4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/lnet/lib-move.c
34 * Data movement routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/pagemap.h>
41 #include <lnet/lib-lnet.h>
42 #include <linux/nsproxy.h>
43 #include <net/net_namespace.h>
45 extern unsigned int lnet_current_net_count;
47 static int local_nid_dist_zero = 1;
48 module_param(local_nid_dist_zero, int, 0444);
49 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
51 struct lnet_send_data {
52 struct lnet_ni *sd_best_ni;
53 struct lnet_peer_ni *sd_best_lpni;
54 struct lnet_peer_ni *sd_final_dst_lpni;
55 struct lnet_peer *sd_peer;
56 struct lnet_peer *sd_gw_peer;
57 struct lnet_peer_ni *sd_gw_lpni;
58 struct lnet_peer_net *sd_peer_net;
59 struct lnet_msg *sd_msg;
60 lnet_nid_t sd_dst_nid;
61 lnet_nid_t sd_src_nid;
62 lnet_nid_t sd_rtr_nid;
68 static inline struct lnet_comm_count *
69 get_stats_counts(struct lnet_element_stats *stats,
70 enum lnet_stats_type stats_type)
73 case LNET_STATS_TYPE_SEND:
74 return &stats->el_send_stats;
75 case LNET_STATS_TYPE_RECV:
76 return &stats->el_recv_stats;
77 case LNET_STATS_TYPE_DROP:
78 return &stats->el_drop_stats;
80 CERROR("Unknown stats type\n");
86 void lnet_incr_stats(struct lnet_element_stats *stats,
87 enum lnet_msg_type msg_type,
88 enum lnet_stats_type stats_type)
90 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
96 atomic_inc(&counts->co_ack_count);
99 atomic_inc(&counts->co_put_count);
102 atomic_inc(&counts->co_get_count);
105 atomic_inc(&counts->co_reply_count);
108 atomic_inc(&counts->co_hello_count);
111 CERROR("There is a BUG in the code. Unknown message type\n");
116 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
117 enum lnet_stats_type stats_type)
119 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
123 return (atomic_read(&counts->co_ack_count) +
124 atomic_read(&counts->co_put_count) +
125 atomic_read(&counts->co_get_count) +
126 atomic_read(&counts->co_reply_count) +
127 atomic_read(&counts->co_hello_count));
130 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
131 struct lnet_comm_count *counts)
133 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
134 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
135 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
136 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
137 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
140 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
141 struct lnet_element_stats *stats)
143 struct lnet_comm_count *counts;
148 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
151 assign_stats(&msg_stats->im_send_stats, counts);
153 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
156 assign_stats(&msg_stats->im_recv_stats, counts);
158 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
161 assign_stats(&msg_stats->im_drop_stats, counts);
165 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
167 struct lnet_test_peer *tp;
168 struct list_head *el;
169 struct list_head *next;
170 struct list_head cull;
172 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
173 if (threshold != 0) {
174 /* Adding a new entry */
175 LIBCFS_ALLOC(tp, sizeof(*tp));
180 tp->tp_threshold = threshold;
183 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
188 /* removing entries */
189 INIT_LIST_HEAD(&cull);
193 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
194 tp = list_entry(el, struct lnet_test_peer, tp_list);
196 if (tp->tp_threshold == 0 || /* needs culling anyway */
197 nid == LNET_NID_ANY || /* removing all entries */
198 tp->tp_nid == nid) { /* matched this one */
199 list_del(&tp->tp_list);
200 list_add(&tp->tp_list, &cull);
206 while (!list_empty(&cull)) {
207 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
209 list_del(&tp->tp_list);
210 LIBCFS_FREE(tp, sizeof(*tp));
216 fail_peer (lnet_nid_t nid, int outgoing)
218 struct lnet_test_peer *tp;
219 struct list_head *el;
220 struct list_head *next;
221 struct list_head cull;
224 INIT_LIST_HEAD(&cull);
226 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
229 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
230 tp = list_entry(el, struct lnet_test_peer, tp_list);
232 if (tp->tp_threshold == 0) {
235 /* only cull zombies on outgoing tests,
236 * since we may be at interrupt priority on
237 * incoming messages. */
238 list_del(&tp->tp_list);
239 list_add(&tp->tp_list, &cull);
244 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
245 nid == tp->tp_nid) { /* fail this peer */
248 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
251 tp->tp_threshold == 0) {
253 list_del(&tp->tp_list);
254 list_add(&tp->tp_list, &cull);
263 while (!list_empty(&cull)) {
264 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
265 list_del(&tp->tp_list);
267 LIBCFS_FREE(tp, sizeof(*tp));
274 lnet_iov_nob(unsigned int niov, struct kvec *iov)
276 unsigned int nob = 0;
278 LASSERT(niov == 0 || iov != NULL);
280 nob += (iov++)->iov_len;
284 EXPORT_SYMBOL(lnet_iov_nob);
287 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
288 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
291 /* NB diov, siov are READ-ONLY */
292 unsigned int this_nob;
297 /* skip complete frags before 'doffset' */
299 while (doffset >= diov->iov_len) {
300 doffset -= diov->iov_len;
306 /* skip complete frags before 'soffset' */
308 while (soffset >= siov->iov_len) {
309 soffset -= siov->iov_len;
318 this_nob = MIN(diov->iov_len - doffset,
319 siov->iov_len - soffset);
320 this_nob = MIN(this_nob, nob);
322 memcpy((char *)diov->iov_base + doffset,
323 (char *)siov->iov_base + soffset, this_nob);
326 if (diov->iov_len > doffset + this_nob) {
334 if (siov->iov_len > soffset + this_nob) {
343 EXPORT_SYMBOL(lnet_copy_iov2iov);
346 lnet_extract_iov(int dst_niov, struct kvec *dst,
347 int src_niov, struct kvec *src,
348 unsigned int offset, unsigned int len)
350 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
351 * for exactly 'len' bytes, and return the number of entries.
352 * NB not destructive to 'src' */
353 unsigned int frag_len;
356 if (len == 0) /* no data => */
357 return (0); /* no frags */
359 LASSERT(src_niov > 0);
360 while (offset >= src->iov_len) { /* skip initial frags */
361 offset -= src->iov_len;
364 LASSERT(src_niov > 0);
369 LASSERT(src_niov > 0);
370 LASSERT((int)niov <= dst_niov);
372 frag_len = src->iov_len - offset;
373 dst->iov_base = ((char *)src->iov_base) + offset;
375 if (len <= frag_len) {
380 dst->iov_len = frag_len;
390 EXPORT_SYMBOL(lnet_extract_iov);
394 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
396 unsigned int nob = 0;
398 LASSERT(niov == 0 || kiov != NULL);
400 nob += (kiov++)->kiov_len;
404 EXPORT_SYMBOL(lnet_kiov_nob);
407 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
408 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
411 /* NB diov, siov are READ-ONLY */
412 unsigned int this_nob;
419 LASSERT (!in_interrupt ());
422 while (doffset >= diov->kiov_len) {
423 doffset -= diov->kiov_len;
430 while (soffset >= siov->kiov_len) {
431 soffset -= siov->kiov_len;
440 this_nob = MIN(diov->kiov_len - doffset,
441 siov->kiov_len - soffset);
442 this_nob = MIN(this_nob, nob);
445 daddr = ((char *)kmap(diov->kiov_page)) +
446 diov->kiov_offset + doffset;
448 saddr = ((char *)kmap(siov->kiov_page)) +
449 siov->kiov_offset + soffset;
451 /* Vanishing risk of kmap deadlock when mapping 2 pages.
452 * However in practice at least one of the kiovs will be mapped
453 * kernel pages and the map/unmap will be NOOPs */
455 memcpy (daddr, saddr, this_nob);
458 if (diov->kiov_len > doffset + this_nob) {
462 kunmap(diov->kiov_page);
469 if (siov->kiov_len > soffset + this_nob) {
473 kunmap(siov->kiov_page);
482 kunmap(diov->kiov_page);
484 kunmap(siov->kiov_page);
486 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
489 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
490 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
493 /* NB iov, kiov are READ-ONLY */
494 unsigned int this_nob;
500 LASSERT (!in_interrupt ());
503 while (iovoffset >= iov->iov_len) {
504 iovoffset -= iov->iov_len;
511 while (kiovoffset >= kiov->kiov_len) {
512 kiovoffset -= kiov->kiov_len;
521 this_nob = MIN(iov->iov_len - iovoffset,
522 kiov->kiov_len - kiovoffset);
523 this_nob = MIN(this_nob, nob);
526 addr = ((char *)kmap(kiov->kiov_page)) +
527 kiov->kiov_offset + kiovoffset;
529 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
532 if (iov->iov_len > iovoffset + this_nob) {
533 iovoffset += this_nob;
540 if (kiov->kiov_len > kiovoffset + this_nob) {
542 kiovoffset += this_nob;
544 kunmap(kiov->kiov_page);
554 kunmap(kiov->kiov_page);
556 EXPORT_SYMBOL(lnet_copy_kiov2iov);
559 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
560 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
563 /* NB kiov, iov are READ-ONLY */
564 unsigned int this_nob;
570 LASSERT (!in_interrupt ());
573 while (kiovoffset >= kiov->kiov_len) {
574 kiovoffset -= kiov->kiov_len;
581 while (iovoffset >= iov->iov_len) {
582 iovoffset -= iov->iov_len;
591 this_nob = MIN(kiov->kiov_len - kiovoffset,
592 iov->iov_len - iovoffset);
593 this_nob = MIN(this_nob, nob);
596 addr = ((char *)kmap(kiov->kiov_page)) +
597 kiov->kiov_offset + kiovoffset;
599 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
602 if (kiov->kiov_len > kiovoffset + this_nob) {
604 kiovoffset += this_nob;
606 kunmap(kiov->kiov_page);
613 if (iov->iov_len > iovoffset + this_nob) {
614 iovoffset += this_nob;
623 kunmap(kiov->kiov_page);
625 EXPORT_SYMBOL(lnet_copy_iov2kiov);
628 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
629 int src_niov, lnet_kiov_t *src,
630 unsigned int offset, unsigned int len)
632 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
633 * for exactly 'len' bytes, and return the number of entries.
634 * NB not destructive to 'src' */
635 unsigned int frag_len;
638 if (len == 0) /* no data => */
639 return (0); /* no frags */
641 LASSERT(src_niov > 0);
642 while (offset >= src->kiov_len) { /* skip initial frags */
643 offset -= src->kiov_len;
646 LASSERT(src_niov > 0);
651 LASSERT(src_niov > 0);
652 LASSERT((int)niov <= dst_niov);
654 frag_len = src->kiov_len - offset;
655 dst->kiov_page = src->kiov_page;
656 dst->kiov_offset = src->kiov_offset + offset;
658 if (len <= frag_len) {
660 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
664 dst->kiov_len = frag_len;
665 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
675 EXPORT_SYMBOL(lnet_extract_kiov);
678 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
679 int delayed, unsigned int offset, unsigned int mlen,
682 unsigned int niov = 0;
683 struct kvec *iov = NULL;
684 lnet_kiov_t *kiov = NULL;
687 LASSERT (!in_interrupt ());
688 LASSERT (mlen == 0 || msg != NULL);
691 LASSERT(msg->msg_receiving);
692 LASSERT(!msg->msg_sending);
693 LASSERT(rlen == msg->msg_len);
694 LASSERT(mlen <= msg->msg_len);
695 LASSERT(msg->msg_offset == offset);
696 LASSERT(msg->msg_wanted == mlen);
698 msg->msg_receiving = 0;
701 niov = msg->msg_niov;
703 kiov = msg->msg_kiov;
706 LASSERT ((iov == NULL) != (kiov == NULL));
710 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
711 niov, iov, kiov, offset, mlen,
714 lnet_finalize(msg, rc);
718 lnet_setpayloadbuffer(struct lnet_msg *msg)
720 struct lnet_libmd *md = msg->msg_md;
722 LASSERT(msg->msg_len > 0);
723 LASSERT(!msg->msg_routing);
725 LASSERT(msg->msg_niov == 0);
726 LASSERT(msg->msg_iov == NULL);
727 LASSERT(msg->msg_kiov == NULL);
729 msg->msg_niov = md->md_niov;
730 if ((md->md_options & LNET_MD_KIOV) != 0)
731 msg->msg_kiov = md->md_iov.kiov;
733 msg->msg_iov = md->md_iov.iov;
737 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
738 unsigned int offset, unsigned int len)
740 msg->msg_type = type;
741 msg->msg_target = target;
743 msg->msg_offset = offset;
746 lnet_setpayloadbuffer(msg);
748 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
749 msg->msg_hdr.type = cpu_to_le32(type);
750 /* dest_nid will be overwritten by lnet_select_pathway() */
751 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
752 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
753 /* src_nid will be set later */
754 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
755 msg->msg_hdr.payload_length = cpu_to_le32(len);
759 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
761 void *priv = msg->msg_private;
764 LASSERT (!in_interrupt ());
765 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
766 (msg->msg_txcredit && msg->msg_peertxcredit));
768 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
770 msg->msg_no_resend = true;
771 lnet_finalize(msg, rc);
776 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
780 LASSERT(!msg->msg_sending);
781 LASSERT(msg->msg_receiving);
782 LASSERT(!msg->msg_rx_ready_delay);
783 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
785 msg->msg_rx_ready_delay = 1;
786 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
789 CERROR("recv from %s / send to %s aborted: "
790 "eager_recv failed %d\n",
791 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
792 libcfs_id2str(msg->msg_target), rc);
793 LASSERT(rc < 0); /* required by my callers */
799 /* NB: returns 1 when alive, 0 when dead, negative when error;
800 * may drop the lnet_net_lock */
802 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
803 struct lnet_msg *msg)
805 if (!lnet_peer_aliveness_enabled(lpni))
809 * If we're resending a message, let's attempt to send it even if
810 * the peer is down to fulfill our resend quota on the message
812 if (msg->msg_retry_count > 0)
815 /* try and send recovery messages irregardless */
816 if (msg->msg_recovery)
819 /* always send any responses */
820 if (msg->msg_type == LNET_MSG_ACK ||
821 msg->msg_type == LNET_MSG_REPLY)
824 return lnet_is_peer_ni_alive(lpni);
828 * \param msg The message to be sent.
829 * \param do_send True if lnet_ni_send() should be called in this function.
830 * lnet_send() is going to lnet_net_unlock immediately after this, so
831 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
833 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
834 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
835 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
836 * \retval -ECANCELED If the MD of the message has been unlinked.
839 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
841 struct lnet_peer_ni *lp = msg->msg_txpeer;
842 struct lnet_ni *ni = msg->msg_txni;
843 int cpt = msg->msg_tx_cpt;
844 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
846 /* non-lnet_send() callers have checked before */
847 LASSERT(!do_send || msg->msg_tx_delayed);
848 LASSERT(!msg->msg_receiving);
849 LASSERT(msg->msg_tx_committed);
850 /* can't get here if we're sending to the loopback interface */
851 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
853 /* NB 'lp' is always the next hop */
854 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
855 lnet_peer_alive_locked(ni, lp, msg) == 0) {
856 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
857 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
859 lnet_net_unlock(cpt);
861 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
863 LNET_STATS_TYPE_DROP);
865 lnet_incr_stats(&msg->msg_txni->ni_stats,
867 LNET_STATS_TYPE_DROP);
869 CNETERR("Dropping message for %s: peer not alive\n",
870 libcfs_id2str(msg->msg_target));
871 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
873 lnet_finalize(msg, -EHOSTUNREACH);
876 return -EHOSTUNREACH;
879 if (msg->msg_md != NULL &&
880 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
881 lnet_net_unlock(cpt);
883 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
884 "called on the MD/ME.\n",
885 libcfs_id2str(msg->msg_target));
887 msg->msg_no_resend = true;
888 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
889 msg, libcfs_id2str(msg->msg_target));
890 lnet_finalize(msg, -ECANCELED);
897 if (!msg->msg_peertxcredit) {
898 spin_lock(&lp->lpni_lock);
899 LASSERT((lp->lpni_txcredits < 0) ==
900 !list_empty(&lp->lpni_txq));
902 msg->msg_peertxcredit = 1;
903 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
904 lp->lpni_txcredits--;
906 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
907 lp->lpni_mintxcredits = lp->lpni_txcredits;
909 if (lp->lpni_txcredits < 0) {
910 msg->msg_tx_delayed = 1;
911 list_add_tail(&msg->msg_list, &lp->lpni_txq);
912 spin_unlock(&lp->lpni_lock);
913 return LNET_CREDIT_WAIT;
915 spin_unlock(&lp->lpni_lock);
918 if (!msg->msg_txcredit) {
919 LASSERT((tq->tq_credits < 0) ==
920 !list_empty(&tq->tq_delayed));
922 msg->msg_txcredit = 1;
924 atomic_dec(&ni->ni_tx_credits);
926 if (tq->tq_credits < tq->tq_credits_min)
927 tq->tq_credits_min = tq->tq_credits;
929 if (tq->tq_credits < 0) {
930 msg->msg_tx_delayed = 1;
931 list_add_tail(&msg->msg_list, &tq->tq_delayed);
932 return LNET_CREDIT_WAIT;
936 /* unset the tx_delay flag as we're going to send it now */
937 msg->msg_tx_delayed = 0;
940 lnet_net_unlock(cpt);
941 lnet_ni_send(ni, msg);
944 return LNET_CREDIT_OK;
948 static struct lnet_rtrbufpool *
949 lnet_msg2bufpool(struct lnet_msg *msg)
951 struct lnet_rtrbufpool *rbp;
954 LASSERT(msg->msg_rx_committed);
956 cpt = msg->msg_rx_cpt;
957 rbp = &the_lnet.ln_rtrpools[cpt][0];
959 LASSERT(msg->msg_len <= LNET_MTU);
960 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
962 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
969 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
971 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
972 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
973 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
974 * received or OK to receive */
975 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
976 struct lnet_peer *lp;
977 struct lnet_rtrbufpool *rbp;
978 struct lnet_rtrbuf *rb;
980 LASSERT(msg->msg_iov == NULL);
981 LASSERT(msg->msg_kiov == NULL);
982 LASSERT(msg->msg_niov == 0);
983 LASSERT(msg->msg_routing);
984 LASSERT(msg->msg_receiving);
985 LASSERT(!msg->msg_sending);
986 LASSERT(lpni->lpni_peer_net);
987 LASSERT(lpni->lpni_peer_net->lpn_peer);
989 lp = lpni->lpni_peer_net->lpn_peer;
991 /* non-lnet_parse callers only receive delayed messages */
992 LASSERT(!do_recv || msg->msg_rx_delayed);
994 if (!msg->msg_peerrtrcredit) {
995 /* lpni_lock protects the credit manipulation */
996 spin_lock(&lpni->lpni_lock);
997 /* lp_lock protects the lp_rtrq */
998 spin_lock(&lp->lp_lock);
1000 msg->msg_peerrtrcredit = 1;
1001 lpni->lpni_rtrcredits--;
1002 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1003 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1005 if (lpni->lpni_rtrcredits < 0) {
1006 /* must have checked eager_recv before here */
1007 LASSERT(msg->msg_rx_ready_delay);
1008 msg->msg_rx_delayed = 1;
1009 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1010 spin_unlock(&lp->lp_lock);
1011 spin_unlock(&lpni->lpni_lock);
1012 return LNET_CREDIT_WAIT;
1014 spin_unlock(&lp->lp_lock);
1015 spin_unlock(&lpni->lpni_lock);
1018 rbp = lnet_msg2bufpool(msg);
1020 if (!msg->msg_rtrcredit) {
1021 msg->msg_rtrcredit = 1;
1023 if (rbp->rbp_credits < rbp->rbp_mincredits)
1024 rbp->rbp_mincredits = rbp->rbp_credits;
1026 if (rbp->rbp_credits < 0) {
1027 /* must have checked eager_recv before here */
1028 LASSERT(msg->msg_rx_ready_delay);
1029 msg->msg_rx_delayed = 1;
1030 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1031 return LNET_CREDIT_WAIT;
1035 LASSERT(!list_empty(&rbp->rbp_bufs));
1036 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1037 list_del(&rb->rb_list);
1039 msg->msg_niov = rbp->rbp_npages;
1040 msg->msg_kiov = &rb->rb_kiov[0];
1042 /* unset the msg-rx_delayed flag since we're receiving the message */
1043 msg->msg_rx_delayed = 0;
1046 int cpt = msg->msg_rx_cpt;
1048 lnet_net_unlock(cpt);
1049 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1050 0, msg->msg_len, msg->msg_len);
1053 return LNET_CREDIT_OK;
1057 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1059 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1060 struct lnet_ni *txni = msg->msg_txni;
1061 struct lnet_msg *msg2;
1063 if (msg->msg_txcredit) {
1064 struct lnet_ni *ni = msg->msg_txni;
1065 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1067 /* give back NI txcredits */
1068 msg->msg_txcredit = 0;
1070 LASSERT((tq->tq_credits < 0) ==
1071 !list_empty(&tq->tq_delayed));
1074 atomic_inc(&ni->ni_tx_credits);
1075 if (tq->tq_credits <= 0) {
1076 msg2 = list_entry(tq->tq_delayed.next,
1077 struct lnet_msg, msg_list);
1078 list_del(&msg2->msg_list);
1080 LASSERT(msg2->msg_txni == ni);
1081 LASSERT(msg2->msg_tx_delayed);
1082 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1084 (void) lnet_post_send_locked(msg2, 1);
1088 if (msg->msg_peertxcredit) {
1089 /* give back peer txcredits */
1090 msg->msg_peertxcredit = 0;
1092 spin_lock(&txpeer->lpni_lock);
1093 LASSERT((txpeer->lpni_txcredits < 0) ==
1094 !list_empty(&txpeer->lpni_txq));
1096 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1097 LASSERT(txpeer->lpni_txqnob >= 0);
1099 txpeer->lpni_txcredits++;
1100 if (txpeer->lpni_txcredits <= 0) {
1103 msg2 = list_entry(txpeer->lpni_txq.next,
1104 struct lnet_msg, msg_list);
1105 list_del(&msg2->msg_list);
1106 spin_unlock(&txpeer->lpni_lock);
1108 LASSERT(msg2->msg_txpeer == txpeer);
1109 LASSERT(msg2->msg_tx_delayed);
1111 msg2_cpt = msg2->msg_tx_cpt;
1114 * The msg_cpt can be different from the msg2_cpt
1115 * so we need to make sure we lock the correct cpt
1117 * Once we call lnet_post_send_locked() it is no
1118 * longer safe to access msg2, since it could've
1119 * been freed by lnet_finalize(), but we still
1120 * need to relock the correct cpt, so we cache the
1121 * msg2_cpt for the purpose of the check that
1122 * follows the call to lnet_pose_send_locked().
1124 if (msg2_cpt != msg->msg_tx_cpt) {
1125 lnet_net_unlock(msg->msg_tx_cpt);
1126 lnet_net_lock(msg2_cpt);
1128 (void) lnet_post_send_locked(msg2, 1);
1129 if (msg2_cpt != msg->msg_tx_cpt) {
1130 lnet_net_unlock(msg2_cpt);
1131 lnet_net_lock(msg->msg_tx_cpt);
1134 spin_unlock(&txpeer->lpni_lock);
1139 msg->msg_txni = NULL;
1140 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1143 if (txpeer != NULL) {
1144 msg->msg_txpeer = NULL;
1145 lnet_peer_ni_decref_locked(txpeer);
1150 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1152 struct lnet_msg *msg;
1154 if (list_empty(&rbp->rbp_msgs))
1156 msg = list_entry(rbp->rbp_msgs.next,
1157 struct lnet_msg, msg_list);
1158 list_del(&msg->msg_list);
1160 (void)lnet_post_routed_recv_locked(msg, 1);
1164 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1166 struct lnet_msg *msg;
1167 struct lnet_msg *tmp;
1169 lnet_net_unlock(cpt);
1171 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1172 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1173 0, 0, 0, msg->msg_hdr.payload_length);
1174 list_del_init(&msg->msg_list);
1175 msg->msg_no_resend = true;
1176 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1177 lnet_finalize(msg, -ECANCELED);
1184 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1186 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1187 struct lnet_peer *lp;
1188 struct lnet_ni *rxni = msg->msg_rxni;
1189 struct lnet_msg *msg2;
1191 if (msg->msg_rtrcredit) {
1192 /* give back global router credits */
1193 struct lnet_rtrbuf *rb;
1194 struct lnet_rtrbufpool *rbp;
1196 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1197 * there until it gets one allocated, or aborts the wait
1199 LASSERT(msg->msg_kiov != NULL);
1201 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1204 msg->msg_kiov = NULL;
1205 msg->msg_rtrcredit = 0;
1207 LASSERT(rbp == lnet_msg2bufpool(msg));
1209 LASSERT((rbp->rbp_credits > 0) ==
1210 !list_empty(&rbp->rbp_bufs));
1212 /* If routing is now turned off, we just drop this buffer and
1213 * don't bother trying to return credits. */
1214 if (!the_lnet.ln_routing) {
1215 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1219 /* It is possible that a user has lowered the desired number of
1220 * buffers in this pool. Make sure we never put back
1221 * more buffers than the stated number. */
1222 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1223 /* Discard this buffer so we don't have too
1225 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1226 rbp->rbp_nbuffers--;
1228 list_add(&rb->rb_list, &rbp->rbp_bufs);
1230 if (rbp->rbp_credits <= 0)
1231 lnet_schedule_blocked_locked(rbp);
1236 if (msg->msg_peerrtrcredit) {
1238 LASSERT(rxpeerni->lpni_peer_net);
1239 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1241 lp = rxpeerni->lpni_peer_net->lpn_peer;
1243 /* give back peer router credits */
1244 msg->msg_peerrtrcredit = 0;
1246 spin_lock(&rxpeerni->lpni_lock);
1247 spin_lock(&lp->lp_lock);
1249 rxpeerni->lpni_rtrcredits++;
1251 /* drop all messages which are queued to be routed on that
1253 if (!the_lnet.ln_routing) {
1254 struct list_head drop;
1255 INIT_LIST_HEAD(&drop);
1256 list_splice_init(&lp->lp_rtrq, &drop);
1257 spin_unlock(&lp->lp_lock);
1258 spin_unlock(&rxpeerni->lpni_lock);
1259 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1260 } else if (!list_empty(&lp->lp_rtrq)) {
1263 msg2 = list_entry(lp->lp_rtrq.next,
1264 struct lnet_msg, msg_list);
1265 list_del(&msg2->msg_list);
1266 msg2_cpt = msg2->msg_rx_cpt;
1267 spin_unlock(&lp->lp_lock);
1268 spin_unlock(&rxpeerni->lpni_lock);
1270 * messages on the lp_rtrq can be from any NID in
1271 * the peer, which means they might have different
1272 * cpts. We need to make sure we lock the right
1275 if (msg2_cpt != msg->msg_rx_cpt) {
1276 lnet_net_unlock(msg->msg_rx_cpt);
1277 lnet_net_lock(msg2_cpt);
1279 (void) lnet_post_routed_recv_locked(msg2, 1);
1280 if (msg2_cpt != msg->msg_rx_cpt) {
1281 lnet_net_unlock(msg2_cpt);
1282 lnet_net_lock(msg->msg_rx_cpt);
1285 spin_unlock(&lp->lp_lock);
1286 spin_unlock(&rxpeerni->lpni_lock);
1290 msg->msg_rxni = NULL;
1291 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1293 if (rxpeerni != NULL) {
1294 msg->msg_rxpeer = NULL;
1295 lnet_peer_ni_decref_locked(rxpeerni);
1300 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1302 if (p1->lpni_txqnob < p2->lpni_txqnob)
1305 if (p1->lpni_txqnob > p2->lpni_txqnob)
1308 if (p1->lpni_txcredits > p2->lpni_txcredits)
1311 if (p1->lpni_txcredits < p2->lpni_txcredits)
1317 static struct lnet_peer_ni *
1318 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1319 struct lnet_peer *peer,
1320 struct lnet_peer_net *peer_net)
1323 * Look at the peer NIs for the destination peer that connect
1324 * to the chosen net. If a peer_ni is preferred when using the
1325 * best_ni to communicate, we use that one. If there is no
1326 * preferred peer_ni, or there are multiple preferred peer_ni,
1327 * the available transmit credits are used. If the transmit
1328 * credits are equal, we round-robin over the peer_ni.
1330 struct lnet_peer_ni *lpni = NULL;
1331 struct lnet_peer_ni *best_lpni = NULL;
1332 int best_lpni_credits = INT_MIN;
1333 bool preferred = false;
1335 int best_lpni_healthv = 0;
1338 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1340 * if the best_ni we've chosen aleady has this lpni
1341 * preferred, then let's use it
1344 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1346 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1347 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1352 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1355 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1356 libcfs_nid2str(lpni->lpni_nid),
1357 lpni->lpni_txcredits, best_lpni_credits,
1358 lpni->lpni_seq, best_lpni->lpni_seq);
1360 /* pick the healthiest peer ni */
1361 if (lpni_healthv < best_lpni_healthv) {
1363 } else if (lpni_healthv > best_lpni_healthv) {
1364 best_lpni_healthv = lpni_healthv;
1365 /* if this is a preferred peer use it */
1366 } else if (!preferred && ni_is_pref) {
1368 } else if (preferred && !ni_is_pref) {
1370 * this is not the preferred peer so let's ignore
1374 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1376 * We already have a peer that has more credits
1377 * available than this one. No need to consider
1378 * this peer further.
1381 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1383 * The best peer found so far and the current peer
1384 * have the same number of available credits let's
1385 * make sure to select between them using Round
1389 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1395 best_lpni_credits = lpni->lpni_txcredits;
1398 /* if we still can't find a peer ni then we can't reach it */
1400 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1401 LNET_NIDNET(dst_nid);
1402 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1403 libcfs_net2str(net_id));
1407 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1408 libcfs_nid2str(best_lpni->lpni_nid));
1414 * Prerequisite: the best_ni should already be set in the sd
1416 static inline struct lnet_peer_ni *
1417 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1420 struct lnet_peer_net *peer_net;
1423 * The gateway is Multi-Rail capable so now we must select the
1426 peer_net = lnet_peer_get_net_locked(peer, net_id);
1429 CERROR("gateway peer %s has no NI on net %s\n",
1430 libcfs_nid2str(peer->lp_primary_nid),
1431 libcfs_net2str(net_id));
1435 return lnet_select_peer_ni(sd->sd_best_ni, sd->sd_dst_nid,
1440 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2,
1441 struct lnet_peer_ni **best_lpni)
1443 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1444 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1445 struct lnet_peer *lp1 = r1->lr_gateway;
1446 struct lnet_peer *lp2 = r2->lr_gateway;
1447 struct lnet_peer_ni *lpni1;
1448 struct lnet_peer_ni *lpni2;
1449 struct lnet_send_data sd;
1452 sd.sd_best_ni = NULL;
1453 sd.sd_dst_nid = LNET_NID_ANY;
1454 lpni1 = lnet_find_best_lpni_on_net(&sd, lp1, r1->lr_lnet);
1455 lpni2 = lnet_find_best_lpni_on_net(&sd, lp2, r2->lr_lnet);
1456 LASSERT(lpni1 && lpni2);
1458 if (r1->lr_priority < r2->lr_priority) {
1463 if (r1->lr_priority > r2->lr_priority) {
1468 if (r1_hops < r2_hops) {
1473 if (r1_hops > r2_hops) {
1478 rc = lnet_compare_peers(lpni1, lpni2);
1482 } else if (rc == -1) {
1487 if (r1->lr_seq - r2->lr_seq <= 0) {
1496 static struct lnet_route *
1497 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1498 lnet_nid_t rtr_nid, struct lnet_route **prev_route,
1499 struct lnet_peer_ni **gwni)
1501 struct lnet_peer_ni *best_gw_ni = NULL;
1502 struct lnet_route *best_route;
1503 struct lnet_route *last_route;
1504 struct lnet_remotenet *rnet;
1505 struct lnet_peer *lp_best;
1506 struct lnet_route *route;
1507 struct lnet_peer *lp;
1510 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1511 * rtr_nid nid, otherwise find the best gateway I can use */
1513 rnet = lnet_find_rnet_locked(remote_net);
1518 best_route = last_route = NULL;
1519 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1520 lp = route->lr_gateway;
1522 if (!lnet_is_route_alive(route))
1525 if (lp_best == NULL) {
1526 best_route = last_route = route;
1530 /* no protection on below fields, but it's harmless */
1531 if (last_route->lr_seq - route->lr_seq < 0)
1534 rc = lnet_compare_routes(route, best_route, &best_gw_ni);
1542 *prev_route = last_route;
1548 static struct lnet_ni *
1549 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1550 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1553 struct lnet_ni *ni = NULL;
1554 unsigned int shortest_distance;
1559 * If there is no peer_ni that we can send to on this network,
1560 * then there is no point in looking for a new best_ni here.
1562 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1565 if (best_ni == NULL) {
1566 shortest_distance = UINT_MAX;
1567 best_credits = INT_MIN;
1570 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1571 best_ni->ni_dev_cpt);
1572 best_credits = atomic_read(&best_ni->ni_tx_credits);
1573 best_healthv = atomic_read(&best_ni->ni_healthv);
1576 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1577 unsigned int distance;
1582 ni_credits = atomic_read(&ni->ni_tx_credits);
1583 ni_healthv = atomic_read(&ni->ni_healthv);
1584 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1587 * calculate the distance from the CPT on which
1588 * the message memory is allocated to the CPT of
1589 * the NI's physical device
1591 distance = cfs_cpt_distance(lnet_cpt_table(),
1595 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1596 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1597 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1598 : "not seleced", best_credits, shortest_distance,
1599 (best_ni) ? best_ni->ni_seq : 0);
1602 * All distances smaller than the NUMA range
1603 * are treated equally.
1605 if (distance < lnet_numa_range)
1606 distance = lnet_numa_range;
1609 * Select on health, shorter distance, available
1610 * credits, then round-robin.
1614 } else if (ni_healthv < best_healthv) {
1616 } else if (ni_healthv > best_healthv) {
1617 best_healthv = ni_healthv;
1619 * If we're going to prefer this ni because it's
1620 * the healthiest, then we should set the
1621 * shortest_distance in the algorithm in case
1622 * there are multiple NIs with the same health but
1623 * different distances.
1625 if (distance < shortest_distance)
1626 shortest_distance = distance;
1627 } else if (distance > shortest_distance) {
1629 } else if (distance < shortest_distance) {
1630 shortest_distance = distance;
1631 } else if (ni_credits < best_credits) {
1633 } else if (ni_credits == best_credits) {
1634 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1638 best_credits = ni_credits;
1641 CDEBUG(D_NET, "selected best_ni %s\n",
1642 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1648 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1649 * because such traffic is required to perform discovery. We therefore
1650 * exclude all GET and PUT on that portal. We also exclude all ACK and
1651 * REPLY traffic, but that is because the portal is not tracked in the
1652 * message structure for these message types. We could restrict this
1653 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1656 lnet_msg_discovery(struct lnet_msg *msg)
1658 if (msg->msg_type == LNET_MSG_PUT) {
1659 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1661 } else if (msg->msg_type == LNET_MSG_GET) {
1662 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1668 #define SRC_SPEC 0x0001
1669 #define SRC_ANY 0x0002
1670 #define LOCAL_DST 0x0004
1671 #define REMOTE_DST 0x0008
1672 #define MR_DST 0x0010
1673 #define NMR_DST 0x0020
1674 #define SND_RESP 0x0040
1676 /* The following to defines are used for return codes */
1677 #define REPEAT_SEND 0x1000
1678 #define PASS_THROUGH 0x2000
1680 /* The different cases lnet_select pathway needs to handle */
1681 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1682 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1683 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1684 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1685 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1686 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1687 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1688 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1691 lnet_handle_lo_send(struct lnet_send_data *sd)
1693 struct lnet_msg *msg = sd->sd_msg;
1694 int cpt = sd->sd_cpt;
1696 /* No send credit hassles with LOLND */
1697 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1698 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1699 if (!msg->msg_routing)
1700 msg->msg_hdr.src_nid =
1701 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1702 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1703 lnet_msg_commit(msg, cpt);
1704 msg->msg_txni = the_lnet.ln_loni;
1706 return LNET_CREDIT_OK;
1710 lnet_handle_send(struct lnet_send_data *sd)
1712 struct lnet_ni *best_ni = sd->sd_best_ni;
1713 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1714 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1715 struct lnet_msg *msg = sd->sd_msg;
1717 __u32 send_case = sd->sd_send_case;
1719 __u32 routing = send_case & REMOTE_DST;
1720 struct lnet_rsp_tracker *rspt;
1723 * Increment sequence number of the selected peer so that we
1724 * pick the next one in Round Robin.
1726 best_lpni->lpni_seq++;
1729 * grab a reference on the peer_ni so it sticks around even if
1730 * we need to drop and relock the lnet_net_lock below.
1732 lnet_peer_ni_addref_locked(best_lpni);
1735 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1736 * message. This ensures that we get a CPT that is correct for
1737 * the NI when the NI has been restricted to a subset of all CPTs.
1738 * If the selected CPT differs from the one currently locked, we
1739 * must unlock and relock the lnet_net_lock(), and then check whether
1740 * the configuration has changed. We don't have a hold on the best_ni
1741 * yet, and it may have vanished.
1743 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1744 if (sd->sd_cpt != cpt2) {
1745 __u32 seq = lnet_get_dlc_seq_locked();
1746 lnet_net_unlock(sd->sd_cpt);
1748 lnet_net_lock(sd->sd_cpt);
1749 if (seq != lnet_get_dlc_seq_locked()) {
1750 lnet_peer_ni_decref_locked(best_lpni);
1756 * store the best_lpni in the message right away to avoid having
1757 * to do the same operation under different conditions
1759 msg->msg_txpeer = best_lpni;
1760 msg->msg_txni = best_ni;
1763 * grab a reference for the best_ni since now it's in use in this
1764 * send. The reference will be dropped in lnet_finalize()
1766 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1769 * Always set the target.nid to the best peer picked. Either the
1770 * NID will be one of the peer NIDs selected, or the same NID as
1771 * what was originally set in the target or it will be the NID of
1772 * a router if this message should be routed
1774 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1777 * lnet_msg_commit assigns the correct cpt to the message, which
1778 * is used to decrement the correct refcount on the ni when it's
1779 * time to return the credits
1781 lnet_msg_commit(msg, sd->sd_cpt);
1784 * If we are routing the message then we keep the src_nid that was
1785 * set by the originator. If we are not routing then we are the
1786 * originator and set it here.
1788 if (!msg->msg_routing)
1789 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1792 msg->msg_target_is_router = 1;
1793 msg->msg_target.pid = LNET_PID_LUSTRE;
1795 * since we're routing we want to ensure that the
1796 * msg_hdr.dest_nid is set to the final destination. When
1797 * the router receives this message it knows how to route
1800 * final_dst_lpni is set at the beginning of the
1801 * lnet_select_pathway() function and is never changed.
1802 * It's safe to use it here.
1804 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1807 * if we're not routing set the dest_nid to the best peer
1808 * ni NID that we picked earlier in the algorithm.
1810 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1814 * if we have response tracker block update it with the next hop
1818 rspt = msg->msg_md->md_rspt_ptr;
1820 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1821 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1822 libcfs_nid2str(rspt->rspt_next_hop_nid));
1826 rc = lnet_post_send_locked(msg, 0);
1829 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1830 libcfs_nid2str(msg->msg_hdr.src_nid),
1831 libcfs_nid2str(msg->msg_txni->ni_nid),
1832 libcfs_nid2str(sd->sd_src_nid),
1833 libcfs_nid2str(msg->msg_hdr.dest_nid),
1834 libcfs_nid2str(sd->sd_dst_nid),
1835 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1836 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1842 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1844 if (sd->sd_send_case & NMR_DST &&
1845 sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1846 sd->sd_msg->msg_type != LNET_MSG_ACK &&
1847 sd->sd_best_lpni->lpni_pref_nnids == 0) {
1848 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1849 libcfs_nid2str(sd->sd_best_ni->ni_nid),
1850 libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1851 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1852 sd->sd_best_ni->ni_nid);
1861 * use the source and destination NIDs as the pathway
1864 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1866 /* the destination lpni is set before we get here. */
1869 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1870 if (!sd->sd_best_ni) {
1871 CERROR("Can't send to %s: src %s is not a "
1872 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1873 libcfs_nid2str(sd->sd_src_nid));
1878 * the preferred NID will only be set for NMR peers
1880 lnet_set_non_mr_pref_nid(sd);
1882 return lnet_handle_send(sd);
1890 * Run the selection algorithm on the peer NIs unless we're sending
1891 * a response, in this case just send to the destination
1894 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1896 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1897 if (!sd->sd_best_ni) {
1898 CERROR("Can't send to %s: src %s is not a "
1899 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1900 libcfs_nid2str(sd->sd_src_nid));
1905 * only run the selection algorithm to pick the peer_ni if we're
1906 * sending a GET or a PUT. Responses are sent to the same
1907 * destination NID provided.
1909 if (!(sd->sd_send_case & SND_RESP)) {
1911 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1912 sd->sd_best_ni->ni_net->net_id);
1915 if (sd->sd_best_lpni &&
1916 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1917 return lnet_handle_lo_send(sd);
1918 else if (sd->sd_best_lpni)
1919 return lnet_handle_send(sd);
1921 CERROR("can't send to %s. no NI on %s\n",
1922 libcfs_nid2str(sd->sd_dst_nid),
1923 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1925 return -EHOSTUNREACH;
1929 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1930 struct lnet_peer *peer,
1931 struct lnet_peer_net *peer_net,
1935 struct lnet_net *local_net;
1936 struct lnet_ni *best_ni;
1938 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1943 * Iterate through the NIs in this local Net and select
1944 * the NI to send from. The selection is determined by
1945 * these 3 criterion in the following priority:
1947 * 2. NI available credits
1950 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1951 peer, peer_net, cpt);
1953 if (incr_seq && best_ni)
1960 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1961 struct lnet_msg *msg, lnet_nid_t rtr_nid,
1964 struct lnet_peer *peer;
1965 lnet_nid_t primary_nid;
1968 lnet_peer_ni_addref_locked(lpni);
1970 peer = lpni->lpni_peer_net->lpn_peer;
1972 if (lnet_peer_gw_discovery(peer)) {
1973 lnet_peer_ni_decref_locked(lpni);
1977 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1978 lnet_peer_ni_decref_locked(lpni);
1982 rc = lnet_discover_peer_locked(lpni, cpt, false);
1984 lnet_peer_ni_decref_locked(lpni);
1987 /* The peer may have changed. */
1988 peer = lpni->lpni_peer_net->lpn_peer;
1989 /* queue message and return */
1990 msg->msg_rtr_nid_param = rtr_nid;
1991 msg->msg_sending = 0;
1992 msg->msg_txpeer = NULL;
1993 spin_lock(&peer->lp_lock);
1994 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1995 spin_unlock(&peer->lp_lock);
1996 lnet_peer_ni_decref_locked(lpni);
1997 primary_nid = peer->lp_primary_nid;
1999 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2000 msg, libcfs_nid2str(primary_nid));
2002 return LNET_DC_WAIT;
2006 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2008 struct lnet_peer_ni **gw_lpni,
2009 struct lnet_peer **gw_peer)
2012 struct lnet_peer *gw;
2013 struct lnet_peer *lp;
2014 struct lnet_peer_net *lpn;
2015 struct lnet_peer_net *best_lpn = NULL;
2016 struct lnet_remotenet *rnet;
2017 struct lnet_route *best_route;
2018 struct lnet_route *last_route;
2019 struct lnet_peer_ni *lpni = NULL;
2020 struct lnet_peer_ni *gwni = NULL;
2021 lnet_nid_t src_nid = sd->sd_src_nid;
2023 /* we've already looked up the initial lpni using dst_nid */
2024 lpni = sd->sd_best_lpni;
2025 /* the peer tree must be in existence */
2026 LASSERT(lpni && lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
2027 lp = lpni->lpni_peer_net->lpn_peer;
2029 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2030 /* is this remote network reachable? */
2031 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2038 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2045 CERROR("peer %s has no available nets \n",
2046 libcfs_nid2str(sd->sd_dst_nid));
2047 return -EHOSTUNREACH;
2050 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd, lp, best_lpn->lpn_net_id);
2051 if (!sd->sd_best_lpni) {
2052 CERROR("peer %s down\n", libcfs_nid2str(sd->sd_dst_nid));
2053 return -EHOSTUNREACH;
2056 best_route = lnet_find_route_locked(NULL, best_lpn->lpn_net_id,
2057 sd->sd_rtr_nid, &last_route,
2060 CERROR("no route to %s from %s\n",
2061 libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
2062 return -EHOSTUNREACH;
2066 CERROR("Internal Error. Route expected to %s from %s\n",
2067 libcfs_nid2str(dst_nid),
2068 libcfs_nid2str(src_nid));
2072 gw = best_route->lr_gateway;
2073 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2076 * Discover this gateway if it hasn't already been discovered.
2077 * This means we might delay the message until discovery has
2080 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2081 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid,
2086 if (!sd->sd_best_ni)
2087 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2088 lnet_peer_get_net_locked(gw,
2089 best_route->lr_lnet),
2093 if (!sd->sd_best_ni) {
2094 CERROR("Internal Error. Expected local ni on %s "
2095 "but non found :%s\n",
2096 libcfs_net2str(best_route->lr_lnet),
2097 libcfs_nid2str(sd->sd_src_nid));
2105 * increment the sequence numbers since now we're sure we're
2106 * going to use this path
2108 LASSERT(best_route && last_route);
2109 best_route->lr_seq = last_route->lr_seq + 1;
2110 best_lpn->lpn_seq++;
2120 * Remote destination
2121 * Non-MR destination
2125 * Remote destination
2128 * The handling of these two cases is similar. Even though the destination
2129 * can be MR or non-MR, we'll deal directly with the router.
2132 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2135 struct lnet_peer_ni *gw_lpni = NULL;
2136 struct lnet_peer *gw_peer = NULL;
2139 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2140 if (!sd->sd_best_ni) {
2141 CERROR("Can't send to %s: src %s is not a "
2142 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2143 libcfs_nid2str(sd->sd_src_nid));
2147 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2152 if (sd->sd_send_case & NMR_DST)
2154 * since the final destination is non-MR let's set its preferred
2155 * NID before we send
2157 lnet_set_non_mr_pref_nid(sd);
2160 * We're going to send to the gw found so let's set its
2163 sd->sd_peer = gw_peer;
2164 sd->sd_best_lpni = gw_lpni;
2166 return lnet_handle_send(sd);
2170 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2173 struct lnet_peer_net *peer_net = NULL;
2174 struct lnet_ni *best_ni = NULL;
2177 * The peer can have multiple interfaces, some of them can be on
2178 * the local network and others on a routed network. We should
2179 * prefer the local network. However if the local network is not
2180 * available then we need to try the routed network
2183 /* go through all the peer nets and find the best_ni */
2184 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2186 * The peer's list of nets can contain non-local nets. We
2187 * want to only examine the local ones.
2189 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2191 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2192 peer_net, md_cpt, false);
2195 * if this is a discovery message and lp_disc_net_id is
2196 * specified then use that net to send the discovery on.
2198 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2204 /* increment sequence number so we can round robin */
2210 static struct lnet_ni *
2211 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2213 struct lnet_ni *best_ni = NULL;
2214 struct lnet_peer_net *peer_net;
2215 struct lnet_peer *peer = sd->sd_peer;
2216 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2217 struct lnet_peer_ni *lpni;
2218 int cpt = sd->sd_cpt;
2221 * We must use a consistent source address when sending to a
2222 * non-MR peer. However, a non-MR peer can have multiple NIDs
2223 * on multiple networks, and we may even need to talk to this
2224 * peer on multiple networks -- certain types of
2225 * load-balancing configuration do this.
2227 * So we need to pick the NI the peer prefers for this
2228 * particular network.
2231 /* Get the target peer_ni */
2232 peer_net = lnet_peer_get_net_locked(peer,
2233 LNET_NIDNET(best_lpni->lpni_nid));
2234 LASSERT(peer_net != NULL);
2235 list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2237 if (lpni->lpni_pref_nnids == 0)
2239 LASSERT(lpni->lpni_pref_nnids == 1);
2240 best_ni = lnet_nid2ni_locked(
2241 lpni->lpni_pref.nid, cpt);
2248 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2250 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2252 struct lnet_ni *best_ni = NULL;
2253 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2256 * We must use a consistent source address when sending to a
2257 * non-MR peer. However, a non-MR peer can have multiple NIDs
2258 * on multiple networks, and we may even need to talk to this
2259 * peer on multiple networks -- certain types of
2260 * load-balancing configuration do this.
2262 * So we need to pick the NI the peer prefers for this
2263 * particular network.
2266 best_ni = lnet_find_existing_preferred_best_ni(sd);
2268 /* if best_ni is still not set just pick one */
2271 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2272 sd->sd_best_lpni->lpni_peer_net,
2273 sd->sd_md_cpt, true);
2274 /* If there is no best_ni we don't have a route */
2276 CERROR("no path to %s from net %s\n",
2277 libcfs_nid2str(best_lpni->lpni_nid),
2278 libcfs_net2str(best_lpni->lpni_net->net_id));
2279 return -EHOSTUNREACH;
2283 sd->sd_best_ni = best_ni;
2285 /* Set preferred NI if necessary. */
2286 lnet_set_non_mr_pref_nid(sd);
2293 * Source not specified
2297 * always use the same source NID for NMR peers
2298 * If we've talked to that peer before then we already have a preferred
2299 * source NI associated with it. Otherwise, we select a preferred local NI
2300 * and store it in the peer
2303 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2307 /* sd->sd_best_lpni is already set to the final destination */
2310 * At this point we should've created the peer ni and peer. If we
2311 * can't find it, then something went wrong. Instead of assert
2312 * output a relevant message and fail the send
2314 if (!sd->sd_best_lpni) {
2315 CERROR("Internal fault. Unable to send msg %s to %s. "
2317 lnet_msgtyp2str(sd->sd_msg->msg_type),
2318 libcfs_nid2str(sd->sd_dst_nid));
2322 rc = lnet_select_preferred_best_ni(sd);
2324 rc = lnet_handle_send(sd);
2330 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2333 * NOTE we've already handled the remote peer case. So we only
2334 * need to worry about the local case here.
2336 * if we're sending a response, ACK or reply, we need to send it
2337 * to the destination NID given to us. At this point we already
2338 * have the peer_ni we're suppose to send to, so just find the
2339 * best_ni on the peer net and use that. Since we're sending to an
2340 * MR peer then we can just run the selection algorithm on our
2341 * local NIs and pick the best one.
2343 if (sd->sd_send_case & SND_RESP) {
2345 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2346 sd->sd_best_lpni->lpni_peer_net,
2347 sd->sd_md_cpt, true);
2349 if (!sd->sd_best_ni) {
2351 * We're not going to deal with not able to send
2352 * a response to the provided final destination
2354 CERROR("Can't send response to %s. "
2355 "No local NI available\n",
2356 libcfs_nid2str(sd->sd_dst_nid));
2357 return -EHOSTUNREACH;
2360 return lnet_handle_send(sd);
2364 * If we get here that means we're sending a fresh request, PUT or
2365 * GET, so we need to run our standard selection algorithm.
2366 * First find the best local interface that's on any of the peer's
2369 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2371 lnet_msg_discovery(sd->sd_msg));
2372 if (sd->sd_best_ni) {
2374 lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2375 sd->sd_best_ni->ni_net->net_id);
2378 * if we're successful in selecting a peer_ni on the local
2379 * network, then send to it. Otherwise fall through and
2380 * try and see if we can reach it over another routed
2383 if (sd->sd_best_lpni &&
2384 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2386 * in case we initially started with a routed
2387 * destination, let's reset to local
2389 sd->sd_send_case &= ~REMOTE_DST;
2390 sd->sd_send_case |= LOCAL_DST;
2391 return lnet_handle_lo_send(sd);
2392 } else if (sd->sd_best_lpni) {
2394 * in case we initially started with a routed
2395 * destination, let's reset to local
2397 sd->sd_send_case &= ~REMOTE_DST;
2398 sd->sd_send_case |= LOCAL_DST;
2399 return lnet_handle_send(sd);
2402 CERROR("Internal Error. Expected to have a best_lpni: "
2404 libcfs_nid2str(sd->sd_src_nid),
2405 libcfs_nid2str(sd->sd_dst_nid));
2411 * Peer doesn't have a local network. Let's see if there is
2412 * a remote network we can reach it on.
2414 return PASS_THROUGH;
2419 * Source NID not specified
2424 * Source NID not speified
2425 * Remote destination
2428 * In both of these cases if we're sending a response, ACK or REPLY, then
2429 * we need to send to the destination NID provided.
2431 * In the remote case let's deal with MR routers.
2436 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2439 struct lnet_peer *gw_peer = NULL;
2440 struct lnet_peer_ni *gw_lpni = NULL;
2443 * handle sending a response to a remote peer here so we don't
2444 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2446 if (sd->sd_send_case & REMOTE_DST &&
2447 sd->sd_send_case & SND_RESP) {
2448 struct lnet_peer_ni *gw;
2449 struct lnet_peer *gw_peer;
2451 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2454 CERROR("Can't send response to %s. "
2455 "No route available\n",
2456 libcfs_nid2str(sd->sd_dst_nid));
2457 return -EHOSTUNREACH;
2458 } else if (rc > 0) {
2462 sd->sd_best_lpni = gw;
2463 sd->sd_peer = gw_peer;
2465 return lnet_handle_send(sd);
2469 * Even though the NID for the peer might not be on a local network,
2470 * since the peer is MR there could be other interfaces on the
2471 * local network. In that case we'd still like to prefer the local
2472 * network over the routed network. If we're unable to do that
2473 * then we select the best router among the different routed networks,
2474 * and if the router is MR then we can deal with it as such.
2476 rc = lnet_handle_any_mr_dsta(sd);
2477 if (rc != PASS_THROUGH)
2481 * Now that we must route to the destination, we must consider the
2482 * MR case, where the destination has multiple interfaces, some of
2483 * which we can route to and others we do not. For this reason we
2484 * need to select the destination which we can route to and if
2485 * there are multiple, we need to round robin.
2487 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2492 sd->sd_send_case &= ~LOCAL_DST;
2493 sd->sd_send_case |= REMOTE_DST;
2495 sd->sd_peer = gw_peer;
2496 sd->sd_best_lpni = gw_lpni;
2498 return lnet_handle_send(sd);
2502 * Source not specified
2503 * Remote destination
2506 * Must send to the specified peer NID using the same source NID that
2507 * we've used before. If it's the first time to talk to that peer then
2508 * find the source NI and assign it as preferred to that peer
2511 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2514 struct lnet_peer_ni *gw_lpni = NULL;
2515 struct lnet_peer *gw_peer = NULL;
2518 * Let's set if we have a preferred NI to talk to this NMR peer
2520 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2523 * find the router and that'll find the best NI if we didn't find
2526 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2532 * set the best_ni we've chosen as the preferred one for
2535 lnet_set_non_mr_pref_nid(sd);
2537 /* we'll be sending to the gw */
2538 sd->sd_best_lpni = gw_lpni;
2539 sd->sd_peer = gw_peer;
2541 return lnet_handle_send(sd);
2545 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2548 * turn off the SND_RESP bit.
2549 * It will be checked in the case handling
2551 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2553 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2554 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2555 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2556 (send_case & MR_DST) ? "MR: " : "NMR: ",
2557 libcfs_nid2str(sd->sd_dst_nid),
2558 (send_case & LOCAL_DST) ? "local" : "routed");
2560 switch (send_case) {
2562 * For all cases where the source is specified, we should always
2563 * use the destination NID, whether it's an MR destination or not,
2564 * since we're continuing a series of related messages for the
2567 case SRC_SPEC_LOCAL_NMR_DST:
2568 return lnet_handle_spec_local_nmr_dst(sd);
2569 case SRC_SPEC_LOCAL_MR_DST:
2570 return lnet_handle_spec_local_mr_dst(sd);
2571 case SRC_SPEC_ROUTER_NMR_DST:
2572 case SRC_SPEC_ROUTER_MR_DST:
2573 return lnet_handle_spec_router_dst(sd);
2574 case SRC_ANY_LOCAL_NMR_DST:
2575 return lnet_handle_any_local_nmr_dst(sd);
2576 case SRC_ANY_LOCAL_MR_DST:
2577 case SRC_ANY_ROUTER_MR_DST:
2578 return lnet_handle_any_mr_dst(sd);
2579 case SRC_ANY_ROUTER_NMR_DST:
2580 return lnet_handle_any_router_nmr_dst(sd);
2582 CERROR("Unknown send case\n");
2588 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2589 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2591 struct lnet_peer_ni *lpni;
2592 struct lnet_peer *peer;
2593 struct lnet_send_data send_data;
2596 __u32 send_case = 0;
2598 memset(&send_data, 0, sizeof(send_data));
2601 * get an initial CPT to use for locking. The idea here is not to
2602 * serialize the calls to select_pathway, so that as many
2603 * operations can run concurrently as possible. To do that we use
2604 * the CPT where this call is being executed. Later on when we
2605 * determine the CPT to use in lnet_message_commit, we switch the
2606 * lock and check if there was any configuration change. If none,
2607 * then we proceed, if there is, then we restart the operation.
2609 cpt = lnet_net_lock_current();
2611 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2612 if (md_cpt == CFS_CPT_ANY)
2618 * If we're being asked to send to the loopback interface, there
2619 * is no need to go through any selection. We can just shortcut
2620 * the entire process and send over lolnd
2622 send_data.sd_msg = msg;
2623 send_data.sd_cpt = cpt;
2624 if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2625 rc = lnet_handle_lo_send(&send_data);
2626 lnet_net_unlock(cpt);
2631 * find an existing peer_ni, or create one and mark it as having been
2632 * created due to network traffic. This call will create the
2633 * peer->peer_net->peer_ni tree.
2635 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2637 lnet_net_unlock(cpt);
2638 return PTR_ERR(lpni);
2642 * Cache the original src_nid. If we need to resend the message
2643 * then we'll need to know whether the src_nid was originally
2644 * specified for this message. If it was originally specified,
2645 * then we need to keep using the same src_nid since it's
2646 * continuing the same sequence of messages.
2648 msg->msg_src_nid_param = src_nid;
2651 * Now that we have a peer_ni, check if we want to discover
2652 * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2653 * trigger discovery.
2655 peer = lpni->lpni_peer_net->lpn_peer;
2656 rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2658 lnet_peer_ni_decref_locked(lpni);
2659 lnet_net_unlock(cpt);
2662 lnet_peer_ni_decref_locked(lpni);
2665 * Identify the different send cases
2667 if (src_nid == LNET_NID_ANY)
2668 send_case |= SRC_ANY;
2670 send_case |= SRC_SPEC;
2672 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2673 send_case |= LOCAL_DST;
2675 send_case |= REMOTE_DST;
2678 * if this is a non-MR peer or if we're recovering a peer ni then
2679 * let's consider this an NMR case so we can hit the destination
2682 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2683 send_case |= NMR_DST;
2685 send_case |= MR_DST;
2687 if (msg->msg_type == LNET_MSG_REPLY ||
2688 msg->msg_type == LNET_MSG_ACK)
2689 send_case |= SND_RESP;
2691 /* assign parameters to the send_data */
2692 send_data.sd_rtr_nid = rtr_nid;
2693 send_data.sd_src_nid = src_nid;
2694 send_data.sd_dst_nid = dst_nid;
2695 send_data.sd_best_lpni = lpni;
2697 * keep a pointer to the final destination in case we're going to
2698 * route, so we'll need to access it later
2700 send_data.sd_final_dst_lpni = lpni;
2701 send_data.sd_peer = peer;
2702 send_data.sd_md_cpt = md_cpt;
2703 send_data.sd_send_case = send_case;
2705 rc = lnet_handle_send_case_locked(&send_data);
2708 * Update the local cpt since send_data.sd_cpt might've been
2709 * updated as a result of calling lnet_handle_send_case_locked().
2711 cpt = send_data.sd_cpt;
2713 if (rc == REPEAT_SEND)
2716 lnet_net_unlock(cpt);
2722 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2724 lnet_nid_t dst_nid = msg->msg_target.nid;
2728 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2729 * but we might want to use pre-determined router for ACK/REPLY
2732 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2733 LASSERT(msg->msg_txpeer == NULL);
2734 LASSERT(msg->msg_txni == NULL);
2735 LASSERT(!msg->msg_sending);
2736 LASSERT(!msg->msg_target_is_router);
2737 LASSERT(!msg->msg_receiving);
2739 msg->msg_sending = 1;
2741 LASSERT(!msg->msg_tx_committed);
2743 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2745 if (rc == -EHOSTUNREACH)
2746 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2748 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2752 if (rc == LNET_CREDIT_OK)
2753 lnet_ni_send(msg->msg_txni, msg);
2755 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2759 enum lnet_mt_event_type {
2760 MT_TYPE_LOCAL_NI = 0,
2764 struct lnet_mt_event_info {
2765 enum lnet_mt_event_type mt_type;
2769 /* called with res_lock held */
2771 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2773 struct lnet_rsp_tracker *rspt;
2776 * msg has a refcount on the MD so the MD is not going away.
2777 * The rspt queue for the cpt is protected by
2778 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2780 if (!md->md_rspt_ptr)
2783 rspt = md->md_rspt_ptr;
2784 md->md_rspt_ptr = NULL;
2787 LASSERT(rspt->rspt_cpt == cpt);
2790 * invalidate the handle to indicate that a response has been
2791 * received, which will then lead the monitor thread to clean up
2794 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2798 lnet_finalize_expired_responses(bool force)
2800 struct lnet_libmd *md;
2801 struct list_head local_queue;
2802 struct lnet_rsp_tracker *rspt, *tmp;
2805 if (the_lnet.ln_mt_rstq == NULL)
2808 cfs_cpt_for_each(i, lnet_cpt_table()) {
2809 INIT_LIST_HEAD(&local_queue);
2812 if (!the_lnet.ln_mt_rstq[i]) {
2816 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2819 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2821 * The rspt mdh will be invalidated when a response
2822 * is received or whenever we want to discard the
2823 * block the monitor thread will walk the queue
2824 * and clean up any rsts with an invalid mdh.
2825 * The monitor thread will walk the queue until
2826 * the first unexpired rspt block. This means that
2827 * some rspt blocks which received their
2828 * corresponding responses will linger in the
2829 * queue until they are cleaned up eventually.
2832 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2834 list_del_init(&rspt->rspt_on_list);
2835 lnet_rspt_free(rspt, i);
2839 if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
2841 struct lnet_peer_ni *lpni;
2844 md = lnet_handle2md(&rspt->rspt_mdh);
2846 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2848 list_del_init(&rspt->rspt_on_list);
2849 lnet_rspt_free(rspt, i);
2852 LASSERT(md->md_rspt_ptr == rspt);
2853 md->md_rspt_ptr = NULL;
2857 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2860 list_del_init(&rspt->rspt_on_list);
2862 nid = rspt->rspt_next_hop_nid;
2864 CNETERR("Response timed out: md = %p: nid = %s\n",
2865 md, libcfs_nid2str(nid));
2866 LNetMDUnlink(rspt->rspt_mdh);
2867 lnet_rspt_free(rspt, i);
2870 * If there is a timeout on the response
2871 * from the next hop decrement its health
2872 * value so that we don't use it
2875 lpni = lnet_find_peer_ni_locked(nid);
2877 lnet_handle_remote_failure_locked(lpni);
2878 lnet_peer_ni_decref_locked(lpni);
2888 if (!list_empty(&local_queue))
2889 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2895 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2897 struct lnet_msg *msg;
2899 while (!list_empty(resendq)) {
2900 struct lnet_peer_ni *lpni;
2902 msg = list_entry(resendq->next, struct lnet_msg,
2905 list_del_init(&msg->msg_list);
2907 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2909 lnet_net_unlock(cpt);
2910 CERROR("Expected that a peer is already created for %s\n",
2911 libcfs_nid2str(msg->msg_hdr.dest_nid));
2912 msg->msg_no_resend = true;
2913 lnet_finalize(msg, -EFAULT);
2916 struct lnet_peer *peer;
2918 lnet_nid_t src_nid = LNET_NID_ANY;
2921 * if this message is not being routed and the
2922 * peer is non-MR then we must use the same
2923 * src_nid that was used in the original send.
2924 * Otherwise if we're routing the message (IE
2925 * we're a router) then we can use any of our
2926 * local interfaces. It doesn't matter to the
2927 * final destination.
2929 peer = lpni->lpni_peer_net->lpn_peer;
2930 if (!msg->msg_routing &&
2931 !lnet_peer_is_multi_rail(peer))
2932 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
2935 * If we originally specified a src NID, then we
2936 * must attempt to reuse it in the resend as well.
2938 if (msg->msg_src_nid_param != LNET_NID_ANY)
2939 src_nid = msg->msg_src_nid_param;
2940 lnet_peer_ni_decref_locked(lpni);
2942 lnet_net_unlock(cpt);
2943 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
2944 libcfs_nid2str(src_nid),
2945 libcfs_id2str(msg->msg_target),
2946 lnet_msgtyp2str(msg->msg_type),
2948 msg->msg_retry_count);
2949 rc = lnet_send(src_nid, msg, LNET_NID_ANY);
2951 CERROR("Error sending %s to %s: %d\n",
2952 lnet_msgtyp2str(msg->msg_type),
2953 libcfs_id2str(msg->msg_target), rc);
2954 msg->msg_no_resend = true;
2955 lnet_finalize(msg, rc);
2959 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
2965 lnet_resend_pending_msgs(void)
2969 cfs_cpt_for_each(i, lnet_cpt_table()) {
2971 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
2976 /* called with cpt and ni_lock held */
2978 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
2980 struct lnet_handle_md recovery_mdh;
2982 LNetInvalidateMDHandle(&recovery_mdh);
2984 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
2986 recovery_mdh = ni->ni_ping_mdh;
2987 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2990 lnet_net_unlock(cpt);
2991 if (!LNetMDHandleIsInvalid(recovery_mdh))
2992 LNetMDUnlink(recovery_mdh);
2998 lnet_recover_local_nis(void)
3000 struct lnet_mt_event_info *ev_info;
3001 struct list_head processed_list;
3002 struct list_head local_queue;
3003 struct lnet_handle_md mdh;
3004 struct lnet_ni *tmp;
3010 INIT_LIST_HEAD(&local_queue);
3011 INIT_LIST_HEAD(&processed_list);
3014 * splice the recovery queue on a local queue. We will iterate
3015 * through the local queue and update it as needed. Once we're
3016 * done with the traversal, we'll splice the local queue back on
3017 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3018 * will be traversed in the next iteration.
3021 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3025 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3027 * if an NI is being deleted or it is now healthy, there
3028 * is no need to keep it around in the recovery queue.
3029 * The monitor thread is the only thread responsible for
3030 * removing the NI from the recovery queue.
3031 * Multiple threads can be adding NIs to the recovery
3034 healthv = atomic_read(&ni->ni_healthv);
3038 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3039 healthv == LNET_MAX_HEALTH_VALUE) {
3040 list_del_init(&ni->ni_recovery);
3041 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3043 lnet_ni_decref_locked(ni, 0);
3049 * if the local NI failed recovery we must unlink the md.
3050 * But we want to keep the local_ni on the recovery queue
3051 * so we can continue the attempts to recover it.
3053 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3054 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3055 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3062 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3063 libcfs_nid2str(ni->ni_nid));
3066 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3067 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3070 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3072 CERROR("out of memory. Can't recover %s\n",
3073 libcfs_nid2str(ni->ni_nid));
3075 ni->ni_recovery_state &=
3076 ~LNET_NI_RECOVERY_PENDING;
3081 mdh = ni->ni_ping_mdh;
3083 * Invalidate the ni mdh in case it's deleted.
3084 * We'll unlink the mdh in this case below.
3086 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3090 * remove the NI from the local queue and drop the
3091 * reference count to it while we're recovering
3092 * it. The reason for that, is that the NI could
3093 * be deleted, and the way the code is structured
3094 * is if we don't drop the NI, then the deletion
3095 * code will enter a loop waiting for the
3096 * reference count to be removed while holding the
3097 * ln_mutex_lock(). When we look up the peer to
3098 * send to in lnet_select_pathway() we will try to
3099 * lock the ln_mutex_lock() as well, leading to
3100 * a deadlock. By dropping the refcount and
3101 * removing it from the list, we allow for the NI
3102 * to be removed, then we use the cached NID to
3103 * look it up again. If it's gone, then we just
3104 * continue examining the rest of the queue.
3107 list_del_init(&ni->ni_recovery);
3108 lnet_ni_decref_locked(ni, 0);
3111 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3112 ev_info->mt_nid = nid;
3113 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3114 ev_info, the_lnet.ln_mt_eqh, true);
3115 /* lookup the nid again */
3117 ni = lnet_nid2ni_locked(nid, 0);
3120 * the NI has been deleted when we dropped
3128 * Same note as in lnet_recover_peer_nis(). When
3129 * we're sending the ping, the NI is free to be
3130 * deleted or manipulated. By this point it
3131 * could've been added back on the recovery queue,
3132 * and a refcount taken on it.
3133 * So we can't just add it blindly again or we'll
3134 * corrupt the queue. We must check under lock if
3135 * it's not on any list and if not then add it
3136 * to the processed list, which will eventually be
3137 * spliced back on to the recovery queue.
3139 ni->ni_ping_mdh = mdh;
3140 if (list_empty(&ni->ni_recovery)) {
3141 list_add_tail(&ni->ni_recovery, &processed_list);
3142 lnet_ni_addref_locked(ni, 0);
3148 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3154 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3155 * reexamined in the next iteration.
3157 list_splice_init(&processed_list, &local_queue);
3159 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3163 static struct list_head **
3164 lnet_create_array_of_queues(void)
3166 struct list_head **qs;
3167 struct list_head *q;
3170 qs = cfs_percpt_alloc(lnet_cpt_table(),
3171 sizeof(struct list_head));
3173 CERROR("Failed to allocate queues\n");
3177 cfs_percpt_for_each(q, i, qs)
3184 lnet_resendqs_create(void)
3186 struct list_head **resendqs;
3187 resendqs = lnet_create_array_of_queues();
3192 lnet_net_lock(LNET_LOCK_EX);
3193 the_lnet.ln_mt_resendqs = resendqs;
3194 lnet_net_unlock(LNET_LOCK_EX);
3200 lnet_clean_local_ni_recoveryq(void)
3204 /* This is only called when the monitor thread has stopped */
3207 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3208 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3209 struct lnet_ni, ni_recovery);
3210 list_del_init(&ni->ni_recovery);
3212 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3214 lnet_ni_decref_locked(ni, 0);
3221 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3224 struct lnet_handle_md recovery_mdh;
3226 LNetInvalidateMDHandle(&recovery_mdh);
3228 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3229 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3230 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3232 spin_unlock(&lpni->lpni_lock);
3233 lnet_net_unlock(cpt);
3234 if (!LNetMDHandleIsInvalid(recovery_mdh))
3235 LNetMDUnlink(recovery_mdh);
3237 spin_lock(&lpni->lpni_lock);
3241 lnet_clean_peer_ni_recoveryq(void)
3243 struct lnet_peer_ni *lpni, *tmp;
3245 lnet_net_lock(LNET_LOCK_EX);
3247 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3249 list_del_init(&lpni->lpni_recovery);
3250 spin_lock(&lpni->lpni_lock);
3251 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3252 spin_unlock(&lpni->lpni_lock);