4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/lnet/lib-move.c
34 * Data movement routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/pagemap.h>
41 #include <lnet/lib-lnet.h>
42 #include <linux/nsproxy.h>
43 #include <net/net_namespace.h>
45 static int local_nid_dist_zero = 1;
46 module_param(local_nid_dist_zero, int, 0444);
47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
49 struct lnet_send_data {
50 struct lnet_ni *sd_best_ni;
51 struct lnet_peer_ni *sd_best_lpni;
52 struct lnet_peer_ni *sd_final_dst_lpni;
53 struct lnet_peer *sd_peer;
54 struct lnet_peer *sd_gw_peer;
55 struct lnet_peer_ni *sd_gw_lpni;
56 struct lnet_peer_net *sd_peer_net;
57 struct lnet_msg *sd_msg;
58 lnet_nid_t sd_dst_nid;
59 lnet_nid_t sd_src_nid;
60 lnet_nid_t sd_rtr_nid;
66 static inline struct lnet_comm_count *
67 get_stats_counts(struct lnet_element_stats *stats,
68 enum lnet_stats_type stats_type)
71 case LNET_STATS_TYPE_SEND:
72 return &stats->el_send_stats;
73 case LNET_STATS_TYPE_RECV:
74 return &stats->el_recv_stats;
75 case LNET_STATS_TYPE_DROP:
76 return &stats->el_drop_stats;
78 CERROR("Unknown stats type\n");
84 void lnet_incr_stats(struct lnet_element_stats *stats,
85 enum lnet_msg_type msg_type,
86 enum lnet_stats_type stats_type)
88 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
94 atomic_inc(&counts->co_ack_count);
97 atomic_inc(&counts->co_put_count);
100 atomic_inc(&counts->co_get_count);
103 atomic_inc(&counts->co_reply_count);
106 atomic_inc(&counts->co_hello_count);
109 CERROR("There is a BUG in the code. Unknown message type\n");
114 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
115 enum lnet_stats_type stats_type)
117 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
121 return (atomic_read(&counts->co_ack_count) +
122 atomic_read(&counts->co_put_count) +
123 atomic_read(&counts->co_get_count) +
124 atomic_read(&counts->co_reply_count) +
125 atomic_read(&counts->co_hello_count));
128 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
129 struct lnet_comm_count *counts)
131 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
132 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
133 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
134 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
135 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
138 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
139 struct lnet_element_stats *stats)
141 struct lnet_comm_count *counts;
146 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
149 assign_stats(&msg_stats->im_send_stats, counts);
151 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
154 assign_stats(&msg_stats->im_recv_stats, counts);
156 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
159 assign_stats(&msg_stats->im_drop_stats, counts);
163 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
165 struct lnet_test_peer *tp;
166 struct list_head *el;
167 struct list_head *next;
168 struct list_head cull;
170 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
171 if (threshold != 0) {
172 /* Adding a new entry */
173 LIBCFS_ALLOC(tp, sizeof(*tp));
178 tp->tp_threshold = threshold;
181 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
186 /* removing entries */
187 INIT_LIST_HEAD(&cull);
191 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
192 tp = list_entry(el, struct lnet_test_peer, tp_list);
194 if (tp->tp_threshold == 0 || /* needs culling anyway */
195 nid == LNET_NID_ANY || /* removing all entries */
196 tp->tp_nid == nid) { /* matched this one */
197 list_del(&tp->tp_list);
198 list_add(&tp->tp_list, &cull);
204 while (!list_empty(&cull)) {
205 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
207 list_del(&tp->tp_list);
208 LIBCFS_FREE(tp, sizeof(*tp));
214 fail_peer (lnet_nid_t nid, int outgoing)
216 struct lnet_test_peer *tp;
217 struct list_head *el;
218 struct list_head *next;
219 struct list_head cull;
222 INIT_LIST_HEAD(&cull);
224 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
227 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
228 tp = list_entry(el, struct lnet_test_peer, tp_list);
230 if (tp->tp_threshold == 0) {
233 /* only cull zombies on outgoing tests,
234 * since we may be at interrupt priority on
235 * incoming messages. */
236 list_del(&tp->tp_list);
237 list_add(&tp->tp_list, &cull);
242 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
243 nid == tp->tp_nid) { /* fail this peer */
246 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
249 tp->tp_threshold == 0) {
251 list_del(&tp->tp_list);
252 list_add(&tp->tp_list, &cull);
261 while (!list_empty(&cull)) {
262 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
263 list_del(&tp->tp_list);
265 LIBCFS_FREE(tp, sizeof(*tp));
272 lnet_iov_nob(unsigned int niov, struct kvec *iov)
274 unsigned int nob = 0;
276 LASSERT(niov == 0 || iov != NULL);
278 nob += (iov++)->iov_len;
282 EXPORT_SYMBOL(lnet_iov_nob);
285 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
286 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
289 /* NB diov, siov are READ-ONLY */
290 unsigned int this_nob;
295 /* skip complete frags before 'doffset' */
297 while (doffset >= diov->iov_len) {
298 doffset -= diov->iov_len;
304 /* skip complete frags before 'soffset' */
306 while (soffset >= siov->iov_len) {
307 soffset -= siov->iov_len;
316 this_nob = MIN(diov->iov_len - doffset,
317 siov->iov_len - soffset);
318 this_nob = MIN(this_nob, nob);
320 memcpy((char *)diov->iov_base + doffset,
321 (char *)siov->iov_base + soffset, this_nob);
324 if (diov->iov_len > doffset + this_nob) {
332 if (siov->iov_len > soffset + this_nob) {
341 EXPORT_SYMBOL(lnet_copy_iov2iov);
344 lnet_extract_iov(int dst_niov, struct kvec *dst,
345 int src_niov, struct kvec *src,
346 unsigned int offset, unsigned int len)
348 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
349 * for exactly 'len' bytes, and return the number of entries.
350 * NB not destructive to 'src' */
351 unsigned int frag_len;
354 if (len == 0) /* no data => */
355 return (0); /* no frags */
357 LASSERT(src_niov > 0);
358 while (offset >= src->iov_len) { /* skip initial frags */
359 offset -= src->iov_len;
362 LASSERT(src_niov > 0);
367 LASSERT(src_niov > 0);
368 LASSERT((int)niov <= dst_niov);
370 frag_len = src->iov_len - offset;
371 dst->iov_base = ((char *)src->iov_base) + offset;
373 if (len <= frag_len) {
378 dst->iov_len = frag_len;
388 EXPORT_SYMBOL(lnet_extract_iov);
392 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
394 unsigned int nob = 0;
396 LASSERT(niov == 0 || kiov != NULL);
398 nob += (kiov++)->kiov_len;
402 EXPORT_SYMBOL(lnet_kiov_nob);
405 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
406 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
409 /* NB diov, siov are READ-ONLY */
410 unsigned int this_nob;
417 LASSERT (!in_interrupt ());
420 while (doffset >= diov->kiov_len) {
421 doffset -= diov->kiov_len;
428 while (soffset >= siov->kiov_len) {
429 soffset -= siov->kiov_len;
438 this_nob = MIN(diov->kiov_len - doffset,
439 siov->kiov_len - soffset);
440 this_nob = MIN(this_nob, nob);
443 daddr = ((char *)kmap(diov->kiov_page)) +
444 diov->kiov_offset + doffset;
446 saddr = ((char *)kmap(siov->kiov_page)) +
447 siov->kiov_offset + soffset;
449 /* Vanishing risk of kmap deadlock when mapping 2 pages.
450 * However in practice at least one of the kiovs will be mapped
451 * kernel pages and the map/unmap will be NOOPs */
453 memcpy (daddr, saddr, this_nob);
456 if (diov->kiov_len > doffset + this_nob) {
460 kunmap(diov->kiov_page);
467 if (siov->kiov_len > soffset + this_nob) {
471 kunmap(siov->kiov_page);
480 kunmap(diov->kiov_page);
482 kunmap(siov->kiov_page);
484 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
487 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
488 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
491 /* NB iov, kiov are READ-ONLY */
492 unsigned int this_nob;
498 LASSERT (!in_interrupt ());
501 while (iovoffset >= iov->iov_len) {
502 iovoffset -= iov->iov_len;
509 while (kiovoffset >= kiov->kiov_len) {
510 kiovoffset -= kiov->kiov_len;
519 this_nob = MIN(iov->iov_len - iovoffset,
520 kiov->kiov_len - kiovoffset);
521 this_nob = MIN(this_nob, nob);
524 addr = ((char *)kmap(kiov->kiov_page)) +
525 kiov->kiov_offset + kiovoffset;
527 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
530 if (iov->iov_len > iovoffset + this_nob) {
531 iovoffset += this_nob;
538 if (kiov->kiov_len > kiovoffset + this_nob) {
540 kiovoffset += this_nob;
542 kunmap(kiov->kiov_page);
552 kunmap(kiov->kiov_page);
554 EXPORT_SYMBOL(lnet_copy_kiov2iov);
557 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
558 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
561 /* NB kiov, iov are READ-ONLY */
562 unsigned int this_nob;
568 LASSERT (!in_interrupt ());
571 while (kiovoffset >= kiov->kiov_len) {
572 kiovoffset -= kiov->kiov_len;
579 while (iovoffset >= iov->iov_len) {
580 iovoffset -= iov->iov_len;
589 this_nob = MIN(kiov->kiov_len - kiovoffset,
590 iov->iov_len - iovoffset);
591 this_nob = MIN(this_nob, nob);
594 addr = ((char *)kmap(kiov->kiov_page)) +
595 kiov->kiov_offset + kiovoffset;
597 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
600 if (kiov->kiov_len > kiovoffset + this_nob) {
602 kiovoffset += this_nob;
604 kunmap(kiov->kiov_page);
611 if (iov->iov_len > iovoffset + this_nob) {
612 iovoffset += this_nob;
621 kunmap(kiov->kiov_page);
623 EXPORT_SYMBOL(lnet_copy_iov2kiov);
626 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
627 int src_niov, lnet_kiov_t *src,
628 unsigned int offset, unsigned int len)
630 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
631 * for exactly 'len' bytes, and return the number of entries.
632 * NB not destructive to 'src' */
633 unsigned int frag_len;
636 if (len == 0) /* no data => */
637 return (0); /* no frags */
639 LASSERT(src_niov > 0);
640 while (offset >= src->kiov_len) { /* skip initial frags */
641 offset -= src->kiov_len;
644 LASSERT(src_niov > 0);
649 LASSERT(src_niov > 0);
650 LASSERT((int)niov <= dst_niov);
652 frag_len = src->kiov_len - offset;
653 dst->kiov_page = src->kiov_page;
654 dst->kiov_offset = src->kiov_offset + offset;
656 if (len <= frag_len) {
658 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
662 dst->kiov_len = frag_len;
663 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
673 EXPORT_SYMBOL(lnet_extract_kiov);
676 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
677 int delayed, unsigned int offset, unsigned int mlen,
680 unsigned int niov = 0;
681 struct kvec *iov = NULL;
682 lnet_kiov_t *kiov = NULL;
685 LASSERT (!in_interrupt ());
686 LASSERT (mlen == 0 || msg != NULL);
689 LASSERT(msg->msg_receiving);
690 LASSERT(!msg->msg_sending);
691 LASSERT(rlen == msg->msg_len);
692 LASSERT(mlen <= msg->msg_len);
693 LASSERT(msg->msg_offset == offset);
694 LASSERT(msg->msg_wanted == mlen);
696 msg->msg_receiving = 0;
699 niov = msg->msg_niov;
701 kiov = msg->msg_kiov;
704 LASSERT ((iov == NULL) != (kiov == NULL));
708 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
709 niov, iov, kiov, offset, mlen,
712 lnet_finalize(msg, rc);
716 lnet_setpayloadbuffer(struct lnet_msg *msg)
718 struct lnet_libmd *md = msg->msg_md;
720 LASSERT(msg->msg_len > 0);
721 LASSERT(!msg->msg_routing);
723 LASSERT(msg->msg_niov == 0);
724 LASSERT(msg->msg_iov == NULL);
725 LASSERT(msg->msg_kiov == NULL);
727 msg->msg_niov = md->md_niov;
728 if ((md->md_options & LNET_MD_KIOV) != 0)
729 msg->msg_kiov = md->md_iov.kiov;
731 msg->msg_iov = md->md_iov.iov;
735 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
736 unsigned int offset, unsigned int len)
738 msg->msg_type = type;
739 msg->msg_target = target;
741 msg->msg_offset = offset;
744 lnet_setpayloadbuffer(msg);
746 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
747 msg->msg_hdr.type = cpu_to_le32(type);
748 /* dest_nid will be overwritten by lnet_select_pathway() */
749 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
750 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
751 /* src_nid will be set later */
752 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
753 msg->msg_hdr.payload_length = cpu_to_le32(len);
757 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
759 void *priv = msg->msg_private;
762 LASSERT (!in_interrupt ());
763 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
764 (msg->msg_txcredit && msg->msg_peertxcredit));
766 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
768 msg->msg_no_resend = true;
769 lnet_finalize(msg, rc);
774 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
778 LASSERT(!msg->msg_sending);
779 LASSERT(msg->msg_receiving);
780 LASSERT(!msg->msg_rx_ready_delay);
781 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
783 msg->msg_rx_ready_delay = 1;
784 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
787 CERROR("recv from %s / send to %s aborted: "
788 "eager_recv failed %d\n",
789 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
790 libcfs_id2str(msg->msg_target), rc);
791 LASSERT(rc < 0); /* required by my callers */
797 /* NB: returns 1 when alive, 0 when dead, negative when error;
798 * may drop the lnet_net_lock */
800 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
801 struct lnet_msg *msg)
803 if (!lnet_peer_aliveness_enabled(lpni))
807 * If we're resending a message, let's attempt to send it even if
808 * the peer is down to fulfill our resend quota on the message
810 if (msg->msg_retry_count > 0)
813 /* try and send recovery messages irregardless */
814 if (msg->msg_recovery)
817 /* always send any responses */
818 if (msg->msg_type == LNET_MSG_ACK ||
819 msg->msg_type == LNET_MSG_REPLY)
822 return lnet_is_peer_ni_alive(lpni);
826 * \param msg The message to be sent.
827 * \param do_send True if lnet_ni_send() should be called in this function.
828 * lnet_send() is going to lnet_net_unlock immediately after this, so
829 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
831 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
832 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
833 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
834 * \retval -ECANCELED If the MD of the message has been unlinked.
837 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
839 struct lnet_peer_ni *lp = msg->msg_txpeer;
840 struct lnet_ni *ni = msg->msg_txni;
841 int cpt = msg->msg_tx_cpt;
842 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
844 /* non-lnet_send() callers have checked before */
845 LASSERT(!do_send || msg->msg_tx_delayed);
846 LASSERT(!msg->msg_receiving);
847 LASSERT(msg->msg_tx_committed);
848 /* can't get here if we're sending to the loopback interface */
849 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
851 /* NB 'lp' is always the next hop */
852 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
853 lnet_peer_alive_locked(ni, lp, msg) == 0) {
854 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
855 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
857 lnet_net_unlock(cpt);
859 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
861 LNET_STATS_TYPE_DROP);
863 lnet_incr_stats(&msg->msg_txni->ni_stats,
865 LNET_STATS_TYPE_DROP);
867 CNETERR("Dropping message for %s: peer not alive\n",
868 libcfs_id2str(msg->msg_target));
869 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
871 lnet_finalize(msg, -EHOSTUNREACH);
874 return -EHOSTUNREACH;
877 if (msg->msg_md != NULL &&
878 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
879 lnet_net_unlock(cpt);
881 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
882 "called on the MD/ME.\n",
883 libcfs_id2str(msg->msg_target));
885 msg->msg_no_resend = true;
886 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
887 msg, libcfs_id2str(msg->msg_target));
888 lnet_finalize(msg, -ECANCELED);
895 if (!msg->msg_peertxcredit) {
896 spin_lock(&lp->lpni_lock);
897 LASSERT((lp->lpni_txcredits < 0) ==
898 !list_empty(&lp->lpni_txq));
900 msg->msg_peertxcredit = 1;
901 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
902 lp->lpni_txcredits--;
904 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
905 lp->lpni_mintxcredits = lp->lpni_txcredits;
907 if (lp->lpni_txcredits < 0) {
908 msg->msg_tx_delayed = 1;
909 list_add_tail(&msg->msg_list, &lp->lpni_txq);
910 spin_unlock(&lp->lpni_lock);
911 return LNET_CREDIT_WAIT;
913 spin_unlock(&lp->lpni_lock);
916 if (!msg->msg_txcredit) {
917 LASSERT((tq->tq_credits < 0) ==
918 !list_empty(&tq->tq_delayed));
920 msg->msg_txcredit = 1;
922 atomic_dec(&ni->ni_tx_credits);
924 if (tq->tq_credits < tq->tq_credits_min)
925 tq->tq_credits_min = tq->tq_credits;
927 if (tq->tq_credits < 0) {
928 msg->msg_tx_delayed = 1;
929 list_add_tail(&msg->msg_list, &tq->tq_delayed);
930 return LNET_CREDIT_WAIT;
934 /* unset the tx_delay flag as we're going to send it now */
935 msg->msg_tx_delayed = 0;
938 lnet_net_unlock(cpt);
939 lnet_ni_send(ni, msg);
942 return LNET_CREDIT_OK;
946 static struct lnet_rtrbufpool *
947 lnet_msg2bufpool(struct lnet_msg *msg)
949 struct lnet_rtrbufpool *rbp;
952 LASSERT(msg->msg_rx_committed);
954 cpt = msg->msg_rx_cpt;
955 rbp = &the_lnet.ln_rtrpools[cpt][0];
957 LASSERT(msg->msg_len <= LNET_MTU);
958 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
960 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
967 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
969 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
970 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
971 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
972 * received or OK to receive */
973 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
974 struct lnet_peer *lp;
975 struct lnet_rtrbufpool *rbp;
976 struct lnet_rtrbuf *rb;
978 LASSERT(msg->msg_iov == NULL);
979 LASSERT(msg->msg_kiov == NULL);
980 LASSERT(msg->msg_niov == 0);
981 LASSERT(msg->msg_routing);
982 LASSERT(msg->msg_receiving);
983 LASSERT(!msg->msg_sending);
984 LASSERT(lpni->lpni_peer_net);
985 LASSERT(lpni->lpni_peer_net->lpn_peer);
987 lp = lpni->lpni_peer_net->lpn_peer;
989 /* non-lnet_parse callers only receive delayed messages */
990 LASSERT(!do_recv || msg->msg_rx_delayed);
992 if (!msg->msg_peerrtrcredit) {
993 /* lpni_lock protects the credit manipulation */
994 spin_lock(&lpni->lpni_lock);
995 /* lp_lock protects the lp_rtrq */
996 spin_lock(&lp->lp_lock);
998 msg->msg_peerrtrcredit = 1;
999 lpni->lpni_rtrcredits--;
1000 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1001 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1003 if (lpni->lpni_rtrcredits < 0) {
1004 /* must have checked eager_recv before here */
1005 LASSERT(msg->msg_rx_ready_delay);
1006 msg->msg_rx_delayed = 1;
1007 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1008 spin_unlock(&lp->lp_lock);
1009 spin_unlock(&lpni->lpni_lock);
1010 return LNET_CREDIT_WAIT;
1012 spin_unlock(&lp->lp_lock);
1013 spin_unlock(&lpni->lpni_lock);
1016 rbp = lnet_msg2bufpool(msg);
1018 if (!msg->msg_rtrcredit) {
1019 msg->msg_rtrcredit = 1;
1021 if (rbp->rbp_credits < rbp->rbp_mincredits)
1022 rbp->rbp_mincredits = rbp->rbp_credits;
1024 if (rbp->rbp_credits < 0) {
1025 /* must have checked eager_recv before here */
1026 LASSERT(msg->msg_rx_ready_delay);
1027 msg->msg_rx_delayed = 1;
1028 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1029 return LNET_CREDIT_WAIT;
1033 LASSERT(!list_empty(&rbp->rbp_bufs));
1034 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1035 list_del(&rb->rb_list);
1037 msg->msg_niov = rbp->rbp_npages;
1038 msg->msg_kiov = &rb->rb_kiov[0];
1040 /* unset the msg-rx_delayed flag since we're receiving the message */
1041 msg->msg_rx_delayed = 0;
1044 int cpt = msg->msg_rx_cpt;
1046 lnet_net_unlock(cpt);
1047 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1048 0, msg->msg_len, msg->msg_len);
1051 return LNET_CREDIT_OK;
1055 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1057 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1058 struct lnet_ni *txni = msg->msg_txni;
1059 struct lnet_msg *msg2;
1061 if (msg->msg_txcredit) {
1062 struct lnet_ni *ni = msg->msg_txni;
1063 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1065 /* give back NI txcredits */
1066 msg->msg_txcredit = 0;
1068 LASSERT((tq->tq_credits < 0) ==
1069 !list_empty(&tq->tq_delayed));
1072 atomic_inc(&ni->ni_tx_credits);
1073 if (tq->tq_credits <= 0) {
1074 msg2 = list_entry(tq->tq_delayed.next,
1075 struct lnet_msg, msg_list);
1076 list_del(&msg2->msg_list);
1078 LASSERT(msg2->msg_txni == ni);
1079 LASSERT(msg2->msg_tx_delayed);
1080 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1082 (void) lnet_post_send_locked(msg2, 1);
1086 if (msg->msg_peertxcredit) {
1087 /* give back peer txcredits */
1088 msg->msg_peertxcredit = 0;
1090 spin_lock(&txpeer->lpni_lock);
1091 LASSERT((txpeer->lpni_txcredits < 0) ==
1092 !list_empty(&txpeer->lpni_txq));
1094 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1095 LASSERT(txpeer->lpni_txqnob >= 0);
1097 txpeer->lpni_txcredits++;
1098 if (txpeer->lpni_txcredits <= 0) {
1101 msg2 = list_entry(txpeer->lpni_txq.next,
1102 struct lnet_msg, msg_list);
1103 list_del(&msg2->msg_list);
1104 spin_unlock(&txpeer->lpni_lock);
1106 LASSERT(msg2->msg_txpeer == txpeer);
1107 LASSERT(msg2->msg_tx_delayed);
1109 msg2_cpt = msg2->msg_tx_cpt;
1112 * The msg_cpt can be different from the msg2_cpt
1113 * so we need to make sure we lock the correct cpt
1115 * Once we call lnet_post_send_locked() it is no
1116 * longer safe to access msg2, since it could've
1117 * been freed by lnet_finalize(), but we still
1118 * need to relock the correct cpt, so we cache the
1119 * msg2_cpt for the purpose of the check that
1120 * follows the call to lnet_pose_send_locked().
1122 if (msg2_cpt != msg->msg_tx_cpt) {
1123 lnet_net_unlock(msg->msg_tx_cpt);
1124 lnet_net_lock(msg2_cpt);
1126 (void) lnet_post_send_locked(msg2, 1);
1127 if (msg2_cpt != msg->msg_tx_cpt) {
1128 lnet_net_unlock(msg2_cpt);
1129 lnet_net_lock(msg->msg_tx_cpt);
1132 spin_unlock(&txpeer->lpni_lock);
1137 msg->msg_txni = NULL;
1138 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1141 if (txpeer != NULL) {
1142 msg->msg_txpeer = NULL;
1143 lnet_peer_ni_decref_locked(txpeer);
1148 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1150 struct lnet_msg *msg;
1152 if (list_empty(&rbp->rbp_msgs))
1154 msg = list_entry(rbp->rbp_msgs.next,
1155 struct lnet_msg, msg_list);
1156 list_del(&msg->msg_list);
1158 (void)lnet_post_routed_recv_locked(msg, 1);
1162 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1164 struct lnet_msg *msg;
1165 struct lnet_msg *tmp;
1167 lnet_net_unlock(cpt);
1169 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1170 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1171 0, 0, 0, msg->msg_hdr.payload_length);
1172 list_del_init(&msg->msg_list);
1173 msg->msg_no_resend = true;
1174 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1175 lnet_finalize(msg, -ECANCELED);
1182 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1184 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1185 struct lnet_peer *lp;
1186 struct lnet_ni *rxni = msg->msg_rxni;
1187 struct lnet_msg *msg2;
1189 if (msg->msg_rtrcredit) {
1190 /* give back global router credits */
1191 struct lnet_rtrbuf *rb;
1192 struct lnet_rtrbufpool *rbp;
1194 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1195 * there until it gets one allocated, or aborts the wait
1197 LASSERT(msg->msg_kiov != NULL);
1199 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1202 msg->msg_kiov = NULL;
1203 msg->msg_rtrcredit = 0;
1205 LASSERT(rbp == lnet_msg2bufpool(msg));
1207 LASSERT((rbp->rbp_credits > 0) ==
1208 !list_empty(&rbp->rbp_bufs));
1210 /* If routing is now turned off, we just drop this buffer and
1211 * don't bother trying to return credits. */
1212 if (!the_lnet.ln_routing) {
1213 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1217 /* It is possible that a user has lowered the desired number of
1218 * buffers in this pool. Make sure we never put back
1219 * more buffers than the stated number. */
1220 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1221 /* Discard this buffer so we don't have too
1223 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1224 rbp->rbp_nbuffers--;
1226 list_add(&rb->rb_list, &rbp->rbp_bufs);
1228 if (rbp->rbp_credits <= 0)
1229 lnet_schedule_blocked_locked(rbp);
1234 if (msg->msg_peerrtrcredit) {
1236 LASSERT(rxpeerni->lpni_peer_net);
1237 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1239 lp = rxpeerni->lpni_peer_net->lpn_peer;
1241 /* give back peer router credits */
1242 msg->msg_peerrtrcredit = 0;
1244 spin_lock(&rxpeerni->lpni_lock);
1245 spin_lock(&lp->lp_lock);
1247 rxpeerni->lpni_rtrcredits++;
1249 /* drop all messages which are queued to be routed on that
1251 if (!the_lnet.ln_routing) {
1252 struct list_head drop;
1253 INIT_LIST_HEAD(&drop);
1254 list_splice_init(&lp->lp_rtrq, &drop);
1255 spin_unlock(&lp->lp_lock);
1256 spin_unlock(&rxpeerni->lpni_lock);
1257 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1258 } else if (!list_empty(&lp->lp_rtrq)) {
1261 msg2 = list_entry(lp->lp_rtrq.next,
1262 struct lnet_msg, msg_list);
1263 list_del(&msg2->msg_list);
1264 msg2_cpt = msg2->msg_rx_cpt;
1265 spin_unlock(&lp->lp_lock);
1266 spin_unlock(&rxpeerni->lpni_lock);
1268 * messages on the lp_rtrq can be from any NID in
1269 * the peer, which means they might have different
1270 * cpts. We need to make sure we lock the right
1273 if (msg2_cpt != msg->msg_rx_cpt) {
1274 lnet_net_unlock(msg->msg_rx_cpt);
1275 lnet_net_lock(msg2_cpt);
1277 (void) lnet_post_routed_recv_locked(msg2, 1);
1278 if (msg2_cpt != msg->msg_rx_cpt) {
1279 lnet_net_unlock(msg2_cpt);
1280 lnet_net_lock(msg->msg_rx_cpt);
1283 spin_unlock(&lp->lp_lock);
1284 spin_unlock(&rxpeerni->lpni_lock);
1288 msg->msg_rxni = NULL;
1289 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1291 if (rxpeerni != NULL) {
1292 msg->msg_rxpeer = NULL;
1293 lnet_peer_ni_decref_locked(rxpeerni);
1298 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1300 if (p1->lpni_txqnob < p2->lpni_txqnob)
1303 if (p1->lpni_txqnob > p2->lpni_txqnob)
1306 if (p1->lpni_txcredits > p2->lpni_txcredits)
1309 if (p1->lpni_txcredits < p2->lpni_txcredits)
1315 static struct lnet_peer_ni *
1316 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1317 struct lnet_peer *peer,
1318 struct lnet_peer_net *peer_net)
1321 * Look at the peer NIs for the destination peer that connect
1322 * to the chosen net. If a peer_ni is preferred when using the
1323 * best_ni to communicate, we use that one. If there is no
1324 * preferred peer_ni, or there are multiple preferred peer_ni,
1325 * the available transmit credits are used. If the transmit
1326 * credits are equal, we round-robin over the peer_ni.
1328 struct lnet_peer_ni *lpni = NULL;
1329 struct lnet_peer_ni *best_lpni = NULL;
1330 int best_lpni_credits = INT_MIN;
1331 bool preferred = false;
1333 int best_lpni_healthv = 0;
1336 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1338 * if the best_ni we've chosen aleady has this lpni
1339 * preferred, then let's use it
1342 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1344 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1345 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1350 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1353 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1354 libcfs_nid2str(lpni->lpni_nid),
1355 lpni->lpni_txcredits, best_lpni_credits,
1356 lpni->lpni_seq, best_lpni->lpni_seq);
1358 /* pick the healthiest peer ni */
1359 if (lpni_healthv < best_lpni_healthv) {
1361 } else if (lpni_healthv > best_lpni_healthv) {
1362 best_lpni_healthv = lpni_healthv;
1363 /* if this is a preferred peer use it */
1364 } else if (!preferred && ni_is_pref) {
1366 } else if (preferred && !ni_is_pref) {
1368 * this is not the preferred peer so let's ignore
1372 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1374 * We already have a peer that has more credits
1375 * available than this one. No need to consider
1376 * this peer further.
1379 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1381 * The best peer found so far and the current peer
1382 * have the same number of available credits let's
1383 * make sure to select between them using Round
1387 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1393 best_lpni_credits = lpni->lpni_txcredits;
1396 /* if we still can't find a peer ni then we can't reach it */
1398 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1399 LNET_NIDNET(dst_nid);
1400 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1401 libcfs_net2str(net_id));
1405 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1406 libcfs_nid2str(best_lpni->lpni_nid));
1412 * Prerequisite: the best_ni should already be set in the sd
1414 static inline struct lnet_peer_ni *
1415 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1418 struct lnet_peer_net *peer_net;
1421 * The gateway is Multi-Rail capable so now we must select the
1424 peer_net = lnet_peer_get_net_locked(peer, net_id);
1427 CERROR("gateway peer %s has no NI on net %s\n",
1428 libcfs_nid2str(peer->lp_primary_nid),
1429 libcfs_net2str(net_id));
1433 return lnet_select_peer_ni(sd->sd_best_ni, sd->sd_dst_nid,
1438 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2,
1439 struct lnet_peer_ni **best_lpni)
1441 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1442 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1443 struct lnet_peer *lp1 = r1->lr_gateway;
1444 struct lnet_peer *lp2 = r2->lr_gateway;
1445 struct lnet_peer_ni *lpni1;
1446 struct lnet_peer_ni *lpni2;
1447 struct lnet_send_data sd;
1450 sd.sd_best_ni = NULL;
1451 sd.sd_dst_nid = LNET_NID_ANY;
1452 lpni1 = lnet_find_best_lpni_on_net(&sd, lp1, r1->lr_lnet);
1453 lpni2 = lnet_find_best_lpni_on_net(&sd, lp2, r2->lr_lnet);
1454 LASSERT(lpni1 && lpni2);
1456 if (r1->lr_priority < r2->lr_priority) {
1461 if (r1->lr_priority > r2->lr_priority) {
1466 if (r1_hops < r2_hops) {
1471 if (r1_hops > r2_hops) {
1476 rc = lnet_compare_peers(lpni1, lpni2);
1480 } else if (rc == -1) {
1485 if (r1->lr_seq - r2->lr_seq <= 0) {
1494 static struct lnet_route *
1495 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1496 lnet_nid_t rtr_nid, struct lnet_route **prev_route,
1497 struct lnet_peer_ni **gwni)
1499 struct lnet_peer_ni *best_gw_ni = NULL;
1500 struct lnet_route *best_route;
1501 struct lnet_route *last_route;
1502 struct lnet_remotenet *rnet;
1503 struct lnet_peer *lp_best;
1504 struct lnet_route *route;
1505 struct lnet_peer *lp;
1508 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1509 * rtr_nid nid, otherwise find the best gateway I can use */
1511 rnet = lnet_find_rnet_locked(remote_net);
1516 best_route = last_route = NULL;
1517 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1518 lp = route->lr_gateway;
1520 if (!lnet_is_route_alive(route))
1523 if (lp_best == NULL) {
1524 best_route = last_route = route;
1528 /* no protection on below fields, but it's harmless */
1529 if (last_route->lr_seq - route->lr_seq < 0)
1532 rc = lnet_compare_routes(route, best_route, &best_gw_ni);
1540 *prev_route = last_route;
1546 static struct lnet_ni *
1547 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1548 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1551 struct lnet_ni *ni = NULL;
1552 unsigned int shortest_distance;
1557 * If there is no peer_ni that we can send to on this network,
1558 * then there is no point in looking for a new best_ni here.
1560 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1563 if (best_ni == NULL) {
1564 shortest_distance = UINT_MAX;
1565 best_credits = INT_MIN;
1568 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1569 best_ni->ni_dev_cpt);
1570 best_credits = atomic_read(&best_ni->ni_tx_credits);
1571 best_healthv = atomic_read(&best_ni->ni_healthv);
1574 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1575 unsigned int distance;
1580 ni_credits = atomic_read(&ni->ni_tx_credits);
1581 ni_healthv = atomic_read(&ni->ni_healthv);
1582 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1585 * calculate the distance from the CPT on which
1586 * the message memory is allocated to the CPT of
1587 * the NI's physical device
1589 distance = cfs_cpt_distance(lnet_cpt_table(),
1593 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1594 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1595 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1596 : "not seleced", best_credits, shortest_distance,
1597 (best_ni) ? best_ni->ni_seq : 0);
1600 * All distances smaller than the NUMA range
1601 * are treated equally.
1603 if (distance < lnet_numa_range)
1604 distance = lnet_numa_range;
1607 * Select on health, shorter distance, available
1608 * credits, then round-robin.
1612 } else if (ni_healthv < best_healthv) {
1614 } else if (ni_healthv > best_healthv) {
1615 best_healthv = ni_healthv;
1617 * If we're going to prefer this ni because it's
1618 * the healthiest, then we should set the
1619 * shortest_distance in the algorithm in case
1620 * there are multiple NIs with the same health but
1621 * different distances.
1623 if (distance < shortest_distance)
1624 shortest_distance = distance;
1625 } else if (distance > shortest_distance) {
1627 } else if (distance < shortest_distance) {
1628 shortest_distance = distance;
1629 } else if (ni_credits < best_credits) {
1631 } else if (ni_credits == best_credits) {
1632 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1636 best_credits = ni_credits;
1639 CDEBUG(D_NET, "selected best_ni %s\n",
1640 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1646 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1647 * because such traffic is required to perform discovery. We therefore
1648 * exclude all GET and PUT on that portal. We also exclude all ACK and
1649 * REPLY traffic, but that is because the portal is not tracked in the
1650 * message structure for these message types. We could restrict this
1651 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1654 lnet_msg_discovery(struct lnet_msg *msg)
1656 if (msg->msg_type == LNET_MSG_PUT) {
1657 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1659 } else if (msg->msg_type == LNET_MSG_GET) {
1660 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1666 #define SRC_SPEC 0x0001
1667 #define SRC_ANY 0x0002
1668 #define LOCAL_DST 0x0004
1669 #define REMOTE_DST 0x0008
1670 #define MR_DST 0x0010
1671 #define NMR_DST 0x0020
1672 #define SND_RESP 0x0040
1674 /* The following to defines are used for return codes */
1675 #define REPEAT_SEND 0x1000
1676 #define PASS_THROUGH 0x2000
1678 /* The different cases lnet_select pathway needs to handle */
1679 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1680 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1681 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1682 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1683 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1684 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1685 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1686 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1689 lnet_handle_lo_send(struct lnet_send_data *sd)
1691 struct lnet_msg *msg = sd->sd_msg;
1692 int cpt = sd->sd_cpt;
1694 /* No send credit hassles with LOLND */
1695 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1696 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1697 if (!msg->msg_routing)
1698 msg->msg_hdr.src_nid =
1699 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1700 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1701 lnet_msg_commit(msg, cpt);
1702 msg->msg_txni = the_lnet.ln_loni;
1704 return LNET_CREDIT_OK;
1708 lnet_handle_send(struct lnet_send_data *sd)
1710 struct lnet_ni *best_ni = sd->sd_best_ni;
1711 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1712 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1713 struct lnet_msg *msg = sd->sd_msg;
1715 __u32 send_case = sd->sd_send_case;
1717 __u32 routing = send_case & REMOTE_DST;
1718 struct lnet_rsp_tracker *rspt;
1721 * Increment sequence number of the selected peer so that we
1722 * pick the next one in Round Robin.
1724 best_lpni->lpni_seq++;
1727 * grab a reference on the peer_ni so it sticks around even if
1728 * we need to drop and relock the lnet_net_lock below.
1730 lnet_peer_ni_addref_locked(best_lpni);
1733 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1734 * message. This ensures that we get a CPT that is correct for
1735 * the NI when the NI has been restricted to a subset of all CPTs.
1736 * If the selected CPT differs from the one currently locked, we
1737 * must unlock and relock the lnet_net_lock(), and then check whether
1738 * the configuration has changed. We don't have a hold on the best_ni
1739 * yet, and it may have vanished.
1741 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1742 if (sd->sd_cpt != cpt2) {
1743 __u32 seq = lnet_get_dlc_seq_locked();
1744 lnet_net_unlock(sd->sd_cpt);
1746 lnet_net_lock(sd->sd_cpt);
1747 if (seq != lnet_get_dlc_seq_locked()) {
1748 lnet_peer_ni_decref_locked(best_lpni);
1754 * store the best_lpni in the message right away to avoid having
1755 * to do the same operation under different conditions
1757 msg->msg_txpeer = best_lpni;
1758 msg->msg_txni = best_ni;
1761 * grab a reference for the best_ni since now it's in use in this
1762 * send. The reference will be dropped in lnet_finalize()
1764 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1767 * Always set the target.nid to the best peer picked. Either the
1768 * NID will be one of the peer NIDs selected, or the same NID as
1769 * what was originally set in the target or it will be the NID of
1770 * a router if this message should be routed
1772 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1775 * lnet_msg_commit assigns the correct cpt to the message, which
1776 * is used to decrement the correct refcount on the ni when it's
1777 * time to return the credits
1779 lnet_msg_commit(msg, sd->sd_cpt);
1782 * If we are routing the message then we keep the src_nid that was
1783 * set by the originator. If we are not routing then we are the
1784 * originator and set it here.
1786 if (!msg->msg_routing)
1787 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1790 msg->msg_target_is_router = 1;
1791 msg->msg_target.pid = LNET_PID_LUSTRE;
1793 * since we're routing we want to ensure that the
1794 * msg_hdr.dest_nid is set to the final destination. When
1795 * the router receives this message it knows how to route
1798 * final_dst_lpni is set at the beginning of the
1799 * lnet_select_pathway() function and is never changed.
1800 * It's safe to use it here.
1802 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1805 * if we're not routing set the dest_nid to the best peer
1806 * ni NID that we picked earlier in the algorithm.
1808 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1812 * if we have response tracker block update it with the next hop
1816 rspt = msg->msg_md->md_rspt_ptr;
1818 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1819 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1820 libcfs_nid2str(rspt->rspt_next_hop_nid));
1824 rc = lnet_post_send_locked(msg, 0);
1827 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1828 libcfs_nid2str(msg->msg_hdr.src_nid),
1829 libcfs_nid2str(msg->msg_txni->ni_nid),
1830 libcfs_nid2str(sd->sd_src_nid),
1831 libcfs_nid2str(msg->msg_hdr.dest_nid),
1832 libcfs_nid2str(sd->sd_dst_nid),
1833 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1834 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1840 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1842 if (sd->sd_send_case & NMR_DST &&
1843 sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1844 sd->sd_msg->msg_type != LNET_MSG_ACK &&
1845 sd->sd_best_lpni->lpni_pref_nnids == 0) {
1846 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1847 libcfs_nid2str(sd->sd_best_ni->ni_nid),
1848 libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1849 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1850 sd->sd_best_ni->ni_nid);
1859 * use the source and destination NIDs as the pathway
1862 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1864 /* the destination lpni is set before we get here. */
1867 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1868 if (!sd->sd_best_ni) {
1869 CERROR("Can't send to %s: src %s is not a "
1870 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1871 libcfs_nid2str(sd->sd_src_nid));
1876 * the preferred NID will only be set for NMR peers
1878 lnet_set_non_mr_pref_nid(sd);
1880 return lnet_handle_send(sd);
1888 * Run the selection algorithm on the peer NIs unless we're sending
1889 * a response, in this case just send to the destination
1892 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1894 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1895 if (!sd->sd_best_ni) {
1896 CERROR("Can't send to %s: src %s is not a "
1897 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1898 libcfs_nid2str(sd->sd_src_nid));
1903 * only run the selection algorithm to pick the peer_ni if we're
1904 * sending a GET or a PUT. Responses are sent to the same
1905 * destination NID provided.
1907 if (!(sd->sd_send_case & SND_RESP)) {
1909 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1910 sd->sd_best_ni->ni_net->net_id);
1913 if (sd->sd_best_lpni &&
1914 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1915 return lnet_handle_lo_send(sd);
1916 else if (sd->sd_best_lpni)
1917 return lnet_handle_send(sd);
1919 CERROR("can't send to %s. no NI on %s\n",
1920 libcfs_nid2str(sd->sd_dst_nid),
1921 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1923 return -EHOSTUNREACH;
1927 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1928 struct lnet_peer *peer,
1929 struct lnet_peer_net *peer_net,
1933 struct lnet_net *local_net;
1934 struct lnet_ni *best_ni;
1936 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1941 * Iterate through the NIs in this local Net and select
1942 * the NI to send from. The selection is determined by
1943 * these 3 criterion in the following priority:
1945 * 2. NI available credits
1948 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1949 peer, peer_net, cpt);
1951 if (incr_seq && best_ni)
1958 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1959 struct lnet_msg *msg, lnet_nid_t rtr_nid,
1962 struct lnet_peer *peer;
1963 lnet_nid_t primary_nid;
1966 lnet_peer_ni_addref_locked(lpni);
1968 peer = lpni->lpni_peer_net->lpn_peer;
1970 if (lnet_peer_gw_discovery(peer)) {
1971 lnet_peer_ni_decref_locked(lpni);
1975 rc = lnet_discover_peer_locked(lpni, cpt, false);
1977 lnet_peer_ni_decref_locked(lpni);
1980 /* The peer may have changed. */
1981 peer = lpni->lpni_peer_net->lpn_peer;
1982 /* queue message and return */
1983 msg->msg_rtr_nid_param = rtr_nid;
1984 msg->msg_sending = 0;
1985 msg->msg_txpeer = NULL;
1986 spin_lock(&peer->lp_lock);
1987 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1988 spin_unlock(&peer->lp_lock);
1989 lnet_peer_ni_decref_locked(lpni);
1990 primary_nid = peer->lp_primary_nid;
1992 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
1993 msg, libcfs_nid2str(primary_nid));
1995 return LNET_DC_WAIT;
1999 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2001 struct lnet_peer_ni **gw_lpni,
2002 struct lnet_peer **gw_peer)
2004 struct lnet_peer *gw;
2005 struct lnet_route *best_route;
2006 struct lnet_route *last_route;
2007 struct lnet_peer_ni *lpni = NULL;
2008 lnet_nid_t src_nid = sd->sd_src_nid;
2010 best_route = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
2011 sd->sd_rtr_nid, &last_route,
2014 CERROR("no route to %s from %s\n",
2015 libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
2016 return -EHOSTUNREACH;
2020 CERROR("Internal Error. Route expected to %s from %s\n",
2021 libcfs_nid2str(dst_nid),
2022 libcfs_nid2str(src_nid));
2026 gw = best_route->lr_gateway;
2027 LASSERT(gw == lpni->lpni_peer_net->lpn_peer);
2030 * Discover this gateway if it hasn't already been discovered.
2031 * This means we might delay the message until discovery has
2034 if (lnet_msg_discovery(sd->sd_msg) &&
2035 !lnet_peer_is_uptodate(gw)) {
2036 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2037 return lnet_initiate_peer_discovery(lpni, sd->sd_msg,
2038 sd->sd_rtr_nid, sd->sd_cpt);
2041 if (!sd->sd_best_ni)
2042 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2043 lnet_peer_get_net_locked(gw,
2044 best_route->lr_lnet),
2048 if (!sd->sd_best_ni) {
2049 CERROR("Internal Error. Expected local ni on %s "
2050 "but non found :%s\n",
2051 libcfs_net2str(best_route->lr_lnet),
2052 libcfs_nid2str(sd->sd_src_nid));
2060 * increment the route sequence number since now we're sure we're
2063 LASSERT(best_route && last_route);
2064 best_route->lr_seq = last_route->lr_seq + 1;
2074 * Remote destination
2075 * Non-MR destination
2079 * Remote destination
2082 * The handling of these two cases is similar. Even though the destination
2083 * can be MR or non-MR, we'll deal directly with the router.
2086 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2089 struct lnet_peer_ni *gw_lpni = NULL;
2090 struct lnet_peer *gw_peer = NULL;
2093 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2094 if (!sd->sd_best_ni) {
2095 CERROR("Can't send to %s: src %s is not a "
2096 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2097 libcfs_nid2str(sd->sd_src_nid));
2101 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2106 if (sd->sd_send_case & NMR_DST)
2108 * since the final destination is non-MR let's set its preferred
2109 * NID before we send
2111 lnet_set_non_mr_pref_nid(sd);
2114 * We're going to send to the gw found so let's set its
2117 sd->sd_peer = gw_peer;
2118 sd->sd_best_lpni = gw_lpni;
2120 return lnet_handle_send(sd);
2124 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
2126 struct lnet_peer_net *peer_net = NULL;
2127 struct lnet_ni *best_ni = NULL;
2130 * The peer can have multiple interfaces, some of them can be on
2131 * the local network and others on a routed network. We should
2132 * prefer the local network. However if the local network is not
2133 * available then we need to try the routed network
2136 /* go through all the peer nets and find the best_ni */
2137 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2139 * The peer's list of nets can contain non-local nets. We
2140 * want to only examine the local ones.
2142 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2144 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2145 peer_net, md_cpt, false);
2149 /* increment sequence number so we can round robin */
2155 static struct lnet_ni *
2156 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2158 struct lnet_ni *best_ni = NULL;
2159 struct lnet_peer_net *peer_net;
2160 struct lnet_peer *peer = sd->sd_peer;
2161 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2162 struct lnet_peer_ni *lpni;
2163 int cpt = sd->sd_cpt;
2166 * We must use a consistent source address when sending to a
2167 * non-MR peer. However, a non-MR peer can have multiple NIDs
2168 * on multiple networks, and we may even need to talk to this
2169 * peer on multiple networks -- certain types of
2170 * load-balancing configuration do this.
2172 * So we need to pick the NI the peer prefers for this
2173 * particular network.
2176 /* Get the target peer_ni */
2177 peer_net = lnet_peer_get_net_locked(peer,
2178 LNET_NIDNET(best_lpni->lpni_nid));
2179 LASSERT(peer_net != NULL);
2180 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2182 if (lpni->lpni_pref_nnids == 0)
2184 LASSERT(lpni->lpni_pref_nnids == 1);
2185 best_ni = lnet_nid2ni_locked(
2186 lpni->lpni_pref.nid, cpt);
2193 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2195 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2197 struct lnet_ni *best_ni = NULL;
2198 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2201 * We must use a consistent source address when sending to a
2202 * non-MR peer. However, a non-MR peer can have multiple NIDs
2203 * on multiple networks, and we may even need to talk to this
2204 * peer on multiple networks -- certain types of
2205 * load-balancing configuration do this.
2207 * So we need to pick the NI the peer prefers for this
2208 * particular network.
2211 best_ni = lnet_find_existing_preferred_best_ni(sd);
2213 /* if best_ni is still not set just pick one */
2216 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2217 sd->sd_best_lpni->lpni_peer_net,
2218 sd->sd_md_cpt, true);
2219 /* If there is no best_ni we don't have a route */
2221 CERROR("no path to %s from net %s\n",
2222 libcfs_nid2str(best_lpni->lpni_nid),
2223 libcfs_net2str(best_lpni->lpni_net->net_id));
2224 return -EHOSTUNREACH;
2228 sd->sd_best_ni = best_ni;
2230 /* Set preferred NI if necessary. */
2231 lnet_set_non_mr_pref_nid(sd);
2238 * Source not specified
2242 * always use the same source NID for NMR peers
2243 * If we've talked to that peer before then we already have a preferred
2244 * source NI associated with it. Otherwise, we select a preferred local NI
2245 * and store it in the peer
2248 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2252 /* sd->sd_best_lpni is already set to the final destination */
2255 * At this point we should've created the peer ni and peer. If we
2256 * can't find it, then something went wrong. Instead of assert
2257 * output a relevant message and fail the send
2259 if (!sd->sd_best_lpni) {
2260 CERROR("Internal fault. Unable to send msg %s to %s. "
2262 lnet_msgtyp2str(sd->sd_msg->msg_type),
2263 libcfs_nid2str(sd->sd_dst_nid));
2267 rc = lnet_select_preferred_best_ni(sd);
2269 rc = lnet_handle_send(sd);
2275 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2278 * NOTE we've already handled the remote peer case. So we only
2279 * need to worry about the local case here.
2281 * if we're sending a response, ACK or reply, we need to send it
2282 * to the destination NID given to us. At this point we already
2283 * have the peer_ni we're suppose to send to, so just find the
2284 * best_ni on the peer net and use that. Since we're sending to an
2285 * MR peer then we can just run the selection algorithm on our
2286 * local NIs and pick the best one.
2288 if (sd->sd_send_case & SND_RESP) {
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);
2294 if (!sd->sd_best_ni) {
2296 * We're not going to deal with not able to send
2297 * a response to the provided final destination
2299 CERROR("Can't send response to %s. "
2300 "No local NI available\n",
2301 libcfs_nid2str(sd->sd_dst_nid));
2302 return -EHOSTUNREACH;
2305 return lnet_handle_send(sd);
2309 * If we get here that means we're sending a fresh request, PUT or
2310 * GET, so we need to run our standard selection algorithm.
2311 * First find the best local interface that's on any of the peer's
2314 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2316 if (sd->sd_best_ni) {
2318 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2319 sd->sd_best_ni->ni_net->net_id);
2322 * if we're successful in selecting a peer_ni on the local
2323 * network, then send to it. Otherwise fall through and
2324 * try and see if we can reach it over another routed
2327 if (sd->sd_best_lpni &&
2328 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2330 * in case we initially started with a routed
2331 * destination, let's reset to local
2333 sd->sd_send_case &= ~REMOTE_DST;
2334 sd->sd_send_case |= LOCAL_DST;
2335 return lnet_handle_lo_send(sd);
2336 } else if (sd->sd_best_lpni) {
2338 * in case we initially started with a routed
2339 * destination, let's reset to local
2341 sd->sd_send_case &= ~REMOTE_DST;
2342 sd->sd_send_case |= LOCAL_DST;
2343 return lnet_handle_send(sd);
2346 CERROR("Internal Error. Expected to have a best_lpni: "
2348 libcfs_nid2str(sd->sd_src_nid),
2349 libcfs_nid2str(sd->sd_dst_nid));
2355 * Peer doesn't have a local network. Let's see if there is
2356 * a remote network we can reach it on.
2358 return PASS_THROUGH;
2363 * Source NID not specified
2368 * Source NID not speified
2369 * Remote destination
2372 * In both of these cases if we're sending a response, ACK or REPLY, then
2373 * we need to send to the destination NID provided.
2375 * In the remote case let's deal with MR routers.
2380 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2383 struct lnet_peer *gw_peer = NULL;
2384 struct lnet_peer_ni *gw_lpni = NULL;
2387 * handle sending a response to a remote peer here so we don't
2388 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2390 if (sd->sd_send_case & REMOTE_DST &&
2391 sd->sd_send_case & SND_RESP) {
2392 struct lnet_peer_ni *gw;
2393 struct lnet_peer *gw_peer;
2395 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2398 CERROR("Can't send response to %s. "
2399 "No route available\n",
2400 libcfs_nid2str(sd->sd_dst_nid));
2401 return -EHOSTUNREACH;
2402 } else if (rc > 0) {
2406 sd->sd_best_lpni = gw;
2407 sd->sd_peer = gw_peer;
2409 return lnet_handle_send(sd);
2413 * Even though the NID for the peer might not be on a local network,
2414 * since the peer is MR there could be other interfaces on the
2415 * local network. In that case we'd still like to prefer the local
2416 * network over the routed network. If we're unable to do that
2417 * then we select the best router among the different routed networks,
2418 * and if the router is MR then we can deal with it as such.
2420 rc = lnet_handle_any_mr_dsta(sd);
2421 if (rc != PASS_THROUGH)
2425 * TODO; One possible enhancement is to run the selection
2426 * algorithm on the peer. However for remote peers the credits are
2427 * not decremented, so we'll be basically going over the peer NIs
2428 * in round robin. An MR router will run the selection algorithm
2429 * on the next-hop interfaces.
2431 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2436 sd->sd_send_case &= ~LOCAL_DST;
2437 sd->sd_send_case |= REMOTE_DST;
2439 sd->sd_peer = gw_peer;
2440 sd->sd_best_lpni = gw_lpni;
2442 return lnet_handle_send(sd);
2446 * Source not specified
2447 * Remote destination
2450 * Must send to the specified peer NID using the same source NID that
2451 * we've used before. If it's the first time to talk to that peer then
2452 * find the source NI and assign it as preferred to that peer
2455 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2458 struct lnet_peer_ni *gw_lpni = NULL;
2459 struct lnet_peer *gw_peer = NULL;
2462 * Let's set if we have a preferred NI to talk to this NMR peer
2464 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2467 * find the router and that'll find the best NI if we didn't find
2470 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2476 * set the best_ni we've chosen as the preferred one for
2479 lnet_set_non_mr_pref_nid(sd);
2481 /* we'll be sending to the gw */
2482 sd->sd_best_lpni = gw_lpni;
2483 sd->sd_peer = gw_peer;
2485 return lnet_handle_send(sd);
2489 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2492 * turn off the SND_RESP bit.
2493 * It will be checked in the case handling
2495 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2497 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2498 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2499 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2500 (send_case & MR_DST) ? "MR: " : "NMR: ",
2501 libcfs_nid2str(sd->sd_dst_nid),
2502 (send_case & LOCAL_DST) ? "local" : "routed");
2504 switch (send_case) {
2506 * For all cases where the source is specified, we should always
2507 * use the destination NID, whether it's an MR destination or not,
2508 * since we're continuing a series of related messages for the
2511 case SRC_SPEC_LOCAL_NMR_DST:
2512 return lnet_handle_spec_local_nmr_dst(sd);
2513 case SRC_SPEC_LOCAL_MR_DST:
2514 return lnet_handle_spec_local_mr_dst(sd);
2515 case SRC_SPEC_ROUTER_NMR_DST:
2516 case SRC_SPEC_ROUTER_MR_DST:
2517 return lnet_handle_spec_router_dst(sd);
2518 case SRC_ANY_LOCAL_NMR_DST:
2519 return lnet_handle_any_local_nmr_dst(sd);
2520 case SRC_ANY_LOCAL_MR_DST:
2521 case SRC_ANY_ROUTER_MR_DST:
2522 return lnet_handle_any_mr_dst(sd);
2523 case SRC_ANY_ROUTER_NMR_DST:
2524 return lnet_handle_any_router_nmr_dst(sd);
2526 CERROR("Unknown send case\n");
2532 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2533 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2535 struct lnet_peer_ni *lpni;
2536 struct lnet_peer *peer;
2537 struct lnet_send_data send_data;
2540 __u32 send_case = 0;
2542 memset(&send_data, 0, sizeof(send_data));
2545 * get an initial CPT to use for locking. The idea here is not to
2546 * serialize the calls to select_pathway, so that as many
2547 * operations can run concurrently as possible. To do that we use
2548 * the CPT where this call is being executed. Later on when we
2549 * determine the CPT to use in lnet_message_commit, we switch the
2550 * lock and check if there was any configuration change. If none,
2551 * then we proceed, if there is, then we restart the operation.
2553 cpt = lnet_net_lock_current();
2555 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2556 if (md_cpt == CFS_CPT_ANY)
2562 * If we're being asked to send to the loopback interface, there
2563 * is no need to go through any selection. We can just shortcut
2564 * the entire process and send over lolnd
2566 send_data.sd_msg = msg;
2567 send_data.sd_cpt = cpt;
2568 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2569 rc = lnet_handle_lo_send(&send_data);
2570 lnet_net_unlock(cpt);
2575 * find an existing peer_ni, or create one and mark it as having been
2576 * created due to network traffic. This call will create the
2577 * peer->peer_net->peer_ni tree.
2579 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2581 lnet_net_unlock(cpt);
2582 return PTR_ERR(lpni);
2586 * Cache the original src_nid. If we need to resend the message
2587 * then we'll need to know whether the src_nid was originally
2588 * specified for this message. If it was originally specified,
2589 * then we need to keep using the same src_nid since it's
2590 * continuing the same sequence of messages.
2592 msg->msg_src_nid_param = src_nid;
2595 * Now that we have a peer_ni, check if we want to discover
2596 * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2597 * trigger discovery.
2599 peer = lpni->lpni_peer_net->lpn_peer;
2600 if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
2601 rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2602 lnet_peer_ni_decref_locked(lpni);
2603 lnet_net_unlock(cpt);
2606 lnet_peer_ni_decref_locked(lpni);
2609 * Identify the different send cases
2611 if (src_nid == LNET_NID_ANY)
2612 send_case |= SRC_ANY;
2614 send_case |= SRC_SPEC;
2616 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2617 send_case |= LOCAL_DST;
2619 send_case |= REMOTE_DST;
2622 * if this is a non-MR peer or if we're recovering a peer ni then
2623 * let's consider this an NMR case so we can hit the destination
2626 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2627 send_case |= NMR_DST;
2629 send_case |= MR_DST;
2631 if (msg->msg_type == LNET_MSG_REPLY ||
2632 msg->msg_type == LNET_MSG_ACK)
2633 send_case |= SND_RESP;
2635 /* assign parameters to the send_data */
2636 send_data.sd_rtr_nid = rtr_nid;
2637 send_data.sd_src_nid = src_nid;
2638 send_data.sd_dst_nid = dst_nid;
2639 send_data.sd_best_lpni = lpni;
2641 * keep a pointer to the final destination in case we're going to
2642 * route, so we'll need to access it later
2644 send_data.sd_final_dst_lpni = lpni;
2645 send_data.sd_peer = peer;
2646 send_data.sd_md_cpt = md_cpt;
2647 send_data.sd_send_case = send_case;
2649 rc = lnet_handle_send_case_locked(&send_data);
2652 * Update the local cpt since send_data.sd_cpt might've been
2653 * updated as a result of calling lnet_handle_send_case_locked().
2655 cpt = send_data.sd_cpt;
2657 if (rc == REPEAT_SEND)
2660 lnet_net_unlock(cpt);
2666 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2668 lnet_nid_t dst_nid = msg->msg_target.nid;
2672 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2673 * but we might want to use pre-determined router for ACK/REPLY
2676 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2677 LASSERT(msg->msg_txpeer == NULL);
2678 LASSERT(msg->msg_txni == NULL);
2679 LASSERT(!msg->msg_sending);
2680 LASSERT(!msg->msg_target_is_router);
2681 LASSERT(!msg->msg_receiving);
2683 msg->msg_sending = 1;
2685 LASSERT(!msg->msg_tx_committed);
2687 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2691 if (rc == LNET_CREDIT_OK)
2692 lnet_ni_send(msg->msg_txni, msg);
2694 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2698 enum lnet_mt_event_type {
2699 MT_TYPE_LOCAL_NI = 0,
2703 struct lnet_mt_event_info {
2704 enum lnet_mt_event_type mt_type;
2708 /* called with res_lock held */
2710 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2712 struct lnet_rsp_tracker *rspt;
2715 * msg has a refcount on the MD so the MD is not going away.
2716 * The rspt queue for the cpt is protected by
2717 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2719 if (!md->md_rspt_ptr)
2722 rspt = md->md_rspt_ptr;
2723 md->md_rspt_ptr = NULL;
2726 LASSERT(rspt->rspt_cpt == cpt);
2729 * invalidate the handle to indicate that a response has been
2730 * received, which will then lead the monitor thread to clean up
2733 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2737 lnet_finalize_expired_responses(bool force)
2739 struct lnet_libmd *md;
2740 struct list_head local_queue;
2741 struct lnet_rsp_tracker *rspt, *tmp;
2744 if (the_lnet.ln_mt_rstq == NULL)
2747 cfs_cpt_for_each(i, lnet_cpt_table()) {
2748 INIT_LIST_HEAD(&local_queue);
2751 if (!the_lnet.ln_mt_rstq[i]) {
2755 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2758 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2760 * The rspt mdh will be invalidated when a response
2761 * is received or whenever we want to discard the
2762 * block the monitor thread will walk the queue
2763 * and clean up any rsts with an invalid mdh.
2764 * The monitor thread will walk the queue until
2765 * the first unexpired rspt block. This means that
2766 * some rspt blocks which received their
2767 * corresponding responses will linger in the
2768 * queue until they are cleaned up eventually.
2771 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2773 list_del_init(&rspt->rspt_on_list);
2774 lnet_rspt_free(rspt, i);
2778 if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
2780 struct lnet_peer_ni *lpni;
2783 md = lnet_handle2md(&rspt->rspt_mdh);
2785 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2787 list_del_init(&rspt->rspt_on_list);
2788 lnet_rspt_free(rspt, i);
2791 LASSERT(md->md_rspt_ptr == rspt);
2792 md->md_rspt_ptr = NULL;
2796 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2799 list_del_init(&rspt->rspt_on_list);
2801 nid = rspt->rspt_next_hop_nid;
2803 CNETERR("Response timed out: md = %p: nid = %s\n",
2804 md, libcfs_nid2str(nid));
2805 LNetMDUnlink(rspt->rspt_mdh);
2806 lnet_rspt_free(rspt, i);
2809 * If there is a timeout on the response
2810 * from the next hop decrement its health
2811 * value so that we don't use it
2814 lpni = lnet_find_peer_ni_locked(nid);
2816 lnet_handle_remote_failure_locked(lpni);
2817 lnet_peer_ni_decref_locked(lpni);
2827 if (!list_empty(&local_queue))
2828 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2834 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2836 struct lnet_msg *msg;
2838 while (!list_empty(resendq)) {
2839 struct lnet_peer_ni *lpni;
2841 msg = list_entry(resendq->next, struct lnet_msg,
2844 list_del_init(&msg->msg_list);
2846 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2848 lnet_net_unlock(cpt);
2849 CERROR("Expected that a peer is already created for %s\n",
2850 libcfs_nid2str(msg->msg_hdr.dest_nid));
2851 msg->msg_no_resend = true;
2852 lnet_finalize(msg, -EFAULT);
2855 struct lnet_peer *peer;
2857 lnet_nid_t src_nid = LNET_NID_ANY;
2860 * if this message is not being routed and the
2861 * peer is non-MR then we must use the same
2862 * src_nid that was used in the original send.
2863 * Otherwise if we're routing the message (IE
2864 * we're a router) then we can use any of our
2865 * local interfaces. It doesn't matter to the
2866 * final destination.
2868 peer = lpni->lpni_peer_net->lpn_peer;
2869 if (!msg->msg_routing &&
2870 !lnet_peer_is_multi_rail(peer))
2871 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
2874 * If we originally specified a src NID, then we
2875 * must attempt to reuse it in the resend as well.
2877 if (msg->msg_src_nid_param != LNET_NID_ANY)
2878 src_nid = msg->msg_src_nid_param;
2879 lnet_peer_ni_decref_locked(lpni);
2881 lnet_net_unlock(cpt);
2882 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
2883 libcfs_nid2str(src_nid),
2884 libcfs_id2str(msg->msg_target),
2885 lnet_msgtyp2str(msg->msg_type),
2887 msg->msg_retry_count);
2888 rc = lnet_send(src_nid, msg, LNET_NID_ANY);
2890 CERROR("Error sending %s to %s: %d\n",
2891 lnet_msgtyp2str(msg->msg_type),
2892 libcfs_id2str(msg->msg_target), rc);
2893 msg->msg_no_resend = true;
2894 lnet_finalize(msg, rc);
2898 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
2904 lnet_resend_pending_msgs(void)
2908 cfs_cpt_for_each(i, lnet_cpt_table()) {
2910 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
2915 /* called with cpt and ni_lock held */
2917 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
2919 struct lnet_handle_md recovery_mdh;
2921 LNetInvalidateMDHandle(&recovery_mdh);
2923 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
2925 recovery_mdh = ni->ni_ping_mdh;
2926 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2929 lnet_net_unlock(cpt);
2930 if (!LNetMDHandleIsInvalid(recovery_mdh))
2931 LNetMDUnlink(recovery_mdh);
2937 lnet_recover_local_nis(void)
2939 struct lnet_mt_event_info *ev_info;
2940 struct list_head processed_list;
2941 struct list_head local_queue;
2942 struct lnet_handle_md mdh;
2943 struct lnet_ni *tmp;
2949 INIT_LIST_HEAD(&local_queue);
2950 INIT_LIST_HEAD(&processed_list);
2953 * splice the recovery queue on a local queue. We will iterate
2954 * through the local queue and update it as needed. Once we're
2955 * done with the traversal, we'll splice the local queue back on
2956 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
2957 * will be traversed in the next iteration.
2960 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
2964 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
2966 * if an NI is being deleted or it is now healthy, there
2967 * is no need to keep it around in the recovery queue.
2968 * The monitor thread is the only thread responsible for
2969 * removing the NI from the recovery queue.
2970 * Multiple threads can be adding NIs to the recovery
2973 healthv = atomic_read(&ni->ni_healthv);
2977 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
2978 healthv == LNET_MAX_HEALTH_VALUE) {
2979 list_del_init(&ni->ni_recovery);
2980 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
2982 lnet_ni_decref_locked(ni, 0);
2988 * if the local NI failed recovery we must unlink the md.
2989 * But we want to keep the local_ni on the recovery queue
2990 * so we can continue the attempts to recover it.
2992 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
2993 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
2994 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3001 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3002 libcfs_nid2str(ni->ni_nid));
3005 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3006 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3009 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3011 CERROR("out of memory. Can't recover %s\n",
3012 libcfs_nid2str(ni->ni_nid));
3014 ni->ni_recovery_state &=
3015 ~LNET_NI_RECOVERY_PENDING;
3020 mdh = ni->ni_ping_mdh;
3022 * Invalidate the ni mdh in case it's deleted.
3023 * We'll unlink the mdh in this case below.
3025 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3029 * remove the NI from the local queue and drop the
3030 * reference count to it while we're recovering
3031 * it. The reason for that, is that the NI could
3032 * be deleted, and the way the code is structured
3033 * is if we don't drop the NI, then the deletion
3034 * code will enter a loop waiting for the
3035 * reference count to be removed while holding the
3036 * ln_mutex_lock(). When we look up the peer to
3037 * send to in lnet_select_pathway() we will try to
3038 * lock the ln_mutex_lock() as well, leading to
3039 * a deadlock. By dropping the refcount and
3040 * removing it from the list, we allow for the NI
3041 * to be removed, then we use the cached NID to
3042 * look it up again. If it's gone, then we just
3043 * continue examining the rest of the queue.
3046 list_del_init(&ni->ni_recovery);
3047 lnet_ni_decref_locked(ni, 0);
3050 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3051 ev_info->mt_nid = nid;
3052 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3053 ev_info, the_lnet.ln_mt_eqh, true);
3054 /* lookup the nid again */
3056 ni = lnet_nid2ni_locked(nid, 0);
3059 * the NI has been deleted when we dropped
3067 * Same note as in lnet_recover_peer_nis(). When
3068 * we're sending the ping, the NI is free to be
3069 * deleted or manipulated. By this point it
3070 * could've been added back on the recovery queue,
3071 * and a refcount taken on it.
3072 * So we can't just add it blindly again or we'll
3073 * corrupt the queue. We must check under lock if
3074 * it's not on any list and if not then add it
3075 * to the processed list, which will eventually be
3076 * spliced back on to the recovery queue.
3078 ni->ni_ping_mdh = mdh;
3079 if (list_empty(&ni->ni_recovery)) {
3080 list_add_tail(&ni->ni_recovery, &processed_list);
3081 lnet_ni_addref_locked(ni, 0);
3087 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3093 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3094 * reexamined in the next iteration.
3096 list_splice_init(&processed_list, &local_queue);
3098 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3102 static struct list_head **
3103 lnet_create_array_of_queues(void)
3105 struct list_head **qs;
3106 struct list_head *q;
3109 qs = cfs_percpt_alloc(lnet_cpt_table(),
3110 sizeof(struct list_head));
3112 CERROR("Failed to allocate queues\n");
3116 cfs_percpt_for_each(q, i, qs)
3123 lnet_resendqs_create(void)
3125 struct list_head **resendqs;
3126 resendqs = lnet_create_array_of_queues();
3131 lnet_net_lock(LNET_LOCK_EX);
3132 the_lnet.ln_mt_resendqs = resendqs;
3133 lnet_net_unlock(LNET_LOCK_EX);
3139 lnet_clean_local_ni_recoveryq(void)
3143 /* This is only called when the monitor thread has stopped */
3146 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3147 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3148 struct lnet_ni, ni_recovery);
3149 list_del_init(&ni->ni_recovery);
3151 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3153 lnet_ni_decref_locked(ni, 0);
3160 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3163 struct lnet_handle_md recovery_mdh;
3165 LNetInvalidateMDHandle(&recovery_mdh);
3167 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3168 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3169 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3171 spin_unlock(&lpni->lpni_lock);
3172 lnet_net_unlock(cpt);
3173 if (!LNetMDHandleIsInvalid(recovery_mdh))
3174 LNetMDUnlink(recovery_mdh);
3176 spin_lock(&lpni->lpni_lock);
3180 lnet_clean_peer_ni_recoveryq(void)
3182 struct lnet_peer_ni *lpni, *tmp;
3184 lnet_net_lock(LNET_LOCK_EX);
3186 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3188 list_del_init(&lpni->lpni_recovery);
3189 spin_lock(&lpni->lpni_lock);
3190 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3191 spin_unlock(&lpni->lpni_lock);
3192 lnet_peer_ni_decref_locked(lpni);
3195 lnet_net_unlock(LNET_LOCK_EX);
3199 lnet_clean_resendqs(void)
3201 struct lnet_msg *msg, *tmp;
3202 struct list_head msgs;
3205 INIT_LIST_HEAD(&msgs);
3207 cfs_cpt_for_each(i, lnet_cpt_table()) {
3209 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3211 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3212 list_del_init(&msg->msg_list);
3213 msg->msg_no_resend = true;
3214 lnet_finalize(msg, -ESHUTDOWN);
3218 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3222 lnet_recover_peer_nis(void)
3224 struct lnet_mt_event_info *ev_info;
3225 struct list_head processed_list;
3226 struct list_head local_queue;
3227 struct lnet_handle_md mdh;
3228 struct lnet_peer_ni *lpni;
3229 struct lnet_peer_ni *tmp;
3234 INIT_LIST_HEAD(&local_queue);
3235 INIT_LIST_HEAD(&processed_list);
3238 * Always use cpt 0 for locking across all interactions with
3239 * ln_mt_peerNIRecovq
3242 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3246 list_for_each_entry_safe(lpni, tmp, &local_queue,
3249 * The same protection strategy is used here as is in the
3250 * local recovery case.
3253 healthv = atomic_read(&lpni->lpni_healthv);
3254 spin_lock(&lpni->lpni_lock);
3255 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3256 healthv == LNET_MAX_HEALTH_VALUE) {
3257 list_del_init(&lpni->lpni_recovery);
3258 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3259 spin_unlock(&lpni->lpni_lock);
3260 lnet_peer_ni_decref_locked(lpni);
3266 * If the peer NI has failed recovery we must unlink the
3267 * md. But we want to keep the peer ni on the recovery
3268 * queue so we can try to continue recovering it
3270 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3271 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3272 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3275 spin_unlock(&lpni->lpni_lock);
3279 * NOTE: we're racing with peer deletion from user space.
3280 * It's possible that a peer is deleted after we check its
3281 * state. In this case the recovery can create a new peer
3283 spin_lock(&lpni->lpni_lock);
3284 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3285 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3286 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3287 spin_unlock(&lpni->lpni_lock);
3289 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3291 CERROR("out of memory. Can't recover %s\n",
3292 libcfs_nid2str(lpni->lpni_nid));
3293 spin_lock(&lpni->lpni_lock);
3294 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3295 spin_unlock(&lpni->lpni_lock);
3299 /* look at the comments in lnet_recover_local_nis() */
3300 mdh = lpni->lpni_recovery_ping_mdh;
3301 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3302 nid = lpni->lpni_nid;
3304 list_del_init(&lpni->lpni_recovery);
3305 lnet_peer_ni_decref_locked(lpni);
3308 ev_info->mt_type = MT_TYPE_PEER_NI;
3309 ev_info->mt_nid = nid;
3310 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3311 ev_info, the_lnet.ln_mt_eqh, true);
3314 * lnet_find_peer_ni_locked() grabs a refcount for
3315 * us. No need to take it explicitly.
3317 lpni = lnet_find_peer_ni_locked(nid);
3324 lpni->lpni_recovery_ping_mdh = mdh;
3326 * While we're unlocked the lpni could've been
3327 * readded on the recovery queue. In this case we
3328 * don't need to add it to the local queue, since
3329 * it's already on there and the thread that added
3330 * it would've incremented the refcount on the
3331 * peer, which means we need to decref the refcount
3332 * that was implicitly grabbed by find_peer_ni_locked.
3333 * Otherwise, if the lpni is still not on
3334 * the recovery queue, then we'll add it to the
3337 if (list_empty(&lpni->lpni_recovery))
3338 list_add_tail(&lpni->lpni_recovery, &processed_list);
3340 lnet_peer_ni_decref_locked(lpni);
3343 spin_lock(&lpni->lpni_lock);
3345 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3347 spin_unlock(&lpni->lpni_lock);
3350 list_splice_init(&processed_list, &local_queue);
3352 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3357 lnet_monitor_thread(void *arg)
3359 time64_t recovery_timeout = 0;
3360 time64_t rsp_timeout = 0;
3365 * The monitor thread takes care of the following:
3366 * 1. Checks the aliveness of routers
3367 * 2. Checks if there are messages on the resend queue to resend
3369 * 3. Check if there are any NIs on the local recovery queue and
3371 * 4. Checks if there are any NIs on the remote recovery queue
3374 cfs_block_allsigs();
3376 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3377 now = ktime_get_real_seconds();
3379 if (lnet_router_checker_active())
3380 lnet_check_routers();
3382 lnet_resend_pending_msgs();
3384 if (now >= rsp_timeout) {
3385 lnet_finalize_expired_responses(false);
3386 rsp_timeout = now + (lnet_transaction_timeout / 2);
3389 if (now >= recovery_timeout) {
3390 lnet_recover_local_nis();
3391 lnet_recover_peer_nis();
3392 recovery_timeout = now + lnet_recovery_interval;
3396 * TODO do we need to check if we should sleep without
3397 * timeout? Technically, an active system will always
3398 * have messages in flight so this check will always
3399 * evaluate to false. And on an idle system do we care
3400 * if we wake up every 1 second? Although, we've seen
3401 * cases where we get a complaint that an idle thread
3402 * is waking up unnecessarily.
3404 interval = min(lnet_recovery_interval,
3405 lnet_transaction_timeout / 2);
3406 wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
3408 cfs_time_seconds(interval));
3412 lnet_net_lock(LNET_LOCK_EX);
3413 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3414 lnet_net_unlock(LNET_LOCK_EX);
3416 /* signal that the monitor thread is exiting */
3417 up(&the_lnet.ln_mt_signal);
3425 * Returns == 0 if success
3426 * Returns > 0 if LNetMDBind or prior fails
3427 * Returns < 0 if LNetGet fails
3430 lnet_send_ping(lnet_nid_t dest_nid,
3431 struct lnet_handle_md *mdh, int nnis,
3432 void *user_data, struct lnet_handle_eq eqh, bool recovery)
3434 struct lnet_md md = { NULL };
3435 struct lnet_process_id id;
3436 struct lnet_ping_buffer *pbuf;
3439 if (dest_nid == LNET_NID_ANY) {
3444 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3450 /* initialize md content */
3451 md.start = &pbuf->pb_info;
3452 md.length = LNET_PING_INFO_SIZE(nnis);
3453 md.threshold = 2; /* GET/REPLY */
3455 md.options = LNET_MD_TRUNCATE;
3456 md.user_ptr = user_data;
3459 rc = LNetMDBind(md, LNET_UNLINK, mdh);
3461 lnet_ping_buffer_decref(pbuf);
3462 CERROR("Can't bind MD: %d\n", rc);
3463 rc = -rc; /* change the rc to positive */
3466 id.pid = LNET_PID_LUSTRE;
3469 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3470 LNET_RESERVED_PORTAL,
3471 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3474 goto fail_unlink_md;
3480 LNetInvalidateMDHandle(mdh);
3486 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3487 int status, bool unlink_event)
3489 lnet_nid_t nid = ev_info->mt_nid;
3491 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3495 ni = lnet_nid2ni_locked(nid, 0);
3501 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3503 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3508 CERROR("local NI (%s) recovery failed with %d\n",
3509 libcfs_nid2str(nid), status);
3513 * need to increment healthv for the ni here, because in
3514 * the lnet_finalize() path we don't have access to this
3515 * NI. And in order to get access to it, we'll need to
3516 * carry forward too much information.
3517 * In the peer case, it'll naturally be incremented
3520 lnet_inc_healthv(&ni->ni_healthv);
3522 struct lnet_peer_ni *lpni;
3525 cpt = lnet_net_lock_current();
3526 lpni = lnet_find_peer_ni_locked(nid);
3528 lnet_net_unlock(cpt);
3531 spin_lock(&lpni->lpni_lock);
3532 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3534 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3535 spin_unlock(&lpni->lpni_lock);
3536 lnet_peer_ni_decref_locked(lpni);
3537 lnet_net_unlock(cpt);
3540 CERROR("peer NI (%s) recovery failed with %d\n",
3541 libcfs_nid2str(nid), status);
3546 lnet_mt_event_handler(struct lnet_event *event)
3548 struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3549 struct lnet_ping_buffer *pbuf;
3551 /* TODO: remove assert */
3552 LASSERT(event->type == LNET_EVENT_REPLY ||
3553 event->type == LNET_EVENT_SEND ||
3554 event->type == LNET_EVENT_UNLINK);
3556 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3559 switch (event->type) {
3560 case LNET_EVENT_UNLINK:
3561 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3562 libcfs_nid2str(ev_info->mt_nid));
3563 case LNET_EVENT_REPLY:
3564 lnet_handle_recovery_reply(ev_info, event->status,
3565 event->type == LNET_EVENT_UNLINK);
3567 case LNET_EVENT_SEND:
3568 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3569 libcfs_nid2str(ev_info->mt_nid),
3570 (event->status) ? "unsuccessfully" :
3571 "successfully", event->status);
3574 CERROR("Unexpected event: %d\n", event->type);
3577 if (event->unlinked) {
3578 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3579 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3580 lnet_ping_buffer_decref(pbuf);
3585 lnet_rsp_tracker_create(void)
3587 struct list_head **rstqs;
3588 rstqs = lnet_create_array_of_queues();
3593 the_lnet.ln_mt_rstq = rstqs;
3599 lnet_rsp_tracker_clean(void)
3601 lnet_finalize_expired_responses(true);
3603 cfs_percpt_free(the_lnet.ln_mt_rstq);
3604 the_lnet.ln_mt_rstq = NULL;
3607 int lnet_monitor_thr_start(void)
3610 struct task_struct *task;
3612 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3615 rc = lnet_resendqs_create();
3619 rc = lnet_rsp_tracker_create();
3623 sema_init(&the_lnet.ln_mt_signal, 0);
3625 lnet_net_lock(LNET_LOCK_EX);
3626 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3627 lnet_net_unlock(LNET_LOCK_EX);
3628 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3631 CERROR("Can't start monitor thread: %d\n", rc);
3638 lnet_net_lock(LNET_LOCK_EX);
3639 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3640 lnet_net_unlock(LNET_LOCK_EX);
3641 /* block until event callback signals exit */
3642 down(&the_lnet.ln_mt_signal);
3644 lnet_net_lock(LNET_LOCK_EX);
3645 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3646 lnet_net_unlock(LNET_LOCK_EX);
3647 lnet_rsp_tracker_clean();
3648 lnet_clean_local_ni_recoveryq();
3649 lnet_clean_peer_ni_recoveryq();
3650 lnet_clean_resendqs();
3651 LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3654 lnet_rsp_tracker_clean();
3655 lnet_clean_local_ni_recoveryq();
3656 lnet_clean_peer_ni_recoveryq();
3657 lnet_clean_resendqs();
3661 void lnet_monitor_thr_stop(void)
3663 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3666 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3667 lnet_net_lock(LNET_LOCK_EX);
3668 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3669 lnet_net_unlock(LNET_LOCK_EX);
3671 /* tell the monitor thread that we're shutting down */
3672 wake_up(&the_lnet.ln_mt_waitq);
3674 /* block until monitor thread signals that it's done */
3675 down(&the_lnet.ln_mt_signal);
3676 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3678 /* perform cleanup tasks */
3679 lnet_rsp_tracker_clean();
3680 lnet_clean_local_ni_recoveryq();
3681 lnet_clean_peer_ni_recoveryq();
3682 lnet_clean_resendqs();
3688 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3692 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3693 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3694 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3695 lnet_net_unlock(cpt);
3697 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3701 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3703 struct lnet_hdr *hdr = &msg->msg_hdr;
3705 if (msg->msg_wanted != 0)
3706 lnet_setpayloadbuffer(msg);
3708 lnet_build_msg_event(msg, LNET_EVENT_PUT);
3710 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
3711 * it back into the ACK during lnet_finalize() */
3712 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3713 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3715 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3716 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3720 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3722 struct lnet_hdr *hdr = &msg->msg_hdr;
3723 struct lnet_match_info info;
3727 /* Convert put fields to host byte order */
3728 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3729 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
3730 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
3732 /* Primary peer NID. */
3733 info.mi_id.nid = msg->msg_initiator;
3734 info.mi_id.pid = hdr->src_pid;
3735 info.mi_opc = LNET_MD_OP_PUT;
3736 info.mi_portal = hdr->msg.put.ptl_index;
3737 info.mi_rlength = hdr->payload_length;
3738 info.mi_roffset = hdr->msg.put.offset;
3739 info.mi_mbits = hdr->msg.put.match_bits;
3740 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3742 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3743 ready_delay = msg->msg_rx_ready_delay;
3746 rc = lnet_ptl_match_md(&info, msg);
3751 case LNET_MATCHMD_OK:
3752 lnet_recv_put(ni, msg);
3755 case LNET_MATCHMD_NONE:
3757 /* no eager_recv or has already called it, should
3758 * have been attached on delayed list */
3761 rc = lnet_ni_eager_recv(ni, msg);
3768 case LNET_MATCHMD_DROP:
3769 CNETERR("Dropping PUT from %s portal %d match %llu"
3770 " offset %d length %d: %d\n",
3771 libcfs_id2str(info.mi_id), info.mi_portal,
3772 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3774 return -ENOENT; /* -ve: OK but no match */
3779 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3781 struct lnet_match_info info;
3782 struct lnet_hdr *hdr = &msg->msg_hdr;
3783 struct lnet_process_id source_id;
3784 struct lnet_handle_wire reply_wmd;
3787 /* Convert get fields to host byte order */
3788 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
3789 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
3790 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
3791 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
3793 source_id.nid = hdr->src_nid;
3794 source_id.pid = hdr->src_pid;
3795 /* Primary peer NID */
3796 info.mi_id.nid = msg->msg_initiator;
3797 info.mi_id.pid = hdr->src_pid;
3798 info.mi_opc = LNET_MD_OP_GET;
3799 info.mi_portal = hdr->msg.get.ptl_index;
3800 info.mi_rlength = hdr->msg.get.sink_length;
3801 info.mi_roffset = hdr->msg.get.src_offset;
3802 info.mi_mbits = hdr->msg.get.match_bits;
3803 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3805 rc = lnet_ptl_match_md(&info, msg);
3806 if (rc == LNET_MATCHMD_DROP) {
3807 CNETERR("Dropping GET from %s portal %d match %llu"
3808 " offset %d length %d\n",
3809 libcfs_id2str(info.mi_id), info.mi_portal,
3810 info.mi_mbits, info.mi_roffset, info.mi_rlength);
3811 return -ENOENT; /* -ve: OK but no match */
3814 LASSERT(rc == LNET_MATCHMD_OK);
3816 lnet_build_msg_event(msg, LNET_EVENT_GET);
3818 reply_wmd = hdr->msg.get.return_wmd;
3820 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3821 msg->msg_offset, msg->msg_wanted);
3823 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3826 /* The LND completes the REPLY from her recv procedure */
3827 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3828 msg->msg_offset, msg->msg_len, msg->msg_len);
3832 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3833 msg->msg_receiving = 0;
3835 rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3837 /* didn't get as far as lnet_ni_send() */
3838 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3839 libcfs_nid2str(ni->ni_nid),
3840 libcfs_id2str(info.mi_id), rc);
3842 lnet_finalize(msg, rc);
3849 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3851 void *private = msg->msg_private;
3852 struct lnet_hdr *hdr = &msg->msg_hdr;
3853 struct lnet_process_id src = {0};
3854 struct lnet_libmd *md;
3859 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3862 src.nid = hdr->src_nid;
3863 src.pid = hdr->src_pid;
3865 /* NB handles only looked up by creator (no flips) */
3866 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
3867 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3868 CNETERR("%s: Dropping REPLY from %s for %s "
3870 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3871 (md == NULL) ? "invalid" : "inactive",
3872 hdr->msg.reply.dst_wmd.wh_interface_cookie,
3873 hdr->msg.reply.dst_wmd.wh_object_cookie);
3874 if (md != NULL && md->md_me != NULL)
3875 CERROR("REPLY MD also attached to portal %d\n",
3876 md->md_me->me_portal);
3878 lnet_res_unlock(cpt);
3879 return -ENOENT; /* -ve: OK but no match */
3882 LASSERT(md->md_offset == 0);
3884 rlength = hdr->payload_length;
3885 mlength = MIN(rlength, (int)md->md_length);
3887 if (mlength < rlength &&
3888 (md->md_options & LNET_MD_TRUNCATE) == 0) {
3889 CNETERR("%s: Dropping REPLY from %s length %d "
3890 "for MD %#llx would overflow (%d)\n",
3891 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3892 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
3894 lnet_res_unlock(cpt);
3895 return -ENOENT; /* -ve: OK but no match */
3898 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
3899 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3900 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
3902 lnet_msg_attach_md(msg, md, 0, mlength);
3905 lnet_setpayloadbuffer(msg);
3907 lnet_res_unlock(cpt);
3909 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3911 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
3916 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
3918 struct lnet_hdr *hdr = &msg->msg_hdr;
3919 struct lnet_process_id src = {0};
3920 struct lnet_libmd *md;
3923 src.nid = hdr->src_nid;
3924 src.pid = hdr->src_pid;
3926 /* Convert ack fields to host byte order */
3927 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
3928 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
3930 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
3933 /* NB handles only looked up by creator (no flips) */
3934 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
3935 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3936 /* Don't moan; this is expected */
3938 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
3939 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3940 (md == NULL) ? "invalid" : "inactive",
3941 hdr->msg.ack.dst_wmd.wh_interface_cookie,
3942 hdr->msg.ack.dst_wmd.wh_object_cookie);
3943 if (md != NULL && md->md_me != NULL)
3944 CERROR("Source MD also attached to portal %d\n",
3945 md->md_me->me_portal);
3947 lnet_res_unlock(cpt);
3948 return -ENOENT; /* -ve! */
3951 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
3952 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3953 hdr->msg.ack.dst_wmd.wh_object_cookie);
3955 lnet_msg_attach_md(msg, md, 0, 0);
3957 lnet_res_unlock(cpt);
3959 lnet_build_msg_event(msg, LNET_EVENT_ACK);
3961 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
3966 * \retval LNET_CREDIT_OK If \a msg is forwarded
3967 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
3968 * \retval -ve error code
3971 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
3975 if (!the_lnet.ln_routing)
3978 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
3979 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
3980 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
3981 msg->msg_rx_ready_delay = 1;
3983 lnet_net_unlock(msg->msg_rx_cpt);
3984 rc = lnet_ni_eager_recv(ni, msg);
3985 lnet_net_lock(msg->msg_rx_cpt);
3990 rc = lnet_post_routed_recv_locked(msg, 0);
3995 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
3999 switch (msg->msg_type) {
4001 rc = lnet_parse_ack(ni, msg);
4004 rc = lnet_parse_put(ni, msg);
4007 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4009 case LNET_MSG_REPLY:
4010 rc = lnet_parse_reply(ni, msg);
4012 default: /* prevent an unused label if !kernel */
4017 LASSERT(rc == 0 || rc == -ENOENT);
4022 lnet_msgtyp2str (int type)
4031 case LNET_MSG_REPLY:
4033 case LNET_MSG_HELLO:
4036 return ("<UNKNOWN>");
4041 lnet_print_hdr(struct lnet_hdr *hdr)
4043 struct lnet_process_id src = {
4044 .nid = hdr->src_nid,
4045 .pid = hdr->src_pid,
4047 struct lnet_process_id dst = {
4048 .nid = hdr->dest_nid,
4049 .pid = hdr->dest_pid,
4051 char *type_str = lnet_msgtyp2str(hdr->type);
4053 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4054 CWARN(" From %s\n", libcfs_id2str(src));
4055 CWARN(" To %s\n", libcfs_id2str(dst));
4057 switch (hdr->type) {
4062 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
4063 "match bits %llu\n",
4064 hdr->msg.put.ptl_index,
4065 hdr->msg.put.ack_wmd.wh_interface_cookie,
4066 hdr->msg.put.ack_wmd.wh_object_cookie,
4067 hdr->msg.put.match_bits);
4068 CWARN(" Length %d, offset %d, hdr data %#llx\n",
4069 hdr->payload_length, hdr->msg.put.offset,
4070 hdr->msg.put.hdr_data);
4074 CWARN(" Ptl index %d, return md %#llx.%#llx, "
4075 "match bits %llu\n", hdr->msg.get.ptl_index,
4076 hdr->msg.get.return_wmd.wh_interface_cookie,
4077 hdr->msg.get.return_wmd.wh_object_cookie,
4078 hdr->msg.get.match_bits);
4079 CWARN(" Length %d, src offset %d\n",
4080 hdr->msg.get.sink_length,
4081 hdr->msg.get.src_offset);
4085 CWARN(" dst md %#llx.%#llx, "
4086 "manipulated length %d\n",
4087 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4088 hdr->msg.ack.dst_wmd.wh_object_cookie,
4089 hdr->msg.ack.mlength);
4092 case LNET_MSG_REPLY:
4093 CWARN(" dst md %#llx.%#llx, "
4095 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4096 hdr->msg.reply.dst_wmd.wh_object_cookie,
4097 hdr->payload_length);
4103 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4104 void *private, int rdma_req)
4109 struct lnet_msg *msg;
4110 lnet_pid_t dest_pid;
4111 lnet_nid_t dest_nid;
4113 struct lnet_peer_ni *lpni;
4114 __u32 payload_length;
4117 LASSERT (!in_interrupt ());
4119 type = le32_to_cpu(hdr->type);
4120 src_nid = le64_to_cpu(hdr->src_nid);
4121 dest_nid = le64_to_cpu(hdr->dest_nid);
4122 dest_pid = le32_to_cpu(hdr->dest_pid);
4123 payload_length = le32_to_cpu(hdr->payload_length);
4125 for_me = (ni->ni_nid == dest_nid);
4126 cpt = lnet_cpt_of_nid(from_nid, ni);
4128 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4129 libcfs_nid2str(dest_nid),
4130 libcfs_nid2str(ni->ni_nid),
4131 libcfs_nid2str(src_nid),
4132 lnet_msgtyp2str(type),
4133 (for_me) ? "for me" : "routed");
4138 if (payload_length > 0) {
4139 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4140 libcfs_nid2str(from_nid),
4141 libcfs_nid2str(src_nid),
4142 lnet_msgtyp2str(type), payload_length);
4148 case LNET_MSG_REPLY:
4149 if (payload_length >
4150 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4151 CERROR("%s, src %s: bad %s payload %d "
4152 "(%d max expected)\n",
4153 libcfs_nid2str(from_nid),
4154 libcfs_nid2str(src_nid),
4155 lnet_msgtyp2str(type),
4157 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4163 CERROR("%s, src %s: Bad message type 0x%x\n",
4164 libcfs_nid2str(from_nid),
4165 libcfs_nid2str(src_nid), type);
4169 if (the_lnet.ln_routing &&
4170 ni->ni_last_alive != ktime_get_real_seconds()) {
4171 /* NB: so far here is the only place to set NI status to "up */
4173 ni->ni_last_alive = ktime_get_real_seconds();
4174 if (ni->ni_status != NULL &&
4175 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
4176 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4180 /* Regard a bad destination NID as a protocol error. Senders should
4181 * know what they're doing; if they don't they're misconfigured, buggy
4182 * or malicious so we chop them off at the knees :) */
4185 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4186 /* should have gone direct */
4187 CERROR("%s, src %s: Bad dest nid %s "
4188 "(should have been sent direct)\n",
4189 libcfs_nid2str(from_nid),
4190 libcfs_nid2str(src_nid),
4191 libcfs_nid2str(dest_nid));
4195 if (lnet_islocalnid(dest_nid)) {
4196 /* dest is another local NI; sender should have used
4197 * this node's NID on its own network */
4198 CERROR("%s, src %s: Bad dest nid %s "
4199 "(it's my nid but on a different network)\n",
4200 libcfs_nid2str(from_nid),
4201 libcfs_nid2str(src_nid),
4202 libcfs_nid2str(dest_nid));
4206 if (rdma_req && type == LNET_MSG_GET) {
4207 CERROR("%s, src %s: Bad optimized GET for %s "
4208 "(final destination must be me)\n",
4209 libcfs_nid2str(from_nid),
4210 libcfs_nid2str(src_nid),
4211 libcfs_nid2str(dest_nid));
4215 if (!the_lnet.ln_routing) {
4216 CERROR("%s, src %s: Dropping message for %s "
4217 "(routing not enabled)\n",
4218 libcfs_nid2str(from_nid),
4219 libcfs_nid2str(src_nid),
4220 libcfs_nid2str(dest_nid));
4225 /* Message looks OK; we're not going to return an error, so we MUST
4226 * call back lnd_recv() come what may... */
4228 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4229 fail_peer(src_nid, 0)) { /* shall we now? */
4230 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4231 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4232 lnet_msgtyp2str(type));
4236 if (!list_empty(&the_lnet.ln_drop_rules) &&
4237 lnet_drop_rule_match(hdr, NULL)) {
4238 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4239 "silent message loss\n",
4240 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4241 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4245 if (lnet_drop_asym_route && for_me &&
4246 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4247 struct lnet_net *net;
4248 struct lnet_remotenet *rnet;
4251 /* we are dealing with a routed message,
4252 * so see if route to reach src_nid goes through from_nid
4255 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4257 lnet_net_unlock(cpt);
4258 CERROR("net %s not found\n",
4259 libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4263 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4265 struct lnet_peer *gw = NULL;
4266 struct lnet_peer_ni *lpni = NULL;
4267 struct lnet_route *route;
4269 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4271 gw = route->lr_gateway;
4272 if (route->lr_lnet != net->net_id)
4275 * if the nid is one of the gateway's NIDs
4276 * then this is a valid gateway
4278 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4279 NULL, lpni)) != NULL) {
4280 if (lpni->lpni_nid == from_nid) {
4287 lnet_net_unlock(cpt);
4289 /* we would not use from_nid to route a message to
4291 * => asymmetric routing detected but forbidden
4293 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4294 libcfs_nid2str(from_nid),
4295 libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4300 msg = lnet_msg_alloc();
4302 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4303 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4304 lnet_msgtyp2str(type));
4308 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4309 * pointers NULL etc */
4311 msg->msg_type = type;
4312 msg->msg_private = private;
4313 msg->msg_receiving = 1;
4314 msg->msg_rdma_get = rdma_req;
4315 msg->msg_len = msg->msg_wanted = payload_length;
4316 msg->msg_offset = 0;
4317 msg->msg_hdr = *hdr;
4318 /* for building message event */
4319 msg->msg_from = from_nid;
4321 msg->msg_target.pid = dest_pid;
4322 msg->msg_target.nid = dest_nid;
4323 msg->msg_routing = 1;
4326 /* convert common msg->hdr fields to host byteorder */
4327 msg->msg_hdr.type = type;
4328 msg->msg_hdr.src_nid = src_nid;
4329 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4330 msg->msg_hdr.dest_nid = dest_nid;
4331 msg->msg_hdr.dest_pid = dest_pid;
4332 msg->msg_hdr.payload_length = payload_length;
4336 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4338 lnet_net_unlock(cpt);
4339 CERROR("%s, src %s: Dropping %s "
4340 "(error %ld looking up sender)\n",
4341 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4342 lnet_msgtyp2str(type), PTR_ERR(lpni));
4344 if (rc == -ESHUTDOWN)
4345 /* We are shutting down. Don't do anything more */
4349 msg->msg_rxpeer = lpni;
4351 lnet_ni_addref_locked(ni, cpt);
4352 /* Multi-Rail: Primary NID of source. */
4353 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4356 * mark the status of this lpni as UP since we received a message
4357 * from it. The ping response reports back the ns_status which is
4358 * marked on the remote as up or down and we cache it here.
4360 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4362 lnet_msg_commit(msg, cpt);
4364 /* message delay simulation */
4365 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4366 lnet_delay_rule_match_locked(hdr, msg))) {
4367 lnet_net_unlock(cpt);
4372 rc = lnet_parse_forward_locked(ni, msg);
4373 lnet_net_unlock(cpt);
4378 if (rc == LNET_CREDIT_OK) {
4379 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4380 0, payload_length, payload_length);
4385 lnet_net_unlock(cpt);
4387 rc = lnet_parse_local(ni, msg);
4393 LASSERT(msg->msg_md == NULL);
4394 lnet_finalize(msg, rc);
4397 lnet_drop_message(ni, cpt, private, payload_length, type);
4400 EXPORT_SYMBOL(lnet_parse);
4403 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4405 while (!list_empty(head)) {
4406 struct lnet_process_id id = {0};
4407 struct lnet_msg *msg;
4409 msg = list_entry(head->next, struct lnet_msg, msg_list);
4410 list_del(&msg->msg_list);
4412 id.nid = msg->msg_hdr.src_nid;
4413 id.pid = msg->msg_hdr.src_pid;
4415 LASSERT(msg->msg_md == NULL);
4416 LASSERT(msg->msg_rx_delayed);
4417 LASSERT(msg->msg_rxpeer != NULL);
4418 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4420 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4421 " offset %d length %d: %s\n",
4423 msg->msg_hdr.msg.put.ptl_index,
4424 msg->msg_hdr.msg.put.match_bits,
4425 msg->msg_hdr.msg.put.offset,
4426 msg->msg_hdr.payload_length, reason);
4428 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4429 * called lnet_drop_message(), so I just hang onto msg as well
4430 * until that's done */
4432 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4433 msg->msg_private, msg->msg_len,
4436 msg->msg_no_resend = true;
4438 * NB: message will not generate event because w/o attached MD,
4439 * but we still should give error code so lnet_msg_decommit()
4440 * can skip counters operations and other checks.
4442 lnet_finalize(msg, -ENOENT);
4447 lnet_recv_delayed_msg_list(struct list_head *head)
4449 while (!list_empty(head)) {
4450 struct lnet_msg *msg;
4451 struct lnet_process_id id;
4453 msg = list_entry(head->next, struct lnet_msg, msg_list);
4454 list_del(&msg->msg_list);
4456 /* md won't disappear under me, since each msg
4457 * holds a ref on it */
4459 id.nid = msg->msg_hdr.src_nid;
4460 id.pid = msg->msg_hdr.src_pid;
4462 LASSERT(msg->msg_rx_delayed);
4463 LASSERT(msg->msg_md != NULL);
4464 LASSERT(msg->msg_rxpeer != NULL);
4465 LASSERT(msg->msg_rxni != NULL);
4466 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4468 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4469 "match %llu offset %d length %d.\n",
4470 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4471 msg->msg_hdr.msg.put.match_bits,
4472 msg->msg_hdr.msg.put.offset,
4473 msg->msg_hdr.payload_length);
4475 lnet_recv_put(msg->msg_rxni, msg);
4480 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4481 struct lnet_libmd *md, struct lnet_handle_md mdh)
4484 bool new_entry = true;
4485 struct lnet_rsp_tracker *local_rspt;
4488 * MD has a refcount taken by message so it's not going away.
4489 * The MD however can be looked up. We need to secure the access
4490 * to the md_rspt_ptr by taking the res_lock.
4491 * The rspt can be accessed without protection up to when it gets
4492 * added to the list.
4496 local_rspt = md->md_rspt_ptr;
4497 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4498 if (local_rspt != NULL) {
4500 * we already have an rspt attached to the md, so we'll
4501 * update the deadline on that one.
4503 LIBCFS_FREE(rspt, sizeof(*rspt));
4507 rspt->rspt_mdh = mdh;
4508 rspt->rspt_cpt = cpt;
4509 /* store the rspt so we can access it when we get the REPLY */
4510 md->md_rspt_ptr = rspt;
4513 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4516 * add to the list of tracked responses. It's added to tail of the
4517 * list in order to expire all the older entries first.
4520 if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
4521 list_del_init(&local_rspt->rspt_on_list);
4522 list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4523 lnet_net_unlock(cpt);
4524 lnet_res_unlock(cpt);
4528 * Initiate an asynchronous PUT operation.
4530 * There are several events associated with a PUT: completion of the send on
4531 * the initiator node (LNET_EVENT_SEND), and when the send completes
4532 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4533 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4534 * used at the target node to indicate the completion of incoming data
4537 * The local events will be logged in the EQ associated with the MD pointed to
4538 * by \a mdh handle. Using a MD without an associated EQ results in these
4539 * events being discarded. In this case, the caller must have another
4540 * mechanism (e.g., a higher level protocol) for determining when it is safe
4541 * to modify the memory region associated with the MD.
4543 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4544 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4546 * \param self Indicates the NID of a local interface through which to send
4547 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4548 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4549 * must be "free floating" (See LNetMDBind()).
4550 * \param ack Controls whether an acknowledgment is requested.
4551 * Acknowledgments are only sent when they are requested by the initiating
4552 * process and the target MD enables them.
4553 * \param target A process identifier for the target process.
4554 * \param portal The index in the \a target's portal table.
4555 * \param match_bits The match bits to use for MD selection at the target
4557 * \param offset The offset into the target MD (only used when the target
4558 * MD has the LNET_MD_MANAGE_REMOTE option set).
4559 * \param hdr_data 64 bits of user data that can be included in the message
4560 * header. This data is written to an event queue entry at the target if an
4561 * EQ is present on the matching MD.
4563 * \retval 0 Success, and only in this case events will be generated
4564 * and logged to EQ (if it exists).
4565 * \retval -EIO Simulated failure.
4566 * \retval -ENOMEM Memory allocation failure.
4567 * \retval -ENOENT Invalid MD object.
4569 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4572 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4573 struct lnet_process_id target, unsigned int portal,
4574 __u64 match_bits, unsigned int offset,
4577 struct lnet_msg *msg;
4578 struct lnet_libmd *md;
4581 struct lnet_rsp_tracker *rspt = NULL;
4583 LASSERT(the_lnet.ln_refcount > 0);
4585 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4586 fail_peer(target.nid, 1)) { /* shall we now? */
4587 CERROR("Dropping PUT to %s: simulated failure\n",
4588 libcfs_id2str(target));
4592 msg = lnet_msg_alloc();
4594 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4595 libcfs_id2str(target));
4598 msg->msg_vmflush = !!memory_pressure_get();
4600 cpt = lnet_cpt_of_cookie(mdh.cookie);
4602 if (ack == LNET_ACK_REQ) {
4603 rspt = lnet_rspt_alloc(cpt);
4605 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4606 libcfs_id2str(target));
4609 INIT_LIST_HEAD(&rspt->rspt_on_list);
4614 md = lnet_handle2md(&mdh);
4615 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4616 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4617 match_bits, portal, libcfs_id2str(target),
4618 md == NULL ? -1 : md->md_threshold);
4619 if (md != NULL && md->md_me != NULL)
4620 CERROR("Source MD also attached to portal %d\n",
4621 md->md_me->me_portal);
4622 lnet_res_unlock(cpt);
4624 LIBCFS_FREE(rspt, sizeof(*rspt));
4629 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4631 lnet_msg_attach_md(msg, md, 0, 0);
4633 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4635 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4636 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4637 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4638 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4640 /* NB handles only looked up by creator (no flips) */
4641 if (ack == LNET_ACK_REQ) {
4642 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4643 the_lnet.ln_interface_cookie;
4644 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4645 md->md_lh.lh_cookie;
4647 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4648 LNET_WIRE_HANDLE_COOKIE_NONE;
4649 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4650 LNET_WIRE_HANDLE_COOKIE_NONE;
4653 lnet_res_unlock(cpt);
4655 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4657 if (ack == LNET_ACK_REQ)
4658 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4660 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4664 rc = lnet_send(self, msg, LNET_NID_ANY);
4667 CNETERR("Error sending PUT to %s: %d\n",
4668 libcfs_id2str(target), rc);
4669 msg->msg_no_resend = true;
4670 lnet_finalize(msg, rc);
4673 /* completion will be signalled by an event */
4676 EXPORT_SYMBOL(LNetPut);
4679 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
4680 * returns a msg for the LND to pass to lnet_finalize() when the sink
4681 * data has been received.
4683 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4684 * lnet_finalize() is called on it, so the LND must call this first
4687 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4689 struct lnet_msg *msg = lnet_msg_alloc();
4690 struct lnet_libmd *getmd = getmsg->msg_md;
4691 struct lnet_process_id peer_id = getmsg->msg_target;
4694 LASSERT(!getmsg->msg_target_is_router);
4695 LASSERT(!getmsg->msg_routing);
4698 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4699 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4703 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4706 LASSERT(getmd->md_refcount > 0);
4708 if (getmd->md_threshold == 0) {
4709 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4710 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4712 lnet_res_unlock(cpt);
4716 LASSERT(getmd->md_offset == 0);
4718 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4719 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4721 /* setup information for lnet_build_msg_event */
4722 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4723 msg->msg_from = peer_id.nid;
4724 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4725 msg->msg_hdr.src_nid = peer_id.nid;
4726 msg->msg_hdr.payload_length = getmd->md_length;
4727 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4729 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4730 lnet_res_unlock(cpt);
4732 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4735 lnet_msg_commit(msg, cpt);
4736 lnet_net_unlock(cpt);
4738 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4743 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4746 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4747 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4748 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4750 lnet_net_unlock(cpt);
4757 EXPORT_SYMBOL(lnet_create_reply_msg);
4760 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4763 /* Set the REPLY length, now the RDMA that elides the REPLY message has
4764 * completed and I know it. */
4765 LASSERT(reply != NULL);
4766 LASSERT(reply->msg_type == LNET_MSG_GET);
4767 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4769 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
4770 * the end of my buffer, I might as well be dead. */
4771 LASSERT(len <= reply->msg_ev.mlength);
4773 reply->msg_ev.mlength = len;
4775 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4778 * Initiate an asynchronous GET operation.
4780 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4781 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4782 * the target node in the REPLY has been written to local MD.
4784 * On the target node, an LNET_EVENT_GET is logged when the GET request
4785 * arrives and is accepted into a MD.
4787 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4788 * \param mdh A handle for the MD that describes the memory into which the
4789 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4791 * \retval 0 Success, and only in this case events will be generated
4792 * and logged to EQ (if it exists) of the MD.
4793 * \retval -EIO Simulated failure.
4794 * \retval -ENOMEM Memory allocation failure.
4795 * \retval -ENOENT Invalid MD object.
4798 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4799 struct lnet_process_id target, unsigned int portal,
4800 __u64 match_bits, unsigned int offset, bool recovery)
4802 struct lnet_msg *msg;
4803 struct lnet_libmd *md;
4804 struct lnet_rsp_tracker *rspt;
4808 LASSERT(the_lnet.ln_refcount > 0);
4810 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4811 fail_peer(target.nid, 1)) /* shall we now? */
4813 CERROR("Dropping GET to %s: simulated failure\n",
4814 libcfs_id2str(target));
4818 msg = lnet_msg_alloc();
4820 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4821 libcfs_id2str(target));
4825 cpt = lnet_cpt_of_cookie(mdh.cookie);
4827 rspt = lnet_rspt_alloc(cpt);
4829 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4830 libcfs_id2str(target));
4833 INIT_LIST_HEAD(&rspt->rspt_on_list);
4835 msg->msg_recovery = recovery;
4839 md = lnet_handle2md(&mdh);
4840 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4841 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4842 match_bits, portal, libcfs_id2str(target),
4843 md == NULL ? -1 : md->md_threshold);
4844 if (md != NULL && md->md_me != NULL)
4845 CERROR("REPLY MD also attached to portal %d\n",
4846 md->md_me->me_portal);
4848 lnet_res_unlock(cpt);
4851 LIBCFS_FREE(rspt, sizeof(*rspt));
4855 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4857 lnet_msg_attach_md(msg, md, 0, 0);
4859 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
4861 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
4862 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
4863 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
4864 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
4866 /* NB handles only looked up by creator (no flips) */
4867 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
4868 the_lnet.ln_interface_cookie;
4869 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
4870 md->md_lh.lh_cookie;
4872 lnet_res_unlock(cpt);
4874 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4876 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4878 rc = lnet_send(self, msg, LNET_NID_ANY);
4880 CNETERR("Error sending GET to %s: %d\n",
4881 libcfs_id2str(target), rc);
4882 msg->msg_no_resend = true;
4883 lnet_finalize(msg, rc);
4886 /* completion will be signalled by an event */
4889 EXPORT_SYMBOL(LNetGet);
4892 * Calculate distance to node at \a dstnid.
4894 * \param dstnid Target NID.
4895 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
4897 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
4900 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
4901 * local_nid_dist_zero is set, which is the default.
4902 * \retval positives Distance to target NID, i.e. number of hops plus one.
4903 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
4906 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
4908 struct list_head *e;
4909 struct lnet_ni *ni = NULL;
4910 struct lnet_remotenet *rnet;
4911 __u32 dstnet = LNET_NIDNET(dstnid);
4915 struct list_head *rn_list;
4917 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
4918 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
4919 * keep order 0 free for 0@lo and order 1 free for a local NID
4922 LASSERT(the_lnet.ln_refcount > 0);
4924 cpt = lnet_net_lock_current();
4926 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
4927 if (ni->ni_nid == dstnid) {
4928 if (srcnidp != NULL)
4930 if (orderp != NULL) {
4931 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
4936 lnet_net_unlock(cpt);
4938 return local_nid_dist_zero ? 0 : 1;
4941 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
4942 /* Check if ni was originally created in
4943 * current net namespace.
4944 * If not, assign order above 0xffff0000,
4945 * to make this ni not a priority. */
4946 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
4947 order += 0xffff0000;
4949 if (srcnidp != NULL)
4950 *srcnidp = ni->ni_nid;
4953 lnet_net_unlock(cpt);
4960 rn_list = lnet_net2rnethash(dstnet);
4961 list_for_each(e, rn_list) {
4962 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
4964 if (rnet->lrn_net == dstnet) {
4965 struct lnet_route *route;
4966 struct lnet_route *shortest = NULL;
4967 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
4970 LASSERT(!list_empty(&rnet->lrn_routes));
4972 list_for_each_entry(route, &rnet->lrn_routes,
4974 route_hops = route->lr_hops;
4975 if (route_hops == LNET_UNDEFINED_HOPS)
4977 if (shortest == NULL ||
4978 route_hops < shortest_hops) {
4980 shortest_hops = route_hops;
4984 LASSERT(shortest != NULL);
4985 hops = shortest_hops;
4986 if (srcnidp != NULL) {
4987 struct lnet_net *net;
4988 net = lnet_get_net_locked(shortest->lr_lnet);
4990 ni = lnet_get_next_ni_locked(net, NULL);
4991 *srcnidp = ni->ni_nid;
4995 lnet_net_unlock(cpt);
5001 lnet_net_unlock(cpt);
5002 return -EHOSTUNREACH;
5004 EXPORT_SYMBOL(LNetDist);