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 */
677 /* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */
679 lnet_ni_query_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp)
681 cfs_time_t last_alive = 0;
683 LASSERT(lnet_peer_aliveness_enabled(lp));
684 LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
686 lnet_net_unlock(lp->lpni_cpt);
687 (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
688 lnet_net_lock(lp->lpni_cpt);
690 lp->lpni_last_query = cfs_time_current();
692 if (last_alive != 0) /* NI has updated timestamp */
693 lp->lpni_last_alive = last_alive;
696 /* NB: always called with lnet_net_lock held */
698 lnet_peer_is_alive (struct lnet_peer_ni *lp, cfs_time_t now)
703 LASSERT (lnet_peer_aliveness_enabled(lp));
706 * Trust lnet_notify() if it has more recent aliveness news, but
707 * ignore the initial assumed death (see lnet_peers_start_down()).
709 if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
710 cfs_time_aftereq(lp->lpni_timestamp, lp->lpni_last_alive))
714 cfs_time_add(lp->lpni_last_alive,
715 cfs_time_seconds(lp->lpni_net->net_tunables.
717 alive = cfs_time_after(deadline, now);
720 * Update obsolete lp_alive except for routers assumed to be dead
721 * initially, because router checker would update aliveness in this
722 * case, and moreover lpni_last_alive at peer creation is assumed.
724 if (alive && !lp->lpni_alive &&
725 !(lnet_isrouter(lp) && lp->lpni_alive_count == 0))
726 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
732 /* NB: returns 1 when alive, 0 when dead, negative when error;
733 * may drop the lnet_net_lock */
735 lnet_peer_alive_locked (struct lnet_ni *ni, struct lnet_peer_ni *lp)
737 cfs_time_t now = cfs_time_current();
739 if (!lnet_peer_aliveness_enabled(lp))
742 if (lnet_peer_is_alive(lp, now))
746 * Peer appears dead, but we should avoid frequent NI queries (at
747 * most once per lnet_queryinterval seconds).
749 if (lp->lpni_last_query != 0) {
750 static const int lnet_queryinterval = 1;
752 cfs_time_t next_query =
753 cfs_time_add(lp->lpni_last_query,
754 cfs_time_seconds(lnet_queryinterval));
756 if (cfs_time_before(now, next_query)) {
758 CWARN("Unexpected aliveness of peer %s: "
760 libcfs_nid2str(lp->lpni_nid),
761 (int)now, (int)next_query,
763 lp->lpni_net->net_tunables.lct_peer_timeout);
768 /* query NI for latest aliveness news */
769 lnet_ni_query_locked(ni, lp);
771 if (lnet_peer_is_alive(lp, now))
774 lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
779 * \param msg The message to be sent.
780 * \param do_send True if lnet_ni_send() should be called in this function.
781 * lnet_send() is going to lnet_net_unlock immediately after this, so
782 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
784 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
785 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
786 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
787 * \retval -ECANCELED If the MD of the message has been unlinked.
790 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
792 struct lnet_peer_ni *lp = msg->msg_txpeer;
793 struct lnet_ni *ni = msg->msg_txni;
794 int cpt = msg->msg_tx_cpt;
795 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
797 /* non-lnet_send() callers have checked before */
798 LASSERT(!do_send || msg->msg_tx_delayed);
799 LASSERT(!msg->msg_receiving);
800 LASSERT(msg->msg_tx_committed);
802 /* NB 'lp' is always the next hop */
803 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
804 lnet_peer_alive_locked(ni, lp) == 0) {
805 the_lnet.ln_counters[cpt]->drop_count++;
806 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
807 lnet_net_unlock(cpt);
809 atomic_inc(&msg->msg_txpeer->lpni_stats.drop_count);
811 atomic_inc(&msg->msg_txni->ni_stats.drop_count);
813 CNETERR("Dropping message for %s: peer not alive\n",
814 libcfs_id2str(msg->msg_target));
816 lnet_finalize(ni, msg, -EHOSTUNREACH);
819 return -EHOSTUNREACH;
822 if (msg->msg_md != NULL &&
823 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
824 lnet_net_unlock(cpt);
826 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
827 "called on the MD/ME.\n",
828 libcfs_id2str(msg->msg_target));
830 lnet_finalize(ni, msg, -ECANCELED);
836 if (!msg->msg_peertxcredit) {
837 LASSERT((lp->lpni_txcredits < 0) ==
838 !list_empty(&lp->lpni_txq));
840 msg->msg_peertxcredit = 1;
841 lp->lpni_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
842 lp->lpni_txcredits--;
844 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
845 lp->lpni_mintxcredits = lp->lpni_txcredits;
847 if (lp->lpni_txcredits < 0) {
848 msg->msg_tx_delayed = 1;
849 list_add_tail(&msg->msg_list, &lp->lpni_txq);
850 return LNET_CREDIT_WAIT;
854 if (!msg->msg_txcredit) {
855 LASSERT((tq->tq_credits < 0) ==
856 !list_empty(&tq->tq_delayed));
858 msg->msg_txcredit = 1;
861 if (tq->tq_credits < tq->tq_credits_min)
862 tq->tq_credits_min = tq->tq_credits;
864 if (tq->tq_credits < 0) {
865 msg->msg_tx_delayed = 1;
866 list_add_tail(&msg->msg_list, &tq->tq_delayed);
867 return LNET_CREDIT_WAIT;
872 lnet_net_unlock(cpt);
873 lnet_ni_send(ni, msg);
876 return LNET_CREDIT_OK;
880 static lnet_rtrbufpool_t *
881 lnet_msg2bufpool(lnet_msg_t *msg)
883 lnet_rtrbufpool_t *rbp;
886 LASSERT(msg->msg_rx_committed);
888 cpt = msg->msg_rx_cpt;
889 rbp = &the_lnet.ln_rtrpools[cpt][0];
891 LASSERT(msg->msg_len <= LNET_MTU);
892 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
894 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
901 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
903 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
904 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
905 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
906 * received or OK to receive */
907 struct lnet_peer_ni *lp = msg->msg_rxpeer;
908 lnet_rtrbufpool_t *rbp;
911 LASSERT (msg->msg_iov == NULL);
912 LASSERT (msg->msg_kiov == NULL);
913 LASSERT (msg->msg_niov == 0);
914 LASSERT (msg->msg_routing);
915 LASSERT (msg->msg_receiving);
916 LASSERT (!msg->msg_sending);
918 /* non-lnet_parse callers only receive delayed messages */
919 LASSERT(!do_recv || msg->msg_rx_delayed);
921 if (!msg->msg_peerrtrcredit) {
922 LASSERT((lp->lpni_rtrcredits < 0) ==
923 !list_empty(&lp->lpni_rtrq));
925 msg->msg_peerrtrcredit = 1;
926 lp->lpni_rtrcredits--;
927 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
928 lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
930 if (lp->lpni_rtrcredits < 0) {
931 /* must have checked eager_recv before here */
932 LASSERT(msg->msg_rx_ready_delay);
933 msg->msg_rx_delayed = 1;
934 list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
935 return LNET_CREDIT_WAIT;
939 rbp = lnet_msg2bufpool(msg);
941 if (!msg->msg_rtrcredit) {
942 msg->msg_rtrcredit = 1;
944 if (rbp->rbp_credits < rbp->rbp_mincredits)
945 rbp->rbp_mincredits = rbp->rbp_credits;
947 if (rbp->rbp_credits < 0) {
948 /* must have checked eager_recv before here */
949 LASSERT(msg->msg_rx_ready_delay);
950 msg->msg_rx_delayed = 1;
951 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
952 return LNET_CREDIT_WAIT;
956 LASSERT(!list_empty(&rbp->rbp_bufs));
957 rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
958 list_del(&rb->rb_list);
960 msg->msg_niov = rbp->rbp_npages;
961 msg->msg_kiov = &rb->rb_kiov[0];
964 int cpt = msg->msg_rx_cpt;
966 lnet_net_unlock(cpt);
967 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
968 0, msg->msg_len, msg->msg_len);
971 return LNET_CREDIT_OK;
975 lnet_return_tx_credits_locked(lnet_msg_t *msg)
977 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
978 struct lnet_ni *txni = msg->msg_txni;
981 if (msg->msg_txcredit) {
982 struct lnet_ni *ni = msg->msg_txni;
983 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
985 /* give back NI txcredits */
986 msg->msg_txcredit = 0;
988 LASSERT((tq->tq_credits < 0) ==
989 !list_empty(&tq->tq_delayed));
992 if (tq->tq_credits <= 0) {
993 msg2 = list_entry(tq->tq_delayed.next,
994 lnet_msg_t, msg_list);
995 list_del(&msg2->msg_list);
997 LASSERT(msg2->msg_txni == ni);
998 LASSERT(msg2->msg_tx_delayed);
1000 (void) lnet_post_send_locked(msg2, 1);
1004 if (msg->msg_peertxcredit) {
1005 /* give back peer txcredits */
1006 msg->msg_peertxcredit = 0;
1008 LASSERT((txpeer->lpni_txcredits < 0) ==
1009 !list_empty(&txpeer->lpni_txq));
1011 txpeer->lpni_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1012 LASSERT (txpeer->lpni_txqnob >= 0);
1014 txpeer->lpni_txcredits++;
1015 if (txpeer->lpni_txcredits <= 0) {
1016 msg2 = list_entry(txpeer->lpni_txq.next,
1017 lnet_msg_t, msg_list);
1018 list_del(&msg2->msg_list);
1020 LASSERT(msg2->msg_txpeer == txpeer);
1021 LASSERT(msg2->msg_tx_delayed);
1023 (void) lnet_post_send_locked(msg2, 1);
1028 msg->msg_txni = NULL;
1029 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1032 if (txpeer != NULL) {
1035 * Once the patch for the health comes in we need to set
1036 * the health of the peer ni to bad when we fail to send
1038 * int status = msg->msg_ev.status;
1040 * lnet_set_peer_ni_health_locked(txpeer, false)
1042 msg->msg_txpeer = NULL;
1043 lnet_peer_ni_decref_locked(txpeer);
1048 lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp)
1052 if (list_empty(&rbp->rbp_msgs))
1054 msg = list_entry(rbp->rbp_msgs.next,
1055 lnet_msg_t, msg_list);
1056 list_del(&msg->msg_list);
1058 (void)lnet_post_routed_recv_locked(msg, 1);
1062 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1066 struct list_head drop;
1068 INIT_LIST_HEAD(&drop);
1070 list_splice_init(list, &drop);
1072 lnet_net_unlock(cpt);
1074 list_for_each_entry_safe(msg, tmp, &drop, msg_list) {
1075 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1076 0, 0, 0, msg->msg_hdr.payload_length);
1077 list_del_init(&msg->msg_list);
1078 lnet_finalize(NULL, msg, -ECANCELED);
1085 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1087 struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1088 struct lnet_ni *rxni = msg->msg_rxni;
1091 if (msg->msg_rtrcredit) {
1092 /* give back global router credits */
1094 lnet_rtrbufpool_t *rbp;
1096 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1097 * there until it gets one allocated, or aborts the wait
1099 LASSERT(msg->msg_kiov != NULL);
1101 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1104 msg->msg_kiov = NULL;
1105 msg->msg_rtrcredit = 0;
1107 LASSERT(rbp == lnet_msg2bufpool(msg));
1109 LASSERT((rbp->rbp_credits > 0) ==
1110 !list_empty(&rbp->rbp_bufs));
1112 /* If routing is now turned off, we just drop this buffer and
1113 * don't bother trying to return credits. */
1114 if (!the_lnet.ln_routing) {
1115 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1119 /* It is possible that a user has lowered the desired number of
1120 * buffers in this pool. Make sure we never put back
1121 * more buffers than the stated number. */
1122 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1123 /* Discard this buffer so we don't have too
1125 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1126 rbp->rbp_nbuffers--;
1128 list_add(&rb->rb_list, &rbp->rbp_bufs);
1130 if (rbp->rbp_credits <= 0)
1131 lnet_schedule_blocked_locked(rbp);
1136 if (msg->msg_peerrtrcredit) {
1137 /* give back peer router credits */
1138 msg->msg_peerrtrcredit = 0;
1140 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1141 !list_empty(&rxpeer->lpni_rtrq));
1143 rxpeer->lpni_rtrcredits++;
1145 /* drop all messages which are queued to be routed on that
1147 if (!the_lnet.ln_routing) {
1148 lnet_drop_routed_msgs_locked(&rxpeer->lpni_rtrq,
1150 } else if (rxpeer->lpni_rtrcredits <= 0) {
1151 msg2 = list_entry(rxpeer->lpni_rtrq.next,
1152 lnet_msg_t, msg_list);
1153 list_del(&msg2->msg_list);
1155 (void) lnet_post_routed_recv_locked(msg2, 1);
1159 msg->msg_rxni = NULL;
1160 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1162 if (rxpeer != NULL) {
1163 msg->msg_rxpeer = NULL;
1164 lnet_peer_ni_decref_locked(rxpeer);
1169 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1171 if (p1->lpni_txqnob < p2->lpni_txqnob)
1174 if (p1->lpni_txqnob > p2->lpni_txqnob)
1177 if (p1->lpni_txcredits > p2->lpni_txcredits)
1180 if (p1->lpni_txcredits < p2->lpni_txcredits)
1187 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1189 struct lnet_peer_ni *p1 = r1->lr_gateway;
1190 struct lnet_peer_ni *p2 = r2->lr_gateway;
1191 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1192 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1195 if (r1->lr_priority < r2->lr_priority)
1198 if (r1->lr_priority > r2->lr_priority)
1201 if (r1_hops < r2_hops)
1204 if (r1_hops > r2_hops)
1207 rc = lnet_compare_peers(p1, p2);
1211 if (r1->lr_seq - r2->lr_seq <= 0)
1217 static struct lnet_peer_ni *
1218 lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
1221 lnet_remotenet_t *rnet;
1222 lnet_route_t *route;
1223 lnet_route_t *best_route;
1224 lnet_route_t *last_route;
1225 struct lnet_peer_ni *lpni_best;
1226 struct lnet_peer_ni *lp;
1229 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1230 * rtr_nid nid, otherwise find the best gateway I can use */
1232 rnet = lnet_find_rnet_locked(LNET_NIDNET(target));
1237 best_route = last_route = NULL;
1238 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1239 lp = route->lr_gateway;
1241 if (!lnet_is_route_alive(route))
1244 if (net != NULL && lp->lpni_net != net)
1247 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1250 if (lpni_best == NULL) {
1251 best_route = last_route = route;
1256 /* no protection on below fields, but it's harmless */
1257 if (last_route->lr_seq - route->lr_seq < 0)
1260 rc = lnet_compare_routes(route, best_route);
1268 /* set sequence number on the best router to the latest sequence + 1
1269 * so we can round-robin all routers, it's race and inaccurate but
1270 * harmless and functional */
1271 if (best_route != NULL)
1272 best_route->lr_seq = last_route->lr_seq + 1;
1277 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
1278 struct lnet_msg *msg, lnet_nid_t rtr_nid, bool *lo_sent)
1280 struct lnet_ni *best_ni = NULL;
1281 struct lnet_peer_ni *best_lpni = NULL;
1282 struct lnet_peer_ni *net_gw = NULL;
1283 struct lnet_peer_ni *best_gw = NULL;
1284 struct lnet_peer_ni *lpni;
1285 struct lnet_peer *peer = NULL;
1286 struct lnet_peer_net *peer_net;
1287 struct lnet_net *local_net;
1288 struct lnet_ni *ni = NULL;
1290 bool routing = false;
1291 bool ni_is_pref = false;
1292 bool preferred = false;
1293 int best_credits = 0;
1295 int best_lpni_credits = INT_MIN;
1297 int shortest_distance = INT_MAX;
1299 bool found_ir = false;
1303 * get an initial CPT to use for locking. The idea here is not to
1304 * serialize the calls to select_pathway, so that as many
1305 * operations can run concurrently as possible. To do that we use
1306 * the CPT where this call is being executed. Later on when we
1307 * determine the CPT to use in lnet_message_commit, we switch the
1308 * lock and check if there was any configuration changes, if none,
1309 * then we proceed, if there is, then we'll need to update the cpt
1310 * and redo the operation.
1312 cpt = lnet_net_lock_current();
1318 shortest_distance = INT_MAX;
1321 if (the_lnet.ln_shutdown) {
1322 lnet_net_unlock(cpt);
1326 if (msg->msg_md != NULL)
1327 /* get the cpt of the MD, used during NUMA based selection */
1328 md_cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
1330 md_cpt = CFS_CPT_ANY;
1333 * initialize the variables which could be reused if we go to
1337 seq = lnet_get_dlc_seq_locked();
1339 peer = lnet_find_or_create_peer_locked(dst_nid, cpt);
1341 lnet_net_unlock(cpt);
1342 return PTR_ERR(peer);
1345 /* If peer is not healthy then can not send anything to it */
1346 if (!lnet_is_peer_healthy_locked(peer)) {
1347 lnet_net_unlock(cpt);
1348 return -EHOSTUNREACH;
1351 if (!peer->lp_multi_rail && lnet_get_num_peer_nis(peer) > 1) {
1352 CERROR("peer %s is declared to be non MR capable, "
1353 "yet configured with more than one NID\n",
1354 libcfs_nid2str(dst_nid));
1359 * STEP 1: first jab at determineing best_ni
1360 * if src_nid is explicitly specified, then best_ni is already
1361 * pre-determiend for us. Otherwise we need to select the best
1362 * one to use later on
1364 if (src_nid != LNET_NID_ANY) {
1365 best_ni = lnet_nid2ni_locked(src_nid, cpt);
1367 lnet_net_unlock(cpt);
1368 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1369 "local nid\n", libcfs_nid2str(dst_nid),
1370 libcfs_nid2str(src_nid));
1374 if (best_ni->ni_net->net_id != LNET_NIDNET(dst_nid)) {
1375 lnet_net_unlock(cpt);
1376 LCONSOLE_WARN("No route to %s via from %s\n",
1377 libcfs_nid2str(dst_nid),
1378 libcfs_nid2str(src_nid));
1387 * Decide whether we need to route to peer_ni.
1388 * Get the local net that I need to be on to be able to directly
1389 * send to that peer.
1391 * a. Find the peer which the dst_nid belongs to.
1392 * b. Iterate through each of the peer_nets/nis to decide
1393 * the best peer/local_ni pair to use
1395 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_on_peer_list) {
1396 if (!lnet_is_peer_net_healthy_locked(peer_net))
1399 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1402 * go through each peer_ni on that peer_net and
1403 * determine the best possible gw to go through
1405 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
1406 lpni_on_peer_net_list) {
1407 net_gw = lnet_find_route_locked(NULL,
1412 * if no route is found for that network then
1413 * move onto the next peer_ni in the peer
1422 rc = lnet_compare_peers(net_gw,
1434 local_net = lnet_get_net_locked
1435 (LNET_NIDNET(best_gw->lpni_nid));
1442 /* no routable net found go on to a different net */
1447 * Iterate through the NIs in this local Net and select
1448 * the NI to send from. The selection is determined by
1449 * these 3 criterion in the following priority:
1451 * 2. NI available credits
1454 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1455 if (!lnet_is_ni_healthy_locked(ni))
1459 * calculate the distance from the cpt on which
1460 * the message memory is allocated to the CPT of
1461 * the NI's physical device
1463 distance = cfs_cpt_distance(lnet_cpt_table(),
1468 * If we already have a closer NI within the NUMA
1469 * range provided, then there is no need to
1470 * consider the current NI. Move on to the next
1473 if (distance > shortest_distance &&
1474 distance > lnet_get_numa_range())
1477 if (distance < shortest_distance &&
1478 distance > lnet_get_numa_range()) {
1480 * The current NI is the closest one that we
1481 * have found, even though it's not in the
1482 * NUMA range specified. This occurs if
1483 * the NUMA range is less than the least
1484 * of the distances in the system.
1485 * In effect NUMA range consideration is
1488 shortest_distance = distance;
1489 } else if ((distance <= shortest_distance &&
1490 distance < lnet_get_numa_range()) ||
1491 distance == shortest_distance) {
1493 * This NI is either within range or it's
1494 * equidistant. In both of these cases we
1495 * would want to select the NI based on
1496 * its available credits first, and then
1499 if (distance <= shortest_distance &&
1500 distance < lnet_get_numa_range()) {
1502 * If this is the first NI that's
1503 * within range, then set the
1504 * shortest distance to the range
1505 * specified by the user. In
1506 * effect we're saying that all
1507 * NIs that fall within this NUMA
1508 * range shall be dealt with as
1509 * having equal NUMA weight. Which
1510 * will mean that we should select
1511 * through that set by their
1512 * available credits first
1513 * followed by Round Robin.
1515 * And since this is the first NI
1516 * in the range, let's just set it
1517 * as our best_ni for now. The
1518 * following NIs found in the
1519 * range will be dealt with as
1520 * mentioned previously.
1522 shortest_distance = lnet_get_numa_range();
1529 * This NI is NUMA equidistant let's
1530 * select using credits followed by Round
1533 if (ni->ni_tx_queues[cpt]->tq_credits <
1536 } else if (ni->ni_tx_queues[cpt]->tq_credits ==
1539 if (best_ni->ni_seq <= ni->ni_seq)
1546 best_credits = ni->ni_tx_queues[cpt]->tq_credits;
1550 * if the peer is not MR capable, then we should always send to it
1551 * using the first NI in the NET we determined.
1553 if (!peer->lp_multi_rail && local_net != NULL)
1554 best_ni = lnet_net2ni_locked(local_net->net_id, cpt);
1557 lnet_net_unlock(cpt);
1558 LCONSOLE_WARN("No local ni found to send from to %s\n",
1559 libcfs_nid2str(dst_nid));
1564 * Now that we selected the NI to use increment its sequence
1565 * number so the Round Robin algorithm will detect that it has
1566 * been used and pick the next NI.
1574 if (best_ni == the_lnet.ln_loni) {
1575 /* No send credit hassles with LOLND */
1576 lnet_ni_addref_locked(best_ni, cpt);
1577 msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
1578 if (!msg->msg_routing)
1579 msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
1580 msg->msg_target.nid = best_ni->ni_nid;
1581 lnet_msg_commit(msg, cpt);
1583 lnet_net_unlock(cpt);
1584 msg->msg_txni = best_ni;
1585 lnet_ni_send(best_ni, msg);
1593 if (msg->msg_type == LNET_MSG_REPLY ||
1594 msg->msg_type == LNET_MSG_ACK) {
1596 * for replies we want to respond on the same peer_ni we
1597 * received the message on if possible. If not, then pick
1598 * a peer_ni to send to
1600 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1602 lnet_peer_ni_decref_locked(best_lpni);
1605 CDEBUG(D_NET, "unable to send msg_type %d to "
1606 "originating %s\n", msg->msg_type,
1607 libcfs_nid2str(dst_nid));
1611 peer_net = lnet_peer_get_net_locked(peer,
1612 best_ni->ni_net->net_id);
1614 * peer_net is not available or the src_nid is explicitly defined
1615 * and the peer_net for that src_nid is unhealthy. find a route to
1616 * the destination nid.
1619 (src_nid != LNET_NID_ANY &&
1620 !lnet_is_peer_net_healthy_locked(peer_net))) {
1621 best_gw = lnet_find_route_locked(best_ni->ni_net,
1625 * if no route is found for that network then
1626 * move onto the next peer_ni in the peer
1629 lnet_net_unlock(cpt);
1630 LCONSOLE_WARN("No route to peer from %s\n",
1631 libcfs_nid2str(best_ni->ni_nid));
1632 return -EHOSTUNREACH;
1635 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1636 libcfs_nid2str(dst_nid),
1637 libcfs_nid2str(best_gw->lpni_nid),
1638 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1640 best_lpni = lnet_find_peer_ni_locked(dst_nid);
1641 LASSERT(best_lpni != NULL);
1642 lnet_peer_ni_decref_locked(best_lpni);
1647 } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
1649 * this peer_net is unhealthy but we still have an opportunity
1650 * to find another peer_net that we can use
1652 __u32 net_id = peer_net->lpn_net_id;
1653 lnet_net_unlock(cpt);
1655 LCONSOLE_WARN("peer net %s unhealthy\n",
1656 libcfs_net2str(net_id));
1661 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1663 * if this peer ni is not healthy just skip it, no point in
1664 * examining it further
1666 if (!lnet_is_peer_ni_healthy_locked(lpni))
1668 ni_is_pref = lnet_peer_is_ni_pref_locked(lpni, best_ni);
1670 /* if this is a preferred peer use it */
1671 if (!preferred && ni_is_pref) {
1673 } else if (preferred && !ni_is_pref) {
1675 * this is not the preferred peer so let's ignore
1679 } if (lpni->lpni_txcredits < best_lpni_credits)
1681 * We already have a peer that has more credits
1682 * available than this one. No need to consider
1683 * this peer further.
1686 else if (lpni->lpni_txcredits == best_lpni_credits) {
1688 * The best peer found so far and the current peer
1689 * have the same number of available credits let's
1690 * make sure to select between them using Round
1694 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1700 best_lpni_credits = lpni->lpni_txcredits;
1704 * Increment sequence number of the peer selected so that we can
1705 * pick the next one in Round Robin.
1707 best_lpni->lpni_seq++;
1709 /* if we still can't find a peer ni then we can't reach it */
1711 __u32 net_id = peer_net->lpn_net_id;
1712 lnet_net_unlock(cpt);
1713 LCONSOLE_WARN("no peer_ni found on peer net %s\n",
1714 libcfs_net2str(net_id));
1715 return -EHOSTUNREACH;
1720 * determine the cpt to use and if it has changed then
1721 * lock the new cpt and check if the config has changed.
1722 * If it has changed then repeat the algorithm since the
1723 * ni or peer list could have changed and the algorithm
1724 * would endup picking a different ni/peer_ni pair.
1726 cpt2 = best_lpni->lpni_cpt;
1728 lnet_net_unlock(cpt);
1731 seq2 = lnet_get_dlc_seq_locked();
1733 lnet_net_unlock(cpt);
1739 * store the best_lpni in the message right away to avoid having
1740 * to do the same operation under different conditions
1742 msg->msg_txpeer = (routing) ? best_gw : best_lpni;
1743 msg->msg_txni = best_ni;
1745 * grab a reference for the best_ni since now it's in use in this
1746 * send. the reference will need to be dropped when the message is
1747 * finished in lnet_finalize()
1749 lnet_ni_addref_locked(msg->msg_txni, cpt);
1750 lnet_peer_ni_addref_locked(msg->msg_txpeer);
1753 * set the destination nid in the message here because it's
1754 * possible that we'd be sending to a different nid than the one
1757 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1760 * Always set the target.nid to the best peer picked. Either the
1761 * nid will be one of the preconfigured NIDs, or the same NID as
1762 * what was originaly set in the target or it will be the NID of
1763 * a router if this message should be routed
1765 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1768 * lnet_msg_commit assigns the correct cpt to the message, which
1769 * is used to decrement the correct refcount on the ni when it's
1770 * time to return the credits
1772 lnet_msg_commit(msg, cpt);
1775 * If we are routing the message then we don't need to overwrite
1776 * the src_nid since it would've been set at the origin. Otherwise
1777 * we are the originator so we need to set it.
1779 if (!msg->msg_routing)
1780 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1783 msg->msg_target_is_router = 1;
1784 msg->msg_target.pid = LNET_PID_LUSTRE;
1787 rc = lnet_post_send_locked(msg, 0);
1789 lnet_net_unlock(cpt);
1795 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1797 lnet_nid_t dst_nid = msg->msg_target.nid;
1799 bool lo_sent = false;
1802 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1803 * but we might want to use pre-determined router for ACK/REPLY
1806 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1807 LASSERT (msg->msg_txpeer == NULL);
1808 LASSERT (!msg->msg_sending);
1809 LASSERT (!msg->msg_target_is_router);
1810 LASSERT (!msg->msg_receiving);
1812 msg->msg_sending = 1;
1814 LASSERT(!msg->msg_tx_committed);
1816 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid, &lo_sent);
1817 if (rc < 0 || lo_sent)
1820 if (rc == LNET_CREDIT_OK)
1821 lnet_ni_send(msg->msg_txni, msg);
1823 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
1828 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1831 the_lnet.ln_counters[cpt]->drop_count++;
1832 the_lnet.ln_counters[cpt]->drop_length += nob;
1833 lnet_net_unlock(cpt);
1835 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1839 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1841 lnet_hdr_t *hdr = &msg->msg_hdr;
1843 if (msg->msg_wanted != 0)
1844 lnet_setpayloadbuffer(msg);
1846 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1848 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1849 * it back into the ACK during lnet_finalize() */
1850 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1851 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1853 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1854 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1858 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1860 lnet_hdr_t *hdr = &msg->msg_hdr;
1861 struct lnet_match_info info;
1865 /* Convert put fields to host byte order */
1866 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1867 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1868 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1870 /* Primary peer NID. */
1871 info.mi_id.nid = msg->msg_initiator;
1872 info.mi_id.pid = hdr->src_pid;
1873 info.mi_opc = LNET_MD_OP_PUT;
1874 info.mi_portal = hdr->msg.put.ptl_index;
1875 info.mi_rlength = hdr->payload_length;
1876 info.mi_roffset = hdr->msg.put.offset;
1877 info.mi_mbits = hdr->msg.put.match_bits;
1878 info.mi_cpt = msg->msg_rxpeer->lpni_cpt;
1880 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
1881 ready_delay = msg->msg_rx_ready_delay;
1884 rc = lnet_ptl_match_md(&info, msg);
1889 case LNET_MATCHMD_OK:
1890 lnet_recv_put(ni, msg);
1893 case LNET_MATCHMD_NONE:
1895 /* no eager_recv or has already called it, should
1896 * have been attached on delayed list */
1899 rc = lnet_ni_eager_recv(ni, msg);
1906 case LNET_MATCHMD_DROP:
1907 CNETERR("Dropping PUT from %s portal %d match %llu"
1908 " offset %d length %d: %d\n",
1909 libcfs_id2str(info.mi_id), info.mi_portal,
1910 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1912 return -ENOENT; /* -ve: OK but no match */
1917 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1919 struct lnet_match_info info;
1920 lnet_hdr_t *hdr = &msg->msg_hdr;
1921 lnet_process_id_t source_id;
1922 struct lnet_handle_wire reply_wmd;
1925 /* Convert get fields to host byte order */
1926 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1927 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1928 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1929 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1931 source_id.nid = hdr->src_nid;
1932 source_id.pid = hdr->src_pid;
1933 /* Primary peer NID */
1934 info.mi_id.nid = msg->msg_initiator;
1935 info.mi_id.pid = hdr->src_pid;
1936 info.mi_opc = LNET_MD_OP_GET;
1937 info.mi_portal = hdr->msg.get.ptl_index;
1938 info.mi_rlength = hdr->msg.get.sink_length;
1939 info.mi_roffset = hdr->msg.get.src_offset;
1940 info.mi_mbits = hdr->msg.get.match_bits;
1942 rc = lnet_ptl_match_md(&info, msg);
1943 if (rc == LNET_MATCHMD_DROP) {
1944 CNETERR("Dropping GET from %s portal %d match %llu"
1945 " offset %d length %d\n",
1946 libcfs_id2str(info.mi_id), info.mi_portal,
1947 info.mi_mbits, info.mi_roffset, info.mi_rlength);
1948 return -ENOENT; /* -ve: OK but no match */
1951 LASSERT(rc == LNET_MATCHMD_OK);
1953 lnet_build_msg_event(msg, LNET_EVENT_GET);
1955 reply_wmd = hdr->msg.get.return_wmd;
1957 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
1958 msg->msg_offset, msg->msg_wanted);
1960 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1963 /* The LND completes the REPLY from her recv procedure */
1964 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1965 msg->msg_offset, msg->msg_len, msg->msg_len);
1969 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1970 msg->msg_receiving = 0;
1972 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
1974 /* didn't get as far as lnet_ni_send() */
1975 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1976 libcfs_nid2str(ni->ni_nid),
1977 libcfs_id2str(info.mi_id), rc);
1979 lnet_finalize(ni, msg, rc);
1986 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1988 void *private = msg->msg_private;
1989 lnet_hdr_t *hdr = &msg->msg_hdr;
1990 lnet_process_id_t src = {0};
1996 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
1999 src.nid = hdr->src_nid;
2000 src.pid = hdr->src_pid;
2002 /* NB handles only looked up by creator (no flips) */
2003 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
2004 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2005 CNETERR("%s: Dropping REPLY from %s for %s "
2007 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2008 (md == NULL) ? "invalid" : "inactive",
2009 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2010 hdr->msg.reply.dst_wmd.wh_object_cookie);
2011 if (md != NULL && md->md_me != NULL)
2012 CERROR("REPLY MD also attached to portal %d\n",
2013 md->md_me->me_portal);
2015 lnet_res_unlock(cpt);
2016 return -ENOENT; /* -ve: OK but no match */
2019 LASSERT(md->md_offset == 0);
2021 rlength = hdr->payload_length;
2022 mlength = MIN(rlength, (int)md->md_length);
2024 if (mlength < rlength &&
2025 (md->md_options & LNET_MD_TRUNCATE) == 0) {
2026 CNETERR("%s: Dropping REPLY from %s length %d "
2027 "for MD %#llx would overflow (%d)\n",
2028 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2029 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2031 lnet_res_unlock(cpt);
2032 return -ENOENT; /* -ve: OK but no match */
2035 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
2036 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2037 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2039 lnet_msg_attach_md(msg, md, 0, mlength);
2042 lnet_setpayloadbuffer(msg);
2044 lnet_res_unlock(cpt);
2046 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2048 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2053 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
2055 lnet_hdr_t *hdr = &msg->msg_hdr;
2056 lnet_process_id_t src = {0};
2060 src.nid = hdr->src_nid;
2061 src.pid = hdr->src_pid;
2063 /* Convert ack fields to host byte order */
2064 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2065 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2067 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
2070 /* NB handles only looked up by creator (no flips) */
2071 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2072 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2073 /* Don't moan; this is expected */
2075 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
2076 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2077 (md == NULL) ? "invalid" : "inactive",
2078 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2079 hdr->msg.ack.dst_wmd.wh_object_cookie);
2080 if (md != NULL && md->md_me != NULL)
2081 CERROR("Source MD also attached to portal %d\n",
2082 md->md_me->me_portal);
2084 lnet_res_unlock(cpt);
2085 return -ENOENT; /* -ve! */
2088 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
2089 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2090 hdr->msg.ack.dst_wmd.wh_object_cookie);
2092 lnet_msg_attach_md(msg, md, 0, 0);
2094 lnet_res_unlock(cpt);
2096 lnet_build_msg_event(msg, LNET_EVENT_ACK);
2098 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2103 * \retval LNET_CREDIT_OK If \a msg is forwarded
2104 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
2105 * \retval -ve error code
2108 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
2112 if (!the_lnet.ln_routing)
2115 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
2116 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2117 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
2118 msg->msg_rx_ready_delay = 1;
2120 lnet_net_unlock(msg->msg_rx_cpt);
2121 rc = lnet_ni_eager_recv(ni, msg);
2122 lnet_net_lock(msg->msg_rx_cpt);
2127 rc = lnet_post_routed_recv_locked(msg, 0);
2132 lnet_parse_local(lnet_ni_t *ni, lnet_msg_t *msg)
2136 switch (msg->msg_type) {
2138 rc = lnet_parse_ack(ni, msg);
2141 rc = lnet_parse_put(ni, msg);
2144 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
2146 case LNET_MSG_REPLY:
2147 rc = lnet_parse_reply(ni, msg);
2149 default: /* prevent an unused label if !kernel */
2154 LASSERT(rc == 0 || rc == -ENOENT);
2159 lnet_msgtyp2str (int type)
2168 case LNET_MSG_REPLY:
2170 case LNET_MSG_HELLO:
2173 return ("<UNKNOWN>");
2178 lnet_print_hdr(lnet_hdr_t * hdr)
2180 lnet_process_id_t src = {0};
2181 lnet_process_id_t dst = {0};
2182 char *type_str = lnet_msgtyp2str(hdr->type);
2184 src.nid = hdr->src_nid;
2185 src.pid = hdr->src_pid;
2187 dst.nid = hdr->dest_nid;
2188 dst.pid = hdr->dest_pid;
2190 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2191 CWARN(" From %s\n", libcfs_id2str(src));
2192 CWARN(" To %s\n", libcfs_id2str(dst));
2194 switch (hdr->type) {
2199 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
2200 "match bits %llu\n",
2201 hdr->msg.put.ptl_index,
2202 hdr->msg.put.ack_wmd.wh_interface_cookie,
2203 hdr->msg.put.ack_wmd.wh_object_cookie,
2204 hdr->msg.put.match_bits);
2205 CWARN(" Length %d, offset %d, hdr data %#llx\n",
2206 hdr->payload_length, hdr->msg.put.offset,
2207 hdr->msg.put.hdr_data);
2211 CWARN(" Ptl index %d, return md %#llx.%#llx, "
2212 "match bits %llu\n", hdr->msg.get.ptl_index,
2213 hdr->msg.get.return_wmd.wh_interface_cookie,
2214 hdr->msg.get.return_wmd.wh_object_cookie,
2215 hdr->msg.get.match_bits);
2216 CWARN(" Length %d, src offset %d\n",
2217 hdr->msg.get.sink_length,
2218 hdr->msg.get.src_offset);
2222 CWARN(" dst md %#llx.%#llx, "
2223 "manipulated length %d\n",
2224 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2225 hdr->msg.ack.dst_wmd.wh_object_cookie,
2226 hdr->msg.ack.mlength);
2229 case LNET_MSG_REPLY:
2230 CWARN(" dst md %#llx.%#llx, "
2232 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2233 hdr->msg.reply.dst_wmd.wh_object_cookie,
2234 hdr->payload_length);
2240 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
2241 void *private, int rdma_req)
2246 struct lnet_msg *msg;
2247 lnet_pid_t dest_pid;
2248 lnet_nid_t dest_nid;
2250 struct lnet_peer_ni *lpni;
2251 __u32 payload_length;
2254 LASSERT (!in_interrupt ());
2256 type = le32_to_cpu(hdr->type);
2257 src_nid = le64_to_cpu(hdr->src_nid);
2258 dest_nid = le64_to_cpu(hdr->dest_nid);
2259 dest_pid = le32_to_cpu(hdr->dest_pid);
2260 payload_length = le32_to_cpu(hdr->payload_length);
2262 for_me = (ni->ni_nid == dest_nid);
2263 cpt = lnet_cpt_of_nid(from_nid, ni);
2268 if (payload_length > 0) {
2269 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2270 libcfs_nid2str(from_nid),
2271 libcfs_nid2str(src_nid),
2272 lnet_msgtyp2str(type), payload_length);
2278 case LNET_MSG_REPLY:
2279 if (payload_length >
2280 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2281 CERROR("%s, src %s: bad %s payload %d "
2282 "(%d max expected)\n",
2283 libcfs_nid2str(from_nid),
2284 libcfs_nid2str(src_nid),
2285 lnet_msgtyp2str(type),
2287 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2293 CERROR("%s, src %s: Bad message type 0x%x\n",
2294 libcfs_nid2str(from_nid),
2295 libcfs_nid2str(src_nid), type);
2299 if (the_lnet.ln_routing &&
2300 ni->ni_last_alive != ktime_get_real_seconds()) {
2301 /* NB: so far here is the only place to set NI status to "up */
2303 ni->ni_last_alive = ktime_get_real_seconds();
2304 if (ni->ni_status != NULL &&
2305 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
2306 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
2310 /* Regard a bad destination NID as a protocol error. Senders should
2311 * know what they're doing; if they don't they're misconfigured, buggy
2312 * or malicious so we chop them off at the knees :) */
2315 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2316 /* should have gone direct */
2317 CERROR("%s, src %s: Bad dest nid %s "
2318 "(should have been sent direct)\n",
2319 libcfs_nid2str(from_nid),
2320 libcfs_nid2str(src_nid),
2321 libcfs_nid2str(dest_nid));
2325 if (lnet_islocalnid(dest_nid)) {
2326 /* dest is another local NI; sender should have used
2327 * this node's NID on its own network */
2328 CERROR("%s, src %s: Bad dest nid %s "
2329 "(it's my nid but on a different network)\n",
2330 libcfs_nid2str(from_nid),
2331 libcfs_nid2str(src_nid),
2332 libcfs_nid2str(dest_nid));
2336 if (rdma_req && type == LNET_MSG_GET) {
2337 CERROR("%s, src %s: Bad optimized GET for %s "
2338 "(final destination must be me)\n",
2339 libcfs_nid2str(from_nid),
2340 libcfs_nid2str(src_nid),
2341 libcfs_nid2str(dest_nid));
2345 if (!the_lnet.ln_routing) {
2346 CERROR("%s, src %s: Dropping message for %s "
2347 "(routing not enabled)\n",
2348 libcfs_nid2str(from_nid),
2349 libcfs_nid2str(src_nid),
2350 libcfs_nid2str(dest_nid));
2355 /* Message looks OK; we're not going to return an error, so we MUST
2356 * call back lnd_recv() come what may... */
2358 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2359 fail_peer(src_nid, 0)) { /* shall we now? */
2360 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2361 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2362 lnet_msgtyp2str(type));
2366 if (!list_empty(&the_lnet.ln_drop_rules) &&
2367 lnet_drop_rule_match(hdr)) {
2368 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
2369 "silent message loss\n",
2370 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2371 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
2376 msg = lnet_msg_alloc();
2378 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2379 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2380 lnet_msgtyp2str(type));
2384 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
2385 * pointers NULL etc */
2387 msg->msg_type = type;
2388 msg->msg_private = private;
2389 msg->msg_receiving = 1;
2390 msg->msg_rdma_get = rdma_req;
2391 msg->msg_len = msg->msg_wanted = payload_length;
2392 msg->msg_offset = 0;
2393 msg->msg_hdr = *hdr;
2394 /* for building message event */
2395 msg->msg_from = from_nid;
2397 msg->msg_target.pid = dest_pid;
2398 msg->msg_target.nid = dest_nid;
2399 msg->msg_routing = 1;
2402 /* convert common msg->hdr fields to host byteorder */
2403 msg->msg_hdr.type = type;
2404 msg->msg_hdr.src_nid = src_nid;
2405 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2406 msg->msg_hdr.dest_nid = dest_nid;
2407 msg->msg_hdr.dest_pid = dest_pid;
2408 msg->msg_hdr.payload_length = payload_length;
2410 /* Multi-Rail: Primary NID of source. */
2411 msg->msg_initiator = lnet_peer_primary_nid(src_nid);
2414 lpni = lnet_nid2peerni_locked(from_nid, cpt);
2416 lnet_net_unlock(cpt);
2417 CERROR("%s, src %s: Dropping %s "
2418 "(error %ld looking up sender)\n",
2419 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2420 lnet_msgtyp2str(type), PTR_ERR(lpni));
2422 if (rc == -ESHUTDOWN)
2423 /* We are shutting down. Don't do anything more */
2427 msg->msg_rxpeer = lpni;
2429 lnet_ni_addref_locked(ni, cpt);
2431 if (lnet_isrouter(msg->msg_rxpeer)) {
2432 lnet_peer_set_alive(msg->msg_rxpeer);
2433 if (avoid_asym_router_failure &&
2434 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
2435 /* received a remote message from router, update
2436 * remote NI status on this router.
2437 * NB: multi-hop routed message will be ignored.
2439 lnet_router_ni_update_locked(msg->msg_rxpeer,
2440 LNET_NIDNET(src_nid));
2444 lnet_msg_commit(msg, cpt);
2446 /* message delay simulation */
2447 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
2448 lnet_delay_rule_match_locked(hdr, msg))) {
2449 lnet_net_unlock(cpt);
2454 rc = lnet_parse_forward_locked(ni, msg);
2455 lnet_net_unlock(cpt);
2460 if (rc == LNET_CREDIT_OK) {
2461 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2462 0, payload_length, payload_length);
2467 lnet_net_unlock(cpt);
2469 rc = lnet_parse_local(ni, msg);
2475 LASSERT(msg->msg_md == NULL);
2476 lnet_finalize(ni, msg, rc);
2479 lnet_drop_message(ni, cpt, private, payload_length);
2482 EXPORT_SYMBOL(lnet_parse);
2485 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
2487 while (!list_empty(head)) {
2488 lnet_process_id_t id = {0};
2491 msg = list_entry(head->next, lnet_msg_t, msg_list);
2492 list_del(&msg->msg_list);
2494 id.nid = msg->msg_hdr.src_nid;
2495 id.pid = msg->msg_hdr.src_pid;
2497 LASSERT(msg->msg_md == NULL);
2498 LASSERT(msg->msg_rx_delayed);
2499 LASSERT(msg->msg_rxpeer != NULL);
2500 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2502 CWARN("Dropping delayed PUT from %s portal %d match %llu"
2503 " offset %d length %d: %s\n",
2505 msg->msg_hdr.msg.put.ptl_index,
2506 msg->msg_hdr.msg.put.match_bits,
2507 msg->msg_hdr.msg.put.offset,
2508 msg->msg_hdr.payload_length, reason);
2510 /* NB I can't drop msg's ref on msg_rxpeer until after I've
2511 * called lnet_drop_message(), so I just hang onto msg as well
2512 * until that's done */
2514 lnet_drop_message(msg->msg_rxni,
2515 msg->msg_rxpeer->lpni_cpt,
2516 msg->msg_private, msg->msg_len);
2518 * NB: message will not generate event because w/o attached MD,
2519 * but we still should give error code so lnet_msg_decommit()
2520 * can skip counters operations and other checks.
2522 lnet_finalize(msg->msg_rxni, msg, -ENOENT);
2527 lnet_recv_delayed_msg_list(struct list_head *head)
2529 while (!list_empty(head)) {
2531 lnet_process_id_t id;
2533 msg = list_entry(head->next, lnet_msg_t, msg_list);
2534 list_del(&msg->msg_list);
2536 /* md won't disappear under me, since each msg
2537 * holds a ref on it */
2539 id.nid = msg->msg_hdr.src_nid;
2540 id.pid = msg->msg_hdr.src_pid;
2542 LASSERT(msg->msg_rx_delayed);
2543 LASSERT(msg->msg_md != NULL);
2544 LASSERT(msg->msg_rxpeer != NULL);
2545 LASSERT(msg->msg_rxni != NULL);
2546 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2548 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2549 "match %llu offset %d length %d.\n",
2550 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2551 msg->msg_hdr.msg.put.match_bits,
2552 msg->msg_hdr.msg.put.offset,
2553 msg->msg_hdr.payload_length);
2555 lnet_recv_put(msg->msg_rxni, msg);
2560 * Initiate an asynchronous PUT operation.
2562 * There are several events associated with a PUT: completion of the send on
2563 * the initiator node (LNET_EVENT_SEND), and when the send completes
2564 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2565 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2566 * used at the target node to indicate the completion of incoming data
2569 * The local events will be logged in the EQ associated with the MD pointed to
2570 * by \a mdh handle. Using a MD without an associated EQ results in these
2571 * events being discarded. In this case, the caller must have another
2572 * mechanism (e.g., a higher level protocol) for determining when it is safe
2573 * to modify the memory region associated with the MD.
2575 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2576 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2578 * \param self Indicates the NID of a local interface through which to send
2579 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2580 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2581 * must be "free floating" (See LNetMDBind()).
2582 * \param ack Controls whether an acknowledgment is requested.
2583 * Acknowledgments are only sent when they are requested by the initiating
2584 * process and the target MD enables them.
2585 * \param target A process identifier for the target process.
2586 * \param portal The index in the \a target's portal table.
2587 * \param match_bits The match bits to use for MD selection at the target
2589 * \param offset The offset into the target MD (only used when the target
2590 * MD has the LNET_MD_MANAGE_REMOTE option set).
2591 * \param hdr_data 64 bits of user data that can be included in the message
2592 * header. This data is written to an event queue entry at the target if an
2593 * EQ is present on the matching MD.
2595 * \retval 0 Success, and only in this case events will be generated
2596 * and logged to EQ (if it exists).
2597 * \retval -EIO Simulated failure.
2598 * \retval -ENOMEM Memory allocation failure.
2599 * \retval -ENOENT Invalid MD object.
2601 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2604 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2605 lnet_process_id_t target, unsigned int portal,
2606 __u64 match_bits, unsigned int offset,
2609 struct lnet_msg *msg;
2610 struct lnet_libmd *md;
2614 LASSERT(the_lnet.ln_refcount > 0);
2616 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2617 fail_peer(target.nid, 1)) { /* shall we now? */
2618 CERROR("Dropping PUT to %s: simulated failure\n",
2619 libcfs_id2str(target));
2623 msg = lnet_msg_alloc();
2625 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2626 libcfs_id2str(target));
2629 msg->msg_vmflush = !!memory_pressure_get();
2631 cpt = lnet_cpt_of_cookie(mdh.cookie);
2634 md = lnet_handle2md(&mdh);
2635 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2636 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
2637 match_bits, portal, libcfs_id2str(target),
2638 md == NULL ? -1 : md->md_threshold);
2639 if (md != NULL && md->md_me != NULL)
2640 CERROR("Source MD also attached to portal %d\n",
2641 md->md_me->me_portal);
2642 lnet_res_unlock(cpt);
2648 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2650 lnet_msg_attach_md(msg, md, 0, 0);
2652 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2654 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2655 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2656 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2657 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2659 /* NB handles only looked up by creator (no flips) */
2660 if (ack == LNET_ACK_REQ) {
2661 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2662 the_lnet.ln_interface_cookie;
2663 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2664 md->md_lh.lh_cookie;
2666 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2667 LNET_WIRE_HANDLE_COOKIE_NONE;
2668 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2669 LNET_WIRE_HANDLE_COOKIE_NONE;
2672 lnet_res_unlock(cpt);
2674 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2676 rc = lnet_send(self, msg, LNET_NID_ANY);
2678 CNETERR("Error sending PUT to %s: %d\n",
2679 libcfs_id2str(target), rc);
2680 lnet_finalize(NULL, msg, rc);
2683 /* completion will be signalled by an event */
2686 EXPORT_SYMBOL(LNetPut);
2689 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2691 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2692 * returns a msg for the LND to pass to lnet_finalize() when the sink
2693 * data has been received.
2695 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2696 * lnet_finalize() is called on it, so the LND must call this first */
2698 struct lnet_msg *msg = lnet_msg_alloc();
2699 struct lnet_libmd *getmd = getmsg->msg_md;
2700 lnet_process_id_t peer_id = getmsg->msg_target;
2703 LASSERT(!getmsg->msg_target_is_router);
2704 LASSERT(!getmsg->msg_routing);
2707 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2708 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2712 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2715 LASSERT(getmd->md_refcount > 0);
2717 if (getmd->md_threshold == 0) {
2718 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
2719 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2721 lnet_res_unlock(cpt);
2725 LASSERT(getmd->md_offset == 0);
2727 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2728 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2730 /* setup information for lnet_build_msg_event */
2731 msg->msg_initiator = lnet_peer_primary_nid(peer_id.nid);
2732 /* Cheaper: msg->msg_initiator = getmsg->msg_txpeer->lp_nid; */
2733 msg->msg_from = peer_id.nid;
2734 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2735 msg->msg_hdr.src_nid = peer_id.nid;
2736 msg->msg_hdr.payload_length = getmd->md_length;
2737 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2739 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2740 lnet_res_unlock(cpt);
2742 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
2745 lnet_msg_commit(msg, cpt);
2746 lnet_net_unlock(cpt);
2748 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2753 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
2756 the_lnet.ln_counters[cpt]->drop_count++;
2757 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2758 lnet_net_unlock(cpt);
2765 EXPORT_SYMBOL(lnet_create_reply_msg);
2768 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2770 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2771 * completed and I know it. */
2772 LASSERT(reply != NULL);
2773 LASSERT(reply->msg_type == LNET_MSG_GET);
2774 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
2776 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2777 * the end of my buffer, I might as well be dead. */
2778 LASSERT(len <= reply->msg_ev.mlength);
2780 reply->msg_ev.mlength = len;
2782 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2785 * Initiate an asynchronous GET operation.
2787 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2788 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2789 * the target node in the REPLY has been written to local MD.
2791 * On the target node, an LNET_EVENT_GET is logged when the GET request
2792 * arrives and is accepted into a MD.
2794 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2795 * \param mdh A handle for the MD that describes the memory into which the
2796 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2798 * \retval 0 Success, and only in this case events will be generated
2799 * and logged to EQ (if it exists) of the MD.
2800 * \retval -EIO Simulated failure.
2801 * \retval -ENOMEM Memory allocation failure.
2802 * \retval -ENOENT Invalid MD object.
2805 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2806 lnet_process_id_t target, unsigned int portal,
2807 __u64 match_bits, unsigned int offset)
2809 struct lnet_msg *msg;
2810 struct lnet_libmd *md;
2814 LASSERT(the_lnet.ln_refcount > 0);
2816 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2817 fail_peer(target.nid, 1)) /* shall we now? */
2819 CERROR("Dropping GET to %s: simulated failure\n",
2820 libcfs_id2str(target));
2824 msg = lnet_msg_alloc();
2826 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2827 libcfs_id2str(target));
2831 cpt = lnet_cpt_of_cookie(mdh.cookie);
2834 md = lnet_handle2md(&mdh);
2835 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2836 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
2837 match_bits, portal, libcfs_id2str(target),
2838 md == NULL ? -1 : md->md_threshold);
2839 if (md != NULL && md->md_me != NULL)
2840 CERROR("REPLY MD also attached to portal %d\n",
2841 md->md_me->me_portal);
2843 lnet_res_unlock(cpt);
2849 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2851 lnet_msg_attach_md(msg, md, 0, 0);
2853 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2855 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2856 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2857 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2858 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2860 /* NB handles only looked up by creator (no flips) */
2861 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2862 the_lnet.ln_interface_cookie;
2863 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2864 md->md_lh.lh_cookie;
2866 lnet_res_unlock(cpt);
2868 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2870 rc = lnet_send(self, msg, LNET_NID_ANY);
2872 CNETERR("Error sending GET to %s: %d\n",
2873 libcfs_id2str(target), rc);
2874 lnet_finalize(NULL, msg, rc);
2877 /* completion will be signalled by an event */
2880 EXPORT_SYMBOL(LNetGet);
2883 * Calculate distance to node at \a dstnid.
2885 * \param dstnid Target NID.
2886 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2888 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2891 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2892 * local_nid_dist_zero is set, which is the default.
2893 * \retval positives Distance to target NID, i.e. number of hops plus one.
2894 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2897 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2899 struct list_head *e;
2900 struct lnet_ni *ni = NULL;
2901 lnet_remotenet_t *rnet;
2902 __u32 dstnet = LNET_NIDNET(dstnid);
2906 struct list_head *rn_list;
2908 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2909 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2910 * keep order 0 free for 0@lo and order 1 free for a local NID
2913 LASSERT(the_lnet.ln_refcount > 0);
2915 cpt = lnet_net_lock_current();
2917 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
2918 if (ni->ni_nid == dstnid) {
2919 if (srcnidp != NULL)
2921 if (orderp != NULL) {
2922 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2927 lnet_net_unlock(cpt);
2929 return local_nid_dist_zero ? 0 : 1;
2932 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2933 /* Check if ni was originally created in
2934 * current net namespace.
2935 * If not, assign order above 0xffff0000,
2936 * to make this ni not a priority. */
2937 if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
2938 order += 0xffff0000;
2940 if (srcnidp != NULL)
2941 *srcnidp = ni->ni_nid;
2944 lnet_net_unlock(cpt);
2951 rn_list = lnet_net2rnethash(dstnet);
2952 list_for_each(e, rn_list) {
2953 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2955 if (rnet->lrn_net == dstnet) {
2956 lnet_route_t *route;
2957 lnet_route_t *shortest = NULL;
2958 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
2961 LASSERT(!list_empty(&rnet->lrn_routes));
2963 list_for_each_entry(route, &rnet->lrn_routes,
2965 route_hops = route->lr_hops;
2966 if (route_hops == LNET_UNDEFINED_HOPS)
2968 if (shortest == NULL ||
2969 route_hops < shortest_hops) {
2971 shortest_hops = route_hops;
2975 LASSERT(shortest != NULL);
2976 hops = shortest_hops;
2977 if (srcnidp != NULL) {
2978 ni = lnet_get_next_ni_locked(
2979 shortest->lr_gateway->lpni_net,
2981 *srcnidp = ni->ni_nid;
2985 lnet_net_unlock(cpt);
2991 lnet_net_unlock(cpt);
2992 return -EHOSTUNREACH;
2994 EXPORT_SYMBOL(LNetDist);