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, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
20 *
21 * GPL HEADER END
22 */
23 /*
24 * Copyright (c) 2012, Intel Corporation.
25 */
26 /*
27 * This file is part of Lustre, http://www.lustre.org/
28 * Lustre is a trademark of Sun Microsystems, Inc.
29 *
30 * lnet/lnet/lib-ptl.c
31 *
32 * portal & match routines
33 *
34 * Author: liang@whamcloud.com
35 */
36
37 #define DEBUG_SUBSYSTEM S_LNET
38
39 #include "../../include/linux/lnet/lib-lnet.h"
40
41 /* NB: add /proc interfaces in upcoming patches */
42 int portal_rotor = LNET_PTL_ROTOR_HASH_RT;
43 module_param(portal_rotor, int, 0644);
44 MODULE_PARM_DESC(portal_rotor, "redirect PUTs to different cpu-partitions");
45
46 static int
lnet_ptl_match_type(unsigned int index,lnet_process_id_t match_id,__u64 mbits,__u64 ignore_bits)47 lnet_ptl_match_type(unsigned int index, lnet_process_id_t match_id,
48 __u64 mbits, __u64 ignore_bits)
49 {
50 struct lnet_portal *ptl = the_lnet.ln_portals[index];
51 int unique;
52
53 unique = ignore_bits == 0 &&
54 match_id.nid != LNET_NID_ANY &&
55 match_id.pid != LNET_PID_ANY;
56
57 LASSERT(!lnet_ptl_is_unique(ptl) || !lnet_ptl_is_wildcard(ptl));
58
59 /* prefer to check w/o any lock */
60 if (likely(lnet_ptl_is_unique(ptl) || lnet_ptl_is_wildcard(ptl)))
61 goto match;
62
63 /* unset, new portal */
64 lnet_ptl_lock(ptl);
65 /* check again with lock */
66 if (unlikely(lnet_ptl_is_unique(ptl) || lnet_ptl_is_wildcard(ptl))) {
67 lnet_ptl_unlock(ptl);
68 goto match;
69 }
70
71 /* still not set */
72 if (unique)
73 lnet_ptl_setopt(ptl, LNET_PTL_MATCH_UNIQUE);
74 else
75 lnet_ptl_setopt(ptl, LNET_PTL_MATCH_WILDCARD);
76
77 lnet_ptl_unlock(ptl);
78
79 return 1;
80
81 match:
82 if ((lnet_ptl_is_unique(ptl) && !unique) ||
83 (lnet_ptl_is_wildcard(ptl) && unique))
84 return 0;
85 return 1;
86 }
87
88 static void
lnet_ptl_enable_mt(struct lnet_portal * ptl,int cpt)89 lnet_ptl_enable_mt(struct lnet_portal *ptl, int cpt)
90 {
91 struct lnet_match_table *mtable = ptl->ptl_mtables[cpt];
92 int i;
93
94 /* with hold of both lnet_res_lock(cpt) and lnet_ptl_lock */
95 LASSERT(lnet_ptl_is_wildcard(ptl));
96
97 mtable->mt_enabled = 1;
98
99 ptl->ptl_mt_maps[ptl->ptl_mt_nmaps] = cpt;
100 for (i = ptl->ptl_mt_nmaps - 1; i >= 0; i--) {
101 LASSERT(ptl->ptl_mt_maps[i] != cpt);
102 if (ptl->ptl_mt_maps[i] < cpt)
103 break;
104
105 /* swap to order */
106 ptl->ptl_mt_maps[i + 1] = ptl->ptl_mt_maps[i];
107 ptl->ptl_mt_maps[i] = cpt;
108 }
109
110 ptl->ptl_mt_nmaps++;
111 }
112
113 static void
lnet_ptl_disable_mt(struct lnet_portal * ptl,int cpt)114 lnet_ptl_disable_mt(struct lnet_portal *ptl, int cpt)
115 {
116 struct lnet_match_table *mtable = ptl->ptl_mtables[cpt];
117 int i;
118
119 /* with hold of both lnet_res_lock(cpt) and lnet_ptl_lock */
120 LASSERT(lnet_ptl_is_wildcard(ptl));
121
122 if (LNET_CPT_NUMBER == 1)
123 return; /* never disable the only match-table */
124
125 mtable->mt_enabled = 0;
126
127 LASSERT(ptl->ptl_mt_nmaps > 0 &&
128 ptl->ptl_mt_nmaps <= LNET_CPT_NUMBER);
129
130 /* remove it from mt_maps */
131 ptl->ptl_mt_nmaps--;
132 for (i = 0; i < ptl->ptl_mt_nmaps; i++) {
133 if (ptl->ptl_mt_maps[i] >= cpt) /* overwrite it */
134 ptl->ptl_mt_maps[i] = ptl->ptl_mt_maps[i + 1];
135 }
136 }
137
138 static int
lnet_try_match_md(lnet_libmd_t * md,struct lnet_match_info * info,struct lnet_msg * msg)139 lnet_try_match_md(lnet_libmd_t *md,
140 struct lnet_match_info *info, struct lnet_msg *msg)
141 {
142 /* ALWAYS called holding the lnet_res_lock, and can't lnet_res_unlock;
143 * lnet_match_blocked_msg() relies on this to avoid races */
144 unsigned int offset;
145 unsigned int mlength;
146 lnet_me_t *me = md->md_me;
147
148 /* MD exhausted */
149 if (lnet_md_exhausted(md))
150 return LNET_MATCHMD_NONE | LNET_MATCHMD_EXHAUSTED;
151
152 /* mismatched MD op */
153 if ((md->md_options & info->mi_opc) == 0)
154 return LNET_MATCHMD_NONE;
155
156 /* mismatched ME nid/pid? */
157 if (me->me_match_id.nid != LNET_NID_ANY &&
158 me->me_match_id.nid != info->mi_id.nid)
159 return LNET_MATCHMD_NONE;
160
161 if (me->me_match_id.pid != LNET_PID_ANY &&
162 me->me_match_id.pid != info->mi_id.pid)
163 return LNET_MATCHMD_NONE;
164
165 /* mismatched ME matchbits? */
166 if (((me->me_match_bits ^ info->mi_mbits) & ~me->me_ignore_bits) != 0)
167 return LNET_MATCHMD_NONE;
168
169 /* Hurrah! This _is_ a match; check it out... */
170
171 if ((md->md_options & LNET_MD_MANAGE_REMOTE) == 0)
172 offset = md->md_offset;
173 else
174 offset = info->mi_roffset;
175
176 if ((md->md_options & LNET_MD_MAX_SIZE) != 0) {
177 mlength = md->md_max_size;
178 LASSERT(md->md_offset + mlength <= md->md_length);
179 } else {
180 mlength = md->md_length - offset;
181 }
182
183 if (info->mi_rlength <= mlength) { /* fits in allowed space */
184 mlength = info->mi_rlength;
185 } else if ((md->md_options & LNET_MD_TRUNCATE) == 0) {
186 /* this packet _really_ is too big */
187 CERROR("Matching packet from %s, match %llu length %d too big: %d left, %d allowed\n",
188 libcfs_id2str(info->mi_id), info->mi_mbits,
189 info->mi_rlength, md->md_length - offset, mlength);
190
191 return LNET_MATCHMD_DROP;
192 }
193
194 /* Commit to this ME/MD */
195 CDEBUG(D_NET, "Incoming %s index %x from %s of "
196 "length %d/%d into md %#llx [%d] + %d\n",
197 (info->mi_opc == LNET_MD_OP_PUT) ? "put" : "get",
198 info->mi_portal, libcfs_id2str(info->mi_id), mlength,
199 info->mi_rlength, md->md_lh.lh_cookie, md->md_niov, offset);
200
201 lnet_msg_attach_md(msg, md, offset, mlength);
202 md->md_offset = offset + mlength;
203
204 if (!lnet_md_exhausted(md))
205 return LNET_MATCHMD_OK;
206
207 /* Auto-unlink NOW, so the ME gets unlinked if required.
208 * We bumped md->md_refcount above so the MD just gets flagged
209 * for unlink when it is finalized. */
210 if ((md->md_flags & LNET_MD_FLAG_AUTO_UNLINK) != 0)
211 lnet_md_unlink(md);
212
213 return LNET_MATCHMD_OK | LNET_MATCHMD_EXHAUSTED;
214 }
215
216 static struct lnet_match_table *
lnet_match2mt(struct lnet_portal * ptl,lnet_process_id_t id,__u64 mbits)217 lnet_match2mt(struct lnet_portal *ptl, lnet_process_id_t id, __u64 mbits)
218 {
219 if (LNET_CPT_NUMBER == 1)
220 return ptl->ptl_mtables[0]; /* the only one */
221
222 /* if it's a unique portal, return match-table hashed by NID */
223 return lnet_ptl_is_unique(ptl) ?
224 ptl->ptl_mtables[lnet_cpt_of_nid(id.nid)] : NULL;
225 }
226
227 struct lnet_match_table *
lnet_mt_of_attach(unsigned int index,lnet_process_id_t id,__u64 mbits,__u64 ignore_bits,lnet_ins_pos_t pos)228 lnet_mt_of_attach(unsigned int index, lnet_process_id_t id,
229 __u64 mbits, __u64 ignore_bits, lnet_ins_pos_t pos)
230 {
231 struct lnet_portal *ptl;
232 struct lnet_match_table *mtable;
233
234 /* NB: called w/o lock */
235 LASSERT(index < the_lnet.ln_nportals);
236
237 if (!lnet_ptl_match_type(index, id, mbits, ignore_bits))
238 return NULL;
239
240 ptl = the_lnet.ln_portals[index];
241
242 mtable = lnet_match2mt(ptl, id, mbits);
243 if (mtable != NULL) /* unique portal or only one match-table */
244 return mtable;
245
246 /* it's a wildcard portal */
247 switch (pos) {
248 default:
249 return NULL;
250 case LNET_INS_BEFORE:
251 case LNET_INS_AFTER:
252 /* posted by no affinity thread, always hash to specific
253 * match-table to avoid buffer stealing which is heavy */
254 return ptl->ptl_mtables[ptl->ptl_index % LNET_CPT_NUMBER];
255 case LNET_INS_LOCAL:
256 /* posted by cpu-affinity thread */
257 return ptl->ptl_mtables[lnet_cpt_current()];
258 }
259 }
260
261 static struct lnet_match_table *
lnet_mt_of_match(struct lnet_match_info * info,struct lnet_msg * msg)262 lnet_mt_of_match(struct lnet_match_info *info, struct lnet_msg *msg)
263 {
264 struct lnet_match_table *mtable;
265 struct lnet_portal *ptl;
266 int nmaps;
267 int rotor;
268 int routed;
269 int cpt;
270
271 /* NB: called w/o lock */
272 LASSERT(info->mi_portal < the_lnet.ln_nportals);
273 ptl = the_lnet.ln_portals[info->mi_portal];
274
275 LASSERT(lnet_ptl_is_wildcard(ptl) || lnet_ptl_is_unique(ptl));
276
277 mtable = lnet_match2mt(ptl, info->mi_id, info->mi_mbits);
278 if (mtable != NULL)
279 return mtable;
280
281 /* it's a wildcard portal */
282 routed = LNET_NIDNET(msg->msg_hdr.src_nid) !=
283 LNET_NIDNET(msg->msg_hdr.dest_nid);
284
285 if (portal_rotor == LNET_PTL_ROTOR_OFF ||
286 (portal_rotor != LNET_PTL_ROTOR_ON && !routed)) {
287 cpt = lnet_cpt_current();
288 if (ptl->ptl_mtables[cpt]->mt_enabled)
289 return ptl->ptl_mtables[cpt];
290 }
291
292 rotor = ptl->ptl_rotor++; /* get round-robin factor */
293 if (portal_rotor == LNET_PTL_ROTOR_HASH_RT && routed)
294 cpt = lnet_cpt_of_nid(msg->msg_hdr.src_nid);
295 else
296 cpt = rotor % LNET_CPT_NUMBER;
297
298 if (!ptl->ptl_mtables[cpt]->mt_enabled) {
299 /* is there any active entry for this portal? */
300 nmaps = ptl->ptl_mt_nmaps;
301 /* map to an active mtable to avoid heavy "stealing" */
302 if (nmaps != 0) {
303 /* NB: there is possibility that ptl_mt_maps is being
304 * changed because we are not under protection of
305 * lnet_ptl_lock, but it shouldn't hurt anything */
306 cpt = ptl->ptl_mt_maps[rotor % nmaps];
307 }
308 }
309
310 return ptl->ptl_mtables[cpt];
311 }
312
313 static int
lnet_mt_test_exhausted(struct lnet_match_table * mtable,int pos)314 lnet_mt_test_exhausted(struct lnet_match_table *mtable, int pos)
315 {
316 __u64 *bmap;
317 int i;
318
319 if (!lnet_ptl_is_wildcard(the_lnet.ln_portals[mtable->mt_portal]))
320 return 0;
321
322 if (pos < 0) { /* check all bits */
323 for (i = 0; i < LNET_MT_EXHAUSTED_BMAP; i++) {
324 if (mtable->mt_exhausted[i] != (__u64)(-1))
325 return 0;
326 }
327 return 1;
328 }
329
330 LASSERT(pos <= LNET_MT_HASH_IGNORE);
331 /* mtable::mt_mhash[pos] is marked as exhausted or not */
332 bmap = &mtable->mt_exhausted[pos >> LNET_MT_BITS_U64];
333 pos &= (1 << LNET_MT_BITS_U64) - 1;
334
335 return ((*bmap) & (1ULL << pos)) != 0;
336 }
337
338 static void
lnet_mt_set_exhausted(struct lnet_match_table * mtable,int pos,int exhausted)339 lnet_mt_set_exhausted(struct lnet_match_table *mtable, int pos, int exhausted)
340 {
341 __u64 *bmap;
342
343 LASSERT(lnet_ptl_is_wildcard(the_lnet.ln_portals[mtable->mt_portal]));
344 LASSERT(pos <= LNET_MT_HASH_IGNORE);
345
346 /* set mtable::mt_mhash[pos] as exhausted/non-exhausted */
347 bmap = &mtable->mt_exhausted[pos >> LNET_MT_BITS_U64];
348 pos &= (1 << LNET_MT_BITS_U64) - 1;
349
350 if (!exhausted)
351 *bmap &= ~(1ULL << pos);
352 else
353 *bmap |= 1ULL << pos;
354 }
355
356 struct list_head *
lnet_mt_match_head(struct lnet_match_table * mtable,lnet_process_id_t id,__u64 mbits)357 lnet_mt_match_head(struct lnet_match_table *mtable,
358 lnet_process_id_t id, __u64 mbits)
359 {
360 struct lnet_portal *ptl = the_lnet.ln_portals[mtable->mt_portal];
361
362 if (lnet_ptl_is_wildcard(ptl)) {
363 return &mtable->mt_mhash[mbits & LNET_MT_HASH_MASK];
364 } else {
365 unsigned long hash = mbits + id.nid + id.pid;
366
367 LASSERT(lnet_ptl_is_unique(ptl));
368 hash = hash_long(hash, LNET_MT_HASH_BITS);
369 return &mtable->mt_mhash[hash];
370 }
371 }
372
373 int
lnet_mt_match_md(struct lnet_match_table * mtable,struct lnet_match_info * info,struct lnet_msg * msg)374 lnet_mt_match_md(struct lnet_match_table *mtable,
375 struct lnet_match_info *info, struct lnet_msg *msg)
376 {
377 struct list_head *head;
378 lnet_me_t *me;
379 lnet_me_t *tmp;
380 int exhausted = 0;
381 int rc;
382
383 /* any ME with ignore bits? */
384 if (!list_empty(&mtable->mt_mhash[LNET_MT_HASH_IGNORE]))
385 head = &mtable->mt_mhash[LNET_MT_HASH_IGNORE];
386 else
387 head = lnet_mt_match_head(mtable, info->mi_id, info->mi_mbits);
388 again:
389 /* NB: only wildcard portal needs to return LNET_MATCHMD_EXHAUSTED */
390 if (lnet_ptl_is_wildcard(the_lnet.ln_portals[mtable->mt_portal]))
391 exhausted = LNET_MATCHMD_EXHAUSTED;
392
393 list_for_each_entry_safe(me, tmp, head, me_list) {
394 /* ME attached but MD not attached yet */
395 if (me->me_md == NULL)
396 continue;
397
398 LASSERT(me == me->me_md->md_me);
399
400 rc = lnet_try_match_md(me->me_md, info, msg);
401 if ((rc & LNET_MATCHMD_EXHAUSTED) == 0)
402 exhausted = 0; /* mlist is not empty */
403
404 if ((rc & LNET_MATCHMD_FINISH) != 0) {
405 /* don't return EXHAUSTED bit because we don't know
406 * whether the mlist is empty or not */
407 return rc & ~LNET_MATCHMD_EXHAUSTED;
408 }
409 }
410
411 if (exhausted == LNET_MATCHMD_EXHAUSTED) { /* @head is exhausted */
412 lnet_mt_set_exhausted(mtable, head - mtable->mt_mhash, 1);
413 if (!lnet_mt_test_exhausted(mtable, -1))
414 exhausted = 0;
415 }
416
417 if (exhausted == 0 && head == &mtable->mt_mhash[LNET_MT_HASH_IGNORE]) {
418 head = lnet_mt_match_head(mtable, info->mi_id, info->mi_mbits);
419 goto again; /* re-check MEs w/o ignore-bits */
420 }
421
422 if (info->mi_opc == LNET_MD_OP_GET ||
423 !lnet_ptl_is_lazy(the_lnet.ln_portals[info->mi_portal]))
424 return LNET_MATCHMD_DROP | exhausted;
425
426 return LNET_MATCHMD_NONE | exhausted;
427 }
428
429 static int
lnet_ptl_match_early(struct lnet_portal * ptl,struct lnet_msg * msg)430 lnet_ptl_match_early(struct lnet_portal *ptl, struct lnet_msg *msg)
431 {
432 int rc;
433
434 /* message arrived before any buffer posting on this portal,
435 * simply delay or drop this message */
436 if (likely(lnet_ptl_is_wildcard(ptl) || lnet_ptl_is_unique(ptl)))
437 return 0;
438
439 lnet_ptl_lock(ptl);
440 /* check it again with hold of lock */
441 if (lnet_ptl_is_wildcard(ptl) || lnet_ptl_is_unique(ptl)) {
442 lnet_ptl_unlock(ptl);
443 return 0;
444 }
445
446 if (lnet_ptl_is_lazy(ptl)) {
447 if (msg->msg_rx_ready_delay) {
448 msg->msg_rx_delayed = 1;
449 list_add_tail(&msg->msg_list,
450 &ptl->ptl_msg_delayed);
451 }
452 rc = LNET_MATCHMD_NONE;
453 } else {
454 rc = LNET_MATCHMD_DROP;
455 }
456
457 lnet_ptl_unlock(ptl);
458 return rc;
459 }
460
461 static int
lnet_ptl_match_delay(struct lnet_portal * ptl,struct lnet_match_info * info,struct lnet_msg * msg)462 lnet_ptl_match_delay(struct lnet_portal *ptl,
463 struct lnet_match_info *info, struct lnet_msg *msg)
464 {
465 int first = ptl->ptl_mt_maps[0]; /* read w/o lock */
466 int rc = 0;
467 int i;
468
469 /* steal buffer from other CPTs, and delay it if nothing to steal,
470 * this function is more expensive than a regular match, but we
471 * don't expect it can happen a lot */
472 LASSERT(lnet_ptl_is_wildcard(ptl));
473
474 for (i = 0; i < LNET_CPT_NUMBER; i++) {
475 struct lnet_match_table *mtable;
476 int cpt;
477
478 cpt = (first + i) % LNET_CPT_NUMBER;
479 mtable = ptl->ptl_mtables[cpt];
480 if (i != 0 && i != LNET_CPT_NUMBER - 1 && !mtable->mt_enabled)
481 continue;
482
483 lnet_res_lock(cpt);
484 lnet_ptl_lock(ptl);
485
486 if (i == 0) { /* the first try, attach on stealing list */
487 list_add_tail(&msg->msg_list,
488 &ptl->ptl_msg_stealing);
489 }
490
491 if (!list_empty(&msg->msg_list)) { /* on stealing list */
492 rc = lnet_mt_match_md(mtable, info, msg);
493
494 if ((rc & LNET_MATCHMD_EXHAUSTED) != 0 &&
495 mtable->mt_enabled)
496 lnet_ptl_disable_mt(ptl, cpt);
497
498 if ((rc & LNET_MATCHMD_FINISH) != 0)
499 list_del_init(&msg->msg_list);
500
501 } else {
502 /* could be matched by lnet_ptl_attach_md()
503 * which is called by another thread */
504 rc = msg->msg_md == NULL ?
505 LNET_MATCHMD_DROP : LNET_MATCHMD_OK;
506 }
507
508 if (!list_empty(&msg->msg_list) && /* not matched yet */
509 (i == LNET_CPT_NUMBER - 1 || /* the last CPT */
510 ptl->ptl_mt_nmaps == 0 || /* no active CPT */
511 (ptl->ptl_mt_nmaps == 1 && /* the only active CPT */
512 ptl->ptl_mt_maps[0] == cpt))) {
513 /* nothing to steal, delay or drop */
514 list_del_init(&msg->msg_list);
515
516 if (lnet_ptl_is_lazy(ptl)) {
517 msg->msg_rx_delayed = 1;
518 list_add_tail(&msg->msg_list,
519 &ptl->ptl_msg_delayed);
520 rc = LNET_MATCHMD_NONE;
521 } else {
522 rc = LNET_MATCHMD_DROP;
523 }
524 }
525
526 lnet_ptl_unlock(ptl);
527 lnet_res_unlock(cpt);
528
529 if ((rc & LNET_MATCHMD_FINISH) != 0 || msg->msg_rx_delayed)
530 break;
531 }
532
533 return rc;
534 }
535
536 int
lnet_ptl_match_md(struct lnet_match_info * info,struct lnet_msg * msg)537 lnet_ptl_match_md(struct lnet_match_info *info, struct lnet_msg *msg)
538 {
539 struct lnet_match_table *mtable;
540 struct lnet_portal *ptl;
541 int rc;
542
543 CDEBUG(D_NET, "Request from %s of length %d into portal %d MB=%#llx\n",
544 libcfs_id2str(info->mi_id), info->mi_rlength, info->mi_portal,
545 info->mi_mbits);
546
547 if (info->mi_portal >= the_lnet.ln_nportals) {
548 CERROR("Invalid portal %d not in [0-%d]\n",
549 info->mi_portal, the_lnet.ln_nportals);
550 return LNET_MATCHMD_DROP;
551 }
552
553 ptl = the_lnet.ln_portals[info->mi_portal];
554 rc = lnet_ptl_match_early(ptl, msg);
555 if (rc != 0) /* matched or delayed early message */
556 return rc;
557
558 mtable = lnet_mt_of_match(info, msg);
559 lnet_res_lock(mtable->mt_cpt);
560
561 if (the_lnet.ln_shutdown) {
562 rc = LNET_MATCHMD_DROP;
563 goto out1;
564 }
565
566 rc = lnet_mt_match_md(mtable, info, msg);
567 if ((rc & LNET_MATCHMD_EXHAUSTED) != 0 && mtable->mt_enabled) {
568 lnet_ptl_lock(ptl);
569 lnet_ptl_disable_mt(ptl, mtable->mt_cpt);
570 lnet_ptl_unlock(ptl);
571 }
572
573 if ((rc & LNET_MATCHMD_FINISH) != 0) /* matched or dropping */
574 goto out1;
575
576 if (!msg->msg_rx_ready_delay)
577 goto out1;
578
579 LASSERT(lnet_ptl_is_lazy(ptl));
580 LASSERT(!msg->msg_rx_delayed);
581
582 /* NB: we don't expect "delay" can happen a lot */
583 if (lnet_ptl_is_unique(ptl) || LNET_CPT_NUMBER == 1) {
584 lnet_ptl_lock(ptl);
585
586 msg->msg_rx_delayed = 1;
587 list_add_tail(&msg->msg_list, &ptl->ptl_msg_delayed);
588
589 lnet_ptl_unlock(ptl);
590 lnet_res_unlock(mtable->mt_cpt);
591
592 } else {
593 lnet_res_unlock(mtable->mt_cpt);
594 rc = lnet_ptl_match_delay(ptl, info, msg);
595 }
596
597 if (msg->msg_rx_delayed) {
598 CDEBUG(D_NET,
599 "Delaying %s from %s ptl %d MB %#llx off %d len %d\n",
600 info->mi_opc == LNET_MD_OP_PUT ? "PUT" : "GET",
601 libcfs_id2str(info->mi_id), info->mi_portal,
602 info->mi_mbits, info->mi_roffset, info->mi_rlength);
603 }
604 goto out0;
605 out1:
606 lnet_res_unlock(mtable->mt_cpt);
607 out0:
608 /* EXHAUSTED bit is only meaningful for internal functions */
609 return rc & ~LNET_MATCHMD_EXHAUSTED;
610 }
611
612 void
lnet_ptl_detach_md(lnet_me_t * me,lnet_libmd_t * md)613 lnet_ptl_detach_md(lnet_me_t *me, lnet_libmd_t *md)
614 {
615 LASSERT(me->me_md == md && md->md_me == me);
616
617 me->me_md = NULL;
618 md->md_me = NULL;
619 }
620
621 /* called with lnet_res_lock held */
622 void
lnet_ptl_attach_md(lnet_me_t * me,lnet_libmd_t * md,struct list_head * matches,struct list_head * drops)623 lnet_ptl_attach_md(lnet_me_t *me, lnet_libmd_t *md,
624 struct list_head *matches, struct list_head *drops)
625 {
626 struct lnet_portal *ptl = the_lnet.ln_portals[me->me_portal];
627 struct lnet_match_table *mtable;
628 struct list_head *head;
629 lnet_msg_t *tmp;
630 lnet_msg_t *msg;
631 int exhausted = 0;
632 int cpt;
633
634 LASSERT(md->md_refcount == 0); /* a brand new MD */
635
636 me->me_md = md;
637 md->md_me = me;
638
639 cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
640 mtable = ptl->ptl_mtables[cpt];
641
642 if (list_empty(&ptl->ptl_msg_stealing) &&
643 list_empty(&ptl->ptl_msg_delayed) &&
644 !lnet_mt_test_exhausted(mtable, me->me_pos))
645 return;
646
647 lnet_ptl_lock(ptl);
648 head = &ptl->ptl_msg_stealing;
649 again:
650 list_for_each_entry_safe(msg, tmp, head, msg_list) {
651 struct lnet_match_info info;
652 lnet_hdr_t *hdr;
653 int rc;
654
655 LASSERT(msg->msg_rx_delayed || head == &ptl->ptl_msg_stealing);
656
657 hdr = &msg->msg_hdr;
658 info.mi_id.nid = hdr->src_nid;
659 info.mi_id.pid = hdr->src_pid;
660 info.mi_opc = LNET_MD_OP_PUT;
661 info.mi_portal = hdr->msg.put.ptl_index;
662 info.mi_rlength = hdr->payload_length;
663 info.mi_roffset = hdr->msg.put.offset;
664 info.mi_mbits = hdr->msg.put.match_bits;
665
666 rc = lnet_try_match_md(md, &info, msg);
667
668 exhausted = (rc & LNET_MATCHMD_EXHAUSTED) != 0;
669 if ((rc & LNET_MATCHMD_NONE) != 0) {
670 if (exhausted)
671 break;
672 continue;
673 }
674
675 /* Hurrah! This _is_ a match */
676 LASSERT((rc & LNET_MATCHMD_FINISH) != 0);
677 list_del_init(&msg->msg_list);
678
679 if (head == &ptl->ptl_msg_stealing) {
680 if (exhausted)
681 break;
682 /* stealing thread will handle the message */
683 continue;
684 }
685
686 if ((rc & LNET_MATCHMD_OK) != 0) {
687 list_add_tail(&msg->msg_list, matches);
688
689 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d match %llu offset %d length %d.\n",
690 libcfs_id2str(info.mi_id),
691 info.mi_portal, info.mi_mbits,
692 info.mi_roffset, info.mi_rlength);
693 } else {
694 list_add_tail(&msg->msg_list, drops);
695 }
696
697 if (exhausted)
698 break;
699 }
700
701 if (!exhausted && head == &ptl->ptl_msg_stealing) {
702 head = &ptl->ptl_msg_delayed;
703 goto again;
704 }
705
706 if (lnet_ptl_is_wildcard(ptl) && !exhausted) {
707 lnet_mt_set_exhausted(mtable, me->me_pos, 0);
708 if (!mtable->mt_enabled)
709 lnet_ptl_enable_mt(ptl, cpt);
710 }
711
712 lnet_ptl_unlock(ptl);
713 }
714
715 static void
lnet_ptl_cleanup(struct lnet_portal * ptl)716 lnet_ptl_cleanup(struct lnet_portal *ptl)
717 {
718 struct lnet_match_table *mtable;
719 int i;
720
721 if (ptl->ptl_mtables == NULL) /* uninitialized portal */
722 return;
723
724 LASSERT(list_empty(&ptl->ptl_msg_delayed));
725 LASSERT(list_empty(&ptl->ptl_msg_stealing));
726 cfs_percpt_for_each(mtable, i, ptl->ptl_mtables) {
727 struct list_head *mhash;
728 lnet_me_t *me;
729 int j;
730
731 if (mtable->mt_mhash == NULL) /* uninitialized match-table */
732 continue;
733
734 mhash = mtable->mt_mhash;
735 /* cleanup ME */
736 for (j = 0; j < LNET_MT_HASH_SIZE + 1; j++) {
737 while (!list_empty(&mhash[j])) {
738 me = list_entry(mhash[j].next,
739 lnet_me_t, me_list);
740 CERROR("Active ME %p on exit\n", me);
741 list_del(&me->me_list);
742 lnet_me_free(me);
743 }
744 }
745 /* the extra entry is for MEs with ignore bits */
746 LIBCFS_FREE(mhash, sizeof(*mhash) * (LNET_MT_HASH_SIZE + 1));
747 }
748
749 cfs_percpt_free(ptl->ptl_mtables);
750 ptl->ptl_mtables = NULL;
751 }
752
753 static int
lnet_ptl_setup(struct lnet_portal * ptl,int index)754 lnet_ptl_setup(struct lnet_portal *ptl, int index)
755 {
756 struct lnet_match_table *mtable;
757 struct list_head *mhash;
758 int i;
759 int j;
760
761 ptl->ptl_mtables = cfs_percpt_alloc(lnet_cpt_table(),
762 sizeof(struct lnet_match_table));
763 if (ptl->ptl_mtables == NULL) {
764 CERROR("Failed to create match table for portal %d\n", index);
765 return -ENOMEM;
766 }
767
768 ptl->ptl_index = index;
769 INIT_LIST_HEAD(&ptl->ptl_msg_delayed);
770 INIT_LIST_HEAD(&ptl->ptl_msg_stealing);
771 spin_lock_init(&ptl->ptl_lock);
772 cfs_percpt_for_each(mtable, i, ptl->ptl_mtables) {
773 /* the extra entry is for MEs with ignore bits */
774 LIBCFS_CPT_ALLOC(mhash, lnet_cpt_table(), i,
775 sizeof(*mhash) * (LNET_MT_HASH_SIZE + 1));
776 if (mhash == NULL) {
777 CERROR("Failed to create match hash for portal %d\n",
778 index);
779 goto failed;
780 }
781
782 memset(&mtable->mt_exhausted[0], -1,
783 sizeof(mtable->mt_exhausted[0]) *
784 LNET_MT_EXHAUSTED_BMAP);
785 mtable->mt_mhash = mhash;
786 for (j = 0; j < LNET_MT_HASH_SIZE + 1; j++)
787 INIT_LIST_HEAD(&mhash[j]);
788
789 mtable->mt_portal = index;
790 mtable->mt_cpt = i;
791 }
792
793 return 0;
794 failed:
795 lnet_ptl_cleanup(ptl);
796 return -ENOMEM;
797 }
798
799 void
lnet_portals_destroy(void)800 lnet_portals_destroy(void)
801 {
802 int i;
803
804 if (the_lnet.ln_portals == NULL)
805 return;
806
807 for (i = 0; i < the_lnet.ln_nportals; i++)
808 lnet_ptl_cleanup(the_lnet.ln_portals[i]);
809
810 cfs_array_free(the_lnet.ln_portals);
811 the_lnet.ln_portals = NULL;
812 }
813
814 int
lnet_portals_create(void)815 lnet_portals_create(void)
816 {
817 int size;
818 int i;
819
820 size = offsetof(struct lnet_portal, ptl_mt_maps[LNET_CPT_NUMBER]);
821
822 the_lnet.ln_nportals = MAX_PORTALS;
823 the_lnet.ln_portals = cfs_array_alloc(the_lnet.ln_nportals, size);
824 if (the_lnet.ln_portals == NULL) {
825 CERROR("Failed to allocate portals table\n");
826 return -ENOMEM;
827 }
828
829 for (i = 0; i < the_lnet.ln_nportals; i++) {
830 if (lnet_ptl_setup(the_lnet.ln_portals[i], i)) {
831 lnet_portals_destroy();
832 return -ENOMEM;
833 }
834 }
835
836 return 0;
837 }
838
839 /**
840 * Turn on the lazy portal attribute. Use with caution!
841 *
842 * This portal attribute only affects incoming PUT requests to the portal,
843 * and is off by default. By default, if there's no matching MD for an
844 * incoming PUT request, it is simply dropped. With the lazy attribute on,
845 * such requests are queued indefinitely until either a matching MD is
846 * posted to the portal or the lazy attribute is turned off.
847 *
848 * It would prevent dropped requests, however it should be regarded as the
849 * last line of defense - i.e. users must keep a close watch on active
850 * buffers on a lazy portal and once it becomes too low post more buffers as
851 * soon as possible. This is because delayed requests usually have detrimental
852 * effects on underlying network connections. A few delayed requests often
853 * suffice to bring an underlying connection to a complete halt, due to flow
854 * control mechanisms.
855 *
856 * There's also a DOS attack risk. If users don't post match-all MDs on a
857 * lazy portal, a malicious peer can easily stop a service by sending some
858 * PUT requests with match bits that won't match any MD. A routed server is
859 * especially vulnerable since the connections to its neighbor routers are
860 * shared among all clients.
861 *
862 * \param portal Index of the portal to enable the lazy attribute on.
863 *
864 * \retval 0 On success.
865 * \retval -EINVAL If \a portal is not a valid index.
866 */
867 int
LNetSetLazyPortal(int portal)868 LNetSetLazyPortal(int portal)
869 {
870 struct lnet_portal *ptl;
871
872 if (portal < 0 || portal >= the_lnet.ln_nportals)
873 return -EINVAL;
874
875 CDEBUG(D_NET, "Setting portal %d lazy\n", portal);
876 ptl = the_lnet.ln_portals[portal];
877
878 lnet_res_lock(LNET_LOCK_EX);
879 lnet_ptl_lock(ptl);
880
881 lnet_ptl_setopt(ptl, LNET_PTL_LAZY);
882
883 lnet_ptl_unlock(ptl);
884 lnet_res_unlock(LNET_LOCK_EX);
885
886 return 0;
887 }
888 EXPORT_SYMBOL(LNetSetLazyPortal);
889
890 /**
891 * Turn off the lazy portal attribute. Delayed requests on the portal,
892 * if any, will be all dropped when this function returns.
893 *
894 * \param portal Index of the portal to disable the lazy attribute on.
895 *
896 * \retval 0 On success.
897 * \retval -EINVAL If \a portal is not a valid index.
898 */
899 int
LNetClearLazyPortal(int portal)900 LNetClearLazyPortal(int portal)
901 {
902 struct lnet_portal *ptl;
903 LIST_HEAD (zombies);
904
905 if (portal < 0 || portal >= the_lnet.ln_nportals)
906 return -EINVAL;
907
908 ptl = the_lnet.ln_portals[portal];
909
910 lnet_res_lock(LNET_LOCK_EX);
911 lnet_ptl_lock(ptl);
912
913 if (!lnet_ptl_is_lazy(ptl)) {
914 lnet_ptl_unlock(ptl);
915 lnet_res_unlock(LNET_LOCK_EX);
916 return 0;
917 }
918
919 if (the_lnet.ln_shutdown)
920 CWARN("Active lazy portal %d on exit\n", portal);
921 else
922 CDEBUG(D_NET, "clearing portal %d lazy\n", portal);
923
924 /* grab all the blocked messages atomically */
925 list_splice_init(&ptl->ptl_msg_delayed, &zombies);
926
927 lnet_ptl_unsetopt(ptl, LNET_PTL_LAZY);
928
929 lnet_ptl_unlock(ptl);
930 lnet_res_unlock(LNET_LOCK_EX);
931
932 lnet_drop_delayed_msg_list(&zombies, "Clearing lazy portal attr");
933
934 return 0;
935 }
936 EXPORT_SYMBOL(LNetClearLazyPortal);
937