4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/lnet/lib-move.c
34 * Data movement routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/pagemap.h>
41 #include <lnet/lib-lnet.h>
42 #include <linux/nsproxy.h>
43 #include <net/net_namespace.h>
45 static int local_nid_dist_zero = 1;
46 module_param(local_nid_dist_zero, int, 0444);
47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
49 struct lnet_send_data {
50 struct lnet_ni *sd_best_ni;
51 struct lnet_peer_ni *sd_best_lpni;
52 struct lnet_peer_ni *sd_final_dst_lpni;
53 struct lnet_peer *sd_peer;
54 struct lnet_peer *sd_gw_peer;
55 struct lnet_peer_ni *sd_gw_lpni;
56 struct lnet_peer_net *sd_peer_net;
57 struct lnet_msg *sd_msg;
58 lnet_nid_t sd_dst_nid;
59 lnet_nid_t sd_src_nid;
60 lnet_nid_t sd_rtr_nid;
67 lnet_msg_is_response(struct lnet_msg *msg)
69 return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
72 static inline struct lnet_comm_count *
73 get_stats_counts(struct lnet_element_stats *stats,
74 enum lnet_stats_type stats_type)
77 case LNET_STATS_TYPE_SEND:
78 return &stats->el_send_stats;
79 case LNET_STATS_TYPE_RECV:
80 return &stats->el_recv_stats;
81 case LNET_STATS_TYPE_DROP:
82 return &stats->el_drop_stats;
84 CERROR("Unknown stats type\n");
90 void lnet_incr_stats(struct lnet_element_stats *stats,
91 enum lnet_msg_type msg_type,
92 enum lnet_stats_type stats_type)
94 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
100 atomic_inc(&counts->co_ack_count);
103 atomic_inc(&counts->co_put_count);
106 atomic_inc(&counts->co_get_count);
109 atomic_inc(&counts->co_reply_count);
112 atomic_inc(&counts->co_hello_count);
115 CERROR("There is a BUG in the code. Unknown message type\n");
120 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
121 enum lnet_stats_type stats_type)
123 struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
127 return (atomic_read(&counts->co_ack_count) +
128 atomic_read(&counts->co_put_count) +
129 atomic_read(&counts->co_get_count) +
130 atomic_read(&counts->co_reply_count) +
131 atomic_read(&counts->co_hello_count));
134 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
135 struct lnet_comm_count *counts)
137 msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
138 msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
139 msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
140 msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
141 msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
144 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
145 struct lnet_element_stats *stats)
147 struct lnet_comm_count *counts;
152 counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
155 assign_stats(&msg_stats->im_send_stats, counts);
157 counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
160 assign_stats(&msg_stats->im_recv_stats, counts);
162 counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
165 assign_stats(&msg_stats->im_drop_stats, counts);
169 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
171 struct lnet_test_peer *tp;
172 struct list_head *el;
173 struct list_head *next;
176 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
177 if (threshold != 0) {
178 /* Adding a new entry */
179 LIBCFS_ALLOC(tp, sizeof(*tp));
184 tp->tp_threshold = threshold;
187 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
194 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
195 tp = list_entry(el, struct lnet_test_peer, tp_list);
197 if (tp->tp_threshold == 0 || /* needs culling anyway */
198 nid == LNET_NID_ANY || /* removing all entries */
199 tp->tp_nid == nid) { /* matched this one */
200 list_move(&tp->tp_list, &cull);
206 while (!list_empty(&cull)) {
207 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
209 list_del(&tp->tp_list);
210 LIBCFS_FREE(tp, sizeof(*tp));
216 fail_peer (lnet_nid_t nid, int outgoing)
218 struct lnet_test_peer *tp;
219 struct list_head *el;
220 struct list_head *next;
224 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
227 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
228 tp = list_entry(el, struct lnet_test_peer, tp_list);
230 if (tp->tp_threshold == 0) {
233 /* only cull zombies on outgoing tests,
234 * since we may be at interrupt priority on
235 * incoming messages. */
236 list_move(&tp->tp_list, &cull);
241 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
242 nid == tp->tp_nid) { /* fail this peer */
245 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
248 tp->tp_threshold == 0) {
250 list_move(&tp->tp_list, &cull);
259 while (!list_empty(&cull)) {
260 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
261 list_del(&tp->tp_list);
263 LIBCFS_FREE(tp, sizeof(*tp));
270 lnet_iov_nob(unsigned int niov, struct kvec *iov)
272 unsigned int nob = 0;
274 LASSERT(niov == 0 || iov != NULL);
276 nob += (iov++)->iov_len;
280 EXPORT_SYMBOL(lnet_iov_nob);
283 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
284 unsigned int nsiov, struct kvec *siov, unsigned int soffset,
287 /* NB diov, siov are READ-ONLY */
288 unsigned int this_nob;
293 /* skip complete frags before 'doffset' */
295 while (doffset >= diov->iov_len) {
296 doffset -= diov->iov_len;
302 /* skip complete frags before 'soffset' */
304 while (soffset >= siov->iov_len) {
305 soffset -= siov->iov_len;
314 this_nob = min3((unsigned int)diov->iov_len - doffset,
315 (unsigned int)siov->iov_len - soffset,
318 memcpy((char *)diov->iov_base + doffset,
319 (char *)siov->iov_base + soffset, this_nob);
322 if (diov->iov_len > doffset + this_nob) {
330 if (siov->iov_len > soffset + this_nob) {
339 EXPORT_SYMBOL(lnet_copy_iov2iov);
342 lnet_extract_iov(int dst_niov, struct kvec *dst,
343 int src_niov, struct kvec *src,
344 unsigned int offset, unsigned int len)
346 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
347 * for exactly 'len' bytes, and return the number of entries.
348 * NB not destructive to 'src' */
349 unsigned int frag_len;
352 if (len == 0) /* no data => */
353 return (0); /* no frags */
355 LASSERT(src_niov > 0);
356 while (offset >= src->iov_len) { /* skip initial frags */
357 offset -= src->iov_len;
360 LASSERT(src_niov > 0);
365 LASSERT(src_niov > 0);
366 LASSERT((int)niov <= dst_niov);
368 frag_len = src->iov_len - offset;
369 dst->iov_base = ((char *)src->iov_base) + offset;
371 if (len <= frag_len) {
376 dst->iov_len = frag_len;
386 EXPORT_SYMBOL(lnet_extract_iov);
390 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
392 unsigned int nob = 0;
394 LASSERT(niov == 0 || kiov != NULL);
396 nob += (kiov++)->bv_len;
400 EXPORT_SYMBOL(lnet_kiov_nob);
403 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
404 unsigned int doffset,
405 unsigned int nsiov, struct bio_vec *siov,
406 unsigned int soffset,
409 /* NB diov, siov are READ-ONLY */
410 unsigned int this_nob;
417 LASSERT (!in_interrupt ());
420 while (doffset >= diov->bv_len) {
421 doffset -= diov->bv_len;
428 while (soffset >= siov->bv_len) {
429 soffset -= siov->bv_len;
438 this_nob = min3(diov->bv_len - doffset,
439 siov->bv_len - soffset,
443 daddr = ((char *)kmap(diov->bv_page)) +
444 diov->bv_offset + doffset;
446 saddr = ((char *)kmap(siov->bv_page)) +
447 siov->bv_offset + soffset;
449 /* Vanishing risk of kmap deadlock when mapping 2 pages.
450 * However in practice at least one of the kiovs will be mapped
451 * kernel pages and the map/unmap will be NOOPs */
453 memcpy (daddr, saddr, this_nob);
456 if (diov->bv_len > doffset + this_nob) {
460 kunmap(diov->bv_page);
467 if (siov->bv_len > soffset + this_nob) {
471 kunmap(siov->bv_page);
480 kunmap(diov->bv_page);
482 kunmap(siov->bv_page);
484 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
487 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
488 unsigned int nkiov, struct bio_vec *kiov,
489 unsigned int kiovoffset,
492 /* NB iov, kiov are READ-ONLY */
493 unsigned int this_nob;
499 LASSERT (!in_interrupt ());
502 while (iovoffset >= iov->iov_len) {
503 iovoffset -= iov->iov_len;
510 while (kiovoffset >= kiov->bv_len) {
511 kiovoffset -= kiov->bv_len;
520 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
521 (unsigned int)kiov->bv_len - kiovoffset,
525 addr = ((char *)kmap(kiov->bv_page)) +
526 kiov->bv_offset + kiovoffset;
528 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
531 if (iov->iov_len > iovoffset + this_nob) {
532 iovoffset += this_nob;
539 if (kiov->bv_len > kiovoffset + this_nob) {
541 kiovoffset += this_nob;
543 kunmap(kiov->bv_page);
553 kunmap(kiov->bv_page);
555 EXPORT_SYMBOL(lnet_copy_kiov2iov);
558 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
559 unsigned int kiovoffset,
560 unsigned int niov, struct kvec *iov, unsigned int iovoffset,
563 /* NB kiov, iov are READ-ONLY */
564 unsigned int this_nob;
570 LASSERT (!in_interrupt ());
573 while (kiovoffset >= kiov->bv_len) {
574 kiovoffset -= kiov->bv_len;
581 while (iovoffset >= iov->iov_len) {
582 iovoffset -= iov->iov_len;
591 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
592 (unsigned int)iov->iov_len - iovoffset,
596 addr = ((char *)kmap(kiov->bv_page)) +
597 kiov->bv_offset + kiovoffset;
599 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
602 if (kiov->bv_len > kiovoffset + this_nob) {
604 kiovoffset += this_nob;
606 kunmap(kiov->bv_page);
613 if (iov->iov_len > iovoffset + this_nob) {
614 iovoffset += this_nob;
623 kunmap(kiov->bv_page);
625 EXPORT_SYMBOL(lnet_copy_iov2kiov);
628 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
629 int src_niov, struct bio_vec *src,
630 unsigned int offset, unsigned int len)
632 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
633 * for exactly 'len' bytes, and return the number of entries.
634 * NB not destructive to 'src' */
635 unsigned int frag_len;
638 if (len == 0) /* no data => */
639 return (0); /* no frags */
641 LASSERT(src_niov > 0);
642 while (offset >= src->bv_len) { /* skip initial frags */
643 offset -= src->bv_len;
646 LASSERT(src_niov > 0);
651 LASSERT(src_niov > 0);
652 LASSERT((int)niov <= dst_niov);
654 frag_len = src->bv_len - offset;
655 dst->bv_page = src->bv_page;
656 dst->bv_offset = src->bv_offset + offset;
658 if (len <= frag_len) {
660 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
664 dst->bv_len = frag_len;
665 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
675 EXPORT_SYMBOL(lnet_extract_kiov);
678 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
679 int delayed, unsigned int offset, unsigned int mlen,
682 unsigned int niov = 0;
683 struct kvec *iov = NULL;
684 struct bio_vec *kiov = NULL;
687 LASSERT (!in_interrupt ());
688 LASSERT (mlen == 0 || msg != NULL);
691 LASSERT(msg->msg_receiving);
692 LASSERT(!msg->msg_sending);
693 LASSERT(rlen == msg->msg_len);
694 LASSERT(mlen <= msg->msg_len);
695 LASSERT(msg->msg_offset == offset);
696 LASSERT(msg->msg_wanted == mlen);
698 msg->msg_receiving = 0;
701 niov = msg->msg_niov;
702 kiov = msg->msg_kiov;
705 LASSERT ((iov == NULL) != (kiov == NULL));
709 rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
710 niov, kiov, offset, mlen,
713 lnet_finalize(msg, rc);
717 lnet_setpayloadbuffer(struct lnet_msg *msg)
719 struct lnet_libmd *md = msg->msg_md;
721 LASSERT(msg->msg_len > 0);
722 LASSERT(!msg->msg_routing);
724 LASSERT(msg->msg_niov == 0);
725 LASSERT(msg->msg_kiov == NULL);
727 msg->msg_niov = md->md_niov;
728 msg->msg_kiov = md->md_kiov;
732 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
733 unsigned int offset, unsigned int len)
735 msg->msg_type = type;
736 msg->msg_target = target;
738 msg->msg_offset = offset;
741 lnet_setpayloadbuffer(msg);
743 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
744 msg->msg_hdr.type = cpu_to_le32(type);
745 /* dest_nid will be overwritten by lnet_select_pathway() */
746 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
747 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
748 /* src_nid will be set later */
749 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
750 msg->msg_hdr.payload_length = cpu_to_le32(len);
754 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
756 void *priv = msg->msg_private;
759 LASSERT(!in_interrupt());
760 LASSERT(ni->ni_nid == LNET_NID_LO_0 ||
761 (msg->msg_txcredit && msg->msg_peertxcredit));
763 rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
765 msg->msg_no_resend = true;
766 lnet_finalize(msg, rc);
771 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
775 LASSERT(!msg->msg_sending);
776 LASSERT(msg->msg_receiving);
777 LASSERT(!msg->msg_rx_ready_delay);
778 LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
780 msg->msg_rx_ready_delay = 1;
781 rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
784 CERROR("recv from %s / send to %s aborted: "
785 "eager_recv failed %d\n",
786 libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
787 libcfs_id2str(msg->msg_target), rc);
788 LASSERT(rc < 0); /* required by my callers */
795 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
799 deadline = lpni->lpni_last_alive +
800 lpni->lpni_net->net_tunables.lct_peer_timeout;
803 * assume peer_ni is alive as long as we're within the configured
812 /* NB: returns 1 when alive, 0 when dead, negative when error;
813 * may drop the lnet_net_lock */
815 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
816 struct lnet_msg *msg)
818 time64_t now = ktime_get_seconds();
820 if (!lnet_peer_aliveness_enabled(lpni))
824 * If we're resending a message, let's attempt to send it even if
825 * the peer is down to fulfill our resend quota on the message
827 if (msg->msg_retry_count > 0)
830 /* try and send recovery messages irregardless */
831 if (msg->msg_recovery)
834 /* always send any responses */
835 if (lnet_msg_is_response(msg))
838 if (!lnet_is_peer_deadline_passed(lpni, now))
841 return lnet_is_peer_ni_alive(lpni);
845 * \param msg The message to be sent.
846 * \param do_send True if lnet_ni_send() should be called in this function.
847 * lnet_send() is going to lnet_net_unlock immediately after this, so
848 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
850 * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
851 * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
852 * \retval -EHOSTUNREACH If the next hop of the message appears dead.
853 * \retval -ECANCELED If the MD of the message has been unlinked.
856 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
858 struct lnet_peer_ni *lp = msg->msg_txpeer;
859 struct lnet_ni *ni = msg->msg_txni;
860 int cpt = msg->msg_tx_cpt;
861 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
863 /* non-lnet_send() callers have checked before */
864 LASSERT(!do_send || msg->msg_tx_delayed);
865 LASSERT(!msg->msg_receiving);
866 LASSERT(msg->msg_tx_committed);
867 /* can't get here if we're sending to the loopback interface */
868 LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
870 /* NB 'lp' is always the next hop */
871 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
872 lnet_peer_alive_locked(ni, lp, msg) == 0) {
873 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
874 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
876 lnet_net_unlock(cpt);
878 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
880 LNET_STATS_TYPE_DROP);
882 lnet_incr_stats(&msg->msg_txni->ni_stats,
884 LNET_STATS_TYPE_DROP);
886 CNETERR("Dropping message for %s: peer not alive\n",
887 libcfs_id2str(msg->msg_target));
888 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
890 lnet_finalize(msg, -EHOSTUNREACH);
893 return -EHOSTUNREACH;
896 if (msg->msg_md != NULL &&
897 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
898 lnet_net_unlock(cpt);
900 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
901 "called on the MD/ME.\n",
902 libcfs_id2str(msg->msg_target));
904 msg->msg_no_resend = true;
905 CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
906 msg, libcfs_id2str(msg->msg_target));
907 lnet_finalize(msg, -ECANCELED);
914 if (!msg->msg_peertxcredit) {
915 spin_lock(&lp->lpni_lock);
916 LASSERT((lp->lpni_txcredits < 0) ==
917 !list_empty(&lp->lpni_txq));
919 msg->msg_peertxcredit = 1;
920 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
921 lp->lpni_txcredits--;
923 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
924 lp->lpni_mintxcredits = lp->lpni_txcredits;
926 if (lp->lpni_txcredits < 0) {
927 msg->msg_tx_delayed = 1;
928 list_add_tail(&msg->msg_list, &lp->lpni_txq);
929 spin_unlock(&lp->lpni_lock);
930 return LNET_CREDIT_WAIT;
932 spin_unlock(&lp->lpni_lock);
935 if (!msg->msg_txcredit) {
936 LASSERT((tq->tq_credits < 0) ==
937 !list_empty(&tq->tq_delayed));
939 msg->msg_txcredit = 1;
941 atomic_dec(&ni->ni_tx_credits);
943 if (tq->tq_credits < tq->tq_credits_min)
944 tq->tq_credits_min = tq->tq_credits;
946 if (tq->tq_credits < 0) {
947 msg->msg_tx_delayed = 1;
948 list_add_tail(&msg->msg_list, &tq->tq_delayed);
949 return LNET_CREDIT_WAIT;
953 /* unset the tx_delay flag as we're going to send it now */
954 msg->msg_tx_delayed = 0;
957 lnet_net_unlock(cpt);
958 lnet_ni_send(ni, msg);
961 return LNET_CREDIT_OK;
965 static struct lnet_rtrbufpool *
966 lnet_msg2bufpool(struct lnet_msg *msg)
968 struct lnet_rtrbufpool *rbp;
971 LASSERT(msg->msg_rx_committed);
973 cpt = msg->msg_rx_cpt;
974 rbp = &the_lnet.ln_rtrpools[cpt][0];
976 LASSERT(msg->msg_len <= LNET_MTU);
977 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
979 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
986 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
988 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
989 * sets do_recv FALSE and I don't do the unlock/send/lock bit.
990 * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
991 * received or OK to receive */
992 struct lnet_peer_ni *lpni = msg->msg_rxpeer;
993 struct lnet_peer *lp;
994 struct lnet_rtrbufpool *rbp;
995 struct lnet_rtrbuf *rb;
997 LASSERT(msg->msg_kiov == NULL);
998 LASSERT(msg->msg_niov == 0);
999 LASSERT(msg->msg_routing);
1000 LASSERT(msg->msg_receiving);
1001 LASSERT(!msg->msg_sending);
1002 LASSERT(lpni->lpni_peer_net);
1003 LASSERT(lpni->lpni_peer_net->lpn_peer);
1005 lp = lpni->lpni_peer_net->lpn_peer;
1007 /* non-lnet_parse callers only receive delayed messages */
1008 LASSERT(!do_recv || msg->msg_rx_delayed);
1010 if (!msg->msg_peerrtrcredit) {
1011 /* lpni_lock protects the credit manipulation */
1012 spin_lock(&lpni->lpni_lock);
1014 msg->msg_peerrtrcredit = 1;
1015 lpni->lpni_rtrcredits--;
1016 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1017 lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1019 if (lpni->lpni_rtrcredits < 0) {
1020 spin_unlock(&lpni->lpni_lock);
1021 /* must have checked eager_recv before here */
1022 LASSERT(msg->msg_rx_ready_delay);
1023 msg->msg_rx_delayed = 1;
1024 /* lp_lock protects the lp_rtrq */
1025 spin_lock(&lp->lp_lock);
1026 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1027 spin_unlock(&lp->lp_lock);
1028 return LNET_CREDIT_WAIT;
1030 spin_unlock(&lpni->lpni_lock);
1033 rbp = lnet_msg2bufpool(msg);
1035 if (!msg->msg_rtrcredit) {
1036 msg->msg_rtrcredit = 1;
1038 if (rbp->rbp_credits < rbp->rbp_mincredits)
1039 rbp->rbp_mincredits = rbp->rbp_credits;
1041 if (rbp->rbp_credits < 0) {
1042 /* must have checked eager_recv before here */
1043 LASSERT(msg->msg_rx_ready_delay);
1044 msg->msg_rx_delayed = 1;
1045 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1046 return LNET_CREDIT_WAIT;
1050 LASSERT(!list_empty(&rbp->rbp_bufs));
1051 rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1052 list_del(&rb->rb_list);
1054 msg->msg_niov = rbp->rbp_npages;
1055 msg->msg_kiov = &rb->rb_kiov[0];
1057 /* unset the msg-rx_delayed flag since we're receiving the message */
1058 msg->msg_rx_delayed = 0;
1061 int cpt = msg->msg_rx_cpt;
1063 lnet_net_unlock(cpt);
1064 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1065 0, msg->msg_len, msg->msg_len);
1068 return LNET_CREDIT_OK;
1072 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1074 struct lnet_peer_ni *txpeer = msg->msg_txpeer;
1075 struct lnet_ni *txni = msg->msg_txni;
1076 struct lnet_msg *msg2;
1078 if (msg->msg_txcredit) {
1079 struct lnet_ni *ni = msg->msg_txni;
1080 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1082 /* give back NI txcredits */
1083 msg->msg_txcredit = 0;
1085 LASSERT((tq->tq_credits < 0) ==
1086 !list_empty(&tq->tq_delayed));
1089 atomic_inc(&ni->ni_tx_credits);
1090 if (tq->tq_credits <= 0) {
1091 msg2 = list_entry(tq->tq_delayed.next,
1092 struct lnet_msg, msg_list);
1093 list_del(&msg2->msg_list);
1095 LASSERT(msg2->msg_txni == ni);
1096 LASSERT(msg2->msg_tx_delayed);
1097 LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1099 (void) lnet_post_send_locked(msg2, 1);
1103 if (msg->msg_peertxcredit) {
1104 /* give back peer txcredits */
1105 msg->msg_peertxcredit = 0;
1107 spin_lock(&txpeer->lpni_lock);
1108 LASSERT((txpeer->lpni_txcredits < 0) ==
1109 !list_empty(&txpeer->lpni_txq));
1111 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1112 LASSERT(txpeer->lpni_txqnob >= 0);
1114 txpeer->lpni_txcredits++;
1115 if (txpeer->lpni_txcredits <= 0) {
1118 msg2 = list_entry(txpeer->lpni_txq.next,
1119 struct lnet_msg, msg_list);
1120 list_del(&msg2->msg_list);
1121 spin_unlock(&txpeer->lpni_lock);
1123 LASSERT(msg2->msg_txpeer == txpeer);
1124 LASSERT(msg2->msg_tx_delayed);
1126 msg2_cpt = msg2->msg_tx_cpt;
1129 * The msg_cpt can be different from the msg2_cpt
1130 * so we need to make sure we lock the correct cpt
1132 * Once we call lnet_post_send_locked() it is no
1133 * longer safe to access msg2, since it could've
1134 * been freed by lnet_finalize(), but we still
1135 * need to relock the correct cpt, so we cache the
1136 * msg2_cpt for the purpose of the check that
1137 * follows the call to lnet_pose_send_locked().
1139 if (msg2_cpt != msg->msg_tx_cpt) {
1140 lnet_net_unlock(msg->msg_tx_cpt);
1141 lnet_net_lock(msg2_cpt);
1143 (void) lnet_post_send_locked(msg2, 1);
1144 if (msg2_cpt != msg->msg_tx_cpt) {
1145 lnet_net_unlock(msg2_cpt);
1146 lnet_net_lock(msg->msg_tx_cpt);
1149 spin_unlock(&txpeer->lpni_lock);
1154 msg->msg_txni = NULL;
1155 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1158 if (txpeer != NULL) {
1159 msg->msg_txpeer = NULL;
1160 lnet_peer_ni_decref_locked(txpeer);
1165 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1167 struct lnet_msg *msg;
1169 if (list_empty(&rbp->rbp_msgs))
1171 msg = list_entry(rbp->rbp_msgs.next,
1172 struct lnet_msg, msg_list);
1173 list_del(&msg->msg_list);
1175 (void)lnet_post_routed_recv_locked(msg, 1);
1179 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1181 struct lnet_msg *msg;
1182 struct lnet_msg *tmp;
1184 lnet_net_unlock(cpt);
1186 list_for_each_entry_safe(msg, tmp, list, msg_list) {
1187 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1188 0, 0, 0, msg->msg_hdr.payload_length);
1189 list_del_init(&msg->msg_list);
1190 msg->msg_no_resend = true;
1191 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1192 lnet_finalize(msg, -ECANCELED);
1199 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1201 struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1202 struct lnet_peer *lp;
1203 struct lnet_ni *rxni = msg->msg_rxni;
1204 struct lnet_msg *msg2;
1206 if (msg->msg_rtrcredit) {
1207 /* give back global router credits */
1208 struct lnet_rtrbuf *rb;
1209 struct lnet_rtrbufpool *rbp;
1211 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1212 * there until it gets one allocated, or aborts the wait
1214 LASSERT(msg->msg_kiov != NULL);
1216 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1219 msg->msg_kiov = NULL;
1220 msg->msg_rtrcredit = 0;
1222 LASSERT(rbp == lnet_msg2bufpool(msg));
1224 LASSERT((rbp->rbp_credits > 0) ==
1225 !list_empty(&rbp->rbp_bufs));
1227 /* If routing is now turned off, we just drop this buffer and
1228 * don't bother trying to return credits. */
1229 if (!the_lnet.ln_routing) {
1230 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1234 /* It is possible that a user has lowered the desired number of
1235 * buffers in this pool. Make sure we never put back
1236 * more buffers than the stated number. */
1237 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1238 /* Discard this buffer so we don't have too
1240 lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1241 rbp->rbp_nbuffers--;
1243 list_add(&rb->rb_list, &rbp->rbp_bufs);
1245 if (rbp->rbp_credits <= 0)
1246 lnet_schedule_blocked_locked(rbp);
1251 if (msg->msg_peerrtrcredit) {
1253 LASSERT(rxpeerni->lpni_peer_net);
1254 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1256 /* give back peer router credits */
1257 msg->msg_peerrtrcredit = 0;
1259 spin_lock(&rxpeerni->lpni_lock);
1260 rxpeerni->lpni_rtrcredits++;
1261 spin_unlock(&rxpeerni->lpni_lock);
1263 lp = rxpeerni->lpni_peer_net->lpn_peer;
1264 spin_lock(&lp->lp_lock);
1266 /* drop all messages which are queued to be routed on that
1268 if (!the_lnet.ln_routing) {
1270 list_splice_init(&lp->lp_rtrq, &drop);
1271 spin_unlock(&lp->lp_lock);
1272 lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1273 } else if (!list_empty(&lp->lp_rtrq)) {
1276 msg2 = list_entry(lp->lp_rtrq.next,
1277 struct lnet_msg, msg_list);
1278 list_del(&msg2->msg_list);
1279 msg2_cpt = msg2->msg_rx_cpt;
1280 spin_unlock(&lp->lp_lock);
1282 * messages on the lp_rtrq can be from any NID in
1283 * the peer, which means they might have different
1284 * cpts. We need to make sure we lock the right
1287 if (msg2_cpt != msg->msg_rx_cpt) {
1288 lnet_net_unlock(msg->msg_rx_cpt);
1289 lnet_net_lock(msg2_cpt);
1291 (void) lnet_post_routed_recv_locked(msg2, 1);
1292 if (msg2_cpt != msg->msg_rx_cpt) {
1293 lnet_net_unlock(msg2_cpt);
1294 lnet_net_lock(msg->msg_rx_cpt);
1297 spin_unlock(&lp->lp_lock);
1301 msg->msg_rxni = NULL;
1302 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1304 if (rxpeerni != NULL) {
1305 msg->msg_rxpeer = NULL;
1306 lnet_peer_ni_decref_locked(rxpeerni);
1311 lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1313 if (p1->lpni_txqnob < p2->lpni_txqnob)
1316 if (p1->lpni_txqnob > p2->lpni_txqnob)
1319 if (p1->lpni_txcredits > p2->lpni_txcredits)
1322 if (p1->lpni_txcredits < p2->lpni_txcredits)
1328 static struct lnet_peer_ni *
1329 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1330 struct lnet_peer *peer,
1331 struct lnet_peer_net *peer_net)
1334 * Look at the peer NIs for the destination peer that connect
1335 * to the chosen net. If a peer_ni is preferred when using the
1336 * best_ni to communicate, we use that one. If there is no
1337 * preferred peer_ni, or there are multiple preferred peer_ni,
1338 * the available transmit credits are used. If the transmit
1339 * credits are equal, we round-robin over the peer_ni.
1341 struct lnet_peer_ni *lpni = NULL;
1342 struct lnet_peer_ni *best_lpni = NULL;
1343 int best_lpni_credits = INT_MIN;
1344 bool preferred = false;
1346 int best_lpni_healthv = 0;
1349 while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1351 * if the best_ni we've chosen aleady has this lpni
1352 * preferred, then let's use it
1355 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1357 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1358 libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1363 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1366 CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1367 libcfs_nid2str(lpni->lpni_nid),
1368 lpni->lpni_txcredits, best_lpni_credits,
1369 lpni->lpni_seq, best_lpni->lpni_seq);
1371 /* pick the healthiest peer ni */
1372 if (lpni_healthv < best_lpni_healthv) {
1374 } else if (lpni_healthv > best_lpni_healthv) {
1375 best_lpni_healthv = lpni_healthv;
1376 /* if this is a preferred peer use it */
1377 } else if (!preferred && ni_is_pref) {
1379 } else if (preferred && !ni_is_pref) {
1381 * this is not the preferred peer so let's ignore
1385 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1387 * We already have a peer that has more credits
1388 * available than this one. No need to consider
1389 * this peer further.
1392 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1394 * The best peer found so far and the current peer
1395 * have the same number of available credits let's
1396 * make sure to select between them using Round
1400 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1406 best_lpni_credits = lpni->lpni_txcredits;
1409 /* if we still can't find a peer ni then we can't reach it */
1411 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1412 LNET_NIDNET(dst_nid);
1413 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1414 libcfs_net2str(net_id));
1418 CDEBUG(D_NET, "sd_best_lpni = %s\n",
1419 libcfs_nid2str(best_lpni->lpni_nid));
1425 * Prerequisite: the best_ni should already be set in the sd
1427 static inline struct lnet_peer_ni *
1428 lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid,
1429 struct lnet_peer *peer, __u32 net_id)
1431 struct lnet_peer_net *peer_net;
1434 * The gateway is Multi-Rail capable so now we must select the
1437 peer_net = lnet_peer_get_net_locked(peer, net_id);
1439 return lnet_select_peer_ni(lni, dst_nid, peer, peer_net);
1442 /* Compare route priorities and hop counts */
1444 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1446 int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1447 int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1449 if (r1->lr_priority < r2->lr_priority)
1452 if (r1->lr_priority > r2->lr_priority)
1455 if (r1_hops < r2_hops)
1458 if (r1_hops > r2_hops)
1464 static struct lnet_route *
1465 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1466 struct lnet_route **prev_route,
1467 struct lnet_peer_ni **gwni)
1469 struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1470 struct lnet_route *best_route;
1471 struct lnet_route *last_route;
1472 struct lnet_route *route;
1475 best_route = last_route = NULL;
1476 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1477 if (!lnet_is_route_alive(route))
1481 * Restrict the selection of the router NI on the src_net
1482 * provided. If the src_net is LNET_NID_ANY, then select
1483 * the best interface available.
1486 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1490 best_route = last_route = route;
1493 CERROR("Gateway %s does not have a peer NI on net %s\n",
1494 libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1495 libcfs_net2str(src_net));
1500 /* no protection on below fields, but it's harmless */
1501 if (last_route->lr_seq - route->lr_seq < 0)
1504 rc = lnet_compare_routes(route, best_route);
1508 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1512 CERROR("Gateway %s does not have a peer NI on net %s\n",
1513 libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1514 libcfs_net2str(src_net));
1524 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1528 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1535 *prev_route = last_route;
1541 static struct lnet_ni *
1542 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1543 struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1546 struct lnet_ni *ni = NULL;
1547 unsigned int shortest_distance;
1552 * If there is no peer_ni that we can send to on this network,
1553 * then there is no point in looking for a new best_ni here.
1555 if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1558 if (best_ni == NULL) {
1559 shortest_distance = UINT_MAX;
1560 best_credits = INT_MIN;
1563 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1564 best_ni->ni_dev_cpt);
1565 best_credits = atomic_read(&best_ni->ni_tx_credits);
1566 best_healthv = atomic_read(&best_ni->ni_healthv);
1569 while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1570 unsigned int distance;
1575 ni_credits = atomic_read(&ni->ni_tx_credits);
1576 ni_healthv = atomic_read(&ni->ni_healthv);
1577 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1580 * calculate the distance from the CPT on which
1581 * the message memory is allocated to the CPT of
1582 * the NI's physical device
1584 distance = cfs_cpt_distance(lnet_cpt_table(),
1588 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1589 libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1590 ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1591 : "not seleced", best_credits, shortest_distance,
1592 (best_ni) ? best_ni->ni_seq : 0);
1595 * All distances smaller than the NUMA range
1596 * are treated equally.
1598 if (distance < lnet_numa_range)
1599 distance = lnet_numa_range;
1602 * Select on health, shorter distance, available
1603 * credits, then round-robin.
1607 } else if (ni_healthv < best_healthv) {
1609 } else if (ni_healthv > best_healthv) {
1610 best_healthv = ni_healthv;
1612 * If we're going to prefer this ni because it's
1613 * the healthiest, then we should set the
1614 * shortest_distance in the algorithm in case
1615 * there are multiple NIs with the same health but
1616 * different distances.
1618 if (distance < shortest_distance)
1619 shortest_distance = distance;
1620 } else if (distance > shortest_distance) {
1622 } else if (distance < shortest_distance) {
1623 shortest_distance = distance;
1624 } else if (ni_credits < best_credits) {
1626 } else if (ni_credits == best_credits) {
1627 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1631 best_credits = ni_credits;
1634 CDEBUG(D_NET, "selected best_ni %s\n",
1635 (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1641 * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1642 * because such traffic is required to perform discovery. We therefore
1643 * exclude all GET and PUT on that portal. We also exclude all ACK and
1644 * REPLY traffic, but that is because the portal is not tracked in the
1645 * message structure for these message types. We could restrict this
1646 * further by also checking for LNET_PROTO_PING_MATCHBITS.
1649 lnet_msg_discovery(struct lnet_msg *msg)
1651 if (msg->msg_type == LNET_MSG_PUT) {
1652 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1654 } else if (msg->msg_type == LNET_MSG_GET) {
1655 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1661 #define SRC_SPEC 0x0001
1662 #define SRC_ANY 0x0002
1663 #define LOCAL_DST 0x0004
1664 #define REMOTE_DST 0x0008
1665 #define MR_DST 0x0010
1666 #define NMR_DST 0x0020
1667 #define SND_RESP 0x0040
1669 /* The following to defines are used for return codes */
1670 #define REPEAT_SEND 0x1000
1671 #define PASS_THROUGH 0x2000
1673 /* The different cases lnet_select pathway needs to handle */
1674 #define SRC_SPEC_LOCAL_MR_DST (SRC_SPEC | LOCAL_DST | MR_DST)
1675 #define SRC_SPEC_ROUTER_MR_DST (SRC_SPEC | REMOTE_DST | MR_DST)
1676 #define SRC_SPEC_LOCAL_NMR_DST (SRC_SPEC | LOCAL_DST | NMR_DST)
1677 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1678 #define SRC_ANY_LOCAL_MR_DST (SRC_ANY | LOCAL_DST | MR_DST)
1679 #define SRC_ANY_ROUTER_MR_DST (SRC_ANY | REMOTE_DST | MR_DST)
1680 #define SRC_ANY_LOCAL_NMR_DST (SRC_ANY | LOCAL_DST | NMR_DST)
1681 #define SRC_ANY_ROUTER_NMR_DST (SRC_ANY | REMOTE_DST | NMR_DST)
1684 lnet_handle_lo_send(struct lnet_send_data *sd)
1686 struct lnet_msg *msg = sd->sd_msg;
1687 int cpt = sd->sd_cpt;
1689 /* No send credit hassles with LOLND */
1690 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1691 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1692 if (!msg->msg_routing)
1693 msg->msg_hdr.src_nid =
1694 cpu_to_le64(the_lnet.ln_loni->ni_nid);
1695 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1696 lnet_msg_commit(msg, cpt);
1697 msg->msg_txni = the_lnet.ln_loni;
1699 return LNET_CREDIT_OK;
1703 lnet_handle_send(struct lnet_send_data *sd)
1705 struct lnet_ni *best_ni = sd->sd_best_ni;
1706 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1707 struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1708 struct lnet_msg *msg = sd->sd_msg;
1710 __u32 send_case = sd->sd_send_case;
1712 __u32 routing = send_case & REMOTE_DST;
1713 struct lnet_rsp_tracker *rspt;
1716 * Increment sequence number of the selected peer so that we
1717 * pick the next one in Round Robin.
1719 best_lpni->lpni_seq++;
1722 * grab a reference on the peer_ni so it sticks around even if
1723 * we need to drop and relock the lnet_net_lock below.
1725 lnet_peer_ni_addref_locked(best_lpni);
1728 * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1729 * message. This ensures that we get a CPT that is correct for
1730 * the NI when the NI has been restricted to a subset of all CPTs.
1731 * If the selected CPT differs from the one currently locked, we
1732 * must unlock and relock the lnet_net_lock(), and then check whether
1733 * the configuration has changed. We don't have a hold on the best_ni
1734 * yet, and it may have vanished.
1736 cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1737 if (sd->sd_cpt != cpt2) {
1738 __u32 seq = lnet_get_dlc_seq_locked();
1739 lnet_net_unlock(sd->sd_cpt);
1741 lnet_net_lock(sd->sd_cpt);
1742 if (seq != lnet_get_dlc_seq_locked()) {
1743 lnet_peer_ni_decref_locked(best_lpni);
1749 * store the best_lpni in the message right away to avoid having
1750 * to do the same operation under different conditions
1752 msg->msg_txpeer = best_lpni;
1753 msg->msg_txni = best_ni;
1756 * grab a reference for the best_ni since now it's in use in this
1757 * send. The reference will be dropped in lnet_finalize()
1759 lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1762 * Always set the target.nid to the best peer picked. Either the
1763 * NID will be one of the peer NIDs selected, or the same NID as
1764 * what was originally set in the target or it will be the NID of
1765 * a router if this message should be routed
1767 msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1770 * lnet_msg_commit assigns the correct cpt to the message, which
1771 * is used to decrement the correct refcount on the ni when it's
1772 * time to return the credits
1774 lnet_msg_commit(msg, sd->sd_cpt);
1777 * If we are routing the message then we keep the src_nid that was
1778 * set by the originator. If we are not routing then we are the
1779 * originator and set it here.
1781 if (!msg->msg_routing)
1782 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1785 msg->msg_target_is_router = 1;
1786 msg->msg_target.pid = LNET_PID_LUSTRE;
1788 * since we're routing we want to ensure that the
1789 * msg_hdr.dest_nid is set to the final destination. When
1790 * the router receives this message it knows how to route
1793 * final_dst_lpni is set at the beginning of the
1794 * lnet_select_pathway() function and is never changed.
1795 * It's safe to use it here.
1797 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1800 * if we're not routing set the dest_nid to the best peer
1801 * ni NID that we picked earlier in the algorithm.
1803 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1807 * if we have response tracker block update it with the next hop
1811 rspt = msg->msg_md->md_rspt_ptr;
1813 rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1814 CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1815 libcfs_nid2str(rspt->rspt_next_hop_nid));
1819 rc = lnet_post_send_locked(msg, 0);
1822 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1823 libcfs_nid2str(msg->msg_hdr.src_nid),
1824 libcfs_nid2str(msg->msg_txni->ni_nid),
1825 libcfs_nid2str(sd->sd_src_nid),
1826 libcfs_nid2str(msg->msg_hdr.dest_nid),
1827 libcfs_nid2str(sd->sd_dst_nid),
1828 libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1829 libcfs_nid2str(sd->sd_rtr_nid),
1830 lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1836 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1837 struct lnet_msg *msg)
1839 if (!lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1840 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1841 libcfs_nid2str(lni->ni_nid),
1842 libcfs_nid2str(lpni->lpni_nid));
1843 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1852 * use the source and destination NIDs as the pathway
1855 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1857 /* the destination lpni is set before we get here. */
1860 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1861 if (!sd->sd_best_ni) {
1862 CERROR("Can't send to %s: src %s is not a "
1863 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1864 libcfs_nid2str(sd->sd_src_nid));
1868 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
1870 return lnet_handle_send(sd);
1878 * Don't run the selection algorithm on the peer NIs. By specifying the
1879 * local NID, we're also saying that we should always use the destination NID
1880 * provided. This handles the case where we should be using the same
1881 * destination NID for the all the messages which belong to the same RPC
1885 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1887 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1888 if (!sd->sd_best_ni) {
1889 CERROR("Can't send to %s: src %s is not a "
1890 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1891 libcfs_nid2str(sd->sd_src_nid));
1895 if (sd->sd_best_lpni &&
1896 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1897 return lnet_handle_lo_send(sd);
1898 else if (sd->sd_best_lpni)
1899 return lnet_handle_send(sd);
1901 CERROR("can't send to %s. no NI on %s\n",
1902 libcfs_nid2str(sd->sd_dst_nid),
1903 libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1905 return -EHOSTUNREACH;
1909 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1910 struct lnet_peer *peer,
1911 struct lnet_peer_net *peer_net,
1915 struct lnet_net *local_net;
1916 struct lnet_ni *best_ni;
1918 local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1923 * Iterate through the NIs in this local Net and select
1924 * the NI to send from. The selection is determined by
1925 * these 3 criterion in the following priority:
1927 * 2. NI available credits
1930 best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1931 peer, peer_net, cpt);
1933 if (incr_seq && best_ni)
1940 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
1943 struct lnet_peer *peer;
1946 lnet_peer_ni_addref_locked(lpni);
1948 peer = lpni->lpni_peer_net->lpn_peer;
1950 if (lnet_peer_gw_discovery(peer)) {
1951 lnet_peer_ni_decref_locked(lpni);
1955 if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1956 lnet_peer_ni_decref_locked(lpni);
1960 rc = lnet_discover_peer_locked(lpni, cpt, false);
1962 lnet_peer_ni_decref_locked(lpni);
1965 /* The peer may have changed. */
1966 peer = lpni->lpni_peer_net->lpn_peer;
1967 spin_lock(&peer->lp_lock);
1968 if (lnet_peer_is_uptodate_locked(peer)) {
1969 spin_unlock(&peer->lp_lock);
1970 lnet_peer_ni_decref_locked(lpni);
1973 /* queue message and return */
1974 msg->msg_sending = 0;
1975 msg->msg_txpeer = NULL;
1976 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1977 spin_unlock(&peer->lp_lock);
1979 lnet_peer_ni_decref_locked(lpni);
1981 CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
1982 msg, libcfs_nid2str(peer->lp_primary_nid));
1984 return LNET_DC_WAIT;
1988 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1990 struct lnet_peer_ni **gw_lpni,
1991 struct lnet_peer **gw_peer)
1995 struct lnet_peer *gw;
1996 struct lnet_peer *lp;
1997 struct lnet_peer_net *lpn;
1998 struct lnet_peer_net *best_lpn = NULL;
1999 struct lnet_remotenet *rnet, *best_rnet = NULL;
2000 struct lnet_route *best_route = NULL;
2001 struct lnet_route *last_route = NULL;
2002 struct lnet_peer_ni *lpni = NULL;
2003 struct lnet_peer_ni *gwni = NULL;
2004 lnet_nid_t src_nid = sd->sd_src_nid;
2006 /* If a router nid was specified then we are replying to a GET or
2007 * sending an ACK. In this case we use the gateway associated with the
2008 * specified router nid.
2010 if (sd->sd_rtr_nid != LNET_NID_ANY) {
2011 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2013 CERROR("No peer NI for gateway %s\n",
2014 libcfs_nid2str(sd->sd_rtr_nid));
2015 return -EHOSTUNREACH;
2017 gw = gwni->lpni_peer_net->lpn_peer;
2018 lnet_peer_ni_decref_locked(gwni);
2019 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2021 /* we've already looked up the initial lpni using dst_nid */
2022 lpni = sd->sd_best_lpni;
2023 /* the peer tree must be in existence */
2024 LASSERT(lpni && lpni->lpni_peer_net &&
2025 lpni->lpni_peer_net->lpn_peer);
2026 lp = lpni->lpni_peer_net->lpn_peer;
2028 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2029 /* is this remote network reachable? */
2030 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2039 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2047 CERROR("peer %s has no available nets\n",
2048 libcfs_nid2str(sd->sd_dst_nid));
2049 return -EHOSTUNREACH;
2052 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni,
2055 best_lpn->lpn_net_id);
2056 if (!sd->sd_best_lpni) {
2057 CERROR("peer %s down\n",
2058 libcfs_nid2str(sd->sd_dst_nid));
2059 return -EHOSTUNREACH;
2063 * We're attempting to round robin over the remote peer
2064 * NI's so update the final destination we selected
2066 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2069 * find the best route. Restrict the selection on the net of the
2070 * local NI if we've already picked the local NI to send from.
2071 * Otherwise, let's pick any route we can find and then find
2072 * a local NI we can reach the route's gateway on. Any route we select
2073 * will be reachable by virtue of the restriction we have when
2076 best_route = lnet_find_route_locked(best_rnet,
2077 LNET_NIDNET(src_nid),
2078 &last_route, &gwni);
2081 CERROR("no route to %s from %s\n",
2082 libcfs_nid2str(dst_nid),
2083 libcfs_nid2str(src_nid));
2084 return -EHOSTUNREACH;
2088 CERROR("Internal Error. Route expected to %s from %s\n",
2089 libcfs_nid2str(dst_nid),
2090 libcfs_nid2str(src_nid));
2094 gw = best_route->lr_gateway;
2095 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2096 local_lnet = best_route->lr_lnet;
2099 * Increment the sequence number of the remote lpni so we
2100 * can round robin over the different interfaces of the
2103 sd->sd_best_lpni->lpni_seq++;
2107 * Discover this gateway if it hasn't already been discovered.
2108 * This means we might delay the message until discovery has
2111 sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2112 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2116 if (!sd->sd_best_ni)
2117 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2118 lnet_peer_get_net_locked(gw,
2123 if (!sd->sd_best_ni) {
2124 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2125 libcfs_net2str(local_lnet),
2126 libcfs_nid2str(sd->sd_src_nid));
2134 * increment the sequence numbers since now we're sure we're
2135 * going to use this path
2137 if (sd->sd_rtr_nid == LNET_NID_ANY) {
2138 LASSERT(best_route && last_route);
2139 best_route->lr_seq = last_route->lr_seq + 1;
2140 best_lpn->lpn_seq++;
2151 * Remote destination
2152 * Non-MR destination
2156 * Remote destination
2159 * The handling of these two cases is similar. Even though the destination
2160 * can be MR or non-MR, we'll deal directly with the router.
2163 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2166 struct lnet_peer_ni *gw_lpni = NULL;
2167 struct lnet_peer *gw_peer = NULL;
2170 sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2171 if (!sd->sd_best_ni) {
2172 CERROR("Can't send to %s: src %s is not a "
2173 "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2174 libcfs_nid2str(sd->sd_src_nid));
2178 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2183 if (sd->sd_send_case & NMR_DST)
2185 * since the final destination is non-MR let's set its preferred
2186 * NID before we send
2188 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2192 * We're going to send to the gw found so let's set its
2195 sd->sd_peer = gw_peer;
2196 sd->sd_best_lpni = gw_lpni;
2198 return lnet_handle_send(sd);
2202 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2205 struct lnet_peer_net *peer_net = NULL;
2206 struct lnet_ni *best_ni = NULL;
2207 int lpn_healthv = 0;
2210 * The peer can have multiple interfaces, some of them can be on
2211 * the local network and others on a routed network. We should
2212 * prefer the local network. However if the local network is not
2213 * available then we need to try the routed network
2216 /* go through all the peer nets and find the best_ni */
2217 list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2219 * The peer's list of nets can contain non-local nets. We
2220 * want to only examine the local ones.
2222 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2225 /* always select the lpn with the best health */
2226 if (lpn_healthv <= peer_net->lpn_healthv)
2227 lpn_healthv = peer_net->lpn_healthv;
2231 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer, peer_net,
2235 * if this is a discovery message and lp_disc_net_id is
2236 * specified then use that net to send the discovery on.
2238 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2244 /* increment sequence number so we can round robin */
2250 static struct lnet_ni *
2251 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2253 struct lnet_ni *best_ni = NULL;
2254 struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2255 struct lnet_peer_ni *lpni_entry;
2258 * We must use a consistent source address when sending to a
2259 * non-MR peer. However, a non-MR peer can have multiple NIDs
2260 * on multiple networks, and we may even need to talk to this
2261 * peer on multiple networks -- certain types of
2262 * load-balancing configuration do this.
2264 * So we need to pick the NI the peer prefers for this
2265 * particular network.
2268 list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2270 if (lpni_entry->lpni_pref_nnids == 0)
2272 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2273 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2280 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2282 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2284 struct lnet_ni *best_ni = NULL;
2285 struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2288 * We must use a consistent source address when sending to a
2289 * non-MR peer. However, a non-MR peer can have multiple NIDs
2290 * on multiple networks, and we may even need to talk to this
2291 * peer on multiple networks -- certain types of
2292 * load-balancing configuration do this.
2294 * So we need to pick the NI the peer prefers for this
2295 * particular network.
2298 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2301 /* if best_ni is still not set just pick one */
2304 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2305 sd->sd_best_lpni->lpni_peer_net,
2306 sd->sd_md_cpt, true);
2307 /* If there is no best_ni we don't have a route */
2309 CERROR("no path to %s from net %s\n",
2310 libcfs_nid2str(best_lpni->lpni_nid),
2311 libcfs_net2str(best_lpni->lpni_net->net_id));
2312 return -EHOSTUNREACH;
2316 sd->sd_best_ni = best_ni;
2318 /* Set preferred NI if necessary. */
2319 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2326 * Source not specified
2330 * always use the same source NID for NMR peers
2331 * If we've talked to that peer before then we already have a preferred
2332 * source NI associated with it. Otherwise, we select a preferred local NI
2333 * and store it in the peer
2336 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2340 /* sd->sd_best_lpni is already set to the final destination */
2343 * At this point we should've created the peer ni and peer. If we
2344 * can't find it, then something went wrong. Instead of assert
2345 * output a relevant message and fail the send
2347 if (!sd->sd_best_lpni) {
2348 CERROR("Internal fault. Unable to send msg %s to %s. "
2350 lnet_msgtyp2str(sd->sd_msg->msg_type),
2351 libcfs_nid2str(sd->sd_dst_nid));
2355 rc = lnet_select_preferred_best_ni(sd);
2357 rc = lnet_handle_send(sd);
2363 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2366 * NOTE we've already handled the remote peer case. So we only
2367 * need to worry about the local case here.
2369 * if we're sending a response, ACK or reply, we need to send it
2370 * to the destination NID given to us. At this point we already
2371 * have the peer_ni we're suppose to send to, so just find the
2372 * best_ni on the peer net and use that. Since we're sending to an
2373 * MR peer then we can just run the selection algorithm on our
2374 * local NIs and pick the best one.
2376 if (sd->sd_send_case & SND_RESP) {
2378 lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2379 sd->sd_best_lpni->lpni_peer_net,
2380 sd->sd_md_cpt, true);
2382 if (!sd->sd_best_ni) {
2384 * We're not going to deal with not able to send
2385 * a response to the provided final destination
2387 CERROR("Can't send response to %s. "
2388 "No local NI available\n",
2389 libcfs_nid2str(sd->sd_dst_nid));
2390 return -EHOSTUNREACH;
2393 return lnet_handle_send(sd);
2397 * If we get here that means we're sending a fresh request, PUT or
2398 * GET, so we need to run our standard selection algorithm.
2399 * First find the best local interface that's on any of the peer's
2402 sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2404 lnet_msg_discovery(sd->sd_msg));
2405 if (sd->sd_best_ni) {
2407 lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid,
2409 sd->sd_best_ni->ni_net->net_id);
2412 * if we're successful in selecting a peer_ni on the local
2413 * network, then send to it. Otherwise fall through and
2414 * try and see if we can reach it over another routed
2417 if (sd->sd_best_lpni &&
2418 sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2420 * in case we initially started with a routed
2421 * destination, let's reset to local
2423 sd->sd_send_case &= ~REMOTE_DST;
2424 sd->sd_send_case |= LOCAL_DST;
2425 return lnet_handle_lo_send(sd);
2426 } else if (sd->sd_best_lpni) {
2428 * in case we initially started with a routed
2429 * destination, let's reset to local
2431 sd->sd_send_case &= ~REMOTE_DST;
2432 sd->sd_send_case |= LOCAL_DST;
2433 return lnet_handle_send(sd);
2436 CERROR("Internal Error. Expected to have a best_lpni: "
2438 libcfs_nid2str(sd->sd_src_nid),
2439 libcfs_nid2str(sd->sd_dst_nid));
2445 * Peer doesn't have a local network. Let's see if there is
2446 * a remote network we can reach it on.
2448 return PASS_THROUGH;
2453 * Source NID not specified
2458 * Source NID not speified
2459 * Remote destination
2462 * In both of these cases if we're sending a response, ACK or REPLY, then
2463 * we need to send to the destination NID provided.
2465 * In the remote case let's deal with MR routers.
2470 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2473 struct lnet_peer *gw_peer = NULL;
2474 struct lnet_peer_ni *gw_lpni = NULL;
2477 * handle sending a response to a remote peer here so we don't
2478 * have to worry about it if we hit lnet_handle_any_mr_dsta()
2480 if (sd->sd_send_case & REMOTE_DST &&
2481 sd->sd_send_case & SND_RESP) {
2482 struct lnet_peer_ni *gw;
2483 struct lnet_peer *gw_peer;
2485 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2488 CERROR("Can't send response to %s. "
2489 "No route available\n",
2490 libcfs_nid2str(sd->sd_dst_nid));
2491 return -EHOSTUNREACH;
2492 } else if (rc > 0) {
2496 sd->sd_best_lpni = gw;
2497 sd->sd_peer = gw_peer;
2499 return lnet_handle_send(sd);
2503 * Even though the NID for the peer might not be on a local network,
2504 * since the peer is MR there could be other interfaces on the
2505 * local network. In that case we'd still like to prefer the local
2506 * network over the routed network. If we're unable to do that
2507 * then we select the best router among the different routed networks,
2508 * and if the router is MR then we can deal with it as such.
2510 rc = lnet_handle_any_mr_dsta(sd);
2511 if (rc != PASS_THROUGH)
2515 * Now that we must route to the destination, we must consider the
2516 * MR case, where the destination has multiple interfaces, some of
2517 * which we can route to and others we do not. For this reason we
2518 * need to select the destination which we can route to and if
2519 * there are multiple, we need to round robin.
2521 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2526 sd->sd_send_case &= ~LOCAL_DST;
2527 sd->sd_send_case |= REMOTE_DST;
2529 sd->sd_peer = gw_peer;
2530 sd->sd_best_lpni = gw_lpni;
2532 return lnet_handle_send(sd);
2536 * Source not specified
2537 * Remote destination
2540 * Must send to the specified peer NID using the same source NID that
2541 * we've used before. If it's the first time to talk to that peer then
2542 * find the source NI and assign it as preferred to that peer
2545 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2548 struct lnet_peer_ni *gw_lpni = NULL;
2549 struct lnet_peer *gw_peer = NULL;
2552 * Let's see if we have a preferred NI to talk to this NMR peer
2554 sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2558 * find the router and that'll find the best NI if we didn't find
2561 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2567 * set the best_ni we've chosen as the preferred one for
2570 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2572 /* we'll be sending to the gw */
2573 sd->sd_best_lpni = gw_lpni;
2574 sd->sd_peer = gw_peer;
2576 return lnet_handle_send(sd);
2580 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2583 * turn off the SND_RESP bit.
2584 * It will be checked in the case handling
2586 __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2588 CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2589 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2590 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2591 (send_case & MR_DST) ? "MR: " : "NMR: ",
2592 libcfs_nid2str(sd->sd_dst_nid),
2593 (send_case & LOCAL_DST) ? "local" : "routed");
2595 switch (send_case) {
2597 * For all cases where the source is specified, we should always
2598 * use the destination NID, whether it's an MR destination or not,
2599 * since we're continuing a series of related messages for the
2602 case SRC_SPEC_LOCAL_NMR_DST:
2603 return lnet_handle_spec_local_nmr_dst(sd);
2604 case SRC_SPEC_LOCAL_MR_DST:
2605 return lnet_handle_spec_local_mr_dst(sd);
2606 case SRC_SPEC_ROUTER_NMR_DST:
2607 case SRC_SPEC_ROUTER_MR_DST:
2608 return lnet_handle_spec_router_dst(sd);
2609 case SRC_ANY_LOCAL_NMR_DST:
2610 return lnet_handle_any_local_nmr_dst(sd);
2611 case SRC_ANY_LOCAL_MR_DST:
2612 case SRC_ANY_ROUTER_MR_DST:
2613 return lnet_handle_any_mr_dst(sd);
2614 case SRC_ANY_ROUTER_NMR_DST:
2615 return lnet_handle_any_router_nmr_dst(sd);
2617 CERROR("Unknown send case\n");
2623 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2624 struct lnet_msg *msg, lnet_nid_t rtr_nid)
2626 struct lnet_peer_ni *lpni;
2627 struct lnet_peer *peer;
2628 struct lnet_send_data send_data;
2631 __u32 send_case = 0;
2633 memset(&send_data, 0, sizeof(send_data));
2636 * get an initial CPT to use for locking. The idea here is not to
2637 * serialize the calls to select_pathway, so that as many
2638 * operations can run concurrently as possible. To do that we use
2639 * the CPT where this call is being executed. Later on when we
2640 * determine the CPT to use in lnet_message_commit, we switch the
2641 * lock and check if there was any configuration change. If none,
2642 * then we proceed, if there is, then we restart the operation.
2644 cpt = lnet_net_lock_current();
2646 md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2647 if (md_cpt == CFS_CPT_ANY)
2653 * If we're being asked to send to the loopback interface, there
2654 * is no need to go through any selection. We can just shortcut
2655 * the entire process and send over lolnd
2657 send_data.sd_msg = msg;
2658 send_data.sd_cpt = cpt;
2659 if (dst_nid == LNET_NID_LO_0) {
2660 rc = lnet_handle_lo_send(&send_data);
2661 lnet_net_unlock(cpt);
2666 * find an existing peer_ni, or create one and mark it as having been
2667 * created due to network traffic. This call will create the
2668 * peer->peer_net->peer_ni tree.
2670 lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2672 lnet_net_unlock(cpt);
2673 return PTR_ERR(lpni);
2677 * Cache the original src_nid and rtr_nid. If we need to resend the
2678 * message then we'll need to know whether the src_nid was originally
2679 * specified for this message. If it was originally specified,
2680 * then we need to keep using the same src_nid since it's
2681 * continuing the same sequence of messages. Similarly, rtr_nid will
2682 * affect our choice of next hop.
2684 msg->msg_src_nid_param = src_nid;
2685 msg->msg_rtr_nid_param = rtr_nid;
2688 * If necessary, perform discovery on the peer that owns this peer_ni.
2689 * Note, this can result in the ownership of this peer_ni changing
2690 * to another peer object.
2692 rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2694 lnet_peer_ni_decref_locked(lpni);
2695 lnet_net_unlock(cpt);
2698 lnet_peer_ni_decref_locked(lpni);
2700 peer = lpni->lpni_peer_net->lpn_peer;
2703 * Identify the different send cases
2705 if (src_nid == LNET_NID_ANY)
2706 send_case |= SRC_ANY;
2708 send_case |= SRC_SPEC;
2710 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2711 send_case |= LOCAL_DST;
2713 send_case |= REMOTE_DST;
2716 * Deal with the peer as NMR in the following cases:
2717 * 1. the peer is NMR
2718 * 2. We're trying to recover a specific peer NI
2719 * 3. I'm a router sending to the final destination
2720 * In this case the source of the message would've
2721 * already selected the final destination so my job
2722 * is to honor the selection.
2724 if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
2725 (msg->msg_routing && (send_case & LOCAL_DST)))
2726 send_case |= NMR_DST;
2728 send_case |= MR_DST;
2730 if (lnet_msg_is_response(msg))
2731 send_case |= SND_RESP;
2733 /* assign parameters to the send_data */
2734 send_data.sd_rtr_nid = rtr_nid;
2735 send_data.sd_src_nid = src_nid;
2736 send_data.sd_dst_nid = dst_nid;
2737 send_data.sd_best_lpni = lpni;
2739 * keep a pointer to the final destination in case we're going to
2740 * route, so we'll need to access it later
2742 send_data.sd_final_dst_lpni = lpni;
2743 send_data.sd_peer = peer;
2744 send_data.sd_md_cpt = md_cpt;
2745 send_data.sd_send_case = send_case;
2747 rc = lnet_handle_send_case_locked(&send_data);
2750 * Update the local cpt since send_data.sd_cpt might've been
2751 * updated as a result of calling lnet_handle_send_case_locked().
2753 cpt = send_data.sd_cpt;
2755 if (rc == REPEAT_SEND)
2758 lnet_net_unlock(cpt);
2764 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2766 lnet_nid_t dst_nid = msg->msg_target.nid;
2770 * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2771 * but we might want to use pre-determined router for ACK/REPLY
2774 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2775 LASSERT(msg->msg_txpeer == NULL);
2776 LASSERT(msg->msg_txni == NULL);
2777 LASSERT(!msg->msg_sending);
2778 LASSERT(!msg->msg_target_is_router);
2779 LASSERT(!msg->msg_receiving);
2781 msg->msg_sending = 1;
2783 LASSERT(!msg->msg_tx_committed);
2785 rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2787 if (rc == -EHOSTUNREACH)
2788 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2790 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2794 if (rc == LNET_CREDIT_OK)
2795 lnet_ni_send(msg->msg_txni, msg);
2797 /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2801 enum lnet_mt_event_type {
2802 MT_TYPE_LOCAL_NI = 0,
2806 struct lnet_mt_event_info {
2807 enum lnet_mt_event_type mt_type;
2811 /* called with res_lock held */
2813 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2815 struct lnet_rsp_tracker *rspt;
2818 * msg has a refcount on the MD so the MD is not going away.
2819 * The rspt queue for the cpt is protected by
2820 * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2822 if (!md->md_rspt_ptr)
2825 rspt = md->md_rspt_ptr;
2828 LASSERT(rspt->rspt_cpt == cpt);
2830 md->md_rspt_ptr = NULL;
2832 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2834 * The monitor thread has invalidated this handle because the
2835 * response timed out, but it failed to lookup the MD. That
2836 * means this response tracker is on the zombie list. We can
2837 * safely remove it under the resource lock (held by caller) and
2838 * free the response tracker block.
2840 list_del(&rspt->rspt_on_list);
2841 lnet_rspt_free(rspt, cpt);
2844 * invalidate the handle to indicate that a response has been
2845 * received, which will then lead the monitor thread to clean up
2848 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2853 lnet_clean_zombie_rstqs(void)
2855 struct lnet_rsp_tracker *rspt, *tmp;
2858 cfs_cpt_for_each(i, lnet_cpt_table()) {
2859 list_for_each_entry_safe(rspt, tmp,
2860 the_lnet.ln_mt_zombie_rstqs[i],
2862 list_del(&rspt->rspt_on_list);
2863 lnet_rspt_free(rspt, i);
2867 cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2871 lnet_finalize_expired_responses(void)
2873 struct lnet_libmd *md;
2874 struct lnet_rsp_tracker *rspt, *tmp;
2878 if (the_lnet.ln_mt_rstq == NULL)
2881 cfs_cpt_for_each(i, lnet_cpt_table()) {
2882 LIST_HEAD(local_queue);
2885 if (!the_lnet.ln_mt_rstq[i]) {
2889 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2894 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2896 * The rspt mdh will be invalidated when a response
2897 * is received or whenever we want to discard the
2898 * block the monitor thread will walk the queue
2899 * and clean up any rsts with an invalid mdh.
2900 * The monitor thread will walk the queue until
2901 * the first unexpired rspt block. This means that
2902 * some rspt blocks which received their
2903 * corresponding responses will linger in the
2904 * queue until they are cleaned up eventually.
2907 if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2909 list_del(&rspt->rspt_on_list);
2910 lnet_rspt_free(rspt, i);
2914 if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2915 the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2916 struct lnet_peer_ni *lpni;
2919 md = lnet_handle2md(&rspt->rspt_mdh);
2921 /* MD has been queued for unlink, but
2922 * rspt hasn't been detached (Note we've
2923 * checked above that the rspt_mdh is
2924 * valid). Since we cannot lookup the MD
2925 * we're unable to detach the rspt
2926 * ourselves. Thus, move the rspt to the
2927 * zombie list where we'll wait for
2929 * 1. The remaining operations on the
2930 * MD to complete. In this case the
2931 * final operation will result in
2932 * lnet_msg_detach_md()->
2933 * lnet_detach_rsp_tracker() where
2934 * we will clean up this response
2936 * 2. LNet to shutdown. In this case
2937 * we'll wait until after all LND Nets
2938 * have shutdown and then we can
2939 * safely free any remaining response
2940 * tracker blocks on the zombie list.
2941 * Note: We need to hold the resource
2942 * lock when adding to the zombie list
2943 * because we may have concurrent access
2944 * with lnet_detach_rsp_tracker().
2946 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2947 list_move(&rspt->rspt_on_list,
2948 the_lnet.ln_mt_zombie_rstqs[i]);
2952 LASSERT(md->md_rspt_ptr == rspt);
2953 md->md_rspt_ptr = NULL;
2956 LNetMDUnlink(rspt->rspt_mdh);
2958 nid = rspt->rspt_next_hop_nid;
2960 list_del(&rspt->rspt_on_list);
2961 lnet_rspt_free(rspt, i);
2963 /* If we're shutting down we just want to clean
2964 * up the rspt blocks
2966 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2970 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2974 "Response timeout: md = %p: nid = %s\n",
2975 md, libcfs_nid2str(nid));
2978 * If there is a timeout on the response
2979 * from the next hop decrement its health
2980 * value so that we don't use it
2983 lpni = lnet_find_peer_ni_locked(nid);
2985 lnet_handle_remote_failure_locked(lpni);
2986 lnet_peer_ni_decref_locked(lpni);
2995 if (!list_empty(&local_queue)) {
2997 list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3004 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3006 struct lnet_msg *msg;
3008 while (!list_empty(resendq)) {
3009 struct lnet_peer_ni *lpni;
3011 msg = list_entry(resendq->next, struct lnet_msg,
3014 list_del_init(&msg->msg_list);
3016 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3018 lnet_net_unlock(cpt);
3019 CERROR("Expected that a peer is already created for %s\n",
3020 libcfs_nid2str(msg->msg_hdr.dest_nid));
3021 msg->msg_no_resend = true;
3022 lnet_finalize(msg, -EFAULT);
3027 lnet_peer_ni_decref_locked(lpni);
3029 lnet_net_unlock(cpt);
3030 CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3031 libcfs_nid2str(msg->msg_src_nid_param),
3032 libcfs_id2str(msg->msg_target),
3033 lnet_msgtyp2str(msg->msg_type),
3035 msg->msg_retry_count);
3036 rc = lnet_send(msg->msg_src_nid_param, msg,
3037 msg->msg_rtr_nid_param);
3039 CERROR("Error sending %s to %s: %d\n",
3040 lnet_msgtyp2str(msg->msg_type),
3041 libcfs_id2str(msg->msg_target), rc);
3042 msg->msg_no_resend = true;
3043 lnet_finalize(msg, rc);
3047 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3053 lnet_resend_pending_msgs(void)
3057 cfs_cpt_for_each(i, lnet_cpt_table()) {
3059 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3064 /* called with cpt and ni_lock held */
3066 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3068 struct lnet_handle_md recovery_mdh;
3070 LNetInvalidateMDHandle(&recovery_mdh);
3072 if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3074 recovery_mdh = ni->ni_ping_mdh;
3075 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3078 lnet_net_unlock(cpt);
3079 if (!LNetMDHandleIsInvalid(recovery_mdh))
3080 LNetMDUnlink(recovery_mdh);
3086 lnet_recover_local_nis(void)
3088 struct lnet_mt_event_info *ev_info;
3089 LIST_HEAD(processed_list);
3090 LIST_HEAD(local_queue);
3091 struct lnet_handle_md mdh;
3092 struct lnet_ni *tmp;
3099 * splice the recovery queue on a local queue. We will iterate
3100 * through the local queue and update it as needed. Once we're
3101 * done with the traversal, we'll splice the local queue back on
3102 * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3103 * will be traversed in the next iteration.
3106 list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3110 list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3112 * if an NI is being deleted or it is now healthy, there
3113 * is no need to keep it around in the recovery queue.
3114 * The monitor thread is the only thread responsible for
3115 * removing the NI from the recovery queue.
3116 * Multiple threads can be adding NIs to the recovery
3119 healthv = atomic_read(&ni->ni_healthv);
3123 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3124 healthv == LNET_MAX_HEALTH_VALUE) {
3125 list_del_init(&ni->ni_recovery);
3126 lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3128 lnet_ni_decref_locked(ni, 0);
3134 * if the local NI failed recovery we must unlink the md.
3135 * But we want to keep the local_ni on the recovery queue
3136 * so we can continue the attempts to recover it.
3138 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3139 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3140 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3147 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3148 libcfs_nid2str(ni->ni_nid));
3151 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3152 ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3155 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3157 CERROR("out of memory. Can't recover %s\n",
3158 libcfs_nid2str(ni->ni_nid));
3160 ni->ni_recovery_state &=
3161 ~LNET_NI_RECOVERY_PENDING;
3166 mdh = ni->ni_ping_mdh;
3168 * Invalidate the ni mdh in case it's deleted.
3169 * We'll unlink the mdh in this case below.
3171 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3175 * remove the NI from the local queue and drop the
3176 * reference count to it while we're recovering
3177 * it. The reason for that, is that the NI could
3178 * be deleted, and the way the code is structured
3179 * is if we don't drop the NI, then the deletion
3180 * code will enter a loop waiting for the
3181 * reference count to be removed while holding the
3182 * ln_mutex_lock(). When we look up the peer to
3183 * send to in lnet_select_pathway() we will try to
3184 * lock the ln_mutex_lock() as well, leading to
3185 * a deadlock. By dropping the refcount and
3186 * removing it from the list, we allow for the NI
3187 * to be removed, then we use the cached NID to
3188 * look it up again. If it's gone, then we just
3189 * continue examining the rest of the queue.
3192 list_del_init(&ni->ni_recovery);
3193 lnet_ni_decref_locked(ni, 0);
3196 ev_info->mt_type = MT_TYPE_LOCAL_NI;
3197 ev_info->mt_nid = nid;
3198 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3199 ev_info, the_lnet.ln_mt_handler,
3201 /* lookup the nid again */
3203 ni = lnet_nid2ni_locked(nid, 0);
3206 * the NI has been deleted when we dropped
3214 * Same note as in lnet_recover_peer_nis(). When
3215 * we're sending the ping, the NI is free to be
3216 * deleted or manipulated. By this point it
3217 * could've been added back on the recovery queue,
3218 * and a refcount taken on it.
3219 * So we can't just add it blindly again or we'll
3220 * corrupt the queue. We must check under lock if
3221 * it's not on any list and if not then add it
3222 * to the processed list, which will eventually be
3223 * spliced back on to the recovery queue.
3225 ni->ni_ping_mdh = mdh;
3226 if (list_empty(&ni->ni_recovery)) {
3227 list_add_tail(&ni->ni_recovery, &processed_list);
3228 lnet_ni_addref_locked(ni, 0);
3234 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3240 * put back the remaining NIs on the ln_mt_localNIRecovq to be
3241 * reexamined in the next iteration.
3243 list_splice_init(&processed_list, &local_queue);
3245 list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3250 lnet_resendqs_create(void)
3252 struct list_head **resendqs;
3253 resendqs = lnet_create_array_of_queues();
3258 lnet_net_lock(LNET_LOCK_EX);
3259 the_lnet.ln_mt_resendqs = resendqs;
3260 lnet_net_unlock(LNET_LOCK_EX);
3266 lnet_clean_local_ni_recoveryq(void)
3270 /* This is only called when the monitor thread has stopped */
3273 while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3274 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3275 struct lnet_ni, ni_recovery);
3276 list_del_init(&ni->ni_recovery);
3278 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3280 lnet_ni_decref_locked(ni, 0);
3287 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3290 struct lnet_handle_md recovery_mdh;
3292 LNetInvalidateMDHandle(&recovery_mdh);
3294 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3295 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3296 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3298 spin_unlock(&lpni->lpni_lock);
3299 lnet_net_unlock(cpt);
3300 if (!LNetMDHandleIsInvalid(recovery_mdh))
3301 LNetMDUnlink(recovery_mdh);
3303 spin_lock(&lpni->lpni_lock);
3307 lnet_clean_peer_ni_recoveryq(void)
3309 struct lnet_peer_ni *lpni, *tmp;
3311 lnet_net_lock(LNET_LOCK_EX);
3313 list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3315 list_del_init(&lpni->lpni_recovery);
3316 spin_lock(&lpni->lpni_lock);
3317 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3318 spin_unlock(&lpni->lpni_lock);
3319 lnet_peer_ni_decref_locked(lpni);
3322 lnet_net_unlock(LNET_LOCK_EX);
3326 lnet_clean_resendqs(void)
3328 struct lnet_msg *msg, *tmp;
3332 cfs_cpt_for_each(i, lnet_cpt_table()) {
3334 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3336 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3337 list_del_init(&msg->msg_list);
3338 msg->msg_no_resend = true;
3339 lnet_finalize(msg, -ESHUTDOWN);
3343 cfs_percpt_free(the_lnet.ln_mt_resendqs);
3347 lnet_recover_peer_nis(void)
3349 struct lnet_mt_event_info *ev_info;
3350 LIST_HEAD(processed_list);
3351 LIST_HEAD(local_queue);
3352 struct lnet_handle_md mdh;
3353 struct lnet_peer_ni *lpni;
3354 struct lnet_peer_ni *tmp;
3360 * Always use cpt 0 for locking across all interactions with
3361 * ln_mt_peerNIRecovq
3364 list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3368 list_for_each_entry_safe(lpni, tmp, &local_queue,
3371 * The same protection strategy is used here as is in the
3372 * local recovery case.
3375 healthv = atomic_read(&lpni->lpni_healthv);
3376 spin_lock(&lpni->lpni_lock);
3377 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3378 healthv == LNET_MAX_HEALTH_VALUE) {
3379 list_del_init(&lpni->lpni_recovery);
3380 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3381 spin_unlock(&lpni->lpni_lock);
3382 lnet_peer_ni_decref_locked(lpni);
3388 * If the peer NI has failed recovery we must unlink the
3389 * md. But we want to keep the peer ni on the recovery
3390 * queue so we can try to continue recovering it
3392 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3393 lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3394 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3397 spin_unlock(&lpni->lpni_lock);
3401 * NOTE: we're racing with peer deletion from user space.
3402 * It's possible that a peer is deleted after we check its
3403 * state. In this case the recovery can create a new peer
3405 spin_lock(&lpni->lpni_lock);
3406 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3407 !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3408 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3409 spin_unlock(&lpni->lpni_lock);
3411 LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3413 CERROR("out of memory. Can't recover %s\n",
3414 libcfs_nid2str(lpni->lpni_nid));
3415 spin_lock(&lpni->lpni_lock);
3416 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3417 spin_unlock(&lpni->lpni_lock);
3421 /* look at the comments in lnet_recover_local_nis() */
3422 mdh = lpni->lpni_recovery_ping_mdh;
3423 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3424 nid = lpni->lpni_nid;
3426 list_del_init(&lpni->lpni_recovery);
3427 lnet_peer_ni_decref_locked(lpni);
3430 ev_info->mt_type = MT_TYPE_PEER_NI;
3431 ev_info->mt_nid = nid;
3432 rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3433 ev_info, the_lnet.ln_mt_handler,
3437 * lnet_find_peer_ni_locked() grabs a refcount for
3438 * us. No need to take it explicitly.
3440 lpni = lnet_find_peer_ni_locked(nid);
3447 lpni->lpni_recovery_ping_mdh = mdh;
3449 * While we're unlocked the lpni could've been
3450 * readded on the recovery queue. In this case we
3451 * don't need to add it to the local queue, since
3452 * it's already on there and the thread that added
3453 * it would've incremented the refcount on the
3454 * peer, which means we need to decref the refcount
3455 * that was implicitly grabbed by find_peer_ni_locked.
3456 * Otherwise, if the lpni is still not on
3457 * the recovery queue, then we'll add it to the
3460 if (list_empty(&lpni->lpni_recovery))
3461 list_add_tail(&lpni->lpni_recovery, &processed_list);
3463 lnet_peer_ni_decref_locked(lpni);
3466 spin_lock(&lpni->lpni_lock);
3468 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3470 spin_unlock(&lpni->lpni_lock);
3473 list_splice_init(&processed_list, &local_queue);
3475 list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3480 lnet_monitor_thread(void *arg)
3482 time64_t recovery_timeout = 0;
3483 time64_t rsp_timeout = 0;
3487 wait_for_completion(&the_lnet.ln_started);
3489 * The monitor thread takes care of the following:
3490 * 1. Checks the aliveness of routers
3491 * 2. Checks if there are messages on the resend queue to resend
3493 * 3. Check if there are any NIs on the local recovery queue and
3495 * 4. Checks if there are any NIs on the remote recovery queue
3498 while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3499 now = ktime_get_real_seconds();
3501 if (lnet_router_checker_active())
3502 lnet_check_routers();
3504 lnet_resend_pending_msgs();
3506 if (now >= rsp_timeout) {
3507 lnet_finalize_expired_responses();
3508 rsp_timeout = now + (lnet_transaction_timeout / 2);
3511 if (now >= recovery_timeout) {
3512 lnet_recover_local_nis();
3513 lnet_recover_peer_nis();
3514 recovery_timeout = now + lnet_recovery_interval;
3518 * TODO do we need to check if we should sleep without
3519 * timeout? Technically, an active system will always
3520 * have messages in flight so this check will always
3521 * evaluate to false. And on an idle system do we care
3522 * if we wake up every 1 second? Although, we've seen
3523 * cases where we get a complaint that an idle thread
3524 * is waking up unnecessarily.
3526 * Take into account the current net_count when you wake
3527 * up for alive router checking, since we need to check
3528 * possibly as many networks as we have configured.
3530 interval = min(lnet_recovery_interval,
3531 min((unsigned int) alive_router_check_interval /
3532 lnet_current_net_count,
3533 lnet_transaction_timeout / 2));
3534 wait_for_completion_interruptible_timeout(
3535 &the_lnet.ln_mt_wait_complete,
3536 cfs_time_seconds(interval));
3537 /* Must re-init the completion before testing anything,
3538 * including ln_mt_state.
3540 reinit_completion(&the_lnet.ln_mt_wait_complete);
3544 lnet_net_lock(LNET_LOCK_EX);
3545 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3546 lnet_net_unlock(LNET_LOCK_EX);
3548 /* signal that the monitor thread is exiting */
3549 up(&the_lnet.ln_mt_signal);
3557 * Returns == 0 if success
3558 * Returns > 0 if LNetMDBind or prior fails
3559 * Returns < 0 if LNetGet fails
3562 lnet_send_ping(lnet_nid_t dest_nid,
3563 struct lnet_handle_md *mdh, int nnis,
3564 void *user_data, lnet_handler_t handler, bool recovery)
3566 struct lnet_md md = { NULL };
3567 struct lnet_process_id id;
3568 struct lnet_ping_buffer *pbuf;
3571 if (dest_nid == LNET_NID_ANY) {
3576 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3582 /* initialize md content */
3583 md.start = &pbuf->pb_info;
3584 md.length = LNET_PING_INFO_SIZE(nnis);
3585 md.threshold = 2; /* GET/REPLY */
3587 md.options = LNET_MD_TRUNCATE;
3588 md.user_ptr = user_data;
3589 md.handler = handler;
3591 rc = LNetMDBind(md, LNET_UNLINK, mdh);
3593 lnet_ping_buffer_decref(pbuf);
3594 CERROR("Can't bind MD: %d\n", rc);
3595 rc = -rc; /* change the rc to positive */
3598 id.pid = LNET_PID_LUSTRE;
3601 rc = LNetGet(LNET_NID_ANY, *mdh, id,
3602 LNET_RESERVED_PORTAL,
3603 LNET_PROTO_PING_MATCHBITS, 0, recovery);
3606 goto fail_unlink_md;
3612 LNetInvalidateMDHandle(mdh);
3618 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3619 int status, bool send, bool unlink_event)
3621 lnet_nid_t nid = ev_info->mt_nid;
3623 if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3627 ni = lnet_nid2ni_locked(nid, 0);
3633 if (!send || (send && status != 0))
3634 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3636 ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3641 CERROR("local NI (%s) recovery failed with %d\n",
3642 libcfs_nid2str(nid), status);
3646 * need to increment healthv for the ni here, because in
3647 * the lnet_finalize() path we don't have access to this
3648 * NI. And in order to get access to it, we'll need to
3649 * carry forward too much information.
3650 * In the peer case, it'll naturally be incremented
3653 lnet_inc_healthv(&ni->ni_healthv,
3654 lnet_health_sensitivity);
3656 struct lnet_peer_ni *lpni;
3659 cpt = lnet_net_lock_current();
3660 lpni = lnet_find_peer_ni_locked(nid);
3662 lnet_net_unlock(cpt);
3665 spin_lock(&lpni->lpni_lock);
3666 if (!send || (send && status != 0))
3667 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3669 lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3670 spin_unlock(&lpni->lpni_lock);
3671 lnet_peer_ni_decref_locked(lpni);
3672 lnet_net_unlock(cpt);
3675 CERROR("peer NI (%s) recovery failed with %d\n",
3676 libcfs_nid2str(nid), status);
3681 lnet_mt_event_handler(struct lnet_event *event)
3683 struct lnet_mt_event_info *ev_info = event->md_user_ptr;
3684 struct lnet_ping_buffer *pbuf;
3686 /* TODO: remove assert */
3687 LASSERT(event->type == LNET_EVENT_REPLY ||
3688 event->type == LNET_EVENT_SEND ||
3689 event->type == LNET_EVENT_UNLINK);
3691 CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3694 switch (event->type) {
3695 case LNET_EVENT_UNLINK:
3696 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3697 libcfs_nid2str(ev_info->mt_nid));
3699 case LNET_EVENT_REPLY:
3700 lnet_handle_recovery_reply(ev_info, event->status, false,
3701 event->type == LNET_EVENT_UNLINK);
3703 case LNET_EVENT_SEND:
3704 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3705 libcfs_nid2str(ev_info->mt_nid),
3706 (event->status) ? "unsuccessfully" :
3707 "successfully", event->status);
3708 lnet_handle_recovery_reply(ev_info, event->status, true, false);
3711 CERROR("Unexpected event: %d\n", event->type);
3714 if (event->unlinked) {
3715 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3716 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
3717 lnet_ping_buffer_decref(pbuf);
3722 lnet_rsp_tracker_create(void)
3724 struct list_head **rstqs;
3725 rstqs = lnet_create_array_of_queues();
3730 the_lnet.ln_mt_rstq = rstqs;
3736 lnet_rsp_tracker_clean(void)
3738 lnet_finalize_expired_responses();
3740 cfs_percpt_free(the_lnet.ln_mt_rstq);
3741 the_lnet.ln_mt_rstq = NULL;
3744 int lnet_monitor_thr_start(void)
3747 struct task_struct *task;
3749 if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3752 rc = lnet_resendqs_create();
3756 rc = lnet_rsp_tracker_create();
3760 sema_init(&the_lnet.ln_mt_signal, 0);
3762 lnet_net_lock(LNET_LOCK_EX);
3763 the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3764 lnet_net_unlock(LNET_LOCK_EX);
3765 task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3768 CERROR("Can't start monitor thread: %d\n", rc);
3775 lnet_net_lock(LNET_LOCK_EX);
3776 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3777 lnet_net_unlock(LNET_LOCK_EX);
3778 /* block until event callback signals exit */
3779 down(&the_lnet.ln_mt_signal);
3781 lnet_net_lock(LNET_LOCK_EX);
3782 the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3783 lnet_net_unlock(LNET_LOCK_EX);
3784 lnet_rsp_tracker_clean();
3785 lnet_clean_local_ni_recoveryq();
3786 lnet_clean_peer_ni_recoveryq();
3787 lnet_clean_resendqs();
3788 the_lnet.ln_mt_handler = NULL;
3791 lnet_rsp_tracker_clean();
3792 lnet_clean_local_ni_recoveryq();
3793 lnet_clean_peer_ni_recoveryq();
3794 lnet_clean_resendqs();
3798 void lnet_monitor_thr_stop(void)
3800 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3803 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3804 lnet_net_lock(LNET_LOCK_EX);
3805 the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3806 lnet_net_unlock(LNET_LOCK_EX);
3808 /* tell the monitor thread that we're shutting down */
3809 complete(&the_lnet.ln_mt_wait_complete);
3811 /* block until monitor thread signals that it's done */
3812 down(&the_lnet.ln_mt_signal);
3813 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3815 /* perform cleanup tasks */
3816 lnet_rsp_tracker_clean();
3817 lnet_clean_local_ni_recoveryq();
3818 lnet_clean_peer_ni_recoveryq();
3819 lnet_clean_resendqs();
3823 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3827 lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3828 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3829 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3830 lnet_net_unlock(cpt);
3832 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3836 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3838 struct lnet_hdr *hdr = &msg->msg_hdr;
3840 if (msg->msg_wanted != 0)
3841 lnet_setpayloadbuffer(msg);
3843 lnet_build_msg_event(msg, LNET_EVENT_PUT);
3845 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
3846 * it back into the ACK during lnet_finalize() */
3847 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3848 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3850 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3851 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3855 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3857 struct lnet_hdr *hdr = &msg->msg_hdr;
3858 struct lnet_match_info info;
3862 /* Convert put fields to host byte order */
3863 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3864 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
3865 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
3867 /* Primary peer NID. */
3868 info.mi_id.nid = msg->msg_initiator;
3869 info.mi_id.pid = hdr->src_pid;
3870 info.mi_opc = LNET_MD_OP_PUT;
3871 info.mi_portal = hdr->msg.put.ptl_index;
3872 info.mi_rlength = hdr->payload_length;
3873 info.mi_roffset = hdr->msg.put.offset;
3874 info.mi_mbits = hdr->msg.put.match_bits;
3875 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3877 msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3878 ready_delay = msg->msg_rx_ready_delay;
3881 rc = lnet_ptl_match_md(&info, msg);
3886 case LNET_MATCHMD_OK:
3887 lnet_recv_put(ni, msg);
3890 case LNET_MATCHMD_NONE:
3892 /* no eager_recv or has already called it, should
3893 * have been attached on delayed list */
3896 rc = lnet_ni_eager_recv(ni, msg);
3903 case LNET_MATCHMD_DROP:
3904 CNETERR("Dropping PUT from %s portal %d match %llu"
3905 " offset %d length %d: %d\n",
3906 libcfs_id2str(info.mi_id), info.mi_portal,
3907 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3909 return -ENOENT; /* -ve: OK but no match */
3914 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3916 struct lnet_match_info info;
3917 struct lnet_hdr *hdr = &msg->msg_hdr;
3918 struct lnet_process_id source_id;
3919 struct lnet_handle_wire reply_wmd;
3922 /* Convert get fields to host byte order */
3923 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
3924 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
3925 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
3926 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
3928 source_id.nid = hdr->src_nid;
3929 source_id.pid = hdr->src_pid;
3930 /* Primary peer NID */
3931 info.mi_id.nid = msg->msg_initiator;
3932 info.mi_id.pid = hdr->src_pid;
3933 info.mi_opc = LNET_MD_OP_GET;
3934 info.mi_portal = hdr->msg.get.ptl_index;
3935 info.mi_rlength = hdr->msg.get.sink_length;
3936 info.mi_roffset = hdr->msg.get.src_offset;
3937 info.mi_mbits = hdr->msg.get.match_bits;
3938 info.mi_cpt = lnet_cpt_of_nid(msg->msg_initiator, ni);
3940 rc = lnet_ptl_match_md(&info, msg);
3941 if (rc == LNET_MATCHMD_DROP) {
3942 CNETERR("Dropping GET from %s portal %d match %llu"
3943 " offset %d length %d\n",
3944 libcfs_id2str(info.mi_id), info.mi_portal,
3945 info.mi_mbits, info.mi_roffset, info.mi_rlength);
3946 return -ENOENT; /* -ve: OK but no match */
3949 LASSERT(rc == LNET_MATCHMD_OK);
3951 lnet_build_msg_event(msg, LNET_EVENT_GET);
3953 reply_wmd = hdr->msg.get.return_wmd;
3955 lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3956 msg->msg_offset, msg->msg_wanted);
3958 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3961 /* The LND completes the REPLY from her recv procedure */
3962 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3963 msg->msg_offset, msg->msg_len, msg->msg_len);
3967 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3968 msg->msg_receiving = 0;
3970 rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3972 /* didn't get as far as lnet_ni_send() */
3973 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3974 libcfs_nid2str(ni->ni_nid),
3975 libcfs_id2str(info.mi_id), rc);
3977 lnet_finalize(msg, rc);
3984 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3986 void *private = msg->msg_private;
3987 struct lnet_hdr *hdr = &msg->msg_hdr;
3988 struct lnet_process_id src = {0};
3989 struct lnet_libmd *md;
3990 unsigned int rlength;
3991 unsigned int mlength;
3994 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3997 src.nid = hdr->src_nid;
3998 src.pid = hdr->src_pid;
4000 /* NB handles only looked up by creator (no flips) */
4001 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4002 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4003 CNETERR("%s: Dropping REPLY from %s for %s "
4005 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4006 (md == NULL) ? "invalid" : "inactive",
4007 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4008 hdr->msg.reply.dst_wmd.wh_object_cookie);
4009 if (md != NULL && md->md_me != NULL)
4010 CERROR("REPLY MD also attached to portal %d\n",
4011 md->md_me->me_portal);
4013 lnet_res_unlock(cpt);
4014 return -ENOENT; /* -ve: OK but no match */
4017 LASSERT(md->md_offset == 0);
4019 rlength = hdr->payload_length;
4020 mlength = min(rlength, md->md_length);
4022 if (mlength < rlength &&
4023 (md->md_options & LNET_MD_TRUNCATE) == 0) {
4024 CNETERR("%s: Dropping REPLY from %s length %d "
4025 "for MD %#llx would overflow (%d)\n",
4026 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4027 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4029 lnet_res_unlock(cpt);
4030 return -ENOENT; /* -ve: OK but no match */
4033 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4034 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4035 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4037 lnet_msg_attach_md(msg, md, 0, mlength);
4040 lnet_setpayloadbuffer(msg);
4042 lnet_res_unlock(cpt);
4044 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4046 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4051 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4053 struct lnet_hdr *hdr = &msg->msg_hdr;
4054 struct lnet_process_id src = {0};
4055 struct lnet_libmd *md;
4058 src.nid = hdr->src_nid;
4059 src.pid = hdr->src_pid;
4061 /* Convert ack fields to host byte order */
4062 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4063 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4065 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4068 /* NB handles only looked up by creator (no flips) */
4069 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4070 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4071 /* Don't moan; this is expected */
4073 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4074 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4075 (md == NULL) ? "invalid" : "inactive",
4076 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4077 hdr->msg.ack.dst_wmd.wh_object_cookie);
4078 if (md != NULL && md->md_me != NULL)
4079 CERROR("Source MD also attached to portal %d\n",
4080 md->md_me->me_portal);
4082 lnet_res_unlock(cpt);
4083 return -ENOENT; /* -ve! */
4086 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4087 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4088 hdr->msg.ack.dst_wmd.wh_object_cookie);
4090 lnet_msg_attach_md(msg, md, 0, 0);
4092 lnet_res_unlock(cpt);
4094 lnet_build_msg_event(msg, LNET_EVENT_ACK);
4096 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4101 * \retval LNET_CREDIT_OK If \a msg is forwarded
4102 * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
4103 * \retval -ve error code
4106 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4110 if (!the_lnet.ln_routing)
4113 if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4114 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4115 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4116 msg->msg_rx_ready_delay = 1;
4118 lnet_net_unlock(msg->msg_rx_cpt);
4119 rc = lnet_ni_eager_recv(ni, msg);
4120 lnet_net_lock(msg->msg_rx_cpt);
4125 rc = lnet_post_routed_recv_locked(msg, 0);
4130 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4134 switch (msg->msg_type) {
4136 rc = lnet_parse_ack(ni, msg);
4139 rc = lnet_parse_put(ni, msg);
4142 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4144 case LNET_MSG_REPLY:
4145 rc = lnet_parse_reply(ni, msg);
4147 default: /* prevent an unused label if !kernel */
4152 LASSERT(rc == 0 || rc == -ENOENT);
4157 lnet_msgtyp2str (int type)
4166 case LNET_MSG_REPLY:
4168 case LNET_MSG_HELLO:
4171 return ("<UNKNOWN>");
4176 lnet_print_hdr(struct lnet_hdr *hdr)
4178 struct lnet_process_id src = {
4179 .nid = hdr->src_nid,
4180 .pid = hdr->src_pid,
4182 struct lnet_process_id dst = {
4183 .nid = hdr->dest_nid,
4184 .pid = hdr->dest_pid,
4186 char *type_str = lnet_msgtyp2str(hdr->type);
4188 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4189 CWARN(" From %s\n", libcfs_id2str(src));
4190 CWARN(" To %s\n", libcfs_id2str(dst));
4192 switch (hdr->type) {
4197 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
4198 "match bits %llu\n",
4199 hdr->msg.put.ptl_index,
4200 hdr->msg.put.ack_wmd.wh_interface_cookie,
4201 hdr->msg.put.ack_wmd.wh_object_cookie,
4202 hdr->msg.put.match_bits);
4203 CWARN(" Length %d, offset %d, hdr data %#llx\n",
4204 hdr->payload_length, hdr->msg.put.offset,
4205 hdr->msg.put.hdr_data);
4209 CWARN(" Ptl index %d, return md %#llx.%#llx, "
4210 "match bits %llu\n", hdr->msg.get.ptl_index,
4211 hdr->msg.get.return_wmd.wh_interface_cookie,
4212 hdr->msg.get.return_wmd.wh_object_cookie,
4213 hdr->msg.get.match_bits);
4214 CWARN(" Length %d, src offset %d\n",
4215 hdr->msg.get.sink_length,
4216 hdr->msg.get.src_offset);
4220 CWARN(" dst md %#llx.%#llx, "
4221 "manipulated length %d\n",
4222 hdr->msg.ack.dst_wmd.wh_interface_cookie,
4223 hdr->msg.ack.dst_wmd.wh_object_cookie,
4224 hdr->msg.ack.mlength);
4227 case LNET_MSG_REPLY:
4228 CWARN(" dst md %#llx.%#llx, "
4230 hdr->msg.reply.dst_wmd.wh_interface_cookie,
4231 hdr->msg.reply.dst_wmd.wh_object_cookie,
4232 hdr->payload_length);
4238 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4239 void *private, int rdma_req)
4241 struct lnet_peer_ni *lpni;
4242 struct lnet_msg *msg;
4243 __u32 payload_length;
4244 lnet_pid_t dest_pid;
4245 lnet_nid_t dest_nid;
4253 LASSERT (!in_interrupt ());
4255 type = le32_to_cpu(hdr->type);
4256 src_nid = le64_to_cpu(hdr->src_nid);
4257 dest_nid = le64_to_cpu(hdr->dest_nid);
4258 dest_pid = le32_to_cpu(hdr->dest_pid);
4259 payload_length = le32_to_cpu(hdr->payload_length);
4261 for_me = (ni->ni_nid == dest_nid);
4262 cpt = lnet_cpt_of_nid(from_nid, ni);
4264 CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4265 libcfs_nid2str(dest_nid),
4266 libcfs_nid2str(ni->ni_nid),
4267 libcfs_nid2str(src_nid),
4268 lnet_msgtyp2str(type),
4269 (for_me) ? "for me" : "routed");
4274 if (payload_length > 0) {
4275 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4276 libcfs_nid2str(from_nid),
4277 libcfs_nid2str(src_nid),
4278 lnet_msgtyp2str(type), payload_length);
4284 case LNET_MSG_REPLY:
4285 if (payload_length >
4286 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4287 CERROR("%s, src %s: bad %s payload %d "
4288 "(%d max expected)\n",
4289 libcfs_nid2str(from_nid),
4290 libcfs_nid2str(src_nid),
4291 lnet_msgtyp2str(type),
4293 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4299 CERROR("%s, src %s: Bad message type 0x%x\n",
4300 libcfs_nid2str(from_nid),
4301 libcfs_nid2str(src_nid), type);
4305 if (the_lnet.ln_routing &&
4306 ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4308 spin_lock(&ni->ni_net->net_lock);
4309 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4310 spin_unlock(&ni->ni_net->net_lock);
4311 if (ni->ni_status != NULL &&
4312 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4313 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4320 lnet_push_update_to_peers(1);
4322 /* Regard a bad destination NID as a protocol error. Senders should
4323 * know what they're doing; if they don't they're misconfigured, buggy
4324 * or malicious so we chop them off at the knees :) */
4327 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4328 /* should have gone direct */
4329 CERROR("%s, src %s: Bad dest nid %s "
4330 "(should have been sent direct)\n",
4331 libcfs_nid2str(from_nid),
4332 libcfs_nid2str(src_nid),
4333 libcfs_nid2str(dest_nid));
4337 if (lnet_islocalnid(dest_nid)) {
4338 /* dest is another local NI; sender should have used
4339 * this node's NID on its own network */
4340 CERROR("%s, src %s: Bad dest nid %s "
4341 "(it's my nid but on a different network)\n",
4342 libcfs_nid2str(from_nid),
4343 libcfs_nid2str(src_nid),
4344 libcfs_nid2str(dest_nid));
4348 if (rdma_req && type == LNET_MSG_GET) {
4349 CERROR("%s, src %s: Bad optimized GET for %s "
4350 "(final destination must be me)\n",
4351 libcfs_nid2str(from_nid),
4352 libcfs_nid2str(src_nid),
4353 libcfs_nid2str(dest_nid));
4357 if (!the_lnet.ln_routing) {
4358 CERROR("%s, src %s: Dropping message for %s "
4359 "(routing not enabled)\n",
4360 libcfs_nid2str(from_nid),
4361 libcfs_nid2str(src_nid),
4362 libcfs_nid2str(dest_nid));
4367 /* Message looks OK; we're not going to return an error, so we MUST
4368 * call back lnd_recv() come what may... */
4370 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4371 fail_peer(src_nid, 0)) { /* shall we now? */
4372 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4373 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4374 lnet_msgtyp2str(type));
4378 if (!list_empty(&the_lnet.ln_drop_rules) &&
4379 lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4381 "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4382 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4383 libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4387 if (lnet_drop_asym_route && for_me &&
4388 LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4389 struct lnet_net *net;
4390 struct lnet_remotenet *rnet;
4393 /* we are dealing with a routed message,
4394 * so see if route to reach src_nid goes through from_nid
4397 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4399 lnet_net_unlock(cpt);
4400 CERROR("net %s not found\n",
4401 libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4405 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4407 struct lnet_peer *gw = NULL;
4408 struct lnet_peer_ni *lpni = NULL;
4409 struct lnet_route *route;
4411 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4413 gw = route->lr_gateway;
4414 if (route->lr_lnet != net->net_id)
4417 * if the nid is one of the gateway's NIDs
4418 * then this is a valid gateway
4420 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4421 NULL, lpni)) != NULL) {
4422 if (lpni->lpni_nid == from_nid) {
4429 lnet_net_unlock(cpt);
4431 /* we would not use from_nid to route a message to
4433 * => asymmetric routing detected but forbidden
4435 CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4436 libcfs_nid2str(from_nid),
4437 libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4442 msg = lnet_msg_alloc();
4444 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4445 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4446 lnet_msgtyp2str(type));
4450 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4451 * pointers NULL etc */
4453 msg->msg_type = type;
4454 msg->msg_private = private;
4455 msg->msg_receiving = 1;
4456 msg->msg_rdma_get = rdma_req;
4457 msg->msg_len = msg->msg_wanted = payload_length;
4458 msg->msg_offset = 0;
4459 msg->msg_hdr = *hdr;
4460 /* for building message event */
4461 msg->msg_from = from_nid;
4463 msg->msg_target.pid = dest_pid;
4464 msg->msg_target.nid = dest_nid;
4465 msg->msg_routing = 1;
4468 /* convert common msg->hdr fields to host byteorder */
4469 msg->msg_hdr.type = type;
4470 msg->msg_hdr.src_nid = src_nid;
4471 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
4472 msg->msg_hdr.dest_nid = dest_nid;
4473 msg->msg_hdr.dest_pid = dest_pid;
4474 msg->msg_hdr.payload_length = payload_length;
4478 lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4480 lnet_net_unlock(cpt);
4481 CERROR("%s, src %s: Dropping %s "
4482 "(error %ld looking up sender)\n",
4483 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4484 lnet_msgtyp2str(type), PTR_ERR(lpni));
4486 if (rc == -ESHUTDOWN)
4487 /* We are shutting down. Don't do anything more */
4492 if (the_lnet.ln_routing)
4493 lpni->lpni_last_alive = ktime_get_seconds();
4495 msg->msg_rxpeer = lpni;
4497 lnet_ni_addref_locked(ni, cpt);
4498 /* Multi-Rail: Primary NID of source. */
4499 msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4502 * mark the status of this lpni as UP since we received a message
4503 * from it. The ping response reports back the ns_status which is
4504 * marked on the remote as up or down and we cache it here.
4506 msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4508 lnet_msg_commit(msg, cpt);
4510 /* message delay simulation */
4511 if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4512 lnet_delay_rule_match_locked(hdr, msg))) {
4513 lnet_net_unlock(cpt);
4518 rc = lnet_parse_forward_locked(ni, msg);
4519 lnet_net_unlock(cpt);
4524 if (rc == LNET_CREDIT_OK) {
4525 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4526 0, payload_length, payload_length);
4531 lnet_net_unlock(cpt);
4533 rc = lnet_parse_local(ni, msg);
4539 LASSERT(msg->msg_md == NULL);
4540 lnet_finalize(msg, rc);
4543 lnet_drop_message(ni, cpt, private, payload_length, type);
4546 EXPORT_SYMBOL(lnet_parse);
4549 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4551 while (!list_empty(head)) {
4552 struct lnet_process_id id = {0};
4553 struct lnet_msg *msg;
4555 msg = list_entry(head->next, struct lnet_msg, msg_list);
4556 list_del(&msg->msg_list);
4558 id.nid = msg->msg_hdr.src_nid;
4559 id.pid = msg->msg_hdr.src_pid;
4561 LASSERT(msg->msg_md == NULL);
4562 LASSERT(msg->msg_rx_delayed);
4563 LASSERT(msg->msg_rxpeer != NULL);
4564 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4566 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4567 " offset %d length %d: %s\n",
4569 msg->msg_hdr.msg.put.ptl_index,
4570 msg->msg_hdr.msg.put.match_bits,
4571 msg->msg_hdr.msg.put.offset,
4572 msg->msg_hdr.payload_length, reason);
4574 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4575 * called lnet_drop_message(), so I just hang onto msg as well
4576 * until that's done */
4578 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4579 msg->msg_private, msg->msg_len,
4582 msg->msg_no_resend = true;
4584 * NB: message will not generate event because w/o attached MD,
4585 * but we still should give error code so lnet_msg_decommit()
4586 * can skip counters operations and other checks.
4588 lnet_finalize(msg, -ENOENT);
4593 lnet_recv_delayed_msg_list(struct list_head *head)
4595 while (!list_empty(head)) {
4596 struct lnet_msg *msg;
4597 struct lnet_process_id id;
4599 msg = list_entry(head->next, struct lnet_msg, msg_list);
4600 list_del(&msg->msg_list);
4602 /* md won't disappear under me, since each msg
4603 * holds a ref on it */
4605 id.nid = msg->msg_hdr.src_nid;
4606 id.pid = msg->msg_hdr.src_pid;
4608 LASSERT(msg->msg_rx_delayed);
4609 LASSERT(msg->msg_md != NULL);
4610 LASSERT(msg->msg_rxpeer != NULL);
4611 LASSERT(msg->msg_rxni != NULL);
4612 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4614 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4615 "match %llu offset %d length %d.\n",
4616 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4617 msg->msg_hdr.msg.put.match_bits,
4618 msg->msg_hdr.msg.put.offset,
4619 msg->msg_hdr.payload_length);
4621 lnet_recv_put(msg->msg_rxni, msg);
4626 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4627 struct lnet_libmd *md, struct lnet_handle_md mdh)
4630 struct lnet_rsp_tracker *local_rspt;
4633 * MD has a refcount taken by message so it's not going away.
4634 * The MD however can be looked up. We need to secure the access
4635 * to the md_rspt_ptr by taking the res_lock.
4636 * The rspt can be accessed without protection up to when it gets
4637 * added to the list.
4641 local_rspt = md->md_rspt_ptr;
4642 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4643 if (local_rspt != NULL) {
4645 * we already have an rspt attached to the md, so we'll
4646 * update the deadline on that one.
4648 lnet_rspt_free(rspt, cpt);
4651 rspt->rspt_mdh = mdh;
4652 rspt->rspt_cpt = cpt;
4653 /* store the rspt so we can access it when we get the REPLY */
4654 md->md_rspt_ptr = rspt;
4657 local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4660 * add to the list of tracked responses. It's added to tail of the
4661 * list in order to expire all the older entries first.
4664 list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4665 lnet_net_unlock(cpt);
4666 lnet_res_unlock(cpt);
4670 * Initiate an asynchronous PUT operation.
4672 * There are several events associated with a PUT: completion of the send on
4673 * the initiator node (LNET_EVENT_SEND), and when the send completes
4674 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4675 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4676 * used at the target node to indicate the completion of incoming data
4679 * The local events will be logged in the EQ associated with the MD pointed to
4680 * by \a mdh handle. Using a MD without an associated EQ results in these
4681 * events being discarded. In this case, the caller must have another
4682 * mechanism (e.g., a higher level protocol) for determining when it is safe
4683 * to modify the memory region associated with the MD.
4685 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4686 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4688 * \param self Indicates the NID of a local interface through which to send
4689 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4690 * \param mdh A handle for the MD that describes the memory to be sent. The MD
4691 * must be "free floating" (See LNetMDBind()).
4692 * \param ack Controls whether an acknowledgment is requested.
4693 * Acknowledgments are only sent when they are requested by the initiating
4694 * process and the target MD enables them.
4695 * \param target A process identifier for the target process.
4696 * \param portal The index in the \a target's portal table.
4697 * \param match_bits The match bits to use for MD selection at the target
4699 * \param offset The offset into the target MD (only used when the target
4700 * MD has the LNET_MD_MANAGE_REMOTE option set).
4701 * \param hdr_data 64 bits of user data that can be included in the message
4702 * header. This data is written to an event queue entry at the target if an
4703 * EQ is present on the matching MD.
4705 * \retval 0 Success, and only in this case events will be generated
4706 * and logged to EQ (if it exists).
4707 * \retval -EIO Simulated failure.
4708 * \retval -ENOMEM Memory allocation failure.
4709 * \retval -ENOENT Invalid MD object.
4711 * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4714 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4715 struct lnet_process_id target, unsigned int portal,
4716 __u64 match_bits, unsigned int offset,
4719 struct lnet_msg *msg;
4720 struct lnet_libmd *md;
4723 struct lnet_rsp_tracker *rspt = NULL;
4725 LASSERT(the_lnet.ln_refcount > 0);
4727 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4728 fail_peer(target.nid, 1)) { /* shall we now? */
4729 CERROR("Dropping PUT to %s: simulated failure\n",
4730 libcfs_id2str(target));
4734 msg = lnet_msg_alloc();
4736 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4737 libcfs_id2str(target));
4740 msg->msg_vmflush = !!memory_pressure_get();
4742 cpt = lnet_cpt_of_cookie(mdh.cookie);
4744 if (ack == LNET_ACK_REQ) {
4745 rspt = lnet_rspt_alloc(cpt);
4747 CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4748 libcfs_id2str(target));
4751 INIT_LIST_HEAD(&rspt->rspt_on_list);
4756 md = lnet_handle2md(&mdh);
4757 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4758 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4759 match_bits, portal, libcfs_id2str(target),
4760 md == NULL ? -1 : md->md_threshold);
4761 if (md != NULL && md->md_me != NULL)
4762 CERROR("Source MD also attached to portal %d\n",
4763 md->md_me->me_portal);
4764 lnet_res_unlock(cpt);
4766 lnet_rspt_free(rspt, cpt);
4771 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4773 lnet_msg_attach_md(msg, md, 0, 0);
4775 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4777 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4778 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4779 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4780 msg->msg_hdr.msg.put.hdr_data = hdr_data;
4782 /* NB handles only looked up by creator (no flips) */
4783 if (ack == LNET_ACK_REQ) {
4784 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4785 the_lnet.ln_interface_cookie;
4786 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4787 md->md_lh.lh_cookie;
4789 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4790 LNET_WIRE_HANDLE_COOKIE_NONE;
4791 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4792 LNET_WIRE_HANDLE_COOKIE_NONE;
4795 lnet_res_unlock(cpt);
4797 lnet_build_msg_event(msg, LNET_EVENT_SEND);
4799 if (ack == LNET_ACK_REQ)
4800 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4802 if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4806 rc = lnet_send(self, msg, LNET_NID_ANY);
4809 CNETERR("Error sending PUT to %s: %d\n",
4810 libcfs_id2str(target), rc);
4811 msg->msg_no_resend = true;
4812 lnet_finalize(msg, rc);
4815 /* completion will be signalled by an event */
4818 EXPORT_SYMBOL(LNetPut);
4821 * The LND can DMA direct to the GET md (i.e. no REPLY msg). This
4822 * returns a msg for the LND to pass to lnet_finalize() when the sink
4823 * data has been received.
4825 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4826 * lnet_finalize() is called on it, so the LND must call this first
4829 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4831 struct lnet_msg *msg = lnet_msg_alloc();
4832 struct lnet_libmd *getmd = getmsg->msg_md;
4833 struct lnet_process_id peer_id = getmsg->msg_target;
4836 LASSERT(!getmsg->msg_target_is_router);
4837 LASSERT(!getmsg->msg_routing);
4840 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4841 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4845 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4848 LASSERT(getmd->md_refcount > 0);
4850 if (getmd->md_threshold == 0) {
4851 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4852 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4854 lnet_res_unlock(cpt);
4858 LASSERT(getmd->md_offset == 0);
4860 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4861 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4863 /* setup information for lnet_build_msg_event */
4864 msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4865 msg->msg_from = peer_id.nid;
4866 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4867 msg->msg_hdr.src_nid = peer_id.nid;
4868 msg->msg_hdr.payload_length = getmd->md_length;
4869 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4871 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4872 lnet_res_unlock(cpt);
4874 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4877 lnet_msg_commit(msg, cpt);
4878 lnet_net_unlock(cpt);
4880 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4885 cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4888 lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4889 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4890 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4892 lnet_net_unlock(cpt);
4899 EXPORT_SYMBOL(lnet_create_reply_msg);
4902 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4905 /* Set the REPLY length, now the RDMA that elides the REPLY message has
4906 * completed and I know it. */
4907 LASSERT(reply != NULL);
4908 LASSERT(reply->msg_type == LNET_MSG_GET);
4909 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4911 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
4912 * the end of my buffer, I might as well be dead. */
4913 LASSERT(len <= reply->msg_ev.mlength);
4915 reply->msg_ev.mlength = len;
4917 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4920 * Initiate an asynchronous GET operation.
4922 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4923 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4924 * the target node in the REPLY has been written to local MD.
4926 * On the target node, an LNET_EVENT_GET is logged when the GET request
4927 * arrives and is accepted into a MD.
4929 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4930 * \param mdh A handle for the MD that describes the memory into which the
4931 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4933 * \retval 0 Success, and only in this case events will be generated
4934 * and logged to EQ (if it exists) of the MD.
4935 * \retval -EIO Simulated failure.
4936 * \retval -ENOMEM Memory allocation failure.
4937 * \retval -ENOENT Invalid MD object.
4940 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4941 struct lnet_process_id target, unsigned int portal,
4942 __u64 match_bits, unsigned int offset, bool recovery)
4944 struct lnet_msg *msg;
4945 struct lnet_libmd *md;
4946 struct lnet_rsp_tracker *rspt;
4950 LASSERT(the_lnet.ln_refcount > 0);
4952 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
4953 fail_peer(target.nid, 1)) /* shall we now? */
4955 CERROR("Dropping GET to %s: simulated failure\n",
4956 libcfs_id2str(target));
4960 msg = lnet_msg_alloc();
4962 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4963 libcfs_id2str(target));
4967 cpt = lnet_cpt_of_cookie(mdh.cookie);
4969 rspt = lnet_rspt_alloc(cpt);
4971 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4972 libcfs_id2str(target));
4975 INIT_LIST_HEAD(&rspt->rspt_on_list);
4977 msg->msg_recovery = recovery;
4981 md = lnet_handle2md(&mdh);
4982 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4983 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4984 match_bits, portal, libcfs_id2str(target),
4985 md == NULL ? -1 : md->md_threshold);
4986 if (md != NULL && md->md_me != NULL)
4987 CERROR("REPLY MD also attached to portal %d\n",
4988 md->md_me->me_portal);
4990 lnet_res_unlock(cpt);
4993 lnet_rspt_free(rspt, cpt);
4997 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4999 lnet_msg_attach_md(msg, md, 0, 0);
5001 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5003 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5004 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5005 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5006 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5008 /* NB handles only looked up by creator (no flips) */
5009 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5010 the_lnet.ln_interface_cookie;
5011 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5012 md->md_lh.lh_cookie;
5014 lnet_res_unlock(cpt);
5016 lnet_build_msg_event(msg, LNET_EVENT_SEND);
5018 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5020 rc = lnet_send(self, msg, LNET_NID_ANY);
5022 CNETERR("Error sending GET to %s: %d\n",
5023 libcfs_id2str(target), rc);
5024 msg->msg_no_resend = true;
5025 lnet_finalize(msg, rc);
5028 /* completion will be signalled by an event */
5031 EXPORT_SYMBOL(LNetGet);
5034 * Calculate distance to node at \a dstnid.
5036 * \param dstnid Target NID.
5037 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5039 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5042 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5043 * local_nid_dist_zero is set, which is the default.
5044 * \retval positives Distance to target NID, i.e. number of hops plus one.
5045 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5048 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5050 struct list_head *e;
5051 struct lnet_ni *ni = NULL;
5052 struct lnet_remotenet *rnet;
5053 __u32 dstnet = LNET_NIDNET(dstnid);
5057 struct list_head *rn_list;
5059 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5060 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5061 * keep order 0 free for 0@lo and order 1 free for a local NID
5064 LASSERT(the_lnet.ln_refcount > 0);
5066 cpt = lnet_net_lock_current();
5068 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5069 if (ni->ni_nid == dstnid) {
5070 if (srcnidp != NULL)
5072 if (orderp != NULL) {
5073 if (dstnid == LNET_NID_LO_0)
5078 lnet_net_unlock(cpt);
5080 return local_nid_dist_zero ? 0 : 1;
5083 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5084 /* Check if ni was originally created in
5085 * current net namespace.
5086 * If not, assign order above 0xffff0000,
5087 * to make this ni not a priority. */
5088 if (current->nsproxy &&
5089 !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5090 order += 0xffff0000;
5091 if (srcnidp != NULL)
5092 *srcnidp = ni->ni_nid;
5095 lnet_net_unlock(cpt);
5102 rn_list = lnet_net2rnethash(dstnet);
5103 list_for_each(e, rn_list) {
5104 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5106 if (rnet->lrn_net == dstnet) {
5107 struct lnet_route *route;
5108 struct lnet_route *shortest = NULL;
5109 __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5112 LASSERT(!list_empty(&rnet->lrn_routes));
5114 list_for_each_entry(route, &rnet->lrn_routes,
5116 route_hops = route->lr_hops;
5117 if (route_hops == LNET_UNDEFINED_HOPS)
5119 if (shortest == NULL ||
5120 route_hops < shortest_hops) {
5122 shortest_hops = route_hops;
5126 LASSERT(shortest != NULL);
5127 hops = shortest_hops;
5128 if (srcnidp != NULL) {
5129 struct lnet_net *net;
5130 net = lnet_get_net_locked(shortest->lr_lnet);
5132 ni = lnet_get_next_ni_locked(net, NULL);
5133 *srcnidp = ni->ni_nid;
5137 lnet_net_unlock(cpt);
5143 lnet_net_unlock(cpt);
5144 return -EHOSTUNREACH;
5146 EXPORT_SYMBOL(LNetDist);