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/
31 * lnet/lnet/lib-move.c
33 * Data movement routines
36 #define DEBUG_SUBSYSTEM S_LNET
38 #include <linux/pagemap.h>
40 #include <lnet/lib-lnet.h>
41 #include <linux/nsproxy.h>
42 #include <net/net_namespace.h>
44 static int local_nid_dist_zero = 1;
45 module_param(local_nid_dist_zero, int, 0444);
46 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
48 struct lnet_send_data {
49 struct lnet_ni *sd_best_ni;
50 struct lnet_peer_ni *sd_best_lpni;
51 struct lnet_peer_ni *sd_final_dst_lpni;
52 struct lnet_peer *sd_peer;
53 struct lnet_peer *sd_gw_peer;
54 struct lnet_peer_ni *sd_gw_lpni;
55 struct lnet_peer_net *sd_peer_net;
56 struct lnet_msg *sd_msg;
57 lnet_nid_t sd_dst_nid;
58 lnet_nid_t sd_src_nid;
59 lnet_nid_t sd_rtr_nid;
66 lnet_msg_is_response(struct lnet_msg *msg)
68 return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
72 lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
74 if (md_options & LNET_MD_NO_TRACK_RESPONSE)
75 /* Explicitly disabled in MD options */
78 if (md_options & LNET_MD_TRACK_RESPONSE)
79 /* Explicity enabled in MD options */
82 if (lnet_response_tracking == 3)
83 /* Enabled for all message types */
86 if (msg_type == LNET_MSG_PUT)
87 return lnet_response_tracking == 2;
89 if (msg_type == LNET_MSG_GET)
90 return lnet_response_tracking == 1;
95 static inline struct lnet_comm_count *
96 get_stats_counts(struct lnet_element_stats *stats,
97 enum lnet_stats_type stats_type)
100 case LNET_STATS_TYPE_SEND:
101 return &stats->el_send_stats;
102 case LNET_STATS_TYPE_RECV:
103 return &stats->el_recv_stats;
104 case LNET_STATS_TYPE_DROP:
105 return &stats->el_drop_stats;
107 CERROR("Unknown stats type\n");
113 void lnet_incr_stats(struct lnet_element_stats *stats,
114 enum lnet_msg_type msg_type,
115 enum lnet_stats_type stats_type)
117 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
123 atomic_inc(&counts->co_ack_count);
126 atomic_inc(&counts->co_put_count);
129 atomic_inc(&counts->co_get_count);
132 atomic_inc(&counts->co_reply_count);
135 atomic_inc(&counts->co_hello_count);
138 CERROR("There is a BUG in the code. Unknown message type\n");
143 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
144 enum lnet_stats_type stats_type)
146 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
150 return (atomic_read(&counts->co_ack_count) +
151 atomic_read(&counts->co_put_count) +
152 atomic_read(&counts->co_get_count) +
153 atomic_read(&counts->co_reply_count) +
154 atomic_read(&counts->co_hello_count));
157 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
158 struct lnet_comm_count *counts)
160 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
161 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
162 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
163 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
164 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
167 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
168 struct lnet_element_stats *stats)
170 struct lnet_comm_count *counts;
175 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
178 assign_stats(&msg_stats->im_send_stats, counts);
180 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
183 assign_stats(&msg_stats->im_recv_stats, counts);
185 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
188 assign_stats(&msg_stats->im_drop_stats, counts);
192 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
194 struct lnet_test_peer *tp;
195 struct list_head *el;
196 struct list_head *next;
199 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
200 if (threshold != 0) {
201 /* Adding a new entry */
202 LIBCFS_ALLOC(tp, sizeof(*tp));
207 tp->tp_threshold = threshold;
210 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
217 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
218 tp = list_entry(el, struct lnet_test_peer, tp_list);
220 if (tp->tp_threshold == 0 || /* needs culling anyway */
221 nid == LNET_NID_ANY || /* removing all entries */
222 tp->tp_nid == nid) { /* matched this one */
223 list_move(&tp->tp_list, &cull);
229 while (!list_empty(&cull)) {
230 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
232 list_del(&tp->tp_list);
233 LIBCFS_FREE(tp, sizeof(*tp));
239 fail_peer (lnet_nid_t nid, int outgoing)
241 struct lnet_test_peer *tp;
242 struct list_head *el;
243 struct list_head *next;
247 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
250 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
251 tp = list_entry(el, struct lnet_test_peer, tp_list);
253 if (tp->tp_threshold == 0) {
256 /* only cull zombies on outgoing tests,
257 * since we may be at interrupt priority on
258 * incoming messages. */
259 list_move(&tp->tp_list, &cull);
264 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
265 nid == tp->tp_nid) { /* fail this peer */
268 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
271 tp->tp_threshold == 0) {
273 list_move(&tp->tp_list, &cull);
282 while (!list_empty(&cull)) {
283 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
284 list_del(&tp->tp_list);
286 LIBCFS_FREE(tp, sizeof(*tp));
293 lnet_iov_nob(unsigned int niov, struct kvec *iov)
295 unsigned int nob = 0;
297 LASSERT(niov == 0 || iov != NULL);
299 nob += (iov++)->iov_len;
303 EXPORT_SYMBOL(lnet_iov_nob);
306 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
307 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
310 /* NB diov, siov are READ-ONLY */
311 unsigned int this_nob;
316 /* skip complete frags before 'doffset' */
318 while (doffset >= diov->iov_len) {
319 doffset -= diov->iov_len;
325 /* skip complete frags before 'soffset' */
327 while (soffset >= siov->iov_len) {
328 soffset -= siov->iov_len;
337 this_nob = min3((unsigned int)diov->iov_len - doffset,
338 (unsigned int)siov->iov_len - soffset,
341 memcpy((char *)diov->iov_base + doffset,
342 (char *)siov->iov_base + soffset, this_nob);
345 if (diov->iov_len > doffset + this_nob) {
353 if (siov->iov_len > soffset + this_nob) {
362 EXPORT_SYMBOL(lnet_copy_iov2iov);
365 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
367 unsigned int nob = 0;
369 LASSERT(niov == 0 || kiov != NULL);
371 nob += (kiov++)->bv_len;
375 EXPORT_SYMBOL(lnet_kiov_nob);
378 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
379 unsigned int doffset,
380 unsigned int nsiov, struct bio_vec *siov,
381 unsigned int soffset,
384 /* NB diov, siov are READ-ONLY */
385 unsigned int this_nob;
392 LASSERT (!in_interrupt ());
395 while (doffset >= diov->bv_len) {
396 doffset -= diov->bv_len;
403 while (soffset >= siov->bv_len) {
404 soffset -= siov->bv_len;
413 this_nob = min3(diov->bv_len - doffset,
414 siov->bv_len - soffset,
418 daddr = ((char *)kmap(diov->bv_page)) +
419 diov->bv_offset + doffset;
421 saddr = ((char *)kmap(siov->bv_page)) +
422 siov->bv_offset + soffset;
424 /* Vanishing risk of kmap deadlock when mapping 2 pages.
425 * However in practice at least one of the kiovs will be mapped
426 * kernel pages and the map/unmap will be NOOPs */
428 memcpy (daddr, saddr, this_nob);
431 if (diov->bv_len > doffset + this_nob) {
435 kunmap(diov->bv_page);
442 if (siov->bv_len > soffset + this_nob) {
446 kunmap(siov->bv_page);
455 kunmap(diov->bv_page);
457 kunmap(siov->bv_page);
459 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
462 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
463 unsigned int nkiov, struct bio_vec *kiov,
464 unsigned int kiovoffset,
467 /* NB iov, kiov are READ-ONLY */
468 unsigned int this_nob;
474 LASSERT (!in_interrupt ());
477 while (iovoffset >= iov->iov_len) {
478 iovoffset -= iov->iov_len;
485 while (kiovoffset >= kiov->bv_len) {
486 kiovoffset -= kiov->bv_len;
495 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
496 (unsigned int)kiov->bv_len - kiovoffset,
500 addr = ((char *)kmap(kiov->bv_page)) +
501 kiov->bv_offset + kiovoffset;
503 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
506 if (iov->iov_len > iovoffset + this_nob) {
507 iovoffset += this_nob;
514 if (kiov->bv_len > kiovoffset + this_nob) {
516 kiovoffset += this_nob;
518 kunmap(kiov->bv_page);
528 kunmap(kiov->bv_page);
530 EXPORT_SYMBOL(lnet_copy_kiov2iov);
533 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
534 unsigned int kiovoffset,
535 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
538 /* NB kiov, iov are READ-ONLY */
539 unsigned int this_nob;
545 LASSERT (!in_interrupt ());
548 while (kiovoffset >= kiov->bv_len) {
549 kiovoffset -= kiov->bv_len;
556 while (iovoffset >= iov->iov_len) {
557 iovoffset -= iov->iov_len;
566 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
567 (unsigned int)iov->iov_len - iovoffset,
571 addr = ((char *)kmap(kiov->bv_page)) +
572 kiov->bv_offset + kiovoffset;
574 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
577 if (kiov->bv_len > kiovoffset + this_nob) {
579 kiovoffset += this_nob;
581 kunmap(kiov->bv_page);
588 if (iov->iov_len > iovoffset + this_nob) {
589 iovoffset += this_nob;
598 kunmap(kiov->bv_page);
600 EXPORT_SYMBOL(lnet_copy_iov2kiov);
603 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
604 int src_niov, struct bio_vec *src,
605 unsigned int offset, unsigned int len)
607 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
608 * for exactly 'len' bytes, and return the number of entries.
609 * NB not destructive to 'src' */
610 unsigned int frag_len;
613 if (len == 0) /* no data => */
614 return (0); /* no frags */
616 LASSERT(src_niov > 0);
617 while (offset >= src->bv_len) { /* skip initial frags */
618 offset -= src->bv_len;
621 LASSERT(src_niov > 0);
626 LASSERT(src_niov > 0);
627 LASSERT((int)niov <= dst_niov);
629 frag_len = src->bv_len - offset;
630 dst->bv_page = src->bv_page;
631 dst->bv_offset = src->bv_offset + offset;
633 if (len <= frag_len) {
635 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
639 dst->bv_len = frag_len;
640 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
650 EXPORT_SYMBOL(lnet_extract_kiov);
653 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
654 int delayed, unsigned int offset, unsigned int mlen,
657 unsigned int niov = 0;
658 struct kvec *iov = NULL;
659 struct bio_vec *kiov = NULL;
662 LASSERT (!in_interrupt ());
663 LASSERT (mlen == 0 || msg != NULL);
666 LASSERT(msg->msg_receiving);
667 LASSERT(!msg->msg_sending);
668 LASSERT(rlen == msg->msg_len);
669 LASSERT(mlen <= msg->msg_len);
670 LASSERT(msg->msg_offset == offset);
671 LASSERT(msg->msg_wanted == mlen);
673 msg->msg_receiving = 0;
676 niov = msg->msg_niov;
677 kiov = msg->msg_kiov;
680 LASSERT ((iov == NULL) != (kiov == NULL));
684 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
685 niov, kiov, offset, mlen,
688 lnet_finalize(msg, rc);
692 lnet_setpayloadbuffer(struct lnet_msg *msg)
694 struct lnet_libmd *md = msg->msg_md;
696 LASSERT(msg->msg_len > 0);
697 LASSERT(!msg->msg_routing);
699 LASSERT(msg->msg_niov == 0);
700 LASSERT(msg->msg_kiov == NULL);
702 msg->msg_niov = md->md_niov;
703 msg->msg_kiov = md->md_kiov;
707 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
708 unsigned int offset, unsigned int len)
710 msg->msg_type = type;
711 msg->msg_target = target;
713 msg->msg_offset = offset;
716 lnet_setpayloadbuffer(msg);
718 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
719 msg->msg_hdr.type = cpu_to_le32(type);
720 /* dest_nid will be overwritten by lnet_select_pathway() */
721 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
722 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
723 /* src_nid will be set later */
724 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
725 msg->msg_hdr.payload_length = cpu_to_le32(len);
729 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
731 void *priv = msg->msg_private;
734 LASSERT(!in_interrupt());
735 LASSERT(ni->ni_nid == LNET_NID_LO_0 ||
736 (msg->msg_txcredit && msg->msg_peertxcredit));
738 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
740 msg->msg_no_resend = true;
741 lnet_finalize(msg, rc);
746 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
750 LASSERT(!msg->msg_sending);
751 LASSERT(msg->msg_receiving);
752 LASSERT(!msg->msg_rx_ready_delay);
753 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
755 msg->msg_rx_ready_delay = 1;
756 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
759 CERROR("recv from %s / send to %s aborted: "
760 "eager_recv failed %d\n",
761 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
762 libcfs_id2str(msg->msg_target), rc);
763 LASSERT(rc < 0); /* required by my callers */
770 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
774 deadline = lpni->lpni_last_alive +
775 lpni->lpni_net->net_tunables.lct_peer_timeout;
778 * assume peer_ni is alive as long as we're within the configured
787 /* NB: returns 1 when alive, 0 when dead, negative when error;
788 * may drop the lnet_net_lock */
790 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
791 struct lnet_msg *msg)
793 time64_t now = ktime_get_seconds();
795 if (!lnet_peer_aliveness_enabled(lpni))
799 * If we're resending a message, let's attempt to send it even if
800 * the peer is down to fulfill our resend quota on the message
802 if (msg->msg_retry_count > 0)
805 /* try and send recovery messages irregardless */
806 if (msg->msg_recovery)
809 /* always send any responses */
810 if (lnet_msg_is_response(msg))
813 if (!lnet_is_peer_deadline_passed(lpni, now))
816 return lnet_is_peer_ni_alive(lpni);
820 * \param msg The message to be sent.
821 * \param do_send True if lnet_ni_send() should be called in this function.
822 * lnet_send() is going to lnet_net_unlock immediately after this, so
823 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
825 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
826 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
827 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
828 * \retval -ECANCELED If the MD of the message has been unlinked.
831 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
833 struct lnet_peer_ni *lp = msg->msg_txpeer;
834 struct lnet_ni *ni = msg->msg_txni;
835 int cpt = msg->msg_tx_cpt;
836 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
838 /* non-lnet_send() callers have checked before */
839 LASSERT(!do_send || msg->msg_tx_delayed);
840 LASSERT(!msg->msg_receiving);
841 LASSERT(msg->msg_tx_committed);
843 /* can't get here if we're sending to the loopback interface */
844 if (the_lnet.ln_loni)
845 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
847 /* NB 'lp' is always the next hop */
848 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
849 lnet_peer_alive_locked(ni, lp, msg) == 0) {
850 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
851 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
853 lnet_net_unlock(cpt);
855 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
857 LNET_STATS_TYPE_DROP);
859 lnet_incr_stats(&msg->msg_txni->ni_stats,
861 LNET_STATS_TYPE_DROP);
863 CNETERR("Dropping message for %s: peer not alive\n",
864 libcfs_id2str(msg->msg_target));
865 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
867 lnet_finalize(msg, -EHOSTUNREACH);
870 return -EHOSTUNREACH;
873 if (msg->msg_md != NULL &&
874 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
875 lnet_net_unlock(cpt);
877 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
878 "called on the MD/ME.\n",
879 libcfs_id2str(msg->msg_target));
881 msg->msg_no_resend = true;
882 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
883 msg, libcfs_id2str(msg->msg_target));
884 lnet_finalize(msg, -ECANCELED);
891 if (!msg->msg_peertxcredit) {
892 spin_lock(&lp->lpni_lock);
893 LASSERT((lp->lpni_txcredits < 0) ==
894 !list_empty(&lp->lpni_txq));
896 msg->msg_peertxcredit = 1;
897 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
898 lp->lpni_txcredits--;
900 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
901 lp->lpni_mintxcredits = lp->lpni_txcredits;
903 if (lp->lpni_txcredits < 0) {
904 msg->msg_tx_delayed = 1;
905 list_add_tail(&msg->msg_list, &lp->lpni_txq);
906 spin_unlock(&lp->lpni_lock);
907 return LNET_CREDIT_WAIT;
909 spin_unlock(&lp->lpni_lock);
912 if (!msg->msg_txcredit) {
913 LASSERT((tq->tq_credits < 0) ==
914 !list_empty(&tq->tq_delayed));
916 msg->msg_txcredit = 1;
918 atomic_dec(&ni->ni_tx_credits);
920 if (tq->tq_credits < tq->tq_credits_min)
921 tq->tq_credits_min = tq->tq_credits;
923 if (tq->tq_credits < 0) {
924 msg->msg_tx_delayed = 1;
925 list_add_tail(&msg->msg_list, &tq->tq_delayed);
926 return LNET_CREDIT_WAIT;
930 /* unset the tx_delay flag as we're going to send it now */
931 msg->msg_tx_delayed = 0;
934 lnet_net_unlock(cpt);
935 lnet_ni_send(ni, msg);
938 return LNET_CREDIT_OK;
942 static struct lnet_rtrbufpool *
943 lnet_msg2bufpool(struct lnet_msg *msg)
945 struct lnet_rtrbufpool *rbp;
948 LASSERT(msg->msg_rx_committed);
950 cpt = msg->msg_rx_cpt;
951 rbp = &the_lnet.ln_rtrpools[cpt][0];
953 LASSERT(msg->msg_len <= LNET_MTU);
954 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
956 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
963 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
965 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
966 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
967 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
968 * received or OK to receive */
969 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
970 struct lnet_peer *lp;
971 struct lnet_rtrbufpool *rbp;
972 struct lnet_rtrbuf *rb;
974 LASSERT(msg->msg_kiov == NULL);
975 LASSERT(msg->msg_niov == 0);
976 LASSERT(msg->msg_routing);
977 LASSERT(msg->msg_receiving);
978 LASSERT(!msg->msg_sending);
979 LASSERT(lpni->lpni_peer_net);
980 LASSERT(lpni->lpni_peer_net->lpn_peer);
982 lp = lpni->lpni_peer_net->lpn_peer;
984 /* non-lnet_parse callers only receive delayed messages */
985 LASSERT(!do_recv || msg->msg_rx_delayed);
987 if (!msg->msg_peerrtrcredit) {
988 /* lpni_lock protects the credit manipulation */
989 spin_lock(&lpni->lpni_lock);
991 msg->msg_peerrtrcredit = 1;
992 lpni->lpni_rtrcredits--;
993 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
994 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
996 if (lpni->lpni_rtrcredits < 0) {
997 spin_unlock(&lpni->lpni_lock);
998 /* must have checked eager_recv before here */
999 LASSERT(msg->msg_rx_ready_delay);
1000 msg->msg_rx_delayed = 1;
1001 /* lp_lock protects the lp_rtrq */
1002 spin_lock(&lp->lp_lock);
1003 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1004 spin_unlock(&lp->lp_lock);
1005 return LNET_CREDIT_WAIT;
1007 spin_unlock(&lpni->lpni_lock);
1010 rbp = lnet_msg2bufpool(msg);
1012 if (!msg->msg_rtrcredit) {
1013 msg->msg_rtrcredit = 1;
1015 if (rbp->rbp_credits < rbp->rbp_mincredits)
1016 rbp->rbp_mincredits = rbp->rbp_credits;
1018 if (rbp->rbp_credits < 0) {
1019 /* must have checked eager_recv before here */
1020 LASSERT(msg->msg_rx_ready_delay);
1021 msg->msg_rx_delayed = 1;
1022 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1023 return LNET_CREDIT_WAIT;
1027 LASSERT(!list_empty(&rbp->rbp_bufs));
1028 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1029 list_del(&rb->rb_list);
1031 msg->msg_niov = rbp->rbp_npages;
1032 msg->msg_kiov = &rb->rb_kiov[0];
1034 /* unset the msg-rx_delayed flag since we're receiving the message */
1035 msg->msg_rx_delayed = 0;
1038 int cpt = msg->msg_rx_cpt;
1040 lnet_net_unlock(cpt);
1041 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1042 0, msg->msg_len, msg->msg_len);
1045 return LNET_CREDIT_OK;
1049 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1051 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1052 struct lnet_ni *txni = msg->msg_txni;
1053 struct lnet_msg *msg2;
1055 if (msg->msg_txcredit) {
1056 struct lnet_ni *ni = msg->msg_txni;
1057 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1059 /* give back NI txcredits */
1060 msg->msg_txcredit = 0;
1062 LASSERT((tq->tq_credits < 0) ==
1063 !list_empty(&tq->tq_delayed));
1066 atomic_inc(&ni->ni_tx_credits);
1067 if (tq->tq_credits <= 0) {
1068 msg2 = list_entry(tq->tq_delayed.next,
1069 struct lnet_msg, msg_list);
1070 list_del(&msg2->msg_list);
1072 LASSERT(msg2->msg_txni == ni);
1073 LASSERT(msg2->msg_tx_delayed);
1074 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1076 (void) lnet_post_send_locked(msg2, 1);
1080 if (msg->msg_peertxcredit) {
1081 /* give back peer txcredits */
1082 msg->msg_peertxcredit = 0;
1084 spin_lock(&txpeer->lpni_lock);
1085 LASSERT((txpeer->lpni_txcredits < 0) ==
1086 !list_empty(&txpeer->lpni_txq));
1088 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1089 LASSERT(txpeer->lpni_txqnob >= 0);
1091 txpeer->lpni_txcredits++;
1092 if (txpeer->lpni_txcredits <= 0) {
1095 msg2 = list_entry(txpeer->lpni_txq.next,
1096 struct lnet_msg, msg_list);
1097 list_del(&msg2->msg_list);
1098 spin_unlock(&txpeer->lpni_lock);
1100 LASSERT(msg2->msg_txpeer == txpeer);
1101 LASSERT(msg2->msg_tx_delayed);
1103 msg2_cpt = msg2->msg_tx_cpt;
1106 * The msg_cpt can be different from the msg2_cpt
1107 * so we need to make sure we lock the correct cpt
1109 * Once we call lnet_post_send_locked() it is no
1110 * longer safe to access msg2, since it could've
1111 * been freed by lnet_finalize(), but we still
1112 * need to relock the correct cpt, so we cache the
1113 * msg2_cpt for the purpose of the check that
1114 * follows the call to lnet_pose_send_locked().
1116 if (msg2_cpt != msg->msg_tx_cpt) {
1117 lnet_net_unlock(msg->msg_tx_cpt);
1118 lnet_net_lock(msg2_cpt);
1120 (void) lnet_post_send_locked(msg2, 1);
1121 if (msg2_cpt != msg->msg_tx_cpt) {
1122 lnet_net_unlock(msg2_cpt);
1123 lnet_net_lock(msg->msg_tx_cpt);
1126 spin_unlock(&txpeer->lpni_lock);
1131 msg->msg_txni = NULL;
1132 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1135 if (txpeer != NULL) {
1136 msg->msg_txpeer = NULL;
1137 lnet_peer_ni_decref_locked(txpeer);
1142 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1144 struct lnet_msg *msg;
1146 if (list_empty(&rbp->rbp_msgs))
1148 msg = list_entry(rbp->rbp_msgs.next,
1149 struct lnet_msg, msg_list);
1150 list_del(&msg->msg_list);
1152 (void)lnet_post_routed_recv_locked(msg, 1);
1156 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1158 struct lnet_msg *msg;
1159 struct lnet_msg *tmp;
1161 lnet_net_unlock(cpt);
1163 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1164 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1165 0, 0, 0, msg->msg_hdr.payload_length);
1166 list_del_init(&msg->msg_list);
1167 msg->msg_no_resend = true;
1168 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1169 lnet_finalize(msg, -ECANCELED);
1176 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1178 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1179 struct lnet_peer *lp;
1180 struct lnet_ni *rxni = msg->msg_rxni;
1181 struct lnet_msg *msg2;
1183 if (msg->msg_rtrcredit) {
1184 /* give back global router credits */
1185 struct lnet_rtrbuf *rb;
1186 struct lnet_rtrbufpool *rbp;
1188 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1189 * there until it gets one allocated, or aborts the wait
1191 LASSERT(msg->msg_kiov != NULL);
1193 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1196 msg->msg_kiov = NULL;
1197 msg->msg_rtrcredit = 0;
1199 LASSERT(rbp == lnet_msg2bufpool(msg));
1201 LASSERT((rbp->rbp_credits > 0) ==
1202 !list_empty(&rbp->rbp_bufs));
1204 /* If routing is now turned off, we just drop this buffer and
1205 * don't bother trying to return credits. */
1206 if (!the_lnet.ln_routing) {
1207 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1211 /* It is possible that a user has lowered the desired number of
1212 * buffers in this pool. Make sure we never put back
1213 * more buffers than the stated number. */
1214 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1215 /* Discard this buffer so we don't have too
1217 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1218 rbp->rbp_nbuffers--;
1220 list_add(&rb->rb_list, &rbp->rbp_bufs);
1222 if (rbp->rbp_credits <= 0)
1223 lnet_schedule_blocked_locked(rbp);
1228 if (msg->msg_peerrtrcredit) {
1230 LASSERT(rxpeerni->lpni_peer_net);
1231 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1233 /* give back peer router credits */
1234 msg->msg_peerrtrcredit = 0;
1236 spin_lock(&rxpeerni->lpni_lock);
1237 rxpeerni->lpni_rtrcredits++;
1238 spin_unlock(&rxpeerni->lpni_lock);
1240 lp = rxpeerni->lpni_peer_net->lpn_peer;
1241 spin_lock(&lp->lp_lock);
1243 /* drop all messages which are queued to be routed on that
1245 if (!the_lnet.ln_routing) {
1247 list_splice_init(&lp->lp_rtrq, &drop);
1248 spin_unlock(&lp->lp_lock);
1249 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1250 } else if (!list_empty(&lp->lp_rtrq)) {
1253 msg2 = list_entry(lp->lp_rtrq.next,
1254 struct lnet_msg, msg_list);
1255 list_del(&msg2->msg_list);
1256 msg2_cpt = msg2->msg_rx_cpt;
1257 spin_unlock(&lp->lp_lock);
1259 * messages on the lp_rtrq can be from any NID in
1260 * the peer, which means they might have different
1261 * cpts. We need to make sure we lock the right
1264 if (msg2_cpt != msg->msg_rx_cpt) {
1265 lnet_net_unlock(msg->msg_rx_cpt);
1266 lnet_net_lock(msg2_cpt);
1268 (void) lnet_post_routed_recv_locked(msg2, 1);
1269 if (msg2_cpt != msg->msg_rx_cpt) {
1270 lnet_net_unlock(msg2_cpt);
1271 lnet_net_lock(msg->msg_rx_cpt);
1274 spin_unlock(&lp->lp_lock);
1278 msg->msg_rxni = NULL;
1279 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1281 if (rxpeerni != NULL) {
1282 msg->msg_rxpeer = NULL;
1283 lnet_peer_ni_decref_locked(rxpeerni);
1287 static struct lnet_peer_ni *
1288 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1289 struct lnet_peer *peer,
1290 struct lnet_peer_ni *best_lpni,
1291 struct lnet_peer_net *peer_net)
1294 * Look at the peer NIs for the destination peer that connect
1295 * to the chosen net. If a peer_ni is preferred when using the
1296 * best_ni to communicate, we use that one. If there is no
1297 * preferred peer_ni, or there are multiple preferred peer_ni,
1298 * the available transmit credits are used. If the transmit
1299 * credits are equal, we round-robin over the peer_ni.
1301 struct lnet_peer_ni *lpni = NULL;
1302 int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1304 int best_lpni_healthv = (best_lpni) ?
1305 atomic_read(&best_lpni->lpni_healthv) : 0;
1306 bool best_lpni_is_preferred = false;
1307 bool lpni_is_preferred;
1309 __u32 lpni_sel_prio;
1310 __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1312 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1314 * if the best_ni we've chosen aleady has this lpni
1315 * preferred, then let's use it
1318 lpni_is_preferred = lnet_peer_is_pref_nid_locked(lpni,
1320 CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1321 libcfs_nid2str(best_ni->ni_nid),
1324 lpni_is_preferred = false;
1327 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1328 lpni_sel_prio = lpni->lpni_sel_priority;
1331 CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1332 libcfs_nid2str(lpni->lpni_nid),
1333 libcfs_nid2str(best_lpni->lpni_nid),
1334 lpni_healthv, best_lpni_healthv,
1335 lpni_sel_prio, best_sel_prio,
1336 lpni->lpni_txcredits, best_lpni_credits,
1337 lpni->lpni_seq, best_lpni->lpni_seq);
1341 /* pick the healthiest peer ni */
1342 if (lpni_healthv < best_lpni_healthv)
1344 else if (lpni_healthv > best_lpni_healthv) {
1345 if (best_lpni_is_preferred)
1346 best_lpni_is_preferred = false;
1350 if (lpni_sel_prio > best_sel_prio)
1352 else if (lpni_sel_prio < best_sel_prio) {
1353 if (best_lpni_is_preferred)
1354 best_lpni_is_preferred = false;
1358 /* if this is a preferred peer use it */
1359 if (!best_lpni_is_preferred && lpni_is_preferred) {
1360 best_lpni_is_preferred = true;
1362 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1363 /* this is not the preferred peer so let's ignore
1369 if (lpni->lpni_txcredits < best_lpni_credits)
1370 /* We already have a peer that has more credits
1371 * available than this one. No need to consider
1372 * this peer further.
1375 else if (lpni->lpni_txcredits > best_lpni_credits)
1378 /* The best peer found so far and the current peer
1379 * have the same number of available credits let's
1380 * make sure to select between them using Round Robin
1382 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1385 best_lpni_is_preferred = lpni_is_preferred;
1386 best_lpni_healthv = lpni_healthv;
1387 best_sel_prio = lpni_sel_prio;
1389 best_lpni_credits = lpni->lpni_txcredits;
1392 /* if we still can't find a peer ni then we can't reach it */
1394 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1395 LNET_NIDNET(dst_nid);
1396 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1397 libcfs_net2str(net_id));
1401 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1402 libcfs_nid2str(best_lpni->lpni_nid));
1408 * Prerequisite: the best_ni should already be set in the sd
1409 * Find the best lpni.
1410 * If the net id is provided then restrict lpni selection on
1411 * that particular net.
1412 * Otherwise find any reachable lpni. When dealing with an MR
1413 * gateway and it has multiple lpnis which we can use
1414 * we want to select the best one from the list of reachable
1417 static inline struct lnet_peer_ni *
1418 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1419 struct lnet_peer *peer, __u32 net_id)
1421 struct lnet_peer_net *peer_net;
1423 /* find the best_lpni on any local network */
1424 if (net_id == LNET_NET_ANY) {
1425 struct lnet_peer_ni *best_lpni = NULL;
1426 struct lnet_peer_net *lpn;
1427 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1428 /* no net specified find any reachable peer ni */
1429 if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1431 best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1437 /* restrict on the specified net */
1438 peer_net = lnet_peer_get_net_locked(peer, net_id);
1440 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1446 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1448 if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1451 if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1454 if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1457 if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1463 /* Compare route priorities and hop counts */
1465 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1467 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1468 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1470 if (r1->lr_priority < r2->lr_priority)
1473 if (r1->lr_priority > r2->lr_priority)
1476 if (r1_hops < r2_hops)
1479 if (r1_hops > r2_hops)
1485 static struct lnet_route *
1486 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1487 struct lnet_peer_ni *remote_lpni,
1488 struct lnet_route **prev_route,
1489 struct lnet_peer_ni **gwni)
1491 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1492 struct lnet_route *best_route;
1493 struct lnet_route *last_route;
1494 struct lnet_route *route;
1496 bool best_rte_is_preferred = false;
1499 CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1500 libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1502 best_route = last_route = NULL;
1503 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1504 if (!lnet_is_route_alive(route))
1506 gw_pnid = route->lr_gateway->lp_primary_nid;
1508 /* no protection on below fields, but it's harmless */
1509 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1512 /* if the best route found is in the preferred list then
1513 * tag it as preferred and use it later on. But if we
1514 * didn't find any routes which are on the preferred list
1515 * then just use the best route possible.
1517 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1519 if (!best_route || (rc && !best_rte_is_preferred)) {
1520 /* Restrict the selection of the router NI on the
1521 * src_net provided. If the src_net is LNET_NID_ANY,
1522 * then select the best interface available.
1524 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1529 "Gateway %s does not have a peer NI on net %s\n",
1530 libcfs_nid2str(gw_pnid),
1531 libcfs_net2str(src_net));
1536 if (rc && !best_rte_is_preferred) {
1537 /* This is the first preferred route we found,
1538 * so it beats any route found previously
1544 best_rte_is_preferred = true;
1545 CDEBUG(D_NET, "preferred gw = %s\n",
1546 libcfs_nid2str(gw_pnid));
1548 } else if ((!rc) && best_rte_is_preferred)
1549 /* The best route we found so far is in the preferred
1550 * list, so it beats any non-preferred route
1555 best_route = last_route = route;
1560 rc = lnet_compare_routes(route, best_route);
1564 /* Restrict the selection of the router NI on the
1565 * src_net provided. If the src_net is LNET_NID_ANY,
1566 * then select the best interface available.
1568 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1573 "Gateway %s does not have a peer NI on net %s\n",
1574 libcfs_nid2str(gw_pnid),
1575 libcfs_net2str(src_net));
1585 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1589 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1596 *prev_route = last_route;
1602 static struct lnet_ni *
1603 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1604 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1607 struct lnet_ni *ni = NULL;
1608 unsigned int shortest_distance;
1611 __u32 best_sel_prio;
1614 * If there is no peer_ni that we can send to on this network,
1615 * then there is no point in looking for a new best_ni here.
1617 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1620 if (best_ni == NULL) {
1621 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1622 shortest_distance = UINT_MAX;
1623 best_credits = INT_MIN;
1626 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1627 best_ni->ni_dev_cpt);
1628 best_credits = atomic_read(&best_ni->ni_tx_credits);
1629 best_healthv = atomic_read(&best_ni->ni_healthv);
1630 best_sel_prio = best_ni->ni_sel_priority;
1633 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1634 unsigned int distance;
1640 ni_credits = atomic_read(&ni->ni_tx_credits);
1641 ni_healthv = atomic_read(&ni->ni_healthv);
1642 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1643 ni_sel_prio = ni->ni_sel_priority;
1646 * calculate the distance from the CPT on which
1647 * the message memory is allocated to the CPT of
1648 * the NI's physical device
1650 distance = cfs_cpt_distance(lnet_cpt_table(),
1655 * All distances smaller than the NUMA range
1656 * are treated equally.
1658 if (distance < lnet_numa_range)
1659 distance = lnet_numa_range;
1662 * Select on health, shorter distance, available
1663 * credits, then round-robin.
1669 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",
1670 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1671 ni->ni_seq, ni_sel_prio,
1672 (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1673 : "not selected", best_credits, shortest_distance,
1674 (best_ni) ? best_ni->ni_seq : 0,
1679 if (ni_healthv < best_healthv)
1681 else if (ni_healthv > best_healthv)
1684 if (ni_sel_prio > best_sel_prio)
1686 else if (ni_sel_prio < best_sel_prio)
1689 if (distance > shortest_distance)
1691 else if (distance < shortest_distance)
1694 if (ni_credits < best_credits)
1696 else if (ni_credits > best_credits)
1699 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1703 best_sel_prio = ni_sel_prio;
1704 shortest_distance = distance;
1705 best_healthv = ni_healthv;
1707 best_credits = ni_credits;
1710 CDEBUG(D_NET, "selected best_ni %s\n",
1711 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1717 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1718 * because such traffic is required to perform discovery. We therefore
1719 * exclude all GET and PUT on that portal. We also exclude all ACK and
1720 * REPLY traffic, but that is because the portal is not tracked in the
1721 * message structure for these message types. We could restrict this
1722 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1725 lnet_msg_discovery(struct lnet_msg *msg)
1727 if (msg->msg_type == LNET_MSG_PUT) {
1728 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1730 } else if (msg->msg_type == LNET_MSG_GET) {
1731 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1737 #define SRC_SPEC 0x0001
1738 #define SRC_ANY 0x0002
1739 #define LOCAL_DST 0x0004
1740 #define REMOTE_DST 0x0008
1741 #define MR_DST 0x0010
1742 #define NMR_DST 0x0020
1743 #define SND_RESP 0x0040
1745 /* The following to defines are used for return codes */
1746 #define REPEAT_SEND 0x1000
1747 #define PASS_THROUGH 0x2000
1749 /* The different cases lnet_select pathway needs to handle */
1750 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1751 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1752 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1753 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1754 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1755 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1756 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1757 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1760 lnet_handle_lo_send(struct lnet_send_data *sd)
1762 struct lnet_msg *msg = sd->sd_msg;
1763 int cpt = sd->sd_cpt;
1765 if (the_lnet.ln_state != LNET_STATE_RUNNING)
1768 /* No send credit hassles with LOLND */
1769 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1770 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1771 if (!msg->msg_routing)
1772 msg->msg_hdr.src_nid =
1773 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1774 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1775 lnet_msg_commit(msg, cpt);
1776 msg->msg_txni = the_lnet.ln_loni;
1778 return LNET_CREDIT_OK;
1782 lnet_handle_send(struct lnet_send_data *sd)
1784 struct lnet_ni *best_ni = sd->sd_best_ni;
1785 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1786 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1787 struct lnet_msg *msg = sd->sd_msg;
1789 __u32 send_case = sd->sd_send_case;
1791 __u32 routing = send_case & REMOTE_DST;
1792 struct lnet_rsp_tracker *rspt;
1794 /* Increment sequence number of the selected peer, peer net,
1795 * local ni and local net so that we pick the next ones
1798 best_lpni->lpni_seq++;
1799 best_lpni->lpni_peer_net->lpn_seq++;
1801 best_ni->ni_net->net_seq++;
1803 CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1804 libcfs_nid2str(best_ni->ni_nid),
1805 best_ni->ni_seq, best_ni->ni_net->net_seq,
1806 atomic_read(&best_ni->ni_tx_credits),
1807 best_ni->ni_sel_priority,
1808 libcfs_nid2str(best_lpni->lpni_nid),
1809 best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1810 best_lpni->lpni_txcredits,
1811 best_lpni->lpni_sel_priority);
1814 * grab a reference on the peer_ni so it sticks around even if
1815 * we need to drop and relock the lnet_net_lock below.
1817 lnet_peer_ni_addref_locked(best_lpni);
1820 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1821 * message. This ensures that we get a CPT that is correct for
1822 * the NI when the NI has been restricted to a subset of all CPTs.
1823 * If the selected CPT differs from the one currently locked, we
1824 * must unlock and relock the lnet_net_lock(), and then check whether
1825 * the configuration has changed. We don't have a hold on the best_ni
1826 * yet, and it may have vanished.
1828 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1829 if (sd->sd_cpt != cpt2) {
1830 __u32 seq = lnet_get_dlc_seq_locked();
1831 lnet_net_unlock(sd->sd_cpt);
1833 lnet_net_lock(sd->sd_cpt);
1834 if (seq != lnet_get_dlc_seq_locked()) {
1835 lnet_peer_ni_decref_locked(best_lpni);
1841 * store the best_lpni in the message right away to avoid having
1842 * to do the same operation under different conditions
1844 msg->msg_txpeer = best_lpni;
1845 msg->msg_txni = best_ni;
1848 * grab a reference for the best_ni since now it's in use in this
1849 * send. The reference will be dropped in lnet_finalize()
1851 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1854 * Always set the target.nid to the best peer picked. Either the
1855 * NID will be one of the peer NIDs selected, or the same NID as
1856 * what was originally set in the target or it will be the NID of
1857 * a router if this message should be routed
1859 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1862 * lnet_msg_commit assigns the correct cpt to the message, which
1863 * is used to decrement the correct refcount on the ni when it's
1864 * time to return the credits
1866 lnet_msg_commit(msg, sd->sd_cpt);
1869 * If we are routing the message then we keep the src_nid that was
1870 * set by the originator. If we are not routing then we are the
1871 * originator and set it here.
1873 if (!msg->msg_routing)
1874 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1877 msg->msg_target_is_router = 1;
1878 msg->msg_target.pid = LNET_PID_LUSTRE;
1880 * since we're routing we want to ensure that the
1881 * msg_hdr.dest_nid is set to the final destination. When
1882 * the router receives this message it knows how to route
1885 * final_dst_lpni is set at the beginning of the
1886 * lnet_select_pathway() function and is never changed.
1887 * It's safe to use it here.
1889 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1892 * if we're not routing set the dest_nid to the best peer
1893 * ni NID that we picked earlier in the algorithm.
1895 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1899 * if we have response tracker block update it with the next hop
1903 rspt = msg->msg_md->md_rspt_ptr;
1905 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1906 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1907 libcfs_nid2str(rspt->rspt_next_hop_nid));
1911 rc = lnet_post_send_locked(msg, 0);
1914 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1915 libcfs_nid2str(msg->msg_hdr.src_nid),
1916 libcfs_nid2str(msg->msg_txni->ni_nid),
1917 libcfs_nid2str(sd->sd_src_nid),
1918 libcfs_nid2str(msg->msg_hdr.dest_nid),
1919 libcfs_nid2str(sd->sd_dst_nid),
1920 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1921 libcfs_nid2str(sd->sd_rtr_nid),
1922 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1928 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1929 struct lnet_msg *msg)
1931 if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1932 !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1933 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1934 libcfs_nid2str(lni->ni_nid),
1935 libcfs_nid2str(lpni->lpni_nid));
1936 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1945 * use the source and destination NIDs as the pathway
1948 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1950 /* the destination lpni is set before we get here. */
1953 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1954 if (!sd->sd_best_ni) {
1955 CERROR("Can't send to %s: src %s is not a "
1956 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1957 libcfs_nid2str(sd->sd_src_nid));
1961 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
1963 return lnet_handle_send(sd);
1971 * Don't run the selection algorithm on the peer NIs. By specifying the
1972 * local NID, we're also saying that we should always use the destination NID
1973 * provided. This handles the case where we should be using the same
1974 * destination NID for the all the messages which belong to the same RPC
1978 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1980 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1981 if (!sd->sd_best_ni) {
1982 CERROR("Can't send to %s: src %s is not a "
1983 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1984 libcfs_nid2str(sd->sd_src_nid));
1988 if (sd->sd_best_lpni &&
1989 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1990 return lnet_handle_lo_send(sd);
1991 else if (sd->sd_best_lpni)
1992 return lnet_handle_send(sd);
1994 CERROR("can't send to %s. no NI on %s\n",
1995 libcfs_nid2str(sd->sd_dst_nid),
1996 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1998 return -EHOSTUNREACH;
2002 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2003 struct lnet_peer *peer,
2004 struct lnet_peer_net *peer_net,
2007 struct lnet_net *local_net;
2008 struct lnet_ni *best_ni;
2010 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2015 * Iterate through the NIs in this local Net and select
2016 * the NI to send from. The selection is determined by
2017 * these 3 criterion in the following priority:
2019 * 2. NI available credits
2022 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2023 peer, peer_net, cpt);
2029 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2032 struct lnet_peer *peer;
2033 struct lnet_peer_ni *new_lpni;
2036 lnet_peer_ni_addref_locked(lpni);
2038 peer = lpni->lpni_peer_net->lpn_peer;
2040 if (lnet_peer_gw_discovery(peer)) {
2041 lnet_peer_ni_decref_locked(lpni);
2045 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2046 lnet_peer_ni_decref_locked(lpni);
2050 rc = lnet_discover_peer_locked(lpni, cpt, false);
2052 lnet_peer_ni_decref_locked(lpni);
2056 new_lpni = lnet_find_peer_ni_locked(lpni->lpni_nid);
2058 lnet_peer_ni_decref_locked(lpni);
2062 peer = new_lpni->lpni_peer_net->lpn_peer;
2063 spin_lock(&peer->lp_lock);
2064 if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2065 /* The peer NI did not change and the peer is up to date.
2066 * Nothing more to do.
2068 spin_unlock(&peer->lp_lock);
2069 lnet_peer_ni_decref_locked(lpni);
2070 lnet_peer_ni_decref_locked(new_lpni);
2073 spin_unlock(&peer->lp_lock);
2075 /* Either the peer NI changed during discovery, or the peer isn't up
2076 * to date. In both cases we want to queue the message on the
2077 * (possibly new) peer's pending queue and queue the peer for discovery
2079 msg->msg_sending = 0;
2080 msg->msg_txpeer = NULL;
2081 lnet_net_unlock(cpt);
2082 lnet_peer_queue_message(peer, msg);
2085 lnet_peer_ni_decref_locked(lpni);
2086 lnet_peer_ni_decref_locked(new_lpni);
2088 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2089 msg, libcfs_nid2str(peer->lp_primary_nid));
2091 return LNET_DC_WAIT;
2095 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2097 struct lnet_peer_ni **gw_lpni,
2098 struct lnet_peer **gw_peer)
2102 struct lnet_peer *gw;
2103 struct lnet_peer *lp;
2104 struct lnet_peer_net *lpn;
2105 struct lnet_peer_net *best_lpn = NULL;
2106 struct lnet_remotenet *rnet, *best_rnet = NULL;
2107 struct lnet_route *best_route = NULL;
2108 struct lnet_route *last_route = NULL;
2109 struct lnet_peer_ni *lpni = NULL;
2110 struct lnet_peer_ni *gwni = NULL;
2111 bool route_found = false;
2112 lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
2113 (sd->sd_best_ni != NULL) ? sd->sd_best_ni->ni_nid :
2115 int best_lpn_healthv = 0;
2116 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2118 CDEBUG(D_NET, "using src nid %s for route restriction\n",
2119 libcfs_nid2str(src_nid));
2121 /* If a router nid was specified then we are replying to a GET or
2122 * sending an ACK. In this case we use the gateway associated with the
2123 * specified router nid.
2125 if (sd->sd_rtr_nid != LNET_NID_ANY) {
2126 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2128 gw = gwni->lpni_peer_net->lpn_peer;
2129 lnet_peer_ni_decref_locked(gwni);
2130 if (gw->lp_rtr_refcount) {
2131 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2135 CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2136 libcfs_nid2str(sd->sd_rtr_nid));
2141 if (sd->sd_msg->msg_routing) {
2142 /* If I'm routing this message then I need to find the
2143 * next hop based on the destination NID
2145 best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
2147 CERROR("Unable to route message to %s - Route table may be misconfigured\n",
2148 libcfs_nid2str(sd->sd_dst_nid));
2149 return -EHOSTUNREACH;
2152 /* we've already looked up the initial lpni using
2155 lpni = sd->sd_best_lpni;
2156 /* the peer tree must be in existence */
2157 LASSERT(lpni && lpni->lpni_peer_net &&
2158 lpni->lpni_peer_net->lpn_peer);
2159 lp = lpni->lpni_peer_net->lpn_peer;
2161 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2162 /* is this remote network reachable? */
2163 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2172 /* select the preferred peer net */
2173 if (best_lpn_healthv > lpn->lpn_healthv)
2175 else if (best_lpn_healthv < lpn->lpn_healthv)
2178 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2180 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2183 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2186 best_lpn_healthv = lpn->lpn_healthv;
2187 best_lpn_sel_prio = lpn->lpn_sel_priority;
2193 CERROR("peer %s has no available nets\n",
2194 libcfs_nid2str(sd->sd_dst_nid));
2195 return -EHOSTUNREACH;
2198 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2201 best_lpn->lpn_net_id);
2202 if (!sd->sd_best_lpni) {
2203 CERROR("peer %s is unreachable\n",
2204 libcfs_nid2str(sd->sd_dst_nid));
2205 return -EHOSTUNREACH;
2208 /* We're attempting to round robin over the remote peer
2209 * NI's so update the final destination we selected
2211 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2213 /* Increment the sequence number of the remote lpni so
2214 * we can round robin over the different interfaces of
2217 sd->sd_best_lpni->lpni_seq++;
2221 * find the best route. Restrict the selection on the net of the
2222 * local NI if we've already picked the local NI to send from.
2223 * Otherwise, let's pick any route we can find and then find
2224 * a local NI we can reach the route's gateway on. Any route we
2225 * select will be reachable by virtue of the restriction we have
2226 * when adding a route.
2228 best_route = lnet_find_route_locked(best_rnet,
2229 LNET_NIDNET(src_nid),
2231 &last_route, &gwni);
2234 CERROR("no route to %s from %s\n",
2235 libcfs_nid2str(dst_nid),
2236 libcfs_nid2str(src_nid));
2237 return -EHOSTUNREACH;
2241 CERROR("Internal Error. Route expected to %s from %s\n",
2242 libcfs_nid2str(dst_nid),
2243 libcfs_nid2str(src_nid));
2247 gw = best_route->lr_gateway;
2248 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2249 local_lnet = best_route->lr_lnet;
2253 * Discover this gateway if it hasn't already been discovered.
2254 * This means we might delay the message until discovery has
2257 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2258 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2262 if (!sd->sd_best_ni)
2263 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2264 lnet_peer_get_net_locked(gw,
2268 if (!sd->sd_best_ni) {
2269 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2270 libcfs_net2str(local_lnet),
2271 libcfs_nid2str(sd->sd_src_nid));
2279 * increment the sequence numbers since now we're sure we're
2280 * going to use this path
2282 if (sd->sd_rtr_nid == LNET_NID_ANY) {
2283 LASSERT(best_route && last_route);
2284 best_route->lr_seq = last_route->lr_seq + 1;
2286 best_lpn->lpn_seq++;
2297 * Remote destination
2298 * Non-MR destination
2302 * Remote destination
2305 * The handling of these two cases is similar. Even though the destination
2306 * can be MR or non-MR, we'll deal directly with the router.
2309 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2312 struct lnet_peer_ni *gw_lpni = NULL;
2313 struct lnet_peer *gw_peer = NULL;
2316 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2317 if (!sd->sd_best_ni) {
2318 CERROR("Can't send to %s: src %s is not a "
2319 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2320 libcfs_nid2str(sd->sd_src_nid));
2324 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2329 if (sd->sd_send_case & NMR_DST)
2331 * since the final destination is non-MR let's set its preferred
2332 * NID before we send
2334 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2338 * We're going to send to the gw found so let's set its
2341 sd->sd_peer = gw_peer;
2342 sd->sd_best_lpni = gw_lpni;
2344 return lnet_handle_send(sd);
2348 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2351 struct lnet_peer_net *lpn = NULL;
2352 struct lnet_peer_net *best_lpn = NULL;
2353 struct lnet_net *net = NULL;
2354 struct lnet_net *best_net = NULL;
2355 struct lnet_ni *best_ni = NULL;
2356 int best_lpn_healthv = 0;
2357 int best_net_healthv = 0;
2359 __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2361 __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2366 * The peer can have multiple interfaces, some of them can be on
2367 * the local network and others on a routed network. We should
2368 * prefer the local network. However if the local network is not
2369 * available then we need to try the routed network
2372 /* go through all the peer nets and find the best_ni */
2373 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2375 * The peer's list of nets can contain non-local nets. We
2376 * want to only examine the local ones.
2378 net = lnet_get_net_locked(lpn->lpn_net_id);
2382 lpn_sel_prio = lpn->lpn_sel_priority;
2383 net_healthv = lnet_get_net_healthv_locked(net);
2384 net_sel_prio = net->net_sel_priority;
2387 * if this is a discovery message and lp_disc_net_id is
2388 * specified then use that net to send the discovery on.
2390 if (peer->lp_disc_net_id == lpn->lpn_net_id &&
2399 /* always select the lpn with the best health */
2400 if (best_lpn_healthv > lpn->lpn_healthv)
2402 else if (best_lpn_healthv < lpn->lpn_healthv)
2405 /* select the preferred peer and local nets */
2406 if (best_lpn_sel_prio < lpn_sel_prio)
2408 else if (best_lpn_sel_prio > lpn_sel_prio)
2411 if (best_net_healthv > net_healthv)
2413 else if (best_net_healthv < net_healthv)
2416 if (best_net_sel_prio < net_sel_prio)
2418 else if (best_net_sel_prio > net_sel_prio)
2421 if (best_lpn->lpn_seq < lpn->lpn_seq)
2423 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2426 /* round robin over the local networks */
2427 if (best_net->net_seq <= net->net_seq)
2431 best_net_healthv = net_healthv;
2432 best_net_sel_prio = net_sel_prio;
2433 best_lpn_healthv = lpn->lpn_healthv;
2434 best_lpn_sel_prio = lpn_sel_prio;
2443 /* Select the best NI on the same net as best_lpn chosen
2446 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer,
2453 static struct lnet_ni *
2454 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2456 struct lnet_ni *best_ni = NULL;
2457 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2458 struct lnet_peer_ni *lpni_entry;
2461 * We must use a consistent source address when sending to a
2462 * non-MR peer. However, a non-MR peer can have multiple NIDs
2463 * on multiple networks, and we may even need to talk to this
2464 * peer on multiple networks -- certain types of
2465 * load-balancing configuration do this.
2467 * So we need to pick the NI the peer prefers for this
2468 * particular network.
2471 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2473 if (lpni_entry->lpni_pref_nnids == 0)
2475 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2476 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2483 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2485 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2487 struct lnet_ni *best_ni = NULL;
2488 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2491 * We must use a consistent source address when sending to a
2492 * non-MR peer. However, a non-MR peer can have multiple NIDs
2493 * on multiple networks, and we may even need to talk to this
2494 * peer on multiple networks -- certain types of
2495 * load-balancing configuration do this.
2497 * So we need to pick the NI the peer prefers for this
2498 * particular network.
2501 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2504 /* if best_ni is still not set just pick one */
2507 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2508 sd->sd_best_lpni->lpni_peer_net,
2510 /* If there is no best_ni we don't have a route */
2512 CERROR("no path to %s from net %s\n",
2513 libcfs_nid2str(best_lpni->lpni_nid),
2514 libcfs_net2str(best_lpni->lpni_net->net_id));
2515 return -EHOSTUNREACH;
2519 sd->sd_best_ni = best_ni;
2521 /* Set preferred NI if necessary. */
2522 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2529 * Source not specified
2533 * always use the same source NID for NMR peers
2534 * If we've talked to that peer before then we already have a preferred
2535 * source NI associated with it. Otherwise, we select a preferred local NI
2536 * and store it in the peer
2539 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2543 /* sd->sd_best_lpni is already set to the final destination */
2546 * At this point we should've created the peer ni and peer. If we
2547 * can't find it, then something went wrong. Instead of assert
2548 * output a relevant message and fail the send
2550 if (!sd->sd_best_lpni) {
2551 CERROR("Internal fault. Unable to send msg %s to %s. "
2553 lnet_msgtyp2str(sd->sd_msg->msg_type),
2554 libcfs_nid2str(sd->sd_dst_nid));
2558 if (sd->sd_msg->msg_routing) {
2559 /* If I'm forwarding this message then I can choose any NI
2560 * on the destination peer net
2562 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2564 sd->sd_best_lpni->lpni_peer_net,
2566 if (!sd->sd_best_ni) {
2567 CERROR("Unable to forward message to %s. No local NI available\n",
2568 libcfs_nid2str(sd->sd_dst_nid));
2572 rc = lnet_select_preferred_best_ni(sd);
2575 rc = lnet_handle_send(sd);
2581 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2584 * NOTE we've already handled the remote peer case. So we only
2585 * need to worry about the local case here.
2587 * if we're sending a response, ACK or reply, we need to send it
2588 * to the destination NID given to us. At this point we already
2589 * have the peer_ni we're suppose to send to, so just find the
2590 * best_ni on the peer net and use that. Since we're sending to an
2591 * MR peer then we can just run the selection algorithm on our
2592 * local NIs and pick the best one.
2594 if (sd->sd_send_case & SND_RESP) {
2596 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2597 sd->sd_best_lpni->lpni_peer_net,
2600 if (!sd->sd_best_ni) {
2602 * We're not going to deal with not able to send
2603 * a response to the provided final destination
2605 CERROR("Can't send response to %s. "
2606 "No local NI available\n",
2607 libcfs_nid2str(sd->sd_dst_nid));
2608 return -EHOSTUNREACH;
2611 return lnet_handle_send(sd);
2615 * If we get here that means we're sending a fresh request, PUT or
2616 * GET, so we need to run our standard selection algorithm.
2617 * First find the best local interface that's on any of the peer's
2620 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2622 lnet_msg_discovery(sd->sd_msg));
2623 if (sd->sd_best_ni) {
2625 lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2627 sd->sd_best_ni->ni_net->net_id);
2630 * if we're successful in selecting a peer_ni on the local
2631 * network, then send to it. Otherwise fall through and
2632 * try and see if we can reach it over another routed
2635 if (sd->sd_best_lpni &&
2636 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2638 * in case we initially started with a routed
2639 * destination, let's reset to local
2641 sd->sd_send_case &= ~REMOTE_DST;
2642 sd->sd_send_case |= LOCAL_DST;
2643 return lnet_handle_lo_send(sd);
2644 } else if (sd->sd_best_lpni) {
2646 * in case we initially started with a routed
2647 * destination, let's reset to local
2649 sd->sd_send_case &= ~REMOTE_DST;
2650 sd->sd_send_case |= LOCAL_DST;
2651 return lnet_handle_send(sd);
2654 CERROR("Internal Error. Expected to have a best_lpni: "
2656 libcfs_nid2str(sd->sd_src_nid),
2657 libcfs_nid2str(sd->sd_dst_nid));
2663 * Peer doesn't have a local network. Let's see if there is
2664 * a remote network we can reach it on.
2666 return PASS_THROUGH;
2671 * Source NID not specified
2676 * Source NID not speified
2677 * Remote destination
2680 * In both of these cases if we're sending a response, ACK or REPLY, then
2681 * we need to send to the destination NID provided.
2683 * In the remote case let's deal with MR routers.
2688 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2691 struct lnet_peer *gw_peer = NULL;
2692 struct lnet_peer_ni *gw_lpni = NULL;
2695 * handle sending a response to a remote peer here so we don't
2696 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2698 if (sd->sd_send_case & REMOTE_DST &&
2699 sd->sd_send_case & SND_RESP) {
2700 struct lnet_peer_ni *gw;
2701 struct lnet_peer *gw_peer;
2703 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2706 CERROR("Can't send response to %s. "
2707 "No route available\n",
2708 libcfs_nid2str(sd->sd_dst_nid));
2709 return -EHOSTUNREACH;
2710 } else if (rc > 0) {
2714 sd->sd_best_lpni = gw;
2715 sd->sd_peer = gw_peer;
2717 return lnet_handle_send(sd);
2721 * Even though the NID for the peer might not be on a local network,
2722 * since the peer is MR there could be other interfaces on the
2723 * local network. In that case we'd still like to prefer the local
2724 * network over the routed network. If we're unable to do that
2725 * then we select the best router among the different routed networks,
2726 * and if the router is MR then we can deal with it as such.
2728 rc = lnet_handle_any_mr_dsta(sd);
2729 if (rc != PASS_THROUGH)
2733 * Now that we must route to the destination, we must consider the
2734 * MR case, where the destination has multiple interfaces, some of
2735 * which we can route to and others we do not. For this reason we
2736 * need to select the destination which we can route to and if
2737 * there are multiple, we need to round robin.
2739 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2744 sd->sd_send_case &= ~LOCAL_DST;
2745 sd->sd_send_case |= REMOTE_DST;
2747 sd->sd_peer = gw_peer;
2748 sd->sd_best_lpni = gw_lpni;
2750 return lnet_handle_send(sd);
2754 * Source not specified
2755 * Remote destination
2758 * Must send to the specified peer NID using the same source NID that
2759 * we've used before. If it's the first time to talk to that peer then
2760 * find the source NI and assign it as preferred to that peer
2763 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2766 struct lnet_peer_ni *gw_lpni = NULL;
2767 struct lnet_peer *gw_peer = NULL;
2770 * Let's see if we have a preferred NI to talk to this NMR peer
2772 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2776 * find the router and that'll find the best NI if we didn't find
2779 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2785 * set the best_ni we've chosen as the preferred one for
2788 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2790 /* we'll be sending to the gw */
2791 sd->sd_best_lpni = gw_lpni;
2792 sd->sd_peer = gw_peer;
2794 return lnet_handle_send(sd);
2798 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2801 * turn off the SND_RESP bit.
2802 * It will be checked in the case handling
2804 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2806 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2807 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2808 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2809 (send_case & MR_DST) ? "MR: " : "NMR: ",
2810 libcfs_nid2str(sd->sd_dst_nid),
2811 (send_case & LOCAL_DST) ? "local" : "routed");
2813 switch (send_case) {
2815 * For all cases where the source is specified, we should always
2816 * use the destination NID, whether it's an MR destination or not,
2817 * since we're continuing a series of related messages for the
2820 case SRC_SPEC_LOCAL_NMR_DST:
2821 return lnet_handle_spec_local_nmr_dst(sd);
2822 case SRC_SPEC_LOCAL_MR_DST:
2823 return lnet_handle_spec_local_mr_dst(sd);
2824 case SRC_SPEC_ROUTER_NMR_DST:
2825 case SRC_SPEC_ROUTER_MR_DST:
2826 return lnet_handle_spec_router_dst(sd);
2827 case SRC_ANY_LOCAL_NMR_DST:
2828 return lnet_handle_any_local_nmr_dst(sd);
2829 case SRC_ANY_LOCAL_MR_DST:
2830 case SRC_ANY_ROUTER_MR_DST:
2831 return lnet_handle_any_mr_dst(sd);
2832 case SRC_ANY_ROUTER_NMR_DST:
2833 return lnet_handle_any_router_nmr_dst(sd);
2835 CERROR("Unknown send case\n");
2841 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2842 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2844 struct lnet_peer_ni *lpni;
2845 struct lnet_peer *peer;
2846 struct lnet_send_data send_data;
2849 __u32 send_case = 0;
2851 bool mr_forwarding_allowed;
2853 memset(&send_data, 0, sizeof(send_data));
2856 * get an initial CPT to use for locking. The idea here is not to
2857 * serialize the calls to select_pathway, so that as many
2858 * operations can run concurrently as possible. To do that we use
2859 * the CPT where this call is being executed. Later on when we
2860 * determine the CPT to use in lnet_message_commit, we switch the
2861 * lock and check if there was any configuration change. If none,
2862 * then we proceed, if there is, then we restart the operation.
2864 cpt = lnet_net_lock_current();
2866 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2867 if (md_cpt == CFS_CPT_ANY)
2873 * If we're being asked to send to the loopback interface, there
2874 * is no need to go through any selection. We can just shortcut
2875 * the entire process and send over lolnd
2877 send_data.sd_msg = msg;
2878 send_data.sd_cpt = cpt;
2879 if (dst_nid == LNET_NID_LO_0) {
2880 rc = lnet_handle_lo_send(&send_data);
2881 lnet_net_unlock(cpt);
2886 * find an existing peer_ni, or create one and mark it as having been
2887 * created due to network traffic. This call will create the
2888 * peer->peer_net->peer_ni tree.
2890 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2892 lnet_net_unlock(cpt);
2893 return PTR_ERR(lpni);
2897 * Cache the original src_nid and rtr_nid. If we need to resend the
2898 * message then we'll need to know whether the src_nid was originally
2899 * specified for this message. If it was originally specified,
2900 * then we need to keep using the same src_nid since it's
2901 * continuing the same sequence of messages. Similarly, rtr_nid will
2902 * affect our choice of next hop.
2904 msg->msg_src_nid_param = src_nid;
2905 msg->msg_rtr_nid_param = rtr_nid;
2908 * If necessary, perform discovery on the peer that owns this peer_ni.
2909 * Note, this can result in the ownership of this peer_ni changing
2910 * to another peer object.
2912 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2914 lnet_peer_ni_decref_locked(lpni);
2915 lnet_net_unlock(cpt);
2918 lnet_peer_ni_decref_locked(lpni);
2920 peer = lpni->lpni_peer_net->lpn_peer;
2923 * Identify the different send cases
2925 if (src_nid == LNET_NID_ANY)
2926 send_case |= SRC_ANY;
2928 send_case |= SRC_SPEC;
2930 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2931 send_case |= LOCAL_DST;
2933 send_case |= REMOTE_DST;
2936 if (msg->msg_routing && (send_case & LOCAL_DST))
2939 /* Determine whether to allow MR forwarding for this message.
2940 * NB: MR forwarding is allowed if the message originator and the
2941 * destination are both MR capable, and the destination lpni that was
2942 * originally chosen by the originator is unhealthy or down.
2943 * We check the MR capability of the destination further below
2945 mr_forwarding_allowed = false;
2947 struct lnet_peer *src_lp;
2948 struct lnet_peer_ni *src_lpni;
2950 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
2952 /* We don't fail the send if we hit any errors here. We'll just
2953 * try to send it via non-multi-rail criteria
2955 if (!IS_ERR(src_lpni)) {
2956 /* Drop ref taken by lnet_nid2peerni_locked() */
2957 lnet_peer_ni_decref_locked(src_lpni);
2958 src_lp = lpni->lpni_peer_net->lpn_peer;
2959 if (lnet_peer_is_multi_rail(src_lp) &&
2960 !lnet_is_peer_ni_alive(lpni))
2961 mr_forwarding_allowed = true;
2964 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
2965 mr_forwarding_allowed ? "allowed" : "not allowed");
2969 * Deal with the peer as NMR in the following cases:
2970 * 1. the peer is NMR
2971 * 2. We're trying to recover a specific peer NI
2972 * 3. I'm a router sending to the final destination and MR forwarding is
2973 * not allowed for this message (as determined above).
2974 * In this case the source of the message would've
2975 * already selected the final destination so my job
2976 * is to honor the selection.
2978 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
2979 (final_hop && !mr_forwarding_allowed))
2980 send_case |= NMR_DST;
2982 send_case |= MR_DST;
2984 if (lnet_msg_is_response(msg))
2985 send_case |= SND_RESP;
2987 /* assign parameters to the send_data */
2988 send_data.sd_rtr_nid = rtr_nid;
2989 send_data.sd_src_nid = src_nid;
2990 send_data.sd_dst_nid = dst_nid;
2991 send_data.sd_best_lpni = lpni;
2993 * keep a pointer to the final destination in case we're going to
2994 * route, so we'll need to access it later
2996 send_data.sd_final_dst_lpni = lpni;
2997 send_data.sd_peer = peer;
2998 send_data.sd_md_cpt = md_cpt;
2999 send_data.sd_send_case = send_case;
3001 rc = lnet_handle_send_case_locked(&send_data);
3004 * Update the local cpt since send_data.sd_cpt might've been
3005 * updated as a result of calling lnet_handle_send_case_locked().
3007 cpt = send_data.sd_cpt;
3009 if (rc == REPEAT_SEND)
3012 lnet_net_unlock(cpt);
3018 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
3020 lnet_nid_t dst_nid = msg->msg_target.nid;
3024 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
3025 * but we might want to use pre-determined router for ACK/REPLY
3028 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3029 LASSERT(msg->msg_txpeer == NULL);
3030 LASSERT(msg->msg_txni == NULL);
3031 LASSERT(!msg->msg_sending);
3032 LASSERT(!msg->msg_target_is_router);
3033 LASSERT(!msg->msg_receiving);
3035 msg->msg_sending = 1;
3037 LASSERT(!msg->msg_tx_committed);
3039 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3041 if (rc == -EHOSTUNREACH)
3042 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3044 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3048 if (rc == LNET_CREDIT_OK)
3049 lnet_ni_send(msg->msg_txni, msg);
3051 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3055 enum lnet_mt_event_type {
3056 MT_TYPE_LOCAL_NI = 0,
3060 struct lnet_mt_event_info {
3061 enum lnet_mt_event_type mt_type;
3065 /* called with res_lock held */
3067 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3069 struct lnet_rsp_tracker *rspt;
3072 * msg has a refcount on the MD so the MD is not going away.
3073 * The rspt queue for the cpt is protected by
3074 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3076 if (!md->md_rspt_ptr)
3079 rspt = md->md_rspt_ptr;
3082 LASSERT(rspt->rspt_cpt == cpt);
3084 md->md_rspt_ptr = NULL;
3086 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3088 * The monitor thread has invalidated this handle because the
3089 * response timed out, but it failed to lookup the MD. That
3090 * means this response tracker is on the zombie list. We can
3091 * safely remove it under the resource lock (held by caller) and
3092 * free the response tracker block.
3094 list_del(&rspt->rspt_on_list);
3095 lnet_rspt_free(rspt, cpt);
3098 * invalidate the handle to indicate that a response has been
3099 * received, which will then lead the monitor thread to clean up
3102 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3107 lnet_clean_zombie_rstqs(void)
3109 struct lnet_rsp_tracker *rspt, *tmp;
3112 cfs_cpt_for_each(i, lnet_cpt_table()) {
3113 list_for_each_entry_safe(rspt, tmp,
3114 the_lnet.ln_mt_zombie_rstqs[i],
3116 list_del(&rspt->rspt_on_list);
3117 lnet_rspt_free(rspt, i);
3121 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3125 lnet_finalize_expired_responses(void)
3127 struct lnet_libmd *md;
3128 struct lnet_rsp_tracker *rspt, *tmp;
3132 if (the_lnet.ln_mt_rstq == NULL)
3135 cfs_cpt_for_each(i, lnet_cpt_table()) {
3136 LIST_HEAD(local_queue);
3139 if (!the_lnet.ln_mt_rstq[i]) {
3143 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3148 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3150 * The rspt mdh will be invalidated when a response
3151 * is received or whenever we want to discard the
3152 * block the monitor thread will walk the queue
3153 * and clean up any rsts with an invalid mdh.
3154 * The monitor thread will walk the queue until
3155 * the first unexpired rspt block. This means that
3156 * some rspt blocks which received their
3157 * corresponding responses will linger in the
3158 * queue until they are cleaned up eventually.
3161 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3163 list_del(&rspt->rspt_on_list);
3164 lnet_rspt_free(rspt, i);
3168 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3169 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3170 struct lnet_peer_ni *lpni;
3173 md = lnet_handle2md(&rspt->rspt_mdh);
3175 /* MD has been queued for unlink, but
3176 * rspt hasn't been detached (Note we've
3177 * checked above that the rspt_mdh is
3178 * valid). Since we cannot lookup the MD
3179 * we're unable to detach the rspt
3180 * ourselves. Thus, move the rspt to the
3181 * zombie list where we'll wait for
3183 * 1. The remaining operations on the
3184 * MD to complete. In this case the
3185 * final operation will result in
3186 * lnet_msg_detach_md()->
3187 * lnet_detach_rsp_tracker() where
3188 * we will clean up this response
3190 * 2. LNet to shutdown. In this case
3191 * we'll wait until after all LND Nets
3192 * have shutdown and then we can
3193 * safely free any remaining response
3194 * tracker blocks on the zombie list.
3195 * Note: We need to hold the resource
3196 * lock when adding to the zombie list
3197 * because we may have concurrent access
3198 * with lnet_detach_rsp_tracker().
3200 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3201 list_move(&rspt->rspt_on_list,
3202 the_lnet.ln_mt_zombie_rstqs[i]);
3206 LASSERT(md->md_rspt_ptr == rspt);
3207 md->md_rspt_ptr = NULL;
3210 LNetMDUnlink(rspt->rspt_mdh);
3212 nid = rspt->rspt_next_hop_nid;
3214 list_del(&rspt->rspt_on_list);
3215 lnet_rspt_free(rspt, i);
3217 /* If we're shutting down we just want to clean
3218 * up the rspt blocks
3220 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3224 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3228 "Response timeout: md = %p: nid = %s\n",
3229 md, libcfs_nid2str(nid));
3232 * If there is a timeout on the response
3233 * from the next hop decrement its health
3234 * value so that we don't use it
3237 lpni = lnet_find_peer_ni_locked(nid);
3239 lnet_handle_remote_failure_locked(lpni);
3240 lnet_peer_ni_decref_locked(lpni);
3249 if (!list_empty(&local_queue)) {
3251 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3258 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3260 struct lnet_msg *msg;
3262 while (!list_empty(resendq)) {
3263 struct lnet_peer_ni *lpni;
3265 msg = list_entry(resendq->next, struct lnet_msg,
3268 list_del_init(&msg->msg_list);
3270 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3272 lnet_net_unlock(cpt);
3273 CERROR("Expected that a peer is already created for %s\n",
3274 libcfs_nid2str(msg->msg_hdr.dest_nid));
3275 msg->msg_no_resend = true;
3276 lnet_finalize(msg, -EFAULT);
3281 lnet_peer_ni_decref_locked(lpni);
3283 lnet_net_unlock(cpt);
3284 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3285 libcfs_nid2str(msg->msg_src_nid_param),
3286 libcfs_id2str(msg->msg_target),
3287 lnet_msgtyp2str(msg->msg_type),
3289 msg->msg_retry_count);
3290 rc = lnet_send(msg->msg_src_nid_param, msg,
3291 msg->msg_rtr_nid_param);
3293 CERROR("Error sending %s to %s: %d\n",
3294 lnet_msgtyp2str(msg->msg_type),
3295 libcfs_id2str(msg->msg_target), rc);
3296 msg->msg_no_resend = true;
3297 lnet_finalize(msg, rc);
3301 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3307 lnet_resend_pending_msgs(void)
3311 cfs_cpt_for_each(i, lnet_cpt_table()) {
3313 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3318 /* called with cpt and ni_lock held */
3320 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3322 struct lnet_handle_md recovery_mdh;
3324 LNetInvalidateMDHandle(&recovery_mdh);
3326 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3328 recovery_mdh = ni->ni_ping_mdh;
3329 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3332 lnet_net_unlock(cpt);
3333 if (!LNetMDHandleIsInvalid(recovery_mdh))
3334 LNetMDUnlink(recovery_mdh);
3340 lnet_recover_local_nis(void)
3342 struct lnet_mt_event_info *ev_info;
3343 LIST_HEAD(processed_list);
3344 LIST_HEAD(local_queue);
3345 struct lnet_handle_md mdh;
3346 struct lnet_ni *tmp;
3354 * splice the recovery queue on a local queue. We will iterate
3355 * through the local queue and update it as needed. Once we're
3356 * done with the traversal, we'll splice the local queue back on
3357 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3358 * will be traversed in the next iteration.
3361 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3365 now = ktime_get_seconds();
3367 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3369 * if an NI is being deleted or it is now healthy, there
3370 * is no need to keep it around in the recovery queue.
3371 * The monitor thread is the only thread responsible for
3372 * removing the NI from the recovery queue.
3373 * Multiple threads can be adding NIs to the recovery
3376 healthv = atomic_read(&ni->ni_healthv);
3380 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3381 healthv == LNET_MAX_HEALTH_VALUE) {
3382 list_del_init(&ni->ni_recovery);
3383 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3385 lnet_ni_decref_locked(ni, 0);
3391 * if the local NI failed recovery we must unlink the md.
3392 * But we want to keep the local_ni on the recovery queue
3393 * so we can continue the attempts to recover it.
3395 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3396 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3397 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3403 if (now < ni->ni_next_ping) {
3410 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3411 libcfs_nid2str(ni->ni_nid));
3414 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3415 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3418 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3420 CERROR("out of memory. Can't recover %s\n",
3421 libcfs_nid2str(ni->ni_nid));
3423 ni->ni_recovery_state &=
3424 ~LNET_NI_RECOVERY_PENDING;
3429 mdh = ni->ni_ping_mdh;
3431 * Invalidate the ni mdh in case it's deleted.
3432 * We'll unlink the mdh in this case below.
3434 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3438 * remove the NI from the local queue and drop the
3439 * reference count to it while we're recovering
3440 * it. The reason for that, is that the NI could
3441 * be deleted, and the way the code is structured
3442 * is if we don't drop the NI, then the deletion
3443 * code will enter a loop waiting for the
3444 * reference count to be removed while holding the
3445 * ln_mutex_lock(). When we look up the peer to
3446 * send to in lnet_select_pathway() we will try to
3447 * lock the ln_mutex_lock() as well, leading to
3448 * a deadlock. By dropping the refcount and
3449 * removing it from the list, we allow for the NI
3450 * to be removed, then we use the cached NID to
3451 * look it up again. If it's gone, then we just
3452 * continue examining the rest of the queue.
3455 list_del_init(&ni->ni_recovery);
3456 lnet_ni_decref_locked(ni, 0);
3459 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3460 ev_info->mt_nid = nid;
3461 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3462 ev_info, the_lnet.ln_mt_handler,
3464 /* lookup the nid again */
3466 ni = lnet_nid2ni_locked(nid, 0);
3469 * the NI has been deleted when we dropped
3476 ni->ni_ping_count++;
3478 ni->ni_ping_mdh = mdh;
3479 lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3484 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3493 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3494 * reexamined in the next iteration.
3496 list_splice_init(&processed_list, &local_queue);
3498 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3503 lnet_resendqs_create(void)
3505 struct list_head **resendqs;
3506 resendqs = lnet_create_array_of_queues();
3511 lnet_net_lock(LNET_LOCK_EX);
3512 the_lnet.ln_mt_resendqs = resendqs;
3513 lnet_net_unlock(LNET_LOCK_EX);
3519 lnet_clean_local_ni_recoveryq(void)
3523 /* This is only called when the monitor thread has stopped */
3526 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3527 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3528 struct lnet_ni, ni_recovery);
3529 list_del_init(&ni->ni_recovery);
3531 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3533 lnet_ni_decref_locked(ni, 0);
3540 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3543 struct lnet_handle_md recovery_mdh;
3545 LNetInvalidateMDHandle(&recovery_mdh);
3547 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3548 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3549 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3551 spin_unlock(&lpni->lpni_lock);
3552 lnet_net_unlock(cpt);
3553 if (!LNetMDHandleIsInvalid(recovery_mdh))
3554 LNetMDUnlink(recovery_mdh);
3556 spin_lock(&lpni->lpni_lock);
3560 lnet_clean_peer_ni_recoveryq(void)
3562 struct lnet_peer_ni *lpni, *tmp;
3564 lnet_net_lock(LNET_LOCK_EX);
3566 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3568 list_del_init(&lpni->lpni_recovery);
3569 spin_lock(&lpni->lpni_lock);
3570 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3571 spin_unlock(&lpni->lpni_lock);
3572 lnet_peer_ni_decref_locked(lpni);
3575 lnet_net_unlock(LNET_LOCK_EX);
3579 lnet_clean_resendqs(void)
3581 struct lnet_msg *msg, *tmp;
3585 cfs_cpt_for_each(i, lnet_cpt_table()) {
3587 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3589 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3590 list_del_init(&msg->msg_list);
3591 msg->msg_no_resend = true;
3592 lnet_finalize(msg, -ESHUTDOWN);
3596 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3600 lnet_recover_peer_nis(void)
3602 struct lnet_mt_event_info *ev_info;
3603 LIST_HEAD(processed_list);
3604 LIST_HEAD(local_queue);
3605 struct lnet_handle_md mdh;
3606 struct lnet_peer_ni *lpni;
3607 struct lnet_peer_ni *tmp;
3614 * Always use cpt 0 for locking across all interactions with
3615 * ln_mt_peerNIRecovq
3618 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3622 now = ktime_get_seconds();
3624 list_for_each_entry_safe(lpni, tmp, &local_queue,
3627 * The same protection strategy is used here as is in the
3628 * local recovery case.
3631 healthv = atomic_read(&lpni->lpni_healthv);
3632 spin_lock(&lpni->lpni_lock);
3633 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3634 healthv == LNET_MAX_HEALTH_VALUE) {
3635 list_del_init(&lpni->lpni_recovery);
3636 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3637 spin_unlock(&lpni->lpni_lock);
3638 lnet_peer_ni_decref_locked(lpni);
3644 * If the peer NI has failed recovery we must unlink the
3645 * md. But we want to keep the peer ni on the recovery
3646 * queue so we can try to continue recovering it
3648 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3649 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3650 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3653 spin_unlock(&lpni->lpni_lock);
3655 if (now < lpni->lpni_next_ping) {
3663 * NOTE: we're racing with peer deletion from user space.
3664 * It's possible that a peer is deleted after we check its
3665 * state. In this case the recovery can create a new peer
3667 spin_lock(&lpni->lpni_lock);
3668 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3669 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3670 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3671 spin_unlock(&lpni->lpni_lock);
3673 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3675 CERROR("out of memory. Can't recover %s\n",
3676 libcfs_nid2str(lpni->lpni_nid));
3677 spin_lock(&lpni->lpni_lock);
3678 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3679 spin_unlock(&lpni->lpni_lock);
3683 /* look at the comments in lnet_recover_local_nis() */
3684 mdh = lpni->lpni_recovery_ping_mdh;
3685 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3686 nid = lpni->lpni_nid;
3688 list_del_init(&lpni->lpni_recovery);
3689 lnet_peer_ni_decref_locked(lpni);
3692 ev_info->mt_type = MT_TYPE_PEER_NI;
3693 ev_info->mt_nid = nid;
3694 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3695 ev_info, the_lnet.ln_mt_handler,
3699 * lnet_find_peer_ni_locked() grabs a refcount for
3700 * us. No need to take it explicitly.
3702 lpni = lnet_find_peer_ni_locked(nid);
3709 lpni->lpni_ping_count++;
3711 lpni->lpni_recovery_ping_mdh = mdh;
3713 lnet_peer_ni_add_to_recoveryq_locked(lpni,
3717 spin_lock(&lpni->lpni_lock);
3718 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3719 spin_unlock(&lpni->lpni_lock);
3722 /* Drop the ref taken by lnet_find_peer_ni_locked() */
3723 lnet_peer_ni_decref_locked(lpni);
3726 spin_unlock(&lpni->lpni_lock);
3729 list_splice_init(&processed_list, &local_queue);
3731 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3736 lnet_monitor_thread(void *arg)
3738 time64_t rsp_timeout = 0;
3741 wait_for_completion(&the_lnet.ln_started);
3743 * The monitor thread takes care of the following:
3744 * 1. Checks the aliveness of routers
3745 * 2. Checks if there are messages on the resend queue to resend
3747 * 3. Check if there are any NIs on the local recovery queue and
3749 * 4. Checks if there are any NIs on the remote recovery queue
3752 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3753 now = ktime_get_real_seconds();
3755 if (lnet_router_checker_active())
3756 lnet_check_routers();
3758 lnet_resend_pending_msgs();
3760 if (now >= rsp_timeout) {
3761 lnet_finalize_expired_responses();
3762 rsp_timeout = now + (lnet_transaction_timeout / 2);
3765 lnet_recover_local_nis();
3766 lnet_recover_peer_nis();
3769 * TODO do we need to check if we should sleep without
3770 * timeout? Technically, an active system will always
3771 * have messages in flight so this check will always
3772 * evaluate to false. And on an idle system do we care
3773 * if we wake up every 1 second? Although, we've seen
3774 * cases where we get a complaint that an idle thread
3775 * is waking up unnecessarily.
3777 wait_for_completion_interruptible_timeout(
3778 &the_lnet.ln_mt_wait_complete,
3779 cfs_time_seconds(1));
3780 /* Must re-init the completion before testing anything,
3781 * including ln_mt_state.
3783 reinit_completion(&the_lnet.ln_mt_wait_complete);
3787 lnet_net_lock(LNET_LOCK_EX);
3788 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3789 lnet_net_unlock(LNET_LOCK_EX);
3791 /* signal that the monitor thread is exiting */
3792 up(&the_lnet.ln_mt_signal);
3800 * Returns == 0 if success
3801 * Returns > 0 if LNetMDBind or prior fails
3802 * Returns < 0 if LNetGet fails
3805 lnet_send_ping(lnet_nid_t dest_nid,
3806 struct lnet_handle_md *mdh, int nnis,
3807 void *user_data, lnet_handler_t handler, bool recovery)
3809 struct lnet_md md = { NULL };
3810 struct lnet_process_id id;
3811 struct lnet_ping_buffer *pbuf;
3814 if (dest_nid == LNET_NID_ANY) {
3819 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3825 /* initialize md content */
3826 md.start = &pbuf->pb_info;
3827 md.length = LNET_PING_INFO_SIZE(nnis);
3828 md.threshold = 2; /* GET/REPLY */
3830 md.options = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3831 md.user_ptr = user_data;
3832 md.handler = handler;
3834 rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3836 lnet_ping_buffer_decref(pbuf);
3837 CERROR("Can't bind MD: %d\n", rc);
3838 rc = -rc; /* change the rc to positive */
3841 id.pid = LNET_PID_LUSTRE;
3844 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3845 LNET_RESERVED_PORTAL,
3846 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3849 goto fail_unlink_md;
3855 LNetInvalidateMDHandle(mdh);
3861 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3862 int status, bool send, bool unlink_event)
3864 lnet_nid_t nid = ev_info->mt_nid;
3866 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3870 ni = lnet_nid2ni_locked(nid, 0);
3876 if (!send || (send && status != 0))
3877 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3879 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3884 CERROR("local NI (%s) recovery failed with %d\n",
3885 libcfs_nid2str(nid), status);
3889 * need to increment healthv for the ni here, because in
3890 * the lnet_finalize() path we don't have access to this
3891 * NI. And in order to get access to it, we'll need to
3892 * carry forward too much information.
3893 * In the peer case, it'll naturally be incremented
3896 lnet_inc_healthv(&ni->ni_healthv,
3897 lnet_health_sensitivity);
3899 struct lnet_peer_ni *lpni;
3902 cpt = lnet_net_lock_current();
3903 lpni = lnet_find_peer_ni_locked(nid);
3905 lnet_net_unlock(cpt);
3908 spin_lock(&lpni->lpni_lock);
3909 if (!send || (send && status != 0))
3910 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3912 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3913 spin_unlock(&lpni->lpni_lock);
3914 lnet_peer_ni_decref_locked(lpni);
3915 lnet_net_unlock(cpt);
3918 CERROR("peer NI (%s) recovery failed with %d\n",
3919 libcfs_nid2str(nid), status);
3924 lnet_mt_event_handler(struct lnet_event *event)
3926 struct lnet_mt_event_info *ev_info = event->md_user_ptr;
3927 struct lnet_ping_buffer *pbuf;
3929 /* TODO: remove assert */
3930 LASSERT(event->type == LNET_EVENT_REPLY ||
3931 event->type == LNET_EVENT_SEND ||
3932 event->type == LNET_EVENT_UNLINK);
3934 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3937 switch (event->type) {
3938 case LNET_EVENT_UNLINK:
3939 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3940 libcfs_nid2str(ev_info->mt_nid));
3942 case LNET_EVENT_REPLY:
3943 lnet_handle_recovery_reply(ev_info, event->status, false,
3944 event->type == LNET_EVENT_UNLINK);
3946 case LNET_EVENT_SEND:
3947 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3948 libcfs_nid2str(ev_info->mt_nid),
3949 (event->status) ? "unsuccessfully" :
3950 "successfully", event->status);
3951 lnet_handle_recovery_reply(ev_info, event->status, true, false);
3954 CERROR("Unexpected event: %d\n", event->type);
3957 if (event->unlinked) {
3958 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3959 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
3960 lnet_ping_buffer_decref(pbuf);
3965 lnet_rsp_tracker_create(void)
3967 struct list_head **rstqs;
3968 rstqs = lnet_create_array_of_queues();
3973 the_lnet.ln_mt_rstq = rstqs;
3979 lnet_rsp_tracker_clean(void)
3981 lnet_finalize_expired_responses();
3983 cfs_percpt_free(the_lnet.ln_mt_rstq);
3984 the_lnet.ln_mt_rstq = NULL;
3987 int lnet_monitor_thr_start(void)
3990 struct task_struct *task;
3992 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3995 rc = lnet_resendqs_create();
3999 rc = lnet_rsp_tracker_create();
4003 sema_init(&the_lnet.ln_mt_signal, 0);
4005 lnet_net_lock(LNET_LOCK_EX);
4006 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4007 lnet_net_unlock(LNET_LOCK_EX);
4008 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4011 CERROR("Can't start monitor thread: %d\n", rc);
4018 lnet_net_lock(LNET_LOCK_EX);
4019 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4020 lnet_net_unlock(LNET_LOCK_EX);
4021 /* block until event callback signals exit */
4022 down(&the_lnet.ln_mt_signal);
4024 lnet_net_lock(LNET_LOCK_EX);
4025 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4026 lnet_net_unlock(LNET_LOCK_EX);
4027 lnet_rsp_tracker_clean();
4028 lnet_clean_local_ni_recoveryq();
4029 lnet_clean_peer_ni_recoveryq();
4030 lnet_clean_resendqs();
4031 the_lnet.ln_mt_handler = NULL;
4034 lnet_rsp_tracker_clean();
4035 lnet_clean_local_ni_recoveryq();
4036 lnet_clean_peer_ni_recoveryq();
4037 lnet_clean_resendqs();
4041 void lnet_monitor_thr_stop(void)
4043 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4046 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4047 lnet_net_lock(LNET_LOCK_EX);
4048 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4049 lnet_net_unlock(LNET_LOCK_EX);
4051 /* tell the monitor thread that we're shutting down */
4052 complete(&the_lnet.ln_mt_wait_complete);
4054 /* block until monitor thread signals that it's done */
4055 down(&the_lnet.ln_mt_signal);
4056 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4058 /* perform cleanup tasks */
4059 lnet_rsp_tracker_clean();
4060 lnet_clean_local_ni_recoveryq();
4061 lnet_clean_peer_ni_recoveryq();
4062 lnet_clean_resendqs();
4066 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4070 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4071 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4072 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4073 lnet_net_unlock(cpt);
4075 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4079 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4081 struct lnet_hdr *hdr = &msg->msg_hdr;
4083 if (msg->msg_wanted != 0)
4084 lnet_setpayloadbuffer(msg);
4086 lnet_build_msg_event(msg, LNET_EVENT_PUT);
4088 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
4089 * it back into the ACK during lnet_finalize() */
4090 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4091 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4093 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4094 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4098 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4100 struct lnet_hdr *hdr = &msg->msg_hdr;
4101 struct lnet_match_info info;
4105 /* Convert put fields to host byte order */
4106 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4107 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
4108 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
4110 /* Primary peer NID. */
4111 info.mi_id.nid = msg->msg_initiator;
4112 info.mi_id.pid = hdr->src_pid;
4113 info.mi_opc = LNET_MD_OP_PUT;
4114 info.mi_portal = hdr->msg.put.ptl_index;
4115 info.mi_rlength = hdr->payload_length;
4116 info.mi_roffset = hdr->msg.put.offset;
4117 info.mi_mbits = hdr->msg.put.match_bits;
4118 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
4120 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4121 ready_delay = msg->msg_rx_ready_delay;
4124 rc = lnet_ptl_match_md(&info, msg);
4129 case LNET_MATCHMD_OK:
4130 lnet_recv_put(ni, msg);
4133 case LNET_MATCHMD_NONE:
4135 /* no eager_recv or has already called it, should
4136 * have been attached on delayed list */
4139 rc = lnet_ni_eager_recv(ni, msg);
4146 case LNET_MATCHMD_DROP:
4147 CNETERR("Dropping PUT from %s portal %d match %llu"
4148 " offset %d length %d: %d\n",
4149 libcfs_id2str(info.mi_id), info.mi_portal,
4150 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4152 return -ENOENT; /* -ve: OK but no match */
4157 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4159 struct lnet_match_info info;
4160 struct lnet_hdr *hdr = &msg->msg_hdr;
4161 struct lnet_process_id source_id;
4162 struct lnet_handle_wire reply_wmd;
4165 /* Convert get fields to host byte order */
4166 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
4167 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
4168 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
4169 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
4171 source_id.nid = hdr->src_nid;
4172 source_id.pid = hdr->src_pid;
4173 /* Primary peer NID */
4174 info.mi_id.nid = msg->msg_initiator;
4175 info.mi_id.pid = hdr->src_pid;
4176 info.mi_opc = LNET_MD_OP_GET;
4177 info.mi_portal = hdr->msg.get.ptl_index;
4178 info.mi_rlength = hdr->msg.get.sink_length;
4179 info.mi_roffset = hdr->msg.get.src_offset;
4180 info.mi_mbits = hdr->msg.get.match_bits;
4181 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
4183 rc = lnet_ptl_match_md(&info, msg);
4184 if (rc == LNET_MATCHMD_DROP) {
4185 CNETERR("Dropping GET from %s portal %d match %llu"
4186 " offset %d length %d\n",
4187 libcfs_id2str(info.mi_id), info.mi_portal,
4188 info.mi_mbits, info.mi_roffset, info.mi_rlength);
4189 return -ENOENT; /* -ve: OK but no match */
4192 LASSERT(rc == LNET_MATCHMD_OK);
4194 lnet_build_msg_event(msg, LNET_EVENT_GET);
4196 reply_wmd = hdr->msg.get.return_wmd;
4198 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
4199 msg->msg_offset, msg->msg_wanted);
4201 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4204 /* The LND completes the REPLY from her recv procedure */
4205 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4206 msg->msg_offset, msg->msg_len, msg->msg_len);
4210 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4211 msg->msg_receiving = 0;
4213 rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
4215 /* didn't get as far as lnet_ni_send() */
4216 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4217 libcfs_nid2str(ni->ni_nid),
4218 libcfs_id2str(info.mi_id), rc);
4220 lnet_finalize(msg, rc);
4227 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4229 void *private = msg->msg_private;
4230 struct lnet_hdr *hdr = &msg->msg_hdr;
4231 struct lnet_process_id src = {0};
4232 struct lnet_libmd *md;
4233 unsigned int rlength;
4234 unsigned int mlength;
4237 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4240 src.nid = hdr->src_nid;
4241 src.pid = hdr->src_pid;
4243 /* NB handles only looked up by creator (no flips) */
4244 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4245 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4246 CNETERR("%s: Dropping REPLY from %s for %s "
4248 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4249 (md == NULL) ? "invalid" : "inactive",
4250 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4251 hdr->msg.reply.dst_wmd.wh_object_cookie);
4252 if (md != NULL && md->md_me != NULL)
4253 CERROR("REPLY MD also attached to portal %d\n",
4254 md->md_me->me_portal);
4256 lnet_res_unlock(cpt);
4257 return -ENOENT; /* -ve: OK but no match */
4260 LASSERT(md->md_offset == 0);
4262 rlength = hdr->payload_length;
4263 mlength = min(rlength, md->md_length);
4265 if (mlength < rlength &&
4266 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4267 CNETERR("%s: Dropping REPLY from %s length %d "
4268 "for MD %#llx would overflow (%d)\n",
4269 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4270 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4272 lnet_res_unlock(cpt);
4273 return -ENOENT; /* -ve: OK but no match */
4276 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4277 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4278 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4280 lnet_msg_attach_md(msg, md, 0, mlength);
4283 lnet_setpayloadbuffer(msg);
4285 lnet_res_unlock(cpt);
4287 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4289 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4294 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4296 struct lnet_hdr *hdr = &msg->msg_hdr;
4297 struct lnet_process_id src = {0};
4298 struct lnet_libmd *md;
4301 src.nid = hdr->src_nid;
4302 src.pid = hdr->src_pid;
4304 /* Convert ack fields to host byte order */
4305 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4306 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4308 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4311 /* NB handles only looked up by creator (no flips) */
4312 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4313 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4314 /* Don't moan; this is expected */
4316 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4317 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4318 (md == NULL) ? "invalid" : "inactive",
4319 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4320 hdr->msg.ack.dst_wmd.wh_object_cookie);
4321 if (md != NULL && md->md_me != NULL)
4322 CERROR("Source MD also attached to portal %d\n",
4323 md->md_me->me_portal);
4325 lnet_res_unlock(cpt);
4326 return -ENOENT; /* -ve! */
4329 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4330 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4331 hdr->msg.ack.dst_wmd.wh_object_cookie);
4333 lnet_msg_attach_md(msg, md, 0, 0);
4335 lnet_res_unlock(cpt);
4337 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4339 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4344 * \retval LNET_CREDIT_OK If \a msg is forwarded
4345 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4346 * \retval -ve error code
4349 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4353 if (!the_lnet.ln_routing)
4356 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4357 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4358 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4359 msg->msg_rx_ready_delay = 1;
4361 lnet_net_unlock(msg->msg_rx_cpt);
4362 rc = lnet_ni_eager_recv(ni, msg);
4363 lnet_net_lock(msg->msg_rx_cpt);
4368 rc = lnet_post_routed_recv_locked(msg, 0);
4373 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4377 switch (msg->msg_type) {
4379 rc = lnet_parse_ack(ni, msg);
4382 rc = lnet_parse_put(ni, msg);
4385 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4387 case LNET_MSG_REPLY:
4388 rc = lnet_parse_reply(ni, msg);
4390 default: /* prevent an unused label if !kernel */
4395 LASSERT(rc == 0 || rc == -ENOENT);
4400 lnet_msgtyp2str (int type)
4409 case LNET_MSG_REPLY:
4411 case LNET_MSG_HELLO:
4414 return ("<UNKNOWN>");
4419 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4420 void *private, int rdma_req)
4422 struct lnet_peer_ni *lpni;
4423 struct lnet_msg *msg;
4424 __u32 payload_length;
4425 lnet_pid_t dest_pid;
4426 lnet_nid_t dest_nid;
4434 LASSERT (!in_interrupt ());
4436 type = le32_to_cpu(hdr->type);
4437 src_nid = le64_to_cpu(hdr->src_nid);
4438 dest_nid = le64_to_cpu(hdr->dest_nid);
4439 dest_pid = le32_to_cpu(hdr->dest_pid);
4440 payload_length = le32_to_cpu(hdr->payload_length);
4442 for_me = (ni->ni_nid == dest_nid);
4443 cpt = lnet_cpt_of_nid(from_nid, ni);
4445 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4446 libcfs_nid2str(dest_nid),
4447 libcfs_nid2str(ni->ni_nid),
4448 libcfs_nid2str(src_nid),
4449 lnet_msgtyp2str(type),
4450 (for_me) ? "for me" : "routed");
4455 if (payload_length > 0) {
4456 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4457 libcfs_nid2str(from_nid),
4458 libcfs_nid2str(src_nid),
4459 lnet_msgtyp2str(type), payload_length);
4465 case LNET_MSG_REPLY:
4466 if (payload_length >
4467 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4468 CERROR("%s, src %s: bad %s payload %d "
4469 "(%d max expected)\n",
4470 libcfs_nid2str(from_nid),
4471 libcfs_nid2str(src_nid),
4472 lnet_msgtyp2str(type),
4474 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4480 CERROR("%s, src %s: Bad message type 0x%x\n",
4481 libcfs_nid2str(from_nid),
4482 libcfs_nid2str(src_nid), type);
4486 if (the_lnet.ln_routing &&
4487 ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4489 spin_lock(&ni->ni_net->net_lock);
4490 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4491 spin_unlock(&ni->ni_net->net_lock);
4492 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4497 lnet_push_update_to_peers(1);
4499 /* Regard a bad destination NID as a protocol error. Senders should
4500 * know what they're doing; if they don't they're misconfigured, buggy
4501 * or malicious so we chop them off at the knees :) */
4504 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4505 /* should have gone direct */
4506 CERROR("%s, src %s: Bad dest nid %s "
4507 "(should have been sent direct)\n",
4508 libcfs_nid2str(from_nid),
4509 libcfs_nid2str(src_nid),
4510 libcfs_nid2str(dest_nid));
4514 if (lnet_islocalnid(dest_nid)) {
4515 /* dest is another local NI; sender should have used
4516 * this node's NID on its own network */
4517 CERROR("%s, src %s: Bad dest nid %s "
4518 "(it's my nid but on a different network)\n",
4519 libcfs_nid2str(from_nid),
4520 libcfs_nid2str(src_nid),
4521 libcfs_nid2str(dest_nid));
4525 if (rdma_req && type == LNET_MSG_GET) {
4526 CERROR("%s, src %s: Bad optimized GET for %s "
4527 "(final destination must be me)\n",
4528 libcfs_nid2str(from_nid),
4529 libcfs_nid2str(src_nid),
4530 libcfs_nid2str(dest_nid));
4534 if (!the_lnet.ln_routing) {
4535 CERROR("%s, src %s: Dropping message for %s "
4536 "(routing not enabled)\n",
4537 libcfs_nid2str(from_nid),
4538 libcfs_nid2str(src_nid),
4539 libcfs_nid2str(dest_nid));
4544 /* Message looks OK; we're not going to return an error, so we MUST
4545 * call back lnd_recv() come what may... */
4547 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4548 fail_peer(src_nid, 0)) { /* shall we now? */
4549 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4550 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4551 lnet_msgtyp2str(type));
4555 if (!list_empty(&the_lnet.ln_drop_rules) &&
4556 lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4558 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4559 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4560 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4564 msg = lnet_msg_alloc();
4566 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4567 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4568 lnet_msgtyp2str(type));
4572 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4573 * pointers NULL etc */
4575 msg->msg_type = type;
4576 msg->msg_private = private;
4577 msg->msg_receiving = 1;
4578 msg->msg_rdma_get = rdma_req;
4579 msg->msg_len = msg->msg_wanted = payload_length;
4580 msg->msg_offset = 0;
4581 msg->msg_hdr = *hdr;
4582 /* for building message event */
4583 msg->msg_from = from_nid;
4585 msg->msg_target.pid = dest_pid;
4586 msg->msg_target.nid = dest_nid;
4587 msg->msg_routing = 1;
4590 /* convert common msg->hdr fields to host byteorder */
4591 msg->msg_hdr.type = type;
4592 msg->msg_hdr.src_nid = src_nid;
4593 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4594 msg->msg_hdr.dest_nid = dest_nid;
4595 msg->msg_hdr.dest_pid = dest_pid;
4596 msg->msg_hdr.payload_length = payload_length;
4600 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4602 lnet_net_unlock(cpt);
4603 CERROR("%s, src %s: Dropping %s "
4604 "(error %ld looking up sender)\n",
4605 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4606 lnet_msgtyp2str(type), PTR_ERR(lpni));
4608 if (rc == -ESHUTDOWN)
4609 /* We are shutting down. Don't do anything more */
4614 /* If this message was forwarded to us from a router then we may need
4615 * to update router aliveness or check for an asymmetrical route
4618 if (((lnet_drop_asym_route && for_me) ||
4619 !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4620 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4621 __u32 src_net_id = LNET_NIDNET(src_nid);
4622 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4623 struct lnet_route *route;
4626 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4627 if (route->lr_net == src_net_id) {
4629 /* If we're transitioning the gateway from
4630 * dead -> alive, and discovery is disabled
4631 * locally or on the gateway, then we need to
4632 * update the cached route aliveness for each
4633 * route to the src_nid's net.
4635 * Otherwise, we're only checking for
4636 * symmetrical route, and we can break the
4639 if (!gw->lp_alive &&
4640 lnet_is_discovery_disabled(gw))
4641 lnet_set_route_aliveness(route, true);
4646 if (lnet_drop_asym_route && for_me && !found) {
4647 lnet_net_unlock(cpt);
4648 /* we would not use from_nid to route a message to
4650 * => asymmetric routing detected but forbidden
4652 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4653 libcfs_nid2str(from_nid),
4654 libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4658 if (!gw->lp_alive) {
4659 struct lnet_peer_net *lpn;
4660 struct lnet_peer_ni *lpni2;
4662 gw->lp_alive = true;
4663 /* Mark all remote NIs on src_nid's net UP */
4664 lpn = lnet_peer_get_net_locked(gw, src_net_id);
4666 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4668 lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4672 lpni->lpni_last_alive = ktime_get_seconds();
4674 msg->msg_rxpeer = lpni;
4676 lnet_ni_addref_locked(ni, cpt);
4677 /* Multi-Rail: Primary NID of source. */
4678 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4681 * mark the status of this lpni as UP since we received a message
4682 * from it. The ping response reports back the ns_status which is
4683 * marked on the remote as up or down and we cache it here.
4685 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4687 lnet_msg_commit(msg, cpt);
4689 /* message delay simulation */
4690 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4691 lnet_delay_rule_match_locked(hdr, msg))) {
4692 lnet_net_unlock(cpt);
4697 rc = lnet_parse_forward_locked(ni, msg);
4698 lnet_net_unlock(cpt);
4703 if (rc == LNET_CREDIT_OK) {
4704 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4705 0, payload_length, payload_length);
4710 lnet_net_unlock(cpt);
4712 rc = lnet_parse_local(ni, msg);
4718 LASSERT(msg->msg_md == NULL);
4719 lnet_finalize(msg, rc);
4722 lnet_drop_message(ni, cpt, private, payload_length, type);
4725 EXPORT_SYMBOL(lnet_parse);
4728 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4730 while (!list_empty(head)) {
4731 struct lnet_process_id id = {0};
4732 struct lnet_msg *msg;
4734 msg = list_entry(head->next, struct lnet_msg, msg_list);
4735 list_del(&msg->msg_list);
4737 id.nid = msg->msg_hdr.src_nid;
4738 id.pid = msg->msg_hdr.src_pid;
4740 LASSERT(msg->msg_md == NULL);
4741 LASSERT(msg->msg_rx_delayed);
4742 LASSERT(msg->msg_rxpeer != NULL);
4743 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4745 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4746 " offset %d length %d: %s\n",
4748 msg->msg_hdr.msg.put.ptl_index,
4749 msg->msg_hdr.msg.put.match_bits,
4750 msg->msg_hdr.msg.put.offset,
4751 msg->msg_hdr.payload_length, reason);
4753 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4754 * called lnet_drop_message(), so I just hang onto msg as well
4755 * until that's done */
4757 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4758 msg->msg_private, msg->msg_len,
4761 msg->msg_no_resend = true;
4763 * NB: message will not generate event because w/o attached MD,
4764 * but we still should give error code so lnet_msg_decommit()
4765 * can skip counters operations and other checks.
4767 lnet_finalize(msg, -ENOENT);
4772 lnet_recv_delayed_msg_list(struct list_head *head)
4774 while (!list_empty(head)) {
4775 struct lnet_msg *msg;
4776 struct lnet_process_id id;
4778 msg = list_entry(head->next, struct lnet_msg, msg_list);
4779 list_del(&msg->msg_list);
4781 /* md won't disappear under me, since each msg
4782 * holds a ref on it */
4784 id.nid = msg->msg_hdr.src_nid;
4785 id.pid = msg->msg_hdr.src_pid;
4787 LASSERT(msg->msg_rx_delayed);
4788 LASSERT(msg->msg_md != NULL);
4789 LASSERT(msg->msg_rxpeer != NULL);
4790 LASSERT(msg->msg_rxni != NULL);
4791 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4793 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4794 "match %llu offset %d length %d.\n",
4795 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4796 msg->msg_hdr.msg.put.match_bits,
4797 msg->msg_hdr.msg.put.offset,
4798 msg->msg_hdr.payload_length);
4800 lnet_recv_put(msg->msg_rxni, msg);
4805 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4806 struct lnet_libmd *md, struct lnet_handle_md mdh)
4809 struct lnet_rsp_tracker *local_rspt;
4812 * MD has a refcount taken by message so it's not going away.
4813 * The MD however can be looked up. We need to secure the access
4814 * to the md_rspt_ptr by taking the res_lock.
4815 * The rspt can be accessed without protection up to when it gets
4816 * added to the list.
4820 local_rspt = md->md_rspt_ptr;
4821 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4822 if (local_rspt != NULL) {
4824 * we already have an rspt attached to the md, so we'll
4825 * update the deadline on that one.
4827 lnet_rspt_free(rspt, cpt);
4830 rspt->rspt_mdh = mdh;
4831 rspt->rspt_cpt = cpt;
4832 /* store the rspt so we can access it when we get the REPLY */
4833 md->md_rspt_ptr = rspt;
4836 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4839 * add to the list of tracked responses. It's added to tail of the
4840 * list in order to expire all the older entries first.
4843 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4844 lnet_net_unlock(cpt);
4845 lnet_res_unlock(cpt);
4849 * Initiate an asynchronous PUT operation.
4851 * There are several events associated with a PUT: completion of the send on
4852 * the initiator node (LNET_EVENT_SEND), and when the send completes
4853 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4854 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4855 * used at the target node to indicate the completion of incoming data
4858 * The local events will be logged in the EQ associated with the MD pointed to
4859 * by \a mdh handle. Using a MD without an associated EQ results in these
4860 * events being discarded. In this case, the caller must have another
4861 * mechanism (e.g., a higher level protocol) for determining when it is safe
4862 * to modify the memory region associated with the MD.
4864 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4865 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4867 * \param self Indicates the NID of a local interface through which to send
4868 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4869 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4870 * must be "free floating" (See LNetMDBind()).
4871 * \param ack Controls whether an acknowledgment is requested.
4872 * Acknowledgments are only sent when they are requested by the initiating
4873 * process and the target MD enables them.
4874 * \param target A process identifier for the target process.
4875 * \param portal The index in the \a target's portal table.
4876 * \param match_bits The match bits to use for MD selection at the target
4878 * \param offset The offset into the target MD (only used when the target
4879 * MD has the LNET_MD_MANAGE_REMOTE option set).
4880 * \param hdr_data 64 bits of user data that can be included in the message
4881 * header. This data is written to an event queue entry at the target if an
4882 * EQ is present on the matching MD.
4884 * \retval 0 Success, and only in this case events will be generated
4885 * and logged to EQ (if it exists).
4886 * \retval -EIO Simulated failure.
4887 * \retval -ENOMEM Memory allocation failure.
4888 * \retval -ENOENT Invalid MD object.
4890 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4893 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4894 struct lnet_process_id target, unsigned int portal,
4895 __u64 match_bits, unsigned int offset,
4898 struct lnet_msg *msg;
4899 struct lnet_libmd *md;
4902 struct lnet_rsp_tracker *rspt = NULL;
4904 LASSERT(the_lnet.ln_refcount > 0);
4906 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4907 fail_peer(target.nid, 1)) { /* shall we now? */
4908 CERROR("Dropping PUT to %s: simulated failure\n",
4909 libcfs_id2str(target));
4913 msg = lnet_msg_alloc();
4915 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4916 libcfs_id2str(target));
4919 msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
4921 cpt = lnet_cpt_of_cookie(mdh.cookie);
4923 if (ack == LNET_ACK_REQ) {
4924 rspt = lnet_rspt_alloc(cpt);
4926 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4927 libcfs_id2str(target));
4930 INIT_LIST_HEAD(&rspt->rspt_on_list);
4935 md = lnet_handle2md(&mdh);
4936 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4937 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4938 match_bits, portal, libcfs_id2str(target),
4939 md == NULL ? -1 : md->md_threshold);
4940 if (md != NULL && md->md_me != NULL)
4941 CERROR("Source MD also attached to portal %d\n",
4942 md->md_me->me_portal);
4943 lnet_res_unlock(cpt);
4946 lnet_rspt_free(rspt, cpt);
4952 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4954 lnet_msg_attach_md(msg, md, 0, 0);
4956 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4958 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4959 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4960 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4961 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4963 /* NB handles only looked up by creator (no flips) */
4964 if (ack == LNET_ACK_REQ) {
4965 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4966 the_lnet.ln_interface_cookie;
4967 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4968 md->md_lh.lh_cookie;
4970 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4971 LNET_WIRE_HANDLE_COOKIE_NONE;
4972 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4973 LNET_WIRE_HANDLE_COOKIE_NONE;
4976 lnet_res_unlock(cpt);
4978 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4980 if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
4982 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4984 lnet_rspt_free(rspt, cpt);
4986 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4990 rc = lnet_send(self, msg, LNET_NID_ANY);
4993 CNETERR("Error sending PUT to %s: %d\n",
4994 libcfs_id2str(target), rc);
4995 msg->msg_no_resend = true;
4996 lnet_finalize(msg, rc);
4999 /* completion will be signalled by an event */
5002 EXPORT_SYMBOL(LNetPut);
5005 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
5006 * returns a msg for the LND to pass to lnet_finalize() when the sink
5007 * data has been received.
5009 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5010 * lnet_finalize() is called on it, so the LND must call this first
5013 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5015 struct lnet_msg *msg = lnet_msg_alloc();
5016 struct lnet_libmd *getmd = getmsg->msg_md;
5017 struct lnet_process_id peer_id = getmsg->msg_target;
5020 LASSERT(!getmsg->msg_target_is_router);
5021 LASSERT(!getmsg->msg_routing);
5024 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5025 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
5029 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5032 LASSERT(getmd->md_refcount > 0);
5034 if (getmd->md_threshold == 0) {
5035 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5036 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
5038 lnet_res_unlock(cpt);
5042 LASSERT(getmd->md_offset == 0);
5044 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5045 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
5047 /* setup information for lnet_build_msg_event */
5048 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5049 msg->msg_from = peer_id.nid;
5050 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5051 msg->msg_hdr.src_nid = peer_id.nid;
5052 msg->msg_hdr.payload_length = getmd->md_length;
5053 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5055 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5056 lnet_res_unlock(cpt);
5058 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
5061 lnet_msg_commit(msg, cpt);
5062 lnet_net_unlock(cpt);
5064 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5069 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
5072 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5073 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5074 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5076 lnet_net_unlock(cpt);
5083 EXPORT_SYMBOL(lnet_create_reply_msg);
5086 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5089 /* Set the REPLY length, now the RDMA that elides the REPLY message has
5090 * completed and I know it. */
5091 LASSERT(reply != NULL);
5092 LASSERT(reply->msg_type == LNET_MSG_GET);
5093 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5095 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
5096 * the end of my buffer, I might as well be dead. */
5097 LASSERT(len <= reply->msg_ev.mlength);
5099 reply->msg_ev.mlength = len;
5101 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5104 * Initiate an asynchronous GET operation.
5106 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5107 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5108 * the target node in the REPLY has been written to local MD.
5110 * On the target node, an LNET_EVENT_GET is logged when the GET request
5111 * arrives and is accepted into a MD.
5113 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5114 * \param mdh A handle for the MD that describes the memory into which the
5115 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5117 * \retval 0 Success, and only in this case events will be generated
5118 * and logged to EQ (if it exists) of the MD.
5119 * \retval -EIO Simulated failure.
5120 * \retval -ENOMEM Memory allocation failure.
5121 * \retval -ENOENT Invalid MD object.
5124 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
5125 struct lnet_process_id target, unsigned int portal,
5126 __u64 match_bits, unsigned int offset, bool recovery)
5128 struct lnet_msg *msg;
5129 struct lnet_libmd *md;
5130 struct lnet_rsp_tracker *rspt;
5134 LASSERT(the_lnet.ln_refcount > 0);
5136 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
5137 fail_peer(target.nid, 1)) /* shall we now? */
5139 CERROR("Dropping GET to %s: simulated failure\n",
5140 libcfs_id2str(target));
5144 msg = lnet_msg_alloc();
5146 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5147 libcfs_id2str(target));
5151 cpt = lnet_cpt_of_cookie(mdh.cookie);
5153 rspt = lnet_rspt_alloc(cpt);
5155 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5156 libcfs_id2str(target));
5159 INIT_LIST_HEAD(&rspt->rspt_on_list);
5161 msg->msg_recovery = recovery;
5165 md = lnet_handle2md(&mdh);
5166 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5167 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5168 match_bits, portal, libcfs_id2str(target),
5169 md == NULL ? -1 : md->md_threshold);
5170 if (md != NULL && md->md_me != NULL)
5171 CERROR("REPLY MD also attached to portal %d\n",
5172 md->md_me->me_portal);
5174 lnet_res_unlock(cpt);
5177 lnet_rspt_free(rspt, cpt);
5181 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5183 lnet_msg_attach_md(msg, md, 0, 0);
5185 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5187 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5188 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5189 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5190 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5192 /* NB handles only looked up by creator (no flips) */
5193 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5194 the_lnet.ln_interface_cookie;
5195 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5196 md->md_lh.lh_cookie;
5198 lnet_res_unlock(cpt);
5200 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5202 if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5203 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5205 lnet_rspt_free(rspt, cpt);
5207 rc = lnet_send(self, msg, LNET_NID_ANY);
5209 CNETERR("Error sending GET to %s: %d\n",
5210 libcfs_id2str(target), rc);
5211 msg->msg_no_resend = true;
5212 lnet_finalize(msg, rc);
5215 /* completion will be signalled by an event */
5218 EXPORT_SYMBOL(LNetGet);
5221 * Calculate distance to node at \a dstnid.
5223 * \param dstnid Target NID.
5224 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5226 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5229 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5230 * local_nid_dist_zero is set, which is the default.
5231 * \retval positives Distance to target NID, i.e. number of hops plus one.
5232 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5235 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5237 struct list_head *e;
5238 struct lnet_ni *ni = NULL;
5239 struct lnet_remotenet *rnet;
5240 __u32 dstnet = LNET_NIDNET(dstnid);
5244 struct list_head *rn_list;
5246 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5247 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5248 * keep order 0 free for 0@lo and order 1 free for a local NID
5251 LASSERT(the_lnet.ln_refcount > 0);
5253 cpt = lnet_net_lock_current();
5255 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5256 if (ni->ni_nid == dstnid) {
5257 if (srcnidp != NULL)
5259 if (orderp != NULL) {
5260 if (dstnid == LNET_NID_LO_0)
5265 lnet_net_unlock(cpt);
5267 return local_nid_dist_zero ? 0 : 1;
5270 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5271 /* Check if ni was originally created in
5272 * current net namespace.
5273 * If not, assign order above 0xffff0000,
5274 * to make this ni not a priority. */
5275 if (current->nsproxy &&
5276 !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5277 order += 0xffff0000;
5278 if (srcnidp != NULL)
5279 *srcnidp = ni->ni_nid;
5282 lnet_net_unlock(cpt);
5289 rn_list = lnet_net2rnethash(dstnet);
5290 list_for_each(e, rn_list) {
5291 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5293 if (rnet->lrn_net == dstnet) {
5294 struct lnet_route *route;
5295 struct lnet_route *shortest = NULL;
5296 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5299 LASSERT(!list_empty(&rnet->lrn_routes));
5301 list_for_each_entry(route, &rnet->lrn_routes,
5303 route_hops = route->lr_hops;
5304 if (route_hops == LNET_UNDEFINED_HOPS)
5306 if (shortest == NULL ||
5307 route_hops < shortest_hops) {
5309 shortest_hops = route_hops;
5313 LASSERT(shortest != NULL);
5314 hops = shortest_hops;
5315 if (srcnidp != NULL) {
5316 struct lnet_net *net;
5317 net = lnet_get_net_locked(shortest->lr_lnet);
5319 ni = lnet_get_next_ni_locked(net, NULL);
5320 *srcnidp = ni->ni_nid;
5324 lnet_net_unlock(cpt);
5330 lnet_net_unlock(cpt);
5331 return -EHOSTUNREACH;
5333 EXPORT_SYMBOL(LNetDist);