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 "../../include/linux/lnet/lib-lnet.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 int
lnet_fail_nid(lnet_nid_t nid,unsigned int threshold)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
fail_peer(lnet_nid_t nid,int outgoing)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
lnet_iov_nob(unsigned int niov,struct iovec * iov)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
lnet_copy_iov2iov(unsigned int ndiov,struct iovec * diov,unsigned int doffset,unsigned int nsiov,struct iovec * siov,unsigned int soffset,unsigned int nob)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
lnet_extract_iov(int dst_niov,struct iovec * dst,int src_niov,struct iovec * src,unsigned int offset,unsigned int len)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
lnet_kiov_nob(unsigned int niov,lnet_kiov_t * kiov)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
lnet_copy_kiov2kiov(unsigned int ndiov,lnet_kiov_t * diov,unsigned int doffset,unsigned int nsiov,lnet_kiov_t * siov,unsigned int soffset,unsigned int nob)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
lnet_copy_kiov2iov(unsigned int niov,struct iovec * iov,unsigned int iovoffset,unsigned int nkiov,lnet_kiov_t * kiov,unsigned int kiovoffset,unsigned int nob)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
lnet_copy_iov2kiov(unsigned int nkiov,lnet_kiov_t * kiov,unsigned int kiovoffset,unsigned int niov,struct iovec * iov,unsigned int iovoffset,unsigned int nob)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
lnet_extract_kiov(int dst_niov,lnet_kiov_t * dst,int src_niov,lnet_kiov_t * src,unsigned int offset,unsigned int len)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
lnet_ni_recv(lnet_ni_t * ni,void * private,lnet_msg_t * msg,int delayed,unsigned int offset,unsigned int mlen,unsigned int rlen)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
lnet_setpayloadbuffer(lnet_msg_t * msg)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
lnet_prep_send(lnet_msg_t * msg,int type,lnet_process_id_t target,unsigned int offset,unsigned int len)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
lnet_ni_send(lnet_ni_t * ni,lnet_msg_t * msg)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
lnet_ni_eager_recv(lnet_ni_t * ni,lnet_msg_t * msg)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
lnet_ni_query_locked(lnet_ni_t * ni,lnet_peer_t * lp)683 lnet_ni_query_locked(lnet_ni_t *ni, lnet_peer_t *lp)
684 {
685 unsigned long 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
lnet_peer_is_alive(lnet_peer_t * lp,unsigned long now)702 lnet_peer_is_alive(lnet_peer_t *lp, unsigned long now)
703 {
704 int alive;
705 unsigned long 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
lnet_peer_alive_locked(lnet_peer_t * lp)735 lnet_peer_alive_locked(lnet_peer_t *lp)
736 {
737 unsigned long 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 unsigned long next_query =
751 cfs_time_add(lp->lp_last_query,
752 cfs_time_seconds(lnet_queryinterval));
753
754 if (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 /**
777 * \param msg The message to be sent.
778 * \param do_send True if lnet_ni_send() should be called in this function.
779 * lnet_send() is going to lnet_net_unlock immediately after this, so
780 * it sets do_send FALSE and I don't do the unlock/send/lock bit.
781 *
782 * \retval 0 If \a msg sent or OK to send.
783 * \retval EAGAIN If \a msg blocked for credit.
784 * \retval EHOSTUNREACH If the next hop of the message appears dead.
785 * \retval ECANCELED If the MD of the message has been unlinked.
786 */
787 static int
lnet_post_send_locked(lnet_msg_t * msg,int do_send)788 lnet_post_send_locked(lnet_msg_t *msg, int do_send)
789 {
790 lnet_peer_t *lp = msg->msg_txpeer;
791 lnet_ni_t *ni = lp->lp_ni;
792 int cpt = msg->msg_tx_cpt;
793 struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
794
795 /* non-lnet_send() callers have checked before */
796 LASSERT(!do_send || msg->msg_tx_delayed);
797 LASSERT(!msg->msg_receiving);
798 LASSERT(msg->msg_tx_committed);
799
800 /* NB 'lp' is always the next hop */
801 if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
802 lnet_peer_alive_locked(lp) == 0) {
803 the_lnet.ln_counters[cpt]->drop_count++;
804 the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
805 lnet_net_unlock(cpt);
806
807 CNETERR("Dropping message for %s: peer not alive\n",
808 libcfs_id2str(msg->msg_target));
809 if (do_send)
810 lnet_finalize(ni, msg, -EHOSTUNREACH);
811
812 lnet_net_lock(cpt);
813 return EHOSTUNREACH;
814 }
815
816 if (msg->msg_md != NULL &&
817 (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
818 lnet_net_unlock(cpt);
819
820 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
821 "called on the MD/ME.\n",
822 libcfs_id2str(msg->msg_target));
823 if (do_send)
824 lnet_finalize(ni, msg, -ECANCELED);
825
826 lnet_net_lock(cpt);
827 return ECANCELED;
828 }
829
830 if (!msg->msg_peertxcredit) {
831 LASSERT((lp->lp_txcredits < 0) ==
832 !list_empty(&lp->lp_txq));
833
834 msg->msg_peertxcredit = 1;
835 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
836 lp->lp_txcredits--;
837
838 if (lp->lp_txcredits < lp->lp_mintxcredits)
839 lp->lp_mintxcredits = lp->lp_txcredits;
840
841 if (lp->lp_txcredits < 0) {
842 msg->msg_tx_delayed = 1;
843 list_add_tail(&msg->msg_list, &lp->lp_txq);
844 return EAGAIN;
845 }
846 }
847
848 if (!msg->msg_txcredit) {
849 LASSERT((tq->tq_credits < 0) ==
850 !list_empty(&tq->tq_delayed));
851
852 msg->msg_txcredit = 1;
853 tq->tq_credits--;
854
855 if (tq->tq_credits < tq->tq_credits_min)
856 tq->tq_credits_min = tq->tq_credits;
857
858 if (tq->tq_credits < 0) {
859 msg->msg_tx_delayed = 1;
860 list_add_tail(&msg->msg_list, &tq->tq_delayed);
861 return EAGAIN;
862 }
863 }
864
865 if (do_send) {
866 lnet_net_unlock(cpt);
867 lnet_ni_send(ni, msg);
868 lnet_net_lock(cpt);
869 }
870 return 0;
871 }
872
873
874 lnet_rtrbufpool_t *
lnet_msg2bufpool(lnet_msg_t * msg)875 lnet_msg2bufpool(lnet_msg_t *msg)
876 {
877 lnet_rtrbufpool_t *rbp;
878 int cpt;
879
880 LASSERT(msg->msg_rx_committed);
881
882 cpt = msg->msg_rx_cpt;
883 rbp = &the_lnet.ln_rtrpools[cpt][0];
884
885 LASSERT(msg->msg_len <= LNET_MTU);
886 while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_CACHE_SIZE) {
887 rbp++;
888 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
889 }
890
891 return rbp;
892 }
893
894 int
lnet_post_routed_recv_locked(lnet_msg_t * msg,int do_recv)895 lnet_post_routed_recv_locked(lnet_msg_t *msg, int do_recv)
896 {
897 /* lnet_parse is going to lnet_net_unlock immediately after this, so it
898 * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
899 * return EAGAIN if msg blocked and 0 if received or OK to receive */
900 lnet_peer_t *lp = msg->msg_rxpeer;
901 lnet_rtrbufpool_t *rbp;
902 lnet_rtrbuf_t *rb;
903
904 LASSERT(msg->msg_iov == NULL);
905 LASSERT(msg->msg_kiov == NULL);
906 LASSERT(msg->msg_niov == 0);
907 LASSERT(msg->msg_routing);
908 LASSERT(msg->msg_receiving);
909 LASSERT(!msg->msg_sending);
910
911 /* non-lnet_parse callers only receive delayed messages */
912 LASSERT(!do_recv || msg->msg_rx_delayed);
913
914 if (!msg->msg_peerrtrcredit) {
915 LASSERT((lp->lp_rtrcredits < 0) ==
916 !list_empty(&lp->lp_rtrq));
917
918 msg->msg_peerrtrcredit = 1;
919 lp->lp_rtrcredits--;
920 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
921 lp->lp_minrtrcredits = lp->lp_rtrcredits;
922
923 if (lp->lp_rtrcredits < 0) {
924 /* must have checked eager_recv before here */
925 LASSERT(msg->msg_rx_ready_delay);
926 msg->msg_rx_delayed = 1;
927 list_add_tail(&msg->msg_list, &lp->lp_rtrq);
928 return EAGAIN;
929 }
930 }
931
932 rbp = lnet_msg2bufpool(msg);
933
934 if (!msg->msg_rtrcredit) {
935 LASSERT((rbp->rbp_credits < 0) ==
936 !list_empty(&rbp->rbp_msgs));
937
938 msg->msg_rtrcredit = 1;
939 rbp->rbp_credits--;
940 if (rbp->rbp_credits < rbp->rbp_mincredits)
941 rbp->rbp_mincredits = rbp->rbp_credits;
942
943 if (rbp->rbp_credits < 0) {
944 /* must have checked eager_recv before here */
945 LASSERT(msg->msg_rx_ready_delay);
946 msg->msg_rx_delayed = 1;
947 list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
948 return EAGAIN;
949 }
950 }
951
952 LASSERT(!list_empty(&rbp->rbp_bufs));
953 rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
954 list_del(&rb->rb_list);
955
956 msg->msg_niov = rbp->rbp_npages;
957 msg->msg_kiov = &rb->rb_kiov[0];
958
959 if (do_recv) {
960 int cpt = msg->msg_rx_cpt;
961
962 lnet_net_unlock(cpt);
963 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
964 0, msg->msg_len, msg->msg_len);
965 lnet_net_lock(cpt);
966 }
967 return 0;
968 }
969
970 void
lnet_return_tx_credits_locked(lnet_msg_t * msg)971 lnet_return_tx_credits_locked(lnet_msg_t *msg)
972 {
973 lnet_peer_t *txpeer = msg->msg_txpeer;
974 lnet_msg_t *msg2;
975
976 if (msg->msg_txcredit) {
977 struct lnet_ni *ni = txpeer->lp_ni;
978 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
979
980 /* give back NI txcredits */
981 msg->msg_txcredit = 0;
982
983 LASSERT((tq->tq_credits < 0) ==
984 !list_empty(&tq->tq_delayed));
985
986 tq->tq_credits++;
987 if (tq->tq_credits <= 0) {
988 msg2 = list_entry(tq->tq_delayed.next,
989 lnet_msg_t, msg_list);
990 list_del(&msg2->msg_list);
991
992 LASSERT(msg2->msg_txpeer->lp_ni == ni);
993 LASSERT(msg2->msg_tx_delayed);
994
995 (void) lnet_post_send_locked(msg2, 1);
996 }
997 }
998
999 if (msg->msg_peertxcredit) {
1000 /* give back peer txcredits */
1001 msg->msg_peertxcredit = 0;
1002
1003 LASSERT((txpeer->lp_txcredits < 0) ==
1004 !list_empty(&txpeer->lp_txq));
1005
1006 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1007 LASSERT(txpeer->lp_txqnob >= 0);
1008
1009 txpeer->lp_txcredits++;
1010 if (txpeer->lp_txcredits <= 0) {
1011 msg2 = list_entry(txpeer->lp_txq.next,
1012 lnet_msg_t, msg_list);
1013 list_del(&msg2->msg_list);
1014
1015 LASSERT(msg2->msg_txpeer == txpeer);
1016 LASSERT(msg2->msg_tx_delayed);
1017
1018 (void) lnet_post_send_locked(msg2, 1);
1019 }
1020 }
1021
1022 if (txpeer != NULL) {
1023 msg->msg_txpeer = NULL;
1024 lnet_peer_decref_locked(txpeer);
1025 }
1026 }
1027
1028 void
lnet_return_rx_credits_locked(lnet_msg_t * msg)1029 lnet_return_rx_credits_locked(lnet_msg_t *msg)
1030 {
1031 lnet_peer_t *rxpeer = msg->msg_rxpeer;
1032 lnet_msg_t *msg2;
1033
1034 if (msg->msg_rtrcredit) {
1035 /* give back global router credits */
1036 lnet_rtrbuf_t *rb;
1037 lnet_rtrbufpool_t *rbp;
1038
1039 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1040 * there until it gets one allocated, or aborts the wait
1041 * itself */
1042 LASSERT(msg->msg_kiov != NULL);
1043
1044 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1045 rbp = rb->rb_pool;
1046 LASSERT(rbp == lnet_msg2bufpool(msg));
1047
1048 msg->msg_kiov = NULL;
1049 msg->msg_rtrcredit = 0;
1050
1051 LASSERT((rbp->rbp_credits < 0) ==
1052 !list_empty(&rbp->rbp_msgs));
1053 LASSERT((rbp->rbp_credits > 0) ==
1054 !list_empty(&rbp->rbp_bufs));
1055
1056 list_add(&rb->rb_list, &rbp->rbp_bufs);
1057 rbp->rbp_credits++;
1058 if (rbp->rbp_credits <= 0) {
1059 msg2 = list_entry(rbp->rbp_msgs.next,
1060 lnet_msg_t, msg_list);
1061 list_del(&msg2->msg_list);
1062
1063 (void) lnet_post_routed_recv_locked(msg2, 1);
1064 }
1065 }
1066
1067 if (msg->msg_peerrtrcredit) {
1068 /* give back peer router credits */
1069 msg->msg_peerrtrcredit = 0;
1070
1071 LASSERT((rxpeer->lp_rtrcredits < 0) ==
1072 !list_empty(&rxpeer->lp_rtrq));
1073
1074 rxpeer->lp_rtrcredits++;
1075 if (rxpeer->lp_rtrcredits <= 0) {
1076 msg2 = list_entry(rxpeer->lp_rtrq.next,
1077 lnet_msg_t, msg_list);
1078 list_del(&msg2->msg_list);
1079
1080 (void) lnet_post_routed_recv_locked(msg2, 1);
1081 }
1082 }
1083 if (rxpeer != NULL) {
1084 msg->msg_rxpeer = NULL;
1085 lnet_peer_decref_locked(rxpeer);
1086 }
1087 }
1088
1089 static int
lnet_compare_routes(lnet_route_t * r1,lnet_route_t * r2)1090 lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
1091 {
1092 lnet_peer_t *p1 = r1->lr_gateway;
1093 lnet_peer_t *p2 = r2->lr_gateway;
1094
1095 if (r1->lr_priority < r2->lr_priority)
1096 return 1;
1097
1098 if (r1->lr_priority > r2->lr_priority)
1099 return -1;
1100
1101 if (r1->lr_hops < r2->lr_hops)
1102 return 1;
1103
1104 if (r1->lr_hops > r2->lr_hops)
1105 return -1;
1106
1107 if (p1->lp_txqnob < p2->lp_txqnob)
1108 return 1;
1109
1110 if (p1->lp_txqnob > p2->lp_txqnob)
1111 return -1;
1112
1113 if (p1->lp_txcredits > p2->lp_txcredits)
1114 return 1;
1115
1116 if (p1->lp_txcredits < p2->lp_txcredits)
1117 return -1;
1118
1119 if (r1->lr_seq - r2->lr_seq <= 0)
1120 return 1;
1121
1122 return -1;
1123 }
1124
1125 static lnet_peer_t *
lnet_find_route_locked(lnet_ni_t * ni,lnet_nid_t target,lnet_nid_t rtr_nid)1126 lnet_find_route_locked(lnet_ni_t *ni, lnet_nid_t target, lnet_nid_t rtr_nid)
1127 {
1128 lnet_remotenet_t *rnet;
1129 lnet_route_t *rtr;
1130 lnet_route_t *rtr_best;
1131 lnet_route_t *rtr_last;
1132 struct lnet_peer *lp_best;
1133 struct lnet_peer *lp;
1134 int rc;
1135
1136 /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1137 * rtr_nid nid, otherwise find the best gateway I can use */
1138
1139 rnet = lnet_find_net_locked(LNET_NIDNET(target));
1140 if (rnet == NULL)
1141 return NULL;
1142
1143 lp_best = NULL;
1144 rtr_best = rtr_last = NULL;
1145 list_for_each_entry(rtr, &rnet->lrn_routes, lr_list) {
1146 lp = rtr->lr_gateway;
1147
1148 if (!lp->lp_alive || /* gateway is down */
1149 ((lp->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0 &&
1150 rtr->lr_downis != 0)) /* NI to target is down */
1151 continue;
1152
1153 if (ni != NULL && lp->lp_ni != ni)
1154 continue;
1155
1156 if (lp->lp_nid == rtr_nid) /* it's pre-determined router */
1157 return lp;
1158
1159 if (lp_best == NULL) {
1160 rtr_best = rtr_last = rtr;
1161 lp_best = lp;
1162 continue;
1163 }
1164
1165 /* no protection on below fields, but it's harmless */
1166 if (rtr_last->lr_seq - rtr->lr_seq < 0)
1167 rtr_last = rtr;
1168
1169 rc = lnet_compare_routes(rtr, rtr_best);
1170 if (rc < 0)
1171 continue;
1172
1173 rtr_best = rtr;
1174 lp_best = lp;
1175 }
1176
1177 /* set sequence number on the best router to the latest sequence + 1
1178 * so we can round-robin all routers, it's race and inaccurate but
1179 * harmless and functional */
1180 if (rtr_best != NULL)
1181 rtr_best->lr_seq = rtr_last->lr_seq + 1;
1182 return lp_best;
1183 }
1184
1185 int
lnet_send(lnet_nid_t src_nid,lnet_msg_t * msg,lnet_nid_t rtr_nid)1186 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
1187 {
1188 lnet_nid_t dst_nid = msg->msg_target.nid;
1189 struct lnet_ni *src_ni;
1190 struct lnet_ni *local_ni;
1191 struct lnet_peer *lp;
1192 int cpt;
1193 int cpt2;
1194 int rc;
1195
1196 /* NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
1197 * but we might want to use pre-determined router for ACK/REPLY
1198 * in the future */
1199 /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
1200 LASSERT(msg->msg_txpeer == NULL);
1201 LASSERT(!msg->msg_sending);
1202 LASSERT(!msg->msg_target_is_router);
1203 LASSERT(!msg->msg_receiving);
1204
1205 msg->msg_sending = 1;
1206
1207 LASSERT(!msg->msg_tx_committed);
1208 cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid);
1209 again:
1210 lnet_net_lock(cpt);
1211
1212 if (the_lnet.ln_shutdown) {
1213 lnet_net_unlock(cpt);
1214 return -ESHUTDOWN;
1215 }
1216
1217 if (src_nid == LNET_NID_ANY) {
1218 src_ni = NULL;
1219 } else {
1220 src_ni = lnet_nid2ni_locked(src_nid, cpt);
1221 if (src_ni == NULL) {
1222 lnet_net_unlock(cpt);
1223 LCONSOLE_WARN("Can't send to %s: src %s is not a "
1224 "local nid\n", libcfs_nid2str(dst_nid),
1225 libcfs_nid2str(src_nid));
1226 return -EINVAL;
1227 }
1228 LASSERT(!msg->msg_routing);
1229 }
1230
1231 /* Is this for someone on a local network? */
1232 local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt);
1233
1234 if (local_ni != NULL) {
1235 if (src_ni == NULL) {
1236 src_ni = local_ni;
1237 src_nid = src_ni->ni_nid;
1238 } else if (src_ni == local_ni) {
1239 lnet_ni_decref_locked(local_ni, cpt);
1240 } else {
1241 lnet_ni_decref_locked(local_ni, cpt);
1242 lnet_ni_decref_locked(src_ni, cpt);
1243 lnet_net_unlock(cpt);
1244 LCONSOLE_WARN("No route to %s via from %s\n",
1245 libcfs_nid2str(dst_nid),
1246 libcfs_nid2str(src_nid));
1247 return -EINVAL;
1248 }
1249
1250 LASSERT(src_nid != LNET_NID_ANY);
1251 lnet_msg_commit(msg, cpt);
1252
1253 if (!msg->msg_routing)
1254 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1255
1256 if (src_ni == the_lnet.ln_loni) {
1257 /* No send credit hassles with LOLND */
1258 lnet_net_unlock(cpt);
1259 lnet_ni_send(src_ni, msg);
1260
1261 lnet_net_lock(cpt);
1262 lnet_ni_decref_locked(src_ni, cpt);
1263 lnet_net_unlock(cpt);
1264 return 0;
1265 }
1266
1267 rc = lnet_nid2peer_locked(&lp, dst_nid, cpt);
1268 /* lp has ref on src_ni; lose mine */
1269 lnet_ni_decref_locked(src_ni, cpt);
1270 if (rc != 0) {
1271 lnet_net_unlock(cpt);
1272 LCONSOLE_WARN("Error %d finding peer %s\n", rc,
1273 libcfs_nid2str(dst_nid));
1274 /* ENOMEM or shutting down */
1275 return rc;
1276 }
1277 LASSERT(lp->lp_ni == src_ni);
1278 } else {
1279 /* sending to a remote network */
1280 lp = lnet_find_route_locked(src_ni, dst_nid, rtr_nid);
1281 if (lp == NULL) {
1282 if (src_ni != NULL)
1283 lnet_ni_decref_locked(src_ni, cpt);
1284 lnet_net_unlock(cpt);
1285
1286 LCONSOLE_WARN("No route to %s via %s "
1287 "(all routers down)\n",
1288 libcfs_id2str(msg->msg_target),
1289 libcfs_nid2str(src_nid));
1290 return -EHOSTUNREACH;
1291 }
1292
1293 /* rtr_nid is LNET_NID_ANY or NID of pre-determined router,
1294 * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't
1295 * pre-determined router, this can happen if router table
1296 * was changed when we release the lock */
1297 if (rtr_nid != lp->lp_nid) {
1298 cpt2 = lnet_cpt_of_nid_locked(lp->lp_nid);
1299 if (cpt2 != cpt) {
1300 if (src_ni != NULL)
1301 lnet_ni_decref_locked(src_ni, cpt);
1302 lnet_net_unlock(cpt);
1303
1304 rtr_nid = lp->lp_nid;
1305 cpt = cpt2;
1306 goto again;
1307 }
1308 }
1309
1310 CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
1311 libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid),
1312 lnet_msgtyp2str(msg->msg_type), msg->msg_len);
1313
1314 if (src_ni == NULL) {
1315 src_ni = lp->lp_ni;
1316 src_nid = src_ni->ni_nid;
1317 } else {
1318 LASSERT(src_ni == lp->lp_ni);
1319 lnet_ni_decref_locked(src_ni, cpt);
1320 }
1321
1322 lnet_peer_addref_locked(lp);
1323
1324 LASSERT(src_nid != LNET_NID_ANY);
1325 lnet_msg_commit(msg, cpt);
1326
1327 if (!msg->msg_routing) {
1328 /* I'm the source and now I know which NI to send on */
1329 msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1330 }
1331
1332 msg->msg_target_is_router = 1;
1333 msg->msg_target.nid = lp->lp_nid;
1334 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1335 }
1336
1337 /* 'lp' is our best choice of peer */
1338
1339 LASSERT(!msg->msg_peertxcredit);
1340 LASSERT(!msg->msg_txcredit);
1341 LASSERT(msg->msg_txpeer == NULL);
1342
1343 msg->msg_txpeer = lp; /* msg takes my ref on lp */
1344
1345 rc = lnet_post_send_locked(msg, 0);
1346 lnet_net_unlock(cpt);
1347
1348 if (rc == EHOSTUNREACH || rc == ECANCELED)
1349 return -rc;
1350
1351 if (rc == 0)
1352 lnet_ni_send(src_ni, msg);
1353
1354 return 0; /* rc == 0 or EAGAIN */
1355 }
1356
1357 static void
lnet_drop_message(lnet_ni_t * ni,int cpt,void * private,unsigned int nob)1358 lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
1359 {
1360 lnet_net_lock(cpt);
1361 the_lnet.ln_counters[cpt]->drop_count++;
1362 the_lnet.ln_counters[cpt]->drop_length += nob;
1363 lnet_net_unlock(cpt);
1364
1365 lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1366 }
1367
1368 static void
lnet_recv_put(lnet_ni_t * ni,lnet_msg_t * msg)1369 lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
1370 {
1371 lnet_hdr_t *hdr = &msg->msg_hdr;
1372
1373 if (msg->msg_wanted != 0)
1374 lnet_setpayloadbuffer(msg);
1375
1376 lnet_build_msg_event(msg, LNET_EVENT_PUT);
1377
1378 /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
1379 * it back into the ACK during lnet_finalize() */
1380 msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1381 (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
1382
1383 lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
1384 msg->msg_offset, msg->msg_wanted, hdr->payload_length);
1385 }
1386
1387 static int
lnet_parse_put(lnet_ni_t * ni,lnet_msg_t * msg)1388 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1389 {
1390 lnet_hdr_t *hdr = &msg->msg_hdr;
1391 struct lnet_match_info info;
1392 int rc;
1393
1394 /* Convert put fields to host byte order */
1395 hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1396 hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1397 hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1398
1399 info.mi_id.nid = hdr->src_nid;
1400 info.mi_id.pid = hdr->src_pid;
1401 info.mi_opc = LNET_MD_OP_PUT;
1402 info.mi_portal = hdr->msg.put.ptl_index;
1403 info.mi_rlength = hdr->payload_length;
1404 info.mi_roffset = hdr->msg.put.offset;
1405 info.mi_mbits = hdr->msg.put.match_bits;
1406
1407 msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL;
1408
1409 again:
1410 rc = lnet_ptl_match_md(&info, msg);
1411 switch (rc) {
1412 default:
1413 LBUG();
1414
1415 case LNET_MATCHMD_OK:
1416 lnet_recv_put(ni, msg);
1417 return 0;
1418
1419 case LNET_MATCHMD_NONE:
1420 if (msg->msg_rx_delayed) /* attached on delayed list */
1421 return 0;
1422
1423 rc = lnet_ni_eager_recv(ni, msg);
1424 if (rc == 0)
1425 goto again;
1426 /* fall through */
1427
1428 case LNET_MATCHMD_DROP:
1429 CNETERR("Dropping PUT from %s portal %d match %llu offset %d length %d: %d\n",
1430 libcfs_id2str(info.mi_id), info.mi_portal,
1431 info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
1432
1433 return ENOENT; /* +ve: OK but no match */
1434 }
1435 }
1436
1437 static int
lnet_parse_get(lnet_ni_t * ni,lnet_msg_t * msg,int rdma_get)1438 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1439 {
1440 struct lnet_match_info info;
1441 lnet_hdr_t *hdr = &msg->msg_hdr;
1442 lnet_handle_wire_t reply_wmd;
1443 int rc;
1444
1445 /* Convert get fields to host byte order */
1446 hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1447 hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1448 hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1449 hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1450
1451 info.mi_id.nid = hdr->src_nid;
1452 info.mi_id.pid = hdr->src_pid;
1453 info.mi_opc = LNET_MD_OP_GET;
1454 info.mi_portal = hdr->msg.get.ptl_index;
1455 info.mi_rlength = hdr->msg.get.sink_length;
1456 info.mi_roffset = hdr->msg.get.src_offset;
1457 info.mi_mbits = hdr->msg.get.match_bits;
1458
1459 rc = lnet_ptl_match_md(&info, msg);
1460 if (rc == LNET_MATCHMD_DROP) {
1461 CNETERR("Dropping GET from %s portal %d match %llu offset %d length %d\n",
1462 libcfs_id2str(info.mi_id), info.mi_portal,
1463 info.mi_mbits, info.mi_roffset, info.mi_rlength);
1464 return ENOENT; /* +ve: OK but no match */
1465 }
1466
1467 LASSERT(rc == LNET_MATCHMD_OK);
1468
1469 lnet_build_msg_event(msg, LNET_EVENT_GET);
1470
1471 reply_wmd = hdr->msg.get.return_wmd;
1472
1473 lnet_prep_send(msg, LNET_MSG_REPLY, info.mi_id,
1474 msg->msg_offset, msg->msg_wanted);
1475
1476 msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1477
1478 if (rdma_get) {
1479 /* The LND completes the REPLY from her recv procedure */
1480 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1481 msg->msg_offset, msg->msg_len, msg->msg_len);
1482 return 0;
1483 }
1484
1485 lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1486 msg->msg_receiving = 0;
1487
1488 rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
1489 if (rc < 0) {
1490 /* didn't get as far as lnet_ni_send() */
1491 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1492 libcfs_nid2str(ni->ni_nid),
1493 libcfs_id2str(info.mi_id), rc);
1494
1495 lnet_finalize(ni, msg, rc);
1496 }
1497
1498 return 0;
1499 }
1500
1501 static int
lnet_parse_reply(lnet_ni_t * ni,lnet_msg_t * msg)1502 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1503 {
1504 void *private = msg->msg_private;
1505 lnet_hdr_t *hdr = &msg->msg_hdr;
1506 lnet_process_id_t src = {0};
1507 lnet_libmd_t *md;
1508 int rlength;
1509 int mlength;
1510 int cpt;
1511
1512 cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
1513 lnet_res_lock(cpt);
1514
1515 src.nid = hdr->src_nid;
1516 src.pid = hdr->src_pid;
1517
1518 /* NB handles only looked up by creator (no flips) */
1519 md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1520 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1521 CNETERR("%s: Dropping REPLY from %s for %s MD %#llx.%#llx\n",
1522 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1523 (md == NULL) ? "invalid" : "inactive",
1524 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1525 hdr->msg.reply.dst_wmd.wh_object_cookie);
1526 if (md != NULL && md->md_me != NULL)
1527 CERROR("REPLY MD also attached to portal %d\n",
1528 md->md_me->me_portal);
1529
1530 lnet_res_unlock(cpt);
1531 return ENOENT; /* +ve: OK but no match */
1532 }
1533
1534 LASSERT(md->md_offset == 0);
1535
1536 rlength = hdr->payload_length;
1537 mlength = MIN(rlength, (int)md->md_length);
1538
1539 if (mlength < rlength &&
1540 (md->md_options & LNET_MD_TRUNCATE) == 0) {
1541 CNETERR("%s: Dropping REPLY from %s length %d for MD %#llx would overflow (%d)\n",
1542 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1543 rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1544 mlength);
1545 lnet_res_unlock(cpt);
1546 return ENOENT; /* +ve: OK but no match */
1547 }
1548
1549 CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
1550 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1551 mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1552
1553 lnet_msg_attach_md(msg, md, 0, mlength);
1554
1555 if (mlength != 0)
1556 lnet_setpayloadbuffer(msg);
1557
1558 lnet_res_unlock(cpt);
1559
1560 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
1561
1562 lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1563 return 0;
1564 }
1565
1566 static int
lnet_parse_ack(lnet_ni_t * ni,lnet_msg_t * msg)1567 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1568 {
1569 lnet_hdr_t *hdr = &msg->msg_hdr;
1570 lnet_process_id_t src = {0};
1571 lnet_libmd_t *md;
1572 int cpt;
1573
1574 src.nid = hdr->src_nid;
1575 src.pid = hdr->src_pid;
1576
1577 /* Convert ack fields to host byte order */
1578 hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1579 hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1580
1581 cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
1582 lnet_res_lock(cpt);
1583
1584 /* NB handles only looked up by creator (no flips) */
1585 md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1586 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
1587 /* Don't moan; this is expected */
1588 CDEBUG(D_NET,
1589 "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
1590 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1591 (md == NULL) ? "invalid" : "inactive",
1592 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1593 hdr->msg.ack.dst_wmd.wh_object_cookie);
1594 if (md != NULL && md->md_me != NULL)
1595 CERROR("Source MD also attached to portal %d\n",
1596 md->md_me->me_portal);
1597
1598 lnet_res_unlock(cpt);
1599 return ENOENT; /* +ve! */
1600 }
1601
1602 CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
1603 libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1604 hdr->msg.ack.dst_wmd.wh_object_cookie);
1605
1606 lnet_msg_attach_md(msg, md, 0, 0);
1607
1608 lnet_res_unlock(cpt);
1609
1610 lnet_build_msg_event(msg, LNET_EVENT_ACK);
1611
1612 lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
1613 return 0;
1614 }
1615
1616 static int
lnet_parse_forward_locked(lnet_ni_t * ni,lnet_msg_t * msg)1617 lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
1618 {
1619 int rc = 0;
1620
1621 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
1622 lnet_msg2bufpool(msg)->rbp_credits <= 0) {
1623 if (ni->ni_lnd->lnd_eager_recv == NULL) {
1624 msg->msg_rx_ready_delay = 1;
1625 } else {
1626 lnet_net_unlock(msg->msg_rx_cpt);
1627 rc = lnet_ni_eager_recv(ni, msg);
1628 lnet_net_lock(msg->msg_rx_cpt);
1629 }
1630 }
1631
1632 if (rc == 0)
1633 rc = lnet_post_routed_recv_locked(msg, 0);
1634 return rc;
1635 }
1636
1637 char *
lnet_msgtyp2str(int type)1638 lnet_msgtyp2str(int type)
1639 {
1640 switch (type) {
1641 case LNET_MSG_ACK:
1642 return "ACK";
1643 case LNET_MSG_PUT:
1644 return "PUT";
1645 case LNET_MSG_GET:
1646 return "GET";
1647 case LNET_MSG_REPLY:
1648 return "REPLY";
1649 case LNET_MSG_HELLO:
1650 return "HELLO";
1651 default:
1652 return "<UNKNOWN>";
1653 }
1654 }
1655 EXPORT_SYMBOL(lnet_msgtyp2str);
1656
1657 void
lnet_print_hdr(lnet_hdr_t * hdr)1658 lnet_print_hdr(lnet_hdr_t *hdr)
1659 {
1660 lnet_process_id_t src = {0};
1661 lnet_process_id_t dst = {0};
1662 char *type_str = lnet_msgtyp2str(hdr->type);
1663
1664 src.nid = hdr->src_nid;
1665 src.pid = hdr->src_pid;
1666
1667 dst.nid = hdr->dest_nid;
1668 dst.pid = hdr->dest_pid;
1669
1670 CWARN("P3 Header at %p of type %s\n", hdr, type_str);
1671 CWARN(" From %s\n", libcfs_id2str(src));
1672 CWARN(" To %s\n", libcfs_id2str(dst));
1673
1674 switch (hdr->type) {
1675 default:
1676 break;
1677
1678 case LNET_MSG_PUT:
1679 CWARN(" Ptl index %d, ack md %#llx.%#llx, "
1680 "match bits %llu\n",
1681 hdr->msg.put.ptl_index,
1682 hdr->msg.put.ack_wmd.wh_interface_cookie,
1683 hdr->msg.put.ack_wmd.wh_object_cookie,
1684 hdr->msg.put.match_bits);
1685 CWARN(" Length %d, offset %d, hdr data %#llx\n",
1686 hdr->payload_length, hdr->msg.put.offset,
1687 hdr->msg.put.hdr_data);
1688 break;
1689
1690 case LNET_MSG_GET:
1691 CWARN(" Ptl index %d, return md %#llx.%#llx, "
1692 "match bits %llu\n", hdr->msg.get.ptl_index,
1693 hdr->msg.get.return_wmd.wh_interface_cookie,
1694 hdr->msg.get.return_wmd.wh_object_cookie,
1695 hdr->msg.get.match_bits);
1696 CWARN(" Length %d, src offset %d\n",
1697 hdr->msg.get.sink_length,
1698 hdr->msg.get.src_offset);
1699 break;
1700
1701 case LNET_MSG_ACK:
1702 CWARN(" dst md %#llx.%#llx, "
1703 "manipulated length %d\n",
1704 hdr->msg.ack.dst_wmd.wh_interface_cookie,
1705 hdr->msg.ack.dst_wmd.wh_object_cookie,
1706 hdr->msg.ack.mlength);
1707 break;
1708
1709 case LNET_MSG_REPLY:
1710 CWARN(" dst md %#llx.%#llx, "
1711 "length %d\n",
1712 hdr->msg.reply.dst_wmd.wh_interface_cookie,
1713 hdr->msg.reply.dst_wmd.wh_object_cookie,
1714 hdr->payload_length);
1715 }
1716
1717 }
1718
1719 int
lnet_parse(lnet_ni_t * ni,lnet_hdr_t * hdr,lnet_nid_t from_nid,void * private,int rdma_req)1720 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
1721 void *private, int rdma_req)
1722 {
1723 int rc = 0;
1724 int cpt;
1725 int for_me;
1726 struct lnet_msg *msg;
1727 lnet_pid_t dest_pid;
1728 lnet_nid_t dest_nid;
1729 lnet_nid_t src_nid;
1730 __u32 payload_length;
1731 __u32 type;
1732
1733 LASSERT(!in_interrupt());
1734
1735 type = le32_to_cpu(hdr->type);
1736 src_nid = le64_to_cpu(hdr->src_nid);
1737 dest_nid = le64_to_cpu(hdr->dest_nid);
1738 dest_pid = le32_to_cpu(hdr->dest_pid);
1739 payload_length = le32_to_cpu(hdr->payload_length);
1740
1741 for_me = (ni->ni_nid == dest_nid);
1742 cpt = lnet_cpt_of_nid(from_nid);
1743
1744 switch (type) {
1745 case LNET_MSG_ACK:
1746 case LNET_MSG_GET:
1747 if (payload_length > 0) {
1748 CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
1749 libcfs_nid2str(from_nid),
1750 libcfs_nid2str(src_nid),
1751 lnet_msgtyp2str(type), payload_length);
1752 return -EPROTO;
1753 }
1754 break;
1755
1756 case LNET_MSG_PUT:
1757 case LNET_MSG_REPLY:
1758 if (payload_length >
1759 (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
1760 CERROR("%s, src %s: bad %s payload %d "
1761 "(%d max expected)\n",
1762 libcfs_nid2str(from_nid),
1763 libcfs_nid2str(src_nid),
1764 lnet_msgtyp2str(type),
1765 payload_length,
1766 for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
1767 return -EPROTO;
1768 }
1769 break;
1770
1771 default:
1772 CERROR("%s, src %s: Bad message type 0x%x\n",
1773 libcfs_nid2str(from_nid),
1774 libcfs_nid2str(src_nid), type);
1775 return -EPROTO;
1776 }
1777
1778 if (the_lnet.ln_routing &&
1779 ni->ni_last_alive != get_seconds()) {
1780 lnet_ni_lock(ni);
1781
1782 /* NB: so far here is the only place to set NI status to "up */
1783 ni->ni_last_alive = get_seconds();
1784 if (ni->ni_status != NULL &&
1785 ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
1786 ni->ni_status->ns_status = LNET_NI_STATUS_UP;
1787 lnet_ni_unlock(ni);
1788 }
1789
1790 /* Regard a bad destination NID as a protocol error. Senders should
1791 * know what they're doing; if they don't they're misconfigured, buggy
1792 * or malicious so we chop them off at the knees :) */
1793
1794 if (!for_me) {
1795 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
1796 /* should have gone direct */
1797 CERROR("%s, src %s: Bad dest nid %s "
1798 "(should have been sent direct)\n",
1799 libcfs_nid2str(from_nid),
1800 libcfs_nid2str(src_nid),
1801 libcfs_nid2str(dest_nid));
1802 return -EPROTO;
1803 }
1804
1805 if (lnet_islocalnid(dest_nid)) {
1806 /* dest is another local NI; sender should have used
1807 * this node's NID on its own network */
1808 CERROR("%s, src %s: Bad dest nid %s "
1809 "(it's my nid but on a different network)\n",
1810 libcfs_nid2str(from_nid),
1811 libcfs_nid2str(src_nid),
1812 libcfs_nid2str(dest_nid));
1813 return -EPROTO;
1814 }
1815
1816 if (rdma_req && type == LNET_MSG_GET) {
1817 CERROR("%s, src %s: Bad optimized GET for %s "
1818 "(final destination must be me)\n",
1819 libcfs_nid2str(from_nid),
1820 libcfs_nid2str(src_nid),
1821 libcfs_nid2str(dest_nid));
1822 return -EPROTO;
1823 }
1824
1825 if (!the_lnet.ln_routing) {
1826 CERROR("%s, src %s: Dropping message for %s "
1827 "(routing not enabled)\n",
1828 libcfs_nid2str(from_nid),
1829 libcfs_nid2str(src_nid),
1830 libcfs_nid2str(dest_nid));
1831 goto drop;
1832 }
1833 }
1834
1835 /* Message looks OK; we're not going to return an error, so we MUST
1836 * call back lnd_recv() come what may... */
1837
1838 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
1839 fail_peer(src_nid, 0)) { /* shall we now? */
1840 CERROR("%s, src %s: Dropping %s to simulate failure\n",
1841 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1842 lnet_msgtyp2str(type));
1843 goto drop;
1844 }
1845
1846 msg = lnet_msg_alloc();
1847 if (msg == NULL) {
1848 CERROR("%s, src %s: Dropping %s (out of memory)\n",
1849 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1850 lnet_msgtyp2str(type));
1851 goto drop;
1852 }
1853
1854 /* msg zeroed in lnet_msg_alloc;
1855 * i.e. flags all clear, pointers NULL etc
1856 */
1857
1858 msg->msg_type = type;
1859 msg->msg_private = private;
1860 msg->msg_receiving = 1;
1861 msg->msg_len = msg->msg_wanted = payload_length;
1862 msg->msg_offset = 0;
1863 msg->msg_hdr = *hdr;
1864 /* for building message event */
1865 msg->msg_from = from_nid;
1866 if (!for_me) {
1867 msg->msg_target.pid = dest_pid;
1868 msg->msg_target.nid = dest_nid;
1869 msg->msg_routing = 1;
1870
1871 } else {
1872 /* convert common msg->hdr fields to host byteorder */
1873 msg->msg_hdr.type = type;
1874 msg->msg_hdr.src_nid = src_nid;
1875 msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
1876 msg->msg_hdr.dest_nid = dest_nid;
1877 msg->msg_hdr.dest_pid = dest_pid;
1878 msg->msg_hdr.payload_length = payload_length;
1879 }
1880
1881 lnet_net_lock(cpt);
1882 rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid, cpt);
1883 if (rc != 0) {
1884 lnet_net_unlock(cpt);
1885 CERROR("%s, src %s: Dropping %s "
1886 "(error %d looking up sender)\n",
1887 libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
1888 lnet_msgtyp2str(type), rc);
1889 lnet_msg_free(msg);
1890 goto drop;
1891 }
1892
1893 lnet_msg_commit(msg, cpt);
1894
1895 if (!for_me) {
1896 rc = lnet_parse_forward_locked(ni, msg);
1897 lnet_net_unlock(cpt);
1898
1899 if (rc < 0)
1900 goto free_drop;
1901 if (rc == 0) {
1902 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1903 0, payload_length, payload_length);
1904 }
1905 return 0;
1906 }
1907
1908 lnet_net_unlock(cpt);
1909
1910 switch (type) {
1911 case LNET_MSG_ACK:
1912 rc = lnet_parse_ack(ni, msg);
1913 break;
1914 case LNET_MSG_PUT:
1915 rc = lnet_parse_put(ni, msg);
1916 break;
1917 case LNET_MSG_GET:
1918 rc = lnet_parse_get(ni, msg, rdma_req);
1919 break;
1920 case LNET_MSG_REPLY:
1921 rc = lnet_parse_reply(ni, msg);
1922 break;
1923 default:
1924 LASSERT(0);
1925 rc = -EPROTO;
1926 goto free_drop; /* prevent an unused label if !kernel */
1927 }
1928
1929 if (rc == 0)
1930 return 0;
1931
1932 LASSERT(rc == ENOENT);
1933
1934 free_drop:
1935 LASSERT(msg->msg_md == NULL);
1936 lnet_finalize(ni, msg, rc);
1937
1938 drop:
1939 lnet_drop_message(ni, cpt, private, payload_length);
1940 return 0;
1941 }
1942 EXPORT_SYMBOL(lnet_parse);
1943
1944 void
lnet_drop_delayed_msg_list(struct list_head * head,char * reason)1945 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
1946 {
1947 while (!list_empty(head)) {
1948 lnet_process_id_t id = {0};
1949 lnet_msg_t *msg;
1950
1951 msg = list_entry(head->next, lnet_msg_t, msg_list);
1952 list_del(&msg->msg_list);
1953
1954 id.nid = msg->msg_hdr.src_nid;
1955 id.pid = msg->msg_hdr.src_pid;
1956
1957 LASSERT(msg->msg_md == NULL);
1958 LASSERT(msg->msg_rx_delayed);
1959 LASSERT(msg->msg_rxpeer != NULL);
1960 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
1961
1962 CWARN("Dropping delayed PUT from %s portal %d match %llu offset %d length %d: %s\n",
1963 libcfs_id2str(id),
1964 msg->msg_hdr.msg.put.ptl_index,
1965 msg->msg_hdr.msg.put.match_bits,
1966 msg->msg_hdr.msg.put.offset,
1967 msg->msg_hdr.payload_length, reason);
1968
1969 /* NB I can't drop msg's ref on msg_rxpeer until after I've
1970 * called lnet_drop_message(), so I just hang onto msg as well
1971 * until that's done */
1972
1973 lnet_drop_message(msg->msg_rxpeer->lp_ni,
1974 msg->msg_rxpeer->lp_cpt,
1975 msg->msg_private, msg->msg_len);
1976 /*
1977 * NB: message will not generate event because w/o attached MD,
1978 * but we still should give error code so lnet_msg_decommit()
1979 * can skip counters operations and other checks.
1980 */
1981 lnet_finalize(msg->msg_rxpeer->lp_ni, msg, -ENOENT);
1982 }
1983 }
1984
1985 void
lnet_recv_delayed_msg_list(struct list_head * head)1986 lnet_recv_delayed_msg_list(struct list_head *head)
1987 {
1988 while (!list_empty(head)) {
1989 lnet_msg_t *msg;
1990 lnet_process_id_t id;
1991
1992 msg = list_entry(head->next, lnet_msg_t, msg_list);
1993 list_del(&msg->msg_list);
1994
1995 /* md won't disappear under me, since each msg
1996 * holds a ref on it */
1997
1998 id.nid = msg->msg_hdr.src_nid;
1999 id.pid = msg->msg_hdr.src_pid;
2000
2001 LASSERT(msg->msg_rx_delayed);
2002 LASSERT(msg->msg_md != NULL);
2003 LASSERT(msg->msg_rxpeer != NULL);
2004 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
2005
2006 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
2007 "match %llu offset %d length %d.\n",
2008 libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
2009 msg->msg_hdr.msg.put.match_bits,
2010 msg->msg_hdr.msg.put.offset,
2011 msg->msg_hdr.payload_length);
2012
2013 lnet_recv_put(msg->msg_rxpeer->lp_ni, msg);
2014 }
2015 }
2016
2017 /**
2018 * Initiate an asynchronous PUT operation.
2019 *
2020 * There are several events associated with a PUT: completion of the send on
2021 * the initiator node (LNET_EVENT_SEND), and when the send completes
2022 * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
2023 * that the operation was accepted by the target. The event LNET_EVENT_PUT is
2024 * used at the target node to indicate the completion of incoming data
2025 * delivery.
2026 *
2027 * The local events will be logged in the EQ associated with the MD pointed to
2028 * by \a mdh handle. Using a MD without an associated EQ results in these
2029 * events being discarded. In this case, the caller must have another
2030 * mechanism (e.g., a higher level protocol) for determining when it is safe
2031 * to modify the memory region associated with the MD.
2032 *
2033 * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
2034 * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
2035 *
2036 * \param self Indicates the NID of a local interface through which to send
2037 * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
2038 * \param mdh A handle for the MD that describes the memory to be sent. The MD
2039 * must be "free floating" (See LNetMDBind()).
2040 * \param ack Controls whether an acknowledgment is requested.
2041 * Acknowledgments are only sent when they are requested by the initiating
2042 * process and the target MD enables them.
2043 * \param target A process identifier for the target process.
2044 * \param portal The index in the \a target's portal table.
2045 * \param match_bits The match bits to use for MD selection at the target
2046 * process.
2047 * \param offset The offset into the target MD (only used when the target
2048 * MD has the LNET_MD_MANAGE_REMOTE option set).
2049 * \param hdr_data 64 bits of user data that can be included in the message
2050 * header. This data is written to an event queue entry at the target if an
2051 * EQ is present on the matching MD.
2052 *
2053 * \retval 0 Success, and only in this case events will be generated
2054 * and logged to EQ (if it exists).
2055 * \retval -EIO Simulated failure.
2056 * \retval -ENOMEM Memory allocation failure.
2057 * \retval -ENOENT Invalid MD object.
2058 *
2059 * \see lnet_event_t::hdr_data and lnet_event_kind_t.
2060 */
2061 int
LNetPut(lnet_nid_t self,lnet_handle_md_t mdh,lnet_ack_req_t ack,lnet_process_id_t target,unsigned int portal,__u64 match_bits,unsigned int offset,__u64 hdr_data)2062 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2063 lnet_process_id_t target, unsigned int portal,
2064 __u64 match_bits, unsigned int offset,
2065 __u64 hdr_data)
2066 {
2067 struct lnet_msg *msg;
2068 struct lnet_libmd *md;
2069 int cpt;
2070 int rc;
2071
2072 LASSERT(the_lnet.ln_init);
2073 LASSERT(the_lnet.ln_refcount > 0);
2074
2075 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2076 fail_peer(target.nid, 1)) { /* shall we now? */
2077 CERROR("Dropping PUT to %s: simulated failure\n",
2078 libcfs_id2str(target));
2079 return -EIO;
2080 }
2081
2082 msg = lnet_msg_alloc();
2083 if (msg == NULL) {
2084 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2085 libcfs_id2str(target));
2086 return -ENOMEM;
2087 }
2088 msg->msg_vmflush = !!memory_pressure_get();
2089
2090 cpt = lnet_cpt_of_cookie(mdh.cookie);
2091 lnet_res_lock(cpt);
2092
2093 md = lnet_handle2md(&mdh);
2094 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2095 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
2096 match_bits, portal, libcfs_id2str(target),
2097 md == NULL ? -1 : md->md_threshold);
2098 if (md != NULL && md->md_me != NULL)
2099 CERROR("Source MD also attached to portal %d\n",
2100 md->md_me->me_portal);
2101 lnet_res_unlock(cpt);
2102
2103 lnet_msg_free(msg);
2104 return -ENOENT;
2105 }
2106
2107 CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2108
2109 lnet_msg_attach_md(msg, md, 0, 0);
2110
2111 lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2112
2113 msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2114 msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2115 msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2116 msg->msg_hdr.msg.put.hdr_data = hdr_data;
2117
2118 /* NB handles only looked up by creator (no flips) */
2119 if (ack == LNET_ACK_REQ) {
2120 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2121 the_lnet.ln_interface_cookie;
2122 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2123 md->md_lh.lh_cookie;
2124 } else {
2125 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2126 LNET_WIRE_HANDLE_COOKIE_NONE;
2127 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2128 LNET_WIRE_HANDLE_COOKIE_NONE;
2129 }
2130
2131 lnet_res_unlock(cpt);
2132
2133 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2134
2135 rc = lnet_send(self, msg, LNET_NID_ANY);
2136 if (rc != 0) {
2137 CNETERR("Error sending PUT to %s: %d\n",
2138 libcfs_id2str(target), rc);
2139 lnet_finalize(NULL, msg, rc);
2140 }
2141
2142 /* completion will be signalled by an event */
2143 return 0;
2144 }
2145 EXPORT_SYMBOL(LNetPut);
2146
2147 lnet_msg_t *
lnet_create_reply_msg(lnet_ni_t * ni,lnet_msg_t * getmsg)2148 lnet_create_reply_msg(lnet_ni_t *ni, lnet_msg_t *getmsg)
2149 {
2150 /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
2151 * returns a msg for the LND to pass to lnet_finalize() when the sink
2152 * data has been received.
2153 *
2154 * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2155 * lnet_finalize() is called on it, so the LND must call this first */
2156
2157 struct lnet_msg *msg = lnet_msg_alloc();
2158 struct lnet_libmd *getmd = getmsg->msg_md;
2159 lnet_process_id_t peer_id = getmsg->msg_target;
2160 int cpt;
2161
2162 LASSERT(!getmsg->msg_target_is_router);
2163 LASSERT(!getmsg->msg_routing);
2164
2165 cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
2166 lnet_res_lock(cpt);
2167
2168 LASSERT(getmd->md_refcount > 0);
2169
2170 if (msg == NULL) {
2171 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
2172 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2173 goto drop;
2174 }
2175
2176 if (getmd->md_threshold == 0) {
2177 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
2178 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2179 getmd);
2180 lnet_res_unlock(cpt);
2181 goto drop;
2182 }
2183
2184 LASSERT(getmd->md_offset == 0);
2185
2186 CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2187 libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2188
2189 /* setup information for lnet_build_msg_event */
2190 msg->msg_from = peer_id.nid;
2191 msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2192 msg->msg_hdr.src_nid = peer_id.nid;
2193 msg->msg_hdr.payload_length = getmd->md_length;
2194 msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
2195
2196 lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
2197 lnet_res_unlock(cpt);
2198
2199 cpt = lnet_cpt_of_nid(peer_id.nid);
2200
2201 lnet_net_lock(cpt);
2202 lnet_msg_commit(msg, cpt);
2203 lnet_net_unlock(cpt);
2204
2205 lnet_build_msg_event(msg, LNET_EVENT_REPLY);
2206
2207 return msg;
2208
2209 drop:
2210 cpt = lnet_cpt_of_nid(peer_id.nid);
2211
2212 lnet_net_lock(cpt);
2213 the_lnet.ln_counters[cpt]->drop_count++;
2214 the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
2215 lnet_net_unlock(cpt);
2216
2217 if (msg != NULL)
2218 lnet_msg_free(msg);
2219
2220 return NULL;
2221 }
2222 EXPORT_SYMBOL(lnet_create_reply_msg);
2223
2224 void
lnet_set_reply_msg_len(lnet_ni_t * ni,lnet_msg_t * reply,unsigned int len)2225 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2226 {
2227 /* Set the REPLY length, now the RDMA that elides the REPLY message has
2228 * completed and I know it. */
2229 LASSERT(reply != NULL);
2230 LASSERT(reply->msg_type == LNET_MSG_GET);
2231 LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
2232
2233 /* NB I trusted my peer to RDMA. If she tells me she's written beyond
2234 * the end of my buffer, I might as well be dead. */
2235 LASSERT(len <= reply->msg_ev.mlength);
2236
2237 reply->msg_ev.mlength = len;
2238 }
2239 EXPORT_SYMBOL(lnet_set_reply_msg_len);
2240
2241 /**
2242 * Initiate an asynchronous GET operation.
2243 *
2244 * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
2245 * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
2246 * the target node in the REPLY has been written to local MD.
2247 *
2248 * On the target node, an LNET_EVENT_GET is logged when the GET request
2249 * arrives and is accepted into a MD.
2250 *
2251 * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
2252 * \param mdh A handle for the MD that describes the memory into which the
2253 * requested data will be received. The MD must be "free floating"
2254 * (See LNetMDBind()).
2255 *
2256 * \retval 0 Success, and only in this case events will be generated
2257 * and logged to EQ (if it exists) of the MD.
2258 * \retval -EIO Simulated failure.
2259 * \retval -ENOMEM Memory allocation failure.
2260 * \retval -ENOENT Invalid MD object.
2261 */
2262 int
LNetGet(lnet_nid_t self,lnet_handle_md_t mdh,lnet_process_id_t target,unsigned int portal,__u64 match_bits,unsigned int offset)2263 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2264 lnet_process_id_t target, unsigned int portal,
2265 __u64 match_bits, unsigned int offset)
2266 {
2267 struct lnet_msg *msg;
2268 struct lnet_libmd *md;
2269 int cpt;
2270 int rc;
2271
2272 LASSERT(the_lnet.ln_init);
2273 LASSERT(the_lnet.ln_refcount > 0);
2274
2275 if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
2276 fail_peer(target.nid, 1)) { /* shall we now? */
2277 CERROR("Dropping GET to %s: simulated failure\n",
2278 libcfs_id2str(target));
2279 return -EIO;
2280 }
2281
2282 msg = lnet_msg_alloc();
2283 if (msg == NULL) {
2284 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2285 libcfs_id2str(target));
2286 return -ENOMEM;
2287 }
2288
2289 cpt = lnet_cpt_of_cookie(mdh.cookie);
2290 lnet_res_lock(cpt);
2291
2292 md = lnet_handle2md(&mdh);
2293 if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
2294 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
2295 match_bits, portal, libcfs_id2str(target),
2296 md == NULL ? -1 : md->md_threshold);
2297 if (md != NULL && md->md_me != NULL)
2298 CERROR("REPLY MD also attached to portal %d\n",
2299 md->md_me->me_portal);
2300
2301 lnet_res_unlock(cpt);
2302
2303 lnet_msg_free(msg);
2304 return -ENOENT;
2305 }
2306
2307 CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2308
2309 lnet_msg_attach_md(msg, md, 0, 0);
2310
2311 lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2312
2313 msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2314 msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2315 msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2316 msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2317
2318 /* NB handles only looked up by creator (no flips) */
2319 msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2320 the_lnet.ln_interface_cookie;
2321 msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2322 md->md_lh.lh_cookie;
2323
2324 lnet_res_unlock(cpt);
2325
2326 lnet_build_msg_event(msg, LNET_EVENT_SEND);
2327
2328 rc = lnet_send(self, msg, LNET_NID_ANY);
2329 if (rc < 0) {
2330 CNETERR("Error sending GET to %s: %d\n",
2331 libcfs_id2str(target), rc);
2332 lnet_finalize(NULL, msg, rc);
2333 }
2334
2335 /* completion will be signalled by an event */
2336 return 0;
2337 }
2338 EXPORT_SYMBOL(LNetGet);
2339
2340 /**
2341 * Calculate distance to node at \a dstnid.
2342 *
2343 * \param dstnid Target NID.
2344 * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
2345 * is saved here.
2346 * \param orderp If not NULL, order of the route to reach \a dstnid is saved
2347 * here.
2348 *
2349 * \retval 0 If \a dstnid belongs to a local interface, and reserved option
2350 * local_nid_dist_zero is set, which is the default.
2351 * \retval positives Distance to target NID, i.e. number of hops plus one.
2352 * \retval -EHOSTUNREACH If \a dstnid is not reachable.
2353 */
2354 int
LNetDist(lnet_nid_t dstnid,lnet_nid_t * srcnidp,__u32 * orderp)2355 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2356 {
2357 struct list_head *e;
2358 struct lnet_ni *ni;
2359 lnet_remotenet_t *rnet;
2360 __u32 dstnet = LNET_NIDNET(dstnid);
2361 int hops;
2362 int cpt;
2363 __u32 order = 2;
2364 struct list_head *rn_list;
2365
2366 /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2367 * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2368 * keep order 0 free for 0@lo and order 1 free for a local NID
2369 * match */
2370
2371 LASSERT(the_lnet.ln_init);
2372 LASSERT(the_lnet.ln_refcount > 0);
2373
2374 cpt = lnet_net_lock_current();
2375
2376 list_for_each(e, &the_lnet.ln_nis) {
2377 ni = list_entry(e, lnet_ni_t, ni_list);
2378
2379 if (ni->ni_nid == dstnid) {
2380 if (srcnidp != NULL)
2381 *srcnidp = dstnid;
2382 if (orderp != NULL) {
2383 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2384 *orderp = 0;
2385 else
2386 *orderp = 1;
2387 }
2388 lnet_net_unlock(cpt);
2389
2390 return local_nid_dist_zero ? 0 : 1;
2391 }
2392
2393 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
2394 if (srcnidp != NULL)
2395 *srcnidp = ni->ni_nid;
2396 if (orderp != NULL)
2397 *orderp = order;
2398 lnet_net_unlock(cpt);
2399 return 1;
2400 }
2401
2402 order++;
2403 }
2404
2405 rn_list = lnet_net2rnethash(dstnet);
2406 list_for_each(e, rn_list) {
2407 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2408
2409 if (rnet->lrn_net == dstnet) {
2410 lnet_route_t *route;
2411 lnet_route_t *shortest = NULL;
2412
2413 LASSERT(!list_empty(&rnet->lrn_routes));
2414
2415 list_for_each_entry(route, &rnet->lrn_routes,
2416 lr_list) {
2417 if (shortest == NULL ||
2418 route->lr_hops < shortest->lr_hops)
2419 shortest = route;
2420 }
2421
2422 LASSERT(shortest != NULL);
2423 hops = shortest->lr_hops;
2424 if (srcnidp != NULL)
2425 *srcnidp = shortest->lr_gateway->lp_ni->ni_nid;
2426 if (orderp != NULL)
2427 *orderp = order;
2428 lnet_net_unlock(cpt);
2429 return hops + 1;
2430 }
2431 order++;
2432 }
2433
2434 lnet_net_unlock(cpt);
2435 return -EHOSTUNREACH;
2436 }
2437 EXPORT_SYMBOL(LNetDist);
2438
2439 /**
2440 * Set the number of asynchronous messages expected from a target process.
2441 *
2442 * This function is only meaningful for userspace callers. It's a no-op when
2443 * called from kernel.
2444 *
2445 * Asynchronous messages are those that can come from a target when the
2446 * userspace process is not waiting for IO to complete; e.g., AST callbacks
2447 * from Lustre servers. Specifying the expected number of such messages
2448 * allows them to be eagerly received when user process is not running in
2449 * LNet; otherwise network errors may occur.
2450 *
2451 * \param id Process ID of the target process.
2452 * \param nasync Number of asynchronous messages expected from the target.
2453 *
2454 * \return 0 on success, and an error code otherwise.
2455 */
2456 int
LNetSetAsync(lnet_process_id_t id,int nasync)2457 LNetSetAsync(lnet_process_id_t id, int nasync)
2458 {
2459 return 0;
2460 }
2461 EXPORT_SYMBOL(LNetSetAsync);
2462