1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/lnet/lib-move.c
38 * Data movement routines
41 #define DEBUG_SUBSYSTEM S_LNET
43 #include <lnet/lib-lnet.h>
45 static int local_nid_dist_zero = 1;
46 CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444,
50 static void lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg);
52 #define LNET_MATCHMD_NONE 0 /* Didn't match */
53 #define LNET_MATCHMD_OK 1 /* Matched OK */
54 #define LNET_MATCHMD_DROP 2 /* Must be discarded */
57 lnet_try_match_md (int index, int op_mask, lnet_process_id_t src,
58 unsigned int rlength, unsigned int roffset,
59 __u64 match_bits, lnet_libmd_t *md, lnet_msg_t *msg,
60 unsigned int *mlength_out, unsigned int *offset_out)
62 /* ALWAYS called holding the LNET_LOCK, and can't LNET_UNLOCK;
63 * lnet_match_blocked_msg() relies on this to avoid races */
66 lnet_me_t *me = md->md_me;
68 /* mismatched MD op */
69 if ((md->md_options & op_mask) == 0)
70 return LNET_MATCHMD_NONE;
73 if (lnet_md_exhausted(md))
74 return LNET_MATCHMD_NONE;
76 /* mismatched ME nid/pid? */
77 if (me->me_match_id.nid != LNET_NID_ANY &&
78 me->me_match_id.nid != src.nid)
79 return LNET_MATCHMD_NONE;
81 if (me->me_match_id.pid != LNET_PID_ANY &&
82 me->me_match_id.pid != src.pid)
83 return LNET_MATCHMD_NONE;
85 /* mismatched ME matchbits? */
86 if (((me->me_match_bits ^ match_bits) & ~me->me_ignore_bits) != 0)
87 return LNET_MATCHMD_NONE;
89 /* Hurrah! This _is_ a match; check it out... */
91 if ((md->md_options & LNET_MD_MANAGE_REMOTE) == 0)
92 offset = md->md_offset;
96 if ((md->md_options & LNET_MD_MAX_SIZE) != 0) {
97 mlength = md->md_max_size;
98 LASSERT (md->md_offset + mlength <= md->md_length);
100 mlength = md->md_length - offset;
103 if (rlength <= mlength) { /* fits in allowed space */
105 } else if ((md->md_options & LNET_MD_TRUNCATE) == 0) {
106 /* this packet _really_ is too big */
107 CERROR("Matching packet from %s, match "LPU64
108 " length %d too big: %d left, %d allowed\n",
109 libcfs_id2str(src), match_bits, rlength,
110 md->md_length - offset, mlength);
112 return LNET_MATCHMD_DROP;
115 /* Commit to this ME/MD */
116 CDEBUG(D_NET, "Incoming %s index %x from %s of "
117 "length %d/%d into md "LPX64" [%d] + %d\n",
118 (op_mask == LNET_MD_OP_PUT) ? "put" : "get",
119 index, libcfs_id2str(src), mlength, rlength,
120 md->md_lh.lh_cookie, md->md_niov, offset);
122 lnet_commit_md(md, msg);
123 md->md_offset = offset + mlength;
125 /* NB Caller will set ev.type and ev.hdr_data */
126 msg->msg_ev.initiator = src;
127 msg->msg_ev.pt_index = index;
128 msg->msg_ev.match_bits = match_bits;
129 msg->msg_ev.rlength = rlength;
130 msg->msg_ev.mlength = mlength;
131 msg->msg_ev.offset = offset;
133 lnet_md_deconstruct(md, &msg->msg_ev.md);
134 lnet_md2handle(&msg->msg_ev.md_handle, md);
136 *offset_out = offset;
137 *mlength_out = mlength;
139 /* Auto-unlink NOW, so the ME gets unlinked if required.
140 * We bumped md->md_refcount above so the MD just gets flagged
141 * for unlink when it is finalized. */
142 if ((md->md_flags & LNET_MD_FLAG_AUTO_UNLINK) != 0 &&
143 lnet_md_exhausted(md)) {
147 return LNET_MATCHMD_OK;
151 lnet_match_md(int index, int op_mask, lnet_process_id_t src,
152 unsigned int rlength, unsigned int roffset,
153 __u64 match_bits, lnet_msg_t *msg,
154 unsigned int *mlength_out, unsigned int *offset_out,
155 lnet_libmd_t **md_out)
157 lnet_portal_t *ptl = &the_lnet.ln_portals[index];
164 CDEBUG (D_NET, "Request from %s of length %d into portal %d "
165 "MB="LPX64"\n", libcfs_id2str(src), rlength, index, match_bits);
167 if (index < 0 || index >= the_lnet.ln_nportals) {
168 CERROR("Invalid portal %d not in [0-%d]\n",
169 index, the_lnet.ln_nportals);
170 return LNET_MATCHMD_DROP;
173 head = lnet_portal_me_head(index, src, match_bits);
174 if (head == NULL) /* nobody posted anything on this portal */
177 cfs_list_for_each_entry_safe_typed (me, tmp, head,
178 lnet_me_t, me_list) {
181 /* ME attached but MD not attached yet */
185 LASSERT (me == md->md_me);
187 rc = lnet_try_match_md(index, op_mask, src, rlength,
188 roffset, match_bits, md, msg,
189 mlength_out, offset_out);
194 case LNET_MATCHMD_NONE:
197 case LNET_MATCHMD_OK:
199 return LNET_MATCHMD_OK;
201 case LNET_MATCHMD_DROP:
202 return LNET_MATCHMD_DROP;
208 if (op_mask == LNET_MD_OP_GET ||
209 !lnet_portal_is_lazy(ptl))
210 return LNET_MATCHMD_DROP;
212 return LNET_MATCHMD_NONE;
216 lnet_fail_nid (lnet_nid_t nid, unsigned int threshold)
218 lnet_test_peer_t *tp;
223 LASSERT (the_lnet.ln_init);
225 if (threshold != 0) {
226 /* Adding a new entry */
227 LIBCFS_ALLOC(tp, sizeof(*tp));
232 tp->tp_threshold = threshold;
235 cfs_list_add_tail (&tp->tp_list, &the_lnet.ln_test_peers);
240 /* removing entries */
241 CFS_INIT_LIST_HEAD (&cull);
245 cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
246 tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
248 if (tp->tp_threshold == 0 || /* needs culling anyway */
249 nid == LNET_NID_ANY || /* removing all entries */
250 tp->tp_nid == nid) /* matched this one */
252 cfs_list_del (&tp->tp_list);
253 cfs_list_add (&tp->tp_list, &cull);
259 while (!cfs_list_empty (&cull)) {
260 tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
262 cfs_list_del (&tp->tp_list);
263 LIBCFS_FREE(tp, sizeof (*tp));
269 fail_peer (lnet_nid_t nid, int outgoing)
271 lnet_test_peer_t *tp;
277 CFS_INIT_LIST_HEAD (&cull);
281 cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
282 tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
284 if (tp->tp_threshold == 0) {
287 /* only cull zombies on outgoing tests,
288 * since we may be at interrupt priority on
289 * incoming messages. */
290 cfs_list_del (&tp->tp_list);
291 cfs_list_add (&tp->tp_list, &cull);
296 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
297 nid == tp->tp_nid) { /* fail this peer */
300 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
303 tp->tp_threshold == 0) {
305 cfs_list_del (&tp->tp_list);
306 cfs_list_add (&tp->tp_list, &cull);
315 while (!cfs_list_empty (&cull)) {
316 tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
317 cfs_list_del (&tp->tp_list);
319 LIBCFS_FREE(tp, sizeof (*tp));
326 lnet_iov_nob (unsigned int niov, struct iovec *iov)
328 unsigned int nob = 0;
331 nob += (iov++)->iov_len;
337 lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
338 unsigned int nsiov, struct iovec *siov, unsigned int soffset,
341 /* NB diov, siov are READ-ONLY */
342 unsigned int this_nob;
347 /* skip complete frags before 'doffset' */
349 while (doffset >= diov->iov_len) {
350 doffset -= diov->iov_len;
356 /* skip complete frags before 'soffset' */
358 while (soffset >= siov->iov_len) {
359 soffset -= siov->iov_len;
368 this_nob = MIN(diov->iov_len - doffset,
369 siov->iov_len - soffset);
370 this_nob = MIN(this_nob, nob);
372 memcpy ((char *)diov->iov_base + doffset,
373 (char *)siov->iov_base + soffset, this_nob);
376 if (diov->iov_len > doffset + this_nob) {
384 if (siov->iov_len > soffset + this_nob) {
395 lnet_extract_iov (int dst_niov, struct iovec *dst,
396 int src_niov, struct iovec *src,
397 unsigned int offset, unsigned int len)
399 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
400 * for exactly 'len' bytes, and return the number of entries.
401 * NB not destructive to 'src' */
402 unsigned int frag_len;
405 if (len == 0) /* no data => */
406 return (0); /* no frags */
408 LASSERT (src_niov > 0);
409 while (offset >= src->iov_len) { /* skip initial frags */
410 offset -= src->iov_len;
413 LASSERT (src_niov > 0);
418 LASSERT (src_niov > 0);
419 LASSERT ((int)niov <= dst_niov);
421 frag_len = src->iov_len - offset;
422 dst->iov_base = ((char *)src->iov_base) + offset;
424 if (len <= frag_len) {
429 dst->iov_len = frag_len;
442 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
449 lnet_copy_kiov2kiov (unsigned int ndkiov, lnet_kiov_t *dkiov, unsigned int doffset,
450 unsigned int nskiov, lnet_kiov_t *skiov, unsigned int soffset,
457 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
458 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
465 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
466 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
473 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
474 int src_niov, lnet_kiov_t *src,
475 unsigned int offset, unsigned int len)
480 #else /* __KERNEL__ */
483 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
485 unsigned int nob = 0;
488 nob += (kiov++)->kiov_len;
494 lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
495 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
498 /* NB diov, siov are READ-ONLY */
499 unsigned int this_nob;
506 LASSERT (!cfs_in_interrupt ());
509 while (doffset >= diov->kiov_len) {
510 doffset -= diov->kiov_len;
517 while (soffset >= siov->kiov_len) {
518 soffset -= siov->kiov_len;
527 this_nob = MIN(diov->kiov_len - doffset,
528 siov->kiov_len - soffset);
529 this_nob = MIN(this_nob, nob);
532 daddr = ((char *)cfs_kmap(diov->kiov_page)) +
533 diov->kiov_offset + doffset;
535 saddr = ((char *)cfs_kmap(siov->kiov_page)) +
536 siov->kiov_offset + soffset;
538 /* Vanishing risk of kmap deadlock when mapping 2 pages.
539 * However in practice at least one of the kiovs will be mapped
540 * kernel pages and the map/unmap will be NOOPs */
542 memcpy (daddr, saddr, this_nob);
545 if (diov->kiov_len > doffset + this_nob) {
549 cfs_kunmap(diov->kiov_page);
556 if (siov->kiov_len > soffset + this_nob) {
560 cfs_kunmap(siov->kiov_page);
569 cfs_kunmap(diov->kiov_page);
571 cfs_kunmap(siov->kiov_page);
575 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
576 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
579 /* NB iov, kiov are READ-ONLY */
580 unsigned int this_nob;
586 LASSERT (!cfs_in_interrupt ());
589 while (iovoffset >= iov->iov_len) {
590 iovoffset -= iov->iov_len;
597 while (kiovoffset >= kiov->kiov_len) {
598 kiovoffset -= kiov->kiov_len;
607 this_nob = MIN(iov->iov_len - iovoffset,
608 kiov->kiov_len - kiovoffset);
609 this_nob = MIN(this_nob, nob);
612 addr = ((char *)cfs_kmap(kiov->kiov_page)) +
613 kiov->kiov_offset + kiovoffset;
615 memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
618 if (iov->iov_len > iovoffset + this_nob) {
619 iovoffset += this_nob;
626 if (kiov->kiov_len > kiovoffset + this_nob) {
628 kiovoffset += this_nob;
630 cfs_kunmap(kiov->kiov_page);
640 cfs_kunmap(kiov->kiov_page);
644 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
645 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
648 /* NB kiov, iov are READ-ONLY */
649 unsigned int this_nob;
655 LASSERT (!cfs_in_interrupt ());
658 while (kiovoffset >= kiov->kiov_len) {
659 kiovoffset -= kiov->kiov_len;
666 while (iovoffset >= iov->iov_len) {
667 iovoffset -= iov->iov_len;
676 this_nob = MIN(kiov->kiov_len - kiovoffset,
677 iov->iov_len - iovoffset);
678 this_nob = MIN(this_nob, nob);
681 addr = ((char *)cfs_kmap(kiov->kiov_page)) +
682 kiov->kiov_offset + kiovoffset;
684 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
687 if (kiov->kiov_len > kiovoffset + this_nob) {
689 kiovoffset += this_nob;
691 cfs_kunmap(kiov->kiov_page);
698 if (iov->iov_len > iovoffset + this_nob) {
699 iovoffset += this_nob;
708 cfs_kunmap(kiov->kiov_page);
712 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
713 int src_niov, lnet_kiov_t *src,
714 unsigned int offset, unsigned int len)
716 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
717 * for exactly 'len' bytes, and return the number of entries.
718 * NB not destructive to 'src' */
719 unsigned int frag_len;
722 if (len == 0) /* no data => */
723 return (0); /* no frags */
725 LASSERT (src_niov > 0);
726 while (offset >= src->kiov_len) { /* skip initial frags */
727 offset -= src->kiov_len;
730 LASSERT (src_niov > 0);
735 LASSERT (src_niov > 0);
736 LASSERT ((int)niov <= dst_niov);
738 frag_len = src->kiov_len - offset;
739 dst->kiov_page = src->kiov_page;
740 dst->kiov_offset = src->kiov_offset + offset;
742 if (len <= frag_len) {
744 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
748 dst->kiov_len = frag_len;
749 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
762 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
763 unsigned int offset, unsigned int mlen, unsigned int rlen)
765 unsigned int niov = 0;
766 struct iovec *iov = NULL;
767 lnet_kiov_t *kiov = NULL;
770 LASSERT (!cfs_in_interrupt ());
771 LASSERT (mlen == 0 || msg != NULL);
774 LASSERT(msg->msg_receiving);
775 LASSERT(!msg->msg_sending);
776 LASSERT(rlen == msg->msg_len);
777 LASSERT(mlen <= msg->msg_len);
779 msg->msg_wanted = mlen;
780 msg->msg_offset = offset;
781 msg->msg_receiving = 0;
784 niov = msg->msg_niov;
786 kiov = msg->msg_kiov;
789 LASSERT ((iov == NULL) != (kiov == NULL));
793 rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
794 niov, iov, kiov, offset, mlen, rlen);
796 lnet_finalize(ni, msg, rc);
800 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
802 lnet_peer_t *p1 = r1->lr_gateway;
803 lnet_peer_t *p2 = r2->lr_gateway;
805 if (r1->lr_hops < r2->lr_hops)
808 if (r1->lr_hops > r2->lr_hops)
811 if (p1->lp_txqnob < p2->lp_txqnob)
814 if (p1->lp_txqnob > p2->lp_txqnob)
817 if (p1->lp_txcredits > p2->lp_txcredits)
820 if (p1->lp_txcredits < p2->lp_txcredits)
828 lnet_setpayloadbuffer(lnet_msg_t *msg)
830 lnet_libmd_t *md = msg->msg_md;
832 LASSERT (msg->msg_len > 0);
833 LASSERT (!msg->msg_routing);
834 LASSERT (md != NULL);
835 LASSERT (msg->msg_niov == 0);
836 LASSERT (msg->msg_iov == NULL);
837 LASSERT (msg->msg_kiov == NULL);
839 msg->msg_niov = md->md_niov;
840 if ((md->md_options & LNET_MD_KIOV) != 0)
841 msg->msg_kiov = md->md_iov.kiov;
843 msg->msg_iov = md->md_iov.iov;
847 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
848 unsigned int offset, unsigned int len)
850 msg->msg_type = type;
851 msg->msg_target = target;
853 msg->msg_offset = offset;
856 lnet_setpayloadbuffer(msg);
858 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
859 msg->msg_hdr.type = cpu_to_le32(type);
860 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
861 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
862 /* src_nid will be set later */
863 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
864 msg->msg_hdr.payload_length = cpu_to_le32(len);
868 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
870 void *priv = msg->msg_private;
873 LASSERT (!cfs_in_interrupt ());
874 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
875 (msg->msg_txcredit && msg->msg_peertxcredit));
877 rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
879 lnet_finalize(ni, msg, rc);
883 lnet_eager_recv_locked(lnet_msg_t *msg)
889 LASSERT (!msg->msg_delayed);
890 msg->msg_delayed = 1;
892 LASSERT (msg->msg_receiving);
893 LASSERT (!msg->msg_sending);
895 peer = msg->msg_rxpeer;
898 if (ni->ni_lnd->lnd_eager_recv != NULL) {
901 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
904 CERROR("recv from %s / send to %s aborted: "
905 "eager_recv failed %d\n",
906 libcfs_nid2str(peer->lp_nid),
907 libcfs_id2str(msg->msg_target), rc);
908 LASSERT (rc < 0); /* required by my callers */
917 /* NB: caller shall hold a ref on 'lp' as I'd drop LNET_LOCK */
919 lnet_ni_peer_alive(lnet_peer_t *lp)
921 cfs_time_t last_alive = 0;
922 lnet_ni_t *ni = lp->lp_ni;
924 LASSERT (lnet_peer_aliveness_enabled(lp));
925 LASSERT (ni->ni_lnd->lnd_query != NULL);
928 (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
931 lp->lp_last_query = cfs_time_current();
933 if (last_alive != 0) /* NI has updated timestamp */
934 lp->lp_last_alive = last_alive;
938 /* NB: always called with LNET_LOCK held */
940 lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
945 LASSERT (lnet_peer_aliveness_enabled(lp));
947 /* Trust lnet_notify() if it has more recent aliveness news, but
948 * ignore the initial assumed death (see lnet_peers_start_down()).
950 if (!lp->lp_alive && lp->lp_alive_count > 0 &&
951 cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive))
954 deadline = cfs_time_add(lp->lp_last_alive,
955 cfs_time_seconds(lp->lp_ni->ni_peertimeout));
956 alive = cfs_time_after(deadline, now);
958 /* Update obsolete lp_alive except for routers assumed to be dead
959 * initially, because router checker would update aliveness in this
960 * case, and moreover lp_last_alive at peer creation is assumed.
962 if (alive && !lp->lp_alive &&
963 !(lnet_isrouter(lp) && lp->lp_alive_count == 0))
964 lnet_notify_locked(lp, 0, 1, lp->lp_last_alive);
970 /* NB: returns 1 when alive, 0 when dead, negative when error;
971 * may drop the LNET_LOCK */
973 lnet_peer_alive_locked (lnet_peer_t *lp)
975 cfs_time_t now = cfs_time_current();
977 if (!lnet_peer_aliveness_enabled(lp))
980 if (lnet_peer_is_alive(lp, now))
983 /* Peer appears dead, but we should avoid frequent NI queries (at
984 * most once per lnet_queryinterval seconds). */
985 if (lp->lp_last_query != 0) {
986 static const int lnet_queryinterval = 1;
988 cfs_time_t next_query =
989 cfs_time_add(lp->lp_last_query,
990 cfs_time_seconds(lnet_queryinterval));
992 if (cfs_time_before(now, next_query)) {
994 CWARN("Unexpected aliveness of peer %s: "
996 libcfs_nid2str(lp->lp_nid),
997 (int)now, (int)next_query,
999 lp->lp_ni->ni_peertimeout);
1004 /* query NI for latest aliveness news */
1005 lnet_ni_peer_alive(lp);
1007 if (lnet_peer_is_alive(lp, now))
1010 lnet_notify_locked(lp, 0, 0, lp->lp_last_alive);
1015 lnet_post_send_locked (lnet_msg_t *msg, int do_send)
1017 /* lnet_send is going to LNET_UNLOCK immediately after this, so it sets
1018 * do_send FALSE and I don't do the unlock/send/lock bit. I return
1019 * EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer appears dead, and
1020 * 0 if sent or OK to send */
1021 lnet_peer_t *lp = msg->msg_txpeer;
1022 lnet_ni_t *ni = lp->lp_ni;
1024 /* non-lnet_send() callers have checked before */
1025 LASSERT (!do_send || msg->msg_delayed);
1026 LASSERT (!msg->msg_receiving);
1028 /* NB 'lp' is always the next hop */
1029 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
1030 lnet_peer_alive_locked(lp) == 0) {
1031 the_lnet.ln_counters.drop_count++;
1032 the_lnet.ln_counters.drop_length += msg->msg_len;
1035 CNETERR("Dropping message for %s: peer not alive\n",
1036 libcfs_id2str(msg->msg_target));
1038 lnet_finalize(ni, msg, -EHOSTUNREACH);
1041 return EHOSTUNREACH;
1044 if (!msg->msg_peertxcredit) {
1045 LASSERT ((lp->lp_txcredits < 0) ==
1046 !cfs_list_empty(&lp->lp_txq));
1048 msg->msg_peertxcredit = 1;
1049 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
1052 if (lp->lp_txcredits < lp->lp_mintxcredits)
1053 lp->lp_mintxcredits = lp->lp_txcredits;
1055 if (lp->lp_txcredits < 0) {
1056 msg->msg_delayed = 1;
1057 cfs_list_add_tail(&msg->msg_list, &lp->lp_txq);
1062 if (!msg->msg_txcredit) {
1063 LASSERT ((ni->ni_txcredits < 0) ==
1064 !cfs_list_empty(&ni->ni_txq));
1066 msg->msg_txcredit = 1;
1069 if (ni->ni_txcredits < ni->ni_mintxcredits)
1070 ni->ni_mintxcredits = ni->ni_txcredits;
1072 if (ni->ni_txcredits < 0) {
1073 msg->msg_delayed = 1;
1074 cfs_list_add_tail(&msg->msg_list, &ni->ni_txq);
1081 lnet_ni_send(ni, msg);
1089 lnet_commit_routedmsg (lnet_msg_t *msg)
1091 /* ALWAYS called holding the LNET_LOCK */
1092 LASSERT (msg->msg_routing);
1094 the_lnet.ln_counters.msgs_alloc++;
1095 if (the_lnet.ln_counters.msgs_alloc >
1096 the_lnet.ln_counters.msgs_max)
1097 the_lnet.ln_counters.msgs_max =
1098 the_lnet.ln_counters.msgs_alloc;
1100 the_lnet.ln_counters.route_count++;
1101 the_lnet.ln_counters.route_length += msg->msg_len;
1103 LASSERT (!msg->msg_onactivelist);
1104 msg->msg_onactivelist = 1;
1105 cfs_list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
1109 lnet_msg2bufpool(lnet_msg_t *msg)
1111 lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[0];
1113 LASSERT (msg->msg_len <= LNET_MTU);
1114 while (msg->msg_len > (unsigned int)rbp->rbp_npages * CFS_PAGE_SIZE) {
1116 LASSERT (rbp < &the_lnet.ln_rtrpools[LNET_NRBPOOLS]);
1123 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
1125 /* lnet_parse is going to LNET_UNLOCK immediately after this, so it
1126 * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
1127 * return EAGAIN if msg blocked and 0 if received or OK to receive */
1128 lnet_peer_t *lp = msg->msg_rxpeer;
1129 lnet_rtrbufpool_t *rbp;
1132 LASSERT (msg->msg_iov == NULL);
1133 LASSERT (msg->msg_kiov == NULL);
1134 LASSERT (msg->msg_niov == 0);
1135 LASSERT (msg->msg_routing);
1136 LASSERT (msg->msg_receiving);
1137 LASSERT (!msg->msg_sending);
1139 /* non-lnet_parse callers only send delayed messages */
1140 LASSERT (!do_recv || msg->msg_delayed);
1142 if (!msg->msg_peerrtrcredit) {
1143 LASSERT ((lp->lp_rtrcredits < 0) ==
1144 !cfs_list_empty(&lp->lp_rtrq));
1146 msg->msg_peerrtrcredit = 1;
1147 lp->lp_rtrcredits--;
1148 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
1149 lp->lp_minrtrcredits = lp->lp_rtrcredits;
1151 if (lp->lp_rtrcredits < 0) {
1152 /* must have checked eager_recv before here */
1153 LASSERT (msg->msg_delayed);
1154 cfs_list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1159 rbp = lnet_msg2bufpool(msg);
1161 if (!msg->msg_rtrcredit) {
1162 LASSERT ((rbp->rbp_credits < 0) ==
1163 !cfs_list_empty(&rbp->rbp_msgs));
1165 msg->msg_rtrcredit = 1;
1167 if (rbp->rbp_credits < rbp->rbp_mincredits)
1168 rbp->rbp_mincredits = rbp->rbp_credits;
1170 if (rbp->rbp_credits < 0) {
1171 /* must have checked eager_recv before here */
1172 LASSERT (msg->msg_delayed);
1173 cfs_list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1178 LASSERT (!cfs_list_empty(&rbp->rbp_bufs));
1179 rb = cfs_list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
1180 cfs_list_del(&rb->rb_list);
1182 msg->msg_niov = rbp->rbp_npages;
1183 msg->msg_kiov = &rb->rb_kiov[0];
1187 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
1188 0, msg->msg_len, msg->msg_len);
1196 lnet_return_credits_locked (lnet_msg_t *msg)
1198 lnet_peer_t *txpeer = msg->msg_txpeer;
1199 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1203 if (msg->msg_txcredit) {
1204 /* give back NI txcredits */
1205 msg->msg_txcredit = 0;
1208 LASSERT((ni->ni_txcredits < 0) == !cfs_list_empty(&ni->ni_txq));
1211 if (ni->ni_txcredits <= 0) {
1212 msg2 = cfs_list_entry(ni->ni_txq.next, lnet_msg_t,
1214 cfs_list_del(&msg2->msg_list);
1216 LASSERT(msg2->msg_txpeer->lp_ni == ni);
1217 LASSERT(msg2->msg_delayed);
1219 (void) lnet_post_send_locked(msg2, 1);
1223 if (msg->msg_peertxcredit) {
1224 /* give back peer txcredits */
1225 msg->msg_peertxcredit = 0;
1227 LASSERT((txpeer->lp_txcredits < 0) ==
1228 !cfs_list_empty(&txpeer->lp_txq));
1230 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1231 LASSERT (txpeer->lp_txqnob >= 0);
1233 txpeer->lp_txcredits++;
1234 if (txpeer->lp_txcredits <= 0) {
1235 msg2 = cfs_list_entry(txpeer->lp_txq.next,
1236 lnet_msg_t, msg_list);
1237 cfs_list_del(&msg2->msg_list);
1239 LASSERT (msg2->msg_txpeer == txpeer);
1240 LASSERT (msg2->msg_delayed);
1242 (void) lnet_post_send_locked(msg2, 1);
1246 if (txpeer != NULL) {
1247 msg->msg_txpeer = NULL;
1248 lnet_peer_decref_locked(txpeer);
1252 if (msg->msg_rtrcredit) {
1253 /* give back global router credits */
1255 lnet_rtrbufpool_t *rbp;
1257 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1258 * there until it gets one allocated, or aborts the wait
1260 LASSERT (msg->msg_kiov != NULL);
1262 rb = cfs_list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1264 LASSERT (rbp == lnet_msg2bufpool(msg));
1266 msg->msg_kiov = NULL;
1267 msg->msg_rtrcredit = 0;
1269 LASSERT((rbp->rbp_credits < 0) ==
1270 !cfs_list_empty(&rbp->rbp_msgs));
1271 LASSERT((rbp->rbp_credits > 0) ==
1272 !cfs_list_empty(&rbp->rbp_bufs));
1274 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1276 if (rbp->rbp_credits <= 0) {
1277 msg2 = cfs_list_entry(rbp->rbp_msgs.next,
1278 lnet_msg_t, msg_list);
1279 cfs_list_del(&msg2->msg_list);
1281 (void) lnet_post_routed_recv_locked(msg2, 1);
1285 if (msg->msg_peerrtrcredit) {
1286 /* give back peer router credits */
1287 msg->msg_peerrtrcredit = 0;
1289 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1290 !cfs_list_empty(&rxpeer->lp_rtrq));
1292 rxpeer->lp_rtrcredits++;
1293 if (rxpeer->lp_rtrcredits <= 0) {
1294 msg2 = cfs_list_entry(rxpeer->lp_rtrq.next,
1295 lnet_msg_t, msg_list);
1296 cfs_list_del(&msg2->msg_list);
1298 (void) lnet_post_routed_recv_locked(msg2, 1);
1302 LASSERT (!msg->msg_rtrcredit);
1303 LASSERT (!msg->msg_peerrtrcredit);
1305 if (rxpeer != NULL) {
1306 msg->msg_rxpeer = NULL;
1307 lnet_peer_decref_locked(rxpeer);
1312 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg)
1314 lnet_nid_t dst_nid = msg->msg_target.nid;
1316 lnet_ni_t *local_ni;
1317 lnet_remotenet_t *rnet;
1318 lnet_route_t *route;
1319 lnet_route_t *best_route;
1325 LASSERT (msg->msg_txpeer == NULL);
1326 LASSERT (!msg->msg_sending);
1327 LASSERT (!msg->msg_target_is_router);
1328 LASSERT (!msg->msg_receiving);
1330 msg->msg_sending = 1;
1332 /* NB! ni != NULL == interface pre-determined (ACK/REPLY) */
1336 if (the_lnet.ln_shutdown) {
1341 if (src_nid == LNET_NID_ANY) {
1344 src_ni = lnet_nid2ni_locked(src_nid);
1345 if (src_ni == NULL) {
1347 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1348 "local nid\n", libcfs_nid2str(dst_nid),
1349 libcfs_nid2str(src_nid));
1352 LASSERT (!msg->msg_routing);
1355 /* Is this for someone on a local network? */
1356 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid));
1358 if (local_ni != NULL) {
1359 if (src_ni == NULL) {
1361 src_nid = src_ni->ni_nid;
1362 } else if (src_ni == local_ni) {
1363 lnet_ni_decref_locked(local_ni);
1365 lnet_ni_decref_locked(local_ni);
1366 lnet_ni_decref_locked(src_ni);
1368 LCONSOLE_WARN("No route to %s via from %s\n",
1369 libcfs_nid2str(dst_nid),
1370 libcfs_nid2str(src_nid));
1374 LASSERT (src_nid != LNET_NID_ANY);
1376 if (!msg->msg_routing)
1377 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1379 if (src_ni == the_lnet.ln_loni) {
1380 /* No send credit hassles with LOLND */
1382 lnet_ni_send(src_ni, msg);
1383 lnet_ni_decref(src_ni);
1387 rc = lnet_nid2peer_locked(&lp, dst_nid);
1388 lnet_ni_decref_locked(src_ni); /* lp has ref on src_ni; lose mine */
1391 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1392 libcfs_nid2str(dst_nid));
1393 /* ENOMEM or shutting down */
1396 LASSERT (lp->lp_ni == src_ni);
1402 * - once application finishes computation, check here to update
1403 * router states before it waits for pending IO in LNetEQPoll
1404 * - recursion breaker: router checker sends no message
1405 * to remote networks */
1406 if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING)
1407 lnet_router_checker();
1411 /* sending to a remote network */
1412 rnet = lnet_find_net_locked(LNET_NIDNET(dst_nid));
1415 lnet_ni_decref_locked(src_ni);
1417 LCONSOLE_WARN("No route to %s\n",
1418 libcfs_id2str(msg->msg_target));
1419 return -EHOSTUNREACH;
1422 /* Find the best gateway I can use */
1425 cfs_list_for_each(tmp, &rnet->lrn_routes) {
1426 route = cfs_list_entry(tmp, lnet_route_t, lr_list);
1427 lp2 = route->lr_gateway;
1429 if (lp2->lp_alive &&
1430 lnet_router_down_ni(lp2, rnet->lrn_net) <= 0 &&
1431 (src_ni == NULL || lp2->lp_ni == src_ni) &&
1433 lnet_compare_routes(route, best_route) > 0)) {
1441 lnet_ni_decref_locked(src_ni);
1444 LCONSOLE_WARN("No route to %s via %s "
1445 "(all routers down)\n",
1446 libcfs_id2str(msg->msg_target),
1447 libcfs_nid2str(src_nid));
1448 return -EHOSTUNREACH;
1451 /* Place selected route at the end of the route list to ensure
1452 * fairness; everything else being equal... */
1453 cfs_list_del(&best_route->lr_list);
1454 cfs_list_add_tail(&best_route->lr_list, &rnet->lrn_routes);
1455 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1456 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1457 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1459 if (src_ni == NULL) {
1461 src_nid = src_ni->ni_nid;
1463 LASSERT (src_ni == lp->lp_ni);
1464 lnet_ni_decref_locked(src_ni);
1467 lnet_peer_addref_locked(lp);
1469 LASSERT (src_nid != LNET_NID_ANY);
1471 if (!msg->msg_routing) {
1472 /* I'm the source and now I know which NI to send on */
1473 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1476 msg->msg_target_is_router = 1;
1477 msg->msg_target.nid = lp->lp_nid;
1478 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1481 /* 'lp' is our best choice of peer */
1483 LASSERT (!msg->msg_peertxcredit);
1484 LASSERT (!msg->msg_txcredit);
1485 LASSERT (msg->msg_txpeer == NULL);
1487 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1489 rc = lnet_post_send_locked(msg, 0);
1492 if (rc == EHOSTUNREACH)
1493 return -EHOSTUNREACH;
1496 lnet_ni_send(src_ni, msg);
1502 lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg)
1504 /* ALWAYS called holding the LNET_LOCK */
1505 /* Here, we commit the MD to a network OP by marking it busy and
1506 * decrementing its threshold. Come what may, the network "owns"
1507 * the MD until a call to lnet_finalize() signals completion. */
1508 LASSERT (!msg->msg_routing);
1513 if (md->md_threshold != LNET_MD_THRESH_INF) {
1514 LASSERT (md->md_threshold > 0);
1518 the_lnet.ln_counters.msgs_alloc++;
1519 if (the_lnet.ln_counters.msgs_alloc >
1520 the_lnet.ln_counters.msgs_max)
1521 the_lnet.ln_counters.msgs_max =
1522 the_lnet.ln_counters.msgs_alloc;
1524 LASSERT (!msg->msg_onactivelist);
1525 msg->msg_onactivelist = 1;
1526 cfs_list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
1530 lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob)
1533 the_lnet.ln_counters.drop_count++;
1534 the_lnet.ln_counters.drop_length += nob;
1537 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1541 lnet_drop_delayed_put(lnet_msg_t *msg, char *reason)
1543 lnet_process_id_t id = {0};
1545 id.nid = msg->msg_hdr.src_nid;
1546 id.pid = msg->msg_hdr.src_pid;
1548 LASSERT (msg->msg_md == NULL);
1549 LASSERT (msg->msg_delayed);
1550 LASSERT (msg->msg_rxpeer != NULL);
1551 LASSERT (msg->msg_hdr.type == LNET_MSG_PUT);
1553 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
1554 " offset %d length %d: %s\n",
1556 msg->msg_hdr.msg.put.ptl_index,
1557 msg->msg_hdr.msg.put.match_bits,
1558 msg->msg_hdr.msg.put.offset,
1559 msg->msg_hdr.payload_length,
1562 /* NB I can't drop msg's ref on msg_rxpeer until after I've
1563 * called lnet_drop_message(), so I just hang onto msg as well
1564 * until that's done */
1566 lnet_drop_message(msg->msg_rxpeer->lp_ni,
1567 msg->msg_private, msg->msg_len);
1571 lnet_peer_decref_locked(msg->msg_rxpeer);
1572 msg->msg_rxpeer = NULL;
1580 * Turn on the lazy portal attribute. Use with caution!
1582 * This portal attribute only affects incoming PUT requests to the portal,
1583 * and is off by default. By default, if there's no matching MD for an
1584 * incoming PUT request, it is simply dropped. With the lazy attribute on,
1585 * such requests are queued indefinitely until either a matching MD is
1586 * posted to the portal or the lazy attribute is turned off.
1588 * It would prevent dropped requests, however it should be regarded as the
1589 * last line of defense - i.e. users must keep a close watch on active
1590 * buffers on a lazy portal and once it becomes too low post more buffers as
1591 * soon as possible. This is because delayed requests usually have detrimental
1592 * effects on underlying network connections. A few delayed requests often
1593 * suffice to bring an underlying connection to a complete halt, due to flow
1594 * control mechanisms.
1596 * There's also a DOS attack risk. If users don't post match-all MDs on a
1597 * lazy portal, a malicious peer can easily stop a service by sending some
1598 * PUT requests with match bits that won't match any MD. A routed server is
1599 * especially vulnerable since the connections to its neighbor routers are
1600 * shared among all clients.
1602 * \param portal Index of the portal to enable the lazy attribute on.
1604 * \retval 0 On success.
1605 * \retval -EINVAL If \a portal is not a valid index.
1608 LNetSetLazyPortal(int portal)
1610 lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1612 if (portal < 0 || portal >= the_lnet.ln_nportals)
1615 CDEBUG(D_NET, "Setting portal %d lazy\n", portal);
1618 lnet_portal_setopt(ptl, LNET_PTL_LAZY);
1625 * Turn off the lazy portal attribute. Delayed requests on the portal,
1626 * if any, will be all dropped when this function returns.
1628 * \param portal Index of the portal to disable the lazy attribute on.
1630 * \retval 0 On success.
1631 * \retval -EINVAL If \a portal is not a valid index.
1634 LNetClearLazyPortal(int portal)
1637 lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1640 if (portal < 0 || portal >= the_lnet.ln_nportals)
1645 if (!lnet_portal_is_lazy(ptl)) {
1650 if (the_lnet.ln_shutdown)
1651 CWARN ("Active lazy portal %d on exit\n", portal);
1653 CDEBUG (D_NET, "clearing portal %d lazy\n", portal);
1655 /* grab all the blocked messages atomically */
1656 cfs_list_add(&zombies, &ptl->ptl_msgq);
1657 cfs_list_del_init(&ptl->ptl_msgq);
1659 ptl->ptl_msgq_version++;
1660 lnet_portal_unsetopt(ptl, LNET_PTL_LAZY);
1664 while (!cfs_list_empty(&zombies)) {
1665 msg = cfs_list_entry(zombies.next, lnet_msg_t, msg_list);
1666 cfs_list_del(&msg->msg_list);
1668 lnet_drop_delayed_put(msg, "Clearing lazy portal attr");
1675 lnet_recv_put(lnet_libmd_t *md, lnet_msg_t *msg, int delayed,
1676 unsigned int offset, unsigned int mlength)
1678 lnet_hdr_t *hdr = &msg->msg_hdr;
1682 the_lnet.ln_counters.recv_count++;
1683 the_lnet.ln_counters.recv_length += mlength;
1688 lnet_setpayloadbuffer(msg);
1690 msg->msg_ev.type = LNET_EVENT_PUT;
1691 msg->msg_ev.target.pid = hdr->dest_pid;
1692 msg->msg_ev.target.nid = hdr->dest_nid;
1693 msg->msg_ev.hdr_data = hdr->msg.put.hdr_data;
1695 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1696 * it back into the ACK during lnet_finalize() */
1697 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1698 (md->md_options & LNET_MD_ACK_DISABLE) == 0);
1700 lnet_ni_recv(msg->msg_rxpeer->lp_ni,
1702 msg, delayed, offset, mlength,
1703 hdr->payload_length);
1706 /* called with LNET_LOCK held */
1708 lnet_match_blocked_msg(lnet_libmd_t *md)
1710 CFS_LIST_HEAD (drops);
1711 CFS_LIST_HEAD (matches);
1716 lnet_me_t *me = md->md_me;
1718 LASSERT (me->me_portal < (unsigned int)the_lnet.ln_nportals);
1720 ptl = &the_lnet.ln_portals[me->me_portal];
1721 if (!lnet_portal_is_lazy(ptl)) {
1722 LASSERT (cfs_list_empty(&ptl->ptl_msgq));
1726 LASSERT (md->md_refcount == 0); /* a brand new MD */
1728 cfs_list_for_each_safe (entry, tmp, &ptl->ptl_msgq) {
1731 unsigned int mlength;
1732 unsigned int offset;
1734 lnet_process_id_t src;
1736 msg = cfs_list_entry(entry, lnet_msg_t, msg_list);
1738 LASSERT (msg->msg_delayed);
1740 hdr = &msg->msg_hdr;
1741 index = hdr->msg.put.ptl_index;
1743 src.nid = hdr->src_nid;
1744 src.pid = hdr->src_pid;
1746 rc = lnet_try_match_md(index, LNET_MD_OP_PUT, src,
1747 hdr->payload_length,
1748 hdr->msg.put.offset,
1749 hdr->msg.put.match_bits,
1750 md, msg, &mlength, &offset);
1752 if (rc == LNET_MATCHMD_NONE)
1755 /* Hurrah! This _is_ a match */
1756 cfs_list_del(&msg->msg_list);
1757 ptl->ptl_msgq_version++;
1759 if (rc == LNET_MATCHMD_OK) {
1760 cfs_list_add_tail(&msg->msg_list, &matches);
1762 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
1763 "match "LPU64" offset %d length %d.\n",
1765 hdr->msg.put.ptl_index,
1766 hdr->msg.put.match_bits,
1767 hdr->msg.put.offset,
1768 hdr->payload_length);
1770 LASSERT (rc == LNET_MATCHMD_DROP);
1772 cfs_list_add_tail(&msg->msg_list, &drops);
1775 if (lnet_md_exhausted(md))
1781 cfs_list_for_each_safe (entry, tmp, &drops) {
1782 msg = cfs_list_entry(entry, lnet_msg_t, msg_list);
1784 cfs_list_del(&msg->msg_list);
1786 lnet_drop_delayed_put(msg, "Bad match");
1789 cfs_list_for_each_safe (entry, tmp, &matches) {
1790 msg = cfs_list_entry(entry, lnet_msg_t, msg_list);
1792 cfs_list_del(&msg->msg_list);
1794 /* md won't disappear under me, since each msg
1795 * holds a ref on it */
1796 lnet_recv_put(md, msg, 1,
1798 msg->msg_ev.mlength);
1805 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1810 lnet_hdr_t *hdr = &msg->msg_hdr;
1811 unsigned int rlength = hdr->payload_length;
1812 unsigned int mlength = 0;
1813 unsigned int offset = 0;
1814 lnet_process_id_t src= {0};
1818 src.nid = hdr->src_nid;
1819 src.pid = hdr->src_pid;
1821 /* Convert put fields to host byte order */
1822 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1823 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1824 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1826 index = hdr->msg.put.ptl_index;
1831 rc = lnet_match_md(index, LNET_MD_OP_PUT, src,
1832 rlength, hdr->msg.put.offset,
1833 hdr->msg.put.match_bits, msg,
1834 &mlength, &offset, &md);
1839 case LNET_MATCHMD_OK:
1841 lnet_recv_put(md, msg, msg->msg_delayed, offset, mlength);
1844 case LNET_MATCHMD_NONE:
1845 ptl = &the_lnet.ln_portals[index];
1846 version = ptl->ptl_ml_version;
1849 if (!msg->msg_delayed)
1850 rc = lnet_eager_recv_locked(msg);
1853 !the_lnet.ln_shutdown &&
1854 lnet_portal_is_lazy(ptl)) {
1855 if (version != ptl->ptl_ml_version)
1858 cfs_list_add_tail(&msg->msg_list, &ptl->ptl_msgq);
1859 ptl->ptl_msgq_version++;
1862 CDEBUG(D_NET, "Delaying PUT from %s portal %d match "
1863 LPU64" offset %d length %d: no match \n",
1864 libcfs_id2str(src), index,
1865 hdr->msg.put.match_bits,
1866 hdr->msg.put.offset, rlength);
1871 case LNET_MATCHMD_DROP:
1872 CNETERR("Dropping PUT from %s portal %d match "LPU64
1873 " offset %d length %d: %d\n",
1874 libcfs_id2str(src), index,
1875 hdr->msg.put.match_bits,
1876 hdr->msg.put.offset, rlength, rc);
1879 return ENOENT; /* +ve: OK but no match */
1884 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1886 lnet_hdr_t *hdr = &msg->msg_hdr;
1887 unsigned int mlength = 0;
1888 unsigned int offset = 0;
1889 lnet_process_id_t src = {0};
1890 lnet_handle_wire_t reply_wmd;
1894 src.nid = hdr->src_nid;
1895 src.pid = hdr->src_pid;
1897 /* Convert get fields to host byte order */
1898 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1899 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1900 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1901 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1905 rc = lnet_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src,
1906 hdr->msg.get.sink_length, hdr->msg.get.src_offset,
1907 hdr->msg.get.match_bits, msg,
1908 &mlength, &offset, &md);
1909 if (rc == LNET_MATCHMD_DROP) {
1910 CNETERR("Dropping GET from %s portal %d match "LPU64
1911 " offset %d length %d\n",
1913 hdr->msg.get.ptl_index,
1914 hdr->msg.get.match_bits,
1915 hdr->msg.get.src_offset,
1916 hdr->msg.get.sink_length);
1918 return ENOENT; /* +ve: OK but no match */
1921 LASSERT (rc == LNET_MATCHMD_OK);
1923 the_lnet.ln_counters.send_count++;
1924 the_lnet.ln_counters.send_length += mlength;
1928 msg->msg_ev.type = LNET_EVENT_GET;
1929 msg->msg_ev.target.pid = hdr->dest_pid;
1930 msg->msg_ev.target.nid = hdr->dest_nid;
1931 msg->msg_ev.hdr_data = 0;
1933 reply_wmd = hdr->msg.get.return_wmd;
1935 lnet_prep_send(msg, LNET_MSG_REPLY, src, offset, mlength);
1937 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1940 /* The LND completes the REPLY from her recv procedure */
1941 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1942 msg->msg_offset, msg->msg_len, msg->msg_len);
1946 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1947 msg->msg_receiving = 0;
1949 rc = lnet_send(ni->ni_nid, msg);
1951 /* didn't get as far as lnet_ni_send() */
1952 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1953 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc);
1955 lnet_finalize(ni, msg, rc);
1962 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1964 void *private = msg->msg_private;
1965 lnet_hdr_t *hdr = &msg->msg_hdr;
1966 lnet_process_id_t src = {0};
1973 src.nid = hdr->src_nid;
1974 src.pid = hdr->src_pid;
1976 /* NB handles only looked up by creator (no flips) */
1977 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1978 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1979 CNETERR("%s: Dropping REPLY from %s for %s "
1980 "MD "LPX64"."LPX64"\n",
1981 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1982 (md == NULL) ? "invalid" : "inactive",
1983 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1984 hdr->msg.reply.dst_wmd.wh_object_cookie);
1985 if (md != NULL && md->md_me != NULL)
1986 CERROR("REPLY MD also attached to portal %d\n",
1987 md->md_me->me_portal);
1990 return ENOENT; /* +ve: OK but no match */
1993 LASSERT (md->md_offset == 0);
1995 rlength = hdr->payload_length;
1996 mlength = MIN(rlength, (int)md->md_length);
1998 if (mlength < rlength &&
1999 (md->md_options & LNET_MD_TRUNCATE) == 0) {
2000 CNETERR("%s: Dropping REPLY from %s length %d "
2001 "for MD "LPX64" would overflow (%d)\n",
2002 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2003 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2006 return ENOENT; /* +ve: OK but no match */
2009 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
2010 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2011 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2013 lnet_commit_md(md, msg);
2016 lnet_setpayloadbuffer(msg);
2018 msg->msg_ev.type = LNET_EVENT_REPLY;
2019 msg->msg_ev.target.pid = hdr->dest_pid;
2020 msg->msg_ev.target.nid = hdr->dest_nid;
2021 msg->msg_ev.initiator = src;
2022 msg->msg_ev.rlength = rlength;
2023 msg->msg_ev.mlength = mlength;
2024 msg->msg_ev.offset = 0;
2026 lnet_md_deconstruct(md, &msg->msg_ev.md);
2027 lnet_md2handle(&msg->msg_ev.md_handle, md);
2029 the_lnet.ln_counters.recv_count++;
2030 the_lnet.ln_counters.recv_length += mlength;
2034 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2039 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
2041 lnet_hdr_t *hdr = &msg->msg_hdr;
2042 lnet_process_id_t src = {0};
2045 src.nid = hdr->src_nid;
2046 src.pid = hdr->src_pid;
2048 /* Convert ack fields to host byte order */
2049 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2050 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2054 /* NB handles only looked up by creator (no flips) */
2055 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2056 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2057 /* Don't moan; this is expected */
2059 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
2060 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2061 (md == NULL) ? "invalid" : "inactive",
2062 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2063 hdr->msg.ack.dst_wmd.wh_object_cookie);
2064 if (md != NULL && md->md_me != NULL)
2065 CERROR("Source MD also attached to portal %d\n",
2066 md->md_me->me_portal);
2069 return ENOENT; /* +ve! */
2072 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
2073 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2074 hdr->msg.ack.dst_wmd.wh_object_cookie);
2076 lnet_commit_md(md, msg);
2078 msg->msg_ev.type = LNET_EVENT_ACK;
2079 msg->msg_ev.target.pid = hdr->dest_pid;
2080 msg->msg_ev.target.nid = hdr->dest_nid;
2081 msg->msg_ev.initiator = src;
2082 msg->msg_ev.mlength = hdr->msg.ack.mlength;
2083 msg->msg_ev.match_bits = hdr->msg.ack.match_bits;
2085 lnet_md_deconstruct(md, &msg->msg_ev.md);
2086 lnet_md2handle(&msg->msg_ev.md_handle, md);
2088 the_lnet.ln_counters.recv_count++;
2092 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2097 lnet_msgtyp2str (int type)
2106 case LNET_MSG_REPLY:
2108 case LNET_MSG_HELLO:
2111 return ("<UNKNOWN>");
2116 lnet_print_hdr(lnet_hdr_t * hdr)
2118 lnet_process_id_t src = {0};
2119 lnet_process_id_t dst = {0};
2120 char *type_str = lnet_msgtyp2str (hdr->type);
2122 src.nid = hdr->src_nid;
2123 src.pid = hdr->src_pid;
2125 dst.nid = hdr->dest_nid;
2126 dst.pid = hdr->dest_pid;
2128 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2129 CWARN(" From %s\n", libcfs_id2str(src));
2130 CWARN(" To %s\n", libcfs_id2str(dst));
2132 switch (hdr->type) {
2137 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
2138 "match bits "LPU64"\n",
2139 hdr->msg.put.ptl_index,
2140 hdr->msg.put.ack_wmd.wh_interface_cookie,
2141 hdr->msg.put.ack_wmd.wh_object_cookie,
2142 hdr->msg.put.match_bits);
2143 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
2144 hdr->payload_length, hdr->msg.put.offset,
2145 hdr->msg.put.hdr_data);
2149 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
2150 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
2151 hdr->msg.get.return_wmd.wh_interface_cookie,
2152 hdr->msg.get.return_wmd.wh_object_cookie,
2153 hdr->msg.get.match_bits);
2154 CWARN(" Length %d, src offset %d\n",
2155 hdr->msg.get.sink_length,
2156 hdr->msg.get.src_offset);
2160 CWARN(" dst md "LPX64"."LPX64", "
2161 "manipulated length %d\n",
2162 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2163 hdr->msg.ack.dst_wmd.wh_object_cookie,
2164 hdr->msg.ack.mlength);
2167 case LNET_MSG_REPLY:
2168 CWARN(" dst md "LPX64"."LPX64", "
2170 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2171 hdr->msg.reply.dst_wmd.wh_object_cookie,
2172 hdr->payload_length);
2178 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
2179 void *private, int rdma_req)
2184 lnet_pid_t dest_pid;
2185 lnet_nid_t dest_nid;
2187 __u32 payload_length;
2190 LASSERT (!cfs_in_interrupt ());
2192 type = le32_to_cpu(hdr->type);
2193 src_nid = le64_to_cpu(hdr->src_nid);
2194 dest_nid = le64_to_cpu(hdr->dest_nid);
2195 dest_pid = le32_to_cpu(hdr->dest_pid);
2196 payload_length = le32_to_cpu(hdr->payload_length);
2198 for_me = (ni->ni_nid == dest_nid);
2203 if (payload_length > 0) {
2204 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2205 libcfs_nid2str(from_nid),
2206 libcfs_nid2str(src_nid),
2207 lnet_msgtyp2str(type), payload_length);
2213 case LNET_MSG_REPLY:
2214 if (payload_length > (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2215 CERROR("%s, src %s: bad %s payload %d "
2216 "(%d max expected)\n",
2217 libcfs_nid2str(from_nid),
2218 libcfs_nid2str(src_nid),
2219 lnet_msgtyp2str(type),
2221 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2227 CERROR("%s, src %s: Bad message type 0x%x\n",
2228 libcfs_nid2str(from_nid),
2229 libcfs_nid2str(src_nid), type);
2233 if (the_lnet.ln_routing) {
2234 cfs_time_t now = cfs_time_current();
2238 ni->ni_last_alive = now;
2239 if (ni->ni_status != NULL &&
2240 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
2241 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
2246 /* Regard a bad destination NID as a protocol error. Senders should
2247 * know what they're doing; if they don't they're misconfigured, buggy
2248 * or malicious so we chop them off at the knees :) */
2251 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2252 /* should have gone direct */
2253 CERROR ("%s, src %s: Bad dest nid %s "
2254 "(should have been sent direct)\n",
2255 libcfs_nid2str(from_nid),
2256 libcfs_nid2str(src_nid),
2257 libcfs_nid2str(dest_nid));
2261 if (lnet_islocalnid(dest_nid)) {
2262 /* dest is another local NI; sender should have used
2263 * this node's NID on its own network */
2264 CERROR ("%s, src %s: Bad dest nid %s "
2265 "(it's my nid but on a different network)\n",
2266 libcfs_nid2str(from_nid),
2267 libcfs_nid2str(src_nid),
2268 libcfs_nid2str(dest_nid));
2272 if (rdma_req && type == LNET_MSG_GET) {
2273 CERROR ("%s, src %s: Bad optimized GET for %s "
2274 "(final destination must be me)\n",
2275 libcfs_nid2str(from_nid),
2276 libcfs_nid2str(src_nid),
2277 libcfs_nid2str(dest_nid));
2281 if (!the_lnet.ln_routing) {
2282 CERROR ("%s, src %s: Dropping message for %s "
2283 "(routing not enabled)\n",
2284 libcfs_nid2str(from_nid),
2285 libcfs_nid2str(src_nid),
2286 libcfs_nid2str(dest_nid));
2291 /* Message looks OK; we're not going to return an error, so we MUST
2292 * call back lnd_recv() come what may... */
2294 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2295 fail_peer (src_nid, 0)) /* shall we now? */
2297 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2298 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2299 lnet_msgtyp2str(type));
2303 msg = lnet_msg_alloc();
2305 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2306 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2307 lnet_msgtyp2str(type));
2311 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
2313 msg->msg_type = type;
2314 msg->msg_private = private;
2315 msg->msg_receiving = 1;
2316 msg->msg_len = msg->msg_wanted = payload_length;
2317 msg->msg_offset = 0;
2318 msg->msg_hdr = *hdr;
2321 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid);
2324 CERROR("%s, src %s: Dropping %s "
2325 "(error %d looking up sender)\n",
2326 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2327 lnet_msgtyp2str(type), rc);
2336 msg->msg_target.pid = dest_pid;
2337 msg->msg_target.nid = dest_nid;
2338 msg->msg_routing = 1;
2339 msg->msg_offset = 0;
2342 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
2343 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2344 rc = lnet_eager_recv_locked(msg);
2350 lnet_commit_routedmsg(msg);
2351 rc = lnet_post_routed_recv_locked(msg, 0);
2355 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2356 0, payload_length, payload_length);
2360 /* convert common msg->hdr fields to host byteorder */
2361 msg->msg_hdr.type = type;
2362 msg->msg_hdr.src_nid = src_nid;
2363 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2364 msg->msg_hdr.dest_nid = dest_nid;
2365 msg->msg_hdr.dest_pid = dest_pid;
2366 msg->msg_hdr.payload_length = payload_length;
2368 msg->msg_ev.sender = from_nid;
2372 rc = lnet_parse_ack(ni, msg);
2375 rc = lnet_parse_put(ni, msg);
2378 rc = lnet_parse_get(ni, msg, rdma_req);
2380 case LNET_MSG_REPLY:
2381 rc = lnet_parse_reply(ni, msg);
2385 goto free_drop; /* prevent an unused label if !kernel */
2391 LASSERT (rc == ENOENT);
2394 LASSERT (msg->msg_md == NULL);
2396 if (msg->msg_rxpeer != NULL) {
2397 lnet_peer_decref_locked(msg->msg_rxpeer);
2398 msg->msg_rxpeer = NULL;
2400 lnet_msg_free(msg); /* expects LNET_LOCK held */
2404 lnet_drop_message(ni, private, payload_length);
2409 * Initiate an asynchronous PUT operation.
2411 * There are several events associated with a PUT: completion of the send on
2412 * the initiator node (LNET_EVENT_SEND), and when the send completes
2413 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2414 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2415 * used at the target node to indicate the completion of incoming data
2418 * The local events will be logged in the EQ associated with the MD pointed to
2419 * by \a mdh handle. Using a MD without an associated EQ results in these
2420 * events being discarded. In this case, the caller must have another
2421 * mechanism (e.g., a higher level protocol) for determining when it is safe
2422 * to modify the memory region associated with the MD.
2424 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2425 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2427 * \param self Indicates the NID of a local interface through which to send
2428 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2429 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2430 * must be "free floating" (See LNetMDBind()).
2431 * \param ack Controls whether an acknowledgment is requested.
2432 * Acknowledgments are only sent when they are requested by the initiating
2433 * process and the target MD enables them.
2434 * \param target A process identifier for the target process.
2435 * \param portal The index in the \a target's portal table.
2436 * \param match_bits The match bits to use for MD selection at the target
2438 * \param offset The offset into the target MD (only used when the target
2439 * MD has the LNET_MD_MANAGE_REMOTE option set).
2440 * \param hdr_data 64 bits of user data that can be included in the message
2441 * header. This data is written to an event queue entry at the target if an
2442 * EQ is present on the matching MD.
2444 * \retval 0 Success, and only in this case events will be generated
2445 * and logged to EQ (if it exists).
2446 * \retval -EIO Simulated failure.
2447 * \retval -ENOMEM Memory allocation failure.
2448 * \retval -ENOENT Invalid MD object.
2450 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2453 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2454 lnet_process_id_t target, unsigned int portal,
2455 __u64 match_bits, unsigned int offset,
2462 LASSERT (the_lnet.ln_init);
2463 LASSERT (the_lnet.ln_refcount > 0);
2465 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2466 fail_peer (target.nid, 1)) /* shall we now? */
2468 CERROR("Dropping PUT to %s: simulated failure\n",
2469 libcfs_id2str(target));
2473 msg = lnet_msg_alloc();
2475 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2476 libcfs_id2str(target));
2479 msg->msg_vmflush = !!cfs_memory_pressure_get();
2483 md = lnet_handle2md(&mdh);
2484 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2487 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2488 match_bits, portal, libcfs_id2str(target),
2489 md == NULL ? -1 : md->md_threshold);
2490 if (md != NULL && md->md_me != NULL)
2491 CERROR("Source MD also attached to portal %d\n",
2492 md->md_me->me_portal);
2498 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2500 lnet_commit_md(md, msg);
2502 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2504 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2505 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2506 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2507 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2509 /* NB handles only looked up by creator (no flips) */
2510 if (ack == LNET_ACK_REQ) {
2511 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2512 the_lnet.ln_interface_cookie;
2513 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2514 md->md_lh.lh_cookie;
2516 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2517 LNET_WIRE_HANDLE_COOKIE_NONE;
2518 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2519 LNET_WIRE_HANDLE_COOKIE_NONE;
2522 msg->msg_ev.type = LNET_EVENT_SEND;
2523 msg->msg_ev.initiator.nid = LNET_NID_ANY;
2524 msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2525 msg->msg_ev.target = target;
2526 msg->msg_ev.sender = LNET_NID_ANY;
2527 msg->msg_ev.pt_index = portal;
2528 msg->msg_ev.match_bits = match_bits;
2529 msg->msg_ev.rlength = md->md_length;
2530 msg->msg_ev.mlength = md->md_length;
2531 msg->msg_ev.offset = offset;
2532 msg->msg_ev.hdr_data = hdr_data;
2534 lnet_md_deconstruct(md, &msg->msg_ev.md);
2535 lnet_md2handle(&msg->msg_ev.md_handle, md);
2537 the_lnet.ln_counters.send_count++;
2538 the_lnet.ln_counters.send_length += md->md_length;
2542 rc = lnet_send(self, msg);
2544 CNETERR( "Error sending PUT to %s: %d\n",
2545 libcfs_id2str(target), rc);
2546 lnet_finalize (NULL, msg, rc);
2549 /* completion will be signalled by an event */
2554 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2556 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2557 * returns a msg for the LND to pass to lnet_finalize() when the sink
2558 * data has been received.
2560 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2561 * lnet_finalize() is called on it, so the LND must call this first */
2563 lnet_msg_t *msg = lnet_msg_alloc();
2564 lnet_libmd_t *getmd = getmsg->msg_md;
2565 lnet_process_id_t peer_id = getmsg->msg_target;
2567 LASSERT (!getmsg->msg_target_is_router);
2568 LASSERT (!getmsg->msg_routing);
2572 LASSERT (getmd->md_refcount > 0);
2575 CERROR ("%s: Dropping REPLY from %s: can't allocate msg\n",
2576 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2580 if (getmd->md_threshold == 0) {
2581 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2582 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2587 LASSERT (getmd->md_offset == 0);
2589 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2590 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2592 lnet_commit_md (getmd, msg);
2594 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2596 msg->msg_ev.type = LNET_EVENT_REPLY;
2597 msg->msg_ev.initiator = peer_id;
2598 msg->msg_ev.sender = peer_id.nid; /* optimized GETs can't be routed */
2599 msg->msg_ev.rlength = msg->msg_ev.mlength = getmd->md_length;
2600 msg->msg_ev.offset = 0;
2602 lnet_md_deconstruct(getmd, &msg->msg_ev.md);
2603 lnet_md2handle(&msg->msg_ev.md_handle, getmd);
2605 the_lnet.ln_counters.recv_count++;
2606 the_lnet.ln_counters.recv_length += getmd->md_length;
2615 the_lnet.ln_counters.drop_count++;
2616 the_lnet.ln_counters.drop_length += getmd->md_length;
2624 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2626 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2627 * completed and I know it. */
2628 LASSERT (reply != NULL);
2629 LASSERT (reply->msg_type == LNET_MSG_GET);
2630 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2632 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2633 * the end of my buffer, I might as well be dead. */
2634 LASSERT (len <= reply->msg_ev.mlength);
2636 reply->msg_ev.mlength = len;
2640 * Initiate an asynchronous GET operation.
2642 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2643 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2644 * the target node in the REPLY has been written to local MD.
2646 * On the target node, an LNET_EVENT_GET is logged when the GET request
2647 * arrives and is accepted into a MD.
2649 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2650 * \param mdh A handle for the MD that describes the memory into which the
2651 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2653 * \retval 0 Success, and only in this case events will be generated
2654 * and logged to EQ (if it exists) of the MD.
2655 * \retval -EIO Simulated failure.
2656 * \retval -ENOMEM Memory allocation failure.
2657 * \retval -ENOENT Invalid MD object.
2660 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2661 lnet_process_id_t target, unsigned int portal,
2662 __u64 match_bits, unsigned int offset)
2668 LASSERT (the_lnet.ln_init);
2669 LASSERT (the_lnet.ln_refcount > 0);
2671 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2672 fail_peer (target.nid, 1)) /* shall we now? */
2674 CERROR("Dropping GET to %s: simulated failure\n",
2675 libcfs_id2str(target));
2679 msg = lnet_msg_alloc();
2681 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2682 libcfs_id2str(target));
2688 md = lnet_handle2md(&mdh);
2689 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2692 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2693 match_bits, portal, libcfs_id2str(target),
2694 md == NULL ? -1 : md->md_threshold);
2695 if (md != NULL && md->md_me != NULL)
2696 CERROR("REPLY MD also attached to portal %d\n",
2697 md->md_me->me_portal);
2703 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2705 lnet_commit_md(md, msg);
2707 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2709 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2710 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2711 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2712 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2714 /* NB handles only looked up by creator (no flips) */
2715 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2716 the_lnet.ln_interface_cookie;
2717 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2718 md->md_lh.lh_cookie;
2720 msg->msg_ev.type = LNET_EVENT_SEND;
2721 msg->msg_ev.initiator.nid = LNET_NID_ANY;
2722 msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2723 msg->msg_ev.target = target;
2724 msg->msg_ev.sender = LNET_NID_ANY;
2725 msg->msg_ev.pt_index = portal;
2726 msg->msg_ev.match_bits = match_bits;
2727 msg->msg_ev.rlength = md->md_length;
2728 msg->msg_ev.mlength = md->md_length;
2729 msg->msg_ev.offset = offset;
2730 msg->msg_ev.hdr_data = 0;
2732 lnet_md_deconstruct(md, &msg->msg_ev.md);
2733 lnet_md2handle(&msg->msg_ev.md_handle, md);
2735 the_lnet.ln_counters.send_count++;
2739 rc = lnet_send(self, msg);
2741 CNETERR( "Error sending GET to %s: %d\n",
2742 libcfs_id2str(target), rc);
2743 lnet_finalize (NULL, msg, rc);
2746 /* completion will be signalled by an event */
2751 * Calculate distance to node at \a dstnid.
2753 * \param dstnid Target NID.
2754 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2756 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2759 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2760 * local_nid_dist_zero is set, which is the default.
2761 * \retval positives Distance to target NID, i.e. number of hops plus one.
2762 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2765 LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2769 lnet_remotenet_t *rnet;
2770 __u32 dstnet = LNET_NIDNET(dstnid);
2774 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2775 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2776 * keep order 0 free for 0@lo and order 1 free for a local NID
2779 LASSERT (the_lnet.ln_init);
2780 LASSERT (the_lnet.ln_refcount > 0);
2784 cfs_list_for_each (e, &the_lnet.ln_nis) {
2785 ni = cfs_list_entry(e, lnet_ni_t, ni_list);
2787 if (ni->ni_nid == dstnid) {
2788 if (srcnidp != NULL)
2790 if (orderp != NULL) {
2791 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2798 return local_nid_dist_zero ? 0 : 1;
2801 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2802 if (srcnidp != NULL)
2803 *srcnidp = ni->ni_nid;
2813 cfs_list_for_each (e, &the_lnet.ln_remote_nets) {
2814 rnet = cfs_list_entry(e, lnet_remotenet_t, lrn_list);
2816 if (rnet->lrn_net == dstnet) {
2817 lnet_route_t *route;
2818 lnet_route_t *shortest = NULL;
2820 LASSERT (!cfs_list_empty(&rnet->lrn_routes));
2822 cfs_list_for_each_entry(route, &rnet->lrn_routes,
2824 if (shortest == NULL ||
2825 route->lr_hops < shortest->lr_hops)
2829 LASSERT (shortest != NULL);
2830 hops = shortest->lr_hops;
2831 if (srcnidp != NULL)
2832 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2842 return -EHOSTUNREACH;
2846 * Set the number of asynchronous messages expected from a target process.
2848 * This function is only meaningful for userspace callers. It's a no-op when
2849 * called from kernel.
2851 * Asynchronous messages are those that can come from a target when the
2852 * userspace process is not waiting for IO to complete; e.g., AST callbacks
2853 * from Lustre servers. Specifying the expected number of such messages
2854 * allows them to be eagerly received when user process is not running in
2855 * LNet; otherwise network errors may occur.
2857 * \param id Process ID of the target process.
2858 * \param nasync Number of asynchronous messages expected from the target.
2860 * \return 0 on success, and an error code otherwise.
2863 LNetSetAsync(lnet_process_id_t id, int nasync)
2869 lnet_remotenet_t *rnet;
2871 lnet_route_t *route;
2878 /* Target on a local network? */
2879 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2881 if (ni->ni_lnd->lnd_setasync != NULL)
2882 rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2887 /* Target on a remote network: apply to routers */
2889 LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2894 /* Snapshot all the router NIDs */
2896 rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2898 cfs_list_for_each(tmp, &rnet->lrn_routes) {
2899 if (nnids == maxnids) {
2901 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2906 route = cfs_list_entry(tmp, lnet_route_t, lr_list);
2907 nids[nnids++] = route->lr_gateway->lp_nid;
2912 /* set async on all the routers */
2913 while (nnids-- > 0) {
2914 id.pid = LUSTRE_SRV_LNET_PID;
2915 id.nid = nids[nnids];
2917 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2921 if (ni->ni_lnd->lnd_setasync != NULL) {
2922 rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2929 LIBCFS_FREE(nids, maxnids * sizeof(*nids));