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;
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);
188 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
189 tp = list_entry(el, struct lnet_test_peer, tp_list);
191 if (tp->tp_threshold == 0 || /* needs culling anyway */
192 nid == LNET_NID_ANY || /* removing all entries */
193 tp->tp_nid == nid) { /* matched this one */
194 list_move(&tp->tp_list, &cull);
200 while (!list_empty(&cull)) {
201 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
203 list_del(&tp->tp_list);
204 LIBCFS_FREE(tp, sizeof(*tp));
210 fail_peer (lnet_nid_t nid, int outgoing)
212 struct lnet_test_peer *tp;
213 struct list_head *el;
214 struct list_head *next;
218 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
221 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
222 tp = list_entry(el, struct lnet_test_peer, tp_list);
224 if (tp->tp_threshold == 0) {
227 /* only cull zombies on outgoing tests,
228 * since we may be at interrupt priority on
229 * incoming messages. */
230 list_move(&tp->tp_list, &cull);
235 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
236 nid == tp->tp_nid) { /* fail this peer */
239 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
242 tp->tp_threshold == 0) {
244 list_move(&tp->tp_list, &cull);
253 while (!list_empty(&cull)) {
254 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
255 list_del(&tp->tp_list);
257 LIBCFS_FREE(tp, sizeof(*tp));
264 lnet_iov_nob(unsigned int niov, struct kvec *iov)
266 unsigned int nob = 0;
268 LASSERT(niov == 0 || iov != NULL);
270 nob += (iov++)->iov_len;
274 EXPORT_SYMBOL(lnet_iov_nob);
277 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
278 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
281 /* NB diov, siov are READ-ONLY */
282 unsigned int this_nob;
287 /* skip complete frags before 'doffset' */
289 while (doffset >= diov->iov_len) {
290 doffset -= diov->iov_len;
296 /* skip complete frags before 'soffset' */
298 while (soffset >= siov->iov_len) {
299 soffset -= siov->iov_len;
308 this_nob = min3((unsigned int)diov->iov_len - doffset,
309 (unsigned int)siov->iov_len - soffset,
312 memcpy((char *)diov->iov_base + doffset,
313 (char *)siov->iov_base + soffset, this_nob);
316 if (diov->iov_len > doffset + this_nob) {
324 if (siov->iov_len > soffset + this_nob) {
333 EXPORT_SYMBOL(lnet_copy_iov2iov);
336 lnet_extract_iov(int dst_niov, struct kvec *dst,
337 int src_niov, struct kvec *src,
338 unsigned int offset, unsigned int len)
340 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
341 * for exactly 'len' bytes, and return the number of entries.
342 * NB not destructive to 'src' */
343 unsigned int frag_len;
346 if (len == 0) /* no data => */
347 return (0); /* no frags */
349 LASSERT(src_niov > 0);
350 while (offset >= src->iov_len) { /* skip initial frags */
351 offset -= src->iov_len;
354 LASSERT(src_niov > 0);
359 LASSERT(src_niov > 0);
360 LASSERT((int)niov <= dst_niov);
362 frag_len = src->iov_len - offset;
363 dst->iov_base = ((char *)src->iov_base) + offset;
365 if (len <= frag_len) {
370 dst->iov_len = frag_len;
380 EXPORT_SYMBOL(lnet_extract_iov);
384 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
386 unsigned int nob = 0;
388 LASSERT(niov == 0 || kiov != NULL);
390 nob += (kiov++)->kiov_len;
394 EXPORT_SYMBOL(lnet_kiov_nob);
397 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
398 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
401 /* NB diov, siov are READ-ONLY */
402 unsigned int this_nob;
409 LASSERT (!in_interrupt ());
412 while (doffset >= diov->kiov_len) {
413 doffset -= diov->kiov_len;
420 while (soffset >= siov->kiov_len) {
421 soffset -= siov->kiov_len;
430 this_nob = min3(diov->kiov_len - doffset,
431 siov->kiov_len - soffset,
435 daddr = ((char *)kmap(diov->kiov_page)) +
436 diov->kiov_offset + doffset;
438 saddr = ((char *)kmap(siov->kiov_page)) +
439 siov->kiov_offset + soffset;
441 /* Vanishing risk of kmap deadlock when mapping 2 pages.
442 * However in practice at least one of the kiovs will be mapped
443 * kernel pages and the map/unmap will be NOOPs */
445 memcpy (daddr, saddr, this_nob);
448 if (diov->kiov_len > doffset + this_nob) {
452 kunmap(diov->kiov_page);
459 if (siov->kiov_len > soffset + this_nob) {
463 kunmap(siov->kiov_page);
472 kunmap(diov->kiov_page);
474 kunmap(siov->kiov_page);
476 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
479 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
480 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
483 /* NB iov, kiov are READ-ONLY */
484 unsigned int this_nob;
490 LASSERT (!in_interrupt ());
493 while (iovoffset >= iov->iov_len) {
494 iovoffset -= iov->iov_len;
501 while (kiovoffset >= kiov->kiov_len) {
502 kiovoffset -= kiov->kiov_len;
511 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
512 (unsigned int)kiov->kiov_len - kiovoffset,
516 addr = ((char *)kmap(kiov->kiov_page)) +
517 kiov->kiov_offset + kiovoffset;
519 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
522 if (iov->iov_len > iovoffset + this_nob) {
523 iovoffset += this_nob;
530 if (kiov->kiov_len > kiovoffset + this_nob) {
532 kiovoffset += this_nob;
534 kunmap(kiov->kiov_page);
544 kunmap(kiov->kiov_page);
546 EXPORT_SYMBOL(lnet_copy_kiov2iov);
549 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
550 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
553 /* NB kiov, iov are READ-ONLY */
554 unsigned int this_nob;
560 LASSERT (!in_interrupt ());
563 while (kiovoffset >= kiov->kiov_len) {
564 kiovoffset -= kiov->kiov_len;
571 while (iovoffset >= iov->iov_len) {
572 iovoffset -= iov->iov_len;
581 this_nob = min3((unsigned int)kiov->kiov_len - kiovoffset,
582 (unsigned int)iov->iov_len - iovoffset,
586 addr = ((char *)kmap(kiov->kiov_page)) +
587 kiov->kiov_offset + kiovoffset;
589 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
592 if (kiov->kiov_len > kiovoffset + this_nob) {
594 kiovoffset += this_nob;
596 kunmap(kiov->kiov_page);
603 if (iov->iov_len > iovoffset + this_nob) {
604 iovoffset += this_nob;
613 kunmap(kiov->kiov_page);
615 EXPORT_SYMBOL(lnet_copy_iov2kiov);
618 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
619 int src_niov, lnet_kiov_t *src,
620 unsigned int offset, unsigned int len)
622 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
623 * for exactly 'len' bytes, and return the number of entries.
624 * NB not destructive to 'src' */
625 unsigned int frag_len;
628 if (len == 0) /* no data => */
629 return (0); /* no frags */
631 LASSERT(src_niov > 0);
632 while (offset >= src->kiov_len) { /* skip initial frags */
633 offset -= src->kiov_len;
636 LASSERT(src_niov > 0);
641 LASSERT(src_niov > 0);
642 LASSERT((int)niov <= dst_niov);
644 frag_len = src->kiov_len - offset;
645 dst->kiov_page = src->kiov_page;
646 dst->kiov_offset = src->kiov_offset + offset;
648 if (len <= frag_len) {
650 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
654 dst->kiov_len = frag_len;
655 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
665 EXPORT_SYMBOL(lnet_extract_kiov);
668 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
669 int delayed, unsigned int offset, unsigned int mlen,
672 unsigned int niov = 0;
673 struct kvec *iov = NULL;
674 lnet_kiov_t *kiov = NULL;
677 LASSERT (!in_interrupt ());
678 LASSERT (mlen == 0 || msg != NULL);
681 LASSERT(msg->msg_receiving);
682 LASSERT(!msg->msg_sending);
683 LASSERT(rlen == msg->msg_len);
684 LASSERT(mlen <= msg->msg_len);
685 LASSERT(msg->msg_offset == offset);
686 LASSERT(msg->msg_wanted == mlen);
688 msg->msg_receiving = 0;
691 niov = msg->msg_niov;
693 kiov = msg->msg_kiov;
696 LASSERT ((iov == NULL) != (kiov == NULL));
700 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
701 niov, iov, kiov, offset, mlen,
704 lnet_finalize(msg, rc);
708 lnet_setpayloadbuffer(struct lnet_msg *msg)
710 struct lnet_libmd *md = msg->msg_md;
712 LASSERT(msg->msg_len > 0);
713 LASSERT(!msg->msg_routing);
715 LASSERT(msg->msg_niov == 0);
716 LASSERT(msg->msg_iov == NULL);
717 LASSERT(msg->msg_kiov == NULL);
719 msg->msg_niov = md->md_niov;
720 if ((md->md_options & LNET_MD_KIOV) != 0)
721 msg->msg_kiov = md->md_iov.kiov;
723 msg->msg_iov = md->md_iov.iov;
727 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
728 unsigned int offset, unsigned int len)
730 msg->msg_type = type;
731 msg->msg_target = target;
733 msg->msg_offset = offset;
736 lnet_setpayloadbuffer(msg);
738 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
739 msg->msg_hdr.type = cpu_to_le32(type);
740 /* dest_nid will be overwritten by lnet_select_pathway() */
741 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
742 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
743 /* src_nid will be set later */
744 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
745 msg->msg_hdr.payload_length = cpu_to_le32(len);
749 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
751 void *priv = msg->msg_private;
754 LASSERT (!in_interrupt ());
755 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
756 (msg->msg_txcredit && msg->msg_peertxcredit));
758 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
760 msg->msg_no_resend = true;
761 lnet_finalize(msg, rc);
766 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
770 LASSERT(!msg->msg_sending);
771 LASSERT(msg->msg_receiving);
772 LASSERT(!msg->msg_rx_ready_delay);
773 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
775 msg->msg_rx_ready_delay = 1;
776 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
779 CERROR("recv from %s / send to %s aborted: "
780 "eager_recv failed %d\n",
781 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
782 libcfs_id2str(msg->msg_target), rc);
783 LASSERT(rc < 0); /* required by my callers */
790 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
794 deadline = lpni->lpni_last_alive +
795 lpni->lpni_net->net_tunables.lct_peer_timeout;
798 * assume peer_ni is alive as long as we're within the configured
807 /* NB: returns 1 when alive, 0 when dead, negative when error;
808 * may drop the lnet_net_lock */
810 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
811 struct lnet_msg *msg)
813 time64_t now = ktime_get_seconds();
815 if (!lnet_peer_aliveness_enabled(lpni))
819 * If we're resending a message, let's attempt to send it even if
820 * the peer is down to fulfill our resend quota on the message
822 if (msg->msg_retry_count > 0)
825 /* try and send recovery messages irregardless */
826 if (msg->msg_recovery)
829 /* always send any responses */
830 if (msg->msg_type == LNET_MSG_ACK ||
831 msg->msg_type == LNET_MSG_REPLY)
834 if (!lnet_is_peer_deadline_passed(lpni, now))
837 return lnet_is_peer_ni_alive(lpni);
841 * \param msg The message to be sent.
842 * \param do_send True if lnet_ni_send() should be called in this function.
843 * lnet_send() is going to lnet_net_unlock immediately after this, so
844 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
846 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
847 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
848 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
849 * \retval -ECANCELED If the MD of the message has been unlinked.
852 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
854 struct lnet_peer_ni *lp = msg->msg_txpeer;
855 struct lnet_ni *ni = msg->msg_txni;
856 int cpt = msg->msg_tx_cpt;
857 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
859 /* non-lnet_send() callers have checked before */
860 LASSERT(!do_send || msg->msg_tx_delayed);
861 LASSERT(!msg->msg_receiving);
862 LASSERT(msg->msg_tx_committed);
863 /* can't get here if we're sending to the loopback interface */
864 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
866 /* NB 'lp' is always the next hop */
867 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
868 lnet_peer_alive_locked(ni, lp, msg) == 0) {
869 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
870 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
872 lnet_net_unlock(cpt);
874 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
876 LNET_STATS_TYPE_DROP);
878 lnet_incr_stats(&msg->msg_txni->ni_stats,
880 LNET_STATS_TYPE_DROP);
882 CNETERR("Dropping message for %s: peer not alive\n",
883 libcfs_id2str(msg->msg_target));
884 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
886 lnet_finalize(msg, -EHOSTUNREACH);
889 return -EHOSTUNREACH;
892 if (msg->msg_md != NULL &&
893 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
894 lnet_net_unlock(cpt);
896 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
897 "called on the MD/ME.\n",
898 libcfs_id2str(msg->msg_target));
900 msg->msg_no_resend = true;
901 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
902 msg, libcfs_id2str(msg->msg_target));
903 lnet_finalize(msg, -ECANCELED);
910 if (!msg->msg_peertxcredit) {
911 spin_lock(&lp->lpni_lock);
912 LASSERT((lp->lpni_txcredits < 0) ==
913 !list_empty(&lp->lpni_txq));
915 msg->msg_peertxcredit = 1;
916 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
917 lp->lpni_txcredits--;
919 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
920 lp->lpni_mintxcredits = lp->lpni_txcredits;
922 if (lp->lpni_txcredits < 0) {
923 msg->msg_tx_delayed = 1;
924 list_add_tail(&msg->msg_list, &lp->lpni_txq);
925 spin_unlock(&lp->lpni_lock);
926 return LNET_CREDIT_WAIT;
928 spin_unlock(&lp->lpni_lock);
931 if (!msg->msg_txcredit) {
932 LASSERT((tq->tq_credits < 0) ==
933 !list_empty(&tq->tq_delayed));
935 msg->msg_txcredit = 1;
937 atomic_dec(&ni->ni_tx_credits);
939 if (tq->tq_credits < tq->tq_credits_min)
940 tq->tq_credits_min = tq->tq_credits;
942 if (tq->tq_credits < 0) {
943 msg->msg_tx_delayed = 1;
944 list_add_tail(&msg->msg_list, &tq->tq_delayed);
945 return LNET_CREDIT_WAIT;
949 /* unset the tx_delay flag as we're going to send it now */
950 msg->msg_tx_delayed = 0;
953 lnet_net_unlock(cpt);
954 lnet_ni_send(ni, msg);
957 return LNET_CREDIT_OK;
961 static struct lnet_rtrbufpool *
962 lnet_msg2bufpool(struct lnet_msg *msg)
964 struct lnet_rtrbufpool *rbp;
967 LASSERT(msg->msg_rx_committed);
969 cpt = msg->msg_rx_cpt;
970 rbp = &the_lnet.ln_rtrpools[cpt][0];
972 LASSERT(msg->msg_len <= LNET_MTU);
973 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
975 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
982 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
984 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
985 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
986 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
987 * received or OK to receive */
988 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
989 struct lnet_peer *lp;
990 struct lnet_rtrbufpool *rbp;
991 struct lnet_rtrbuf *rb;
993 LASSERT(msg->msg_iov == NULL);
994 LASSERT(msg->msg_kiov == NULL);
995 LASSERT(msg->msg_niov == 0);
996 LASSERT(msg->msg_routing);
997 LASSERT(msg->msg_receiving);
998 LASSERT(!msg->msg_sending);
999 LASSERT(lpni->lpni_peer_net);
1000 LASSERT(lpni->lpni_peer_net->lpn_peer);
1002 lp = lpni->lpni_peer_net->lpn_peer;
1004 /* non-lnet_parse callers only receive delayed messages */
1005 LASSERT(!do_recv || msg->msg_rx_delayed);
1007 if (!msg->msg_peerrtrcredit) {
1008 /* lpni_lock protects the credit manipulation */
1009 spin_lock(&lpni->lpni_lock);
1010 /* lp_lock protects the lp_rtrq */
1011 spin_lock(&lp->lp_lock);
1013 msg->msg_peerrtrcredit = 1;
1014 lpni->lpni_rtrcredits--;
1015 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1016 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1018 if (lpni->lpni_rtrcredits < 0) {
1019 /* must have checked eager_recv before here */
1020 LASSERT(msg->msg_rx_ready_delay);
1021 msg->msg_rx_delayed = 1;
1022 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1023 spin_unlock(&lp->lp_lock);
1024 spin_unlock(&lpni->lpni_lock);
1025 return LNET_CREDIT_WAIT;
1027 spin_unlock(&lp->lp_lock);
1028 spin_unlock(&lpni->lpni_lock);
1031 rbp = lnet_msg2bufpool(msg);
1033 if (!msg->msg_rtrcredit) {
1034 msg->msg_rtrcredit = 1;
1036 if (rbp->rbp_credits < rbp->rbp_mincredits)
1037 rbp->rbp_mincredits = rbp->rbp_credits;
1039 if (rbp->rbp_credits < 0) {
1040 /* must have checked eager_recv before here */
1041 LASSERT(msg->msg_rx_ready_delay);
1042 msg->msg_rx_delayed = 1;
1043 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1044 return LNET_CREDIT_WAIT;
1048 LASSERT(!list_empty(&rbp->rbp_bufs));
1049 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1050 list_del(&rb->rb_list);
1052 msg->msg_niov = rbp->rbp_npages;
1053 msg->msg_kiov = &rb->rb_kiov[0];
1055 /* unset the msg-rx_delayed flag since we're receiving the message */
1056 msg->msg_rx_delayed = 0;
1059 int cpt = msg->msg_rx_cpt;
1061 lnet_net_unlock(cpt);
1062 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1063 0, msg->msg_len, msg->msg_len);
1066 return LNET_CREDIT_OK;
1070 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1072 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1073 struct lnet_ni *txni = msg->msg_txni;
1074 struct lnet_msg *msg2;
1076 if (msg->msg_txcredit) {
1077 struct lnet_ni *ni = msg->msg_txni;
1078 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1080 /* give back NI txcredits */
1081 msg->msg_txcredit = 0;
1083 LASSERT((tq->tq_credits < 0) ==
1084 !list_empty(&tq->tq_delayed));
1087 atomic_inc(&ni->ni_tx_credits);
1088 if (tq->tq_credits <= 0) {
1089 msg2 = list_entry(tq->tq_delayed.next,
1090 struct lnet_msg, msg_list);
1091 list_del(&msg2->msg_list);
1093 LASSERT(msg2->msg_txni == ni);
1094 LASSERT(msg2->msg_tx_delayed);
1095 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1097 (void) lnet_post_send_locked(msg2, 1);
1101 if (msg->msg_peertxcredit) {
1102 /* give back peer txcredits */
1103 msg->msg_peertxcredit = 0;
1105 spin_lock(&txpeer->lpni_lock);
1106 LASSERT((txpeer->lpni_txcredits < 0) ==
1107 !list_empty(&txpeer->lpni_txq));
1109 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1110 LASSERT(txpeer->lpni_txqnob >= 0);
1112 txpeer->lpni_txcredits++;
1113 if (txpeer->lpni_txcredits <= 0) {
1116 msg2 = list_entry(txpeer->lpni_txq.next,
1117 struct lnet_msg, msg_list);
1118 list_del(&msg2->msg_list);
1119 spin_unlock(&txpeer->lpni_lock);
1121 LASSERT(msg2->msg_txpeer == txpeer);
1122 LASSERT(msg2->msg_tx_delayed);
1124 msg2_cpt = msg2->msg_tx_cpt;
1127 * The msg_cpt can be different from the msg2_cpt
1128 * so we need to make sure we lock the correct cpt
1130 * Once we call lnet_post_send_locked() it is no
1131 * longer safe to access msg2, since it could've
1132 * been freed by lnet_finalize(), but we still
1133 * need to relock the correct cpt, so we cache the
1134 * msg2_cpt for the purpose of the check that
1135 * follows the call to lnet_pose_send_locked().
1137 if (msg2_cpt != msg->msg_tx_cpt) {
1138 lnet_net_unlock(msg->msg_tx_cpt);
1139 lnet_net_lock(msg2_cpt);
1141 (void) lnet_post_send_locked(msg2, 1);
1142 if (msg2_cpt != msg->msg_tx_cpt) {
1143 lnet_net_unlock(msg2_cpt);
1144 lnet_net_lock(msg->msg_tx_cpt);
1147 spin_unlock(&txpeer->lpni_lock);
1152 msg->msg_txni = NULL;
1153 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1156 if (txpeer != NULL) {
1157 msg->msg_txpeer = NULL;
1158 lnet_peer_ni_decref_locked(txpeer);
1163 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1165 struct lnet_msg *msg;
1167 if (list_empty(&rbp->rbp_msgs))
1169 msg = list_entry(rbp->rbp_msgs.next,
1170 struct lnet_msg, msg_list);
1171 list_del(&msg->msg_list);
1173 (void)lnet_post_routed_recv_locked(msg, 1);
1177 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1179 struct lnet_msg *msg;
1180 struct lnet_msg *tmp;
1182 lnet_net_unlock(cpt);
1184 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1185 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1186 0, 0, 0, msg->msg_hdr.payload_length);
1187 list_del_init(&msg->msg_list);
1188 msg->msg_no_resend = true;
1189 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1190 lnet_finalize(msg, -ECANCELED);
1197 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1199 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1200 struct lnet_peer *lp;
1201 struct lnet_ni *rxni = msg->msg_rxni;
1202 struct lnet_msg *msg2;
1204 if (msg->msg_rtrcredit) {
1205 /* give back global router credits */
1206 struct lnet_rtrbuf *rb;
1207 struct lnet_rtrbufpool *rbp;
1209 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1210 * there until it gets one allocated, or aborts the wait
1212 LASSERT(msg->msg_kiov != NULL);
1214 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1217 msg->msg_kiov = NULL;
1218 msg->msg_rtrcredit = 0;
1220 LASSERT(rbp == lnet_msg2bufpool(msg));
1222 LASSERT((rbp->rbp_credits > 0) ==
1223 !list_empty(&rbp->rbp_bufs));
1225 /* If routing is now turned off, we just drop this buffer and
1226 * don't bother trying to return credits. */
1227 if (!the_lnet.ln_routing) {
1228 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1232 /* It is possible that a user has lowered the desired number of
1233 * buffers in this pool. Make sure we never put back
1234 * more buffers than the stated number. */
1235 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1236 /* Discard this buffer so we don't have too
1238 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1239 rbp->rbp_nbuffers--;
1241 list_add(&rb->rb_list, &rbp->rbp_bufs);
1243 if (rbp->rbp_credits <= 0)
1244 lnet_schedule_blocked_locked(rbp);
1249 if (msg->msg_peerrtrcredit) {
1251 LASSERT(rxpeerni->lpni_peer_net);
1252 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1254 lp = rxpeerni->lpni_peer_net->lpn_peer;
1256 /* give back peer router credits */
1257 msg->msg_peerrtrcredit = 0;
1259 spin_lock(&rxpeerni->lpni_lock);
1260 spin_lock(&lp->lp_lock);
1262 rxpeerni->lpni_rtrcredits++;
1264 /* drop all messages which are queued to be routed on that
1266 if (!the_lnet.ln_routing) {
1268 list_splice_init(&lp->lp_rtrq, &drop);
1269 spin_unlock(&lp->lp_lock);
1270 spin_unlock(&rxpeerni->lpni_lock);
1271 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1272 } else if (!list_empty(&lp->lp_rtrq)) {
1275 msg2 = list_entry(lp->lp_rtrq.next,
1276 struct lnet_msg, msg_list);
1277 list_del(&msg2->msg_list);
1278 msg2_cpt = msg2->msg_rx_cpt;
1279 spin_unlock(&lp->lp_lock);
1280 spin_unlock(&rxpeerni->lpni_lock);
1282 * messages on the lp_rtrq can be from any NID in
1283 * the peer, which means they might have different
1284 * cpts. We need to make sure we lock the right
1287 if (msg2_cpt != msg->msg_rx_cpt) {
1288 lnet_net_unlock(msg->msg_rx_cpt);
1289 lnet_net_lock(msg2_cpt);
1291 (void) lnet_post_routed_recv_locked(msg2, 1);
1292 if (msg2_cpt != msg->msg_rx_cpt) {
1293 lnet_net_unlock(msg2_cpt);
1294 lnet_net_lock(msg->msg_rx_cpt);
1297 spin_unlock(&lp->lp_lock);
1298 spin_unlock(&rxpeerni->lpni_lock);
1302 msg->msg_rxni = NULL;
1303 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1305 if (rxpeerni != NULL) {
1306 msg->msg_rxpeer = NULL;
1307 lnet_peer_ni_decref_locked(rxpeerni);
1312 lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1314 if (p1->lpni_txqnob < p2->lpni_txqnob)
1317 if (p1->lpni_txqnob > p2->lpni_txqnob)
1320 if (p1->lpni_txcredits > p2->lpni_txcredits)
1323 if (p1->lpni_txcredits < p2->lpni_txcredits)
1329 static struct lnet_peer_ni *
1330 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1331 struct lnet_peer *peer,
1332 struct lnet_peer_net *peer_net)
1335 * Look at the peer NIs for the destination peer that connect
1336 * to the chosen net. If a peer_ni is preferred when using the
1337 * best_ni to communicate, we use that one. If there is no
1338 * preferred peer_ni, or there are multiple preferred peer_ni,
1339 * the available transmit credits are used. If the transmit
1340 * credits are equal, we round-robin over the peer_ni.
1342 struct lnet_peer_ni *lpni = NULL;
1343 struct lnet_peer_ni *best_lpni = NULL;
1344 int best_lpni_credits = INT_MIN;
1345 bool preferred = false;
1347 int best_lpni_healthv = 0;
1350 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1352 * if the best_ni we've chosen aleady has this lpni
1353 * preferred, then let's use it
1356 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1358 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1359 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1364 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1367 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1368 libcfs_nid2str(lpni->lpni_nid),
1369 lpni->lpni_txcredits, best_lpni_credits,
1370 lpni->lpni_seq, best_lpni->lpni_seq);
1372 /* pick the healthiest peer ni */
1373 if (lpni_healthv < best_lpni_healthv) {
1375 } else if (lpni_healthv > best_lpni_healthv) {
1376 best_lpni_healthv = lpni_healthv;
1377 /* if this is a preferred peer use it */
1378 } else if (!preferred && ni_is_pref) {
1380 } else if (preferred && !ni_is_pref) {
1382 * this is not the preferred peer so let's ignore
1386 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1388 * We already have a peer that has more credits
1389 * available than this one. No need to consider
1390 * this peer further.
1393 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1395 * The best peer found so far and the current peer
1396 * have the same number of available credits let's
1397 * make sure to select between them using Round
1401 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1407 best_lpni_credits = lpni->lpni_txcredits;
1410 /* if we still can't find a peer ni then we can't reach it */
1412 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1413 LNET_NIDNET(dst_nid);
1414 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1415 libcfs_net2str(net_id));
1419 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1420 libcfs_nid2str(best_lpni->lpni_nid));
1426 * Prerequisite: the best_ni should already be set in the sd
1428 static inline struct lnet_peer_ni *
1429 lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid,
1430 struct lnet_peer *peer, __u32 net_id)
1432 struct lnet_peer_net *peer_net;
1435 * The gateway is Multi-Rail capable so now we must select the
1438 peer_net = lnet_peer_get_net_locked(peer, net_id);
1441 CERROR("gateway peer %s has no NI on net %s\n",
1442 libcfs_nid2str(peer->lp_primary_nid),
1443 libcfs_net2str(net_id));
1447 return lnet_select_peer_ni(lni, dst_nid, peer, peer_net);
1450 /* Compare route priorities and hop counts */
1452 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1454 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1455 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1457 if (r1->lr_priority < r2->lr_priority)
1460 if (r1->lr_priority > r2->lr_priority)
1463 if (r1_hops < r2_hops)
1466 if (r1_hops > r2_hops)
1472 static struct lnet_route *
1473 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1474 struct lnet_route **prev_route,
1475 struct lnet_peer_ni **gwni)
1477 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1478 struct lnet_route *best_route;
1479 struct lnet_route *last_route;
1480 struct lnet_route *route;
1483 __u32 any_net = LNET_NIDNET(LNET_NID_ANY);
1485 best_route = last_route = NULL;
1486 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1487 if (!lnet_is_route_alive(route))
1490 /* If the src_net is specified then we need to find an lpni
1493 restrict_net = src_net == any_net ? route->lr_lnet : src_net;
1495 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1499 best_route = last_route = route;
1502 CERROR("Gateway %s does not have a peer NI on net %s\n",
1503 libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1504 libcfs_net2str(restrict_net));
1509 /* no protection on below fields, but it's harmless */
1510 if (last_route->lr_seq - route->lr_seq < 0)
1513 rc = lnet_compare_routes(route, best_route);
1517 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1521 CERROR("Gateway %s does not have a peer NI on net %s\n",
1522 libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1523 libcfs_net2str(restrict_net));
1533 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1537 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1544 *prev_route = last_route;
1550 static struct lnet_ni *
1551 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1552 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1555 struct lnet_ni *ni = NULL;
1556 unsigned int shortest_distance;
1561 * If there is no peer_ni that we can send to on this network,
1562 * then there is no point in looking for a new best_ni here.
1564 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1567 if (best_ni == NULL) {
1568 shortest_distance = UINT_MAX;
1569 best_credits = INT_MIN;
1572 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1573 best_ni->ni_dev_cpt);
1574 best_credits = atomic_read(&best_ni->ni_tx_credits);
1575 best_healthv = atomic_read(&best_ni->ni_healthv);
1578 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1579 unsigned int distance;
1584 ni_credits = atomic_read(&ni->ni_tx_credits);
1585 ni_healthv = atomic_read(&ni->ni_healthv);
1586 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1589 * calculate the distance from the CPT on which
1590 * the message memory is allocated to the CPT of
1591 * the NI's physical device
1593 distance = cfs_cpt_distance(lnet_cpt_table(),
1597 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1598 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1599 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1600 : "not seleced", best_credits, shortest_distance,
1601 (best_ni) ? best_ni->ni_seq : 0);
1604 * All distances smaller than the NUMA range
1605 * are treated equally.
1607 if (distance < lnet_numa_range)
1608 distance = lnet_numa_range;
1611 * Select on health, shorter distance, available
1612 * credits, then round-robin.
1616 } else if (ni_healthv < best_healthv) {
1618 } else if (ni_healthv > best_healthv) {
1619 best_healthv = ni_healthv;
1621 * If we're going to prefer this ni because it's
1622 * the healthiest, then we should set the
1623 * shortest_distance in the algorithm in case
1624 * there are multiple NIs with the same health but
1625 * different distances.
1627 if (distance < shortest_distance)
1628 shortest_distance = distance;
1629 } else if (distance > shortest_distance) {
1631 } else if (distance < shortest_distance) {
1632 shortest_distance = distance;
1633 } else if (ni_credits < best_credits) {
1635 } else if (ni_credits == best_credits) {
1636 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1640 best_credits = ni_credits;
1643 CDEBUG(D_NET, "selected best_ni %s\n",
1644 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1650 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1651 * because such traffic is required to perform discovery. We therefore
1652 * exclude all GET and PUT on that portal. We also exclude all ACK and
1653 * REPLY traffic, but that is because the portal is not tracked in the
1654 * message structure for these message types. We could restrict this
1655 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1658 lnet_msg_discovery(struct lnet_msg *msg)
1660 if (msg->msg_type == LNET_MSG_PUT) {
1661 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1663 } else if (msg->msg_type == LNET_MSG_GET) {
1664 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1670 #define SRC_SPEC 0x0001
1671 #define SRC_ANY 0x0002
1672 #define LOCAL_DST 0x0004
1673 #define REMOTE_DST 0x0008
1674 #define MR_DST 0x0010
1675 #define NMR_DST 0x0020
1676 #define SND_RESP 0x0040
1678 /* The following to defines are used for return codes */
1679 #define REPEAT_SEND 0x1000
1680 #define PASS_THROUGH 0x2000
1682 /* The different cases lnet_select pathway needs to handle */
1683 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1684 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1685 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1686 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1687 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1688 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1689 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1690 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1693 lnet_handle_lo_send(struct lnet_send_data *sd)
1695 struct lnet_msg *msg = sd->sd_msg;
1696 int cpt = sd->sd_cpt;
1698 /* No send credit hassles with LOLND */
1699 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1700 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1701 if (!msg->msg_routing)
1702 msg->msg_hdr.src_nid =
1703 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1704 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1705 lnet_msg_commit(msg, cpt);
1706 msg->msg_txni = the_lnet.ln_loni;
1708 return LNET_CREDIT_OK;
1712 lnet_handle_send(struct lnet_send_data *sd)
1714 struct lnet_ni *best_ni = sd->sd_best_ni;
1715 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1716 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1717 struct lnet_msg *msg = sd->sd_msg;
1719 __u32 send_case = sd->sd_send_case;
1721 __u32 routing = send_case & REMOTE_DST;
1722 struct lnet_rsp_tracker *rspt;
1725 * Increment sequence number of the selected peer so that we
1726 * pick the next one in Round Robin.
1728 best_lpni->lpni_seq++;
1731 * grab a reference on the peer_ni so it sticks around even if
1732 * we need to drop and relock the lnet_net_lock below.
1734 lnet_peer_ni_addref_locked(best_lpni);
1737 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1738 * message. This ensures that we get a CPT that is correct for
1739 * the NI when the NI has been restricted to a subset of all CPTs.
1740 * If the selected CPT differs from the one currently locked, we
1741 * must unlock and relock the lnet_net_lock(), and then check whether
1742 * the configuration has changed. We don't have a hold on the best_ni
1743 * yet, and it may have vanished.
1745 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1746 if (sd->sd_cpt != cpt2) {
1747 __u32 seq = lnet_get_dlc_seq_locked();
1748 lnet_net_unlock(sd->sd_cpt);
1750 lnet_net_lock(sd->sd_cpt);
1751 if (seq != lnet_get_dlc_seq_locked()) {
1752 lnet_peer_ni_decref_locked(best_lpni);
1758 * store the best_lpni in the message right away to avoid having
1759 * to do the same operation under different conditions
1761 msg->msg_txpeer = best_lpni;
1762 msg->msg_txni = best_ni;
1765 * grab a reference for the best_ni since now it's in use in this
1766 * send. The reference will be dropped in lnet_finalize()
1768 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1771 * Always set the target.nid to the best peer picked. Either the
1772 * NID will be one of the peer NIDs selected, or the same NID as
1773 * what was originally set in the target or it will be the NID of
1774 * a router if this message should be routed
1776 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1779 * lnet_msg_commit assigns the correct cpt to the message, which
1780 * is used to decrement the correct refcount on the ni when it's
1781 * time to return the credits
1783 lnet_msg_commit(msg, sd->sd_cpt);
1786 * If we are routing the message then we keep the src_nid that was
1787 * set by the originator. If we are not routing then we are the
1788 * originator and set it here.
1790 if (!msg->msg_routing)
1791 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1794 msg->msg_target_is_router = 1;
1795 msg->msg_target.pid = LNET_PID_LUSTRE;
1797 * since we're routing we want to ensure that the
1798 * msg_hdr.dest_nid is set to the final destination. When
1799 * the router receives this message it knows how to route
1802 * final_dst_lpni is set at the beginning of the
1803 * lnet_select_pathway() function and is never changed.
1804 * It's safe to use it here.
1806 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1809 * if we're not routing set the dest_nid to the best peer
1810 * ni NID that we picked earlier in the algorithm.
1812 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1816 * if we have response tracker block update it with the next hop
1820 rspt = msg->msg_md->md_rspt_ptr;
1822 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1823 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1824 libcfs_nid2str(rspt->rspt_next_hop_nid));
1828 rc = lnet_post_send_locked(msg, 0);
1831 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1832 libcfs_nid2str(msg->msg_hdr.src_nid),
1833 libcfs_nid2str(msg->msg_txni->ni_nid),
1834 libcfs_nid2str(sd->sd_src_nid),
1835 libcfs_nid2str(msg->msg_hdr.dest_nid),
1836 libcfs_nid2str(sd->sd_dst_nid),
1837 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1838 libcfs_nid2str(sd->sd_rtr_nid),
1839 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1845 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1846 struct lnet_msg *msg)
1848 if (msg->msg_type != LNET_MSG_REPLY && msg->msg_type != LNET_MSG_ACK &&
1849 lpni->lpni_pref_nnids == 0) {
1850 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1851 libcfs_nid2str(lni->ni_nid),
1852 libcfs_nid2str(lpni->lpni_nid));
1853 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1862 * use the source and destination NIDs as the pathway
1865 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1867 /* the destination lpni is set before we get here. */
1870 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1871 if (!sd->sd_best_ni) {
1872 CERROR("Can't send to %s: src %s is not a "
1873 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1874 libcfs_nid2str(sd->sd_src_nid));
1878 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
1880 return lnet_handle_send(sd);
1888 * Don't run the selection algorithm on the peer NIs. By specifying the
1889 * local NID, we're also saying that we should always use the destination NID
1890 * provided. This handles the case where we should be using the same
1891 * destination NID for the all the messages which belong to the same RPC
1895 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1897 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1898 if (!sd->sd_best_ni) {
1899 CERROR("Can't send to %s: src %s is not a "
1900 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1901 libcfs_nid2str(sd->sd_src_nid));
1905 if (sd->sd_best_lpni &&
1906 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1907 return lnet_handle_lo_send(sd);
1908 else if (sd->sd_best_lpni)
1909 return lnet_handle_send(sd);
1911 CERROR("can't send to %s. no NI on %s\n",
1912 libcfs_nid2str(sd->sd_dst_nid),
1913 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1915 return -EHOSTUNREACH;
1919 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1920 struct lnet_peer *peer,
1921 struct lnet_peer_net *peer_net,
1925 struct lnet_net *local_net;
1926 struct lnet_ni *best_ni;
1928 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1933 * Iterate through the NIs in this local Net and select
1934 * the NI to send from. The selection is determined by
1935 * these 3 criterion in the following priority:
1937 * 2. NI available credits
1940 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1941 peer, peer_net, cpt);
1943 if (incr_seq && best_ni)
1950 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1951 struct lnet_msg *msg, lnet_nid_t rtr_nid,
1954 struct lnet_peer *peer;
1955 lnet_nid_t primary_nid;
1958 lnet_peer_ni_addref_locked(lpni);
1960 peer = lpni->lpni_peer_net->lpn_peer;
1962 if (lnet_peer_gw_discovery(peer)) {
1963 lnet_peer_ni_decref_locked(lpni);
1967 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1968 lnet_peer_ni_decref_locked(lpni);
1972 rc = lnet_discover_peer_locked(lpni, cpt, false);
1974 lnet_peer_ni_decref_locked(lpni);
1977 /* The peer may have changed. */
1978 peer = lpni->lpni_peer_net->lpn_peer;
1979 spin_lock(&peer->lp_lock);
1980 if (lnet_peer_is_uptodate_locked(peer)) {
1981 spin_unlock(&peer->lp_lock);
1982 lnet_peer_ni_decref_locked(lpni);
1985 /* queue message and return */
1986 msg->msg_rtr_nid_param = rtr_nid;
1987 msg->msg_sending = 0;
1988 msg->msg_txpeer = NULL;
1989 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1990 primary_nid = peer->lp_primary_nid;
1991 spin_unlock(&peer->lp_lock);
1993 lnet_peer_ni_decref_locked(lpni);
1995 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
1996 msg, libcfs_nid2str(primary_nid));
1998 return LNET_DC_WAIT;
2002 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2004 struct lnet_peer_ni **gw_lpni,
2005 struct lnet_peer **gw_peer)
2009 struct lnet_peer *gw;
2010 struct lnet_peer *lp;
2011 struct lnet_peer_net *lpn;
2012 struct lnet_peer_net *best_lpn = NULL;
2013 struct lnet_remotenet *rnet, *best_rnet = NULL;
2014 struct lnet_route *best_route = NULL;
2015 struct lnet_route *last_route = NULL;
2016 struct lnet_peer_ni *lpni = NULL;
2017 struct lnet_peer_ni *gwni = NULL;
2018 lnet_nid_t src_nid = sd->sd_src_nid;
2020 /* If a router nid was specified then we are replying to a GET or
2021 * sending an ACK. In this case we use the gateway associated with the
2022 * specified router nid.
2024 if (sd->sd_rtr_nid != LNET_NID_ANY) {
2025 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2027 CERROR("No peer NI for gateway %s\n",
2028 libcfs_nid2str(sd->sd_rtr_nid));
2029 return -EHOSTUNREACH;
2031 gw = gwni->lpni_peer_net->lpn_peer;
2032 lnet_peer_ni_decref_locked(gwni);
2033 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2035 /* we've already looked up the initial lpni using dst_nid */
2036 lpni = sd->sd_best_lpni;
2037 /* the peer tree must be in existence */
2038 LASSERT(lpni && lpni->lpni_peer_net &&
2039 lpni->lpni_peer_net->lpn_peer);
2040 lp = lpni->lpni_peer_net->lpn_peer;
2042 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2043 /* is this remote network reachable? */
2044 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2053 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2061 CERROR("peer %s has no available nets\n",
2062 libcfs_nid2str(sd->sd_dst_nid));
2063 return -EHOSTUNREACH;
2066 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni,
2069 best_lpn->lpn_net_id);
2070 if (!sd->sd_best_lpni) {
2071 CERROR("peer %s down\n",
2072 libcfs_nid2str(sd->sd_dst_nid));
2073 return -EHOSTUNREACH;
2076 best_route = lnet_find_route_locked(best_rnet,
2077 LNET_NIDNET(src_nid),
2078 &last_route, &gwni);
2080 CERROR("no route to %s from %s\n",
2081 libcfs_nid2str(dst_nid),
2082 libcfs_nid2str(src_nid));
2083 return -EHOSTUNREACH;
2087 CERROR("Internal Error. Route expected to %s from %s\n",
2088 libcfs_nid2str(dst_nid),
2089 libcfs_nid2str(src_nid));
2093 gw = best_route->lr_gateway;
2094 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2095 local_lnet = best_route->lr_lnet;
2100 * Discover this gateway if it hasn't already been discovered.
2101 * This means we might delay the message until discovery has
2104 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2105 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid,
2110 if (!sd->sd_best_ni)
2111 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2112 lnet_peer_get_net_locked(gw,
2117 if (!sd->sd_best_ni) {
2118 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2119 libcfs_net2str(local_lnet),
2120 libcfs_nid2str(sd->sd_src_nid));
2128 * increment the sequence numbers since now we're sure we're
2129 * going to use this path
2131 if (sd->sd_rtr_nid == LNET_NID_ANY) {
2132 LASSERT(best_route && last_route);
2133 best_route->lr_seq = last_route->lr_seq + 1;
2134 best_lpn->lpn_seq++;
2145 * Remote destination
2146 * Non-MR destination
2150 * Remote destination
2153 * The handling of these two cases is similar. Even though the destination
2154 * can be MR or non-MR, we'll deal directly with the router.
2157 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2160 struct lnet_peer_ni *gw_lpni = NULL;
2161 struct lnet_peer *gw_peer = NULL;
2164 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2165 if (!sd->sd_best_ni) {
2166 CERROR("Can't send to %s: src %s is not a "
2167 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2168 libcfs_nid2str(sd->sd_src_nid));
2172 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2177 if (sd->sd_send_case & NMR_DST)
2179 * since the final destination is non-MR let's set its preferred
2180 * NID before we send
2182 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2186 * We're going to send to the gw found so let's set its
2189 sd->sd_peer = gw_peer;
2190 sd->sd_best_lpni = gw_lpni;
2192 return lnet_handle_send(sd);
2196 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2199 struct lnet_peer_net *peer_net = NULL;
2200 struct lnet_ni *best_ni = NULL;
2203 * The peer can have multiple interfaces, some of them can be on
2204 * the local network and others on a routed network. We should
2205 * prefer the local network. However if the local network is not
2206 * available then we need to try the routed network
2209 /* go through all the peer nets and find the best_ni */
2210 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2212 * The peer's list of nets can contain non-local nets. We
2213 * want to only examine the local ones.
2215 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2217 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2218 peer_net, md_cpt, false);
2221 * if this is a discovery message and lp_disc_net_id is
2222 * specified then use that net to send the discovery on.
2224 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2230 /* increment sequence number so we can round robin */
2236 static struct lnet_ni *
2237 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2239 struct lnet_ni *best_ni = NULL;
2240 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2241 struct lnet_peer_ni *lpni_entry;
2244 * We must use a consistent source address when sending to a
2245 * non-MR peer. However, a non-MR peer can have multiple NIDs
2246 * on multiple networks, and we may even need to talk to this
2247 * peer on multiple networks -- certain types of
2248 * load-balancing configuration do this.
2250 * So we need to pick the NI the peer prefers for this
2251 * particular network.
2254 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2256 if (lpni_entry->lpni_pref_nnids == 0)
2258 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2259 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2266 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2268 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2270 struct lnet_ni *best_ni = NULL;
2271 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2274 * We must use a consistent source address when sending to a
2275 * non-MR peer. However, a non-MR peer can have multiple NIDs
2276 * on multiple networks, and we may even need to talk to this
2277 * peer on multiple networks -- certain types of
2278 * load-balancing configuration do this.
2280 * So we need to pick the NI the peer prefers for this
2281 * particular network.
2284 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2287 /* if best_ni is still not set just pick one */
2290 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2291 sd->sd_best_lpni->lpni_peer_net,
2292 sd->sd_md_cpt, true);
2293 /* If there is no best_ni we don't have a route */
2295 CERROR("no path to %s from net %s\n",
2296 libcfs_nid2str(best_lpni->lpni_nid),
2297 libcfs_net2str(best_lpni->lpni_net->net_id));
2298 return -EHOSTUNREACH;
2302 sd->sd_best_ni = best_ni;
2304 /* Set preferred NI if necessary. */
2305 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2312 * Source not specified
2316 * always use the same source NID for NMR peers
2317 * If we've talked to that peer before then we already have a preferred
2318 * source NI associated with it. Otherwise, we select a preferred local NI
2319 * and store it in the peer
2322 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2326 /* sd->sd_best_lpni is already set to the final destination */
2329 * At this point we should've created the peer ni and peer. If we
2330 * can't find it, then something went wrong. Instead of assert
2331 * output a relevant message and fail the send
2333 if (!sd->sd_best_lpni) {
2334 CERROR("Internal fault. Unable to send msg %s to %s. "
2336 lnet_msgtyp2str(sd->sd_msg->msg_type),
2337 libcfs_nid2str(sd->sd_dst_nid));
2341 rc = lnet_select_preferred_best_ni(sd);
2343 rc = lnet_handle_send(sd);
2349 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2352 * NOTE we've already handled the remote peer case. So we only
2353 * need to worry about the local case here.
2355 * if we're sending a response, ACK or reply, we need to send it
2356 * to the destination NID given to us. At this point we already
2357 * have the peer_ni we're suppose to send to, so just find the
2358 * best_ni on the peer net and use that. Since we're sending to an
2359 * MR peer then we can just run the selection algorithm on our
2360 * local NIs and pick the best one.
2362 if (sd->sd_send_case & SND_RESP) {
2364 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2365 sd->sd_best_lpni->lpni_peer_net,
2366 sd->sd_md_cpt, true);
2368 if (!sd->sd_best_ni) {
2370 * We're not going to deal with not able to send
2371 * a response to the provided final destination
2373 CERROR("Can't send response to %s. "
2374 "No local NI available\n",
2375 libcfs_nid2str(sd->sd_dst_nid));
2376 return -EHOSTUNREACH;
2379 return lnet_handle_send(sd);
2383 * If we get here that means we're sending a fresh request, PUT or
2384 * GET, so we need to run our standard selection algorithm.
2385 * First find the best local interface that's on any of the peer's
2388 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2390 lnet_msg_discovery(sd->sd_msg));
2391 if (sd->sd_best_ni) {
2393 lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid,
2395 sd->sd_best_ni->ni_net->net_id);
2398 * if we're successful in selecting a peer_ni on the local
2399 * network, then send to it. Otherwise fall through and
2400 * try and see if we can reach it over another routed
2403 if (sd->sd_best_lpni &&
2404 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2406 * in case we initially started with a routed
2407 * destination, let's reset to local
2409 sd->sd_send_case &= ~REMOTE_DST;
2410 sd->sd_send_case |= LOCAL_DST;
2411 return lnet_handle_lo_send(sd);
2412 } else if (sd->sd_best_lpni) {
2414 * in case we initially started with a routed
2415 * destination, let's reset to local
2417 sd->sd_send_case &= ~REMOTE_DST;
2418 sd->sd_send_case |= LOCAL_DST;
2419 return lnet_handle_send(sd);
2422 CERROR("Internal Error. Expected to have a best_lpni: "
2424 libcfs_nid2str(sd->sd_src_nid),
2425 libcfs_nid2str(sd->sd_dst_nid));
2431 * Peer doesn't have a local network. Let's see if there is
2432 * a remote network we can reach it on.
2434 return PASS_THROUGH;
2439 * Source NID not specified
2444 * Source NID not speified
2445 * Remote destination
2448 * In both of these cases if we're sending a response, ACK or REPLY, then
2449 * we need to send to the destination NID provided.
2451 * In the remote case let's deal with MR routers.
2456 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2459 struct lnet_peer *gw_peer = NULL;
2460 struct lnet_peer_ni *gw_lpni = NULL;
2463 * handle sending a response to a remote peer here so we don't
2464 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2466 if (sd->sd_send_case & REMOTE_DST &&
2467 sd->sd_send_case & SND_RESP) {
2468 struct lnet_peer_ni *gw;
2469 struct lnet_peer *gw_peer;
2471 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2474 CERROR("Can't send response to %s. "
2475 "No route available\n",
2476 libcfs_nid2str(sd->sd_dst_nid));
2477 return -EHOSTUNREACH;
2478 } else if (rc > 0) {
2482 sd->sd_best_lpni = gw;
2483 sd->sd_peer = gw_peer;
2485 return lnet_handle_send(sd);
2489 * Even though the NID for the peer might not be on a local network,
2490 * since the peer is MR there could be other interfaces on the
2491 * local network. In that case we'd still like to prefer the local
2492 * network over the routed network. If we're unable to do that
2493 * then we select the best router among the different routed networks,
2494 * and if the router is MR then we can deal with it as such.
2496 rc = lnet_handle_any_mr_dsta(sd);
2497 if (rc != PASS_THROUGH)
2501 * Now that we must route to the destination, we must consider the
2502 * MR case, where the destination has multiple interfaces, some of
2503 * which we can route to and others we do not. For this reason we
2504 * need to select the destination which we can route to and if
2505 * there are multiple, we need to round robin.
2507 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2512 sd->sd_send_case &= ~LOCAL_DST;
2513 sd->sd_send_case |= REMOTE_DST;
2515 sd->sd_peer = gw_peer;
2516 sd->sd_best_lpni = gw_lpni;
2518 return lnet_handle_send(sd);
2522 * Source not specified
2523 * Remote destination
2526 * Must send to the specified peer NID using the same source NID that
2527 * we've used before. If it's the first time to talk to that peer then
2528 * find the source NI and assign it as preferred to that peer
2531 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2534 struct lnet_peer_ni *gw_lpni = NULL;
2535 struct lnet_peer *gw_peer = NULL;
2538 * Let's see if we have a preferred NI to talk to this NMR peer
2540 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2544 * find the router and that'll find the best NI if we didn't find
2547 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2553 * set the best_ni we've chosen as the preferred one for
2556 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2558 /* we'll be sending to the gw */
2559 sd->sd_best_lpni = gw_lpni;
2560 sd->sd_peer = gw_peer;
2562 return lnet_handle_send(sd);
2566 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2569 * turn off the SND_RESP bit.
2570 * It will be checked in the case handling
2572 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2574 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2575 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2576 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2577 (send_case & MR_DST) ? "MR: " : "NMR: ",
2578 libcfs_nid2str(sd->sd_dst_nid),
2579 (send_case & LOCAL_DST) ? "local" : "routed");
2581 switch (send_case) {
2583 * For all cases where the source is specified, we should always
2584 * use the destination NID, whether it's an MR destination or not,
2585 * since we're continuing a series of related messages for the
2588 case SRC_SPEC_LOCAL_NMR_DST:
2589 return lnet_handle_spec_local_nmr_dst(sd);
2590 case SRC_SPEC_LOCAL_MR_DST:
2591 return lnet_handle_spec_local_mr_dst(sd);
2592 case SRC_SPEC_ROUTER_NMR_DST:
2593 case SRC_SPEC_ROUTER_MR_DST:
2594 return lnet_handle_spec_router_dst(sd);
2595 case SRC_ANY_LOCAL_NMR_DST:
2596 return lnet_handle_any_local_nmr_dst(sd);
2597 case SRC_ANY_LOCAL_MR_DST:
2598 case SRC_ANY_ROUTER_MR_DST:
2599 return lnet_handle_any_mr_dst(sd);
2600 case SRC_ANY_ROUTER_NMR_DST:
2601 return lnet_handle_any_router_nmr_dst(sd);
2603 CERROR("Unknown send case\n");
2609 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2610 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2612 struct lnet_peer_ni *lpni;
2613 struct lnet_peer *peer;
2614 struct lnet_send_data send_data;
2617 __u32 send_case = 0;
2619 memset(&send_data, 0, sizeof(send_data));
2622 * get an initial CPT to use for locking. The idea here is not to
2623 * serialize the calls to select_pathway, so that as many
2624 * operations can run concurrently as possible. To do that we use
2625 * the CPT where this call is being executed. Later on when we
2626 * determine the CPT to use in lnet_message_commit, we switch the
2627 * lock and check if there was any configuration change. If none,
2628 * then we proceed, if there is, then we restart the operation.
2630 cpt = lnet_net_lock_current();
2632 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2633 if (md_cpt == CFS_CPT_ANY)
2639 * If we're being asked to send to the loopback interface, there
2640 * is no need to go through any selection. We can just shortcut
2641 * the entire process and send over lolnd
2643 send_data.sd_msg = msg;
2644 send_data.sd_cpt = cpt;
2645 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2646 rc = lnet_handle_lo_send(&send_data);
2647 lnet_net_unlock(cpt);
2652 * find an existing peer_ni, or create one and mark it as having been
2653 * created due to network traffic. This call will create the
2654 * peer->peer_net->peer_ni tree.
2656 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2658 lnet_net_unlock(cpt);
2659 return PTR_ERR(lpni);
2663 * Cache the original src_nid. If we need to resend the message
2664 * then we'll need to know whether the src_nid was originally
2665 * specified for this message. If it was originally specified,
2666 * then we need to keep using the same src_nid since it's
2667 * continuing the same sequence of messages.
2669 msg->msg_src_nid_param = src_nid;
2672 * If necessary, perform discovery on the peer that owns this peer_ni.
2673 * Note, this can result in the ownership of this peer_ni changing
2674 * to another peer object.
2676 rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2678 lnet_peer_ni_decref_locked(lpni);
2679 lnet_net_unlock(cpt);
2682 lnet_peer_ni_decref_locked(lpni);
2684 peer = lpni->lpni_peer_net->lpn_peer;
2687 * Identify the different send cases
2689 if (src_nid == LNET_NID_ANY)
2690 send_case |= SRC_ANY;
2692 send_case |= SRC_SPEC;
2694 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2695 send_case |= LOCAL_DST;
2697 send_case |= REMOTE_DST;
2700 * if this is a non-MR peer or if we're recovering a peer ni then
2701 * let's consider this an NMR case so we can hit the destination
2704 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2705 send_case |= NMR_DST;
2707 send_case |= MR_DST;
2709 if (msg->msg_type == LNET_MSG_REPLY ||
2710 msg->msg_type == LNET_MSG_ACK)
2711 send_case |= SND_RESP;
2713 /* assign parameters to the send_data */
2714 send_data.sd_rtr_nid = rtr_nid;
2715 send_data.sd_src_nid = src_nid;
2716 send_data.sd_dst_nid = dst_nid;
2717 send_data.sd_best_lpni = lpni;
2719 * keep a pointer to the final destination in case we're going to
2720 * route, so we'll need to access it later
2722 send_data.sd_final_dst_lpni = lpni;
2723 send_data.sd_peer = peer;
2724 send_data.sd_md_cpt = md_cpt;
2725 send_data.sd_send_case = send_case;
2727 rc = lnet_handle_send_case_locked(&send_data);
2730 * Update the local cpt since send_data.sd_cpt might've been
2731 * updated as a result of calling lnet_handle_send_case_locked().
2733 cpt = send_data.sd_cpt;
2735 if (rc == REPEAT_SEND)
2738 lnet_net_unlock(cpt);
2744 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2746 lnet_nid_t dst_nid = msg->msg_target.nid;
2750 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2751 * but we might want to use pre-determined router for ACK/REPLY
2754 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2755 LASSERT(msg->msg_txpeer == NULL);
2756 LASSERT(msg->msg_txni == NULL);
2757 LASSERT(!msg->msg_sending);
2758 LASSERT(!msg->msg_target_is_router);
2759 LASSERT(!msg->msg_receiving);
2761 msg->msg_sending = 1;
2763 LASSERT(!msg->msg_tx_committed);
2765 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2767 if (rc == -EHOSTUNREACH)
2768 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2770 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2774 if (rc == LNET_CREDIT_OK)
2775 lnet_ni_send(msg->msg_txni, msg);
2777 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2781 enum lnet_mt_event_type {
2782 MT_TYPE_LOCAL_NI = 0,
2786 struct lnet_mt_event_info {
2787 enum lnet_mt_event_type mt_type;
2791 /* called with res_lock held */
2793 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2795 struct lnet_rsp_tracker *rspt;
2798 * msg has a refcount on the MD so the MD is not going away.
2799 * The rspt queue for the cpt is protected by
2800 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2802 if (!md->md_rspt_ptr)
2805 rspt = md->md_rspt_ptr;
2808 LASSERT(rspt->rspt_cpt == cpt);
2810 md->md_rspt_ptr = NULL;
2812 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2814 * The monitor thread has invalidated this handle because the
2815 * response timed out, but it failed to lookup the MD. That
2816 * means this response tracker is on the zombie list. We can
2817 * safely remove it under the resource lock (held by caller) and
2818 * free the response tracker block.
2820 list_del(&rspt->rspt_on_list);
2821 lnet_rspt_free(rspt, cpt);
2824 * invalidate the handle to indicate that a response has been
2825 * received, which will then lead the monitor thread to clean up
2828 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2833 lnet_clean_zombie_rstqs(void)
2835 struct lnet_rsp_tracker *rspt, *tmp;
2838 cfs_cpt_for_each(i, lnet_cpt_table()) {
2839 list_for_each_entry_safe(rspt, tmp,
2840 the_lnet.ln_mt_zombie_rstqs[i],
2842 list_del(&rspt->rspt_on_list);
2843 lnet_rspt_free(rspt, i);
2847 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2851 lnet_finalize_expired_responses(void)
2853 struct lnet_libmd *md;
2854 struct lnet_rsp_tracker *rspt, *tmp;
2858 if (the_lnet.ln_mt_rstq == NULL)
2861 cfs_cpt_for_each(i, lnet_cpt_table()) {
2862 LIST_HEAD(local_queue);
2865 if (!the_lnet.ln_mt_rstq[i]) {
2869 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2874 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2876 * The rspt mdh will be invalidated when a response
2877 * is received or whenever we want to discard the
2878 * block the monitor thread will walk the queue
2879 * and clean up any rsts with an invalid mdh.
2880 * The monitor thread will walk the queue until
2881 * the first unexpired rspt block. This means that
2882 * some rspt blocks which received their
2883 * corresponding responses will linger in the
2884 * queue until they are cleaned up eventually.
2887 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2889 list_del(&rspt->rspt_on_list);
2890 lnet_rspt_free(rspt, i);
2894 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2895 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2896 struct lnet_peer_ni *lpni;
2899 md = lnet_handle2md(&rspt->rspt_mdh);
2901 /* MD has been queued for unlink, but
2902 * rspt hasn't been detached (Note we've
2903 * checked above that the rspt_mdh is
2904 * valid). Since we cannot lookup the MD
2905 * we're unable to detach the rspt
2906 * ourselves. Thus, move the rspt to the
2907 * zombie list where we'll wait for
2909 * 1. The remaining operations on the
2910 * MD to complete. In this case the
2911 * final operation will result in
2912 * lnet_msg_detach_md()->
2913 * lnet_detach_rsp_tracker() where
2914 * we will clean up this response
2916 * 2. LNet to shutdown. In this case
2917 * we'll wait until after all LND Nets
2918 * have shutdown and then we can
2919 * safely free any remaining response
2920 * tracker blocks on the zombie list.
2921 * Note: We need to hold the resource
2922 * lock when adding to the zombie list
2923 * because we may have concurrent access
2924 * with lnet_detach_rsp_tracker().
2926 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2927 list_move(&rspt->rspt_on_list,
2928 the_lnet.ln_mt_zombie_rstqs[i]);
2932 LASSERT(md->md_rspt_ptr == rspt);
2933 md->md_rspt_ptr = NULL;
2936 LNetMDUnlink(rspt->rspt_mdh);
2938 nid = rspt->rspt_next_hop_nid;
2940 list_del(&rspt->rspt_on_list);
2941 lnet_rspt_free(rspt, i);
2943 /* If we're shutting down we just want to clean
2944 * up the rspt blocks
2946 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2950 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2954 "Response timeout: md = %p: nid = %s\n",
2955 md, libcfs_nid2str(nid));
2958 * If there is a timeout on the response
2959 * from the next hop decrement its health
2960 * value so that we don't use it
2963 lpni = lnet_find_peer_ni_locked(nid);
2965 lnet_handle_remote_failure_locked(lpni);
2966 lnet_peer_ni_decref_locked(lpni);
2975 if (!list_empty(&local_queue)) {
2977 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2984 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2986 struct lnet_msg *msg;
2988 while (!list_empty(resendq)) {
2989 struct lnet_peer_ni *lpni;
2991 msg = list_entry(resendq->next, struct lnet_msg,
2994 list_del_init(&msg->msg_list);
2996 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2998 lnet_net_unlock(cpt);
2999 CERROR("Expected that a peer is already created for %s\n",
3000 libcfs_nid2str(msg->msg_hdr.dest_nid));
3001 msg->msg_no_resend = true;
3002 lnet_finalize(msg, -EFAULT);
3005 struct lnet_peer *peer;
3007 lnet_nid_t src_nid = LNET_NID_ANY;
3010 * if this message is not being routed and the
3011 * peer is non-MR then we must use the same
3012 * src_nid that was used in the original send.
3013 * Otherwise if we're routing the message (IE
3014 * we're a router) then we can use any of our
3015 * local interfaces. It doesn't matter to the
3016 * final destination.
3018 peer = lpni->lpni_peer_net->lpn_peer;
3019 if (!msg->msg_routing &&
3020 !lnet_peer_is_multi_rail(peer))
3021 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
3024 * If we originally specified a src NID, then we
3025 * must attempt to reuse it in the resend as well.
3027 if (msg->msg_src_nid_param != LNET_NID_ANY)
3028 src_nid = msg->msg_src_nid_param;
3029 lnet_peer_ni_decref_locked(lpni);
3031 lnet_net_unlock(cpt);
3032 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3033 libcfs_nid2str(src_nid),
3034 libcfs_id2str(msg->msg_target),
3035 lnet_msgtyp2str(msg->msg_type),
3037 msg->msg_retry_count);
3038 rc = lnet_send(src_nid, msg, LNET_NID_ANY);
3040 CERROR("Error sending %s to %s: %d\n",
3041 lnet_msgtyp2str(msg->msg_type),
3042 libcfs_id2str(msg->msg_target), rc);
3043 msg->msg_no_resend = true;
3044 lnet_finalize(msg, rc);
3048 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3054 lnet_resend_pending_msgs(void)
3058 cfs_cpt_for_each(i, lnet_cpt_table()) {
3060 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3065 /* called with cpt and ni_lock held */
3067 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3069 struct lnet_handle_md recovery_mdh;
3071 LNetInvalidateMDHandle(&recovery_mdh);
3073 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3075 recovery_mdh = ni->ni_ping_mdh;
3076 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3079 lnet_net_unlock(cpt);
3080 if (!LNetMDHandleIsInvalid(recovery_mdh))
3081 LNetMDUnlink(recovery_mdh);
3087 lnet_recover_local_nis(void)
3089 struct lnet_mt_event_info *ev_info;
3090 LIST_HEAD(processed_list);
3091 LIST_HEAD(local_queue);
3092 struct lnet_handle_md mdh;
3093 struct lnet_ni *tmp;
3100 * splice the recovery queue on a local queue. We will iterate
3101 * through the local queue and update it as needed. Once we're
3102 * done with the traversal, we'll splice the local queue back on
3103 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3104 * will be traversed in the next iteration.
3107 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3111 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3113 * if an NI is being deleted or it is now healthy, there
3114 * is no need to keep it around in the recovery queue.
3115 * The monitor thread is the only thread responsible for
3116 * removing the NI from the recovery queue.
3117 * Multiple threads can be adding NIs to the recovery
3120 healthv = atomic_read(&ni->ni_healthv);
3124 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3125 healthv == LNET_MAX_HEALTH_VALUE) {
3126 list_del_init(&ni->ni_recovery);
3127 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3129 lnet_ni_decref_locked(ni, 0);
3135 * if the local NI failed recovery we must unlink the md.
3136 * But we want to keep the local_ni on the recovery queue
3137 * so we can continue the attempts to recover it.
3139 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3140 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3141 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3148 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3149 libcfs_nid2str(ni->ni_nid));
3152 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3153 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3156 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3158 CERROR("out of memory. Can't recover %s\n",
3159 libcfs_nid2str(ni->ni_nid));
3161 ni->ni_recovery_state &=
3162 ~LNET_NI_RECOVERY_PENDING;
3167 mdh = ni->ni_ping_mdh;
3169 * Invalidate the ni mdh in case it's deleted.
3170 * We'll unlink the mdh in this case below.
3172 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3176 * remove the NI from the local queue and drop the
3177 * reference count to it while we're recovering
3178 * it. The reason for that, is that the NI could
3179 * be deleted, and the way the code is structured
3180 * is if we don't drop the NI, then the deletion
3181 * code will enter a loop waiting for the
3182 * reference count to be removed while holding the
3183 * ln_mutex_lock(). When we look up the peer to
3184 * send to in lnet_select_pathway() we will try to
3185 * lock the ln_mutex_lock() as well, leading to
3186 * a deadlock. By dropping the refcount and
3187 * removing it from the list, we allow for the NI
3188 * to be removed, then we use the cached NID to
3189 * look it up again. If it's gone, then we just
3190 * continue examining the rest of the queue.
3193 list_del_init(&ni->ni_recovery);
3194 lnet_ni_decref_locked(ni, 0);
3197 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3198 ev_info->mt_nid = nid;
3199 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3200 ev_info, the_lnet.ln_mt_eqh, true);
3201 /* lookup the nid again */
3203 ni = lnet_nid2ni_locked(nid, 0);
3206 * the NI has been deleted when we dropped
3214 * Same note as in lnet_recover_peer_nis(). When
3215 * we're sending the ping, the NI is free to be
3216 * deleted or manipulated. By this point it
3217 * could've been added back on the recovery queue,
3218 * and a refcount taken on it.
3219 * So we can't just add it blindly again or we'll
3220 * corrupt the queue. We must check under lock if
3221 * it's not on any list and if not then add it
3222 * to the processed list, which will eventually be
3223 * spliced back on to the recovery queue.
3225 ni->ni_ping_mdh = mdh;
3226 if (list_empty(&ni->ni_recovery)) {
3227 list_add_tail(&ni->ni_recovery, &processed_list);
3228 lnet_ni_addref_locked(ni, 0);
3234 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3240 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3241 * reexamined in the next iteration.
3243 list_splice_init(&processed_list, &local_queue);
3245 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3250 lnet_resendqs_create(void)
3252 struct list_head **resendqs;
3253 resendqs = lnet_create_array_of_queues();
3258 lnet_net_lock(LNET_LOCK_EX);
3259 the_lnet.ln_mt_resendqs = resendqs;
3260 lnet_net_unlock(LNET_LOCK_EX);
3266 lnet_clean_local_ni_recoveryq(void)
3270 /* This is only called when the monitor thread has stopped */
3273 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3274 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3275 struct lnet_ni, ni_recovery);
3276 list_del_init(&ni->ni_recovery);
3278 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3280 lnet_ni_decref_locked(ni, 0);
3287 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3290 struct lnet_handle_md recovery_mdh;
3292 LNetInvalidateMDHandle(&recovery_mdh);
3294 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3295 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3296 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3298 spin_unlock(&lpni->lpni_lock);
3299 lnet_net_unlock(cpt);
3300 if (!LNetMDHandleIsInvalid(recovery_mdh))
3301 LNetMDUnlink(recovery_mdh);
3303 spin_lock(&lpni->lpni_lock);
3307 lnet_clean_peer_ni_recoveryq(void)
3309 struct lnet_peer_ni *lpni, *tmp;
3311 lnet_net_lock(LNET_LOCK_EX);
3313 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3315 list_del_init(&lpni->lpni_recovery);
3316 spin_lock(&lpni->lpni_lock);
3317 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3318 spin_unlock(&lpni->lpni_lock);
3319 lnet_peer_ni_decref_locked(lpni);
3322 lnet_net_unlock(LNET_LOCK_EX);
3326 lnet_clean_resendqs(void)
3328 struct lnet_msg *msg, *tmp;
3332 cfs_cpt_for_each(i, lnet_cpt_table()) {
3334 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3336 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3337 list_del_init(&msg->msg_list);
3338 msg->msg_no_resend = true;
3339 lnet_finalize(msg, -ESHUTDOWN);
3343 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3347 lnet_recover_peer_nis(void)
3349 struct lnet_mt_event_info *ev_info;
3350 LIST_HEAD(processed_list);
3351 LIST_HEAD(local_queue);
3352 struct lnet_handle_md mdh;
3353 struct lnet_peer_ni *lpni;
3354 struct lnet_peer_ni *tmp;
3360 * Always use cpt 0 for locking across all interactions with
3361 * ln_mt_peerNIRecovq
3364 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3368 list_for_each_entry_safe(lpni, tmp, &local_queue,
3371 * The same protection strategy is used here as is in the
3372 * local recovery case.
3375 healthv = atomic_read(&lpni->lpni_healthv);
3376 spin_lock(&lpni->lpni_lock);
3377 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3378 healthv == LNET_MAX_HEALTH_VALUE) {
3379 list_del_init(&lpni->lpni_recovery);
3380 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3381 spin_unlock(&lpni->lpni_lock);
3382 lnet_peer_ni_decref_locked(lpni);
3388 * If the peer NI has failed recovery we must unlink the
3389 * md. But we want to keep the peer ni on the recovery
3390 * queue so we can try to continue recovering it
3392 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3393 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3394 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3397 spin_unlock(&lpni->lpni_lock);
3401 * NOTE: we're racing with peer deletion from user space.
3402 * It's possible that a peer is deleted after we check its
3403 * state. In this case the recovery can create a new peer
3405 spin_lock(&lpni->lpni_lock);
3406 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3407 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3408 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3409 spin_unlock(&lpni->lpni_lock);
3411 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3413 CERROR("out of memory. Can't recover %s\n",
3414 libcfs_nid2str(lpni->lpni_nid));
3415 spin_lock(&lpni->lpni_lock);
3416 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3417 spin_unlock(&lpni->lpni_lock);
3421 /* look at the comments in lnet_recover_local_nis() */
3422 mdh = lpni->lpni_recovery_ping_mdh;
3423 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3424 nid = lpni->lpni_nid;
3426 list_del_init(&lpni->lpni_recovery);
3427 lnet_peer_ni_decref_locked(lpni);
3430 ev_info->mt_type = MT_TYPE_PEER_NI;
3431 ev_info->mt_nid = nid;
3432 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3433 ev_info, the_lnet.ln_mt_eqh, true);
3436 * lnet_find_peer_ni_locked() grabs a refcount for
3437 * us. No need to take it explicitly.
3439 lpni = lnet_find_peer_ni_locked(nid);
3446 lpni->lpni_recovery_ping_mdh = mdh;
3448 * While we're unlocked the lpni could've been
3449 * readded on the recovery queue. In this case we
3450 * don't need to add it to the local queue, since
3451 * it's already on there and the thread that added
3452 * it would've incremented the refcount on the
3453 * peer, which means we need to decref the refcount
3454 * that was implicitly grabbed by find_peer_ni_locked.
3455 * Otherwise, if the lpni is still not on
3456 * the recovery queue, then we'll add it to the
3459 if (list_empty(&lpni->lpni_recovery))
3460 list_add_tail(&lpni->lpni_recovery, &processed_list);
3462 lnet_peer_ni_decref_locked(lpni);
3465 spin_lock(&lpni->lpni_lock);
3467 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3469 spin_unlock(&lpni->lpni_lock);
3472 list_splice_init(&processed_list, &local_queue);
3474 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3479 lnet_monitor_thread(void *arg)
3481 time64_t recovery_timeout = 0;
3482 time64_t rsp_timeout = 0;
3486 wait_for_completion(&the_lnet.ln_started);
3488 * The monitor thread takes care of the following:
3489 * 1. Checks the aliveness of routers
3490 * 2. Checks if there are messages on the resend queue to resend
3492 * 3. Check if there are any NIs on the local recovery queue and
3494 * 4. Checks if there are any NIs on the remote recovery queue
3497 cfs_block_allsigs();
3499 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3500 now = ktime_get_real_seconds();
3502 if (lnet_router_checker_active())
3503 lnet_check_routers();
3505 lnet_resend_pending_msgs();
3507 if (now >= rsp_timeout) {
3508 lnet_finalize_expired_responses();
3509 rsp_timeout = now + (lnet_transaction_timeout / 2);
3512 if (now >= recovery_timeout) {
3513 lnet_recover_local_nis();
3514 lnet_recover_peer_nis();
3515 recovery_timeout = now + lnet_recovery_interval;
3519 * TODO do we need to check if we should sleep without
3520 * timeout? Technically, an active system will always
3521 * have messages in flight so this check will always
3522 * evaluate to false. And on an idle system do we care
3523 * if we wake up every 1 second? Although, we've seen
3524 * cases where we get a complaint that an idle thread
3525 * is waking up unnecessarily.
3527 * Take into account the current net_count when you wake
3528 * up for alive router checking, since we need to check
3529 * possibly as many networks as we have configured.
3531 interval = min(lnet_recovery_interval,
3532 min((unsigned int) alive_router_check_interval /
3533 lnet_current_net_count,
3534 lnet_transaction_timeout / 2));
3535 wait_for_completion_interruptible_timeout(
3536 &the_lnet.ln_mt_wait_complete,
3537 cfs_time_seconds(interval));
3538 /* Must re-init the completion before testing anything,
3539 * including ln_mt_state.
3541 reinit_completion(&the_lnet.ln_mt_wait_complete);
3545 lnet_net_lock(LNET_LOCK_EX);
3546 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3547 lnet_net_unlock(LNET_LOCK_EX);
3549 /* signal that the monitor thread is exiting */
3550 up(&the_lnet.ln_mt_signal);
3558 * Returns == 0 if success
3559 * Returns > 0 if LNetMDBind or prior fails
3560 * Returns < 0 if LNetGet fails
3563 lnet_send_ping(lnet_nid_t dest_nid,
3564 struct lnet_handle_md *mdh, int nnis,
3565 void *user_data, struct lnet_handle_eq eqh, bool recovery)
3567 struct lnet_md md = { NULL };
3568 struct lnet_process_id id;
3569 struct lnet_ping_buffer *pbuf;
3572 if (dest_nid == LNET_NID_ANY) {
3577 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3583 /* initialize md content */
3584 md.start = &pbuf->pb_info;
3585 md.length = LNET_PING_INFO_SIZE(nnis);
3586 md.threshold = 2; /* GET/REPLY */
3588 md.options = LNET_MD_TRUNCATE;
3589 md.user_ptr = user_data;
3592 rc = LNetMDBind(md, LNET_UNLINK, mdh);
3594 lnet_ping_buffer_decref(pbuf);
3595 CERROR("Can't bind MD: %d\n", rc);
3596 rc = -rc; /* change the rc to positive */
3599 id.pid = LNET_PID_LUSTRE;
3602 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3603 LNET_RESERVED_PORTAL,
3604 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3607 goto fail_unlink_md;
3613 LNetInvalidateMDHandle(mdh);
3619 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3620 int status, bool unlink_event)
3622 lnet_nid_t nid = ev_info->mt_nid;
3624 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3628 ni = lnet_nid2ni_locked(nid, 0);
3634 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3636 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3641 CERROR("local NI (%s) recovery failed with %d\n",
3642 libcfs_nid2str(nid), status);
3646 * need to increment healthv for the ni here, because in
3647 * the lnet_finalize() path we don't have access to this
3648 * NI. And in order to get access to it, we'll need to
3649 * carry forward too much information.
3650 * In the peer case, it'll naturally be incremented
3653 lnet_inc_healthv(&ni->ni_healthv);
3655 struct lnet_peer_ni *lpni;
3658 cpt = lnet_net_lock_current();
3659 lpni = lnet_find_peer_ni_locked(nid);
3661 lnet_net_unlock(cpt);
3664 spin_lock(&lpni->lpni_lock);
3665 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3667 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3668 spin_unlock(&lpni->lpni_lock);
3669 lnet_peer_ni_decref_locked(lpni);
3670 lnet_net_unlock(cpt);
3673 CERROR("peer NI (%s) recovery failed with %d\n",
3674 libcfs_nid2str(nid), status);
3679 lnet_mt_event_handler(struct lnet_event *event)
3681 struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3682 struct lnet_ping_buffer *pbuf;
3684 /* TODO: remove assert */
3685 LASSERT(event->type == LNET_EVENT_REPLY ||
3686 event->type == LNET_EVENT_SEND ||
3687 event->type == LNET_EVENT_UNLINK);
3689 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3692 switch (event->type) {
3693 case LNET_EVENT_UNLINK:
3694 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3695 libcfs_nid2str(ev_info->mt_nid));
3697 case LNET_EVENT_REPLY:
3698 lnet_handle_recovery_reply(ev_info, event->status,
3699 event->type == LNET_EVENT_UNLINK);
3701 case LNET_EVENT_SEND:
3702 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3703 libcfs_nid2str(ev_info->mt_nid),
3704 (event->status) ? "unsuccessfully" :
3705 "successfully", event->status);
3708 CERROR("Unexpected event: %d\n", event->type);
3711 if (event->unlinked) {
3712 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3713 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3714 lnet_ping_buffer_decref(pbuf);
3719 lnet_rsp_tracker_create(void)
3721 struct list_head **rstqs;
3722 rstqs = lnet_create_array_of_queues();
3727 the_lnet.ln_mt_rstq = rstqs;
3733 lnet_rsp_tracker_clean(void)
3735 lnet_finalize_expired_responses();
3737 cfs_percpt_free(the_lnet.ln_mt_rstq);
3738 the_lnet.ln_mt_rstq = NULL;
3741 int lnet_monitor_thr_start(void)
3744 struct task_struct *task;
3746 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3749 rc = lnet_resendqs_create();
3753 rc = lnet_rsp_tracker_create();
3757 sema_init(&the_lnet.ln_mt_signal, 0);
3759 lnet_net_lock(LNET_LOCK_EX);
3760 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3761 lnet_net_unlock(LNET_LOCK_EX);
3762 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3765 CERROR("Can't start monitor thread: %d\n", rc);
3772 lnet_net_lock(LNET_LOCK_EX);
3773 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3774 lnet_net_unlock(LNET_LOCK_EX);
3775 /* block until event callback signals exit */
3776 down(&the_lnet.ln_mt_signal);
3778 lnet_net_lock(LNET_LOCK_EX);
3779 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3780 lnet_net_unlock(LNET_LOCK_EX);
3781 lnet_rsp_tracker_clean();
3782 lnet_clean_local_ni_recoveryq();
3783 lnet_clean_peer_ni_recoveryq();
3784 lnet_clean_resendqs();
3785 LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3788 lnet_rsp_tracker_clean();
3789 lnet_clean_local_ni_recoveryq();
3790 lnet_clean_peer_ni_recoveryq();
3791 lnet_clean_resendqs();
3795 void lnet_monitor_thr_stop(void)
3797 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3800 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3801 lnet_net_lock(LNET_LOCK_EX);
3802 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3803 lnet_net_unlock(LNET_LOCK_EX);
3805 /* tell the monitor thread that we're shutting down */
3806 complete(&the_lnet.ln_mt_wait_complete);
3808 /* block until monitor thread signals that it's done */
3809 down(&the_lnet.ln_mt_signal);
3810 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3812 /* perform cleanup tasks */
3813 lnet_rsp_tracker_clean();
3814 lnet_clean_local_ni_recoveryq();
3815 lnet_clean_peer_ni_recoveryq();
3816 lnet_clean_resendqs();
3820 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3824 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3825 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3826 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3827 lnet_net_unlock(cpt);
3829 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3833 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3835 struct lnet_hdr *hdr = &msg->msg_hdr;
3837 if (msg->msg_wanted != 0)
3838 lnet_setpayloadbuffer(msg);
3840 lnet_build_msg_event(msg, LNET_EVENT_PUT);
3842 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
3843 * it back into the ACK during lnet_finalize() */
3844 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3845 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3847 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3848 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3852 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3854 struct lnet_hdr *hdr = &msg->msg_hdr;
3855 struct lnet_match_info info;
3859 /* Convert put fields to host byte order */
3860 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3861 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
3862 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
3864 /* Primary peer NID. */
3865 info.mi_id.nid = msg->msg_initiator;
3866 info.mi_id.pid = hdr->src_pid;
3867 info.mi_opc = LNET_MD_OP_PUT;
3868 info.mi_portal = hdr->msg.put.ptl_index;
3869 info.mi_rlength = hdr->payload_length;
3870 info.mi_roffset = hdr->msg.put.offset;
3871 info.mi_mbits = hdr->msg.put.match_bits;
3872 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3874 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3875 ready_delay = msg->msg_rx_ready_delay;
3878 rc = lnet_ptl_match_md(&info, msg);
3883 case LNET_MATCHMD_OK:
3884 lnet_recv_put(ni, msg);
3887 case LNET_MATCHMD_NONE:
3889 /* no eager_recv or has already called it, should
3890 * have been attached on delayed list */
3893 rc = lnet_ni_eager_recv(ni, msg);
3900 case LNET_MATCHMD_DROP:
3901 CNETERR("Dropping PUT from %s portal %d match %llu"
3902 " offset %d length %d: %d\n",
3903 libcfs_id2str(info.mi_id), info.mi_portal,
3904 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3906 return -ENOENT; /* -ve: OK but no match */
3911 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3913 struct lnet_match_info info;
3914 struct lnet_hdr *hdr = &msg->msg_hdr;
3915 struct lnet_process_id source_id;
3916 struct lnet_handle_wire reply_wmd;
3919 /* Convert get fields to host byte order */
3920 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
3921 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
3922 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
3923 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
3925 source_id.nid = hdr->src_nid;
3926 source_id.pid = hdr->src_pid;
3927 /* Primary peer NID */
3928 info.mi_id.nid = msg->msg_initiator;
3929 info.mi_id.pid = hdr->src_pid;
3930 info.mi_opc = LNET_MD_OP_GET;
3931 info.mi_portal = hdr->msg.get.ptl_index;
3932 info.mi_rlength = hdr->msg.get.sink_length;
3933 info.mi_roffset = hdr->msg.get.src_offset;
3934 info.mi_mbits = hdr->msg.get.match_bits;
3935 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3937 rc = lnet_ptl_match_md(&info, msg);
3938 if (rc == LNET_MATCHMD_DROP) {
3939 CNETERR("Dropping GET from %s portal %d match %llu"
3940 " offset %d length %d\n",
3941 libcfs_id2str(info.mi_id), info.mi_portal,
3942 info.mi_mbits, info.mi_roffset, info.mi_rlength);
3943 return -ENOENT; /* -ve: OK but no match */
3946 LASSERT(rc == LNET_MATCHMD_OK);
3948 lnet_build_msg_event(msg, LNET_EVENT_GET);
3950 reply_wmd = hdr->msg.get.return_wmd;
3952 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3953 msg->msg_offset, msg->msg_wanted);
3955 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3958 /* The LND completes the REPLY from her recv procedure */
3959 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3960 msg->msg_offset, msg->msg_len, msg->msg_len);
3964 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3965 msg->msg_receiving = 0;
3967 rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3969 /* didn't get as far as lnet_ni_send() */
3970 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3971 libcfs_nid2str(ni->ni_nid),
3972 libcfs_id2str(info.mi_id), rc);
3974 lnet_finalize(msg, rc);
3981 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3983 void *private = msg->msg_private;
3984 struct lnet_hdr *hdr = &msg->msg_hdr;
3985 struct lnet_process_id src = {0};
3986 struct lnet_libmd *md;
3987 unsigned int rlength;
3988 unsigned int mlength;
3991 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3994 src.nid = hdr->src_nid;
3995 src.pid = hdr->src_pid;
3997 /* NB handles only looked up by creator (no flips) */
3998 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
3999 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4000 CNETERR("%s: Dropping REPLY from %s for %s "
4002 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4003 (md == NULL) ? "invalid" : "inactive",
4004 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4005 hdr->msg.reply.dst_wmd.wh_object_cookie);
4006 if (md != NULL && md->md_me != NULL)
4007 CERROR("REPLY MD also attached to portal %d\n",
4008 md->md_me->me_portal);
4010 lnet_res_unlock(cpt);
4011 return -ENOENT; /* -ve: OK but no match */
4014 LASSERT(md->md_offset == 0);
4016 rlength = hdr->payload_length;
4017 mlength = min(rlength, md->md_length);
4019 if (mlength < rlength &&
4020 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4021 CNETERR("%s: Dropping REPLY from %s length %d "
4022 "for MD %#llx would overflow (%d)\n",
4023 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4024 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4026 lnet_res_unlock(cpt);
4027 return -ENOENT; /* -ve: OK but no match */
4030 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4031 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4032 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4034 lnet_msg_attach_md(msg, md, 0, mlength);
4037 lnet_setpayloadbuffer(msg);
4039 lnet_res_unlock(cpt);
4041 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4043 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4048 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4050 struct lnet_hdr *hdr = &msg->msg_hdr;
4051 struct lnet_process_id src = {0};
4052 struct lnet_libmd *md;
4055 src.nid = hdr->src_nid;
4056 src.pid = hdr->src_pid;
4058 /* Convert ack fields to host byte order */
4059 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4060 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4062 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4065 /* NB handles only looked up by creator (no flips) */
4066 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4067 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4068 /* Don't moan; this is expected */
4070 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4071 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4072 (md == NULL) ? "invalid" : "inactive",
4073 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4074 hdr->msg.ack.dst_wmd.wh_object_cookie);
4075 if (md != NULL && md->md_me != NULL)
4076 CERROR("Source MD also attached to portal %d\n",
4077 md->md_me->me_portal);
4079 lnet_res_unlock(cpt);
4080 return -ENOENT; /* -ve! */
4083 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4084 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4085 hdr->msg.ack.dst_wmd.wh_object_cookie);
4087 lnet_msg_attach_md(msg, md, 0, 0);
4089 lnet_res_unlock(cpt);
4091 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4093 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4098 * \retval LNET_CREDIT_OK If \a msg is forwarded
4099 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4100 * \retval -ve error code
4103 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4107 if (!the_lnet.ln_routing)
4110 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4111 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4112 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4113 msg->msg_rx_ready_delay = 1;
4115 lnet_net_unlock(msg->msg_rx_cpt);
4116 rc = lnet_ni_eager_recv(ni, msg);
4117 lnet_net_lock(msg->msg_rx_cpt);
4122 rc = lnet_post_routed_recv_locked(msg, 0);
4127 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4131 switch (msg->msg_type) {
4133 rc = lnet_parse_ack(ni, msg);
4136 rc = lnet_parse_put(ni, msg);
4139 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4141 case LNET_MSG_REPLY:
4142 rc = lnet_parse_reply(ni, msg);
4144 default: /* prevent an unused label if !kernel */
4149 LASSERT(rc == 0 || rc == -ENOENT);
4154 lnet_msgtyp2str (int type)
4163 case LNET_MSG_REPLY:
4165 case LNET_MSG_HELLO:
4168 return ("<UNKNOWN>");
4173 lnet_print_hdr(struct lnet_hdr *hdr)
4175 struct lnet_process_id src = {
4176 .nid = hdr->src_nid,
4177 .pid = hdr->src_pid,
4179 struct lnet_process_id dst = {
4180 .nid = hdr->dest_nid,
4181 .pid = hdr->dest_pid,
4183 char *type_str = lnet_msgtyp2str(hdr->type);
4185 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4186 CWARN(" From %s\n", libcfs_id2str(src));
4187 CWARN(" To %s\n", libcfs_id2str(dst));
4189 switch (hdr->type) {
4194 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
4195 "match bits %llu\n",
4196 hdr->msg.put.ptl_index,
4197 hdr->msg.put.ack_wmd.wh_interface_cookie,
4198 hdr->msg.put.ack_wmd.wh_object_cookie,
4199 hdr->msg.put.match_bits);
4200 CWARN(" Length %d, offset %d, hdr data %#llx\n",
4201 hdr->payload_length, hdr->msg.put.offset,
4202 hdr->msg.put.hdr_data);
4206 CWARN(" Ptl index %d, return md %#llx.%#llx, "
4207 "match bits %llu\n", hdr->msg.get.ptl_index,
4208 hdr->msg.get.return_wmd.wh_interface_cookie,
4209 hdr->msg.get.return_wmd.wh_object_cookie,
4210 hdr->msg.get.match_bits);
4211 CWARN(" Length %d, src offset %d\n",
4212 hdr->msg.get.sink_length,
4213 hdr->msg.get.src_offset);
4217 CWARN(" dst md %#llx.%#llx, "
4218 "manipulated length %d\n",
4219 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4220 hdr->msg.ack.dst_wmd.wh_object_cookie,
4221 hdr->msg.ack.mlength);
4224 case LNET_MSG_REPLY:
4225 CWARN(" dst md %#llx.%#llx, "
4227 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4228 hdr->msg.reply.dst_wmd.wh_object_cookie,
4229 hdr->payload_length);
4235 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4236 void *private, int rdma_req)
4238 struct lnet_peer_ni *lpni;
4239 struct lnet_msg *msg;
4240 __u32 payload_length;
4241 lnet_pid_t dest_pid;
4242 lnet_nid_t dest_nid;
4250 LASSERT (!in_interrupt ());
4252 type = le32_to_cpu(hdr->type);
4253 src_nid = le64_to_cpu(hdr->src_nid);
4254 dest_nid = le64_to_cpu(hdr->dest_nid);
4255 dest_pid = le32_to_cpu(hdr->dest_pid);
4256 payload_length = le32_to_cpu(hdr->payload_length);
4258 for_me = (ni->ni_nid == dest_nid);
4259 cpt = lnet_cpt_of_nid(from_nid, ni);
4261 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4262 libcfs_nid2str(dest_nid),
4263 libcfs_nid2str(ni->ni_nid),
4264 libcfs_nid2str(src_nid),
4265 lnet_msgtyp2str(type),
4266 (for_me) ? "for me" : "routed");
4271 if (payload_length > 0) {
4272 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4273 libcfs_nid2str(from_nid),
4274 libcfs_nid2str(src_nid),
4275 lnet_msgtyp2str(type), payload_length);
4281 case LNET_MSG_REPLY:
4282 if (payload_length >
4283 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4284 CERROR("%s, src %s: bad %s payload %d "
4285 "(%d max expected)\n",
4286 libcfs_nid2str(from_nid),
4287 libcfs_nid2str(src_nid),
4288 lnet_msgtyp2str(type),
4290 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4296 CERROR("%s, src %s: Bad message type 0x%x\n",
4297 libcfs_nid2str(from_nid),
4298 libcfs_nid2str(src_nid), type);
4302 if (the_lnet.ln_routing &&
4303 ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4305 spin_lock(&ni->ni_net->net_lock);
4306 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4307 spin_unlock(&ni->ni_net->net_lock);
4308 if (ni->ni_status != NULL &&
4309 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4310 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4317 lnet_push_update_to_peers(1);
4319 /* Regard a bad destination NID as a protocol error. Senders should
4320 * know what they're doing; if they don't they're misconfigured, buggy
4321 * or malicious so we chop them off at the knees :) */
4324 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4325 /* should have gone direct */
4326 CERROR("%s, src %s: Bad dest nid %s "
4327 "(should have been sent direct)\n",
4328 libcfs_nid2str(from_nid),
4329 libcfs_nid2str(src_nid),
4330 libcfs_nid2str(dest_nid));
4334 if (lnet_islocalnid(dest_nid)) {
4335 /* dest is another local NI; sender should have used
4336 * this node's NID on its own network */
4337 CERROR("%s, src %s: Bad dest nid %s "
4338 "(it's my nid but on a different network)\n",
4339 libcfs_nid2str(from_nid),
4340 libcfs_nid2str(src_nid),
4341 libcfs_nid2str(dest_nid));
4345 if (rdma_req && type == LNET_MSG_GET) {
4346 CERROR("%s, src %s: Bad optimized GET for %s "
4347 "(final destination must be me)\n",
4348 libcfs_nid2str(from_nid),
4349 libcfs_nid2str(src_nid),
4350 libcfs_nid2str(dest_nid));
4354 if (!the_lnet.ln_routing) {
4355 CERROR("%s, src %s: Dropping message for %s "
4356 "(routing not enabled)\n",
4357 libcfs_nid2str(from_nid),
4358 libcfs_nid2str(src_nid),
4359 libcfs_nid2str(dest_nid));
4364 /* Message looks OK; we're not going to return an error, so we MUST
4365 * call back lnd_recv() come what may... */
4367 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4368 fail_peer(src_nid, 0)) { /* shall we now? */
4369 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4370 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4371 lnet_msgtyp2str(type));
4375 if (!list_empty(&the_lnet.ln_drop_rules) &&
4376 lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4378 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4379 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4380 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4384 if (lnet_drop_asym_route && for_me &&
4385 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4386 struct lnet_net *net;
4387 struct lnet_remotenet *rnet;
4390 /* we are dealing with a routed message,
4391 * so see if route to reach src_nid goes through from_nid
4394 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4396 lnet_net_unlock(cpt);
4397 CERROR("net %s not found\n",
4398 libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4402 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4404 struct lnet_peer *gw = NULL;
4405 struct lnet_peer_ni *lpni = NULL;
4406 struct lnet_route *route;
4408 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4410 gw = route->lr_gateway;
4411 if (route->lr_lnet != net->net_id)
4414 * if the nid is one of the gateway's NIDs
4415 * then this is a valid gateway
4417 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4418 NULL, lpni)) != NULL) {
4419 if (lpni->lpni_nid == from_nid) {
4426 lnet_net_unlock(cpt);
4428 /* we would not use from_nid to route a message to
4430 * => asymmetric routing detected but forbidden
4432 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4433 libcfs_nid2str(from_nid),
4434 libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4439 msg = lnet_msg_alloc();
4441 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4442 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4443 lnet_msgtyp2str(type));
4447 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4448 * pointers NULL etc */
4450 msg->msg_type = type;
4451 msg->msg_private = private;
4452 msg->msg_receiving = 1;
4453 msg->msg_rdma_get = rdma_req;
4454 msg->msg_len = msg->msg_wanted = payload_length;
4455 msg->msg_offset = 0;
4456 msg->msg_hdr = *hdr;
4457 /* for building message event */
4458 msg->msg_from = from_nid;
4460 msg->msg_target.pid = dest_pid;
4461 msg->msg_target.nid = dest_nid;
4462 msg->msg_routing = 1;
4465 /* convert common msg->hdr fields to host byteorder */
4466 msg->msg_hdr.type = type;
4467 msg->msg_hdr.src_nid = src_nid;
4468 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4469 msg->msg_hdr.dest_nid = dest_nid;
4470 msg->msg_hdr.dest_pid = dest_pid;
4471 msg->msg_hdr.payload_length = payload_length;
4475 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4477 lnet_net_unlock(cpt);
4478 CERROR("%s, src %s: Dropping %s "
4479 "(error %ld looking up sender)\n",
4480 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4481 lnet_msgtyp2str(type), PTR_ERR(lpni));
4483 if (rc == -ESHUTDOWN)
4484 /* We are shutting down. Don't do anything more */
4489 if (the_lnet.ln_routing)
4490 lpni->lpni_last_alive = ktime_get_seconds();
4492 msg->msg_rxpeer = lpni;
4494 lnet_ni_addref_locked(ni, cpt);
4495 /* Multi-Rail: Primary NID of source. */
4496 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4499 * mark the status of this lpni as UP since we received a message
4500 * from it. The ping response reports back the ns_status which is
4501 * marked on the remote as up or down and we cache it here.
4503 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4505 lnet_msg_commit(msg, cpt);
4507 /* message delay simulation */
4508 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4509 lnet_delay_rule_match_locked(hdr, msg))) {
4510 lnet_net_unlock(cpt);
4515 rc = lnet_parse_forward_locked(ni, msg);
4516 lnet_net_unlock(cpt);
4521 if (rc == LNET_CREDIT_OK) {
4522 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4523 0, payload_length, payload_length);
4528 lnet_net_unlock(cpt);
4530 rc = lnet_parse_local(ni, msg);
4536 LASSERT(msg->msg_md == NULL);
4537 lnet_finalize(msg, rc);
4540 lnet_drop_message(ni, cpt, private, payload_length, type);
4543 EXPORT_SYMBOL(lnet_parse);
4546 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4548 while (!list_empty(head)) {
4549 struct lnet_process_id id = {0};
4550 struct lnet_msg *msg;
4552 msg = list_entry(head->next, struct lnet_msg, msg_list);
4553 list_del(&msg->msg_list);
4555 id.nid = msg->msg_hdr.src_nid;
4556 id.pid = msg->msg_hdr.src_pid;
4558 LASSERT(msg->msg_md == NULL);
4559 LASSERT(msg->msg_rx_delayed);
4560 LASSERT(msg->msg_rxpeer != NULL);
4561 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4563 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4564 " offset %d length %d: %s\n",
4566 msg->msg_hdr.msg.put.ptl_index,
4567 msg->msg_hdr.msg.put.match_bits,
4568 msg->msg_hdr.msg.put.offset,
4569 msg->msg_hdr.payload_length, reason);
4571 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4572 * called lnet_drop_message(), so I just hang onto msg as well
4573 * until that's done */
4575 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4576 msg->msg_private, msg->msg_len,
4579 msg->msg_no_resend = true;
4581 * NB: message will not generate event because w/o attached MD,
4582 * but we still should give error code so lnet_msg_decommit()
4583 * can skip counters operations and other checks.
4585 lnet_finalize(msg, -ENOENT);
4590 lnet_recv_delayed_msg_list(struct list_head *head)
4592 while (!list_empty(head)) {
4593 struct lnet_msg *msg;
4594 struct lnet_process_id id;
4596 msg = list_entry(head->next, struct lnet_msg, msg_list);
4597 list_del(&msg->msg_list);
4599 /* md won't disappear under me, since each msg
4600 * holds a ref on it */
4602 id.nid = msg->msg_hdr.src_nid;
4603 id.pid = msg->msg_hdr.src_pid;
4605 LASSERT(msg->msg_rx_delayed);
4606 LASSERT(msg->msg_md != NULL);
4607 LASSERT(msg->msg_rxpeer != NULL);
4608 LASSERT(msg->msg_rxni != NULL);
4609 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4611 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4612 "match %llu offset %d length %d.\n",
4613 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4614 msg->msg_hdr.msg.put.match_bits,
4615 msg->msg_hdr.msg.put.offset,
4616 msg->msg_hdr.payload_length);
4618 lnet_recv_put(msg->msg_rxni, msg);
4623 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4624 struct lnet_libmd *md, struct lnet_handle_md mdh)
4627 struct lnet_rsp_tracker *local_rspt;
4630 * MD has a refcount taken by message so it's not going away.
4631 * The MD however can be looked up. We need to secure the access
4632 * to the md_rspt_ptr by taking the res_lock.
4633 * The rspt can be accessed without protection up to when it gets
4634 * added to the list.
4638 local_rspt = md->md_rspt_ptr;
4639 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4640 if (local_rspt != NULL) {
4642 * we already have an rspt attached to the md, so we'll
4643 * update the deadline on that one.
4645 lnet_rspt_free(rspt, cpt);
4648 rspt->rspt_mdh = mdh;
4649 rspt->rspt_cpt = cpt;
4650 /* store the rspt so we can access it when we get the REPLY */
4651 md->md_rspt_ptr = rspt;
4654 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4657 * add to the list of tracked responses. It's added to tail of the
4658 * list in order to expire all the older entries first.
4661 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4662 lnet_net_unlock(cpt);
4663 lnet_res_unlock(cpt);
4667 * Initiate an asynchronous PUT operation.
4669 * There are several events associated with a PUT: completion of the send on
4670 * the initiator node (LNET_EVENT_SEND), and when the send completes
4671 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4672 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4673 * used at the target node to indicate the completion of incoming data
4676 * The local events will be logged in the EQ associated with the MD pointed to
4677 * by \a mdh handle. Using a MD without an associated EQ results in these
4678 * events being discarded. In this case, the caller must have another
4679 * mechanism (e.g., a higher level protocol) for determining when it is safe
4680 * to modify the memory region associated with the MD.
4682 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4683 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4685 * \param self Indicates the NID of a local interface through which to send
4686 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4687 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4688 * must be "free floating" (See LNetMDBind()).
4689 * \param ack Controls whether an acknowledgment is requested.
4690 * Acknowledgments are only sent when they are requested by the initiating
4691 * process and the target MD enables them.
4692 * \param target A process identifier for the target process.
4693 * \param portal The index in the \a target's portal table.
4694 * \param match_bits The match bits to use for MD selection at the target
4696 * \param offset The offset into the target MD (only used when the target
4697 * MD has the LNET_MD_MANAGE_REMOTE option set).
4698 * \param hdr_data 64 bits of user data that can be included in the message
4699 * header. This data is written to an event queue entry at the target if an
4700 * EQ is present on the matching MD.
4702 * \retval 0 Success, and only in this case events will be generated
4703 * and logged to EQ (if it exists).
4704 * \retval -EIO Simulated failure.
4705 * \retval -ENOMEM Memory allocation failure.
4706 * \retval -ENOENT Invalid MD object.
4708 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4711 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4712 struct lnet_process_id target, unsigned int portal,
4713 __u64 match_bits, unsigned int offset,
4716 struct lnet_msg *msg;
4717 struct lnet_libmd *md;
4720 struct lnet_rsp_tracker *rspt = NULL;
4722 LASSERT(the_lnet.ln_refcount > 0);
4724 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4725 fail_peer(target.nid, 1)) { /* shall we now? */
4726 CERROR("Dropping PUT to %s: simulated failure\n",
4727 libcfs_id2str(target));
4731 msg = lnet_msg_alloc();
4733 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4734 libcfs_id2str(target));
4737 msg->msg_vmflush = !!memory_pressure_get();
4739 cpt = lnet_cpt_of_cookie(mdh.cookie);
4741 if (ack == LNET_ACK_REQ) {
4742 rspt = lnet_rspt_alloc(cpt);
4744 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4745 libcfs_id2str(target));
4748 INIT_LIST_HEAD(&rspt->rspt_on_list);
4753 md = lnet_handle2md(&mdh);
4754 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4755 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4756 match_bits, portal, libcfs_id2str(target),
4757 md == NULL ? -1 : md->md_threshold);
4758 if (md != NULL && md->md_me != NULL)
4759 CERROR("Source MD also attached to portal %d\n",
4760 md->md_me->me_portal);
4761 lnet_res_unlock(cpt);
4763 lnet_rspt_free(rspt, cpt);
4768 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4770 lnet_msg_attach_md(msg, md, 0, 0);
4772 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4774 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4775 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4776 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4777 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4779 /* NB handles only looked up by creator (no flips) */
4780 if (ack == LNET_ACK_REQ) {
4781 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4782 the_lnet.ln_interface_cookie;
4783 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4784 md->md_lh.lh_cookie;
4786 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4787 LNET_WIRE_HANDLE_COOKIE_NONE;
4788 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4789 LNET_WIRE_HANDLE_COOKIE_NONE;
4792 lnet_res_unlock(cpt);
4794 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4796 if (ack == LNET_ACK_REQ)
4797 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4799 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4803 rc = lnet_send(self, msg, LNET_NID_ANY);
4806 CNETERR("Error sending PUT to %s: %d\n",
4807 libcfs_id2str(target), rc);
4808 msg->msg_no_resend = true;
4809 lnet_finalize(msg, rc);
4812 /* completion will be signalled by an event */
4815 EXPORT_SYMBOL(LNetPut);
4818 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
4819 * returns a msg for the LND to pass to lnet_finalize() when the sink
4820 * data has been received.
4822 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4823 * lnet_finalize() is called on it, so the LND must call this first
4826 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4828 struct lnet_msg *msg = lnet_msg_alloc();
4829 struct lnet_libmd *getmd = getmsg->msg_md;
4830 struct lnet_process_id peer_id = getmsg->msg_target;
4833 LASSERT(!getmsg->msg_target_is_router);
4834 LASSERT(!getmsg->msg_routing);
4837 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4838 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4842 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4845 LASSERT(getmd->md_refcount > 0);
4847 if (getmd->md_threshold == 0) {
4848 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4849 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4851 lnet_res_unlock(cpt);
4855 LASSERT(getmd->md_offset == 0);
4857 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4858 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4860 /* setup information for lnet_build_msg_event */
4861 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4862 msg->msg_from = peer_id.nid;
4863 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4864 msg->msg_hdr.src_nid = peer_id.nid;
4865 msg->msg_hdr.payload_length = getmd->md_length;
4866 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4868 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4869 lnet_res_unlock(cpt);
4871 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4874 lnet_msg_commit(msg, cpt);
4875 lnet_net_unlock(cpt);
4877 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4882 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4885 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4886 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4887 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4889 lnet_net_unlock(cpt);
4896 EXPORT_SYMBOL(lnet_create_reply_msg);
4899 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4902 /* Set the REPLY length, now the RDMA that elides the REPLY message has
4903 * completed and I know it. */
4904 LASSERT(reply != NULL);
4905 LASSERT(reply->msg_type == LNET_MSG_GET);
4906 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4908 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
4909 * the end of my buffer, I might as well be dead. */
4910 LASSERT(len <= reply->msg_ev.mlength);
4912 reply->msg_ev.mlength = len;
4914 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4917 * Initiate an asynchronous GET operation.
4919 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4920 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4921 * the target node in the REPLY has been written to local MD.
4923 * On the target node, an LNET_EVENT_GET is logged when the GET request
4924 * arrives and is accepted into a MD.
4926 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4927 * \param mdh A handle for the MD that describes the memory into which the
4928 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4930 * \retval 0 Success, and only in this case events will be generated
4931 * and logged to EQ (if it exists) of the MD.
4932 * \retval -EIO Simulated failure.
4933 * \retval -ENOMEM Memory allocation failure.
4934 * \retval -ENOENT Invalid MD object.
4937 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4938 struct lnet_process_id target, unsigned int portal,
4939 __u64 match_bits, unsigned int offset, bool recovery)
4941 struct lnet_msg *msg;
4942 struct lnet_libmd *md;
4943 struct lnet_rsp_tracker *rspt;
4947 LASSERT(the_lnet.ln_refcount > 0);
4949 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4950 fail_peer(target.nid, 1)) /* shall we now? */
4952 CERROR("Dropping GET to %s: simulated failure\n",
4953 libcfs_id2str(target));
4957 msg = lnet_msg_alloc();
4959 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4960 libcfs_id2str(target));
4964 cpt = lnet_cpt_of_cookie(mdh.cookie);
4966 rspt = lnet_rspt_alloc(cpt);
4968 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4969 libcfs_id2str(target));
4972 INIT_LIST_HEAD(&rspt->rspt_on_list);
4974 msg->msg_recovery = recovery;
4978 md = lnet_handle2md(&mdh);
4979 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4980 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4981 match_bits, portal, libcfs_id2str(target),
4982 md == NULL ? -1 : md->md_threshold);
4983 if (md != NULL && md->md_me != NULL)
4984 CERROR("REPLY MD also attached to portal %d\n",
4985 md->md_me->me_portal);
4987 lnet_res_unlock(cpt);
4990 lnet_rspt_free(rspt, cpt);
4994 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4996 lnet_msg_attach_md(msg, md, 0, 0);
4998 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5000 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5001 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5002 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5003 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5005 /* NB handles only looked up by creator (no flips) */
5006 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5007 the_lnet.ln_interface_cookie;
5008 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5009 md->md_lh.lh_cookie;
5011 lnet_res_unlock(cpt);
5013 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5015 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5017 rc = lnet_send(self, msg, LNET_NID_ANY);
5019 CNETERR("Error sending GET to %s: %d\n",
5020 libcfs_id2str(target), rc);
5021 msg->msg_no_resend = true;
5022 lnet_finalize(msg, rc);
5025 /* completion will be signalled by an event */
5028 EXPORT_SYMBOL(LNetGet);
5031 * Calculate distance to node at \a dstnid.
5033 * \param dstnid Target NID.
5034 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5036 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5039 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5040 * local_nid_dist_zero is set, which is the default.
5041 * \retval positives Distance to target NID, i.e. number of hops plus one.
5042 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5045 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5047 struct list_head *e;
5048 struct lnet_ni *ni = NULL;
5049 struct lnet_remotenet *rnet;
5050 __u32 dstnet = LNET_NIDNET(dstnid);
5054 struct list_head *rn_list;
5056 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5057 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5058 * keep order 0 free for 0@lo and order 1 free for a local NID
5061 LASSERT(the_lnet.ln_refcount > 0);
5063 cpt = lnet_net_lock_current();
5065 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5066 if (ni->ni_nid == dstnid) {
5067 if (srcnidp != NULL)
5069 if (orderp != NULL) {
5070 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5075 lnet_net_unlock(cpt);
5077 return local_nid_dist_zero ? 0 : 1;
5080 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5081 /* Check if ni was originally created in
5082 * current net namespace.
5083 * If not, assign order above 0xffff0000,
5084 * to make this ni not a priority. */
5085 if (current->nsproxy &&
5086 !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5087 order += 0xffff0000;
5088 if (srcnidp != NULL)
5089 *srcnidp = ni->ni_nid;
5092 lnet_net_unlock(cpt);
5099 rn_list = lnet_net2rnethash(dstnet);
5100 list_for_each(e, rn_list) {
5101 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5103 if (rnet->lrn_net == dstnet) {
5104 struct lnet_route *route;
5105 struct lnet_route *shortest = NULL;
5106 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5109 LASSERT(!list_empty(&rnet->lrn_routes));
5111 list_for_each_entry(route, &rnet->lrn_routes,
5113 route_hops = route->lr_hops;
5114 if (route_hops == LNET_UNDEFINED_HOPS)
5116 if (shortest == NULL ||
5117 route_hops < shortest_hops) {
5119 shortest_hops = route_hops;
5123 LASSERT(shortest != NULL);
5124 hops = shortest_hops;
5125 if (srcnidp != NULL) {
5126 struct lnet_net *net;
5127 net = lnet_get_net_locked(shortest->lr_lnet);
5129 ni = lnet_get_next_ni_locked(net, NULL);
5130 *srcnidp = ni->ni_nid;
5134 lnet_net_unlock(cpt);
5140 lnet_net_unlock(cpt);
5141 return -EHOSTUNREACH;
5143 EXPORT_SYMBOL(LNetDist);