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 2008 Sun Microsystems, Inc. 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];
163 CDEBUG (D_NET, "Request from %s of length %d into portal %d "
164 "MB="LPX64"\n", libcfs_id2str(src), rlength, index, match_bits);
166 if (index < 0 || index >= the_lnet.ln_nportals) {
167 CERROR("Invalid portal %d not in [0-%d]\n",
168 index, the_lnet.ln_nportals);
169 return LNET_MATCHMD_DROP;
172 cfs_list_for_each_entry_safe_typed (me, tmp, &ptl->ptl_ml,
173 lnet_me_t, me_list) {
176 /* ME attached but MD not attached yet */
180 LASSERT (me == md->md_me);
182 rc = lnet_try_match_md(index, op_mask, src, rlength,
183 roffset, match_bits, md, msg,
184 mlength_out, offset_out);
189 case LNET_MATCHMD_NONE:
192 case LNET_MATCHMD_OK:
194 return LNET_MATCHMD_OK;
196 case LNET_MATCHMD_DROP:
197 return LNET_MATCHMD_DROP;
202 if (op_mask == LNET_MD_OP_GET ||
203 (ptl->ptl_options & LNET_PTL_LAZY) == 0)
204 return LNET_MATCHMD_DROP;
206 return LNET_MATCHMD_NONE;
210 lnet_fail_nid (lnet_nid_t nid, unsigned int threshold)
212 lnet_test_peer_t *tp;
213 struct list_head *el;
214 struct list_head *next;
215 struct list_head cull;
217 LASSERT (the_lnet.ln_init);
219 if (threshold != 0) {
220 /* Adding a new entry */
221 LIBCFS_ALLOC(tp, sizeof(*tp));
226 tp->tp_threshold = threshold;
229 list_add_tail (&tp->tp_list, &the_lnet.ln_test_peers);
234 /* removing entries */
235 CFS_INIT_LIST_HEAD (&cull);
239 list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
240 tp = list_entry (el, lnet_test_peer_t, tp_list);
242 if (tp->tp_threshold == 0 || /* needs culling anyway */
243 nid == LNET_NID_ANY || /* removing all entries */
244 tp->tp_nid == nid) /* matched this one */
246 list_del (&tp->tp_list);
247 list_add (&tp->tp_list, &cull);
253 while (!list_empty (&cull)) {
254 tp = list_entry (cull.next, lnet_test_peer_t, tp_list);
256 list_del (&tp->tp_list);
257 LIBCFS_FREE(tp, sizeof (*tp));
263 fail_peer (lnet_nid_t nid, int outgoing)
265 lnet_test_peer_t *tp;
266 struct list_head *el;
267 struct list_head *next;
268 struct list_head cull;
271 CFS_INIT_LIST_HEAD (&cull);
275 list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
276 tp = list_entry (el, lnet_test_peer_t, tp_list);
278 if (tp->tp_threshold == 0) {
281 /* only cull zombies on outgoing tests,
282 * since we may be at interrupt priority on
283 * incoming messages. */
284 list_del (&tp->tp_list);
285 list_add (&tp->tp_list, &cull);
290 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
291 nid == tp->tp_nid) { /* fail this peer */
294 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
297 tp->tp_threshold == 0) {
299 list_del (&tp->tp_list);
300 list_add (&tp->tp_list, &cull);
309 while (!list_empty (&cull)) {
310 tp = list_entry (cull.next, lnet_test_peer_t, tp_list);
311 list_del (&tp->tp_list);
313 LIBCFS_FREE(tp, sizeof (*tp));
320 lnet_iov_nob (unsigned int niov, struct iovec *iov)
322 unsigned int nob = 0;
325 nob += (iov++)->iov_len;
331 lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
332 unsigned int nsiov, struct iovec *siov, unsigned int soffset,
335 /* NB diov, siov are READ-ONLY */
336 unsigned int this_nob;
341 /* skip complete frags before 'doffset' */
343 while (doffset >= diov->iov_len) {
344 doffset -= diov->iov_len;
350 /* skip complete frags before 'soffset' */
352 while (soffset >= siov->iov_len) {
353 soffset -= siov->iov_len;
362 this_nob = MIN(diov->iov_len - doffset,
363 siov->iov_len - soffset);
364 this_nob = MIN(this_nob, nob);
366 memcpy ((char *)diov->iov_base + doffset,
367 (char *)siov->iov_base + soffset, this_nob);
370 if (diov->iov_len > doffset + this_nob) {
378 if (siov->iov_len > soffset + this_nob) {
389 lnet_extract_iov (int dst_niov, struct iovec *dst,
390 int src_niov, struct iovec *src,
391 unsigned int offset, unsigned int len)
393 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
394 * for exactly 'len' bytes, and return the number of entries.
395 * NB not destructive to 'src' */
396 unsigned int frag_len;
399 if (len == 0) /* no data => */
400 return (0); /* no frags */
402 LASSERT (src_niov > 0);
403 while (offset >= src->iov_len) { /* skip initial frags */
404 offset -= src->iov_len;
407 LASSERT (src_niov > 0);
412 LASSERT (src_niov > 0);
413 LASSERT ((int)niov <= dst_niov);
415 frag_len = src->iov_len - offset;
416 dst->iov_base = ((char *)src->iov_base) + offset;
418 if (len <= frag_len) {
423 dst->iov_len = frag_len;
436 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
443 lnet_copy_kiov2kiov (unsigned int ndkiov, lnet_kiov_t *dkiov, unsigned int doffset,
444 unsigned int nskiov, lnet_kiov_t *skiov, unsigned int soffset,
451 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
452 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
459 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
460 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
467 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
468 int src_niov, lnet_kiov_t *src,
469 unsigned int offset, unsigned int len)
474 #else /* __KERNEL__ */
477 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
479 unsigned int nob = 0;
482 nob += (kiov++)->kiov_len;
488 lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
489 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
492 /* NB diov, siov are READ-ONLY */
493 unsigned int this_nob;
500 LASSERT (!in_interrupt ());
503 while (doffset >= diov->kiov_len) {
504 doffset -= diov->kiov_len;
511 while (soffset >= siov->kiov_len) {
512 soffset -= siov->kiov_len;
521 this_nob = MIN(diov->kiov_len - doffset,
522 siov->kiov_len - soffset);
523 this_nob = MIN(this_nob, nob);
526 daddr = ((char *)cfs_kmap(diov->kiov_page)) +
527 diov->kiov_offset + doffset;
529 saddr = ((char *)cfs_kmap(siov->kiov_page)) +
530 siov->kiov_offset + soffset;
532 /* Vanishing risk of kmap deadlock when mapping 2 pages.
533 * However in practice at least one of the kiovs will be mapped
534 * kernel pages and the map/unmap will be NOOPs */
536 memcpy (daddr, saddr, this_nob);
539 if (diov->kiov_len > doffset + this_nob) {
543 cfs_kunmap(diov->kiov_page);
550 if (siov->kiov_len > soffset + this_nob) {
554 cfs_kunmap(siov->kiov_page);
563 cfs_kunmap(diov->kiov_page);
565 cfs_kunmap(siov->kiov_page);
569 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
570 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
573 /* NB iov, kiov are READ-ONLY */
574 unsigned int this_nob;
580 LASSERT (!in_interrupt ());
583 while (iovoffset >= iov->iov_len) {
584 iovoffset -= iov->iov_len;
591 while (kiovoffset >= kiov->kiov_len) {
592 kiovoffset -= kiov->kiov_len;
601 this_nob = MIN(iov->iov_len - iovoffset,
602 kiov->kiov_len - kiovoffset);
603 this_nob = MIN(this_nob, nob);
606 addr = ((char *)cfs_kmap(kiov->kiov_page)) +
607 kiov->kiov_offset + kiovoffset;
609 memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
612 if (iov->iov_len > iovoffset + this_nob) {
613 iovoffset += this_nob;
620 if (kiov->kiov_len > kiovoffset + this_nob) {
622 kiovoffset += this_nob;
624 cfs_kunmap(kiov->kiov_page);
634 cfs_kunmap(kiov->kiov_page);
638 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
639 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
642 /* NB kiov, iov are READ-ONLY */
643 unsigned int this_nob;
649 LASSERT (!in_interrupt ());
652 while (kiovoffset >= kiov->kiov_len) {
653 kiovoffset -= kiov->kiov_len;
660 while (iovoffset >= iov->iov_len) {
661 iovoffset -= iov->iov_len;
670 this_nob = MIN(kiov->kiov_len - kiovoffset,
671 iov->iov_len - iovoffset);
672 this_nob = MIN(this_nob, nob);
675 addr = ((char *)cfs_kmap(kiov->kiov_page)) +
676 kiov->kiov_offset + kiovoffset;
678 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
681 if (kiov->kiov_len > kiovoffset + this_nob) {
683 kiovoffset += this_nob;
685 cfs_kunmap(kiov->kiov_page);
692 if (iov->iov_len > iovoffset + this_nob) {
693 iovoffset += this_nob;
702 cfs_kunmap(kiov->kiov_page);
706 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
707 int src_niov, lnet_kiov_t *src,
708 unsigned int offset, unsigned int len)
710 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
711 * for exactly 'len' bytes, and return the number of entries.
712 * NB not destructive to 'src' */
713 unsigned int frag_len;
716 if (len == 0) /* no data => */
717 return (0); /* no frags */
719 LASSERT (src_niov > 0);
720 while (offset >= src->kiov_len) { /* skip initial frags */
721 offset -= src->kiov_len;
724 LASSERT (src_niov > 0);
729 LASSERT (src_niov > 0);
730 LASSERT ((int)niov <= dst_niov);
732 frag_len = src->kiov_len - offset;
733 dst->kiov_page = src->kiov_page;
734 dst->kiov_offset = src->kiov_offset + offset;
736 if (len <= frag_len) {
738 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
742 dst->kiov_len = frag_len;
743 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
756 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
757 unsigned int offset, unsigned int mlen, unsigned int rlen)
759 unsigned int niov = 0;
760 struct iovec *iov = NULL;
761 lnet_kiov_t *kiov = NULL;
764 LASSERT (!in_interrupt ());
765 LASSERT (mlen == 0 || msg != NULL);
768 LASSERT(msg->msg_receiving);
769 LASSERT(!msg->msg_sending);
770 LASSERT(rlen == msg->msg_len);
771 LASSERT(mlen <= msg->msg_len);
773 msg->msg_wanted = mlen;
774 msg->msg_offset = offset;
775 msg->msg_receiving = 0;
778 niov = msg->msg_niov;
780 kiov = msg->msg_kiov;
783 LASSERT ((iov == NULL) != (kiov == NULL));
787 rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
788 niov, iov, kiov, offset, mlen, rlen);
790 lnet_finalize(ni, msg, rc);
794 lnet_compare_routers(lnet_peer_t *p1, lnet_peer_t *p2)
796 if (p1->lp_txqnob < p2->lp_txqnob)
799 if (p1->lp_txqnob > p2->lp_txqnob)
802 if (p1->lp_txcredits > p2->lp_txcredits)
805 if (p1->lp_txcredits < p2->lp_txcredits)
813 lnet_setpayloadbuffer(lnet_msg_t *msg)
815 lnet_libmd_t *md = msg->msg_md;
817 LASSERT (msg->msg_len > 0);
818 LASSERT (!msg->msg_routing);
819 LASSERT (md != NULL);
820 LASSERT (msg->msg_niov == 0);
821 LASSERT (msg->msg_iov == NULL);
822 LASSERT (msg->msg_kiov == NULL);
824 msg->msg_niov = md->md_niov;
825 if ((md->md_options & LNET_MD_KIOV) != 0)
826 msg->msg_kiov = md->md_iov.kiov;
828 msg->msg_iov = md->md_iov.iov;
832 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
833 unsigned int offset, unsigned int len)
835 msg->msg_type = type;
836 msg->msg_target = target;
838 msg->msg_offset = offset;
841 lnet_setpayloadbuffer(msg);
843 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
844 msg->msg_hdr.type = cpu_to_le32(type);
845 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
846 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
847 /* src_nid will be set later */
848 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
849 msg->msg_hdr.payload_length = cpu_to_le32(len);
853 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
855 void *priv = msg->msg_private;
858 LASSERT (!in_interrupt ());
859 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
860 (msg->msg_txcredit && msg->msg_peertxcredit));
862 rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
864 lnet_finalize(ni, msg, rc);
868 lnet_eager_recv_locked(lnet_msg_t *msg)
874 LASSERT (!msg->msg_delayed);
875 msg->msg_delayed = 1;
877 LASSERT (msg->msg_receiving);
878 LASSERT (!msg->msg_sending);
880 peer = msg->msg_rxpeer;
883 if (ni->ni_lnd->lnd_eager_recv != NULL) {
886 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
889 CERROR("recv from %s / send to %s aborted: "
890 "eager_recv failed %d\n",
891 libcfs_nid2str(peer->lp_nid),
892 libcfs_id2str(msg->msg_target), rc);
893 LASSERT (rc < 0); /* required by my callers */
902 /* NB: caller shall hold a ref on 'lp' as I'd drop LNET_LOCK */
904 lnet_ni_peer_alive(lnet_peer_t *lp)
906 time_t last_alive = 0;
907 lnet_ni_t *ni = lp->lp_ni;
909 LASSERT (ni != NULL);
910 LASSERT (ni->ni_peertimeout > 0);
911 LASSERT (ni->ni_lnd->lnd_query != NULL);
914 (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
917 lp->lp_last_query = cfs_time_current_sec();
919 if (last_alive != 0) /* NI has updated timestamp */
920 lp->lp_last_alive = last_alive;
924 /* NB: always called with LNET_LOCK held */
926 lnet_peer_is_alive (lnet_peer_t *lp, time_t now)
928 lnet_ni_t *ni = lp->lp_ni;
932 LASSERT (ni != NULL);
933 LASSERT (ni->ni_peertimeout > 0);
935 if (!lp->lp_alive && lp->lp_alive_count > 0 &&
936 cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive))
939 deadline = cfs_time_add(lp->lp_last_alive, ni->ni_peertimeout);
940 alive = cfs_time_after(deadline, now);
941 if (alive && !lp->lp_alive) /* update obsolete lp_alive */
942 lnet_notify_locked(lp, 0, 1, lp->lp_last_alive);
947 /* don't query LND about aliveness of a dead peer more frequently than: */
948 static int lnet_queryinterval = 1; /* 1 second */
950 /* NB: returns 1 when alive, 0 when dead, negative when error;
951 * may drop the LNET_LOCK */
953 lnet_peer_alive_locked (lnet_peer_t *lp)
955 lnet_ni_t *ni = lp->lp_ni;
956 time_t now = cfs_time_current_sec();
958 LASSERT (ni != NULL);
960 if (ni->ni_peertimeout <= 0) /* disabled */
963 if (lnet_peer_is_alive(lp, now))
966 /* peer appears dead, should we query right now? */
967 if (lp->lp_last_query != 0) {
969 cfs_time_add(lp->lp_last_query,
972 if (cfs_time_before(now, deadline)) {
974 CWARN("Unexpected aliveness of peer %s: "
976 libcfs_nid2str(lp->lp_nid),
977 (int)now, (int)deadline,
978 lnet_queryinterval, ni->ni_peertimeout);
983 /* query LND for latest aliveness news */
984 lnet_ni_peer_alive(lp);
986 if (lnet_peer_is_alive(lp, now))
989 lnet_notify_locked(lp, 0, 0, lp->lp_last_alive);
994 lnet_post_send_locked (lnet_msg_t *msg, int do_send)
996 /* lnet_send is going to LNET_UNLOCK immediately after this, so it sets
997 * do_send FALSE and I don't do the unlock/send/lock bit. I return
998 * EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer appears dead, and
999 * 0 if sent or OK to send */
1000 lnet_peer_t *lp = msg->msg_txpeer;
1001 lnet_ni_t *ni = lp->lp_ni;
1003 /* non-lnet_send() callers have checked before */
1004 LASSERT (!do_send || msg->msg_delayed);
1005 LASSERT (!msg->msg_receiving);
1007 /* NB 'lp' is always the next hop */
1008 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
1009 lnet_peer_alive_locked(lp) == 0) {
1012 CDEBUG(D_NETERROR, "Dropping message for %s: peer not alive\n",
1013 libcfs_id2str(msg->msg_target));
1015 lnet_finalize(ni, msg, -EHOSTUNREACH);
1018 return EHOSTUNREACH;
1021 if (!msg->msg_peertxcredit) {
1022 LASSERT ((lp->lp_txcredits < 0) == !list_empty(&lp->lp_txq));
1024 msg->msg_peertxcredit = 1;
1025 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
1028 if (lp->lp_txcredits < lp->lp_mintxcredits)
1029 lp->lp_mintxcredits = lp->lp_txcredits;
1031 if (lp->lp_txcredits < 0) {
1032 msg->msg_delayed = 1;
1033 list_add_tail (&msg->msg_list, &lp->lp_txq);
1038 if (!msg->msg_txcredit) {
1039 LASSERT ((ni->ni_txcredits < 0) == !list_empty(&ni->ni_txq));
1041 msg->msg_txcredit = 1;
1044 if (ni->ni_txcredits < ni->ni_mintxcredits)
1045 ni->ni_mintxcredits = ni->ni_txcredits;
1047 if (ni->ni_txcredits < 0) {
1048 msg->msg_delayed = 1;
1049 list_add_tail (&msg->msg_list, &ni->ni_txq);
1056 lnet_ni_send(ni, msg);
1064 lnet_commit_routedmsg (lnet_msg_t *msg)
1066 /* ALWAYS called holding the LNET_LOCK */
1067 LASSERT (msg->msg_routing);
1069 the_lnet.ln_counters.msgs_alloc++;
1070 if (the_lnet.ln_counters.msgs_alloc >
1071 the_lnet.ln_counters.msgs_max)
1072 the_lnet.ln_counters.msgs_max =
1073 the_lnet.ln_counters.msgs_alloc;
1075 the_lnet.ln_counters.route_count++;
1076 the_lnet.ln_counters.route_length += msg->msg_len;
1078 LASSERT (!msg->msg_onactivelist);
1079 msg->msg_onactivelist = 1;
1080 list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
1084 lnet_msg2bufpool(lnet_msg_t *msg)
1086 lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[0];
1088 LASSERT (msg->msg_len <= LNET_MTU);
1089 while (msg->msg_len > (unsigned int)rbp->rbp_npages * CFS_PAGE_SIZE) {
1091 LASSERT (rbp < &the_lnet.ln_rtrpools[LNET_NRBPOOLS]);
1098 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
1100 /* lnet_parse is going to LNET_UNLOCK immediately after this, so it
1101 * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
1102 * return EAGAIN if msg blocked and 0 if received or OK to receive */
1103 lnet_peer_t *lp = msg->msg_rxpeer;
1104 lnet_rtrbufpool_t *rbp;
1107 LASSERT (msg->msg_iov == NULL);
1108 LASSERT (msg->msg_kiov == NULL);
1109 LASSERT (msg->msg_niov == 0);
1110 LASSERT (msg->msg_routing);
1111 LASSERT (msg->msg_receiving);
1112 LASSERT (!msg->msg_sending);
1114 /* non-lnet_parse callers only send delayed messages */
1115 LASSERT (!do_recv || msg->msg_delayed);
1117 if (!msg->msg_peerrtrcredit) {
1118 LASSERT ((lp->lp_rtrcredits < 0) == !list_empty(&lp->lp_rtrq));
1120 msg->msg_peerrtrcredit = 1;
1121 lp->lp_rtrcredits--;
1122 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
1123 lp->lp_minrtrcredits = lp->lp_rtrcredits;
1125 if (lp->lp_rtrcredits < 0) {
1126 /* must have checked eager_recv before here */
1127 LASSERT (msg->msg_delayed);
1128 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1133 rbp = lnet_msg2bufpool(msg);
1135 if (!msg->msg_rtrcredit) {
1136 LASSERT ((rbp->rbp_credits < 0) == !list_empty(&rbp->rbp_msgs));
1138 msg->msg_rtrcredit = 1;
1140 if (rbp->rbp_credits < rbp->rbp_mincredits)
1141 rbp->rbp_mincredits = rbp->rbp_credits;
1143 if (rbp->rbp_credits < 0) {
1144 /* must have checked eager_recv before here */
1145 LASSERT (msg->msg_delayed);
1146 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1151 LASSERT (!list_empty(&rbp->rbp_bufs));
1152 rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
1153 list_del(&rb->rb_list);
1155 msg->msg_niov = rbp->rbp_npages;
1156 msg->msg_kiov = &rb->rb_kiov[0];
1160 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
1161 0, msg->msg_len, msg->msg_len);
1169 lnet_return_credits_locked (lnet_msg_t *msg)
1171 lnet_peer_t *txpeer = msg->msg_txpeer;
1172 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1176 if (msg->msg_txcredit) {
1177 /* give back NI txcredits */
1178 msg->msg_txcredit = 0;
1181 LASSERT((ni->ni_txcredits < 0) == !list_empty(&ni->ni_txq));
1184 if (ni->ni_txcredits <= 0) {
1185 msg2 = list_entry(ni->ni_txq.next, lnet_msg_t, msg_list);
1186 list_del(&msg2->msg_list);
1188 LASSERT(msg2->msg_txpeer->lp_ni == ni);
1189 LASSERT(msg2->msg_delayed);
1191 (void) lnet_post_send_locked(msg2, 1);
1195 if (msg->msg_peertxcredit) {
1196 /* give back peer txcredits */
1197 msg->msg_peertxcredit = 0;
1199 LASSERT((txpeer->lp_txcredits < 0) == !list_empty(&txpeer->lp_txq));
1201 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1202 LASSERT (txpeer->lp_txqnob >= 0);
1204 txpeer->lp_txcredits++;
1205 if (txpeer->lp_txcredits <= 0) {
1206 msg2 = list_entry(txpeer->lp_txq.next,
1207 lnet_msg_t, msg_list);
1208 list_del(&msg2->msg_list);
1210 LASSERT (msg2->msg_txpeer == txpeer);
1211 LASSERT (msg2->msg_delayed);
1213 (void) lnet_post_send_locked(msg2, 1);
1217 if (txpeer != NULL) {
1218 msg->msg_txpeer = NULL;
1219 lnet_peer_decref_locked(txpeer);
1223 if (msg->msg_rtrcredit) {
1224 /* give back global router credits */
1226 lnet_rtrbufpool_t *rbp;
1228 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1229 * there until it gets one allocated, or aborts the wait
1231 LASSERT (msg->msg_kiov != NULL);
1233 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1235 LASSERT (rbp == lnet_msg2bufpool(msg));
1237 msg->msg_kiov = NULL;
1238 msg->msg_rtrcredit = 0;
1240 LASSERT((rbp->rbp_credits < 0) == !list_empty(&rbp->rbp_msgs));
1241 LASSERT((rbp->rbp_credits > 0) == !list_empty(&rbp->rbp_bufs));
1243 list_add(&rb->rb_list, &rbp->rbp_bufs);
1245 if (rbp->rbp_credits <= 0) {
1246 msg2 = list_entry(rbp->rbp_msgs.next,
1247 lnet_msg_t, msg_list);
1248 list_del(&msg2->msg_list);
1250 (void) lnet_post_routed_recv_locked(msg2, 1);
1254 if (msg->msg_peerrtrcredit) {
1255 /* give back peer router credits */
1256 msg->msg_peerrtrcredit = 0;
1258 LASSERT((rxpeer->lp_rtrcredits < 0) == !list_empty(&rxpeer->lp_rtrq));
1260 rxpeer->lp_rtrcredits++;
1261 if (rxpeer->lp_rtrcredits <= 0) {
1262 msg2 = list_entry(rxpeer->lp_rtrq.next,
1263 lnet_msg_t, msg_list);
1264 list_del(&msg2->msg_list);
1266 (void) lnet_post_routed_recv_locked(msg2, 1);
1270 LASSERT (!msg->msg_rtrcredit);
1271 LASSERT (!msg->msg_peerrtrcredit);
1273 if (rxpeer != NULL) {
1274 msg->msg_rxpeer = NULL;
1275 lnet_peer_decref_locked(rxpeer);
1280 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg)
1282 lnet_nid_t dst_nid = msg->msg_target.nid;
1284 lnet_ni_t *local_ni;
1285 lnet_remotenet_t *rnet;
1286 lnet_route_t *route;
1287 lnet_route_t *best_route;
1288 struct list_head *tmp;
1293 LASSERT (msg->msg_txpeer == NULL);
1294 LASSERT (!msg->msg_sending);
1295 LASSERT (!msg->msg_target_is_router);
1296 LASSERT (!msg->msg_receiving);
1298 msg->msg_sending = 1;
1300 /* NB! ni != NULL == interface pre-determined (ACK/REPLY) */
1304 if (the_lnet.ln_shutdown) {
1309 if (src_nid == LNET_NID_ANY) {
1312 src_ni = lnet_nid2ni_locked(src_nid);
1313 if (src_ni == NULL) {
1315 CERROR("Can't send to %s: src %s is not a local nid\n",
1316 libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1319 LASSERT (!msg->msg_routing);
1322 /* Is this for someone on a local network? */
1323 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid));
1325 if (local_ni != NULL) {
1326 if (src_ni == NULL) {
1328 src_nid = src_ni->ni_nid;
1329 } else if (src_ni == local_ni) {
1330 lnet_ni_decref_locked(local_ni);
1332 lnet_ni_decref_locked(local_ni);
1333 lnet_ni_decref_locked(src_ni);
1335 CERROR("no route to %s via from %s\n",
1336 libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1340 LASSERT (src_nid != LNET_NID_ANY);
1342 if (!msg->msg_routing)
1343 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1345 if (src_ni == the_lnet.ln_loni) {
1346 /* No send credit hassles with LOLND */
1348 lnet_ni_send(src_ni, msg);
1349 lnet_ni_decref(src_ni);
1353 rc = lnet_nid2peer_locked(&lp, dst_nid);
1354 lnet_ni_decref_locked(src_ni); /* lp has ref on src_ni; lose mine */
1357 CERROR("Error %d finding peer %s\n", rc,
1358 libcfs_nid2str(dst_nid));
1359 /* ENOMEM or shutting down */
1362 LASSERT (lp->lp_ni == src_ni);
1364 /* sending to a remote network */
1365 rnet = lnet_find_net_locked(LNET_NIDNET(dst_nid));
1368 lnet_ni_decref_locked(src_ni);
1370 CERROR("No route to %s\n", libcfs_id2str(msg->msg_target));
1371 return -EHOSTUNREACH;
1374 /* Find the best gateway I can use */
1377 list_for_each(tmp, &rnet->lrn_routes) {
1378 route = list_entry(tmp, lnet_route_t, lr_list);
1379 lp2 = route->lr_gateway;
1381 if (lp2->lp_alive &&
1382 (src_ni == NULL || lp2->lp_ni == src_ni) &&
1383 (lp == NULL || lnet_compare_routers(lp2, lp) > 0)) {
1391 lnet_ni_decref_locked(src_ni);
1393 CERROR("No route to %s (all routers down)\n",
1394 libcfs_id2str(msg->msg_target));
1395 return -EHOSTUNREACH;
1398 /* Place selected route at the end of the route list to ensure
1399 * fairness; everything else being equal... */
1400 list_del(&best_route->lr_list);
1401 list_add_tail(&best_route->lr_list, &rnet->lrn_routes);
1403 if (src_ni == NULL) {
1405 src_nid = src_ni->ni_nid;
1407 LASSERT (src_ni == lp->lp_ni);
1408 lnet_ni_decref_locked(src_ni);
1411 lnet_peer_addref_locked(lp);
1413 LASSERT (src_nid != LNET_NID_ANY);
1415 if (!msg->msg_routing) {
1416 /* I'm the source and now I know which NI to send on */
1417 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1420 msg->msg_target_is_router = 1;
1421 msg->msg_target.nid = lp->lp_nid;
1422 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1425 /* 'lp' is our best choice of peer */
1427 LASSERT (!msg->msg_peertxcredit);
1428 LASSERT (!msg->msg_txcredit);
1429 LASSERT (msg->msg_txpeer == NULL);
1431 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1433 rc = lnet_post_send_locked(msg, 0);
1436 if (rc == EHOSTUNREACH)
1437 return -EHOSTUNREACH;
1440 lnet_ni_send(src_ni, msg);
1446 lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg)
1448 /* ALWAYS called holding the LNET_LOCK */
1449 /* Here, we commit the MD to a network OP by marking it busy and
1450 * decrementing its threshold. Come what may, the network "owns"
1451 * the MD until a call to lnet_finalize() signals completion. */
1452 LASSERT (!msg->msg_routing);
1457 if (md->md_threshold != LNET_MD_THRESH_INF) {
1458 LASSERT (md->md_threshold > 0);
1462 the_lnet.ln_counters.msgs_alloc++;
1463 if (the_lnet.ln_counters.msgs_alloc >
1464 the_lnet.ln_counters.msgs_max)
1465 the_lnet.ln_counters.msgs_max =
1466 the_lnet.ln_counters.msgs_alloc;
1468 LASSERT (!msg->msg_onactivelist);
1469 msg->msg_onactivelist = 1;
1470 list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
1474 lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob)
1477 the_lnet.ln_counters.drop_count++;
1478 the_lnet.ln_counters.drop_length += nob;
1481 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1485 lnet_drop_delayed_put(lnet_msg_t *msg, char *reason)
1487 lnet_process_id_t id = {0};
1489 id.nid = msg->msg_hdr.src_nid;
1490 id.pid = msg->msg_hdr.src_pid;
1492 LASSERT (msg->msg_md == NULL);
1493 LASSERT (msg->msg_delayed);
1494 LASSERT (msg->msg_rxpeer != NULL);
1495 LASSERT (msg->msg_hdr.type == LNET_MSG_PUT);
1497 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
1498 " offset %d length %d: %s\n",
1500 msg->msg_hdr.msg.put.ptl_index,
1501 msg->msg_hdr.msg.put.match_bits,
1502 msg->msg_hdr.msg.put.offset,
1503 msg->msg_hdr.payload_length,
1506 /* NB I can't drop msg's ref on msg_rxpeer until after I've
1507 * called lnet_drop_message(), so I just hang onto msg as well
1508 * until that's done */
1510 lnet_drop_message(msg->msg_rxpeer->lp_ni,
1511 msg->msg_private, msg->msg_len);
1515 lnet_peer_decref_locked(msg->msg_rxpeer);
1516 msg->msg_rxpeer = NULL;
1524 LNetSetLazyPortal(int portal)
1526 lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1528 if (portal < 0 || portal >= the_lnet.ln_nportals)
1531 CDEBUG(D_NET, "Setting portal %d lazy\n", portal);
1535 ptl->ptl_options |= LNET_PTL_LAZY;
1543 LNetClearLazyPortal(int portal)
1545 struct list_head zombies;
1546 lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1549 if (portal < 0 || portal >= the_lnet.ln_nportals)
1554 if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) {
1559 if (the_lnet.ln_shutdown)
1560 CWARN ("Active lazy portal %d on exit\n", portal);
1562 CDEBUG (D_NET, "clearing portal %d lazy\n", portal);
1564 /* grab all the blocked messages atomically */
1565 list_add(&zombies, &ptl->ptl_msgq);
1566 list_del_init(&ptl->ptl_msgq);
1568 ptl->ptl_msgq_version++;
1569 ptl->ptl_options &= ~LNET_PTL_LAZY;
1573 while (!list_empty(&zombies)) {
1574 msg = list_entry(zombies.next, lnet_msg_t, msg_list);
1575 list_del(&msg->msg_list);
1577 lnet_drop_delayed_put(msg, "Clearing lazy portal attr");
1584 lnet_recv_put(lnet_libmd_t *md, lnet_msg_t *msg, int delayed,
1585 unsigned int offset, unsigned int mlength)
1587 lnet_hdr_t *hdr = &msg->msg_hdr;
1591 the_lnet.ln_counters.recv_count++;
1592 the_lnet.ln_counters.recv_length += mlength;
1597 lnet_setpayloadbuffer(msg);
1599 msg->msg_ev.type = LNET_EVENT_PUT;
1600 msg->msg_ev.target.pid = hdr->dest_pid;
1601 msg->msg_ev.target.nid = hdr->dest_nid;
1602 msg->msg_ev.hdr_data = hdr->msg.put.hdr_data;
1604 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1605 * it back into the ACK during lnet_finalize() */
1606 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1607 (md->md_options & LNET_MD_ACK_DISABLE) == 0);
1609 lnet_ni_recv(msg->msg_rxpeer->lp_ni,
1611 msg, delayed, offset, mlength,
1612 hdr->payload_length);
1615 /* called with LNET_LOCK held */
1617 lnet_match_blocked_msg(lnet_libmd_t *md)
1619 CFS_LIST_HEAD (drops);
1620 CFS_LIST_HEAD (matches);
1621 struct list_head *tmp;
1622 struct list_head *entry;
1624 lnet_me_t *me = md->md_me;
1625 lnet_portal_t *ptl = &the_lnet.ln_portals[me->me_portal];
1627 LASSERT (me->me_portal < (unsigned int)the_lnet.ln_nportals);
1629 if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) {
1630 LASSERT (list_empty(&ptl->ptl_msgq));
1634 LASSERT (md->md_refcount == 0); /* a brand new MD */
1636 list_for_each_safe (entry, tmp, &ptl->ptl_msgq) {
1639 unsigned int mlength;
1640 unsigned int offset;
1642 lnet_process_id_t src;
1644 msg = list_entry(entry, lnet_msg_t, msg_list);
1646 LASSERT (msg->msg_delayed);
1648 hdr = &msg->msg_hdr;
1649 index = hdr->msg.put.ptl_index;
1651 src.nid = hdr->src_nid;
1652 src.pid = hdr->src_pid;
1654 rc = lnet_try_match_md(index, LNET_MD_OP_PUT, src,
1655 hdr->payload_length,
1656 hdr->msg.put.offset,
1657 hdr->msg.put.match_bits,
1658 md, msg, &mlength, &offset);
1660 if (rc == LNET_MATCHMD_NONE)
1663 /* Hurrah! This _is_ a match */
1664 list_del(&msg->msg_list);
1665 ptl->ptl_msgq_version++;
1667 if (rc == LNET_MATCHMD_OK) {
1668 list_add_tail(&msg->msg_list, &matches);
1670 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
1671 "match "LPU64" offset %d length %d.\n",
1673 hdr->msg.put.ptl_index,
1674 hdr->msg.put.match_bits,
1675 hdr->msg.put.offset,
1676 hdr->payload_length);
1678 LASSERT (rc == LNET_MATCHMD_DROP);
1680 list_add_tail(&msg->msg_list, &drops);
1683 if (lnet_md_exhausted(md))
1689 list_for_each_safe (entry, tmp, &drops) {
1690 msg = list_entry(entry, lnet_msg_t, msg_list);
1692 list_del(&msg->msg_list);
1694 lnet_drop_delayed_put(msg, "Bad match");
1697 list_for_each_safe (entry, tmp, &matches) {
1698 msg = list_entry(entry, lnet_msg_t, msg_list);
1700 list_del(&msg->msg_list);
1702 /* md won't disappear under me, since each msg
1703 * holds a ref on it */
1704 lnet_recv_put(md, msg, 1,
1706 msg->msg_ev.mlength);
1713 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1718 lnet_hdr_t *hdr = &msg->msg_hdr;
1719 unsigned int rlength = hdr->payload_length;
1720 unsigned int mlength = 0;
1721 unsigned int offset = 0;
1722 lnet_process_id_t src= {0};
1726 src.nid = hdr->src_nid;
1727 src.pid = hdr->src_pid;
1729 /* Convert put fields to host byte order */
1730 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1731 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1732 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1734 index = hdr->msg.put.ptl_index;
1735 ptl = &the_lnet.ln_portals[index];
1740 rc = lnet_match_md(index, LNET_MD_OP_PUT, src,
1741 rlength, hdr->msg.put.offset,
1742 hdr->msg.put.match_bits, msg,
1743 &mlength, &offset, &md);
1748 case LNET_MATCHMD_OK:
1750 lnet_recv_put(md, msg, msg->msg_delayed, offset, mlength);
1753 case LNET_MATCHMD_NONE:
1754 version = ptl->ptl_ml_version;
1757 if (!msg->msg_delayed)
1758 rc = lnet_eager_recv_locked(msg);
1761 !the_lnet.ln_shutdown &&
1762 ((ptl->ptl_options & LNET_PTL_LAZY) != 0)) {
1763 if (version != ptl->ptl_ml_version)
1766 list_add_tail(&msg->msg_list, &ptl->ptl_msgq);
1767 ptl->ptl_msgq_version++;
1770 CDEBUG(D_NET, "Delaying PUT from %s portal %d match "
1771 LPU64" offset %d length %d: no match \n",
1772 libcfs_id2str(src), index,
1773 hdr->msg.put.match_bits,
1774 hdr->msg.put.offset, rlength);
1779 case LNET_MATCHMD_DROP:
1781 "Dropping PUT from %s portal %d match "LPU64
1782 " offset %d length %d: %d\n",
1783 libcfs_id2str(src), index,
1784 hdr->msg.put.match_bits,
1785 hdr->msg.put.offset, rlength, rc);
1788 return ENOENT; /* +ve: OK but no match */
1793 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1795 lnet_hdr_t *hdr = &msg->msg_hdr;
1796 unsigned int mlength = 0;
1797 unsigned int offset = 0;
1798 lnet_process_id_t src = {0};
1799 lnet_handle_wire_t reply_wmd;
1803 src.nid = hdr->src_nid;
1804 src.pid = hdr->src_pid;
1806 /* Convert get fields to host byte order */
1807 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1808 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1809 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1810 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1814 rc = lnet_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src,
1815 hdr->msg.get.sink_length, hdr->msg.get.src_offset,
1816 hdr->msg.get.match_bits, msg,
1817 &mlength, &offset, &md);
1818 if (rc == LNET_MATCHMD_DROP) {
1820 "Dropping GET from %s portal %d match "LPU64
1821 " offset %d length %d\n",
1823 hdr->msg.get.ptl_index,
1824 hdr->msg.get.match_bits,
1825 hdr->msg.get.src_offset,
1826 hdr->msg.get.sink_length);
1828 return ENOENT; /* +ve: OK but no match */
1831 LASSERT (rc == LNET_MATCHMD_OK);
1833 the_lnet.ln_counters.send_count++;
1834 the_lnet.ln_counters.send_length += mlength;
1838 msg->msg_ev.type = LNET_EVENT_GET;
1839 msg->msg_ev.target.pid = hdr->dest_pid;
1840 msg->msg_ev.target.nid = hdr->dest_nid;
1841 msg->msg_ev.hdr_data = 0;
1843 reply_wmd = hdr->msg.get.return_wmd;
1845 lnet_prep_send(msg, LNET_MSG_REPLY, src, offset, mlength);
1847 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1850 /* The LND completes the REPLY from her recv procedure */
1851 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1852 msg->msg_offset, msg->msg_len, msg->msg_len);
1856 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1857 msg->msg_receiving = 0;
1859 rc = lnet_send(ni->ni_nid, msg);
1861 /* didn't get as far as lnet_ni_send() */
1862 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1863 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc);
1865 lnet_finalize(ni, msg, rc);
1872 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1874 void *private = msg->msg_private;
1875 lnet_hdr_t *hdr = &msg->msg_hdr;
1876 lnet_process_id_t src = {0};
1883 src.nid = hdr->src_nid;
1884 src.pid = hdr->src_pid;
1886 /* NB handles only looked up by creator (no flips) */
1887 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1888 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1889 CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s for %s "
1890 "MD "LPX64"."LPX64"\n",
1891 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1892 (md == NULL) ? "invalid" : "inactive",
1893 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1894 hdr->msg.reply.dst_wmd.wh_object_cookie);
1895 if (md != NULL && md->md_me != NULL)
1896 CERROR("REPLY MD also attached to portal %d\n",
1897 md->md_me->me_portal);
1900 return ENOENT; /* +ve: OK but no match */
1903 LASSERT (md->md_offset == 0);
1905 rlength = hdr->payload_length;
1906 mlength = MIN(rlength, (int)md->md_length);
1908 if (mlength < rlength &&
1909 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1910 CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s length %d "
1911 "for MD "LPX64" would overflow (%d)\n",
1912 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1913 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1916 return ENOENT; /* +ve: OK but no match */
1919 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1920 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1921 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1923 lnet_commit_md(md, msg);
1926 lnet_setpayloadbuffer(msg);
1928 msg->msg_ev.type = LNET_EVENT_REPLY;
1929 msg->msg_ev.target.pid = hdr->dest_pid;
1930 msg->msg_ev.target.nid = hdr->dest_nid;
1931 msg->msg_ev.initiator = src;
1932 msg->msg_ev.rlength = rlength;
1933 msg->msg_ev.mlength = mlength;
1934 msg->msg_ev.offset = 0;
1936 lnet_md_deconstruct(md, &msg->msg_ev.md);
1937 lnet_md2handle(&msg->msg_ev.md_handle, md);
1939 the_lnet.ln_counters.recv_count++;
1940 the_lnet.ln_counters.recv_length += mlength;
1944 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1949 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1951 lnet_hdr_t *hdr = &msg->msg_hdr;
1952 lnet_process_id_t src = {0};
1955 src.nid = hdr->src_nid;
1956 src.pid = hdr->src_pid;
1958 /* Convert ack fields to host byte order */
1959 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1960 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1964 /* NB handles only looked up by creator (no flips) */
1965 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1966 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1967 /* Don't moan; this is expected */
1969 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1970 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1971 (md == NULL) ? "invalid" : "inactive",
1972 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1973 hdr->msg.ack.dst_wmd.wh_object_cookie);
1974 if (md != NULL && md->md_me != NULL)
1975 CERROR("Source MD also attached to portal %d\n",
1976 md->md_me->me_portal);
1979 return ENOENT; /* +ve! */
1982 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1983 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1984 hdr->msg.ack.dst_wmd.wh_object_cookie);
1986 lnet_commit_md(md, msg);
1988 msg->msg_ev.type = LNET_EVENT_ACK;
1989 msg->msg_ev.target.pid = hdr->dest_pid;
1990 msg->msg_ev.target.nid = hdr->dest_nid;
1991 msg->msg_ev.initiator = src;
1992 msg->msg_ev.mlength = hdr->msg.ack.mlength;
1993 msg->msg_ev.match_bits = hdr->msg.ack.match_bits;
1995 lnet_md_deconstruct(md, &msg->msg_ev.md);
1996 lnet_md2handle(&msg->msg_ev.md_handle, md);
1998 the_lnet.ln_counters.recv_count++;
2002 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2007 lnet_msgtyp2str (int type)
2016 case LNET_MSG_REPLY:
2018 case LNET_MSG_HELLO:
2021 return ("<UNKNOWN>");
2026 lnet_print_hdr(lnet_hdr_t * hdr)
2028 lnet_process_id_t src = {0};
2029 lnet_process_id_t dst = {0};
2030 char *type_str = lnet_msgtyp2str (hdr->type);
2032 src.nid = hdr->src_nid;
2033 src.pid = hdr->src_pid;
2035 dst.nid = hdr->dest_nid;
2036 dst.pid = hdr->dest_pid;
2038 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2039 CWARN(" From %s\n", libcfs_id2str(src));
2040 CWARN(" To %s\n", libcfs_id2str(dst));
2042 switch (hdr->type) {
2047 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
2048 "match bits "LPU64"\n",
2049 hdr->msg.put.ptl_index,
2050 hdr->msg.put.ack_wmd.wh_interface_cookie,
2051 hdr->msg.put.ack_wmd.wh_object_cookie,
2052 hdr->msg.put.match_bits);
2053 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
2054 hdr->payload_length, hdr->msg.put.offset,
2055 hdr->msg.put.hdr_data);
2059 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
2060 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
2061 hdr->msg.get.return_wmd.wh_interface_cookie,
2062 hdr->msg.get.return_wmd.wh_object_cookie,
2063 hdr->msg.get.match_bits);
2064 CWARN(" Length %d, src offset %d\n",
2065 hdr->msg.get.sink_length,
2066 hdr->msg.get.src_offset);
2070 CWARN(" dst md "LPX64"."LPX64", "
2071 "manipulated length %d\n",
2072 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2073 hdr->msg.ack.dst_wmd.wh_object_cookie,
2074 hdr->msg.ack.mlength);
2077 case LNET_MSG_REPLY:
2078 CWARN(" dst md "LPX64"."LPX64", "
2080 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2081 hdr->msg.reply.dst_wmd.wh_object_cookie,
2082 hdr->payload_length);
2089 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
2090 void *private, int rdma_req)
2095 lnet_pid_t dest_pid;
2096 lnet_nid_t dest_nid;
2098 __u32 payload_length;
2101 LASSERT (!in_interrupt ());
2103 type = le32_to_cpu(hdr->type);
2104 src_nid = le64_to_cpu(hdr->src_nid);
2105 dest_nid = le64_to_cpu(hdr->dest_nid);
2106 dest_pid = le32_to_cpu(hdr->dest_pid);
2107 payload_length = le32_to_cpu(hdr->payload_length);
2109 for_me = (ni->ni_nid == dest_nid);
2114 if (payload_length > 0) {
2115 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2116 libcfs_nid2str(from_nid),
2117 libcfs_nid2str(src_nid),
2118 lnet_msgtyp2str(type), payload_length);
2124 case LNET_MSG_REPLY:
2125 if (payload_length > (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2126 CERROR("%s, src %s: bad %s payload %d "
2127 "(%d max expected)\n",
2128 libcfs_nid2str(from_nid),
2129 libcfs_nid2str(src_nid),
2130 lnet_msgtyp2str(type),
2132 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2138 CERROR("%s, src %s: Bad message type 0x%x\n",
2139 libcfs_nid2str(from_nid),
2140 libcfs_nid2str(src_nid), type);
2144 /* Regard a bad destination NID as a protocol error. Senders should
2145 * know what they're doing; if they don't they're misconfigured, buggy
2146 * or malicious so we chop them off at the knees :) */
2149 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2150 /* should have gone direct */
2151 CERROR ("%s, src %s: Bad dest nid %s "
2152 "(should have been sent direct)\n",
2153 libcfs_nid2str(from_nid),
2154 libcfs_nid2str(src_nid),
2155 libcfs_nid2str(dest_nid));
2159 if (lnet_islocalnid(dest_nid)) {
2160 /* dest is another local NI; sender should have used
2161 * this node's NID on its own network */
2162 CERROR ("%s, src %s: Bad dest nid %s "
2163 "(it's my nid but on a different network)\n",
2164 libcfs_nid2str(from_nid),
2165 libcfs_nid2str(src_nid),
2166 libcfs_nid2str(dest_nid));
2170 if (rdma_req && type == LNET_MSG_GET) {
2171 CERROR ("%s, src %s: Bad optimized GET for %s "
2172 "(final destination must be me)\n",
2173 libcfs_nid2str(from_nid),
2174 libcfs_nid2str(src_nid),
2175 libcfs_nid2str(dest_nid));
2179 if (!the_lnet.ln_routing) {
2180 CERROR ("%s, src %s: Dropping message for %s "
2181 "(routing not enabled)\n",
2182 libcfs_nid2str(from_nid),
2183 libcfs_nid2str(src_nid),
2184 libcfs_nid2str(dest_nid));
2189 /* Message looks OK; we're not going to return an error, so we MUST
2190 * call back lnd_recv() come what may... */
2192 if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2193 fail_peer (src_nid, 0)) /* shall we now? */
2195 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2196 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2197 lnet_msgtyp2str(type));
2201 msg = lnet_msg_alloc();
2203 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2204 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2205 lnet_msgtyp2str(type));
2209 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
2211 msg->msg_type = type;
2212 msg->msg_private = private;
2213 msg->msg_receiving = 1;
2214 msg->msg_len = msg->msg_wanted = payload_length;
2215 msg->msg_offset = 0;
2216 msg->msg_hdr = *hdr;
2219 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid);
2222 CERROR("%s, src %s: Dropping %s "
2223 "(error %d looking up sender)\n",
2224 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2225 lnet_msgtyp2str(type), rc);
2234 msg->msg_target.pid = dest_pid;
2235 msg->msg_target.nid = dest_nid;
2236 msg->msg_routing = 1;
2237 msg->msg_offset = 0;
2240 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
2241 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2242 rc = lnet_eager_recv_locked(msg);
2248 lnet_commit_routedmsg(msg);
2249 rc = lnet_post_routed_recv_locked(msg, 0);
2253 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2254 0, payload_length, payload_length);
2258 /* convert common msg->hdr fields to host byteorder */
2259 msg->msg_hdr.type = type;
2260 msg->msg_hdr.src_nid = src_nid;
2261 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2262 msg->msg_hdr.dest_nid = dest_nid;
2263 msg->msg_hdr.dest_pid = dest_pid;
2264 msg->msg_hdr.payload_length = payload_length;
2266 msg->msg_ev.sender = from_nid;
2270 rc = lnet_parse_ack(ni, msg);
2273 rc = lnet_parse_put(ni, msg);
2276 rc = lnet_parse_get(ni, msg, rdma_req);
2278 case LNET_MSG_REPLY:
2279 rc = lnet_parse_reply(ni, msg);
2283 goto free_drop; /* prevent an unused label if !kernel */
2289 LASSERT (rc == ENOENT);
2292 LASSERT (msg->msg_md == NULL);
2294 if (msg->msg_rxpeer != NULL) {
2295 lnet_peer_decref_locked(msg->msg_rxpeer);
2296 msg->msg_rxpeer = NULL;
2298 lnet_msg_free(msg); /* expects LNET_LOCK held */
2302 lnet_drop_message(ni, private, payload_length);
2307 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2308 lnet_process_id_t target, unsigned int portal,
2309 __u64 match_bits, unsigned int offset,
2316 LASSERT (the_lnet.ln_init);
2317 LASSERT (the_lnet.ln_refcount > 0);
2319 if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2320 fail_peer (target.nid, 1)) /* shall we now? */
2322 CERROR("Dropping PUT to %s: simulated failure\n",
2323 libcfs_id2str(target));
2327 msg = lnet_msg_alloc();
2329 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2330 libcfs_id2str(target));
2336 md = lnet_handle2md(&mdh);
2337 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2340 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2341 match_bits, portal, libcfs_id2str(target),
2342 md == NULL ? -1 : md->md_threshold);
2343 if (md != NULL && md->md_me != NULL)
2344 CERROR("Source MD also attached to portal %d\n",
2345 md->md_me->me_portal);
2351 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2353 lnet_commit_md(md, msg);
2355 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2357 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2358 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2359 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2360 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2362 /* NB handles only looked up by creator (no flips) */
2363 if (ack == LNET_ACK_REQ) {
2364 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2365 the_lnet.ln_interface_cookie;
2366 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2367 md->md_lh.lh_cookie;
2369 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2370 LNET_WIRE_HANDLE_COOKIE_NONE;
2371 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2372 LNET_WIRE_HANDLE_COOKIE_NONE;
2375 msg->msg_ev.type = LNET_EVENT_SEND;
2376 msg->msg_ev.initiator.nid = LNET_NID_ANY;
2377 msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2378 msg->msg_ev.target = target;
2379 msg->msg_ev.sender = LNET_NID_ANY;
2380 msg->msg_ev.pt_index = portal;
2381 msg->msg_ev.match_bits = match_bits;
2382 msg->msg_ev.rlength = md->md_length;
2383 msg->msg_ev.mlength = md->md_length;
2384 msg->msg_ev.offset = offset;
2385 msg->msg_ev.hdr_data = hdr_data;
2387 lnet_md_deconstruct(md, &msg->msg_ev.md);
2388 lnet_md2handle(&msg->msg_ev.md_handle, md);
2390 the_lnet.ln_counters.send_count++;
2391 the_lnet.ln_counters.send_length += md->md_length;
2395 rc = lnet_send(self, msg);
2397 CERROR("Error sending PUT to %s: %d\n",
2398 libcfs_id2str(target), rc);
2399 lnet_finalize (NULL, msg, rc);
2402 /* completion will be signalled by an event */
2407 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2409 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2410 * returns a msg for the LND to pass to lnet_finalize() when the sink
2411 * data has been received.
2413 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2414 * lnet_finalize() is called on it, so the LND must call this first */
2416 lnet_msg_t *msg = lnet_msg_alloc();
2417 lnet_libmd_t *getmd = getmsg->msg_md;
2418 lnet_process_id_t peer_id = getmsg->msg_target;
2420 LASSERT (!getmsg->msg_target_is_router);
2421 LASSERT (!getmsg->msg_routing);
2425 LASSERT (getmd->md_refcount > 0);
2428 CERROR ("%s: Dropping REPLY from %s: can't allocate msg\n",
2429 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2433 if (getmd->md_threshold == 0) {
2434 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2435 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2440 LASSERT (getmd->md_offset == 0);
2442 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2443 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2445 lnet_commit_md (getmd, msg);
2447 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2449 msg->msg_ev.type = LNET_EVENT_REPLY;
2450 msg->msg_ev.initiator = peer_id;
2451 msg->msg_ev.sender = peer_id.nid; /* optimized GETs can't be routed */
2452 msg->msg_ev.rlength = msg->msg_ev.mlength = getmd->md_length;
2453 msg->msg_ev.offset = 0;
2455 lnet_md_deconstruct(getmd, &msg->msg_ev.md);
2456 lnet_md2handle(&msg->msg_ev.md_handle, getmd);
2458 the_lnet.ln_counters.recv_count++;
2459 the_lnet.ln_counters.recv_length += getmd->md_length;
2468 the_lnet.ln_counters.drop_count++;
2469 the_lnet.ln_counters.drop_length += getmd->md_length;
2477 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2479 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2480 * completed and I know it. */
2481 LASSERT (reply != NULL);
2482 LASSERT (reply->msg_type == LNET_MSG_GET);
2483 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2485 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2486 * the end of my buffer, I might as well be dead. */
2487 LASSERT (len <= reply->msg_ev.mlength);
2489 reply->msg_ev.mlength = len;
2493 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2494 lnet_process_id_t target, unsigned int portal,
2495 __u64 match_bits, unsigned int offset)
2501 LASSERT (the_lnet.ln_init);
2502 LASSERT (the_lnet.ln_refcount > 0);
2504 if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2505 fail_peer (target.nid, 1)) /* shall we now? */
2507 CERROR("Dropping GET to %s: simulated failure\n",
2508 libcfs_id2str(target));
2512 msg = lnet_msg_alloc();
2514 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2515 libcfs_id2str(target));
2521 md = lnet_handle2md(&mdh);
2522 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2525 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2526 match_bits, portal, libcfs_id2str(target),
2527 md == NULL ? -1 : md->md_threshold);
2528 if (md != NULL && md->md_me != NULL)
2529 CERROR("REPLY MD also attached to portal %d\n",
2530 md->md_me->me_portal);
2536 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2538 lnet_commit_md(md, msg);
2540 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2542 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2543 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2544 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2545 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2547 /* NB handles only looked up by creator (no flips) */
2548 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2549 the_lnet.ln_interface_cookie;
2550 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2551 md->md_lh.lh_cookie;
2553 msg->msg_ev.type = LNET_EVENT_SEND;
2554 msg->msg_ev.initiator.nid = LNET_NID_ANY;
2555 msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2556 msg->msg_ev.target = target;
2557 msg->msg_ev.sender = LNET_NID_ANY;
2558 msg->msg_ev.pt_index = portal;
2559 msg->msg_ev.match_bits = match_bits;
2560 msg->msg_ev.rlength = md->md_length;
2561 msg->msg_ev.mlength = md->md_length;
2562 msg->msg_ev.offset = offset;
2563 msg->msg_ev.hdr_data = 0;
2565 lnet_md_deconstruct(md, &msg->msg_ev.md);
2566 lnet_md2handle(&msg->msg_ev.md_handle, md);
2568 the_lnet.ln_counters.send_count++;
2572 rc = lnet_send(self, msg);
2574 CERROR("error sending GET to %s: %d\n",
2575 libcfs_id2str(target), rc);
2576 lnet_finalize (NULL, msg, rc);
2579 /* completion will be signalled by an event */
2584 LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2586 struct list_head *e;
2588 lnet_route_t *route;
2589 lnet_remotenet_t *rnet;
2590 __u32 dstnet = LNET_NIDNET(dstnid);
2594 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2595 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2596 * keep order 0 free for 0@lo and order 1 free for a local NID
2599 LASSERT (the_lnet.ln_init);
2600 LASSERT (the_lnet.ln_refcount > 0);
2604 list_for_each (e, &the_lnet.ln_nis) {
2605 ni = list_entry(e, lnet_ni_t, ni_list);
2607 if (ni->ni_nid == dstnid) {
2608 if (srcnidp != NULL)
2610 if (orderp != NULL) {
2611 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2618 return local_nid_dist_zero ? 0 : 1;
2621 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2622 if (srcnidp != NULL)
2623 *srcnidp = ni->ni_nid;
2633 list_for_each (e, &the_lnet.ln_remote_nets) {
2634 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2636 if (rnet->lrn_net == dstnet) {
2637 LASSERT (!list_empty(&rnet->lrn_routes));
2638 route = list_entry(rnet->lrn_routes.next,
2639 lnet_route_t, lr_list);
2640 hops = rnet->lrn_hops;
2641 if (srcnidp != NULL)
2642 *srcnidp = route->lr_gateway->lp_ni->ni_nid;
2652 return -EHOSTUNREACH;
2656 LNetSetAsync(lnet_process_id_t id, int nasync)
2662 lnet_remotenet_t *rnet;
2663 struct list_head *tmp;
2664 lnet_route_t *route;
2671 /* Target on a local network? */
2673 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2675 if (ni->ni_lnd->lnd_setasync != NULL)
2676 rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2681 /* Target on a remote network: apply to routers */
2683 LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2688 /* Snapshot all the router NIDs */
2690 rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2692 list_for_each(tmp, &rnet->lrn_routes) {
2693 if (nnids == maxnids) {
2695 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2700 route = list_entry(tmp, lnet_route_t, lr_list);
2701 nids[nnids++] = route->lr_gateway->lp_nid;
2706 /* set async on all the routers */
2707 while (nnids-- > 0) {
2708 id.pid = LUSTRE_SRV_LNET_PID;
2709 id.nid = nids[nnids];
2711 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2715 if (ni->ni_lnd->lnd_setasync != NULL) {
2716 rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2723 LIBCFS_FREE(nids, maxnids * sizeof(*nids));