projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge tag 'regulator-v3.13' of git://git.kernel.org/pub/scm/linux/kernel/git/broonie...
[deliverable/linux.git] / drivers / staging / lustre / 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.sun.com/software/products/lustre/docs/GPLv2.pdf
19 *
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
22 * have any questions.
23 *
24 * GPL HEADER END
25 */
26 /*
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
29 *
30 * Copyright (c) 2011, 2012, Intel Corporation.
31 */
32 /*
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
35 *
36 * lnet/lnet/lib-move.c
37 *
38 * Data movement routines
39 */
40
41 #define DEBUG_SUBSYSTEM S_LNET
42
43 #include <linux/lnet/lib-lnet.h>
44
45 static int local_nid_dist_zero = 1;
46 CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444,
47 "Reserved");
48
49 int
50 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
51 {
52 lnet_test_peer_t *tp;
53 struct list_head *el;
54 struct list_head *next;
55 struct list_head cull;
56
57 LASSERT(the_lnet.ln_init);
58
59 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
60 if (threshold != 0) {
61 /* Adding a new entry */
62 LIBCFS_ALLOC(tp, sizeof(*tp));
63 if (tp == NULL)
64 return -ENOMEM;
65
66 tp->tp_nid = nid;
67 tp->tp_threshold = threshold;
68
69 lnet_net_lock(0);
70 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
71 lnet_net_unlock(0);
72 return 0;
73 }
74
75 /* removing entries */
76 INIT_LIST_HEAD(&cull);
77
78 lnet_net_lock(0);
79
80 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
81 tp = list_entry(el, lnet_test_peer_t, tp_list);
82
83 if (tp->tp_threshold == 0 || /* needs culling anyway */
84 nid == LNET_NID_ANY || /* removing all entries */
85 tp->tp_nid == nid) { /* matched this one */
86 list_del(&tp->tp_list);
87 list_add(&tp->tp_list, &cull);
88 }
89 }
90
91 lnet_net_unlock(0);
92
93 while (!list_empty(&cull)) {
94 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
95
96 list_del(&tp->tp_list);
97 LIBCFS_FREE(tp, sizeof(*tp));
98 }
99 return 0;
100 }
101
102 static int
103 fail_peer(lnet_nid_t nid, int outgoing)
104 {
105 lnet_test_peer_t *tp;
106 struct list_head *el;
107 struct list_head *next;
108 struct list_head cull;
109 int fail = 0;
110
111 INIT_LIST_HEAD(&cull);
112
113 /* NB: use lnet_net_lock(0) to serialize operations on test peers */
114 lnet_net_lock(0);
115
116 list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
117 tp = list_entry(el, lnet_test_peer_t, tp_list);
118
119 if (tp->tp_threshold == 0) {
120 /* zombie entry */
121 if (outgoing) {
122 /* only cull zombies on outgoing tests,
123 * since we may be at interrupt priority on
124 * incoming messages. */
125 list_del(&tp->tp_list);
126 list_add(&tp->tp_list, &cull);
127 }
128 continue;
129 }
130
131 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
132 nid == tp->tp_nid) { /* fail this peer */
133 fail = 1;
134
135 if (tp->tp_threshold != LNET_MD_THRESH_INF) {
136 tp->tp_threshold--;
137 if (outgoing &&
138 tp->tp_threshold == 0) {
139 /* see above */
140 list_del(&tp->tp_list);
141 list_add(&tp->tp_list, &cull);
142 }
143 }
144 break;
145 }
146 }
147
148 lnet_net_unlock(0);
149
150 while (!list_empty(&cull)) {
151 tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
152 list_del(&tp->tp_list);
153
154 LIBCFS_FREE(tp, sizeof(*tp));
155 }
156
157 return fail;
158 }
159
160 unsigned int
161 lnet_iov_nob(unsigned int niov, struct iovec *iov)
162 {
163 unsigned int nob = 0;
164
165 while (niov-- > 0)
166 nob += (iov++)->iov_len;
167
168 return nob;
169 }
170 EXPORT_SYMBOL(lnet_iov_nob);
171
172 void
173 lnet_copy_iov2iov(unsigned int ndiov, struct iovec *diov, unsigned int doffset,
174 unsigned int nsiov, struct iovec *siov, unsigned int soffset,
175 unsigned int nob)
176 {
177 /* NB diov, siov are READ-ONLY */
178 unsigned int this_nob;
179
180 if (nob == 0)
181 return;
182
183 /* skip complete frags before 'doffset' */
184 LASSERT(ndiov > 0);
185 while (doffset >= diov->iov_len) {
186 doffset -= diov->iov_len;
187 diov++;
188 ndiov--;
189 LASSERT(ndiov > 0);
190 }
191
192 /* skip complete frags before 'soffset' */
193 LASSERT(nsiov > 0);
194 while (soffset >= siov->iov_len) {
195 soffset -= siov->iov_len;
196 siov++;
197 nsiov--;
198 LASSERT(nsiov > 0);
199 }
200
201 do {
202 LASSERT(ndiov > 0);
203 LASSERT(nsiov > 0);
204 this_nob = MIN(diov->iov_len - doffset,
205 siov->iov_len - soffset);
206 this_nob = MIN(this_nob, nob);
207
208 memcpy((char *)diov->iov_base + doffset,
209 (char *)siov->iov_base + soffset, this_nob);
210 nob -= this_nob;
211
212 if (diov->iov_len > doffset + this_nob) {
213 doffset += this_nob;
214 } else {
215 diov++;
216 ndiov--;
217 doffset = 0;
218 }
219
220 if (siov->iov_len > soffset + this_nob) {
221 soffset += this_nob;
222 } else {
223 siov++;
224 nsiov--;
225 soffset = 0;
226 }
227 } while (nob > 0);
228 }
229 EXPORT_SYMBOL(lnet_copy_iov2iov);
230
231 int
232 lnet_extract_iov(int dst_niov, struct iovec *dst,
233 int src_niov, struct iovec *src,
234 unsigned int offset, unsigned int len)
235 {
236 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
237 * for exactly 'len' bytes, and return the number of entries.
238 * NB not destructive to 'src' */
239 unsigned int frag_len;
240 unsigned int niov;
241
242 if (len == 0) /* no data => */
243 return 0; /* no frags */
244
245 LASSERT(src_niov > 0);
246 while (offset >= src->iov_len) { /* skip initial frags */
247 offset -= src->iov_len;
248 src_niov--;
249 src++;
250 LASSERT(src_niov > 0);
251 }
252
253 niov = 1;
254 for (;;) {
255 LASSERT(src_niov > 0);
256 LASSERT((int)niov <= dst_niov);
257
258 frag_len = src->iov_len - offset;
259 dst->iov_base = ((char *)src->iov_base) + offset;
260
261 if (len <= frag_len) {
262 dst->iov_len = len;
263 return niov;
264 }
265
266 dst->iov_len = frag_len;
267
268 len -= frag_len;
269 dst++;
270 src++;
271 niov++;
272 src_niov--;
273 offset = 0;
274 }
275 }
276 EXPORT_SYMBOL(lnet_extract_iov);
277
278
279 unsigned int
280 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
281 {
282 unsigned int nob = 0;
283
284 while (niov-- > 0)
285 nob += (kiov++)->kiov_len;
286
287 return nob;
288 }
289 EXPORT_SYMBOL(lnet_kiov_nob);
290
291 void
292 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
293 unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
294 unsigned int nob)
295 {
296 /* NB diov, siov are READ-ONLY */
297 unsigned int this_nob;
298 char *daddr = NULL;
299 char *saddr = NULL;
300
301 if (nob == 0)
302 return;
303
304 LASSERT(!in_interrupt());
305
306 LASSERT(ndiov > 0);
307 while (doffset >= diov->kiov_len) {
308 doffset -= diov->kiov_len;
309 diov++;
310 ndiov--;
311 LASSERT(ndiov > 0);
312 }
313
314 LASSERT(nsiov > 0);
315 while (soffset >= siov->kiov_len) {
316 soffset -= siov->kiov_len;
317 siov++;
318 nsiov--;
319 LASSERT(nsiov > 0);
320 }
321
322 do {
323 LASSERT(ndiov > 0);
324 LASSERT(nsiov > 0);
325 this_nob = MIN(diov->kiov_len - doffset,
326 siov->kiov_len - soffset);
327 this_nob = MIN(this_nob, nob);
328
329 if (daddr == NULL)
330 daddr = ((char *)kmap(diov->kiov_page)) +
331 diov->kiov_offset + doffset;
332 if (saddr == NULL)
333 saddr = ((char *)kmap(siov->kiov_page)) +
334 siov->kiov_offset + soffset;
335
336 /* Vanishing risk of kmap deadlock when mapping 2 pages.
337 * However in practice at least one of the kiovs will be mapped
338 * kernel pages and the map/unmap will be NOOPs */
339
340 memcpy(daddr, saddr, this_nob);
341 nob -= this_nob;
342
343 if (diov->kiov_len > doffset + this_nob) {
344 daddr += this_nob;
345 doffset += this_nob;
346 } else {
347 kunmap(diov->kiov_page);
348 daddr = NULL;
349 diov++;
350 ndiov--;
351 doffset = 0;
352 }
353
354 if (siov->kiov_len > soffset + this_nob) {
355 saddr += this_nob;
356 soffset += this_nob;
357 } else {
358 kunmap(siov->kiov_page);
359 saddr = NULL;
360 siov++;
361 nsiov--;
362 soffset = 0;
363 }
364 } while (nob > 0);
365
366 if (daddr != NULL)
367 kunmap(diov->kiov_page);
368 if (saddr != NULL)
369 kunmap(siov->kiov_page);
370 }
371 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
372
373 void
374 lnet_copy_kiov2iov(unsigned int niov, struct iovec *iov, unsigned int iovoffset,
375 unsigned int nkiov, lnet_kiov_t *kiov,
376 unsigned int kiovoffset, unsigned int nob)
377 {
378 /* NB iov, kiov are READ-ONLY */
379 unsigned int this_nob;
380 char *addr = NULL;
381
382 if (nob == 0)
383 return;
384
385 LASSERT(!in_interrupt());
386
387 LASSERT(niov > 0);
388 while (iovoffset >= iov->iov_len) {
389 iovoffset -= iov->iov_len;
390 iov++;
391 niov--;
392 LASSERT(niov > 0);
393 }
394
395 LASSERT(nkiov > 0);
396 while (kiovoffset >= kiov->kiov_len) {
397 kiovoffset -= kiov->kiov_len;
398 kiov++;
399 nkiov--;
400 LASSERT(nkiov > 0);
401 }
402
403 do {
404 LASSERT(niov > 0);
405 LASSERT(nkiov > 0);
406 this_nob = MIN(iov->iov_len - iovoffset,
407 kiov->kiov_len - kiovoffset);
408 this_nob = MIN(this_nob, nob);
409
410 if (addr == NULL)
411 addr = ((char *)kmap(kiov->kiov_page)) +
412 kiov->kiov_offset + kiovoffset;
413
414 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
415 nob -= this_nob;
416
417 if (iov->iov_len > iovoffset + this_nob) {
418 iovoffset += this_nob;
419 } else {
420 iov++;
421 niov--;
422 iovoffset = 0;
423 }
424
425 if (kiov->kiov_len > kiovoffset + this_nob) {
426 addr += this_nob;
427 kiovoffset += this_nob;
428 } else {
429 kunmap(kiov->kiov_page);
430 addr = NULL;
431 kiov++;
432 nkiov--;
433 kiovoffset = 0;
434 }
435
436 } while (nob > 0);
437
438 if (addr != NULL)
439 kunmap(kiov->kiov_page);
440 }
441 EXPORT_SYMBOL(lnet_copy_kiov2iov);
442
443 void
444 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov,
445 unsigned int kiovoffset, unsigned int niov,
446 struct iovec *iov, unsigned int iovoffset,
447 unsigned int nob)
448 {
449 /* NB kiov, iov are READ-ONLY */
450 unsigned int this_nob;
451 char *addr = NULL;
452
453 if (nob == 0)
454 return;
455
456 LASSERT(!in_interrupt());
457
458 LASSERT(nkiov > 0);
459 while (kiovoffset >= kiov->kiov_len) {
460 kiovoffset -= kiov->kiov_len;
461 kiov++;
462 nkiov--;
463 LASSERT(nkiov > 0);
464 }
465
466 LASSERT(niov > 0);
467 while (iovoffset >= iov->iov_len) {
468 iovoffset -= iov->iov_len;
469 iov++;
470 niov--;
471 LASSERT(niov > 0);
472 }
473
474 do {
475 LASSERT(nkiov > 0);
476 LASSERT(niov > 0);
477 this_nob = MIN(kiov->kiov_len - kiovoffset,
478 iov->iov_len - iovoffset);
479 this_nob = MIN(this_nob, nob);
480
481 if (addr == NULL)
482 addr = ((char *)kmap(kiov->kiov_page)) +
483 kiov->kiov_offset + kiovoffset;
484
485 memcpy(addr, (char *)iov->iov_base + iovoffset, this_nob);
486 nob -= this_nob;
487
488 if (kiov->kiov_len > kiovoffset + this_nob) {
489 addr += this_nob;
490 kiovoffset += this_nob;
491 } else {
492 kunmap(kiov->kiov_page);
493 addr = NULL;
494 kiov++;
495 nkiov--;
496 kiovoffset = 0;
497 }
498
499 if (iov->iov_len > iovoffset + this_nob) {
500 iovoffset += this_nob;
501 } else {
502 iov++;
503 niov--;
504 iovoffset = 0;
505 }
506 } while (nob > 0);
507
508 if (addr != NULL)
509 kunmap(kiov->kiov_page);
510 }
511 EXPORT_SYMBOL(lnet_copy_iov2kiov);
512
513 int
514 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
515 int src_niov, lnet_kiov_t *src,
516 unsigned int offset, unsigned int len)
517 {
518 /* Initialise 'dst' to the subset of 'src' starting at 'offset',
519 * for exactly 'len' bytes, and return the number of entries.
520 * NB not destructive to 'src' */
521 unsigned int frag_len;
522 unsigned int niov;
523
524 if (len == 0) /* no data => */
525 return 0; /* no frags */
526
527 LASSERT(src_niov > 0);
528 while (offset >= src->kiov_len) { /* skip initial frags */
529 offset -= src->kiov_len;
530 src_niov--;
531 src++;
532 LASSERT(src_niov > 0);
533 }
534
535 niov = 1;
536 for (;;) {
537 LASSERT(src_niov > 0);
538 LASSERT((int)niov <= dst_niov);
539
540 frag_len = src->kiov_len - offset;
541 dst->kiov_page = src->kiov_page;
542 dst->kiov_offset = src->kiov_offset + offset;
543
544 if (len <= frag_len) {
545 dst->kiov_len = len;
546 LASSERT(dst->kiov_offset + dst->kiov_len
547 <= PAGE_CACHE_SIZE);
548 return niov;
549 }
550
551 dst->kiov_len = frag_len;
552 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE);
553
554 len -= frag_len;
555 dst++;
556 src++;
557 niov++;
558 src_niov--;
559 offset = 0;
560 }
561 }
562 EXPORT_SYMBOL(lnet_extract_kiov);
563
564 void
565 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
566 unsigned int offset, unsigned int mlen, unsigned int rlen)
567 {
568 unsigned int niov = 0;
569 struct iovec *iov = NULL;
570 lnet_kiov_t *kiov = NULL;
571 int rc;
572
573 LASSERT(!in_interrupt());
574 LASSERT(mlen == 0 || msg != NULL);
575
576 if (msg != NULL) {
577 LASSERT(msg->msg_receiving);
578 LASSERT(!msg->msg_sending);
579 LASSERT(rlen == msg->msg_len);
580 LASSERT(mlen <= msg->msg_len);
581 LASSERT(msg->msg_offset == offset);
582 LASSERT(msg->msg_wanted == mlen);
583
584 msg->msg_receiving = 0;
585
586 if (mlen != 0) {
587 niov = msg->msg_niov;
588 iov = msg->msg_iov;
589 kiov = msg->msg_kiov;
590
591 LASSERT(niov > 0);
592 LASSERT((iov == NULL) != (kiov == NULL));
593 }
594 }
595
596 rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
597 niov, iov, kiov, offset, mlen, rlen);
598 if (rc < 0)
599 lnet_finalize(ni, msg, rc);
600 }
601
602 void
603 lnet_setpayloadbuffer(lnet_msg_t *msg)
604 {
605 lnet_libmd_t *md = msg->msg_md;
606
607 LASSERT(msg->msg_len > 0);
608 LASSERT(!msg->msg_routing);
609 LASSERT(md != NULL);
610 LASSERT(msg->msg_niov == 0);
611 LASSERT(msg->msg_iov == NULL);
612 LASSERT(msg->msg_kiov == NULL);
613
614 msg->msg_niov = md->md_niov;
615 if ((md->md_options & LNET_MD_KIOV) != 0)
616 msg->msg_kiov = md->md_iov.kiov;
617 else
618 msg->msg_iov = md->md_iov.iov;
619 }
620
621 void
622 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
623 unsigned int offset, unsigned int len)
624 {
625 msg->msg_type = type;
626 msg->msg_target = target;
627 msg->msg_len = len;
628 msg->msg_offset = offset;
629
630 if (len != 0)
631 lnet_setpayloadbuffer(msg);
632
633 memset(&msg->msg_hdr, 0, sizeof(msg->msg_hdr));
634 msg->msg_hdr.type = cpu_to_le32(type);
635 msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
636 msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
637 /* src_nid will be set later */
638 msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
639 msg->msg_hdr.payload_length = cpu_to_le32(len);
640 }
641
642 void
643 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
644 {
645 void *priv = msg->msg_private;
646 int rc;
647
648 LASSERT(!in_interrupt());
649 LASSERT(LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
650 (msg->msg_txcredit && msg->msg_peertxcredit));
651
652 rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
653 if (rc < 0)
654 lnet_finalize(ni, msg, rc);
655 }
656
657 int
658 lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
659 {
660 int rc;
661
662 LASSERT(!msg->msg_sending);
663 LASSERT(msg->msg_receiving);
664 LASSERT(!msg->msg_rx_ready_delay);
665 LASSERT(ni->ni_lnd->lnd_eager_recv != NULL);
666
667 msg->msg_rx_ready_delay = 1;
668 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
669 &msg->msg_private);
670 if (rc != 0) {
671 CERROR("recv from %s / send to %s aborted: "
672 "eager_recv failed %d\n",
673 libcfs_nid2str(msg->msg_rxpeer->lp_nid),
674 libcfs_id2str(msg->msg_target), rc);
675 LASSERT(rc < 0); /* required by my callers */
676 }
677
678 return rc;
679 }
680
681 /* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */
682 void
683 lnet_ni_query_locked(lnet_ni_t *ni, lnet_peer_t *lp)
684 {
685 cfs_time_t last_alive = 0;
686
687 LASSERT(lnet_peer_aliveness_enabled(lp));
688 LASSERT(ni->ni_lnd->lnd_query != NULL);
689
690 lnet_net_unlock(lp->lp_cpt);
691 (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive);
692 lnet_net_lock(lp->lp_cpt);
693
694 lp->lp_last_query = cfs_time_current();
695
696 if (last_alive != 0) /* NI has updated timestamp */
697 lp->lp_last_alive = last_alive;
698 }
699
700 /* NB: always called with lnet_net_lock held */
701 static inline int
702 lnet_peer_is_alive(lnet_peer_t *lp, cfs_time_t now)
703 {
704 int alive;
705 cfs_time_t deadline;
706
707 LASSERT(lnet_peer_aliveness_enabled(lp));
708
709 /* Trust lnet_notify() if it has more recent aliveness news, but
710 * ignore the initial assumed death (see lnet_peers_start_down()).
711 */
712 if (!lp->lp_alive && lp->lp_alive_count > 0 &&
713 cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive))
714 return 0;
715
716 deadline = cfs_time_add(lp->lp_last_alive,
717 cfs_time_seconds(lp->lp_ni->ni_peertimeout));
718 alive = cfs_time_after(deadline, now);
719
720 /* Update obsolete lp_alive except for routers assumed to be dead
721 * initially, because router checker would update aliveness in this
722 * case, and moreover lp_last_alive at peer creation is assumed.
723 */
724 if (alive && !lp->lp_alive &&
725 !(lnet_isrouter(lp) && lp->lp_alive_count == 0))
726 lnet_notify_locked(lp, 0, 1, lp->lp_last_alive);
727
728 return alive;
729 }
730
731
732 /* NB: returns 1 when alive, 0 when dead, negative when error;
733 * may drop the lnet_net_lock */
734 int
735 lnet_peer_alive_locked(lnet_peer_t *lp)
736 {
737 cfs_time_t now = cfs_time_current();
738
739 if (!lnet_peer_aliveness_enabled(lp))
740 return -ENODEV;
741
742 if (lnet_peer_is_alive(lp, now))
743 return 1;
744
745 /* Peer appears dead, but we should avoid frequent NI queries (at
746 * most once per lnet_queryinterval seconds). */
747 if (lp->lp_last_query != 0) {
748 static const int lnet_queryinterval = 1;
749
750 cfs_time_t next_query =
751 cfs_time_add(lp->lp_last_query,
752 cfs_time_seconds(lnet_queryinterval));
753
754 if (cfs_time_before(now, next_query)) {
755 if (lp->lp_alive)
756 CWARN("Unexpected aliveness of peer %s: "
757 "%d < %d (%d/%d)\n",
758 libcfs_nid2str(lp->lp_nid),
759 (int)now, (int)next_query,
760 lnet_queryinterval,
761 lp->lp_ni->ni_peertimeout);
762 return 0;
763 }
764 }
765
766 /* query NI for latest aliveness news */
767 lnet_ni_query_locked(lp->lp_ni, lp);
768
769 if (lnet_peer_is_alive(lp, now))
770 return 1;
771
772 lnet_notify_locked(lp, 0, 0, lp->lp_last_alive);
773 return 0;
774 }
775
776 int
777 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
778 {
779 /* lnet_send is going to lnet_net_unlock immediately after this,
780 * so it sets do_send FALSE and I don't do the unlock/send/lock bit.
781 * I return EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer
782 * appears dead, and 0 if sent or OK to send */
783 struct lnet_peer *lp = msg->msg_txpeer;
784 struct lnet_ni *ni = lp->lp_ni;
785 struct lnet_tx_queue *tq;
786 int cpt;
787
788 /* non-lnet_send() callers have checked before */
789 LASSERT(!do_send || msg->msg_tx_delayed);
790 LASSERT(!msg->msg_receiving);
791 LASSERT(msg->msg_tx_committed);
792
793 cpt = msg->msg_tx_cpt;
794 tq = ni->ni_tx_queues[cpt];
795
796 /* NB 'lp' is always the next hop */
797 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
798 lnet_peer_alive_locked(lp) == 0) {
799 the_lnet.ln_counters[cpt]->drop_count++;
800 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
801 lnet_net_unlock(cpt);
802
803 CNETERR("Dropping message for %s: peer not alive\n",
804 libcfs_id2str(msg->msg_target));
805 if (do_send)
806 lnet_finalize(ni, msg, -EHOSTUNREACH);
807
808 lnet_net_lock(cpt);
809 return EHOSTUNREACH;
810 }
811
812 if (!msg->msg_peertxcredit) {
813 LASSERT((lp->lp_txcredits < 0) ==
814 !list_empty(&lp->lp_txq));
815
816 msg->msg_peertxcredit = 1;
817 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
818 lp->lp_txcredits--;
819
820 if (lp->lp_txcredits < lp->lp_mintxcredits)
821 lp->lp_mintxcredits = lp->lp_txcredits;
822
823 if (lp->lp_txcredits < 0) {
824 msg->msg_tx_delayed = 1;
825 list_add_tail(&msg->msg_list, &lp->lp_txq);
826 return EAGAIN;
827 }
828 }
829
830 if (!msg->msg_txcredit) {
831 LASSERT((tq->tq_credits < 0) ==
832 !list_empty(&tq->tq_delayed));
833
834 msg->msg_txcredit = 1;
835 tq->tq_credits--;
836
837 if (tq->tq_credits < tq->tq_credits_min)
838 tq->tq_credits_min = tq->tq_credits;
839
840 if (tq->tq_credits < 0) {
841 msg->msg_tx_delayed = 1;
842 list_add_tail(&msg->msg_list, &tq->tq_delayed);
843 return EAGAIN;
844 }
845 }
846
847 if (do_send) {
848 lnet_net_unlock(cpt);
849 lnet_ni_send(ni, msg);
850 lnet_net_lock(cpt);
851 }
852 return 0;
853 }
854
855
856 lnet_rtrbufpool_t *
857 lnet_msg2bufpool(lnet_msg_t *msg)
858 {
859 lnet_rtrbufpool_t *rbp;
860 int cpt;
861
862 LASSERT(msg->msg_rx_committed);
863
864 cpt = msg->msg_rx_cpt;
865 rbp = &the_lnet.ln_rtrpools[cpt][0];
866
867 LASSERT(msg->msg_len <= LNET_MTU);
868 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_CACHE_SIZE) {
869 rbp++;
870 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
871 }
872
873 return rbp;
874 }
875
876 int
877 lnet_post_routed_recv_locked(lnet_msg_t *msg, int do_recv)
878 {
879 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
880 * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
881 * return EAGAIN if msg blocked and 0 if received or OK to receive */
882 lnet_peer_t *lp = msg->msg_rxpeer;
883 lnet_rtrbufpool_t *rbp;
884 lnet_rtrbuf_t *rb;
885
886 LASSERT(msg->msg_iov == NULL);
887 LASSERT(msg->msg_kiov == NULL);
888 LASSERT(msg->msg_niov == 0);
889 LASSERT(msg->msg_routing);
890 LASSERT(msg->msg_receiving);
891 LASSERT(!msg->msg_sending);
892
893 /* non-lnet_parse callers only receive delayed messages */
894 LASSERT(!do_recv || msg->msg_rx_delayed);
895
896 if (!msg->msg_peerrtrcredit) {
897 LASSERT((lp->lp_rtrcredits < 0) ==
898 !list_empty(&lp->lp_rtrq));
899
900 msg->msg_peerrtrcredit = 1;
901 lp->lp_rtrcredits--;
902 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
903 lp->lp_minrtrcredits = lp->lp_rtrcredits;
904
905 if (lp->lp_rtrcredits < 0) {
906 /* must have checked eager_recv before here */
907 LASSERT(msg->msg_rx_ready_delay);
908 msg->msg_rx_delayed = 1;
909 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
910 return EAGAIN;
911 }
912 }
913
914 rbp = lnet_msg2bufpool(msg);
915
916 if (!msg->msg_rtrcredit) {
917 LASSERT((rbp->rbp_credits < 0) ==
918 !list_empty(&rbp->rbp_msgs));
919
920 msg->msg_rtrcredit = 1;
921 rbp->rbp_credits--;
922 if (rbp->rbp_credits < rbp->rbp_mincredits)
923 rbp->rbp_mincredits = rbp->rbp_credits;
924
925 if (rbp->rbp_credits < 0) {
926 /* must have checked eager_recv before here */
927 LASSERT(msg->msg_rx_ready_delay);
928 msg->msg_rx_delayed = 1;
929 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
930 return EAGAIN;
931 }
932 }
933
934 LASSERT(!list_empty(&rbp->rbp_bufs));
935 rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
936 list_del(&rb->rb_list);
937
938 msg->msg_niov = rbp->rbp_npages;
939 msg->msg_kiov = &rb->rb_kiov[0];
940
941 if (do_recv) {
942 int cpt = msg->msg_rx_cpt;
943
944 lnet_net_unlock(cpt);
945 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
946 0, msg->msg_len, msg->msg_len);
947 lnet_net_lock(cpt);
948 }
949 return 0;
950 }
951
952 void
953 lnet_return_tx_credits_locked(lnet_msg_t *msg)
954 {
955 lnet_peer_t *txpeer = msg->msg_txpeer;
956 lnet_msg_t *msg2;
957
958 if (msg->msg_txcredit) {
959 struct lnet_ni *ni = txpeer->lp_ni;
960 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
961
962 /* give back NI txcredits */
963 msg->msg_txcredit = 0;
964
965 LASSERT((tq->tq_credits < 0) ==
966 !list_empty(&tq->tq_delayed));
967
968 tq->tq_credits++;
969 if (tq->tq_credits <= 0) {
970 msg2 = list_entry(tq->tq_delayed.next,
971 lnet_msg_t, msg_list);
972 list_del(&msg2->msg_list);
973
974 LASSERT(msg2->msg_txpeer->lp_ni == ni);
975 LASSERT(msg2->msg_tx_delayed);
976
977 (void) lnet_post_send_locked(msg2, 1);
978 }
979 }
980
981 if (msg->msg_peertxcredit) {
982 /* give back peer txcredits */
983 msg->msg_peertxcredit = 0;
984
985 LASSERT((txpeer->lp_txcredits < 0) ==
986 !list_empty(&txpeer->lp_txq));
987
988 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
989 LASSERT(txpeer->lp_txqnob >= 0);
990
991 txpeer->lp_txcredits++;
992 if (txpeer->lp_txcredits <= 0) {
993 msg2 = list_entry(txpeer->lp_txq.next,
994 lnet_msg_t, msg_list);
995 list_del(&msg2->msg_list);
996
997 LASSERT(msg2->msg_txpeer == txpeer);
998 LASSERT(msg2->msg_tx_delayed);
999
1000 (void) lnet_post_send_locked(msg2, 1);
1001 }
1002 }
1003
1004 if (txpeer != NULL) {
1005 msg->msg_txpeer = NULL;
1006 lnet_peer_decref_locked(txpeer);
1007 }
1008 }
1009
1010 void
1011 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1012 {
1013 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1014 lnet_msg_t *msg2;
1015
1016 if (msg->msg_rtrcredit) {
1017 /* give back global router credits */
1018 lnet_rtrbuf_t *rb;
1019 lnet_rtrbufpool_t *rbp;
1020
1021 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1022 * there until it gets one allocated, or aborts the wait
1023 * itself */
1024 LASSERT(msg->msg_kiov != NULL);
1025
1026 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1027 rbp = rb->rb_pool;
1028 LASSERT(rbp == lnet_msg2bufpool(msg));
1029
1030 msg->msg_kiov = NULL;
1031 msg->msg_rtrcredit = 0;
1032
1033 LASSERT((rbp->rbp_credits < 0) ==
1034 !list_empty(&rbp->rbp_msgs));
1035 LASSERT((rbp->rbp_credits > 0) ==
1036 !list_empty(&rbp->rbp_bufs));
1037
1038 list_add(&rb->rb_list, &rbp->rbp_bufs);
1039 rbp->rbp_credits++;
1040 if (rbp->rbp_credits <= 0) {
1041 msg2 = list_entry(rbp->rbp_msgs.next,
1042 lnet_msg_t, msg_list);
1043 list_del(&msg2->msg_list);
1044
1045 (void) lnet_post_routed_recv_locked(msg2, 1);
1046 }
1047 }
1048
1049 if (msg->msg_peerrtrcredit) {
1050 /* give back peer router credits */
1051 msg->msg_peerrtrcredit = 0;
1052
1053 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1054 !list_empty(&rxpeer->lp_rtrq));
1055
1056 rxpeer->lp_rtrcredits++;
1057 if (rxpeer->lp_rtrcredits <= 0) {
1058 msg2 = list_entry(rxpeer->lp_rtrq.next,
1059 lnet_msg_t, msg_list);
1060 list_del(&msg2->msg_list);
1061
1062 (void) lnet_post_routed_recv_locked(msg2, 1);
1063 }
1064 }
1065 if (rxpeer != NULL) {
1066 msg->msg_rxpeer = NULL;
1067 lnet_peer_decref_locked(rxpeer);
1068 }
1069 }
1070
1071 static int
1072 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1073 {
1074 lnet_peer_t *p1 = r1->lr_gateway;
1075 lnet_peer_t *p2 = r2->lr_gateway;
1076
1077 if (r1->lr_hops < r2->lr_hops)
1078 return 1;
1079
1080 if (r1->lr_hops > r2->lr_hops)
1081 return -1;
1082
1083 if (p1->lp_txqnob < p2->lp_txqnob)
1084 return 1;
1085
1086 if (p1->lp_txqnob > p2->lp_txqnob)
1087 return -1;
1088
1089 if (p1->lp_txcredits > p2->lp_txcredits)
1090 return 1;
1091
1092 if (p1->lp_txcredits < p2->lp_txcredits)
1093 return -1;
1094
1095 if (r1->lr_seq - r2->lr_seq <= 0)
1096 return 1;
1097
1098 return -1;
1099 }
1100
1101 static lnet_peer_t *
1102 lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target, lnet_nid_t rtr_nid)
1103 {
1104 lnet_remotenet_t *rnet;
1105 lnet_route_t *rtr;
1106 lnet_route_t *rtr_best;
1107 lnet_route_t *rtr_last;
1108 struct lnet_peer *lp_best;
1109 struct lnet_peer *lp;
1110 int rc;
1111
1112 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1113 * rtr_nid nid, otherwise find the best gateway I can use */
1114
1115 rnet = lnet_find_net_locked(LNET_NIDNET(target));
1116 if (rnet == NULL)
1117 return NULL;
1118
1119 lp_best = NULL;
1120 rtr_best = rtr_last = NULL;
1121 list_for_each_entry(rtr, &rnet->lrn_routes, lr_list) {
1122 lp = rtr->lr_gateway;
1123
1124 if (!lp->lp_alive || /* gateway is down */
1125 ((lp->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0 &&
1126 rtr->lr_downis != 0)) /* NI to target is down */
1127 continue;
1128
1129 if (ni != NULL && lp->lp_ni != ni)
1130 continue;
1131
1132 if (lp->lp_nid == rtr_nid) /* it's pre-determined router */
1133 return lp;
1134
1135 if (lp_best == NULL) {
1136 rtr_best = rtr_last = rtr;
1137 lp_best = lp;
1138 continue;
1139 }
1140
1141 /* no protection on below fields, but it's harmless */
1142 if (rtr_last->lr_seq - rtr->lr_seq < 0)
1143 rtr_last = rtr;
1144
1145 rc = lnet_compare_routes(rtr, rtr_best);
1146 if (rc < 0)
1147 continue;
1148
1149 rtr_best = rtr;
1150 lp_best = lp;
1151 }
1152
1153 /* set sequence number on the best router to the latest sequence + 1
1154 * so we can round-robin all routers, it's race and inaccurate but
1155 * harmless and functional */
1156 if (rtr_best != NULL)
1157 rtr_best->lr_seq = rtr_last->lr_seq + 1;
1158 return lp_best;
1159 }
1160
1161 int
1162 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1163 {
1164 lnet_nid_t dst_nid = msg->msg_target.nid;
1165 struct lnet_ni *src_ni;
1166 struct lnet_ni *local_ni;
1167 struct lnet_peer *lp;
1168 int cpt;
1169 int cpt2;
1170 int rc;
1171
1172 /* NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1173 * but we might want to use pre-determined router for ACK/REPLY
1174 * in the future */
1175 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1176 LASSERT(msg->msg_txpeer == NULL);
1177 LASSERT(!msg->msg_sending);
1178 LASSERT(!msg->msg_target_is_router);
1179 LASSERT(!msg->msg_receiving);
1180
1181 msg->msg_sending = 1;
1182
1183 LASSERT(!msg->msg_tx_committed);
1184 cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid);
1185 again:
1186 lnet_net_lock(cpt);
1187
1188 if (the_lnet.ln_shutdown) {
1189 lnet_net_unlock(cpt);
1190 return -ESHUTDOWN;
1191 }
1192
1193 if (src_nid == LNET_NID_ANY) {
1194 src_ni = NULL;
1195 } else {
1196 src_ni = lnet_nid2ni_locked(src_nid, cpt);
1197 if (src_ni == NULL) {
1198 lnet_net_unlock(cpt);
1199 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1200 "local nid\n", libcfs_nid2str(dst_nid),
1201 libcfs_nid2str(src_nid));
1202 return -EINVAL;
1203 }
1204 LASSERT(!msg->msg_routing);
1205 }
1206
1207 /* Is this for someone on a local network? */
1208 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt);
1209
1210 if (local_ni != NULL) {
1211 if (src_ni == NULL) {
1212 src_ni = local_ni;
1213 src_nid = src_ni->ni_nid;
1214 } else if (src_ni == local_ni) {
1215 lnet_ni_decref_locked(local_ni, cpt);
1216 } else {
1217 lnet_ni_decref_locked(local_ni, cpt);
1218 lnet_ni_decref_locked(src_ni, cpt);
1219 lnet_net_unlock(cpt);
1220 LCONSOLE_WARN("No route to %s via from %s\n",
1221 libcfs_nid2str(dst_nid),
1222 libcfs_nid2str(src_nid));
1223 return -EINVAL;
1224 }
1225
1226 LASSERT(src_nid != LNET_NID_ANY);
1227 lnet_msg_commit(msg, cpt);
1228
1229 if (!msg->msg_routing)
1230 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1231
1232 if (src_ni == the_lnet.ln_loni) {
1233 /* No send credit hassles with LOLND */
1234 lnet_net_unlock(cpt);
1235 lnet_ni_send(src_ni, msg);
1236
1237 lnet_net_lock(cpt);
1238 lnet_ni_decref_locked(src_ni, cpt);
1239 lnet_net_unlock(cpt);
1240 return 0;
1241 }
1242
1243 rc = lnet_nid2peer_locked(&lp, dst_nid, cpt);
1244 /* lp has ref on src_ni; lose mine */
1245 lnet_ni_decref_locked(src_ni, cpt);
1246 if (rc != 0) {
1247 lnet_net_unlock(cpt);
1248 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1249 libcfs_nid2str(dst_nid));
1250 /* ENOMEM or shutting down */
1251 return rc;
1252 }
1253 LASSERT(lp->lp_ni == src_ni);
1254 } else {
1255 /* sending to a remote network */
1256 lp = lnet_find_route_locked(src_ni, dst_nid, rtr_nid);
1257 if (lp == NULL) {
1258 if (src_ni != NULL)
1259 lnet_ni_decref_locked(src_ni, cpt);
1260 lnet_net_unlock(cpt);
1261
1262 LCONSOLE_WARN("No route to %s via %s "
1263 "(all routers down)\n",
1264 libcfs_id2str(msg->msg_target),
1265 libcfs_nid2str(src_nid));
1266 return -EHOSTUNREACH;
1267 }
1268
1269 /* rtr_nid is LNET_NID_ANY or NID of pre-determined router,
1270 * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't
1271 * pre-determined router, this can happen if router table
1272 * was changed when we release the lock */
1273 if (rtr_nid != lp->lp_nid) {
1274 cpt2 = lnet_cpt_of_nid_locked(lp->lp_nid);
1275 if (cpt2 != cpt) {
1276 if (src_ni != NULL)
1277 lnet_ni_decref_locked(src_ni, cpt);
1278 lnet_net_unlock(cpt);
1279
1280 rtr_nid = lp->lp_nid;
1281 cpt = cpt2;
1282 goto again;
1283 }
1284 }
1285
1286 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1287 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1288 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1289
1290 if (src_ni == NULL) {
1291 src_ni = lp->lp_ni;
1292 src_nid = src_ni->ni_nid;
1293 } else {
1294 LASSERT(src_ni == lp->lp_ni);
1295 lnet_ni_decref_locked(src_ni, cpt);
1296 }
1297
1298 lnet_peer_addref_locked(lp);
1299
1300 LASSERT(src_nid != LNET_NID_ANY);
1301 lnet_msg_commit(msg, cpt);
1302
1303 if (!msg->msg_routing) {
1304 /* I'm the source and now I know which NI to send on */
1305 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1306 }
1307
1308 msg->msg_target_is_router = 1;
1309 msg->msg_target.nid = lp->lp_nid;
1310 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1311 }
1312
1313 /* 'lp' is our best choice of peer */
1314
1315 LASSERT(!msg->msg_peertxcredit);
1316 LASSERT(!msg->msg_txcredit);
1317 LASSERT(msg->msg_txpeer == NULL);
1318
1319 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1320
1321 rc = lnet_post_send_locked(msg, 0);
1322 lnet_net_unlock(cpt);
1323
1324 if (rc == EHOSTUNREACH)
1325 return -EHOSTUNREACH;
1326
1327 if (rc == 0)
1328 lnet_ni_send(src_ni, msg);
1329
1330 return 0;
1331 }
1332
1333 static void
1334 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1335 {
1336 lnet_net_lock(cpt);
1337 the_lnet.ln_counters[cpt]->drop_count++;
1338 the_lnet.ln_counters[cpt]->drop_length += nob;
1339 lnet_net_unlock(cpt);
1340
1341 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1342 }
1343
1344 static void
1345 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1346 {
1347 lnet_hdr_t *hdr = &msg->msg_hdr;
1348
1349 if (msg->msg_wanted != 0)
1350 lnet_setpayloadbuffer(msg);
1351
1352 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1353
1354 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1355 * it back into the ACK during lnet_finalize() */
1356 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1357 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1358
1359 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1360 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1361 }
1362
1363 static int
1364 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1365 {
1366 lnet_hdr_t *hdr = &msg->msg_hdr;
1367 struct lnet_match_info info;
1368 int rc;
1369
1370 /* Convert put fields to host byte order */
1371 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1372 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1373 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1374
1375 info.mi_id.nid = hdr->src_nid;
1376 info.mi_id.pid = hdr->src_pid;
1377 info.mi_opc = LNET_MD_OP_PUT;
1378 info.mi_portal = hdr->msg.put.ptl_index;
1379 info.mi_rlength = hdr->payload_length;
1380 info.mi_roffset = hdr->msg.put.offset;
1381 info.mi_mbits = hdr->msg.put.match_bits;
1382
1383 msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL;
1384
1385 again:
1386 rc = lnet_ptl_match_md(&info, msg);
1387 switch (rc) {
1388 default:
1389 LBUG();
1390
1391 case LNET_MATCHMD_OK:
1392 lnet_recv_put(ni, msg);
1393 return 0;
1394
1395 case LNET_MATCHMD_NONE:
1396 if (msg->msg_rx_delayed) /* attached on delayed list */
1397 return 0;
1398
1399 rc = lnet_ni_eager_recv(ni, msg);
1400 if (rc == 0)
1401 goto again;
1402 /* fall through */
1403
1404 case LNET_MATCHMD_DROP:
1405 CNETERR("Dropping PUT from %s portal %d match "LPU64
1406 " offset %d length %d: %d\n",
1407 libcfs_id2str(info.mi_id), info.mi_portal,
1408 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1409
1410 return ENOENT; /* +ve: OK but no match */
1411 }
1412 }
1413
1414 static int
1415 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1416 {
1417 struct lnet_match_info info;
1418 lnet_hdr_t *hdr = &msg->msg_hdr;
1419 lnet_handle_wire_t reply_wmd;
1420 int rc;
1421
1422 /* Convert get fields to host byte order */
1423 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1424 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1425 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1426 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1427
1428 info.mi_id.nid = hdr->src_nid;
1429 info.mi_id.pid = hdr->src_pid;
1430 info.mi_opc = LNET_MD_OP_GET;
1431 info.mi_portal = hdr->msg.get.ptl_index;
1432 info.mi_rlength = hdr->msg.get.sink_length;
1433 info.mi_roffset = hdr->msg.get.src_offset;
1434 info.mi_mbits = hdr->msg.get.match_bits;
1435
1436 rc = lnet_ptl_match_md(&info, msg);
1437 if (rc == LNET_MATCHMD_DROP) {
1438 CNETERR("Dropping GET from %s portal %d match "LPU64
1439 " offset %d length %d\n",
1440 libcfs_id2str(info.mi_id), info.mi_portal,
1441 info.mi_mbits, info.mi_roffset, info.mi_rlength);
1442 return ENOENT; /* +ve: OK but no match */
1443 }
1444
1445 LASSERT(rc == LNET_MATCHMD_OK);
1446
1447 lnet_build_msg_event(msg, LNET_EVENT_GET);
1448
1449 reply_wmd = hdr->msg.get.return_wmd;
1450
1451 lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
1452 msg->msg_offset, msg->msg_wanted);
1453
1454 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1455
1456 if (rdma_get) {
1457 /* The LND completes the REPLY from her recv procedure */
1458 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1459 msg->msg_offset, msg->msg_len, msg->msg_len);
1460 return 0;
1461 }
1462
1463 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1464 msg->msg_receiving = 0;
1465
1466 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
1467 if (rc < 0) {
1468 /* didn't get as far as lnet_ni_send() */
1469 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1470 libcfs_nid2str(ni->ni_nid),
1471 libcfs_id2str(info.mi_id), rc);
1472
1473 lnet_finalize(ni, msg, rc);
1474 }
1475
1476 return 0;
1477 }
1478
1479 static int
1480 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1481 {
1482 void *private = msg->msg_private;
1483 lnet_hdr_t *hdr = &msg->msg_hdr;
1484 lnet_process_id_t src = {0};
1485 lnet_libmd_t *md;
1486 int rlength;
1487 int mlength;
1488 int cpt;
1489
1490 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
1491 lnet_res_lock(cpt);
1492
1493 src.nid = hdr->src_nid;
1494 src.pid = hdr->src_pid;
1495
1496 /* NB handles only looked up by creator (no flips) */
1497 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1498 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1499 CNETERR("%s: Dropping REPLY from %s for %s "
1500 "MD "LPX64"."LPX64"\n",
1501 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1502 (md == NULL) ? "invalid" : "inactive",
1503 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1504 hdr->msg.reply.dst_wmd.wh_object_cookie);
1505 if (md != NULL && md->md_me != NULL)
1506 CERROR("REPLY MD also attached to portal %d\n",
1507 md->md_me->me_portal);
1508
1509 lnet_res_unlock(cpt);
1510 return ENOENT; /* +ve: OK but no match */
1511 }
1512
1513 LASSERT(md->md_offset == 0);
1514
1515 rlength = hdr->payload_length;
1516 mlength = MIN(rlength, (int)md->md_length);
1517
1518 if (mlength < rlength &&
1519 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1520 CNETERR("%s: Dropping REPLY from %s length %d "
1521 "for MD "LPX64" would overflow (%d)\n",
1522 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1523 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1524 mlength);
1525 lnet_res_unlock(cpt);
1526 return ENOENT; /* +ve: OK but no match */
1527 }
1528
1529 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1530 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1531 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1532
1533 lnet_msg_attach_md(msg, md, 0, mlength);
1534
1535 if (mlength != 0)
1536 lnet_setpayloadbuffer(msg);
1537
1538 lnet_res_unlock(cpt);
1539
1540 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
1541
1542 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1543 return 0;
1544 }
1545
1546 static int
1547 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1548 {
1549 lnet_hdr_t *hdr = &msg->msg_hdr;
1550 lnet_process_id_t src = {0};
1551 lnet_libmd_t *md;
1552 int cpt;
1553
1554 src.nid = hdr->src_nid;
1555 src.pid = hdr->src_pid;
1556
1557 /* Convert ack fields to host byte order */
1558 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1559 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1560
1561 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
1562 lnet_res_lock(cpt);
1563
1564 /* NB handles only looked up by creator (no flips) */
1565 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1566 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1567 /* Don't moan; this is expected */
1568 CDEBUG(D_NET,
1569 "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1570 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1571 (md == NULL) ? "invalid" : "inactive",
1572 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1573 hdr->msg.ack.dst_wmd.wh_object_cookie);
1574 if (md != NULL && md->md_me != NULL)
1575 CERROR("Source MD also attached to portal %d\n",
1576 md->md_me->me_portal);
1577
1578 lnet_res_unlock(cpt);
1579 return ENOENT; /* +ve! */
1580 }
1581
1582 CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1583 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1584 hdr->msg.ack.dst_wmd.wh_object_cookie);
1585
1586 lnet_msg_attach_md(msg, md, 0, 0);
1587
1588 lnet_res_unlock(cpt);
1589
1590 lnet_build_msg_event(msg, LNET_EVENT_ACK);
1591
1592 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
1593 return 0;
1594 }
1595
1596 static int
1597 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
1598 {
1599 int rc = 0;
1600
1601 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
1602 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
1603 if (ni->ni_lnd->lnd_eager_recv == NULL) {
1604 msg->msg_rx_ready_delay = 1;
1605 } else {
1606 lnet_net_unlock(msg->msg_rx_cpt);
1607 rc = lnet_ni_eager_recv(ni, msg);
1608 lnet_net_lock(msg->msg_rx_cpt);
1609 }
1610 }
1611
1612 if (rc == 0)
1613 rc = lnet_post_routed_recv_locked(msg, 0);
1614 return rc;
1615 }
1616
1617 char *
1618 lnet_msgtyp2str(int type)
1619 {
1620 switch (type) {
1621 case LNET_MSG_ACK:
1622 return "ACK";
1623 case LNET_MSG_PUT:
1624 return "PUT";
1625 case LNET_MSG_GET:
1626 return "GET";
1627 case LNET_MSG_REPLY:
1628 return "REPLY";
1629 case LNET_MSG_HELLO:
1630 return "HELLO";
1631 default:
1632 return "<UNKNOWN>";
1633 }
1634 }
1635 EXPORT_SYMBOL(lnet_msgtyp2str);
1636
1637 void
1638 lnet_print_hdr(lnet_hdr_t *hdr)
1639 {
1640 lnet_process_id_t src = {0};
1641 lnet_process_id_t dst = {0};
1642 char *type_str = lnet_msgtyp2str(hdr->type);
1643
1644 src.nid = hdr->src_nid;
1645 src.pid = hdr->src_pid;
1646
1647 dst.nid = hdr->dest_nid;
1648 dst.pid = hdr->dest_pid;
1649
1650 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
1651 CWARN(" From %s\n", libcfs_id2str(src));
1652 CWARN(" To %s\n", libcfs_id2str(dst));
1653
1654 switch (hdr->type) {
1655 default:
1656 break;
1657
1658 case LNET_MSG_PUT:
1659 CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
1660 "match bits "LPU64"\n",
1661 hdr->msg.put.ptl_index,
1662 hdr->msg.put.ack_wmd.wh_interface_cookie,
1663 hdr->msg.put.ack_wmd.wh_object_cookie,
1664 hdr->msg.put.match_bits);
1665 CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
1666 hdr->payload_length, hdr->msg.put.offset,
1667 hdr->msg.put.hdr_data);
1668 break;
1669
1670 case LNET_MSG_GET:
1671 CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
1672 "match bits "LPU64"\n", hdr->msg.get.ptl_index,
1673 hdr->msg.get.return_wmd.wh_interface_cookie,
1674 hdr->msg.get.return_wmd.wh_object_cookie,
1675 hdr->msg.get.match_bits);
1676 CWARN(" Length %d, src offset %d\n",
1677 hdr->msg.get.sink_length,
1678 hdr->msg.get.src_offset);
1679 break;
1680
1681 case LNET_MSG_ACK:
1682 CWARN(" dst md "LPX64"."LPX64", "
1683 "manipulated length %d\n",
1684 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1685 hdr->msg.ack.dst_wmd.wh_object_cookie,
1686 hdr->msg.ack.mlength);
1687 break;
1688
1689 case LNET_MSG_REPLY:
1690 CWARN(" dst md "LPX64"."LPX64", "
1691 "length %d\n",
1692 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1693 hdr->msg.reply.dst_wmd.wh_object_cookie,
1694 hdr->payload_length);
1695 }
1696
1697 }
1698
1699 int
1700 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
1701 void *private, int rdma_req)
1702 {
1703 int rc = 0;
1704 int cpt;
1705 int for_me;
1706 struct lnet_msg *msg;
1707 lnet_pid_t dest_pid;
1708 lnet_nid_t dest_nid;
1709 lnet_nid_t src_nid;
1710 __u32 payload_length;
1711 __u32 type;
1712
1713 LASSERT(!in_interrupt());
1714
1715 type = le32_to_cpu(hdr->type);
1716 src_nid = le64_to_cpu(hdr->src_nid);
1717 dest_nid = le64_to_cpu(hdr->dest_nid);
1718 dest_pid = le32_to_cpu(hdr->dest_pid);
1719 payload_length = le32_to_cpu(hdr->payload_length);
1720
1721 for_me = (ni->ni_nid == dest_nid);
1722 cpt = lnet_cpt_of_nid(from_nid);
1723
1724 switch (type) {
1725 case LNET_MSG_ACK:
1726 case LNET_MSG_GET:
1727 if (payload_length > 0) {
1728 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
1729 libcfs_nid2str(from_nid),
1730 libcfs_nid2str(src_nid),
1731 lnet_msgtyp2str(type), payload_length);
1732 return -EPROTO;
1733 }
1734 break;
1735
1736 case LNET_MSG_PUT:
1737 case LNET_MSG_REPLY:
1738 if (payload_length >
1739 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
1740 CERROR("%s, src %s: bad %s payload %d "
1741 "(%d max expected)\n",
1742 libcfs_nid2str(from_nid),
1743 libcfs_nid2str(src_nid),
1744 lnet_msgtyp2str(type),
1745 payload_length,
1746 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
1747 return -EPROTO;
1748 }
1749 break;
1750
1751 default:
1752 CERROR("%s, src %s: Bad message type 0x%x\n",
1753 libcfs_nid2str(from_nid),
1754 libcfs_nid2str(src_nid), type);
1755 return -EPROTO;
1756 }
1757
1758 if (the_lnet.ln_routing &&
1759 ni->ni_last_alive != cfs_time_current_sec()) {
1760 lnet_ni_lock(ni);
1761
1762 /* NB: so far here is the only place to set NI status to "up */
1763 ni->ni_last_alive = cfs_time_current_sec();
1764 if (ni->ni_status != NULL &&
1765 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
1766 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
1767 lnet_ni_unlock(ni);
1768 }
1769
1770 /* Regard a bad destination NID as a protocol error. Senders should
1771 * know what they're doing; if they don't they're misconfigured, buggy
1772 * or malicious so we chop them off at the knees :) */
1773
1774 if (!for_me) {
1775 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
1776 /* should have gone direct */
1777 CERROR("%s, src %s: Bad dest nid %s "
1778 "(should have been sent direct)\n",
1779 libcfs_nid2str(from_nid),
1780 libcfs_nid2str(src_nid),
1781 libcfs_nid2str(dest_nid));
1782 return -EPROTO;
1783 }
1784
1785 if (lnet_islocalnid(dest_nid)) {
1786 /* dest is another local NI; sender should have used
1787 * this node's NID on its own network */
1788 CERROR("%s, src %s: Bad dest nid %s "
1789 "(it's my nid but on a different network)\n",
1790 libcfs_nid2str(from_nid),
1791 libcfs_nid2str(src_nid),
1792 libcfs_nid2str(dest_nid));
1793 return -EPROTO;
1794 }
1795
1796 if (rdma_req && type == LNET_MSG_GET) {
1797 CERROR("%s, src %s: Bad optimized GET for %s "
1798 "(final destination must be me)\n",
1799 libcfs_nid2str(from_nid),
1800 libcfs_nid2str(src_nid),
1801 libcfs_nid2str(dest_nid));
1802 return -EPROTO;
1803 }
1804
1805 if (!the_lnet.ln_routing) {
1806 CERROR("%s, src %s: Dropping message for %s "
1807 "(routing not enabled)\n",
1808 libcfs_nid2str(from_nid),
1809 libcfs_nid2str(src_nid),
1810 libcfs_nid2str(dest_nid));
1811 goto drop;
1812 }
1813 }
1814
1815 /* Message looks OK; we're not going to return an error, so we MUST
1816 * call back lnd_recv() come what may... */
1817
1818 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
1819 fail_peer(src_nid, 0)) { /* shall we now? */
1820 CERROR("%s, src %s: Dropping %s to simulate failure\n",
1821 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1822 lnet_msgtyp2str(type));
1823 goto drop;
1824 }
1825
1826 msg = lnet_msg_alloc();
1827 if (msg == NULL) {
1828 CERROR("%s, src %s: Dropping %s (out of memory)\n",
1829 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1830 lnet_msgtyp2str(type));
1831 goto drop;
1832 }
1833
1834 /* msg zeroed in lnet_msg_alloc;
1835 * i.e. flags all clear, pointers NULL etc
1836 */
1837
1838 msg->msg_type = type;
1839 msg->msg_private = private;
1840 msg->msg_receiving = 1;
1841 msg->msg_len = msg->msg_wanted = payload_length;
1842 msg->msg_offset = 0;
1843 msg->msg_hdr = *hdr;
1844 /* for building message event */
1845 msg->msg_from = from_nid;
1846 if (!for_me) {
1847 msg->msg_target.pid = dest_pid;
1848 msg->msg_target.nid = dest_nid;
1849 msg->msg_routing = 1;
1850
1851 } else {
1852 /* convert common msg->hdr fields to host byteorder */
1853 msg->msg_hdr.type = type;
1854 msg->msg_hdr.src_nid = src_nid;
1855 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
1856 msg->msg_hdr.dest_nid = dest_nid;
1857 msg->msg_hdr.dest_pid = dest_pid;
1858 msg->msg_hdr.payload_length = payload_length;
1859 }
1860
1861 lnet_net_lock(cpt);
1862 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid, cpt);
1863 if (rc != 0) {
1864 lnet_net_unlock(cpt);
1865 CERROR("%s, src %s: Dropping %s "
1866 "(error %d looking up sender)\n",
1867 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1868 lnet_msgtyp2str(type), rc);
1869 lnet_msg_free(msg);
1870 goto drop;
1871 }
1872
1873 lnet_msg_commit(msg, cpt);
1874
1875 if (!for_me) {
1876 rc = lnet_parse_forward_locked(ni, msg);
1877 lnet_net_unlock(cpt);
1878
1879 if (rc < 0)
1880 goto free_drop;
1881 if (rc == 0) {
1882 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1883 0, payload_length, payload_length);
1884 }
1885 return 0;
1886 }
1887
1888 lnet_net_unlock(cpt);
1889
1890 switch (type) {
1891 case LNET_MSG_ACK:
1892 rc = lnet_parse_ack(ni, msg);
1893 break;
1894 case LNET_MSG_PUT:
1895 rc = lnet_parse_put(ni, msg);
1896 break;
1897 case LNET_MSG_GET:
1898 rc = lnet_parse_get(ni, msg, rdma_req);
1899 break;
1900 case LNET_MSG_REPLY:
1901 rc = lnet_parse_reply(ni, msg);
1902 break;
1903 default:
1904 LASSERT(0);
1905 rc = -EPROTO;
1906 goto free_drop; /* prevent an unused label if !kernel */
1907 }
1908
1909 if (rc == 0)
1910 return 0;
1911
1912 LASSERT(rc == ENOENT);
1913
1914 free_drop:
1915 LASSERT(msg->msg_md == NULL);
1916 lnet_finalize(ni, msg, rc);
1917
1918 drop:
1919 lnet_drop_message(ni, cpt, private, payload_length);
1920 return 0;
1921 }
1922 EXPORT_SYMBOL(lnet_parse);
1923
1924 void
1925 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
1926 {
1927 while (!list_empty(head)) {
1928 lnet_process_id_t id = {0};
1929 lnet_msg_t *msg;
1930
1931 msg = list_entry(head->next, lnet_msg_t, msg_list);
1932 list_del(&msg->msg_list);
1933
1934 id.nid = msg->msg_hdr.src_nid;
1935 id.pid = msg->msg_hdr.src_pid;
1936
1937 LASSERT(msg->msg_md == NULL);
1938 LASSERT(msg->msg_rx_delayed);
1939 LASSERT(msg->msg_rxpeer != NULL);
1940 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
1941
1942 CWARN("Dropping delayed PUT from %s portal %d match "LPU64
1943 " offset %d length %d: %s\n",
1944 libcfs_id2str(id),
1945 msg->msg_hdr.msg.put.ptl_index,
1946 msg->msg_hdr.msg.put.match_bits,
1947 msg->msg_hdr.msg.put.offset,
1948 msg->msg_hdr.payload_length, reason);
1949
1950 /* NB I can't drop msg's ref on msg_rxpeer until after I've
1951 * called lnet_drop_message(), so I just hang onto msg as well
1952 * until that's done */
1953
1954 lnet_drop_message(msg->msg_rxpeer->lp_ni,
1955 msg->msg_rxpeer->lp_cpt,
1956 msg->msg_private, msg->msg_len);
1957 /*
1958 * NB: message will not generate event because w/o attached MD,
1959 * but we still should give error code so lnet_msg_decommit()
1960 * can skip counters operations and other checks.
1961 */
1962 lnet_finalize(msg->msg_rxpeer->lp_ni, msg, -ENOENT);
1963 }
1964 }
1965
1966 void
1967 lnet_recv_delayed_msg_list(struct list_head *head)
1968 {
1969 while (!list_empty(head)) {
1970 lnet_msg_t *msg;
1971 lnet_process_id_t id;
1972
1973 msg = list_entry(head->next, lnet_msg_t, msg_list);
1974 list_del(&msg->msg_list);
1975
1976 /* md won't disappear under me, since each msg
1977 * holds a ref on it */
1978
1979 id.nid = msg->msg_hdr.src_nid;
1980 id.pid = msg->msg_hdr.src_pid;
1981
1982 LASSERT(msg->msg_rx_delayed);
1983 LASSERT(msg->msg_md != NULL);
1984 LASSERT(msg->msg_rxpeer != NULL);
1985 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
1986
1987 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
1988 "match "LPU64" offset %d length %d.\n",
1989 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
1990 msg->msg_hdr.msg.put.match_bits,
1991 msg->msg_hdr.msg.put.offset,
1992 msg->msg_hdr.payload_length);
1993
1994 lnet_recv_put(msg->msg_rxpeer->lp_ni, msg);
1995 }
1996 }
1997
1998 /**
1999 * Initiate an asynchronous PUT operation.
2000 *
2001 * There are several events associated with a PUT: completion of the send on
2002 * the initiator node (LNET_EVENT_SEND), and when the send completes
2003 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2004 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2005 * used at the target node to indicate the completion of incoming data
2006 * delivery.
2007 *
2008 * The local events will be logged in the EQ associated with the MD pointed to
2009 * by \a mdh handle. Using a MD without an associated EQ results in these
2010 * events being discarded. In this case, the caller must have another
2011 * mechanism (e.g., a higher level protocol) for determining when it is safe
2012 * to modify the memory region associated with the MD.
2013 *
2014 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2015 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2016 *
2017 * \param self Indicates the NID of a local interface through which to send
2018 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2019 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2020 * must be "free floating" (See LNetMDBind()).
2021 * \param ack Controls whether an acknowledgment is requested.
2022 * Acknowledgments are only sent when they are requested by the initiating
2023 * process and the target MD enables them.
2024 * \param target A process identifier for the target process.
2025 * \param portal The index in the \a target's portal table.
2026 * \param match_bits The match bits to use for MD selection at the target
2027 * process.
2028 * \param offset The offset into the target MD (only used when the target
2029 * MD has the LNET_MD_MANAGE_REMOTE option set).
2030 * \param hdr_data 64 bits of user data that can be included in the message
2031 * header. This data is written to an event queue entry at the target if an
2032 * EQ is present on the matching MD.
2033 *
2034 * \retval 0 Success, and only in this case events will be generated
2035 * and logged to EQ (if it exists).
2036 * \retval -EIO Simulated failure.
2037 * \retval -ENOMEM Memory allocation failure.
2038 * \retval -ENOENT Invalid MD object.
2039 *
2040 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2041 */
2042 int
2043 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2044 lnet_process_id_t target, unsigned int portal,
2045 __u64 match_bits, unsigned int offset,
2046 __u64 hdr_data)
2047 {
2048 struct lnet_msg *msg;
2049 struct lnet_libmd *md;
2050 int cpt;
2051 int rc;
2052
2053 LASSERT(the_lnet.ln_init);
2054 LASSERT(the_lnet.ln_refcount > 0);
2055
2056 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2057 fail_peer(target.nid, 1)) { /* shall we now? */
2058 CERROR("Dropping PUT to %s: simulated failure\n",
2059 libcfs_id2str(target));
2060 return -EIO;
2061 }
2062
2063 msg = lnet_msg_alloc();
2064 if (msg == NULL) {
2065 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2066 libcfs_id2str(target));
2067 return -ENOMEM;
2068 }
2069 msg->msg_vmflush = !!memory_pressure_get();
2070
2071 cpt = lnet_cpt_of_cookie(mdh.cookie);
2072 lnet_res_lock(cpt);
2073
2074 md = lnet_handle2md(&mdh);
2075 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2076 CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
2077 match_bits, portal, libcfs_id2str(target),
2078 md == NULL ? -1 : md->md_threshold);
2079 if (md != NULL && md->md_me != NULL)
2080 CERROR("Source MD also attached to portal %d\n",
2081 md->md_me->me_portal);
2082 lnet_res_unlock(cpt);
2083
2084 lnet_msg_free(msg);
2085 return -ENOENT;
2086 }
2087
2088 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2089
2090 lnet_msg_attach_md(msg, md, 0, 0);
2091
2092 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2093
2094 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2095 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2096 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2097 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2098
2099 /* NB handles only looked up by creator (no flips) */
2100 if (ack == LNET_ACK_REQ) {
2101 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2102 the_lnet.ln_interface_cookie;
2103 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2104 md->md_lh.lh_cookie;
2105 } else {
2106 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2107 LNET_WIRE_HANDLE_COOKIE_NONE;
2108 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2109 LNET_WIRE_HANDLE_COOKIE_NONE;
2110 }
2111
2112 lnet_res_unlock(cpt);
2113
2114 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2115
2116 rc = lnet_send(self, msg, LNET_NID_ANY);
2117 if (rc != 0) {
2118 CNETERR("Error sending PUT to %s: %d\n",
2119 libcfs_id2str(target), rc);
2120 lnet_finalize(NULL, msg, rc);
2121 }
2122
2123 /* completion will be signalled by an event */
2124 return 0;
2125 }
2126 EXPORT_SYMBOL(LNetPut);
2127
2128 lnet_msg_t *
2129 lnet_create_reply_msg(lnet_ni_t *ni, lnet_msg_t *getmsg)
2130 {
2131 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2132 * returns a msg for the LND to pass to lnet_finalize() when the sink
2133 * data has been received.
2134 *
2135 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2136 * lnet_finalize() is called on it, so the LND must call this first */
2137
2138 struct lnet_msg *msg = lnet_msg_alloc();
2139 struct lnet_libmd *getmd = getmsg->msg_md;
2140 lnet_process_id_t peer_id = getmsg->msg_target;
2141 int cpt;
2142
2143 LASSERT(!getmsg->msg_target_is_router);
2144 LASSERT(!getmsg->msg_routing);
2145
2146 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2147 lnet_res_lock(cpt);
2148
2149 LASSERT(getmd->md_refcount > 0);
2150
2151 if (msg == NULL) {
2152 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2153 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2154 goto drop;
2155 }
2156
2157 if (getmd->md_threshold == 0) {
2158 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
2159 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2160 getmd);
2161 lnet_res_unlock(cpt);
2162 goto drop;
2163 }
2164
2165 LASSERT(getmd->md_offset == 0);
2166
2167 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2168 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2169
2170 /* setup information for lnet_build_msg_event */
2171 msg->msg_from = peer_id.nid;
2172 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2173 msg->msg_hdr.src_nid = peer_id.nid;
2174 msg->msg_hdr.payload_length = getmd->md_length;
2175 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2176
2177 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2178 lnet_res_unlock(cpt);
2179
2180 cpt = lnet_cpt_of_nid(peer_id.nid);
2181
2182 lnet_net_lock(cpt);
2183 lnet_msg_commit(msg, cpt);
2184 lnet_net_unlock(cpt);
2185
2186 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2187
2188 return msg;
2189
2190 drop:
2191 cpt = lnet_cpt_of_nid(peer_id.nid);
2192
2193 lnet_net_lock(cpt);
2194 the_lnet.ln_counters[cpt]->drop_count++;
2195 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2196 lnet_net_unlock(cpt);
2197
2198 if (msg != NULL)
2199 lnet_msg_free(msg);
2200
2201 return NULL;
2202 }
2203 EXPORT_SYMBOL(lnet_create_reply_msg);
2204
2205 void
2206 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2207 {
2208 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2209 * completed and I know it. */
2210 LASSERT(reply != NULL);
2211 LASSERT(reply->msg_type == LNET_MSG_GET);
2212 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
2213
2214 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2215 * the end of my buffer, I might as well be dead. */
2216 LASSERT(len <= reply->msg_ev.mlength);
2217
2218 reply->msg_ev.mlength = len;
2219 }
2220 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2221
2222 /**
2223 * Initiate an asynchronous GET operation.
2224 *
2225 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2226 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2227 * the target node in the REPLY has been written to local MD.
2228 *
2229 * On the target node, an LNET_EVENT_GET is logged when the GET request
2230 * arrives and is accepted into a MD.
2231 *
2232 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2233 * \param mdh A handle for the MD that describes the memory into which the
2234 * requested data will be received. The MD must be "free floating"
2235 * (See LNetMDBind()).
2236 *
2237 * \retval 0 Success, and only in this case events will be generated
2238 * and logged to EQ (if it exists) of the MD.
2239 * \retval -EIO Simulated failure.
2240 * \retval -ENOMEM Memory allocation failure.
2241 * \retval -ENOENT Invalid MD object.
2242 */
2243 int
2244 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2245 lnet_process_id_t target, unsigned int portal,
2246 __u64 match_bits, unsigned int offset)
2247 {
2248 struct lnet_msg *msg;
2249 struct lnet_libmd *md;
2250 int cpt;
2251 int rc;
2252
2253 LASSERT(the_lnet.ln_init);
2254 LASSERT(the_lnet.ln_refcount > 0);
2255
2256 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2257 fail_peer(target.nid, 1)) { /* shall we now? */
2258 CERROR("Dropping GET to %s: simulated failure\n",
2259 libcfs_id2str(target));
2260 return -EIO;
2261 }
2262
2263 msg = lnet_msg_alloc();
2264 if (msg == NULL) {
2265 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2266 libcfs_id2str(target));
2267 return -ENOMEM;
2268 }
2269
2270 cpt = lnet_cpt_of_cookie(mdh.cookie);
2271 lnet_res_lock(cpt);
2272
2273 md = lnet_handle2md(&mdh);
2274 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2275 CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
2276 match_bits, portal, libcfs_id2str(target),
2277 md == NULL ? -1 : md->md_threshold);
2278 if (md != NULL && md->md_me != NULL)
2279 CERROR("REPLY MD also attached to portal %d\n",
2280 md->md_me->me_portal);
2281
2282 lnet_res_unlock(cpt);
2283
2284 lnet_msg_free(msg);
2285
2286 return -ENOENT;
2287 }
2288
2289 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2290
2291 lnet_msg_attach_md(msg, md, 0, 0);
2292
2293 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2294
2295 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2296 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2297 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2298 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2299
2300 /* NB handles only looked up by creator (no flips) */
2301 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2302 the_lnet.ln_interface_cookie;
2303 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2304 md->md_lh.lh_cookie;
2305
2306 lnet_res_unlock(cpt);
2307
2308 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2309
2310 rc = lnet_send(self, msg, LNET_NID_ANY);
2311 if (rc < 0) {
2312 CNETERR("Error sending GET to %s: %d\n",
2313 libcfs_id2str(target), rc);
2314 lnet_finalize(NULL, msg, rc);
2315 }
2316
2317 /* completion will be signalled by an event */
2318 return 0;
2319 }
2320 EXPORT_SYMBOL(LNetGet);
2321
2322 /**
2323 * Calculate distance to node at \a dstnid.
2324 *
2325 * \param dstnid Target NID.
2326 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2327 * is saved here.
2328 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2329 * here.
2330 *
2331 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2332 * local_nid_dist_zero is set, which is the default.
2333 * \retval positives Distance to target NID, i.e. number of hops plus one.
2334 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2335 */
2336 int
2337 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2338 {
2339 struct list_head *e;
2340 struct lnet_ni *ni;
2341 lnet_remotenet_t *rnet;
2342 __u32 dstnet = LNET_NIDNET(dstnid);
2343 int hops;
2344 int cpt;
2345 __u32 order = 2;
2346 struct list_head *rn_list;
2347
2348 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2349 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2350 * keep order 0 free for 0@lo and order 1 free for a local NID
2351 * match */
2352
2353 LASSERT(the_lnet.ln_init);
2354 LASSERT(the_lnet.ln_refcount > 0);
2355
2356 cpt = lnet_net_lock_current();
2357
2358 list_for_each(e, &the_lnet.ln_nis) {
2359 ni = list_entry(e, lnet_ni_t, ni_list);
2360
2361 if (ni->ni_nid == dstnid) {
2362 if (srcnidp != NULL)
2363 *srcnidp = dstnid;
2364 if (orderp != NULL) {
2365 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2366 *orderp = 0;
2367 else
2368 *orderp = 1;
2369 }
2370 lnet_net_unlock(cpt);
2371
2372 return local_nid_dist_zero ? 0 : 1;
2373 }
2374
2375 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2376 if (srcnidp != NULL)
2377 *srcnidp = ni->ni_nid;
2378 if (orderp != NULL)
2379 *orderp = order;
2380 lnet_net_unlock(cpt);
2381 return 1;
2382 }
2383
2384 order++;
2385 }
2386
2387 rn_list = lnet_net2rnethash(dstnet);
2388 list_for_each(e, rn_list) {
2389 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2390
2391 if (rnet->lrn_net == dstnet) {
2392 lnet_route_t *route;
2393 lnet_route_t *shortest = NULL;
2394
2395 LASSERT(!list_empty(&rnet->lrn_routes));
2396
2397 list_for_each_entry(route, &rnet->lrn_routes,
2398 lr_list) {
2399 if (shortest == NULL ||
2400 route->lr_hops < shortest->lr_hops)
2401 shortest = route;
2402 }
2403
2404 LASSERT(shortest != NULL);
2405 hops = shortest->lr_hops;
2406 if (srcnidp != NULL)
2407 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2408 if (orderp != NULL)
2409 *orderp = order;
2410 lnet_net_unlock(cpt);
2411 return hops + 1;
2412 }
2413 order++;
2414 }
2415
2416 lnet_net_unlock(cpt);
2417 return -EHOSTUNREACH;
2418 }
2419 EXPORT_SYMBOL(LNetDist);
2420
2421 /**
2422 * Set the number of asynchronous messages expected from a target process.
2423 *
2424 * This function is only meaningful for userspace callers. It's a no-op when
2425 * called from kernel.
2426 *
2427 * Asynchronous messages are those that can come from a target when the
2428 * userspace process is not waiting for IO to complete; e.g., AST callbacks
2429 * from Lustre servers. Specifying the expected number of such messages
2430 * allows them to be eagerly received when user process is not running in
2431 * LNet; otherwise network errors may occur.
2432 *
2433 * \param id Process ID of the target process.
2434 * \param nasync Number of asynchronous messages expected from the target.
2435 *
2436 * \return 0 on success, and an error code otherwise.
2437 */
2438 int
2439 LNetSetAsync(lnet_process_id_t id, int nasync)
2440 {
2441 return 0;
2442 }
2443 EXPORT_SYMBOL(LNetSetAsync);
Linux kernel - Deliverables
This page took 0.129338 seconds and 5 git commands to generate.