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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/lnet/lib-move.c
34 * Data movement routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/pagemap.h>
41 #include <lnet/lib-lnet.h>
42 #include <linux/nsproxy.h>
43 #include <net/net_namespace.h>
45 static int local_nid_dist_zero = 1;
46 module_param(local_nid_dist_zero, int, 0444);
47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
49 struct lnet_send_data {
50 struct lnet_ni *sd_best_ni;
51 struct lnet_peer_ni *sd_best_lpni;
52 struct lnet_peer_ni *sd_final_dst_lpni;
53 struct lnet_peer *sd_peer;
54 struct lnet_peer *sd_gw_peer;
55 struct lnet_peer_ni *sd_gw_lpni;
56 struct lnet_peer_net *sd_peer_net;
57 struct lnet_msg *sd_msg;
58 lnet_nid_t sd_dst_nid;
59 lnet_nid_t sd_src_nid;
60 lnet_nid_t sd_rtr_nid;
67 lnet_msg_is_response(struct lnet_msg *msg)
69 return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
72 static inline struct lnet_comm_count *
73 get_stats_counts(struct lnet_element_stats *stats,
74 enum lnet_stats_type stats_type)
77 case LNET_STATS_TYPE_SEND:
78 return &stats->el_send_stats;
79 case LNET_STATS_TYPE_RECV:
80 return &stats->el_recv_stats;
81 case LNET_STATS_TYPE_DROP:
82 return &stats->el_drop_stats;
84 CERROR("Unknown stats type\n");
90 void lnet_incr_stats(struct lnet_element_stats *stats,
91 enum lnet_msg_type msg_type,
92 enum lnet_stats_type stats_type)
94 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
100 atomic_inc(&counts->co_ack_count);
103 atomic_inc(&counts->co_put_count);
106 atomic_inc(&counts->co_get_count);
109 atomic_inc(&counts->co_reply_count);
112 atomic_inc(&counts->co_hello_count);
115 CERROR("There is a BUG in the code. Unknown message type\n");
120 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
121 enum lnet_stats_type stats_type)
123 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
127 return (atomic_read(&counts->co_ack_count) +
128 atomic_read(&counts->co_put_count) +
129 atomic_read(&counts->co_get_count) +
130 atomic_read(&counts->co_reply_count) +
131 atomic_read(&counts->co_hello_count));
134 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
135 struct lnet_comm_count *counts)
137 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
138 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
139 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
140 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
141 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
144 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
145 struct lnet_element_stats *stats)
147 struct lnet_comm_count *counts;
152 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
155 assign_stats(&msg_stats->im_send_stats, counts);
157 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
160 assign_stats(&msg_stats->im_recv_stats, counts);
162 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
165 assign_stats(&msg_stats->im_drop_stats, counts);
169 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
171 struct lnet_test_peer *tp;
172 struct list_head *el;
173 struct list_head *next;
176 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
177 if (threshold != 0) {
178 /* Adding a new entry */
179 LIBCFS_ALLOC(tp, sizeof(*tp));
184 tp->tp_threshold = threshold;
187 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
194 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
195 tp = list_entry(el, struct lnet_test_peer, tp_list);
197 if (tp->tp_threshold == 0 || /* needs culling anyway */
198 nid == LNET_NID_ANY || /* removing all entries */
199 tp->tp_nid == nid) { /* matched this one */
200 list_move(&tp->tp_list, &cull);
206 while (!list_empty(&cull)) {
207 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
209 list_del(&tp->tp_list);
210 LIBCFS_FREE(tp, sizeof(*tp));
216 fail_peer (lnet_nid_t nid, int outgoing)
218 struct lnet_test_peer *tp;
219 struct list_head *el;
220 struct list_head *next;
224 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
227 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
228 tp = list_entry(el, struct lnet_test_peer, tp_list);
230 if (tp->tp_threshold == 0) {
233 /* only cull zombies on outgoing tests,
234 * since we may be at interrupt priority on
235 * incoming messages. */
236 list_move(&tp->tp_list, &cull);
241 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
242 nid == tp->tp_nid) { /* fail this peer */
245 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
248 tp->tp_threshold == 0) {
250 list_move(&tp->tp_list, &cull);
259 while (!list_empty(&cull)) {
260 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
261 list_del(&tp->tp_list);
263 LIBCFS_FREE(tp, sizeof(*tp));
270 lnet_iov_nob(unsigned int niov, struct kvec *iov)
272 unsigned int nob = 0;
274 LASSERT(niov == 0 || iov != NULL);
276 nob += (iov++)->iov_len;
280 EXPORT_SYMBOL(lnet_iov_nob);
283 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
284 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
287 /* NB diov, siov are READ-ONLY */
288 unsigned int this_nob;
293 /* skip complete frags before 'doffset' */
295 while (doffset >= diov->iov_len) {
296 doffset -= diov->iov_len;
302 /* skip complete frags before 'soffset' */
304 while (soffset >= siov->iov_len) {
305 soffset -= siov->iov_len;
314 this_nob = min3((unsigned int)diov->iov_len - doffset,
315 (unsigned int)siov->iov_len - soffset,
318 memcpy((char *)diov->iov_base + doffset,
319 (char *)siov->iov_base + soffset, this_nob);
322 if (diov->iov_len > doffset + this_nob) {
330 if (siov->iov_len > soffset + this_nob) {
339 EXPORT_SYMBOL(lnet_copy_iov2iov);
342 lnet_extract_iov(int dst_niov, struct kvec *dst,
343 int src_niov, struct kvec *src,
344 unsigned int offset, unsigned int len)
346 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
347 * for exactly 'len' bytes, and return the number of entries.
348 * NB not destructive to 'src' */
349 unsigned int frag_len;
352 if (len == 0) /* no data => */
353 return (0); /* no frags */
355 LASSERT(src_niov > 0);
356 while (offset >= src->iov_len) { /* skip initial frags */
357 offset -= src->iov_len;
360 LASSERT(src_niov > 0);
365 LASSERT(src_niov > 0);
366 LASSERT((int)niov <= dst_niov);
368 frag_len = src->iov_len - offset;
369 dst->iov_base = ((char *)src->iov_base) + offset;
371 if (len <= frag_len) {
376 dst->iov_len = frag_len;
386 EXPORT_SYMBOL(lnet_extract_iov);
390 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
392 unsigned int nob = 0;
394 LASSERT(niov == 0 || kiov != NULL);
396 nob += (kiov++)->bv_len;
400 EXPORT_SYMBOL(lnet_kiov_nob);
403 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
404 unsigned int doffset,
405 unsigned int nsiov, struct bio_vec *siov,
406 unsigned int soffset,
409 /* NB diov, siov are READ-ONLY */
410 unsigned int this_nob;
417 LASSERT (!in_interrupt ());
420 while (doffset >= diov->bv_len) {
421 doffset -= diov->bv_len;
428 while (soffset >= siov->bv_len) {
429 soffset -= siov->bv_len;
438 this_nob = min3(diov->bv_len - doffset,
439 siov->bv_len - soffset,
443 daddr = ((char *)kmap(diov->bv_page)) +
444 diov->bv_offset + doffset;
446 saddr = ((char *)kmap(siov->bv_page)) +
447 siov->bv_offset + soffset;
449 /* Vanishing risk of kmap deadlock when mapping 2 pages.
450 * However in practice at least one of the kiovs will be mapped
451 * kernel pages and the map/unmap will be NOOPs */
453 memcpy (daddr, saddr, this_nob);
456 if (diov->bv_len > doffset + this_nob) {
460 kunmap(diov->bv_page);
467 if (siov->bv_len > soffset + this_nob) {
471 kunmap(siov->bv_page);
480 kunmap(diov->bv_page);
482 kunmap(siov->bv_page);
484 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
487 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
488 unsigned int nkiov, struct bio_vec *kiov,
489 unsigned int kiovoffset,
492 /* NB iov, kiov are READ-ONLY */
493 unsigned int this_nob;
499 LASSERT (!in_interrupt ());
502 while (iovoffset >= iov->iov_len) {
503 iovoffset -= iov->iov_len;
510 while (kiovoffset >= kiov->bv_len) {
511 kiovoffset -= kiov->bv_len;
520 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
521 (unsigned int)kiov->bv_len - kiovoffset,
525 addr = ((char *)kmap(kiov->bv_page)) +
526 kiov->bv_offset + kiovoffset;
528 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
531 if (iov->iov_len > iovoffset + this_nob) {
532 iovoffset += this_nob;
539 if (kiov->bv_len > kiovoffset + this_nob) {
541 kiovoffset += this_nob;
543 kunmap(kiov->bv_page);
553 kunmap(kiov->bv_page);
555 EXPORT_SYMBOL(lnet_copy_kiov2iov);
558 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
559 unsigned int kiovoffset,
560 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
563 /* NB kiov, iov are READ-ONLY */
564 unsigned int this_nob;
570 LASSERT (!in_interrupt ());
573 while (kiovoffset >= kiov->bv_len) {
574 kiovoffset -= kiov->bv_len;
581 while (iovoffset >= iov->iov_len) {
582 iovoffset -= iov->iov_len;
591 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
592 (unsigned int)iov->iov_len - iovoffset,
596 addr = ((char *)kmap(kiov->bv_page)) +
597 kiov->bv_offset + kiovoffset;
599 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
602 if (kiov->bv_len > kiovoffset + this_nob) {
604 kiovoffset += this_nob;
606 kunmap(kiov->bv_page);
613 if (iov->iov_len > iovoffset + this_nob) {
614 iovoffset += this_nob;
623 kunmap(kiov->bv_page);
625 EXPORT_SYMBOL(lnet_copy_iov2kiov);
628 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
629 int src_niov, struct bio_vec *src,
630 unsigned int offset, unsigned int len)
632 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
633 * for exactly 'len' bytes, and return the number of entries.
634 * NB not destructive to 'src' */
635 unsigned int frag_len;
638 if (len == 0) /* no data => */
639 return (0); /* no frags */
641 LASSERT(src_niov > 0);
642 while (offset >= src->bv_len) { /* skip initial frags */
643 offset -= src->bv_len;
646 LASSERT(src_niov > 0);
651 LASSERT(src_niov > 0);
652 LASSERT((int)niov <= dst_niov);
654 frag_len = src->bv_len - offset;
655 dst->bv_page = src->bv_page;
656 dst->bv_offset = src->bv_offset + offset;
658 if (len <= frag_len) {
660 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
664 dst->bv_len = frag_len;
665 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
675 EXPORT_SYMBOL(lnet_extract_kiov);
678 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
679 int delayed, unsigned int offset, unsigned int mlen,
682 unsigned int niov = 0;
683 struct kvec *iov = NULL;
684 struct bio_vec *kiov = NULL;
687 LASSERT (!in_interrupt ());
688 LASSERT (mlen == 0 || msg != NULL);
691 LASSERT(msg->msg_receiving);
692 LASSERT(!msg->msg_sending);
693 LASSERT(rlen == msg->msg_len);
694 LASSERT(mlen <= msg->msg_len);
695 LASSERT(msg->msg_offset == offset);
696 LASSERT(msg->msg_wanted == mlen);
698 msg->msg_receiving = 0;
701 niov = msg->msg_niov;
703 kiov = msg->msg_kiov;
706 LASSERT ((iov == NULL) != (kiov == NULL));
710 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
711 niov, iov, kiov, offset, mlen,
714 lnet_finalize(msg, rc);
718 lnet_setpayloadbuffer(struct lnet_msg *msg)
720 struct lnet_libmd *md = msg->msg_md;
722 LASSERT(msg->msg_len > 0);
723 LASSERT(!msg->msg_routing);
725 LASSERT(msg->msg_niov == 0);
726 LASSERT(msg->msg_iov == NULL);
727 LASSERT(msg->msg_kiov == NULL);
729 msg->msg_niov = md->md_niov;
730 if ((md->md_options & LNET_MD_KIOV) != 0)
731 msg->msg_kiov = md->md_iov.kiov;
733 msg->msg_iov = md->md_iov.iov;
737 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
738 unsigned int offset, unsigned int len)
740 msg->msg_type = type;
741 msg->msg_target = target;
743 msg->msg_offset = offset;
746 lnet_setpayloadbuffer(msg);
748 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
749 msg->msg_hdr.type = cpu_to_le32(type);
750 /* dest_nid will be overwritten by lnet_select_pathway() */
751 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
752 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
753 /* src_nid will be set later */
754 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
755 msg->msg_hdr.payload_length = cpu_to_le32(len);
759 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
761 void *priv = msg->msg_private;
764 LASSERT (!in_interrupt ());
765 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
766 (msg->msg_txcredit && msg->msg_peertxcredit));
768 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
770 msg->msg_no_resend = true;
771 lnet_finalize(msg, rc);
776 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
780 LASSERT(!msg->msg_sending);
781 LASSERT(msg->msg_receiving);
782 LASSERT(!msg->msg_rx_ready_delay);
783 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
785 msg->msg_rx_ready_delay = 1;
786 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
789 CERROR("recv from %s / send to %s aborted: "
790 "eager_recv failed %d\n",
791 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
792 libcfs_id2str(msg->msg_target), rc);
793 LASSERT(rc < 0); /* required by my callers */
800 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
804 deadline = lpni->lpni_last_alive +
805 lpni->lpni_net->net_tunables.lct_peer_timeout;
808 * assume peer_ni is alive as long as we're within the configured
817 /* NB: returns 1 when alive, 0 when dead, negative when error;
818 * may drop the lnet_net_lock */
820 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
821 struct lnet_msg *msg)
823 time64_t now = ktime_get_seconds();
825 if (!lnet_peer_aliveness_enabled(lpni))
829 * If we're resending a message, let's attempt to send it even if
830 * the peer is down to fulfill our resend quota on the message
832 if (msg->msg_retry_count > 0)
835 /* try and send recovery messages irregardless */
836 if (msg->msg_recovery)
839 /* always send any responses */
840 if (lnet_msg_is_response(msg))
843 if (!lnet_is_peer_deadline_passed(lpni, now))
846 return lnet_is_peer_ni_alive(lpni);
850 * \param msg The message to be sent.
851 * \param do_send True if lnet_ni_send() should be called in this function.
852 * lnet_send() is going to lnet_net_unlock immediately after this, so
853 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
855 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
856 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
857 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
858 * \retval -ECANCELED If the MD of the message has been unlinked.
861 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
863 struct lnet_peer_ni *lp = msg->msg_txpeer;
864 struct lnet_ni *ni = msg->msg_txni;
865 int cpt = msg->msg_tx_cpt;
866 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
868 /* non-lnet_send() callers have checked before */
869 LASSERT(!do_send || msg->msg_tx_delayed);
870 LASSERT(!msg->msg_receiving);
871 LASSERT(msg->msg_tx_committed);
872 /* can't get here if we're sending to the loopback interface */
873 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
875 /* NB 'lp' is always the next hop */
876 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
877 lnet_peer_alive_locked(ni, lp, msg) == 0) {
878 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
879 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
881 lnet_net_unlock(cpt);
883 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
885 LNET_STATS_TYPE_DROP);
887 lnet_incr_stats(&msg->msg_txni->ni_stats,
889 LNET_STATS_TYPE_DROP);
891 CNETERR("Dropping message for %s: peer not alive\n",
892 libcfs_id2str(msg->msg_target));
893 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
895 lnet_finalize(msg, -EHOSTUNREACH);
898 return -EHOSTUNREACH;
901 if (msg->msg_md != NULL &&
902 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
903 lnet_net_unlock(cpt);
905 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
906 "called on the MD/ME.\n",
907 libcfs_id2str(msg->msg_target));
909 msg->msg_no_resend = true;
910 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
911 msg, libcfs_id2str(msg->msg_target));
912 lnet_finalize(msg, -ECANCELED);
919 if (!msg->msg_peertxcredit) {
920 spin_lock(&lp->lpni_lock);
921 LASSERT((lp->lpni_txcredits < 0) ==
922 !list_empty(&lp->lpni_txq));
924 msg->msg_peertxcredit = 1;
925 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
926 lp->lpni_txcredits--;
928 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
929 lp->lpni_mintxcredits = lp->lpni_txcredits;
931 if (lp->lpni_txcredits < 0) {
932 msg->msg_tx_delayed = 1;
933 list_add_tail(&msg->msg_list, &lp->lpni_txq);
934 spin_unlock(&lp->lpni_lock);
935 return LNET_CREDIT_WAIT;
937 spin_unlock(&lp->lpni_lock);
940 if (!msg->msg_txcredit) {
941 LASSERT((tq->tq_credits < 0) ==
942 !list_empty(&tq->tq_delayed));
944 msg->msg_txcredit = 1;
946 atomic_dec(&ni->ni_tx_credits);
948 if (tq->tq_credits < tq->tq_credits_min)
949 tq->tq_credits_min = tq->tq_credits;
951 if (tq->tq_credits < 0) {
952 msg->msg_tx_delayed = 1;
953 list_add_tail(&msg->msg_list, &tq->tq_delayed);
954 return LNET_CREDIT_WAIT;
958 /* unset the tx_delay flag as we're going to send it now */
959 msg->msg_tx_delayed = 0;
962 lnet_net_unlock(cpt);
963 lnet_ni_send(ni, msg);
966 return LNET_CREDIT_OK;
970 static struct lnet_rtrbufpool *
971 lnet_msg2bufpool(struct lnet_msg *msg)
973 struct lnet_rtrbufpool *rbp;
976 LASSERT(msg->msg_rx_committed);
978 cpt = msg->msg_rx_cpt;
979 rbp = &the_lnet.ln_rtrpools[cpt][0];
981 LASSERT(msg->msg_len <= LNET_MTU);
982 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
984 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
991 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
993 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
994 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
995 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
996 * received or OK to receive */
997 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
998 struct lnet_peer *lp;
999 struct lnet_rtrbufpool *rbp;
1000 struct lnet_rtrbuf *rb;
1002 LASSERT(msg->msg_iov == NULL);
1003 LASSERT(msg->msg_kiov == NULL);
1004 LASSERT(msg->msg_niov == 0);
1005 LASSERT(msg->msg_routing);
1006 LASSERT(msg->msg_receiving);
1007 LASSERT(!msg->msg_sending);
1008 LASSERT(lpni->lpni_peer_net);
1009 LASSERT(lpni->lpni_peer_net->lpn_peer);
1011 lp = lpni->lpni_peer_net->lpn_peer;
1013 /* non-lnet_parse callers only receive delayed messages */
1014 LASSERT(!do_recv || msg->msg_rx_delayed);
1016 if (!msg->msg_peerrtrcredit) {
1017 /* lpni_lock protects the credit manipulation */
1018 spin_lock(&lpni->lpni_lock);
1020 msg->msg_peerrtrcredit = 1;
1021 lpni->lpni_rtrcredits--;
1022 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1023 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1025 if (lpni->lpni_rtrcredits < 0) {
1026 spin_unlock(&lpni->lpni_lock);
1027 /* must have checked eager_recv before here */
1028 LASSERT(msg->msg_rx_ready_delay);
1029 msg->msg_rx_delayed = 1;
1030 /* lp_lock protects the lp_rtrq */
1031 spin_lock(&lp->lp_lock);
1032 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1033 spin_unlock(&lp->lp_lock);
1034 return LNET_CREDIT_WAIT;
1036 spin_unlock(&lpni->lpni_lock);
1039 rbp = lnet_msg2bufpool(msg);
1041 if (!msg->msg_rtrcredit) {
1042 msg->msg_rtrcredit = 1;
1044 if (rbp->rbp_credits < rbp->rbp_mincredits)
1045 rbp->rbp_mincredits = rbp->rbp_credits;
1047 if (rbp->rbp_credits < 0) {
1048 /* must have checked eager_recv before here */
1049 LASSERT(msg->msg_rx_ready_delay);
1050 msg->msg_rx_delayed = 1;
1051 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1052 return LNET_CREDIT_WAIT;
1056 LASSERT(!list_empty(&rbp->rbp_bufs));
1057 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1058 list_del(&rb->rb_list);
1060 msg->msg_niov = rbp->rbp_npages;
1061 msg->msg_kiov = &rb->rb_kiov[0];
1063 /* unset the msg-rx_delayed flag since we're receiving the message */
1064 msg->msg_rx_delayed = 0;
1067 int cpt = msg->msg_rx_cpt;
1069 lnet_net_unlock(cpt);
1070 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1071 0, msg->msg_len, msg->msg_len);
1074 return LNET_CREDIT_OK;
1078 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1080 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1081 struct lnet_ni *txni = msg->msg_txni;
1082 struct lnet_msg *msg2;
1084 if (msg->msg_txcredit) {
1085 struct lnet_ni *ni = msg->msg_txni;
1086 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1088 /* give back NI txcredits */
1089 msg->msg_txcredit = 0;
1091 LASSERT((tq->tq_credits < 0) ==
1092 !list_empty(&tq->tq_delayed));
1095 atomic_inc(&ni->ni_tx_credits);
1096 if (tq->tq_credits <= 0) {
1097 msg2 = list_entry(tq->tq_delayed.next,
1098 struct lnet_msg, msg_list);
1099 list_del(&msg2->msg_list);
1101 LASSERT(msg2->msg_txni == ni);
1102 LASSERT(msg2->msg_tx_delayed);
1103 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1105 (void) lnet_post_send_locked(msg2, 1);
1109 if (msg->msg_peertxcredit) {
1110 /* give back peer txcredits */
1111 msg->msg_peertxcredit = 0;
1113 spin_lock(&txpeer->lpni_lock);
1114 LASSERT((txpeer->lpni_txcredits < 0) ==
1115 !list_empty(&txpeer->lpni_txq));
1117 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1118 LASSERT(txpeer->lpni_txqnob >= 0);
1120 txpeer->lpni_txcredits++;
1121 if (txpeer->lpni_txcredits <= 0) {
1124 msg2 = list_entry(txpeer->lpni_txq.next,
1125 struct lnet_msg, msg_list);
1126 list_del(&msg2->msg_list);
1127 spin_unlock(&txpeer->lpni_lock);
1129 LASSERT(msg2->msg_txpeer == txpeer);
1130 LASSERT(msg2->msg_tx_delayed);
1132 msg2_cpt = msg2->msg_tx_cpt;
1135 * The msg_cpt can be different from the msg2_cpt
1136 * so we need to make sure we lock the correct cpt
1138 * Once we call lnet_post_send_locked() it is no
1139 * longer safe to access msg2, since it could've
1140 * been freed by lnet_finalize(), but we still
1141 * need to relock the correct cpt, so we cache the
1142 * msg2_cpt for the purpose of the check that
1143 * follows the call to lnet_pose_send_locked().
1145 if (msg2_cpt != msg->msg_tx_cpt) {
1146 lnet_net_unlock(msg->msg_tx_cpt);
1147 lnet_net_lock(msg2_cpt);
1149 (void) lnet_post_send_locked(msg2, 1);
1150 if (msg2_cpt != msg->msg_tx_cpt) {
1151 lnet_net_unlock(msg2_cpt);
1152 lnet_net_lock(msg->msg_tx_cpt);
1155 spin_unlock(&txpeer->lpni_lock);
1160 msg->msg_txni = NULL;
1161 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1164 if (txpeer != NULL) {
1165 msg->msg_txpeer = NULL;
1166 lnet_peer_ni_decref_locked(txpeer);
1171 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1173 struct lnet_msg *msg;
1175 if (list_empty(&rbp->rbp_msgs))
1177 msg = list_entry(rbp->rbp_msgs.next,
1178 struct lnet_msg, msg_list);
1179 list_del(&msg->msg_list);
1181 (void)lnet_post_routed_recv_locked(msg, 1);
1185 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1187 struct lnet_msg *msg;
1188 struct lnet_msg *tmp;
1190 lnet_net_unlock(cpt);
1192 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1193 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1194 0, 0, 0, msg->msg_hdr.payload_length);
1195 list_del_init(&msg->msg_list);
1196 msg->msg_no_resend = true;
1197 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1198 lnet_finalize(msg, -ECANCELED);
1205 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1207 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1208 struct lnet_peer *lp;
1209 struct lnet_ni *rxni = msg->msg_rxni;
1210 struct lnet_msg *msg2;
1212 if (msg->msg_rtrcredit) {
1213 /* give back global router credits */
1214 struct lnet_rtrbuf *rb;
1215 struct lnet_rtrbufpool *rbp;
1217 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1218 * there until it gets one allocated, or aborts the wait
1220 LASSERT(msg->msg_kiov != NULL);
1222 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1225 msg->msg_kiov = NULL;
1226 msg->msg_rtrcredit = 0;
1228 LASSERT(rbp == lnet_msg2bufpool(msg));
1230 LASSERT((rbp->rbp_credits > 0) ==
1231 !list_empty(&rbp->rbp_bufs));
1233 /* If routing is now turned off, we just drop this buffer and
1234 * don't bother trying to return credits. */
1235 if (!the_lnet.ln_routing) {
1236 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1240 /* It is possible that a user has lowered the desired number of
1241 * buffers in this pool. Make sure we never put back
1242 * more buffers than the stated number. */
1243 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1244 /* Discard this buffer so we don't have too
1246 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1247 rbp->rbp_nbuffers--;
1249 list_add(&rb->rb_list, &rbp->rbp_bufs);
1251 if (rbp->rbp_credits <= 0)
1252 lnet_schedule_blocked_locked(rbp);
1257 if (msg->msg_peerrtrcredit) {
1259 LASSERT(rxpeerni->lpni_peer_net);
1260 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1262 /* give back peer router credits */
1263 msg->msg_peerrtrcredit = 0;
1265 spin_lock(&rxpeerni->lpni_lock);
1266 rxpeerni->lpni_rtrcredits++;
1267 spin_unlock(&rxpeerni->lpni_lock);
1269 lp = rxpeerni->lpni_peer_net->lpn_peer;
1270 spin_lock(&lp->lp_lock);
1272 /* drop all messages which are queued to be routed on that
1274 if (!the_lnet.ln_routing) {
1276 list_splice_init(&lp->lp_rtrq, &drop);
1277 spin_unlock(&lp->lp_lock);
1278 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1279 } else if (!list_empty(&lp->lp_rtrq)) {
1282 msg2 = list_entry(lp->lp_rtrq.next,
1283 struct lnet_msg, msg_list);
1284 list_del(&msg2->msg_list);
1285 msg2_cpt = msg2->msg_rx_cpt;
1286 spin_unlock(&lp->lp_lock);
1288 * messages on the lp_rtrq can be from any NID in
1289 * the peer, which means they might have different
1290 * cpts. We need to make sure we lock the right
1293 if (msg2_cpt != msg->msg_rx_cpt) {
1294 lnet_net_unlock(msg->msg_rx_cpt);
1295 lnet_net_lock(msg2_cpt);
1297 (void) lnet_post_routed_recv_locked(msg2, 1);
1298 if (msg2_cpt != msg->msg_rx_cpt) {
1299 lnet_net_unlock(msg2_cpt);
1300 lnet_net_lock(msg->msg_rx_cpt);
1303 spin_unlock(&lp->lp_lock);
1307 msg->msg_rxni = NULL;
1308 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1310 if (rxpeerni != NULL) {
1311 msg->msg_rxpeer = NULL;
1312 lnet_peer_ni_decref_locked(rxpeerni);
1317 lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1319 if (p1->lpni_txqnob < p2->lpni_txqnob)
1322 if (p1->lpni_txqnob > p2->lpni_txqnob)
1325 if (p1->lpni_txcredits > p2->lpni_txcredits)
1328 if (p1->lpni_txcredits < p2->lpni_txcredits)
1334 static struct lnet_peer_ni *
1335 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1336 struct lnet_peer *peer,
1337 struct lnet_peer_net *peer_net)
1340 * Look at the peer NIs for the destination peer that connect
1341 * to the chosen net. If a peer_ni is preferred when using the
1342 * best_ni to communicate, we use that one. If there is no
1343 * preferred peer_ni, or there are multiple preferred peer_ni,
1344 * the available transmit credits are used. If the transmit
1345 * credits are equal, we round-robin over the peer_ni.
1347 struct lnet_peer_ni *lpni = NULL;
1348 struct lnet_peer_ni *best_lpni = NULL;
1349 int best_lpni_credits = INT_MIN;
1350 bool preferred = false;
1352 int best_lpni_healthv = 0;
1355 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1357 * if the best_ni we've chosen aleady has this lpni
1358 * preferred, then let's use it
1361 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1363 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1364 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1369 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1372 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1373 libcfs_nid2str(lpni->lpni_nid),
1374 lpni->lpni_txcredits, best_lpni_credits,
1375 lpni->lpni_seq, best_lpni->lpni_seq);
1377 /* pick the healthiest peer ni */
1378 if (lpni_healthv < best_lpni_healthv) {
1380 } else if (lpni_healthv > best_lpni_healthv) {
1381 best_lpni_healthv = lpni_healthv;
1382 /* if this is a preferred peer use it */
1383 } else if (!preferred && ni_is_pref) {
1385 } else if (preferred && !ni_is_pref) {
1387 * this is not the preferred peer so let's ignore
1391 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1393 * We already have a peer that has more credits
1394 * available than this one. No need to consider
1395 * this peer further.
1398 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1400 * The best peer found so far and the current peer
1401 * have the same number of available credits let's
1402 * make sure to select between them using Round
1406 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1412 best_lpni_credits = lpni->lpni_txcredits;
1415 /* if we still can't find a peer ni then we can't reach it */
1417 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1418 LNET_NIDNET(dst_nid);
1419 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1420 libcfs_net2str(net_id));
1424 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1425 libcfs_nid2str(best_lpni->lpni_nid));
1431 * Prerequisite: the best_ni should already be set in the sd
1433 static inline struct lnet_peer_ni *
1434 lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid,
1435 struct lnet_peer *peer, __u32 net_id)
1437 struct lnet_peer_net *peer_net;
1440 * The gateway is Multi-Rail capable so now we must select the
1443 peer_net = lnet_peer_get_net_locked(peer, net_id);
1445 return lnet_select_peer_ni(lni, dst_nid, peer, peer_net);
1448 /* Compare route priorities and hop counts */
1450 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1452 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1453 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1455 if (r1->lr_priority < r2->lr_priority)
1458 if (r1->lr_priority > r2->lr_priority)
1461 if (r1_hops < r2_hops)
1464 if (r1_hops > r2_hops)
1470 static struct lnet_route *
1471 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1472 struct lnet_route **prev_route,
1473 struct lnet_peer_ni **gwni)
1475 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1476 struct lnet_route *best_route;
1477 struct lnet_route *last_route;
1478 struct lnet_route *route;
1481 best_route = last_route = NULL;
1482 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1483 if (!lnet_is_route_alive(route))
1487 * Restrict the selection of the router NI on the src_net
1488 * provided. If the src_net is LNET_NID_ANY, then select
1489 * the best interface available.
1492 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1496 best_route = last_route = route;
1499 CERROR("Gateway %s does not have a peer NI on net %s\n",
1500 libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1501 libcfs_net2str(src_net));
1506 /* no protection on below fields, but it's harmless */
1507 if (last_route->lr_seq - route->lr_seq < 0)
1510 rc = lnet_compare_routes(route, best_route);
1514 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1518 CERROR("Gateway %s does not have a peer NI on net %s\n",
1519 libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1520 libcfs_net2str(src_net));
1530 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1534 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1541 *prev_route = last_route;
1547 static struct lnet_ni *
1548 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1549 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1552 struct lnet_ni *ni = NULL;
1553 unsigned int shortest_distance;
1558 * If there is no peer_ni that we can send to on this network,
1559 * then there is no point in looking for a new best_ni here.
1561 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1564 if (best_ni == NULL) {
1565 shortest_distance = UINT_MAX;
1566 best_credits = INT_MIN;
1569 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1570 best_ni->ni_dev_cpt);
1571 best_credits = atomic_read(&best_ni->ni_tx_credits);
1572 best_healthv = atomic_read(&best_ni->ni_healthv);
1575 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1576 unsigned int distance;
1581 ni_credits = atomic_read(&ni->ni_tx_credits);
1582 ni_healthv = atomic_read(&ni->ni_healthv);
1583 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1586 * calculate the distance from the CPT on which
1587 * the message memory is allocated to the CPT of
1588 * the NI's physical device
1590 distance = cfs_cpt_distance(lnet_cpt_table(),
1594 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1595 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1596 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1597 : "not seleced", best_credits, shortest_distance,
1598 (best_ni) ? best_ni->ni_seq : 0);
1601 * All distances smaller than the NUMA range
1602 * are treated equally.
1604 if (distance < lnet_numa_range)
1605 distance = lnet_numa_range;
1608 * Select on health, shorter distance, available
1609 * credits, then round-robin.
1613 } else if (ni_healthv < best_healthv) {
1615 } else if (ni_healthv > best_healthv) {
1616 best_healthv = ni_healthv;
1618 * If we're going to prefer this ni because it's
1619 * the healthiest, then we should set the
1620 * shortest_distance in the algorithm in case
1621 * there are multiple NIs with the same health but
1622 * different distances.
1624 if (distance < shortest_distance)
1625 shortest_distance = distance;
1626 } else if (distance > shortest_distance) {
1628 } else if (distance < shortest_distance) {
1629 shortest_distance = distance;
1630 } else if (ni_credits < best_credits) {
1632 } else if (ni_credits == best_credits) {
1633 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1637 best_credits = ni_credits;
1640 CDEBUG(D_NET, "selected best_ni %s\n",
1641 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1647 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1648 * because such traffic is required to perform discovery. We therefore
1649 * exclude all GET and PUT on that portal. We also exclude all ACK and
1650 * REPLY traffic, but that is because the portal is not tracked in the
1651 * message structure for these message types. We could restrict this
1652 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1655 lnet_msg_discovery(struct lnet_msg *msg)
1657 if (msg->msg_type == LNET_MSG_PUT) {
1658 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1660 } else if (msg->msg_type == LNET_MSG_GET) {
1661 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1667 #define SRC_SPEC 0x0001
1668 #define SRC_ANY 0x0002
1669 #define LOCAL_DST 0x0004
1670 #define REMOTE_DST 0x0008
1671 #define MR_DST 0x0010
1672 #define NMR_DST 0x0020
1673 #define SND_RESP 0x0040
1675 /* The following to defines are used for return codes */
1676 #define REPEAT_SEND 0x1000
1677 #define PASS_THROUGH 0x2000
1679 /* The different cases lnet_select pathway needs to handle */
1680 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1681 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1682 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1683 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1684 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1685 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1686 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1687 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1690 lnet_handle_lo_send(struct lnet_send_data *sd)
1692 struct lnet_msg *msg = sd->sd_msg;
1693 int cpt = sd->sd_cpt;
1695 /* No send credit hassles with LOLND */
1696 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1697 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1698 if (!msg->msg_routing)
1699 msg->msg_hdr.src_nid =
1700 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1701 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1702 lnet_msg_commit(msg, cpt);
1703 msg->msg_txni = the_lnet.ln_loni;
1705 return LNET_CREDIT_OK;
1709 lnet_handle_send(struct lnet_send_data *sd)
1711 struct lnet_ni *best_ni = sd->sd_best_ni;
1712 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1713 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1714 struct lnet_msg *msg = sd->sd_msg;
1716 __u32 send_case = sd->sd_send_case;
1718 __u32 routing = send_case & REMOTE_DST;
1719 struct lnet_rsp_tracker *rspt;
1722 * Increment sequence number of the selected peer so that we
1723 * pick the next one in Round Robin.
1725 best_lpni->lpni_seq++;
1728 * grab a reference on the peer_ni so it sticks around even if
1729 * we need to drop and relock the lnet_net_lock below.
1731 lnet_peer_ni_addref_locked(best_lpni);
1734 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1735 * message. This ensures that we get a CPT that is correct for
1736 * the NI when the NI has been restricted to a subset of all CPTs.
1737 * If the selected CPT differs from the one currently locked, we
1738 * must unlock and relock the lnet_net_lock(), and then check whether
1739 * the configuration has changed. We don't have a hold on the best_ni
1740 * yet, and it may have vanished.
1742 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1743 if (sd->sd_cpt != cpt2) {
1744 __u32 seq = lnet_get_dlc_seq_locked();
1745 lnet_net_unlock(sd->sd_cpt);
1747 lnet_net_lock(sd->sd_cpt);
1748 if (seq != lnet_get_dlc_seq_locked()) {
1749 lnet_peer_ni_decref_locked(best_lpni);
1755 * store the best_lpni in the message right away to avoid having
1756 * to do the same operation under different conditions
1758 msg->msg_txpeer = best_lpni;
1759 msg->msg_txni = best_ni;
1762 * grab a reference for the best_ni since now it's in use in this
1763 * send. The reference will be dropped in lnet_finalize()
1765 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1768 * Always set the target.nid to the best peer picked. Either the
1769 * NID will be one of the peer NIDs selected, or the same NID as
1770 * what was originally set in the target or it will be the NID of
1771 * a router if this message should be routed
1773 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1776 * lnet_msg_commit assigns the correct cpt to the message, which
1777 * is used to decrement the correct refcount on the ni when it's
1778 * time to return the credits
1780 lnet_msg_commit(msg, sd->sd_cpt);
1783 * If we are routing the message then we keep the src_nid that was
1784 * set by the originator. If we are not routing then we are the
1785 * originator and set it here.
1787 if (!msg->msg_routing)
1788 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1791 msg->msg_target_is_router = 1;
1792 msg->msg_target.pid = LNET_PID_LUSTRE;
1794 * since we're routing we want to ensure that the
1795 * msg_hdr.dest_nid is set to the final destination. When
1796 * the router receives this message it knows how to route
1799 * final_dst_lpni is set at the beginning of the
1800 * lnet_select_pathway() function and is never changed.
1801 * It's safe to use it here.
1803 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1806 * if we're not routing set the dest_nid to the best peer
1807 * ni NID that we picked earlier in the algorithm.
1809 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1813 * if we have response tracker block update it with the next hop
1817 rspt = msg->msg_md->md_rspt_ptr;
1819 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1820 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1821 libcfs_nid2str(rspt->rspt_next_hop_nid));
1825 rc = lnet_post_send_locked(msg, 0);
1828 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1829 libcfs_nid2str(msg->msg_hdr.src_nid),
1830 libcfs_nid2str(msg->msg_txni->ni_nid),
1831 libcfs_nid2str(sd->sd_src_nid),
1832 libcfs_nid2str(msg->msg_hdr.dest_nid),
1833 libcfs_nid2str(sd->sd_dst_nid),
1834 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1835 libcfs_nid2str(sd->sd_rtr_nid),
1836 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1842 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1843 struct lnet_msg *msg)
1845 if (!lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1846 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1847 libcfs_nid2str(lni->ni_nid),
1848 libcfs_nid2str(lpni->lpni_nid));
1849 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1858 * use the source and destination NIDs as the pathway
1861 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1863 /* the destination lpni is set before we get here. */
1866 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1867 if (!sd->sd_best_ni) {
1868 CERROR("Can't send to %s: src %s is not a "
1869 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1870 libcfs_nid2str(sd->sd_src_nid));
1874 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
1876 return lnet_handle_send(sd);
1884 * Don't run the selection algorithm on the peer NIs. By specifying the
1885 * local NID, we're also saying that we should always use the destination NID
1886 * provided. This handles the case where we should be using the same
1887 * destination NID for the all the messages which belong to the same RPC
1891 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1893 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1894 if (!sd->sd_best_ni) {
1895 CERROR("Can't send to %s: src %s is not a "
1896 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1897 libcfs_nid2str(sd->sd_src_nid));
1901 if (sd->sd_best_lpni &&
1902 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1903 return lnet_handle_lo_send(sd);
1904 else if (sd->sd_best_lpni)
1905 return lnet_handle_send(sd);
1907 CERROR("can't send to %s. no NI on %s\n",
1908 libcfs_nid2str(sd->sd_dst_nid),
1909 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1911 return -EHOSTUNREACH;
1915 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1916 struct lnet_peer *peer,
1917 struct lnet_peer_net *peer_net,
1921 struct lnet_net *local_net;
1922 struct lnet_ni *best_ni;
1924 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1929 * Iterate through the NIs in this local Net and select
1930 * the NI to send from. The selection is determined by
1931 * these 3 criterion in the following priority:
1933 * 2. NI available credits
1936 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1937 peer, peer_net, cpt);
1939 if (incr_seq && best_ni)
1946 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
1949 struct lnet_peer *peer;
1952 lnet_peer_ni_addref_locked(lpni);
1954 peer = lpni->lpni_peer_net->lpn_peer;
1956 if (lnet_peer_gw_discovery(peer)) {
1957 lnet_peer_ni_decref_locked(lpni);
1961 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1962 lnet_peer_ni_decref_locked(lpni);
1966 rc = lnet_discover_peer_locked(lpni, cpt, false);
1968 lnet_peer_ni_decref_locked(lpni);
1971 /* The peer may have changed. */
1972 peer = lpni->lpni_peer_net->lpn_peer;
1973 spin_lock(&peer->lp_lock);
1974 if (lnet_peer_is_uptodate_locked(peer)) {
1975 spin_unlock(&peer->lp_lock);
1976 lnet_peer_ni_decref_locked(lpni);
1979 /* queue message and return */
1980 msg->msg_sending = 0;
1981 msg->msg_txpeer = NULL;
1982 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1983 spin_unlock(&peer->lp_lock);
1985 lnet_peer_ni_decref_locked(lpni);
1987 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
1988 msg, libcfs_nid2str(peer->lp_primary_nid));
1990 return LNET_DC_WAIT;
1994 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1996 struct lnet_peer_ni **gw_lpni,
1997 struct lnet_peer **gw_peer)
2001 struct lnet_peer *gw;
2002 struct lnet_peer *lp;
2003 struct lnet_peer_net *lpn;
2004 struct lnet_peer_net *best_lpn = NULL;
2005 struct lnet_remotenet *rnet, *best_rnet = NULL;
2006 struct lnet_route *best_route = NULL;
2007 struct lnet_route *last_route = NULL;
2008 struct lnet_peer_ni *lpni = NULL;
2009 struct lnet_peer_ni *gwni = NULL;
2010 lnet_nid_t src_nid = sd->sd_src_nid;
2012 /* If a router nid was specified then we are replying to a GET or
2013 * sending an ACK. In this case we use the gateway associated with the
2014 * specified router nid.
2016 if (sd->sd_rtr_nid != LNET_NID_ANY) {
2017 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2019 CERROR("No peer NI for gateway %s\n",
2020 libcfs_nid2str(sd->sd_rtr_nid));
2021 return -EHOSTUNREACH;
2023 gw = gwni->lpni_peer_net->lpn_peer;
2024 lnet_peer_ni_decref_locked(gwni);
2025 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2027 /* we've already looked up the initial lpni using dst_nid */
2028 lpni = sd->sd_best_lpni;
2029 /* the peer tree must be in existence */
2030 LASSERT(lpni && lpni->lpni_peer_net &&
2031 lpni->lpni_peer_net->lpn_peer);
2032 lp = lpni->lpni_peer_net->lpn_peer;
2034 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2035 /* is this remote network reachable? */
2036 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2045 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2053 CERROR("peer %s has no available nets\n",
2054 libcfs_nid2str(sd->sd_dst_nid));
2055 return -EHOSTUNREACH;
2058 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni,
2061 best_lpn->lpn_net_id);
2062 if (!sd->sd_best_lpni) {
2063 CERROR("peer %s down\n",
2064 libcfs_nid2str(sd->sd_dst_nid));
2065 return -EHOSTUNREACH;
2069 * We're attempting to round robin over the remote peer
2070 * NI's so update the final destination we selected
2072 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2075 * find the best route. Restrict the selection on the net of the
2076 * local NI if we've already picked the local NI to send from.
2077 * Otherwise, let's pick any route we can find and then find
2078 * a local NI we can reach the route's gateway on. Any route we select
2079 * will be reachable by virtue of the restriction we have when
2082 best_route = lnet_find_route_locked(best_rnet,
2083 LNET_NIDNET(src_nid),
2084 &last_route, &gwni);
2087 CERROR("no route to %s from %s\n",
2088 libcfs_nid2str(dst_nid),
2089 libcfs_nid2str(src_nid));
2090 return -EHOSTUNREACH;
2094 CERROR("Internal Error. Route expected to %s from %s\n",
2095 libcfs_nid2str(dst_nid),
2096 libcfs_nid2str(src_nid));
2100 gw = best_route->lr_gateway;
2101 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2102 local_lnet = best_route->lr_lnet;
2105 * Increment the sequence number of the remote lpni so we
2106 * can round robin over the different interfaces of the
2109 sd->sd_best_lpni->lpni_seq++;
2113 * Discover this gateway if it hasn't already been discovered.
2114 * This means we might delay the message until discovery has
2117 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2118 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2122 if (!sd->sd_best_ni)
2123 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2124 lnet_peer_get_net_locked(gw,
2129 if (!sd->sd_best_ni) {
2130 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2131 libcfs_net2str(local_lnet),
2132 libcfs_nid2str(sd->sd_src_nid));
2140 * increment the sequence numbers since now we're sure we're
2141 * going to use this path
2143 if (sd->sd_rtr_nid == LNET_NID_ANY) {
2144 LASSERT(best_route && last_route);
2145 best_route->lr_seq = last_route->lr_seq + 1;
2146 best_lpn->lpn_seq++;
2157 * Remote destination
2158 * Non-MR destination
2162 * Remote destination
2165 * The handling of these two cases is similar. Even though the destination
2166 * can be MR or non-MR, we'll deal directly with the router.
2169 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2172 struct lnet_peer_ni *gw_lpni = NULL;
2173 struct lnet_peer *gw_peer = NULL;
2176 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2177 if (!sd->sd_best_ni) {
2178 CERROR("Can't send to %s: src %s is not a "
2179 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2180 libcfs_nid2str(sd->sd_src_nid));
2184 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2189 if (sd->sd_send_case & NMR_DST)
2191 * since the final destination is non-MR let's set its preferred
2192 * NID before we send
2194 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2198 * We're going to send to the gw found so let's set its
2201 sd->sd_peer = gw_peer;
2202 sd->sd_best_lpni = gw_lpni;
2204 return lnet_handle_send(sd);
2208 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2211 struct lnet_peer_net *peer_net = NULL;
2212 struct lnet_ni *best_ni = NULL;
2213 int lpn_healthv = 0;
2216 * The peer can have multiple interfaces, some of them can be on
2217 * the local network and others on a routed network. We should
2218 * prefer the local network. However if the local network is not
2219 * available then we need to try the routed network
2222 /* go through all the peer nets and find the best_ni */
2223 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2225 * The peer's list of nets can contain non-local nets. We
2226 * want to only examine the local ones.
2228 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2231 /* always select the lpn with the best health */
2232 if (lpn_healthv <= peer_net->lpn_healthv)
2233 lpn_healthv = peer_net->lpn_healthv;
2237 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer, peer_net,
2241 * if this is a discovery message and lp_disc_net_id is
2242 * specified then use that net to send the discovery on.
2244 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2250 /* increment sequence number so we can round robin */
2256 static struct lnet_ni *
2257 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2259 struct lnet_ni *best_ni = NULL;
2260 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2261 struct lnet_peer_ni *lpni_entry;
2264 * We must use a consistent source address when sending to a
2265 * non-MR peer. However, a non-MR peer can have multiple NIDs
2266 * on multiple networks, and we may even need to talk to this
2267 * peer on multiple networks -- certain types of
2268 * load-balancing configuration do this.
2270 * So we need to pick the NI the peer prefers for this
2271 * particular network.
2274 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2276 if (lpni_entry->lpni_pref_nnids == 0)
2278 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2279 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2286 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2288 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2290 struct lnet_ni *best_ni = NULL;
2291 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2294 * We must use a consistent source address when sending to a
2295 * non-MR peer. However, a non-MR peer can have multiple NIDs
2296 * on multiple networks, and we may even need to talk to this
2297 * peer on multiple networks -- certain types of
2298 * load-balancing configuration do this.
2300 * So we need to pick the NI the peer prefers for this
2301 * particular network.
2304 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2307 /* if best_ni is still not set just pick one */
2310 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2311 sd->sd_best_lpni->lpni_peer_net,
2312 sd->sd_md_cpt, true);
2313 /* If there is no best_ni we don't have a route */
2315 CERROR("no path to %s from net %s\n",
2316 libcfs_nid2str(best_lpni->lpni_nid),
2317 libcfs_net2str(best_lpni->lpni_net->net_id));
2318 return -EHOSTUNREACH;
2322 sd->sd_best_ni = best_ni;
2324 /* Set preferred NI if necessary. */
2325 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2332 * Source not specified
2336 * always use the same source NID for NMR peers
2337 * If we've talked to that peer before then we already have a preferred
2338 * source NI associated with it. Otherwise, we select a preferred local NI
2339 * and store it in the peer
2342 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2346 /* sd->sd_best_lpni is already set to the final destination */
2349 * At this point we should've created the peer ni and peer. If we
2350 * can't find it, then something went wrong. Instead of assert
2351 * output a relevant message and fail the send
2353 if (!sd->sd_best_lpni) {
2354 CERROR("Internal fault. Unable to send msg %s to %s. "
2356 lnet_msgtyp2str(sd->sd_msg->msg_type),
2357 libcfs_nid2str(sd->sd_dst_nid));
2361 rc = lnet_select_preferred_best_ni(sd);
2363 rc = lnet_handle_send(sd);
2369 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2372 * NOTE we've already handled the remote peer case. So we only
2373 * need to worry about the local case here.
2375 * if we're sending a response, ACK or reply, we need to send it
2376 * to the destination NID given to us. At this point we already
2377 * have the peer_ni we're suppose to send to, so just find the
2378 * best_ni on the peer net and use that. Since we're sending to an
2379 * MR peer then we can just run the selection algorithm on our
2380 * local NIs and pick the best one.
2382 if (sd->sd_send_case & SND_RESP) {
2384 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2385 sd->sd_best_lpni->lpni_peer_net,
2386 sd->sd_md_cpt, true);
2388 if (!sd->sd_best_ni) {
2390 * We're not going to deal with not able to send
2391 * a response to the provided final destination
2393 CERROR("Can't send response to %s. "
2394 "No local NI available\n",
2395 libcfs_nid2str(sd->sd_dst_nid));
2396 return -EHOSTUNREACH;
2399 return lnet_handle_send(sd);
2403 * If we get here that means we're sending a fresh request, PUT or
2404 * GET, so we need to run our standard selection algorithm.
2405 * First find the best local interface that's on any of the peer's
2408 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2410 lnet_msg_discovery(sd->sd_msg));
2411 if (sd->sd_best_ni) {
2413 lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid,
2415 sd->sd_best_ni->ni_net->net_id);
2418 * if we're successful in selecting a peer_ni on the local
2419 * network, then send to it. Otherwise fall through and
2420 * try and see if we can reach it over another routed
2423 if (sd->sd_best_lpni &&
2424 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2426 * in case we initially started with a routed
2427 * destination, let's reset to local
2429 sd->sd_send_case &= ~REMOTE_DST;
2430 sd->sd_send_case |= LOCAL_DST;
2431 return lnet_handle_lo_send(sd);
2432 } else if (sd->sd_best_lpni) {
2434 * in case we initially started with a routed
2435 * destination, let's reset to local
2437 sd->sd_send_case &= ~REMOTE_DST;
2438 sd->sd_send_case |= LOCAL_DST;
2439 return lnet_handle_send(sd);
2442 CERROR("Internal Error. Expected to have a best_lpni: "
2444 libcfs_nid2str(sd->sd_src_nid),
2445 libcfs_nid2str(sd->sd_dst_nid));
2451 * Peer doesn't have a local network. Let's see if there is
2452 * a remote network we can reach it on.
2454 return PASS_THROUGH;
2459 * Source NID not specified
2464 * Source NID not speified
2465 * Remote destination
2468 * In both of these cases if we're sending a response, ACK or REPLY, then
2469 * we need to send to the destination NID provided.
2471 * In the remote case let's deal with MR routers.
2476 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2479 struct lnet_peer *gw_peer = NULL;
2480 struct lnet_peer_ni *gw_lpni = NULL;
2483 * handle sending a response to a remote peer here so we don't
2484 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2486 if (sd->sd_send_case & REMOTE_DST &&
2487 sd->sd_send_case & SND_RESP) {
2488 struct lnet_peer_ni *gw;
2489 struct lnet_peer *gw_peer;
2491 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2494 CERROR("Can't send response to %s. "
2495 "No route available\n",
2496 libcfs_nid2str(sd->sd_dst_nid));
2497 return -EHOSTUNREACH;
2498 } else if (rc > 0) {
2502 sd->sd_best_lpni = gw;
2503 sd->sd_peer = gw_peer;
2505 return lnet_handle_send(sd);
2509 * Even though the NID for the peer might not be on a local network,
2510 * since the peer is MR there could be other interfaces on the
2511 * local network. In that case we'd still like to prefer the local
2512 * network over the routed network. If we're unable to do that
2513 * then we select the best router among the different routed networks,
2514 * and if the router is MR then we can deal with it as such.
2516 rc = lnet_handle_any_mr_dsta(sd);
2517 if (rc != PASS_THROUGH)
2521 * Now that we must route to the destination, we must consider the
2522 * MR case, where the destination has multiple interfaces, some of
2523 * which we can route to and others we do not. For this reason we
2524 * need to select the destination which we can route to and if
2525 * there are multiple, we need to round robin.
2527 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2532 sd->sd_send_case &= ~LOCAL_DST;
2533 sd->sd_send_case |= REMOTE_DST;
2535 sd->sd_peer = gw_peer;
2536 sd->sd_best_lpni = gw_lpni;
2538 return lnet_handle_send(sd);
2542 * Source not specified
2543 * Remote destination
2546 * Must send to the specified peer NID using the same source NID that
2547 * we've used before. If it's the first time to talk to that peer then
2548 * find the source NI and assign it as preferred to that peer
2551 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2554 struct lnet_peer_ni *gw_lpni = NULL;
2555 struct lnet_peer *gw_peer = NULL;
2558 * Let's see if we have a preferred NI to talk to this NMR peer
2560 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2564 * find the router and that'll find the best NI if we didn't find
2567 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2573 * set the best_ni we've chosen as the preferred one for
2576 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2578 /* we'll be sending to the gw */
2579 sd->sd_best_lpni = gw_lpni;
2580 sd->sd_peer = gw_peer;
2582 return lnet_handle_send(sd);
2586 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2589 * turn off the SND_RESP bit.
2590 * It will be checked in the case handling
2592 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2594 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2595 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2596 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2597 (send_case & MR_DST) ? "MR: " : "NMR: ",
2598 libcfs_nid2str(sd->sd_dst_nid),
2599 (send_case & LOCAL_DST) ? "local" : "routed");
2601 switch (send_case) {
2603 * For all cases where the source is specified, we should always
2604 * use the destination NID, whether it's an MR destination or not,
2605 * since we're continuing a series of related messages for the
2608 case SRC_SPEC_LOCAL_NMR_DST:
2609 return lnet_handle_spec_local_nmr_dst(sd);
2610 case SRC_SPEC_LOCAL_MR_DST:
2611 return lnet_handle_spec_local_mr_dst(sd);
2612 case SRC_SPEC_ROUTER_NMR_DST:
2613 case SRC_SPEC_ROUTER_MR_DST:
2614 return lnet_handle_spec_router_dst(sd);
2615 case SRC_ANY_LOCAL_NMR_DST:
2616 return lnet_handle_any_local_nmr_dst(sd);
2617 case SRC_ANY_LOCAL_MR_DST:
2618 case SRC_ANY_ROUTER_MR_DST:
2619 return lnet_handle_any_mr_dst(sd);
2620 case SRC_ANY_ROUTER_NMR_DST:
2621 return lnet_handle_any_router_nmr_dst(sd);
2623 CERROR("Unknown send case\n");
2629 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2630 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2632 struct lnet_peer_ni *lpni;
2633 struct lnet_peer *peer;
2634 struct lnet_send_data send_data;
2637 __u32 send_case = 0;
2639 memset(&send_data, 0, sizeof(send_data));
2642 * get an initial CPT to use for locking. The idea here is not to
2643 * serialize the calls to select_pathway, so that as many
2644 * operations can run concurrently as possible. To do that we use
2645 * the CPT where this call is being executed. Later on when we
2646 * determine the CPT to use in lnet_message_commit, we switch the
2647 * lock and check if there was any configuration change. If none,
2648 * then we proceed, if there is, then we restart the operation.
2650 cpt = lnet_net_lock_current();
2652 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2653 if (md_cpt == CFS_CPT_ANY)
2659 * If we're being asked to send to the loopback interface, there
2660 * is no need to go through any selection. We can just shortcut
2661 * the entire process and send over lolnd
2663 send_data.sd_msg = msg;
2664 send_data.sd_cpt = cpt;
2665 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2666 rc = lnet_handle_lo_send(&send_data);
2667 lnet_net_unlock(cpt);
2672 * find an existing peer_ni, or create one and mark it as having been
2673 * created due to network traffic. This call will create the
2674 * peer->peer_net->peer_ni tree.
2676 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2678 lnet_net_unlock(cpt);
2679 return PTR_ERR(lpni);
2683 * Cache the original src_nid and rtr_nid. If we need to resend the
2684 * message then we'll need to know whether the src_nid was originally
2685 * specified for this message. If it was originally specified,
2686 * then we need to keep using the same src_nid since it's
2687 * continuing the same sequence of messages. Similarly, rtr_nid will
2688 * affect our choice of next hop.
2690 msg->msg_src_nid_param = src_nid;
2691 msg->msg_rtr_nid_param = rtr_nid;
2694 * If necessary, perform discovery on the peer that owns this peer_ni.
2695 * Note, this can result in the ownership of this peer_ni changing
2696 * to another peer object.
2698 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2700 lnet_peer_ni_decref_locked(lpni);
2701 lnet_net_unlock(cpt);
2704 lnet_peer_ni_decref_locked(lpni);
2706 peer = lpni->lpni_peer_net->lpn_peer;
2709 * Identify the different send cases
2711 if (src_nid == LNET_NID_ANY)
2712 send_case |= SRC_ANY;
2714 send_case |= SRC_SPEC;
2716 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2717 send_case |= LOCAL_DST;
2719 send_case |= REMOTE_DST;
2722 * Deal with the peer as NMR in the following cases:
2723 * 1. the peer is NMR
2724 * 2. We're trying to recover a specific peer NI
2725 * 3. I'm a router sending to the final destination
2726 * In this case the source of the message would've
2727 * already selected the final destination so my job
2728 * is to honor the selection.
2730 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
2731 (msg->msg_routing && (send_case & LOCAL_DST)))
2732 send_case |= NMR_DST;
2734 send_case |= MR_DST;
2736 if (lnet_msg_is_response(msg))
2737 send_case |= SND_RESP;
2739 /* assign parameters to the send_data */
2740 send_data.sd_rtr_nid = rtr_nid;
2741 send_data.sd_src_nid = src_nid;
2742 send_data.sd_dst_nid = dst_nid;
2743 send_data.sd_best_lpni = lpni;
2745 * keep a pointer to the final destination in case we're going to
2746 * route, so we'll need to access it later
2748 send_data.sd_final_dst_lpni = lpni;
2749 send_data.sd_peer = peer;
2750 send_data.sd_md_cpt = md_cpt;
2751 send_data.sd_send_case = send_case;
2753 rc = lnet_handle_send_case_locked(&send_data);
2756 * Update the local cpt since send_data.sd_cpt might've been
2757 * updated as a result of calling lnet_handle_send_case_locked().
2759 cpt = send_data.sd_cpt;
2761 if (rc == REPEAT_SEND)
2764 lnet_net_unlock(cpt);
2770 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2772 lnet_nid_t dst_nid = msg->msg_target.nid;
2776 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2777 * but we might want to use pre-determined router for ACK/REPLY
2780 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2781 LASSERT(msg->msg_txpeer == NULL);
2782 LASSERT(msg->msg_txni == NULL);
2783 LASSERT(!msg->msg_sending);
2784 LASSERT(!msg->msg_target_is_router);
2785 LASSERT(!msg->msg_receiving);
2787 msg->msg_sending = 1;
2789 LASSERT(!msg->msg_tx_committed);
2791 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2793 if (rc == -EHOSTUNREACH)
2794 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2796 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2800 if (rc == LNET_CREDIT_OK)
2801 lnet_ni_send(msg->msg_txni, msg);
2803 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2807 enum lnet_mt_event_type {
2808 MT_TYPE_LOCAL_NI = 0,
2812 struct lnet_mt_event_info {
2813 enum lnet_mt_event_type mt_type;
2817 /* called with res_lock held */
2819 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2821 struct lnet_rsp_tracker *rspt;
2824 * msg has a refcount on the MD so the MD is not going away.
2825 * The rspt queue for the cpt is protected by
2826 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2828 if (!md->md_rspt_ptr)
2831 rspt = md->md_rspt_ptr;
2834 LASSERT(rspt->rspt_cpt == cpt);
2836 md->md_rspt_ptr = NULL;
2838 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2840 * The monitor thread has invalidated this handle because the
2841 * response timed out, but it failed to lookup the MD. That
2842 * means this response tracker is on the zombie list. We can
2843 * safely remove it under the resource lock (held by caller) and
2844 * free the response tracker block.
2846 list_del(&rspt->rspt_on_list);
2847 lnet_rspt_free(rspt, cpt);
2850 * invalidate the handle to indicate that a response has been
2851 * received, which will then lead the monitor thread to clean up
2854 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2859 lnet_clean_zombie_rstqs(void)
2861 struct lnet_rsp_tracker *rspt, *tmp;
2864 cfs_cpt_for_each(i, lnet_cpt_table()) {
2865 list_for_each_entry_safe(rspt, tmp,
2866 the_lnet.ln_mt_zombie_rstqs[i],
2868 list_del(&rspt->rspt_on_list);
2869 lnet_rspt_free(rspt, i);
2873 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2877 lnet_finalize_expired_responses(void)
2879 struct lnet_libmd *md;
2880 struct lnet_rsp_tracker *rspt, *tmp;
2884 if (the_lnet.ln_mt_rstq == NULL)
2887 cfs_cpt_for_each(i, lnet_cpt_table()) {
2888 LIST_HEAD(local_queue);
2891 if (!the_lnet.ln_mt_rstq[i]) {
2895 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2900 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2902 * The rspt mdh will be invalidated when a response
2903 * is received or whenever we want to discard the
2904 * block the monitor thread will walk the queue
2905 * and clean up any rsts with an invalid mdh.
2906 * The monitor thread will walk the queue until
2907 * the first unexpired rspt block. This means that
2908 * some rspt blocks which received their
2909 * corresponding responses will linger in the
2910 * queue until they are cleaned up eventually.
2913 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2915 list_del(&rspt->rspt_on_list);
2916 lnet_rspt_free(rspt, i);
2920 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2921 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2922 struct lnet_peer_ni *lpni;
2925 md = lnet_handle2md(&rspt->rspt_mdh);
2927 /* MD has been queued for unlink, but
2928 * rspt hasn't been detached (Note we've
2929 * checked above that the rspt_mdh is
2930 * valid). Since we cannot lookup the MD
2931 * we're unable to detach the rspt
2932 * ourselves. Thus, move the rspt to the
2933 * zombie list where we'll wait for
2935 * 1. The remaining operations on the
2936 * MD to complete. In this case the
2937 * final operation will result in
2938 * lnet_msg_detach_md()->
2939 * lnet_detach_rsp_tracker() where
2940 * we will clean up this response
2942 * 2. LNet to shutdown. In this case
2943 * we'll wait until after all LND Nets
2944 * have shutdown and then we can
2945 * safely free any remaining response
2946 * tracker blocks on the zombie list.
2947 * Note: We need to hold the resource
2948 * lock when adding to the zombie list
2949 * because we may have concurrent access
2950 * with lnet_detach_rsp_tracker().
2952 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2953 list_move(&rspt->rspt_on_list,
2954 the_lnet.ln_mt_zombie_rstqs[i]);
2958 LASSERT(md->md_rspt_ptr == rspt);
2959 md->md_rspt_ptr = NULL;
2962 LNetMDUnlink(rspt->rspt_mdh);
2964 nid = rspt->rspt_next_hop_nid;
2966 list_del(&rspt->rspt_on_list);
2967 lnet_rspt_free(rspt, i);
2969 /* If we're shutting down we just want to clean
2970 * up the rspt blocks
2972 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2976 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2980 "Response timeout: md = %p: nid = %s\n",
2981 md, libcfs_nid2str(nid));
2984 * If there is a timeout on the response
2985 * from the next hop decrement its health
2986 * value so that we don't use it
2989 lpni = lnet_find_peer_ni_locked(nid);
2991 lnet_handle_remote_failure_locked(lpni);
2992 lnet_peer_ni_decref_locked(lpni);
3001 if (!list_empty(&local_queue)) {
3003 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3010 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3012 struct lnet_msg *msg;
3014 while (!list_empty(resendq)) {
3015 struct lnet_peer_ni *lpni;
3017 msg = list_entry(resendq->next, struct lnet_msg,
3020 list_del_init(&msg->msg_list);
3022 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3024 lnet_net_unlock(cpt);
3025 CERROR("Expected that a peer is already created for %s\n",
3026 libcfs_nid2str(msg->msg_hdr.dest_nid));
3027 msg->msg_no_resend = true;
3028 lnet_finalize(msg, -EFAULT);
3033 lnet_peer_ni_decref_locked(lpni);
3035 lnet_net_unlock(cpt);
3036 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3037 libcfs_nid2str(msg->msg_src_nid_param),
3038 libcfs_id2str(msg->msg_target),
3039 lnet_msgtyp2str(msg->msg_type),
3041 msg->msg_retry_count);
3042 rc = lnet_send(msg->msg_src_nid_param, msg,
3043 msg->msg_rtr_nid_param);
3045 CERROR("Error sending %s to %s: %d\n",
3046 lnet_msgtyp2str(msg->msg_type),
3047 libcfs_id2str(msg->msg_target), rc);
3048 msg->msg_no_resend = true;
3049 lnet_finalize(msg, rc);
3053 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3059 lnet_resend_pending_msgs(void)
3063 cfs_cpt_for_each(i, lnet_cpt_table()) {
3065 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3070 /* called with cpt and ni_lock held */
3072 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3074 struct lnet_handle_md recovery_mdh;
3076 LNetInvalidateMDHandle(&recovery_mdh);
3078 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3080 recovery_mdh = ni->ni_ping_mdh;
3081 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3084 lnet_net_unlock(cpt);
3085 if (!LNetMDHandleIsInvalid(recovery_mdh))
3086 LNetMDUnlink(recovery_mdh);
3092 lnet_recover_local_nis(void)
3094 struct lnet_mt_event_info *ev_info;
3095 LIST_HEAD(processed_list);
3096 LIST_HEAD(local_queue);
3097 struct lnet_handle_md mdh;
3098 struct lnet_ni *tmp;
3105 * splice the recovery queue on a local queue. We will iterate
3106 * through the local queue and update it as needed. Once we're
3107 * done with the traversal, we'll splice the local queue back on
3108 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3109 * will be traversed in the next iteration.
3112 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3116 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3118 * if an NI is being deleted or it is now healthy, there
3119 * is no need to keep it around in the recovery queue.
3120 * The monitor thread is the only thread responsible for
3121 * removing the NI from the recovery queue.
3122 * Multiple threads can be adding NIs to the recovery
3125 healthv = atomic_read(&ni->ni_healthv);
3129 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3130 healthv == LNET_MAX_HEALTH_VALUE) {
3131 list_del_init(&ni->ni_recovery);
3132 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3134 lnet_ni_decref_locked(ni, 0);
3140 * if the local NI failed recovery we must unlink the md.
3141 * But we want to keep the local_ni on the recovery queue
3142 * so we can continue the attempts to recover it.
3144 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3145 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3146 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3153 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3154 libcfs_nid2str(ni->ni_nid));
3157 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3158 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3161 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3163 CERROR("out of memory. Can't recover %s\n",
3164 libcfs_nid2str(ni->ni_nid));
3166 ni->ni_recovery_state &=
3167 ~LNET_NI_RECOVERY_PENDING;
3172 mdh = ni->ni_ping_mdh;
3174 * Invalidate the ni mdh in case it's deleted.
3175 * We'll unlink the mdh in this case below.
3177 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3181 * remove the NI from the local queue and drop the
3182 * reference count to it while we're recovering
3183 * it. The reason for that, is that the NI could
3184 * be deleted, and the way the code is structured
3185 * is if we don't drop the NI, then the deletion
3186 * code will enter a loop waiting for the
3187 * reference count to be removed while holding the
3188 * ln_mutex_lock(). When we look up the peer to
3189 * send to in lnet_select_pathway() we will try to
3190 * lock the ln_mutex_lock() as well, leading to
3191 * a deadlock. By dropping the refcount and
3192 * removing it from the list, we allow for the NI
3193 * to be removed, then we use the cached NID to
3194 * look it up again. If it's gone, then we just
3195 * continue examining the rest of the queue.
3198 list_del_init(&ni->ni_recovery);
3199 lnet_ni_decref_locked(ni, 0);
3202 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3203 ev_info->mt_nid = nid;
3204 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3205 ev_info, the_lnet.ln_mt_eq, true);
3206 /* lookup the nid again */
3208 ni = lnet_nid2ni_locked(nid, 0);
3211 * the NI has been deleted when we dropped
3219 * Same note as in lnet_recover_peer_nis(). When
3220 * we're sending the ping, the NI is free to be
3221 * deleted or manipulated. By this point it
3222 * could've been added back on the recovery queue,
3223 * and a refcount taken on it.
3224 * So we can't just add it blindly again or we'll
3225 * corrupt the queue. We must check under lock if
3226 * it's not on any list and if not then add it
3227 * to the processed list, which will eventually be
3228 * spliced back on to the recovery queue.
3230 ni->ni_ping_mdh = mdh;
3231 if (list_empty(&ni->ni_recovery)) {
3232 list_add_tail(&ni->ni_recovery, &processed_list);
3233 lnet_ni_addref_locked(ni, 0);
3239 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3245 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3246 * reexamined in the next iteration.
3248 list_splice_init(&processed_list, &local_queue);
3250 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3255 lnet_resendqs_create(void)
3257 struct list_head **resendqs;
3258 resendqs = lnet_create_array_of_queues();
3263 lnet_net_lock(LNET_LOCK_EX);
3264 the_lnet.ln_mt_resendqs = resendqs;
3265 lnet_net_unlock(LNET_LOCK_EX);
3271 lnet_clean_local_ni_recoveryq(void)
3275 /* This is only called when the monitor thread has stopped */
3278 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3279 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3280 struct lnet_ni, ni_recovery);
3281 list_del_init(&ni->ni_recovery);
3283 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3285 lnet_ni_decref_locked(ni, 0);
3292 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3295 struct lnet_handle_md recovery_mdh;
3297 LNetInvalidateMDHandle(&recovery_mdh);
3299 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3300 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3301 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3303 spin_unlock(&lpni->lpni_lock);
3304 lnet_net_unlock(cpt);
3305 if (!LNetMDHandleIsInvalid(recovery_mdh))
3306 LNetMDUnlink(recovery_mdh);
3308 spin_lock(&lpni->lpni_lock);
3312 lnet_clean_peer_ni_recoveryq(void)
3314 struct lnet_peer_ni *lpni, *tmp;
3316 lnet_net_lock(LNET_LOCK_EX);
3318 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3320 list_del_init(&lpni->lpni_recovery);
3321 spin_lock(&lpni->lpni_lock);
3322 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3323 spin_unlock(&lpni->lpni_lock);
3324 lnet_peer_ni_decref_locked(lpni);
3327 lnet_net_unlock(LNET_LOCK_EX);
3331 lnet_clean_resendqs(void)
3333 struct lnet_msg *msg, *tmp;
3337 cfs_cpt_for_each(i, lnet_cpt_table()) {
3339 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3341 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3342 list_del_init(&msg->msg_list);
3343 msg->msg_no_resend = true;
3344 lnet_finalize(msg, -ESHUTDOWN);
3348 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3352 lnet_recover_peer_nis(void)
3354 struct lnet_mt_event_info *ev_info;
3355 LIST_HEAD(processed_list);
3356 LIST_HEAD(local_queue);
3357 struct lnet_handle_md mdh;
3358 struct lnet_peer_ni *lpni;
3359 struct lnet_peer_ni *tmp;
3365 * Always use cpt 0 for locking across all interactions with
3366 * ln_mt_peerNIRecovq
3369 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3373 list_for_each_entry_safe(lpni, tmp, &local_queue,
3376 * The same protection strategy is used here as is in the
3377 * local recovery case.
3380 healthv = atomic_read(&lpni->lpni_healthv);
3381 spin_lock(&lpni->lpni_lock);
3382 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3383 healthv == LNET_MAX_HEALTH_VALUE) {
3384 list_del_init(&lpni->lpni_recovery);
3385 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3386 spin_unlock(&lpni->lpni_lock);
3387 lnet_peer_ni_decref_locked(lpni);
3393 * If the peer NI has failed recovery we must unlink the
3394 * md. But we want to keep the peer ni on the recovery
3395 * queue so we can try to continue recovering it
3397 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3398 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3399 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3402 spin_unlock(&lpni->lpni_lock);
3406 * NOTE: we're racing with peer deletion from user space.
3407 * It's possible that a peer is deleted after we check its
3408 * state. In this case the recovery can create a new peer
3410 spin_lock(&lpni->lpni_lock);
3411 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3412 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3413 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3414 spin_unlock(&lpni->lpni_lock);
3416 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3418 CERROR("out of memory. Can't recover %s\n",
3419 libcfs_nid2str(lpni->lpni_nid));
3420 spin_lock(&lpni->lpni_lock);
3421 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3422 spin_unlock(&lpni->lpni_lock);
3426 /* look at the comments in lnet_recover_local_nis() */
3427 mdh = lpni->lpni_recovery_ping_mdh;
3428 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3429 nid = lpni->lpni_nid;
3431 list_del_init(&lpni->lpni_recovery);
3432 lnet_peer_ni_decref_locked(lpni);
3435 ev_info->mt_type = MT_TYPE_PEER_NI;
3436 ev_info->mt_nid = nid;
3437 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3438 ev_info, the_lnet.ln_mt_eq, true);
3441 * lnet_find_peer_ni_locked() grabs a refcount for
3442 * us. No need to take it explicitly.
3444 lpni = lnet_find_peer_ni_locked(nid);
3451 lpni->lpni_recovery_ping_mdh = mdh;
3453 * While we're unlocked the lpni could've been
3454 * readded on the recovery queue. In this case we
3455 * don't need to add it to the local queue, since
3456 * it's already on there and the thread that added
3457 * it would've incremented the refcount on the
3458 * peer, which means we need to decref the refcount
3459 * that was implicitly grabbed by find_peer_ni_locked.
3460 * Otherwise, if the lpni is still not on
3461 * the recovery queue, then we'll add it to the
3464 if (list_empty(&lpni->lpni_recovery))
3465 list_add_tail(&lpni->lpni_recovery, &processed_list);
3467 lnet_peer_ni_decref_locked(lpni);
3470 spin_lock(&lpni->lpni_lock);
3472 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3474 spin_unlock(&lpni->lpni_lock);
3477 list_splice_init(&processed_list, &local_queue);
3479 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3484 lnet_monitor_thread(void *arg)
3486 time64_t recovery_timeout = 0;
3487 time64_t rsp_timeout = 0;
3491 wait_for_completion(&the_lnet.ln_started);
3493 * The monitor thread takes care of the following:
3494 * 1. Checks the aliveness of routers
3495 * 2. Checks if there are messages on the resend queue to resend
3497 * 3. Check if there are any NIs on the local recovery queue and
3499 * 4. Checks if there are any NIs on the remote recovery queue
3502 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3503 now = ktime_get_real_seconds();
3505 if (lnet_router_checker_active())
3506 lnet_check_routers();
3508 lnet_resend_pending_msgs();
3510 if (now >= rsp_timeout) {
3511 lnet_finalize_expired_responses();
3512 rsp_timeout = now + (lnet_transaction_timeout / 2);
3515 if (now >= recovery_timeout) {
3516 lnet_recover_local_nis();
3517 lnet_recover_peer_nis();
3518 recovery_timeout = now + lnet_recovery_interval;
3522 * TODO do we need to check if we should sleep without
3523 * timeout? Technically, an active system will always
3524 * have messages in flight so this check will always
3525 * evaluate to false. And on an idle system do we care
3526 * if we wake up every 1 second? Although, we've seen
3527 * cases where we get a complaint that an idle thread
3528 * is waking up unnecessarily.
3530 * Take into account the current net_count when you wake
3531 * up for alive router checking, since we need to check
3532 * possibly as many networks as we have configured.
3534 interval = min(lnet_recovery_interval,
3535 min((unsigned int) alive_router_check_interval /
3536 lnet_current_net_count,
3537 lnet_transaction_timeout / 2));
3538 wait_for_completion_interruptible_timeout(
3539 &the_lnet.ln_mt_wait_complete,
3540 cfs_time_seconds(interval));
3541 /* Must re-init the completion before testing anything,
3542 * including ln_mt_state.
3544 reinit_completion(&the_lnet.ln_mt_wait_complete);
3548 lnet_net_lock(LNET_LOCK_EX);
3549 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3550 lnet_net_unlock(LNET_LOCK_EX);
3552 /* signal that the monitor thread is exiting */
3553 up(&the_lnet.ln_mt_signal);
3561 * Returns == 0 if success
3562 * Returns > 0 if LNetMDBind or prior fails
3563 * Returns < 0 if LNetGet fails
3566 lnet_send_ping(lnet_nid_t dest_nid,
3567 struct lnet_handle_md *mdh, int nnis,
3568 void *user_data, struct lnet_eq *eq, bool recovery)
3570 struct lnet_md md = { NULL };
3571 struct lnet_process_id id;
3572 struct lnet_ping_buffer *pbuf;
3575 if (dest_nid == LNET_NID_ANY) {
3580 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3586 /* initialize md content */
3587 md.start = &pbuf->pb_info;
3588 md.length = LNET_PING_INFO_SIZE(nnis);
3589 md.threshold = 2; /* GET/REPLY */
3591 md.options = LNET_MD_TRUNCATE;
3592 md.user_ptr = user_data;
3595 rc = LNetMDBind(md, LNET_UNLINK, mdh);
3597 lnet_ping_buffer_decref(pbuf);
3598 CERROR("Can't bind MD: %d\n", rc);
3599 rc = -rc; /* change the rc to positive */
3602 id.pid = LNET_PID_LUSTRE;
3605 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3606 LNET_RESERVED_PORTAL,
3607 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3610 goto fail_unlink_md;
3616 LNetInvalidateMDHandle(mdh);
3622 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3623 int status, bool unlink_event)
3625 lnet_nid_t nid = ev_info->mt_nid;
3627 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3631 ni = lnet_nid2ni_locked(nid, 0);
3637 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3639 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3644 CERROR("local NI (%s) recovery failed with %d\n",
3645 libcfs_nid2str(nid), status);
3649 * need to increment healthv for the ni here, because in
3650 * the lnet_finalize() path we don't have access to this
3651 * NI. And in order to get access to it, we'll need to
3652 * carry forward too much information.
3653 * In the peer case, it'll naturally be incremented
3656 lnet_inc_healthv(&ni->ni_healthv,
3657 lnet_health_sensitivity);
3659 struct lnet_peer_ni *lpni;
3662 cpt = lnet_net_lock_current();
3663 lpni = lnet_find_peer_ni_locked(nid);
3665 lnet_net_unlock(cpt);
3668 spin_lock(&lpni->lpni_lock);
3669 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3671 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3672 spin_unlock(&lpni->lpni_lock);
3673 lnet_peer_ni_decref_locked(lpni);
3674 lnet_net_unlock(cpt);
3677 CERROR("peer NI (%s) recovery failed with %d\n",
3678 libcfs_nid2str(nid), status);
3683 lnet_mt_event_handler(struct lnet_event *event)
3685 struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3686 struct lnet_ping_buffer *pbuf;
3688 /* TODO: remove assert */
3689 LASSERT(event->type == LNET_EVENT_REPLY ||
3690 event->type == LNET_EVENT_SEND ||
3691 event->type == LNET_EVENT_UNLINK);
3693 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3696 switch (event->type) {
3697 case LNET_EVENT_UNLINK:
3698 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3699 libcfs_nid2str(ev_info->mt_nid));
3701 case LNET_EVENT_REPLY:
3702 lnet_handle_recovery_reply(ev_info, event->status,
3703 event->type == LNET_EVENT_UNLINK);
3705 case LNET_EVENT_SEND:
3706 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3707 libcfs_nid2str(ev_info->mt_nid),
3708 (event->status) ? "unsuccessfully" :
3709 "successfully", event->status);
3712 CERROR("Unexpected event: %d\n", event->type);
3715 if (event->unlinked) {
3716 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3717 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3718 lnet_ping_buffer_decref(pbuf);
3723 lnet_rsp_tracker_create(void)
3725 struct list_head **rstqs;
3726 rstqs = lnet_create_array_of_queues();
3731 the_lnet.ln_mt_rstq = rstqs;
3737 lnet_rsp_tracker_clean(void)
3739 lnet_finalize_expired_responses();
3741 cfs_percpt_free(the_lnet.ln_mt_rstq);
3742 the_lnet.ln_mt_rstq = NULL;
3745 int lnet_monitor_thr_start(void)
3748 struct task_struct *task;
3750 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3753 rc = lnet_resendqs_create();
3757 rc = lnet_rsp_tracker_create();
3761 sema_init(&the_lnet.ln_mt_signal, 0);
3763 lnet_net_lock(LNET_LOCK_EX);
3764 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3765 lnet_net_unlock(LNET_LOCK_EX);
3766 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3769 CERROR("Can't start monitor thread: %d\n", rc);
3776 lnet_net_lock(LNET_LOCK_EX);
3777 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3778 lnet_net_unlock(LNET_LOCK_EX);
3779 /* block until event callback signals exit */
3780 down(&the_lnet.ln_mt_signal);
3782 lnet_net_lock(LNET_LOCK_EX);
3783 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3784 lnet_net_unlock(LNET_LOCK_EX);
3785 lnet_rsp_tracker_clean();
3786 lnet_clean_local_ni_recoveryq();
3787 lnet_clean_peer_ni_recoveryq();
3788 lnet_clean_resendqs();
3789 the_lnet.ln_mt_eq = NULL;
3792 lnet_rsp_tracker_clean();
3793 lnet_clean_local_ni_recoveryq();
3794 lnet_clean_peer_ni_recoveryq();
3795 lnet_clean_resendqs();
3799 void lnet_monitor_thr_stop(void)
3801 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3804 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3805 lnet_net_lock(LNET_LOCK_EX);
3806 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3807 lnet_net_unlock(LNET_LOCK_EX);
3809 /* tell the monitor thread that we're shutting down */
3810 complete(&the_lnet.ln_mt_wait_complete);
3812 /* block until monitor thread signals that it's done */
3813 down(&the_lnet.ln_mt_signal);
3814 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3816 /* perform cleanup tasks */
3817 lnet_rsp_tracker_clean();
3818 lnet_clean_local_ni_recoveryq();
3819 lnet_clean_peer_ni_recoveryq();
3820 lnet_clean_resendqs();
3824 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3828 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3829 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3830 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3831 lnet_net_unlock(cpt);
3833 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3837 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3839 struct lnet_hdr *hdr = &msg->msg_hdr;
3841 if (msg->msg_wanted != 0)
3842 lnet_setpayloadbuffer(msg);
3844 lnet_build_msg_event(msg, LNET_EVENT_PUT);
3846 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
3847 * it back into the ACK during lnet_finalize() */
3848 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3849 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3851 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3852 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3856 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3858 struct lnet_hdr *hdr = &msg->msg_hdr;
3859 struct lnet_match_info info;
3863 /* Convert put fields to host byte order */
3864 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3865 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
3866 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
3868 /* Primary peer NID. */
3869 info.mi_id.nid = msg->msg_initiator;
3870 info.mi_id.pid = hdr->src_pid;
3871 info.mi_opc = LNET_MD_OP_PUT;
3872 info.mi_portal = hdr->msg.put.ptl_index;
3873 info.mi_rlength = hdr->payload_length;
3874 info.mi_roffset = hdr->msg.put.offset;
3875 info.mi_mbits = hdr->msg.put.match_bits;
3876 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3878 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3879 ready_delay = msg->msg_rx_ready_delay;
3882 rc = lnet_ptl_match_md(&info, msg);
3887 case LNET_MATCHMD_OK:
3888 lnet_recv_put(ni, msg);
3891 case LNET_MATCHMD_NONE:
3893 /* no eager_recv or has already called it, should
3894 * have been attached on delayed list */
3897 rc = lnet_ni_eager_recv(ni, msg);
3904 case LNET_MATCHMD_DROP:
3905 CNETERR("Dropping PUT from %s portal %d match %llu"
3906 " offset %d length %d: %d\n",
3907 libcfs_id2str(info.mi_id), info.mi_portal,
3908 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3910 return -ENOENT; /* -ve: OK but no match */
3915 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3917 struct lnet_match_info info;
3918 struct lnet_hdr *hdr = &msg->msg_hdr;
3919 struct lnet_process_id source_id;
3920 struct lnet_handle_wire reply_wmd;
3923 /* Convert get fields to host byte order */
3924 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
3925 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
3926 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
3927 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
3929 source_id.nid = hdr->src_nid;
3930 source_id.pid = hdr->src_pid;
3931 /* Primary peer NID */
3932 info.mi_id.nid = msg->msg_initiator;
3933 info.mi_id.pid = hdr->src_pid;
3934 info.mi_opc = LNET_MD_OP_GET;
3935 info.mi_portal = hdr->msg.get.ptl_index;
3936 info.mi_rlength = hdr->msg.get.sink_length;
3937 info.mi_roffset = hdr->msg.get.src_offset;
3938 info.mi_mbits = hdr->msg.get.match_bits;
3939 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3941 rc = lnet_ptl_match_md(&info, msg);
3942 if (rc == LNET_MATCHMD_DROP) {
3943 CNETERR("Dropping GET from %s portal %d match %llu"
3944 " offset %d length %d\n",
3945 libcfs_id2str(info.mi_id), info.mi_portal,
3946 info.mi_mbits, info.mi_roffset, info.mi_rlength);
3947 return -ENOENT; /* -ve: OK but no match */
3950 LASSERT(rc == LNET_MATCHMD_OK);
3952 lnet_build_msg_event(msg, LNET_EVENT_GET);
3954 reply_wmd = hdr->msg.get.return_wmd;
3956 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3957 msg->msg_offset, msg->msg_wanted);
3959 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3962 /* The LND completes the REPLY from her recv procedure */
3963 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3964 msg->msg_offset, msg->msg_len, msg->msg_len);
3968 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3969 msg->msg_receiving = 0;
3971 rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3973 /* didn't get as far as lnet_ni_send() */
3974 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3975 libcfs_nid2str(ni->ni_nid),
3976 libcfs_id2str(info.mi_id), rc);
3978 lnet_finalize(msg, rc);
3985 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3987 void *private = msg->msg_private;
3988 struct lnet_hdr *hdr = &msg->msg_hdr;
3989 struct lnet_process_id src = {0};
3990 struct lnet_libmd *md;
3991 unsigned int rlength;
3992 unsigned int mlength;
3995 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3998 src.nid = hdr->src_nid;
3999 src.pid = hdr->src_pid;
4001 /* NB handles only looked up by creator (no flips) */
4002 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4003 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4004 CNETERR("%s: Dropping REPLY from %s for %s "
4006 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4007 (md == NULL) ? "invalid" : "inactive",
4008 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4009 hdr->msg.reply.dst_wmd.wh_object_cookie);
4010 if (md != NULL && md->md_me != NULL)
4011 CERROR("REPLY MD also attached to portal %d\n",
4012 md->md_me->me_portal);
4014 lnet_res_unlock(cpt);
4015 return -ENOENT; /* -ve: OK but no match */
4018 LASSERT(md->md_offset == 0);
4020 rlength = hdr->payload_length;
4021 mlength = min(rlength, md->md_length);
4023 if (mlength < rlength &&
4024 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4025 CNETERR("%s: Dropping REPLY from %s length %d "
4026 "for MD %#llx would overflow (%d)\n",
4027 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4028 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4030 lnet_res_unlock(cpt);
4031 return -ENOENT; /* -ve: OK but no match */
4034 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4035 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4036 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4038 lnet_msg_attach_md(msg, md, 0, mlength);
4041 lnet_setpayloadbuffer(msg);
4043 lnet_res_unlock(cpt);
4045 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4047 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4052 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4054 struct lnet_hdr *hdr = &msg->msg_hdr;
4055 struct lnet_process_id src = {0};
4056 struct lnet_libmd *md;
4059 src.nid = hdr->src_nid;
4060 src.pid = hdr->src_pid;
4062 /* Convert ack fields to host byte order */
4063 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4064 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4066 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4069 /* NB handles only looked up by creator (no flips) */
4070 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4071 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4072 /* Don't moan; this is expected */
4074 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4075 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4076 (md == NULL) ? "invalid" : "inactive",
4077 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4078 hdr->msg.ack.dst_wmd.wh_object_cookie);
4079 if (md != NULL && md->md_me != NULL)
4080 CERROR("Source MD also attached to portal %d\n",
4081 md->md_me->me_portal);
4083 lnet_res_unlock(cpt);
4084 return -ENOENT; /* -ve! */
4087 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4088 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4089 hdr->msg.ack.dst_wmd.wh_object_cookie);
4091 lnet_msg_attach_md(msg, md, 0, 0);
4093 lnet_res_unlock(cpt);
4095 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4097 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4102 * \retval LNET_CREDIT_OK If \a msg is forwarded
4103 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4104 * \retval -ve error code
4107 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4111 if (!the_lnet.ln_routing)
4114 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4115 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4116 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4117 msg->msg_rx_ready_delay = 1;
4119 lnet_net_unlock(msg->msg_rx_cpt);
4120 rc = lnet_ni_eager_recv(ni, msg);
4121 lnet_net_lock(msg->msg_rx_cpt);
4126 rc = lnet_post_routed_recv_locked(msg, 0);
4131 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4135 switch (msg->msg_type) {
4137 rc = lnet_parse_ack(ni, msg);
4140 rc = lnet_parse_put(ni, msg);
4143 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4145 case LNET_MSG_REPLY:
4146 rc = lnet_parse_reply(ni, msg);
4148 default: /* prevent an unused label if !kernel */
4153 LASSERT(rc == 0 || rc == -ENOENT);
4158 lnet_msgtyp2str (int type)
4167 case LNET_MSG_REPLY:
4169 case LNET_MSG_HELLO:
4172 return ("<UNKNOWN>");
4177 lnet_print_hdr(struct lnet_hdr *hdr)
4179 struct lnet_process_id src = {
4180 .nid = hdr->src_nid,
4181 .pid = hdr->src_pid,
4183 struct lnet_process_id dst = {
4184 .nid = hdr->dest_nid,
4185 .pid = hdr->dest_pid,
4187 char *type_str = lnet_msgtyp2str(hdr->type);
4189 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4190 CWARN(" From %s\n", libcfs_id2str(src));
4191 CWARN(" To %s\n", libcfs_id2str(dst));
4193 switch (hdr->type) {
4198 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
4199 "match bits %llu\n",
4200 hdr->msg.put.ptl_index,
4201 hdr->msg.put.ack_wmd.wh_interface_cookie,
4202 hdr->msg.put.ack_wmd.wh_object_cookie,
4203 hdr->msg.put.match_bits);
4204 CWARN(" Length %d, offset %d, hdr data %#llx\n",
4205 hdr->payload_length, hdr->msg.put.offset,
4206 hdr->msg.put.hdr_data);
4210 CWARN(" Ptl index %d, return md %#llx.%#llx, "
4211 "match bits %llu\n", hdr->msg.get.ptl_index,
4212 hdr->msg.get.return_wmd.wh_interface_cookie,
4213 hdr->msg.get.return_wmd.wh_object_cookie,
4214 hdr->msg.get.match_bits);
4215 CWARN(" Length %d, src offset %d\n",
4216 hdr->msg.get.sink_length,
4217 hdr->msg.get.src_offset);
4221 CWARN(" dst md %#llx.%#llx, "
4222 "manipulated length %d\n",
4223 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4224 hdr->msg.ack.dst_wmd.wh_object_cookie,
4225 hdr->msg.ack.mlength);
4228 case LNET_MSG_REPLY:
4229 CWARN(" dst md %#llx.%#llx, "
4231 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4232 hdr->msg.reply.dst_wmd.wh_object_cookie,
4233 hdr->payload_length);
4239 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4240 void *private, int rdma_req)
4242 struct lnet_peer_ni *lpni;
4243 struct lnet_msg *msg;
4244 __u32 payload_length;
4245 lnet_pid_t dest_pid;
4246 lnet_nid_t dest_nid;
4254 LASSERT (!in_interrupt ());
4256 type = le32_to_cpu(hdr->type);
4257 src_nid = le64_to_cpu(hdr->src_nid);
4258 dest_nid = le64_to_cpu(hdr->dest_nid);
4259 dest_pid = le32_to_cpu(hdr->dest_pid);
4260 payload_length = le32_to_cpu(hdr->payload_length);
4262 for_me = (ni->ni_nid == dest_nid);
4263 cpt = lnet_cpt_of_nid(from_nid, ni);
4265 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4266 libcfs_nid2str(dest_nid),
4267 libcfs_nid2str(ni->ni_nid),
4268 libcfs_nid2str(src_nid),
4269 lnet_msgtyp2str(type),
4270 (for_me) ? "for me" : "routed");
4275 if (payload_length > 0) {
4276 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4277 libcfs_nid2str(from_nid),
4278 libcfs_nid2str(src_nid),
4279 lnet_msgtyp2str(type), payload_length);
4285 case LNET_MSG_REPLY:
4286 if (payload_length >
4287 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4288 CERROR("%s, src %s: bad %s payload %d "
4289 "(%d max expected)\n",
4290 libcfs_nid2str(from_nid),
4291 libcfs_nid2str(src_nid),
4292 lnet_msgtyp2str(type),
4294 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4300 CERROR("%s, src %s: Bad message type 0x%x\n",
4301 libcfs_nid2str(from_nid),
4302 libcfs_nid2str(src_nid), type);
4306 if (the_lnet.ln_routing &&
4307 ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4309 spin_lock(&ni->ni_net->net_lock);
4310 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4311 spin_unlock(&ni->ni_net->net_lock);
4312 if (ni->ni_status != NULL &&
4313 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4314 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4321 lnet_push_update_to_peers(1);
4323 /* Regard a bad destination NID as a protocol error. Senders should
4324 * know what they're doing; if they don't they're misconfigured, buggy
4325 * or malicious so we chop them off at the knees :) */
4328 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4329 /* should have gone direct */
4330 CERROR("%s, src %s: Bad dest nid %s "
4331 "(should have been sent direct)\n",
4332 libcfs_nid2str(from_nid),
4333 libcfs_nid2str(src_nid),
4334 libcfs_nid2str(dest_nid));
4338 if (lnet_islocalnid(dest_nid)) {
4339 /* dest is another local NI; sender should have used
4340 * this node's NID on its own network */
4341 CERROR("%s, src %s: Bad dest nid %s "
4342 "(it's my nid but on a different network)\n",
4343 libcfs_nid2str(from_nid),
4344 libcfs_nid2str(src_nid),
4345 libcfs_nid2str(dest_nid));
4349 if (rdma_req && type == LNET_MSG_GET) {
4350 CERROR("%s, src %s: Bad optimized GET for %s "
4351 "(final destination must be me)\n",
4352 libcfs_nid2str(from_nid),
4353 libcfs_nid2str(src_nid),
4354 libcfs_nid2str(dest_nid));
4358 if (!the_lnet.ln_routing) {
4359 CERROR("%s, src %s: Dropping message for %s "
4360 "(routing not enabled)\n",
4361 libcfs_nid2str(from_nid),
4362 libcfs_nid2str(src_nid),
4363 libcfs_nid2str(dest_nid));
4368 /* Message looks OK; we're not going to return an error, so we MUST
4369 * call back lnd_recv() come what may... */
4371 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4372 fail_peer(src_nid, 0)) { /* shall we now? */
4373 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4374 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4375 lnet_msgtyp2str(type));
4379 if (!list_empty(&the_lnet.ln_drop_rules) &&
4380 lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4382 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4383 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4384 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4388 if (lnet_drop_asym_route && for_me &&
4389 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4390 struct lnet_net *net;
4391 struct lnet_remotenet *rnet;
4394 /* we are dealing with a routed message,
4395 * so see if route to reach src_nid goes through from_nid
4398 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4400 lnet_net_unlock(cpt);
4401 CERROR("net %s not found\n",
4402 libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4406 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4408 struct lnet_peer *gw = NULL;
4409 struct lnet_peer_ni *lpni = NULL;
4410 struct lnet_route *route;
4412 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4414 gw = route->lr_gateway;
4415 if (route->lr_lnet != net->net_id)
4418 * if the nid is one of the gateway's NIDs
4419 * then this is a valid gateway
4421 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4422 NULL, lpni)) != NULL) {
4423 if (lpni->lpni_nid == from_nid) {
4430 lnet_net_unlock(cpt);
4432 /* we would not use from_nid to route a message to
4434 * => asymmetric routing detected but forbidden
4436 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4437 libcfs_nid2str(from_nid),
4438 libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4443 msg = lnet_msg_alloc();
4445 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4446 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4447 lnet_msgtyp2str(type));
4451 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4452 * pointers NULL etc */
4454 msg->msg_type = type;
4455 msg->msg_private = private;
4456 msg->msg_receiving = 1;
4457 msg->msg_rdma_get = rdma_req;
4458 msg->msg_len = msg->msg_wanted = payload_length;
4459 msg->msg_offset = 0;
4460 msg->msg_hdr = *hdr;
4461 /* for building message event */
4462 msg->msg_from = from_nid;
4464 msg->msg_target.pid = dest_pid;
4465 msg->msg_target.nid = dest_nid;
4466 msg->msg_routing = 1;
4469 /* convert common msg->hdr fields to host byteorder */
4470 msg->msg_hdr.type = type;
4471 msg->msg_hdr.src_nid = src_nid;
4472 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4473 msg->msg_hdr.dest_nid = dest_nid;
4474 msg->msg_hdr.dest_pid = dest_pid;
4475 msg->msg_hdr.payload_length = payload_length;
4479 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4481 lnet_net_unlock(cpt);
4482 CERROR("%s, src %s: Dropping %s "
4483 "(error %ld looking up sender)\n",
4484 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4485 lnet_msgtyp2str(type), PTR_ERR(lpni));
4487 if (rc == -ESHUTDOWN)
4488 /* We are shutting down. Don't do anything more */
4493 if (the_lnet.ln_routing)
4494 lpni->lpni_last_alive = ktime_get_seconds();
4496 msg->msg_rxpeer = lpni;
4498 lnet_ni_addref_locked(ni, cpt);
4499 /* Multi-Rail: Primary NID of source. */
4500 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4503 * mark the status of this lpni as UP since we received a message
4504 * from it. The ping response reports back the ns_status which is
4505 * marked on the remote as up or down and we cache it here.
4507 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4509 lnet_msg_commit(msg, cpt);
4511 /* message delay simulation */
4512 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4513 lnet_delay_rule_match_locked(hdr, msg))) {
4514 lnet_net_unlock(cpt);
4519 rc = lnet_parse_forward_locked(ni, msg);
4520 lnet_net_unlock(cpt);
4525 if (rc == LNET_CREDIT_OK) {
4526 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4527 0, payload_length, payload_length);
4532 lnet_net_unlock(cpt);
4534 rc = lnet_parse_local(ni, msg);
4540 LASSERT(msg->msg_md == NULL);
4541 lnet_finalize(msg, rc);
4544 lnet_drop_message(ni, cpt, private, payload_length, type);
4547 EXPORT_SYMBOL(lnet_parse);
4550 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4552 while (!list_empty(head)) {
4553 struct lnet_process_id id = {0};
4554 struct lnet_msg *msg;
4556 msg = list_entry(head->next, struct lnet_msg, msg_list);
4557 list_del(&msg->msg_list);
4559 id.nid = msg->msg_hdr.src_nid;
4560 id.pid = msg->msg_hdr.src_pid;
4562 LASSERT(msg->msg_md == NULL);
4563 LASSERT(msg->msg_rx_delayed);
4564 LASSERT(msg->msg_rxpeer != NULL);
4565 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4567 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4568 " offset %d length %d: %s\n",
4570 msg->msg_hdr.msg.put.ptl_index,
4571 msg->msg_hdr.msg.put.match_bits,
4572 msg->msg_hdr.msg.put.offset,
4573 msg->msg_hdr.payload_length, reason);
4575 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4576 * called lnet_drop_message(), so I just hang onto msg as well
4577 * until that's done */
4579 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4580 msg->msg_private, msg->msg_len,
4583 msg->msg_no_resend = true;
4585 * NB: message will not generate event because w/o attached MD,
4586 * but we still should give error code so lnet_msg_decommit()
4587 * can skip counters operations and other checks.
4589 lnet_finalize(msg, -ENOENT);
4594 lnet_recv_delayed_msg_list(struct list_head *head)
4596 while (!list_empty(head)) {
4597 struct lnet_msg *msg;
4598 struct lnet_process_id id;
4600 msg = list_entry(head->next, struct lnet_msg, msg_list);
4601 list_del(&msg->msg_list);
4603 /* md won't disappear under me, since each msg
4604 * holds a ref on it */
4606 id.nid = msg->msg_hdr.src_nid;
4607 id.pid = msg->msg_hdr.src_pid;
4609 LASSERT(msg->msg_rx_delayed);
4610 LASSERT(msg->msg_md != NULL);
4611 LASSERT(msg->msg_rxpeer != NULL);
4612 LASSERT(msg->msg_rxni != NULL);
4613 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4615 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4616 "match %llu offset %d length %d.\n",
4617 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4618 msg->msg_hdr.msg.put.match_bits,
4619 msg->msg_hdr.msg.put.offset,
4620 msg->msg_hdr.payload_length);
4622 lnet_recv_put(msg->msg_rxni, msg);
4627 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4628 struct lnet_libmd *md, struct lnet_handle_md mdh)
4631 struct lnet_rsp_tracker *local_rspt;
4634 * MD has a refcount taken by message so it's not going away.
4635 * The MD however can be looked up. We need to secure the access
4636 * to the md_rspt_ptr by taking the res_lock.
4637 * The rspt can be accessed without protection up to when it gets
4638 * added to the list.
4642 local_rspt = md->md_rspt_ptr;
4643 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4644 if (local_rspt != NULL) {
4646 * we already have an rspt attached to the md, so we'll
4647 * update the deadline on that one.
4649 lnet_rspt_free(rspt, cpt);
4652 rspt->rspt_mdh = mdh;
4653 rspt->rspt_cpt = cpt;
4654 /* store the rspt so we can access it when we get the REPLY */
4655 md->md_rspt_ptr = rspt;
4658 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4661 * add to the list of tracked responses. It's added to tail of the
4662 * list in order to expire all the older entries first.
4665 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4666 lnet_net_unlock(cpt);
4667 lnet_res_unlock(cpt);
4671 * Initiate an asynchronous PUT operation.
4673 * There are several events associated with a PUT: completion of the send on
4674 * the initiator node (LNET_EVENT_SEND), and when the send completes
4675 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4676 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4677 * used at the target node to indicate the completion of incoming data
4680 * The local events will be logged in the EQ associated with the MD pointed to
4681 * by \a mdh handle. Using a MD without an associated EQ results in these
4682 * events being discarded. In this case, the caller must have another
4683 * mechanism (e.g., a higher level protocol) for determining when it is safe
4684 * to modify the memory region associated with the MD.
4686 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4687 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4689 * \param self Indicates the NID of a local interface through which to send
4690 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4691 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4692 * must be "free floating" (See LNetMDBind()).
4693 * \param ack Controls whether an acknowledgment is requested.
4694 * Acknowledgments are only sent when they are requested by the initiating
4695 * process and the target MD enables them.
4696 * \param target A process identifier for the target process.
4697 * \param portal The index in the \a target's portal table.
4698 * \param match_bits The match bits to use for MD selection at the target
4700 * \param offset The offset into the target MD (only used when the target
4701 * MD has the LNET_MD_MANAGE_REMOTE option set).
4702 * \param hdr_data 64 bits of user data that can be included in the message
4703 * header. This data is written to an event queue entry at the target if an
4704 * EQ is present on the matching MD.
4706 * \retval 0 Success, and only in this case events will be generated
4707 * and logged to EQ (if it exists).
4708 * \retval -EIO Simulated failure.
4709 * \retval -ENOMEM Memory allocation failure.
4710 * \retval -ENOENT Invalid MD object.
4712 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4715 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4716 struct lnet_process_id target, unsigned int portal,
4717 __u64 match_bits, unsigned int offset,
4720 struct lnet_msg *msg;
4721 struct lnet_libmd *md;
4724 struct lnet_rsp_tracker *rspt = NULL;
4726 LASSERT(the_lnet.ln_refcount > 0);
4728 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4729 fail_peer(target.nid, 1)) { /* shall we now? */
4730 CERROR("Dropping PUT to %s: simulated failure\n",
4731 libcfs_id2str(target));
4735 msg = lnet_msg_alloc();
4737 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4738 libcfs_id2str(target));
4741 msg->msg_vmflush = !!memory_pressure_get();
4743 cpt = lnet_cpt_of_cookie(mdh.cookie);
4745 if (ack == LNET_ACK_REQ) {
4746 rspt = lnet_rspt_alloc(cpt);
4748 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4749 libcfs_id2str(target));
4752 INIT_LIST_HEAD(&rspt->rspt_on_list);
4757 md = lnet_handle2md(&mdh);
4758 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4759 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4760 match_bits, portal, libcfs_id2str(target),
4761 md == NULL ? -1 : md->md_threshold);
4762 if (md != NULL && md->md_me != NULL)
4763 CERROR("Source MD also attached to portal %d\n",
4764 md->md_me->me_portal);
4765 lnet_res_unlock(cpt);
4767 lnet_rspt_free(rspt, cpt);
4772 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4774 lnet_msg_attach_md(msg, md, 0, 0);
4776 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4778 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4779 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4780 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4781 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4783 /* NB handles only looked up by creator (no flips) */
4784 if (ack == LNET_ACK_REQ) {
4785 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4786 the_lnet.ln_interface_cookie;
4787 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4788 md->md_lh.lh_cookie;
4790 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4791 LNET_WIRE_HANDLE_COOKIE_NONE;
4792 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4793 LNET_WIRE_HANDLE_COOKIE_NONE;
4796 lnet_res_unlock(cpt);
4798 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4800 if (ack == LNET_ACK_REQ)
4801 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4803 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4807 rc = lnet_send(self, msg, LNET_NID_ANY);
4810 CNETERR("Error sending PUT to %s: %d\n",
4811 libcfs_id2str(target), rc);
4812 msg->msg_no_resend = true;
4813 lnet_finalize(msg, rc);
4816 /* completion will be signalled by an event */
4819 EXPORT_SYMBOL(LNetPut);
4822 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
4823 * returns a msg for the LND to pass to lnet_finalize() when the sink
4824 * data has been received.
4826 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4827 * lnet_finalize() is called on it, so the LND must call this first
4830 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4832 struct lnet_msg *msg = lnet_msg_alloc();
4833 struct lnet_libmd *getmd = getmsg->msg_md;
4834 struct lnet_process_id peer_id = getmsg->msg_target;
4837 LASSERT(!getmsg->msg_target_is_router);
4838 LASSERT(!getmsg->msg_routing);
4841 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4842 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4846 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4849 LASSERT(getmd->md_refcount > 0);
4851 if (getmd->md_threshold == 0) {
4852 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4853 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4855 lnet_res_unlock(cpt);
4859 LASSERT(getmd->md_offset == 0);
4861 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4862 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4864 /* setup information for lnet_build_msg_event */
4865 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4866 msg->msg_from = peer_id.nid;
4867 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4868 msg->msg_hdr.src_nid = peer_id.nid;
4869 msg->msg_hdr.payload_length = getmd->md_length;
4870 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4872 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4873 lnet_res_unlock(cpt);
4875 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4878 lnet_msg_commit(msg, cpt);
4879 lnet_net_unlock(cpt);
4881 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4886 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4889 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4890 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4891 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4893 lnet_net_unlock(cpt);
4900 EXPORT_SYMBOL(lnet_create_reply_msg);
4903 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4906 /* Set the REPLY length, now the RDMA that elides the REPLY message has
4907 * completed and I know it. */
4908 LASSERT(reply != NULL);
4909 LASSERT(reply->msg_type == LNET_MSG_GET);
4910 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4912 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
4913 * the end of my buffer, I might as well be dead. */
4914 LASSERT(len <= reply->msg_ev.mlength);
4916 reply->msg_ev.mlength = len;
4918 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4921 * Initiate an asynchronous GET operation.
4923 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4924 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4925 * the target node in the REPLY has been written to local MD.
4927 * On the target node, an LNET_EVENT_GET is logged when the GET request
4928 * arrives and is accepted into a MD.
4930 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4931 * \param mdh A handle for the MD that describes the memory into which the
4932 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4934 * \retval 0 Success, and only in this case events will be generated
4935 * and logged to EQ (if it exists) of the MD.
4936 * \retval -EIO Simulated failure.
4937 * \retval -ENOMEM Memory allocation failure.
4938 * \retval -ENOENT Invalid MD object.
4941 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4942 struct lnet_process_id target, unsigned int portal,
4943 __u64 match_bits, unsigned int offset, bool recovery)
4945 struct lnet_msg *msg;
4946 struct lnet_libmd *md;
4947 struct lnet_rsp_tracker *rspt;
4951 LASSERT(the_lnet.ln_refcount > 0);
4953 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4954 fail_peer(target.nid, 1)) /* shall we now? */
4956 CERROR("Dropping GET to %s: simulated failure\n",
4957 libcfs_id2str(target));
4961 msg = lnet_msg_alloc();
4963 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4964 libcfs_id2str(target));
4968 cpt = lnet_cpt_of_cookie(mdh.cookie);
4970 rspt = lnet_rspt_alloc(cpt);
4972 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4973 libcfs_id2str(target));
4976 INIT_LIST_HEAD(&rspt->rspt_on_list);
4978 msg->msg_recovery = recovery;
4982 md = lnet_handle2md(&mdh);
4983 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4984 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4985 match_bits, portal, libcfs_id2str(target),
4986 md == NULL ? -1 : md->md_threshold);
4987 if (md != NULL && md->md_me != NULL)
4988 CERROR("REPLY MD also attached to portal %d\n",
4989 md->md_me->me_portal);
4991 lnet_res_unlock(cpt);
4994 lnet_rspt_free(rspt, cpt);
4998 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5000 lnet_msg_attach_md(msg, md, 0, 0);
5002 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5004 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5005 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5006 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5007 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5009 /* NB handles only looked up by creator (no flips) */
5010 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5011 the_lnet.ln_interface_cookie;
5012 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5013 md->md_lh.lh_cookie;
5015 lnet_res_unlock(cpt);
5017 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5019 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5021 rc = lnet_send(self, msg, LNET_NID_ANY);
5023 CNETERR("Error sending GET to %s: %d\n",
5024 libcfs_id2str(target), rc);
5025 msg->msg_no_resend = true;
5026 lnet_finalize(msg, rc);
5029 /* completion will be signalled by an event */
5032 EXPORT_SYMBOL(LNetGet);
5035 * Calculate distance to node at \a dstnid.
5037 * \param dstnid Target NID.
5038 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5040 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5043 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5044 * local_nid_dist_zero is set, which is the default.
5045 * \retval positives Distance to target NID, i.e. number of hops plus one.
5046 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5049 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5051 struct list_head *e;
5052 struct lnet_ni *ni = NULL;
5053 struct lnet_remotenet *rnet;
5054 __u32 dstnet = LNET_NIDNET(dstnid);
5058 struct list_head *rn_list;
5060 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5061 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5062 * keep order 0 free for 0@lo and order 1 free for a local NID
5065 LASSERT(the_lnet.ln_refcount > 0);
5067 cpt = lnet_net_lock_current();
5069 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5070 if (ni->ni_nid == dstnid) {
5071 if (srcnidp != NULL)
5073 if (orderp != NULL) {
5074 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5079 lnet_net_unlock(cpt);
5081 return local_nid_dist_zero ? 0 : 1;
5084 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5085 /* Check if ni was originally created in
5086 * current net namespace.
5087 * If not, assign order above 0xffff0000,
5088 * to make this ni not a priority. */
5089 if (current->nsproxy &&
5090 !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5091 order += 0xffff0000;
5092 if (srcnidp != NULL)
5093 *srcnidp = ni->ni_nid;
5096 lnet_net_unlock(cpt);
5103 rn_list = lnet_net2rnethash(dstnet);
5104 list_for_each(e, rn_list) {
5105 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5107 if (rnet->lrn_net == dstnet) {
5108 struct lnet_route *route;
5109 struct lnet_route *shortest = NULL;
5110 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5113 LASSERT(!list_empty(&rnet->lrn_routes));
5115 list_for_each_entry(route, &rnet->lrn_routes,
5117 route_hops = route->lr_hops;
5118 if (route_hops == LNET_UNDEFINED_HOPS)
5120 if (shortest == NULL ||
5121 route_hops < shortest_hops) {
5123 shortest_hops = route_hops;
5127 LASSERT(shortest != NULL);
5128 hops = shortest_hops;
5129 if (srcnidp != NULL) {
5130 struct lnet_net *net;
5131 net = lnet_get_net_locked(shortest->lr_lnet);
5133 ni = lnet_get_next_ni_locked(net, NULL);
5134 *srcnidp = ni->ni_nid;
5138 lnet_net_unlock(cpt);
5144 lnet_net_unlock(cpt);
5145 return -EHOSTUNREACH;
5147 EXPORT_SYMBOL(LNetDist);