4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/lnet/lib-move.c
34 * Data movement routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/pagemap.h>
41 #include <lnet/lib-lnet.h>
42 #include <linux/nsproxy.h>
43 #include <net/net_namespace.h>
45 static int local_nid_dist_zero = 1;
46 module_param(local_nid_dist_zero, int, 0444);
47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
49 struct lnet_send_data {
50 struct lnet_ni *sd_best_ni;
51 struct lnet_peer_ni *sd_best_lpni;
52 struct lnet_peer_ni *sd_final_dst_lpni;
53 struct lnet_peer *sd_peer;
54 struct lnet_peer *sd_gw_peer;
55 struct lnet_peer_ni *sd_gw_lpni;
56 struct lnet_peer_net *sd_peer_net;
57 struct lnet_msg *sd_msg;
58 lnet_nid_t sd_dst_nid;
59 lnet_nid_t sd_src_nid;
60 lnet_nid_t sd_rtr_nid;
67 lnet_msg_is_response(struct lnet_msg *msg)
69 return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
73 lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
75 if (md_options & LNET_MD_NO_TRACK_RESPONSE)
76 /* Explicitly disabled in MD options */
79 if (md_options & LNET_MD_TRACK_RESPONSE)
80 /* Explicity enabled in MD options */
83 if (lnet_response_tracking == 3)
84 /* Enabled for all message types */
87 if (msg_type == LNET_MSG_PUT)
88 return lnet_response_tracking == 2;
90 if (msg_type == LNET_MSG_GET)
91 return lnet_response_tracking == 1;
96 static inline struct lnet_comm_count *
97 get_stats_counts(struct lnet_element_stats *stats,
98 enum lnet_stats_type stats_type)
100 switch (stats_type) {
101 case LNET_STATS_TYPE_SEND:
102 return &stats->el_send_stats;
103 case LNET_STATS_TYPE_RECV:
104 return &stats->el_recv_stats;
105 case LNET_STATS_TYPE_DROP:
106 return &stats->el_drop_stats;
108 CERROR("Unknown stats type\n");
114 void lnet_incr_stats(struct lnet_element_stats *stats,
115 enum lnet_msg_type msg_type,
116 enum lnet_stats_type stats_type)
118 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
124 atomic_inc(&counts->co_ack_count);
127 atomic_inc(&counts->co_put_count);
130 atomic_inc(&counts->co_get_count);
133 atomic_inc(&counts->co_reply_count);
136 atomic_inc(&counts->co_hello_count);
139 CERROR("There is a BUG in the code. Unknown message type\n");
144 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
145 enum lnet_stats_type stats_type)
147 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
151 return (atomic_read(&counts->co_ack_count) +
152 atomic_read(&counts->co_put_count) +
153 atomic_read(&counts->co_get_count) +
154 atomic_read(&counts->co_reply_count) +
155 atomic_read(&counts->co_hello_count));
158 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
159 struct lnet_comm_count *counts)
161 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
162 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
163 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
164 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
165 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
168 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
169 struct lnet_element_stats *stats)
171 struct lnet_comm_count *counts;
176 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
179 assign_stats(&msg_stats->im_send_stats, counts);
181 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
184 assign_stats(&msg_stats->im_recv_stats, counts);
186 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
189 assign_stats(&msg_stats->im_drop_stats, counts);
193 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
195 struct lnet_test_peer *tp;
196 struct list_head *el;
197 struct list_head *next;
200 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
201 if (threshold != 0) {
202 /* Adding a new entry */
203 LIBCFS_ALLOC(tp, sizeof(*tp));
208 tp->tp_threshold = threshold;
211 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
218 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
219 tp = list_entry(el, struct lnet_test_peer, tp_list);
221 if (tp->tp_threshold == 0 || /* needs culling anyway */
222 nid == LNET_NID_ANY || /* removing all entries */
223 tp->tp_nid == nid) { /* matched this one */
224 list_move(&tp->tp_list, &cull);
230 while (!list_empty(&cull)) {
231 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
233 list_del(&tp->tp_list);
234 LIBCFS_FREE(tp, sizeof(*tp));
240 fail_peer (lnet_nid_t nid, int outgoing)
242 struct lnet_test_peer *tp;
243 struct list_head *el;
244 struct list_head *next;
248 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
251 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
252 tp = list_entry(el, struct lnet_test_peer, tp_list);
254 if (tp->tp_threshold == 0) {
257 /* only cull zombies on outgoing tests,
258 * since we may be at interrupt priority on
259 * incoming messages. */
260 list_move(&tp->tp_list, &cull);
265 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
266 nid == tp->tp_nid) { /* fail this peer */
269 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
272 tp->tp_threshold == 0) {
274 list_move(&tp->tp_list, &cull);
283 while (!list_empty(&cull)) {
284 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
285 list_del(&tp->tp_list);
287 LIBCFS_FREE(tp, sizeof(*tp));
294 lnet_iov_nob(unsigned int niov, struct kvec *iov)
296 unsigned int nob = 0;
298 LASSERT(niov == 0 || iov != NULL);
300 nob += (iov++)->iov_len;
304 EXPORT_SYMBOL(lnet_iov_nob);
307 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
308 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
311 /* NB diov, siov are READ-ONLY */
312 unsigned int this_nob;
317 /* skip complete frags before 'doffset' */
319 while (doffset >= diov->iov_len) {
320 doffset -= diov->iov_len;
326 /* skip complete frags before 'soffset' */
328 while (soffset >= siov->iov_len) {
329 soffset -= siov->iov_len;
338 this_nob = min3((unsigned int)diov->iov_len - doffset,
339 (unsigned int)siov->iov_len - soffset,
342 memcpy((char *)diov->iov_base + doffset,
343 (char *)siov->iov_base + soffset, this_nob);
346 if (diov->iov_len > doffset + this_nob) {
354 if (siov->iov_len > soffset + this_nob) {
363 EXPORT_SYMBOL(lnet_copy_iov2iov);
366 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
368 unsigned int nob = 0;
370 LASSERT(niov == 0 || kiov != NULL);
372 nob += (kiov++)->bv_len;
376 EXPORT_SYMBOL(lnet_kiov_nob);
379 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
380 unsigned int doffset,
381 unsigned int nsiov, struct bio_vec *siov,
382 unsigned int soffset,
385 /* NB diov, siov are READ-ONLY */
386 unsigned int this_nob;
393 LASSERT (!in_interrupt ());
396 while (doffset >= diov->bv_len) {
397 doffset -= diov->bv_len;
404 while (soffset >= siov->bv_len) {
405 soffset -= siov->bv_len;
414 this_nob = min3(diov->bv_len - doffset,
415 siov->bv_len - soffset,
419 daddr = ((char *)kmap(diov->bv_page)) +
420 diov->bv_offset + doffset;
422 saddr = ((char *)kmap(siov->bv_page)) +
423 siov->bv_offset + soffset;
425 /* Vanishing risk of kmap deadlock when mapping 2 pages.
426 * However in practice at least one of the kiovs will be mapped
427 * kernel pages and the map/unmap will be NOOPs */
429 memcpy (daddr, saddr, this_nob);
432 if (diov->bv_len > doffset + this_nob) {
436 kunmap(diov->bv_page);
443 if (siov->bv_len > soffset + this_nob) {
447 kunmap(siov->bv_page);
456 kunmap(diov->bv_page);
458 kunmap(siov->bv_page);
460 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
463 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
464 unsigned int nkiov, struct bio_vec *kiov,
465 unsigned int kiovoffset,
468 /* NB iov, kiov are READ-ONLY */
469 unsigned int this_nob;
475 LASSERT (!in_interrupt ());
478 while (iovoffset >= iov->iov_len) {
479 iovoffset -= iov->iov_len;
486 while (kiovoffset >= kiov->bv_len) {
487 kiovoffset -= kiov->bv_len;
496 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
497 (unsigned int)kiov->bv_len - kiovoffset,
501 addr = ((char *)kmap(kiov->bv_page)) +
502 kiov->bv_offset + kiovoffset;
504 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
507 if (iov->iov_len > iovoffset + this_nob) {
508 iovoffset += this_nob;
515 if (kiov->bv_len > kiovoffset + this_nob) {
517 kiovoffset += this_nob;
519 kunmap(kiov->bv_page);
529 kunmap(kiov->bv_page);
531 EXPORT_SYMBOL(lnet_copy_kiov2iov);
534 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
535 unsigned int kiovoffset,
536 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
539 /* NB kiov, iov are READ-ONLY */
540 unsigned int this_nob;
546 LASSERT (!in_interrupt ());
549 while (kiovoffset >= kiov->bv_len) {
550 kiovoffset -= kiov->bv_len;
557 while (iovoffset >= iov->iov_len) {
558 iovoffset -= iov->iov_len;
567 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
568 (unsigned int)iov->iov_len - iovoffset,
572 addr = ((char *)kmap(kiov->bv_page)) +
573 kiov->bv_offset + kiovoffset;
575 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
578 if (kiov->bv_len > kiovoffset + this_nob) {
580 kiovoffset += this_nob;
582 kunmap(kiov->bv_page);
589 if (iov->iov_len > iovoffset + this_nob) {
590 iovoffset += this_nob;
599 kunmap(kiov->bv_page);
601 EXPORT_SYMBOL(lnet_copy_iov2kiov);
604 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
605 int src_niov, struct bio_vec *src,
606 unsigned int offset, unsigned int len)
608 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
609 * for exactly 'len' bytes, and return the number of entries.
610 * NB not destructive to 'src' */
611 unsigned int frag_len;
614 if (len == 0) /* no data => */
615 return (0); /* no frags */
617 LASSERT(src_niov > 0);
618 while (offset >= src->bv_len) { /* skip initial frags */
619 offset -= src->bv_len;
622 LASSERT(src_niov > 0);
627 LASSERT(src_niov > 0);
628 LASSERT((int)niov <= dst_niov);
630 frag_len = src->bv_len - offset;
631 dst->bv_page = src->bv_page;
632 dst->bv_offset = src->bv_offset + offset;
634 if (len <= frag_len) {
636 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
640 dst->bv_len = frag_len;
641 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
651 EXPORT_SYMBOL(lnet_extract_kiov);
654 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
655 int delayed, unsigned int offset, unsigned int mlen,
658 unsigned int niov = 0;
659 struct kvec *iov = NULL;
660 struct bio_vec *kiov = NULL;
663 LASSERT (!in_interrupt ());
664 LASSERT (mlen == 0 || msg != NULL);
667 LASSERT(msg->msg_receiving);
668 LASSERT(!msg->msg_sending);
669 LASSERT(rlen == msg->msg_len);
670 LASSERT(mlen <= msg->msg_len);
671 LASSERT(msg->msg_offset == offset);
672 LASSERT(msg->msg_wanted == mlen);
674 msg->msg_receiving = 0;
677 niov = msg->msg_niov;
678 kiov = msg->msg_kiov;
681 LASSERT ((iov == NULL) != (kiov == NULL));
685 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
686 niov, kiov, offset, mlen,
689 lnet_finalize(msg, rc);
693 lnet_setpayloadbuffer(struct lnet_msg *msg)
695 struct lnet_libmd *md = msg->msg_md;
697 LASSERT(msg->msg_len > 0);
698 LASSERT(!msg->msg_routing);
700 LASSERT(msg->msg_niov == 0);
701 LASSERT(msg->msg_kiov == NULL);
703 msg->msg_niov = md->md_niov;
704 msg->msg_kiov = md->md_kiov;
708 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
709 unsigned int offset, unsigned int len)
711 msg->msg_type = type;
712 msg->msg_target = target;
714 msg->msg_offset = offset;
717 lnet_setpayloadbuffer(msg);
719 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
720 msg->msg_hdr.type = cpu_to_le32(type);
721 /* dest_nid will be overwritten by lnet_select_pathway() */
722 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
723 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
724 /* src_nid will be set later */
725 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
726 msg->msg_hdr.payload_length = cpu_to_le32(len);
730 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
732 void *priv = msg->msg_private;
735 LASSERT(!in_interrupt());
736 LASSERT(ni->ni_nid == LNET_NID_LO_0 ||
737 (msg->msg_txcredit && msg->msg_peertxcredit));
739 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
741 msg->msg_no_resend = true;
742 lnet_finalize(msg, rc);
747 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
751 LASSERT(!msg->msg_sending);
752 LASSERT(msg->msg_receiving);
753 LASSERT(!msg->msg_rx_ready_delay);
754 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
756 msg->msg_rx_ready_delay = 1;
757 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
760 CERROR("recv from %s / send to %s aborted: "
761 "eager_recv failed %d\n",
762 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
763 libcfs_id2str(msg->msg_target), rc);
764 LASSERT(rc < 0); /* required by my callers */
771 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
775 deadline = lpni->lpni_last_alive +
776 lpni->lpni_net->net_tunables.lct_peer_timeout;
779 * assume peer_ni is alive as long as we're within the configured
788 /* NB: returns 1 when alive, 0 when dead, negative when error;
789 * may drop the lnet_net_lock */
791 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
792 struct lnet_msg *msg)
794 time64_t now = ktime_get_seconds();
796 if (!lnet_peer_aliveness_enabled(lpni))
800 * If we're resending a message, let's attempt to send it even if
801 * the peer is down to fulfill our resend quota on the message
803 if (msg->msg_retry_count > 0)
806 /* try and send recovery messages irregardless */
807 if (msg->msg_recovery)
810 /* always send any responses */
811 if (lnet_msg_is_response(msg))
814 if (!lnet_is_peer_deadline_passed(lpni, now))
817 return lnet_is_peer_ni_alive(lpni);
821 * \param msg The message to be sent.
822 * \param do_send True if lnet_ni_send() should be called in this function.
823 * lnet_send() is going to lnet_net_unlock immediately after this, so
824 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
826 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
827 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
828 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
829 * \retval -ECANCELED If the MD of the message has been unlinked.
832 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
834 struct lnet_peer_ni *lp = msg->msg_txpeer;
835 struct lnet_ni *ni = msg->msg_txni;
836 int cpt = msg->msg_tx_cpt;
837 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
839 /* non-lnet_send() callers have checked before */
840 LASSERT(!do_send || msg->msg_tx_delayed);
841 LASSERT(!msg->msg_receiving);
842 LASSERT(msg->msg_tx_committed);
843 /* can't get here if we're sending to the loopback interface */
844 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
846 /* NB 'lp' is always the next hop */
847 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
848 lnet_peer_alive_locked(ni, lp, msg) == 0) {
849 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
850 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
852 lnet_net_unlock(cpt);
854 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
856 LNET_STATS_TYPE_DROP);
858 lnet_incr_stats(&msg->msg_txni->ni_stats,
860 LNET_STATS_TYPE_DROP);
862 CNETERR("Dropping message for %s: peer not alive\n",
863 libcfs_id2str(msg->msg_target));
864 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
866 lnet_finalize(msg, -EHOSTUNREACH);
869 return -EHOSTUNREACH;
872 if (msg->msg_md != NULL &&
873 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
874 lnet_net_unlock(cpt);
876 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
877 "called on the MD/ME.\n",
878 libcfs_id2str(msg->msg_target));
880 msg->msg_no_resend = true;
881 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
882 msg, libcfs_id2str(msg->msg_target));
883 lnet_finalize(msg, -ECANCELED);
890 if (!msg->msg_peertxcredit) {
891 spin_lock(&lp->lpni_lock);
892 LASSERT((lp->lpni_txcredits < 0) ==
893 !list_empty(&lp->lpni_txq));
895 msg->msg_peertxcredit = 1;
896 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
897 lp->lpni_txcredits--;
899 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
900 lp->lpni_mintxcredits = lp->lpni_txcredits;
902 if (lp->lpni_txcredits < 0) {
903 msg->msg_tx_delayed = 1;
904 list_add_tail(&msg->msg_list, &lp->lpni_txq);
905 spin_unlock(&lp->lpni_lock);
906 return LNET_CREDIT_WAIT;
908 spin_unlock(&lp->lpni_lock);
911 if (!msg->msg_txcredit) {
912 LASSERT((tq->tq_credits < 0) ==
913 !list_empty(&tq->tq_delayed));
915 msg->msg_txcredit = 1;
917 atomic_dec(&ni->ni_tx_credits);
919 if (tq->tq_credits < tq->tq_credits_min)
920 tq->tq_credits_min = tq->tq_credits;
922 if (tq->tq_credits < 0) {
923 msg->msg_tx_delayed = 1;
924 list_add_tail(&msg->msg_list, &tq->tq_delayed);
925 return LNET_CREDIT_WAIT;
929 /* unset the tx_delay flag as we're going to send it now */
930 msg->msg_tx_delayed = 0;
933 lnet_net_unlock(cpt);
934 lnet_ni_send(ni, msg);
937 return LNET_CREDIT_OK;
941 static struct lnet_rtrbufpool *
942 lnet_msg2bufpool(struct lnet_msg *msg)
944 struct lnet_rtrbufpool *rbp;
947 LASSERT(msg->msg_rx_committed);
949 cpt = msg->msg_rx_cpt;
950 rbp = &the_lnet.ln_rtrpools[cpt][0];
952 LASSERT(msg->msg_len <= LNET_MTU);
953 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
955 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
962 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
964 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
965 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
966 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
967 * received or OK to receive */
968 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
969 struct lnet_peer *lp;
970 struct lnet_rtrbufpool *rbp;
971 struct lnet_rtrbuf *rb;
973 LASSERT(msg->msg_kiov == NULL);
974 LASSERT(msg->msg_niov == 0);
975 LASSERT(msg->msg_routing);
976 LASSERT(msg->msg_receiving);
977 LASSERT(!msg->msg_sending);
978 LASSERT(lpni->lpni_peer_net);
979 LASSERT(lpni->lpni_peer_net->lpn_peer);
981 lp = lpni->lpni_peer_net->lpn_peer;
983 /* non-lnet_parse callers only receive delayed messages */
984 LASSERT(!do_recv || msg->msg_rx_delayed);
986 if (!msg->msg_peerrtrcredit) {
987 /* lpni_lock protects the credit manipulation */
988 spin_lock(&lpni->lpni_lock);
990 msg->msg_peerrtrcredit = 1;
991 lpni->lpni_rtrcredits--;
992 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
993 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
995 if (lpni->lpni_rtrcredits < 0) {
996 spin_unlock(&lpni->lpni_lock);
997 /* must have checked eager_recv before here */
998 LASSERT(msg->msg_rx_ready_delay);
999 msg->msg_rx_delayed = 1;
1000 /* lp_lock protects the lp_rtrq */
1001 spin_lock(&lp->lp_lock);
1002 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1003 spin_unlock(&lp->lp_lock);
1004 return LNET_CREDIT_WAIT;
1006 spin_unlock(&lpni->lpni_lock);
1009 rbp = lnet_msg2bufpool(msg);
1011 if (!msg->msg_rtrcredit) {
1012 msg->msg_rtrcredit = 1;
1014 if (rbp->rbp_credits < rbp->rbp_mincredits)
1015 rbp->rbp_mincredits = rbp->rbp_credits;
1017 if (rbp->rbp_credits < 0) {
1018 /* must have checked eager_recv before here */
1019 LASSERT(msg->msg_rx_ready_delay);
1020 msg->msg_rx_delayed = 1;
1021 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1022 return LNET_CREDIT_WAIT;
1026 LASSERT(!list_empty(&rbp->rbp_bufs));
1027 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1028 list_del(&rb->rb_list);
1030 msg->msg_niov = rbp->rbp_npages;
1031 msg->msg_kiov = &rb->rb_kiov[0];
1033 /* unset the msg-rx_delayed flag since we're receiving the message */
1034 msg->msg_rx_delayed = 0;
1037 int cpt = msg->msg_rx_cpt;
1039 lnet_net_unlock(cpt);
1040 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1041 0, msg->msg_len, msg->msg_len);
1044 return LNET_CREDIT_OK;
1048 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1050 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1051 struct lnet_ni *txni = msg->msg_txni;
1052 struct lnet_msg *msg2;
1054 if (msg->msg_txcredit) {
1055 struct lnet_ni *ni = msg->msg_txni;
1056 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1058 /* give back NI txcredits */
1059 msg->msg_txcredit = 0;
1061 LASSERT((tq->tq_credits < 0) ==
1062 !list_empty(&tq->tq_delayed));
1065 atomic_inc(&ni->ni_tx_credits);
1066 if (tq->tq_credits <= 0) {
1067 msg2 = list_entry(tq->tq_delayed.next,
1068 struct lnet_msg, msg_list);
1069 list_del(&msg2->msg_list);
1071 LASSERT(msg2->msg_txni == ni);
1072 LASSERT(msg2->msg_tx_delayed);
1073 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1075 (void) lnet_post_send_locked(msg2, 1);
1079 if (msg->msg_peertxcredit) {
1080 /* give back peer txcredits */
1081 msg->msg_peertxcredit = 0;
1083 spin_lock(&txpeer->lpni_lock);
1084 LASSERT((txpeer->lpni_txcredits < 0) ==
1085 !list_empty(&txpeer->lpni_txq));
1087 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1088 LASSERT(txpeer->lpni_txqnob >= 0);
1090 txpeer->lpni_txcredits++;
1091 if (txpeer->lpni_txcredits <= 0) {
1094 msg2 = list_entry(txpeer->lpni_txq.next,
1095 struct lnet_msg, msg_list);
1096 list_del(&msg2->msg_list);
1097 spin_unlock(&txpeer->lpni_lock);
1099 LASSERT(msg2->msg_txpeer == txpeer);
1100 LASSERT(msg2->msg_tx_delayed);
1102 msg2_cpt = msg2->msg_tx_cpt;
1105 * The msg_cpt can be different from the msg2_cpt
1106 * so we need to make sure we lock the correct cpt
1108 * Once we call lnet_post_send_locked() it is no
1109 * longer safe to access msg2, since it could've
1110 * been freed by lnet_finalize(), but we still
1111 * need to relock the correct cpt, so we cache the
1112 * msg2_cpt for the purpose of the check that
1113 * follows the call to lnet_pose_send_locked().
1115 if (msg2_cpt != msg->msg_tx_cpt) {
1116 lnet_net_unlock(msg->msg_tx_cpt);
1117 lnet_net_lock(msg2_cpt);
1119 (void) lnet_post_send_locked(msg2, 1);
1120 if (msg2_cpt != msg->msg_tx_cpt) {
1121 lnet_net_unlock(msg2_cpt);
1122 lnet_net_lock(msg->msg_tx_cpt);
1125 spin_unlock(&txpeer->lpni_lock);
1130 msg->msg_txni = NULL;
1131 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1134 if (txpeer != NULL) {
1135 msg->msg_txpeer = NULL;
1136 lnet_peer_ni_decref_locked(txpeer);
1141 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1143 struct lnet_msg *msg;
1145 if (list_empty(&rbp->rbp_msgs))
1147 msg = list_entry(rbp->rbp_msgs.next,
1148 struct lnet_msg, msg_list);
1149 list_del(&msg->msg_list);
1151 (void)lnet_post_routed_recv_locked(msg, 1);
1155 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1157 struct lnet_msg *msg;
1158 struct lnet_msg *tmp;
1160 lnet_net_unlock(cpt);
1162 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1163 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1164 0, 0, 0, msg->msg_hdr.payload_length);
1165 list_del_init(&msg->msg_list);
1166 msg->msg_no_resend = true;
1167 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1168 lnet_finalize(msg, -ECANCELED);
1175 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1177 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1178 struct lnet_peer *lp;
1179 struct lnet_ni *rxni = msg->msg_rxni;
1180 struct lnet_msg *msg2;
1182 if (msg->msg_rtrcredit) {
1183 /* give back global router credits */
1184 struct lnet_rtrbuf *rb;
1185 struct lnet_rtrbufpool *rbp;
1187 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1188 * there until it gets one allocated, or aborts the wait
1190 LASSERT(msg->msg_kiov != NULL);
1192 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1195 msg->msg_kiov = NULL;
1196 msg->msg_rtrcredit = 0;
1198 LASSERT(rbp == lnet_msg2bufpool(msg));
1200 LASSERT((rbp->rbp_credits > 0) ==
1201 !list_empty(&rbp->rbp_bufs));
1203 /* If routing is now turned off, we just drop this buffer and
1204 * don't bother trying to return credits. */
1205 if (!the_lnet.ln_routing) {
1206 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1210 /* It is possible that a user has lowered the desired number of
1211 * buffers in this pool. Make sure we never put back
1212 * more buffers than the stated number. */
1213 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1214 /* Discard this buffer so we don't have too
1216 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1217 rbp->rbp_nbuffers--;
1219 list_add(&rb->rb_list, &rbp->rbp_bufs);
1221 if (rbp->rbp_credits <= 0)
1222 lnet_schedule_blocked_locked(rbp);
1227 if (msg->msg_peerrtrcredit) {
1229 LASSERT(rxpeerni->lpni_peer_net);
1230 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1232 /* give back peer router credits */
1233 msg->msg_peerrtrcredit = 0;
1235 spin_lock(&rxpeerni->lpni_lock);
1236 rxpeerni->lpni_rtrcredits++;
1237 spin_unlock(&rxpeerni->lpni_lock);
1239 lp = rxpeerni->lpni_peer_net->lpn_peer;
1240 spin_lock(&lp->lp_lock);
1242 /* drop all messages which are queued to be routed on that
1244 if (!the_lnet.ln_routing) {
1246 list_splice_init(&lp->lp_rtrq, &drop);
1247 spin_unlock(&lp->lp_lock);
1248 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1249 } else if (!list_empty(&lp->lp_rtrq)) {
1252 msg2 = list_entry(lp->lp_rtrq.next,
1253 struct lnet_msg, msg_list);
1254 list_del(&msg2->msg_list);
1255 msg2_cpt = msg2->msg_rx_cpt;
1256 spin_unlock(&lp->lp_lock);
1258 * messages on the lp_rtrq can be from any NID in
1259 * the peer, which means they might have different
1260 * cpts. We need to make sure we lock the right
1263 if (msg2_cpt != msg->msg_rx_cpt) {
1264 lnet_net_unlock(msg->msg_rx_cpt);
1265 lnet_net_lock(msg2_cpt);
1267 (void) lnet_post_routed_recv_locked(msg2, 1);
1268 if (msg2_cpt != msg->msg_rx_cpt) {
1269 lnet_net_unlock(msg2_cpt);
1270 lnet_net_lock(msg->msg_rx_cpt);
1273 spin_unlock(&lp->lp_lock);
1277 msg->msg_rxni = NULL;
1278 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1280 if (rxpeerni != NULL) {
1281 msg->msg_rxpeer = NULL;
1282 lnet_peer_ni_decref_locked(rxpeerni);
1286 static struct lnet_peer_ni *
1287 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1288 struct lnet_peer *peer,
1289 struct lnet_peer_ni *best_lpni,
1290 struct lnet_peer_net *peer_net)
1293 * Look at the peer NIs for the destination peer that connect
1294 * to the chosen net. If a peer_ni is preferred when using the
1295 * best_ni to communicate, we use that one. If there is no
1296 * preferred peer_ni, or there are multiple preferred peer_ni,
1297 * the available transmit credits are used. If the transmit
1298 * credits are equal, we round-robin over the peer_ni.
1300 struct lnet_peer_ni *lpni = NULL;
1301 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1303 int best_lpni_healthv = (best_lpni) ?
1304 atomic_read(&best_lpni->lpni_healthv) : 0;
1305 bool best_lpni_is_preferred = false;
1306 bool lpni_is_preferred;
1308 __u32 lpni_sel_prio;
1309 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1311 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1313 * if the best_ni we've chosen aleady has this lpni
1314 * preferred, then let's use it
1317 lpni_is_preferred = lnet_peer_is_pref_nid_locked(lpni,
1319 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1320 libcfs_nid2str(best_ni->ni_nid),
1323 lpni_is_preferred = false;
1326 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1327 lpni_sel_prio = lpni->lpni_sel_priority;
1330 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1331 libcfs_nid2str(lpni->lpni_nid),
1332 libcfs_nid2str(best_lpni->lpni_nid),
1333 lpni_healthv, best_lpni_healthv,
1334 lpni_sel_prio, best_sel_prio,
1335 lpni->lpni_txcredits, best_lpni_credits,
1336 lpni->lpni_seq, best_lpni->lpni_seq);
1340 /* pick the healthiest peer ni */
1341 if (lpni_healthv < best_lpni_healthv)
1343 else if (lpni_healthv > best_lpni_healthv) {
1344 if (best_lpni_is_preferred)
1345 best_lpni_is_preferred = false;
1349 if (lpni_sel_prio > best_sel_prio)
1351 else if (lpni_sel_prio < best_sel_prio) {
1352 if (best_lpni_is_preferred)
1353 best_lpni_is_preferred = false;
1357 /* if this is a preferred peer use it */
1358 if (!best_lpni_is_preferred && lpni_is_preferred) {
1359 best_lpni_is_preferred = true;
1361 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1362 /* this is not the preferred peer so let's ignore
1368 if (lpni->lpni_txcredits < best_lpni_credits)
1369 /* We already have a peer that has more credits
1370 * available than this one. No need to consider
1371 * this peer further.
1374 else if (lpni->lpni_txcredits > best_lpni_credits)
1377 /* The best peer found so far and the current peer
1378 * have the same number of available credits let's
1379 * make sure to select between them using Round Robin
1381 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1384 best_lpni_is_preferred = lpni_is_preferred;
1385 best_lpni_healthv = lpni_healthv;
1386 best_sel_prio = lpni_sel_prio;
1388 best_lpni_credits = lpni->lpni_txcredits;
1391 /* if we still can't find a peer ni then we can't reach it */
1393 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1394 LNET_NIDNET(dst_nid);
1395 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1396 libcfs_net2str(net_id));
1400 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1401 libcfs_nid2str(best_lpni->lpni_nid));
1407 * Prerequisite: the best_ni should already be set in the sd
1408 * Find the best lpni.
1409 * If the net id is provided then restrict lpni selection on
1410 * that particular net.
1411 * Otherwise find any reachable lpni. When dealing with an MR
1412 * gateway and it has multiple lpnis which we can use
1413 * we want to select the best one from the list of reachable
1416 static inline struct lnet_peer_ni *
1417 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1418 struct lnet_peer *peer, __u32 net_id)
1420 struct lnet_peer_net *peer_net;
1422 /* find the best_lpni on any local network */
1423 if (net_id == LNET_NET_ANY) {
1424 struct lnet_peer_ni *best_lpni = NULL;
1425 struct lnet_peer_net *lpn;
1426 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1427 /* no net specified find any reachable peer ni */
1428 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1430 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1436 /* restrict on the specified net */
1437 peer_net = lnet_peer_get_net_locked(peer, net_id);
1439 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1445 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1447 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1450 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1453 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1456 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1462 /* Compare route priorities and hop counts */
1464 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1466 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1467 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1469 if (r1->lr_priority < r2->lr_priority)
1472 if (r1->lr_priority > r2->lr_priority)
1475 if (r1_hops < r2_hops)
1478 if (r1_hops > r2_hops)
1484 static struct lnet_route *
1485 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1486 struct lnet_peer_ni *remote_lpni,
1487 struct lnet_route **prev_route,
1488 struct lnet_peer_ni **gwni)
1490 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1491 struct lnet_route *best_route;
1492 struct lnet_route *last_route;
1493 struct lnet_route *route;
1495 bool best_rte_is_preferred = false;
1498 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1499 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1501 best_route = last_route = NULL;
1502 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1503 if (!lnet_is_route_alive(route))
1505 gw_pnid = route->lr_gateway->lp_primary_nid;
1507 /* no protection on below fields, but it's harmless */
1508 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1511 /* if the best route found is in the preferred list then
1512 * tag it as preferred and use it later on. But if we
1513 * didn't find any routes which are on the preferred list
1514 * then just use the best route possible.
1516 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1518 if (!best_route || (rc && !best_rte_is_preferred)) {
1519 /* Restrict the selection of the router NI on the
1520 * src_net provided. If the src_net is LNET_NID_ANY,
1521 * then select the best interface available.
1523 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1528 "Gateway %s does not have a peer NI on net %s\n",
1529 libcfs_nid2str(gw_pnid),
1530 libcfs_net2str(src_net));
1535 if (rc && !best_rte_is_preferred) {
1536 /* This is the first preferred route we found,
1537 * so it beats any route found previously
1543 best_rte_is_preferred = true;
1544 CDEBUG(D_NET, "preferred gw = %s\n",
1545 libcfs_nid2str(gw_pnid));
1547 } else if ((!rc) && best_rte_is_preferred)
1548 /* The best route we found so far is in the preferred
1549 * list, so it beats any non-preferred route
1554 best_route = last_route = route;
1559 rc = lnet_compare_routes(route, best_route);
1563 /* Restrict the selection of the router NI on the
1564 * src_net provided. If the src_net is LNET_NID_ANY,
1565 * then select the best interface available.
1567 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1572 "Gateway %s does not have a peer NI on net %s\n",
1573 libcfs_nid2str(gw_pnid),
1574 libcfs_net2str(src_net));
1584 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1588 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1595 *prev_route = last_route;
1601 static struct lnet_ni *
1602 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1603 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1606 struct lnet_ni *ni = NULL;
1607 unsigned int shortest_distance;
1610 __u32 best_sel_prio;
1613 * If there is no peer_ni that we can send to on this network,
1614 * then there is no point in looking for a new best_ni here.
1616 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1619 if (best_ni == NULL) {
1620 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1621 shortest_distance = UINT_MAX;
1622 best_credits = INT_MIN;
1625 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1626 best_ni->ni_dev_cpt);
1627 best_credits = atomic_read(&best_ni->ni_tx_credits);
1628 best_healthv = atomic_read(&best_ni->ni_healthv);
1629 best_sel_prio = best_ni->ni_sel_priority;
1632 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1633 unsigned int distance;
1639 ni_credits = atomic_read(&ni->ni_tx_credits);
1640 ni_healthv = atomic_read(&ni->ni_healthv);
1641 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1642 ni_sel_prio = ni->ni_sel_priority;
1645 * calculate the distance from the CPT on which
1646 * the message memory is allocated to the CPT of
1647 * the NI's physical device
1649 distance = cfs_cpt_distance(lnet_cpt_table(),
1654 * All distances smaller than the NUMA range
1655 * are treated equally.
1657 if (distance < lnet_numa_range)
1658 distance = lnet_numa_range;
1661 * Select on health, shorter distance, available
1662 * credits, then round-robin.
1668 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d, p:%u] with best_ni %s [c:%d, d:%d, s:%d, p:%u]\n",
1669 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1670 ni->ni_seq, ni_sel_prio,
1671 (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1672 : "not selected", best_credits, shortest_distance,
1673 (best_ni) ? best_ni->ni_seq : 0,
1678 if (ni_healthv < best_healthv)
1680 else if (ni_healthv > best_healthv)
1683 if (ni_sel_prio > best_sel_prio)
1685 else if (ni_sel_prio < best_sel_prio)
1688 if (distance > shortest_distance)
1690 else if (distance < shortest_distance)
1693 if (ni_credits < best_credits)
1695 else if (ni_credits > best_credits)
1698 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1702 best_sel_prio = ni_sel_prio;
1703 shortest_distance = distance;
1704 best_healthv = ni_healthv;
1706 best_credits = ni_credits;
1709 CDEBUG(D_NET, "selected best_ni %s\n",
1710 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1716 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1717 * because such traffic is required to perform discovery. We therefore
1718 * exclude all GET and PUT on that portal. We also exclude all ACK and
1719 * REPLY traffic, but that is because the portal is not tracked in the
1720 * message structure for these message types. We could restrict this
1721 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1724 lnet_msg_discovery(struct lnet_msg *msg)
1726 if (msg->msg_type == LNET_MSG_PUT) {
1727 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1729 } else if (msg->msg_type == LNET_MSG_GET) {
1730 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1736 #define SRC_SPEC 0x0001
1737 #define SRC_ANY 0x0002
1738 #define LOCAL_DST 0x0004
1739 #define REMOTE_DST 0x0008
1740 #define MR_DST 0x0010
1741 #define NMR_DST 0x0020
1742 #define SND_RESP 0x0040
1744 /* The following to defines are used for return codes */
1745 #define REPEAT_SEND 0x1000
1746 #define PASS_THROUGH 0x2000
1748 /* The different cases lnet_select pathway needs to handle */
1749 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1750 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1751 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1752 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1753 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1754 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1755 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1756 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1759 lnet_handle_lo_send(struct lnet_send_data *sd)
1761 struct lnet_msg *msg = sd->sd_msg;
1762 int cpt = sd->sd_cpt;
1764 /* No send credit hassles with LOLND */
1765 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1766 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1767 if (!msg->msg_routing)
1768 msg->msg_hdr.src_nid =
1769 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1770 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1771 lnet_msg_commit(msg, cpt);
1772 msg->msg_txni = the_lnet.ln_loni;
1774 return LNET_CREDIT_OK;
1778 lnet_handle_send(struct lnet_send_data *sd)
1780 struct lnet_ni *best_ni = sd->sd_best_ni;
1781 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1782 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1783 struct lnet_msg *msg = sd->sd_msg;
1785 __u32 send_case = sd->sd_send_case;
1787 __u32 routing = send_case & REMOTE_DST;
1788 struct lnet_rsp_tracker *rspt;
1790 /* Increment sequence number of the selected peer, peer net,
1791 * local ni and local net so that we pick the next ones
1794 best_lpni->lpni_seq++;
1795 best_lpni->lpni_peer_net->lpn_seq++;
1797 best_ni->ni_net->net_seq++;
1799 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1800 libcfs_nid2str(best_ni->ni_nid),
1801 best_ni->ni_seq, best_ni->ni_net->net_seq,
1802 atomic_read(&best_ni->ni_tx_credits),
1803 best_ni->ni_sel_priority,
1804 libcfs_nid2str(best_lpni->lpni_nid),
1805 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1806 best_lpni->lpni_txcredits,
1807 best_lpni->lpni_sel_priority);
1810 * grab a reference on the peer_ni so it sticks around even if
1811 * we need to drop and relock the lnet_net_lock below.
1813 lnet_peer_ni_addref_locked(best_lpni);
1816 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1817 * message. This ensures that we get a CPT that is correct for
1818 * the NI when the NI has been restricted to a subset of all CPTs.
1819 * If the selected CPT differs from the one currently locked, we
1820 * must unlock and relock the lnet_net_lock(), and then check whether
1821 * the configuration has changed. We don't have a hold on the best_ni
1822 * yet, and it may have vanished.
1824 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1825 if (sd->sd_cpt != cpt2) {
1826 __u32 seq = lnet_get_dlc_seq_locked();
1827 lnet_net_unlock(sd->sd_cpt);
1829 lnet_net_lock(sd->sd_cpt);
1830 if (seq != lnet_get_dlc_seq_locked()) {
1831 lnet_peer_ni_decref_locked(best_lpni);
1837 * store the best_lpni in the message right away to avoid having
1838 * to do the same operation under different conditions
1840 msg->msg_txpeer = best_lpni;
1841 msg->msg_txni = best_ni;
1844 * grab a reference for the best_ni since now it's in use in this
1845 * send. The reference will be dropped in lnet_finalize()
1847 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1850 * Always set the target.nid to the best peer picked. Either the
1851 * NID will be one of the peer NIDs selected, or the same NID as
1852 * what was originally set in the target or it will be the NID of
1853 * a router if this message should be routed
1855 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1858 * lnet_msg_commit assigns the correct cpt to the message, which
1859 * is used to decrement the correct refcount on the ni when it's
1860 * time to return the credits
1862 lnet_msg_commit(msg, sd->sd_cpt);
1865 * If we are routing the message then we keep the src_nid that was
1866 * set by the originator. If we are not routing then we are the
1867 * originator and set it here.
1869 if (!msg->msg_routing)
1870 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1873 msg->msg_target_is_router = 1;
1874 msg->msg_target.pid = LNET_PID_LUSTRE;
1876 * since we're routing we want to ensure that the
1877 * msg_hdr.dest_nid is set to the final destination. When
1878 * the router receives this message it knows how to route
1881 * final_dst_lpni is set at the beginning of the
1882 * lnet_select_pathway() function and is never changed.
1883 * It's safe to use it here.
1885 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1888 * if we're not routing set the dest_nid to the best peer
1889 * ni NID that we picked earlier in the algorithm.
1891 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1895 * if we have response tracker block update it with the next hop
1899 rspt = msg->msg_md->md_rspt_ptr;
1901 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1902 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1903 libcfs_nid2str(rspt->rspt_next_hop_nid));
1907 rc = lnet_post_send_locked(msg, 0);
1910 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1911 libcfs_nid2str(msg->msg_hdr.src_nid),
1912 libcfs_nid2str(msg->msg_txni->ni_nid),
1913 libcfs_nid2str(sd->sd_src_nid),
1914 libcfs_nid2str(msg->msg_hdr.dest_nid),
1915 libcfs_nid2str(sd->sd_dst_nid),
1916 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1917 libcfs_nid2str(sd->sd_rtr_nid),
1918 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1924 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1925 struct lnet_msg *msg)
1927 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1928 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1929 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1930 libcfs_nid2str(lni->ni_nid),
1931 libcfs_nid2str(lpni->lpni_nid));
1932 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1941 * use the source and destination NIDs as the pathway
1944 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1946 /* the destination lpni is set before we get here. */
1949 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1950 if (!sd->sd_best_ni) {
1951 CERROR("Can't send to %s: src %s is not a "
1952 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1953 libcfs_nid2str(sd->sd_src_nid));
1957 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
1959 return lnet_handle_send(sd);
1967 * Don't run the selection algorithm on the peer NIs. By specifying the
1968 * local NID, we're also saying that we should always use the destination NID
1969 * provided. This handles the case where we should be using the same
1970 * destination NID for the all the messages which belong to the same RPC
1974 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1976 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1977 if (!sd->sd_best_ni) {
1978 CERROR("Can't send to %s: src %s is not a "
1979 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1980 libcfs_nid2str(sd->sd_src_nid));
1984 if (sd->sd_best_lpni &&
1985 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1986 return lnet_handle_lo_send(sd);
1987 else if (sd->sd_best_lpni)
1988 return lnet_handle_send(sd);
1990 CERROR("can't send to %s. no NI on %s\n",
1991 libcfs_nid2str(sd->sd_dst_nid),
1992 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1994 return -EHOSTUNREACH;
1998 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1999 struct lnet_peer *peer,
2000 struct lnet_peer_net *peer_net,
2003 struct lnet_net *local_net;
2004 struct lnet_ni *best_ni;
2006 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2011 * Iterate through the NIs in this local Net and select
2012 * the NI to send from. The selection is determined by
2013 * these 3 criterion in the following priority:
2015 * 2. NI available credits
2018 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2019 peer, peer_net, cpt);
2025 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2028 struct lnet_peer *peer;
2031 lnet_peer_ni_addref_locked(lpni);
2033 peer = lpni->lpni_peer_net->lpn_peer;
2035 if (lnet_peer_gw_discovery(peer)) {
2036 lnet_peer_ni_decref_locked(lpni);
2040 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2041 lnet_peer_ni_decref_locked(lpni);
2045 rc = lnet_discover_peer_locked(lpni, cpt, false);
2047 lnet_peer_ni_decref_locked(lpni);
2050 /* The peer may have changed. */
2051 peer = lpni->lpni_peer_net->lpn_peer;
2052 spin_lock(&peer->lp_lock);
2053 if (lnet_peer_is_uptodate_locked(peer)) {
2054 spin_unlock(&peer->lp_lock);
2055 lnet_peer_ni_decref_locked(lpni);
2058 /* queue message and return */
2059 msg->msg_sending = 0;
2060 msg->msg_txpeer = NULL;
2061 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2062 spin_unlock(&peer->lp_lock);
2064 lnet_peer_ni_decref_locked(lpni);
2066 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2067 msg, libcfs_nid2str(peer->lp_primary_nid));
2069 return LNET_DC_WAIT;
2073 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2075 struct lnet_peer_ni **gw_lpni,
2076 struct lnet_peer **gw_peer)
2080 struct lnet_peer *gw;
2081 struct lnet_peer *lp;
2082 struct lnet_peer_net *lpn;
2083 struct lnet_peer_net *best_lpn = NULL;
2084 struct lnet_remotenet *rnet, *best_rnet = NULL;
2085 struct lnet_route *best_route = NULL;
2086 struct lnet_route *last_route = NULL;
2087 struct lnet_peer_ni *lpni = NULL;
2088 struct lnet_peer_ni *gwni = NULL;
2089 bool route_found = false;
2090 lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
2091 (sd->sd_best_ni != NULL) ? sd->sd_best_ni->ni_nid :
2093 int best_lpn_healthv = 0;
2094 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2096 CDEBUG(D_NET, "using src nid %s for route restriction\n",
2097 libcfs_nid2str(src_nid));
2099 /* If a router nid was specified then we are replying to a GET or
2100 * sending an ACK. In this case we use the gateway associated with the
2101 * specified router nid.
2103 if (sd->sd_rtr_nid != LNET_NID_ANY) {
2104 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2106 gw = gwni->lpni_peer_net->lpn_peer;
2107 lnet_peer_ni_decref_locked(gwni);
2108 if (gw->lp_rtr_refcount) {
2109 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2113 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2114 libcfs_nid2str(sd->sd_rtr_nid));
2119 if (sd->sd_msg->msg_routing) {
2120 /* If I'm routing this message then I need to find the
2121 * next hop based on the destination NID
2123 best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
2125 CERROR("Unable to route message to %s - Route table may be misconfigured\n",
2126 libcfs_nid2str(sd->sd_dst_nid));
2127 return -EHOSTUNREACH;
2130 /* we've already looked up the initial lpni using
2133 lpni = sd->sd_best_lpni;
2134 /* the peer tree must be in existence */
2135 LASSERT(lpni && lpni->lpni_peer_net &&
2136 lpni->lpni_peer_net->lpn_peer);
2137 lp = lpni->lpni_peer_net->lpn_peer;
2139 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2140 /* is this remote network reachable? */
2141 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2150 /* select the preferred peer net */
2151 if (best_lpn_healthv > lpn->lpn_healthv)
2153 else if (best_lpn_healthv < lpn->lpn_healthv)
2156 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2158 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2161 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2164 best_lpn_healthv = lpn->lpn_healthv;
2165 best_lpn_sel_prio = lpn->lpn_sel_priority;
2171 CERROR("peer %s has no available nets\n",
2172 libcfs_nid2str(sd->sd_dst_nid));
2173 return -EHOSTUNREACH;
2176 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2179 best_lpn->lpn_net_id);
2180 if (!sd->sd_best_lpni) {
2181 CERROR("peer %s is unreachable\n",
2182 libcfs_nid2str(sd->sd_dst_nid));
2183 return -EHOSTUNREACH;
2186 /* We're attempting to round robin over the remote peer
2187 * NI's so update the final destination we selected
2189 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2191 /* Increment the sequence number of the remote lpni so
2192 * we can round robin over the different interfaces of
2195 sd->sd_best_lpni->lpni_seq++;
2199 * find the best route. Restrict the selection on the net of the
2200 * local NI if we've already picked the local NI to send from.
2201 * Otherwise, let's pick any route we can find and then find
2202 * a local NI we can reach the route's gateway on. Any route we
2203 * select will be reachable by virtue of the restriction we have
2204 * when adding a route.
2206 best_route = lnet_find_route_locked(best_rnet,
2207 LNET_NIDNET(src_nid),
2209 &last_route, &gwni);
2212 CERROR("no route to %s from %s\n",
2213 libcfs_nid2str(dst_nid),
2214 libcfs_nid2str(src_nid));
2215 return -EHOSTUNREACH;
2219 CERROR("Internal Error. Route expected to %s from %s\n",
2220 libcfs_nid2str(dst_nid),
2221 libcfs_nid2str(src_nid));
2225 gw = best_route->lr_gateway;
2226 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2227 local_lnet = best_route->lr_lnet;
2231 * Discover this gateway if it hasn't already been discovered.
2232 * This means we might delay the message until discovery has
2235 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2236 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2240 if (!sd->sd_best_ni)
2241 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2242 lnet_peer_get_net_locked(gw,
2246 if (!sd->sd_best_ni) {
2247 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2248 libcfs_net2str(local_lnet),
2249 libcfs_nid2str(sd->sd_src_nid));
2257 * increment the sequence numbers since now we're sure we're
2258 * going to use this path
2260 if (sd->sd_rtr_nid == LNET_NID_ANY) {
2261 LASSERT(best_route && last_route);
2262 best_route->lr_seq = last_route->lr_seq + 1;
2264 best_lpn->lpn_seq++;
2275 * Remote destination
2276 * Non-MR destination
2280 * Remote destination
2283 * The handling of these two cases is similar. Even though the destination
2284 * can be MR or non-MR, we'll deal directly with the router.
2287 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2290 struct lnet_peer_ni *gw_lpni = NULL;
2291 struct lnet_peer *gw_peer = NULL;
2294 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2295 if (!sd->sd_best_ni) {
2296 CERROR("Can't send to %s: src %s is not a "
2297 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2298 libcfs_nid2str(sd->sd_src_nid));
2302 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2307 if (sd->sd_send_case & NMR_DST)
2309 * since the final destination is non-MR let's set its preferred
2310 * NID before we send
2312 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2316 * We're going to send to the gw found so let's set its
2319 sd->sd_peer = gw_peer;
2320 sd->sd_best_lpni = gw_lpni;
2322 return lnet_handle_send(sd);
2326 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2329 struct lnet_peer_net *lpn = NULL;
2330 struct lnet_peer_net *best_lpn = NULL;
2331 struct lnet_net *net = NULL;
2332 struct lnet_net *best_net = NULL;
2333 struct lnet_ni *best_ni = NULL;
2334 int best_lpn_healthv = 0;
2335 int best_net_healthv = 0;
2337 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2339 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2344 * The peer can have multiple interfaces, some of them can be on
2345 * the local network and others on a routed network. We should
2346 * prefer the local network. However if the local network is not
2347 * available then we need to try the routed network
2350 /* go through all the peer nets and find the best_ni */
2351 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2353 * The peer's list of nets can contain non-local nets. We
2354 * want to only examine the local ones.
2356 net = lnet_get_net_locked(lpn->lpn_net_id);
2360 lpn_sel_prio = lpn->lpn_sel_priority;
2361 net_healthv = lnet_get_net_healthv_locked(net);
2362 net_sel_prio = net->net_sel_priority;
2365 * if this is a discovery message and lp_disc_net_id is
2366 * specified then use that net to send the discovery on.
2368 if (peer->lp_disc_net_id == lpn->lpn_net_id &&
2377 /* always select the lpn with the best health */
2378 if (best_lpn_healthv > lpn->lpn_healthv)
2380 else if (best_lpn_healthv < lpn->lpn_healthv)
2383 /* select the preferred peer and local nets */
2384 if (best_lpn_sel_prio < lpn_sel_prio)
2386 else if (best_lpn_sel_prio > lpn_sel_prio)
2389 if (best_net_healthv > net_healthv)
2391 else if (best_net_healthv < net_healthv)
2394 if (best_net_sel_prio < net_sel_prio)
2396 else if (best_net_sel_prio > net_sel_prio)
2399 if (best_lpn->lpn_seq < lpn->lpn_seq)
2401 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2404 /* round robin over the local networks */
2405 if (best_net->net_seq <= net->net_seq)
2409 best_net_healthv = net_healthv;
2410 best_net_sel_prio = net_sel_prio;
2411 best_lpn_healthv = lpn->lpn_healthv;
2412 best_lpn_sel_prio = lpn_sel_prio;
2421 /* Select the best NI on the same net as best_lpn chosen
2424 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer,
2431 static struct lnet_ni *
2432 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2434 struct lnet_ni *best_ni = NULL;
2435 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2436 struct lnet_peer_ni *lpni_entry;
2439 * We must use a consistent source address when sending to a
2440 * non-MR peer. However, a non-MR peer can have multiple NIDs
2441 * on multiple networks, and we may even need to talk to this
2442 * peer on multiple networks -- certain types of
2443 * load-balancing configuration do this.
2445 * So we need to pick the NI the peer prefers for this
2446 * particular network.
2449 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2451 if (lpni_entry->lpni_pref_nnids == 0)
2453 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2454 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2461 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2463 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2465 struct lnet_ni *best_ni = NULL;
2466 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2469 * We must use a consistent source address when sending to a
2470 * non-MR peer. However, a non-MR peer can have multiple NIDs
2471 * on multiple networks, and we may even need to talk to this
2472 * peer on multiple networks -- certain types of
2473 * load-balancing configuration do this.
2475 * So we need to pick the NI the peer prefers for this
2476 * particular network.
2479 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2482 /* if best_ni is still not set just pick one */
2485 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2486 sd->sd_best_lpni->lpni_peer_net,
2488 /* If there is no best_ni we don't have a route */
2490 CERROR("no path to %s from net %s\n",
2491 libcfs_nid2str(best_lpni->lpni_nid),
2492 libcfs_net2str(best_lpni->lpni_net->net_id));
2493 return -EHOSTUNREACH;
2497 sd->sd_best_ni = best_ni;
2499 /* Set preferred NI if necessary. */
2500 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2507 * Source not specified
2511 * always use the same source NID for NMR peers
2512 * If we've talked to that peer before then we already have a preferred
2513 * source NI associated with it. Otherwise, we select a preferred local NI
2514 * and store it in the peer
2517 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2521 /* sd->sd_best_lpni is already set to the final destination */
2524 * At this point we should've created the peer ni and peer. If we
2525 * can't find it, then something went wrong. Instead of assert
2526 * output a relevant message and fail the send
2528 if (!sd->sd_best_lpni) {
2529 CERROR("Internal fault. Unable to send msg %s to %s. "
2531 lnet_msgtyp2str(sd->sd_msg->msg_type),
2532 libcfs_nid2str(sd->sd_dst_nid));
2536 if (sd->sd_msg->msg_routing) {
2537 /* If I'm forwarding this message then I can choose any NI
2538 * on the destination peer net
2540 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2542 sd->sd_best_lpni->lpni_peer_net,
2544 if (!sd->sd_best_ni) {
2545 CERROR("Unable to forward message to %s. No local NI available\n",
2546 libcfs_nid2str(sd->sd_dst_nid));
2550 rc = lnet_select_preferred_best_ni(sd);
2553 rc = lnet_handle_send(sd);
2559 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2562 * NOTE we've already handled the remote peer case. So we only
2563 * need to worry about the local case here.
2565 * if we're sending a response, ACK or reply, we need to send it
2566 * to the destination NID given to us. At this point we already
2567 * have the peer_ni we're suppose to send to, so just find the
2568 * best_ni on the peer net and use that. Since we're sending to an
2569 * MR peer then we can just run the selection algorithm on our
2570 * local NIs and pick the best one.
2572 if (sd->sd_send_case & SND_RESP) {
2574 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2575 sd->sd_best_lpni->lpni_peer_net,
2578 if (!sd->sd_best_ni) {
2580 * We're not going to deal with not able to send
2581 * a response to the provided final destination
2583 CERROR("Can't send response to %s. "
2584 "No local NI available\n",
2585 libcfs_nid2str(sd->sd_dst_nid));
2586 return -EHOSTUNREACH;
2589 return lnet_handle_send(sd);
2593 * If we get here that means we're sending a fresh request, PUT or
2594 * GET, so we need to run our standard selection algorithm.
2595 * First find the best local interface that's on any of the peer's
2598 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2600 lnet_msg_discovery(sd->sd_msg));
2601 if (sd->sd_best_ni) {
2603 lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2605 sd->sd_best_ni->ni_net->net_id);
2608 * if we're successful in selecting a peer_ni on the local
2609 * network, then send to it. Otherwise fall through and
2610 * try and see if we can reach it over another routed
2613 if (sd->sd_best_lpni &&
2614 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2616 * in case we initially started with a routed
2617 * destination, let's reset to local
2619 sd->sd_send_case &= ~REMOTE_DST;
2620 sd->sd_send_case |= LOCAL_DST;
2621 return lnet_handle_lo_send(sd);
2622 } else if (sd->sd_best_lpni) {
2624 * in case we initially started with a routed
2625 * destination, let's reset to local
2627 sd->sd_send_case &= ~REMOTE_DST;
2628 sd->sd_send_case |= LOCAL_DST;
2629 return lnet_handle_send(sd);
2632 CERROR("Internal Error. Expected to have a best_lpni: "
2634 libcfs_nid2str(sd->sd_src_nid),
2635 libcfs_nid2str(sd->sd_dst_nid));
2641 * Peer doesn't have a local network. Let's see if there is
2642 * a remote network we can reach it on.
2644 return PASS_THROUGH;
2649 * Source NID not specified
2654 * Source NID not speified
2655 * Remote destination
2658 * In both of these cases if we're sending a response, ACK or REPLY, then
2659 * we need to send to the destination NID provided.
2661 * In the remote case let's deal with MR routers.
2666 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2669 struct lnet_peer *gw_peer = NULL;
2670 struct lnet_peer_ni *gw_lpni = NULL;
2673 * handle sending a response to a remote peer here so we don't
2674 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2676 if (sd->sd_send_case & REMOTE_DST &&
2677 sd->sd_send_case & SND_RESP) {
2678 struct lnet_peer_ni *gw;
2679 struct lnet_peer *gw_peer;
2681 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2684 CERROR("Can't send response to %s. "
2685 "No route available\n",
2686 libcfs_nid2str(sd->sd_dst_nid));
2687 return -EHOSTUNREACH;
2688 } else if (rc > 0) {
2692 sd->sd_best_lpni = gw;
2693 sd->sd_peer = gw_peer;
2695 return lnet_handle_send(sd);
2699 * Even though the NID for the peer might not be on a local network,
2700 * since the peer is MR there could be other interfaces on the
2701 * local network. In that case we'd still like to prefer the local
2702 * network over the routed network. If we're unable to do that
2703 * then we select the best router among the different routed networks,
2704 * and if the router is MR then we can deal with it as such.
2706 rc = lnet_handle_any_mr_dsta(sd);
2707 if (rc != PASS_THROUGH)
2711 * Now that we must route to the destination, we must consider the
2712 * MR case, where the destination has multiple interfaces, some of
2713 * which we can route to and others we do not. For this reason we
2714 * need to select the destination which we can route to and if
2715 * there are multiple, we need to round robin.
2717 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2722 sd->sd_send_case &= ~LOCAL_DST;
2723 sd->sd_send_case |= REMOTE_DST;
2725 sd->sd_peer = gw_peer;
2726 sd->sd_best_lpni = gw_lpni;
2728 return lnet_handle_send(sd);
2732 * Source not specified
2733 * Remote destination
2736 * Must send to the specified peer NID using the same source NID that
2737 * we've used before. If it's the first time to talk to that peer then
2738 * find the source NI and assign it as preferred to that peer
2741 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2744 struct lnet_peer_ni *gw_lpni = NULL;
2745 struct lnet_peer *gw_peer = NULL;
2748 * Let's see if we have a preferred NI to talk to this NMR peer
2750 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2754 * find the router and that'll find the best NI if we didn't find
2757 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2763 * set the best_ni we've chosen as the preferred one for
2766 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2768 /* we'll be sending to the gw */
2769 sd->sd_best_lpni = gw_lpni;
2770 sd->sd_peer = gw_peer;
2772 return lnet_handle_send(sd);
2776 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2779 * turn off the SND_RESP bit.
2780 * It will be checked in the case handling
2782 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2784 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2785 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2786 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2787 (send_case & MR_DST) ? "MR: " : "NMR: ",
2788 libcfs_nid2str(sd->sd_dst_nid),
2789 (send_case & LOCAL_DST) ? "local" : "routed");
2791 switch (send_case) {
2793 * For all cases where the source is specified, we should always
2794 * use the destination NID, whether it's an MR destination or not,
2795 * since we're continuing a series of related messages for the
2798 case SRC_SPEC_LOCAL_NMR_DST:
2799 return lnet_handle_spec_local_nmr_dst(sd);
2800 case SRC_SPEC_LOCAL_MR_DST:
2801 return lnet_handle_spec_local_mr_dst(sd);
2802 case SRC_SPEC_ROUTER_NMR_DST:
2803 case SRC_SPEC_ROUTER_MR_DST:
2804 return lnet_handle_spec_router_dst(sd);
2805 case SRC_ANY_LOCAL_NMR_DST:
2806 return lnet_handle_any_local_nmr_dst(sd);
2807 case SRC_ANY_LOCAL_MR_DST:
2808 case SRC_ANY_ROUTER_MR_DST:
2809 return lnet_handle_any_mr_dst(sd);
2810 case SRC_ANY_ROUTER_NMR_DST:
2811 return lnet_handle_any_router_nmr_dst(sd);
2813 CERROR("Unknown send case\n");
2819 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2820 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2822 struct lnet_peer_ni *lpni;
2823 struct lnet_peer *peer;
2824 struct lnet_send_data send_data;
2827 __u32 send_case = 0;
2829 bool mr_forwarding_allowed;
2831 memset(&send_data, 0, sizeof(send_data));
2834 * get an initial CPT to use for locking. The idea here is not to
2835 * serialize the calls to select_pathway, so that as many
2836 * operations can run concurrently as possible. To do that we use
2837 * the CPT where this call is being executed. Later on when we
2838 * determine the CPT to use in lnet_message_commit, we switch the
2839 * lock and check if there was any configuration change. If none,
2840 * then we proceed, if there is, then we restart the operation.
2842 cpt = lnet_net_lock_current();
2844 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2845 if (md_cpt == CFS_CPT_ANY)
2851 * If we're being asked to send to the loopback interface, there
2852 * is no need to go through any selection. We can just shortcut
2853 * the entire process and send over lolnd
2855 send_data.sd_msg = msg;
2856 send_data.sd_cpt = cpt;
2857 if (dst_nid == LNET_NID_LO_0) {
2858 rc = lnet_handle_lo_send(&send_data);
2859 lnet_net_unlock(cpt);
2864 * find an existing peer_ni, or create one and mark it as having been
2865 * created due to network traffic. This call will create the
2866 * peer->peer_net->peer_ni tree.
2868 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2870 lnet_net_unlock(cpt);
2871 return PTR_ERR(lpni);
2875 * Cache the original src_nid and rtr_nid. If we need to resend the
2876 * message then we'll need to know whether the src_nid was originally
2877 * specified for this message. If it was originally specified,
2878 * then we need to keep using the same src_nid since it's
2879 * continuing the same sequence of messages. Similarly, rtr_nid will
2880 * affect our choice of next hop.
2882 msg->msg_src_nid_param = src_nid;
2883 msg->msg_rtr_nid_param = rtr_nid;
2886 * If necessary, perform discovery on the peer that owns this peer_ni.
2887 * Note, this can result in the ownership of this peer_ni changing
2888 * to another peer object.
2890 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2892 lnet_peer_ni_decref_locked(lpni);
2893 lnet_net_unlock(cpt);
2896 lnet_peer_ni_decref_locked(lpni);
2898 peer = lpni->lpni_peer_net->lpn_peer;
2901 * Identify the different send cases
2903 if (src_nid == LNET_NID_ANY)
2904 send_case |= SRC_ANY;
2906 send_case |= SRC_SPEC;
2908 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2909 send_case |= LOCAL_DST;
2911 send_case |= REMOTE_DST;
2914 if (msg->msg_routing && (send_case & LOCAL_DST))
2917 /* Determine whether to allow MR forwarding for this message.
2918 * NB: MR forwarding is allowed if the message originator and the
2919 * destination are both MR capable, and the destination lpni that was
2920 * originally chosen by the originator is unhealthy or down.
2921 * We check the MR capability of the destination further below
2923 mr_forwarding_allowed = false;
2925 struct lnet_peer *src_lp;
2926 struct lnet_peer_ni *src_lpni;
2928 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
2930 /* We don't fail the send if we hit any errors here. We'll just
2931 * try to send it via non-multi-rail criteria
2933 if (!IS_ERR(src_lpni)) {
2934 /* Drop ref taken by lnet_nid2peerni_locked() */
2935 lnet_peer_ni_decref_locked(src_lpni);
2936 src_lp = lpni->lpni_peer_net->lpn_peer;
2937 if (lnet_peer_is_multi_rail(src_lp) &&
2938 !lnet_is_peer_ni_alive(lpni))
2939 mr_forwarding_allowed = true;
2942 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
2943 mr_forwarding_allowed ? "allowed" : "not allowed");
2947 * Deal with the peer as NMR in the following cases:
2948 * 1. the peer is NMR
2949 * 2. We're trying to recover a specific peer NI
2950 * 3. I'm a router sending to the final destination and MR forwarding is
2951 * not allowed for this message (as determined above).
2952 * In this case the source of the message would've
2953 * already selected the final destination so my job
2954 * is to honor the selection.
2956 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
2957 (final_hop && !mr_forwarding_allowed))
2958 send_case |= NMR_DST;
2960 send_case |= MR_DST;
2962 if (lnet_msg_is_response(msg))
2963 send_case |= SND_RESP;
2965 /* assign parameters to the send_data */
2966 send_data.sd_rtr_nid = rtr_nid;
2967 send_data.sd_src_nid = src_nid;
2968 send_data.sd_dst_nid = dst_nid;
2969 send_data.sd_best_lpni = lpni;
2971 * keep a pointer to the final destination in case we're going to
2972 * route, so we'll need to access it later
2974 send_data.sd_final_dst_lpni = lpni;
2975 send_data.sd_peer = peer;
2976 send_data.sd_md_cpt = md_cpt;
2977 send_data.sd_send_case = send_case;
2979 rc = lnet_handle_send_case_locked(&send_data);
2982 * Update the local cpt since send_data.sd_cpt might've been
2983 * updated as a result of calling lnet_handle_send_case_locked().
2985 cpt = send_data.sd_cpt;
2987 if (rc == REPEAT_SEND)
2990 lnet_net_unlock(cpt);
2996 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2998 lnet_nid_t dst_nid = msg->msg_target.nid;
3002 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
3003 * but we might want to use pre-determined router for ACK/REPLY
3006 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3007 LASSERT(msg->msg_txpeer == NULL);
3008 LASSERT(msg->msg_txni == NULL);
3009 LASSERT(!msg->msg_sending);
3010 LASSERT(!msg->msg_target_is_router);
3011 LASSERT(!msg->msg_receiving);
3013 msg->msg_sending = 1;
3015 LASSERT(!msg->msg_tx_committed);
3017 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3019 if (rc == -EHOSTUNREACH)
3020 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3022 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3026 if (rc == LNET_CREDIT_OK)
3027 lnet_ni_send(msg->msg_txni, msg);
3029 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3033 enum lnet_mt_event_type {
3034 MT_TYPE_LOCAL_NI = 0,
3038 struct lnet_mt_event_info {
3039 enum lnet_mt_event_type mt_type;
3043 /* called with res_lock held */
3045 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3047 struct lnet_rsp_tracker *rspt;
3050 * msg has a refcount on the MD so the MD is not going away.
3051 * The rspt queue for the cpt is protected by
3052 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3054 if (!md->md_rspt_ptr)
3057 rspt = md->md_rspt_ptr;
3060 LASSERT(rspt->rspt_cpt == cpt);
3062 md->md_rspt_ptr = NULL;
3064 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3066 * The monitor thread has invalidated this handle because the
3067 * response timed out, but it failed to lookup the MD. That
3068 * means this response tracker is on the zombie list. We can
3069 * safely remove it under the resource lock (held by caller) and
3070 * free the response tracker block.
3072 list_del(&rspt->rspt_on_list);
3073 lnet_rspt_free(rspt, cpt);
3076 * invalidate the handle to indicate that a response has been
3077 * received, which will then lead the monitor thread to clean up
3080 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3085 lnet_clean_zombie_rstqs(void)
3087 struct lnet_rsp_tracker *rspt, *tmp;
3090 cfs_cpt_for_each(i, lnet_cpt_table()) {
3091 list_for_each_entry_safe(rspt, tmp,
3092 the_lnet.ln_mt_zombie_rstqs[i],
3094 list_del(&rspt->rspt_on_list);
3095 lnet_rspt_free(rspt, i);
3099 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3103 lnet_finalize_expired_responses(void)
3105 struct lnet_libmd *md;
3106 struct lnet_rsp_tracker *rspt, *tmp;
3110 if (the_lnet.ln_mt_rstq == NULL)
3113 cfs_cpt_for_each(i, lnet_cpt_table()) {
3114 LIST_HEAD(local_queue);
3117 if (!the_lnet.ln_mt_rstq[i]) {
3121 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3126 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3128 * The rspt mdh will be invalidated when a response
3129 * is received or whenever we want to discard the
3130 * block the monitor thread will walk the queue
3131 * and clean up any rsts with an invalid mdh.
3132 * The monitor thread will walk the queue until
3133 * the first unexpired rspt block. This means that
3134 * some rspt blocks which received their
3135 * corresponding responses will linger in the
3136 * queue until they are cleaned up eventually.
3139 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3141 list_del(&rspt->rspt_on_list);
3142 lnet_rspt_free(rspt, i);
3146 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3147 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3148 struct lnet_peer_ni *lpni;
3151 md = lnet_handle2md(&rspt->rspt_mdh);
3153 /* MD has been queued for unlink, but
3154 * rspt hasn't been detached (Note we've
3155 * checked above that the rspt_mdh is
3156 * valid). Since we cannot lookup the MD
3157 * we're unable to detach the rspt
3158 * ourselves. Thus, move the rspt to the
3159 * zombie list where we'll wait for
3161 * 1. The remaining operations on the
3162 * MD to complete. In this case the
3163 * final operation will result in
3164 * lnet_msg_detach_md()->
3165 * lnet_detach_rsp_tracker() where
3166 * we will clean up this response
3168 * 2. LNet to shutdown. In this case
3169 * we'll wait until after all LND Nets
3170 * have shutdown and then we can
3171 * safely free any remaining response
3172 * tracker blocks on the zombie list.
3173 * Note: We need to hold the resource
3174 * lock when adding to the zombie list
3175 * because we may have concurrent access
3176 * with lnet_detach_rsp_tracker().
3178 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3179 list_move(&rspt->rspt_on_list,
3180 the_lnet.ln_mt_zombie_rstqs[i]);
3184 LASSERT(md->md_rspt_ptr == rspt);
3185 md->md_rspt_ptr = NULL;
3188 LNetMDUnlink(rspt->rspt_mdh);
3190 nid = rspt->rspt_next_hop_nid;
3192 list_del(&rspt->rspt_on_list);
3193 lnet_rspt_free(rspt, i);
3195 /* If we're shutting down we just want to clean
3196 * up the rspt blocks
3198 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3202 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;