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, 2016, 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 <lnet/lib-lnet.h>
40 #include <linux/nsproxy.h>
41 #include <net/net_namespace.h>
43 static int local_nid_dist_zero = 1;
44 module_param(local_nid_dist_zero, int, 0444);
45 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
48 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
52 struct list_head *next;
53 struct list_head cull;
55 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
57 /* Adding a new entry */
58 LIBCFS_ALLOC(tp, sizeof(*tp));
63 tp->tp_threshold = threshold;
66 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
71 /* removing entries */
72 INIT_LIST_HEAD(&cull);
76 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
77 tp = list_entry(el, lnet_test_peer_t, tp_list);
79 if (tp->tp_threshold == 0 || /* needs culling anyway */
80 nid == LNET_NID_ANY || /* removing all entries */
81 tp->tp_nid == nid) { /* matched this one */
82 list_del(&tp->tp_list);
83 list_add(&tp->tp_list, &cull);
89 while (!list_empty(&cull)) {
90 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
92 list_del(&tp->tp_list);
93 LIBCFS_FREE(tp, sizeof(*tp));
99 fail_peer (lnet_nid_t nid, int outgoing)
101 lnet_test_peer_t *tp;
102 struct list_head *el;
103 struct list_head *next;
104 struct list_head cull;
107 INIT_LIST_HEAD(&cull);
109 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
112 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
113 tp = list_entry(el, lnet_test_peer_t, tp_list);
115 if (tp->tp_threshold == 0) {
118 /* only cull zombies on outgoing tests,
119 * since we may be at interrupt priority on
120 * incoming messages. */
121 list_del(&tp->tp_list);
122 list_add(&tp->tp_list, &cull);
127 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
128 nid == tp->tp_nid) { /* fail this peer */
131 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
134 tp->tp_threshold == 0) {
136 list_del(&tp->tp_list);
137 list_add(&tp->tp_list, &cull);
146 while (!list_empty(&cull)) {
147 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
148 list_del(&tp->tp_list);
150 LIBCFS_FREE(tp, sizeof(*tp));
157 lnet_iov_nob(unsigned int niov, struct kvec *iov)
159 unsigned int nob = 0;
161 LASSERT(niov == 0 || iov != NULL);
163 nob += (iov++)->iov_len;
167 EXPORT_SYMBOL(lnet_iov_nob);
170 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
171 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
174 /* NB diov, siov are READ-ONLY */
175 unsigned int this_nob;
180 /* skip complete frags before 'doffset' */
182 while (doffset >= diov->iov_len) {
183 doffset -= diov->iov_len;
189 /* skip complete frags before 'soffset' */
191 while (soffset >= siov->iov_len) {
192 soffset -= siov->iov_len;
201 this_nob = MIN(diov->iov_len - doffset,
202 siov->iov_len - soffset);
203 this_nob = MIN(this_nob, nob);
205 memcpy((char *)diov->iov_base + doffset,
206 (char *)siov->iov_base + soffset, this_nob);
209 if (diov->iov_len > doffset + this_nob) {
217 if (siov->iov_len > soffset + this_nob) {
226 EXPORT_SYMBOL(lnet_copy_iov2iov);
229 lnet_extract_iov(int dst_niov, struct kvec *dst,
230 int src_niov, struct kvec *src,
231 unsigned int offset, unsigned int len)
233 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
234 * for exactly 'len' bytes, and return the number of entries.
235 * NB not destructive to 'src' */
236 unsigned int frag_len;
239 if (len == 0) /* no data => */
240 return (0); /* no frags */
242 LASSERT(src_niov > 0);
243 while (offset >= src->iov_len) { /* skip initial frags */
244 offset -= src->iov_len;
247 LASSERT(src_niov > 0);
252 LASSERT(src_niov > 0);
253 LASSERT((int)niov <= dst_niov);
255 frag_len = src->iov_len - offset;
256 dst->iov_base = ((char *)src->iov_base) + offset;
258 if (len <= frag_len) {
263 dst->iov_len = frag_len;
273 EXPORT_SYMBOL(lnet_extract_iov);
277 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
279 unsigned int nob = 0;
281 LASSERT(niov == 0 || kiov != NULL);
283 nob += (kiov++)->kiov_len;
287 EXPORT_SYMBOL(lnet_kiov_nob);
290 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
291 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
294 /* NB diov, siov are READ-ONLY */
295 unsigned int this_nob;
302 LASSERT (!in_interrupt ());
305 while (doffset >= diov->kiov_len) {
306 doffset -= diov->kiov_len;
313 while (soffset >= siov->kiov_len) {
314 soffset -= siov->kiov_len;
323 this_nob = MIN(diov->kiov_len - doffset,
324 siov->kiov_len - soffset);
325 this_nob = MIN(this_nob, nob);
328 daddr = ((char *)kmap(diov->kiov_page)) +
329 diov->kiov_offset + doffset;
331 saddr = ((char *)kmap(siov->kiov_page)) +
332 siov->kiov_offset + soffset;
334 /* Vanishing risk of kmap deadlock when mapping 2 pages.
335 * However in practice at least one of the kiovs will be mapped
336 * kernel pages and the map/unmap will be NOOPs */
338 memcpy (daddr, saddr, this_nob);
341 if (diov->kiov_len > doffset + this_nob) {
345 kunmap(diov->kiov_page);
352 if (siov->kiov_len > soffset + this_nob) {
356 kunmap(siov->kiov_page);
365 kunmap(diov->kiov_page);
367 kunmap(siov->kiov_page);
369 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
372 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
373 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
376 /* NB iov, kiov are READ-ONLY */
377 unsigned int this_nob;
383 LASSERT (!in_interrupt ());
386 while (iovoffset >= iov->iov_len) {
387 iovoffset -= iov->iov_len;
394 while (kiovoffset >= kiov->kiov_len) {
395 kiovoffset -= kiov->kiov_len;
404 this_nob = MIN(iov->iov_len - iovoffset,
405 kiov->kiov_len - kiovoffset);
406 this_nob = MIN(this_nob, nob);
409 addr = ((char *)kmap(kiov->kiov_page)) +
410 kiov->kiov_offset + kiovoffset;
412 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
415 if (iov->iov_len > iovoffset + this_nob) {
416 iovoffset += this_nob;
423 if (kiov->kiov_len > kiovoffset + this_nob) {
425 kiovoffset += this_nob;
427 kunmap(kiov->kiov_page);
437 kunmap(kiov->kiov_page);
439 EXPORT_SYMBOL(lnet_copy_kiov2iov);
442 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
443 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
446 /* NB kiov, iov are READ-ONLY */
447 unsigned int this_nob;
453 LASSERT (!in_interrupt ());
456 while (kiovoffset >= kiov->kiov_len) {
457 kiovoffset -= kiov->kiov_len;
464 while (iovoffset >= iov->iov_len) {
465 iovoffset -= iov->iov_len;
474 this_nob = MIN(kiov->kiov_len - kiovoffset,
475 iov->iov_len - iovoffset);
476 this_nob = MIN(this_nob, nob);
479 addr = ((char *)kmap(kiov->kiov_page)) +
480 kiov->kiov_offset + kiovoffset;
482 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
485 if (kiov->kiov_len > kiovoffset + this_nob) {
487 kiovoffset += this_nob;
489 kunmap(kiov->kiov_page);
496 if (iov->iov_len > iovoffset + this_nob) {
497 iovoffset += this_nob;
506 kunmap(kiov->kiov_page);
508 EXPORT_SYMBOL(lnet_copy_iov2kiov);
511 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
512 int src_niov, lnet_kiov_t *src,
513 unsigned int offset, unsigned int len)
515 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
516 * for exactly 'len' bytes, and return the number of entries.
517 * NB not destructive to 'src' */
518 unsigned int frag_len;
521 if (len == 0) /* no data => */
522 return (0); /* no frags */
524 LASSERT(src_niov > 0);
525 while (offset >= src->kiov_len) { /* skip initial frags */
526 offset -= src->kiov_len;
529 LASSERT(src_niov > 0);
534 LASSERT(src_niov > 0);
535 LASSERT((int)niov <= dst_niov);
537 frag_len = src->kiov_len - offset;
538 dst->kiov_page = src->kiov_page;
539 dst->kiov_offset = src->kiov_offset + offset;
541 if (len <= frag_len) {
543 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
547 dst->kiov_len = frag_len;
548 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
558 EXPORT_SYMBOL(lnet_extract_kiov);
561 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
562 unsigned int offset, unsigned int mlen, unsigned int rlen)
564 unsigned int niov = 0;
565 struct kvec *iov = NULL;
566 lnet_kiov_t *kiov = NULL;
569 LASSERT (!in_interrupt ());
570 LASSERT (mlen == 0 || msg != NULL);
573 LASSERT(msg->msg_receiving);
574 LASSERT(!msg->msg_sending);
575 LASSERT(rlen == msg->msg_len);
576 LASSERT(mlen <= msg->msg_len);
577 LASSERT(msg->msg_offset == offset);
578 LASSERT(msg->msg_wanted == mlen);
580 msg->msg_receiving = 0;
583 niov = msg->msg_niov;
585 kiov = msg->msg_kiov;
588 LASSERT ((iov == NULL) != (kiov == NULL));
592 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
593 niov, iov, kiov, offset, mlen,
596 lnet_finalize(ni, msg, rc);
600 lnet_setpayloadbuffer(lnet_msg_t *msg)
602 lnet_libmd_t *md = msg->msg_md;
604 LASSERT(msg->msg_len > 0);
605 LASSERT(!msg->msg_routing);
607 LASSERT(msg->msg_niov == 0);
608 LASSERT(msg->msg_iov == NULL);
609 LASSERT(msg->msg_kiov == NULL);
611 msg->msg_niov = md->md_niov;
612 if ((md->md_options & LNET_MD_KIOV) != 0)
613 msg->msg_kiov = md->md_iov.kiov;
615 msg->msg_iov = md->md_iov.iov;
619 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
620 unsigned int offset, unsigned int len)
622 msg->msg_type = type;
623 msg->msg_target = target;
625 msg->msg_offset = offset;
628 lnet_setpayloadbuffer(msg);
630 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
631 msg->msg_hdr.type = cpu_to_le32(type);
632 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
633 /* src_nid will be set later */
634 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
635 msg->msg_hdr.payload_length = cpu_to_le32(len);
639 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
641 void *priv = msg->msg_private;
644 LASSERT (!in_interrupt ());
645 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
646 (msg->msg_txcredit && msg->msg_peertxcredit));
648 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
650 lnet_finalize(ni, msg, rc);
654 lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
658 LASSERT(!msg->msg_sending);
659 LASSERT(msg->msg_receiving);
660 LASSERT(!msg->msg_rx_ready_delay);
661 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
663 msg->msg_rx_ready_delay = 1;
664 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
667 CERROR("recv from %s / send to %s aborted: "
668 "eager_recv failed %d\n",
669 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
670 libcfs_id2str(msg->msg_target), rc);
671 LASSERT(rc < 0); /* required by my callers */
678 * This function can be called from two paths:
679 * 1. when sending a message
680 * 2. when decommiting a message (lnet_msg_decommit_tx())
681 * In both these cases the peer_ni should have it's reference count
682 * acquired by the caller and therefore it is safe to drop the spin
683 * lock before calling lnd_query()
686 lnet_ni_query_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp)
688 cfs_time_t last_alive = 0;
689 int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
691 LASSERT(lnet_peer_aliveness_enabled(lp));
692 LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
694 lnet_net_unlock(cpt);
695 (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
698 lp->lpni_last_query = cfs_time_current();
700 if (last_alive != 0) /* NI has updated timestamp */
701 lp->lpni_last_alive = last_alive;
704 /* NB: always called with lnet_net_lock held */
706 lnet_peer_is_alive (struct lnet_peer_ni *lp, cfs_time_t now)
711 LASSERT (lnet_peer_aliveness_enabled(lp));
714 * Trust lnet_notify() if it has more recent aliveness news, but
715 * ignore the initial assumed death (see lnet_peers_start_down()).
717 spin_lock(&lp->lpni_lock);
718 if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
719 cfs_time_aftereq(lp->lpni_timestamp, lp->lpni_last_alive)) {
720 spin_unlock(&lp->lpni_lock);
725 cfs_time_add(lp->lpni_last_alive,
726 cfs_time_seconds(lp->lpni_net->net_tunables.
728 alive = cfs_time_after(deadline, now);
731 * Update obsolete lp_alive except for routers assumed to be dead
732 * initially, because router checker would update aliveness in this
733 * case, and moreover lpni_last_alive at peer creation is assumed.
735 if (alive && !lp->lpni_alive &&
736 !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
737 spin_unlock(&lp->lpni_lock);
738 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
740 spin_unlock(&lp->lpni_lock);
747 /* NB: returns 1 when alive, 0 when dead, negative when error;
748 * may drop the lnet_net_lock */
750 lnet_peer_alive_locked (struct lnet_ni *ni, struct lnet_peer_ni *lp)
752 cfs_time_t now = cfs_time_current();
754 if (!lnet_peer_aliveness_enabled(lp))
757 if (lnet_peer_is_alive(lp, now))
761 * Peer appears dead, but we should avoid frequent NI queries (at
762 * most once per lnet_queryinterval seconds).
764 if (lp->lpni_last_query != 0) {
765 static const int lnet_queryinterval = 1;
767 cfs_time_t next_query =
768 cfs_time_add(lp->lpni_last_query,
769 cfs_time_seconds(lnet_queryinterval));
771 if (cfs_time_before(now, next_query)) {
773 CWARN("Unexpected aliveness of peer %s: "
775 libcfs_nid2str(lp->lpni_nid),
776 (int)now, (int)next_query,
778 lp->lpni_net->net_tunables.lct_peer_timeout);
783 /* query NI for latest aliveness news */
784 lnet_ni_query_locked(ni, lp);
786 if (lnet_peer_is_alive(lp, now))
789 lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
794 * \param msg The message to be sent.
795 * \param do_send True if lnet_ni_send() should be called in this function.
796 * lnet_send() is going to lnet_net_unlock immediately after this, so
797 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
799 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
800 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
801 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
802 * \retval -ECANCELED If the MD of the message has been unlinked.
805 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
807 struct lnet_peer_ni *lp = msg->msg_txpeer;
808 struct lnet_ni *ni = msg->msg_txni;
809 int cpt = msg->msg_tx_cpt;
810 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
812 /* non-lnet_send() callers have checked before */
813 LASSERT(!do_send || msg->msg_tx_delayed);
814 LASSERT(!msg->msg_receiving);
815 LASSERT(msg->msg_tx_committed);
817 /* NB 'lp' is always the next hop */
818 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
819 lnet_peer_alive_locked(ni, lp) == 0) {
820 the_lnet.ln_counters[cpt]->drop_count++;
821 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
822 lnet_net_unlock(cpt);
824 atomic_inc(&msg->msg_txpeer->lpni_stats.drop_count);
826 atomic_inc(&msg->msg_txni->ni_stats.drop_count);
828 CNETERR("Dropping message for %s: peer not alive\n",
829 libcfs_id2str(msg->msg_target));
831 lnet_finalize(ni, msg, -EHOSTUNREACH);
834 return -EHOSTUNREACH;
837 if (msg->msg_md != NULL &&
838 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
839 lnet_net_unlock(cpt);
841 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
842 "called on the MD/ME.\n",
843 libcfs_id2str(msg->msg_target));
845 lnet_finalize(ni, msg, -ECANCELED);
851 if (!msg->msg_peertxcredit) {
852 spin_lock(&lp->lpni_lock);
853 LASSERT((lp->lpni_txcredits < 0) ==
854 !list_empty(&lp->lpni_txq));
856 msg->msg_peertxcredit = 1;
857 lp->lpni_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
858 lp->lpni_txcredits--;
860 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
861 lp->lpni_mintxcredits = lp->lpni_txcredits;
863 if (lp->lpni_txcredits < 0) {
864 msg->msg_tx_delayed = 1;
865 list_add_tail(&msg->msg_list, &lp->lpni_txq);
866 spin_unlock(&lp->lpni_lock);
867 return LNET_CREDIT_WAIT;
869 spin_unlock(&lp->lpni_lock);
872 if (!msg->msg_txcredit) {
873 LASSERT((tq->tq_credits < 0) ==
874 !list_empty(&tq->tq_delayed));
876 msg->msg_txcredit = 1;
878 atomic_dec(&ni->ni_tx_credits);
880 if (tq->tq_credits < tq->tq_credits_min)
881 tq->tq_credits_min = tq->tq_credits;
883 if (tq->tq_credits < 0) {
884 msg->msg_tx_delayed = 1;
885 list_add_tail(&msg->msg_list, &tq->tq_delayed);
886 return LNET_CREDIT_WAIT;
891 lnet_net_unlock(cpt);
892 lnet_ni_send(ni, msg);
895 return LNET_CREDIT_OK;
899 static lnet_rtrbufpool_t *
900 lnet_msg2bufpool(lnet_msg_t *msg)
902 lnet_rtrbufpool_t *rbp;
905 LASSERT(msg->msg_rx_committed);
907 cpt = msg->msg_rx_cpt;
908 rbp = &the_lnet.ln_rtrpools[cpt][0];
910 LASSERT(msg->msg_len <= LNET_MTU);
911 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
913 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
920 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
922 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
923 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
924 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
925 * received or OK to receive */
926 struct lnet_peer_ni *lp = msg->msg_rxpeer;
927 lnet_rtrbufpool_t *rbp;
930 LASSERT (msg->msg_iov == NULL);
931 LASSERT (msg->msg_kiov == NULL);
932 LASSERT (msg->msg_niov == 0);
933 LASSERT (msg->msg_routing);
934 LASSERT (msg->msg_receiving);
935 LASSERT (!msg->msg_sending);
937 /* non-lnet_parse callers only receive delayed messages */
938 LASSERT(!do_recv || msg->msg_rx_delayed);
940 if (!msg->msg_peerrtrcredit) {
941 spin_lock(&lp->lpni_lock);
942 LASSERT((lp->lpni_rtrcredits < 0) ==
943 !list_empty(&lp->lpni_rtrq));
945 msg->msg_peerrtrcredit = 1;
946 lp->lpni_rtrcredits--;
947 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
948 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
950 if (lp->lpni_rtrcredits < 0) {
951 /* must have checked eager_recv before here */
952 LASSERT(msg->msg_rx_ready_delay);
953 msg->msg_rx_delayed = 1;
954 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
955 spin_unlock(&lp->lpni_lock);
956 return LNET_CREDIT_WAIT;
958 spin_unlock(&lp->lpni_lock);
961 rbp = lnet_msg2bufpool(msg);
963 if (!msg->msg_rtrcredit) {
964 msg->msg_rtrcredit = 1;
966 if (rbp->rbp_credits < rbp->rbp_mincredits)
967 rbp->rbp_mincredits = rbp->rbp_credits;
969 if (rbp->rbp_credits < 0) {
970 /* must have checked eager_recv before here */
971 LASSERT(msg->msg_rx_ready_delay);
972 msg->msg_rx_delayed = 1;
973 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
974 return LNET_CREDIT_WAIT;
978 LASSERT(!list_empty(&rbp->rbp_bufs));
979 rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
980 list_del(&rb->rb_list);
982 msg->msg_niov = rbp->rbp_npages;
983 msg->msg_kiov = &rb->rb_kiov[0];
986 int cpt = msg->msg_rx_cpt;
988 lnet_net_unlock(cpt);
989 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
990 0, msg->msg_len, msg->msg_len);
993 return LNET_CREDIT_OK;
997 lnet_return_tx_credits_locked(lnet_msg_t *msg)
999 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1000 struct lnet_ni *txni = msg->msg_txni;
1003 if (msg->msg_txcredit) {
1004 struct lnet_ni *ni = msg->msg_txni;
1005 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1007 /* give back NI txcredits */
1008 msg->msg_txcredit = 0;
1010 LASSERT((tq->tq_credits < 0) ==
1011 !list_empty(&tq->tq_delayed));
1014 atomic_inc(&ni->ni_tx_credits);
1015 if (tq->tq_credits <= 0) {
1016 msg2 = list_entry(tq->tq_delayed.next,
1017 lnet_msg_t, msg_list);
1018 list_del(&msg2->msg_list);
1020 LASSERT(msg2->msg_txni == ni);
1021 LASSERT(msg2->msg_tx_delayed);
1022 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1024 (void) lnet_post_send_locked(msg2, 1);
1028 if (msg->msg_peertxcredit) {
1029 /* give back peer txcredits */
1030 msg->msg_peertxcredit = 0;
1032 spin_lock(&txpeer->lpni_lock);
1033 LASSERT((txpeer->lpni_txcredits < 0) ==
1034 !list_empty(&txpeer->lpni_txq));
1036 txpeer->lpni_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1037 LASSERT(txpeer->lpni_txqnob >= 0);
1039 txpeer->lpni_txcredits++;
1040 if (txpeer->lpni_txcredits <= 0) {
1041 msg2 = list_entry(txpeer->lpni_txq.next,
1042 lnet_msg_t, msg_list);
1043 list_del(&msg2->msg_list);
1044 spin_unlock(&txpeer->lpni_lock);
1046 LASSERT(msg2->msg_txpeer == txpeer);
1047 LASSERT(msg2->msg_tx_delayed);
1049 if (msg2->msg_tx_cpt != msg->msg_tx_cpt) {
1050 lnet_net_unlock(msg->msg_tx_cpt);
1051 lnet_net_lock(msg2->msg_tx_cpt);
1053 (void) lnet_post_send_locked(msg2, 1);
1054 if (msg2->msg_tx_cpt != msg->msg_tx_cpt) {
1055 lnet_net_unlock(msg2->msg_tx_cpt);
1056 lnet_net_lock(msg->msg_tx_cpt);
1059 spin_unlock(&txpeer->lpni_lock);
1064 msg->msg_txni = NULL;
1065 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1068 if (txpeer != NULL) {
1071 * Once the patch for the health comes in we need to set
1072 * the health of the peer ni to bad when we fail to send
1074 * int status = msg->msg_ev.status;
1076 * lnet_set_peer_ni_health_locked(txpeer, false)
1078 msg->msg_txpeer = NULL;
1079 lnet_peer_ni_decref_locked(txpeer);
1084 lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp)
1088 if (list_empty(&rbp->rbp_msgs))
1090 msg = list_entry(rbp->rbp_msgs.next,
1091 lnet_msg_t, msg_list);
1092 list_del(&msg->msg_list);
1094 (void)lnet_post_routed_recv_locked(msg, 1);
1098 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1103 lnet_net_unlock(cpt);
1105 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1106 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1107 0, 0, 0, msg->msg_hdr.payload_length);
1108 list_del_init(&msg->msg_list);
1109 lnet_finalize(NULL, msg, -ECANCELED);
1116 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1118 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1119 struct lnet_ni *rxni = msg->msg_rxni;
1122 if (msg->msg_rtrcredit) {
1123 /* give back global router credits */
1125 lnet_rtrbufpool_t *rbp;
1127 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1128 * there until it gets one allocated, or aborts the wait
1130 LASSERT(msg->msg_kiov != NULL);
1132 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1135 msg->msg_kiov = NULL;
1136 msg->msg_rtrcredit = 0;
1138 LASSERT(rbp == lnet_msg2bufpool(msg));
1140 LASSERT((rbp->rbp_credits > 0) ==
1141 !list_empty(&rbp->rbp_bufs));
1143 /* If routing is now turned off, we just drop this buffer and
1144 * don't bother trying to return credits. */
1145 if (!the_lnet.ln_routing) {
1146 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1150 /* It is possible that a user has lowered the desired number of
1151 * buffers in this pool. Make sure we never put back
1152 * more buffers than the stated number. */
1153 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1154 /* Discard this buffer so we don't have too
1156 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1157 rbp->rbp_nbuffers--;
1159 list_add(&rb->rb_list, &rbp->rbp_bufs);
1161 if (rbp->rbp_credits <= 0)
1162 lnet_schedule_blocked_locked(rbp);
1167 if (msg->msg_peerrtrcredit) {
1168 /* give back peer router credits */
1169 msg->msg_peerrtrcredit = 0;
1171 spin_lock(&rxpeer->lpni_lock);
1172 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1173 !list_empty(&rxpeer->lpni_rtrq));
1175 rxpeer->lpni_rtrcredits++;
1177 /* drop all messages which are queued to be routed on that
1179 if (!the_lnet.ln_routing) {
1180 struct list_head drop;
1181 INIT_LIST_HEAD(&drop);
1182 list_splice_init(&rxpeer->lpni_rtrq, &drop);
1183 spin_unlock(&rxpeer->lpni_lock);
1184 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1185 } else if (rxpeer->lpni_rtrcredits <= 0) {
1186 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1187 lnet_msg_t, msg_list);
1188 list_del(&msg2->msg_list);
1189 spin_unlock(&rxpeer->lpni_lock);
1190 (void) lnet_post_routed_recv_locked(msg2, 1);
1192 spin_unlock(&rxpeer->lpni_lock);
1196 msg->msg_rxni = NULL;
1197 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1199 if (rxpeer != NULL) {
1200 msg->msg_rxpeer = NULL;
1201 lnet_peer_ni_decref_locked(rxpeer);
1206 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1208 if (p1->lpni_txqnob < p2->lpni_txqnob)
1211 if (p1->lpni_txqnob > p2->lpni_txqnob)
1214 if (p1->lpni_txcredits > p2->lpni_txcredits)
1217 if (p1->lpni_txcredits < p2->lpni_txcredits)
1224 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1226 struct lnet_peer_ni *p1 = r1->lr_gateway;
1227 struct lnet_peer_ni *p2 = r2->lr_gateway;
1228 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1229 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1232 if (r1->lr_priority < r2->lr_priority)
1235 if (r1->lr_priority > r2->lr_priority)
1238 if (r1_hops < r2_hops)
1241 if (r1_hops > r2_hops)
1244 rc = lnet_compare_peers(p1, p2);
1248 if (r1->lr_seq - r2->lr_seq <= 0)
1254 static struct lnet_peer_ni *
1255 lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
1258 lnet_remotenet_t *rnet;
1259 lnet_route_t *route;
1260 lnet_route_t *best_route;
1261 lnet_route_t *last_route;
1262 struct lnet_peer_ni *lpni_best;
1263 struct lnet_peer_ni *lp;
1266 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1267 * rtr_nid nid, otherwise find the best gateway I can use */
1269 rnet = lnet_find_rnet_locked(LNET_NIDNET(target));
1274 best_route = last_route = NULL;
1275 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1276 lp = route->lr_gateway;
1278 if (!lnet_is_route_alive(route))
1281 if (net != NULL && lp->lpni_net != net)
1284 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1287 if (lpni_best == NULL) {
1288 best_route = last_route = route;
1293 /* no protection on below fields, but it's harmless */
1294 if (last_route->lr_seq - route->lr_seq < 0)
1297 rc = lnet_compare_routes(route, best_route);
1305 /* set sequence number on the best router to the latest sequence + 1
1306 * so we can round-robin all routers, it's race and inaccurate but
1307 * harmless and functional */
1308 if (best_route != NULL)
1309 best_route->lr_seq = last_route->lr_seq + 1;
1314 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
1315 struct lnet_msg *msg, lnet_nid_t rtr_nid, bool *lo_sent)
1317 struct lnet_ni *best_ni = NULL;
1318 struct lnet_peer_ni *best_lpni = NULL;
1319 struct lnet_peer_ni *net_gw = NULL;
1320 struct lnet_peer_ni *best_gw = NULL;
1321 struct lnet_peer_ni *lpni;
1322 struct lnet_peer *peer = NULL;
1323 struct lnet_peer_net *peer_net;
1324 struct lnet_net *local_net;
1325 struct lnet_ni *ni = NULL;
1327 bool routing = false;
1328 bool ni_is_pref = false;
1329 bool preferred = false;
1330 int best_credits = 0;
1332 int best_lpni_credits = INT_MIN;
1334 int shortest_distance = INT_MAX;
1336 bool found_ir = false;
1340 * get an initial CPT to use for locking. The idea here is not to
1341 * serialize the calls to select_pathway, so that as many
1342 * operations can run concurrently as possible. To do that we use
1343 * the CPT where this call is being executed. Later on when we
1344 * determine the CPT to use in lnet_message_commit, we switch the
1345 * lock and check if there was any configuration changes, if none,
1346 * then we proceed, if there is, then we'll need to update the cpt
1347 * and redo the operation.
1349 cpt = lnet_net_lock_current();
1355 shortest_distance = INT_MAX;
1358 if (the_lnet.ln_shutdown) {
1359 lnet_net_unlock(cpt);
1363 if (msg->msg_md != NULL)
1364 /* get the cpt of the MD, used during NUMA based selection */
1365 md_cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
1367 md_cpt = CFS_CPT_ANY;
1370 * initialize the variables which could be reused if we go to
1374 seq = lnet_get_dlc_seq_locked();
1376 peer = lnet_find_or_create_peer_locked(dst_nid, cpt);
1378 lnet_net_unlock(cpt);
1379 return PTR_ERR(peer);
1382 /* If peer is not healthy then can not send anything to it */
1383 if (!lnet_is_peer_healthy_locked(peer)) {
1384 lnet_net_unlock(cpt);
1385 return -EHOSTUNREACH;
1388 if (!peer->lp_multi_rail && lnet_get_num_peer_nis(peer) > 1) {
1389 CERROR("peer %s is declared to be non MR capable, "
1390 "yet configured with more than one NID\n",
1391 libcfs_nid2str(dst_nid));
1396 * STEP 1: first jab at determineing best_ni
1397 * if src_nid is explicitly specified, then best_ni is already
1398 * pre-determiend for us. Otherwise we need to select the best
1399 * one to use later on
1401 if (src_nid != LNET_NID_ANY) {
1402 best_ni = lnet_nid2ni_locked(src_nid, cpt);
1404 lnet_net_unlock(cpt);
1405 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1406 "local nid\n", libcfs_nid2str(dst_nid),
1407 libcfs_nid2str(src_nid));
1411 if (best_ni->ni_net->net_id != LNET_NIDNET(dst_nid)) {
1412 lnet_net_unlock(cpt);
1413 LCONSOLE_WARN("No route to %s via from %s\n",
1414 libcfs_nid2str(dst_nid),
1415 libcfs_nid2str(src_nid));
1424 * Decide whether we need to route to peer_ni.
1425 * Get the local net that I need to be on to be able to directly
1426 * send to that peer.
1428 * a. Find the peer which the dst_nid belongs to.
1429 * b. Iterate through each of the peer_nets/nis to decide
1430 * the best peer/local_ni pair to use
1432 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_on_peer_list) {
1433 if (!lnet_is_peer_net_healthy_locked(peer_net))
1436 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1439 * go through each peer_ni on that peer_net and
1440 * determine the best possible gw to go through
1442 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
1443 lpni_on_peer_net_list) {
1444 net_gw = lnet_find_route_locked(NULL,
1449 * if no route is found for that network then
1450 * move onto the next peer_ni in the peer
1459 rc = lnet_compare_peers(net_gw,
1471 local_net = lnet_get_net_locked
1472 (LNET_NIDNET(best_gw->lpni_nid));
1479 /* no routable net found go on to a different net */
1484 * Iterate through the NIs in this local Net and select
1485 * the NI to send from. The selection is determined by
1486 * these 3 criterion in the following priority:
1488 * 2. NI available credits
1491 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1494 if (!lnet_is_ni_healthy_locked(ni))
1497 ni_credits = atomic_read(&ni->ni_tx_credits);
1500 * calculate the distance from the cpt on which
1501 * the message memory is allocated to the CPT of
1502 * the NI's physical device
1504 distance = cfs_cpt_distance(lnet_cpt_table(),
1509 * If we already have a closer NI within the NUMA
1510 * range provided, then there is no need to
1511 * consider the current NI. Move on to the next
1514 if (distance > shortest_distance &&
1515 distance > lnet_get_numa_range())
1518 if (distance < shortest_distance &&
1519 distance > lnet_get_numa_range()) {
1521 * The current NI is the closest one that we
1522 * have found, even though it's not in the
1523 * NUMA range specified. This occurs if
1524 * the NUMA range is less than the least
1525 * of the distances in the system.
1526 * In effect NUMA range consideration is
1529 shortest_distance = distance;
1530 } else if ((distance <= shortest_distance &&
1531 distance < lnet_get_numa_range()) ||
1532 distance == shortest_distance) {
1534 * This NI is either within range or it's
1535 * equidistant. In both of these cases we
1536 * would want to select the NI based on
1537 * its available credits first, and then
1540 if (distance <= shortest_distance &&
1541 distance < lnet_get_numa_range()) {
1543 * If this is the first NI that's
1544 * within range, then set the
1545 * shortest distance to the range
1546 * specified by the user. In
1547 * effect we're saying that all
1548 * NIs that fall within this NUMA
1549 * range shall be dealt with as
1550 * having equal NUMA weight. Which
1551 * will mean that we should select
1552 * through that set by their
1553 * available credits first
1554 * followed by Round Robin.
1556 * And since this is the first NI
1557 * in the range, let's just set it
1558 * as our best_ni for now. The
1559 * following NIs found in the
1560 * range will be dealt with as
1561 * mentioned previously.
1563 shortest_distance = lnet_get_numa_range();
1570 * This NI is NUMA equidistant let's
1571 * select using credits followed by Round
1574 if (ni_credits < best_credits) {
1576 } else if (ni_credits == best_credits) {
1578 if (best_ni->ni_seq <= ni->ni_seq)
1585 best_credits = ni_credits;
1589 * if the peer is not MR capable, then we should always send to it
1590 * using the first NI in the NET we determined.
1592 if (!peer->lp_multi_rail && local_net != NULL)
1593 best_ni = lnet_net2ni_locked(local_net->net_id, cpt);
1596 lnet_net_unlock(cpt);
1597 LCONSOLE_WARN("No local ni found to send from to %s\n",
1598 libcfs_nid2str(dst_nid));
1603 * Now that we selected the NI to use increment its sequence
1604 * number so the Round Robin algorithm will detect that it has
1605 * been used and pick the next NI.
1613 if (best_ni == the_lnet.ln_loni) {
1614 /* No send credit hassles with LOLND */
1615 lnet_ni_addref_locked(best_ni, cpt);
1616 msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
1617 if (!msg->msg_routing)
1618 msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
1619 msg->msg_target.nid = best_ni->ni_nid;
1620 lnet_msg_commit(msg, cpt);
1622 lnet_net_unlock(cpt);
1623 msg->msg_txni = best_ni;
1624 lnet_ni_send(best_ni, msg);
1632 if (msg->msg_type == LNET_MSG_REPLY ||
1633 msg->msg_type == LNET_MSG_ACK) {
1635 * for replies we want to respond on the same peer_ni we
1636 * received the message on if possible. If not, then pick
1637 * a peer_ni to send to
1639 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1641 lnet_peer_ni_decref_locked(best_lpni);
1644 CDEBUG(D_NET, "unable to send msg_type %d to "
1645 "originating %s\n", msg->msg_type,
1646 libcfs_nid2str(dst_nid));
1650 peer_net = lnet_peer_get_net_locked(peer,
1651 best_ni->ni_net->net_id);
1653 * peer_net is not available or the src_nid is explicitly defined
1654 * and the peer_net for that src_nid is unhealthy. find a route to
1655 * the destination nid.
1658 (src_nid != LNET_NID_ANY &&
1659 !lnet_is_peer_net_healthy_locked(peer_net))) {
1660 best_gw = lnet_find_route_locked(best_ni->ni_net,
1664 * if no route is found for that network then
1665 * move onto the next peer_ni in the peer
1668 lnet_net_unlock(cpt);
1669 LCONSOLE_WARN("No route to peer from %s\n",
1670 libcfs_nid2str(best_ni->ni_nid));
1671 return -EHOSTUNREACH;
1674 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1675 libcfs_nid2str(dst_nid),
1676 libcfs_nid2str(best_gw->lpni_nid),
1677 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1679 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1680 LASSERT(best_lpni != NULL);
1681 lnet_peer_ni_decref_locked(best_lpni);
1686 } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
1688 * this peer_net is unhealthy but we still have an opportunity
1689 * to find another peer_net that we can use
1691 __u32 net_id = peer_net->lpn_net_id;
1692 lnet_net_unlock(cpt);
1694 LCONSOLE_WARN("peer net %s unhealthy\n",
1695 libcfs_net2str(net_id));
1700 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1702 * if this peer ni is not healthy just skip it, no point in
1703 * examining it further
1705 if (!lnet_is_peer_ni_healthy_locked(lpni))
1707 ni_is_pref = lnet_peer_is_ni_pref_locked(lpni, best_ni);
1709 /* if this is a preferred peer use it */
1710 if (!preferred && ni_is_pref) {
1712 } else if (preferred && !ni_is_pref) {
1714 * this is not the preferred peer so let's ignore
1718 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1720 * We already have a peer that has more credits
1721 * available than this one. No need to consider
1722 * this peer further.
1725 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1727 * The best peer found so far and the current peer
1728 * have the same number of available credits let's
1729 * make sure to select between them using Round
1733 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1739 best_lpni_credits = lpni->lpni_txcredits;
1743 * Increment sequence number of the peer selected so that we can
1744 * pick the next one in Round Robin.
1746 best_lpni->lpni_seq++;
1748 /* if we still can't find a peer ni then we can't reach it */
1750 __u32 net_id = peer_net->lpn_net_id;
1751 lnet_net_unlock(cpt);
1752 LCONSOLE_WARN("no peer_ni found on peer net %s\n",
1753 libcfs_net2str(net_id));
1754 return -EHOSTUNREACH;
1759 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1760 * message. This ensures that we get a CPT that is correct for
1761 * the NI when the NI has been restricted to a subset of all CPTs.
1762 * If the selected CPT differs from the one currently locked, we
1763 * must unlock and relock the lnet_net_lock(), and then check whether
1764 * the configuration has changed. We don't have a hold on the best_ni
1765 * or best_peer_ni yet, and they may have vanished.
1767 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1769 lnet_net_unlock(cpt);
1772 seq2 = lnet_get_dlc_seq_locked();
1774 lnet_net_unlock(cpt);
1780 * store the best_lpni in the message right away to avoid having
1781 * to do the same operation under different conditions
1783 msg->msg_txpeer = (routing) ? best_gw : best_lpni;
1784 msg->msg_txni = best_ni;
1786 * grab a reference for the best_ni since now it's in use in this
1787 * send. the reference will need to be dropped when the message is
1788 * finished in lnet_finalize()
1790 lnet_ni_addref_locked(msg->msg_txni, cpt);
1791 lnet_peer_ni_addref_locked(msg->msg_txpeer);
1794 * set the destination nid in the message here because it's
1795 * possible that we'd be sending to a different nid than the one
1798 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1801 * Always set the target.nid to the best peer picked. Either the
1802 * nid will be one of the preconfigured NIDs, or the same NID as
1803 * what was originaly set in the target or it will be the NID of
1804 * a router if this message should be routed
1806 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1809 * lnet_msg_commit assigns the correct cpt to the message, which
1810 * is used to decrement the correct refcount on the ni when it's
1811 * time to return the credits
1813 lnet_msg_commit(msg, cpt);
1816 * If we are routing the message then we don't need to overwrite
1817 * the src_nid since it would've been set at the origin. Otherwise
1818 * we are the originator so we need to set it.
1820 if (!msg->msg_routing)
1821 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1824 msg->msg_target_is_router = 1;
1825 msg->msg_target.pid = LNET_PID_LUSTRE;
1828 rc = lnet_post_send_locked(msg, 0);
1830 lnet_net_unlock(cpt);
1836 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1838 lnet_nid_t dst_nid = msg->msg_target.nid;
1840 bool lo_sent = false;
1843 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1844 * but we might want to use pre-determined router for ACK/REPLY
1847 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1848 LASSERT (msg->msg_txpeer == NULL);
1849 LASSERT (!msg->msg_sending);
1850 LASSERT (!msg->msg_target_is_router);
1851 LASSERT (!msg->msg_receiving);
1853 msg->msg_sending = 1;
1855 LASSERT(!msg->msg_tx_committed);
1857 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid, &lo_sent);
1858 if (rc < 0 || lo_sent)
1861 if (rc == LNET_CREDIT_OK)
1862 lnet_ni_send(msg->msg_txni, msg);
1864 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
1869 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1872 the_lnet.ln_counters[cpt]->drop_count++;
1873 the_lnet.ln_counters[cpt]->drop_length += nob;
1874 lnet_net_unlock(cpt);
1876 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1880 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1882 lnet_hdr_t *hdr = &msg->msg_hdr;
1884 if (msg->msg_wanted != 0)
1885 lnet_setpayloadbuffer(msg);
1887 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1889 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1890 * it back into the ACK during lnet_finalize() */
1891 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1892 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1894 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1895 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1899 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1901 lnet_hdr_t *hdr = &msg->msg_hdr;
1902 struct lnet_match_info info;
1906 /* Convert put fields to host byte order */
1907 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1908 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1909 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1911 /* Primary peer NID. */
1912 info.mi_id.nid = msg->msg_initiator;
1913 info.mi_id.pid = hdr->src_pid;
1914 info.mi_opc = LNET_MD_OP_PUT;
1915 info.mi_portal = hdr->msg.put.ptl_index;
1916 info.mi_rlength = hdr->payload_length;
1917 info.mi_roffset = hdr->msg.put.offset;
1918 info.mi_mbits = hdr->msg.put.match_bits;
1919 info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
1921 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
1922 ready_delay = msg->msg_rx_ready_delay;
1925 rc = lnet_ptl_match_md(&info, msg);
1930 case LNET_MATCHMD_OK:
1931 lnet_recv_put(ni, msg);
1934 case LNET_MATCHMD_NONE:
1936 /* no eager_recv or has already called it, should
1937 * have been attached on delayed list */
1940 rc = lnet_ni_eager_recv(ni, msg);
1947 case LNET_MATCHMD_DROP:
1948 CNETERR("Dropping PUT from %s portal %d match %llu"
1949 " offset %d length %d: %d\n",
1950 libcfs_id2str(info.mi_id), info.mi_portal,
1951 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1953 return -ENOENT; /* -ve: OK but no match */
1958 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1960 struct lnet_match_info info;
1961 lnet_hdr_t *hdr = &msg->msg_hdr;
1962 lnet_process_id_t source_id;
1963 struct lnet_handle_wire reply_wmd;
1966 /* Convert get fields to host byte order */
1967 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1968 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1969 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1970 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1972 source_id.nid = hdr->src_nid;
1973 source_id.pid = hdr->src_pid;
1974 /* Primary peer NID */
1975 info.mi_id.nid = msg->msg_initiator;
1976 info.mi_id.pid = hdr->src_pid;
1977 info.mi_opc = LNET_MD_OP_GET;
1978 info.mi_portal = hdr->msg.get.ptl_index;
1979 info.mi_rlength = hdr->msg.get.sink_length;
1980 info.mi_roffset = hdr->msg.get.src_offset;
1981 info.mi_mbits = hdr->msg.get.match_bits;
1983 rc = lnet_ptl_match_md(&info, msg);
1984 if (rc == LNET_MATCHMD_DROP) {
1985 CNETERR("Dropping GET from %s portal %d match %llu"
1986 " offset %d length %d\n",
1987 libcfs_id2str(info.mi_id), info.mi_portal,
1988 info.mi_mbits, info.mi_roffset, info.mi_rlength);
1989 return -ENOENT; /* -ve: OK but no match */
1992 LASSERT(rc == LNET_MATCHMD_OK);
1994 lnet_build_msg_event(msg, LNET_EVENT_GET);
1996 reply_wmd = hdr->msg.get.return_wmd;
1998 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
1999 msg->msg_offset, msg->msg_wanted);
2001 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
2004 /* The LND completes the REPLY from her recv procedure */
2005 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2006 msg->msg_offset, msg->msg_len, msg->msg_len);
2010 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
2011 msg->msg_receiving = 0;
2013 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
2015 /* didn't get as far as lnet_ni_send() */
2016 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
2017 libcfs_nid2str(ni->ni_nid),
2018 libcfs_id2str(info.mi_id), rc);
2020 lnet_finalize(ni, msg, rc);
2027 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
2029 void *private = msg->msg_private;
2030 lnet_hdr_t *hdr = &msg->msg_hdr;
2031 lnet_process_id_t src = {0};
2037 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
2040 src.nid = hdr->src_nid;
2041 src.pid = hdr->src_pid;
2043 /* NB handles only looked up by creator (no flips) */
2044 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
2045 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2046 CNETERR("%s: Dropping REPLY from %s for %s "
2048 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2049 (md == NULL) ? "invalid" : "inactive",
2050 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2051 hdr->msg.reply.dst_wmd.wh_object_cookie);
2052 if (md != NULL && md->md_me != NULL)
2053 CERROR("REPLY MD also attached to portal %d\n",
2054 md->md_me->me_portal);
2056 lnet_res_unlock(cpt);
2057 return -ENOENT; /* -ve: OK but no match */
2060 LASSERT(md->md_offset == 0);
2062 rlength = hdr->payload_length;
2063 mlength = MIN(rlength, (int)md->md_length);
2065 if (mlength < rlength &&
2066 (md->md_options & LNET_MD_TRUNCATE) == 0) {
2067 CNETERR("%s: Dropping REPLY from %s length %d "
2068 "for MD %#llx would overflow (%d)\n",
2069 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2070 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2072 lnet_res_unlock(cpt);
2073 return -ENOENT; /* -ve: OK but no match */
2076 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
2077 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2078 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2080 lnet_msg_attach_md(msg, md, 0, mlength);
2083 lnet_setpayloadbuffer(msg);
2085 lnet_res_unlock(cpt);
2087 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2089 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2094 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
2096 lnet_hdr_t *hdr = &msg->msg_hdr;
2097 lnet_process_id_t src = {0};
2101 src.nid = hdr->src_nid;
2102 src.pid = hdr->src_pid;
2104 /* Convert ack fields to host byte order */
2105 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2106 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2108 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
2111 /* NB handles only looked up by creator (no flips) */
2112 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2113 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2114 /* Don't moan; this is expected */
2116 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
2117 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2118 (md == NULL) ? "invalid" : "inactive",
2119 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2120 hdr->msg.ack.dst_wmd.wh_object_cookie);
2121 if (md != NULL && md->md_me != NULL)
2122 CERROR("Source MD also attached to portal %d\n",
2123 md->md_me->me_portal);
2125 lnet_res_unlock(cpt);
2126 return -ENOENT; /* -ve! */
2129 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
2130 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2131 hdr->msg.ack.dst_wmd.wh_object_cookie);
2133 lnet_msg_attach_md(msg, md, 0, 0);
2135 lnet_res_unlock(cpt);
2137 lnet_build_msg_event(msg, LNET_EVENT_ACK);
2139 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2144 * \retval LNET_CREDIT_OK If \a msg is forwarded
2145 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
2146 * \retval -ve error code
2149 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
2153 if (!the_lnet.ln_routing)
2156 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
2157 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2158 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
2159 msg->msg_rx_ready_delay = 1;
2161 lnet_net_unlock(msg->msg_rx_cpt);
2162 rc = lnet_ni_eager_recv(ni, msg);
2163 lnet_net_lock(msg->msg_rx_cpt);
2168 rc = lnet_post_routed_recv_locked(msg, 0);
2173 lnet_parse_local(lnet_ni_t *ni, lnet_msg_t *msg)
2177 switch (msg->msg_type) {
2179 rc = lnet_parse_ack(ni, msg);
2182 rc = lnet_parse_put(ni, msg);
2185 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
2187 case LNET_MSG_REPLY:
2188 rc = lnet_parse_reply(ni, msg);
2190 default: /* prevent an unused label if !kernel */
2195 LASSERT(rc == 0 || rc == -ENOENT);
2200 lnet_msgtyp2str (int type)
2209 case LNET_MSG_REPLY:
2211 case LNET_MSG_HELLO:
2214 return ("<UNKNOWN>");
2219 lnet_print_hdr(lnet_hdr_t * hdr)
2221 lnet_process_id_t src = {0};
2222 lnet_process_id_t dst = {0};
2223 char *type_str = lnet_msgtyp2str(hdr->type);
2225 src.nid = hdr->src_nid;
2226 src.pid = hdr->src_pid;
2228 dst.nid = hdr->dest_nid;
2229 dst.pid = hdr->dest_pid;
2231 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2232 CWARN(" From %s\n", libcfs_id2str(src));
2233 CWARN(" To %s\n", libcfs_id2str(dst));
2235 switch (hdr->type) {
2240 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
2241 "match bits %llu\n",
2242 hdr->msg.put.ptl_index,
2243 hdr->msg.put.ack_wmd.wh_interface_cookie,
2244 hdr->msg.put.ack_wmd.wh_object_cookie,
2245 hdr->msg.put.match_bits);
2246 CWARN(" Length %d, offset %d, hdr data %#llx\n",
2247 hdr->payload_length, hdr->msg.put.offset,
2248 hdr->msg.put.hdr_data);
2252 CWARN(" Ptl index %d, return md %#llx.%#llx, "
2253 "match bits %llu\n", hdr->msg.get.ptl_index,
2254 hdr->msg.get.return_wmd.wh_interface_cookie,
2255 hdr->msg.get.return_wmd.wh_object_cookie,
2256 hdr->msg.get.match_bits);
2257 CWARN(" Length %d, src offset %d\n",
2258 hdr->msg.get.sink_length,
2259 hdr->msg.get.src_offset);
2263 CWARN(" dst md %#llx.%#llx, "
2264 "manipulated length %d\n",
2265 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2266 hdr->msg.ack.dst_wmd.wh_object_cookie,
2267 hdr->msg.ack.mlength);
2270 case LNET_MSG_REPLY:
2271 CWARN(" dst md %#llx.%#llx, "
2273 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2274 hdr->msg.reply.dst_wmd.wh_object_cookie,
2275 hdr->payload_length);
2281 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
2282 void *private, int rdma_req)
2287 struct lnet_msg *msg;
2288 lnet_pid_t dest_pid;
2289 lnet_nid_t dest_nid;
2291 struct lnet_peer_ni *lpni;
2292 __u32 payload_length;
2295 LASSERT (!in_interrupt ());
2297 type = le32_to_cpu(hdr->type);
2298 src_nid = le64_to_cpu(hdr->src_nid);
2299 dest_nid = le64_to_cpu(hdr->dest_nid);
2300 dest_pid = le32_to_cpu(hdr->dest_pid);
2301 payload_length = le32_to_cpu(hdr->payload_length);
2303 for_me = (ni->ni_nid == dest_nid);
2304 cpt = lnet_cpt_of_nid(from_nid, ni);
2309 if (payload_length > 0) {
2310 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2311 libcfs_nid2str(from_nid),
2312 libcfs_nid2str(src_nid),
2313 lnet_msgtyp2str(type), payload_length);
2319 case LNET_MSG_REPLY:
2320 if (payload_length >
2321 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2322 CERROR("%s, src %s: bad %s payload %d "
2323 "(%d max expected)\n",
2324 libcfs_nid2str(from_nid),
2325 libcfs_nid2str(src_nid),
2326 lnet_msgtyp2str(type),
2328 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2334 CERROR("%s, src %s: Bad message type 0x%x\n",
2335 libcfs_nid2str(from_nid),
2336 libcfs_nid2str(src_nid), type);
2340 if (the_lnet.ln_routing &&
2341 ni->ni_last_alive != ktime_get_real_seconds()) {
2342 /* NB: so far here is the only place to set NI status to "up */
2344 ni->ni_last_alive = ktime_get_real_seconds();
2345 if (ni->ni_status != NULL &&
2346 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
2347 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
2351 /* Regard a bad destination NID as a protocol error. Senders should
2352 * know what they're doing; if they don't they're misconfigured, buggy
2353 * or malicious so we chop them off at the knees :) */
2356 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2357 /* should have gone direct */
2358 CERROR("%s, src %s: Bad dest nid %s "
2359 "(should have been sent direct)\n",
2360 libcfs_nid2str(from_nid),
2361 libcfs_nid2str(src_nid),
2362 libcfs_nid2str(dest_nid));
2366 if (lnet_islocalnid(dest_nid)) {
2367 /* dest is another local NI; sender should have used
2368 * this node's NID on its own network */
2369 CERROR("%s, src %s: Bad dest nid %s "
2370 "(it's my nid but on a different network)\n",
2371 libcfs_nid2str(from_nid),
2372 libcfs_nid2str(src_nid),
2373 libcfs_nid2str(dest_nid));
2377 if (rdma_req && type == LNET_MSG_GET) {
2378 CERROR("%s, src %s: Bad optimized GET for %s "
2379 "(final destination must be me)\n",
2380 libcfs_nid2str(from_nid),
2381 libcfs_nid2str(src_nid),
2382 libcfs_nid2str(dest_nid));
2386 if (!the_lnet.ln_routing) {
2387 CERROR("%s, src %s: Dropping message for %s "
2388 "(routing not enabled)\n",
2389 libcfs_nid2str(from_nid),
2390 libcfs_nid2str(src_nid),
2391 libcfs_nid2str(dest_nid));
2396 /* Message looks OK; we're not going to return an error, so we MUST
2397 * call back lnd_recv() come what may... */
2399 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2400 fail_peer(src_nid, 0)) { /* shall we now? */
2401 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2402 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2403 lnet_msgtyp2str(type));
2407 if (!list_empty(&the_lnet.ln_drop_rules) &&
2408 lnet_drop_rule_match(hdr)) {
2409 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
2410 "silent message loss\n",
2411 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2412 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
2417 msg = lnet_msg_alloc();
2419 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2420 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2421 lnet_msgtyp2str(type));
2425 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
2426 * pointers NULL etc */
2428 msg->msg_type = type;
2429 msg->msg_private = private;
2430 msg->msg_receiving = 1;
2431 msg->msg_rdma_get = rdma_req;
2432 msg->msg_len = msg->msg_wanted = payload_length;
2433 msg->msg_offset = 0;
2434 msg->msg_hdr = *hdr;
2435 /* for building message event */
2436 msg->msg_from = from_nid;
2438 msg->msg_target.pid = dest_pid;
2439 msg->msg_target.nid = dest_nid;
2440 msg->msg_routing = 1;
2443 /* convert common msg->hdr fields to host byteorder */
2444 msg->msg_hdr.type = type;
2445 msg->msg_hdr.src_nid = src_nid;
2446 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2447 msg->msg_hdr.dest_nid = dest_nid;
2448 msg->msg_hdr.dest_pid = dest_pid;
2449 msg->msg_hdr.payload_length = payload_length;
2451 /* Multi-Rail: Primary NID of source. */
2452 msg->msg_initiator = lnet_peer_primary_nid(src_nid);
2455 lpni = lnet_nid2peerni_locked(from_nid, cpt);
2457 lnet_net_unlock(cpt);
2458 CERROR("%s, src %s: Dropping %s "
2459 "(error %ld looking up sender)\n",
2460 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2461 lnet_msgtyp2str(type), PTR_ERR(lpni));
2463 if (rc == -ESHUTDOWN)
2464 /* We are shutting down. Don't do anything more */
2468 msg->msg_rxpeer = lpni;
2470 lnet_ni_addref_locked(ni, cpt);
2472 if (lnet_isrouter(msg->msg_rxpeer)) {
2473 lnet_peer_set_alive(msg->msg_rxpeer);
2474 if (avoid_asym_router_failure &&
2475 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
2476 /* received a remote message from router, update
2477 * remote NI status on this router.
2478 * NB: multi-hop routed message will be ignored.
2480 lnet_router_ni_update_locked(msg->msg_rxpeer,
2481 LNET_NIDNET(src_nid));
2485 lnet_msg_commit(msg, cpt);
2487 /* message delay simulation */
2488 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
2489 lnet_delay_rule_match_locked(hdr, msg))) {
2490 lnet_net_unlock(cpt);
2495 rc = lnet_parse_forward_locked(ni, msg);
2496 lnet_net_unlock(cpt);
2501 if (rc == LNET_CREDIT_OK) {
2502 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2503 0, payload_length, payload_length);
2508 lnet_net_unlock(cpt);
2510 rc = lnet_parse_local(ni, msg);
2516 LASSERT(msg->msg_md == NULL);
2517 lnet_finalize(ni, msg, rc);
2520 lnet_drop_message(ni, cpt, private, payload_length);
2523 EXPORT_SYMBOL(lnet_parse);
2526 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
2528 while (!list_empty(head)) {
2529 lnet_process_id_t id = {0};
2532 msg = list_entry(head->next, lnet_msg_t, msg_list);
2533 list_del(&msg->msg_list);
2535 id.nid = msg->msg_hdr.src_nid;
2536 id.pid = msg->msg_hdr.src_pid;
2538 LASSERT(msg->msg_md == NULL);
2539 LASSERT(msg->msg_rx_delayed);
2540 LASSERT(msg->msg_rxpeer != NULL);
2541 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2543 CWARN("Dropping delayed PUT from %s portal %d match %llu"
2544 " offset %d length %d: %s\n",
2546 msg->msg_hdr.msg.put.ptl_index,
2547 msg->msg_hdr.msg.put.match_bits,
2548 msg->msg_hdr.msg.put.offset,
2549 msg->msg_hdr.payload_length, reason);
2551 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2552 * called lnet_drop_message(), so I just hang onto msg as well
2553 * until that's done */
2555 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
2556 msg->msg_private, msg->msg_len);
2558 * NB: message will not generate event because w/o attached MD,
2559 * but we still should give error code so lnet_msg_decommit()
2560 * can skip counters operations and other checks.
2562 lnet_finalize(msg->msg_rxni, msg, -ENOENT);
2567 lnet_recv_delayed_msg_list(struct list_head *head)
2569 while (!list_empty(head)) {
2571 lnet_process_id_t id;
2573 msg = list_entry(head->next, lnet_msg_t, msg_list);
2574 list_del(&msg->msg_list);
2576 /* md won't disappear under me, since each msg
2577 * holds a ref on it */
2579 id.nid = msg->msg_hdr.src_nid;
2580 id.pid = msg->msg_hdr.src_pid;
2582 LASSERT(msg->msg_rx_delayed);
2583 LASSERT(msg->msg_md != NULL);
2584 LASSERT(msg->msg_rxpeer != NULL);
2585 LASSERT(msg->msg_rxni != NULL);
2586 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2588 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2589 "match %llu offset %d length %d.\n",
2590 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2591 msg->msg_hdr.msg.put.match_bits,
2592 msg->msg_hdr.msg.put.offset,
2593 msg->msg_hdr.payload_length);
2595 lnet_recv_put(msg->msg_rxni, msg);
2600 * Initiate an asynchronous PUT operation.
2602 * There are several events associated with a PUT: completion of the send on
2603 * the initiator node (LNET_EVENT_SEND), and when the send completes
2604 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2605 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2606 * used at the target node to indicate the completion of incoming data
2609 * The local events will be logged in the EQ associated with the MD pointed to
2610 * by \a mdh handle. Using a MD without an associated EQ results in these
2611 * events being discarded. In this case, the caller must have another
2612 * mechanism (e.g., a higher level protocol) for determining when it is safe
2613 * to modify the memory region associated with the MD.
2615 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2616 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2618 * \param self Indicates the NID of a local interface through which to send
2619 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2620 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2621 * must be "free floating" (See LNetMDBind()).
2622 * \param ack Controls whether an acknowledgment is requested.
2623 * Acknowledgments are only sent when they are requested by the initiating
2624 * process and the target MD enables them.
2625 * \param target A process identifier for the target process.
2626 * \param portal The index in the \a target's portal table.
2627 * \param match_bits The match bits to use for MD selection at the target
2629 * \param offset The offset into the target MD (only used when the target
2630 * MD has the LNET_MD_MANAGE_REMOTE option set).
2631 * \param hdr_data 64 bits of user data that can be included in the message
2632 * header. This data is written to an event queue entry at the target if an
2633 * EQ is present on the matching MD.
2635 * \retval 0 Success, and only in this case events will be generated
2636 * and logged to EQ (if it exists).
2637 * \retval -EIO Simulated failure.
2638 * \retval -ENOMEM Memory allocation failure.
2639 * \retval -ENOENT Invalid MD object.
2641 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2644 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2645 lnet_process_id_t target, unsigned int portal,
2646 __u64 match_bits, unsigned int offset,
2649 struct lnet_msg *msg;
2650 struct lnet_libmd *md;
2654 LASSERT(the_lnet.ln_refcount > 0);
2656 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2657 fail_peer(target.nid, 1)) { /* shall we now? */
2658 CERROR("Dropping PUT to %s: simulated failure\n",
2659 libcfs_id2str(target));
2663 msg = lnet_msg_alloc();
2665 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2666 libcfs_id2str(target));
2669 msg->msg_vmflush = !!memory_pressure_get();
2671 cpt = lnet_cpt_of_cookie(mdh.cookie);
2674 md = lnet_handle2md(&mdh);
2675 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2676 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
2677 match_bits, portal, libcfs_id2str(target),
2678 md == NULL ? -1 : md->md_threshold);
2679 if (md != NULL && md->md_me != NULL)
2680 CERROR("Source MD also attached to portal %d\n",
2681 md->md_me->me_portal);
2682 lnet_res_unlock(cpt);
2688 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2690 lnet_msg_attach_md(msg, md, 0, 0);
2692 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2694 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2695 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2696 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2697 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2699 /* NB handles only looked up by creator (no flips) */
2700 if (ack == LNET_ACK_REQ) {
2701 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2702 the_lnet.ln_interface_cookie;
2703 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2704 md->md_lh.lh_cookie;
2706 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2707 LNET_WIRE_HANDLE_COOKIE_NONE;
2708 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2709 LNET_WIRE_HANDLE_COOKIE_NONE;
2712 lnet_res_unlock(cpt);
2714 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2716 rc = lnet_send(self, msg, LNET_NID_ANY);
2718 CNETERR("Error sending PUT to %s: %d\n",
2719 libcfs_id2str(target), rc);
2720 lnet_finalize(NULL, msg, rc);
2723 /* completion will be signalled by an event */
2726 EXPORT_SYMBOL(LNetPut);
2729 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2731 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2732 * returns a msg for the LND to pass to lnet_finalize() when the sink
2733 * data has been received.
2735 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2736 * lnet_finalize() is called on it, so the LND must call this first */
2738 struct lnet_msg *msg = lnet_msg_alloc();
2739 struct lnet_libmd *getmd = getmsg->msg_md;
2740 lnet_process_id_t peer_id = getmsg->msg_target;
2743 LASSERT(!getmsg->msg_target_is_router);
2744 LASSERT(!getmsg->msg_routing);
2747 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2748 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2752 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2755 LASSERT(getmd->md_refcount > 0);
2757 if (getmd->md_threshold == 0) {
2758 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
2759 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2761 lnet_res_unlock(cpt);
2765 LASSERT(getmd->md_offset == 0);
2767 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2768 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2770 /* setup information for lnet_build_msg_event */
2771 msg->msg_initiator = lnet_peer_primary_nid(peer_id.nid);
2772 /* Cheaper: msg->msg_initiator = getmsg->msg_txpeer->lp_nid; */
2773 msg->msg_from = peer_id.nid;
2774 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2775 msg->msg_hdr.src_nid = peer_id.nid;
2776 msg->msg_hdr.payload_length = getmd->md_length;
2777 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2779 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2780 lnet_res_unlock(cpt);
2782 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
2785 lnet_msg_commit(msg, cpt);
2786 lnet_net_unlock(cpt);
2788 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2793 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
2796 the_lnet.ln_counters[cpt]->drop_count++;
2797 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2798 lnet_net_unlock(cpt);
2805 EXPORT_SYMBOL(lnet_create_reply_msg);
2808 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2810 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2811 * completed and I know it. */
2812 LASSERT(reply != NULL);
2813 LASSERT(reply->msg_type == LNET_MSG_GET);
2814 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
2816 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2817 * the end of my buffer, I might as well be dead. */
2818 LASSERT(len <= reply->msg_ev.mlength);
2820 reply->msg_ev.mlength = len;
2822 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2825 * Initiate an asynchronous GET operation.
2827 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2828 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2829 * the target node in the REPLY has been written to local MD.
2831 * On the target node, an LNET_EVENT_GET is logged when the GET request
2832 * arrives and is accepted into a MD.
2834 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2835 * \param mdh A handle for the MD that describes the memory into which the
2836 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2838 * \retval 0 Success, and only in this case events will be generated
2839 * and logged to EQ (if it exists) of the MD.
2840 * \retval -EIO Simulated failure.
2841 * \retval -ENOMEM Memory allocation failure.
2842 * \retval -ENOENT Invalid MD object.
2845 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2846 lnet_process_id_t target, unsigned int portal,
2847 __u64 match_bits, unsigned int offset)
2849 struct lnet_msg *msg;
2850 struct lnet_libmd *md;
2854 LASSERT(the_lnet.ln_refcount > 0);
2856 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2857 fail_peer(target.nid, 1)) /* shall we now? */
2859 CERROR("Dropping GET to %s: simulated failure\n",
2860 libcfs_id2str(target));
2864 msg = lnet_msg_alloc();
2866 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2867 libcfs_id2str(target));
2871 cpt = lnet_cpt_of_cookie(mdh.cookie);
2874 md = lnet_handle2md(&mdh);
2875 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2876 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
2877 match_bits, portal, libcfs_id2str(target),
2878 md == NULL ? -1 : md->md_threshold);
2879 if (md != NULL && md->md_me != NULL)
2880 CERROR("REPLY MD also attached to portal %d\n",
2881 md->md_me->me_portal);
2883 lnet_res_unlock(cpt);
2889 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2891 lnet_msg_attach_md(msg, md, 0, 0);
2893 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2895 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2896 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2897 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2898 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2900 /* NB handles only looked up by creator (no flips) */
2901 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2902 the_lnet.ln_interface_cookie;
2903 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2904 md->md_lh.lh_cookie;
2906 lnet_res_unlock(cpt);
2908 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2910 rc = lnet_send(self, msg, LNET_NID_ANY);
2912 CNETERR("Error sending GET to %s: %d\n",
2913 libcfs_id2str(target), rc);
2914 lnet_finalize(NULL, msg, rc);
2917 /* completion will be signalled by an event */
2920 EXPORT_SYMBOL(LNetGet);
2923 * Calculate distance to node at \a dstnid.
2925 * \param dstnid Target NID.
2926 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2928 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2931 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2932 * local_nid_dist_zero is set, which is the default.
2933 * \retval positives Distance to target NID, i.e. number of hops plus one.
2934 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2937 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2939 struct list_head *e;
2940 struct lnet_ni *ni = NULL;
2941 lnet_remotenet_t *rnet;
2942 __u32 dstnet = LNET_NIDNET(dstnid);
2946 struct list_head *rn_list;
2948 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2949 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2950 * keep order 0 free for 0@lo and order 1 free for a local NID
2953 LASSERT(the_lnet.ln_refcount > 0);
2955 cpt = lnet_net_lock_current();
2957 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
2958 if (ni->ni_nid == dstnid) {
2959 if (srcnidp != NULL)
2961 if (orderp != NULL) {
2962 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2967 lnet_net_unlock(cpt);
2969 return local_nid_dist_zero ? 0 : 1;
2972 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2973 /* Check if ni was originally created in
2974 * current net namespace.
2975 * If not, assign order above 0xffff0000,
2976 * to make this ni not a priority. */
2977 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
2978 order += 0xffff0000;
2980 if (srcnidp != NULL)
2981 *srcnidp = ni->ni_nid;
2984 lnet_net_unlock(cpt);
2991 rn_list = lnet_net2rnethash(dstnet);
2992 list_for_each(e, rn_list) {
2993 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2995 if (rnet->lrn_net == dstnet) {
2996 lnet_route_t *route;
2997 lnet_route_t *shortest = NULL;
2998 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
3001 LASSERT(!list_empty(&rnet->lrn_routes));
3003 list_for_each_entry(route, &rnet->lrn_routes,
3005 route_hops = route->lr_hops;
3006 if (route_hops == LNET_UNDEFINED_HOPS)
3008 if (shortest == NULL ||
3009 route_hops < shortest_hops) {
3011 shortest_hops = route_hops;
3015 LASSERT(shortest != NULL);
3016 hops = shortest_hops;
3017 if (srcnidp != NULL) {
3018 ni = lnet_get_next_ni_locked(
3019 shortest->lr_gateway->lpni_net,
3021 *srcnidp = ni->ni_nid;
3025 lnet_net_unlock(cpt);
3031 lnet_net_unlock(cpt);
3032 return -EHOSTUNREACH;
3034 EXPORT_SYMBOL(LNetDist);