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;
66 static inline struct lnet_comm_count *
67 get_stats_counts(struct lnet_element_stats *stats,
68 enum lnet_stats_type stats_type)
71 case LNET_STATS_TYPE_SEND:
72 return &stats->el_send_stats;
73 case LNET_STATS_TYPE_RECV:
74 return &stats->el_recv_stats;
75 case LNET_STATS_TYPE_DROP:
76 return &stats->el_drop_stats;
78 CERROR("Unknown stats type\n");
84 void lnet_incr_stats(struct lnet_element_stats *stats,
85 enum lnet_msg_type msg_type,
86 enum lnet_stats_type stats_type)
88 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
94 atomic_inc(&counts->co_ack_count);
97 atomic_inc(&counts->co_put_count);
100 atomic_inc(&counts->co_get_count);
103 atomic_inc(&counts->co_reply_count);
106 atomic_inc(&counts->co_hello_count);
109 CERROR("There is a BUG in the code. Unknown message type\n");
114 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
115 enum lnet_stats_type stats_type)
117 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
121 return (atomic_read(&counts->co_ack_count) +
122 atomic_read(&counts->co_put_count) +
123 atomic_read(&counts->co_get_count) +
124 atomic_read(&counts->co_reply_count) +
125 atomic_read(&counts->co_hello_count));
128 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
129 struct lnet_comm_count *counts)
131 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
132 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
133 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
134 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
135 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
138 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
139 struct lnet_element_stats *stats)
141 struct lnet_comm_count *counts;
146 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
149 assign_stats(&msg_stats->im_send_stats, counts);
151 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
154 assign_stats(&msg_stats->im_recv_stats, counts);
156 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
159 assign_stats(&msg_stats->im_drop_stats, counts);
163 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
165 struct lnet_test_peer *tp;
166 struct list_head *el;
167 struct list_head *next;
168 struct list_head cull;
170 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
171 if (threshold != 0) {
172 /* Adding a new entry */
173 LIBCFS_ALLOC(tp, sizeof(*tp));
178 tp->tp_threshold = threshold;
181 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
186 /* removing entries */
187 INIT_LIST_HEAD(&cull);
191 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
192 tp = list_entry(el, struct lnet_test_peer, tp_list);
194 if (tp->tp_threshold == 0 || /* needs culling anyway */
195 nid == LNET_NID_ANY || /* removing all entries */
196 tp->tp_nid == nid) { /* matched this one */
197 list_move(&tp->tp_list, &cull);
203 while (!list_empty(&cull)) {
204 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
206 list_del(&tp->tp_list);
207 LIBCFS_FREE(tp, sizeof(*tp));
213 fail_peer (lnet_nid_t nid, int outgoing)
215 struct lnet_test_peer *tp;
216 struct list_head *el;
217 struct list_head *next;
218 struct list_head cull;
221 INIT_LIST_HEAD(&cull);
223 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
226 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
227 tp = list_entry(el, struct lnet_test_peer, tp_list);
229 if (tp->tp_threshold == 0) {
232 /* only cull zombies on outgoing tests,
233 * since we may be at interrupt priority on
234 * incoming messages. */
235 list_move(&tp->tp_list, &cull);
240 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
241 nid == tp->tp_nid) { /* fail this peer */
244 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
247 tp->tp_threshold == 0) {
249 list_move(&tp->tp_list, &cull);
258 while (!list_empty(&cull)) {
259 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
260 list_del(&tp->tp_list);
262 LIBCFS_FREE(tp, sizeof(*tp));
269 lnet_iov_nob(unsigned int niov, struct kvec *iov)
271 unsigned int nob = 0;
273 LASSERT(niov == 0 || iov != NULL);
275 nob += (iov++)->iov_len;
279 EXPORT_SYMBOL(lnet_iov_nob);
282 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
283 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
286 /* NB diov, siov are READ-ONLY */
287 unsigned int this_nob;
292 /* skip complete frags before 'doffset' */
294 while (doffset >= diov->iov_len) {
295 doffset -= diov->iov_len;
301 /* skip complete frags before 'soffset' */
303 while (soffset >= siov->iov_len) {
304 soffset -= siov->iov_len;
313 this_nob = MIN(diov->iov_len - doffset,
314 siov->iov_len - soffset);
315 this_nob = MIN(this_nob, nob);
317 memcpy((char *)diov->iov_base + doffset,
318 (char *)siov->iov_base + soffset, this_nob);
321 if (diov->iov_len > doffset + this_nob) {
329 if (siov->iov_len > soffset + this_nob) {
338 EXPORT_SYMBOL(lnet_copy_iov2iov);
341 lnet_extract_iov(int dst_niov, struct kvec *dst,
342 int src_niov, struct kvec *src,
343 unsigned int offset, unsigned int len)
345 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
346 * for exactly 'len' bytes, and return the number of entries.
347 * NB not destructive to 'src' */
348 unsigned int frag_len;
351 if (len == 0) /* no data => */
352 return (0); /* no frags */
354 LASSERT(src_niov > 0);
355 while (offset >= src->iov_len) { /* skip initial frags */
356 offset -= src->iov_len;
359 LASSERT(src_niov > 0);
364 LASSERT(src_niov > 0);
365 LASSERT((int)niov <= dst_niov);
367 frag_len = src->iov_len - offset;
368 dst->iov_base = ((char *)src->iov_base) + offset;
370 if (len <= frag_len) {
375 dst->iov_len = frag_len;
385 EXPORT_SYMBOL(lnet_extract_iov);
389 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
391 unsigned int nob = 0;
393 LASSERT(niov == 0 || kiov != NULL);
395 nob += (kiov++)->kiov_len;
399 EXPORT_SYMBOL(lnet_kiov_nob);
402 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
403 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
406 /* NB diov, siov are READ-ONLY */
407 unsigned int this_nob;
414 LASSERT (!in_interrupt ());
417 while (doffset >= diov->kiov_len) {
418 doffset -= diov->kiov_len;
425 while (soffset >= siov->kiov_len) {
426 soffset -= siov->kiov_len;
435 this_nob = MIN(diov->kiov_len - doffset,
436 siov->kiov_len - soffset);
437 this_nob = MIN(this_nob, nob);
440 daddr = ((char *)kmap(diov->kiov_page)) +
441 diov->kiov_offset + doffset;
443 saddr = ((char *)kmap(siov->kiov_page)) +
444 siov->kiov_offset + soffset;
446 /* Vanishing risk of kmap deadlock when mapping 2 pages.
447 * However in practice at least one of the kiovs will be mapped
448 * kernel pages and the map/unmap will be NOOPs */
450 memcpy (daddr, saddr, this_nob);
453 if (diov->kiov_len > doffset + this_nob) {
457 kunmap(diov->kiov_page);
464 if (siov->kiov_len > soffset + this_nob) {
468 kunmap(siov->kiov_page);
477 kunmap(diov->kiov_page);
479 kunmap(siov->kiov_page);
481 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
484 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
485 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
488 /* NB iov, kiov are READ-ONLY */
489 unsigned int this_nob;
495 LASSERT (!in_interrupt ());
498 while (iovoffset >= iov->iov_len) {
499 iovoffset -= iov->iov_len;
506 while (kiovoffset >= kiov->kiov_len) {
507 kiovoffset -= kiov->kiov_len;
516 this_nob = MIN(iov->iov_len - iovoffset,
517 kiov->kiov_len - kiovoffset);
518 this_nob = MIN(this_nob, nob);
521 addr = ((char *)kmap(kiov->kiov_page)) +
522 kiov->kiov_offset + kiovoffset;
524 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
527 if (iov->iov_len > iovoffset + this_nob) {
528 iovoffset += this_nob;
535 if (kiov->kiov_len > kiovoffset + this_nob) {
537 kiovoffset += this_nob;
539 kunmap(kiov->kiov_page);
549 kunmap(kiov->kiov_page);
551 EXPORT_SYMBOL(lnet_copy_kiov2iov);
554 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
555 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
558 /* NB kiov, iov are READ-ONLY */
559 unsigned int this_nob;
565 LASSERT (!in_interrupt ());
568 while (kiovoffset >= kiov->kiov_len) {
569 kiovoffset -= kiov->kiov_len;
576 while (iovoffset >= iov->iov_len) {
577 iovoffset -= iov->iov_len;
586 this_nob = MIN(kiov->kiov_len - kiovoffset,
587 iov->iov_len - iovoffset);
588 this_nob = MIN(this_nob, nob);
591 addr = ((char *)kmap(kiov->kiov_page)) +
592 kiov->kiov_offset + kiovoffset;
594 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
597 if (kiov->kiov_len > kiovoffset + this_nob) {
599 kiovoffset += this_nob;
601 kunmap(kiov->kiov_page);
608 if (iov->iov_len > iovoffset + this_nob) {
609 iovoffset += this_nob;
618 kunmap(kiov->kiov_page);
620 EXPORT_SYMBOL(lnet_copy_iov2kiov);
623 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
624 int src_niov, lnet_kiov_t *src,
625 unsigned int offset, unsigned int len)
627 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
628 * for exactly 'len' bytes, and return the number of entries.
629 * NB not destructive to 'src' */
630 unsigned int frag_len;
633 if (len == 0) /* no data => */
634 return (0); /* no frags */
636 LASSERT(src_niov > 0);
637 while (offset >= src->kiov_len) { /* skip initial frags */
638 offset -= src->kiov_len;
641 LASSERT(src_niov > 0);
646 LASSERT(src_niov > 0);
647 LASSERT((int)niov <= dst_niov);
649 frag_len = src->kiov_len - offset;
650 dst->kiov_page = src->kiov_page;
651 dst->kiov_offset = src->kiov_offset + offset;
653 if (len <= frag_len) {
655 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
659 dst->kiov_len = frag_len;
660 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
670 EXPORT_SYMBOL(lnet_extract_kiov);
673 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
674 int delayed, unsigned int offset, unsigned int mlen,
677 unsigned int niov = 0;
678 struct kvec *iov = NULL;
679 lnet_kiov_t *kiov = NULL;
682 LASSERT (!in_interrupt ());
683 LASSERT (mlen == 0 || msg != NULL);
686 LASSERT(msg->msg_receiving);
687 LASSERT(!msg->msg_sending);
688 LASSERT(rlen == msg->msg_len);
689 LASSERT(mlen <= msg->msg_len);
690 LASSERT(msg->msg_offset == offset);
691 LASSERT(msg->msg_wanted == mlen);
693 msg->msg_receiving = 0;
696 niov = msg->msg_niov;
698 kiov = msg->msg_kiov;
701 LASSERT ((iov == NULL) != (kiov == NULL));
705 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
706 niov, iov, kiov, offset, mlen,
709 lnet_finalize(msg, rc);
713 lnet_setpayloadbuffer(struct lnet_msg *msg)
715 struct lnet_libmd *md = msg->msg_md;
717 LASSERT(msg->msg_len > 0);
718 LASSERT(!msg->msg_routing);
720 LASSERT(msg->msg_niov == 0);
721 LASSERT(msg->msg_iov == NULL);
722 LASSERT(msg->msg_kiov == NULL);
724 msg->msg_niov = md->md_niov;
725 if ((md->md_options & LNET_MD_KIOV) != 0)
726 msg->msg_kiov = md->md_iov.kiov;
728 msg->msg_iov = md->md_iov.iov;
732 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
733 unsigned int offset, unsigned int len)
735 msg->msg_type = type;
736 msg->msg_target = target;
738 msg->msg_offset = offset;
741 lnet_setpayloadbuffer(msg);
743 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
744 msg->msg_hdr.type = cpu_to_le32(type);
745 /* dest_nid will be overwritten by lnet_select_pathway() */
746 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
747 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
748 /* src_nid will be set later */
749 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
750 msg->msg_hdr.payload_length = cpu_to_le32(len);
754 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
756 void *priv = msg->msg_private;
759 LASSERT (!in_interrupt ());
760 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
761 (msg->msg_txcredit && msg->msg_peertxcredit));
763 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
765 msg->msg_no_resend = true;
766 lnet_finalize(msg, rc);
771 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
775 LASSERT(!msg->msg_sending);
776 LASSERT(msg->msg_receiving);
777 LASSERT(!msg->msg_rx_ready_delay);
778 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
780 msg->msg_rx_ready_delay = 1;
781 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
784 CERROR("recv from %s / send to %s aborted: "
785 "eager_recv failed %d\n",
786 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
787 libcfs_id2str(msg->msg_target), rc);
788 LASSERT(rc < 0); /* required by my callers */
795 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
799 deadline = lpni->lpni_last_alive +
800 lpni->lpni_net->net_tunables.lct_peer_timeout;
803 * assume peer_ni is alive as long as we're within the configured
812 /* NB: returns 1 when alive, 0 when dead, negative when error;
813 * may drop the lnet_net_lock */
815 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
816 struct lnet_msg *msg)
818 time64_t now = ktime_get_seconds();
820 if (!lnet_peer_aliveness_enabled(lpni))
824 * If we're resending a message, let's attempt to send it even if
825 * the peer is down to fulfill our resend quota on the message
827 if (msg->msg_retry_count > 0)
830 /* try and send recovery messages irregardless */
831 if (msg->msg_recovery)
834 /* always send any responses */
835 if (msg->msg_type == LNET_MSG_ACK ||
836 msg->msg_type == LNET_MSG_REPLY)
839 if (!lnet_is_peer_deadline_passed(lpni, now))
842 return lnet_is_peer_ni_alive(lpni);
846 * \param msg The message to be sent.
847 * \param do_send True if lnet_ni_send() should be called in this function.
848 * lnet_send() is going to lnet_net_unlock immediately after this, so
849 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
851 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
852 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
853 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
854 * \retval -ECANCELED If the MD of the message has been unlinked.
857 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
859 struct lnet_peer_ni *lp = msg->msg_txpeer;
860 struct lnet_ni *ni = msg->msg_txni;
861 int cpt = msg->msg_tx_cpt;
862 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
864 /* non-lnet_send() callers have checked before */
865 LASSERT(!do_send || msg->msg_tx_delayed);
866 LASSERT(!msg->msg_receiving);
867 LASSERT(msg->msg_tx_committed);
868 /* can't get here if we're sending to the loopback interface */
869 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
871 /* NB 'lp' is always the next hop */
872 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
873 lnet_peer_alive_locked(ni, lp, msg) == 0) {
874 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
875 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
877 lnet_net_unlock(cpt);
879 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
881 LNET_STATS_TYPE_DROP);
883 lnet_incr_stats(&msg->msg_txni->ni_stats,
885 LNET_STATS_TYPE_DROP);
887 CNETERR("Dropping message for %s: peer not alive\n",
888 libcfs_id2str(msg->msg_target));
889 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
891 lnet_finalize(msg, -EHOSTUNREACH);
894 return -EHOSTUNREACH;
897 if (msg->msg_md != NULL &&
898 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
899 lnet_net_unlock(cpt);
901 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
902 "called on the MD/ME.\n",
903 libcfs_id2str(msg->msg_target));
905 msg->msg_no_resend = true;
906 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
907 msg, libcfs_id2str(msg->msg_target));
908 lnet_finalize(msg, -ECANCELED);
915 if (!msg->msg_peertxcredit) {
916 spin_lock(&lp->lpni_lock);
917 LASSERT((lp->lpni_txcredits < 0) ==
918 !list_empty(&lp->lpni_txq));
920 msg->msg_peertxcredit = 1;
921 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
922 lp->lpni_txcredits--;
924 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
925 lp->lpni_mintxcredits = lp->lpni_txcredits;
927 if (lp->lpni_txcredits < 0) {
928 msg->msg_tx_delayed = 1;
929 list_add_tail(&msg->msg_list, &lp->lpni_txq);
930 spin_unlock(&lp->lpni_lock);
931 return LNET_CREDIT_WAIT;
933 spin_unlock(&lp->lpni_lock);
936 if (!msg->msg_txcredit) {
937 LASSERT((tq->tq_credits < 0) ==
938 !list_empty(&tq->tq_delayed));
940 msg->msg_txcredit = 1;
942 atomic_dec(&ni->ni_tx_credits);
944 if (tq->tq_credits < tq->tq_credits_min)
945 tq->tq_credits_min = tq->tq_credits;
947 if (tq->tq_credits < 0) {
948 msg->msg_tx_delayed = 1;
949 list_add_tail(&msg->msg_list, &tq->tq_delayed);
950 return LNET_CREDIT_WAIT;
954 /* unset the tx_delay flag as we're going to send it now */
955 msg->msg_tx_delayed = 0;
958 lnet_net_unlock(cpt);
959 lnet_ni_send(ni, msg);
962 return LNET_CREDIT_OK;
966 static struct lnet_rtrbufpool *
967 lnet_msg2bufpool(struct lnet_msg *msg)
969 struct lnet_rtrbufpool *rbp;
972 LASSERT(msg->msg_rx_committed);
974 cpt = msg->msg_rx_cpt;
975 rbp = &the_lnet.ln_rtrpools[cpt][0];
977 LASSERT(msg->msg_len <= LNET_MTU);
978 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
980 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
987 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
989 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
990 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
991 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
992 * received or OK to receive */
993 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
994 struct lnet_peer *lp;
995 struct lnet_rtrbufpool *rbp;
996 struct lnet_rtrbuf *rb;
998 LASSERT(msg->msg_iov == NULL);
999 LASSERT(msg->msg_kiov == NULL);
1000 LASSERT(msg->msg_niov == 0);
1001 LASSERT(msg->msg_routing);
1002 LASSERT(msg->msg_receiving);
1003 LASSERT(!msg->msg_sending);
1004 LASSERT(lpni->lpni_peer_net);
1005 LASSERT(lpni->lpni_peer_net->lpn_peer);
1007 lp = lpni->lpni_peer_net->lpn_peer;
1009 /* non-lnet_parse callers only receive delayed messages */
1010 LASSERT(!do_recv || msg->msg_rx_delayed);
1012 if (!msg->msg_peerrtrcredit) {
1013 /* lpni_lock protects the credit manipulation */
1014 spin_lock(&lpni->lpni_lock);
1015 /* lp_lock protects the lp_rtrq */
1016 spin_lock(&lp->lp_lock);
1018 msg->msg_peerrtrcredit = 1;
1019 lpni->lpni_rtrcredits--;
1020 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1021 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1023 if (lpni->lpni_rtrcredits < 0) {
1024 /* must have checked eager_recv before here */
1025 LASSERT(msg->msg_rx_ready_delay);
1026 msg->msg_rx_delayed = 1;
1027 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1028 spin_unlock(&lp->lp_lock);
1029 spin_unlock(&lpni->lpni_lock);
1030 return LNET_CREDIT_WAIT;
1032 spin_unlock(&lp->lp_lock);
1033 spin_unlock(&lpni->lpni_lock);
1036 rbp = lnet_msg2bufpool(msg);
1038 if (!msg->msg_rtrcredit) {
1039 msg->msg_rtrcredit = 1;
1041 if (rbp->rbp_credits < rbp->rbp_mincredits)
1042 rbp->rbp_mincredits = rbp->rbp_credits;
1044 if (rbp->rbp_credits < 0) {
1045 /* must have checked eager_recv before here */
1046 LASSERT(msg->msg_rx_ready_delay);
1047 msg->msg_rx_delayed = 1;
1048 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1049 return LNET_CREDIT_WAIT;
1053 LASSERT(!list_empty(&rbp->rbp_bufs));
1054 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1055 list_del(&rb->rb_list);
1057 msg->msg_niov = rbp->rbp_npages;
1058 msg->msg_kiov = &rb->rb_kiov[0];
1060 /* unset the msg-rx_delayed flag since we're receiving the message */
1061 msg->msg_rx_delayed = 0;
1064 int cpt = msg->msg_rx_cpt;
1066 lnet_net_unlock(cpt);
1067 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1068 0, msg->msg_len, msg->msg_len);
1071 return LNET_CREDIT_OK;
1075 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1077 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1078 struct lnet_ni *txni = msg->msg_txni;
1079 struct lnet_msg *msg2;
1081 if (msg->msg_txcredit) {
1082 struct lnet_ni *ni = msg->msg_txni;
1083 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1085 /* give back NI txcredits */
1086 msg->msg_txcredit = 0;
1088 LASSERT((tq->tq_credits < 0) ==
1089 !list_empty(&tq->tq_delayed));
1092 atomic_inc(&ni->ni_tx_credits);
1093 if (tq->tq_credits <= 0) {
1094 msg2 = list_entry(tq->tq_delayed.next,
1095 struct lnet_msg, msg_list);
1096 list_del(&msg2->msg_list);
1098 LASSERT(msg2->msg_txni == ni);
1099 LASSERT(msg2->msg_tx_delayed);
1100 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1102 (void) lnet_post_send_locked(msg2, 1);
1106 if (msg->msg_peertxcredit) {
1107 /* give back peer txcredits */
1108 msg->msg_peertxcredit = 0;
1110 spin_lock(&txpeer->lpni_lock);
1111 LASSERT((txpeer->lpni_txcredits < 0) ==
1112 !list_empty(&txpeer->lpni_txq));
1114 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1115 LASSERT(txpeer->lpni_txqnob >= 0);
1117 txpeer->lpni_txcredits++;
1118 if (txpeer->lpni_txcredits <= 0) {
1121 msg2 = list_entry(txpeer->lpni_txq.next,
1122 struct lnet_msg, msg_list);
1123 list_del(&msg2->msg_list);
1124 spin_unlock(&txpeer->lpni_lock);
1126 LASSERT(msg2->msg_txpeer == txpeer);
1127 LASSERT(msg2->msg_tx_delayed);
1129 msg2_cpt = msg2->msg_tx_cpt;
1132 * The msg_cpt can be different from the msg2_cpt
1133 * so we need to make sure we lock the correct cpt
1135 * Once we call lnet_post_send_locked() it is no
1136 * longer safe to access msg2, since it could've
1137 * been freed by lnet_finalize(), but we still
1138 * need to relock the correct cpt, so we cache the
1139 * msg2_cpt for the purpose of the check that
1140 * follows the call to lnet_pose_send_locked().
1142 if (msg2_cpt != msg->msg_tx_cpt) {
1143 lnet_net_unlock(msg->msg_tx_cpt);
1144 lnet_net_lock(msg2_cpt);
1146 (void) lnet_post_send_locked(msg2, 1);
1147 if (msg2_cpt != msg->msg_tx_cpt) {
1148 lnet_net_unlock(msg2_cpt);
1149 lnet_net_lock(msg->msg_tx_cpt);
1152 spin_unlock(&txpeer->lpni_lock);
1157 msg->msg_txni = NULL;
1158 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1161 if (txpeer != NULL) {
1162 msg->msg_txpeer = NULL;
1163 lnet_peer_ni_decref_locked(txpeer);
1168 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1170 struct lnet_msg *msg;
1172 if (list_empty(&rbp->rbp_msgs))
1174 msg = list_entry(rbp->rbp_msgs.next,
1175 struct lnet_msg, msg_list);
1176 list_del(&msg->msg_list);
1178 (void)lnet_post_routed_recv_locked(msg, 1);
1182 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1184 struct lnet_msg *msg;
1185 struct lnet_msg *tmp;
1187 lnet_net_unlock(cpt);
1189 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1190 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1191 0, 0, 0, msg->msg_hdr.payload_length);
1192 list_del_init(&msg->msg_list);
1193 msg->msg_no_resend = true;
1194 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1195 lnet_finalize(msg, -ECANCELED);
1202 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1204 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1205 struct lnet_peer *lp;
1206 struct lnet_ni *rxni = msg->msg_rxni;
1207 struct lnet_msg *msg2;
1209 if (msg->msg_rtrcredit) {
1210 /* give back global router credits */
1211 struct lnet_rtrbuf *rb;
1212 struct lnet_rtrbufpool *rbp;
1214 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1215 * there until it gets one allocated, or aborts the wait
1217 LASSERT(msg->msg_kiov != NULL);
1219 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1222 msg->msg_kiov = NULL;
1223 msg->msg_rtrcredit = 0;
1225 LASSERT(rbp == lnet_msg2bufpool(msg));
1227 LASSERT((rbp->rbp_credits > 0) ==
1228 !list_empty(&rbp->rbp_bufs));
1230 /* If routing is now turned off, we just drop this buffer and
1231 * don't bother trying to return credits. */
1232 if (!the_lnet.ln_routing) {
1233 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1237 /* It is possible that a user has lowered the desired number of
1238 * buffers in this pool. Make sure we never put back
1239 * more buffers than the stated number. */
1240 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1241 /* Discard this buffer so we don't have too
1243 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1244 rbp->rbp_nbuffers--;
1246 list_add(&rb->rb_list, &rbp->rbp_bufs);
1248 if (rbp->rbp_credits <= 0)
1249 lnet_schedule_blocked_locked(rbp);
1254 if (msg->msg_peerrtrcredit) {
1256 LASSERT(rxpeerni->lpni_peer_net);
1257 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1259 lp = rxpeerni->lpni_peer_net->lpn_peer;
1261 /* give back peer router credits */
1262 msg->msg_peerrtrcredit = 0;
1264 spin_lock(&rxpeerni->lpni_lock);
1265 spin_lock(&lp->lp_lock);
1267 rxpeerni->lpni_rtrcredits++;
1269 /* drop all messages which are queued to be routed on that
1271 if (!the_lnet.ln_routing) {
1272 struct list_head drop;
1273 INIT_LIST_HEAD(&drop);
1274 list_splice_init(&lp->lp_rtrq, &drop);
1275 spin_unlock(&lp->lp_lock);
1276 spin_unlock(&rxpeerni->lpni_lock);
1277 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1278 } else if (!list_empty(&lp->lp_rtrq)) {
1281 msg2 = list_entry(lp->lp_rtrq.next,
1282 struct lnet_msg, msg_list);
1283 list_del(&msg2->msg_list);
1284 msg2_cpt = msg2->msg_rx_cpt;
1285 spin_unlock(&lp->lp_lock);
1286 spin_unlock(&rxpeerni->lpni_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);
1304 spin_unlock(&rxpeerni->lpni_lock);
1308 msg->msg_rxni = NULL;
1309 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1311 if (rxpeerni != NULL) {
1312 msg->msg_rxpeer = NULL;
1313 lnet_peer_ni_decref_locked(rxpeerni);
1318 lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1320 if (p1->lpni_txqnob < p2->lpni_txqnob)
1323 if (p1->lpni_txqnob > p2->lpni_txqnob)
1326 if (p1->lpni_txcredits > p2->lpni_txcredits)
1329 if (p1->lpni_txcredits < p2->lpni_txcredits)
1335 static struct lnet_peer_ni *
1336 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1337 struct lnet_peer *peer,
1338 struct lnet_peer_net *peer_net)
1341 * Look at the peer NIs for the destination peer that connect
1342 * to the chosen net. If a peer_ni is preferred when using the
1343 * best_ni to communicate, we use that one. If there is no
1344 * preferred peer_ni, or there are multiple preferred peer_ni,
1345 * the available transmit credits are used. If the transmit
1346 * credits are equal, we round-robin over the peer_ni.
1348 struct lnet_peer_ni *lpni = NULL;
1349 struct lnet_peer_ni *best_lpni = NULL;
1350 int best_lpni_credits = INT_MIN;
1351 bool preferred = false;
1353 int best_lpni_healthv = 0;
1356 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1358 * if the best_ni we've chosen aleady has this lpni
1359 * preferred, then let's use it
1362 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1364 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1365 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1370 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1373 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1374 libcfs_nid2str(lpni->lpni_nid),
1375 lpni->lpni_txcredits, best_lpni_credits,
1376 lpni->lpni_seq, best_lpni->lpni_seq);
1378 /* pick the healthiest peer ni */
1379 if (lpni_healthv < best_lpni_healthv) {
1381 } else if (lpni_healthv > best_lpni_healthv) {
1382 best_lpni_healthv = lpni_healthv;
1383 /* if this is a preferred peer use it */
1384 } else if (!preferred && ni_is_pref) {
1386 } else if (preferred && !ni_is_pref) {
1388 * this is not the preferred peer so let's ignore
1392 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1394 * We already have a peer that has more credits
1395 * available than this one. No need to consider
1396 * this peer further.
1399 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1401 * The best peer found so far and the current peer
1402 * have the same number of available credits let's
1403 * make sure to select between them using Round
1407 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1413 best_lpni_credits = lpni->lpni_txcredits;
1416 /* if we still can't find a peer ni then we can't reach it */
1418 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1419 LNET_NIDNET(dst_nid);
1420 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1421 libcfs_net2str(net_id));
1425 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1426 libcfs_nid2str(best_lpni->lpni_nid));
1432 * Prerequisite: the best_ni should already be set in the sd
1434 static inline struct lnet_peer_ni *
1435 lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid,
1436 struct lnet_peer *peer, __u32 net_id)
1438 struct lnet_peer_net *peer_net;
1441 * The gateway is Multi-Rail capable so now we must select the
1444 peer_net = lnet_peer_get_net_locked(peer, net_id);
1447 CERROR("gateway peer %s has no NI on net %s\n",
1448 libcfs_nid2str(peer->lp_primary_nid),
1449 libcfs_net2str(net_id));
1453 return lnet_select_peer_ni(lni, dst_nid, peer, peer_net);
1456 /* Compare route priorities and hop counts */
1458 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1460 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1461 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1463 if (r1->lr_priority < r2->lr_priority)
1466 if (r1->lr_priority > r2->lr_priority)
1469 if (r1_hops < r2_hops)
1472 if (r1_hops > r2_hops)
1478 static struct lnet_route *
1479 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1480 struct lnet_route **prev_route,
1481 struct lnet_peer_ni **gwni)
1483 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1484 struct lnet_route *best_route;
1485 struct lnet_route *last_route;
1486 struct lnet_route *route;
1489 __u32 any_net = LNET_NIDNET(LNET_NID_ANY);
1491 best_route = last_route = NULL;
1492 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1493 if (!lnet_is_route_alive(route))
1496 /* If the src_net is specified then we need to find an lpni
1499 restrict_net = src_net == any_net ? route->lr_lnet : src_net;
1501 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1505 best_route = last_route = route;
1508 CERROR("Gateway %s does not have a peer NI on net %s\n",
1509 libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1510 libcfs_net2str(restrict_net));
1515 /* no protection on below fields, but it's harmless */
1516 if (last_route->lr_seq - route->lr_seq < 0)
1519 rc = lnet_compare_routes(route, best_route);
1523 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1527 CERROR("Gateway %s does not have a peer NI on net %s\n",
1528 libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1529 libcfs_net2str(restrict_net));
1539 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1543 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1550 *prev_route = last_route;
1556 static struct lnet_ni *
1557 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1558 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1561 struct lnet_ni *ni = NULL;
1562 unsigned int shortest_distance;
1567 * If there is no peer_ni that we can send to on this network,
1568 * then there is no point in looking for a new best_ni here.
1570 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1573 if (best_ni == NULL) {
1574 shortest_distance = UINT_MAX;
1575 best_credits = INT_MIN;
1578 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1579 best_ni->ni_dev_cpt);
1580 best_credits = atomic_read(&best_ni->ni_tx_credits);
1581 best_healthv = atomic_read(&best_ni->ni_healthv);
1584 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1585 unsigned int distance;
1590 ni_credits = atomic_read(&ni->ni_tx_credits);
1591 ni_healthv = atomic_read(&ni->ni_healthv);
1592 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1595 * calculate the distance from the CPT on which
1596 * the message memory is allocated to the CPT of
1597 * the NI's physical device
1599 distance = cfs_cpt_distance(lnet_cpt_table(),
1603 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1604 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1605 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1606 : "not seleced", best_credits, shortest_distance,
1607 (best_ni) ? best_ni->ni_seq : 0);
1610 * All distances smaller than the NUMA range
1611 * are treated equally.
1613 if (distance < lnet_numa_range)
1614 distance = lnet_numa_range;
1617 * Select on health, shorter distance, available
1618 * credits, then round-robin.
1622 } else if (ni_healthv < best_healthv) {
1624 } else if (ni_healthv > best_healthv) {
1625 best_healthv = ni_healthv;
1627 * If we're going to prefer this ni because it's
1628 * the healthiest, then we should set the
1629 * shortest_distance in the algorithm in case
1630 * there are multiple NIs with the same health but
1631 * different distances.
1633 if (distance < shortest_distance)
1634 shortest_distance = distance;
1635 } else if (distance > shortest_distance) {
1637 } else if (distance < shortest_distance) {
1638 shortest_distance = distance;
1639 } else if (ni_credits < best_credits) {
1641 } else if (ni_credits == best_credits) {
1642 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1646 best_credits = ni_credits;
1649 CDEBUG(D_NET, "selected best_ni %s\n",
1650 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1656 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1657 * because such traffic is required to perform discovery. We therefore
1658 * exclude all GET and PUT on that portal. We also exclude all ACK and
1659 * REPLY traffic, but that is because the portal is not tracked in the
1660 * message structure for these message types. We could restrict this
1661 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1664 lnet_msg_discovery(struct lnet_msg *msg)
1666 if (msg->msg_type == LNET_MSG_PUT) {
1667 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1669 } else if (msg->msg_type == LNET_MSG_GET) {
1670 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1676 #define SRC_SPEC 0x0001
1677 #define SRC_ANY 0x0002
1678 #define LOCAL_DST 0x0004
1679 #define REMOTE_DST 0x0008
1680 #define MR_DST 0x0010
1681 #define NMR_DST 0x0020
1682 #define SND_RESP 0x0040
1684 /* The following to defines are used for return codes */
1685 #define REPEAT_SEND 0x1000
1686 #define PASS_THROUGH 0x2000
1688 /* The different cases lnet_select pathway needs to handle */
1689 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1690 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1691 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1692 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1693 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1694 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1695 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1696 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1699 lnet_handle_lo_send(struct lnet_send_data *sd)
1701 struct lnet_msg *msg = sd->sd_msg;
1702 int cpt = sd->sd_cpt;
1704 /* No send credit hassles with LOLND */
1705 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1706 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1707 if (!msg->msg_routing)
1708 msg->msg_hdr.src_nid =
1709 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1710 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1711 lnet_msg_commit(msg, cpt);
1712 msg->msg_txni = the_lnet.ln_loni;
1714 return LNET_CREDIT_OK;
1718 lnet_handle_send(struct lnet_send_data *sd)
1720 struct lnet_ni *best_ni = sd->sd_best_ni;
1721 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1722 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1723 struct lnet_msg *msg = sd->sd_msg;
1725 __u32 send_case = sd->sd_send_case;
1727 __u32 routing = send_case & REMOTE_DST;
1728 struct lnet_rsp_tracker *rspt;
1731 * Increment sequence number of the selected peer so that we
1732 * pick the next one in Round Robin.
1734 best_lpni->lpni_seq++;
1737 * grab a reference on the peer_ni so it sticks around even if
1738 * we need to drop and relock the lnet_net_lock below.
1740 lnet_peer_ni_addref_locked(best_lpni);
1743 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1744 * message. This ensures that we get a CPT that is correct for
1745 * the NI when the NI has been restricted to a subset of all CPTs.
1746 * If the selected CPT differs from the one currently locked, we
1747 * must unlock and relock the lnet_net_lock(), and then check whether
1748 * the configuration has changed. We don't have a hold on the best_ni
1749 * yet, and it may have vanished.
1751 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1752 if (sd->sd_cpt != cpt2) {
1753 __u32 seq = lnet_get_dlc_seq_locked();
1754 lnet_net_unlock(sd->sd_cpt);
1756 lnet_net_lock(sd->sd_cpt);
1757 if (seq != lnet_get_dlc_seq_locked()) {
1758 lnet_peer_ni_decref_locked(best_lpni);
1764 * store the best_lpni in the message right away to avoid having
1765 * to do the same operation under different conditions
1767 msg->msg_txpeer = best_lpni;
1768 msg->msg_txni = best_ni;
1771 * grab a reference for the best_ni since now it's in use in this
1772 * send. The reference will be dropped in lnet_finalize()
1774 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1777 * Always set the target.nid to the best peer picked. Either the
1778 * NID will be one of the peer NIDs selected, or the same NID as
1779 * what was originally set in the target or it will be the NID of
1780 * a router if this message should be routed
1782 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1785 * lnet_msg_commit assigns the correct cpt to the message, which
1786 * is used to decrement the correct refcount on the ni when it's
1787 * time to return the credits
1789 lnet_msg_commit(msg, sd->sd_cpt);
1792 * If we are routing the message then we keep the src_nid that was
1793 * set by the originator. If we are not routing then we are the
1794 * originator and set it here.
1796 if (!msg->msg_routing)
1797 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1800 msg->msg_target_is_router = 1;
1801 msg->msg_target.pid = LNET_PID_LUSTRE;
1803 * since we're routing we want to ensure that the
1804 * msg_hdr.dest_nid is set to the final destination. When
1805 * the router receives this message it knows how to route
1808 * final_dst_lpni is set at the beginning of the
1809 * lnet_select_pathway() function and is never changed.
1810 * It's safe to use it here.
1812 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1815 * if we're not routing set the dest_nid to the best peer
1816 * ni NID that we picked earlier in the algorithm.
1818 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1822 * if we have response tracker block update it with the next hop
1826 rspt = msg->msg_md->md_rspt_ptr;
1828 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1829 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1830 libcfs_nid2str(rspt->rspt_next_hop_nid));
1834 rc = lnet_post_send_locked(msg, 0);
1837 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1838 libcfs_nid2str(msg->msg_hdr.src_nid),
1839 libcfs_nid2str(msg->msg_txni->ni_nid),
1840 libcfs_nid2str(sd->sd_src_nid),
1841 libcfs_nid2str(msg->msg_hdr.dest_nid),
1842 libcfs_nid2str(sd->sd_dst_nid),
1843 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1844 libcfs_nid2str(sd->sd_rtr_nid),
1845 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1851 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1853 if (sd->sd_send_case & NMR_DST &&
1854 sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1855 sd->sd_msg->msg_type != LNET_MSG_ACK &&
1856 sd->sd_best_lpni->lpni_pref_nnids == 0) {
1857 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1858 libcfs_nid2str(sd->sd_best_ni->ni_nid),
1859 libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1860 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1861 sd->sd_best_ni->ni_nid);
1870 * use the source and destination NIDs as the pathway
1873 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1875 /* the destination lpni is set before we get here. */
1878 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1879 if (!sd->sd_best_ni) {
1880 CERROR("Can't send to %s: src %s is not a "
1881 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1882 libcfs_nid2str(sd->sd_src_nid));
1887 * the preferred NID will only be set for NMR peers
1889 lnet_set_non_mr_pref_nid(sd);
1891 return lnet_handle_send(sd);
1899 * Don't run the selection algorithm on the peer NIs. By specifying the
1900 * local NID, we're also saying that we should always use the destination NID
1901 * provided. This handles the case where we should be using the same
1902 * destination NID for the all the messages which belong to the same RPC
1906 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1908 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1909 if (!sd->sd_best_ni) {
1910 CERROR("Can't send to %s: src %s is not a "
1911 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1912 libcfs_nid2str(sd->sd_src_nid));
1916 if (sd->sd_best_lpni &&
1917 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1918 return lnet_handle_lo_send(sd);
1919 else if (sd->sd_best_lpni)
1920 return lnet_handle_send(sd);
1922 CERROR("can't send to %s. no NI on %s\n",
1923 libcfs_nid2str(sd->sd_dst_nid),
1924 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1926 return -EHOSTUNREACH;
1930 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1931 struct lnet_peer *peer,
1932 struct lnet_peer_net *peer_net,
1936 struct lnet_net *local_net;
1937 struct lnet_ni *best_ni;
1939 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1944 * Iterate through the NIs in this local Net and select
1945 * the NI to send from. The selection is determined by
1946 * these 3 criterion in the following priority:
1948 * 2. NI available credits
1951 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1952 peer, peer_net, cpt);
1954 if (incr_seq && best_ni)
1961 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1962 struct lnet_msg *msg, lnet_nid_t rtr_nid,
1965 struct lnet_peer *peer;
1966 lnet_nid_t primary_nid;
1969 lnet_peer_ni_addref_locked(lpni);
1971 peer = lpni->lpni_peer_net->lpn_peer;
1973 if (lnet_peer_gw_discovery(peer)) {
1974 lnet_peer_ni_decref_locked(lpni);
1978 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1979 lnet_peer_ni_decref_locked(lpni);
1983 rc = lnet_discover_peer_locked(lpni, cpt, false);
1985 lnet_peer_ni_decref_locked(lpni);
1988 /* The peer may have changed. */
1989 peer = lpni->lpni_peer_net->lpn_peer;
1990 spin_lock(&peer->lp_lock);
1991 if (lnet_peer_is_uptodate_locked(peer)) {
1992 spin_unlock(&peer->lp_lock);
1993 lnet_peer_ni_decref_locked(lpni);
1996 /* queue message and return */
1997 msg->msg_rtr_nid_param = rtr_nid;
1998 msg->msg_sending = 0;
1999 msg->msg_txpeer = NULL;
2000 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2001 primary_nid = peer->lp_primary_nid;
2002 spin_unlock(&peer->lp_lock);
2004 lnet_peer_ni_decref_locked(lpni);
2006 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2007 msg, libcfs_nid2str(primary_nid));
2009 return LNET_DC_WAIT;
2013 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2015 struct lnet_peer_ni **gw_lpni,
2016 struct lnet_peer **gw_peer)
2020 struct lnet_peer *gw;
2021 struct lnet_peer *lp;
2022 struct lnet_peer_net *lpn;
2023 struct lnet_peer_net *best_lpn = NULL;
2024 struct lnet_remotenet *rnet, *best_rnet = NULL;
2025 struct lnet_route *best_route = NULL;
2026 struct lnet_route *last_route = NULL;
2027 struct lnet_peer_ni *lpni = NULL;
2028 struct lnet_peer_ni *gwni = NULL;
2029 lnet_nid_t src_nid = sd->sd_src_nid;
2031 /* If a router nid was specified then we are replying to a GET or
2032 * sending an ACK. In this case we use the gateway associated with the
2033 * specified router nid.
2035 if (sd->sd_rtr_nid != LNET_NID_ANY) {
2036 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2038 CERROR("No peer NI for gateway %s\n",
2039 libcfs_nid2str(sd->sd_rtr_nid));
2040 return -EHOSTUNREACH;
2042 gw = gwni->lpni_peer_net->lpn_peer;
2043 lnet_peer_ni_decref_locked(gwni);
2044 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2046 /* we've already looked up the initial lpni using dst_nid */
2047 lpni = sd->sd_best_lpni;
2048 /* the peer tree must be in existence */
2049 LASSERT(lpni && lpni->lpni_peer_net &&
2050 lpni->lpni_peer_net->lpn_peer);
2051 lp = lpni->lpni_peer_net->lpn_peer;
2053 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2054 /* is this remote network reachable? */
2055 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2064 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2072 CERROR("peer %s has no available nets\n",
2073 libcfs_nid2str(sd->sd_dst_nid));
2074 return -EHOSTUNREACH;
2077 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni,
2080 best_lpn->lpn_net_id);
2081 if (!sd->sd_best_lpni) {
2082 CERROR("peer %s down\n",
2083 libcfs_nid2str(sd->sd_dst_nid));
2084 return -EHOSTUNREACH;
2087 best_route = lnet_find_route_locked(best_rnet,
2088 LNET_NIDNET(src_nid),
2089 &last_route, &gwni);
2091 CERROR("no route to %s from %s\n",
2092 libcfs_nid2str(dst_nid),
2093 libcfs_nid2str(src_nid));
2094 return -EHOSTUNREACH;
2098 CERROR("Internal Error. Route expected to %s from %s\n",
2099 libcfs_nid2str(dst_nid),
2100 libcfs_nid2str(src_nid));
2104 gw = best_route->lr_gateway;
2105 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2106 local_lnet = best_route->lr_lnet;
2111 * Discover this gateway if it hasn't already been discovered.
2112 * This means we might delay the message until discovery has
2115 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2116 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid,
2121 if (!sd->sd_best_ni)
2122 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2123 lnet_peer_get_net_locked(gw,
2128 if (!sd->sd_best_ni) {
2129 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2130 libcfs_net2str(local_lnet),
2131 libcfs_nid2str(sd->sd_src_nid));
2139 * increment the sequence numbers since now we're sure we're
2140 * going to use this path
2142 if (sd->sd_rtr_nid == LNET_NID_ANY) {
2143 LASSERT(best_route && last_route);
2144 best_route->lr_seq = last_route->lr_seq + 1;
2145 best_lpn->lpn_seq++;
2156 * Remote destination
2157 * Non-MR destination
2161 * Remote destination
2164 * The handling of these two cases is similar. Even though the destination
2165 * can be MR or non-MR, we'll deal directly with the router.
2168 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2171 struct lnet_peer_ni *gw_lpni = NULL;
2172 struct lnet_peer *gw_peer = NULL;
2175 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2176 if (!sd->sd_best_ni) {
2177 CERROR("Can't send to %s: src %s is not a "
2178 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2179 libcfs_nid2str(sd->sd_src_nid));
2183 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2188 if (sd->sd_send_case & NMR_DST)
2190 * since the final destination is non-MR let's set its preferred
2191 * NID before we send
2193 lnet_set_non_mr_pref_nid(sd);
2196 * We're going to send to the gw found so let's set its
2199 sd->sd_peer = gw_peer;
2200 sd->sd_best_lpni = gw_lpni;
2202 return lnet_handle_send(sd);
2206 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2209 struct lnet_peer_net *peer_net = NULL;
2210 struct lnet_ni *best_ni = NULL;
2213 * The peer can have multiple interfaces, some of them can be on
2214 * the local network and others on a routed network. We should
2215 * prefer the local network. However if the local network is not
2216 * available then we need to try the routed network
2219 /* go through all the peer nets and find the best_ni */
2220 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2222 * The peer's list of nets can contain non-local nets. We
2223 * want to only examine the local ones.
2225 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2227 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2228 peer_net, md_cpt, false);
2231 * if this is a discovery message and lp_disc_net_id is
2232 * specified then use that net to send the discovery on.
2234 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2240 /* increment sequence number so we can round robin */
2246 static struct lnet_ni *
2247 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2249 struct lnet_ni *best_ni = NULL;
2250 struct lnet_peer_net *peer_net;
2251 struct lnet_peer *peer = sd->sd_peer;
2252 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2253 struct lnet_peer_ni *lpni;
2254 int cpt = sd->sd_cpt;
2257 * We must use a consistent source address when sending to a
2258 * non-MR peer. However, a non-MR peer can have multiple NIDs
2259 * on multiple networks, and we may even need to talk to this
2260 * peer on multiple networks -- certain types of
2261 * load-balancing configuration do this.
2263 * So we need to pick the NI the peer prefers for this
2264 * particular network.
2267 /* Get the target peer_ni */
2268 peer_net = lnet_peer_get_net_locked(peer,
2269 LNET_NIDNET(best_lpni->lpni_nid));
2270 LASSERT(peer_net != NULL);
2271 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2273 if (lpni->lpni_pref_nnids == 0)
2275 LASSERT(lpni->lpni_pref_nnids == 1);
2276 best_ni = lnet_nid2ni_locked(
2277 lpni->lpni_pref.nid, cpt);
2284 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2286 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2288 struct lnet_ni *best_ni = NULL;
2289 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2292 * We must use a consistent source address when sending to a
2293 * non-MR peer. However, a non-MR peer can have multiple NIDs
2294 * on multiple networks, and we may even need to talk to this
2295 * peer on multiple networks -- certain types of
2296 * load-balancing configuration do this.
2298 * So we need to pick the NI the peer prefers for this
2299 * particular network.
2302 best_ni = lnet_find_existing_preferred_best_ni(sd);
2304 /* if best_ni is still not set just pick one */
2307 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2308 sd->sd_best_lpni->lpni_peer_net,
2309 sd->sd_md_cpt, true);
2310 /* If there is no best_ni we don't have a route */
2312 CERROR("no path to %s from net %s\n",
2313 libcfs_nid2str(best_lpni->lpni_nid),
2314 libcfs_net2str(best_lpni->lpni_net->net_id));
2315 return -EHOSTUNREACH;
2319 sd->sd_best_ni = best_ni;
2321 /* Set preferred NI if necessary. */
2322 lnet_set_non_mr_pref_nid(sd);
2329 * Source not specified
2333 * always use the same source NID for NMR peers
2334 * If we've talked to that peer before then we already have a preferred
2335 * source NI associated with it. Otherwise, we select a preferred local NI
2336 * and store it in the peer
2339 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2343 /* sd->sd_best_lpni is already set to the final destination */
2346 * At this point we should've created the peer ni and peer. If we
2347 * can't find it, then something went wrong. Instead of assert
2348 * output a relevant message and fail the send
2350 if (!sd->sd_best_lpni) {
2351 CERROR("Internal fault. Unable to send msg %s to %s. "
2353 lnet_msgtyp2str(sd->sd_msg->msg_type),
2354 libcfs_nid2str(sd->sd_dst_nid));
2358 rc = lnet_select_preferred_best_ni(sd);
2360 rc = lnet_handle_send(sd);
2366 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2369 * NOTE we've already handled the remote peer case. So we only
2370 * need to worry about the local case here.
2372 * if we're sending a response, ACK or reply, we need to send it
2373 * to the destination NID given to us. At this point we already
2374 * have the peer_ni we're suppose to send to, so just find the
2375 * best_ni on the peer net and use that. Since we're sending to an
2376 * MR peer then we can just run the selection algorithm on our
2377 * local NIs and pick the best one.
2379 if (sd->sd_send_case & SND_RESP) {
2381 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2382 sd->sd_best_lpni->lpni_peer_net,
2383 sd->sd_md_cpt, true);
2385 if (!sd->sd_best_ni) {
2387 * We're not going to deal with not able to send
2388 * a response to the provided final destination
2390 CERROR("Can't send response to %s. "
2391 "No local NI available\n",
2392 libcfs_nid2str(sd->sd_dst_nid));
2393 return -EHOSTUNREACH;
2396 return lnet_handle_send(sd);
2400 * If we get here that means we're sending a fresh request, PUT or
2401 * GET, so we need to run our standard selection algorithm.
2402 * First find the best local interface that's on any of the peer's
2405 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2407 lnet_msg_discovery(sd->sd_msg));
2408 if (sd->sd_best_ni) {
2410 lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid,
2412 sd->sd_best_ni->ni_net->net_id);
2415 * if we're successful in selecting a peer_ni on the local
2416 * network, then send to it. Otherwise fall through and
2417 * try and see if we can reach it over another routed
2420 if (sd->sd_best_lpni &&
2421 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2423 * in case we initially started with a routed
2424 * destination, let's reset to local
2426 sd->sd_send_case &= ~REMOTE_DST;
2427 sd->sd_send_case |= LOCAL_DST;
2428 return lnet_handle_lo_send(sd);
2429 } else if (sd->sd_best_lpni) {
2431 * in case we initially started with a routed
2432 * destination, let's reset to local
2434 sd->sd_send_case &= ~REMOTE_DST;
2435 sd->sd_send_case |= LOCAL_DST;
2436 return lnet_handle_send(sd);
2439 CERROR("Internal Error. Expected to have a best_lpni: "
2441 libcfs_nid2str(sd->sd_src_nid),
2442 libcfs_nid2str(sd->sd_dst_nid));
2448 * Peer doesn't have a local network. Let's see if there is
2449 * a remote network we can reach it on.
2451 return PASS_THROUGH;
2456 * Source NID not specified
2461 * Source NID not speified
2462 * Remote destination
2465 * In both of these cases if we're sending a response, ACK or REPLY, then
2466 * we need to send to the destination NID provided.
2468 * In the remote case let's deal with MR routers.
2473 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2476 struct lnet_peer *gw_peer = NULL;
2477 struct lnet_peer_ni *gw_lpni = NULL;
2480 * handle sending a response to a remote peer here so we don't
2481 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2483 if (sd->sd_send_case & REMOTE_DST &&
2484 sd->sd_send_case & SND_RESP) {
2485 struct lnet_peer_ni *gw;
2486 struct lnet_peer *gw_peer;
2488 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2491 CERROR("Can't send response to %s. "
2492 "No route available\n",
2493 libcfs_nid2str(sd->sd_dst_nid));
2494 return -EHOSTUNREACH;
2495 } else if (rc > 0) {
2499 sd->sd_best_lpni = gw;
2500 sd->sd_peer = gw_peer;
2502 return lnet_handle_send(sd);
2506 * Even though the NID for the peer might not be on a local network,
2507 * since the peer is MR there could be other interfaces on the
2508 * local network. In that case we'd still like to prefer the local
2509 * network over the routed network. If we're unable to do that
2510 * then we select the best router among the different routed networks,
2511 * and if the router is MR then we can deal with it as such.
2513 rc = lnet_handle_any_mr_dsta(sd);
2514 if (rc != PASS_THROUGH)
2518 * Now that we must route to the destination, we must consider the
2519 * MR case, where the destination has multiple interfaces, some of
2520 * which we can route to and others we do not. For this reason we
2521 * need to select the destination which we can route to and if
2522 * there are multiple, we need to round robin.
2524 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2529 sd->sd_send_case &= ~LOCAL_DST;
2530 sd->sd_send_case |= REMOTE_DST;
2532 sd->sd_peer = gw_peer;
2533 sd->sd_best_lpni = gw_lpni;
2535 return lnet_handle_send(sd);
2539 * Source not specified
2540 * Remote destination
2543 * Must send to the specified peer NID using the same source NID that
2544 * we've used before. If it's the first time to talk to that peer then
2545 * find the source NI and assign it as preferred to that peer
2548 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2551 struct lnet_peer_ni *gw_lpni = NULL;
2552 struct lnet_peer *gw_peer = NULL;
2555 * Let's set if we have a preferred NI to talk to this NMR peer
2557 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2560 * find the router and that'll find the best NI if we didn't find
2563 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2569 * set the best_ni we've chosen as the preferred one for
2572 lnet_set_non_mr_pref_nid(sd);
2574 /* we'll be sending to the gw */
2575 sd->sd_best_lpni = gw_lpni;
2576 sd->sd_peer = gw_peer;
2578 return lnet_handle_send(sd);
2582 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2585 * turn off the SND_RESP bit.
2586 * It will be checked in the case handling
2588 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2590 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2591 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2592 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2593 (send_case & MR_DST) ? "MR: " : "NMR: ",
2594 libcfs_nid2str(sd->sd_dst_nid),
2595 (send_case & LOCAL_DST) ? "local" : "routed");
2597 switch (send_case) {
2599 * For all cases where the source is specified, we should always
2600 * use the destination NID, whether it's an MR destination or not,
2601 * since we're continuing a series of related messages for the
2604 case SRC_SPEC_LOCAL_NMR_DST:
2605 return lnet_handle_spec_local_nmr_dst(sd);
2606 case SRC_SPEC_LOCAL_MR_DST:
2607 return lnet_handle_spec_local_mr_dst(sd);
2608 case SRC_SPEC_ROUTER_NMR_DST:
2609 case SRC_SPEC_ROUTER_MR_DST:
2610 return lnet_handle_spec_router_dst(sd);
2611 case SRC_ANY_LOCAL_NMR_DST:
2612 return lnet_handle_any_local_nmr_dst(sd);
2613 case SRC_ANY_LOCAL_MR_DST:
2614 case SRC_ANY_ROUTER_MR_DST:
2615 return lnet_handle_any_mr_dst(sd);
2616 case SRC_ANY_ROUTER_NMR_DST:
2617 return lnet_handle_any_router_nmr_dst(sd);
2619 CERROR("Unknown send case\n");
2625 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2626 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2628 struct lnet_peer_ni *lpni;
2629 struct lnet_peer *peer;
2630 struct lnet_send_data send_data;
2633 __u32 send_case = 0;
2635 memset(&send_data, 0, sizeof(send_data));
2638 * get an initial CPT to use for locking. The idea here is not to
2639 * serialize the calls to select_pathway, so that as many
2640 * operations can run concurrently as possible. To do that we use
2641 * the CPT where this call is being executed. Later on when we
2642 * determine the CPT to use in lnet_message_commit, we switch the
2643 * lock and check if there was any configuration change. If none,
2644 * then we proceed, if there is, then we restart the operation.
2646 cpt = lnet_net_lock_current();
2648 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2649 if (md_cpt == CFS_CPT_ANY)
2655 * If we're sending to ourselves then there is no need to go through
2656 * any selection. We can shortcut the entire process and send over
2659 * However, we make two exceptions to this rule:
2660 * 1. If the src_nid is specified then our API defines that we must send
2661 * via that interface.
2662 * 2. Recovery messages must be sent to the lnet_ni that is being
2665 send_data.sd_msg = msg;
2666 send_data.sd_cpt = cpt;
2667 if (src_nid == LNET_NID_ANY && !msg->msg_recovery &&
2668 lnet_nid2ni_locked(dst_nid, cpt)) {
2669 rc = lnet_handle_lo_send(&send_data);
2670 lnet_net_unlock(cpt);
2675 * find an existing peer_ni, or create one and mark it as having been
2676 * created due to network traffic. This call will create the
2677 * peer->peer_net->peer_ni tree.
2679 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2681 lnet_net_unlock(cpt);
2682 return PTR_ERR(lpni);
2686 * Cache the original src_nid. If we need to resend the message
2687 * then we'll need to know whether the src_nid was originally
2688 * specified for this message. If it was originally specified,
2689 * then we need to keep using the same src_nid since it's
2690 * continuing the same sequence of messages.
2692 msg->msg_src_nid_param = src_nid;
2695 * If necessary, perform discovery on the peer that owns this peer_ni.
2696 * Note, this can result in the ownership of this peer_ni changing
2697 * to another peer object.
2699 rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2701 lnet_peer_ni_decref_locked(lpni);
2702 lnet_net_unlock(cpt);
2705 lnet_peer_ni_decref_locked(lpni);
2707 peer = lpni->lpni_peer_net->lpn_peer;
2710 * Identify the different send cases
2712 if (src_nid == LNET_NID_ANY)
2713 send_case |= SRC_ANY;
2715 send_case |= SRC_SPEC;
2717 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2718 send_case |= LOCAL_DST;
2720 send_case |= REMOTE_DST;
2723 * if this is a non-MR peer or if we're recovering a peer ni then
2724 * let's consider this an NMR case so we can hit the destination
2727 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2728 send_case |= NMR_DST;
2730 send_case |= MR_DST;
2732 if (msg->msg_type == LNET_MSG_REPLY ||
2733 msg->msg_type == LNET_MSG_ACK)
2734 send_case |= SND_RESP;
2736 /* assign parameters to the send_data */
2737 send_data.sd_rtr_nid = rtr_nid;
2738 send_data.sd_src_nid = src_nid;
2739 send_data.sd_dst_nid = dst_nid;
2740 send_data.sd_best_lpni = lpni;
2742 * keep a pointer to the final destination in case we're going to
2743 * route, so we'll need to access it later
2745 send_data.sd_final_dst_lpni = lpni;
2746 send_data.sd_peer = peer;
2747 send_data.sd_md_cpt = md_cpt;
2748 send_data.sd_send_case = send_case;
2750 rc = lnet_handle_send_case_locked(&send_data);
2753 * Update the local cpt since send_data.sd_cpt might've been
2754 * updated as a result of calling lnet_handle_send_case_locked().
2756 cpt = send_data.sd_cpt;
2758 if (rc == REPEAT_SEND)
2761 lnet_net_unlock(cpt);
2767 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2769 lnet_nid_t dst_nid = msg->msg_target.nid;
2773 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2774 * but we might want to use pre-determined router for ACK/REPLY
2777 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2778 LASSERT(msg->msg_txpeer == NULL);
2779 LASSERT(msg->msg_txni == NULL);
2780 LASSERT(!msg->msg_sending);
2781 LASSERT(!msg->msg_target_is_router);
2782 LASSERT(!msg->msg_receiving);
2784 msg->msg_sending = 1;
2786 LASSERT(!msg->msg_tx_committed);
2788 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2790 if (rc == -EHOSTUNREACH)
2791 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2793 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2797 if (rc == LNET_CREDIT_OK)
2798 lnet_ni_send(msg->msg_txni, msg);
2800 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2804 enum lnet_mt_event_type {
2805 MT_TYPE_LOCAL_NI = 0,
2809 struct lnet_mt_event_info {
2810 enum lnet_mt_event_type mt_type;
2814 /* called with res_lock held */
2816 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2818 struct lnet_rsp_tracker *rspt;
2821 * msg has a refcount on the MD so the MD is not going away.
2822 * The rspt queue for the cpt is protected by
2823 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2825 if (!md->md_rspt_ptr)
2828 rspt = md->md_rspt_ptr;
2831 LASSERT(rspt->rspt_cpt == cpt);
2833 md->md_rspt_ptr = NULL;
2835 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2837 * The monitor thread has invalidated this handle because the
2838 * response timed out, but it failed to lookup the MD. That
2839 * means this response tracker is on the zombie list. We can
2840 * safely remove it under the resource lock (held by caller) and
2841 * free the response tracker block.
2843 list_del(&rspt->rspt_on_list);
2844 lnet_rspt_free(rspt, cpt);
2847 * invalidate the handle to indicate that a response has been
2848 * received, which will then lead the monitor thread to clean up
2851 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2856 lnet_clean_zombie_rstqs(void)
2858 struct lnet_rsp_tracker *rspt, *tmp;
2861 cfs_cpt_for_each(i, lnet_cpt_table()) {
2862 list_for_each_entry_safe(rspt, tmp,
2863 the_lnet.ln_mt_zombie_rstqs[i],
2865 list_del(&rspt->rspt_on_list);
2866 lnet_rspt_free(rspt, i);
2870 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2874 lnet_finalize_expired_responses(void)
2876 struct lnet_libmd *md;
2877 struct list_head local_queue;
2878 struct lnet_rsp_tracker *rspt, *tmp;
2882 if (the_lnet.ln_mt_rstq == NULL)
2885 cfs_cpt_for_each(i, lnet_cpt_table()) {
2886 INIT_LIST_HEAD(&local_queue);
2889 if (!the_lnet.ln_mt_rstq[i]) {
2893 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2898 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2900 * The rspt mdh will be invalidated when a response
2901 * is received or whenever we want to discard the
2902 * block the monitor thread will walk the queue
2903 * and clean up any rsts with an invalid mdh.
2904 * The monitor thread will walk the queue until
2905 * the first unexpired rspt block. This means that
2906 * some rspt blocks which received their
2907 * corresponding responses will linger in the
2908 * queue until they are cleaned up eventually.
2911 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2913 list_del(&rspt->rspt_on_list);
2914 lnet_rspt_free(rspt, i);
2918 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2919 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2920 struct lnet_peer_ni *lpni;
2923 md = lnet_handle2md(&rspt->rspt_mdh);
2925 /* MD has been queued for unlink, but
2926 * rspt hasn't been detached (Note we've
2927 * checked above that the rspt_mdh is
2928 * valid). Since we cannot lookup the MD
2929 * we're unable to detach the rspt
2930 * ourselves. Thus, move the rspt to the
2931 * zombie list where we'll wait for
2933 * 1. The remaining operations on the
2934 * MD to complete. In this case the
2935 * final operation will result in
2936 * lnet_msg_detach_md()->
2937 * lnet_detach_rsp_tracker() where
2938 * we will clean up this response
2940 * 2. LNet to shutdown. In this case
2941 * we'll wait until after all LND Nets
2942 * have shutdown and then we can
2943 * safely free any remaining response
2944 * tracker blocks on the zombie list.
2945 * Note: We need to hold the resource
2946 * lock when adding to the zombie list
2947 * because we may have concurrent access
2948 * with lnet_detach_rsp_tracker().
2950 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2951 list_move(&rspt->rspt_on_list,
2952 the_lnet.ln_mt_zombie_rstqs[i]);
2956 LASSERT(md->md_rspt_ptr == rspt);
2957 md->md_rspt_ptr = NULL;
2960 LNetMDUnlink(rspt->rspt_mdh);
2962 nid = rspt->rspt_next_hop_nid;
2964 list_del(&rspt->rspt_on_list);
2965 lnet_rspt_free(rspt, i);
2967 /* If we're shutting down we just want to clean
2968 * up the rspt blocks
2970 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2974 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2978 "Response timeout: md = %p: nid = %s\n",
2979 md, libcfs_nid2str(nid));
2982 * If there is a timeout on the response
2983 * from the next hop decrement its health
2984 * value so that we don't use it
2987 lpni = lnet_find_peer_ni_locked(nid);
2989 lnet_handle_remote_failure_locked(lpni);
2990 lnet_peer_ni_decref_locked(lpni);
2999 if (!list_empty(&local_queue)) {
3001 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3008 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3010 struct lnet_msg *msg;
3012 while (!list_empty(resendq)) {
3013 struct lnet_peer_ni *lpni;
3015 msg = list_entry(resendq->next, struct lnet_msg,
3018 list_del_init(&msg->msg_list);
3020 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3022 lnet_net_unlock(cpt);
3023 CERROR("Expected that a peer is already created for %s\n",
3024 libcfs_nid2str(msg->msg_hdr.dest_nid));
3025 msg->msg_no_resend = true;
3026 lnet_finalize(msg, -EFAULT);
3029 struct lnet_peer *peer;
3031 lnet_nid_t src_nid = LNET_NID_ANY;
3034 * if this message is not being routed and the
3035 * peer is non-MR then we must use the same
3036 * src_nid that was used in the original send.
3037 * Otherwise if we're routing the message (IE
3038 * we're a router) then we can use any of our
3039 * local interfaces. It doesn't matter to the
3040 * final destination.
3042 peer = lpni->lpni_peer_net->lpn_peer;
3043 if (!msg->msg_routing &&
3044 !lnet_peer_is_multi_rail(peer))
3045 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
3048 * If we originally specified a src NID, then we
3049 * must attempt to reuse it in the resend as well.
3051 if (msg->msg_src_nid_param != LNET_NID_ANY)
3052 src_nid = msg->msg_src_nid_param;
3053 lnet_peer_ni_decref_locked(lpni);
3055 lnet_net_unlock(cpt);
3056 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3057 libcfs_nid2str(src_nid),
3058 libcfs_id2str(msg->msg_target),
3059 lnet_msgtyp2str(msg->msg_type),
3061 msg->msg_retry_count);
3062 rc = lnet_send(src_nid, msg, LNET_NID_ANY);
3064 CERROR("Error sending %s to %s: %d\n",
3065 lnet_msgtyp2str(msg->msg_type),
3066 libcfs_id2str(msg->msg_target), rc);
3067 msg->msg_no_resend = true;
3068 lnet_finalize(msg, rc);
3072 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3078 lnet_resend_pending_msgs(void)
3082 cfs_cpt_for_each(i, lnet_cpt_table()) {
3084 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3089 /* called with cpt and ni_lock held */
3091 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3093 struct lnet_handle_md recovery_mdh;
3095 LNetInvalidateMDHandle(&recovery_mdh);
3097 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3099 recovery_mdh = ni->ni_ping_mdh;
3100 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3103 lnet_net_unlock(cpt);
3104 if (!LNetMDHandleIsInvalid(recovery_mdh))
3105 LNetMDUnlink(recovery_mdh);
3111 lnet_recover_local_nis(void)
3113 struct lnet_mt_event_info *ev_info;
3114 struct list_head processed_list;
3115 struct list_head local_queue;
3116 struct lnet_handle_md mdh;
3117 struct lnet_ni *tmp;
3123 INIT_LIST_HEAD(&local_queue);
3124 INIT_LIST_HEAD(&processed_list);
3127 * splice the recovery queue on a local queue. We will iterate
3128 * through the local queue and update it as needed. Once we're
3129 * done with the traversal, we'll splice the local queue back on
3130 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3131 * will be traversed in the next iteration.
3134 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3138 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3140 * if an NI is being deleted or it is now healthy, there
3141 * is no need to keep it around in the recovery queue.
3142 * The monitor thread is the only thread responsible for
3143 * removing the NI from the recovery queue.
3144 * Multiple threads can be adding NIs to the recovery
3147 healthv = atomic_read(&ni->ni_healthv);
3151 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3152 healthv == LNET_MAX_HEALTH_VALUE) {
3153 list_del_init(&ni->ni_recovery);
3154 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3156 lnet_ni_decref_locked(ni, 0);
3162 * if the local NI failed recovery we must unlink the md.
3163 * But we want to keep the local_ni on the recovery queue
3164 * so we can continue the attempts to recover it.
3166 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3167 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3168 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3175 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3176 libcfs_nid2str(ni->ni_nid));
3179 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3180 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3183 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3185 CERROR("out of memory. Can't recover %s\n",
3186 libcfs_nid2str(ni->ni_nid));
3188 ni->ni_recovery_state &=
3189 ~LNET_NI_RECOVERY_PENDING;
3194 mdh = ni->ni_ping_mdh;
3196 * Invalidate the ni mdh in case it's deleted.
3197 * We'll unlink the mdh in this case below.
3199 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3203 * remove the NI from the local queue and drop the
3204 * reference count to it while we're recovering
3205 * it. The reason for that, is that the NI could
3206 * be deleted, and the way the code is structured
3207 * is if we don't drop the NI, then the deletion
3208 * code will enter a loop waiting for the
3209 * reference count to be removed while holding the
3210 * ln_mutex_lock(). When we look up the peer to
3211 * send to in lnet_select_pathway() we will try to
3212 * lock the ln_mutex_lock() as well, leading to
3213 * a deadlock. By dropping the refcount and
3214 * removing it from the list, we allow for the NI
3215 * to be removed, then we use the cached NID to
3216 * look it up again. If it's gone, then we just
3217 * continue examining the rest of the queue.
3220 list_del_init(&ni->ni_recovery);
3221 lnet_ni_decref_locked(ni, 0);
3224 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3225 ev_info->mt_nid = nid;
3226 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3227 ev_info, the_lnet.ln_mt_eqh, true);
3228 /* lookup the nid again */
3230 ni = lnet_nid2ni_locked(nid, 0);
3233 * the NI has been deleted when we dropped
3241 * Same note as in lnet_recover_peer_nis(). When
3242 * we're sending the ping, the NI is free to be
3243 * deleted or manipulated. By this point it
3244 * could've been added back on the recovery queue,
3245 * and a refcount taken on it.
3246 * So we can't just add it blindly again or we'll
3247 * corrupt the queue. We must check under lock if
3248 * it's not on any list and if not then add it
3249 * to the processed list, which will eventually be
3250 * spliced back on to the recovery queue.
3252 ni->ni_ping_mdh = mdh;
3253 if (list_empty(&ni->ni_recovery)) {
3254 list_add_tail(&ni->ni_recovery, &processed_list);
3255 lnet_ni_addref_locked(ni, 0);
3261 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3267 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3268 * reexamined in the next iteration.
3270 list_splice_init(&processed_list, &local_queue);
3272 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3277 lnet_resendqs_create(void)
3279 struct list_head **resendqs;
3280 resendqs = lnet_create_array_of_queues();
3285 lnet_net_lock(LNET_LOCK_EX);
3286 the_lnet.ln_mt_resendqs = resendqs;
3287 lnet_net_unlock(LNET_LOCK_EX);
3293 lnet_clean_local_ni_recoveryq(void)
3297 /* This is only called when the monitor thread has stopped */
3300 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3301 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3302 struct lnet_ni, ni_recovery);
3303 list_del_init(&ni->ni_recovery);
3305 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3307 lnet_ni_decref_locked(ni, 0);
3314 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3317 struct lnet_handle_md recovery_mdh;
3319 LNetInvalidateMDHandle(&recovery_mdh);
3321 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3322 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3323 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3325 spin_unlock(&lpni->lpni_lock);
3326 lnet_net_unlock(cpt);
3327 if (!LNetMDHandleIsInvalid(recovery_mdh))
3328 LNetMDUnlink(recovery_mdh);
3330 spin_lock(&lpni->lpni_lock);
3334 lnet_clean_peer_ni_recoveryq(void)
3336 struct lnet_peer_ni *lpni, *tmp;
3338 lnet_net_lock(LNET_LOCK_EX);
3340 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3342 list_del_init(&lpni->lpni_recovery);
3343 spin_lock(&lpni->lpni_lock);
3344 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3345 spin_unlock(&lpni->lpni_lock);
3346 lnet_peer_ni_decref_locked(lpni);
3349 lnet_net_unlock(LNET_LOCK_EX);
3353 lnet_clean_resendqs(void)
3355 struct lnet_msg *msg, *tmp;
3356 struct list_head msgs;
3359 INIT_LIST_HEAD(&msgs);
3361 cfs_cpt_for_each(i, lnet_cpt_table()) {
3363 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3365 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3366 list_del_init(&msg->msg_list);
3367 msg->msg_no_resend = true;
3368 lnet_finalize(msg, -ESHUTDOWN);
3372 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3376 lnet_recover_peer_nis(void)
3378 struct lnet_mt_event_info *ev_info;
3379 struct list_head processed_list;
3380 struct list_head local_queue;
3381 struct lnet_handle_md mdh;
3382 struct lnet_peer_ni *lpni;
3383 struct lnet_peer_ni *tmp;
3388 INIT_LIST_HEAD(&local_queue);
3389 INIT_LIST_HEAD(&processed_list);
3392 * Always use cpt 0 for locking across all interactions with
3393 * ln_mt_peerNIRecovq
3396 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3400 list_for_each_entry_safe(lpni, tmp, &local_queue,
3403 * The same protection strategy is used here as is in the
3404 * local recovery case.
3407 healthv = atomic_read(&lpni->lpni_healthv);
3408 spin_lock(&lpni->lpni_lock);
3409 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3410 healthv == LNET_MAX_HEALTH_VALUE) {
3411 list_del_init(&lpni->lpni_recovery);
3412 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3413 spin_unlock(&lpni->lpni_lock);
3414 lnet_peer_ni_decref_locked(lpni);
3420 * If the peer NI has failed recovery we must unlink the
3421 * md. But we want to keep the peer ni on the recovery
3422 * queue so we can try to continue recovering it
3424 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3425 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3426 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3429 spin_unlock(&lpni->lpni_lock);
3433 * NOTE: we're racing with peer deletion from user space.
3434 * It's possible that a peer is deleted after we check its
3435 * state. In this case the recovery can create a new peer
3437 spin_lock(&lpni->lpni_lock);
3438 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3439 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3440 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3441 spin_unlock(&lpni->lpni_lock);
3443 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3445 CERROR("out of memory. Can't recover %s\n",
3446 libcfs_nid2str(lpni->lpni_nid));
3447 spin_lock(&lpni->lpni_lock);
3448 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3449 spin_unlock(&lpni->lpni_lock);
3453 /* look at the comments in lnet_recover_local_nis() */
3454 mdh = lpni->lpni_recovery_ping_mdh;
3455 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3456 nid = lpni->lpni_nid;
3458 list_del_init(&lpni->lpni_recovery);
3459 lnet_peer_ni_decref_locked(lpni);
3462 ev_info->mt_type = MT_TYPE_PEER_NI;
3463 ev_info->mt_nid = nid;
3464 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3465 ev_info, the_lnet.ln_mt_eqh, true);
3468 * lnet_find_peer_ni_locked() grabs a refcount for
3469 * us. No need to take it explicitly.
3471 lpni = lnet_find_peer_ni_locked(nid);
3478 lpni->lpni_recovery_ping_mdh = mdh;
3480 * While we're unlocked the lpni could've been
3481 * readded on the recovery queue. In this case we
3482 * don't need to add it to the local queue, since
3483 * it's already on there and the thread that added
3484 * it would've incremented the refcount on the
3485 * peer, which means we need to decref the refcount
3486 * that was implicitly grabbed by find_peer_ni_locked.
3487 * Otherwise, if the lpni is still not on
3488 * the recovery queue, then we'll add it to the
3491 if (list_empty(&lpni->lpni_recovery))
3492 list_add_tail(&lpni->lpni_recovery, &processed_list);
3494 lnet_peer_ni_decref_locked(lpni);
3497 spin_lock(&lpni->lpni_lock);
3499 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3501 spin_unlock(&lpni->lpni_lock);
3504 list_splice_init(&processed_list, &local_queue);
3506 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3511 lnet_monitor_thread(void *arg)
3513 time64_t recovery_timeout = 0;
3514 time64_t rsp_timeout = 0;
3518 wait_for_completion(&the_lnet.ln_started);
3520 * The monitor thread takes care of the following:
3521 * 1. Checks the aliveness of routers
3522 * 2. Checks if there are messages on the resend queue to resend
3524 * 3. Check if there are any NIs on the local recovery queue and
3526 * 4. Checks if there are any NIs on the remote recovery queue
3529 cfs_block_allsigs();
3531 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3532 now = ktime_get_real_seconds();
3534 if (lnet_router_checker_active())
3535 lnet_check_routers();
3537 lnet_resend_pending_msgs();
3539 if (now >= rsp_timeout) {
3540 lnet_finalize_expired_responses();
3541 rsp_timeout = now + (lnet_transaction_timeout / 2);
3544 if (now >= recovery_timeout) {
3545 lnet_recover_local_nis();
3546 lnet_recover_peer_nis();
3547 recovery_timeout = now + lnet_recovery_interval;
3551 * TODO do we need to check if we should sleep without
3552 * timeout? Technically, an active system will always
3553 * have messages in flight so this check will always
3554 * evaluate to false. And on an idle system do we care
3555 * if we wake up every 1 second? Although, we've seen
3556 * cases where we get a complaint that an idle thread
3557 * is waking up unnecessarily.
3559 * Take into account the current net_count when you wake
3560 * up for alive router checking, since we need to check
3561 * possibly as many networks as we have configured.
3563 interval = min(lnet_recovery_interval,
3564 min((unsigned int) alive_router_check_interval /
3565 lnet_current_net_count,
3566 lnet_transaction_timeout / 2));
3567 wait_for_completion_interruptible_timeout(
3568 &the_lnet.ln_mt_wait_complete,
3569 cfs_time_seconds(interval));
3570 /* Must re-init the completion before testing anything,
3571 * including ln_mt_state.
3573 reinit_completion(&the_lnet.ln_mt_wait_complete);
3577 lnet_net_lock(LNET_LOCK_EX);
3578 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3579 lnet_net_unlock(LNET_LOCK_EX);
3581 /* signal that the monitor thread is exiting */
3582 up(&the_lnet.ln_mt_signal);
3590 * Returns == 0 if success
3591 * Returns > 0 if LNetMDBind or prior fails
3592 * Returns < 0 if LNetGet fails
3595 lnet_send_ping(lnet_nid_t dest_nid,
3596 struct lnet_handle_md *mdh, int nnis,
3597 void *user_data, struct lnet_handle_eq eqh, bool recovery)
3599 struct lnet_md md = { NULL };
3600 struct lnet_process_id id;
3601 struct lnet_ping_buffer *pbuf;
3604 if (dest_nid == LNET_NID_ANY) {
3609 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3615 /* initialize md content */
3616 md.start = &pbuf->pb_info;
3617 md.length = LNET_PING_INFO_SIZE(nnis);
3618 md.threshold = 2; /* GET/REPLY */
3620 md.options = LNET_MD_TRUNCATE;
3621 md.user_ptr = user_data;
3624 rc = LNetMDBind(md, LNET_UNLINK, mdh);
3626 lnet_ping_buffer_decref(pbuf);
3627 CERROR("Can't bind MD: %d\n", rc);
3628 rc = -rc; /* change the rc to positive */
3631 id.pid = LNET_PID_LUSTRE;
3634 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3635 LNET_RESERVED_PORTAL,
3636 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3639 goto fail_unlink_md;
3645 LNetInvalidateMDHandle(mdh);
3651 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3652 int status, bool unlink_event)
3654 lnet_nid_t nid = ev_info->mt_nid;
3656 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3660 ni = lnet_nid2ni_locked(nid, 0);
3666 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3668 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3673 CERROR("local NI (%s) recovery failed with %d\n",
3674 libcfs_nid2str(nid), status);
3678 * need to increment healthv for the ni here, because in
3679 * the lnet_finalize() path we don't have access to this
3680 * NI. And in order to get access to it, we'll need to
3681 * carry forward too much information.
3682 * In the peer case, it'll naturally be incremented
3685 lnet_inc_healthv(&ni->ni_healthv);
3687 struct lnet_peer_ni *lpni;
3690 cpt = lnet_net_lock_current();
3691 lpni = lnet_find_peer_ni_locked(nid);
3693 lnet_net_unlock(cpt);
3696 spin_lock(&lpni->lpni_lock);
3697 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3699 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3700 spin_unlock(&lpni->lpni_lock);
3701 lnet_peer_ni_decref_locked(lpni);
3702 lnet_net_unlock(cpt);
3705 CERROR("peer NI (%s) recovery failed with %d\n",
3706 libcfs_nid2str(nid), status);
3711 lnet_mt_event_handler(struct lnet_event *event)
3713 struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3714 struct lnet_ping_buffer *pbuf;
3716 /* TODO: remove assert */
3717 LASSERT(event->type == LNET_EVENT_REPLY ||
3718 event->type == LNET_EVENT_SEND ||
3719 event->type == LNET_EVENT_UNLINK);
3721 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3724 switch (event->type) {
3725 case LNET_EVENT_UNLINK:
3726 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3727 libcfs_nid2str(ev_info->mt_nid));
3729 case LNET_EVENT_REPLY:
3730 lnet_handle_recovery_reply(ev_info, event->status,
3731 event->type == LNET_EVENT_UNLINK);
3733 case LNET_EVENT_SEND:
3734 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3735 libcfs_nid2str(ev_info->mt_nid),
3736 (event->status) ? "unsuccessfully" :
3737 "successfully", event->status);
3740 CERROR("Unexpected event: %d\n", event->type);
3743 if (event->unlinked) {
3744 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3745 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3746 lnet_ping_buffer_decref(pbuf);
3751 lnet_rsp_tracker_create(void)
3753 struct list_head **rstqs;
3754 rstqs = lnet_create_array_of_queues();
3759 the_lnet.ln_mt_rstq = rstqs;
3765 lnet_rsp_tracker_clean(void)
3767 lnet_finalize_expired_responses();
3769 cfs_percpt_free(the_lnet.ln_mt_rstq);
3770 the_lnet.ln_mt_rstq = NULL;
3773 int lnet_monitor_thr_start(void)
3776 struct task_struct *task;
3778 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3781 rc = lnet_resendqs_create();
3785 rc = lnet_rsp_tracker_create();
3789 sema_init(&the_lnet.ln_mt_signal, 0);
3791 lnet_net_lock(LNET_LOCK_EX);
3792 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3793 lnet_net_unlock(LNET_LOCK_EX);
3794 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3797 CERROR("Can't start monitor thread: %d\n", rc);
3804 lnet_net_lock(LNET_LOCK_EX);
3805 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3806 lnet_net_unlock(LNET_LOCK_EX);
3807 /* block until event callback signals exit */
3808 down(&the_lnet.ln_mt_signal);
3810 lnet_net_lock(LNET_LOCK_EX);
3811 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3812 lnet_net_unlock(LNET_LOCK_EX);
3813 lnet_rsp_tracker_clean();
3814 lnet_clean_local_ni_recoveryq();
3815 lnet_clean_peer_ni_recoveryq();
3816 lnet_clean_resendqs();
3817 LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3820 lnet_rsp_tracker_clean();
3821 lnet_clean_local_ni_recoveryq();
3822 lnet_clean_peer_ni_recoveryq();
3823 lnet_clean_resendqs();
3827 void lnet_monitor_thr_stop(void)
3829 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3832 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3833 lnet_net_lock(LNET_LOCK_EX);
3834 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3835 lnet_net_unlock(LNET_LOCK_EX);
3837 /* tell the monitor thread that we're shutting down */
3838 complete(&the_lnet.ln_mt_wait_complete);
3840 /* block until monitor thread signals that it's done */
3841 down(&the_lnet.ln_mt_signal);
3842 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3844 /* perform cleanup tasks */
3845 lnet_rsp_tracker_clean();
3846 lnet_clean_local_ni_recoveryq();
3847 lnet_clean_peer_ni_recoveryq();
3848 lnet_clean_resendqs();
3852 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3856 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3857 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3858 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3859 lnet_net_unlock(cpt);
3861 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3865 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3867 struct lnet_hdr *hdr = &msg->msg_hdr;
3869 if (msg->msg_wanted != 0)
3870 lnet_setpayloadbuffer(msg);
3872 lnet_build_msg_event(msg, LNET_EVENT_PUT);
3874 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
3875 * it back into the ACK during lnet_finalize() */
3876 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3877 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3879 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3880 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3884 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3886 struct lnet_hdr *hdr = &msg->msg_hdr;
3887 struct lnet_match_info info;
3891 /* Convert put fields to host byte order */
3892 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3893 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
3894 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
3896 /* Primary peer NID. */
3897 info.mi_id.nid = msg->msg_initiator;
3898 info.mi_id.pid = hdr->src_pid;
3899 info.mi_opc = LNET_MD_OP_PUT;
3900 info.mi_portal = hdr->msg.put.ptl_index;
3901 info.mi_rlength = hdr->payload_length;
3902 info.mi_roffset = hdr->msg.put.offset;
3903 info.mi_mbits = hdr->msg.put.match_bits;
3904 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3906 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3907 ready_delay = msg->msg_rx_ready_delay;
3910 rc = lnet_ptl_match_md(&info, msg);
3915 case LNET_MATCHMD_OK:
3916 lnet_recv_put(ni, msg);
3919 case LNET_MATCHMD_NONE:
3921 /* no eager_recv or has already called it, should
3922 * have been attached on delayed list */
3925 rc = lnet_ni_eager_recv(ni, msg);
3932 case LNET_MATCHMD_DROP:
3933 CNETERR("Dropping PUT from %s portal %d match %llu"
3934 " offset %d length %d: %d\n",
3935 libcfs_id2str(info.mi_id), info.mi_portal,
3936 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3938 return -ENOENT; /* -ve: OK but no match */
3943 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3945 struct lnet_match_info info;
3946 struct lnet_hdr *hdr = &msg->msg_hdr;
3947 struct lnet_process_id source_id;
3948 struct lnet_handle_wire reply_wmd;
3951 /* Convert get fields to host byte order */
3952 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
3953 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
3954 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
3955 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
3957 source_id.nid = hdr->src_nid;
3958 source_id.pid = hdr->src_pid;
3959 /* Primary peer NID */
3960 info.mi_id.nid = msg->msg_initiator;
3961 info.mi_id.pid = hdr->src_pid;
3962 info.mi_opc = LNET_MD_OP_GET;
3963 info.mi_portal = hdr->msg.get.ptl_index;
3964 info.mi_rlength = hdr->msg.get.sink_length;
3965 info.mi_roffset = hdr->msg.get.src_offset;
3966 info.mi_mbits = hdr->msg.get.match_bits;
3967 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3969 rc = lnet_ptl_match_md(&info, msg);
3970 if (rc == LNET_MATCHMD_DROP) {
3971 CNETERR("Dropping GET from %s portal %d match %llu"
3972 " offset %d length %d\n",
3973 libcfs_id2str(info.mi_id), info.mi_portal,
3974 info.mi_mbits, info.mi_roffset, info.mi_rlength);
3975 return -ENOENT; /* -ve: OK but no match */
3978 LASSERT(rc == LNET_MATCHMD_OK);
3980 lnet_build_msg_event(msg, LNET_EVENT_GET);
3982 reply_wmd = hdr->msg.get.return_wmd;
3984 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3985 msg->msg_offset, msg->msg_wanted);
3987 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3990 /* The LND completes the REPLY from her recv procedure */
3991 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3992 msg->msg_offset, msg->msg_len, msg->msg_len);
3996 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3997 msg->msg_receiving = 0;
3999 rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
4001 /* didn't get as far as lnet_ni_send() */
4002 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4003 libcfs_nid2str(ni->ni_nid),
4004 libcfs_id2str(info.mi_id), rc);
4006 lnet_finalize(msg, rc);
4013 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4015 void *private = msg->msg_private;
4016 struct lnet_hdr *hdr = &msg->msg_hdr;
4017 struct lnet_process_id src = {0};
4018 struct lnet_libmd *md;
4023 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4026 src.nid = hdr->src_nid;
4027 src.pid = hdr->src_pid;
4029 /* NB handles only looked up by creator (no flips) */
4030 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4031 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4032 CNETERR("%s: Dropping REPLY from %s for %s "
4034 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4035 (md == NULL) ? "invalid" : "inactive",
4036 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4037 hdr->msg.reply.dst_wmd.wh_object_cookie);
4038 if (md != NULL && md->md_me != NULL)
4039 CERROR("REPLY MD also attached to portal %d\n",
4040 md->md_me->me_portal);
4042 lnet_res_unlock(cpt);
4043 return -ENOENT; /* -ve: OK but no match */
4046 LASSERT(md->md_offset == 0);
4048 rlength = hdr->payload_length;
4049 mlength = MIN(rlength, (int)md->md_length);
4051 if (mlength < rlength &&
4052 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4053 CNETERR("%s: Dropping REPLY from %s length %d "
4054 "for MD %#llx would overflow (%d)\n",
4055 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4056 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4058 lnet_res_unlock(cpt);
4059 return -ENOENT; /* -ve: OK but no match */
4062 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4063 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4064 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4066 lnet_msg_attach_md(msg, md, 0, mlength);
4069 lnet_setpayloadbuffer(msg);
4071 lnet_res_unlock(cpt);
4073 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4075 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4080 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4082 struct lnet_hdr *hdr = &msg->msg_hdr;
4083 struct lnet_process_id src = {0};
4084 struct lnet_libmd *md;
4087 src.nid = hdr->src_nid;
4088 src.pid = hdr->src_pid;
4090 /* Convert ack fields to host byte order */
4091 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4092 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4094 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4097 /* NB handles only looked up by creator (no flips) */
4098 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4099 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4100 /* Don't moan; this is expected */
4102 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4103 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4104 (md == NULL) ? "invalid" : "inactive",
4105 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4106 hdr->msg.ack.dst_wmd.wh_object_cookie);
4107 if (md != NULL && md->md_me != NULL)
4108 CERROR("Source MD also attached to portal %d\n",
4109 md->md_me->me_portal);
4111 lnet_res_unlock(cpt);
4112 return -ENOENT; /* -ve! */
4115 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4116 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4117 hdr->msg.ack.dst_wmd.wh_object_cookie);
4119 lnet_msg_attach_md(msg, md, 0, 0);
4121 lnet_res_unlock(cpt);
4123 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4125 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4130 * \retval LNET_CREDIT_OK If \a msg is forwarded
4131 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4132 * \retval -ve error code
4135 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4139 if (!the_lnet.ln_routing)
4142 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4143 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4144 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4145 msg->msg_rx_ready_delay = 1;
4147 lnet_net_unlock(msg->msg_rx_cpt);
4148 rc = lnet_ni_eager_recv(ni, msg);
4149 lnet_net_lock(msg->msg_rx_cpt);
4154 rc = lnet_post_routed_recv_locked(msg, 0);
4159 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4163 switch (msg->msg_type) {
4165 rc = lnet_parse_ack(ni, msg);
4168 rc = lnet_parse_put(ni, msg);
4171 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4173 case LNET_MSG_REPLY:
4174 rc = lnet_parse_reply(ni, msg);
4176 default: /* prevent an unused label if !kernel */
4181 LASSERT(rc == 0 || rc == -ENOENT);
4186 lnet_msgtyp2str (int type)
4195 case LNET_MSG_REPLY:
4197 case LNET_MSG_HELLO:
4200 return ("<UNKNOWN>");
4205 lnet_print_hdr(struct lnet_hdr *hdr)
4207 struct lnet_process_id src = {
4208 .nid = hdr->src_nid,
4209 .pid = hdr->src_pid,
4211 struct lnet_process_id dst = {
4212 .nid = hdr->dest_nid,
4213 .pid = hdr->dest_pid,
4215 char *type_str = lnet_msgtyp2str(hdr->type);
4217 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4218 CWARN(" From %s\n", libcfs_id2str(src));
4219 CWARN(" To %s\n", libcfs_id2str(dst));
4221 switch (hdr->type) {
4226 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
4227 "match bits %llu\n",
4228 hdr->msg.put.ptl_index,
4229 hdr->msg.put.ack_wmd.wh_interface_cookie,
4230 hdr->msg.put.ack_wmd.wh_object_cookie,
4231 hdr->msg.put.match_bits);
4232 CWARN(" Length %d, offset %d, hdr data %#llx\n",
4233 hdr->payload_length, hdr->msg.put.offset,
4234 hdr->msg.put.hdr_data);
4238 CWARN(" Ptl index %d, return md %#llx.%#llx, "
4239 "match bits %llu\n", hdr->msg.get.ptl_index,
4240 hdr->msg.get.return_wmd.wh_interface_cookie,
4241 hdr->msg.get.return_wmd.wh_object_cookie,
4242 hdr->msg.get.match_bits);
4243 CWARN(" Length %d, src offset %d\n",
4244 hdr->msg.get.sink_length,
4245 hdr->msg.get.src_offset);
4249 CWARN(" dst md %#llx.%#llx, "
4250 "manipulated length %d\n",
4251 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4252 hdr->msg.ack.dst_wmd.wh_object_cookie,
4253 hdr->msg.ack.mlength);
4256 case LNET_MSG_REPLY:
4257 CWARN(" dst md %#llx.%#llx, "
4259 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4260 hdr->msg.reply.dst_wmd.wh_object_cookie,
4261 hdr->payload_length);
4267 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4268 void *private, int rdma_req)
4270 struct lnet_peer_ni *lpni;
4271 struct lnet_msg *msg;
4272 __u32 payload_length;
4273 lnet_pid_t dest_pid;
4274 lnet_nid_t dest_nid;
4282 LASSERT (!in_interrupt ());
4284 type = le32_to_cpu(hdr->type);
4285 src_nid = le64_to_cpu(hdr->src_nid);
4286 dest_nid = le64_to_cpu(hdr->dest_nid);
4287 dest_pid = le32_to_cpu(hdr->dest_pid);
4288 payload_length = le32_to_cpu(hdr->payload_length);
4290 for_me = (ni->ni_nid == dest_nid);
4291 cpt = lnet_cpt_of_nid(from_nid, ni);
4293 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4294 libcfs_nid2str(dest_nid),
4295 libcfs_nid2str(ni->ni_nid),
4296 libcfs_nid2str(src_nid),
4297 lnet_msgtyp2str(type),
4298 (for_me) ? "for me" : "routed");
4303 if (payload_length > 0) {
4304 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4305 libcfs_nid2str(from_nid),
4306 libcfs_nid2str(src_nid),
4307 lnet_msgtyp2str(type), payload_length);
4313 case LNET_MSG_REPLY:
4314 if (payload_length >
4315 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4316 CERROR("%s, src %s: bad %s payload %d "
4317 "(%d max expected)\n",
4318 libcfs_nid2str(from_nid),
4319 libcfs_nid2str(src_nid),
4320 lnet_msgtyp2str(type),
4322 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4328 CERROR("%s, src %s: Bad message type 0x%x\n",
4329 libcfs_nid2str(from_nid),
4330 libcfs_nid2str(src_nid), type);
4334 if (the_lnet.ln_routing &&
4335 ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4337 spin_lock(&ni->ni_net->net_lock);
4338 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4339 spin_unlock(&ni->ni_net->net_lock);
4340 if (ni->ni_status != NULL &&
4341 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4342 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4349 lnet_push_update_to_peers(1);
4351 /* Regard a bad destination NID as a protocol error. Senders should
4352 * know what they're doing; if they don't they're misconfigured, buggy
4353 * or malicious so we chop them off at the knees :) */
4356 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4357 /* should have gone direct */
4358 CERROR("%s, src %s: Bad dest nid %s "
4359 "(should have been sent direct)\n",
4360 libcfs_nid2str(from_nid),
4361 libcfs_nid2str(src_nid),
4362 libcfs_nid2str(dest_nid));
4366 if (lnet_islocalnid(dest_nid)) {
4367 /* dest is another local NI; sender should have used
4368 * this node's NID on its own network */
4369 CERROR("%s, src %s: Bad dest nid %s "
4370 "(it's my nid but on a different network)\n",
4371 libcfs_nid2str(from_nid),
4372 libcfs_nid2str(src_nid),
4373 libcfs_nid2str(dest_nid));
4377 if (rdma_req && type == LNET_MSG_GET) {
4378 CERROR("%s, src %s: Bad optimized GET for %s "
4379 "(final destination must be me)\n",
4380 libcfs_nid2str(from_nid),
4381 libcfs_nid2str(src_nid),
4382 libcfs_nid2str(dest_nid));
4386 if (!the_lnet.ln_routing) {
4387 CERROR("%s, src %s: Dropping message for %s "
4388 "(routing not enabled)\n",
4389 libcfs_nid2str(from_nid),
4390 libcfs_nid2str(src_nid),
4391 libcfs_nid2str(dest_nid));
4396 /* Message looks OK; we're not going to return an error, so we MUST
4397 * call back lnd_recv() come what may... */
4399 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4400 fail_peer(src_nid, 0)) { /* shall we now? */
4401 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4402 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4403 lnet_msgtyp2str(type));
4407 if (!list_empty(&the_lnet.ln_drop_rules) &&
4408 lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4409 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4410 "silent message loss\n",
4411 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4412 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4416 if (lnet_drop_asym_route && for_me &&
4417 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4418 struct lnet_net *net;
4419 struct lnet_remotenet *rnet;
4422 /* we are dealing with a routed message,
4423 * so see if route to reach src_nid goes through from_nid
4426 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4428 lnet_net_unlock(cpt);
4429 CERROR("net %s not found\n",
4430 libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4434 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4436 struct lnet_peer *gw = NULL;
4437 struct lnet_peer_ni *lpni = NULL;
4438 struct lnet_route *route;
4440 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4442 gw = route->lr_gateway;
4443 if (route->lr_lnet != net->net_id)
4446 * if the nid is one of the gateway's NIDs
4447 * then this is a valid gateway
4449 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4450 NULL, lpni)) != NULL) {
4451 if (lpni->lpni_nid == from_nid) {
4458 lnet_net_unlock(cpt);
4460 /* we would not use from_nid to route a message to
4462 * => asymmetric routing detected but forbidden
4464 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4465 libcfs_nid2str(from_nid),
4466 libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4471 msg = lnet_msg_alloc();
4473 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4474 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4475 lnet_msgtyp2str(type));
4479 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4480 * pointers NULL etc */
4482 msg->msg_type = type;
4483 msg->msg_private = private;
4484 msg->msg_receiving = 1;
4485 msg->msg_rdma_get = rdma_req;
4486 msg->msg_len = msg->msg_wanted = payload_length;
4487 msg->msg_offset = 0;
4488 msg->msg_hdr = *hdr;
4489 /* for building message event */
4490 msg->msg_from = from_nid;
4492 msg->msg_target.pid = dest_pid;
4493 msg->msg_target.nid = dest_nid;
4494 msg->msg_routing = 1;
4497 /* convert common msg->hdr fields to host byteorder */
4498 msg->msg_hdr.type = type;
4499 msg->msg_hdr.src_nid = src_nid;
4500 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4501 msg->msg_hdr.dest_nid = dest_nid;
4502 msg->msg_hdr.dest_pid = dest_pid;
4503 msg->msg_hdr.payload_length = payload_length;
4507 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4509 lnet_net_unlock(cpt);
4510 CERROR("%s, src %s: Dropping %s "
4511 "(error %ld looking up sender)\n",
4512 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4513 lnet_msgtyp2str(type), PTR_ERR(lpni));
4515 if (rc == -ESHUTDOWN)
4516 /* We are shutting down. Don't do anything more */
4521 if (the_lnet.ln_routing)
4522 lpni->lpni_last_alive = ktime_get_seconds();
4524 msg->msg_rxpeer = lpni;
4526 lnet_ni_addref_locked(ni, cpt);
4527 /* Multi-Rail: Primary NID of source. */
4528 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4531 * mark the status of this lpni as UP since we received a message
4532 * from it. The ping response reports back the ns_status which is
4533 * marked on the remote as up or down and we cache it here.
4535 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4537 lnet_msg_commit(msg, cpt);
4539 /* message delay simulation */
4540 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4541 lnet_delay_rule_match_locked(hdr, msg))) {
4542 lnet_net_unlock(cpt);
4547 rc = lnet_parse_forward_locked(ni, msg);
4548 lnet_net_unlock(cpt);
4553 if (rc == LNET_CREDIT_OK) {
4554 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4555 0, payload_length, payload_length);
4560 lnet_net_unlock(cpt);
4562 rc = lnet_parse_local(ni, msg);
4568 LASSERT(msg->msg_md == NULL);
4569 lnet_finalize(msg, rc);
4572 lnet_drop_message(ni, cpt, private, payload_length, type);
4575 EXPORT_SYMBOL(lnet_parse);
4578 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4580 while (!list_empty(head)) {
4581 struct lnet_process_id id = {0};
4582 struct lnet_msg *msg;
4584 msg = list_entry(head->next, struct lnet_msg, msg_list);
4585 list_del(&msg->msg_list);
4587 id.nid = msg->msg_hdr.src_nid;
4588 id.pid = msg->msg_hdr.src_pid;
4590 LASSERT(msg->msg_md == NULL);
4591 LASSERT(msg->msg_rx_delayed);
4592 LASSERT(msg->msg_rxpeer != NULL);
4593 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4595 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4596 " offset %d length %d: %s\n",
4598 msg->msg_hdr.msg.put.ptl_index,
4599 msg->msg_hdr.msg.put.match_bits,
4600 msg->msg_hdr.msg.put.offset,
4601 msg->msg_hdr.payload_length, reason);
4603 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4604 * called lnet_drop_message(), so I just hang onto msg as well
4605 * until that's done */
4607 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4608 msg->msg_private, msg->msg_len,
4611 msg->msg_no_resend = true;
4613 * NB: message will not generate event because w/o attached MD,
4614 * but we still should give error code so lnet_msg_decommit()
4615 * can skip counters operations and other checks.
4617 lnet_finalize(msg, -ENOENT);
4622 lnet_recv_delayed_msg_list(struct list_head *head)
4624 while (!list_empty(head)) {
4625 struct lnet_msg *msg;
4626 struct lnet_process_id id;
4628 msg = list_entry(head->next, struct lnet_msg, msg_list);
4629 list_del(&msg->msg_list);
4631 /* md won't disappear under me, since each msg
4632 * holds a ref on it */
4634 id.nid = msg->msg_hdr.src_nid;
4635 id.pid = msg->msg_hdr.src_pid;
4637 LASSERT(msg->msg_rx_delayed);
4638 LASSERT(msg->msg_md != NULL);
4639 LASSERT(msg->msg_rxpeer != NULL);
4640 LASSERT(msg->msg_rxni != NULL);
4641 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4643 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4644 "match %llu offset %d length %d.\n",
4645 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4646 msg->msg_hdr.msg.put.match_bits,
4647 msg->msg_hdr.msg.put.offset,
4648 msg->msg_hdr.payload_length);
4650 lnet_recv_put(msg->msg_rxni, msg);
4655 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4656 struct lnet_libmd *md, struct lnet_handle_md mdh)
4659 struct lnet_rsp_tracker *local_rspt;
4662 * MD has a refcount taken by message so it's not going away.
4663 * The MD however can be looked up. We need to secure the access
4664 * to the md_rspt_ptr by taking the res_lock.
4665 * The rspt can be accessed without protection up to when it gets
4666 * added to the list.
4670 local_rspt = md->md_rspt_ptr;
4671 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4672 if (local_rspt != NULL) {
4674 * we already have an rspt attached to the md, so we'll
4675 * update the deadline on that one.
4677 lnet_rspt_free(rspt, cpt);
4680 rspt->rspt_mdh = mdh;
4681 rspt->rspt_cpt = cpt;
4682 /* store the rspt so we can access it when we get the REPLY */
4683 md->md_rspt_ptr = rspt;
4686 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4689 * add to the list of tracked responses. It's added to tail of the
4690 * list in order to expire all the older entries first.
4693 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4694 lnet_net_unlock(cpt);
4695 lnet_res_unlock(cpt);
4699 * Initiate an asynchronous PUT operation.
4701 * There are several events associated with a PUT: completion of the send on
4702 * the initiator node (LNET_EVENT_SEND), and when the send completes
4703 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4704 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4705 * used at the target node to indicate the completion of incoming data
4708 * The local events will be logged in the EQ associated with the MD pointed to
4709 * by \a mdh handle. Using a MD without an associated EQ results in these
4710 * events being discarded. In this case, the caller must have another
4711 * mechanism (e.g., a higher level protocol) for determining when it is safe
4712 * to modify the memory region associated with the MD.
4714 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4715 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4717 * \param self Indicates the NID of a local interface through which to send
4718 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4719 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4720 * must be "free floating" (See LNetMDBind()).
4721 * \param ack Controls whether an acknowledgment is requested.
4722 * Acknowledgments are only sent when they are requested by the initiating
4723 * process and the target MD enables them.
4724 * \param target A process identifier for the target process.
4725 * \param portal The index in the \a target's portal table.
4726 * \param match_bits The match bits to use for MD selection at the target
4728 * \param offset The offset into the target MD (only used when the target
4729 * MD has the LNET_MD_MANAGE_REMOTE option set).
4730 * \param hdr_data 64 bits of user data that can be included in the message
4731 * header. This data is written to an event queue entry at the target if an
4732 * EQ is present on the matching MD.
4734 * \retval 0 Success, and only in this case events will be generated
4735 * and logged to EQ (if it exists).
4736 * \retval -EIO Simulated failure.
4737 * \retval -ENOMEM Memory allocation failure.
4738 * \retval -ENOENT Invalid MD object.
4740 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4743 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4744 struct lnet_process_id target, unsigned int portal,
4745 __u64 match_bits, unsigned int offset,
4748 struct lnet_msg *msg;
4749 struct lnet_libmd *md;
4752 struct lnet_rsp_tracker *rspt = NULL;
4754 LASSERT(the_lnet.ln_refcount > 0);
4756 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4757 fail_peer(target.nid, 1)) { /* shall we now? */
4758 CERROR("Dropping PUT to %s: simulated failure\n",
4759 libcfs_id2str(target));
4763 msg = lnet_msg_alloc();
4765 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4766 libcfs_id2str(target));
4769 msg->msg_vmflush = !!memory_pressure_get();
4771 cpt = lnet_cpt_of_cookie(mdh.cookie);
4773 if (ack == LNET_ACK_REQ) {
4774 rspt = lnet_rspt_alloc(cpt);
4776 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4777 libcfs_id2str(target));
4780 INIT_LIST_HEAD(&rspt->rspt_on_list);
4785 md = lnet_handle2md(&mdh);
4786 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4787 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4788 match_bits, portal, libcfs_id2str(target),
4789 md == NULL ? -1 : md->md_threshold);
4790 if (md != NULL && md->md_me != NULL)
4791 CERROR("Source MD also attached to portal %d\n",
4792 md->md_me->me_portal);
4793 lnet_res_unlock(cpt);
4795 lnet_rspt_free(rspt, cpt);
4800 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4802 lnet_msg_attach_md(msg, md, 0, 0);
4804 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4806 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4807 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4808 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4809 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4811 /* NB handles only looked up by creator (no flips) */
4812 if (ack == LNET_ACK_REQ) {
4813 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4814 the_lnet.ln_interface_cookie;
4815 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4816 md->md_lh.lh_cookie;
4818 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4819 LNET_WIRE_HANDLE_COOKIE_NONE;
4820 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4821 LNET_WIRE_HANDLE_COOKIE_NONE;
4824 lnet_res_unlock(cpt);
4826 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4828 if (ack == LNET_ACK_REQ)
4829 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4831 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4835 rc = lnet_send(self, msg, LNET_NID_ANY);
4838 CNETERR("Error sending PUT to %s: %d\n",
4839 libcfs_id2str(target), rc);
4840 msg->msg_no_resend = true;
4841 lnet_finalize(msg, rc);
4844 /* completion will be signalled by an event */
4847 EXPORT_SYMBOL(LNetPut);
4850 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
4851 * returns a msg for the LND to pass to lnet_finalize() when the sink
4852 * data has been received.
4854 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4855 * lnet_finalize() is called on it, so the LND must call this first
4858 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4860 struct lnet_msg *msg = lnet_msg_alloc();
4861 struct lnet_libmd *getmd = getmsg->msg_md;
4862 struct lnet_process_id peer_id = getmsg->msg_target;
4865 LASSERT(!getmsg->msg_target_is_router);
4866 LASSERT(!getmsg->msg_routing);
4869 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4870 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4874 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4877 LASSERT(getmd->md_refcount > 0);
4879 if (getmd->md_threshold == 0) {
4880 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4881 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4883 lnet_res_unlock(cpt);
4887 LASSERT(getmd->md_offset == 0);
4889 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4890 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4892 /* setup information for lnet_build_msg_event */
4893 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4894 msg->msg_from = peer_id.nid;
4895 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4896 msg->msg_hdr.src_nid = peer_id.nid;
4897 msg->msg_hdr.payload_length = getmd->md_length;
4898 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4900 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4901 lnet_res_unlock(cpt);
4903 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4906 lnet_msg_commit(msg, cpt);
4907 lnet_net_unlock(cpt);
4909 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4914 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4917 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4918 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4919 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4921 lnet_net_unlock(cpt);
4928 EXPORT_SYMBOL(lnet_create_reply_msg);
4931 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4934 /* Set the REPLY length, now the RDMA that elides the REPLY message has
4935 * completed and I know it. */
4936 LASSERT(reply != NULL);
4937 LASSERT(reply->msg_type == LNET_MSG_GET);
4938 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4940 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
4941 * the end of my buffer, I might as well be dead. */
4942 LASSERT(len <= reply->msg_ev.mlength);
4944 reply->msg_ev.mlength = len;
4946 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4949 * Initiate an asynchronous GET operation.
4951 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4952 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4953 * the target node in the REPLY has been written to local MD.
4955 * On the target node, an LNET_EVENT_GET is logged when the GET request
4956 * arrives and is accepted into a MD.
4958 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4959 * \param mdh A handle for the MD that describes the memory into which the
4960 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4962 * \retval 0 Success, and only in this case events will be generated
4963 * and logged to EQ (if it exists) of the MD.
4964 * \retval -EIO Simulated failure.
4965 * \retval -ENOMEM Memory allocation failure.
4966 * \retval -ENOENT Invalid MD object.
4969 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4970 struct lnet_process_id target, unsigned int portal,
4971 __u64 match_bits, unsigned int offset, bool recovery)
4973 struct lnet_msg *msg;
4974 struct lnet_libmd *md;
4975 struct lnet_rsp_tracker *rspt;
4979 LASSERT(the_lnet.ln_refcount > 0);
4981 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4982 fail_peer(target.nid, 1)) /* shall we now? */
4984 CERROR("Dropping GET to %s: simulated failure\n",
4985 libcfs_id2str(target));
4989 msg = lnet_msg_alloc();
4991 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4992 libcfs_id2str(target));
4996 cpt = lnet_cpt_of_cookie(mdh.cookie);
4998 rspt = lnet_rspt_alloc(cpt);
5000 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5001 libcfs_id2str(target));
5004 INIT_LIST_HEAD(&rspt->rspt_on_list);
5006 msg->msg_recovery = recovery;
5010 md = lnet_handle2md(&mdh);
5011 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5012 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5013 match_bits, portal, libcfs_id2str(target),
5014 md == NULL ? -1 : md->md_threshold);
5015 if (md != NULL && md->md_me != NULL)
5016 CERROR("REPLY MD also attached to portal %d\n",
5017 md->md_me->me_portal);
5019 lnet_res_unlock(cpt);
5022 lnet_rspt_free(rspt, cpt);
5026 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5028 lnet_msg_attach_md(msg, md, 0, 0);
5030 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5032 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5033 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5034 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5035 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5037 /* NB handles only looked up by creator (no flips) */
5038 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5039 the_lnet.ln_interface_cookie;
5040 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5041 md->md_lh.lh_cookie;
5043 lnet_res_unlock(cpt);
5045 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5047 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5049 rc = lnet_send(self, msg, LNET_NID_ANY);
5051 CNETERR("Error sending GET to %s: %d\n",
5052 libcfs_id2str(target), rc);
5053 msg->msg_no_resend = true;
5054 lnet_finalize(msg, rc);
5057 /* completion will be signalled by an event */
5060 EXPORT_SYMBOL(LNetGet);
5063 * Calculate distance to node at \a dstnid.
5065 * \param dstnid Target NID.
5066 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5068 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5071 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5072 * local_nid_dist_zero is set, which is the default.
5073 * \retval positives Distance to target NID, i.e. number of hops plus one.
5074 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5077 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5079 struct list_head *e;
5080 struct lnet_ni *ni = NULL;
5081 struct lnet_remotenet *rnet;
5082 __u32 dstnet = LNET_NIDNET(dstnid);
5086 struct list_head *rn_list;
5088 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5089 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5090 * keep order 0 free for 0@lo and order 1 free for a local NID
5093 LASSERT(the_lnet.ln_refcount > 0);
5095 cpt = lnet_net_lock_current();
5097 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5098 if (ni->ni_nid == dstnid) {
5099 if (srcnidp != NULL)
5101 if (orderp != NULL) {
5102 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5107 lnet_net_unlock(cpt);
5109 return local_nid_dist_zero ? 0 : 1;
5112 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5113 /* Check if ni was originally created in
5114 * current net namespace.
5115 * If not, assign order above 0xffff0000,
5116 * to make this ni not a priority. */
5117 if (current->nsproxy &&
5118 !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5119 order += 0xffff0000;
5120 if (srcnidp != NULL)
5121 *srcnidp = ni->ni_nid;
5124 lnet_net_unlock(cpt);
5131 rn_list = lnet_net2rnethash(dstnet);
5132 list_for_each(e, rn_list) {
5133 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5135 if (rnet->lrn_net == dstnet) {
5136 struct lnet_route *route;
5137 struct lnet_route *shortest = NULL;
5138 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5141 LASSERT(!list_empty(&rnet->lrn_routes));
5143 list_for_each_entry(route, &rnet->lrn_routes,
5145 route_hops = route->lr_hops;
5146 if (route_hops == LNET_UNDEFINED_HOPS)
5148 if (shortest == NULL ||
5149 route_hops < shortest_hops) {
5151 shortest_hops = route_hops;
5155 LASSERT(shortest != NULL);
5156 hops = shortest_hops;
5157 if (srcnidp != NULL) {
5158 struct lnet_net *net;
5159 net = lnet_get_net_locked(shortest->lr_lnet);
5161 ni = lnet_get_next_ni_locked(net, NULL);
5162 *srcnidp = ni->ni_nid;
5166 lnet_net_unlock(cpt);
5172 lnet_net_unlock(cpt);
5173 return -EHOSTUNREACH;
5175 EXPORT_SYMBOL(LNetDist);