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);
1368 * - once application finishes computation, check here to update
1369 * router states before it waits for pending IO in LNetEQPoll
1370 * - recursion breaker: router checker sends no message
1371 * to remote networks */
1372 if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING)
1373 lnet_router_checker();
1377 /* sending to a remote network */
1378 rnet = lnet_find_net_locked(LNET_NIDNET(dst_nid));
1381 lnet_ni_decref_locked(src_ni);
1383 CERROR("No route to %s\n", libcfs_id2str(msg->msg_target));
1384 return -EHOSTUNREACH;
1387 /* Find the best gateway I can use */
1390 list_for_each(tmp, &rnet->lrn_routes) {
1391 route = list_entry(tmp, lnet_route_t, lr_list);
1392 lp2 = route->lr_gateway;
1394 if (lp2->lp_alive &&
1395 (src_ni == NULL || lp2->lp_ni == src_ni) &&
1396 (lp == NULL || lnet_compare_routers(lp2, lp) > 0)) {
1404 lnet_ni_decref_locked(src_ni);
1406 CERROR("No route to %s (all routers down)\n",
1407 libcfs_id2str(msg->msg_target));
1408 return -EHOSTUNREACH;
1411 /* Place selected route at the end of the route list to ensure
1412 * fairness; everything else being equal... */
1413 list_del(&best_route->lr_list);
1414 list_add_tail(&best_route->lr_list, &rnet->lrn_routes);
1416 if (src_ni == NULL) {
1418 src_nid = src_ni->ni_nid;
1420 LASSERT (src_ni == lp->lp_ni);
1421 lnet_ni_decref_locked(src_ni);
1424 lnet_peer_addref_locked(lp);
1426 LASSERT (src_nid != LNET_NID_ANY);
1428 if (!msg->msg_routing) {
1429 /* I'm the source and now I know which NI to send on */
1430 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1433 msg->msg_target_is_router = 1;
1434 msg->msg_target.nid = lp->lp_nid;
1435 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1438 /* 'lp' is our best choice of peer */
1440 LASSERT (!msg->msg_peertxcredit);
1441 LASSERT (!msg->msg_txcredit);
1442 LASSERT (msg->msg_txpeer == NULL);
1444 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1446 rc = lnet_post_send_locked(msg, 0);
1449 if (rc == EHOSTUNREACH)
1450 return -EHOSTUNREACH;
1453 lnet_ni_send(src_ni, msg);
1459 lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg)
1461 /* ALWAYS called holding the LNET_LOCK */
1462 /* Here, we commit the MD to a network OP by marking it busy and
1463 * decrementing its threshold. Come what may, the network "owns"
1464 * the MD until a call to lnet_finalize() signals completion. */
1465 LASSERT (!msg->msg_routing);
1470 if (md->md_threshold != LNET_MD_THRESH_INF) {
1471 LASSERT (md->md_threshold > 0);
1475 the_lnet.ln_counters.msgs_alloc++;
1476 if (the_lnet.ln_counters.msgs_alloc >
1477 the_lnet.ln_counters.msgs_max)
1478 the_lnet.ln_counters.msgs_max =
1479 the_lnet.ln_counters.msgs_alloc;
1481 LASSERT (!msg->msg_onactivelist);
1482 msg->msg_onactivelist = 1;
1483 list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
1487 lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob)
1490 the_lnet.ln_counters.drop_count++;
1491 the_lnet.ln_counters.drop_length += nob;
1494 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1498 lnet_drop_delayed_put(lnet_msg_t *msg, char *reason)
1500 lnet_process_id_t id = {0};
1502 id.nid = msg->msg_hdr.src_nid;
1503 id.pid = msg->msg_hdr.src_pid;
1505 LASSERT (msg->msg_md == NULL);
1506 LASSERT (msg->msg_delayed);
1507 LASSERT (msg->msg_rxpeer != NULL);
1508 LASSERT (msg->msg_hdr.type == LNET_MSG_PUT);
1510 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
1511 " offset %d length %d: %s\n",
1513 msg->msg_hdr.msg.put.ptl_index,
1514 msg->msg_hdr.msg.put.match_bits,
1515 msg->msg_hdr.msg.put.offset,
1516 msg->msg_hdr.payload_length,
1519 /* NB I can't drop msg's ref on msg_rxpeer until after I've
1520 * called lnet_drop_message(), so I just hang onto msg as well
1521 * until that's done */
1523 lnet_drop_message(msg->msg_rxpeer->lp_ni,
1524 msg->msg_private, msg->msg_len);
1528 lnet_peer_decref_locked(msg->msg_rxpeer);
1529 msg->msg_rxpeer = NULL;
1537 LNetSetLazyPortal(int portal)
1539 lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1541 if (portal < 0 || portal >= the_lnet.ln_nportals)
1544 CDEBUG(D_NET, "Setting portal %d lazy\n", portal);
1548 ptl->ptl_options |= LNET_PTL_LAZY;
1556 LNetClearLazyPortal(int portal)
1558 struct list_head zombies;
1559 lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1562 if (portal < 0 || portal >= the_lnet.ln_nportals)
1567 if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) {
1572 if (the_lnet.ln_shutdown)
1573 CWARN ("Active lazy portal %d on exit\n", portal);
1575 CDEBUG (D_NET, "clearing portal %d lazy\n", portal);
1577 /* grab all the blocked messages atomically */
1578 list_add(&zombies, &ptl->ptl_msgq);
1579 list_del_init(&ptl->ptl_msgq);
1581 ptl->ptl_msgq_version++;
1582 ptl->ptl_options &= ~LNET_PTL_LAZY;
1586 while (!list_empty(&zombies)) {
1587 msg = list_entry(zombies.next, lnet_msg_t, msg_list);
1588 list_del(&msg->msg_list);
1590 lnet_drop_delayed_put(msg, "Clearing lazy portal attr");
1597 lnet_recv_put(lnet_libmd_t *md, lnet_msg_t *msg, int delayed,
1598 unsigned int offset, unsigned int mlength)
1600 lnet_hdr_t *hdr = &msg->msg_hdr;
1604 the_lnet.ln_counters.recv_count++;
1605 the_lnet.ln_counters.recv_length += mlength;
1610 lnet_setpayloadbuffer(msg);
1612 msg->msg_ev.type = LNET_EVENT_PUT;
1613 msg->msg_ev.target.pid = hdr->dest_pid;
1614 msg->msg_ev.target.nid = hdr->dest_nid;
1615 msg->msg_ev.hdr_data = hdr->msg.put.hdr_data;
1617 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1618 * it back into the ACK during lnet_finalize() */
1619 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1620 (md->md_options & LNET_MD_ACK_DISABLE) == 0);
1622 lnet_ni_recv(msg->msg_rxpeer->lp_ni,
1624 msg, delayed, offset, mlength,
1625 hdr->payload_length);
1628 /* called with LNET_LOCK held */
1630 lnet_match_blocked_msg(lnet_libmd_t *md)
1632 CFS_LIST_HEAD (drops);
1633 CFS_LIST_HEAD (matches);
1634 struct list_head *tmp;
1635 struct list_head *entry;
1637 lnet_me_t *me = md->md_me;
1638 lnet_portal_t *ptl = &the_lnet.ln_portals[me->me_portal];
1640 LASSERT (me->me_portal < (unsigned int)the_lnet.ln_nportals);
1642 if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) {
1643 LASSERT (list_empty(&ptl->ptl_msgq));
1647 LASSERT (md->md_refcount == 0); /* a brand new MD */
1649 list_for_each_safe (entry, tmp, &ptl->ptl_msgq) {
1652 unsigned int mlength;
1653 unsigned int offset;
1655 lnet_process_id_t src;
1657 msg = list_entry(entry, lnet_msg_t, msg_list);
1659 LASSERT (msg->msg_delayed);
1661 hdr = &msg->msg_hdr;
1662 index = hdr->msg.put.ptl_index;
1664 src.nid = hdr->src_nid;
1665 src.pid = hdr->src_pid;
1667 rc = lnet_try_match_md(index, LNET_MD_OP_PUT, src,
1668 hdr->payload_length,
1669 hdr->msg.put.offset,
1670 hdr->msg.put.match_bits,
1671 md, msg, &mlength, &offset);
1673 if (rc == LNET_MATCHMD_NONE)
1676 /* Hurrah! This _is_ a match */
1677 list_del(&msg->msg_list);
1678 ptl->ptl_msgq_version++;
1680 if (rc == LNET_MATCHMD_OK) {
1681 list_add_tail(&msg->msg_list, &matches);
1683 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
1684 "match "LPU64" offset %d length %d.\n",
1686 hdr->msg.put.ptl_index,
1687 hdr->msg.put.match_bits,
1688 hdr->msg.put.offset,
1689 hdr->payload_length);
1691 LASSERT (rc == LNET_MATCHMD_DROP);
1693 list_add_tail(&msg->msg_list, &drops);
1696 if (lnet_md_exhausted(md))
1702 list_for_each_safe (entry, tmp, &drops) {
1703 msg = list_entry(entry, lnet_msg_t, msg_list);
1705 list_del(&msg->msg_list);
1707 lnet_drop_delayed_put(msg, "Bad match");
1710 list_for_each_safe (entry, tmp, &matches) {
1711 msg = list_entry(entry, lnet_msg_t, msg_list);
1713 list_del(&msg->msg_list);
1715 /* md won't disappear under me, since each msg
1716 * holds a ref on it */
1717 lnet_recv_put(md, msg, 1,
1719 msg->msg_ev.mlength);
1726 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1731 lnet_hdr_t *hdr = &msg->msg_hdr;
1732 unsigned int rlength = hdr->payload_length;
1733 unsigned int mlength = 0;
1734 unsigned int offset = 0;
1735 lnet_process_id_t src= {0};
1739 src.nid = hdr->src_nid;
1740 src.pid = hdr->src_pid;
1742 /* Convert put fields to host byte order */
1743 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1744 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1745 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1747 index = hdr->msg.put.ptl_index;
1748 ptl = &the_lnet.ln_portals[index];
1753 rc = lnet_match_md(index, LNET_MD_OP_PUT, src,
1754 rlength, hdr->msg.put.offset,
1755 hdr->msg.put.match_bits, msg,
1756 &mlength, &offset, &md);
1761 case LNET_MATCHMD_OK:
1763 lnet_recv_put(md, msg, msg->msg_delayed, offset, mlength);
1766 case LNET_MATCHMD_NONE:
1767 version = ptl->ptl_ml_version;
1770 if (!msg->msg_delayed)
1771 rc = lnet_eager_recv_locked(msg);
1774 !the_lnet.ln_shutdown &&
1775 ((ptl->ptl_options & LNET_PTL_LAZY) != 0)) {
1776 if (version != ptl->ptl_ml_version)
1779 list_add_tail(&msg->msg_list, &ptl->ptl_msgq);
1780 ptl->ptl_msgq_version++;
1783 CDEBUG(D_NET, "Delaying PUT from %s portal %d match "
1784 LPU64" offset %d length %d: no match \n",
1785 libcfs_id2str(src), index,
1786 hdr->msg.put.match_bits,
1787 hdr->msg.put.offset, rlength);
1792 case LNET_MATCHMD_DROP:
1794 "Dropping PUT from %s portal %d match "LPU64
1795 " offset %d length %d: %d\n",
1796 libcfs_id2str(src), index,
1797 hdr->msg.put.match_bits,
1798 hdr->msg.put.offset, rlength, rc);
1801 return ENOENT; /* +ve: OK but no match */
1806 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1808 lnet_hdr_t *hdr = &msg->msg_hdr;
1809 unsigned int mlength = 0;
1810 unsigned int offset = 0;
1811 lnet_process_id_t src = {0};
1812 lnet_handle_wire_t reply_wmd;
1816 src.nid = hdr->src_nid;
1817 src.pid = hdr->src_pid;
1819 /* Convert get fields to host byte order */
1820 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1821 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1822 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1823 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1827 rc = lnet_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src,
1828 hdr->msg.get.sink_length, hdr->msg.get.src_offset,
1829 hdr->msg.get.match_bits, msg,
1830 &mlength, &offset, &md);
1831 if (rc == LNET_MATCHMD_DROP) {
1833 "Dropping GET from %s portal %d match "LPU64
1834 " offset %d length %d\n",
1836 hdr->msg.get.ptl_index,
1837 hdr->msg.get.match_bits,
1838 hdr->msg.get.src_offset,
1839 hdr->msg.get.sink_length);
1841 return ENOENT; /* +ve: OK but no match */
1844 LASSERT (rc == LNET_MATCHMD_OK);
1846 the_lnet.ln_counters.send_count++;
1847 the_lnet.ln_counters.send_length += mlength;
1851 msg->msg_ev.type = LNET_EVENT_GET;
1852 msg->msg_ev.target.pid = hdr->dest_pid;
1853 msg->msg_ev.target.nid = hdr->dest_nid;
1854 msg->msg_ev.hdr_data = 0;
1856 reply_wmd = hdr->msg.get.return_wmd;
1858 lnet_prep_send(msg, LNET_MSG_REPLY, src, offset, mlength);
1860 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1863 /* The LND completes the REPLY from her recv procedure */
1864 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1865 msg->msg_offset, msg->msg_len, msg->msg_len);
1869 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1870 msg->msg_receiving = 0;
1872 rc = lnet_send(ni->ni_nid, msg);
1874 /* didn't get as far as lnet_ni_send() */
1875 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1876 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc);
1878 lnet_finalize(ni, msg, rc);
1885 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1887 void *private = msg->msg_private;
1888 lnet_hdr_t *hdr = &msg->msg_hdr;
1889 lnet_process_id_t src = {0};
1896 src.nid = hdr->src_nid;
1897 src.pid = hdr->src_pid;
1899 /* NB handles only looked up by creator (no flips) */
1900 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1901 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1902 CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s for %s "
1903 "MD "LPX64"."LPX64"\n",
1904 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1905 (md == NULL) ? "invalid" : "inactive",
1906 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1907 hdr->msg.reply.dst_wmd.wh_object_cookie);
1908 if (md != NULL && md->md_me != NULL)
1909 CERROR("REPLY MD also attached to portal %d\n",
1910 md->md_me->me_portal);
1913 return ENOENT; /* +ve: OK but no match */
1916 LASSERT (md->md_offset == 0);
1918 rlength = hdr->payload_length;
1919 mlength = MIN(rlength, (int)md->md_length);
1921 if (mlength < rlength &&
1922 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1923 CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s length %d "
1924 "for MD "LPX64" would overflow (%d)\n",
1925 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1926 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1929 return ENOENT; /* +ve: OK but no match */
1932 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1933 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1934 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1936 lnet_commit_md(md, msg);
1939 lnet_setpayloadbuffer(msg);
1941 msg->msg_ev.type = LNET_EVENT_REPLY;
1942 msg->msg_ev.target.pid = hdr->dest_pid;
1943 msg->msg_ev.target.nid = hdr->dest_nid;
1944 msg->msg_ev.initiator = src;
1945 msg->msg_ev.rlength = rlength;
1946 msg->msg_ev.mlength = mlength;
1947 msg->msg_ev.offset = 0;
1949 lnet_md_deconstruct(md, &msg->msg_ev.md);
1950 lnet_md2handle(&msg->msg_ev.md_handle, md);
1952 the_lnet.ln_counters.recv_count++;
1953 the_lnet.ln_counters.recv_length += mlength;
1957 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1962 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1964 lnet_hdr_t *hdr = &msg->msg_hdr;
1965 lnet_process_id_t src = {0};
1968 src.nid = hdr->src_nid;
1969 src.pid = hdr->src_pid;
1971 /* Convert ack fields to host byte order */
1972 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1973 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1977 /* NB handles only looked up by creator (no flips) */
1978 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1979 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1980 /* Don't moan; this is expected */
1982 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1983 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1984 (md == NULL) ? "invalid" : "inactive",
1985 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1986 hdr->msg.ack.dst_wmd.wh_object_cookie);
1987 if (md != NULL && md->md_me != NULL)
1988 CERROR("Source MD also attached to portal %d\n",
1989 md->md_me->me_portal);
1992 return ENOENT; /* +ve! */
1995 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1996 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1997 hdr->msg.ack.dst_wmd.wh_object_cookie);
1999 lnet_commit_md(md, msg);
2001 msg->msg_ev.type = LNET_EVENT_ACK;
2002 msg->msg_ev.target.pid = hdr->dest_pid;
2003 msg->msg_ev.target.nid = hdr->dest_nid;
2004 msg->msg_ev.initiator = src;
2005 msg->msg_ev.mlength = hdr->msg.ack.mlength;
2006 msg->msg_ev.match_bits = hdr->msg.ack.match_bits;
2008 lnet_md_deconstruct(md, &msg->msg_ev.md);
2009 lnet_md2handle(&msg->msg_ev.md_handle, md);
2011 the_lnet.ln_counters.recv_count++;
2015 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2020 lnet_msgtyp2str (int type)
2029 case LNET_MSG_REPLY:
2031 case LNET_MSG_HELLO:
2034 return ("<UNKNOWN>");
2039 lnet_print_hdr(lnet_hdr_t * hdr)
2041 lnet_process_id_t src = {0};
2042 lnet_process_id_t dst = {0};
2043 char *type_str = lnet_msgtyp2str (hdr->type);
2045 src.nid = hdr->src_nid;
2046 src.pid = hdr->src_pid;
2048 dst.nid = hdr->dest_nid;
2049 dst.pid = hdr->dest_pid;
2051 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2052 CWARN(" From %s\n", libcfs_id2str(src));
2053 CWARN(" To %s\n", libcfs_id2str(dst));
2055 switch (hdr->type) {
2060 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
2061 "match bits "LPU64"\n",
2062 hdr->msg.put.ptl_index,
2063 hdr->msg.put.ack_wmd.wh_interface_cookie,
2064 hdr->msg.put.ack_wmd.wh_object_cookie,
2065 hdr->msg.put.match_bits);
2066 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
2067 hdr->payload_length, hdr->msg.put.offset,
2068 hdr->msg.put.hdr_data);
2072 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
2073 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
2074 hdr->msg.get.return_wmd.wh_interface_cookie,
2075 hdr->msg.get.return_wmd.wh_object_cookie,
2076 hdr->msg.get.match_bits);
2077 CWARN(" Length %d, src offset %d\n",
2078 hdr->msg.get.sink_length,
2079 hdr->msg.get.src_offset);
2083 CWARN(" dst md "LPX64"."LPX64", "
2084 "manipulated length %d\n",
2085 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2086 hdr->msg.ack.dst_wmd.wh_object_cookie,
2087 hdr->msg.ack.mlength);
2090 case LNET_MSG_REPLY:
2091 CWARN(" dst md "LPX64"."LPX64", "
2093 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2094 hdr->msg.reply.dst_wmd.wh_object_cookie,
2095 hdr->payload_length);
2102 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
2103 void *private, int rdma_req)
2108 lnet_pid_t dest_pid;
2109 lnet_nid_t dest_nid;
2111 __u32 payload_length;
2114 LASSERT (!in_interrupt ());
2116 type = le32_to_cpu(hdr->type);
2117 src_nid = le64_to_cpu(hdr->src_nid);
2118 dest_nid = le64_to_cpu(hdr->dest_nid);
2119 dest_pid = le32_to_cpu(hdr->dest_pid);
2120 payload_length = le32_to_cpu(hdr->payload_length);
2122 for_me = (ni->ni_nid == dest_nid);
2127 if (payload_length > 0) {
2128 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2129 libcfs_nid2str(from_nid),
2130 libcfs_nid2str(src_nid),
2131 lnet_msgtyp2str(type), payload_length);
2137 case LNET_MSG_REPLY:
2138 if (payload_length > (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2139 CERROR("%s, src %s: bad %s payload %d "
2140 "(%d max expected)\n",
2141 libcfs_nid2str(from_nid),
2142 libcfs_nid2str(src_nid),
2143 lnet_msgtyp2str(type),
2145 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2151 CERROR("%s, src %s: Bad message type 0x%x\n",
2152 libcfs_nid2str(from_nid),
2153 libcfs_nid2str(src_nid), type);
2157 /* Regard a bad destination NID as a protocol error. Senders should
2158 * know what they're doing; if they don't they're misconfigured, buggy
2159 * or malicious so we chop them off at the knees :) */
2162 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2163 /* should have gone direct */
2164 CERROR ("%s, src %s: Bad dest nid %s "
2165 "(should have been sent direct)\n",
2166 libcfs_nid2str(from_nid),
2167 libcfs_nid2str(src_nid),
2168 libcfs_nid2str(dest_nid));
2172 if (lnet_islocalnid(dest_nid)) {
2173 /* dest is another local NI; sender should have used
2174 * this node's NID on its own network */
2175 CERROR ("%s, src %s: Bad dest nid %s "
2176 "(it's my nid but on a different network)\n",
2177 libcfs_nid2str(from_nid),
2178 libcfs_nid2str(src_nid),
2179 libcfs_nid2str(dest_nid));
2183 if (rdma_req && type == LNET_MSG_GET) {
2184 CERROR ("%s, src %s: Bad optimized GET for %s "
2185 "(final destination must be me)\n",
2186 libcfs_nid2str(from_nid),
2187 libcfs_nid2str(src_nid),
2188 libcfs_nid2str(dest_nid));
2192 if (!the_lnet.ln_routing) {
2193 CERROR ("%s, src %s: Dropping message for %s "
2194 "(routing not enabled)\n",
2195 libcfs_nid2str(from_nid),
2196 libcfs_nid2str(src_nid),
2197 libcfs_nid2str(dest_nid));
2202 /* Message looks OK; we're not going to return an error, so we MUST
2203 * call back lnd_recv() come what may... */
2205 if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2206 fail_peer (src_nid, 0)) /* shall we now? */
2208 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2209 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2210 lnet_msgtyp2str(type));
2214 msg = lnet_msg_alloc();
2216 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2217 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2218 lnet_msgtyp2str(type));
2222 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
2224 msg->msg_type = type;
2225 msg->msg_private = private;
2226 msg->msg_receiving = 1;
2227 msg->msg_len = msg->msg_wanted = payload_length;
2228 msg->msg_offset = 0;
2229 msg->msg_hdr = *hdr;
2232 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid);
2235 CERROR("%s, src %s: Dropping %s "
2236 "(error %d looking up sender)\n",
2237 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2238 lnet_msgtyp2str(type), rc);
2247 msg->msg_target.pid = dest_pid;
2248 msg->msg_target.nid = dest_nid;
2249 msg->msg_routing = 1;
2250 msg->msg_offset = 0;
2253 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
2254 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2255 rc = lnet_eager_recv_locked(msg);
2261 lnet_commit_routedmsg(msg);
2262 rc = lnet_post_routed_recv_locked(msg, 0);
2266 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2267 0, payload_length, payload_length);
2271 /* convert common msg->hdr fields to host byteorder */
2272 msg->msg_hdr.type = type;
2273 msg->msg_hdr.src_nid = src_nid;
2274 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2275 msg->msg_hdr.dest_nid = dest_nid;
2276 msg->msg_hdr.dest_pid = dest_pid;
2277 msg->msg_hdr.payload_length = payload_length;
2279 msg->msg_ev.sender = from_nid;
2283 rc = lnet_parse_ack(ni, msg);
2286 rc = lnet_parse_put(ni, msg);
2289 rc = lnet_parse_get(ni, msg, rdma_req);
2291 case LNET_MSG_REPLY:
2292 rc = lnet_parse_reply(ni, msg);
2296 goto free_drop; /* prevent an unused label if !kernel */
2302 LASSERT (rc == ENOENT);
2305 LASSERT (msg->msg_md == NULL);
2307 if (msg->msg_rxpeer != NULL) {
2308 lnet_peer_decref_locked(msg->msg_rxpeer);
2309 msg->msg_rxpeer = NULL;
2311 lnet_msg_free(msg); /* expects LNET_LOCK held */
2315 lnet_drop_message(ni, private, payload_length);
2320 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2321 lnet_process_id_t target, unsigned int portal,
2322 __u64 match_bits, unsigned int offset,
2329 LASSERT (the_lnet.ln_init);
2330 LASSERT (the_lnet.ln_refcount > 0);
2332 if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2333 fail_peer (target.nid, 1)) /* shall we now? */
2335 CERROR("Dropping PUT to %s: simulated failure\n",
2336 libcfs_id2str(target));
2340 msg = lnet_msg_alloc();
2342 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2343 libcfs_id2str(target));
2349 md = lnet_handle2md(&mdh);
2350 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2353 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2354 match_bits, portal, libcfs_id2str(target),
2355 md == NULL ? -1 : md->md_threshold);
2356 if (md != NULL && md->md_me != NULL)
2357 CERROR("Source MD also attached to portal %d\n",
2358 md->md_me->me_portal);
2364 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2366 lnet_commit_md(md, msg);
2368 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2370 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2371 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2372 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2373 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2375 /* NB handles only looked up by creator (no flips) */
2376 if (ack == LNET_ACK_REQ) {
2377 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2378 the_lnet.ln_interface_cookie;
2379 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2380 md->md_lh.lh_cookie;
2382 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2383 LNET_WIRE_HANDLE_COOKIE_NONE;
2384 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2385 LNET_WIRE_HANDLE_COOKIE_NONE;
2388 msg->msg_ev.type = LNET_EVENT_SEND;
2389 msg->msg_ev.initiator.nid = LNET_NID_ANY;
2390 msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2391 msg->msg_ev.target = target;
2392 msg->msg_ev.sender = LNET_NID_ANY;
2393 msg->msg_ev.pt_index = portal;
2394 msg->msg_ev.match_bits = match_bits;
2395 msg->msg_ev.rlength = md->md_length;
2396 msg->msg_ev.mlength = md->md_length;
2397 msg->msg_ev.offset = offset;
2398 msg->msg_ev.hdr_data = hdr_data;
2400 lnet_md_deconstruct(md, &msg->msg_ev.md);
2401 lnet_md2handle(&msg->msg_ev.md_handle, md);
2403 the_lnet.ln_counters.send_count++;
2404 the_lnet.ln_counters.send_length += md->md_length;
2408 rc = lnet_send(self, msg);
2410 CERROR("Error sending PUT to %s: %d\n",
2411 libcfs_id2str(target), rc);
2412 lnet_finalize (NULL, msg, rc);
2415 /* completion will be signalled by an event */
2420 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2422 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2423 * returns a msg for the LND to pass to lnet_finalize() when the sink
2424 * data has been received.
2426 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2427 * lnet_finalize() is called on it, so the LND must call this first */
2429 lnet_msg_t *msg = lnet_msg_alloc();
2430 lnet_libmd_t *getmd = getmsg->msg_md;
2431 lnet_process_id_t peer_id = getmsg->msg_target;
2433 LASSERT (!getmsg->msg_target_is_router);
2434 LASSERT (!getmsg->msg_routing);
2438 LASSERT (getmd->md_refcount > 0);
2441 CERROR ("%s: Dropping REPLY from %s: can't allocate msg\n",
2442 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2446 if (getmd->md_threshold == 0) {
2447 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2448 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2453 LASSERT (getmd->md_offset == 0);
2455 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2456 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2458 lnet_commit_md (getmd, msg);
2460 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2462 msg->msg_ev.type = LNET_EVENT_REPLY;
2463 msg->msg_ev.initiator = peer_id;
2464 msg->msg_ev.sender = peer_id.nid; /* optimized GETs can't be routed */
2465 msg->msg_ev.rlength = msg->msg_ev.mlength = getmd->md_length;
2466 msg->msg_ev.offset = 0;
2468 lnet_md_deconstruct(getmd, &msg->msg_ev.md);
2469 lnet_md2handle(&msg->msg_ev.md_handle, getmd);
2471 the_lnet.ln_counters.recv_count++;
2472 the_lnet.ln_counters.recv_length += getmd->md_length;
2481 the_lnet.ln_counters.drop_count++;
2482 the_lnet.ln_counters.drop_length += getmd->md_length;
2490 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2492 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2493 * completed and I know it. */
2494 LASSERT (reply != NULL);
2495 LASSERT (reply->msg_type == LNET_MSG_GET);
2496 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2498 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2499 * the end of my buffer, I might as well be dead. */
2500 LASSERT (len <= reply->msg_ev.mlength);
2502 reply->msg_ev.mlength = len;
2506 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2507 lnet_process_id_t target, unsigned int portal,
2508 __u64 match_bits, unsigned int offset)
2514 LASSERT (the_lnet.ln_init);
2515 LASSERT (the_lnet.ln_refcount > 0);
2517 if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2518 fail_peer (target.nid, 1)) /* shall we now? */
2520 CERROR("Dropping GET to %s: simulated failure\n",
2521 libcfs_id2str(target));
2525 msg = lnet_msg_alloc();
2527 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2528 libcfs_id2str(target));
2534 md = lnet_handle2md(&mdh);
2535 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2538 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2539 match_bits, portal, libcfs_id2str(target),
2540 md == NULL ? -1 : md->md_threshold);
2541 if (md != NULL && md->md_me != NULL)
2542 CERROR("REPLY MD also attached to portal %d\n",
2543 md->md_me->me_portal);
2549 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2551 lnet_commit_md(md, msg);
2553 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2555 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2556 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2557 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2558 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2560 /* NB handles only looked up by creator (no flips) */
2561 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2562 the_lnet.ln_interface_cookie;
2563 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2564 md->md_lh.lh_cookie;
2566 msg->msg_ev.type = LNET_EVENT_SEND;
2567 msg->msg_ev.initiator.nid = LNET_NID_ANY;
2568 msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2569 msg->msg_ev.target = target;
2570 msg->msg_ev.sender = LNET_NID_ANY;
2571 msg->msg_ev.pt_index = portal;
2572 msg->msg_ev.match_bits = match_bits;
2573 msg->msg_ev.rlength = md->md_length;
2574 msg->msg_ev.mlength = md->md_length;
2575 msg->msg_ev.offset = offset;
2576 msg->msg_ev.hdr_data = 0;
2578 lnet_md_deconstruct(md, &msg->msg_ev.md);
2579 lnet_md2handle(&msg->msg_ev.md_handle, md);
2581 the_lnet.ln_counters.send_count++;
2585 rc = lnet_send(self, msg);
2587 CERROR("error sending GET to %s: %d\n",
2588 libcfs_id2str(target), rc);
2589 lnet_finalize (NULL, msg, rc);
2592 /* completion will be signalled by an event */
2597 LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2599 struct list_head *e;
2601 lnet_route_t *route;
2602 lnet_remotenet_t *rnet;
2603 __u32 dstnet = LNET_NIDNET(dstnid);
2607 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2608 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2609 * keep order 0 free for 0@lo and order 1 free for a local NID
2612 LASSERT (the_lnet.ln_init);
2613 LASSERT (the_lnet.ln_refcount > 0);
2617 list_for_each (e, &the_lnet.ln_nis) {
2618 ni = list_entry(e, lnet_ni_t, ni_list);
2620 if (ni->ni_nid == dstnid) {
2621 if (srcnidp != NULL)
2623 if (orderp != NULL) {
2624 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2631 return local_nid_dist_zero ? 0 : 1;
2634 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2635 if (srcnidp != NULL)
2636 *srcnidp = ni->ni_nid;
2646 list_for_each (e, &the_lnet.ln_remote_nets) {
2647 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2649 if (rnet->lrn_net == dstnet) {
2650 LASSERT (!list_empty(&rnet->lrn_routes));
2651 route = list_entry(rnet->lrn_routes.next,
2652 lnet_route_t, lr_list);
2653 hops = rnet->lrn_hops;
2654 if (srcnidp != NULL)
2655 *srcnidp = route->lr_gateway->lp_ni->ni_nid;
2665 return -EHOSTUNREACH;
2669 LNetSetAsync(lnet_process_id_t id, int nasync)
2675 lnet_remotenet_t *rnet;
2676 struct list_head *tmp;
2677 lnet_route_t *route;
2684 /* Target on a local network? */
2686 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2688 if (ni->ni_lnd->lnd_setasync != NULL)
2689 rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2694 /* Target on a remote network: apply to routers */
2696 LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2701 /* Snapshot all the router NIDs */
2703 rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2705 list_for_each(tmp, &rnet->lrn_routes) {
2706 if (nnids == maxnids) {
2708 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2713 route = list_entry(tmp, lnet_route_t, lr_list);
2714 nids[nnids++] = route->lr_gateway->lp_nid;
2719 /* set async on all the routers */
2720 while (nnids-- > 0) {
2721 id.pid = LUSTRE_SRV_LNET_PID;
2722 id.nid = nids[nnids];
2724 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2728 if (ni->ni_lnd->lnd_setasync != NULL) {
2729 rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2736 LIBCFS_FREE(nids, maxnids * sizeof(*nids));