not active then discover the gateway here.
2310 * This ensures we are able to take advantage of multi-rail routing, but
2311 * if the router checker is active then we do not unecessarily delay
2312 * messages while the gateway is being checked by the dedicated monitor
2315 * NB: We're only checking the alive_router_check_interval here, rather
2316 * than calling lnet_router_checker_active(), because the other
2317 * conditions that are checked by that function are either
2318 * irrelevant (the_lnet.ln_routing) or must be true (list of routers
2321 if (alive_router_check_interval <= 0) {
2322 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2327 if (!sd->sd_best_ni) {
2328 lpn = gwni->lpni_peer_net;
2329 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2332 if (!sd->sd_best_ni) {
2333 CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2334 libcfs_net2str(lpn->lpn_net_id),
2335 libcfs_nidstr(&sd->sd_src_nid));
2344 * increment the sequence numbers since now we're sure we're
2345 * going to use this path
2347 if (LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2348 LASSERT(best_route && last_route);
2349 best_route->lr_seq = last_route->lr_seq + 1;
2351 best_lpn->lpn_seq++;
2362 * Remote destination
2363 * Non-MR destination
2367 * Remote destination
2370 * The handling of these two cases is similar. Even though the destination
2371 * can be MR or non-MR, we'll deal directly with the router.
2374 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2377 struct lnet_peer_ni *gw_lpni = NULL;
2378 struct lnet_peer *gw_peer = NULL;
2381 sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2382 if (!sd->sd_best_ni) {
2383 CERROR("Can't send to %s: src %s is not a local nid\n",
2384 libcfs_nidstr(&sd->sd_dst_nid),
2385 libcfs_nidstr(&sd->sd_src_nid));
2389 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2390 &gw_lpni, &gw_peer);
2394 if (sd->sd_send_case & NMR_DST)
2396 * since the final destination is non-MR let's set its preferred
2397 * NID before we send
2399 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2403 * We're going to send to the gw found so let's set its
2406 sd->sd_peer = gw_peer;
2407 sd->sd_best_lpni = gw_lpni;
2409 return lnet_handle_send(sd);
2413 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2414 struct lnet_msg *msg, bool discovery)
2416 struct lnet_peer_net *lpn = NULL;
2417 struct lnet_peer_net *best_lpn = NULL;
2418 struct lnet_net *net = NULL;
2419 struct lnet_net *best_net = NULL;
2420 struct lnet_ni *best_ni = NULL;
2421 int best_lpn_healthv = 0;
2422 int best_net_healthv = 0;
2424 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2426 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2431 * The peer can have multiple interfaces, some of them can be on
2432 * the local network and others on a routed network. We should
2433 * prefer the local network. However if the local network is not
2434 * available then we need to try the routed network
2437 /* go through all the peer nets and find the best_ni */
2438 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2440 * The peer's list of nets can contain non-local nets. We
2441 * want to only examine the local ones.
2443 net = lnet_get_net_locked(lpn->lpn_net_id);
2447 lpn_sel_prio = lpn->lpn_sel_priority;
2448 net_healthv = lnet_get_net_healthv_locked(net);
2449 net_sel_prio = net->net_sel_priority;
2452 * if this is a discovery message and lp_disc_net_id is
2453 * specified then use that net to send the discovery on.
2455 if (peer->lp_disc_net_id == lpn->lpn_net_id &&
2464 /* always select the lpn with the best health */
2465 if (best_lpn_healthv > lpn->lpn_healthv)
2467 else if (best_lpn_healthv < lpn->lpn_healthv)
2470 /* select the preferred peer and local nets */
2471 if (best_lpn_sel_prio < lpn_sel_prio)
2473 else if (best_lpn_sel_prio > lpn_sel_prio)
2476 if (best_net_healthv > net_healthv)
2478 else if (best_net_healthv < net_healthv)
2481 if (best_net_sel_prio < net_sel_prio)
2483 else if (best_net_sel_prio > net_sel_prio)
2486 if (best_lpn->lpn_seq < lpn->lpn_seq)
2488 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2491 /* round robin over the local networks */
2492 if (best_net->net_seq <= net->net_seq)
2496 best_net_healthv = net_healthv;
2497 best_net_sel_prio = net_sel_prio;
2498 best_lpn_healthv = lpn->lpn_healthv;
2499 best_lpn_sel_prio = lpn_sel_prio;
2508 /* Select the best NI on the same net as best_lpn chosen
2511 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2518 static struct lnet_ni *
2519 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2521 struct lnet_ni *best_ni = NULL;
2522 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2523 struct lnet_peer_ni *lpni_entry;
2526 * We must use a consistent source address when sending to a
2527 * non-MR peer. However, a non-MR peer can have multiple NIDs
2528 * on multiple networks, and we may even need to talk to this
2529 * peer on multiple networks -- certain types of
2530 * load-balancing configuration do this.
2532 * So we need to pick the NI the peer prefers for this
2533 * particular network.
2536 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2538 if (lpni_entry->lpni_pref_nnids == 0)
2540 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2541 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2549 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2551 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2553 struct lnet_ni *best_ni = NULL;
2554 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2557 * We must use a consistent source address when sending to a
2558 * non-MR peer. However, a non-MR peer can have multiple NIDs
2559 * on multiple networks, and we may even need to talk to this
2560 * peer on multiple networks -- certain types of
2561 * load-balancing configuration do this.
2563 * So we need to pick the NI the peer prefers for this
2564 * particular network.
2567 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2570 /* if best_ni is still not set just pick one */
2573 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2574 sd->sd_best_lpni->lpni_peer_net,
2577 /* If there is no best_ni we don't have a route */
2579 CERROR("no path to %s from net %s\n",
2580 libcfs_nidstr(&best_lpni->lpni_nid),
2581 libcfs_net2str(best_lpni->lpni_net->net_id));
2582 return -EHOSTUNREACH;
2586 sd->sd_best_ni = best_ni;
2588 /* Set preferred NI if necessary. */
2589 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2596 * Source not specified
2600 * always use the same source NID for NMR peers
2601 * If we've talked to that peer before then we already have a preferred
2602 * source NI associated with it. Otherwise, we select a preferred local NI
2603 * and store it in the peer
2606 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2610 /* sd->sd_best_lpni is already set to the final destination */
2613 * At this point we should've created the peer ni and peer. If we
2614 * can't find it, then something went wrong. Instead of assert
2615 * output a relevant message and fail the send
2617 if (!sd->sd_best_lpni) {
2618 CERROR("Internal fault. Unable to send msg %s to %s. NID not known\n",
2619 lnet_msgtyp2str(sd->sd_msg->msg_type),
2620 libcfs_nidstr(&sd->sd_dst_nid));
2624 if (sd->sd_msg->msg_routing) {
2625 /* If I'm forwarding this message then I can choose any NI
2626 * on the destination peer net
2628 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2630 sd->sd_best_lpni->lpni_peer_net,
2633 if (!sd->sd_best_ni) {
2634 CERROR("Unable to forward message to %s. No local NI available\n",
2635 libcfs_nidstr(&sd->sd_dst_nid));
2639 rc = lnet_select_preferred_best_ni(sd);
2642 rc = lnet_handle_send(sd);
2648 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2651 * NOTE we've already handled the remote peer case. So we only
2652 * need to worry about the local case here.
2654 * if we're sending a response, ACK or reply, we need to send it
2655 * to the destination NID given to us. At this point we already
2656 * have the peer_ni we're suppose to send to, so just find the
2657 * best_ni on the peer net and use that. Since we're sending to an
2658 * MR peer then we can just run the selection algorithm on our
2659 * local NIs and pick the best one.
2661 if (sd->sd_send_case & SND_RESP) {
2663 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2664 sd->sd_best_lpni->lpni_peer_net,
2668 if (!sd->sd_best_ni) {
2670 * We're not going to deal with not able to send
2671 * a response to the provided final destination
2673 CERROR("Can't send response to %s. No local NI available\n",
2674 libcfs_nidstr(&sd->sd_dst_nid));
2675 return -EHOSTUNREACH;
2678 return lnet_handle_send(sd);
2682 * If we get here that means we're sending a fresh request, PUT or
2683 * GET, so we need to run our standard selection algorithm.
2684 * First find the best local interface that's on any of the peer's
2687 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2690 lnet_msg_discovery(sd->sd_msg));
2691 if (sd->sd_best_ni) {
2693 lnet_find_best_lpni(sd->sd_best_ni,
2694 lnet_nid_to_nid4(&sd->sd_dst_nid),
2696 sd->sd_best_ni->ni_net->net_id);
2699 * if we're successful in selecting a peer_ni on the local
2700 * network, then send to it. Otherwise fall through and
2701 * try and see if we can reach it over another routed
2704 if (sd->sd_best_lpni &&
2705 nid_same(&sd->sd_best_lpni->lpni_nid,
2706 &the_lnet.ln_loni->ni_nid)) {
2708 * in case we initially started with a routed
2709 * destination, let's reset to local
2711 sd->sd_send_case &= ~REMOTE_DST;
2712 sd->sd_send_case |= LOCAL_DST;
2713 return lnet_handle_lo_send(sd);
2714 } else if (sd->sd_best_lpni) {
2716 * in case we initially started with a routed
2717 * destination, let's reset to local
2719 sd->sd_send_case &= ~REMOTE_DST;
2720 sd->sd_send_case |= LOCAL_DST;
2721 return lnet_handle_send(sd);
2724 CERROR("Internal Error. Expected to have a best_lpni: "
2726 libcfs_nidstr(&sd->sd_src_nid),
2727 libcfs_nidstr(&sd->sd_dst_nid));
2733 * Peer doesn't have a local network. Let's see if there is
2734 * a remote network we can reach it on.
2736 return PASS_THROUGH;
2741 * Source NID not specified
2746 * Source NID not speified
2747 * Remote destination
2750 * In both of these cases if we're sending a response, ACK or REPLY, then
2751 * we need to send to the destination NID provided.
2753 * In the remote case let's deal with MR routers.
2758 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2761 struct lnet_peer *gw_peer = NULL;
2762 struct lnet_peer_ni *gw_lpni = NULL;
2765 * handle sending a response to a remote peer here so we don't
2766 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2768 if (sd->sd_send_case & REMOTE_DST &&
2769 sd->sd_send_case & SND_RESP) {
2770 struct lnet_peer_ni *gw;
2771 struct lnet_peer *gw_peer;
2773 rc = lnet_handle_find_routed_path(
2774 sd, &sd->sd_dst_nid, &gw, &gw_peer);
2776 CERROR("Can't send response to %s. No route available\n",
2777 libcfs_nidstr(&sd->sd_dst_nid));
2778 return -EHOSTUNREACH;
2779 } else if (rc > 0) {
2783 sd->sd_best_lpni = gw;
2784 sd->sd_peer = gw_peer;
2786 return lnet_handle_send(sd);
2790 * Even though the NID for the peer might not be on a local network,
2791 * since the peer is MR there could be other interfaces on the
2792 * local network. In that case we'd still like to prefer the local
2793 * network over the routed network. If we're unable to do that
2794 * then we select the best router among the different routed networks,
2795 * and if the router is MR then we can deal with it as such.
2797 rc = lnet_handle_any_mr_dsta(sd);
2798 if (rc != PASS_THROUGH)
2802 * Now that we must route to the destination, we must consider the
2803 * MR case, where the destination has multiple interfaces, some of
2804 * which we can route to and others we do not. For this reason we
2805 * need to select the destination which we can route to and if
2806 * there are multiple, we need to round robin.
2808 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2809 &gw_lpni, &gw_peer);
2813 sd->sd_send_case &= ~LOCAL_DST;
2814 sd->sd_send_case |= REMOTE_DST;
2816 sd->sd_peer = gw_peer;
2817 sd->sd_best_lpni = gw_lpni;
2819 return lnet_handle_send(sd);
2823 * Source not specified
2824 * Remote destination
2827 * Must send to the specified peer NID using the same source NID that
2828 * we've used before. If it's the first time to talk to that peer then
2829 * find the source NI and assign it as preferred to that peer
2832 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2835 struct lnet_peer_ni *gw_lpni = NULL;
2836 struct lnet_peer *gw_peer = NULL;
2839 * Let's see if we have a preferred NI to talk to this NMR peer
2841 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2845 * find the router and that'll find the best NI if we didn't find
2848 rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid, &gw_lpni,
2854 * set the best_ni we've chosen as the preferred one for
2857 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2859 /* we'll be sending to the gw */
2860 sd->sd_best_lpni = gw_lpni;
2861 sd->sd_peer = gw_peer;
2863 return lnet_handle_send(sd);
2867 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2870 * turn off the SND_RESP bit.
2871 * It will be checked in the case handling
2873 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2875 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2876 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2877 (send_case & SRC_SPEC) ? libcfs_nidstr(&sd->sd_src_nid) : "",
2878 (send_case & MR_DST) ? "MR: " : "NMR: ",
2879 libcfs_nidstr(&sd->sd_dst_nid),
2880 (send_case & LOCAL_DST) ? "local" : "routed");
2882 switch (send_case) {
2884 * For all cases where the source is specified, we should always
2885 * use the destination NID, whether it's an MR destination or not,
2886 * since we're continuing a series of related messages for the
2889 case SRC_SPEC_LOCAL_NMR_DST:
2890 return lnet_handle_spec_local_nmr_dst(sd);
2891 case SRC_SPEC_LOCAL_MR_DST:
2892 return lnet_handle_spec_local_mr_dst(sd);
2893 case SRC_SPEC_ROUTER_NMR_DST:
2894 case SRC_SPEC_ROUTER_MR_DST:
2895 return lnet_handle_spec_router_dst(sd);
2896 case SRC_ANY_LOCAL_NMR_DST:
2897 return lnet_handle_any_local_nmr_dst(sd);
2898 case SRC_ANY_LOCAL_MR_DST:
2899 case SRC_ANY_ROUTER_MR_DST:
2900 return lnet_handle_any_mr_dst(sd);
2901 case SRC_ANY_ROUTER_NMR_DST:
2902 return lnet_handle_any_router_nmr_dst(sd);
2904 CERROR("Unknown send case\n");
2910 lnet_select_pathway(struct lnet_nid *src_nid,
2911 struct lnet_nid *dst_nid,
2912 struct lnet_msg *msg,
2913 struct lnet_nid *rtr_nid)
2915 struct lnet_peer_ni *lpni;
2916 struct lnet_peer *peer;
2917 struct lnet_send_data send_data;
2920 __u32 send_case = 0;
2922 bool mr_forwarding_allowed;
2924 memset(&send_data, 0, sizeof(send_data));
2927 * get an initial CPT to use for locking. The idea here is not to
2928 * serialize the calls to select_pathway, so that as many
2929 * operations can run concurrently as possible. To do that we use
2930 * the CPT where this call is being executed. Later on when we
2931 * determine the CPT to use in lnet_message_commit, we switch the
2932 * lock and check if there was any configuration change. If none,
2933 * then we proceed, if there is, then we restart the operation.
2935 cpt = lnet_net_lock_current();
2937 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2938 if (md_cpt == CFS_CPT_ANY)
2944 * If we're being asked to send to the loopback interface, there
2945 * is no need to go through any selection. We can just shortcut
2946 * the entire process and send over lolnd
2948 send_data.sd_msg = msg;
2949 send_data.sd_cpt = cpt;
2950 if (nid_is_lo0(dst_nid)) {
2951 rc = lnet_handle_lo_send(&send_data);
2952 lnet_net_unlock(cpt);
2957 * find an existing peer_ni, or create one and mark it as having been
2958 * created due to network traffic. This call will create the
2959 * peer->peer_net->peer_ni tree.
2961 lpni = lnet_peerni_by_nid_locked(dst_nid, NULL, cpt);
2963 lnet_net_unlock(cpt);
2964 return PTR_ERR(lpni);
2968 * Cache the original src_nid and rtr_nid. If we need to resend the
2969 * message then we'll need to know whether the src_nid was originally
2970 * specified for this message. If it was originally specified,
2971 * then we need to keep using the same src_nid since it's
2972 * continuing the same sequence of messages. Similarly, rtr_nid will
2973 * affect our choice of next hop.
2976 msg->msg_src_nid_param = *src_nid;
2978 msg->msg_src_nid_param = LNET_ANY_NID;
2980 msg->msg_rtr_nid_param = *rtr_nid;
2982 msg->msg_rtr_nid_param = LNET_ANY_NID;
2985 * If necessary, perform discovery on the peer that owns this peer_ni.
2986 * Note, this can result in the ownership of this peer_ni changing
2987 * to another peer object.
2989 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2991 lnet_peer_ni_decref_locked(lpni);
2992 lnet_net_unlock(cpt);
2995 lnet_peer_ni_decref_locked(lpni);
2997 peer = lpni->lpni_peer_net->lpn_peer;
3000 * Identify the different send cases
3002 if (!src_nid || LNET_NID_IS_ANY(src_nid)) {
3003 send_case |= SRC_ANY;
3004 if (lnet_get_net_locked(LNET_NID_NET(dst_nid)))
3005 send_case |= LOCAL_DST;
3007 send_case |= REMOTE_DST;
3009 send_case |= SRC_SPEC;
3010 if (LNET_NID_NET(src_nid) == LNET_NID_NET(dst_nid))
3011 send_case |= LOCAL_DST;
3013 send_case |= REMOTE_DST;
3017 if (msg->msg_routing && (send_case & LOCAL_DST))
3020 /* Determine whether to allow MR forwarding for this message.
3021 * NB: MR forwarding is allowed if the message originator and the
3022 * destination are both MR capable, and the destination lpni that was
3023 * originally chosen by the originator is unhealthy or down.
3024 * We check the MR capability of the destination further below
3026 mr_forwarding_allowed = false;
3028 struct lnet_peer *src_lp;
3029 struct lnet_peer_ni *src_lpni;
3031 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
3033 /* We don't fail the send if we hit any errors here. We'll just
3034 * try to send it via non-multi-rail criteria
3036 if (!IS_ERR(src_lpni)) {
3037 /* Drop ref taken by lnet_nid2peerni_locked() */
3038 lnet_peer_ni_decref_locked(src_lpni);
3039 src_lp = lpni->lpni_peer_net->lpn_peer;
3040 if (lnet_peer_is_multi_rail(src_lp) &&
3041 !lnet_is_peer_ni_alive(lpni))
3042 mr_forwarding_allowed = true;
3045 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3046 mr_forwarding_allowed ? "allowed" : "not allowed");
3050 * Deal with the peer as NMR in the following cases:
3051 * 1. the peer is NMR
3052 * 2. We're trying to recover a specific peer NI
3053 * 3. I'm a router sending to the final destination and MR forwarding is
3054 * not allowed for this message (as determined above).
3055 * In this case the source of the message would've
3056 * already selected the final destination so my job
3057 * is to honor the selection.
3059 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3060 (final_hop && !mr_forwarding_allowed))
3061 send_case |= NMR_DST;
3063 send_case |= MR_DST;
3065 if (lnet_msg_is_response(msg))
3066 send_case |= SND_RESP;
3068 /* assign parameters to the send_data */
3070 send_data.sd_rtr_nid = *rtr_nid;
3072 send_data.sd_rtr_nid = LNET_ANY_NID;
3074 send_data.sd_src_nid = *src_nid;
3076 send_data.sd_src_nid = LNET_ANY_NID;
3077 send_data.sd_dst_nid = *dst_nid;
3078 send_data.sd_best_lpni = lpni;
3080 * keep a pointer to the final destination in case we're going to
3081 * route, so we'll need to access it later
3083 send_data.sd_final_dst_lpni = lpni;
3084 send_data.sd_peer = peer;
3085 send_data.sd_md_cpt = md_cpt;
3086 send_data.sd_send_case = send_case;
3088 rc = lnet_handle_send_case_locked(&send_data);
3091 * Update the local cpt since send_data.sd_cpt might've been
3092 * updated as a result of calling lnet_handle_send_case_locked().
3094 cpt = send_data.sd_cpt;
3096 if (rc == REPEAT_SEND)
3099 lnet_net_unlock(cpt);
3105 lnet_send(struct lnet_nid *src_nid, struct lnet_msg *msg,
3106 struct lnet_nid *rtr_nid)
3108 struct lnet_nid *dst_nid = &msg->msg_target.nid;
3111 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3112 LASSERT(msg->msg_txpeer == NULL);
3113 LASSERT(msg->msg_txni == NULL);
3114 LASSERT(!msg->msg_sending);
3115 LASSERT(!msg->msg_target_is_router);
3116 LASSERT(!msg->msg_receiving);
3118 msg->msg_sending = 1;
3120 LASSERT(!msg->msg_tx_committed);
3122 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3124 if (rc == -EHOSTUNREACH)
3125 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3127 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3131 if (rc == LNET_CREDIT_OK)
3132 lnet_ni_send(msg->msg_txni, msg);
3134 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3138 enum lnet_mt_event_type {
3139 MT_TYPE_LOCAL_NI = 0,
3143 struct lnet_mt_event_info {
3144 enum lnet_mt_event_type mt_type;
3148 /* called with res_lock held */
3150 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3152 struct lnet_rsp_tracker *rspt;
3155 * msg has a refcount on the MD so the MD is not going away.
3156 * The rspt queue for the cpt is protected by
3157 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3159 if (!md->md_rspt_ptr)
3162 rspt = md->md_rspt_ptr;
3165 LASSERT(rspt->rspt_cpt == cpt);
3167 md->md_rspt_ptr = NULL;
3169 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3171 * The monitor thread has invalidated this handle because the
3172 * response timed out, but it failed to lookup the MD. That
3173 * means this response tracker is on the zombie list. We can
3174 * safely remove it under the resource lock (held by caller) and
3175 * free the response tracker block.
3177 list_del(&rspt->rspt_on_list);
3178 lnet_rspt_free(rspt, cpt);
3181 * invalidate the handle to indicate that a response has been
3182 * received, which will then lead the monitor thread to clean up
3185 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3190 lnet_clean_zombie_rstqs(void)
3192 struct lnet_rsp_tracker *rspt, *tmp;
3195 cfs_cpt_for_each(i, lnet_cpt_table()) {
3196 list_for_each_entry_safe(rspt, tmp,
3197 the_lnet.ln_mt_zombie_rstqs[i],
3199 list_del(&rspt->rspt_on_list);
3200 lnet_rspt_free(rspt, i);
3204 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3208 lnet_finalize_expired_responses(void)
3210 struct lnet_libmd *md;
3211 struct lnet_rsp_tracker *rspt, *tmp;
3215 if (the_lnet.ln_mt_rstq == NULL)
3218 cfs_cpt_for_each(i, lnet_cpt_table()) {
3219 LIST_HEAD(local_queue);
3222 if (!the_lnet.ln_mt_rstq[i]) {
3226 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3231 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3233 * The rspt mdh will be invalidated when a response
3234 * is received or whenever we want to discard the
3235 * block the monitor thread will walk the queue
3236 * and clean up any rsts with an invalid mdh.
3237 * The monitor thread will walk the queue until
3238 * the first unexpired rspt block. This means that
3239 * some rspt blocks which received their
3240 * corresponding responses will linger in the
3241 * queue until they are cleaned up eventually.
3244 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3246 list_del(&rspt->rspt_on_list);
3247 lnet_rspt_free(rspt, i);
3251 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3252 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3253 struct lnet_peer_ni *lpni;
3254 struct lnet_nid nid;
3256 md = lnet_handle2md(&rspt->rspt_mdh);
3258 /* MD has been queued for unlink, but
3259 * rspt hasn't been detached (Note we've
3260 * checked above that the rspt_mdh is
3261 * valid). Since we cannot lookup the MD
3262 * we're unable to detach the rspt
3263 * ourselves. Thus, move the rspt to the
3264 * zombie list where we'll wait for
3266 * 1. The remaining operations on the
3267 * MD to complete. In this case the
3268 * final operation will result in
3269 * lnet_msg_detach_md()->
3270 * lnet_detach_rsp_tracker() where
3271 * we will clean up this response
3273 * 2. LNet to shutdown. In this case
3274 * we'll wait until after all LND Nets
3275 * have shutdown and then we can
3276 * safely free any remaining response
3277 * tracker blocks on the zombie list.
3278 * Note: We need to hold the resource
3279 * lock when adding to the zombie list
3280 * because we may have concurrent access
3281 * with lnet_detach_rsp_tracker().
3283 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3284 list_move(&rspt->rspt_on_list,
3285 the_lnet.ln_mt_zombie_rstqs[i]);
3289 LASSERT(md->md_rspt_ptr == rspt);
3290 md->md_rspt_ptr = NULL;
3293 LNetMDUnlink(rspt->rspt_mdh);
3295 nid = rspt->rspt_next_hop_nid;
3297 list_del(&rspt->rspt_on_list);
3298 lnet_rspt_free(rspt, i);
3300 /* If we're shutting down we just want to clean
3301 * up the rspt blocks
3303 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3307 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3311 "Response timeout: md = %p: nid = %s\n",
3312 md, libcfs_nidstr(&nid));
3315 * If there is a timeout on the response
3316 * from the next hop decrement its health
3317 * value so that we don't use it
3320 lpni = lnet_peer_ni_find_locked(&nid);
3322 lnet_handle_remote_failure_locked(lpni);
3323 lnet_peer_ni_decref_locked(lpni);
3332 if (!list_empty(&local_queue)) {
3334 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3341 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3343 struct lnet_msg *msg;
3345 while (!list_empty(resendq)) {
3346 struct lnet_peer_ni *lpni;
3348 msg = list_entry(resendq->next, struct lnet_msg,
3351 list_del_init(&msg->msg_list);
3353 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3355 lnet_net_unlock(cpt);
3356 CERROR("Expected that a peer is already created for %s\n",
3357 libcfs_nid2str(msg->msg_hdr.dest_nid));
3358 msg->msg_no_resend = true;
3359 lnet_finalize(msg, -EFAULT);
3364 lnet_peer_ni_decref_locked(lpni);
3366 lnet_net_unlock(cpt);
3367 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3368 libcfs_nidstr(&msg->msg_src_nid_param),
3369 libcfs_idstr(&msg->msg_target),
3370 lnet_msgtyp2str(msg->msg_type),
3372 msg->msg_retry_count);
3373 rc = lnet_send(&msg->msg_src_nid_param, msg,
3374 &msg->msg_rtr_nid_param);
3376 CERROR("Error sending %s to %s: %d\n",
3377 lnet_msgtyp2str(msg->msg_type),
3378 libcfs_idstr(&msg->msg_target), rc);
3379 msg->msg_no_resend = true;
3380 lnet_finalize(msg, rc);
3384 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3390 lnet_resend_pending_msgs(void)
3394 cfs_cpt_for_each(i, lnet_cpt_table()) {
3396 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3401 /* called with cpt and ni_lock held */
3403 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3405 struct lnet_handle_md recovery_mdh;
3407 LNetInvalidateMDHandle(&recovery_mdh);
3409 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3411 recovery_mdh = ni->ni_ping_mdh;
3412 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3415 lnet_net_unlock(cpt);
3416 if (!LNetMDHandleIsInvalid(recovery_mdh))
3417 LNetMDUnlink(recovery_mdh);
3423 lnet_recover_local_nis(void)
3425 struct lnet_mt_event_info *ev_info;
3426 LIST_HEAD(processed_list);
3427 LIST_HEAD(local_queue);
3428 struct lnet_handle_md mdh;
3429 struct lnet_ni *tmp;
3437 * splice the recovery queue on a local queue. We will iterate
3438 * through the local queue and update it as needed. Once we're
3439 * done with the traversal, we'll splice the local queue back on
3440 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3441 * will be traversed in the next iteration.
3444 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3448 now = ktime_get_seconds();
3450 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3452 * if an NI is being deleted or it is now healthy, there
3453 * is no need to keep it around in the recovery queue.
3454 * The monitor thread is the only thread responsible for
3455 * removing the NI from the recovery queue.
3456 * Multiple threads can be adding NIs to the recovery
3459 healthv = atomic_read(&ni->ni_healthv);
3463 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3464 healthv == LNET_MAX_HEALTH_VALUE) {
3465 list_del_init(&ni->ni_recovery);
3466 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3468 lnet_ni_decref_locked(ni, 0);
3474 * if the local NI failed recovery we must unlink the md.
3475 * But we want to keep the local_ni on the recovery queue
3476 * so we can continue the attempts to recover it.
3478 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3479 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3480 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3486 if (now < ni->ni_next_ping) {
3493 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3494 libcfs_nidstr(&ni->ni_nid));
3497 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3498 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3501 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3503 CERROR("out of memory. Can't recover %s\n",
3504 libcfs_nidstr(&ni->ni_nid));
3506 ni->ni_recovery_state &=
3507 ~LNET_NI_RECOVERY_PENDING;
3512 mdh = ni->ni_ping_mdh;
3514 * Invalidate the ni mdh in case it's deleted.
3515 * We'll unlink the mdh in this case below.
3517 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3518 /* FIXME need to handle large-addr nid */
3519 nid = lnet_nid_to_nid4(&ni->ni_nid);
3522 * remove the NI from the local queue and drop the
3523 * reference count to it while we're recovering
3524 * it. The reason for that, is that the NI could
3525 * be deleted, and the way the code is structured
3526 * is if we don't drop the NI, then the deletion
3527 * code will enter a loop waiting for the
3528 * reference count to be removed while holding the
3529 * ln_mutex_lock(). When we look up the peer to
3530 * send to in lnet_select_pathway() we will try to
3531 * lock the ln_mutex_lock() as well, leading to
3532 * a deadlock. By dropping the refcount and
3533 * removing it from the list, we allow for the NI
3534 * to be removed, then we use the cached NID to
3535 * look it up again. If it's gone, then we just
3536 * continue examining the rest of the queue.
3539 list_del_init(&ni->ni_recovery);
3540 lnet_ni_decref_locked(ni, 0);
3543 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3544 ev_info->mt_nid = nid;
3545 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3546 ev_info, the_lnet.ln_mt_handler,
3548 /* lookup the nid again */
3550 ni = lnet_nid2ni_locked(nid, 0);
3553 * the NI has been deleted when we dropped
3560 ni->ni_ping_count++;
3562 ni->ni_ping_mdh = mdh;
3563 lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3568 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3577 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3578 * reexamined in the next iteration.
3580 list_splice_init(&processed_list, &local_queue);
3582 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3587 lnet_resendqs_create(void)
3589 struct list_head **resendqs;
3590 resendqs = lnet_create_array_of_queues();
3595 lnet_net_lock(LNET_LOCK_EX);
3596 the_lnet.ln_mt_resendqs = resendqs;
3597 lnet_net_unlock(LNET_LOCK_EX);
3603 lnet_clean_local_ni_recoveryq(void)
3607 /* This is only called when the monitor thread has stopped */
3610 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3611 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3612 struct lnet_ni, ni_recovery);
3613 list_del_init(&ni->ni_recovery);
3615 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3617 lnet_ni_decref_locked(ni, 0);
3624 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3627 struct lnet_handle_md recovery_mdh;
3629 LNetInvalidateMDHandle(&recovery_mdh);
3631 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3632 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3633 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3635 spin_unlock(&lpni->lpni_lock);
3636 lnet_net_unlock(cpt);
3637 if (!LNetMDHandleIsInvalid(recovery_mdh))
3638 LNetMDUnlink(recovery_mdh);
3640 spin_lock(&lpni->lpni_lock);
3644 lnet_clean_peer_ni_recoveryq(void)
3646 struct lnet_peer_ni *lpni, *tmp;
3648 lnet_net_lock(LNET_LOCK_EX);
3650 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3652 list_del_init(&lpni->lpni_recovery);
3653 spin_lock(&lpni->lpni_lock);
3654 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3655 spin_unlock(&lpni->lpni_lock);
3656 lnet_peer_ni_decref_locked(lpni);
3659 lnet_net_unlock(LNET_LOCK_EX);
3663 lnet_clean_resendqs(void)
3665 struct lnet_msg *msg, *tmp;
3669 cfs_cpt_for_each(i, lnet_cpt_table()) {
3671 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3673 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3674 list_del_init(&msg->msg_list);
3675 msg->msg_no_resend = true;
3676 lnet_finalize(msg, -ESHUTDOWN);
3680 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3684 lnet_recover_peer_nis(void)
3686 struct lnet_mt_event_info *ev_info;
3687 LIST_HEAD(processed_list);
3688 LIST_HEAD(local_queue);
3689 struct lnet_handle_md mdh;
3690 struct lnet_peer_ni *lpni;
3691 struct lnet_peer_ni *tmp;
3698 * Always use cpt 0 for locking across all interactions with
3699 * ln_mt_peerNIRecovq
3702 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3706 now = ktime_get_seconds();
3708 list_for_each_entry_safe(lpni, tmp, &local_queue,
3711 * The same protection strategy is used here as is in the
3712 * local recovery case.
3715 healthv = atomic_read(&lpni->lpni_healthv);
3716 spin_lock(&lpni->lpni_lock);
3717 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3718 healthv == LNET_MAX_HEALTH_VALUE) {
3719 list_del_init(&lpni->lpni_recovery);
3720 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3721 spin_unlock(&lpni->lpni_lock);
3722 lnet_peer_ni_decref_locked(lpni);
3728 * If the peer NI has failed recovery we must unlink the
3729 * md. But we want to keep the peer ni on the recovery
3730 * queue so we can try to continue recovering it
3732 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3733 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3734 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3737 spin_unlock(&lpni->lpni_lock);
3739 if (now < lpni->lpni_next_ping) {
3747 * NOTE: we're racing with peer deletion from user space.
3748 * It's possible that a peer is deleted after we check its
3749 * state. In this case the recovery can create a new peer
3751 spin_lock(&lpni->lpni_lock);
3752 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3753 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3754 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3755 spin_unlock(&lpni->lpni_lock);
3757 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3759 CERROR("out of memory. Can't recover %s\n",
3760 libcfs_nidstr(&lpni->lpni_nid));
3761 spin_lock(&lpni->lpni_lock);
3762 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3763 spin_unlock(&lpni->lpni_lock);
3767 /* look at the comments in lnet_recover_local_nis() */
3768 mdh = lpni->lpni_recovery_ping_mdh;
3769 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3770 /* FIXME handle large-addr nid */
3771 nid = lnet_nid_to_nid4(&lpni->lpni_nid);
3773 list_del_init(&lpni->lpni_recovery);
3774 lnet_peer_ni_decref_locked(lpni);
3777 ev_info->mt_type = MT_TYPE_PEER_NI;
3778 ev_info->mt_nid = nid;
3779 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3780 ev_info, the_lnet.ln_mt_handler,
3784 * lnet_find_peer_ni_locked() grabs a refcount for
3785 * us. No need to take it explicitly.
3787 lpni = lnet_find_peer_ni_locked(nid);
3794 lpni->lpni_ping_count++;
3796 lpni->lpni_recovery_ping_mdh = mdh;
3798 lnet_peer_ni_add_to_recoveryq_locked(lpni,
3802 spin_lock(&lpni->lpni_lock);
3803 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3804 spin_unlock(&lpni->lpni_lock);
3807 /* Drop the ref taken by lnet_find_peer_ni_locked() */
3808 lnet_peer_ni_decref_locked(lpni);
3811 spin_unlock(&lpni->lpni_lock);
3814 list_splice_init(&processed_list, &local_queue);
3816 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3821 lnet_monitor_thread(void *arg)
3823 time64_t rsp_timeout = 0;
3826 wait_for_completion(&the_lnet.ln_started);
3828 * The monitor thread takes care of the following:
3829 * 1. Checks the aliveness of routers
3830 * 2. Checks if there are messages on the resend queue to resend
3832 * 3. Check if there are any NIs on the local recovery queue and
3834 * 4. Checks if there are any NIs on the remote recovery queue
3837 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3838 now = ktime_get_real_seconds();
3840 if (lnet_router_checker_active())
3841 lnet_check_routers();
3843 lnet_resend_pending_msgs();
3845 if (now >= rsp_timeout) {
3846 lnet_finalize_expired_responses();
3847 rsp_timeout = now + (lnet_transaction_timeout / 2);
3850 lnet_recover_local_nis();
3851 lnet_recover_peer_nis();
3854 * TODO do we need to check if we should sleep without
3855 * timeout? Technically, an active system will always
3856 * have messages in flight so this check will always
3857 * evaluate to false. And on an idle system do we care
3858 * if we wake up every 1 second? Although, we've seen
3859 * cases where we get a complaint that an idle thread
3860 * is waking up unnecessarily.
3862 wait_for_completion_interruptible_timeout(
3863 &the_lnet.ln_mt_wait_complete,
3864 cfs_time_seconds(1));
3865 /* Must re-init the completion before testing anything,
3866 * including ln_mt_state.
3868 reinit_completion(&the_lnet.ln_mt_wait_complete);
3872 lnet_net_lock(LNET_LOCK_EX);
3873 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3874 lnet_net_unlock(LNET_LOCK_EX);
3876 /* signal that the monitor thread is exiting */
3877 up(&the_lnet.ln_mt_signal);
3885 * Returns == 0 if success
3886 * Returns > 0 if LNetMDBind or prior fails
3887 * Returns < 0 if LNetGet fails
3890 lnet_send_ping(lnet_nid_t dest_nid,
3891 struct lnet_handle_md *mdh, int nnis,
3892 void *user_data, lnet_handler_t handler, bool recovery)
3894 struct lnet_md md = { NULL };
3895 struct lnet_process_id id;
3896 struct lnet_ping_buffer *pbuf;
3899 if (dest_nid == LNET_NID_ANY) {
3904 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3910 /* initialize md content */
3911 md.start = &pbuf->pb_info;
3912 md.length = LNET_PING_INFO_SIZE(nnis);
3913 md.threshold = 2; /* GET/REPLY */
3915 md.options = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3916 md.user_ptr = user_data;
3917 md.handler = handler;
3919 rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3921 lnet_ping_buffer_decref(pbuf);
3922 CERROR("Can't bind MD: %d\n", rc);
3923 rc = -rc; /* change the rc to positive */
3926 id.pid = LNET_PID_LUSTRE;
3929 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3930 LNET_RESERVED_PORTAL,
3931 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3934 goto fail_unlink_md;
3940 LNetInvalidateMDHandle(mdh);
3946 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3947 int status, bool send, bool unlink_event)
3949 lnet_nid_t nid = ev_info->mt_nid;
3951 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3955 ni = lnet_nid2ni_locked(nid, 0);
3961 if (!send || (send && status != 0))
3962 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3964 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3969 CERROR("local NI (%s) recovery failed with %d\n",
3970 libcfs_nid2str(nid), status);
3974 * need to increment healthv for the ni here, because in
3975 * the lnet_finalize() path we don't have access to this
3976 * NI. And in order to get access to it, we'll need to
3977 * carry forward too much information.
3978 * In the peer case, it'll naturally be incremented
3981 lnet_inc_healthv(&ni->ni_healthv,
3982 lnet_health_sensitivity);
3984 struct lnet_peer_ni *lpni;
3987 cpt = lnet_net_lock_current();
3988 lpni = lnet_find_peer_ni_locked(nid);
3990 lnet_net_unlock(cpt);
3993 spin_lock(&lpni->lpni_lock);
3994 if (!send || (send && status != 0))
3995 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3997 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3998 spin_unlock(&lpni->lpni_lock);
3999 lnet_peer_ni_decref_locked(lpni);
4000 lnet_net_unlock(cpt);
4003 CERROR("peer NI (%s) recovery failed with %d\n",
4004 libcfs_nid2str(nid), status);
4009 lnet_mt_event_handler(struct lnet_event *event)
4011 struct lnet_mt_event_info *ev_info = event->md_user_ptr;
4012 struct lnet_ping_buffer *pbuf;
4014 /* TODO: remove assert */
4015 LASSERT(event->type == LNET_EVENT_REPLY ||
4016 event->type == LNET_EVENT_SEND ||
4017 event->type == LNET_EVENT_UNLINK);
4019 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
4022 switch (event->type) {
4023 case LNET_EVENT_UNLINK:
4024 CDEBUG(D_NET, "%s recovery ping unlinked\n",
4025 libcfs_nid2str(ev_info->mt_nid));
4027 case LNET_EVENT_REPLY:
4028 lnet_handle_recovery_reply(ev_info, event->status, false,
4029 event->type == LNET_EVENT_UNLINK);
4031 case LNET_EVENT_SEND:
4032 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4033 libcfs_nid2str(ev_info->mt_nid),
4034 (event->status) ? "unsuccessfully" :
4035 "successfully", event->status);
4036 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4039 CERROR("Unexpected event: %d\n", event->type);
4042 if (event->unlinked) {
4043 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4044 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4045 lnet_ping_buffer_decref(pbuf);
4050 lnet_rsp_tracker_create(void)
4052 struct list_head **rstqs;
4053 rstqs = lnet_create_array_of_queues();
4058 the_lnet.ln_mt_rstq = rstqs;
4064 lnet_rsp_tracker_clean(void)
4066 lnet_finalize_expired_responses();
4068 cfs_percpt_free(the_lnet.ln_mt_rstq);
4069 the_lnet.ln_mt_rstq = NULL;
4072 int lnet_monitor_thr_start(void)
4075 struct task_struct *task;
4077 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4080 rc = lnet_resendqs_create();
4084 rc = lnet_rsp_tracker_create();
4088 sema_init(&the_lnet.ln_mt_signal, 0);
4090 lnet_net_lock(LNET_LOCK_EX);
4091 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4092 lnet_net_unlock(LNET_LOCK_EX);
4093 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4096 CERROR("Can't start monitor thread: %d\n", rc);
4103 lnet_net_lock(LNET_LOCK_EX);
4104 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4105 lnet_net_unlock(LNET_LOCK_EX);
4106 /* block until event callback signals exit */
4107 down(&the_lnet.ln_mt_signal);
4109 lnet_net_lock(LNET_LOCK_EX);
4110 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4111 lnet_net_unlock(LNET_LOCK_EX);
4112 lnet_rsp_tracker_clean();
4113 lnet_clean_local_ni_recoveryq();
4114 lnet_clean_peer_ni_recoveryq();
4115 lnet_clean_resendqs();
4116 the_lnet.ln_mt_handler = NULL;
4119 lnet_rsp_tracker_clean();
4120 lnet_clean_local_ni_recoveryq();
4121 lnet_clean_peer_ni_recoveryq();
4122 lnet_clean_resendqs();
4126 void lnet_monitor_thr_stop(void)
4128 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4131 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4132 lnet_net_lock(LNET_LOCK_EX);
4133 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4134 lnet_net_unlock(LNET_LOCK_EX);
4136 /* tell the monitor thread that we're shutting down */
4137 complete(&the_lnet.ln_mt_wait_complete);
4139 /* block until monitor thread signals that it's done */
4140 down(&the_lnet.ln_mt_signal);
4141 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4143 /* perform cleanup tasks */
4144 lnet_rsp_tracker_clean();
4145 lnet_clean_local_ni_recoveryq();
4146 lnet_clean_peer_ni_recoveryq();
4147 lnet_clean_resendqs();
4151 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4155 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4156 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4157 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4158 lnet_net_unlock(cpt);
4160 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4164 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4166 struct lnet_hdr *hdr = &msg->msg_hdr;
4168 if (msg->msg_wanted != 0)
4169 lnet_setpayloadbuffer(msg);
4171 lnet_build_msg_event(msg, LNET_EVENT_PUT);
4173 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
4174 * it back into the ACK during lnet_finalize() */
4175 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4176 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4178 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4179 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4183 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4185 struct lnet_hdr *hdr = &msg->msg_hdr;
4186 struct lnet_match_info info;
4190 /* Convert put fields to host byte order */
4191 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4192 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
4193 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
4195 /* Primary peer NID. */
4196 info.mi_id.nid = msg->msg_initiator;
4197 info.mi_id.pid = hdr->src_pid;
4198 info.mi_opc = LNET_MD_OP_PUT;
4199 info.mi_portal = hdr->msg.put.ptl_index;
4200 info.mi_rlength = hdr->payload_length;
4201 info.mi_roffset = hdr->msg.put.offset;
4202 info.mi_mbits = hdr->msg.put.match_bits;
4203 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4205 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4206 ready_delay = msg->msg_rx_ready_delay;
4209 rc = lnet_ptl_match_md(&info, msg);
4214 case LNET_MATCHMD_OK:
4215 lnet_recv_put(ni, msg);
4218 case LNET_MATCHMD_NONE:
4220 /* no eager_recv or has already called it, should
4221 * have been attached on delayed list */
4224 rc = lnet_ni_eager_recv(ni, msg);
4231 case LNET_MATCHMD_DROP:
4232 CNETERR("Dropping PUT from %s portal %d match %llu"
4233 " offset %d length %d: %d\n",
4234 libcfs_idstr(&info.mi_id), info.mi_portal,
4235 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4237 return -ENOENT; /* -ve: OK but no match */
4242 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4244 struct lnet_match_info info;
4245 struct lnet_hdr *hdr = &msg->msg_hdr;
4246 struct lnet_process_id source_id;
4247 struct lnet_handle_wire reply_wmd;
4250 /* Convert get fields to host byte order */
4251 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
4252 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
4253 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
4254 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
4256 source_id.nid = hdr->src_nid;
4257 source_id.pid = hdr->src_pid;
4258 /* Primary peer NID */
4259 info.mi_id.nid = msg->msg_initiator;
4260 info.mi_id.pid = hdr->src_pid;
4261 info.mi_opc = LNET_MD_OP_GET;
4262 info.mi_portal = hdr->msg.get.ptl_index;
4263 info.mi_rlength = hdr->msg.get.sink_length;
4264 info.mi_roffset = hdr->msg.get.src_offset;
4265 info.mi_mbits = hdr->msg.get.match_bits;
4266 info.mi_cpt = lnet_nid2cpt(&msg->msg_initiator, ni);
4268 rc = lnet_ptl_match_md(&info, msg);
4269 if (rc == LNET_MATCHMD_DROP) {
4270 CNETERR("Dropping GET from %s portal %d match %llu"
4271 " offset %d length %d\n",
4272 libcfs_idstr(&info.mi_id), info.mi_portal,
4273 info.mi_mbits, info.mi_roffset, info.mi_rlength);
4274 return -ENOENT; /* -ve: OK but no match */
4277 LASSERT(rc == LNET_MATCHMD_OK);
4279 lnet_build_msg_event(msg, LNET_EVENT_GET);
4281 reply_wmd = hdr->msg.get.return_wmd;
4283 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
4284 msg->msg_offset, msg->msg_wanted);
4286 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4289 /* The LND completes the REPLY from her recv procedure */
4290 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4291 msg->msg_offset, msg->msg_len, msg->msg_len);
4295 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4296 msg->msg_receiving = 0;
4298 rc = lnet_send(&ni->ni_nid, msg, &msg->msg_from);
4300 /* didn't get as far as lnet_ni_send() */
4301 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4302 libcfs_nidstr(&ni->ni_nid),
4303 libcfs_idstr(&info.mi_id), rc);
4305 lnet_finalize(msg, rc);
4312 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4314 void *private = msg->msg_private;
4315 struct lnet_hdr *hdr = &msg->msg_hdr;
4316 struct lnet_process_id src = {0};
4317 struct lnet_libmd *md;
4318 unsigned int rlength;
4319 unsigned int mlength;
4322 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4325 src.nid = hdr->src_nid;
4326 src.pid = hdr->src_pid;
4328 /* NB handles only looked up by creator (no flips) */
4329 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4330 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4331 CNETERR("%s: Dropping REPLY from %s for %s "
4333 libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4334 (md == NULL) ? "invalid" : "inactive",
4335 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4336 hdr->msg.reply.dst_wmd.wh_object_cookie);
4337 if (md != NULL && md->md_me != NULL)
4338 CERROR("REPLY MD also attached to portal %d\n",
4339 md->md_me->me_portal);
4341 lnet_res_unlock(cpt);
4342 return -ENOENT; /* -ve: OK but no match */
4345 LASSERT(md->md_offset == 0);
4347 rlength = hdr->payload_length;
4348 mlength = min(rlength, md->md_length);
4350 if (mlength < rlength &&
4351 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4352 CNETERR("%s: Dropping REPLY from %s length %d "
4353 "for MD %#llx would overflow (%d)\n",
4354 libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4355 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4357 lnet_res_unlock(cpt);
4358 return -ENOENT; /* -ve: OK but no match */
4361 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4362 libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4363 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4365 lnet_msg_attach_md(msg, md, 0, mlength);
4368 lnet_setpayloadbuffer(msg);
4370 lnet_res_unlock(cpt);
4372 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4374 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4379 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4381 struct lnet_hdr *hdr = &msg->msg_hdr;
4382 struct lnet_process_id src = {0};
4383 struct lnet_libmd *md;
4386 src.nid = hdr->src_nid;
4387 src.pid = hdr->src_pid;
4389 /* Convert ack fields to host byte order */
4390 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4391 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4393 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4396 /* NB handles only looked up by creator (no flips) */
4397 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4398 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4399 /* Don't moan; this is expected */
4401 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4402 libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4403 (md == NULL) ? "invalid" : "inactive",
4404 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4405 hdr->msg.ack.dst_wmd.wh_object_cookie);
4406 if (md != NULL && md->md_me != NULL)
4407 CERROR("Source MD also attached to portal %d\n",
4408 md->md_me->me_portal);
4410 lnet_res_unlock(cpt);
4411 return -ENOENT; /* -ve! */
4414 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4415 libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4416 hdr->msg.ack.dst_wmd.wh_object_cookie);
4418 lnet_msg_attach_md(msg, md, 0, 0);
4420 lnet_res_unlock(cpt);
4422 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4424 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4429 * \retval LNET_CREDIT_OK If \a msg is forwarded
4430 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4431 * \retval -ve error code
4434 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4438 if (!the_lnet.ln_routing)
4441 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4442 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4443 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4444 msg->msg_rx_ready_delay = 1;
4446 lnet_net_unlock(msg->msg_rx_cpt);
4447 rc = lnet_ni_eager_recv(ni, msg);
4448 lnet_net_lock(msg->msg_rx_cpt);
4453 rc = lnet_post_routed_recv_locked(msg, 0);
4458 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4462 switch (msg->msg_type) {
4464 rc = lnet_parse_ack(ni, msg);
4467 rc = lnet_parse_put(ni, msg);
4470 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4472 case LNET_MSG_REPLY:
4473 rc = lnet_parse_reply(ni, msg);
4475 default: /* prevent an unused label if !kernel */
4480 LASSERT(rc == 0 || rc == -ENOENT);
4485 lnet_msgtyp2str (int type)
4494 case LNET_MSG_REPLY:
4496 case LNET_MSG_HELLO:
4499 return ("<UNKNOWN>");
4502 EXPORT_SYMBOL(lnet_msgtyp2str);
4505 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid4,
4506 void *private, int rdma_req)
4508 struct lnet_peer_ni *lpni;
4509 struct lnet_msg *msg;
4510 __u32 payload_length;
4511 lnet_pid_t dest_pid;
4512 lnet_nid_t dest_nid;
4514 struct lnet_nid from_nid;
4521 LASSERT (!in_interrupt ());
4523 lnet_nid4_to_nid(from_nid4, &from_nid);
4525 type = le32_to_cpu(hdr->type);
4526 src_nid = le64_to_cpu(hdr->src_nid);
4527 dest_nid = le64_to_cpu(hdr->dest_nid);
4528 dest_pid = le32_to_cpu(hdr->dest_pid);
4529 payload_length = le32_to_cpu(hdr->payload_length);
4531 /* FIXME handle large-addr nids */
4532 for_me = (lnet_nid_to_nid4(&ni->ni_nid) == dest_nid);
4533 cpt = lnet_cpt_of_nid(from_nid4, ni);
4535 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4536 libcfs_nid2str(dest_nid),
4537 libcfs_nidstr(&ni->ni_nid),
4538 libcfs_nid2str(src_nid),
4539 lnet_msgtyp2str(type),
4540 (for_me) ? "for me" : "routed");
4545 if (payload_length > 0) {
4546 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4547 libcfs_nid2str(from_nid4),
4548 libcfs_nid2str(src_nid),
4549 lnet_msgtyp2str(type), payload_length);
4555 case LNET_MSG_REPLY:
4556 if (payload_length >
4557 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4558 CERROR("%s, src %s: bad %s payload %d "
4559 "(%d max expected)\n",
4560 libcfs_nid2str(from_nid4),
4561 libcfs_nid2str(src_nid),
4562 lnet_msgtyp2str(type),
4564 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4570 CERROR("%s, src %s: Bad message type 0x%x\n",
4571 libcfs_nid2str(from_nid4),
4572 libcfs_nid2str(src_nid), type);
4576 if (the_lnet.ln_routing &&
4577 ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4579 spin_lock(&ni->ni_net->net_lock);
4580 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4581 spin_unlock(&ni->ni_net->net_lock);
4582 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4587 lnet_push_update_to_peers(1);
4589 /* Regard a bad destination NID as a protocol error. Senders should
4590 * know what they're doing; if they don't they're misconfigured, buggy
4591 * or malicious so we chop them off at the knees :) */
4594 if (LNET_NIDNET(dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4595 /* should have gone direct */
4596 CERROR("%s, src %s: Bad dest nid %s "
4597 "(should have been sent direct)\n",
4598 libcfs_nid2str(from_nid4),
4599 libcfs_nid2str(src_nid),
4600 libcfs_nid2str(dest_nid));
4604 if (lnet_islocalnid4(dest_nid)) {
4605 /* dest is another local NI; sender should have used
4606 * this node's NID on its own network */
4607 CERROR("%s, src %s: Bad dest nid %s "
4608 "(it's my nid but on a different network)\n",
4609 libcfs_nid2str(from_nid4),
4610 libcfs_nid2str(src_nid),
4611 libcfs_nid2str(dest_nid));
4615 if (rdma_req && type == LNET_MSG_GET) {
4616 CERROR("%s, src %s: Bad optimized GET for %s "
4617 "(final destination must be me)\n",
4618 libcfs_nid2str(from_nid4),
4619 libcfs_nid2str(src_nid),
4620 libcfs_nid2str(dest_nid));
4624 if (!the_lnet.ln_routing) {
4625 CERROR("%s, src %s: Dropping message for %s "
4626 "(routing not enabled)\n",
4627 libcfs_nid2str(from_nid4),
4628 libcfs_nid2str(src_nid),
4629 libcfs_nid2str(dest_nid));
4634 /* Message looks OK; we're not going to return an error, so we MUST
4635 * call back lnd_recv() come what may... */
4637 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4638 fail_peer(src_nid, 0)) { /* shall we now? */
4639 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4640 libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4641 lnet_msgtyp2str(type));
4645 /* FIXME need to support large-addr nid */
4646 if (!list_empty(&the_lnet.ln_drop_rules) &&
4647 lnet_drop_rule_match(hdr, lnet_nid_to_nid4(&ni->ni_nid), NULL)) {
4649 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4650 libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4651 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4655 msg = lnet_msg_alloc();
4657 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4658 libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4659 lnet_msgtyp2str(type));
4663 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4664 * pointers NULL etc */
4666 msg->msg_type = type;
4667 msg->msg_private = private;
4668 msg->msg_receiving = 1;
4669 msg->msg_rdma_get = rdma_req;
4670 msg->msg_len = msg->msg_wanted = payload_length;
4671 msg->msg_offset = 0;
4672 msg->msg_hdr = *hdr;
4673 /* for building message event */
4674 msg->msg_from = from_nid;
4676 msg->msg_target.pid = dest_pid;
4677 lnet_nid4_to_nid(dest_nid, &msg->msg_target.nid);
4678 msg->msg_routing = 1;
4681 /* convert common msg->hdr fields to host byteorder */
4682 msg->msg_hdr.type = type;
4683 msg->msg_hdr.src_nid = src_nid;
4684 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4685 msg->msg_hdr.dest_nid = dest_nid;
4686 msg->msg_hdr.dest_pid = dest_pid;
4687 msg->msg_hdr.payload_length = payload_length;
4691 lpni = lnet_peerni_by_nid_locked(&from_nid, &ni->ni_nid, cpt);
4693 lnet_net_unlock(cpt);
4695 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4696 libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4697 lnet_msgtyp2str(type), rc);
4699 if (rc == -ESHUTDOWN)
4700 /* We are shutting down. Don't do anything more */
4705 /* If this message was forwarded to us from a router then we may need
4706 * to update router aliveness or check for an asymmetrical route
4709 if (((lnet_drop_asym_route && for_me) ||
4710 !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4711 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid4)) {
4712 __u32 src_net_id = LNET_NIDNET(src_nid);
4713 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4714 struct lnet_route *route;
4717 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4718 if (route->lr_net == src_net_id) {
4720 /* If we're transitioning the gateway from
4721 * dead -> alive, and discovery is disabled
4722 * locally or on the gateway, then we need to
4723 * update the cached route aliveness for each
4724 * route to the src_nid's net.
4726 * Otherwise, we're only checking for
4727 * symmetrical route, and we can break the
4730 if (!gw->lp_alive &&
4731 lnet_is_discovery_disabled(gw))
4732 lnet_set_route_aliveness(route, true);
4737 if (lnet_drop_asym_route && for_me && !found) {
4738 /* Drop ref taken by lnet_nid2peerni_locked() */
4739 lnet_peer_ni_decref_locked(lpni);
4740 lnet_net_unlock(cpt);
4741 /* we would not use from_nid to route a message to
4743 * => asymmetric routing detected but forbidden
4745 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4746 libcfs_nid2str(from_nid4),
4747 libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4751 if (!gw->lp_alive) {
4752 struct lnet_peer_net *lpn;
4753 struct lnet_peer_ni *lpni2;
4755 gw->lp_alive = true;
4756 /* Mark all remote NIs on src_nid's net UP */
4757 lpn = lnet_peer_get_net_locked(gw, src_net_id);
4759 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4761 lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4765 lpni->lpni_last_alive = ktime_get_seconds();
4767 msg->msg_rxpeer = lpni;
4769 lnet_ni_addref_locked(ni, cpt);
4770 /* Multi-Rail: Primary NID of source. */
4771 lnet_peer_primary_nid_locked(src_nid, &msg->msg_initiator);
4774 * mark the status of this lpni as UP since we received a message
4775 * from it. The ping response reports back the ns_status which is
4776 * marked on the remote as up or down and we cache it here.
4778 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4780 lnet_msg_commit(msg, cpt);
4782 /* message delay simulation */
4783 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4784 lnet_delay_rule_match_locked(hdr, msg))) {
4785 lnet_net_unlock(cpt);
4790 rc = lnet_parse_forward_locked(ni, msg);
4791 lnet_net_unlock(cpt);
4796 if (rc == LNET_CREDIT_OK) {
4797 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4798 0, payload_length, payload_length);
4803 lnet_net_unlock(cpt);
4805 rc = lnet_parse_local(ni, msg);
4811 LASSERT(msg->msg_md == NULL);
4812 lnet_finalize(msg, rc);
4815 lnet_drop_message(ni, cpt, private, payload_length, type);
4818 EXPORT_SYMBOL(lnet_parse);
4821 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4823 while (!list_empty(head)) {
4824 struct lnet_process_id id = {0};
4825 struct lnet_msg *msg;
4827 msg = list_entry(head->next, struct lnet_msg, msg_list);
4828 list_del(&msg->msg_list);
4830 id.nid = msg->msg_hdr.src_nid;
4831 id.pid = msg->msg_hdr.src_pid;
4833 LASSERT(msg->msg_md == NULL);
4834 LASSERT(msg->msg_rx_delayed);
4835 LASSERT(msg->msg_rxpeer != NULL);
4836 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4838 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4839 " offset %d length %d: %s\n",
4841 msg->msg_hdr.msg.put.ptl_index,
4842 msg->msg_hdr.msg.put.match_bits,
4843 msg->msg_hdr.msg.put.offset,
4844 msg->msg_hdr.payload_length, reason);
4846 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4847 * called lnet_drop_message(), so I just hang onto msg as well
4848 * until that's done */
4850 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4851 msg->msg_private, msg->msg_len,
4854 msg->msg_no_resend = true;
4856 * NB: message will not generate event because w/o attached MD,
4857 * but we still should give error code so lnet_msg_decommit()
4858 * can skip counters operations and other checks.
4860 lnet_finalize(msg, -ENOENT);
4865 lnet_recv_delayed_msg_list(struct list_head *head)
4867 while (!list_empty(head)) {
4868 struct lnet_msg *msg;
4869 struct lnet_process_id id;
4871 msg = list_entry(head->next, struct lnet_msg, msg_list);
4872 list_del(&msg->msg_list);
4874 /* md won't disappear under me, since each msg
4875 * holds a ref on it */
4877 id.nid = msg->msg_hdr.src_nid;
4878 id.pid = msg->msg_hdr.src_pid;
4880 LASSERT(msg->msg_rx_delayed);
4881 LASSERT(msg->msg_md != NULL);
4882 LASSERT(msg->msg_rxpeer != NULL);
4883 LASSERT(msg->msg_rxni != NULL);
4884 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4886 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4887 "match %llu offset %d length %d.\n",
4888 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4889 msg->msg_hdr.msg.put.match_bits,
4890 msg->msg_hdr.msg.put.offset,
4891 msg->msg_hdr.payload_length);
4893 lnet_recv_put(msg->msg_rxni, msg);
4898 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4899 struct lnet_libmd *md, struct lnet_handle_md mdh)
4902 struct lnet_rsp_tracker *local_rspt;
4905 * MD has a refcount taken by message so it's not going away.
4906 * The MD however can be looked up. We need to secure the access
4907 * to the md_rspt_ptr by taking the res_lock.
4908 * The rspt can be accessed without protection up to when it gets
4909 * added to the list.
4913 local_rspt = md->md_rspt_ptr;
4914 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4915 if (local_rspt != NULL) {
4917 * we already have an rspt attached to the md, so we'll
4918 * update the deadline on that one.
4920 lnet_rspt_free(rspt, cpt);
4923 rspt->rspt_mdh = mdh;
4924 rspt->rspt_cpt = cpt;
4925 /* store the rspt so we can access it when we get the REPLY */
4926 md->md_rspt_ptr = rspt;
4929 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4932 * add to the list of tracked responses. It's added to tail of the
4933 * list in order to expire all the older entries first.
4936 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4937 lnet_net_unlock(cpt);
4938 lnet_res_unlock(cpt);
4942 * Initiate an asynchronous PUT operation.
4944 * There are several events associated with a PUT: completion of the send on
4945 * the initiator node (LNET_EVENT_SEND), and when the send completes
4946 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4947 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4948 * used at the target node to indicate the completion of incoming data
4951 * The local events will be logged in the EQ associated with the MD pointed to
4952 * by \a mdh handle. Using a MD without an associated EQ results in these
4953 * events being discarded. In this case, the caller must have another
4954 * mechanism (e.g., a higher level protocol) for determining when it is safe
4955 * to modify the memory region associated with the MD.
4957 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4958 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4960 * \param self Indicates the NID of a local interface through which to send
4961 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4962 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4963 * must be "free floating" (See LNetMDBind()).
4964 * \param ack Controls whether an acknowledgment is requested.
4965 * Acknowledgments are only sent when they are requested by the initiating
4966 * process and the target MD enables them.
4967 * \param target A process identifier for the target process.
4968 * \param portal The index in the \a target's portal table.
4969 * \param match_bits The match bits to use for MD selection at the target
4971 * \param offset The offset into the target MD (only used when the target
4972 * MD has the LNET_MD_MANAGE_REMOTE option set).
4973 * \param hdr_data 64 bits of user data that can be included in the message
4974 * header. This data is written to an event queue entry at the target if an
4975 * EQ is present on the matching MD.
4977 * \retval 0 Success, and only in this case events will be generated
4978 * and logged to EQ (if it exists).
4979 * \retval -EIO Simulated failure.
4980 * \retval -ENOMEM Memory allocation failure.
4981 * \retval -ENOENT Invalid MD object.
4983 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4986 LNetPut(lnet_nid_t self4, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4987 struct lnet_process_id target, unsigned int portal,
4988 __u64 match_bits, unsigned int offset,
4991 struct lnet_msg *msg;
4992 struct lnet_libmd *md;
4995 struct lnet_rsp_tracker *rspt = NULL;
4996 struct lnet_nid self;
4998 LASSERT(the_lnet.ln_refcount > 0);
5000 lnet_nid4_to_nid(self4, &self);
5002 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5003 fail_peer(target.nid, 1)) { /* shall we now? */
5004 CERROR("Dropping PUT to %s: simulated failure\n",
5005 libcfs_id2str(target));
5009 msg = lnet_msg_alloc();
5011 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
5012 libcfs_id2str(target));
5015 msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
5017 cpt = lnet_cpt_of_cookie(mdh.cookie);
5019 if (ack == LNET_ACK_REQ) {
5020 rspt = lnet_rspt_alloc(cpt);
5022 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
5023 libcfs_id2str(target));
5026 INIT_LIST_HEAD(&rspt->rspt_on_list);
5031 md = lnet_handle2md(&mdh);
5032 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5033 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5034 match_bits, portal, libcfs_id2str(target),
5035 md == NULL ? -1 : md->md_threshold);
5036 if (md != NULL && md->md_me != NULL)
5037 CERROR("Source MD also attached to portal %d\n",
5038 md->md_me->me_portal);
5039 lnet_res_unlock(cpt);
5042 lnet_rspt_free(rspt, cpt);
5048 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
5050 lnet_msg_attach_md(msg, md, 0, 0);
5052 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5054 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5055 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5056 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5057 msg->msg_hdr.msg.put.hdr_data = hdr_data;
5059 /* NB handles only looked up by creator (no flips) */
5060 if (ack == LNET_ACK_REQ) {
5061 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5062 the_lnet.ln_interface_cookie;
5063 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5064 md->md_lh.lh_cookie;
5066 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5067 LNET_WIRE_HANDLE_COOKIE_NONE;
5068 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5069 LNET_WIRE_HANDLE_COOKIE_NONE;
5072 lnet_res_unlock(cpt);
5074 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5076 if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5078 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5080 lnet_rspt_free(rspt, cpt);
5082 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5086 rc = lnet_send(&self, msg, NULL);
5089 CNETERR("Error sending PUT to %s: %d\n",
5090 libcfs_id2str(target), rc);
5091 msg->msg_no_resend = true;
5092 lnet_finalize(msg, rc);
5095 /* completion will be signalled by an event */
5098 EXPORT_SYMBOL(LNetPut);
5101 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
5102 * returns a msg for the LND to pass to lnet_finalize() when the sink
5103 * data has been received.
5105 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5106 * lnet_finalize() is called on it, so the LND must call this first
5109 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5111 struct lnet_msg *msg = lnet_msg_alloc();
5112 struct lnet_libmd *getmd = getmsg->msg_md;
5113 struct lnet_processid *peer_id = &getmsg->msg_target;
5116 LASSERT(!getmsg->msg_target_is_router);
5117 LASSERT(!getmsg->msg_routing);
5120 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5121 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id));
5125 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5128 LASSERT(getmd->md_refcount > 0);
5130 if (getmd->md_threshold == 0) {
5131 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5132 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id),
5134 lnet_res_unlock(cpt);
5138 LASSERT(getmd->md_offset == 0);
5140 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5141 libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id), getmd);
5143 /* setup information for lnet_build_msg_event */
5144 msg->msg_initiator =
5145 getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5146 msg->msg_from = peer_id->nid;
5147 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5148 msg->msg_hdr.src_nid = lnet_nid_to_nid4(&peer_id->nid);
5149 msg->msg_hdr.payload_length = getmd->md_length;
5150 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5152 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5153 lnet_res_unlock(cpt);
5155 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5158 lnet_msg_commit(msg, cpt);
5159 lnet_net_unlock(cpt);
5161 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5166 cpt = lnet_nid2cpt(&peer_id->nid, ni);
5169 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5170 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5171 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5173 lnet_net_unlock(cpt);
5180 EXPORT_SYMBOL(lnet_create_reply_msg);
5183 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5186 /* Set the REPLY length, now the RDMA that elides the REPLY message has
5187 * completed and I know it. */
5188 LASSERT(reply != NULL);
5189 LASSERT(reply->msg_type == LNET_MSG_GET);
5190 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5192 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
5193 * the end of my buffer, I might as well be dead. */
5194 LASSERT(len <= reply->msg_ev.mlength);
5196 reply->msg_ev.mlength = len;
5198 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5201 * Initiate an asynchronous GET operation.
5203 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5204 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5205 * the target node in the REPLY has been written to local MD.
5207 * On the target node, an LNET_EVENT_GET is logged when the GET request
5208 * arrives and is accepted into a MD.
5210 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5211 * \param mdh A handle for the MD that describes the memory into which the
5212 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5214 * \retval 0 Success, and only in this case events will be generated
5215 * and logged to EQ (if it exists) of the MD.
5216 * \retval -EIO Simulated failure.
5217 * \retval -ENOMEM Memory allocation failure.
5218 * \retval -ENOENT Invalid MD object.
5221 LNetGet(lnet_nid_t self4, struct lnet_handle_md mdh,
5222 struct lnet_process_id target, unsigned int portal,
5223 __u64 match_bits, unsigned int offset, bool recovery)
5225 struct lnet_msg *msg;
5226 struct lnet_libmd *md;
5227 struct lnet_rsp_tracker *rspt;
5230 struct lnet_nid self;
5232 LASSERT(the_lnet.ln_refcount > 0);
5234 lnet_nid4_to_nid(self4, &self);
5236 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5237 fail_peer(target.nid, 1)) /* shall we now? */
5239 CERROR("Dropping GET to %s: simulated failure\n",
5240 libcfs_id2str(target));
5244 msg = lnet_msg_alloc();
5246 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5247 libcfs_id2str(target));
5251 cpt = lnet_cpt_of_cookie(mdh.cookie);
5253 rspt = lnet_rspt_alloc(cpt);
5255 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5256 libcfs_id2str(target));
5259 INIT_LIST_HEAD(&rspt->rspt_on_list);
5261 msg->msg_recovery = recovery;
5265 md = lnet_handle2md(&mdh);
5266 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5267 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5268 match_bits, portal, libcfs_id2str(target),
5269 md == NULL ? -1 : md->md_threshold);
5270 if (md != NULL && md->md_me != NULL)
5271 CERROR("REPLY MD also attached to portal %d\n",
5272 md->md_me->me_portal);
5274 lnet_res_unlock(cpt);
5277 lnet_rspt_free(rspt, cpt);
5281 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5283 lnet_msg_attach_md(msg, md, 0, 0);
5285 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5287 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5288 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5289 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5290 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5292 /* NB handles only looked up by creator (no flips) */
5293 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5294 the_lnet.ln_interface_cookie;
5295 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5296 md->md_lh.lh_cookie;
5298 lnet_res_unlock(cpt);
5300 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5302 if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5303 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5305 lnet_rspt_free(rspt, cpt);
5307 rc = lnet_send(&self, msg, NULL);
5309 CNETERR("Error sending GET to %s: %d\n",
5310 libcfs_id2str(target), rc);
5311 msg->msg_no_resend = true;
5312 lnet_finalize(msg, rc);
5315 /* completion will be signalled by an event */
5318 EXPORT_SYMBOL(LNetGet);
5321 * Calculate distance to node at \a dstnid.
5323 * \param dstnid Target NID.
5324 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5326 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5329 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5330 * local_nid_dist_zero is set, which is the default.
5331 * \retval positives Distance to target NID, i.e. number of hops plus one.
5332 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5335 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5337 struct list_head *e;
5338 struct lnet_ni *ni = NULL;
5339 struct lnet_remotenet *rnet;
5340 __u32 dstnet = LNET_NIDNET(dstnid);
5344 struct list_head *rn_list;
5345 bool matched_dstnet = false;
5347 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5348 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5349 * keep order 0 free for 0@lo and order 1 free for a local NID
5352 LASSERT(the_lnet.ln_refcount > 0);
5354 cpt = lnet_net_lock_current();
5356 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5357 /* FIXME support large-addr nid */
5358 if (lnet_nid_to_nid4(&ni->ni_nid) == dstnid) {
5359 if (srcnidp != NULL)
5361 if (orderp != NULL) {
5362 if (dstnid == LNET_NID_LO_0)
5367 lnet_net_unlock(cpt);
5369 return local_nid_dist_zero ? 0 : 1;
5372 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5373 matched_dstnet = true;
5374 /* We matched the destination net, but we may have
5375 * additional local NIs to inspect.
5377 * We record the nid and order as appropriate, but
5378 * they may be overwritten if we match local NI above.
5381 /* FIXME support large-addr nids */
5382 *srcnidp = lnet_nid_to_nid4(&ni->ni_nid);
5385 /* Check if ni was originally created in
5386 * current net namespace.
5387 * If not, assign order above 0xffff0000,
5388 * to make this ni not a priority.
5390 if (current->nsproxy &&
5391 !net_eq(ni->ni_net_ns,
5392 current->nsproxy->net_ns))
5393 *orderp = order + 0xffff0000;
5402 if (matched_dstnet) {
5403 lnet_net_unlock(cpt);
5407 rn_list = lnet_net2rnethash(dstnet);
5408 list_for_each(e, rn_list) {
5409 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5411 if (rnet->lrn_net == dstnet) {
5412 struct lnet_route *route;
5413 struct lnet_route *shortest = NULL;
5414 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5417 LASSERT(!list_empty(&rnet->lrn_routes));
5419 list_for_each_entry(route, &rnet->lrn_routes,
5421 route_hops = route->lr_hops;
5422 if (route_hops == LNET_UNDEFINED_HOPS)
5424 if (shortest == NULL ||
5425 route_hops < shortest_hops) {
5427 shortest_hops = route_hops;
5431 LASSERT(shortest != NULL);
5432 hops = shortest_hops;
5433 if (srcnidp != NULL) {
5434 struct lnet_net *net;
5435 net = lnet_get_net_locked(shortest->lr_lnet);
5437 ni = lnet_get_next_ni_locked(net, NULL);
5438 /* FIXME support large-addr nids */
5439 *srcnidp = lnet_nid_to_nid4(&ni->ni_nid);
5443 lnet_net_unlock(cpt);
5449 lnet_net_unlock(cpt);
5450 return -EHOSTUNREACH;
5452 EXPORT_SYMBOL(LNetDist);