Whamcloud - gitweb
3e2624cb5db784381a240edda3cdf64696a520fe
[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
942         /* NB 'lp' is always the next hop */
943         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
944             lnet_peer_alive_locked(ni, lp, msg) == 0) {
945                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
946                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
947                         msg->msg_len;
948                 lnet_net_unlock(cpt);
949                 if (msg->msg_txpeer)
950                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
951                                         msg->msg_type,
952                                         LNET_STATS_TYPE_DROP);
953                 if (msg->msg_txni)
954                         lnet_incr_stats(&msg->msg_txni->ni_stats,
955                                         msg->msg_type,
956                                         LNET_STATS_TYPE_DROP);
957
958                 CNETERR("Dropping message for %s: peer not alive\n",
959                         libcfs_id2str(msg->msg_target));
960                 msg->msg_health_status = LNET_MSG_STATUS_LOCAL_DROPPED;
961                 if (do_send)
962                         lnet_finalize(msg, -EHOSTUNREACH);
963
964                 lnet_net_lock(cpt);
965                 return -EHOSTUNREACH;
966         }
967
968         if (msg->msg_md != NULL &&
969             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
970                 lnet_net_unlock(cpt);
971
972                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
973                         "called on the MD/ME.\n",
974                         libcfs_id2str(msg->msg_target));
975                 if (do_send) {
976                         msg->msg_no_resend = true;
977                         lnet_finalize(msg, -ECANCELED);
978                 }
979
980                 lnet_net_lock(cpt);
981                 return -ECANCELED;
982         }
983
984         if (!msg->msg_peertxcredit) {
985                 spin_lock(&lp->lpni_lock);
986                 LASSERT((lp->lpni_txcredits < 0) ==
987                         !list_empty(&lp->lpni_txq));
988
989                 msg->msg_peertxcredit = 1;
990                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
991                 lp->lpni_txcredits--;
992
993                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
994                         lp->lpni_mintxcredits = lp->lpni_txcredits;
995
996                 if (lp->lpni_txcredits < 0) {
997                         msg->msg_tx_delayed = 1;
998                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
999                         spin_unlock(&lp->lpni_lock);
1000                         return LNET_CREDIT_WAIT;
1001                 }
1002                 spin_unlock(&lp->lpni_lock);
1003         }
1004
1005         if (!msg->msg_txcredit) {
1006                 LASSERT((tq->tq_credits < 0) ==
1007                         !list_empty(&tq->tq_delayed));
1008
1009                 msg->msg_txcredit = 1;
1010                 tq->tq_credits--;
1011                 atomic_dec(&ni->ni_tx_credits);
1012
1013                 if (tq->tq_credits < tq->tq_credits_min)
1014                         tq->tq_credits_min = tq->tq_credits;
1015
1016                 if (tq->tq_credits < 0) {
1017                         msg->msg_tx_delayed = 1;
1018                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
1019                         return LNET_CREDIT_WAIT;
1020                 }
1021         }
1022
1023         /* unset the tx_delay flag as we're going to send it now */
1024         msg->msg_tx_delayed = 0;
1025
1026         if (do_send) {
1027                 lnet_net_unlock(cpt);
1028                 lnet_ni_send(ni, msg);
1029                 lnet_net_lock(cpt);
1030         }
1031         return LNET_CREDIT_OK;
1032 }
1033
1034
1035 static struct lnet_rtrbufpool *
1036 lnet_msg2bufpool(struct lnet_msg *msg)
1037 {
1038         struct lnet_rtrbufpool  *rbp;
1039         int                     cpt;
1040
1041         LASSERT(msg->msg_rx_committed);
1042
1043         cpt = msg->msg_rx_cpt;
1044         rbp = &the_lnet.ln_rtrpools[cpt][0];
1045
1046         LASSERT(msg->msg_len <= LNET_MTU);
1047         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
1048                 rbp++;
1049                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
1050         }
1051
1052         return rbp;
1053 }
1054
1055 static int
1056 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
1057 {
1058         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
1059          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
1060          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
1061          * received or OK to receive */
1062         struct lnet_peer_ni *lp = msg->msg_rxpeer;
1063         struct lnet_rtrbufpool *rbp;
1064         struct lnet_rtrbuf *rb;
1065
1066         LASSERT (msg->msg_iov == NULL);
1067         LASSERT (msg->msg_kiov == NULL);
1068         LASSERT (msg->msg_niov == 0);
1069         LASSERT (msg->msg_routing);
1070         LASSERT (msg->msg_receiving);
1071         LASSERT (!msg->msg_sending);
1072
1073         /* non-lnet_parse callers only receive delayed messages */
1074         LASSERT(!do_recv || msg->msg_rx_delayed);
1075
1076         if (!msg->msg_peerrtrcredit) {
1077                 spin_lock(&lp->lpni_lock);
1078                 LASSERT((lp->lpni_rtrcredits < 0) ==
1079                         !list_empty(&lp->lpni_rtrq));
1080
1081                 msg->msg_peerrtrcredit = 1;
1082                 lp->lpni_rtrcredits--;
1083                 if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
1084                         lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
1085
1086                 if (lp->lpni_rtrcredits < 0) {
1087                         /* must have checked eager_recv before here */
1088                         LASSERT(msg->msg_rx_ready_delay);
1089                         msg->msg_rx_delayed = 1;
1090                         list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
1091                         spin_unlock(&lp->lpni_lock);
1092                         return LNET_CREDIT_WAIT;
1093                 }
1094                 spin_unlock(&lp->lpni_lock);
1095         }
1096
1097         rbp = lnet_msg2bufpool(msg);
1098
1099         if (!msg->msg_rtrcredit) {
1100                 msg->msg_rtrcredit = 1;
1101                 rbp->rbp_credits--;
1102                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1103                         rbp->rbp_mincredits = rbp->rbp_credits;
1104
1105                 if (rbp->rbp_credits < 0) {
1106                         /* must have checked eager_recv before here */
1107                         LASSERT(msg->msg_rx_ready_delay);
1108                         msg->msg_rx_delayed = 1;
1109                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1110                         return LNET_CREDIT_WAIT;
1111                 }
1112         }
1113
1114         LASSERT(!list_empty(&rbp->rbp_bufs));
1115         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1116         list_del(&rb->rb_list);
1117
1118         msg->msg_niov = rbp->rbp_npages;
1119         msg->msg_kiov = &rb->rb_kiov[0];
1120
1121         /* unset the msg-rx_delayed flag since we're receiving the message */
1122         msg->msg_rx_delayed = 0;
1123
1124         if (do_recv) {
1125                 int cpt = msg->msg_rx_cpt;
1126
1127                 lnet_net_unlock(cpt);
1128                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1129                              0, msg->msg_len, msg->msg_len);
1130                 lnet_net_lock(cpt);
1131         }
1132         return LNET_CREDIT_OK;
1133 }
1134
1135 void
1136 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1137 {
1138         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1139         struct lnet_ni          *txni = msg->msg_txni;
1140         struct lnet_msg         *msg2;
1141
1142         if (msg->msg_txcredit) {
1143                 struct lnet_ni       *ni = msg->msg_txni;
1144                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1145
1146                 /* give back NI txcredits */
1147                 msg->msg_txcredit = 0;
1148
1149                 LASSERT((tq->tq_credits < 0) ==
1150                         !list_empty(&tq->tq_delayed));
1151
1152                 tq->tq_credits++;
1153                 atomic_inc(&ni->ni_tx_credits);
1154                 if (tq->tq_credits <= 0) {
1155                         msg2 = list_entry(tq->tq_delayed.next,
1156                                           struct lnet_msg, msg_list);
1157                         list_del(&msg2->msg_list);
1158
1159                         LASSERT(msg2->msg_txni == ni);
1160                         LASSERT(msg2->msg_tx_delayed);
1161                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1162
1163                         (void) lnet_post_send_locked(msg2, 1);
1164                 }
1165         }
1166
1167         if (msg->msg_peertxcredit) {
1168                 /* give back peer txcredits */
1169                 msg->msg_peertxcredit = 0;
1170
1171                 spin_lock(&txpeer->lpni_lock);
1172                 LASSERT((txpeer->lpni_txcredits < 0) ==
1173                         !list_empty(&txpeer->lpni_txq));
1174
1175                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1176                 LASSERT(txpeer->lpni_txqnob >= 0);
1177
1178                 txpeer->lpni_txcredits++;
1179                 if (txpeer->lpni_txcredits <= 0) {
1180                         int msg2_cpt;
1181
1182                         msg2 = list_entry(txpeer->lpni_txq.next,
1183                                               struct lnet_msg, msg_list);
1184                         list_del(&msg2->msg_list);
1185                         spin_unlock(&txpeer->lpni_lock);
1186
1187                         LASSERT(msg2->msg_txpeer == txpeer);
1188                         LASSERT(msg2->msg_tx_delayed);
1189
1190                         msg2_cpt = msg2->msg_tx_cpt;
1191
1192                         /*
1193                          * The msg_cpt can be different from the msg2_cpt
1194                          * so we need to make sure we lock the correct cpt
1195                          * for msg2.
1196                          * Once we call lnet_post_send_locked() it is no
1197                          * longer safe to access msg2, since it could've
1198                          * been freed by lnet_finalize(), but we still
1199                          * need to relock the correct cpt, so we cache the
1200                          * msg2_cpt for the purpose of the check that
1201                          * follows the call to lnet_pose_send_locked().
1202                          */
1203                         if (msg2_cpt != msg->msg_tx_cpt) {
1204                                 lnet_net_unlock(msg->msg_tx_cpt);
1205                                 lnet_net_lock(msg2_cpt);
1206                         }
1207                         (void) lnet_post_send_locked(msg2, 1);
1208                         if (msg2_cpt != msg->msg_tx_cpt) {
1209                                 lnet_net_unlock(msg2_cpt);
1210                                 lnet_net_lock(msg->msg_tx_cpt);
1211                         }
1212                 } else {
1213                         spin_unlock(&txpeer->lpni_lock);
1214                 }
1215         }
1216
1217         if (txni != NULL) {
1218                 msg->msg_txni = NULL;
1219                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1220         }
1221
1222         if (txpeer != NULL) {
1223                 msg->msg_txpeer = NULL;
1224                 lnet_peer_ni_decref_locked(txpeer);
1225         }
1226 }
1227
1228 void
1229 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1230 {
1231         struct lnet_msg *msg;
1232
1233         if (list_empty(&rbp->rbp_msgs))
1234                 return;
1235         msg = list_entry(rbp->rbp_msgs.next,
1236                          struct lnet_msg, msg_list);
1237         list_del(&msg->msg_list);
1238
1239         (void)lnet_post_routed_recv_locked(msg, 1);
1240 }
1241
1242 void
1243 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1244 {
1245         struct lnet_msg *msg;
1246         struct lnet_msg *tmp;
1247
1248         lnet_net_unlock(cpt);
1249
1250         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1251                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1252                              0, 0, 0, msg->msg_hdr.payload_length);
1253                 list_del_init(&msg->msg_list);
1254                 msg->msg_no_resend = true;
1255                 lnet_finalize(msg, -ECANCELED);
1256         }
1257
1258         lnet_net_lock(cpt);
1259 }
1260
1261 void
1262 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1263 {
1264         struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
1265         struct lnet_ni *rxni = msg->msg_rxni;
1266         struct lnet_msg *msg2;
1267
1268         if (msg->msg_rtrcredit) {
1269                 /* give back global router credits */
1270                 struct lnet_rtrbuf *rb;
1271                 struct lnet_rtrbufpool *rbp;
1272
1273                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1274                  * there until it gets one allocated, or aborts the wait
1275                  * itself */
1276                 LASSERT(msg->msg_kiov != NULL);
1277
1278                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1279                 rbp = rb->rb_pool;
1280
1281                 msg->msg_kiov = NULL;
1282                 msg->msg_rtrcredit = 0;
1283
1284                 LASSERT(rbp == lnet_msg2bufpool(msg));
1285
1286                 LASSERT((rbp->rbp_credits > 0) ==
1287                         !list_empty(&rbp->rbp_bufs));
1288
1289                 /* If routing is now turned off, we just drop this buffer and
1290                  * don't bother trying to return credits.  */
1291                 if (!the_lnet.ln_routing) {
1292                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1293                         goto routing_off;
1294                 }
1295
1296                 /* It is possible that a user has lowered the desired number of
1297                  * buffers in this pool.  Make sure we never put back
1298                  * more buffers than the stated number. */
1299                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1300                         /* Discard this buffer so we don't have too
1301                          * many. */
1302                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1303                         rbp->rbp_nbuffers--;
1304                 } else {
1305                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1306                         rbp->rbp_credits++;
1307                         if (rbp->rbp_credits <= 0)
1308                                 lnet_schedule_blocked_locked(rbp);
1309                 }
1310         }
1311
1312 routing_off:
1313         if (msg->msg_peerrtrcredit) {
1314                 /* give back peer router credits */
1315                 msg->msg_peerrtrcredit = 0;
1316
1317                 spin_lock(&rxpeer->lpni_lock);
1318                 LASSERT((rxpeer->lpni_rtrcredits < 0) ==
1319                         !list_empty(&rxpeer->lpni_rtrq));
1320
1321                 rxpeer->lpni_rtrcredits++;
1322
1323                 /* drop all messages which are queued to be routed on that
1324                  * peer. */
1325                 if (!the_lnet.ln_routing) {
1326                         struct list_head drop;
1327                         INIT_LIST_HEAD(&drop);
1328                         list_splice_init(&rxpeer->lpni_rtrq, &drop);
1329                         spin_unlock(&rxpeer->lpni_lock);
1330                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1331                 } else if (rxpeer->lpni_rtrcredits <= 0) {
1332                         msg2 = list_entry(rxpeer->lpni_rtrq.next,
1333                                           struct lnet_msg, msg_list);
1334                         list_del(&msg2->msg_list);
1335                         spin_unlock(&rxpeer->lpni_lock);
1336                         (void) lnet_post_routed_recv_locked(msg2, 1);
1337                 } else {
1338                         spin_unlock(&rxpeer->lpni_lock);
1339                 }
1340         }
1341         if (rxni != NULL) {
1342                 msg->msg_rxni = NULL;
1343                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1344         }
1345         if (rxpeer != NULL) {
1346                 msg->msg_rxpeer = NULL;
1347                 lnet_peer_ni_decref_locked(rxpeer);
1348         }
1349 }
1350
1351 static int
1352 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1353 {
1354         if (p1->lpni_txqnob < p2->lpni_txqnob)
1355                 return 1;
1356
1357         if (p1->lpni_txqnob > p2->lpni_txqnob)
1358                 return -1;
1359
1360         if (p1->lpni_txcredits > p2->lpni_txcredits)
1361                 return 1;
1362
1363         if (p1->lpni_txcredits < p2->lpni_txcredits)
1364                 return -1;
1365
1366         return 0;
1367 }
1368
1369 static int
1370 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1371 {
1372         struct lnet_peer_ni *p1 = r1->lr_gateway;
1373         struct lnet_peer_ni *p2 = r2->lr_gateway;
1374         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1375         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1376         int rc;
1377
1378         if (r1->lr_priority < r2->lr_priority)
1379                 return 1;
1380
1381         if (r1->lr_priority > r2->lr_priority)
1382                 return -1;
1383
1384         if (r1_hops < r2_hops)
1385                 return 1;
1386
1387         if (r1_hops > r2_hops)
1388                 return -1;
1389
1390         rc = lnet_compare_peers(p1, p2);
1391         if (rc)
1392                 return rc;
1393
1394         if (r1->lr_seq - r2->lr_seq <= 0)
1395                 return 1;
1396
1397         return -1;
1398 }
1399
1400 static struct lnet_peer_ni *
1401 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1402                        lnet_nid_t rtr_nid)
1403 {
1404         struct lnet_remotenet   *rnet;
1405         struct lnet_route               *route;
1406         struct lnet_route               *best_route;
1407         struct lnet_route               *last_route;
1408         struct lnet_peer_ni     *lpni_best;
1409         struct lnet_peer_ni     *lp;
1410         int                     rc;
1411
1412         /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1413          * rtr_nid nid, otherwise find the best gateway I can use */
1414
1415         rnet = lnet_find_rnet_locked(remote_net);
1416         if (rnet == NULL)
1417                 return NULL;
1418
1419         lpni_best = NULL;
1420         best_route = last_route = NULL;
1421         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1422                 lp = route->lr_gateway;
1423
1424                 if (!lnet_is_route_alive(route))
1425                         continue;
1426
1427                 if (net != NULL && lp->lpni_net != net)
1428                         continue;
1429
1430                 if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
1431                         return lp;
1432
1433                 if (lpni_best == NULL) {
1434                         best_route = last_route = route;
1435                         lpni_best = lp;
1436                         continue;
1437                 }
1438
1439                 /* no protection on below fields, but it's harmless */
1440                 if (last_route->lr_seq - route->lr_seq < 0)
1441                         last_route = route;
1442
1443                 rc = lnet_compare_routes(route, best_route);
1444                 if (rc < 0)
1445                         continue;
1446
1447                 best_route = route;
1448                 lpni_best = lp;
1449         }
1450
1451         /* set sequence number on the best router to the latest sequence + 1
1452          * so we can round-robin all routers, it's race and inaccurate but
1453          * harmless and functional  */
1454         if (best_route != NULL)
1455                 best_route->lr_seq = last_route->lr_seq + 1;
1456         return lpni_best;
1457 }
1458
1459 static struct lnet_ni *
1460 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1461                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1462                  int md_cpt)
1463 {
1464         struct lnet_ni *ni = NULL;
1465         unsigned int shortest_distance;
1466         int best_credits;
1467         int best_healthv;
1468
1469         /*
1470          * If there is no peer_ni that we can send to on this network,
1471          * then there is no point in looking for a new best_ni here.
1472         */
1473         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1474                 return best_ni;
1475
1476         if (best_ni == NULL) {
1477                 shortest_distance = UINT_MAX;
1478                 best_credits = INT_MIN;
1479                 best_healthv = 0;
1480         } else {
1481                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1482                                                      best_ni->ni_dev_cpt);
1483                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1484                 best_healthv = atomic_read(&best_ni->ni_healthv);
1485         }
1486
1487         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1488                 unsigned int distance;
1489                 int ni_credits;
1490                 int ni_healthv;
1491                 int ni_fatal;
1492
1493                 ni_credits = atomic_read(&ni->ni_tx_credits);
1494                 ni_healthv = atomic_read(&ni->ni_healthv);
1495                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1496
1497                 /*
1498                  * calculate the distance from the CPT on which
1499                  * the message memory is allocated to the CPT of
1500                  * the NI's physical device
1501                  */
1502                 distance = cfs_cpt_distance(lnet_cpt_table(),
1503                                             md_cpt,
1504                                             ni->ni_dev_cpt);
1505
1506                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1507                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1508                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1509                         : "not seleced", best_credits, shortest_distance,
1510                         (best_ni) ? best_ni->ni_seq : 0);
1511
1512                 /*
1513                  * All distances smaller than the NUMA range
1514                  * are treated equally.
1515                  */
1516                 if (distance < lnet_numa_range)
1517                         distance = lnet_numa_range;
1518
1519                 /*
1520                  * Select on health, shorter distance, available
1521                  * credits, then round-robin.
1522                  */
1523                 if (ni_fatal) {
1524                         continue;
1525                 } else if (ni_healthv < best_healthv) {
1526                         continue;
1527                 } else if (ni_healthv > best_healthv) {
1528                         best_healthv = ni_healthv;
1529                         /*
1530                          * If we're going to prefer this ni because it's
1531                          * the healthiest, then we should set the
1532                          * shortest_distance in the algorithm in case
1533                          * there are multiple NIs with the same health but
1534                          * different distances.
1535                          */
1536                         if (distance < shortest_distance)
1537                                 shortest_distance = distance;
1538                 } else if (distance > shortest_distance) {
1539                         continue;
1540                 } else if (distance < shortest_distance) {
1541                         shortest_distance = distance;
1542                 } else if (ni_credits < best_credits) {
1543                         continue;
1544                 } else if (ni_credits == best_credits) {
1545                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1546                                 continue;
1547                 }
1548                 best_ni = ni;
1549                 best_credits = ni_credits;
1550         }
1551
1552         CDEBUG(D_NET, "selected best_ni %s\n",
1553                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1554
1555         return best_ni;
1556 }
1557
1558 /*
1559  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1560  * because such traffic is required to perform discovery. We therefore
1561  * exclude all GET and PUT on that portal. We also exclude all ACK and
1562  * REPLY traffic, but that is because the portal is not tracked in the
1563  * message structure for these message types. We could restrict this
1564  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1565  */
1566 static bool
1567 lnet_msg_discovery(struct lnet_msg *msg)
1568 {
1569         if (msg->msg_type == LNET_MSG_PUT) {
1570                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1571                         return true;
1572         } else if (msg->msg_type == LNET_MSG_GET) {
1573                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1574                         return true;
1575         }
1576         return false;
1577 }
1578
1579 #define SRC_SPEC        0x0001
1580 #define SRC_ANY         0x0002
1581 #define LOCAL_DST       0x0004
1582 #define REMOTE_DST      0x0008
1583 #define MR_DST          0x0010
1584 #define NMR_DST         0x0020
1585 #define SND_RESP        0x0040
1586
1587 /* The following to defines are used for return codes */
1588 #define REPEAT_SEND     0x1000
1589 #define PASS_THROUGH    0x2000
1590
1591 /* The different cases lnet_select pathway needs to handle */
1592 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1593 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1594 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1595 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1596 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1597 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1598 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1599 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1600
1601 static int
1602 lnet_handle_send(struct lnet_send_data *sd)
1603 {
1604         struct lnet_ni *best_ni = sd->sd_best_ni;
1605         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1606         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1607         struct lnet_msg *msg = sd->sd_msg;
1608         int cpt2;
1609         __u32 send_case = sd->sd_send_case;
1610         int rc;
1611         __u32 routing = send_case & REMOTE_DST;
1612          struct lnet_rsp_tracker *rspt;
1613
1614         /*
1615          * Increment sequence number of the selected peer so that we
1616          * pick the next one in Round Robin.
1617          */
1618         best_lpni->lpni_seq++;
1619
1620         /*
1621          * grab a reference on the peer_ni so it sticks around even if
1622          * we need to drop and relock the lnet_net_lock below.
1623          */
1624         lnet_peer_ni_addref_locked(best_lpni);
1625
1626         /*
1627          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1628          * message. This ensures that we get a CPT that is correct for
1629          * the NI when the NI has been restricted to a subset of all CPTs.
1630          * If the selected CPT differs from the one currently locked, we
1631          * must unlock and relock the lnet_net_lock(), and then check whether
1632          * the configuration has changed. We don't have a hold on the best_ni
1633          * yet, and it may have vanished.
1634          */
1635         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1636         if (sd->sd_cpt != cpt2) {
1637                 __u32 seq = lnet_get_dlc_seq_locked();
1638                 lnet_net_unlock(sd->sd_cpt);
1639                 sd->sd_cpt = cpt2;
1640                 lnet_net_lock(sd->sd_cpt);
1641                 if (seq != lnet_get_dlc_seq_locked()) {
1642                         lnet_peer_ni_decref_locked(best_lpni);
1643                         return REPEAT_SEND;
1644                 }
1645         }
1646
1647         /*
1648          * store the best_lpni in the message right away to avoid having
1649          * to do the same operation under different conditions
1650          */
1651         msg->msg_txpeer = best_lpni;
1652         msg->msg_txni = best_ni;
1653
1654         /*
1655          * grab a reference for the best_ni since now it's in use in this
1656          * send. The reference will be dropped in lnet_finalize()
1657          */
1658         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1659
1660         /*
1661          * Always set the target.nid to the best peer picked. Either the
1662          * NID will be one of the peer NIDs selected, or the same NID as
1663          * what was originally set in the target or it will be the NID of
1664          * a router if this message should be routed
1665          */
1666         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1667
1668         /*
1669          * lnet_msg_commit assigns the correct cpt to the message, which
1670          * is used to decrement the correct refcount on the ni when it's
1671          * time to return the credits
1672          */
1673         lnet_msg_commit(msg, sd->sd_cpt);
1674
1675         /*
1676          * If we are routing the message then we keep the src_nid that was
1677          * set by the originator. If we are not routing then we are the
1678          * originator and set it here.
1679          */
1680         if (!msg->msg_routing)
1681                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1682
1683         if (routing) {
1684                 msg->msg_target_is_router = 1;
1685                 msg->msg_target.pid = LNET_PID_LUSTRE;
1686                 /*
1687                  * since we're routing we want to ensure that the
1688                  * msg_hdr.dest_nid is set to the final destination. When
1689                  * the router receives this message it knows how to route
1690                  * it.
1691                  *
1692                  * final_dst_lpni is set at the beginning of the
1693                  * lnet_select_pathway() function and is never changed.
1694                  * It's safe to use it here.
1695                  */
1696                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1697         } else {
1698                 /*
1699                  * if we're not routing set the dest_nid to the best peer
1700                  * ni NID that we picked earlier in the algorithm.
1701                  */
1702                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1703         }
1704
1705         /*
1706          * if we have response tracker block update it with the next hop
1707          * nid
1708          */
1709         if (msg->msg_md) {
1710                 rspt = msg->msg_md->md_rspt_ptr;
1711                 if (rspt) {
1712                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1713                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1714                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1715                 }
1716         }
1717
1718         rc = lnet_post_send_locked(msg, 0);
1719
1720         if (!rc)
1721                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1722                        libcfs_nid2str(msg->msg_hdr.src_nid),
1723                        libcfs_nid2str(msg->msg_txni->ni_nid),
1724                        libcfs_nid2str(sd->sd_src_nid),
1725                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1726                        libcfs_nid2str(sd->sd_dst_nid),
1727                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1728                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1729
1730         return rc;
1731 }
1732
1733 static struct lnet_peer_ni *
1734 lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer,
1735                     struct lnet_peer_net *peer_net)
1736 {
1737         /*
1738          * Look at the peer NIs for the destination peer that connect
1739          * to the chosen net. If a peer_ni is preferred when using the
1740          * best_ni to communicate, we use that one. If there is no
1741          * preferred peer_ni, or there are multiple preferred peer_ni,
1742          * the available transmit credits are used. If the transmit
1743          * credits are equal, we round-robin over the peer_ni.
1744          */
1745         struct lnet_peer_ni *lpni = NULL;
1746         struct lnet_peer_ni *best_lpni = NULL;
1747         struct lnet_ni *best_ni = sd->sd_best_ni;
1748         lnet_nid_t dst_nid = sd->sd_dst_nid;
1749         int best_lpni_credits = INT_MIN;
1750         bool preferred = false;
1751         bool ni_is_pref;
1752         int best_lpni_healthv = 0;
1753         int lpni_healthv;
1754
1755         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1756                 /*
1757                  * if the best_ni we've chosen aleady has this lpni
1758                  * preferred, then let's use it
1759                  */
1760                 ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1761                                                           best_ni->ni_nid);
1762
1763                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1764
1765                 CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1766                        libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1767
1768                 if (best_lpni)
1769                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1770                                 libcfs_nid2str(lpni->lpni_nid),
1771                                 lpni->lpni_txcredits, best_lpni_credits,
1772                                 lpni->lpni_seq, best_lpni->lpni_seq);
1773
1774                 /* pick the healthiest peer ni */
1775                 if (lpni_healthv < best_lpni_healthv) {
1776                         continue;
1777                 } else if (lpni_healthv > best_lpni_healthv) {
1778                         best_lpni_healthv = lpni_healthv;
1779                 /* if this is a preferred peer use it */
1780                 } else if (!preferred && ni_is_pref) {
1781                         preferred = true;
1782                 } else if (preferred && !ni_is_pref) {
1783                         /*
1784                          * this is not the preferred peer so let's ignore
1785                          * it.
1786                          */
1787                         continue;
1788                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1789                         /*
1790                          * We already have a peer that has more credits
1791                          * available than this one. No need to consider
1792                          * this peer further.
1793                          */
1794                         continue;
1795                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1796                         /*
1797                          * The best peer found so far and the current peer
1798                          * have the same number of available credits let's
1799                          * make sure to select between them using Round
1800                          * Robin
1801                          */
1802                         if (best_lpni) {
1803                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1804                                         continue;
1805                         }
1806                 }
1807
1808                 best_lpni = lpni;
1809                 best_lpni_credits = lpni->lpni_txcredits;
1810         }
1811
1812         /* if we still can't find a peer ni then we can't reach it */
1813         if (!best_lpni) {
1814                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1815                         LNET_NIDNET(dst_nid);
1816                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1817                                 libcfs_net2str(net_id));
1818                 return NULL;
1819         }
1820
1821         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1822                libcfs_nid2str(best_lpni->lpni_nid));
1823
1824         return best_lpni;
1825 }
1826
1827 /*
1828  * Prerequisite: the best_ni should already be set in the sd
1829  */
1830 static inline struct lnet_peer_ni *
1831 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1832                            __u32 net_id)
1833 {
1834         struct lnet_peer_net *peer_net;
1835
1836         /*
1837          * The gateway is Multi-Rail capable so now we must select the
1838          * proper peer_ni
1839          */
1840         peer_net = lnet_peer_get_net_locked(peer, net_id);
1841
1842         if (!peer_net) {
1843                 CERROR("gateway peer %s has no NI on net %s\n",
1844                        libcfs_nid2str(peer->lp_primary_nid),
1845                        libcfs_net2str(net_id));
1846                 return NULL;
1847         }
1848
1849         return lnet_select_peer_ni(sd, peer, peer_net);
1850 }
1851
1852 static inline void
1853 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1854 {
1855         if (sd->sd_send_case & NMR_DST &&
1856             sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1857             sd->sd_msg->msg_type != LNET_MSG_ACK &&
1858             sd->sd_best_lpni->lpni_pref_nnids == 0) {
1859                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1860                        libcfs_nid2str(sd->sd_best_ni->ni_nid),
1861                        libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1862                 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1863                                                  sd->sd_best_ni->ni_nid);
1864         }
1865 }
1866
1867 /*
1868  * Source Specified
1869  * Local Destination
1870  * non-mr peer
1871  *
1872  * use the source and destination NIDs as the pathway
1873  */
1874 static int
1875 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1876 {
1877         /* the destination lpni is set before we get here. */
1878
1879         /* find local NI */
1880         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1881         if (!sd->sd_best_ni) {
1882                 CERROR("Can't send to %s: src %s is not a "
1883                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1884                                 libcfs_nid2str(sd->sd_src_nid));
1885                 return -EINVAL;
1886         }
1887
1888         /*
1889          * the preferred NID will only be set for NMR peers
1890          */
1891         lnet_set_non_mr_pref_nid(sd);
1892
1893         return lnet_handle_send(sd);
1894 }
1895
1896 /*
1897  * Source Specified
1898  * Local Destination
1899  * MR Peer
1900  *
1901  * Run the selection algorithm on the peer NIs unless we're sending
1902  * a response, in this case just send to the destination
1903  */
1904 static int
1905 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1906 {
1907         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1908         if (!sd->sd_best_ni) {
1909                 CERROR("Can't send to %s: src %s is not a "
1910                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1911                                 libcfs_nid2str(sd->sd_src_nid));
1912                 return -EINVAL;
1913         }
1914
1915         /*
1916          * only run the selection algorithm to pick the peer_ni if we're
1917          * sending a GET or a PUT. Responses are sent to the same
1918          * destination NID provided.
1919          */
1920         if (!(sd->sd_send_case & SND_RESP)) {
1921                 sd->sd_best_lpni =
1922                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1923                                              sd->sd_best_ni->ni_net->net_id);
1924         }
1925
1926         if (sd->sd_best_lpni)
1927                 return lnet_handle_send(sd);
1928
1929         CERROR("can't send to %s. no NI on %s\n",
1930                libcfs_nid2str(sd->sd_dst_nid),
1931                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1932
1933         return -EHOSTUNREACH;
1934 }
1935
1936 struct lnet_ni *
1937 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1938                               struct lnet_peer *peer,
1939                               struct lnet_peer_net *peer_net,
1940                               int cpt,
1941                               bool incr_seq)
1942 {
1943         struct lnet_net *local_net;
1944         struct lnet_ni *best_ni;
1945
1946         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1947         if (!local_net)
1948                 return NULL;
1949
1950         /*
1951          * Iterate through the NIs in this local Net and select
1952          * the NI to send from. The selection is determined by
1953          * these 3 criterion in the following priority:
1954          *      1. NUMA
1955          *      2. NI available credits
1956          *      3. Round Robin
1957          */
1958         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1959                                    peer, peer_net, cpt);
1960
1961         if (incr_seq && best_ni)
1962                 best_ni->ni_seq++;
1963
1964         return best_ni;
1965 }
1966
1967 static int
1968 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1969                              lnet_nid_t dst_nid,
1970                              struct lnet_peer_ni **gw_lpni,
1971                              struct lnet_peer **gw_peer)
1972 {
1973         struct lnet_peer_ni *gw;
1974         lnet_nid_t src_nid = sd->sd_src_nid;
1975
1976         gw = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
1977                                     sd->sd_rtr_nid);
1978         if (!gw) {
1979                 CERROR("no route to %s from %s\n",
1980                        libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1981                 return -EHOSTUNREACH;
1982         }
1983
1984         /* get the peer of the gw_ni */
1985         LASSERT(gw->lpni_peer_net);
1986         LASSERT(gw->lpni_peer_net->lpn_peer);
1987
1988         *gw_peer = gw->lpni_peer_net->lpn_peer;
1989
1990         if (!sd->sd_best_ni)
1991                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, *gw_peer,
1992                                         gw->lpni_peer_net,
1993                                         sd->sd_md_cpt,
1994                                         true);
1995
1996         if (!sd->sd_best_ni) {
1997                 CERROR("Internal Error. Expected local ni on %s "
1998                        "but non found :%s\n",
1999                        libcfs_net2str(gw->lpni_peer_net->lpn_net_id),
2000                        libcfs_nid2str(sd->sd_src_nid));
2001                 return -EFAULT;
2002         }
2003
2004         /*
2005          * if gw is MR let's find its best peer_ni
2006          */
2007         if (lnet_peer_is_multi_rail(*gw_peer)) {
2008                 gw = lnet_find_best_lpni_on_net(sd, *gw_peer,
2009                                                 sd->sd_best_ni->ni_net->net_id);
2010                 /*
2011                  * We've already verified that the gw has an NI on that
2012                  * desired net, but we're not finding it. Something is
2013                  * wrong.
2014                  */
2015                 if (!gw) {
2016                         CERROR("Internal Error. Route expected to %s from %s\n",
2017                                 libcfs_nid2str(dst_nid),
2018                                 libcfs_nid2str(src_nid));
2019                         return -EFAULT;
2020                 }
2021         }
2022
2023         *gw_lpni = gw;
2024
2025         return 0;
2026 }
2027
2028 /*
2029  * Handle two cases:
2030  *
2031  * Case 1:
2032  *  Source specified
2033  *  Remote destination
2034  *  Non-MR destination
2035  *
2036  * Case 2:
2037  *  Source specified
2038  *  Remote destination
2039  *  MR destination
2040  *
2041  * The handling of these two cases is similar. Even though the destination
2042  * can be MR or non-MR, we'll deal directly with the router.
2043  */
2044 static int
2045 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2046 {
2047         int rc;
2048         struct lnet_peer_ni *gw_lpni = NULL;
2049         struct lnet_peer *gw_peer = NULL;
2050
2051         /* find local NI */
2052         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2053         if (!sd->sd_best_ni) {
2054                 CERROR("Can't send to %s: src %s is not a "
2055                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2056                                 libcfs_nid2str(sd->sd_src_nid));
2057                 return -EINVAL;
2058         }
2059
2060         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2061                                      &gw_peer);
2062         if (rc < 0)
2063                 return rc;
2064
2065         if (sd->sd_send_case & NMR_DST)
2066                 /*
2067                 * since the final destination is non-MR let's set its preferred
2068                 * NID before we send
2069                 */
2070                 lnet_set_non_mr_pref_nid(sd);
2071
2072         /*
2073          * We're going to send to the gw found so let's set its
2074          * info
2075          */
2076         sd->sd_peer = gw_peer;
2077         sd->sd_best_lpni = gw_lpni;
2078
2079         return lnet_handle_send(sd);
2080 }
2081
2082 struct lnet_ni *
2083 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
2084 {
2085         struct lnet_peer_net *peer_net = NULL;
2086         struct lnet_ni *best_ni = NULL;
2087
2088         /*
2089          * The peer can have multiple interfaces, some of them can be on
2090          * the local network and others on a routed network. We should
2091          * prefer the local network. However if the local network is not
2092          * available then we need to try the routed network
2093          */
2094
2095         /* go through all the peer nets and find the best_ni */
2096         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2097                 /*
2098                  * The peer's list of nets can contain non-local nets. We
2099                  * want to only examine the local ones.
2100                  */
2101                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2102                         continue;
2103                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2104                                                    peer_net, md_cpt, false);
2105         }
2106
2107         if (best_ni)
2108                 /* increment sequence number so we can round robin */
2109                 best_ni->ni_seq++;
2110
2111         return best_ni;
2112 }
2113
2114 static struct lnet_ni *
2115 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2116 {
2117         struct lnet_ni *best_ni = NULL;
2118         struct lnet_peer_net *peer_net;
2119         struct lnet_peer *peer = sd->sd_peer;
2120         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2121         struct lnet_peer_ni *lpni;
2122         int cpt = sd->sd_cpt;
2123
2124         /*
2125          * We must use a consistent source address when sending to a
2126          * non-MR peer. However, a non-MR peer can have multiple NIDs
2127          * on multiple networks, and we may even need to talk to this
2128          * peer on multiple networks -- certain types of
2129          * load-balancing configuration do this.
2130          *
2131          * So we need to pick the NI the peer prefers for this
2132          * particular network.
2133          */
2134
2135         /* Get the target peer_ni */
2136         peer_net = lnet_peer_get_net_locked(peer,
2137                         LNET_NIDNET(best_lpni->lpni_nid));
2138         LASSERT(peer_net != NULL);
2139         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2140                                 lpni_peer_nis) {
2141                 if (lpni->lpni_pref_nnids == 0)
2142                         continue;
2143                 LASSERT(lpni->lpni_pref_nnids == 1);
2144                 best_ni = lnet_nid2ni_locked(
2145                                 lpni->lpni_pref.nid, cpt);
2146                 break;
2147         }
2148
2149         return best_ni;
2150 }
2151
2152 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2153 static int
2154 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2155 {
2156         struct lnet_ni *best_ni = NULL;
2157         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2158
2159         /*
2160          * We must use a consistent source address when sending to a
2161          * non-MR peer. However, a non-MR peer can have multiple NIDs
2162          * on multiple networks, and we may even need to talk to this
2163          * peer on multiple networks -- certain types of
2164          * load-balancing configuration do this.
2165          *
2166          * So we need to pick the NI the peer prefers for this
2167          * particular network.
2168          */
2169
2170         best_ni = lnet_find_existing_preferred_best_ni(sd);
2171
2172         /* if best_ni is still not set just pick one */
2173         if (!best_ni) {
2174                 best_ni =
2175                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2176                                                 sd->sd_best_lpni->lpni_peer_net,
2177                                                 sd->sd_md_cpt, true);
2178                 /* If there is no best_ni we don't have a route */
2179                 if (!best_ni) {
2180                         CERROR("no path to %s from net %s\n",
2181                                 libcfs_nid2str(best_lpni->lpni_nid),
2182                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2183                         return -EHOSTUNREACH;
2184                 }
2185         }
2186
2187         sd->sd_best_ni = best_ni;
2188
2189         /* Set preferred NI if necessary. */
2190         lnet_set_non_mr_pref_nid(sd);
2191
2192         return 0;
2193 }
2194
2195
2196 /*
2197  * Source not specified
2198  * Local destination
2199  * Non-MR Peer
2200  *
2201  * always use the same source NID for NMR peers
2202  * If we've talked to that peer before then we already have a preferred
2203  * source NI associated with it. Otherwise, we select a preferred local NI
2204  * and store it in the peer
2205  */
2206 static int
2207 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2208 {
2209         int rc;
2210
2211         /* sd->sd_best_lpni is already set to the final destination */
2212
2213         /*
2214          * At this point we should've created the peer ni and peer. If we
2215          * can't find it, then something went wrong. Instead of assert
2216          * output a relevant message and fail the send
2217          */
2218         if (!sd->sd_best_lpni) {
2219                 CERROR("Internal fault. Unable to send msg %s to %s. "
2220                        "NID not known\n",
2221                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2222                        libcfs_nid2str(sd->sd_dst_nid));
2223                 return -EFAULT;
2224         }
2225
2226         rc = lnet_select_preferred_best_ni(sd);
2227         if (!rc)
2228                 rc = lnet_handle_send(sd);
2229
2230         return rc;
2231 }
2232
2233 static int
2234 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2235 {
2236         /*
2237          * NOTE we've already handled the remote peer case. So we only
2238          * need to worry about the local case here.
2239          *
2240          * if we're sending a response, ACK or reply, we need to send it
2241          * to the destination NID given to us. At this point we already
2242          * have the peer_ni we're suppose to send to, so just find the
2243          * best_ni on the peer net and use that. Since we're sending to an
2244          * MR peer then we can just run the selection algorithm on our
2245          * local NIs and pick the best one.
2246          */
2247         if (sd->sd_send_case & SND_RESP) {
2248                 sd->sd_best_ni =
2249                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2250                                                 sd->sd_best_lpni->lpni_peer_net,
2251                                                 sd->sd_md_cpt, true);
2252
2253                 if (!sd->sd_best_ni) {
2254                         /*
2255                          * We're not going to deal with not able to send
2256                          * a response to the provided final destination
2257                          */
2258                         CERROR("Can't send response to %s. "
2259                                "No local NI available\n",
2260                                 libcfs_nid2str(sd->sd_dst_nid));
2261                         return -EHOSTUNREACH;
2262                 }
2263
2264                 return lnet_handle_send(sd);
2265         }
2266
2267         /*
2268          * If we get here that means we're sending a fresh request, PUT or
2269          * GET, so we need to run our standard selection algorithm.
2270          * First find the best local interface that's on any of the peer's
2271          * networks.
2272          */
2273         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2274                                                         sd->sd_md_cpt);
2275         if (sd->sd_best_ni) {
2276                 sd->sd_best_lpni =
2277                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2278                                              sd->sd_best_ni->ni_net->net_id);
2279
2280                 /*
2281                  * if we're successful in selecting a peer_ni on the local
2282                  * network, then send to it. Otherwise fall through and
2283                  * try and see if we can reach it over another routed
2284                  * network
2285                  */
2286                 if (sd->sd_best_lpni) {
2287                         /*
2288                          * in case we initially started with a routed
2289                          * destination, let's reset to local
2290                          */
2291                         sd->sd_send_case &= ~REMOTE_DST;
2292                         sd->sd_send_case |= LOCAL_DST;
2293                         return lnet_handle_send(sd);
2294                 }
2295
2296                 CERROR("Internal Error. Expected to have a best_lpni: "
2297                        "%s -> %s\n",
2298                        libcfs_nid2str(sd->sd_src_nid),
2299                        libcfs_nid2str(sd->sd_dst_nid));
2300
2301                 return -EFAULT;
2302         }
2303
2304         /*
2305          * Peer doesn't have a local network. Let's see if there is
2306          * a remote network we can reach it on.
2307          */
2308         return PASS_THROUGH;
2309 }
2310
2311 /*
2312  * Case 1:
2313  *      Source NID not specified
2314  *      Local destination
2315  *      MR peer
2316  *
2317  * Case 2:
2318  *      Source NID not speified
2319  *      Remote destination
2320  *      MR peer
2321  *
2322  * In both of these cases if we're sending a response, ACK or REPLY, then
2323  * we need to send to the destination NID provided.
2324  *
2325  * In the remote case let's deal with MR routers.
2326  *
2327  */
2328
2329 static int
2330 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2331 {
2332         int rc = 0;
2333         struct lnet_peer *gw_peer = NULL;
2334         struct lnet_peer_ni *gw_lpni = NULL;
2335
2336         /*
2337          * handle sending a response to a remote peer here so we don't
2338          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2339          */
2340         if (sd->sd_send_case & REMOTE_DST &&
2341             sd->sd_send_case & SND_RESP) {
2342                 struct lnet_peer_ni *gw;
2343                 struct lnet_peer *gw_peer;
2344
2345                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2346                                                   &gw_peer);
2347                 if (rc < 0) {
2348                         CERROR("Can't send response to %s. "
2349                                "No route available\n",
2350                                 libcfs_nid2str(sd->sd_dst_nid));
2351                         return -EHOSTUNREACH;
2352                 }
2353
2354                 sd->sd_best_lpni = gw;
2355                 sd->sd_peer = gw_peer;
2356
2357                 return lnet_handle_send(sd);
2358         }
2359
2360         /*
2361          * Even though the NID for the peer might not be on a local network,
2362          * since the peer is MR there could be other interfaces on the
2363          * local network. In that case we'd still like to prefer the local
2364          * network over the routed network. If we're unable to do that
2365          * then we select the best router among the different routed networks,
2366          * and if the router is MR then we can deal with it as such.
2367          */
2368         rc = lnet_handle_any_mr_dsta(sd);
2369         if (rc != PASS_THROUGH)
2370                 return rc;
2371
2372         /*
2373          * TODO; One possible enhancement is to run the selection
2374          * algorithm on the peer. However for remote peers the credits are
2375          * not decremented, so we'll be basically going over the peer NIs
2376          * in round robin. An MR router will run the selection algorithm
2377          * on the next-hop interfaces.
2378          */
2379         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2380                                           &gw_peer);
2381         if (rc < 0)
2382                 return rc;
2383
2384         sd->sd_send_case &= ~LOCAL_DST;
2385         sd->sd_send_case |= REMOTE_DST;
2386
2387         sd->sd_peer = gw_peer;
2388         sd->sd_best_lpni = gw_lpni;
2389
2390         return lnet_handle_send(sd);
2391 }
2392
2393 /*
2394  * Source not specified
2395  * Remote destination
2396  * Non-MR peer
2397  *
2398  * Must send to the specified peer NID using the same source NID that
2399  * we've used before. If it's the first time to talk to that peer then
2400  * find the source NI and assign it as preferred to that peer
2401  */
2402 static int
2403 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2404 {
2405         int rc;
2406         struct lnet_peer_ni *gw_lpni = NULL;
2407         struct lnet_peer *gw_peer = NULL;
2408
2409         /*
2410          * Let's set if we have a preferred NI to talk to this NMR peer
2411          */
2412         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2413
2414         /*
2415          * find the router and that'll find the best NI if we didn't find
2416          * it already.
2417          */
2418         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2419                                           &gw_peer);
2420         if (rc < 0)
2421                 return rc;
2422
2423         /*
2424          * set the best_ni we've chosen as the preferred one for
2425          * this peer
2426          */
2427         lnet_set_non_mr_pref_nid(sd);
2428
2429         /* we'll be sending to the gw */
2430         sd->sd_best_lpni = gw_lpni;
2431         sd->sd_peer = gw_peer;
2432
2433         return lnet_handle_send(sd);
2434 }
2435
2436 static int
2437 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2438 {
2439         /*
2440          * turn off the SND_RESP bit.
2441          * It will be checked in the case handling
2442          */
2443         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2444
2445         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2446                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2447                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2448                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2449                 libcfs_nid2str(sd->sd_dst_nid),
2450                 (send_case & LOCAL_DST) ? "local" : "routed");
2451
2452         switch (send_case) {
2453         /*
2454          * For all cases where the source is specified, we should always
2455          * use the destination NID, whether it's an MR destination or not,
2456          * since we're continuing a series of related messages for the
2457          * same RPC
2458          */
2459         case SRC_SPEC_LOCAL_NMR_DST:
2460                 return lnet_handle_spec_local_nmr_dst(sd);
2461         case SRC_SPEC_LOCAL_MR_DST:
2462                 return lnet_handle_spec_local_mr_dst(sd);
2463         case SRC_SPEC_ROUTER_NMR_DST:
2464         case SRC_SPEC_ROUTER_MR_DST:
2465                 return lnet_handle_spec_router_dst(sd);
2466         case SRC_ANY_LOCAL_NMR_DST:
2467                 return lnet_handle_any_local_nmr_dst(sd);
2468         case SRC_ANY_LOCAL_MR_DST:
2469         case SRC_ANY_ROUTER_MR_DST:
2470                 return lnet_handle_any_mr_dst(sd);
2471         case SRC_ANY_ROUTER_NMR_DST:
2472                 return lnet_handle_any_router_nmr_dst(sd);
2473         default:
2474                 CERROR("Unknown send case\n");
2475                 return -1;
2476         }
2477 }
2478
2479 static int
2480 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2481                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2482 {
2483         struct lnet_peer_ni     *lpni;
2484         struct lnet_peer        *peer;
2485         struct lnet_send_data   send_data;
2486         int                     cpt, rc;
2487         int                     md_cpt;
2488         __u32                   send_case = 0;
2489
2490         memset(&send_data, 0, sizeof(send_data));
2491
2492         /*
2493          * get an initial CPT to use for locking. The idea here is not to
2494          * serialize the calls to select_pathway, so that as many
2495          * operations can run concurrently as possible. To do that we use
2496          * the CPT where this call is being executed. Later on when we
2497          * determine the CPT to use in lnet_message_commit, we switch the
2498          * lock and check if there was any configuration change.  If none,
2499          * then we proceed, if there is, then we restart the operation.
2500          */
2501         cpt = lnet_net_lock_current();
2502
2503         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2504         if (md_cpt == CFS_CPT_ANY)
2505                 md_cpt = cpt;
2506
2507 again:
2508
2509         /*
2510          * If we're being asked to send to the loopback interface, there
2511          * is no need to go through any selection. We can just shortcut
2512          * the entire process and send over lolnd
2513          */
2514         if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2515                 /* No send credit hassles with LOLND */
2516                 lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
2517                 msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
2518                 if (!msg->msg_routing)
2519                         msg->msg_hdr.src_nid =
2520                                 cpu_to_le64(the_lnet.ln_loni->ni_nid);
2521                 msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
2522                 lnet_msg_commit(msg, cpt);
2523                 msg->msg_txni = the_lnet.ln_loni;
2524                 lnet_net_unlock(cpt);
2525
2526                 return LNET_CREDIT_OK;
2527         }
2528
2529         /*
2530          * find an existing peer_ni, or create one and mark it as having been
2531          * created due to network traffic. This call will create the
2532          * peer->peer_net->peer_ni tree.
2533          */
2534         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2535         if (IS_ERR(lpni)) {
2536                 lnet_net_unlock(cpt);
2537                 return PTR_ERR(lpni);
2538         }
2539
2540         /*
2541          * Cache the original src_nid. If we need to resend the message
2542          * then we'll need to know whether the src_nid was originally
2543          * specified for this message. If it was originally specified,
2544          * then we need to keep using the same src_nid since it's
2545          * continuing the same sequence of messages.
2546          */
2547         msg->msg_src_nid_param = src_nid;
2548
2549         /*
2550          * Now that we have a peer_ni, check if we want to discover
2551          * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2552          * trigger discovery.
2553          */
2554         peer = lpni->lpni_peer_net->lpn_peer;
2555         if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
2556                 lnet_nid_t primary_nid;
2557                 rc = lnet_discover_peer_locked(lpni, cpt, false);
2558                 if (rc) {
2559                         lnet_peer_ni_decref_locked(lpni);
2560                         lnet_net_unlock(cpt);
2561                         return rc;
2562                 }
2563                 /* The peer may have changed. */
2564                 peer = lpni->lpni_peer_net->lpn_peer;
2565                 /* queue message and return */
2566                 msg->msg_rtr_nid_param = rtr_nid;
2567                 msg->msg_sending = 0;
2568                 list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2569                 lnet_peer_ni_decref_locked(lpni);
2570                 primary_nid = peer->lp_primary_nid;
2571                 lnet_net_unlock(cpt);
2572
2573                 CDEBUG(D_NET, "%s pending discovery\n",
2574                        libcfs_nid2str(primary_nid));
2575
2576                 return LNET_DC_WAIT;
2577         }
2578         lnet_peer_ni_decref_locked(lpni);
2579
2580         /*
2581          * Identify the different send cases
2582          */
2583         if (src_nid == LNET_NID_ANY)
2584                 send_case |= SRC_ANY;
2585         else
2586                 send_case |= SRC_SPEC;
2587
2588         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2589                 send_case |= LOCAL_DST;
2590         else
2591                 send_case |= REMOTE_DST;
2592
2593         /*
2594          * if this is a non-MR peer or if we're recovering a peer ni then
2595          * let's consider this an NMR case so we can hit the destination
2596          * NID.
2597          */
2598         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2599                 send_case |= NMR_DST;
2600         else
2601                 send_case |= MR_DST;
2602
2603         if (msg->msg_type == LNET_MSG_REPLY ||
2604             msg->msg_type == LNET_MSG_ACK)
2605                 send_case |= SND_RESP;
2606
2607         /* assign parameters to the send_data */
2608         send_data.sd_msg = msg;
2609         send_data.sd_rtr_nid = rtr_nid;
2610         send_data.sd_src_nid = src_nid;
2611         send_data.sd_dst_nid = dst_nid;
2612         send_data.sd_best_lpni = lpni;
2613         /*
2614          * keep a pointer to the final destination in case we're going to
2615          * route, so we'll need to access it later
2616          */
2617         send_data.sd_final_dst_lpni = lpni;
2618         send_data.sd_peer = peer;
2619         send_data.sd_md_cpt = md_cpt;
2620         send_data.sd_cpt = cpt;
2621         send_data.sd_send_case = send_case;
2622
2623         rc = lnet_handle_send_case_locked(&send_data);
2624
2625         if (rc == REPEAT_SEND)
2626                 goto again;
2627
2628         lnet_net_unlock(send_data.sd_cpt);
2629
2630         return rc;
2631 }
2632
2633 int
2634 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2635 {
2636         lnet_nid_t              dst_nid = msg->msg_target.nid;
2637         int                     rc;
2638
2639         /*
2640          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2641          * but we might want to use pre-determined router for ACK/REPLY
2642          * in the future
2643          */
2644         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2645         LASSERT(msg->msg_txpeer == NULL);
2646         LASSERT(msg->msg_txni == NULL);
2647         LASSERT(!msg->msg_sending);
2648         LASSERT(!msg->msg_target_is_router);
2649         LASSERT(!msg->msg_receiving);
2650
2651         msg->msg_sending = 1;
2652
2653         LASSERT(!msg->msg_tx_committed);
2654
2655         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2656         if (rc < 0)
2657                 return rc;
2658
2659         if (rc == LNET_CREDIT_OK)
2660                 lnet_ni_send(msg->msg_txni, msg);
2661
2662         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2663         return 0;
2664 }
2665
2666 enum lnet_mt_event_type {
2667         MT_TYPE_LOCAL_NI = 0,
2668         MT_TYPE_PEER_NI
2669 };
2670
2671 struct lnet_mt_event_info {
2672         enum lnet_mt_event_type mt_type;
2673         lnet_nid_t mt_nid;
2674 };
2675
2676 void
2677 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2678 {
2679         struct lnet_rsp_tracker *rspt;
2680
2681         /*
2682          * msg has a refcount on the MD so the MD is not going away.
2683          * The rspt queue for the cpt is protected by
2684          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2685          */
2686         lnet_res_lock(cpt);
2687         if (!md->md_rspt_ptr) {
2688                 lnet_res_unlock(cpt);
2689                 return;
2690         }
2691         rspt = md->md_rspt_ptr;
2692         md->md_rspt_ptr = NULL;
2693
2694         /* debug code */
2695         LASSERT(rspt->rspt_cpt == cpt);
2696
2697         /*
2698          * invalidate the handle to indicate that a response has been
2699          * received, which will then lead the monitor thread to clean up
2700          * the rspt block.
2701          */
2702         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2703         lnet_res_unlock(cpt);
2704 }
2705
2706 static void
2707 lnet_finalize_expired_responses(bool force)
2708 {
2709         struct lnet_libmd *md;
2710         struct list_head local_queue;
2711         struct lnet_rsp_tracker *rspt, *tmp;
2712         int i;
2713
2714         if (the_lnet.ln_mt_rstq == NULL)
2715                 return;
2716
2717         cfs_cpt_for_each(i, lnet_cpt_table()) {
2718                 INIT_LIST_HEAD(&local_queue);
2719
2720                 lnet_net_lock(i);
2721                 if (!the_lnet.ln_mt_rstq[i]) {
2722                         lnet_net_unlock(i);
2723                         continue;
2724                 }
2725                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2726                 lnet_net_unlock(i);
2727
2728                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2729                         /*
2730                          * The rspt mdh will be invalidated when a response
2731                          * is received or whenever we want to discard the
2732                          * block the monitor thread will walk the queue
2733                          * and clean up any rsts with an invalid mdh.
2734                          * The monitor thread will walk the queue until
2735                          * the first unexpired rspt block. This means that
2736                          * some rspt blocks which received their
2737                          * corresponding responses will linger in the
2738                          * queue until they are cleaned up eventually.
2739                          */
2740                         lnet_res_lock(i);
2741                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2742                                 lnet_res_unlock(i);
2743                                 list_del_init(&rspt->rspt_on_list);
2744                                 lnet_rspt_free(rspt, i);
2745                                 continue;
2746                         }
2747
2748                         if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
2749                             force) {
2750                                 struct lnet_peer_ni *lpni;
2751                                 lnet_nid_t nid;
2752
2753                                 md = lnet_handle2md(&rspt->rspt_mdh);
2754                                 if (!md) {
2755                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2756                                         lnet_res_unlock(i);
2757                                         list_del_init(&rspt->rspt_on_list);
2758                                         lnet_rspt_free(rspt, i);
2759                                         continue;
2760                                 }
2761                                 LASSERT(md->md_rspt_ptr == rspt);
2762                                 md->md_rspt_ptr = NULL;
2763                                 lnet_res_unlock(i);
2764
2765                                 lnet_net_lock(i);
2766                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2767                                 lnet_net_unlock(i);
2768
2769                                 list_del_init(&rspt->rspt_on_list);
2770
2771                                 nid = rspt->rspt_next_hop_nid;
2772
2773                                 CNETERR("Response timed out: md = %p: nid = %s\n",
2774                                         md, libcfs_nid2str(nid));
2775                                 LNetMDUnlink(rspt->rspt_mdh);
2776                                 lnet_rspt_free(rspt, i);
2777
2778                                 /*
2779                                  * If there is a timeout on the response
2780                                  * from the next hop decrement its health
2781                                  * value so that we don't use it
2782                                  */
2783                                 lnet_net_lock(0);
2784                                 lpni = lnet_find_peer_ni_locked(nid);
2785                                 if (lpni) {
2786                                         lnet_handle_remote_failure_locked(lpni);
2787                                         lnet_peer_ni_decref_locked(lpni);
2788                                 }
2789                                 lnet_net_unlock(0);
2790                         } else {
2791                                 lnet_res_unlock(i);
2792                                 break;
2793                         }
2794                 }
2795
2796                 lnet_net_lock(i);
2797                 if (!list_empty(&local_queue))
2798                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2799                 lnet_net_unlock(i);
2800         }
2801 }
2802
2803 static void
2804 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2805 {
2806         struct lnet_msg *msg;
2807
2808         while (!list_empty(resendq)) {
2809                 struct lnet_peer_ni *lpni;
2810
2811                 msg = list_entry(resendq->next, struct lnet_msg,
2812                                  msg_list);
2813
2814                 list_del_init(&msg->msg_list);
2815
2816                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2817                 if (!lpni) {
2818                         lnet_net_unlock(cpt);
2819                         CERROR("Expected that a peer is already created for %s\n",
2820                                libcfs_nid2str(msg->msg_hdr.dest_nid));
2821                         msg->msg_no_resend = true;
2822                         lnet_finalize(msg, -EFAULT);
2823                         lnet_net_lock(cpt);
2824                 } else {
2825                         struct lnet_peer *peer;
2826                         int rc;
2827                         lnet_nid_t src_nid = LNET_NID_ANY;
2828
2829                         /*
2830                          * if this message is not being routed and the
2831                          * peer is non-MR then we must use the same
2832                          * src_nid that was used in the original send.
2833                          * Otherwise if we're routing the message (IE
2834                          * we're a router) then we can use any of our
2835                          * local interfaces. It doesn't matter to the
2836                          * final destination.
2837                          */
2838                         peer = lpni->lpni_peer_net->lpn_peer;
2839                         if (!msg->msg_routing &&
2840                             !lnet_peer_is_multi_rail(peer))
2841                                 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
2842
2843                         /*
2844                          * If we originally specified a src NID, then we
2845                          * must attempt to reuse it in the resend as well.
2846                          */
2847                         if (msg->msg_src_nid_param != LNET_NID_ANY)
2848                                 src_nid = msg->msg_src_nid_param;
2849                         lnet_peer_ni_decref_locked(lpni);
2850
2851                         lnet_net_unlock(cpt);
2852                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
2853                                libcfs_nid2str(src_nid),
2854                                libcfs_id2str(msg->msg_target),
2855                                lnet_msgtyp2str(msg->msg_type),
2856                                msg->msg_recovery,
2857                                msg->msg_retry_count);
2858                         rc = lnet_send(src_nid, msg, LNET_NID_ANY);
2859                         if (rc) {
2860                                 CERROR("Error sending %s to %s: %d\n",
2861                                        lnet_msgtyp2str(msg->msg_type),
2862                                        libcfs_id2str(msg->msg_target), rc);
2863                                 msg->msg_no_resend = true;
2864                                 lnet_finalize(msg, rc);
2865                         }
2866                         lnet_net_lock(cpt);
2867                         if (!rc)
2868                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
2869                 }
2870         }
2871 }
2872
2873 static void
2874 lnet_resend_pending_msgs(void)
2875 {
2876         int i;
2877
2878         cfs_cpt_for_each(i, lnet_cpt_table()) {
2879                 lnet_net_lock(i);
2880                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
2881                 lnet_net_unlock(i);
2882         }
2883 }
2884
2885 /* called with cpt and ni_lock held */
2886 static void
2887 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
2888 {
2889         struct lnet_handle_md recovery_mdh;
2890
2891         LNetInvalidateMDHandle(&recovery_mdh);
2892
2893         if (ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING || force) {
2894                 recovery_mdh = ni->ni_ping_mdh;
2895                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2896         }
2897         lnet_ni_unlock(ni);
2898         lnet_net_unlock(cpt);
2899         if (!LNetMDHandleIsInvalid(recovery_mdh))
2900                 LNetMDUnlink(recovery_mdh);
2901         lnet_net_lock(cpt);
2902         lnet_ni_lock(ni);
2903 }
2904
2905 static void
2906 lnet_recover_local_nis(void)
2907 {
2908         struct lnet_mt_event_info *ev_info;
2909         struct list_head processed_list;
2910         struct list_head local_queue;
2911         struct lnet_handle_md mdh;
2912         struct lnet_ni *tmp;
2913         struct lnet_ni *ni;
2914         lnet_nid_t nid;
2915         int healthv;
2916         int rc;
2917
2918         INIT_LIST_HEAD(&local_queue);
2919         INIT_LIST_HEAD(&processed_list);
2920
2921         /*
2922          * splice the recovery queue on a local queue. We will iterate
2923          * through the local queue and update it as needed. Once we're
2924          * done with the traversal, we'll splice the local queue back on
2925          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
2926          * will be traversed in the next iteration.
2927          */
2928         lnet_net_lock(0);
2929         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
2930                          &local_queue);
2931         lnet_net_unlock(0);
2932
2933         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
2934                 /*
2935                  * if an NI is being deleted or it is now healthy, there
2936                  * is no need to keep it around in the recovery queue.
2937                  * The monitor thread is the only thread responsible for
2938                  * removing the NI from the recovery queue.
2939                  * Multiple threads can be adding NIs to the recovery
2940                  * queue.
2941                  */
2942                 healthv = atomic_read(&ni->ni_healthv);
2943
2944                 lnet_net_lock(0);
2945                 lnet_ni_lock(ni);
2946                 if (!(ni->ni_state & LNET_NI_STATE_ACTIVE) ||
2947                     healthv == LNET_MAX_HEALTH_VALUE) {
2948                         list_del_init(&ni->ni_recovery);
2949                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
2950                         lnet_ni_unlock(ni);
2951                         lnet_ni_decref_locked(ni, 0);
2952                         lnet_net_unlock(0);
2953                         continue;
2954                 }
2955
2956                 /*
2957                  * if the local NI failed recovery we must unlink the md.
2958                  * But we want to keep the local_ni on the recovery queue
2959                  * so we can continue the attempts to recover it.
2960                  */
2961                 if (ni->ni_state & LNET_NI_STATE_RECOVERY_FAILED) {
2962                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
2963                         ni->ni_state &= ~LNET_NI_STATE_RECOVERY_FAILED;
2964                 }
2965
2966                 lnet_ni_unlock(ni);
2967                 lnet_net_unlock(0);
2968
2969
2970                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
2971                        libcfs_nid2str(ni->ni_nid));
2972
2973                 lnet_ni_lock(ni);
2974                 if (!(ni->ni_state & LNET_NI_STATE_RECOVERY_PENDING)) {
2975                         ni->ni_state |= LNET_NI_STATE_RECOVERY_PENDING;
2976                         lnet_ni_unlock(ni);
2977
2978                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
2979                         if (!ev_info) {
2980                                 CERROR("out of memory. Can't recover %s\n",
2981                                        libcfs_nid2str(ni->ni_nid));
2982                                 lnet_ni_lock(ni);
2983                                 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
2984                                 lnet_ni_unlock(ni);
2985                                 continue;
2986                         }
2987
2988                         mdh = ni->ni_ping_mdh;
2989                         /*
2990                          * Invalidate the ni mdh in case it's deleted.
2991                          * We'll unlink the mdh in this case below.
2992                          */
2993                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
2994                         nid = ni->ni_nid;
2995
2996                         /*
2997                          * remove the NI from the local queue and drop the
2998                          * reference count to it while we're recovering
2999                          * it. The reason for that, is that the NI could
3000                          * be deleted, and the way the code is structured
3001                          * is if we don't drop the NI, then the deletion
3002                          * code will enter a loop waiting for the
3003                          * reference count to be removed while holding the
3004                          * ln_mutex_lock(). When we look up the peer to
3005                          * send to in lnet_select_pathway() we will try to
3006                          * lock the ln_mutex_lock() as well, leading to
3007                          * a deadlock. By dropping the refcount and
3008                          * removing it from the list, we allow for the NI
3009                          * to be removed, then we use the cached NID to
3010                          * look it up again. If it's gone, then we just
3011                          * continue examining the rest of the queue.
3012                          */
3013                         lnet_net_lock(0);
3014                         list_del_init(&ni->ni_recovery);
3015                         lnet_ni_decref_locked(ni, 0);
3016                         lnet_net_unlock(0);
3017
3018                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3019                         ev_info->mt_nid = nid;
3020                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3021                                             ev_info, the_lnet.ln_mt_eqh, true);
3022                         /* lookup the nid again */
3023                         lnet_net_lock(0);
3024                         ni = lnet_nid2ni_locked(nid, 0);
3025                         if (!ni) {
3026                                 /*
3027                                  * the NI has been deleted when we dropped
3028                                  * the ref count
3029                                  */
3030                                 lnet_net_unlock(0);
3031                                 LNetMDUnlink(mdh);
3032                                 continue;
3033                         }
3034                         /*
3035                          * Same note as in lnet_recover_peer_nis(). When
3036                          * we're sending the ping, the NI is free to be
3037                          * deleted or manipulated. By this point it
3038                          * could've been added back on the recovery queue,
3039                          * and a refcount taken on it.
3040                          * So we can't just add it blindly again or we'll
3041                          * corrupt the queue. We must check under lock if
3042                          * it's not on any list and if not then add it
3043                          * to the processed list, which will eventually be
3044                          * spliced back on to the recovery queue.
3045                          */
3046                         ni->ni_ping_mdh = mdh;
3047                         if (list_empty(&ni->ni_recovery)) {
3048                                 list_add_tail(&ni->ni_recovery, &processed_list);
3049                                 lnet_ni_addref_locked(ni, 0);
3050                         }
3051                         lnet_net_unlock(0);
3052
3053                         lnet_ni_lock(ni);
3054                         if (rc)
3055                                 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
3056                 }
3057                 lnet_ni_unlock(ni);
3058         }
3059
3060         /*
3061          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3062          * reexamined in the next iteration.
3063          */
3064         list_splice_init(&processed_list, &local_queue);
3065         lnet_net_lock(0);
3066         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3067         lnet_net_unlock(0);
3068 }
3069
3070 static struct list_head **
3071 lnet_create_array_of_queues(void)
3072 {
3073         struct list_head **qs;
3074         struct list_head *q;
3075         int i;
3076
3077         qs = cfs_percpt_alloc(lnet_cpt_table(),
3078                               sizeof(struct list_head));
3079         if (!qs) {
3080                 CERROR("Failed to allocate queues\n");
3081                 return NULL;
3082         }
3083
3084         cfs_percpt_for_each(q, i, qs)
3085                 INIT_LIST_HEAD(q);
3086
3087         return qs;
3088 }
3089
3090 static int
3091 lnet_resendqs_create(void)
3092 {
3093         struct list_head **resendqs;
3094         resendqs = lnet_create_array_of_queues();
3095
3096         if (!resendqs)
3097                 return -ENOMEM;
3098
3099         lnet_net_lock(LNET_LOCK_EX);
3100         the_lnet.ln_mt_resendqs = resendqs;
3101         lnet_net_unlock(LNET_LOCK_EX);
3102
3103         return 0;
3104 }
3105
3106 static void
3107 lnet_clean_local_ni_recoveryq(void)
3108 {
3109         struct lnet_ni *ni;
3110
3111         /* This is only called when the monitor thread has stopped */
3112         lnet_net_lock(0);
3113
3114         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3115                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3116                                 struct lnet_ni, ni_recovery);
3117                 list_del_init(&ni->ni_recovery);
3118                 lnet_ni_lock(ni);
3119                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3120                 lnet_ni_unlock(ni);
3121                 lnet_ni_decref_locked(ni, 0);
3122         }
3123
3124         lnet_net_unlock(0);
3125 }
3126
3127 static void
3128 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3129                                      bool force)
3130 {
3131         struct lnet_handle_md recovery_mdh;
3132
3133         LNetInvalidateMDHandle(&recovery_mdh);
3134
3135         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3136                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3137                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3138         }
3139         spin_unlock(&lpni->lpni_lock);
3140         lnet_net_unlock(cpt);
3141         if (!LNetMDHandleIsInvalid(recovery_mdh))
3142                 LNetMDUnlink(recovery_mdh);
3143         lnet_net_lock(cpt);
3144         spin_lock(&lpni->lpni_lock);
3145 }
3146
3147 static void
3148 lnet_clean_peer_ni_recoveryq(void)
3149 {
3150         struct lnet_peer_ni *lpni, *tmp;
3151
3152         lnet_net_lock(LNET_LOCK_EX);
3153
3154         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3155                                  lpni_recovery) {
3156                 list_del_init(&lpni->lpni_recovery);
3157                 spin_lock(&lpni->lpni_lock);
3158                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3159                 spin_unlock(&lpni->lpni_lock);
3160                 lnet_peer_ni_decref_locked(lpni);
3161         }
3162
3163         lnet_net_unlock(LNET_LOCK_EX);
3164 }
3165
3166 static void
3167 lnet_clean_resendqs(void)
3168 {
3169         struct lnet_msg *msg, *tmp;
3170         struct list_head msgs;
3171         int i;
3172
3173         INIT_LIST_HEAD(&msgs);
3174
3175         cfs_cpt_for_each(i, lnet_cpt_table()) {
3176                 lnet_net_lock(i);
3177                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3178                 lnet_net_unlock(i);
3179                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3180                         list_del_init(&msg->msg_list);
3181                         msg->msg_no_resend = true;
3182                         lnet_finalize(msg, -ESHUTDOWN);
3183                 }
3184         }
3185
3186         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3187 }
3188
3189 static void
3190 lnet_recover_peer_nis(void)
3191 {
3192         struct lnet_mt_event_info *ev_info;
3193         struct list_head processed_list;
3194         struct list_head local_queue;
3195         struct lnet_handle_md mdh;
3196         struct lnet_peer_ni *lpni;
3197         struct lnet_peer_ni *tmp;
3198         lnet_nid_t nid;
3199         int healthv;
3200         int rc;
3201
3202         INIT_LIST_HEAD(&local_queue);
3203         INIT_LIST_HEAD(&processed_list);
3204
3205         /*
3206          * Always use cpt 0 for locking across all interactions with
3207          * ln_mt_peerNIRecovq
3208          */
3209         lnet_net_lock(0);
3210         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3211                          &local_queue);
3212         lnet_net_unlock(0);
3213
3214         list_for_each_entry_safe(lpni, tmp, &local_queue,
3215                                  lpni_recovery) {
3216                 /*
3217                  * The same protection strategy is used here as is in the
3218                  * local recovery case.
3219                  */
3220                 lnet_net_lock(0);
3221                 healthv = atomic_read(&lpni->lpni_healthv);
3222                 spin_lock(&lpni->lpni_lock);
3223                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3224                     healthv == LNET_MAX_HEALTH_VALUE) {
3225                         list_del_init(&lpni->lpni_recovery);
3226                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3227                         spin_unlock(&lpni->lpni_lock);
3228                         lnet_peer_ni_decref_locked(lpni);
3229                         lnet_net_unlock(0);
3230                         continue;
3231                 }
3232
3233                 /*
3234                  * If the peer NI has failed recovery we must unlink the
3235                  * md. But we want to keep the peer ni on the recovery
3236                  * queue so we can try to continue recovering it
3237                  */
3238                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3239                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3240                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3241                 }
3242
3243                 spin_unlock(&lpni->lpni_lock);
3244                 lnet_net_unlock(0);
3245
3246                 /*
3247                  * NOTE: we're racing with peer deletion from user space.
3248                  * It's possible that a peer is deleted after we check its
3249                  * state. In this case the recovery can create a new peer
3250                  */
3251                 spin_lock(&lpni->lpni_lock);
3252                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3253                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3254                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3255                         spin_unlock(&lpni->lpni_lock);
3256
3257                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3258                         if (!ev_info) {
3259                                 CERROR("out of memory. Can't recover %s\n",
3260                                        libcfs_nid2str(lpni->lpni_nid));
3261                                 spin_lock(&lpni->lpni_lock);
3262                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3263                                 spin_unlock(&lpni->lpni_lock);
3264                                 continue;
3265                         }
3266
3267                         /* look at the comments in lnet_recover_local_nis() */
3268                         mdh = lpni->lpni_recovery_ping_mdh;
3269                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3270                         nid = lpni->lpni_nid;
3271                         lnet_net_lock(0);
3272                         list_del_init(&lpni->lpni_recovery);
3273                         lnet_peer_ni_decref_locked(lpni);
3274                         lnet_net_unlock(0);
3275
3276                         ev_info->mt_type = MT_TYPE_PEER_NI;
3277                         ev_info->mt_nid = nid;
3278                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3279                                             ev_info, the_lnet.ln_mt_eqh, true);
3280                         lnet_net_lock(0);
3281                         /*
3282                          * lnet_find_peer_ni_locked() grabs a refcount for
3283                          * us. No need to take it explicitly.
3284                          */
3285                         lpni = lnet_find_peer_ni_locked(nid);
3286                         if (!lpni) {
3287                                 lnet_net_unlock(0);
3288                                 LNetMDUnlink(mdh);
3289                                 continue;
3290                         }
3291
3292                         lpni->lpni_recovery_ping_mdh = mdh;
3293                         /*
3294                          * While we're unlocked the lpni could've been
3295                          * readded on the recovery queue. In this case we
3296                          * don't need to add it to the local queue, since
3297                          * it's already on there and the thread that added
3298                          * it would've incremented the refcount on the
3299                          * peer, which means we need to decref the refcount
3300                          * that was implicitly grabbed by find_peer_ni_locked.
3301                          * Otherwise, if the lpni is still not on
3302                          * the recovery queue, then we'll add it to the
3303                          * processed list.
3304                          */
3305                         if (list_empty(&lpni->lpni_recovery))
3306                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3307                         else
3308                                 lnet_peer_ni_decref_locked(lpni);
3309                         lnet_net_unlock(0);
3310
3311                         spin_lock(&lpni->lpni_lock);
3312                         if (rc)
3313                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3314                 }
3315                 spin_unlock(&lpni->lpni_lock);
3316         }
3317
3318         list_splice_init(&processed_list, &local_queue);
3319         lnet_net_lock(0);
3320         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3321         lnet_net_unlock(0);
3322 }
3323
3324 static int
3325 lnet_monitor_thread(void *arg)
3326 {
3327         time64_t recovery_timeout = 0;
3328         time64_t rsp_timeout = 0;
3329         int interval;
3330         time64_t now;
3331
3332         /*
3333          * The monitor thread takes care of the following:
3334          *  1. Checks the aliveness of routers
3335          *  2. Checks if there are messages on the resend queue to resend
3336          *     them.
3337          *  3. Check if there are any NIs on the local recovery queue and
3338          *     pings them
3339          *  4. Checks if there are any NIs on the remote recovery queue
3340          *     and pings them.
3341          */
3342         cfs_block_allsigs();
3343
3344         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3345                 now = ktime_get_real_seconds();
3346
3347                 if (lnet_router_checker_active())
3348                         lnet_check_routers();
3349
3350                 lnet_resend_pending_msgs();
3351
3352                 if (now >= rsp_timeout) {
3353                         lnet_finalize_expired_responses(false);
3354                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3355                 }
3356
3357                 if (now >= recovery_timeout) {
3358                         lnet_recover_local_nis();
3359                         lnet_recover_peer_nis();
3360                         recovery_timeout = now + lnet_recovery_interval;
3361                 }
3362
3363                 /*
3364                  * TODO do we need to check if we should sleep without
3365                  * timeout?  Technically, an active system will always
3366                  * have messages in flight so this check will always
3367                  * evaluate to false. And on an idle system do we care
3368                  * if we wake up every 1 second? Although, we've seen
3369                  * cases where we get a complaint that an idle thread
3370                  * is waking up unnecessarily.
3371                  */
3372                 interval = min(lnet_recovery_interval,
3373                                lnet_transaction_timeout / 2);
3374                 wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
3375                                                 false,
3376                                                 cfs_time_seconds(interval));
3377         }
3378
3379         /* clean up the router checker */
3380         lnet_prune_rc_data(1);
3381
3382         /* Shutting down */
3383         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3384
3385         /* signal that the monitor thread is exiting */
3386         up(&the_lnet.ln_mt_signal);
3387
3388         return 0;
3389 }
3390
3391 /*
3392  * lnet_send_ping
3393  * Sends a ping.
3394  * Returns == 0 if success
3395  * Returns > 0 if LNetMDBind or prior fails
3396  * Returns < 0 if LNetGet fails
3397  */
3398 int
3399 lnet_send_ping(lnet_nid_t dest_nid,
3400                struct lnet_handle_md *mdh, int nnis,
3401                void *user_data, struct lnet_handle_eq eqh, bool recovery)
3402 {
3403         struct lnet_md md = { NULL };
3404         struct lnet_process_id id;
3405         struct lnet_ping_buffer *pbuf;
3406         int rc;
3407
3408         if (dest_nid == LNET_NID_ANY) {
3409                 rc = -EHOSTUNREACH;
3410                 goto fail_error;
3411         }
3412
3413         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3414         if (!pbuf) {
3415                 rc = ENOMEM;
3416                 goto fail_error;
3417         }
3418
3419         /* initialize md content */
3420         md.start     = &pbuf->pb_info;
3421         md.length    = LNET_PING_INFO_SIZE(nnis);
3422         md.threshold = 2; /* GET/REPLY */
3423         md.max_size  = 0;
3424         md.options   = LNET_MD_TRUNCATE;
3425         md.user_ptr  = user_data;
3426         md.eq_handle = eqh;
3427
3428         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3429         if (rc) {
3430                 lnet_ping_buffer_decref(pbuf);
3431                 CERROR("Can't bind MD: %d\n", rc);
3432                 rc = -rc; /* change the rc to positive */
3433                 goto fail_error;
3434         }
3435         id.pid = LNET_PID_LUSTRE;
3436         id.nid = dest_nid;
3437
3438         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3439                      LNET_RESERVED_PORTAL,
3440                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3441
3442         if (rc)
3443                 goto fail_unlink_md;
3444
3445         return 0;
3446
3447 fail_unlink_md:
3448         LNetMDUnlink(*mdh);
3449         LNetInvalidateMDHandle(mdh);
3450 fail_error:
3451         return rc;
3452 }
3453
3454 static void
3455 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3456                            int status)
3457 {
3458         lnet_nid_t nid = ev_info->mt_nid;
3459
3460         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3461                 struct lnet_ni *ni;
3462
3463                 lnet_net_lock(0);
3464                 ni = lnet_nid2ni_locked(nid, 0);
3465                 if (!ni) {
3466                         lnet_net_unlock(0);
3467                         return;
3468                 }
3469                 lnet_ni_lock(ni);
3470                 ni->ni_state &= ~LNET_NI_STATE_RECOVERY_PENDING;
3471                 if (status)
3472                         ni->ni_state |= LNET_NI_STATE_RECOVERY_FAILED;
3473                 lnet_ni_unlock(ni);
3474                 lnet_net_unlock(0);
3475
3476                 if (status != 0) {
3477                         CERROR("local NI (%s) recovery failed with %d\n",
3478                                libcfs_nid2str(nid), status);
3479                         return;
3480                 }
3481                 /*
3482                  * need to increment healthv for the ni here, because in
3483                  * the lnet_finalize() path we don't have access to this
3484                  * NI. And in order to get access to it, we'll need to
3485                  * carry forward too much information.
3486                  * In the peer case, it'll naturally be incremented
3487                  */
3488                 lnet_inc_healthv(&ni->ni_healthv);
3489         } else {
3490                 struct lnet_peer_ni *lpni;
3491                 int cpt;
3492
3493                 cpt = lnet_net_lock_current();
3494                 lpni = lnet_find_peer_ni_locked(nid);
3495                 if (!lpni) {
3496                         lnet_net_unlock(cpt);
3497                         return;
3498                 }
3499                 spin_lock(&lpni->lpni_lock);
3500                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3501                 if (status)
3502                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3503                 spin_unlock(&lpni->lpni_lock);
3504                 lnet_peer_ni_decref_locked(lpni);
3505                 lnet_net_unlock(cpt);
3506
3507                 if (status != 0)
3508                         CERROR("peer NI (%s) recovery failed with %d\n",
3509                                libcfs_nid2str(nid), status);
3510         }
3511 }
3512
3513 static void
3514 lnet_mt_event_handler(struct lnet_event *event)
3515 {
3516         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3517         struct lnet_ping_buffer *pbuf;
3518
3519         /* TODO: remove assert */
3520         LASSERT(event->type == LNET_EVENT_REPLY ||
3521                 event->type == LNET_EVENT_SEND ||
3522                 event->type == LNET_EVENT_UNLINK);
3523
3524         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3525                event->status);
3526
3527         switch (event->type) {
3528         case LNET_EVENT_UNLINK:
3529                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3530                        libcfs_nid2str(ev_info->mt_nid));
3531         case LNET_EVENT_REPLY:
3532                 lnet_handle_recovery_reply(ev_info, event->status);
3533                 break;
3534         case LNET_EVENT_SEND:
3535                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3536                                libcfs_nid2str(ev_info->mt_nid),
3537                                (event->status) ? "unsuccessfully" :
3538                                "successfully", event->status);
3539                 lnet_handle_recovery_reply(ev_info, event->status);
3540                 break;
3541         default:
3542                 CERROR("Unexpected event: %d\n", event->type);
3543                 break;
3544         }
3545         if (event->unlinked) {
3546                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3547                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3548                 lnet_ping_buffer_decref(pbuf);
3549         }
3550 }
3551
3552 static int
3553 lnet_rsp_tracker_create(void)
3554 {
3555         struct list_head **rstqs;
3556         rstqs = lnet_create_array_of_queues();
3557
3558         if (!rstqs)
3559                 return -ENOMEM;
3560
3561         the_lnet.ln_mt_rstq = rstqs;
3562
3563         return 0;
3564 }
3565
3566 static void
3567 lnet_rsp_tracker_clean(void)
3568 {
3569         lnet_finalize_expired_responses(true);
3570
3571         cfs_percpt_free(the_lnet.ln_mt_rstq);
3572         the_lnet.ln_mt_rstq = NULL;
3573 }
3574
3575 int lnet_monitor_thr_start(void)
3576 {
3577         int rc = 0;
3578         struct task_struct *task;
3579
3580         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3581                 return -EALREADY;
3582
3583         rc = lnet_resendqs_create();
3584         if (rc)
3585                 return rc;
3586
3587         rc = lnet_rsp_tracker_create();
3588         if (rc)
3589                 goto clean_queues;
3590
3591         rc = LNetEQAlloc(0, lnet_mt_event_handler, &the_lnet.ln_mt_eqh);
3592         if (rc != 0) {
3593                 CERROR("Can't allocate monitor thread EQ: %d\n", rc);
3594                 goto clean_queues;
3595         }
3596
3597         /* Pre monitor thread start processing */
3598         rc = lnet_router_pre_mt_start();
3599         if (rc)
3600                 goto free_mem;
3601
3602         sema_init(&the_lnet.ln_mt_signal, 0);
3603
3604         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3605         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3606         if (IS_ERR(task)) {
3607                 rc = PTR_ERR(task);
3608                 CERROR("Can't start monitor thread: %d\n", rc);
3609                 goto clean_thread;
3610         }
3611
3612         /* post monitor thread start processing */
3613         lnet_router_post_mt_start();
3614
3615         return 0;
3616
3617 clean_thread:
3618         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3619         /* block until event callback signals exit */
3620         down(&the_lnet.ln_mt_signal);
3621         /* clean up */
3622         lnet_router_cleanup();
3623 free_mem:
3624         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3625         lnet_rsp_tracker_clean();
3626         lnet_clean_local_ni_recoveryq();
3627         lnet_clean_peer_ni_recoveryq();
3628         lnet_clean_resendqs();
3629         LNetEQFree(the_lnet.ln_mt_eqh);
3630         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3631         return rc;
3632 clean_queues:
3633         lnet_rsp_tracker_clean();
3634         lnet_clean_local_ni_recoveryq();
3635         lnet_clean_peer_ni_recoveryq();
3636         lnet_clean_resendqs();
3637         return rc;
3638 }
3639
3640 void lnet_monitor_thr_stop(void)
3641 {
3642         int rc;
3643
3644         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3645                 return;
3646
3647         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3648         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3649
3650         /* tell the monitor thread that we're shutting down */
3651         wake_up(&the_lnet.ln_mt_waitq);
3652
3653         /* block until monitor thread signals that it's done */
3654         down(&the_lnet.ln_mt_signal);
3655         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3656
3657         /* perform cleanup tasks */
3658         lnet_router_cleanup();
3659         lnet_rsp_tracker_clean();
3660         lnet_clean_local_ni_recoveryq();
3661         lnet_clean_peer_ni_recoveryq();
3662         lnet_clean_resendqs();
3663         rc = LNetEQFree(the_lnet.ln_mt_eqh);
3664         LASSERT(rc == 0);
3665         return;
3666 }
3667
3668 void
3669 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3670                   __u32 msg_type)
3671 {
3672         lnet_net_lock(cpt);
3673         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3674         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3675         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3676         lnet_net_unlock(cpt);
3677
3678         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3679 }
3680
3681 static void
3682 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3683 {
3684         struct lnet_hdr *hdr = &msg->msg_hdr;
3685
3686         if (msg->msg_wanted != 0)
3687                 lnet_setpayloadbuffer(msg);
3688
3689         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3690
3691         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3692          * it back into the ACK during lnet_finalize() */
3693         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3694                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3695
3696         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3697                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3698 }
3699
3700 static int
3701 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3702 {
3703         struct lnet_hdr         *hdr = &msg->msg_hdr;
3704         struct lnet_match_info  info;
3705         int                     rc;
3706         bool                    ready_delay;
3707
3708         /* Convert put fields to host byte order */
3709         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3710         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3711         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3712
3713         /* Primary peer NID. */
3714         info.mi_id.nid  = msg->msg_initiator;
3715         info.mi_id.pid  = hdr->src_pid;
3716         info.mi_opc     = LNET_MD_OP_PUT;
3717         info.mi_portal  = hdr->msg.put.ptl_index;
3718         info.mi_rlength = hdr->payload_length;
3719         info.mi_roffset = hdr->msg.put.offset;
3720         info.mi_mbits   = hdr->msg.put.match_bits;
3721         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
3722
3723         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3724         ready_delay = msg->msg_rx_ready_delay;
3725
3726  again:
3727         rc = lnet_ptl_match_md(&info, msg);
3728         switch (rc) {
3729         default:
3730                 LBUG();
3731
3732         case LNET_MATCHMD_OK:
3733                 lnet_recv_put(ni, msg);
3734                 return 0;
3735
3736         case LNET_MATCHMD_NONE:
3737                 if (ready_delay)
3738                         /* no eager_recv or has already called it, should
3739                          * have been attached on delayed list */
3740                         return 0;
3741
3742                 rc = lnet_ni_eager_recv(ni, msg);
3743                 if (rc == 0) {
3744                         ready_delay = true;
3745                         goto again;
3746                 }
3747                 /* fall through */
3748
3749         case LNET_MATCHMD_DROP:
3750                 CNETERR("Dropping PUT from %s portal %d match %llu"
3751                         " offset %d length %d: %d\n",
3752                         libcfs_id2str(info.mi_id), info.mi_portal,
3753                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3754
3755                 return -ENOENT; /* -ve: OK but no match */
3756         }
3757 }
3758
3759 static int
3760 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3761 {
3762         struct lnet_match_info info;
3763         struct lnet_hdr *hdr = &msg->msg_hdr;
3764         struct lnet_process_id source_id;
3765         struct lnet_handle_wire reply_wmd;
3766         int rc;
3767
3768         /* Convert get fields to host byte order */
3769         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3770         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3771         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3772         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3773
3774         source_id.nid = hdr->src_nid;
3775         source_id.pid = hdr->src_pid;
3776         /* Primary peer NID */
3777         info.mi_id.nid  = msg->msg_initiator;
3778         info.mi_id.pid  = hdr->src_pid;
3779         info.mi_opc     = LNET_MD_OP_GET;
3780         info.mi_portal  = hdr->msg.get.ptl_index;