4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
31 * This file is part of Lustre, http://www.lustre.org/
32 * Lustre is a trademark of Sun Microsystems, Inc.
34 * lnet/lnet/lib-move.c
36 * Data movement routines
39 #define DEBUG_SUBSYSTEM S_LNET
41 #include <lnet/lib-lnet.h>
43 static int local_nid_dist_zero = 1;
44 CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444,
48 static void lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg);
50 #define LNET_MATCHMD_NONE 0 /* Didn't match */
51 #define LNET_MATCHMD_OK 1 /* Matched OK */
52 #define LNET_MATCHMD_DROP 2 /* Must be discarded */
55 lnet_try_match_md (int index, int op_mask, lnet_process_id_t src,
56 unsigned int rlength, unsigned int roffset,
57 __u64 match_bits, lnet_libmd_t *md, lnet_msg_t *msg,
58 unsigned int *mlength_out, unsigned int *offset_out)
60 /* ALWAYS called holding the LNET_LOCK, and can't LNET_UNLOCK;
61 * lnet_match_blocked_msg() relies on this to avoid races */
64 lnet_me_t *me = md->md_me;
66 /* mismatched MD op */
67 if ((md->md_options & op_mask) == 0)
68 return LNET_MATCHMD_NONE;
71 if (lnet_md_exhausted(md))
72 return LNET_MATCHMD_NONE;
74 /* mismatched ME nid/pid? */
75 if (me->me_match_id.nid != LNET_NID_ANY &&
76 me->me_match_id.nid != src.nid)
77 return LNET_MATCHMD_NONE;
79 if (me->me_match_id.pid != LNET_PID_ANY &&
80 me->me_match_id.pid != src.pid)
81 return LNET_MATCHMD_NONE;
83 /* mismatched ME matchbits? */
84 if (((me->me_match_bits ^ match_bits) & ~me->me_ignore_bits) != 0)
85 return LNET_MATCHMD_NONE;
87 /* Hurrah! This _is_ a match; check it out... */
89 if ((md->md_options & LNET_MD_MANAGE_REMOTE) == 0)
90 offset = md->md_offset;
94 if ((md->md_options & LNET_MD_MAX_SIZE) != 0) {
95 mlength = md->md_max_size;
96 LASSERT (md->md_offset + mlength <= md->md_length);
98 mlength = md->md_length - offset;
101 if (rlength <= mlength) { /* fits in allowed space */
103 } else if ((md->md_options & LNET_MD_TRUNCATE) == 0) {
104 /* this packet _really_ is too big */
105 CERROR("Matching packet from %s, match "LPU64
106 " length %d too big: %d left, %d allowed\n",
107 libcfs_id2str(src), match_bits, rlength,
108 md->md_length - offset, mlength);
110 return LNET_MATCHMD_DROP;
113 /* Commit to this ME/MD */
114 CDEBUG(D_NET, "Incoming %s index %x from %s of "
115 "length %d/%d into md "LPX64" [%d] + %d\n",
116 (op_mask == LNET_MD_OP_PUT) ? "put" : "get",
117 index, libcfs_id2str(src), mlength, rlength,
118 md->md_lh.lh_cookie, md->md_niov, offset);
120 lnet_commit_md(md, msg);
121 md->md_offset = offset + mlength;
123 /* NB Caller will set ev.type and ev.hdr_data */
124 msg->msg_ev.initiator = src;
125 msg->msg_ev.pt_index = index;
126 msg->msg_ev.match_bits = match_bits;
127 msg->msg_ev.rlength = rlength;
128 msg->msg_ev.mlength = mlength;
129 msg->msg_ev.offset = offset;
131 lnet_md_deconstruct(md, &msg->msg_ev.md);
132 lnet_md2handle(&msg->msg_ev.md_handle, md);
134 *offset_out = offset;
135 *mlength_out = mlength;
137 /* Auto-unlink NOW, so the ME gets unlinked if required.
138 * We bumped md->md_refcount above so the MD just gets flagged
139 * for unlink when it is finalized. */
140 if ((md->md_flags & LNET_MD_FLAG_AUTO_UNLINK) != 0 &&
141 lnet_md_exhausted(md)) {
145 return LNET_MATCHMD_OK;
149 lnet_match_md(int index, int op_mask, lnet_process_id_t src,
150 unsigned int rlength, unsigned int roffset,
151 __u64 match_bits, lnet_msg_t *msg,
152 unsigned int *mlength_out, unsigned int *offset_out,
153 lnet_libmd_t **md_out)
155 lnet_portal_t *ptl = &the_lnet.ln_portals[index];
162 CDEBUG (D_NET, "Request from %s of length %d into portal %d "
163 "MB="LPX64"\n", libcfs_id2str(src), rlength, index, match_bits);
165 if (index < 0 || index >= the_lnet.ln_nportals) {
166 CERROR("Invalid portal %d not in [0-%d]\n",
167 index, the_lnet.ln_nportals);
168 return LNET_MATCHMD_DROP;
171 head = lnet_portal_me_head(index, src, match_bits);
172 if (head == NULL) /* nobody posted anything on this portal */
175 cfs_list_for_each_entry_safe_typed (me, tmp, head,
176 lnet_me_t, me_list) {
179 /* ME attached but MD not attached yet */
183 LASSERT (me == md->md_me);
185 rc = lnet_try_match_md(index, op_mask, src, rlength,
186 roffset, match_bits, md, msg,
187 mlength_out, offset_out);
192 case LNET_MATCHMD_NONE:
195 case LNET_MATCHMD_OK:
197 return LNET_MATCHMD_OK;
199 case LNET_MATCHMD_DROP:
200 return LNET_MATCHMD_DROP;
206 if (op_mask == LNET_MD_OP_GET ||
207 !lnet_portal_is_lazy(ptl))
208 return LNET_MATCHMD_DROP;
210 return LNET_MATCHMD_NONE;
214 lnet_fail_nid (lnet_nid_t nid, unsigned int threshold)
216 lnet_test_peer_t *tp;
221 LASSERT (the_lnet.ln_init);
223 if (threshold != 0) {
224 /* Adding a new entry */
225 LIBCFS_ALLOC(tp, sizeof(*tp));
230 tp->tp_threshold = threshold;
233 cfs_list_add_tail (&tp->tp_list, &the_lnet.ln_test_peers);
238 /* removing entries */
239 CFS_INIT_LIST_HEAD (&cull);
243 cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
244 tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
246 if (tp->tp_threshold == 0 || /* needs culling anyway */
247 nid == LNET_NID_ANY || /* removing all entries */
248 tp->tp_nid == nid) /* matched this one */
250 cfs_list_del (&tp->tp_list);
251 cfs_list_add (&tp->tp_list, &cull);
257 while (!cfs_list_empty (&cull)) {
258 tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
260 cfs_list_del (&tp->tp_list);
261 LIBCFS_FREE(tp, sizeof (*tp));
267 fail_peer (lnet_nid_t nid, int outgoing)
269 lnet_test_peer_t *tp;
275 CFS_INIT_LIST_HEAD (&cull);
279 cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
280 tp = cfs_list_entry (el, lnet_test_peer_t, tp_list);
282 if (tp->tp_threshold == 0) {
285 /* only cull zombies on outgoing tests,
286 * since we may be at interrupt priority on
287 * incoming messages. */
288 cfs_list_del (&tp->tp_list);
289 cfs_list_add (&tp->tp_list, &cull);
294 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
295 nid == tp->tp_nid) { /* fail this peer */
298 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
301 tp->tp_threshold == 0) {
303 cfs_list_del (&tp->tp_list);
304 cfs_list_add (&tp->tp_list, &cull);
313 while (!cfs_list_empty (&cull)) {
314 tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list);
315 cfs_list_del (&tp->tp_list);
317 LIBCFS_FREE(tp, sizeof (*tp));
324 lnet_iov_nob (unsigned int niov, struct iovec *iov)
326 unsigned int nob = 0;
329 nob += (iov++)->iov_len;
335 lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
336 unsigned int nsiov, struct iovec *siov, unsigned int soffset,
339 /* NB diov, siov are READ-ONLY */
340 unsigned int this_nob;
345 /* skip complete frags before 'doffset' */
347 while (doffset >= diov->iov_len) {
348 doffset -= diov->iov_len;
354 /* skip complete frags before 'soffset' */
356 while (soffset >= siov->iov_len) {
357 soffset -= siov->iov_len;
366 this_nob = MIN(diov->iov_len - doffset,
367 siov->iov_len - soffset);
368 this_nob = MIN(this_nob, nob);
370 memcpy ((char *)diov->iov_base + doffset,
371 (char *)siov->iov_base + soffset, this_nob);
374 if (diov->iov_len > doffset + this_nob) {
382 if (siov->iov_len > soffset + this_nob) {
393 lnet_extract_iov (int dst_niov, struct iovec *dst,
394 int src_niov, struct iovec *src,
395 unsigned int offset, unsigned int len)
397 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
398 * for exactly 'len' bytes, and return the number of entries.
399 * NB not destructive to 'src' */
400 unsigned int frag_len;
403 if (len == 0) /* no data => */
404 return (0); /* no frags */
406 LASSERT (src_niov > 0);
407 while (offset >= src->iov_len) { /* skip initial frags */
408 offset -= src->iov_len;
411 LASSERT (src_niov > 0);
416 LASSERT (src_niov > 0);
417 LASSERT ((int)niov <= dst_niov);
419 frag_len = src->iov_len - offset;
420 dst->iov_base = ((char *)src->iov_base) + offset;
422 if (len <= frag_len) {
427 dst->iov_len = frag_len;
440 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
447 lnet_copy_kiov2kiov (unsigned int ndkiov, lnet_kiov_t *dkiov, unsigned int doffset,
448 unsigned int nskiov, lnet_kiov_t *skiov, unsigned int soffset,
455 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
456 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
463 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
464 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
471 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
472 int src_niov, lnet_kiov_t *src,
473 unsigned int offset, unsigned int len)
478 #else /* __KERNEL__ */
481 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
483 unsigned int nob = 0;
486 nob += (kiov++)->kiov_len;
492 lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
493 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
496 /* NB diov, siov are READ-ONLY */
497 unsigned int this_nob;
504 LASSERT (!cfs_in_interrupt ());
507 while (doffset >= diov->kiov_len) {
508 doffset -= diov->kiov_len;
515 while (soffset >= siov->kiov_len) {
516 soffset -= siov->kiov_len;
525 this_nob = MIN(diov->kiov_len - doffset,
526 siov->kiov_len - soffset);
527 this_nob = MIN(this_nob, nob);
530 daddr = ((char *)cfs_kmap(diov->kiov_page)) +
531 diov->kiov_offset + doffset;
533 saddr = ((char *)cfs_kmap(siov->kiov_page)) +
534 siov->kiov_offset + soffset;
536 /* Vanishing risk of kmap deadlock when mapping 2 pages.
537 * However in practice at least one of the kiovs will be mapped
538 * kernel pages and the map/unmap will be NOOPs */
540 memcpy (daddr, saddr, this_nob);
543 if (diov->kiov_len > doffset + this_nob) {
547 cfs_kunmap(diov->kiov_page);
554 if (siov->kiov_len > soffset + this_nob) {
558 cfs_kunmap(siov->kiov_page);
567 cfs_kunmap(diov->kiov_page);
569 cfs_kunmap(siov->kiov_page);
573 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
574 unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
577 /* NB iov, kiov are READ-ONLY */
578 unsigned int this_nob;
584 LASSERT (!cfs_in_interrupt ());
587 while (iovoffset >= iov->iov_len) {
588 iovoffset -= iov->iov_len;
595 while (kiovoffset >= kiov->kiov_len) {
596 kiovoffset -= kiov->kiov_len;
605 this_nob = MIN(iov->iov_len - iovoffset,
606 kiov->kiov_len - kiovoffset);
607 this_nob = MIN(this_nob, nob);
610 addr = ((char *)cfs_kmap(kiov->kiov_page)) +
611 kiov->kiov_offset + kiovoffset;
613 memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
616 if (iov->iov_len > iovoffset + this_nob) {
617 iovoffset += this_nob;
624 if (kiov->kiov_len > kiovoffset + this_nob) {
626 kiovoffset += this_nob;
628 cfs_kunmap(kiov->kiov_page);
638 cfs_kunmap(kiov->kiov_page);
642 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
643 unsigned int niov, struct iovec *iov, unsigned int iovoffset,
646 /* NB kiov, iov are READ-ONLY */
647 unsigned int this_nob;
653 LASSERT (!cfs_in_interrupt ());
656 while (kiovoffset >= kiov->kiov_len) {
657 kiovoffset -= kiov->kiov_len;
664 while (iovoffset >= iov->iov_len) {
665 iovoffset -= iov->iov_len;
674 this_nob = MIN(kiov->kiov_len - kiovoffset,
675 iov->iov_len - iovoffset);
676 this_nob = MIN(this_nob, nob);
679 addr = ((char *)cfs_kmap(kiov->kiov_page)) +
680 kiov->kiov_offset + kiovoffset;
682 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
685 if (kiov->kiov_len > kiovoffset + this_nob) {
687 kiovoffset += this_nob;
689 cfs_kunmap(kiov->kiov_page);
696 if (iov->iov_len > iovoffset + this_nob) {
697 iovoffset += this_nob;
706 cfs_kunmap(kiov->kiov_page);
710 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
711 int src_niov, lnet_kiov_t *src,
712 unsigned int offset, unsigned int len)
714 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
715 * for exactly 'len' bytes, and return the number of entries.
716 * NB not destructive to 'src' */
717 unsigned int frag_len;
720 if (len == 0) /* no data => */
721 return (0); /* no frags */
723 LASSERT (src_niov > 0);
724 while (offset >= src->kiov_len) { /* skip initial frags */
725 offset -= src->kiov_len;
728 LASSERT (src_niov > 0);
733 LASSERT (src_niov > 0);
734 LASSERT ((int)niov <= dst_niov);
736 frag_len = src->kiov_len - offset;
737 dst->kiov_page = src->kiov_page;
738 dst->kiov_offset = src->kiov_offset + offset;
740 if (len <= frag_len) {
742 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
746 dst->kiov_len = frag_len;
747 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
760 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
761 unsigned int offset, unsigned int mlen, unsigned int rlen)
763 unsigned int niov = 0;
764 struct iovec *iov = NULL;
765 lnet_kiov_t *kiov = NULL;
768 LASSERT (!cfs_in_interrupt ());
769 LASSERT (mlen == 0 || msg != NULL);
772 LASSERT(msg->msg_receiving);
773 LASSERT(!msg->msg_sending);
774 LASSERT(rlen == msg->msg_len);
775 LASSERT(mlen <= msg->msg_len);
777 msg->msg_wanted = mlen;
778 msg->msg_offset = offset;
779 msg->msg_receiving = 0;
782 niov = msg->msg_niov;
784 kiov = msg->msg_kiov;
787 LASSERT ((iov == NULL) != (kiov == NULL));
791 rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
792 niov, iov, kiov, offset, mlen, rlen);
794 lnet_finalize(ni, msg, rc);
798 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
800 lnet_peer_t *p1 = r1->lr_gateway;
801 lnet_peer_t *p2 = r2->lr_gateway;
803 if (r1->lr_hops < r2->lr_hops)
806 if (r1->lr_hops > r2->lr_hops)
809 if (p1->lp_txqnob < p2->lp_txqnob)
812 if (p1->lp_txqnob > p2->lp_txqnob)
815 if (p1->lp_txcredits > p2->lp_txcredits)
818 if (p1->lp_txcredits < p2->lp_txcredits)
826 lnet_setpayloadbuffer(lnet_msg_t *msg)
828 lnet_libmd_t *md = msg->msg_md;
830 LASSERT (msg->msg_len > 0);
831 LASSERT (!msg->msg_routing);
832 LASSERT (md != NULL);
833 LASSERT (msg->msg_niov == 0);
834 LASSERT (msg->msg_iov == NULL);
835 LASSERT (msg->msg_kiov == NULL);
837 msg->msg_niov = md->md_niov;
838 if ((md->md_options & LNET_MD_KIOV) != 0)
839 msg->msg_kiov = md->md_iov.kiov;
841 msg->msg_iov = md->md_iov.iov;
845 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
846 unsigned int offset, unsigned int len)
848 msg->msg_type = type;
849 msg->msg_target = target;
851 msg->msg_offset = offset;
854 lnet_setpayloadbuffer(msg);
856 memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
857 msg->msg_hdr.type = cpu_to_le32(type);
858 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
859 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
860 /* src_nid will be set later */
861 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
862 msg->msg_hdr.payload_length = cpu_to_le32(len);
866 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
868 void *priv = msg->msg_private;
871 LASSERT (!cfs_in_interrupt ());
872 LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
873 (msg->msg_txcredit && msg->msg_peertxcredit));
875 rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
877 lnet_finalize(ni, msg, rc);
881 lnet_eager_recv_locked(lnet_msg_t *msg)
887 LASSERT (!msg->msg_delayed);
888 msg->msg_delayed = 1;
890 LASSERT (msg->msg_receiving);
891 LASSERT (!msg->msg_sending);
893 peer = msg->msg_rxpeer;
896 if (ni->ni_lnd->lnd_eager_recv != NULL) {
899 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
902 CERROR("recv from %s / send to %s aborted: "
903 "eager_recv failed %d\n",
904 libcfs_nid2str(peer->lp_nid),
905 libcfs_id2str(msg->msg_target), rc);
906 LASSERT (rc < 0); /* required by my callers */
915 /* NB: caller shall hold a ref on 'lp' as I'd drop LNET_LOCK */
917 lnet_ni_peer_alive(lnet_peer_t *lp)
919 cfs_time_t last_alive = 0;
920 lnet_ni_t *ni = lp->lp_ni;
922 LASSERT (lnet_peer_aliveness_enabled(lp));
923 LASSERT (ni->ni_lnd->lnd_query != NULL);
924 LASSERT (the_lnet.ln_routing == 1);
927 (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
930 lp->lp_last_query = cfs_time_current();
932 if (last_alive != 0) /* NI has updated timestamp */
933 lp->lp_last_alive = last_alive;
936 /* NB: always called with LNET_LOCK held */
938 lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
943 LASSERT (lnet_peer_aliveness_enabled(lp));
944 LASSERT (the_lnet.ln_routing == 1);
946 /* Trust lnet_notify() if it has more recent aliveness news, but
947 * ignore the initial assumed death (see lnet_peers_start_down()).
949 if (!lp->lp_alive && lp->lp_alive_count > 0 &&
950 cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive))
953 deadline = cfs_time_add(lp->lp_last_alive,
954 cfs_time_seconds(lp->lp_ni->ni_peertimeout));
955 alive = cfs_time_after(deadline, now);
957 /* Update obsolete lp_alive except for routers assumed to be dead
958 * initially, because router checker would update aliveness in this
959 * case, and moreover lp_last_alive at peer creation is assumed.
961 if (alive && !lp->lp_alive &&
962 !(lnet_isrouter(lp) && lp->lp_alive_count == 0))
963 lnet_notify_locked(lp, 0, 1, lp->lp_last_alive);
969 /* NB: returns 1 when alive, 0 when dead, negative when error;
970 * may drop the LNET_LOCK */
972 lnet_peer_alive_locked (lnet_peer_t *lp)
974 cfs_time_t now = cfs_time_current();
976 /* LU-630: only router checks peer health. */
977 if (the_lnet.ln_routing == 0)
980 if (!lnet_peer_aliveness_enabled(lp))
983 if (lnet_peer_is_alive(lp, now))
986 /* Peer appears dead, but we should avoid frequent NI queries (at
987 * most once per lnet_queryinterval seconds). */
988 if (lp->lp_last_query != 0) {
989 static const int lnet_queryinterval = 1;
991 cfs_time_t next_query =
992 cfs_time_add(lp->lp_last_query,
993 cfs_time_seconds(lnet_queryinterval));
995 if (cfs_time_before(now, next_query)) {
997 CWARN("Unexpected aliveness of peer %s: "
999 libcfs_nid2str(lp->lp_nid),
1000 (int)now, (int)next_query,
1002 lp->lp_ni->ni_peertimeout);
1007 /* query NI for latest aliveness news */
1008 lnet_ni_peer_alive(lp);
1010 if (lnet_peer_is_alive(lp, now))
1013 lnet_notify_locked(lp, 0, 0, lp->lp_last_alive);
1018 lnet_post_send_locked (lnet_msg_t *msg, int do_send)
1020 /* lnet_send is going to LNET_UNLOCK immediately after this, so it sets
1021 * do_send FALSE and I don't do the unlock/send/lock bit. I return
1022 * EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer appears dead, and
1023 * 0 if sent or OK to send */
1024 lnet_peer_t *lp = msg->msg_txpeer;
1025 lnet_ni_t *ni = lp->lp_ni;
1027 /* non-lnet_send() callers have checked before */
1028 LASSERT (!do_send || msg->msg_delayed);
1029 LASSERT (!msg->msg_receiving);
1031 /* NB 'lp' is always the next hop */
1032 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
1033 lnet_peer_alive_locked(lp) == 0) {
1034 the_lnet.ln_counters.drop_count++;
1035 the_lnet.ln_counters.drop_length += msg->msg_len;
1038 CNETERR("Dropping message for %s: peer not alive\n",
1039 libcfs_id2str(msg->msg_target));
1041 lnet_finalize(ni, msg, -EHOSTUNREACH);
1044 return EHOSTUNREACH;
1047 if (!msg->msg_peertxcredit) {
1048 LASSERT ((lp->lp_txcredits < 0) ==
1049 !cfs_list_empty(&lp->lp_txq));
1051 msg->msg_peertxcredit = 1;
1052 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
1055 if (lp->lp_txcredits < lp->lp_mintxcredits)
1056 lp->lp_mintxcredits = lp->lp_txcredits;
1058 if (lp->lp_txcredits < 0) {
1059 msg->msg_delayed = 1;
1060 cfs_list_add_tail(&msg->msg_list, &lp->lp_txq);
1065 if (!msg->msg_txcredit) {
1066 LASSERT ((ni->ni_txcredits < 0) ==
1067 !cfs_list_empty(&ni->ni_txq));
1069 msg->msg_txcredit = 1;
1072 if (ni->ni_txcredits < ni->ni_mintxcredits)
1073 ni->ni_mintxcredits = ni->ni_txcredits;
1075 if (ni->ni_txcredits < 0) {
1076 msg->msg_delayed = 1;
1077 cfs_list_add_tail(&msg->msg_list, &ni->ni_txq);
1084 lnet_ni_send(ni, msg);
1092 lnet_commit_routedmsg (lnet_msg_t *msg)
1094 /* ALWAYS called holding the LNET_LOCK */
1095 LASSERT (msg->msg_routing);
1097 the_lnet.ln_counters.msgs_alloc++;
1098 if (the_lnet.ln_counters.msgs_alloc >
1099 the_lnet.ln_counters.msgs_max)
1100 the_lnet.ln_counters.msgs_max =
1101 the_lnet.ln_counters.msgs_alloc;
1103 the_lnet.ln_counters.route_count++;
1104 the_lnet.ln_counters.route_length += msg->msg_len;
1106 LASSERT (!msg->msg_onactivelist);
1107 msg->msg_onactivelist = 1;
1108 cfs_list_add(&msg->msg_activelist,
1109 &the_lnet.ln_msg_container.msc_active);
1113 lnet_msg2bufpool(lnet_msg_t *msg)
1115 lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[0];
1117 LASSERT (msg->msg_len <= LNET_MTU);
1118 while (msg->msg_len > (unsigned int)rbp->rbp_npages * CFS_PAGE_SIZE) {
1120 LASSERT (rbp < &the_lnet.ln_rtrpools[LNET_NRBPOOLS]);
1127 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
1129 /* lnet_parse is going to LNET_UNLOCK immediately after this, so it
1130 * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
1131 * return EAGAIN if msg blocked and 0 if received or OK to receive */
1132 lnet_peer_t *lp = msg->msg_rxpeer;
1133 lnet_rtrbufpool_t *rbp;
1136 LASSERT (msg->msg_iov == NULL);
1137 LASSERT (msg->msg_kiov == NULL);
1138 LASSERT (msg->msg_niov == 0);
1139 LASSERT (msg->msg_routing);
1140 LASSERT (msg->msg_receiving);
1141 LASSERT (!msg->msg_sending);
1143 /* non-lnet_parse callers only send delayed messages */
1144 LASSERT (!do_recv || msg->msg_delayed);
1146 if (!msg->msg_peerrtrcredit) {
1147 LASSERT ((lp->lp_rtrcredits < 0) ==
1148 !cfs_list_empty(&lp->lp_rtrq));
1150 msg->msg_peerrtrcredit = 1;
1151 lp->lp_rtrcredits--;
1152 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
1153 lp->lp_minrtrcredits = lp->lp_rtrcredits;
1155 if (lp->lp_rtrcredits < 0) {
1156 /* must have checked eager_recv before here */
1157 LASSERT (msg->msg_delayed);
1158 cfs_list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1163 rbp = lnet_msg2bufpool(msg);
1165 if (!msg->msg_rtrcredit) {
1166 LASSERT ((rbp->rbp_credits < 0) ==
1167 !cfs_list_empty(&rbp->rbp_msgs));
1169 msg->msg_rtrcredit = 1;
1171 if (rbp->rbp_credits < rbp->rbp_mincredits)
1172 rbp->rbp_mincredits = rbp->rbp_credits;
1174 if (rbp->rbp_credits < 0) {
1175 /* must have checked eager_recv before here */
1176 LASSERT (msg->msg_delayed);
1177 cfs_list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1182 LASSERT (!cfs_list_empty(&rbp->rbp_bufs));
1183 rb = cfs_list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
1184 cfs_list_del(&rb->rb_list);
1186 msg->msg_niov = rbp->rbp_npages;
1187 msg->msg_kiov = &rb->rb_kiov[0];
1191 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
1192 0, msg->msg_len, msg->msg_len);
1200 lnet_return_credits_locked (lnet_msg_t *msg)
1202 lnet_peer_t *txpeer = msg->msg_txpeer;
1203 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1207 if (msg->msg_txcredit) {
1208 /* give back NI txcredits */
1209 msg->msg_txcredit = 0;
1212 LASSERT((ni->ni_txcredits < 0) == !cfs_list_empty(&ni->ni_txq));
1215 if (ni->ni_txcredits <= 0) {
1216 msg2 = cfs_list_entry(ni->ni_txq.next, lnet_msg_t,
1218 cfs_list_del(&msg2->msg_list);
1220 LASSERT(msg2->msg_txpeer->lp_ni == ni);
1221 LASSERT(msg2->msg_delayed);
1223 (void) lnet_post_send_locked(msg2, 1);
1227 if (msg->msg_peertxcredit) {
1228 /* give back peer txcredits */
1229 msg->msg_peertxcredit = 0;
1231 LASSERT((txpeer->lp_txcredits < 0) ==
1232 !cfs_list_empty(&txpeer->lp_txq));
1234 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1235 LASSERT (txpeer->lp_txqnob >= 0);
1237 txpeer->lp_txcredits++;
1238 if (txpeer->lp_txcredits <= 0) {
1239 msg2 = cfs_list_entry(txpeer->lp_txq.next,
1240 lnet_msg_t, msg_list);
1241 cfs_list_del(&msg2->msg_list);
1243 LASSERT (msg2->msg_txpeer == txpeer);
1244 LASSERT (msg2->msg_delayed);
1246 (void) lnet_post_send_locked(msg2, 1);
1250 if (txpeer != NULL) {
1251 msg->msg_txpeer = NULL;
1252 lnet_peer_decref_locked(txpeer);
1256 if (msg->msg_rtrcredit) {
1257 /* give back global router credits */
1259 lnet_rtrbufpool_t *rbp;
1261 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1262 * there until it gets one allocated, or aborts the wait
1264 LASSERT (msg->msg_kiov != NULL);
1266 rb = cfs_list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1268 LASSERT (rbp == lnet_msg2bufpool(msg));
1270 msg->msg_kiov = NULL;
1271 msg->msg_rtrcredit = 0;
1273 LASSERT((rbp->rbp_credits < 0) ==
1274 !cfs_list_empty(&rbp->rbp_msgs));
1275 LASSERT((rbp->rbp_credits > 0) ==
1276 !cfs_list_empty(&rbp->rbp_bufs));
1278 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1280 if (rbp->rbp_credits <= 0) {
1281 msg2 = cfs_list_entry(rbp->rbp_msgs.next,
1282 lnet_msg_t, msg_list);
1283 cfs_list_del(&msg2->msg_list);
1285 (void) lnet_post_routed_recv_locked(msg2, 1);
1289 if (msg->msg_peerrtrcredit) {
1290 /* give back peer router credits */
1291 msg->msg_peerrtrcredit = 0;
1293 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1294 !cfs_list_empty(&rxpeer->lp_rtrq));
1296 rxpeer->lp_rtrcredits++;
1297 if (rxpeer->lp_rtrcredits <= 0) {
1298 msg2 = cfs_list_entry(rxpeer->lp_rtrq.next,
1299 lnet_msg_t, msg_list);
1300 cfs_list_del(&msg2->msg_list);
1302 (void) lnet_post_routed_recv_locked(msg2, 1);
1306 LASSERT (!msg->msg_rtrcredit);
1307 LASSERT (!msg->msg_peerrtrcredit);
1309 if (rxpeer != NULL) {
1310 msg->msg_rxpeer = NULL;
1311 lnet_peer_decref_locked(rxpeer);
1316 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg)
1318 lnet_nid_t dst_nid = msg->msg_target.nid;
1320 lnet_ni_t *local_ni;
1321 lnet_remotenet_t *rnet;
1322 lnet_route_t *route;
1323 lnet_route_t *best_route;
1329 LASSERT (msg->msg_txpeer == NULL);
1330 LASSERT (!msg->msg_sending);
1331 LASSERT (!msg->msg_target_is_router);
1332 LASSERT (!msg->msg_receiving);
1334 msg->msg_sending = 1;
1336 /* NB! ni != NULL == interface pre-determined (ACK/REPLY) */
1340 if (the_lnet.ln_shutdown) {
1345 if (src_nid == LNET_NID_ANY) {
1348 src_ni = lnet_nid2ni_locked(src_nid);
1349 if (src_ni == NULL) {
1351 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1352 "local nid\n", libcfs_nid2str(dst_nid),
1353 libcfs_nid2str(src_nid));
1356 LASSERT (!msg->msg_routing);
1359 /* Is this for someone on a local network? */
1360 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid));
1362 if (local_ni != NULL) {
1363 if (src_ni == NULL) {
1365 src_nid = src_ni->ni_nid;
1366 } else if (src_ni == local_ni) {
1367 lnet_ni_decref_locked(local_ni);
1369 lnet_ni_decref_locked(local_ni);
1370 lnet_ni_decref_locked(src_ni);
1372 LCONSOLE_WARN("No route to %s via from %s\n",
1373 libcfs_nid2str(dst_nid),
1374 libcfs_nid2str(src_nid));
1378 LASSERT (src_nid != LNET_NID_ANY);
1380 if (!msg->msg_routing)
1381 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1383 if (src_ni == the_lnet.ln_loni) {
1384 /* No send credit hassles with LOLND */
1386 lnet_ni_send(src_ni, msg);
1387 lnet_ni_decref(src_ni);
1391 rc = lnet_nid2peer_locked(&lp, dst_nid);
1392 lnet_ni_decref_locked(src_ni); /* lp has ref on src_ni; lose mine */
1395 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1396 libcfs_nid2str(dst_nid));
1397 /* ENOMEM or shutting down */
1400 LASSERT (lp->lp_ni == src_ni);
1406 * - once application finishes computation, check here to update
1407 * router states before it waits for pending IO in LNetEQPoll
1408 * - recursion breaker: router checker sends no message
1409 * to remote networks */
1410 if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING)
1411 lnet_router_checker();
1415 /* sending to a remote network */
1416 rnet = lnet_find_net_locked(LNET_NIDNET(dst_nid));
1419 lnet_ni_decref_locked(src_ni);
1421 LCONSOLE_WARN("No route to %s\n",
1422 libcfs_id2str(msg->msg_target));
1423 return -EHOSTUNREACH;
1426 /* Find the best gateway I can use */
1429 cfs_list_for_each(tmp, &rnet->lrn_routes) {
1430 route = cfs_list_entry(tmp, lnet_route_t, lr_list);
1431 lp2 = route->lr_gateway;
1433 if (lp2->lp_alive &&
1434 lnet_router_down_ni(lp2, rnet->lrn_net) <= 0 &&
1435 (src_ni == NULL || lp2->lp_ni == src_ni) &&
1437 lnet_compare_routes(route, best_route) > 0)) {
1445 lnet_ni_decref_locked(src_ni);
1448 LCONSOLE_WARN("No route to %s via %s "
1449 "(all routers down)\n",
1450 libcfs_id2str(msg->msg_target),
1451 libcfs_nid2str(src_nid));
1452 return -EHOSTUNREACH;
1455 /* Place selected route at the end of the route list to ensure
1456 * fairness; everything else being equal... */
1457 cfs_list_del(&best_route->lr_list);
1458 cfs_list_add_tail(&best_route->lr_list, &rnet->lrn_routes);
1459 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1460 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1461 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1463 if (src_ni == NULL) {
1465 src_nid = src_ni->ni_nid;
1467 LASSERT (src_ni == lp->lp_ni);
1468 lnet_ni_decref_locked(src_ni);
1471 lnet_peer_addref_locked(lp);
1473 LASSERT (src_nid != LNET_NID_ANY);
1475 if (!msg->msg_routing) {
1476 /* I'm the source and now I know which NI to send on */
1477 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1480 msg->msg_target_is_router = 1;
1481 msg->msg_target.nid = lp->lp_nid;
1482 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1485 /* 'lp' is our best choice of peer */
1487 LASSERT (!msg->msg_peertxcredit);
1488 LASSERT (!msg->msg_txcredit);
1489 LASSERT (msg->msg_txpeer == NULL);
1491 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1493 rc = lnet_post_send_locked(msg, 0);
1496 if (rc == EHOSTUNREACH)
1497 return -EHOSTUNREACH;
1500 lnet_ni_send(src_ni, msg);
1506 lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg)
1508 /* ALWAYS called holding the LNET_LOCK */
1509 /* Here, we commit the MD to a network OP by marking it busy and
1510 * decrementing its threshold. Come what may, the network "owns"
1511 * the MD until a call to lnet_finalize() signals completion. */
1512 LASSERT (!msg->msg_routing);
1517 if (md->md_threshold != LNET_MD_THRESH_INF) {
1518 LASSERT (md->md_threshold > 0);
1522 the_lnet.ln_counters.msgs_alloc++;
1523 if (the_lnet.ln_counters.msgs_alloc >
1524 the_lnet.ln_counters.msgs_max)
1525 the_lnet.ln_counters.msgs_max =
1526 the_lnet.ln_counters.msgs_alloc;
1528 LASSERT (!msg->msg_onactivelist);
1529 msg->msg_onactivelist = 1;
1530 cfs_list_add(&msg->msg_activelist,
1531 &the_lnet.ln_msg_container.msc_active);
1535 lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob)
1538 the_lnet.ln_counters.drop_count++;
1539 the_lnet.ln_counters.drop_length += nob;
1542 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1546 lnet_drop_delayed_put(lnet_msg_t *msg, char *reason)
1548 lnet_process_id_t id = {0};
1550 id.nid = msg->msg_hdr.src_nid;
1551 id.pid = msg->msg_hdr.src_pid;
1553 LASSERT (msg->msg_md == NULL);
1554 LASSERT (msg->msg_delayed);
1555 LASSERT (msg->msg_rxpeer != NULL);
1556 LASSERT (msg->msg_hdr.type == LNET_MSG_PUT);
1558 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
1559 " offset %d length %d: %s\n",
1561 msg->msg_hdr.msg.put.ptl_index,
1562 msg->msg_hdr.msg.put.match_bits,
1563 msg->msg_hdr.msg.put.offset,
1564 msg->msg_hdr.payload_length,
1567 /* NB I can't drop msg's ref on msg_rxpeer until after I've
1568 * called lnet_drop_message(), so I just hang onto msg as well
1569 * until that's done */
1571 lnet_drop_message(msg->msg_rxpeer->lp_ni,
1572 msg->msg_private, msg->msg_len);
1576 lnet_peer_decref_locked(msg->msg_rxpeer);
1577 msg->msg_rxpeer = NULL;
1579 lnet_msg_free_locked(msg);
1585 * Turn on the lazy portal attribute. Use with caution!
1587 * This portal attribute only affects incoming PUT requests to the portal,
1588 * and is off by default. By default, if there's no matching MD for an
1589 * incoming PUT request, it is simply dropped. With the lazy attribute on,
1590 * such requests are queued indefinitely until either a matching MD is
1591 * posted to the portal or the lazy attribute is turned off.
1593 * It would prevent dropped requests, however it should be regarded as the
1594 * last line of defense - i.e. users must keep a close watch on active
1595 * buffers on a lazy portal and once it becomes too low post more buffers as
1596 * soon as possible. This is because delayed requests usually have detrimental
1597 * effects on underlying network connections. A few delayed requests often
1598 * suffice to bring an underlying connection to a complete halt, due to flow
1599 * control mechanisms.
1601 * There's also a DOS attack risk. If users don't post match-all MDs on a
1602 * lazy portal, a malicious peer can easily stop a service by sending some
1603 * PUT requests with match bits that won't match any MD. A routed server is
1604 * especially vulnerable since the connections to its neighbor routers are
1605 * shared among all clients.
1607 * \param portal Index of the portal to enable the lazy attribute on.
1609 * \retval 0 On success.
1610 * \retval -EINVAL If \a portal is not a valid index.
1613 LNetSetLazyPortal(int portal)
1615 lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1617 if (portal < 0 || portal >= the_lnet.ln_nportals)
1620 CDEBUG(D_NET, "Setting portal %d lazy\n", portal);
1623 lnet_portal_setopt(ptl, LNET_PTL_LAZY);
1630 * Turn off the lazy portal attribute. Delayed requests on the portal,
1631 * if any, will be all dropped when this function returns.
1633 * \param portal Index of the portal to disable the lazy attribute on.
1635 * \retval 0 On success.
1636 * \retval -EINVAL If \a portal is not a valid index.
1639 LNetClearLazyPortal(int portal)
1642 lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1645 if (portal < 0 || portal >= the_lnet.ln_nportals)
1650 if (!lnet_portal_is_lazy(ptl)) {
1655 if (the_lnet.ln_shutdown)
1656 CWARN ("Active lazy portal %d on exit\n", portal);
1658 CDEBUG (D_NET, "clearing portal %d lazy\n", portal);
1660 /* grab all the blocked messages atomically */
1661 cfs_list_add(&zombies, &ptl->ptl_msgq);
1662 cfs_list_del_init(&ptl->ptl_msgq);
1664 ptl->ptl_msgq_version++;
1665 lnet_portal_unsetopt(ptl, LNET_PTL_LAZY);
1669 while (!cfs_list_empty(&zombies)) {
1670 msg = cfs_list_entry(zombies.next, lnet_msg_t, msg_list);
1671 cfs_list_del(&msg->msg_list);
1673 lnet_drop_delayed_put(msg, "Clearing lazy portal attr");
1680 lnet_recv_put(lnet_libmd_t *md, lnet_msg_t *msg, int delayed,
1681 unsigned int offset, unsigned int mlength)
1683 lnet_hdr_t *hdr = &msg->msg_hdr;
1687 the_lnet.ln_counters.recv_count++;
1688 the_lnet.ln_counters.recv_length += mlength;
1693 lnet_setpayloadbuffer(msg);
1695 msg->msg_ev.type = LNET_EVENT_PUT;
1696 msg->msg_ev.target.pid = hdr->dest_pid;
1697 msg->msg_ev.target.nid = hdr->dest_nid;
1698 msg->msg_ev.hdr_data = hdr->msg.put.hdr_data;
1700 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1701 * it back into the ACK during lnet_finalize() */
1702 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1703 (md->md_options & LNET_MD_ACK_DISABLE) == 0);
1705 lnet_ni_recv(msg->msg_rxpeer->lp_ni,
1707 msg, delayed, offset, mlength,
1708 hdr->payload_length);
1711 /* called with LNET_LOCK held */
1713 lnet_match_blocked_msg(lnet_libmd_t *md)
1715 CFS_LIST_HEAD (drops);
1716 CFS_LIST_HEAD (matches);
1721 lnet_me_t *me = md->md_me;
1723 LASSERT (me->me_portal < (unsigned int)the_lnet.ln_nportals);
1725 ptl = &the_lnet.ln_portals[me->me_portal];
1726 if (!lnet_portal_is_lazy(ptl)) {
1727 LASSERT (cfs_list_empty(&ptl->ptl_msgq));
1731 LASSERT (md->md_refcount == 0); /* a brand new MD */
1733 cfs_list_for_each_safe (entry, tmp, &ptl->ptl_msgq) {
1736 unsigned int mlength;
1737 unsigned int offset;
1739 lnet_process_id_t src;
1741 msg = cfs_list_entry(entry, lnet_msg_t, msg_list);
1743 LASSERT (msg->msg_delayed);
1745 hdr = &msg->msg_hdr;
1746 index = hdr->msg.put.ptl_index;
1748 src.nid = hdr->src_nid;
1749 src.pid = hdr->src_pid;
1751 rc = lnet_try_match_md(index, LNET_MD_OP_PUT, src,
1752 hdr->payload_length,
1753 hdr->msg.put.offset,
1754 hdr->msg.put.match_bits,
1755 md, msg, &mlength, &offset);
1757 if (rc == LNET_MATCHMD_NONE)
1760 /* Hurrah! This _is_ a match */
1761 cfs_list_del(&msg->msg_list);
1762 ptl->ptl_msgq_version++;
1764 if (rc == LNET_MATCHMD_OK) {
1765 cfs_list_add_tail(&msg->msg_list, &matches);
1767 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
1768 "match "LPU64" offset %d length %d.\n",
1770 hdr->msg.put.ptl_index,
1771 hdr->msg.put.match_bits,
1772 hdr->msg.put.offset,
1773 hdr->payload_length);
1775 LASSERT (rc == LNET_MATCHMD_DROP);
1777 cfs_list_add_tail(&msg->msg_list, &drops);
1780 if (lnet_md_exhausted(md))
1786 cfs_list_for_each_safe (entry, tmp, &drops) {
1787 msg = cfs_list_entry(entry, lnet_msg_t, msg_list);
1789 cfs_list_del(&msg->msg_list);
1791 lnet_drop_delayed_put(msg, "Bad match");
1794 cfs_list_for_each_safe (entry, tmp, &matches) {
1795 msg = cfs_list_entry(entry, lnet_msg_t, msg_list);
1797 cfs_list_del(&msg->msg_list);
1799 /* md won't disappear under me, since each msg
1800 * holds a ref on it */
1801 lnet_recv_put(md, msg, 1,
1803 msg->msg_ev.mlength);
1810 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1815 lnet_hdr_t *hdr = &msg->msg_hdr;
1816 unsigned int rlength = hdr->payload_length;
1817 unsigned int mlength = 0;
1818 unsigned int offset = 0;
1819 lnet_process_id_t src= {0};
1823 src.nid = hdr->src_nid;
1824 src.pid = hdr->src_pid;
1826 /* Convert put fields to host byte order */
1827 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1828 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1829 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1831 index = hdr->msg.put.ptl_index;
1836 rc = lnet_match_md(index, LNET_MD_OP_PUT, src,
1837 rlength, hdr->msg.put.offset,
1838 hdr->msg.put.match_bits, msg,
1839 &mlength, &offset, &md);
1844 case LNET_MATCHMD_OK:
1846 lnet_recv_put(md, msg, msg->msg_delayed, offset, mlength);
1849 case LNET_MATCHMD_NONE:
1850 ptl = &the_lnet.ln_portals[index];
1851 version = ptl->ptl_ml_version;
1854 if (!msg->msg_delayed)
1855 rc = lnet_eager_recv_locked(msg);
1858 !the_lnet.ln_shutdown &&
1859 lnet_portal_is_lazy(ptl)) {
1860 if (version != ptl->ptl_ml_version)
1863 cfs_list_add_tail(&msg->msg_list, &ptl->ptl_msgq);
1864 ptl->ptl_msgq_version++;
1867 CDEBUG(D_NET, "Delaying PUT from %s portal %d match "
1868 LPU64" offset %d length %d: no match \n",
1869 libcfs_id2str(src), index,
1870 hdr->msg.put.match_bits,
1871 hdr->msg.put.offset, rlength);
1876 case LNET_MATCHMD_DROP:
1877 CNETERR("Dropping PUT from %s portal %d match "LPU64
1878 " offset %d length %d: %d\n",
1879 libcfs_id2str(src), index,
1880 hdr->msg.put.match_bits,
1881 hdr->msg.put.offset, rlength, rc);
1884 return ENOENT; /* +ve: OK but no match */
1889 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1891 lnet_hdr_t *hdr = &msg->msg_hdr;
1892 unsigned int mlength = 0;
1893 unsigned int offset = 0;
1894 lnet_process_id_t src = {0};
1895 lnet_handle_wire_t reply_wmd;
1899 src.nid = hdr->src_nid;
1900 src.pid = hdr->src_pid;
1902 /* Convert get fields to host byte order */
1903 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1904 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1905 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1906 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1910 rc = lnet_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src,
1911 hdr->msg.get.sink_length, hdr->msg.get.src_offset,
1912 hdr->msg.get.match_bits, msg,
1913 &mlength, &offset, &md);
1914 if (rc == LNET_MATCHMD_DROP) {
1915 CNETERR("Dropping GET from %s portal %d match "LPU64
1916 " offset %d length %d\n",
1918 hdr->msg.get.ptl_index,
1919 hdr->msg.get.match_bits,
1920 hdr->msg.get.src_offset,
1921 hdr->msg.get.sink_length);
1923 return ENOENT; /* +ve: OK but no match */
1926 LASSERT (rc == LNET_MATCHMD_OK);
1928 the_lnet.ln_counters.send_count++;
1929 the_lnet.ln_counters.send_length += mlength;
1933 msg->msg_ev.type = LNET_EVENT_GET;
1934 msg->msg_ev.target.pid = hdr->dest_pid;
1935 msg->msg_ev.target.nid = hdr->dest_nid;
1936 msg->msg_ev.hdr_data = 0;
1938 reply_wmd = hdr->msg.get.return_wmd;
1940 lnet_prep_send(msg, LNET_MSG_REPLY, src, offset, mlength);
1942 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1945 /* The LND completes the REPLY from her recv procedure */
1946 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1947 msg->msg_offset, msg->msg_len, msg->msg_len);
1951 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1952 msg->msg_receiving = 0;
1954 rc = lnet_send(ni->ni_nid, msg);
1956 /* didn't get as far as lnet_ni_send() */
1957 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1958 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc);
1960 lnet_finalize(ni, msg, rc);
1967 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1969 void *private = msg->msg_private;
1970 lnet_hdr_t *hdr = &msg->msg_hdr;
1971 lnet_process_id_t src = {0};
1978 src.nid = hdr->src_nid;
1979 src.pid = hdr->src_pid;
1981 /* NB handles only looked up by creator (no flips) */
1982 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1983 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1984 CNETERR("%s: Dropping REPLY from %s for %s "
1985 "MD "LPX64"."LPX64"\n",
1986 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1987 (md == NULL) ? "invalid" : "inactive",
1988 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1989 hdr->msg.reply.dst_wmd.wh_object_cookie);
1990 if (md != NULL && md->md_me != NULL)
1991 CERROR("REPLY MD also attached to portal %d\n",
1992 md->md_me->me_portal);
1995 return ENOENT; /* +ve: OK but no match */
1998 LASSERT (md->md_offset == 0);
2000 rlength = hdr->payload_length;
2001 mlength = MIN(rlength, (int)md->md_length);
2003 if (mlength < rlength &&
2004 (md->md_options & LNET_MD_TRUNCATE) == 0) {
2005 CNETERR("%s: Dropping REPLY from %s length %d "
2006 "for MD "LPX64" would overflow (%d)\n",
2007 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2008 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
2011 return ENOENT; /* +ve: OK but no match */
2014 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
2015 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2016 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
2018 lnet_commit_md(md, msg);
2021 lnet_setpayloadbuffer(msg);
2023 msg->msg_ev.type = LNET_EVENT_REPLY;
2024 msg->msg_ev.target.pid = hdr->dest_pid;
2025 msg->msg_ev.target.nid = hdr->dest_nid;
2026 msg->msg_ev.initiator = src;
2027 msg->msg_ev.rlength = rlength;
2028 msg->msg_ev.mlength = mlength;
2029 msg->msg_ev.offset = 0;
2031 lnet_md_deconstruct(md, &msg->msg_ev.md);
2032 lnet_md2handle(&msg->msg_ev.md_handle, md);
2034 the_lnet.ln_counters.recv_count++;
2035 the_lnet.ln_counters.recv_length += mlength;
2039 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
2044 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
2046 lnet_hdr_t *hdr = &msg->msg_hdr;
2047 lnet_process_id_t src = {0};
2050 src.nid = hdr->src_nid;
2051 src.pid = hdr->src_pid;
2053 /* Convert ack fields to host byte order */
2054 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
2055 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
2059 /* NB handles only looked up by creator (no flips) */
2060 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
2061 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2062 /* Don't moan; this is expected */
2064 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
2065 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2066 (md == NULL) ? "invalid" : "inactive",
2067 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2068 hdr->msg.ack.dst_wmd.wh_object_cookie);
2069 if (md != NULL && md->md_me != NULL)
2070 CERROR("Source MD also attached to portal %d\n",
2071 md->md_me->me_portal);
2074 return ENOENT; /* +ve! */
2077 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
2078 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
2079 hdr->msg.ack.dst_wmd.wh_object_cookie);
2081 lnet_commit_md(md, msg);
2083 msg->msg_ev.type = LNET_EVENT_ACK;
2084 msg->msg_ev.target.pid = hdr->dest_pid;
2085 msg->msg_ev.target.nid = hdr->dest_nid;
2086 msg->msg_ev.initiator = src;
2087 msg->msg_ev.mlength = hdr->msg.ack.mlength;
2088 msg->msg_ev.match_bits = hdr->msg.ack.match_bits;
2090 lnet_md_deconstruct(md, &msg->msg_ev.md);
2091 lnet_md2handle(&msg->msg_ev.md_handle, md);
2093 the_lnet.ln_counters.recv_count++;
2097 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
2102 lnet_msgtyp2str (int type)
2111 case LNET_MSG_REPLY:
2113 case LNET_MSG_HELLO:
2116 return ("<UNKNOWN>");
2121 lnet_print_hdr(lnet_hdr_t * hdr)
2123 lnet_process_id_t src = {0};
2124 lnet_process_id_t dst = {0};
2125 char *type_str = lnet_msgtyp2str (hdr->type);
2127 src.nid = hdr->src_nid;
2128 src.pid = hdr->src_pid;
2130 dst.nid = hdr->dest_nid;
2131 dst.pid = hdr->dest_pid;
2133 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
2134 CWARN(" From %s\n", libcfs_id2str(src));
2135 CWARN(" To %s\n", libcfs_id2str(dst));
2137 switch (hdr->type) {
2142 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
2143 "match bits "LPU64"\n",
2144 hdr->msg.put.ptl_index,
2145 hdr->msg.put.ack_wmd.wh_interface_cookie,
2146 hdr->msg.put.ack_wmd.wh_object_cookie,
2147 hdr->msg.put.match_bits);
2148 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
2149 hdr->payload_length, hdr->msg.put.offset,
2150 hdr->msg.put.hdr_data);
2154 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
2155 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
2156 hdr->msg.get.return_wmd.wh_interface_cookie,
2157 hdr->msg.get.return_wmd.wh_object_cookie,
2158 hdr->msg.get.match_bits);
2159 CWARN(" Length %d, src offset %d\n",
2160 hdr->msg.get.sink_length,
2161 hdr->msg.get.src_offset);
2165 CWARN(" dst md "LPX64"."LPX64", "
2166 "manipulated length %d\n",
2167 hdr->msg.ack.dst_wmd.wh_interface_cookie,
2168 hdr->msg.ack.dst_wmd.wh_object_cookie,
2169 hdr->msg.ack.mlength);
2172 case LNET_MSG_REPLY:
2173 CWARN(" dst md "LPX64"."LPX64", "
2175 hdr->msg.reply.dst_wmd.wh_interface_cookie,
2176 hdr->msg.reply.dst_wmd.wh_object_cookie,
2177 hdr->payload_length);
2183 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
2184 void *private, int rdma_req)
2189 lnet_pid_t dest_pid;
2190 lnet_nid_t dest_nid;
2192 __u32 payload_length;
2195 LASSERT (!cfs_in_interrupt ());
2197 type = le32_to_cpu(hdr->type);
2198 src_nid = le64_to_cpu(hdr->src_nid);
2199 dest_nid = le64_to_cpu(hdr->dest_nid);
2200 dest_pid = le32_to_cpu(hdr->dest_pid);
2201 payload_length = le32_to_cpu(hdr->payload_length);
2203 for_me = (ni->ni_nid == dest_nid);
2208 if (payload_length > 0) {
2209 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
2210 libcfs_nid2str(from_nid),
2211 libcfs_nid2str(src_nid),
2212 lnet_msgtyp2str(type), payload_length);
2218 case LNET_MSG_REPLY:
2219 if (payload_length > (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
2220 CERROR("%s, src %s: bad %s payload %d "
2221 "(%d max expected)\n",
2222 libcfs_nid2str(from_nid),
2223 libcfs_nid2str(src_nid),
2224 lnet_msgtyp2str(type),
2226 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
2232 CERROR("%s, src %s: Bad message type 0x%x\n",
2233 libcfs_nid2str(from_nid),
2234 libcfs_nid2str(src_nid), type);
2238 if (the_lnet.ln_routing) {
2239 cfs_time_t now = cfs_time_current();
2243 ni->ni_last_alive = now;
2244 if (ni->ni_status != NULL &&
2245 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
2246 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
2251 /* Regard a bad destination NID as a protocol error. Senders should
2252 * know what they're doing; if they don't they're misconfigured, buggy
2253 * or malicious so we chop them off at the knees :) */
2256 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2257 /* should have gone direct */
2258 CERROR ("%s, src %s: Bad dest nid %s "
2259 "(should have been sent direct)\n",
2260 libcfs_nid2str(from_nid),
2261 libcfs_nid2str(src_nid),
2262 libcfs_nid2str(dest_nid));
2266 if (lnet_islocalnid(dest_nid)) {
2267 /* dest is another local NI; sender should have used
2268 * this node's NID on its own network */
2269 CERROR ("%s, src %s: Bad dest nid %s "
2270 "(it's my nid but on a different network)\n",
2271 libcfs_nid2str(from_nid),
2272 libcfs_nid2str(src_nid),
2273 libcfs_nid2str(dest_nid));
2277 if (rdma_req && type == LNET_MSG_GET) {
2278 CERROR ("%s, src %s: Bad optimized GET for %s "
2279 "(final destination must be me)\n",
2280 libcfs_nid2str(from_nid),
2281 libcfs_nid2str(src_nid),
2282 libcfs_nid2str(dest_nid));
2286 if (!the_lnet.ln_routing) {
2287 CERROR ("%s, src %s: Dropping message for %s "
2288 "(routing not enabled)\n",
2289 libcfs_nid2str(from_nid),
2290 libcfs_nid2str(src_nid),
2291 libcfs_nid2str(dest_nid));
2296 /* Message looks OK; we're not going to return an error, so we MUST
2297 * call back lnd_recv() come what may... */
2299 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2300 fail_peer (src_nid, 0)) /* shall we now? */
2302 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2303 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2304 lnet_msgtyp2str(type));
2308 msg = lnet_msg_alloc();
2310 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2311 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2312 lnet_msgtyp2str(type));
2316 /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
2318 msg->msg_type = type;
2319 msg->msg_private = private;
2320 msg->msg_receiving = 1;
2321 msg->msg_len = msg->msg_wanted = payload_length;
2322 msg->msg_offset = 0;
2323 msg->msg_hdr = *hdr;
2326 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid);
2329 CERROR("%s, src %s: Dropping %s "
2330 "(error %d looking up sender)\n",
2331 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2332 lnet_msgtyp2str(type), rc);
2341 msg->msg_target.pid = dest_pid;
2342 msg->msg_target.nid = dest_nid;
2343 msg->msg_routing = 1;
2344 msg->msg_offset = 0;
2347 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
2348 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2349 rc = lnet_eager_recv_locked(msg);
2355 lnet_commit_routedmsg(msg);
2356 rc = lnet_post_routed_recv_locked(msg, 0);
2360 lnet_ni_recv(ni, msg->msg_private, msg, 0,
2361 0, payload_length, payload_length);
2365 /* convert common msg->hdr fields to host byteorder */
2366 msg->msg_hdr.type = type;
2367 msg->msg_hdr.src_nid = src_nid;
2368 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2369 msg->msg_hdr.dest_nid = dest_nid;
2370 msg->msg_hdr.dest_pid = dest_pid;
2371 msg->msg_hdr.payload_length = payload_length;
2373 msg->msg_ev.sender = from_nid;
2377 rc = lnet_parse_ack(ni, msg);
2380 rc = lnet_parse_put(ni, msg);
2383 rc = lnet_parse_get(ni, msg, rdma_req);
2385 case LNET_MSG_REPLY:
2386 rc = lnet_parse_reply(ni, msg);
2390 goto free_drop; /* prevent an unused label if !kernel */
2396 LASSERT (rc == ENOENT);
2399 LASSERT (msg->msg_md == NULL);
2401 if (msg->msg_rxpeer != NULL) {
2402 lnet_peer_decref_locked(msg->msg_rxpeer);
2403 msg->msg_rxpeer = NULL;
2405 lnet_msg_free_locked(msg); /* expects LNET_LOCK held */
2409 lnet_drop_message(ni, private, payload_length);
2414 * Initiate an asynchronous PUT operation.
2416 * There are several events associated with a PUT: completion of the send on
2417 * the initiator node (LNET_EVENT_SEND), and when the send completes
2418 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2419 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2420 * used at the target node to indicate the completion of incoming data
2423 * The local events will be logged in the EQ associated with the MD pointed to
2424 * by \a mdh handle. Using a MD without an associated EQ results in these
2425 * events being discarded. In this case, the caller must have another
2426 * mechanism (e.g., a higher level protocol) for determining when it is safe
2427 * to modify the memory region associated with the MD.
2429 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2430 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2432 * \param self Indicates the NID of a local interface through which to send
2433 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2434 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2435 * must be "free floating" (See LNetMDBind()).
2436 * \param ack Controls whether an acknowledgment is requested.
2437 * Acknowledgments are only sent when they are requested by the initiating
2438 * process and the target MD enables them.
2439 * \param target A process identifier for the target process.
2440 * \param portal The index in the \a target's portal table.
2441 * \param match_bits The match bits to use for MD selection at the target
2443 * \param offset The offset into the target MD (only used when the target
2444 * MD has the LNET_MD_MANAGE_REMOTE option set).
2445 * \param hdr_data 64 bits of user data that can be included in the message
2446 * header. This data is written to an event queue entry at the target if an
2447 * EQ is present on the matching MD.
2449 * \retval 0 Success, and only in this case events will be generated
2450 * and logged to EQ (if it exists).
2451 * \retval -EIO Simulated failure.
2452 * \retval -ENOMEM Memory allocation failure.
2453 * \retval -ENOENT Invalid MD object.
2455 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2458 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2459 lnet_process_id_t target, unsigned int portal,
2460 __u64 match_bits, unsigned int offset,
2467 LASSERT (the_lnet.ln_init);
2468 LASSERT (the_lnet.ln_refcount > 0);
2470 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2471 fail_peer (target.nid, 1)) /* shall we now? */
2473 CERROR("Dropping PUT to %s: simulated failure\n",
2474 libcfs_id2str(target));
2478 msg = lnet_msg_alloc();
2480 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2481 libcfs_id2str(target));
2484 msg->msg_vmflush = !!cfs_memory_pressure_get();
2488 md = lnet_handle2md(&mdh);
2489 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2490 lnet_msg_free_locked(msg);
2492 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2493 match_bits, portal, libcfs_id2str(target),
2494 md == NULL ? -1 : md->md_threshold);
2495 if (md != NULL && md->md_me != NULL)
2496 CERROR("Source MD also attached to portal %d\n",
2497 md->md_me->me_portal);
2503 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2505 lnet_commit_md(md, msg);
2507 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2509 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2510 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2511 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2512 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2514 /* NB handles only looked up by creator (no flips) */
2515 if (ack == LNET_ACK_REQ) {
2516 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2517 the_lnet.ln_interface_cookie;
2518 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2519 md->md_lh.lh_cookie;
2521 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2522 LNET_WIRE_HANDLE_COOKIE_NONE;
2523 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2524 LNET_WIRE_HANDLE_COOKIE_NONE;
2527 msg->msg_ev.type = LNET_EVENT_SEND;
2528 msg->msg_ev.initiator.nid = LNET_NID_ANY;
2529 msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2530 msg->msg_ev.target = target;
2531 msg->msg_ev.sender = LNET_NID_ANY;
2532 msg->msg_ev.pt_index = portal;
2533 msg->msg_ev.match_bits = match_bits;
2534 msg->msg_ev.rlength = md->md_length;
2535 msg->msg_ev.mlength = md->md_length;
2536 msg->msg_ev.offset = offset;
2537 msg->msg_ev.hdr_data = hdr_data;
2539 lnet_md_deconstruct(md, &msg->msg_ev.md);
2540 lnet_md2handle(&msg->msg_ev.md_handle, md);
2542 the_lnet.ln_counters.send_count++;
2543 the_lnet.ln_counters.send_length += md->md_length;
2547 rc = lnet_send(self, msg);
2549 CNETERR( "Error sending PUT to %s: %d\n",
2550 libcfs_id2str(target), rc);
2551 lnet_finalize (NULL, msg, rc);
2554 /* completion will be signalled by an event */
2559 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2561 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2562 * returns a msg for the LND to pass to lnet_finalize() when the sink
2563 * data has been received.
2565 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2566 * lnet_finalize() is called on it, so the LND must call this first */
2568 lnet_msg_t *msg = lnet_msg_alloc();
2569 lnet_libmd_t *getmd = getmsg->msg_md;
2570 lnet_process_id_t peer_id = getmsg->msg_target;
2572 LASSERT (!getmsg->msg_target_is_router);
2573 LASSERT (!getmsg->msg_routing);
2577 LASSERT (getmd->md_refcount > 0);
2580 CERROR ("%s: Dropping REPLY from %s: can't allocate msg\n",
2581 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2585 if (getmd->md_threshold == 0) {
2586 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2587 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2592 LASSERT (getmd->md_offset == 0);
2594 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2595 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2597 lnet_commit_md (getmd, msg);
2599 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2601 msg->msg_ev.type = LNET_EVENT_REPLY;
2602 msg->msg_ev.initiator = peer_id;
2603 msg->msg_ev.sender = peer_id.nid; /* optimized GETs can't be routed */
2604 msg->msg_ev.rlength = msg->msg_ev.mlength = getmd->md_length;
2605 msg->msg_ev.offset = 0;
2607 lnet_md_deconstruct(getmd, &msg->msg_ev.md);
2608 lnet_md2handle(&msg->msg_ev.md_handle, getmd);
2610 the_lnet.ln_counters.recv_count++;
2611 the_lnet.ln_counters.recv_length += getmd->md_length;
2618 lnet_msg_free_locked(msg);
2620 the_lnet.ln_counters.drop_count++;
2621 the_lnet.ln_counters.drop_length += getmd->md_length;
2629 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2631 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2632 * completed and I know it. */
2633 LASSERT (reply != NULL);
2634 LASSERT (reply->msg_type == LNET_MSG_GET);
2635 LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2637 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2638 * the end of my buffer, I might as well be dead. */
2639 LASSERT (len <= reply->msg_ev.mlength);
2641 reply->msg_ev.mlength = len;
2645 * Initiate an asynchronous GET operation.
2647 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2648 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2649 * the target node in the REPLY has been written to local MD.
2651 * On the target node, an LNET_EVENT_GET is logged when the GET request
2652 * arrives and is accepted into a MD.
2654 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2655 * \param mdh A handle for the MD that describes the memory into which the
2656 * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
2658 * \retval 0 Success, and only in this case events will be generated
2659 * and logged to EQ (if it exists) of the MD.
2660 * \retval -EIO Simulated failure.
2661 * \retval -ENOMEM Memory allocation failure.
2662 * \retval -ENOENT Invalid MD object.
2665 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2666 lnet_process_id_t target, unsigned int portal,
2667 __u64 match_bits, unsigned int offset)
2673 LASSERT (the_lnet.ln_init);
2674 LASSERT (the_lnet.ln_refcount > 0);
2676 if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2677 fail_peer (target.nid, 1)) /* shall we now? */
2679 CERROR("Dropping GET to %s: simulated failure\n",
2680 libcfs_id2str(target));
2684 msg = lnet_msg_alloc();
2686 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2687 libcfs_id2str(target));
2693 md = lnet_handle2md(&mdh);
2694 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2695 lnet_msg_free_locked(msg);
2697 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2698 match_bits, portal, libcfs_id2str(target),
2699 md == NULL ? -1 : md->md_threshold);
2700 if (md != NULL && md->md_me != NULL)
2701 CERROR("REPLY MD also attached to portal %d\n",
2702 md->md_me->me_portal);
2708 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2710 lnet_commit_md(md, msg);
2712 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2714 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2715 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2716 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2717 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2719 /* NB handles only looked up by creator (no flips) */
2720 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2721 the_lnet.ln_interface_cookie;
2722 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2723 md->md_lh.lh_cookie;
2725 msg->msg_ev.type = LNET_EVENT_SEND;
2726 msg->msg_ev.initiator.nid = LNET_NID_ANY;
2727 msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2728 msg->msg_ev.target = target;
2729 msg->msg_ev.sender = LNET_NID_ANY;
2730 msg->msg_ev.pt_index = portal;
2731 msg->msg_ev.match_bits = match_bits;
2732 msg->msg_ev.rlength = md->md_length;
2733 msg->msg_ev.mlength = md->md_length;
2734 msg->msg_ev.offset = offset;
2735 msg->msg_ev.hdr_data = 0;
2737 lnet_md_deconstruct(md, &msg->msg_ev.md);
2738 lnet_md2handle(&msg->msg_ev.md_handle, md);
2740 the_lnet.ln_counters.send_count++;
2744 rc = lnet_send(self, msg);
2746 CNETERR( "Error sending GET to %s: %d\n",
2747 libcfs_id2str(target), rc);
2748 lnet_finalize (NULL, msg, rc);
2751 /* completion will be signalled by an event */
2756 * Calculate distance to node at \a dstnid.
2758 * \param dstnid Target NID.
2759 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2761 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2764 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2765 * local_nid_dist_zero is set, which is the default.
2766 * \retval positives Distance to target NID, i.e. number of hops plus one.
2767 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2770 LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2774 lnet_remotenet_t *rnet;
2775 __u32 dstnet = LNET_NIDNET(dstnid);
2779 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2780 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2781 * keep order 0 free for 0@lo and order 1 free for a local NID
2784 LASSERT (the_lnet.ln_init);
2785 LASSERT (the_lnet.ln_refcount > 0);
2789 cfs_list_for_each (e, &the_lnet.ln_nis) {
2790 ni = cfs_list_entry(e, lnet_ni_t, ni_list);
2792 if (ni->ni_nid == dstnid) {
2793 if (srcnidp != NULL)
2795 if (orderp != NULL) {
2796 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2803 return local_nid_dist_zero ? 0 : 1;
2806 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2807 if (srcnidp != NULL)
2808 *srcnidp = ni->ni_nid;
2818 cfs_list_for_each (e, &the_lnet.ln_remote_nets) {
2819 rnet = cfs_list_entry(e, lnet_remotenet_t, lrn_list);
2821 if (rnet->lrn_net == dstnet) {
2822 lnet_route_t *route;
2823 lnet_route_t *shortest = NULL;
2825 LASSERT (!cfs_list_empty(&rnet->lrn_routes));
2827 cfs_list_for_each_entry(route, &rnet->lrn_routes,
2829 if (shortest == NULL ||
2830 route->lr_hops < shortest->lr_hops)
2834 LASSERT (shortest != NULL);
2835 hops = shortest->lr_hops;
2836 if (srcnidp != NULL)
2837 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2847 return -EHOSTUNREACH;
2851 * Set the number of asynchronous messages expected from a target process.
2853 * This function is only meaningful for userspace callers. It's a no-op when
2854 * called from kernel.
2856 * Asynchronous messages are those that can come from a target when the
2857 * userspace process is not waiting for IO to complete; e.g., AST callbacks
2858 * from Lustre servers. Specifying the expected number of such messages
2859 * allows them to be eagerly received when user process is not running in
2860 * LNet; otherwise network errors may occur.
2862 * \param id Process ID of the target process.
2863 * \param nasync Number of asynchronous messages expected from the target.
2865 * \return 0 on success, and an error code otherwise.
2868 LNetSetAsync(lnet_process_id_t id, int nasync)
2874 lnet_remotenet_t *rnet;
2876 lnet_route_t *route;
2883 /* Target on a local network? */
2884 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2886 if (ni->ni_lnd->lnd_setasync != NULL)
2887 rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2892 /* Target on a remote network: apply to routers */
2894 LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2899 /* Snapshot all the router NIDs */
2901 rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2903 cfs_list_for_each(tmp, &rnet->lrn_routes) {
2904 if (nnids == maxnids) {
2906 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2911 route = cfs_list_entry(tmp, lnet_route_t, lr_list);
2912 nids[nnids++] = route->lr_gateway->lp_nid;
2917 /* set async on all the routers */
2918 while (nnids-- > 0) {
2919 id.pid = LUSTRE_SRV_LNET_PID;
2920 id.nid = nids[nnids];
2922 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2926 if (ni->ni_lnd->lnd_setasync != NULL) {
2927 rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2934 LIBCFS_FREE(nids, maxnids * sizeof(*nids));