Whamcloud - gitweb
LU-12339 lnet: select LO interface for sending
[fs/lustre-release.git] / lnet / lnet / lib-move.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lnet/lnet/lib-move.c
33  *
34  * Data movement routines
35  */
36
37 #define DEBUG_SUBSYSTEM S_LNET
38
39 #include <linux/pagemap.h>
40
41 #include <lnet/lib-lnet.h>
42 #include <linux/nsproxy.h>
43 #include <net/net_namespace.h>
44
45 static int local_nid_dist_zero = 1;
46 module_param(local_nid_dist_zero, int, 0444);
47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
48
49 struct lnet_send_data {
50         struct lnet_ni *sd_best_ni;
51         struct lnet_peer_ni *sd_best_lpni;
52         struct lnet_peer_ni *sd_final_dst_lpni;
53         struct lnet_peer *sd_peer;
54         struct lnet_peer *sd_gw_peer;
55         struct lnet_peer_ni *sd_gw_lpni;
56         struct lnet_peer_net *sd_peer_net;
57         struct lnet_msg *sd_msg;
58         lnet_nid_t sd_dst_nid;
59         lnet_nid_t sd_src_nid;
60         lnet_nid_t sd_rtr_nid;
61         int sd_cpt;
62         int sd_md_cpt;
63         __u32 sd_send_case;
64 };
65
66 static inline struct lnet_comm_count *
67 get_stats_counts(struct lnet_element_stats *stats,
68                  enum lnet_stats_type stats_type)
69 {
70         switch (stats_type) {
71         case LNET_STATS_TYPE_SEND:
72                 return &stats->el_send_stats;
73         case LNET_STATS_TYPE_RECV:
74                 return &stats->el_recv_stats;
75         case LNET_STATS_TYPE_DROP:
76                 return &stats->el_drop_stats;
77         default:
78                 CERROR("Unknown stats type\n");
79         }
80
81         return NULL;
82 }
83
84 void lnet_incr_stats(struct lnet_element_stats *stats,
85                      enum lnet_msg_type msg_type,
86                      enum lnet_stats_type stats_type)
87 {
88         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
89         if (!counts)
90                 return;
91
92         switch (msg_type) {
93         case LNET_MSG_ACK:
94                 atomic_inc(&counts->co_ack_count);
95                 break;
96         case LNET_MSG_PUT:
97                 atomic_inc(&counts->co_put_count);
98                 break;
99         case LNET_MSG_GET:
100                 atomic_inc(&counts->co_get_count);
101                 break;
102         case LNET_MSG_REPLY:
103                 atomic_inc(&counts->co_reply_count);
104                 break;
105         case LNET_MSG_HELLO:
106                 atomic_inc(&counts->co_hello_count);
107                 break;
108         default:
109                 CERROR("There is a BUG in the code. Unknown message type\n");
110                 break;
111         }
112 }
113
114 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
115                      enum lnet_stats_type stats_type)
116 {
117         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
118         if (!counts)
119                 return 0;
120
121         return (atomic_read(&counts->co_ack_count) +
122                 atomic_read(&counts->co_put_count) +
123                 atomic_read(&counts->co_get_count) +
124                 atomic_read(&counts->co_reply_count) +
125                 atomic_read(&counts->co_hello_count));
126 }
127
128 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
129                                 struct lnet_comm_count *counts)
130 {
131         msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
132         msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
133         msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
134         msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
135         msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
136 }
137
138 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
139                               struct lnet_element_stats *stats)
140 {
141         struct lnet_comm_count *counts;
142
143         LASSERT(msg_stats);
144         LASSERT(stats);
145
146         counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
147         if (!counts)
148                 return;
149         assign_stats(&msg_stats->im_send_stats, counts);
150
151         counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
152         if (!counts)
153                 return;
154         assign_stats(&msg_stats->im_recv_stats, counts);
155
156         counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
157         if (!counts)
158                 return;
159         assign_stats(&msg_stats->im_drop_stats, counts);
160 }
161
162 int
163 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
164 {
165         struct lnet_test_peer *tp;
166         struct list_head *el;
167         struct list_head *next;
168         struct list_head  cull;
169
170         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
171         if (threshold != 0) {
172                 /* Adding a new entry */
173                 LIBCFS_ALLOC(tp, sizeof(*tp));
174                 if (tp == NULL)
175                         return -ENOMEM;
176
177                 tp->tp_nid = nid;
178                 tp->tp_threshold = threshold;
179
180                 lnet_net_lock(0);
181                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
182                 lnet_net_unlock(0);
183                 return 0;
184         }
185
186         /* removing entries */
187         INIT_LIST_HEAD(&cull);
188
189         lnet_net_lock(0);
190
191         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
192                 tp = list_entry(el, struct lnet_test_peer, tp_list);
193
194                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
195                     nid == LNET_NID_ANY ||      /* removing all entries */
196                     tp->tp_nid == nid) {        /* matched this one */
197                         list_del(&tp->tp_list);
198                         list_add(&tp->tp_list, &cull);
199                 }
200         }
201
202         lnet_net_unlock(0);
203
204         while (!list_empty(&cull)) {
205                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
206
207                 list_del(&tp->tp_list);
208                 LIBCFS_FREE(tp, sizeof(*tp));
209         }
210         return 0;
211 }
212
213 static int
214 fail_peer (lnet_nid_t nid, int outgoing)
215 {
216         struct lnet_test_peer *tp;
217         struct list_head *el;
218         struct list_head *next;
219         struct list_head  cull;
220         int               fail = 0;
221
222         INIT_LIST_HEAD(&cull);
223
224         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
225         lnet_net_lock(0);
226
227         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
228                 tp = list_entry(el, struct lnet_test_peer, tp_list);
229
230                 if (tp->tp_threshold == 0) {
231                         /* zombie entry */
232                         if (outgoing) {
233                                 /* only cull zombies on outgoing tests,
234                                  * since we may be at interrupt priority on
235                                  * incoming messages. */
236                                 list_del(&tp->tp_list);
237                                 list_add(&tp->tp_list, &cull);
238                         }
239                         continue;
240                 }
241
242                 if (tp->tp_nid == LNET_NID_ANY ||       /* fail every peer */
243                     nid == tp->tp_nid) {                /* fail this peer */
244                         fail = 1;
245
246                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
247                                 tp->tp_threshold--;
248                                 if (outgoing &&
249                                     tp->tp_threshold == 0) {
250                                         /* see above */
251                                         list_del(&tp->tp_list);
252                                         list_add(&tp->tp_list, &cull);
253                                 }
254                         }
255                         break;
256                 }
257         }
258
259         lnet_net_unlock(0);
260
261         while (!list_empty(&cull)) {
262                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
263                 list_del(&tp->tp_list);
264
265                 LIBCFS_FREE(tp, sizeof(*tp));
266         }
267
268         return fail;
269 }
270
271 unsigned int
272 lnet_iov_nob(unsigned int niov, struct kvec *iov)
273 {
274         unsigned int nob = 0;
275
276         LASSERT(niov == 0 || iov != NULL);
277         while (niov-- > 0)
278                 nob += (iov++)->iov_len;
279
280         return (nob);
281 }
282 EXPORT_SYMBOL(lnet_iov_nob);
283
284 void
285 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
286                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
287                   unsigned int nob)
288 {
289         /* NB diov, siov are READ-ONLY */
290         unsigned int  this_nob;
291
292         if (nob == 0)
293                 return;
294
295         /* skip complete frags before 'doffset' */
296         LASSERT(ndiov > 0);
297         while (doffset >= diov->iov_len) {
298                 doffset -= diov->iov_len;
299                 diov++;
300                 ndiov--;
301                 LASSERT(ndiov > 0);
302         }
303
304         /* skip complete frags before 'soffset' */
305         LASSERT(nsiov > 0);
306         while (soffset >= siov->iov_len) {
307                 soffset -= siov->iov_len;
308                 siov++;
309                 nsiov--;
310                 LASSERT(nsiov > 0);
311         }
312
313         do {
314                 LASSERT(ndiov > 0);
315                 LASSERT(nsiov > 0);
316                 this_nob = MIN(diov->iov_len - doffset,
317                                siov->iov_len - soffset);
318                 this_nob = MIN(this_nob, nob);
319
320                 memcpy((char *)diov->iov_base + doffset,
321                        (char *)siov->iov_base + soffset, this_nob);
322                 nob -= this_nob;
323
324                 if (diov->iov_len > doffset + this_nob) {
325                         doffset += this_nob;
326                 } else {
327                         diov++;
328                         ndiov--;
329                         doffset = 0;
330                 }
331
332                 if (siov->iov_len > soffset + this_nob) {
333                         soffset += this_nob;
334                 } else {
335                         siov++;
336                         nsiov--;
337                         soffset = 0;
338                 }
339         } while (nob > 0);
340 }
341 EXPORT_SYMBOL(lnet_copy_iov2iov);
342
343 int
344 lnet_extract_iov(int dst_niov, struct kvec *dst,
345                  int src_niov, struct kvec *src,
346                  unsigned int offset, unsigned int len)
347 {
348         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
349          * for exactly 'len' bytes, and return the number of entries.
350          * NB not destructive to 'src' */
351         unsigned int    frag_len;
352         unsigned int    niov;
353
354         if (len == 0)                           /* no data => */
355                 return (0);                     /* no frags */
356
357         LASSERT(src_niov > 0);
358         while (offset >= src->iov_len) {      /* skip initial frags */
359                 offset -= src->iov_len;
360                 src_niov--;
361                 src++;
362                 LASSERT(src_niov > 0);
363         }
364
365         niov = 1;
366         for (;;) {
367                 LASSERT(src_niov > 0);
368                 LASSERT((int)niov <= dst_niov);
369
370                 frag_len = src->iov_len - offset;
371                 dst->iov_base = ((char *)src->iov_base) + offset;
372
373                 if (len <= frag_len) {
374                         dst->iov_len = len;
375                         return (niov);
376                 }
377
378                 dst->iov_len = frag_len;
379
380                 len -= frag_len;
381                 dst++;
382                 src++;
383                 niov++;
384                 src_niov--;
385                 offset = 0;
386         }
387 }
388 EXPORT_SYMBOL(lnet_extract_iov);
389
390
391 unsigned int
392 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
393 {
394         unsigned int  nob = 0;
395
396         LASSERT(niov == 0 || kiov != NULL);
397         while (niov-- > 0)
398                 nob += (kiov++)->kiov_len;
399
400         return (nob);
401 }
402 EXPORT_SYMBOL(lnet_kiov_nob);
403
404 void
405 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
406                     unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
407                     unsigned int nob)
408 {
409         /* NB diov, siov are READ-ONLY */
410         unsigned int    this_nob;
411         char           *daddr = NULL;
412         char           *saddr = NULL;
413
414         if (nob == 0)
415                 return;
416
417         LASSERT (!in_interrupt ());
418
419         LASSERT (ndiov > 0);
420         while (doffset >= diov->kiov_len) {
421                 doffset -= diov->kiov_len;
422                 diov++;
423                 ndiov--;
424                 LASSERT(ndiov > 0);
425         }
426
427         LASSERT(nsiov > 0);
428         while (soffset >= siov->kiov_len) {
429                 soffset -= siov->kiov_len;
430                 siov++;
431                 nsiov--;
432                 LASSERT(nsiov > 0);
433         }
434
435         do {
436                 LASSERT(ndiov > 0);
437                 LASSERT(nsiov > 0);
438                 this_nob = MIN(diov->kiov_len - doffset,
439                                siov->kiov_len - soffset);
440                 this_nob = MIN(this_nob, nob);
441
442                 if (daddr == NULL)
443                         daddr = ((char *)kmap(diov->kiov_page)) +
444                                 diov->kiov_offset + doffset;
445                 if (saddr == NULL)
446                         saddr = ((char *)kmap(siov->kiov_page)) +
447                                 siov->kiov_offset + soffset;
448
449                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
450                  * However in practice at least one of the kiovs will be mapped
451                  * kernel pages and the map/unmap will be NOOPs */
452
453                 memcpy (daddr, saddr, this_nob);
454                 nob -= this_nob;
455
456                 if (diov->kiov_len > doffset + this_nob) {
457                         daddr += this_nob;
458                         doffset += this_nob;
459                 } else {
460                         kunmap(diov->kiov_page);
461                         daddr = NULL;
462                         diov++;
463                         ndiov--;
464                         doffset = 0;
465                 }
466
467                 if (siov->kiov_len > soffset + this_nob) {
468                         saddr += this_nob;
469                         soffset += this_nob;
470                 } else {
471                         kunmap(siov->kiov_page);
472                         saddr = NULL;
473                         siov++;
474                         nsiov--;
475                         soffset = 0;
476                 }
477         } while (nob > 0);
478
479         if (daddr != NULL)
480                 kunmap(diov->kiov_page);
481         if (saddr != NULL)
482                 kunmap(siov->kiov_page);
483 }
484 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
485
486 void
487 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
488                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
489                     unsigned int nob)
490 {
491         /* NB iov, kiov are READ-ONLY */
492         unsigned int    this_nob;
493         char           *addr = NULL;
494
495         if (nob == 0)
496                 return;
497
498         LASSERT (!in_interrupt ());
499
500         LASSERT (niov > 0);
501         while (iovoffset >= iov->iov_len) {
502                 iovoffset -= iov->iov_len;
503                 iov++;
504                 niov--;
505                 LASSERT(niov > 0);
506         }
507
508         LASSERT(nkiov > 0);
509         while (kiovoffset >= kiov->kiov_len) {
510                 kiovoffset -= kiov->kiov_len;
511                 kiov++;
512                 nkiov--;
513                 LASSERT(nkiov > 0);
514         }
515
516         do {
517                 LASSERT(niov > 0);
518                 LASSERT(nkiov > 0);
519                 this_nob = MIN(iov->iov_len - iovoffset,
520                                kiov->kiov_len - kiovoffset);
521                 this_nob = MIN(this_nob, nob);
522
523                 if (addr == NULL)
524                         addr = ((char *)kmap(kiov->kiov_page)) +
525                                 kiov->kiov_offset + kiovoffset;
526
527                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
528                 nob -= this_nob;
529
530                 if (iov->iov_len > iovoffset + this_nob) {
531                         iovoffset += this_nob;
532                 } else {
533                         iov++;
534                         niov--;
535                         iovoffset = 0;
536                 }
537
538                 if (kiov->kiov_len > kiovoffset + this_nob) {
539                         addr += this_nob;
540                         kiovoffset += this_nob;
541                 } else {
542                         kunmap(kiov->kiov_page);
543                         addr = NULL;
544                         kiov++;
545                         nkiov--;
546                         kiovoffset = 0;
547                 }
548
549         } while (nob > 0);
550
551         if (addr != NULL)
552                 kunmap(kiov->kiov_page);
553 }
554 EXPORT_SYMBOL(lnet_copy_kiov2iov);
555
556 void
557 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
558                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
559                    unsigned int nob)
560 {
561         /* NB kiov, iov are READ-ONLY */
562         unsigned int    this_nob;
563         char           *addr = NULL;
564
565         if (nob == 0)
566                 return;
567
568         LASSERT (!in_interrupt ());
569
570         LASSERT (nkiov > 0);
571         while (kiovoffset >= kiov->kiov_len) {
572                 kiovoffset -= kiov->kiov_len;
573                 kiov++;
574                 nkiov--;
575                 LASSERT(nkiov > 0);
576         }
577
578         LASSERT(niov > 0);
579         while (iovoffset >= iov->iov_len) {
580                 iovoffset -= iov->iov_len;
581                 iov++;
582                 niov--;
583                 LASSERT(niov > 0);
584         }
585
586         do {
587                 LASSERT(nkiov > 0);
588                 LASSERT(niov > 0);
589                 this_nob = MIN(kiov->kiov_len - kiovoffset,
590                                iov->iov_len - iovoffset);
591                 this_nob = MIN(this_nob, nob);
592
593                 if (addr == NULL)
594                         addr = ((char *)kmap(kiov->kiov_page)) +
595                                 kiov->kiov_offset + kiovoffset;
596
597                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
598                 nob -= this_nob;
599
600                 if (kiov->kiov_len > kiovoffset + this_nob) {
601                         addr += this_nob;
602                         kiovoffset += this_nob;
603                 } else {
604                         kunmap(kiov->kiov_page);
605                         addr = NULL;
606                         kiov++;
607                         nkiov--;
608                         kiovoffset = 0;
609                 }
610
611                 if (iov->iov_len > iovoffset + this_nob) {
612                         iovoffset += this_nob;
613                 } else {
614                         iov++;
615                         niov--;
616                         iovoffset = 0;
617                 }
618         } while (nob > 0);
619
620         if (addr != NULL)
621                 kunmap(kiov->kiov_page);
622 }
623 EXPORT_SYMBOL(lnet_copy_iov2kiov);
624
625 int
626 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
627                   int src_niov, lnet_kiov_t *src,
628                   unsigned int offset, unsigned int len)
629 {
630         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
631          * for exactly 'len' bytes, and return the number of entries.
632          * NB not destructive to 'src' */
633         unsigned int    frag_len;
634         unsigned int    niov;
635
636         if (len == 0)                           /* no data => */
637                 return (0);                     /* no frags */
638
639         LASSERT(src_niov > 0);
640         while (offset >= src->kiov_len) {      /* skip initial frags */
641                 offset -= src->kiov_len;
642                 src_niov--;
643                 src++;
644                 LASSERT(src_niov > 0);
645         }
646
647         niov = 1;
648         for (;;) {
649                 LASSERT(src_niov > 0);
650                 LASSERT((int)niov <= dst_niov);
651
652                 frag_len = src->kiov_len - offset;
653                 dst->kiov_page = src->kiov_page;
654                 dst->kiov_offset = src->kiov_offset + offset;
655
656                 if (len <= frag_len) {
657                         dst->kiov_len = len;
658                         LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
659                         return niov;
660                 }
661
662                 dst->kiov_len = frag_len;
663                 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
664
665                 len -= frag_len;
666                 dst++;
667                 src++;
668                 niov++;
669                 src_niov--;
670                 offset = 0;
671         }
672 }
673 EXPORT_SYMBOL(lnet_extract_kiov);
674
675 void
676 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
677              int delayed, unsigned int offset, unsigned int mlen,
678              unsigned int rlen)
679 {
680         unsigned int  niov = 0;
681         struct kvec *iov = NULL;
682         lnet_kiov_t  *kiov = NULL;
683         int           rc;
684
685         LASSERT (!in_interrupt ());
686         LASSERT (mlen == 0 || msg != NULL);
687
688         if (msg != NULL) {
689                 LASSERT(msg->msg_receiving);
690                 LASSERT(!msg->msg_sending);
691                 LASSERT(rlen == msg->msg_len);
692                 LASSERT(mlen <= msg->msg_len);
693                 LASSERT(msg->msg_offset == offset);
694                 LASSERT(msg->msg_wanted == mlen);
695
696                 msg->msg_receiving = 0;
697
698                 if (mlen != 0) {
699                         niov = msg->msg_niov;
700                         iov  = msg->msg_iov;
701                         kiov = msg->msg_kiov;
702
703                         LASSERT (niov > 0);
704                         LASSERT ((iov == NULL) != (kiov == NULL));
705                 }
706         }
707
708         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
709                                              niov, iov, kiov, offset, mlen,
710                                              rlen);
711         if (rc < 0)
712                 lnet_finalize(msg, rc);
713 }
714
715 static void
716 lnet_setpayloadbuffer(struct lnet_msg *msg)
717 {
718         struct lnet_libmd *md = msg->msg_md;
719
720         LASSERT(msg->msg_len > 0);
721         LASSERT(!msg->msg_routing);
722         LASSERT(md != NULL);
723         LASSERT(msg->msg_niov == 0);
724         LASSERT(msg->msg_iov == NULL);
725         LASSERT(msg->msg_kiov == NULL);
726
727         msg->msg_niov = md->md_niov;
728         if ((md->md_options & LNET_MD_KIOV) != 0)
729                 msg->msg_kiov = md->md_iov.kiov;
730         else
731                 msg->msg_iov = md->md_iov.iov;
732 }
733
734 void
735 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
736                unsigned int offset, unsigned int len)
737 {
738         msg->msg_type = type;
739         msg->msg_target = target;
740         msg->msg_len = len;
741         msg->msg_offset = offset;
742
743         if (len != 0)
744                 lnet_setpayloadbuffer(msg);
745
746         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
747         msg->msg_hdr.type           = cpu_to_le32(type);
748         /* dest_nid will be overwritten by lnet_select_pathway() */
749         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
750         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
751         /* src_nid will be set later */
752         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
753         msg->msg_hdr.payload_length = cpu_to_le32(len);
754 }
755
756 static void
757 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
758 {
759         void   *priv = msg->msg_private;
760         int rc;
761
762         LASSERT (!in_interrupt ());
763         LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
764                  (msg->msg_txcredit && msg->msg_peertxcredit));
765
766         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
767         if (rc < 0) {
768                 msg->msg_no_resend = true;
769                 lnet_finalize(msg, rc);
770         }
771 }
772
773 static int
774 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
775 {
776         int     rc;
777
778         LASSERT(!msg->msg_sending);
779         LASSERT(msg->msg_receiving);
780         LASSERT(!msg->msg_rx_ready_delay);
781         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
782
783         msg->msg_rx_ready_delay = 1;
784         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
785                                                   &msg->msg_private);
786         if (rc != 0) {
787                 CERROR("recv from %s / send to %s aborted: "
788                        "eager_recv failed %d\n",
789                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
790                        libcfs_id2str(msg->msg_target), rc);
791                 LASSERT(rc < 0); /* required by my callers */
792         }
793
794         return rc;
795 }
796
797 /*
798  * This function can be called from two paths:
799  *      1. when sending a message
800  *      2. when decommiting a message (lnet_msg_decommit_tx())
801  * In both these cases the peer_ni should have it's reference count
802  * acquired by the caller and therefore it is safe to drop the spin
803  * lock before calling lnd_query()
804  */
805 static void
806 lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
807 {
808         time64_t last_alive = 0;
809         int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
810
811         LASSERT(lnet_peer_aliveness_enabled(lp));
812         LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
813
814         lnet_net_unlock(cpt);
815         (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
816         lnet_net_lock(cpt);
817
818         lp->lpni_last_query = ktime_get_seconds();
819
820         if (last_alive != 0) /* NI has updated timestamp */
821                 lp->lpni_last_alive = last_alive;
822 }
823
824 /* NB: always called with lnet_net_lock held */
825 static inline int
826 lnet_peer_is_alive(struct lnet_peer_ni *lp, time64_t now)
827 {
828         int alive;
829         time64_t deadline;
830
831         LASSERT (lnet_peer_aliveness_enabled(lp));
832
833         /*
834          * Trust lnet_notify() if it has more recent aliveness news, but
835          * ignore the initial assumed death (see lnet_peers_start_down()).
836          */
837         spin_lock(&lp->lpni_lock);
838         if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
839             lp->lpni_timestamp >= lp->lpni_last_alive) {
840                 spin_unlock(&lp->lpni_lock);
841                 return 0;
842         }
843
844         deadline = lp->lpni_last_alive +
845                    lp->lpni_net->net_tunables.lct_peer_timeout;
846         alive = deadline > now;
847
848         /*
849          * Update obsolete lp_alive except for routers assumed to be dead
850          * initially, because router checker would update aliveness in this
851          * case, and moreover lpni_last_alive at peer creation is assumed.
852          */
853         if (alive && !lp->lpni_alive &&
854             !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
855                 spin_unlock(&lp->lpni_lock);
856                 lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
857         } else {
858                 spin_unlock(&lp->lpni_lock);
859         }
860
861         return alive;
862 }
863
864
865 /* NB: returns 1 when alive, 0 when dead, negative when error;
866  *     may drop the lnet_net_lock */
867 static int
868 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp,
869                        struct lnet_msg *msg)
870 {
871         time64_t now = ktime_get_seconds();
872
873         if (!lnet_peer_aliveness_enabled(lp))
874                 return -ENODEV;
875
876         if (lnet_peer_is_alive(lp, now))
877                 return 1;
878
879         /*
880          * If we're resending a message, let's attempt to send it even if
881          * the peer is down to fulfill our resend quota on the message
882          */
883         if (msg->msg_retry_count > 0)
884                 return 1;
885
886         /*
887          * Peer appears dead, but we should avoid frequent NI queries (at
888          * most once per lnet_queryinterval seconds).
889          */
890         if (lp->lpni_last_query != 0) {
891                 static const int lnet_queryinterval = 1;
892                 time64_t next_query;
893
894                 next_query = lp->lpni_last_query + lnet_queryinterval;
895
896                 if (now < next_query) {
897                         if (lp->lpni_alive)
898                                 CWARN("Unexpected aliveness of peer %s: "
899                                       "%lld < %lld (%d/%d)\n",
900                                       libcfs_nid2str(lp->lpni_nid),
901                                       now, next_query,
902                                       lnet_queryinterval,
903                                       lp->lpni_net->net_tunables.lct_peer_timeout);
904                         return 0;
905                 }
906         }
907
908         /* query NI for latest aliveness news */
909         lnet_ni_query_locked(ni, lp);
910
911         if (lnet_peer_is_alive(lp, now))
912                 return 1;
913
914         lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
915         return 0;
916 }
917
918 /**
919  * \param msg The message to be sent.
920  * \param do_send True if lnet_ni_send() should be called in this function.
921  *        lnet_send() is going to lnet_net_unlock immediately after this, so
922  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
923  *
924  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
925  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
926  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
927  * \retval -ECANCELED If the MD of the message has been unlinked.
928  */
929 static int
930 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
931 {
932         struct lnet_peer_ni     *lp = msg->msg_txpeer;
933         struct lnet_ni          *ni = msg->msg_txni;
934         int                     cpt = msg->msg_tx_cpt;
935         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
936
937         /* non-lnet_send() callers have checked before */
938         LASSERT(!do_send || msg->msg_tx_delayed);
939         LASSERT(!msg->msg_receiving);
940         LASSERT(msg->msg_tx_committed);
941         /* can't get here if we're sending to the loopback interface */
942         LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
943
944         /* NB 'lp' is always the next hop */
945         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
946             lnet_peer_alive_locked(ni, lp, msg) == 0) {
947                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
948                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
949                         msg->msg_len;
950                 lnet_net_unlock(cpt);
951                 if (msg->msg_txpeer)
952                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
953                                         msg->msg_type,
954                                         LNET_STATS_TYPE_DROP);
955                 if (msg->msg_txni)
956                         lnet_incr_stats(&msg->msg_txni->ni_stats,
957                                         msg->msg_type,
958                                         LNET_STATS_TYPE_DROP);
959
960                 CNETERR("Dropping message for %s: peer not alive\n",
961                         libcfs_id2str(msg->msg_target));
962                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
963                 if (do_send)
964                         lnet_finalize(msg, -EHOSTUNREACH);
965
966                 lnet_net_lock(cpt);
967                 return -EHOSTUNREACH;
968         }
969
970         if (msg->msg_md != NULL &&
971             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
972                 lnet_net_unlock(cpt);
973
974                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
975                         "called on the MD/ME.\n",
976                         libcfs_id2str(msg->msg_target));
977                 if (do_send) {
978                         msg->msg_no_resend = true;
979                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
980                                msg, libcfs_id2str(msg->msg_target));
981                         lnet_finalize(msg, -ECANCELED);
982                 }
983
984                 lnet_net_lock(cpt);
985                 return -ECANCELED;
986         }
987
988         if (!msg->msg_peertxcredit) {
989                 spin_lock(&lp->lpni_lock);
990                 LASSERT((lp->lpni_txcredits < 0) ==
991                         !list_empty(&lp->lpni_txq));
992
993                 msg->msg_peertxcredit = 1;
994                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
995                 lp->lpni_txcredits--;
996
997                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
998                         lp->lpni_mintxcredits = lp->lpni_txcredits;
999
1000                 if (lp->lpni_txcredits < 0) {
1001                         msg->msg_tx_delayed = 1;
1002                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
1003                         spin_unlock(&lp->lpni_lock);
1004                         return LNET_CREDIT_WAIT;
1005                 }
1006                 spin_unlock(&lp->lpni_lock);
1007         }
1008
1009         if (!msg->msg_txcredit) {
1010                 LASSERT((tq->tq_credits < 0) ==
1011                         !list_empty(&tq->tq_delayed));
1012
1013                 msg->msg_txcredit = 1;
1014                 tq->tq_credits--;
1015                 atomic_dec(&ni->ni_tx_credits);
1016
1017                 if (tq->tq_credits < tq->tq_credits_min)
1018                         tq->tq_credits_min = tq->tq_credits;
1019
1020                 if (tq->tq_credits < 0) {
1021                         msg->msg_tx_delayed = 1;
1022                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
1023                         return LNET_CREDIT_WAIT;
1024                 }
1025         }
1026
1027         /* unset the tx_delay flag as we're going to send it now */
1028         msg->msg_tx_delayed = 0;
1029
1030         if (do_send) {
1031                 lnet_net_unlock(cpt);
1032                 lnet_ni_send(ni, msg);
1033                 lnet_net_lock(cpt);
1034         }
1035         return LNET_CREDIT_OK;
1036 }
1037
1038
1039 static struct lnet_rtrbufpool *
1040 lnet_msg2bufpool(struct lnet_msg *msg)
1041 {
1042         struct lnet_rtrbufpool  *rbp;
1043         int                     cpt;
1044
1045         LASSERT(msg->msg_rx_committed);
1046
1047         cpt = msg->msg_rx_cpt;
1048         rbp = &the_lnet.ln_rtrpools[cpt][0];
1049
1050         LASSERT(msg->msg_len <= LNET_MTU);
1051         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
1052                 rbp++;
1053                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
1054         }
1055
1056         return rbp;
1057 }
1058
1059 static int
1060 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
1061 {
1062         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
1063          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
1064          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
1065          * received or OK to receive */
1066         struct lnet_peer_ni *lp = msg->msg_rxpeer;
1067         struct lnet_rtrbufpool *rbp;
1068         struct lnet_rtrbuf *rb;
1069
1070         LASSERT (msg->msg_iov == NULL);
1071         LASSERT (msg->msg_kiov == NULL);
1072         LASSERT (msg->msg_niov == 0);
1073         LASSERT (msg->msg_routing);
1074         LASSERT (msg->msg_receiving);
1075         LASSERT (!msg->msg_sending);
1076
1077         /* non-lnet_parse callers only receive delayed messages */
1078         LASSERT(!do_recv || msg->msg_rx_delayed);
1079
1080         if (!msg->msg_peerrtrcredit) {
1081                 spin_lock(&lp->lpni_lock);
1082                 LASSERT((lp->lpni_rtrcredits < 0) ==
1083                         !list_empty(&lp->lpni_rtrq));
1084
1085                 msg->msg_peerrtrcredit = 1;
1086                 lp->lpni_rtrcredits--;
1087                 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
1088                         lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
1089
1090                 if (lp->lpni_rtrcredits < 0) {
1091                         /* must have checked eager_recv before here */
1092                         LASSERT(msg->msg_rx_ready_delay);
1093                         msg->msg_rx_delayed = 1;
1094                         list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
1095                         spin_unlock(&lp->lpni_lock);
1096                         return LNET_CREDIT_WAIT;
1097                 }
1098                 spin_unlock(&lp->lpni_lock);
1099         }
1100
1101         rbp = lnet_msg2bufpool(msg);
1102
1103         if (!msg->msg_rtrcredit) {
1104                 msg->msg_rtrcredit = 1;
1105                 rbp->rbp_credits--;
1106                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1107                         rbp->rbp_mincredits = rbp->rbp_credits;
1108
1109                 if (rbp->rbp_credits < 0) {
1110                         /* must have checked eager_recv before here */
1111                         LASSERT(msg->msg_rx_ready_delay);
1112                         msg->msg_rx_delayed = 1;
1113                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1114                         return LNET_CREDIT_WAIT;
1115                 }
1116         }
1117
1118         LASSERT(!list_empty(&rbp->rbp_bufs));
1119         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1120         list_del(&rb->rb_list);
1121
1122         msg->msg_niov = rbp->rbp_npages;
1123         msg->msg_kiov = &rb->rb_kiov[0];
1124
1125         /* unset the msg-rx_delayed flag since we're receiving the message */
1126         msg->msg_rx_delayed = 0;
1127
1128         if (do_recv) {
1129                 int cpt = msg->msg_rx_cpt;
1130
1131                 lnet_net_unlock(cpt);
1132                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1133                              0, msg->msg_len, msg->msg_len);
1134                 lnet_net_lock(cpt);
1135         }
1136         return LNET_CREDIT_OK;
1137 }
1138
1139 void
1140 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1141 {
1142         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1143         struct lnet_ni          *txni = msg->msg_txni;
1144         struct lnet_msg         *msg2;
1145
1146         if (msg->msg_txcredit) {
1147                 struct lnet_ni       *ni = msg->msg_txni;
1148                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1149
1150                 /* give back NI txcredits */
1151                 msg->msg_txcredit = 0;
1152
1153                 LASSERT((tq->tq_credits < 0) ==
1154                         !list_empty(&tq->tq_delayed));
1155
1156                 tq->tq_credits++;
1157                 atomic_inc(&ni->ni_tx_credits);
1158                 if (tq->tq_credits <= 0) {
1159                         msg2 = list_entry(tq->tq_delayed.next,
1160                                           struct lnet_msg, msg_list);
1161                         list_del(&msg2->msg_list);
1162
1163                         LASSERT(msg2->msg_txni == ni);
1164                         LASSERT(msg2->msg_tx_delayed);
1165                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1166
1167                         (void) lnet_post_send_locked(msg2, 1);
1168                 }
1169         }
1170
1171         if (msg->msg_peertxcredit) {
1172                 /* give back peer txcredits */
1173                 msg->msg_peertxcredit = 0;
1174
1175                 spin_lock(&txpeer->lpni_lock);
1176                 LASSERT((txpeer->lpni_txcredits < 0) ==
1177                         !list_empty(&txpeer->lpni_txq));
1178
1179                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1180                 LASSERT(txpeer->lpni_txqnob >= 0);
1181
1182                 txpeer->lpni_txcredits++;
1183                 if (txpeer->lpni_txcredits <= 0) {
1184                         int msg2_cpt;
1185
1186                         msg2 = list_entry(txpeer->lpni_txq.next,
1187                                               struct lnet_msg, msg_list);
1188                         list_del(&msg2->msg_list);
1189                         spin_unlock(&txpeer->lpni_lock);
1190
1191                         LASSERT(msg2->msg_txpeer == txpeer);
1192                         LASSERT(msg2->msg_tx_delayed);
1193
1194                         msg2_cpt = msg2->msg_tx_cpt;
1195
1196                         /*
1197                          * The msg_cpt can be different from the msg2_cpt
1198                          * so we need to make sure we lock the correct cpt
1199                          * for msg2.
1200                          * Once we call lnet_post_send_locked() it is no
1201                          * longer safe to access msg2, since it could've
1202                          * been freed by lnet_finalize(), but we still
1203                          * need to relock the correct cpt, so we cache the
1204                          * msg2_cpt for the purpose of the check that
1205                          * follows the call to lnet_pose_send_locked().
1206                          */
1207                         if (msg2_cpt != msg->msg_tx_cpt) {
1208                                 lnet_net_unlock(msg->msg_tx_cpt);
1209                                 lnet_net_lock(msg2_cpt);
1210                         }
1211                         (void) lnet_post_send_locked(msg2, 1);
1212                         if (msg2_cpt != msg->msg_tx_cpt) {
1213                                 lnet_net_unlock(msg2_cpt);
1214                                 lnet_net_lock(msg->msg_tx_cpt);
1215                         }
1216                 } else {
1217                         spin_unlock(&txpeer->lpni_lock);
1218                 }
1219         }
1220
1221         if (txni != NULL) {
1222                 msg->msg_txni = NULL;
1223                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1224         }
1225
1226         if (txpeer != NULL) {
1227                 msg->msg_txpeer = NULL;
1228                 lnet_peer_ni_decref_locked(txpeer);
1229         }
1230 }
1231
1232 void
1233 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1234 {
1235         struct lnet_msg *msg;
1236
1237         if (list_empty(&rbp->rbp_msgs))
1238                 return;
1239         msg = list_entry(rbp->rbp_msgs.next,
1240                          struct lnet_msg, msg_list);
1241         list_del(&msg->msg_list);
1242
1243         (void)lnet_post_routed_recv_locked(msg, 1);
1244 }
1245
1246 void
1247 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1248 {
1249         struct lnet_msg *msg;
1250         struct lnet_msg *tmp;
1251
1252         lnet_net_unlock(cpt);
1253
1254         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1255                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1256                              0, 0, 0, msg->msg_hdr.payload_length);
1257                 list_del_init(&msg->msg_list);
1258                 msg->msg_no_resend = true;
1259                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1260                 lnet_finalize(msg, -ECANCELED);
1261         }
1262
1263         lnet_net_lock(cpt);
1264 }
1265
1266 void
1267 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1268 {
1269         struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1270         struct lnet_ni *rxni = msg->msg_rxni;
1271         struct lnet_msg *msg2;
1272
1273         if (msg->msg_rtrcredit) {
1274                 /* give back global router credits */
1275                 struct lnet_rtrbuf *rb;
1276                 struct lnet_rtrbufpool *rbp;
1277
1278                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1279                  * there until it gets one allocated, or aborts the wait
1280                  * itself */
1281                 LASSERT(msg->msg_kiov != NULL);
1282
1283                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1284                 rbp = rb->rb_pool;
1285
1286                 msg->msg_kiov = NULL;
1287                 msg->msg_rtrcredit = 0;
1288
1289                 LASSERT(rbp == lnet_msg2bufpool(msg));
1290
1291                 LASSERT((rbp->rbp_credits > 0) ==
1292                         !list_empty(&rbp->rbp_bufs));
1293
1294                 /* If routing is now turned off, we just drop this buffer and
1295                  * don't bother trying to return credits.  */
1296                 if (!the_lnet.ln_routing) {
1297                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1298                         goto routing_off;
1299                 }
1300
1301                 /* It is possible that a user has lowered the desired number of
1302                  * buffers in this pool.  Make sure we never put back
1303                  * more buffers than the stated number. */
1304                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1305                         /* Discard this buffer so we don't have too
1306                          * many. */
1307                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1308                         rbp->rbp_nbuffers--;
1309                 } else {
1310                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1311                         rbp->rbp_credits++;
1312                         if (rbp->rbp_credits <= 0)
1313                                 lnet_schedule_blocked_locked(rbp);
1314                 }
1315         }
1316
1317 routing_off:
1318         if (msg->msg_peerrtrcredit) {
1319                 /* give back peer router credits */
1320                 msg->msg_peerrtrcredit = 0;
1321
1322                 spin_lock(&rxpeer->lpni_lock);
1323                 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1324                         !list_empty(&rxpeer->lpni_rtrq));
1325
1326                 rxpeer->lpni_rtrcredits++;
1327
1328                 /* drop all messages which are queued to be routed on that
1329                  * peer. */
1330                 if (!the_lnet.ln_routing) {
1331                         struct list_head drop;
1332                         INIT_LIST_HEAD(&drop);
1333                         list_splice_init(&rxpeer->lpni_rtrq, &drop);
1334                         spin_unlock(&rxpeer->lpni_lock);
1335                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1336                 } else if (rxpeer->lpni_rtrcredits <= 0) {
1337                         msg2 = list_entry(rxpeer->lpni_rtrq.next,
1338                                           struct lnet_msg, msg_list);
1339                         list_del(&msg2->msg_list);
1340                         spin_unlock(&rxpeer->lpni_lock);
1341                         (void) lnet_post_routed_recv_locked(msg2, 1);
1342                 } else {
1343                         spin_unlock(&rxpeer->lpni_lock);
1344                 }
1345         }
1346         if (rxni != NULL) {
1347                 msg->msg_rxni = NULL;
1348                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1349         }
1350         if (rxpeer != NULL) {
1351                 msg->msg_rxpeer = NULL;
1352                 lnet_peer_ni_decref_locked(rxpeer);
1353         }
1354 }
1355
1356 static int
1357 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1358 {
1359         if (p1->lpni_txqnob < p2->lpni_txqnob)
1360                 return 1;
1361
1362         if (p1->lpni_txqnob > p2->lpni_txqnob)
1363                 return -1;
1364
1365         if (p1->lpni_txcredits > p2->lpni_txcredits)
1366                 return 1;
1367
1368         if (p1->lpni_txcredits < p2->lpni_txcredits)
1369                 return -1;
1370
1371         return 0;
1372 }
1373
1374 static int
1375 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1376 {
1377         struct lnet_peer_ni *p1 = r1->lr_gateway;
1378         struct lnet_peer_ni *p2 = r2->lr_gateway;
1379         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1380         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1381         int rc;
1382
1383         if (r1->lr_priority < r2->lr_priority)
1384                 return 1;
1385
1386         if (r1->lr_priority > r2->lr_priority)
1387                 return -1;
1388
1389         if (r1_hops < r2_hops)
1390                 return 1;
1391
1392         if (r1_hops > r2_hops)
1393                 return -1;
1394
1395         rc = lnet_compare_peers(p1, p2);
1396         if (rc)
1397                 return rc;
1398
1399         if (r1->lr_seq - r2->lr_seq <= 0)
1400                 return 1;
1401
1402         return -1;
1403 }
1404
1405 static struct lnet_peer_ni *
1406 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1407                        lnet_nid_t rtr_nid)
1408 {
1409         struct lnet_remotenet   *rnet;
1410         struct lnet_route               *route;
1411         struct lnet_route               *best_route;
1412         struct lnet_route               *last_route;
1413         struct lnet_peer_ni     *lpni_best;
1414         struct lnet_peer_ni     *lp;
1415         int                     rc;
1416
1417         /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1418          * rtr_nid nid, otherwise find the best gateway I can use */
1419
1420         rnet = lnet_find_rnet_locked(remote_net);
1421         if (rnet == NULL)
1422                 return NULL;
1423
1424         lpni_best = NULL;
1425         best_route = last_route = NULL;
1426         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1427                 lp = route->lr_gateway;
1428
1429                 if (!lnet_is_route_alive(route))
1430                         continue;
1431
1432                 if (net != NULL && lp->lpni_net != net)
1433                         continue;
1434
1435                 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1436                         return lp;
1437
1438                 if (lpni_best == NULL) {
1439                         best_route = last_route = route;
1440                         lpni_best = lp;
1441                         continue;
1442                 }
1443
1444                 /* no protection on below fields, but it's harmless */
1445                 if (last_route->lr_seq - route->lr_seq < 0)
1446                         last_route = route;
1447
1448                 rc = lnet_compare_routes(route, best_route);
1449                 if (rc < 0)
1450                         continue;
1451
1452                 best_route = route;
1453                 lpni_best = lp;
1454         }
1455
1456         /* set sequence number on the best router to the latest sequence + 1
1457          * so we can round-robin all routers, it's race and inaccurate but
1458          * harmless and functional  */
1459         if (best_route != NULL)
1460                 best_route->lr_seq = last_route->lr_seq + 1;
1461         return lpni_best;
1462 }
1463
1464 static struct lnet_ni *
1465 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1466                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1467                  int md_cpt)
1468 {
1469         struct lnet_ni *ni = NULL;
1470         unsigned int shortest_distance;
1471         int best_credits;
1472         int best_healthv;
1473
1474         /*
1475          * If there is no peer_ni that we can send to on this network,
1476          * then there is no point in looking for a new best_ni here.
1477         */
1478         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1479                 return best_ni;
1480
1481         if (best_ni == NULL) {
1482                 shortest_distance = UINT_MAX;
1483                 best_credits = INT_MIN;
1484                 best_healthv = 0;
1485         } else {
1486                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1487                                                      best_ni->ni_dev_cpt);
1488                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1489                 best_healthv = atomic_read(&best_ni->ni_healthv);
1490         }
1491
1492         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1493                 unsigned int distance;
1494                 int ni_credits;
1495                 int ni_healthv;
1496                 int ni_fatal;
1497
1498                 ni_credits = atomic_read(&ni->ni_tx_credits);
1499                 ni_healthv = atomic_read(&ni->ni_healthv);
1500                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1501
1502                 /*
1503                  * calculate the distance from the CPT on which
1504                  * the message memory is allocated to the CPT of
1505                  * the NI's physical device
1506                  */
1507                 distance = cfs_cpt_distance(lnet_cpt_table(),
1508                                             md_cpt,
1509                                             ni->ni_dev_cpt);
1510
1511                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1512                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1513                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1514                         : "not seleced", best_credits, shortest_distance,
1515                         (best_ni) ? best_ni->ni_seq : 0);
1516
1517                 /*
1518                  * All distances smaller than the NUMA range
1519                  * are treated equally.
1520                  */
1521                 if (distance < lnet_numa_range)
1522                         distance = lnet_numa_range;
1523
1524                 /*
1525                  * Select on health, shorter distance, available
1526                  * credits, then round-robin.
1527                  */
1528                 if (ni_fatal) {
1529                         continue;
1530                 } else if (ni_healthv < best_healthv) {
1531                         continue;
1532                 } else if (ni_healthv > best_healthv) {
1533                         best_healthv = ni_healthv;
1534                         /*
1535                          * If we're going to prefer this ni because it's
1536                          * the healthiest, then we should set the
1537                          * shortest_distance in the algorithm in case
1538                          * there are multiple NIs with the same health but
1539                          * different distances.
1540                          */
1541                         if (distance < shortest_distance)
1542                                 shortest_distance = distance;
1543                 } else if (distance > shortest_distance) {
1544                         continue;
1545                 } else if (distance < shortest_distance) {
1546                         shortest_distance = distance;
1547                 } else if (ni_credits < best_credits) {
1548                         continue;
1549                 } else if (ni_credits == best_credits) {
1550                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1551                                 continue;
1552                 }
1553                 best_ni = ni;
1554                 best_credits = ni_credits;
1555         }
1556
1557         CDEBUG(D_NET, "selected best_ni %s\n",
1558                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1559
1560         return best_ni;
1561 }
1562
1563 /*
1564  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1565  * because such traffic is required to perform discovery. We therefore
1566  * exclude all GET and PUT on that portal. We also exclude all ACK and
1567  * REPLY traffic, but that is because the portal is not tracked in the
1568  * message structure for these message types. We could restrict this
1569  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1570  */
1571 static bool
1572 lnet_msg_discovery(struct lnet_msg *msg)
1573 {
1574         if (msg->msg_type == LNET_MSG_PUT) {
1575                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1576                         return true;
1577         } else if (msg->msg_type == LNET_MSG_GET) {
1578                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1579                         return true;
1580         }
1581         return false;
1582 }
1583
1584 #define SRC_SPEC        0x0001
1585 #define SRC_ANY         0x0002
1586 #define LOCAL_DST       0x0004
1587 #define REMOTE_DST      0x0008
1588 #define MR_DST          0x0010
1589 #define NMR_DST         0x0020
1590 #define SND_RESP        0x0040
1591
1592 /* The following to defines are used for return codes */
1593 #define REPEAT_SEND     0x1000
1594 #define PASS_THROUGH    0x2000
1595
1596 /* The different cases lnet_select pathway needs to handle */
1597 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1598 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1599 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1600 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1601 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1602 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1603 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1604 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1605
1606 static int
1607 lnet_handle_lo_send(struct lnet_send_data *sd)
1608 {
1609         struct lnet_msg *msg = sd->sd_msg;
1610         int cpt = sd->sd_cpt;
1611
1612         /* No send credit hassles with LOLND */
1613         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1614         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1615         if (!msg->msg_routing)
1616                 msg->msg_hdr.src_nid =
1617                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1618         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1619         lnet_msg_commit(msg, cpt);
1620         msg->msg_txni = the_lnet.ln_loni;
1621
1622         return LNET_CREDIT_OK;
1623 }
1624
1625 static int
1626 lnet_handle_send(struct lnet_send_data *sd)
1627 {
1628         struct lnet_ni *best_ni = sd->sd_best_ni;
1629         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1630         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1631         struct lnet_msg *msg = sd->sd_msg;
1632         int cpt2;
1633         __u32 send_case = sd->sd_send_case;
1634         int rc;
1635         __u32 routing = send_case & REMOTE_DST;
1636          struct lnet_rsp_tracker *rspt;
1637
1638         /*
1639          * Increment sequence number of the selected peer so that we
1640          * pick the next one in Round Robin.
1641          */
1642         best_lpni->lpni_seq++;
1643
1644         /*
1645          * grab a reference on the peer_ni so it sticks around even if
1646          * we need to drop and relock the lnet_net_lock below.
1647          */
1648         lnet_peer_ni_addref_locked(best_lpni);
1649
1650         /*
1651          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1652          * message. This ensures that we get a CPT that is correct for
1653          * the NI when the NI has been restricted to a subset of all CPTs.
1654          * If the selected CPT differs from the one currently locked, we
1655          * must unlock and relock the lnet_net_lock(), and then check whether
1656          * the configuration has changed. We don't have a hold on the best_ni
1657          * yet, and it may have vanished.
1658          */
1659         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1660         if (sd->sd_cpt != cpt2) {
1661                 __u32 seq = lnet_get_dlc_seq_locked();
1662                 lnet_net_unlock(sd->sd_cpt);
1663                 sd->sd_cpt = cpt2;
1664                 lnet_net_lock(sd->sd_cpt);
1665                 if (seq != lnet_get_dlc_seq_locked()) {
1666                         lnet_peer_ni_decref_locked(best_lpni);
1667                         return REPEAT_SEND;
1668                 }
1669         }
1670
1671         /*
1672          * store the best_lpni in the message right away to avoid having
1673          * to do the same operation under different conditions
1674          */
1675         msg->msg_txpeer = best_lpni;
1676         msg->msg_txni = best_ni;
1677
1678         /*
1679          * grab a reference for the best_ni since now it's in use in this
1680          * send. The reference will be dropped in lnet_finalize()
1681          */
1682         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1683
1684         /*
1685          * Always set the target.nid to the best peer picked. Either the
1686          * NID will be one of the peer NIDs selected, or the same NID as
1687          * what was originally set in the target or it will be the NID of
1688          * a router if this message should be routed
1689          */
1690         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1691
1692         /*
1693          * lnet_msg_commit assigns the correct cpt to the message, which
1694          * is used to decrement the correct refcount on the ni when it's
1695          * time to return the credits
1696          */
1697         lnet_msg_commit(msg, sd->sd_cpt);
1698
1699         /*
1700          * If we are routing the message then we keep the src_nid that was
1701          * set by the originator. If we are not routing then we are the
1702          * originator and set it here.
1703          */
1704         if (!msg->msg_routing)
1705                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1706
1707         if (routing) {
1708                 msg->msg_target_is_router = 1;
1709                 msg->msg_target.pid = LNET_PID_LUSTRE;
1710                 /*
1711                  * since we're routing we want to ensure that the
1712                  * msg_hdr.dest_nid is set to the final destination. When
1713                  * the router receives this message it knows how to route
1714                  * it.
1715                  *
1716                  * final_dst_lpni is set at the beginning of the
1717                  * lnet_select_pathway() function and is never changed.
1718                  * It's safe to use it here.
1719                  */
1720                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1721         } else {
1722                 /*
1723                  * if we're not routing set the dest_nid to the best peer
1724                  * ni NID that we picked earlier in the algorithm.
1725                  */
1726                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1727         }
1728
1729         /*
1730          * if we have response tracker block update it with the next hop
1731          * nid
1732          */
1733         if (msg->msg_md) {
1734                 rspt = msg->msg_md->md_rspt_ptr;
1735                 if (rspt) {
1736                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1737                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1738                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1739                 }
1740         }
1741
1742         rc = lnet_post_send_locked(msg, 0);
1743
1744         if (!rc)
1745                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1746                        libcfs_nid2str(msg->msg_hdr.src_nid),
1747                        libcfs_nid2str(msg->msg_txni->ni_nid),
1748                        libcfs_nid2str(sd->sd_src_nid),
1749                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1750                        libcfs_nid2str(sd->sd_dst_nid),
1751                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1752                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1753
1754         return rc;
1755 }
1756
1757 static struct lnet_peer_ni *
1758 lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer,
1759                     struct lnet_peer_net *peer_net)
1760 {
1761         /*
1762          * Look at the peer NIs for the destination peer that connect
1763          * to the chosen net. If a peer_ni is preferred when using the
1764          * best_ni to communicate, we use that one. If there is no
1765          * preferred peer_ni, or there are multiple preferred peer_ni,
1766          * the available transmit credits are used. If the transmit
1767          * credits are equal, we round-robin over the peer_ni.
1768          */
1769         struct lnet_peer_ni *lpni = NULL;
1770         struct lnet_peer_ni *best_lpni = NULL;
1771         struct lnet_ni *best_ni = sd->sd_best_ni;
1772         lnet_nid_t dst_nid = sd->sd_dst_nid;
1773         int best_lpni_credits = INT_MIN;
1774         bool preferred = false;
1775         bool ni_is_pref;
1776         int best_lpni_healthv = 0;
1777         int lpni_healthv;
1778
1779         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1780                 /*
1781                  * if the best_ni we've chosen aleady has this lpni
1782                  * preferred, then let's use it
1783                  */
1784                 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1785                                                           best_ni->ni_nid);
1786
1787                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1788
1789                 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1790                        libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1791
1792                 if (best_lpni)
1793                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1794                                 libcfs_nid2str(lpni->lpni_nid),
1795                                 lpni->lpni_txcredits, best_lpni_credits,
1796                                 lpni->lpni_seq, best_lpni->lpni_seq);
1797
1798                 /* pick the healthiest peer ni */
1799                 if (lpni_healthv < best_lpni_healthv) {
1800                         continue;
1801                 } else if (lpni_healthv > best_lpni_healthv) {
1802                         best_lpni_healthv = lpni_healthv;
1803                 /* if this is a preferred peer use it */
1804                 } else if (!preferred && ni_is_pref) {
1805                         preferred = true;
1806                 } else if (preferred && !ni_is_pref) {
1807                         /*
1808                          * this is not the preferred peer so let's ignore
1809                          * it.
1810                          */
1811                         continue;
1812                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1813                         /*
1814                          * We already have a peer that has more credits
1815                          * available than this one. No need to consider
1816                          * this peer further.
1817                          */
1818                         continue;
1819                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1820                         /*
1821                          * The best peer found so far and the current peer
1822                          * have the same number of available credits let's
1823                          * make sure to select between them using Round
1824                          * Robin
1825                          */
1826                         if (best_lpni) {
1827                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1828                                         continue;
1829                         }
1830                 }
1831
1832                 best_lpni = lpni;
1833                 best_lpni_credits = lpni->lpni_txcredits;
1834         }
1835
1836         /* if we still can't find a peer ni then we can't reach it */
1837         if (!best_lpni) {
1838                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1839                         LNET_NIDNET(dst_nid);
1840                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1841                                 libcfs_net2str(net_id));
1842                 return NULL;
1843         }
1844
1845         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1846                libcfs_nid2str(best_lpni->lpni_nid));
1847
1848         return best_lpni;
1849 }
1850
1851 /*
1852  * Prerequisite: the best_ni should already be set in the sd
1853  */
1854 static inline struct lnet_peer_ni *
1855 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1856                            __u32 net_id)
1857 {
1858         struct lnet_peer_net *peer_net;
1859
1860         /*
1861          * The gateway is Multi-Rail capable so now we must select the
1862          * proper peer_ni
1863          */
1864         peer_net = lnet_peer_get_net_locked(peer, net_id);
1865
1866         if (!peer_net) {
1867                 CERROR("gateway peer %s has no NI on net %s\n",
1868                        libcfs_nid2str(peer->lp_primary_nid),
1869                        libcfs_net2str(net_id));
1870                 return NULL;
1871         }
1872
1873         return lnet_select_peer_ni(sd, peer, peer_net);
1874 }
1875
1876 static inline void
1877 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1878 {
1879         if (sd->sd_send_case & NMR_DST &&
1880             sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1881             sd->sd_msg->msg_type != LNET_MSG_ACK &&
1882             sd->sd_best_lpni->lpni_pref_nnids == 0) {
1883                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1884                        libcfs_nid2str(sd->sd_best_ni->ni_nid),
1885                        libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1886                 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1887                                                  sd->sd_best_ni->ni_nid);
1888         }
1889 }
1890
1891 /*
1892  * Source Specified
1893  * Local Destination
1894  * non-mr peer
1895  *
1896  * use the source and destination NIDs as the pathway
1897  */
1898 static int
1899 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1900 {
1901         /* the destination lpni is set before we get here. */
1902
1903         /* find local NI */
1904         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1905         if (!sd->sd_best_ni) {
1906                 CERROR("Can't send to %s: src %s is not a "
1907                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1908                                 libcfs_nid2str(sd->sd_src_nid));
1909                 return -EINVAL;
1910         }
1911
1912         /*
1913          * the preferred NID will only be set for NMR peers
1914          */
1915         lnet_set_non_mr_pref_nid(sd);
1916
1917         return lnet_handle_send(sd);
1918 }
1919
1920 /*
1921  * Source Specified
1922  * Local Destination
1923  * MR Peer
1924  *
1925  * Run the selection algorithm on the peer NIs unless we're sending
1926  * a response, in this case just send to the destination
1927  */
1928 static int
1929 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1930 {
1931         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1932         if (!sd->sd_best_ni) {
1933                 CERROR("Can't send to %s: src %s is not a "
1934                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1935                                 libcfs_nid2str(sd->sd_src_nid));
1936                 return -EINVAL;
1937         }
1938
1939         /*
1940          * only run the selection algorithm to pick the peer_ni if we're
1941          * sending a GET or a PUT. Responses are sent to the same
1942          * destination NID provided.
1943          */
1944         if (!(sd->sd_send_case & SND_RESP)) {
1945                 sd->sd_best_lpni =
1946                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1947                                              sd->sd_best_ni->ni_net->net_id);
1948         }
1949
1950         if (sd->sd_best_lpni &&
1951             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1952                 return lnet_handle_lo_send(sd);
1953         else if (sd->sd_best_lpni)
1954                 return lnet_handle_send(sd);
1955
1956         CERROR("can't send to %s. no NI on %s\n",
1957                libcfs_nid2str(sd->sd_dst_nid),
1958                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1959
1960         return -EHOSTUNREACH;
1961 }
1962
1963 struct lnet_ni *
1964 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1965                               struct lnet_peer *peer,
1966                               struct lnet_peer_net *peer_net,
1967                               int cpt,
1968                               bool incr_seq)
1969 {
1970         struct lnet_net *local_net;
1971         struct lnet_ni *best_ni;
1972
1973         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1974         if (!local_net)
1975                 return NULL;
1976
1977         /*
1978          * Iterate through the NIs in this local Net and select
1979          * the NI to send from. The selection is determined by
1980          * these 3 criterion in the following priority:
1981          *      1. NUMA
1982          *      2. NI available credits
1983          *      3. Round Robin
1984          */
1985         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1986                                    peer, peer_net, cpt);
1987
1988         if (incr_seq && best_ni)
1989                 best_ni->ni_seq++;
1990
1991         return best_ni;
1992 }
1993
1994 static int
1995 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1996                              lnet_nid_t dst_nid,
1997                              struct lnet_peer_ni **gw_lpni,
1998                              struct lnet_peer **gw_peer)
1999 {
2000         struct lnet_peer_ni *gw;
2001         lnet_nid_t src_nid = sd->sd_src_nid;
2002
2003         gw = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
2004                                     sd->sd_rtr_nid);
2005         if (!gw) {
2006                 CERROR("no route to %s from %s\n",
2007                        libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
2008                 return -EHOSTUNREACH;
2009         }
2010
2011         /* get the peer of the gw_ni */
2012         LASSERT(gw->lpni_peer_net);
2013         LASSERT(gw->lpni_peer_net->lpn_peer);
2014
2015         *gw_peer = gw->lpni_peer_net->lpn_peer;
2016
2017         if (!sd->sd_best_ni)
2018                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, *gw_peer,
2019                                         gw->lpni_peer_net,
2020                                         sd->sd_md_cpt,
2021                                         true);
2022
2023         if (!sd->sd_best_ni) {
2024                 CERROR("Internal Error. Expected local ni on %s "
2025                        "but non found :%s\n",
2026                        libcfs_net2str(gw->lpni_peer_net->lpn_net_id),
2027                        libcfs_nid2str(sd->sd_src_nid));
2028                 return -EFAULT;
2029         }
2030
2031         /*
2032          * if gw is MR let's find its best peer_ni
2033          */
2034         if (lnet_peer_is_multi_rail(*gw_peer)) {
2035                 gw = lnet_find_best_lpni_on_net(sd, *gw_peer,
2036                                                 sd->sd_best_ni->ni_net->net_id);
2037                 /*
2038                  * We've already verified that the gw has an NI on that
2039                  * desired net, but we're not finding it. Something is
2040                  * wrong.
2041                  */
2042                 if (!gw) {
2043                         CERROR("Internal Error. Route expected to %s from %s\n",
2044                                 libcfs_nid2str(dst_nid),
2045                                 libcfs_nid2str(src_nid));
2046                         return -EFAULT;
2047                 }
2048         }
2049
2050         *gw_lpni = gw;
2051
2052         return 0;
2053 }
2054
2055 /*
2056  * Handle two cases:
2057  *
2058  * Case 1:
2059  *  Source specified
2060  *  Remote destination
2061  *  Non-MR destination
2062  *
2063  * Case 2:
2064  *  Source specified
2065  *  Remote destination
2066  *  MR destination
2067  *
2068  * The handling of these two cases is similar. Even though the destination
2069  * can be MR or non-MR, we'll deal directly with the router.
2070  */
2071 static int
2072 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2073 {
2074         int rc;
2075         struct lnet_peer_ni *gw_lpni = NULL;
2076         struct lnet_peer *gw_peer = NULL;
2077
2078         /* find local NI */
2079         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2080         if (!sd->sd_best_ni) {
2081                 CERROR("Can't send to %s: src %s is not a "
2082                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2083                                 libcfs_nid2str(sd->sd_src_nid));
2084                 return -EINVAL;
2085         }
2086
2087         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2088                                      &gw_peer);
2089         if (rc < 0)
2090                 return rc;
2091
2092         if (sd->sd_send_case & NMR_DST)
2093                 /*
2094                 * since the final destination is non-MR let's set its preferred
2095                 * NID before we send
2096                 */
2097                 lnet_set_non_mr_pref_nid(sd);
2098
2099         /*
2100          * We're going to send to the gw found so let's set its
2101          * info
2102          */
2103         sd->sd_peer = gw_peer;
2104         sd->sd_best_lpni = gw_lpni;
2105
2106         return lnet_handle_send(sd);
2107 }
2108
2109 struct lnet_ni *
2110 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
2111 {
2112         struct lnet_peer_net *peer_net = NULL;
2113         struct lnet_ni *best_ni = NULL;
2114
2115         /*
2116          * The peer can have multiple interfaces, some of them can be on
2117          * the local network and others on a routed network. We should
2118          * prefer the local network. However if the local network is not
2119          * available then we need to try the routed network
2120          */
2121
2122         /* go through all the peer nets and find the best_ni */
2123         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2124                 /*
2125                  * The peer's list of nets can contain non-local nets. We
2126                  * want to only examine the local ones.
2127                  */
2128                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2129                         continue;
2130                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2131                                                    peer_net, md_cpt, false);
2132         }
2133
2134         if (best_ni)
2135                 /* increment sequence number so we can round robin */
2136                 best_ni->ni_seq++;
2137
2138         return best_ni;
2139 }
2140
2141 static struct lnet_ni *
2142 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2143 {
2144         struct lnet_ni *best_ni = NULL;
2145         struct lnet_peer_net *peer_net;
2146         struct lnet_peer *peer = sd->sd_peer;
2147         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2148         struct lnet_peer_ni *lpni;
2149         int cpt = sd->sd_cpt;
2150
2151         /*
2152          * We must use a consistent source address when sending to a
2153          * non-MR peer. However, a non-MR peer can have multiple NIDs
2154          * on multiple networks, and we may even need to talk to this
2155          * peer on multiple networks -- certain types of
2156          * load-balancing configuration do this.
2157          *
2158          * So we need to pick the NI the peer prefers for this
2159          * particular network.
2160          */
2161
2162         /* Get the target peer_ni */
2163         peer_net = lnet_peer_get_net_locked(peer,
2164                         LNET_NIDNET(best_lpni->lpni_nid));
2165         LASSERT(peer_net != NULL);
2166         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2167                                 lpni_peer_nis) {
2168                 if (lpni->lpni_pref_nnids == 0)
2169                         continue;
2170                 LASSERT(lpni->lpni_pref_nnids == 1);
2171                 best_ni = lnet_nid2ni_locked(
2172                                 lpni->lpni_pref.nid, cpt);
2173                 break;
2174         }
2175
2176         return best_ni;
2177 }
2178
2179 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2180 static int
2181 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2182 {
2183         struct lnet_ni *best_ni = NULL;
2184         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2185
2186         /*
2187          * We must use a consistent source address when sending to a
2188          * non-MR peer. However, a non-MR peer can have multiple NIDs
2189          * on multiple networks, and we may even need to talk to this
2190          * peer on multiple networks -- certain types of
2191          * load-balancing configuration do this.
2192          *
2193          * So we need to pick the NI the peer prefers for this
2194          * particular network.
2195          */
2196
2197         best_ni = lnet_find_existing_preferred_best_ni(sd);
2198
2199         /* if best_ni is still not set just pick one */
2200         if (!best_ni) {
2201                 best_ni =
2202                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2203                                                 sd->sd_best_lpni->lpni_peer_net,
2204                                                 sd->sd_md_cpt, true);
2205                 /* If there is no best_ni we don't have a route */
2206                 if (!best_ni) {
2207                         CERROR("no path to %s from net %s\n",
2208                                 libcfs_nid2str(best_lpni->lpni_nid),
2209                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2210                         return -EHOSTUNREACH;
2211                 }
2212         }
2213
2214         sd->sd_best_ni = best_ni;
2215
2216         /* Set preferred NI if necessary. */
2217         lnet_set_non_mr_pref_nid(sd);
2218
2219         return 0;
2220 }
2221
2222
2223 /*
2224  * Source not specified
2225  * Local destination
2226  * Non-MR Peer
2227  *
2228  * always use the same source NID for NMR peers
2229  * If we've talked to that peer before then we already have a preferred
2230  * source NI associated with it. Otherwise, we select a preferred local NI
2231  * and store it in the peer
2232  */
2233 static int
2234 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2235 {
2236         int rc;
2237
2238         /* sd->sd_best_lpni is already set to the final destination */
2239
2240         /*
2241          * At this point we should've created the peer ni and peer. If we
2242          * can't find it, then something went wrong. Instead of assert
2243          * output a relevant message and fail the send
2244          */
2245         if (!sd->sd_best_lpni) {
2246                 CERROR("Internal fault. Unable to send msg %s to %s. "
2247                        "NID not known\n",
2248                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2249                        libcfs_nid2str(sd->sd_dst_nid));
2250                 return -EFAULT;
2251         }
2252
2253         rc = lnet_select_preferred_best_ni(sd);
2254         if (!rc)
2255                 rc = lnet_handle_send(sd);
2256
2257         return rc;
2258 }
2259
2260 static int
2261 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2262 {
2263         /*
2264          * NOTE we've already handled the remote peer case. So we only
2265          * need to worry about the local case here.
2266          *
2267          * if we're sending a response, ACK or reply, we need to send it
2268          * to the destination NID given to us. At this point we already
2269          * have the peer_ni we're suppose to send to, so just find the
2270          * best_ni on the peer net and use that. Since we're sending to an
2271          * MR peer then we can just run the selection algorithm on our
2272          * local NIs and pick the best one.
2273          */
2274         if (sd->sd_send_case & SND_RESP) {
2275                 sd->sd_best_ni =
2276                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2277                                                 sd->sd_best_lpni->lpni_peer_net,
2278                                                 sd->sd_md_cpt, true);
2279
2280                 if (!sd->sd_best_ni) {
2281                         /*
2282                          * We're not going to deal with not able to send
2283                          * a response to the provided final destination
2284                          */
2285                         CERROR("Can't send response to %s. "
2286                                "No local NI available\n",
2287                                 libcfs_nid2str(sd->sd_dst_nid));
2288                         return -EHOSTUNREACH;
2289                 }
2290
2291                 return lnet_handle_send(sd);
2292         }
2293
2294         /*
2295          * If we get here that means we're sending a fresh request, PUT or
2296          * GET, so we need to run our standard selection algorithm.
2297          * First find the best local interface that's on any of the peer's
2298          * networks.
2299          */
2300         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2301                                                         sd->sd_md_cpt);
2302         if (sd->sd_best_ni) {
2303                 sd->sd_best_lpni =
2304                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2305                                              sd->sd_best_ni->ni_net->net_id);
2306
2307                 /*
2308                  * if we're successful in selecting a peer_ni on the local
2309                  * network, then send to it. Otherwise fall through and
2310                  * try and see if we can reach it over another routed
2311                  * network
2312                  */
2313                 if (sd->sd_best_lpni &&
2314                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2315                         /*
2316                          * in case we initially started with a routed
2317                          * destination, let's reset to local
2318                          */
2319                         sd->sd_send_case &= ~REMOTE_DST;
2320                         sd->sd_send_case |= LOCAL_DST;
2321                         return lnet_handle_lo_send(sd);
2322                 } else if (sd->sd_best_lpni) {
2323                         /*
2324                          * in case we initially started with a routed
2325                          * destination, let's reset to local
2326                          */
2327                         sd->sd_send_case &= ~REMOTE_DST;
2328                         sd->sd_send_case |= LOCAL_DST;
2329                         return lnet_handle_send(sd);
2330                 }
2331
2332                 CERROR("Internal Error. Expected to have a best_lpni: "
2333                        "%s -> %s\n",
2334                        libcfs_nid2str(sd->sd_src_nid),
2335                        libcfs_nid2str(sd->sd_dst_nid));
2336
2337                 return -EFAULT;
2338         }
2339
2340         /*
2341          * Peer doesn't have a local network. Let's see if there is
2342          * a remote network we can reach it on.
2343          */
2344         return PASS_THROUGH;
2345 }
2346
2347 /*
2348  * Case 1:
2349  *      Source NID not specified
2350  *      Local destination
2351  *      MR peer
2352  *
2353  * Case 2:
2354  *      Source NID not speified
2355  *      Remote destination
2356  *      MR peer
2357  *
2358  * In both of these cases if we're sending a response, ACK or REPLY, then
2359  * we need to send to the destination NID provided.
2360  *
2361  * In the remote case let's deal with MR routers.
2362  *
2363  */
2364
2365 static int
2366 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2367 {
2368         int rc = 0;
2369         struct lnet_peer *gw_peer = NULL;
2370         struct lnet_peer_ni *gw_lpni = NULL;
2371
2372         /*
2373          * handle sending a response to a remote peer here so we don't
2374          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2375          */
2376         if (sd->sd_send_case & REMOTE_DST &&
2377             sd->sd_send_case & SND_RESP) {
2378                 struct lnet_peer_ni *gw;
2379                 struct lnet_peer *gw_peer;
2380
2381                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2382                                                   &gw_peer);
2383                 if (rc < 0) {
2384                         CERROR("Can't send response to %s. "
2385                                "No route available\n",
2386                                 libcfs_nid2str(sd->sd_dst_nid));
2387                         return -EHOSTUNREACH;
2388                 }
2389
2390                 sd->sd_best_lpni = gw;
2391                 sd->sd_peer = gw_peer;
2392
2393                 return lnet_handle_send(sd);
2394         }
2395
2396         /*
2397          * Even though the NID for the peer might not be on a local network,
2398          * since the peer is MR there could be other interfaces on the
2399          * local network. In that case we'd still like to prefer the local
2400          * network over the routed network. If we're unable to do that
2401          * then we select the best router among the different routed networks,
2402          * and if the router is MR then we can deal with it as such.
2403          */
2404         rc = lnet_handle_any_mr_dsta(sd);
2405         if (rc != PASS_THROUGH)
2406                 return rc;
2407
2408         /*
2409          * TODO; One possible enhancement is to run the selection
2410          * algorithm on the peer. However for remote peers the credits are
2411          * not decremented, so we'll be basically going over the peer NIs
2412          * in round robin. An MR router will run the selection algorithm
2413          * on the next-hop interfaces.
2414          */
2415         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2416                                           &gw_peer);
2417         if (rc < 0)
2418                 return rc;
2419
2420         sd->sd_send_case &= ~LOCAL_DST;
2421         sd->sd_send_case |= REMOTE_DST;
2422
2423         sd->sd_peer = gw_peer;
2424         sd->sd_best_lpni = gw_lpni;
2425
2426         return lnet_handle_send(sd);
2427 }
2428
2429 /*
2430  * Source not specified
2431  * Remote destination
2432  * Non-MR peer
2433  *
2434  * Must send to the specified peer NID using the same source NID that
2435  * we've used before. If it's the first time to talk to that peer then
2436  * find the source NI and assign it as preferred to that peer
2437  */
2438 static int
2439 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2440 {
2441         int rc;
2442         struct lnet_peer_ni *gw_lpni = NULL;
2443         struct lnet_peer *gw_peer = NULL;
2444
2445         /*
2446          * Let's set if we have a preferred NI to talk to this NMR peer
2447          */
2448         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2449
2450         /*
2451          * find the router and that'll find the best NI if we didn't find
2452          * it already.
2453          */
2454         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2455                                           &gw_peer);
2456         if (rc < 0)
2457                 return rc;
2458
2459         /*
2460          * set the best_ni we've chosen as the preferred one for
2461          * this peer
2462          */
2463         lnet_set_non_mr_pref_nid(sd);
2464
2465         /* we'll be sending to the gw */
2466         sd->sd_best_lpni = gw_lpni;
2467         sd->sd_peer = gw_peer;
2468
2469         return lnet_handle_send(sd);
2470 }
2471
2472 static int
2473 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2474 {
2475         /*
2476          * turn off the SND_RESP bit.
2477          * It will be checked in the case handling
2478          */
2479         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2480
2481         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2482                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2483                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2484                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2485                 libcfs_nid2str(sd->sd_dst_nid),
2486                 (send_case & LOCAL_DST) ? "local" : "routed");
2487
2488         switch (send_case) {
2489         /*
2490          * For all cases where the source is specified, we should always
2491          * use the destination NID, whether it's an MR destination or not,
2492          * since we're continuing a series of related messages for the
2493          * same RPC
2494          */
2495         case SRC_SPEC_LOCAL_NMR_DST:
2496                 return lnet_handle_spec_local_nmr_dst(sd);
2497         case SRC_SPEC_LOCAL_MR_DST:
2498                 return lnet_handle_spec_local_mr_dst(sd);
2499         case SRC_SPEC_ROUTER_NMR_DST:
2500         case SRC_SPEC_ROUTER_MR_DST:
2501                 return lnet_handle_spec_router_dst(sd);
2502         case SRC_ANY_LOCAL_NMR_DST:
2503                 return lnet_handle_any_local_nmr_dst(sd);
2504         case SRC_ANY_LOCAL_MR_DST:
2505         case SRC_ANY_ROUTER_MR_DST:
2506                 return lnet_handle_any_mr_dst(sd);
2507         case SRC_ANY_ROUTER_NMR_DST:
2508                 return lnet_handle_any_router_nmr_dst(sd);
2509         default:
2510                 CERROR("Unknown send case\n");
2511                 return -1;
2512         }
2513 }
2514
2515 static int
2516 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2517                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2518 {
2519         struct lnet_peer_ni     *lpni;
2520         struct lnet_peer        *peer;
2521         struct lnet_send_data   send_data;
2522         int                     cpt, rc;
2523         int                     md_cpt;
2524         __u32                   send_case = 0;
2525
2526         memset(&send_data, 0, sizeof(send_data));
2527
2528         /*
2529          * get an initial CPT to use for locking. The idea here is not to
2530          * serialize the calls to select_pathway, so that as many
2531          * operations can run concurrently as possible. To do that we use
2532          * the CPT where this call is being executed. Later on when we
2533          * determine the CPT to use in lnet_message_commit, we switch the
2534          * lock and check if there was any configuration change.  If none,
2535          * then we proceed, if there is, then we restart the operation.
2536          */
2537         cpt = lnet_net_lock_current();
2538
2539         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2540         if (md_cpt == CFS_CPT_ANY)
2541                 md_cpt = cpt;
2542
2543 again:
2544
2545         /*
2546          * If we're being asked to send to the loopback interface, there
2547          * is no need to go through any selection. We can just shortcut
2548          * the entire process and send over lolnd
2549          */
2550         send_data.sd_msg = msg;
2551         send_data.sd_cpt = cpt;
2552         if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2553                 rc = lnet_handle_lo_send(&send_data);
2554                 lnet_net_unlock(cpt);
2555                 return rc;
2556         }
2557
2558         /*
2559          * find an existing peer_ni, or create one and mark it as having been
2560          * created due to network traffic. This call will create the
2561          * peer->peer_net->peer_ni tree.
2562          */
2563         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2564         if (IS_ERR(lpni)) {
2565                 lnet_net_unlock(cpt);
2566                 return PTR_ERR(lpni);
2567         }
2568
2569         /*
2570          * Cache the original src_nid. If we need to resend the message
2571          * then we'll need to know whether the src_nid was originally
2572          * specified for this message. If it was originally specified,
2573          * then we need to keep using the same src_nid since it's
2574          * continuing the same sequence of messages.
2575          */
2576         msg->msg_src_nid_param = src_nid;
2577
2578         /*
2579          * Now that we have a peer_ni, check if we want to discover
2580          * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2581          * trigger discovery.
2582          */
2583         peer = lpni->lpni_peer_net->lpn_peer;
2584         if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
2585                 lnet_nid_t primary_nid;
2586                 rc = lnet_discover_peer_locked(lpni, cpt, false);
2587                 if (rc) {
2588                         lnet_peer_ni_decref_locked(lpni);
2589                         lnet_net_unlock(cpt);
2590                         return rc;
2591                 }
2592                 /* The peer may have changed. */
2593                 peer = lpni->lpni_peer_net->lpn_peer;
2594                 /* queue message and return */
2595                 msg->msg_rtr_nid_param = rtr_nid;
2596                 msg->msg_sending = 0;
2597                 spin_lock(&peer->lp_lock);
2598                 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2599                 spin_unlock(&peer->lp_lock);
2600                 lnet_peer_ni_decref_locked(lpni);
2601                 primary_nid = peer->lp_primary_nid;
2602                 lnet_net_unlock(cpt);
2603
2604                 CDEBUG(D_NET, "%s pending discovery\n",
2605                        libcfs_nid2str(primary_nid));
2606
2607                 return LNET_DC_WAIT;
2608         }
2609         lnet_peer_ni_decref_locked(lpni);
2610
2611         /*
2612          * Identify the different send cases
2613          */
2614         if (src_nid == LNET_NID_ANY)
2615                 send_case |= SRC_ANY;
2616         else
2617                 send_case |= SRC_SPEC;
2618
2619         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2620                 send_case |= LOCAL_DST;
2621         else
2622                 send_case |= REMOTE_DST;
2623
2624         /*
2625          * if this is a non-MR peer or if we're recovering a peer ni then
2626          * let's consider this an NMR case so we can hit the destination
2627          * NID.
2628          */
2629         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2630                 send_case |= NMR_DST;
2631         else
2632                 send_case |= MR_DST;
2633
2634         if (msg->msg_type == LNET_MSG_REPLY ||
2635             msg->msg_type == LNET_MSG_ACK)
2636                 send_case |= SND_RESP;
2637
2638         /* assign parameters to the send_data */
2639         send_data.sd_rtr_nid = rtr_nid;
2640         send_data.sd_src_nid = src_nid;
2641         send_data.sd_dst_nid = dst_nid;
2642         send_data.sd_best_lpni = lpni;
2643         /*
2644          * keep a pointer to the final destination in case we're going to
2645          * route, so we'll need to access it later
2646          */
2647         send_data.sd_final_dst_lpni = lpni;
2648         send_data.sd_peer = peer;
2649         send_data.sd_md_cpt = md_cpt;
2650         send_data.sd_send_case = send_case;
2651
2652         rc = lnet_handle_send_case_locked(&send_data);
2653
2654         /*
2655          * Update the local cpt since send_data.sd_cpt might've been
2656          * updated as a result of calling lnet_handle_send_case_locked().
2657          */
2658         cpt = send_data.sd_cpt;
2659
2660         if (rc == REPEAT_SEND)
2661                 goto again;
2662
2663         lnet_net_unlock(cpt);
2664
2665         return rc;
2666 }
2667
2668 int
2669 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2670 {
2671         lnet_nid_t              dst_nid = msg->msg_target.nid;
2672         int                     rc;
2673
2674         /*
2675          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2676          * but we might want to use pre-determined router for ACK/REPLY
2677          * in the future
2678          */
2679         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2680         LASSERT(msg->msg_txpeer == NULL);
2681         LASSERT(msg->msg_txni == NULL);
2682         LASSERT(!msg->msg_sending);
2683         LASSERT(!msg->msg_target_is_router);
2684         LASSERT(!msg->msg_receiving);
2685
2686         msg->msg_sending = 1;
2687
2688         LASSERT(!msg->msg_tx_committed);
2689
2690         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2691         if (rc < 0)
2692                 return rc;
2693
2694         if (rc == LNET_CREDIT_OK)
2695                 lnet_ni_send(msg->msg_txni, msg);
2696
2697         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2698         return 0;
2699 }
2700
2701 enum lnet_mt_event_type {
2702         MT_TYPE_LOCAL_NI = 0,
2703         MT_TYPE_PEER_NI
2704 };
2705
2706 struct lnet_mt_event_info {
2707         enum lnet_mt_event_type mt_type;
2708         lnet_nid_t mt_nid;
2709 };
2710
2711 /* called with res_lock held */
2712 void
2713 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2714 {
2715         struct lnet_rsp_tracker *rspt;
2716
2717         /*
2718          * msg has a refcount on the MD so the MD is not going away.
2719          * The rspt queue for the cpt is protected by
2720          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2721          */
2722         if (!md->md_rspt_ptr)
2723                 return;
2724
2725         rspt = md->md_rspt_ptr;
2726         md->md_rspt_ptr = NULL;
2727
2728         /* debug code */
2729         LASSERT(rspt->rspt_cpt == cpt);
2730
2731         /*
2732          * invalidate the handle to indicate that a response has been
2733          * received, which will then lead the monitor thread to clean up
2734          * the rspt block.
2735          */
2736         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2737 }
2738
2739 static void
2740 lnet_finalize_expired_responses(bool force)
2741 {
2742         struct lnet_libmd *md;
2743         struct list_head local_queue;
2744         struct lnet_rsp_tracker *rspt, *tmp;
2745         int i;
2746
2747         if (the_lnet.ln_mt_rstq == NULL)
2748                 return;
2749
2750         cfs_cpt_for_each(i, lnet_cpt_table()) {
2751                 INIT_LIST_HEAD(&local_queue);
2752
2753                 lnet_net_lock(i);
2754                 if (!the_lnet.ln_mt_rstq[i]) {
2755                         lnet_net_unlock(i);
2756                         continue;
2757                 }
2758                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2759                 lnet_net_unlock(i);
2760
2761                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2762                         /*
2763                          * The rspt mdh will be invalidated when a response
2764                          * is received or whenever we want to discard the
2765                          * block the monitor thread will walk the queue
2766                          * and clean up any rsts with an invalid mdh.
2767                          * The monitor thread will walk the queue until
2768                          * the first unexpired rspt block. This means that
2769                          * some rspt blocks which received their
2770                          * corresponding responses will linger in the
2771                          * queue until they are cleaned up eventually.
2772                          */
2773                         lnet_res_lock(i);
2774                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2775                                 lnet_res_unlock(i);
2776                                 list_del_init(&rspt->rspt_on_list);
2777                                 lnet_rspt_free(rspt, i);
2778                                 continue;
2779                         }
2780
2781                         if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
2782                             force) {
2783                                 struct lnet_peer_ni *lpni;
2784                                 lnet_nid_t nid;
2785
2786                                 md = lnet_handle2md(&rspt->rspt_mdh);
2787                                 if (!md) {
2788                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2789                                         lnet_res_unlock(i);
2790                                         list_del_init(&rspt->rspt_on_list);
2791                                         lnet_rspt_free(rspt, i);
2792                                         continue;
2793                                 }
2794                                 LASSERT(md->md_rspt_ptr == rspt);
2795                                 md->md_rspt_ptr = NULL;
2796                                 lnet_res_unlock(i);
2797
2798                                 lnet_net_lock(i);
2799                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2800                                 lnet_net_unlock(i);
2801
2802                                 list_del_init(&rspt->rspt_on_list);
2803
2804                                 nid = rspt->rspt_next_hop_nid;
2805
2806                                 CNETERR("Response timed out: md = %p: nid = %s\n",
2807                                         md, libcfs_nid2str(nid));
2808                                 LNetMDUnlink(rspt->rspt_mdh);
2809                                 lnet_rspt_free(rspt, i);
2810
2811                                 /*
2812                                  * If there is a timeout on the response
2813                                  * from the next hop decrement its health
2814                                  * value so that we don't use it
2815                                  */
2816                                 lnet_net_lock(0);
2817                                 lpni = lnet_find_peer_ni_locked(nid);
2818                                 if (lpni) {
2819                                         lnet_handle_remote_failure_locked(lpni);
2820                                         lnet_peer_ni_decref_locked(lpni);
2821                                 }
2822                                 lnet_net_unlock(0);
2823                         } else {
2824                                 lnet_res_unlock(i);
2825                                 break;
2826                         }
2827                 }
2828
2829                 lnet_net_lock(i);
2830                 if (!list_empty(&local_queue))
2831                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2832                 lnet_net_unlock(i);
2833         }
2834 }
2835
2836 static void
2837 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2838 {
2839         struct lnet_msg *msg;
2840
2841         while (!list_empty(resendq)) {
2842                 struct lnet_peer_ni *lpni;
2843
2844                 msg = list_entry(resendq->next, struct lnet_msg,
2845                                  msg_list);
2846
2847                 list_del_init(&msg->msg_list);
2848
2849                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2850                 if (!lpni) {
2851                         lnet_net_unlock(cpt);
2852                         CERROR("Expected that a peer is already created for %s\n",
2853                                libcfs_nid2str(msg->msg_hdr.dest_nid));
2854                         msg->msg_no_resend = true;
2855                         lnet_finalize(msg, -EFAULT);
2856                         lnet_net_lock(cpt);
2857                 } else {
2858                         struct lnet_peer *peer;
2859                         int rc;
2860                         lnet_nid_t src_nid = LNET_NID_ANY;
2861
2862                         /*
2863                          * if this message is not being routed and the
2864                          * peer is non-MR then we must use the same
2865                          * src_nid that was used in the original send.
2866                          * Otherwise if we're routing the message (IE
2867                          * we're a router) then we can use any of our
2868                          * local interfaces. It doesn't matter to the
2869                          * final destination.
2870                          */
2871                         peer = lpni->lpni_peer_net->lpn_peer;
2872                         if (!msg->msg_routing &&
2873                             !lnet_peer_is_multi_rail(peer))
2874                                 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
2875
2876                         /*
2877                          * If we originally specified a src NID, then we
2878                          * must attempt to reuse it in the resend as well.
2879                          */
2880                         if (msg->msg_src_nid_param != LNET_NID_ANY)
2881                                 src_nid = msg->msg_src_nid_param;
2882                         lnet_peer_ni_decref_locked(lpni);
2883
2884                         lnet_net_unlock(cpt);
2885                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
2886                                libcfs_nid2str(src_nid),
2887                                libcfs_id2str(msg->msg_target),
2888                                lnet_msgtyp2str(msg->msg_type),
2889                                msg->msg_recovery,
2890                                msg->msg_retry_count);
2891                         rc = lnet_send(src_nid, msg, LNET_NID_ANY);
2892                         if (rc) {
2893                                 CERROR("Error sending %s to %s: %d\n",
2894                                        lnet_msgtyp2str(msg->msg_type),
2895                                        libcfs_id2str(msg->msg_target), rc);
2896                                 msg->msg_no_resend = true;
2897                                 lnet_finalize(msg, rc);
2898                         }
2899                         lnet_net_lock(cpt);
2900                         if (!rc)
2901                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
2902                 }
2903         }
2904 }
2905
2906 static void
2907 lnet_resend_pending_msgs(void)
2908 {
2909         int i;
2910
2911         cfs_cpt_for_each(i, lnet_cpt_table()) {
2912                 lnet_net_lock(i);
2913                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
2914                 lnet_net_unlock(i);
2915         }
2916 }
2917
2918 /* called with cpt and ni_lock held */
2919 static void
2920 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
2921 {
2922         struct lnet_handle_md recovery_mdh;
2923
2924         LNetInvalidateMDHandle(&recovery_mdh);
2925
2926         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
2927             force) {
2928                 recovery_mdh = ni->ni_ping_mdh;
2929                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2930         }
2931         lnet_ni_unlock(ni);
2932         lnet_net_unlock(cpt);
2933         if (!LNetMDHandleIsInvalid(recovery_mdh))
2934                 LNetMDUnlink(recovery_mdh);
2935         lnet_net_lock(cpt);
2936         lnet_ni_lock(ni);
2937 }
2938
2939 static void
2940 lnet_recover_local_nis(void)
2941 {
2942         struct lnet_mt_event_info *ev_info;
2943         struct list_head processed_list;
2944         struct list_head local_queue;
2945         struct lnet_handle_md mdh;
2946         struct lnet_ni *tmp;
2947         struct lnet_ni *ni;
2948         lnet_nid_t nid;
2949         int healthv;
2950         int rc;
2951
2952         INIT_LIST_HEAD(&local_queue);
2953         INIT_LIST_HEAD(&processed_list);
2954
2955         /*
2956          * splice the recovery queue on a local queue. We will iterate
2957          * through the local queue and update it as needed. Once we're
2958          * done with the traversal, we'll splice the local queue back on
2959          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
2960          * will be traversed in the next iteration.
2961          */
2962         lnet_net_lock(0);
2963         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
2964                          &local_queue);
2965         lnet_net_unlock(0);
2966
2967         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
2968                 /*
2969                  * if an NI is being deleted or it is now healthy, there
2970                  * is no need to keep it around in the recovery queue.
2971                  * The monitor thread is the only thread responsible for
2972                  * removing the NI from the recovery queue.
2973                  * Multiple threads can be adding NIs to the recovery
2974                  * queue.
2975                  */
2976                 healthv = atomic_read(&ni->ni_healthv);
2977
2978                 lnet_net_lock(0);
2979                 lnet_ni_lock(ni);
2980                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
2981                     healthv == LNET_MAX_HEALTH_VALUE) {
2982                         list_del_init(&ni->ni_recovery);
2983                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
2984                         lnet_ni_unlock(ni);
2985                         lnet_ni_decref_locked(ni, 0);
2986                         lnet_net_unlock(0);
2987                         continue;
2988                 }
2989
2990                 /*
2991                  * if the local NI failed recovery we must unlink the md.
2992                  * But we want to keep the local_ni on the recovery queue
2993                  * so we can continue the attempts to recover it.
2994                  */
2995                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
2996                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
2997                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
2998                 }
2999
3000                 lnet_ni_unlock(ni);
3001                 lnet_net_unlock(0);
3002
3003
3004                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3005                        libcfs_nid2str(ni->ni_nid));
3006
3007                 lnet_ni_lock(ni);
3008                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3009                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3010                         lnet_ni_unlock(ni);
3011
3012                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3013                         if (!ev_info) {
3014                                 CERROR("out of memory. Can't recover %s\n",
3015                                        libcfs_nid2str(ni->ni_nid));
3016                                 lnet_ni_lock(ni);
3017                                 ni->ni_recovery_state &=
3018                                   ~LNET_NI_RECOVERY_PENDING;
3019                                 lnet_ni_unlock(ni);
3020                                 continue;
3021                         }
3022
3023                         mdh = ni->ni_ping_mdh;
3024                         /*
3025                          * Invalidate the ni mdh in case it's deleted.
3026                          * We'll unlink the mdh in this case below.
3027                          */
3028                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3029                         nid = ni->ni_nid;
3030
3031                         /*
3032                          * remove the NI from the local queue and drop the
3033                          * reference count to it while we're recovering
3034                          * it. The reason for that, is that the NI could
3035                          * be deleted, and the way the code is structured
3036                          * is if we don't drop the NI, then the deletion
3037                          * code will enter a loop waiting for the
3038                          * reference count to be removed while holding the
3039                          * ln_mutex_lock(). When we look up the peer to
3040                          * send to in lnet_select_pathway() we will try to
3041                          * lock the ln_mutex_lock() as well, leading to
3042                          * a deadlock. By dropping the refcount and
3043                          * removing it from the list, we allow for the NI
3044                          * to be removed, then we use the cached NID to
3045                          * look it up again. If it's gone, then we just
3046                          * continue examining the rest of the queue.
3047                          */
3048                         lnet_net_lock(0);
3049                         list_del_init(&ni->ni_recovery);
3050                         lnet_ni_decref_locked(ni, 0);
3051                         lnet_net_unlock(0);
3052
3053                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3054                         ev_info->mt_nid = nid;
3055                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3056                                             ev_info, the_lnet.ln_mt_eqh, true);
3057                         /* lookup the nid again */
3058                         lnet_net_lock(0);
3059                         ni = lnet_nid2ni_locked(nid, 0);
3060                         if (!ni) {
3061                                 /*
3062                                  * the NI has been deleted when we dropped
3063                                  * the ref count
3064                                  */
3065                                 lnet_net_unlock(0);
3066                                 LNetMDUnlink(mdh);
3067                                 continue;
3068                         }
3069                         /*
3070                          * Same note as in lnet_recover_peer_nis(). When
3071                          * we're sending the ping, the NI is free to be
3072                          * deleted or manipulated. By this point it
3073                          * could've been added back on the recovery queue,
3074                          * and a refcount taken on it.
3075                          * So we can't just add it blindly again or we'll
3076                          * corrupt the queue. We must check under lock if
3077                          * it's not on any list and if not then add it
3078                          * to the processed list, which will eventually be
3079                          * spliced back on to the recovery queue.
3080                          */
3081                         ni->ni_ping_mdh = mdh;
3082                         if (list_empty(&ni->ni_recovery)) {
3083                                 list_add_tail(&ni->ni_recovery, &processed_list);
3084                                 lnet_ni_addref_locked(ni, 0);
3085                         }
3086                         lnet_net_unlock(0);
3087
3088                         lnet_ni_lock(ni);
3089                         if (rc)
3090                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3091                 }
3092                 lnet_ni_unlock(ni);
3093         }
3094
3095         /*
3096          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3097          * reexamined in the next iteration.
3098          */
3099         list_splice_init(&processed_list, &local_queue);
3100         lnet_net_lock(0);
3101         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3102         lnet_net_unlock(0);
3103 }
3104
3105 static struct list_head **
3106 lnet_create_array_of_queues(void)
3107 {
3108         struct list_head **qs;
3109         struct list_head *q;
3110         int i;
3111
3112         qs = cfs_percpt_alloc(lnet_cpt_table(),
3113                               sizeof(struct list_head));
3114         if (!qs) {
3115                 CERROR("Failed to allocate queues\n");
3116                 return NULL;
3117         }
3118
3119         cfs_percpt_for_each(q, i, qs)
3120                 INIT_LIST_HEAD(q);
3121
3122         return qs;
3123 }
3124
3125 static int
3126 lnet_resendqs_create(void)
3127 {
3128         struct list_head **resendqs;
3129         resendqs = lnet_create_array_of_queues();
3130
3131         if (!resendqs)
3132                 return -ENOMEM;
3133
3134         lnet_net_lock(LNET_LOCK_EX);
3135         the_lnet.ln_mt_resendqs = resendqs;
3136         lnet_net_unlock(LNET_LOCK_EX);
3137
3138         return 0;
3139 }
3140
3141 static void
3142 lnet_clean_local_ni_recoveryq(void)
3143 {
3144         struct lnet_ni *ni;
3145
3146         /* This is only called when the monitor thread has stopped */
3147         lnet_net_lock(0);
3148
3149         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3150                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3151                                 struct lnet_ni, ni_recovery);
3152                 list_del_init(&ni->ni_recovery);
3153                 lnet_ni_lock(ni);
3154                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3155                 lnet_ni_unlock(ni);
3156                 lnet_ni_decref_locked(ni, 0);
3157         }
3158
3159         lnet_net_unlock(0);
3160 }
3161
3162 static void
3163 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3164                                      bool force)
3165 {
3166         struct lnet_handle_md recovery_mdh;
3167
3168         LNetInvalidateMDHandle(&recovery_mdh);
3169
3170         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3171                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3172                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3173         }
3174         spin_unlock(&lpni->lpni_lock);
3175         lnet_net_unlock(cpt);
3176         if (!LNetMDHandleIsInvalid(recovery_mdh))
3177                 LNetMDUnlink(recovery_mdh);
3178         lnet_net_lock(cpt);
3179         spin_lock(&lpni->lpni_lock);
3180 }
3181
3182 static void
3183 lnet_clean_peer_ni_recoveryq(void)
3184 {
3185         struct lnet_peer_ni *lpni, *tmp;
3186
3187         lnet_net_lock(LNET_LOCK_EX);
3188
3189         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3190                                  lpni_recovery) {
3191                 list_del_init(&lpni->lpni_recovery);
3192                 spin_lock(&lpni->lpni_lock);
3193                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3194                 spin_unlock(&lpni->lpni_lock);
3195                 lnet_peer_ni_decref_locked(lpni);
3196         }
3197
3198         lnet_net_unlock(LNET_LOCK_EX);
3199 }
3200
3201 static void
3202 lnet_clean_resendqs(void)
3203 {
3204         struct lnet_msg *msg, *tmp;
3205         struct list_head msgs;
3206         int i;
3207
3208         INIT_LIST_HEAD(&msgs);
3209
3210         cfs_cpt_for_each(i, lnet_cpt_table()) {
3211                 lnet_net_lock(i);
3212                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3213                 lnet_net_unlock(i);
3214                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3215                         list_del_init(&msg->msg_list);
3216                         msg->msg_no_resend = true;
3217                         lnet_finalize(msg, -ESHUTDOWN);
3218                 }
3219         }
3220
3221         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3222 }
3223
3224 static void
3225 lnet_recover_peer_nis(void)
3226 {
3227         struct lnet_mt_event_info *ev_info;
3228         struct list_head processed_list;
3229         struct list_head local_queue;
3230         struct lnet_handle_md mdh;
3231         struct lnet_peer_ni *lpni;
3232         struct lnet_peer_ni *tmp;
3233         lnet_nid_t nid;
3234         int healthv;
3235         int rc;
3236
3237         INIT_LIST_HEAD(&local_queue);
3238         INIT_LIST_HEAD(&processed_list);
3239
3240         /*
3241          * Always use cpt 0 for locking across all interactions with
3242          * ln_mt_peerNIRecovq
3243          */
3244         lnet_net_lock(0);
3245         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3246                          &local_queue);
3247         lnet_net_unlock(0);
3248
3249         list_for_each_entry_safe(lpni, tmp, &local_queue,
3250                                  lpni_recovery) {
3251                 /*
3252                  * The same protection strategy is used here as is in the
3253                  * local recovery case.
3254                  */
3255                 lnet_net_lock(0);
3256                 healthv = atomic_read(&lpni->lpni_healthv);
3257                 spin_lock(&lpni->lpni_lock);
3258                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3259                     healthv == LNET_MAX_HEALTH_VALUE) {
3260                         list_del_init(&lpni->lpni_recovery);
3261                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3262                         spin_unlock(&lpni->lpni_lock);
3263                         lnet_peer_ni_decref_locked(lpni);
3264                         lnet_net_unlock(0);
3265                         continue;
3266                 }
3267
3268                 /*
3269                  * If the peer NI has failed recovery we must unlink the
3270                  * md. But we want to keep the peer ni on the recovery
3271                  * queue so we can try to continue recovering it
3272                  */
3273                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3274                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3275                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3276                 }
3277
3278                 spin_unlock(&lpni->lpni_lock);
3279                 lnet_net_unlock(0);
3280
3281                 /*
3282                  * NOTE: we're racing with peer deletion from user space.
3283                  * It's possible that a peer is deleted after we check its
3284                  * state. In this case the recovery can create a new peer
3285                  */
3286                 spin_lock(&lpni->lpni_lock);
3287                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3288                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3289                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3290                         spin_unlock(&lpni->lpni_lock);
3291
3292                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3293                         if (!ev_info) {
3294                                 CERROR("out of memory. Can't recover %s\n",
3295                                        libcfs_nid2str(lpni->lpni_nid));
3296                                 spin_lock(&lpni->lpni_lock);
3297                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3298                                 spin_unlock(&lpni->lpni_lock);
3299                                 continue;
3300                         }
3301
3302                         /* look at the comments in lnet_recover_local_nis() */
3303                         mdh = lpni->lpni_recovery_ping_mdh;
3304                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3305                         nid = lpni->lpni_nid;
3306                         lnet_net_lock(0);
3307                         list_del_init(&lpni->lpni_recovery);
3308                         lnet_peer_ni_decref_locked(lpni);
3309                         lnet_net_unlock(0);
3310
3311                         ev_info->mt_type = MT_TYPE_PEER_NI;
3312                         ev_info->mt_nid = nid;
3313                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3314                                             ev_info, the_lnet.ln_mt_eqh, true);
3315                         lnet_net_lock(0);
3316                         /*
3317                          * lnet_find_peer_ni_locked() grabs a refcount for
3318                          * us. No need to take it explicitly.
3319                          */
3320                         lpni = lnet_find_peer_ni_locked(nid);
3321                         if (!lpni) {
3322                                 lnet_net_unlock(0);
3323                                 LNetMDUnlink(mdh);
3324                                 continue;
3325                         }
3326
3327                         lpni->lpni_recovery_ping_mdh = mdh;
3328                         /*
3329                          * While we're unlocked the lpni could've been
3330                          * readded on the recovery queue. In this case we
3331                          * don't need to add it to the local queue, since
3332                          * it's already on there and the thread that added
3333                          * it would've incremented the refcount on the
3334                          * peer, which means we need to decref the refcount
3335                          * that was implicitly grabbed by find_peer_ni_locked.
3336                          * Otherwise, if the lpni is still not on
3337                          * the recovery queue, then we'll add it to the
3338                          * processed list.
3339                          */
3340                         if (list_empty(&lpni->lpni_recovery))
3341                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3342                         else
3343                                 lnet_peer_ni_decref_locked(lpni);
3344                         lnet_net_unlock(0);
3345
3346                         spin_lock(&lpni->lpni_lock);
3347                         if (rc)
3348                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3349                 }
3350                 spin_unlock(&lpni->lpni_lock);
3351         }
3352
3353         list_splice_init(&processed_list, &local_queue);
3354         lnet_net_lock(0);
3355         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3356         lnet_net_unlock(0);
3357 }
3358
3359 static int
3360 lnet_monitor_thread(void *arg)
3361 {
3362         time64_t recovery_timeout = 0;
3363         time64_t rsp_timeout = 0;
3364         int interval;
3365         time64_t now;
3366
3367         /*
3368          * The monitor thread takes care of the following:
3369          *  1. Checks the aliveness of routers
3370          *  2. Checks if there are messages on the resend queue to resend
3371          *     them.
3372          *  3. Check if there are any NIs on the local recovery queue and
3373          *     pings them
3374          *  4. Checks if there are any NIs on the remote recovery queue
3375          *     and pings them.
3376          */
3377         cfs_block_allsigs();
3378
3379         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3380                 now = ktime_get_real_seconds();
3381
3382                 if (lnet_router_checker_active())
3383                         lnet_check_routers();
3384
3385                 lnet_resend_pending_msgs();
3386
3387                 if (now >= rsp_timeout) {
3388                         lnet_finalize_expired_responses(false);
3389                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3390                 }
3391
3392                 if (now >= recovery_timeout) {
3393                         lnet_recover_local_nis();
3394                         lnet_recover_peer_nis();
3395                         recovery_timeout = now + lnet_recovery_interval;
3396                 }
3397
3398                 /*
3399                  * TODO do we need to check if we should sleep without
3400                  * timeout?  Technically, an active system will always
3401                  * have messages in flight so this check will always
3402                  * evaluate to false. And on an idle system do we care
3403                  * if we wake up every 1 second? Although, we've seen
3404                  * cases where we get a complaint that an idle thread
3405                  * is waking up unnecessarily.
3406                  */
3407                 interval = min(lnet_recovery_interval,
3408                                lnet_transaction_timeout / 2);
3409                 wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
3410                                                 false,
3411                                                 cfs_time_seconds(interval));
3412         }
3413
3414         /* clean up the router checker */
3415         lnet_prune_rc_data(1);
3416
3417         /* Shutting down */
3418         lnet_net_lock(LNET_LOCK_EX);
3419         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3420         lnet_net_unlock(LNET_LOCK_EX);
3421
3422         /* signal that the monitor thread is exiting */
3423         up(&the_lnet.ln_mt_signal);
3424
3425         return 0;
3426 }
3427
3428 /*
3429  * lnet_send_ping
3430  * Sends a ping.
3431  * Returns == 0 if success
3432  * Returns > 0 if LNetMDBind or prior fails
3433  * Returns < 0 if LNetGet fails
3434  */
3435 int
3436 lnet_send_ping(lnet_nid_t dest_nid,
3437                struct lnet_handle_md *mdh, int nnis,
3438                void *user_data, struct lnet_handle_eq eqh, bool recovery)
3439 {
3440         struct lnet_md md = { NULL };
3441         struct lnet_process_id id;
3442         struct lnet_ping_buffer *pbuf;
3443         int rc;
3444
3445         if (dest_nid == LNET_NID_ANY) {
3446                 rc = -EHOSTUNREACH;
3447                 goto fail_error;
3448         }
3449
3450         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3451         if (!pbuf) {
3452                 rc = ENOMEM;
3453                 goto fail_error;
3454         }
3455
3456         /* initialize md content */
3457         md.start     = &pbuf->pb_info;
3458         md.length    = LNET_PING_INFO_SIZE(nnis);
3459         md.threshold = 2; /* GET/REPLY */
3460         md.max_size  = 0;
3461         md.options   = LNET_MD_TRUNCATE;
3462         md.user_ptr  = user_data;
3463         md.eq_handle = eqh;
3464
3465         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3466         if (rc) {
3467                 lnet_ping_buffer_decref(pbuf);
3468                 CERROR("Can't bind MD: %d\n", rc);
3469                 rc = -rc; /* change the rc to positive */
3470                 goto fail_error;
3471         }
3472         id.pid = LNET_PID_LUSTRE;
3473         id.nid = dest_nid;
3474
3475         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3476                      LNET_RESERVED_PORTAL,
3477                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3478
3479         if (rc)
3480                 goto fail_unlink_md;
3481
3482         return 0;
3483
3484 fail_unlink_md:
3485         LNetMDUnlink(*mdh);
3486         LNetInvalidateMDHandle(mdh);
3487 fail_error:
3488         return rc;
3489 }
3490
3491 static void
3492 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3493                            int status, bool unlink_event)
3494 {
3495         lnet_nid_t nid = ev_info->mt_nid;
3496
3497         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3498                 struct lnet_ni *ni;
3499
3500                 lnet_net_lock(0);
3501                 ni = lnet_nid2ni_locked(nid, 0);
3502                 if (!ni) {
3503                         lnet_net_unlock(0);
3504                         return;
3505                 }
3506                 lnet_ni_lock(ni);
3507                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3508                 if (status)
3509                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3510                 lnet_ni_unlock(ni);
3511                 lnet_net_unlock(0);
3512
3513                 if (status != 0) {
3514                         CERROR("local NI (%s) recovery failed with %d\n",
3515                                libcfs_nid2str(nid), status);
3516                         return;
3517                 }
3518                 /*
3519                  * need to increment healthv for the ni here, because in
3520                  * the lnet_finalize() path we don't have access to this
3521                  * NI. And in order to get access to it, we'll need to
3522                  * carry forward too much information.
3523                  * In the peer case, it'll naturally be incremented
3524                  */
3525                 if (!unlink_event)
3526                         lnet_inc_healthv(&ni->ni_healthv);
3527         } else {
3528                 struct lnet_peer_ni *lpni;
3529                 int cpt;
3530
3531                 cpt = lnet_net_lock_current();
3532                 lpni = lnet_find_peer_ni_locked(nid);
3533                 if (!lpni) {
3534                         lnet_net_unlock(cpt);
3535                         return;
3536                 }
3537                 spin_lock(&lpni->lpni_lock);
3538                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3539                 if (status)
3540                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3541                 spin_unlock(&lpni->lpni_lock);
3542                 lnet_peer_ni_decref_locked(lpni);
3543                 lnet_net_unlock(cpt);
3544
3545                 if (status != 0)
3546                         CERROR("peer NI (%s) recovery failed with %d\n",
3547                                libcfs_nid2str(nid), status);
3548         }
3549 }
3550
3551 void
3552 lnet_mt_event_handler(struct lnet_event *event)
3553 {
3554         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3555         struct lnet_ping_buffer *pbuf;
3556
3557         /* TODO: remove assert */
3558         LASSERT(event->type == LNET_EVENT_REPLY ||
3559                 event->type == LNET_EVENT_SEND ||
3560                 event->type == LNET_EVENT_UNLINK);
3561
3562         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3563                event->status);
3564
3565         switch (event->type) {
3566         case LNET_EVENT_UNLINK:
3567                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3568                        libcfs_nid2str(ev_info->mt_nid));
3569         case LNET_EVENT_REPLY:
3570                 lnet_handle_recovery_reply(ev_info, event->status,
3571                                            event->type == LNET_EVENT_UNLINK);
3572                 break;
3573         case LNET_EVENT_SEND:
3574                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3575                                libcfs_nid2str(ev_info->mt_nid),
3576                                (event->status) ? "unsuccessfully" :
3577                                "successfully", event->status);
3578                 break;
3579         default:
3580                 CERROR("Unexpected event: %d\n", event->type);
3581                 break;
3582         }
3583         if (event->unlinked) {
3584                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3585                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3586                 lnet_ping_buffer_decref(pbuf);
3587         }
3588 }
3589
3590 static int
3591 lnet_rsp_tracker_create(void)
3592 {
3593         struct list_head **rstqs;
3594         rstqs = lnet_create_array_of_queues();
3595
3596         if (!rstqs)
3597                 return -ENOMEM;
3598
3599         the_lnet.ln_mt_rstq = rstqs;
3600
3601         return 0;
3602 }
3603
3604 static void
3605 lnet_rsp_tracker_clean(void)
3606 {
3607         lnet_finalize_expired_responses(true);
3608
3609         cfs_percpt_free(the_lnet.ln_mt_rstq);
3610         the_lnet.ln_mt_rstq = NULL;
3611 }
3612
3613 int lnet_monitor_thr_start(void)
3614 {
3615         int rc = 0;
3616         struct task_struct *task;
3617
3618         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3619                 return -EALREADY;
3620
3621         rc = lnet_resendqs_create();
3622         if (rc)
3623                 return rc;
3624
3625         rc = lnet_rsp_tracker_create();
3626         if (rc)
3627                 goto clean_queues;
3628
3629         /* Pre monitor thread start processing */
3630         rc = lnet_router_pre_mt_start();
3631         if (rc)
3632                 goto free_mem;
3633
3634         sema_init(&the_lnet.ln_mt_signal, 0);
3635
3636         lnet_net_lock(LNET_LOCK_EX);
3637         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3638         lnet_net_unlock(LNET_LOCK_EX);
3639         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3640         if (IS_ERR(task)) {
3641                 rc = PTR_ERR(task);
3642                 CERROR("Can't start monitor thread: %d\n", rc);
3643                 goto clean_thread;
3644         }
3645
3646         /* post monitor thread start processing */
3647         lnet_router_post_mt_start();
3648
3649         return 0;
3650
3651 clean_thread:
3652         lnet_net_lock(LNET_LOCK_EX);
3653         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3654         lnet_net_unlock(LNET_LOCK_EX);
3655         /* block until event callback signals exit */
3656         down(&the_lnet.ln_mt_signal);
3657         /* clean up */
3658         lnet_router_cleanup();
3659 free_mem:
3660         lnet_net_lock(LNET_LOCK_EX);
3661         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3662         lnet_net_unlock(LNET_LOCK_EX);
3663         lnet_rsp_tracker_clean();
3664         lnet_clean_local_ni_recoveryq();
3665         lnet_clean_peer_ni_recoveryq();
3666         lnet_clean_resendqs();
3667         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3668         return rc;
3669 clean_queues:
3670         lnet_rsp_tracker_clean();
3671         lnet_clean_local_ni_recoveryq();
3672         lnet_clean_peer_ni_recoveryq();
3673         lnet_clean_resendqs();
3674         return rc;
3675 }
3676
3677 void lnet_monitor_thr_stop(void)
3678 {
3679         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3680                 return;
3681
3682         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3683         lnet_net_lock(LNET_LOCK_EX);
3684         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3685         lnet_net_unlock(LNET_LOCK_EX);
3686
3687         /* tell the monitor thread that we're shutting down */
3688         wake_up(&the_lnet.ln_mt_waitq);
3689
3690         /* block until monitor thread signals that it's done */
3691         down(&the_lnet.ln_mt_signal);
3692         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3693
3694         /* perform cleanup tasks */
3695         lnet_router_cleanup();
3696         lnet_rsp_tracker_clean();
3697         lnet_clean_local_ni_recoveryq();
3698         lnet_clean_peer_ni_recoveryq();
3699         lnet_clean_resendqs();
3700
3701         return;
3702 }
3703
3704 void
3705 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3706                   __u32 msg_type)
3707 {
3708         lnet_net_lock(cpt);
3709         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3710         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3711         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3712         lnet_net_unlock(cpt);
3713
3714         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3715 }
3716
3717 static void
3718 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3719 {
3720         struct lnet_hdr *hdr = &msg->msg_hdr;
3721
3722         if (msg->msg_wanted != 0)
3723                 lnet_setpayloadbuffer(msg);
3724
3725         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3726
3727         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3728          * it back into the ACK during lnet_finalize() */
3729         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3730                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3731
3732         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3733                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3734 }
3735
3736 static int
3737 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3738 {
3739         struct lnet_hdr         *hdr = &msg->msg_hdr;
3740         struct lnet_match_info  info;
3741         int                     rc;
3742         bool                    ready_delay;
3743
3744         /* Convert put fields to host byte order */
3745         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3746         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3747         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3748
3749         /* Primary peer NID. */
3750         info.mi_id.nid  = msg->msg_initiator;
3751         info.mi_id.pid  = hdr->src_pid;
3752         info.mi_opc     = LNET_MD_OP_PUT;
3753         info.mi_portal  = hdr->msg.put.ptl_index;
3754         info.mi_rlength = hdr->payload_length;
3755         info.mi_roffset = hdr->msg.put.offset;
3756         info.mi_mbits   = hdr->msg.put.match_bits;
3757         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3758
3759         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3760         ready_delay = msg->msg_rx_ready_delay;
3761
3762  again:
3763         rc = lnet_ptl_match_md(&info, msg);
3764         switch (rc) {
3765         default:
3766                 LBUG();
3767
3768         case LNET_MATCHMD_OK:
3769                 lnet_recv_put(ni, msg);
3770                 return 0;
3771
3772         case LNET_MATCHMD_NONE:
3773                 if (ready_delay)
3774                         /* no eager_recv or has already called it, should
3775                          * have been attached on delayed list */
3776                         return 0;
3777
3778                 rc = lnet_ni_eager_recv(ni, msg);
3779                 if (rc == 0) {
3780                         ready_delay = true;
3781                         goto again;
3782                 }
3783                 /* fall through */
3784
3785         case LNET_MATCHMD_DROP:
3786                 CNETERR("Dropping PUT from %s portal %d match %llu"
3787                         " offset %d length %d: %d\n",
3788                         libcfs_id2str(info.mi_id), info.mi_portal,
3789                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3790
3791                 return -ENOENT; /* -ve: OK but no match */
3792         }
3793 }
3794
3795 static int
3796 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3797 {
3798         struct lnet_match_info info;
3799         struct lnet_hdr *hdr = &msg->msg_hdr;
3800         struct lnet_process_id source_id;
3801         struct lnet_handle_wire reply_wmd;
3802         int rc;
3803
3804         /* Convert get fields to host byte order */
3805         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3806         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3807         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3808         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3809
3810         source_id.nid = hdr->src_nid;
3811         source_id.pid = hdr->src_pid;
3812         /* Primary peer NID */
3813         info.mi_id.nid  = msg->msg_initiator;
3814         info.mi_id.pid  = hdr->src_pid;
3815         info.mi_opc     = LNET_MD_OP_GET;
3816         info.mi_portal  = hdr->msg.get.ptl_index;
3817         info.mi_rlength = hdr->msg.get.sink_length;
3818         info.mi_roffset = hdr->msg.get.src_offset;
3819         info.mi_mbits   = hdr->msg.get.match_bits;
3820         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3821
3822         rc = lnet_ptl_match_md(&info, msg);
3823         if (rc == LNET_MATCHMD_DROP) {
3824                 CNETERR("Dropping GET from %s portal %d match %llu"
3825                         " offset %d length %d\n",
3826                         libcfs_id2str(info.mi_id), info.mi_portal,
3827                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3828                 return -ENOENT; /* -ve: OK but no match */
3829         }
3830
3831         LASSERT(rc == LNET_MATCHMD_OK);
3832
3833         lnet_build_msg_event(msg, LNET_EVENT_GET);
3834
3835         reply_wmd = hdr->msg.get.return_wmd;
3836
3837         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3838                        msg->msg_offset, msg->msg_wanted);
3839
3840         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3841
3842         if (rdma_get) {
3843                 /* The LND completes the REPLY from her recv procedure */
3844                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3845                              msg->msg_offset, msg->msg_len, msg->msg_len);
3846                 return 0;
3847         }
3848
3849         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3850         msg->msg_receiving = 0;
3851
3852         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3853         if (rc < 0) {
3854                 /* didn't get as far as lnet_ni_send() */
3855                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3856                        libcfs_nid2str(ni->ni_nid),
3857                        libcfs_id2str(info.mi_id), rc);
3858
3859                 lnet_finalize(msg, rc);
3860         }
3861
3862         return 0;
3863 }
3864
3865 static int
3866 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3867 {
3868         void *private = msg->msg_private;
3869         struct lnet_hdr *hdr = &msg->msg_hdr;
3870         struct lnet_process_id src = {0};
3871         struct lnet_libmd *md;
3872         int rlength;
3873         int mlength;
3874         int cpt;
3875
3876         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3877         lnet_res_lock(cpt);
3878
3879         src.nid = hdr->src_nid;
3880         src.pid = hdr->src_pid;
3881
3882         /* NB handles only looked up by creator (no flips) */
3883         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
3884         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3885                 CNETERR("%s: Dropping REPLY from %s for %s "
3886                         "MD %#llx.%#llx\n",
3887                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3888                         (md == NULL) ? "invalid" : "inactive",
3889                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
3890                         hdr->msg.reply.dst_wmd.wh_object_cookie);
3891                 if (md != NULL && md->md_me != NULL)
3892                         CERROR("REPLY MD also attached to portal %d\n",
3893                                md->md_me->me_portal);
3894
3895                 lnet_res_unlock(cpt);
3896                 return -ENOENT; /* -ve: OK but no match */
3897         }
3898
3899         LASSERT(md->md_offset == 0);
3900
3901         rlength = hdr->payload_length;
3902         mlength = MIN(rlength, (int)md->md_length);
3903
3904         if (mlength < rlength &&
3905             (md->md_options & LNET_MD_TRUNCATE) == 0) {
3906                 CNETERR("%s: Dropping REPLY from %s length %d "
3907                         "for MD %#llx would overflow (%d)\n",
3908                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3909                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
3910                         mlength);
3911                 lnet_res_unlock(cpt);
3912                 return -ENOENT; /* -ve: OK but no match */
3913         }
3914
3915         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
3916                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3917                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
3918
3919         lnet_msg_attach_md(msg, md, 0, mlength);
3920
3921         if (mlength != 0)
3922                 lnet_setpayloadbuffer(msg);
3923
3924         lnet_res_unlock(cpt);
3925
3926         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
3927
3928         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
3929         return 0;
3930 }
3931
3932 static int
3933 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
3934 {
3935         struct lnet_hdr *hdr = &msg->msg_hdr;
3936         struct lnet_process_id src = {0};
3937         struct lnet_libmd *md;
3938         int cpt;
3939
3940         src.nid = hdr->src_nid;
3941         src.pid = hdr->src_pid;
3942
3943         /* Convert ack fields to host byte order */
3944         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
3945         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
3946
3947         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
3948         lnet_res_lock(cpt);
3949
3950         /* NB handles only looked up by creator (no flips) */
3951         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
3952         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3953                 /* Don't moan; this is expected */
3954                 CDEBUG(D_NET,
3955                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
3956                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3957                        (md == NULL) ? "invalid" : "inactive",
3958                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
3959                        hdr->msg.ack.dst_wmd.wh_object_cookie);
3960                 if (md != NULL && md->md_me != NULL)
3961                         CERROR("Source MD also attached to portal %d\n",
3962                                md->md_me->me_portal);
3963
3964                 lnet_res_unlock(cpt);
3965                 return -ENOENT;                  /* -ve! */
3966         }
3967
3968         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
3969                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3970                hdr->msg.ack.dst_wmd.wh_object_cookie);
3971
3972         lnet_msg_attach_md(msg, md, 0, 0);
3973
3974         lnet_res_unlock(cpt);
3975
3976         lnet_build_msg_event(msg, LNET_EVENT_ACK);
3977
3978         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
3979         return 0;
3980 }
3981
3982 /**
3983  * \retval LNET_CREDIT_OK       If \a msg is forwarded
3984  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
3985  * \retval -ve                  error code
3986  */
3987 int
3988 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
3989 {
3990         int     rc = 0;
3991
3992         if (!the_lnet.ln_routing)
3993                 return -ECANCELED;
3994
3995         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
3996             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
3997                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
3998                         msg->msg_rx_ready_delay = 1;
3999                 } else {
4000                         lnet_net_unlock(msg->msg_rx_cpt);
4001                         rc = lnet_ni_eager_recv(ni, msg);
4002                         lnet_net_lock(msg->msg_rx_cpt);
4003                 }
4004         }
4005
4006         if (rc == 0)
4007                 rc = lnet_post_routed_recv_locked(msg, 0);
4008         return rc;
4009 }
4010
4011 int
4012 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4013 {
4014         int     rc;
4015
4016         switch (msg->msg_type) {
4017         case LNET_MSG_ACK:
4018                 rc = lnet_parse_ack(ni, msg);
4019                 break;
4020         case LNET_MSG_PUT:
4021                 rc = lnet_parse_put(ni, msg);
4022                 break;
4023         case LNET_MSG_GET:
4024                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4025                 break;
4026         case LNET_MSG_REPLY:
4027                 rc = lnet_parse_reply(ni, msg);
4028                 break;
4029         default: /* prevent an unused label if !kernel */
4030                 LASSERT(0);
4031                 return -EPROTO;
4032         }
4033
4034         LASSERT(rc == 0 || rc == -ENOENT);
4035         return rc;
4036 }
4037
4038 char *
4039 lnet_msgtyp2str (int type)
4040 {
4041         switch (type) {
4042         case LNET_MSG_ACK:
4043                 return ("ACK");
4044         case LNET_MSG_PUT:
4045                 return ("PUT");
4046         case LNET_MSG_GET:
4047                 return ("GET");
4048         case LNET_MSG_REPLY:
4049                 return ("REPLY");
4050         case LNET_MSG_HELLO:
4051                 return ("HELLO");
4052         default:
4053                 return ("<UNKNOWN>");
4054         }
4055 }
4056
4057 void
4058 lnet_print_hdr(struct lnet_hdr *hdr)
4059 {
4060         struct lnet_process_id src = {
4061                 .nid = hdr->src_nid,
4062                 .pid = hdr->src_pid,
4063         };
4064         struct lnet_process_id dst = {
4065                 .nid = hdr->dest_nid,
4066                 .pid = hdr->dest_pid,
4067         };
4068         char *type_str = lnet_msgtyp2str(hdr->type);
4069
4070         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4071         CWARN("    From %s\n", libcfs_id2str(src));
4072         CWARN("    To   %s\n", libcfs_id2str(dst));
4073
4074         switch (hdr->type) {
4075         default:
4076                 break;
4077
4078         case LNET_MSG_PUT:
4079                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4080                       "match bits %llu\n",
4081                       hdr->msg.put.ptl_index,
4082                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4083                       hdr->msg.put.ack_wmd.wh_object_cookie,
4084                       hdr->msg.put.match_bits);
4085                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4086                       hdr->payload_length, hdr->msg.put.offset,
4087                       hdr->msg.put.hdr_data);
4088                 break;
4089
4090         case LNET_MSG_GET:
4091                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4092                       "match bits %llu\n", hdr->msg.get.ptl_index,
4093                       hdr->msg.get.return_wmd.wh_interface_cookie,
4094                       hdr->msg.get.return_wmd.wh_object_cookie,
4095                       hdr->msg.get.match_bits);
4096                 CWARN("    Length %d, src offset %d\n",
4097                       hdr->msg.get.sink_length,
4098                       hdr->msg.get.src_offset);
4099                 break;
4100
4101         case LNET_MSG_ACK:
4102                 CWARN("    dst md %#llx.%#llx, "
4103                       "manipulated length %d\n",
4104                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4105                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4106                       hdr->msg.ack.mlength);
4107                 break;
4108
4109         case LNET_MSG_REPLY:
4110                 CWARN("    dst md %#llx.%#llx, "
4111                       "length %d\n",
4112                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4113                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4114                       hdr->payload_length);
4115         }
4116
4117 }
4118
4119 int
4120 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4121            void *private, int rdma_req)
4122 {
4123         int             rc = 0;
4124         int             cpt;
4125         int             for_me;
4126         struct lnet_msg *msg;
4127         lnet_pid_t     dest_pid;
4128         lnet_nid_t     dest_nid;
4129         lnet_nid_t     src_nid;
4130         struct lnet_peer_ni *lpni;
4131         __u32          payload_length;
4132         __u32          type;
4133
4134         LASSERT (!in_interrupt ());
4135
4136         type = le32_to_cpu(hdr->type);
4137         src_nid = le64_to_cpu(hdr->src_nid);
4138         dest_nid = le64_to_cpu(hdr->dest_nid);
4139         dest_pid = le32_to_cpu(hdr->dest_pid);
4140         payload_length = le32_to_cpu(hdr->payload_length);
4141
4142         for_me = (ni->ni_nid == dest_nid);
4143         cpt = lnet_cpt_of_nid(from_nid, ni);
4144
4145         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4146                 libcfs_nid2str(dest_nid),
4147                 libcfs_nid2str(ni->ni_nid),
4148                 libcfs_nid2str(src_nid),
4149                 lnet_msgtyp2str(type),
4150                 (for_me) ? "for me" : "routed");
4151
4152         switch (type) {
4153         case LNET_MSG_ACK:
4154         case LNET_MSG_GET:
4155                 if (payload_length > 0) {
4156                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4157                                libcfs_nid2str(from_nid),
4158                                libcfs_nid2str(src_nid),
4159                                lnet_msgtyp2str(type), payload_length);
4160                         return -EPROTO;
4161                 }
4162                 break;
4163
4164         case LNET_MSG_PUT:
4165         case LNET_MSG_REPLY:
4166                 if (payload_length >
4167                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4168                         CERROR("%s, src %s: bad %s payload %d "
4169                                "(%d max expected)\n",
4170                                libcfs_nid2str(from_nid),
4171                                libcfs_nid2str(src_nid),
4172                                lnet_msgtyp2str(type),
4173                                payload_length,
4174                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4175                         return -EPROTO;
4176                 }
4177                 break;
4178
4179         default:
4180                 CERROR("%s, src %s: Bad message type 0x%x\n",
4181                        libcfs_nid2str(from_nid),
4182                        libcfs_nid2str(src_nid), type);
4183                 return -EPROTO;
4184         }
4185
4186         if (the_lnet.ln_routing &&
4187             ni->ni_last_alive != ktime_get_real_seconds()) {
4188                 /* NB: so far here is the only place to set NI status to "up */
4189                 lnet_ni_lock(ni);
4190                 ni->ni_last_alive = ktime_get_real_seconds();
4191                 if (ni->ni_status != NULL &&
4192                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
4193                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4194                 lnet_ni_unlock(ni);
4195         }
4196
4197         /* Regard a bad destination NID as a protocol error.  Senders should
4198          * know what they're doing; if they don't they're misconfigured, buggy
4199          * or malicious so we chop them off at the knees :) */
4200
4201         if (!for_me) {
4202                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4203                         /* should have gone direct */
4204                         CERROR("%s, src %s: Bad dest nid %s "
4205                                "(should have been sent direct)\n",
4206                                 libcfs_nid2str(from_nid),
4207                                 libcfs_nid2str(src_nid),
4208                                 libcfs_nid2str(dest_nid));
4209                         return -EPROTO;
4210                 }
4211
4212                 if (lnet_islocalnid(dest_nid)) {
4213                         /* dest is another local NI; sender should have used
4214                          * this node's NID on its own network */
4215                         CERROR("%s, src %s: Bad dest nid %s "
4216                                "(it's my nid but on a different network)\n",
4217                                 libcfs_nid2str(from_nid),
4218                                 libcfs_nid2str(src_nid),
4219                                 libcfs_nid2str(dest_nid));
4220                         return -EPROTO;
4221                 }
4222
4223                 if (rdma_req && type == LNET_MSG_GET) {
4224                         CERROR("%s, src %s: Bad optimized GET for %s "
4225                                "(final destination must be me)\n",
4226                                 libcfs_nid2str(from_nid),
4227                                 libcfs_nid2str(src_nid),
4228                                 libcfs_nid2str(dest_nid));
4229                         return -EPROTO;
4230                 }
4231
4232                 if (!the_lnet.ln_routing) {
4233                         CERROR("%s, src %s: Dropping message for %s "
4234                                "(routing not enabled)\n",
4235                                 libcfs_nid2str(from_nid),
4236                                 libcfs_nid2str(src_nid),
4237                                 libcfs_nid2str(dest_nid));
4238                         goto drop;
4239                 }
4240         }
4241
4242         /* Message looks OK; we're not going to return an error, so we MUST
4243          * call back lnd_recv() come what may... */
4244
4245         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4246             fail_peer(src_nid, 0)) {                    /* shall we now? */
4247                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4248                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4249                        lnet_msgtyp2str(type));
4250                 goto drop;
4251         }
4252
4253         if (!list_empty(&the_lnet.ln_drop_rules) &&
4254             lnet_drop_rule_match(hdr, NULL)) {
4255                 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4256                               "silent message loss\n",
4257                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4258                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4259                 goto drop;
4260         }
4261
4262         if (lnet_drop_asym_route && for_me &&
4263             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4264                 struct lnet_net *net;
4265                 struct lnet_remotenet *rnet;
4266                 bool found = true;
4267
4268                 /* we are dealing with a routed message,
4269                  * so see if route to reach src_nid goes through from_nid
4270                  */
4271                 lnet_net_lock(cpt);
4272                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4273                 if (!net) {
4274                         lnet_net_unlock(cpt);
4275                         CERROR("net %s not found\n",
4276                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4277                         return -EPROTO;
4278                 }
4279
4280                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4281                 if (rnet) {
4282                         struct lnet_peer_ni *gw = NULL;
4283                         struct lnet_route *route;
4284
4285                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4286                                 found = false;
4287                                 gw = route->lr_gateway;
4288                                 if (gw->lpni_net != net)
4289                                         continue;
4290                                 if (gw->lpni_nid == from_nid) {
4291                                         found = true;
4292                                         break;
4293                                 }
4294                         }
4295                 }
4296                 lnet_net_unlock(cpt);
4297                 if (!found) {
4298                         /* we would not use from_nid to route a message to
4299                          * src_nid
4300                          * => asymmetric routing detected but forbidden
4301                          */
4302                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4303                                libcfs_nid2str(from_nid),
4304                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4305                         goto drop;
4306                 }
4307         }
4308
4309         msg = lnet_msg_alloc();
4310         if (msg == NULL) {
4311                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4312                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4313                        lnet_msgtyp2str(type));
4314                 goto drop;
4315         }
4316
4317         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4318          * pointers NULL etc */
4319
4320         msg->msg_type = type;
4321         msg->msg_private = private;
4322         msg->msg_receiving = 1;
4323         msg->msg_rdma_get = rdma_req;
4324         msg->msg_len = msg->msg_wanted = payload_length;
4325         msg->msg_offset = 0;
4326         msg->msg_hdr = *hdr;
4327         /* for building message event */
4328         msg->msg_from = from_nid;
4329         if (!for_me) {
4330                 msg->msg_target.pid     = dest_pid;
4331                 msg->msg_target.nid     = dest_nid;
4332                 msg->msg_routing        = 1;
4333
4334         } else {
4335                 /* convert common msg->hdr fields to host byteorder */
4336                 msg->msg_hdr.type       = type;
4337                 msg->msg_hdr.src_nid    = src_nid;
4338                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4339                 msg->msg_hdr.dest_nid   = dest_nid;
4340                 msg->msg_hdr.dest_pid   = dest_pid;
4341                 msg->msg_hdr.payload_length = payload_length;
4342         }
4343
4344         lnet_net_lock(cpt);
4345         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4346         if (IS_ERR(lpni)) {
4347                 lnet_net_unlock(cpt);
4348                 CERROR("%s, src %s: Dropping %s "
4349                        "(error %ld looking up sender)\n",
4350                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4351                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4352                 lnet_msg_free(msg);
4353                 if (rc == -ESHUTDOWN)
4354                         /* We are shutting down.  Don't do anything more */
4355                         return 0;
4356                 goto drop;
4357         }
4358         msg->msg_rxpeer = lpni;
4359         msg->msg_rxni = ni;
4360         lnet_ni_addref_locked(ni, cpt);
4361         /* Multi-Rail: Primary NID of source. */
4362         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4363
4364         if (lnet_isrouter(msg->msg_rxpeer)) {
4365                 lnet_peer_set_alive(msg->msg_rxpeer);
4366                 if (avoid_asym_router_failure &&
4367                     LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4368                         /* received a remote message from router, update
4369                          * remote NI status on this router.
4370                          * NB: multi-hop routed message will be ignored.
4371                          */
4372                         lnet_router_ni_update_locked(msg->msg_rxpeer,
4373                                                      LNET_NIDNET(src_nid));
4374                 }
4375         }
4376
4377         lnet_msg_commit(msg, cpt);
4378
4379         /* message delay simulation */
4380         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4381                      lnet_delay_rule_match_locked(hdr, msg))) {
4382                 lnet_net_unlock(cpt);
4383                 return 0;
4384         }
4385
4386         if (!for_me) {
4387                 rc = lnet_parse_forward_locked(ni, msg);
4388                 lnet_net_unlock(cpt);
4389
4390                 if (rc < 0)
4391                         goto free_drop;
4392
4393                 if (rc == LNET_CREDIT_OK) {
4394                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4395                                      0, payload_length, payload_length);
4396                 }
4397                 return 0;
4398         }
4399
4400         lnet_net_unlock(cpt);
4401
4402         rc = lnet_parse_local(ni, msg);
4403         if (rc != 0)
4404                 goto free_drop;
4405         return 0;
4406
4407  free_drop:
4408         LASSERT(msg->msg_md == NULL);
4409         lnet_finalize(msg, rc);
4410
4411  drop:
4412         lnet_drop_message(ni, cpt, private, payload_length, type);
4413         return 0;
4414 }
4415 EXPORT_SYMBOL(lnet_parse);
4416
4417 void
4418 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4419 {
4420         while (!list_empty(head)) {
4421                 struct lnet_process_id id = {0};
4422                 struct lnet_msg *msg;
4423
4424                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4425                 list_del(&msg->msg_list);
4426
4427                 id.nid = msg->msg_hdr.src_nid;
4428                 id.pid = msg->msg_hdr.src_pid;
4429
4430                 LASSERT(msg->msg_md == NULL);
4431                 LASSERT(msg->msg_rx_delayed);
4432                 LASSERT(msg->msg_rxpeer != NULL);
4433                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4434
4435                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4436                       " offset %d length %d: %s\n",
4437                       libcfs_id2str(id),
4438                       msg->msg_hdr.msg.put.ptl_index,
4439                       msg->msg_hdr.msg.put.match_bits,
4440                       msg->msg_hdr.msg.put.offset,
4441                       msg->msg_hdr.payload_length, reason);
4442
4443                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4444                  * called lnet_drop_message(), so I just hang onto msg as well
4445                  * until that's done */
4446
4447                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4448                                   msg->msg_private, msg->msg_len,
4449                                   msg->msg_type);
4450
4451                 msg->msg_no_resend = true;
4452                 /*
4453                  * NB: message will not generate event because w/o attached MD,
4454                  * but we still should give error code so lnet_msg_decommit()
4455                  * can skip counters operations and other checks.
4456                  */
4457                 lnet_finalize(msg, -ENOENT);
4458         }
4459 }
4460
4461 void
4462 lnet_recv_delayed_msg_list(struct list_head *head)
4463 {
4464         while (!list_empty(head)) {
4465                 struct lnet_msg *msg;
4466                 struct lnet_process_id id;
4467
4468                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4469                 list_del(&msg->msg_list);
4470
4471                 /* md won't disappear under me, since each msg
4472                  * holds a ref on it */
4473
4474                 id.nid = msg->msg_hdr.src_nid;
4475                 id.pid = msg->msg_hdr.src_pid;
4476
4477                 LASSERT(msg->msg_rx_delayed);
4478                 LASSERT(msg->msg_md != NULL);
4479                 LASSERT(msg->msg_rxpeer != NULL);
4480                 LASSERT(msg->msg_rxni != NULL);
4481                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4482
4483                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4484                        "match %llu offset %d length %d.\n",
4485                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4486                         msg->msg_hdr.msg.put.match_bits,
4487                         msg->msg_hdr.msg.put.offset,
4488                         msg->msg_hdr.payload_length);
4489
4490                 lnet_recv_put(msg->msg_rxni, msg);
4491         }
4492 }
4493
4494 static void
4495 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4496                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4497 {
4498         s64 timeout_ns;
4499         bool new_entry = true;
4500         struct lnet_rsp_tracker *local_rspt;
4501
4502         /*
4503          * MD has a refcount taken by message so it's not going away.
4504          * The MD however can be looked up. We need to secure the access
4505          * to the md_rspt_ptr by taking the res_lock.
4506          * The rspt can be accessed without protection up to when it gets
4507          * added to the list.
4508          */
4509
4510         lnet_res_lock(cpt);
4511         local_rspt = md->md_rspt_ptr;
4512         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4513         if (local_rspt != NULL) {
4514                 /*
4515                  * we already have an rspt attached to the md, so we'll
4516                  * update the deadline on that one.
4517                  */
4518                 LIBCFS_FREE(rspt, sizeof(*rspt));
4519                 new_entry = false;
4520         } else {
4521                 /* new md */
4522                 rspt->rspt_mdh = mdh;
4523                 rspt->rspt_cpt = cpt;
4524                 /* store the rspt so we can access it when we get the REPLY */
4525                 md->md_rspt_ptr = rspt;
4526                 local_rspt = rspt;
4527         }
4528         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4529
4530         /*
4531          * add to the list of tracked responses. It's added to tail of the
4532          * list in order to expire all the older entries first.
4533          */
4534         lnet_net_lock(cpt);
4535         if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
4536                 list_del_init(&local_rspt->rspt_on_list);
4537         list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4538         lnet_net_unlock(cpt);
4539         lnet_res_unlock(cpt);
4540 }
4541
4542 /**
4543  * Initiate an asynchronous PUT operation.
4544  *
4545  * There are several events associated with a PUT: completion of the send on
4546  * the initiator node (LNET_EVENT_SEND), and when the send completes
4547  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4548  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4549  * used at the target node to indicate the completion of incoming data
4550  * delivery.
4551  *
4552  * The local events will be logged in the EQ associated with the MD pointed to
4553  * by \a mdh handle. Using a MD without an associated EQ results in these
4554  * events being discarded. In this case, the caller must have another
4555  * mechanism (e.g., a higher level protocol) for determining when it is safe
4556  * to modify the memory region associated with the MD.
4557  *
4558  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4559  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4560  *
4561  * \param self Indicates the NID of a local interface through which to send
4562  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4563  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4564  * must be "free floating" (See LNetMDBind()).
4565  * \param ack Controls whether an acknowledgment is requested.
4566  * Acknowledgments are only sent when they are requested by the initiating
4567  * process and the target MD enables them.
4568  * \param target A process identifier for the target process.
4569  * \param portal The index in the \a target's portal table.
4570  * \param match_bits The match bits to use for MD selection at the target
4571  * process.
4572  * \param offset The offset into the target MD (only used when the target
4573  * MD has the LNET_MD_MANAGE_REMOTE option set).
4574  * \param hdr_data 64 bits of user data that can be included in the message
4575  * header. This data is written to an event queue entry at the target if an
4576  * EQ is present on the matching MD.
4577  *
4578  * \retval  0      Success, and only in this case events will be generated
4579  * and logged to EQ (if it exists).
4580  * \retval -EIO    Simulated failure.
4581  * \retval -ENOMEM Memory allocation failure.
4582  * \retval -ENOENT Invalid MD object.
4583  *
4584  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4585  */
4586 int
4587 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4588         struct lnet_process_id target, unsigned int portal,
4589         __u64 match_bits, unsigned int offset,
4590         __u64 hdr_data)
4591 {
4592         struct lnet_msg *msg;
4593         struct lnet_libmd *md;
4594         int cpt;
4595         int rc;
4596         struct lnet_rsp_tracker *rspt = NULL;
4597
4598         LASSERT(the_lnet.ln_refcount > 0);
4599
4600         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4601             fail_peer(target.nid, 1)) {                 /* shall we now? */
4602                 CERROR("Dropping PUT to %s: simulated failure\n",
4603                        libcfs_id2str(target));
4604                 return -EIO;
4605         }
4606
4607         msg = lnet_msg_alloc();
4608         if (msg == NULL) {
4609                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4610                        libcfs_id2str(target));
4611                 return -ENOMEM;
4612         }
4613         msg->msg_vmflush = !!memory_pressure_get();
4614
4615         cpt = lnet_cpt_of_cookie(mdh.cookie);
4616
4617         if (ack == LNET_ACK_REQ) {
4618                 rspt = lnet_rspt_alloc(cpt);
4619                 if (!rspt) {
4620                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4621                                 libcfs_id2str(target));
4622                         return -ENOMEM;
4623                 }
4624                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4625         }
4626
4627         lnet_res_lock(cpt);
4628
4629         md = lnet_handle2md(&mdh);
4630         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4631                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4632                        match_bits, portal, libcfs_id2str(target),
4633                        md == NULL ? -1 : md->md_threshold);
4634                 if (md != NULL && md->md_me != NULL)
4635                         CERROR("Source MD also attached to portal %d\n",
4636                                md->md_me->me_portal);
4637                 lnet_res_unlock(cpt);
4638
4639                 LIBCFS_FREE(rspt, sizeof(*rspt));
4640                 lnet_msg_free(msg);
4641                 return -ENOENT;
4642         }
4643
4644         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4645
4646         lnet_msg_attach_md(msg, md, 0, 0);
4647
4648         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4649
4650         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4651         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4652         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4653         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4654
4655         /* NB handles only looked up by creator (no flips) */
4656         if (ack == LNET_ACK_REQ) {
4657                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4658                         the_lnet.ln_interface_cookie;
4659                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4660                         md->md_lh.lh_cookie;
4661         } else {
4662                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4663                         LNET_WIRE_HANDLE_COOKIE_NONE;
4664                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4665                         LNET_WIRE_HANDLE_COOKIE_NONE;
4666         }
4667
4668         lnet_res_unlock(cpt);
4669
4670         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4671
4672         if (ack == LNET_ACK_REQ)
4673                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4674
4675         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4676                                  CFS_FAIL_ONCE))
4677                 rc = -EIO;
4678         else
4679                 rc = lnet_send(self, msg, LNET_NID_ANY);
4680
4681         if (rc != 0) {
4682                 CNETERR("Error sending PUT to %s: %d\n",
4683                         libcfs_id2str(target), rc);
4684                 msg->msg_no_resend = true;
4685                 lnet_finalize(msg, rc);
4686         }
4687
4688         /* completion will be signalled by an event */
4689         return 0;
4690 }
4691 EXPORT_SYMBOL(LNetPut);
4692
4693 /*
4694  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4695  * returns a msg for the LND to pass to lnet_finalize() when the sink
4696  * data has been received.
4697  *
4698  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4699  * lnet_finalize() is called on it, so the LND must call this first
4700  */
4701 struct lnet_msg *
4702 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4703 {
4704         struct lnet_msg *msg = lnet_msg_alloc();
4705         struct lnet_libmd *getmd = getmsg->msg_md;
4706         struct lnet_process_id peer_id = getmsg->msg_target;
4707         int cpt;
4708
4709         LASSERT(!getmsg->msg_target_is_router);
4710         LASSERT(!getmsg->msg_routing);
4711
4712         if (msg == NULL) {
4713                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4714                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4715                 goto drop;
4716         }
4717
4718         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4719         lnet_res_lock(cpt);
4720
4721         LASSERT(getmd->md_refcount > 0);
4722
4723         if (getmd->md_threshold == 0) {
4724                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4725                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4726                         getmd);
4727                 lnet_res_unlock(cpt);
4728                 goto drop;
4729         }
4730
4731         LASSERT(getmd->md_offset == 0);
4732
4733         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4734                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4735
4736         /* setup information for lnet_build_msg_event */
4737         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4738         msg->msg_from = peer_id.nid;
4739         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4740         msg->msg_hdr.src_nid = peer_id.nid;
4741         msg->msg_hdr.payload_length = getmd->md_length;
4742         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4743
4744         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4745         lnet_res_unlock(cpt);
4746
4747         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4748
4749         lnet_net_lock(cpt);
4750         lnet_msg_commit(msg, cpt);
4751         lnet_net_unlock(cpt);
4752
4753         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4754
4755         return msg;
4756
4757  drop:
4758         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4759
4760         lnet_net_lock(cpt);
4761         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4762         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4763         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4764                 getmd->md_length;
4765         lnet_net_unlock(cpt);
4766
4767         if (msg != NULL)
4768                 lnet_msg_free(msg);
4769
4770         return NULL;
4771 }
4772 EXPORT_SYMBOL(lnet_create_reply_msg);
4773
4774 void
4775 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4776                        unsigned int len)
4777 {
4778         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4779          * completed and I know it. */
4780         LASSERT(reply != NULL);
4781         LASSERT(reply->msg_type == LNET_MSG_GET);
4782         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4783
4784         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4785          * the end of my buffer, I might as well be dead. */
4786         LASSERT(len <= reply->msg_ev.mlength);
4787
4788         reply->msg_ev.mlength = len;
4789 }
4790 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4791
4792 /**
4793  * Initiate an asynchronous GET operation.
4794  *
4795  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4796  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4797  * the target node in the REPLY has been written to local MD.
4798  *
4799  * On the target node, an LNET_EVENT_GET is logged when the GET request
4800  * arrives and is accepted into a MD.
4801  *
4802  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4803  * \param mdh A handle for the MD that describes the memory into which the
4804  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4805  *
4806  * \retval  0      Success, and only in this case events will be generated
4807  * and logged to EQ (if it exists) of the MD.
4808  * \retval -EIO    Simulated failure.
4809  * \retval -ENOMEM Memory allocation failure.
4810  * \retval -ENOENT Invalid MD object.
4811  */
4812 int
4813 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4814         struct lnet_process_id target, unsigned int portal,
4815         __u64 match_bits, unsigned int offset, bool recovery)
4816 {
4817         struct lnet_msg *msg;
4818         struct lnet_libmd *md;
4819         struct lnet_rsp_tracker *rspt;
4820         int cpt;
4821         int rc;
4822
4823         LASSERT(the_lnet.ln_refcount > 0);
4824
4825         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4826             fail_peer(target.nid, 1))                   /* shall we now? */
4827         {
4828                 CERROR("Dropping GET to %s: simulated failure\n",
4829                        libcfs_id2str(target));
4830                 return -EIO;
4831         }
4832
4833         msg = lnet_msg_alloc();
4834         if (!msg) {
4835                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4836                        libcfs_id2str(target));
4837                 return -ENOMEM;
4838         }
4839
4840         cpt = lnet_cpt_of_cookie(mdh.cookie);
4841
4842         rspt = lnet_rspt_alloc(cpt);
4843         if (!rspt) {
4844                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4845                        libcfs_id2str(target));
4846                 return -ENOMEM;
4847         }
4848         INIT_LIST_HEAD(&rspt->rspt_on_list);
4849
4850         msg->msg_recovery = recovery;
4851
4852         lnet_res_lock(cpt);
4853
4854         md = lnet_handle2md(&mdh);
4855         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4856                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4857                        match_bits, portal, libcfs_id2str(target),
4858                        md == NULL ? -1 : md->md_threshold);
4859                 if (md != NULL && md->md_me != NULL)
4860                         CERROR("REPLY MD also attached to portal %d\n",
4861                                md->md_me->me_portal);
4862
4863                 lnet_res_unlock(cpt);
4864
4865                 lnet_msg_free(msg);
4866                 LIBCFS_FREE(rspt, sizeof(*rspt));
4867                 return -ENOENT;
4868         }
4869
4870         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4871
4872         lnet_msg_attach_md(msg, md, 0, 0);
4873
4874         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
4875
4876         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
4877         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
4878         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
4879         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
4880
4881         /* NB handles only looked up by creator (no flips) */
4882         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
4883                 the_lnet.ln_interface_cookie;
4884         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
4885                 md->md_lh.lh_cookie;
4886
4887         lnet_res_unlock(cpt);
4888
4889         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4890
4891         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4892
4893         rc = lnet_send(self, msg, LNET_NID_ANY);
4894         if (rc < 0) {
4895                 CNETERR("Error sending GET to %s: %d\n",
4896                         libcfs_id2str(target), rc);
4897                 msg->msg_no_resend = true;
4898                 lnet_finalize(msg, rc);
4899         }
4900
4901         /* completion will be signalled by an event */
4902         return 0;
4903 }
4904 EXPORT_SYMBOL(LNetGet);
4905
4906 /**
4907  * Calculate distance to node at \a dstnid.
4908  *
4909  * \param dstnid Target NID.
4910  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
4911  * is saved here.
4912  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
4913  * here.
4914  *
4915  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
4916  * local_nid_dist_zero is set, which is the default.
4917  * \retval positives Distance to target NID, i.e. number of hops plus one.
4918  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
4919  */
4920 int
4921 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
4922 {
4923         struct list_head        *e;
4924         struct lnet_ni *ni = NULL;
4925         struct lnet_remotenet *rnet;
4926         __u32                   dstnet = LNET_NIDNET(dstnid);
4927         int                     hops;
4928         int                     cpt;
4929         __u32                   order = 2;
4930         struct list_head        *rn_list;
4931
4932         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
4933          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
4934          * keep order 0 free for 0@lo and order 1 free for a local NID
4935          * match */
4936
4937         LASSERT(the_lnet.ln_refcount > 0);
4938
4939         cpt = lnet_net_lock_current();
4940
4941         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
4942                 if (ni->ni_nid == dstnid) {
4943                         if (srcnidp != NULL)
4944                                 *srcnidp = dstnid;
4945                         if (orderp != NULL) {
4946                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
4947                                         *orderp = 0;
4948                                 else
4949                                         *orderp = 1;
4950                         }
4951                         lnet_net_unlock(cpt);
4952
4953                         return local_nid_dist_zero ? 0 : 1;
4954                 }
4955
4956                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
4957                         /* Check if ni was originally created in
4958                          * current net namespace.
4959                          * If not, assign order above 0xffff0000,
4960                          * to make this ni not a priority. */
4961                         if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
4962                                 order += 0xffff0000;
4963
4964                         if (srcnidp != NULL)
4965                                 *srcnidp = ni->ni_nid;
4966                         if (orderp != NULL)
4967                                 *orderp = order;
4968                         lnet_net_unlock(cpt);
4969                         return 1;
4970                 }
4971
4972                 order++;
4973         }
4974
4975         rn_list = lnet_net2rnethash(dstnet);
4976         list_for_each(e, rn_list) {
4977                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
4978
4979                 if (rnet->lrn_net == dstnet) {
4980                         struct lnet_route *route;
4981                         struct lnet_route *shortest = NULL;
4982                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
4983                         __u32 route_hops;
4984
4985                         LASSERT(!list_empty(&rnet->lrn_routes));
4986
4987                         list_for_each_entry(route, &rnet->lrn_routes,
4988                                             lr_list) {
4989                                 route_hops = route->lr_hops;
4990                                 if (route_hops == LNET_UNDEFINED_HOPS)
4991                                         route_hops = 1;
4992                                 if (shortest == NULL ||
4993                                     route_hops < shortest_hops) {
4994                                         shortest = route;
4995                                         shortest_hops = route_hops;
4996                                 }
4997                         }
4998
4999                         LASSERT(shortest != NULL);
5000                         hops = shortest_hops;
5001                         if (srcnidp != NULL) {
5002                                 ni = lnet_get_next_ni_locked(
5003                                         shortest->lr_gateway->lpni_net,
5004                                         NULL);
5005                                 *srcnidp = ni->ni_nid;
5006                         }
5007                         if (orderp != NULL)
5008                                 *orderp = order;
5009                         lnet_net_unlock(cpt);
5010                         return hops + 1;
5011                 }
5012                 order++;
5013         }
5014
5015         lnet_net_unlock(cpt);
5016         return -EHOSTUNREACH;
5017 }
5018 EXPORT_SYMBOL(LNetDist);