4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
24 * Copyright (c) 2012, Intel Corporation.
27 * This file is part of Lustre, http://www.lustre.org/
28 * Lustre is a trademark of Sun Microsystems, Inc.
32 * portal & match routines
34 * Author: liang@whamcloud.com
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/lnet/lib-lnet.h>
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");
47 lnet_ptl_match_type(unsigned int index
, lnet_process_id_t match_id
,
48 __u64 mbits
, __u64 ignore_bits
)
50 struct lnet_portal
*ptl
= the_lnet
.ln_portals
[index
];
53 unique
= ignore_bits
== 0 &&
54 match_id
.nid
!= LNET_NID_ANY
&&
55 match_id
.pid
!= LNET_PID_ANY
;
57 LASSERT(!lnet_ptl_is_unique(ptl
) || !lnet_ptl_is_wildcard(ptl
));
59 /* prefer to check w/o any lock */
60 if (likely(lnet_ptl_is_unique(ptl
) || lnet_ptl_is_wildcard(ptl
)))
63 /* unset, new portal */
65 /* check again with lock */
66 if (unlikely(lnet_ptl_is_unique(ptl
) || lnet_ptl_is_wildcard(ptl
))) {
73 lnet_ptl_setopt(ptl
, LNET_PTL_MATCH_UNIQUE
);
75 lnet_ptl_setopt(ptl
, LNET_PTL_MATCH_WILDCARD
);
82 if ((lnet_ptl_is_unique(ptl
) && !unique
) ||
83 (lnet_ptl_is_wildcard(ptl
) && unique
))
89 lnet_ptl_enable_mt(struct lnet_portal
*ptl
, int cpt
)
91 struct lnet_match_table
*mtable
= ptl
->ptl_mtables
[cpt
];
94 /* with hold of both lnet_res_lock(cpt) and lnet_ptl_lock */
95 LASSERT(lnet_ptl_is_wildcard(ptl
));
97 mtable
->mt_enabled
= 1;
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
)
106 ptl
->ptl_mt_maps
[i
+ 1] = ptl
->ptl_mt_maps
[i
];
107 ptl
->ptl_mt_maps
[i
] = cpt
;
114 lnet_ptl_disable_mt(struct lnet_portal
*ptl
, int cpt
)
116 struct lnet_match_table
*mtable
= ptl
->ptl_mtables
[cpt
];
119 /* with hold of both lnet_res_lock(cpt) and lnet_ptl_lock */
120 LASSERT(lnet_ptl_is_wildcard(ptl
));
122 if (LNET_CPT_NUMBER
== 1)
123 return; /* never disable the only match-table */
125 mtable
->mt_enabled
= 0;
127 LASSERT(ptl
->ptl_mt_nmaps
> 0 &&
128 ptl
->ptl_mt_nmaps
<= LNET_CPT_NUMBER
);
130 /* remove it from mt_maps */
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];
139 lnet_try_match_md(lnet_libmd_t
*md
,
140 struct lnet_match_info
*info
, struct lnet_msg
*msg
)
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 */
145 unsigned int mlength
;
146 lnet_me_t
*me
= md
->md_me
;
149 if (lnet_md_exhausted(md
))
150 return LNET_MATCHMD_NONE
| LNET_MATCHMD_EXHAUSTED
;
152 /* mismatched MD op */
153 if ((md
->md_options
& info
->mi_opc
) == 0)
154 return LNET_MATCHMD_NONE
;
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
;
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
;
165 /* mismatched ME matchbits? */
166 if (((me
->me_match_bits
^ info
->mi_mbits
) & ~me
->me_ignore_bits
) != 0)
167 return LNET_MATCHMD_NONE
;
169 /* Hurrah! This _is_ a match; check it out... */
171 if ((md
->md_options
& LNET_MD_MANAGE_REMOTE
) == 0)
172 offset
= md
->md_offset
;
174 offset
= info
->mi_roffset
;
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
);
180 mlength
= md
->md_length
- offset
;
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 "LPU64
188 " length %d too big: %d left, %d allowed\n",
189 libcfs_id2str(info
->mi_id
), info
->mi_mbits
,
190 info
->mi_rlength
, md
->md_length
- offset
, mlength
);
192 return LNET_MATCHMD_DROP
;
195 /* Commit to this ME/MD */
196 CDEBUG(D_NET
, "Incoming %s index %x from %s of "
197 "length %d/%d into md "LPX64
" [%d] + %d\n",
198 (info
->mi_opc
== LNET_MD_OP_PUT
) ? "put" : "get",
199 info
->mi_portal
, libcfs_id2str(info
->mi_id
), mlength
,
200 info
->mi_rlength
, md
->md_lh
.lh_cookie
, md
->md_niov
, offset
);
202 lnet_msg_attach_md(msg
, md
, offset
, mlength
);
203 md
->md_offset
= offset
+ mlength
;
205 if (!lnet_md_exhausted(md
))
206 return LNET_MATCHMD_OK
;
208 /* Auto-unlink NOW, so the ME gets unlinked if required.
209 * We bumped md->md_refcount above so the MD just gets flagged
210 * for unlink when it is finalized. */
211 if ((md
->md_flags
& LNET_MD_FLAG_AUTO_UNLINK
) != 0)
214 return LNET_MATCHMD_OK
| LNET_MATCHMD_EXHAUSTED
;
217 static struct lnet_match_table
*
218 lnet_match2mt(struct lnet_portal
*ptl
, lnet_process_id_t id
, __u64 mbits
)
220 if (LNET_CPT_NUMBER
== 1)
221 return ptl
->ptl_mtables
[0]; /* the only one */
223 /* if it's a unique portal, return match-table hashed by NID */
224 return lnet_ptl_is_unique(ptl
) ?
225 ptl
->ptl_mtables
[lnet_cpt_of_nid(id
.nid
)] : NULL
;
228 struct lnet_match_table
*
229 lnet_mt_of_attach(unsigned int index
, lnet_process_id_t id
,
230 __u64 mbits
, __u64 ignore_bits
, lnet_ins_pos_t pos
)
232 struct lnet_portal
*ptl
;
233 struct lnet_match_table
*mtable
;
235 /* NB: called w/o lock */
236 LASSERT(index
< the_lnet
.ln_nportals
);
238 if (!lnet_ptl_match_type(index
, id
, mbits
, ignore_bits
))
241 ptl
= the_lnet
.ln_portals
[index
];
243 mtable
= lnet_match2mt(ptl
, id
, mbits
);
244 if (mtable
!= NULL
) /* unique portal or only one match-table */
247 /* it's a wildcard portal */
251 case LNET_INS_BEFORE
:
253 /* posted by no affinity thread, always hash to specific
254 * match-table to avoid buffer stealing which is heavy */
255 return ptl
->ptl_mtables
[ptl
->ptl_index
% LNET_CPT_NUMBER
];
257 /* posted by cpu-affinity thread */
258 return ptl
->ptl_mtables
[lnet_cpt_current()];
262 static struct lnet_match_table
*
263 lnet_mt_of_match(struct lnet_match_info
*info
, struct lnet_msg
*msg
)
265 struct lnet_match_table
*mtable
;
266 struct lnet_portal
*ptl
;
272 /* NB: called w/o lock */
273 LASSERT(info
->mi_portal
< the_lnet
.ln_nportals
);
274 ptl
= the_lnet
.ln_portals
[info
->mi_portal
];
276 LASSERT(lnet_ptl_is_wildcard(ptl
) || lnet_ptl_is_unique(ptl
));
278 mtable
= lnet_match2mt(ptl
, info
->mi_id
, info
->mi_mbits
);
282 /* it's a wildcard portal */
283 routed
= LNET_NIDNET(msg
->msg_hdr
.src_nid
) !=
284 LNET_NIDNET(msg
->msg_hdr
.dest_nid
);
286 if (portal_rotor
== LNET_PTL_ROTOR_OFF
||
287 (portal_rotor
!= LNET_PTL_ROTOR_ON
&& !routed
)) {
288 cpt
= lnet_cpt_current();
289 if (ptl
->ptl_mtables
[cpt
]->mt_enabled
)
290 return ptl
->ptl_mtables
[cpt
];
293 rotor
= ptl
->ptl_rotor
++; /* get round-robin factor */
294 if (portal_rotor
== LNET_PTL_ROTOR_HASH_RT
&& routed
)
295 cpt
= lnet_cpt_of_nid(msg
->msg_hdr
.src_nid
);
297 cpt
= rotor
% LNET_CPT_NUMBER
;
299 if (!ptl
->ptl_mtables
[cpt
]->mt_enabled
) {
300 /* is there any active entry for this portal? */
301 nmaps
= ptl
->ptl_mt_nmaps
;
302 /* map to an active mtable to avoid heavy "stealing" */
304 /* NB: there is possibility that ptl_mt_maps is being
305 * changed because we are not under protection of
306 * lnet_ptl_lock, but it shouldn't hurt anything */
307 cpt
= ptl
->ptl_mt_maps
[rotor
% nmaps
];
311 return ptl
->ptl_mtables
[cpt
];
315 lnet_mt_test_exhausted(struct lnet_match_table
*mtable
, int pos
)
320 if (!lnet_ptl_is_wildcard(the_lnet
.ln_portals
[mtable
->mt_portal
]))
323 if (pos
< 0) { /* check all bits */
324 for (i
= 0; i
< LNET_MT_EXHAUSTED_BMAP
; i
++) {
325 if (mtable
->mt_exhausted
[i
] != (__u64
)(-1))
331 LASSERT(pos
<= LNET_MT_HASH_IGNORE
);
332 /* mtable::mt_mhash[pos] is marked as exhausted or not */
333 bmap
= &mtable
->mt_exhausted
[pos
>> LNET_MT_BITS_U64
];
334 pos
&= (1 << LNET_MT_BITS_U64
) - 1;
336 return ((*bmap
) & (1ULL << pos
)) != 0;
340 lnet_mt_set_exhausted(struct lnet_match_table
*mtable
, int pos
, int exhausted
)
344 LASSERT(lnet_ptl_is_wildcard(the_lnet
.ln_portals
[mtable
->mt_portal
]));
345 LASSERT(pos
<= LNET_MT_HASH_IGNORE
);
347 /* set mtable::mt_mhash[pos] as exhausted/non-exhausted */
348 bmap
= &mtable
->mt_exhausted
[pos
>> LNET_MT_BITS_U64
];
349 pos
&= (1 << LNET_MT_BITS_U64
) - 1;
352 *bmap
&= ~(1ULL << pos
);
354 *bmap
|= 1ULL << pos
;
358 lnet_mt_match_head(struct lnet_match_table
*mtable
,
359 lnet_process_id_t id
, __u64 mbits
)
361 struct lnet_portal
*ptl
= the_lnet
.ln_portals
[mtable
->mt_portal
];
363 if (lnet_ptl_is_wildcard(ptl
)) {
364 return &mtable
->mt_mhash
[mbits
& LNET_MT_HASH_MASK
];
366 unsigned long hash
= mbits
+ id
.nid
+ id
.pid
;
368 LASSERT(lnet_ptl_is_unique(ptl
));
369 hash
= hash_long(hash
, LNET_MT_HASH_BITS
);
370 return &mtable
->mt_mhash
[hash
];
375 lnet_mt_match_md(struct lnet_match_table
*mtable
,
376 struct lnet_match_info
*info
, struct lnet_msg
*msg
)
378 struct list_head
*head
;
384 /* any ME with ignore bits? */
385 if (!list_empty(&mtable
->mt_mhash
[LNET_MT_HASH_IGNORE
]))
386 head
= &mtable
->mt_mhash
[LNET_MT_HASH_IGNORE
];
388 head
= lnet_mt_match_head(mtable
, info
->mi_id
, info
->mi_mbits
);
390 /* NB: only wildcard portal needs to return LNET_MATCHMD_EXHAUSTED */
391 if (lnet_ptl_is_wildcard(the_lnet
.ln_portals
[mtable
->mt_portal
]))
392 exhausted
= LNET_MATCHMD_EXHAUSTED
;
394 list_for_each_entry_safe(me
, tmp
, head
, me_list
) {
395 /* ME attached but MD not attached yet */
396 if (me
->me_md
== NULL
)
399 LASSERT(me
== me
->me_md
->md_me
);
401 rc
= lnet_try_match_md(me
->me_md
, info
, msg
);
402 if ((rc
& LNET_MATCHMD_EXHAUSTED
) == 0)
403 exhausted
= 0; /* mlist is not empty */
405 if ((rc
& LNET_MATCHMD_FINISH
) != 0) {
406 /* don't return EXHAUSTED bit because we don't know
407 * whether the mlist is empty or not */
408 return rc
& ~LNET_MATCHMD_EXHAUSTED
;
412 if (exhausted
== LNET_MATCHMD_EXHAUSTED
) { /* @head is exhausted */
413 lnet_mt_set_exhausted(mtable
, head
- mtable
->mt_mhash
, 1);
414 if (!lnet_mt_test_exhausted(mtable
, -1))
418 if (exhausted
== 0 && head
== &mtable
->mt_mhash
[LNET_MT_HASH_IGNORE
]) {
419 head
= lnet_mt_match_head(mtable
, info
->mi_id
, info
->mi_mbits
);
420 goto again
; /* re-check MEs w/o ignore-bits */
423 if (info
->mi_opc
== LNET_MD_OP_GET
||
424 !lnet_ptl_is_lazy(the_lnet
.ln_portals
[info
->mi_portal
]))
425 return LNET_MATCHMD_DROP
| exhausted
;
427 return LNET_MATCHMD_NONE
| exhausted
;
431 lnet_ptl_match_early(struct lnet_portal
*ptl
, struct lnet_msg
*msg
)
435 /* message arrived before any buffer posting on this portal,
436 * simply delay or drop this message */
437 if (likely(lnet_ptl_is_wildcard(ptl
) || lnet_ptl_is_unique(ptl
)))
441 /* check it again with hold of lock */
442 if (lnet_ptl_is_wildcard(ptl
) || lnet_ptl_is_unique(ptl
)) {
443 lnet_ptl_unlock(ptl
);
447 if (lnet_ptl_is_lazy(ptl
)) {
448 if (msg
->msg_rx_ready_delay
) {
449 msg
->msg_rx_delayed
= 1;
450 list_add_tail(&msg
->msg_list
,
451 &ptl
->ptl_msg_delayed
);
453 rc
= LNET_MATCHMD_NONE
;
455 rc
= LNET_MATCHMD_DROP
;
458 lnet_ptl_unlock(ptl
);
463 lnet_ptl_match_delay(struct lnet_portal
*ptl
,
464 struct lnet_match_info
*info
, struct lnet_msg
*msg
)
466 int first
= ptl
->ptl_mt_maps
[0]; /* read w/o lock */
470 /* steal buffer from other CPTs, and delay it if nothing to steal,
471 * this function is more expensive than a regular match, but we
472 * don't expect it can happen a lot */
473 LASSERT(lnet_ptl_is_wildcard(ptl
));
475 for (i
= 0; i
< LNET_CPT_NUMBER
; i
++) {
476 struct lnet_match_table
*mtable
;
479 cpt
= (first
+ i
) % LNET_CPT_NUMBER
;
480 mtable
= ptl
->ptl_mtables
[cpt
];
481 if (i
!= 0 && i
!= LNET_CPT_NUMBER
- 1 && !mtable
->mt_enabled
)
487 if (i
== 0) { /* the first try, attach on stealing list */
488 list_add_tail(&msg
->msg_list
,
489 &ptl
->ptl_msg_stealing
);
492 if (!list_empty(&msg
->msg_list
)) { /* on stealing list */
493 rc
= lnet_mt_match_md(mtable
, info
, msg
);
495 if ((rc
& LNET_MATCHMD_EXHAUSTED
) != 0 &&
497 lnet_ptl_disable_mt(ptl
, cpt
);
499 if ((rc
& LNET_MATCHMD_FINISH
) != 0)
500 list_del_init(&msg
->msg_list
);
503 /* could be matched by lnet_ptl_attach_md()
504 * which is called by another thread */
505 rc
= msg
->msg_md
== NULL
?
506 LNET_MATCHMD_DROP
: LNET_MATCHMD_OK
;
509 if (!list_empty(&msg
->msg_list
) && /* not matched yet */
510 (i
== LNET_CPT_NUMBER
- 1 || /* the last CPT */
511 ptl
->ptl_mt_nmaps
== 0 || /* no active CPT */
512 (ptl
->ptl_mt_nmaps
== 1 && /* the only active CPT */
513 ptl
->ptl_mt_maps
[0] == cpt
))) {
514 /* nothing to steal, delay or drop */
515 list_del_init(&msg
->msg_list
);
517 if (lnet_ptl_is_lazy(ptl
)) {
518 msg
->msg_rx_delayed
= 1;
519 list_add_tail(&msg
->msg_list
,
520 &ptl
->ptl_msg_delayed
);
521 rc
= LNET_MATCHMD_NONE
;
523 rc
= LNET_MATCHMD_DROP
;
527 lnet_ptl_unlock(ptl
);
528 lnet_res_unlock(cpt
);
530 if ((rc
& LNET_MATCHMD_FINISH
) != 0 || msg
->msg_rx_delayed
)
538 lnet_ptl_match_md(struct lnet_match_info
*info
, struct lnet_msg
*msg
)
540 struct lnet_match_table
*mtable
;
541 struct lnet_portal
*ptl
;
544 CDEBUG(D_NET
, "Request from %s of length %d into portal %d "
545 "MB="LPX64
"\n", libcfs_id2str(info
->mi_id
),
546 info
->mi_rlength
, info
->mi_portal
, info
->mi_mbits
);
548 if (info
->mi_portal
>= the_lnet
.ln_nportals
) {
549 CERROR("Invalid portal %d not in [0-%d]\n",
550 info
->mi_portal
, the_lnet
.ln_nportals
);
551 return LNET_MATCHMD_DROP
;
554 ptl
= the_lnet
.ln_portals
[info
->mi_portal
];
555 rc
= lnet_ptl_match_early(ptl
, msg
);
556 if (rc
!= 0) /* matched or delayed early message */
559 mtable
= lnet_mt_of_match(info
, msg
);
560 lnet_res_lock(mtable
->mt_cpt
);
562 if (the_lnet
.ln_shutdown
) {
563 rc
= LNET_MATCHMD_DROP
;
567 rc
= lnet_mt_match_md(mtable
, info
, msg
);
568 if ((rc
& LNET_MATCHMD_EXHAUSTED
) != 0 && mtable
->mt_enabled
) {
570 lnet_ptl_disable_mt(ptl
, mtable
->mt_cpt
);
571 lnet_ptl_unlock(ptl
);
574 if ((rc
& LNET_MATCHMD_FINISH
) != 0) /* matched or dropping */
577 if (!msg
->msg_rx_ready_delay
)
580 LASSERT(lnet_ptl_is_lazy(ptl
));
581 LASSERT(!msg
->msg_rx_delayed
);
583 /* NB: we don't expect "delay" can happen a lot */
584 if (lnet_ptl_is_unique(ptl
) || LNET_CPT_NUMBER
== 1) {
587 msg
->msg_rx_delayed
= 1;
588 list_add_tail(&msg
->msg_list
, &ptl
->ptl_msg_delayed
);
590 lnet_ptl_unlock(ptl
);
591 lnet_res_unlock(mtable
->mt_cpt
);
594 lnet_res_unlock(mtable
->mt_cpt
);
595 rc
= lnet_ptl_match_delay(ptl
, info
, msg
);
598 if (msg
->msg_rx_delayed
) {
600 "Delaying %s from %s ptl %d MB "LPX64
" off %d len %d\n",
601 info
->mi_opc
== LNET_MD_OP_PUT
? "PUT" : "GET",
602 libcfs_id2str(info
->mi_id
), info
->mi_portal
,
603 info
->mi_mbits
, info
->mi_roffset
, info
->mi_rlength
);
607 lnet_res_unlock(mtable
->mt_cpt
);
609 /* EXHAUSTED bit is only meaningful for internal functions */
610 return rc
& ~LNET_MATCHMD_EXHAUSTED
;
614 lnet_ptl_detach_md(lnet_me_t
*me
, lnet_libmd_t
*md
)
616 LASSERT(me
->me_md
== md
&& md
->md_me
== me
);
622 /* called with lnet_res_lock held */
624 lnet_ptl_attach_md(lnet_me_t
*me
, lnet_libmd_t
*md
,
625 struct list_head
*matches
, struct list_head
*drops
)
627 struct lnet_portal
*ptl
= the_lnet
.ln_portals
[me
->me_portal
];
628 struct lnet_match_table
*mtable
;
629 struct list_head
*head
;
635 LASSERT(md
->md_refcount
== 0); /* a brand new MD */
640 cpt
= lnet_cpt_of_cookie(md
->md_lh
.lh_cookie
);
641 mtable
= ptl
->ptl_mtables
[cpt
];
643 if (list_empty(&ptl
->ptl_msg_stealing
) &&
644 list_empty(&ptl
->ptl_msg_delayed
) &&
645 !lnet_mt_test_exhausted(mtable
, me
->me_pos
))
649 head
= &ptl
->ptl_msg_stealing
;
651 list_for_each_entry_safe(msg
, tmp
, head
, msg_list
) {
652 struct lnet_match_info info
;
656 LASSERT(msg
->msg_rx_delayed
|| head
== &ptl
->ptl_msg_stealing
);
659 info
.mi_id
.nid
= hdr
->src_nid
;
660 info
.mi_id
.pid
= hdr
->src_pid
;
661 info
.mi_opc
= LNET_MD_OP_PUT
;
662 info
.mi_portal
= hdr
->msg
.put
.ptl_index
;
663 info
.mi_rlength
= hdr
->payload_length
;
664 info
.mi_roffset
= hdr
->msg
.put
.offset
;
665 info
.mi_mbits
= hdr
->msg
.put
.match_bits
;
667 rc
= lnet_try_match_md(md
, &info
, msg
);
669 exhausted
= (rc
& LNET_MATCHMD_EXHAUSTED
) != 0;
670 if ((rc
& LNET_MATCHMD_NONE
) != 0) {
676 /* Hurrah! This _is_ a match */
677 LASSERT((rc
& LNET_MATCHMD_FINISH
) != 0);
678 list_del_init(&msg
->msg_list
);
680 if (head
== &ptl
->ptl_msg_stealing
) {
683 /* stealing thread will handle the message */
687 if ((rc
& LNET_MATCHMD_OK
) != 0) {
688 list_add_tail(&msg
->msg_list
, matches
);
690 CDEBUG(D_NET
, "Resuming delayed PUT from %s portal %d "
691 "match "LPU64
" offset %d length %d.\n",
692 libcfs_id2str(info
.mi_id
),
693 info
.mi_portal
, info
.mi_mbits
,
694 info
.mi_roffset
, info
.mi_rlength
);
696 list_add_tail(&msg
->msg_list
, drops
);
703 if (!exhausted
&& head
== &ptl
->ptl_msg_stealing
) {
704 head
= &ptl
->ptl_msg_delayed
;
708 if (lnet_ptl_is_wildcard(ptl
) && !exhausted
) {
709 lnet_mt_set_exhausted(mtable
, me
->me_pos
, 0);
710 if (!mtable
->mt_enabled
)
711 lnet_ptl_enable_mt(ptl
, cpt
);
714 lnet_ptl_unlock(ptl
);
718 lnet_ptl_cleanup(struct lnet_portal
*ptl
)
720 struct lnet_match_table
*mtable
;
723 if (ptl
->ptl_mtables
== NULL
) /* uninitialized portal */
726 LASSERT(list_empty(&ptl
->ptl_msg_delayed
));
727 LASSERT(list_empty(&ptl
->ptl_msg_stealing
));
728 cfs_percpt_for_each(mtable
, i
, ptl
->ptl_mtables
) {
729 struct list_head
*mhash
;
733 if (mtable
->mt_mhash
== NULL
) /* uninitialized match-table */
736 mhash
= mtable
->mt_mhash
;
738 for (j
= 0; j
< LNET_MT_HASH_SIZE
+ 1; j
++) {
739 while (!list_empty(&mhash
[j
])) {
740 me
= list_entry(mhash
[j
].next
,
742 CERROR("Active ME %p on exit\n", me
);
743 list_del(&me
->me_list
);
747 /* the extra entry is for MEs with ignore bits */
748 LIBCFS_FREE(mhash
, sizeof(*mhash
) * (LNET_MT_HASH_SIZE
+ 1));
751 cfs_percpt_free(ptl
->ptl_mtables
);
752 ptl
->ptl_mtables
= NULL
;
756 lnet_ptl_setup(struct lnet_portal
*ptl
, int index
)
758 struct lnet_match_table
*mtable
;
759 struct list_head
*mhash
;
763 ptl
->ptl_mtables
= cfs_percpt_alloc(lnet_cpt_table(),
764 sizeof(struct lnet_match_table
));
765 if (ptl
->ptl_mtables
== NULL
) {
766 CERROR("Failed to create match table for portal %d\n", index
);
770 ptl
->ptl_index
= index
;
771 INIT_LIST_HEAD(&ptl
->ptl_msg_delayed
);
772 INIT_LIST_HEAD(&ptl
->ptl_msg_stealing
);
773 spin_lock_init(&ptl
->ptl_lock
);
774 cfs_percpt_for_each(mtable
, i
, ptl
->ptl_mtables
) {
775 /* the extra entry is for MEs with ignore bits */
776 LIBCFS_CPT_ALLOC(mhash
, lnet_cpt_table(), i
,
777 sizeof(*mhash
) * (LNET_MT_HASH_SIZE
+ 1));
779 CERROR("Failed to create match hash for portal %d\n",
784 memset(&mtable
->mt_exhausted
[0], -1,
785 sizeof(mtable
->mt_exhausted
[0]) *
786 LNET_MT_EXHAUSTED_BMAP
);
787 mtable
->mt_mhash
= mhash
;
788 for (j
= 0; j
< LNET_MT_HASH_SIZE
+ 1; j
++)
789 INIT_LIST_HEAD(&mhash
[j
]);
791 mtable
->mt_portal
= index
;
797 lnet_ptl_cleanup(ptl
);
802 lnet_portals_destroy(void)
806 if (the_lnet
.ln_portals
== NULL
)
809 for (i
= 0; i
< the_lnet
.ln_nportals
; i
++)
810 lnet_ptl_cleanup(the_lnet
.ln_portals
[i
]);
812 cfs_array_free(the_lnet
.ln_portals
);
813 the_lnet
.ln_portals
= NULL
;
817 lnet_portals_create(void)
822 size
= offsetof(struct lnet_portal
, ptl_mt_maps
[LNET_CPT_NUMBER
]);
824 the_lnet
.ln_nportals
= MAX_PORTALS
;
825 the_lnet
.ln_portals
= cfs_array_alloc(the_lnet
.ln_nportals
, size
);
826 if (the_lnet
.ln_portals
== NULL
) {
827 CERROR("Failed to allocate portals table\n");
831 for (i
= 0; i
< the_lnet
.ln_nportals
; i
++) {
832 if (lnet_ptl_setup(the_lnet
.ln_portals
[i
], i
)) {
833 lnet_portals_destroy();
842 * Turn on the lazy portal attribute. Use with caution!
844 * This portal attribute only affects incoming PUT requests to the portal,
845 * and is off by default. By default, if there's no matching MD for an
846 * incoming PUT request, it is simply dropped. With the lazy attribute on,
847 * such requests are queued indefinitely until either a matching MD is
848 * posted to the portal or the lazy attribute is turned off.
850 * It would prevent dropped requests, however it should be regarded as the
851 * last line of defense - i.e. users must keep a close watch on active
852 * buffers on a lazy portal and once it becomes too low post more buffers as
853 * soon as possible. This is because delayed requests usually have detrimental
854 * effects on underlying network connections. A few delayed requests often
855 * suffice to bring an underlying connection to a complete halt, due to flow
856 * control mechanisms.
858 * There's also a DOS attack risk. If users don't post match-all MDs on a
859 * lazy portal, a malicious peer can easily stop a service by sending some
860 * PUT requests with match bits that won't match any MD. A routed server is
861 * especially vulnerable since the connections to its neighbor routers are
862 * shared among all clients.
864 * \param portal Index of the portal to enable the lazy attribute on.
866 * \retval 0 On success.
867 * \retval -EINVAL If \a portal is not a valid index.
870 LNetSetLazyPortal(int portal
)
872 struct lnet_portal
*ptl
;
874 if (portal
< 0 || portal
>= the_lnet
.ln_nportals
)
877 CDEBUG(D_NET
, "Setting portal %d lazy\n", portal
);
878 ptl
= the_lnet
.ln_portals
[portal
];
880 lnet_res_lock(LNET_LOCK_EX
);
883 lnet_ptl_setopt(ptl
, LNET_PTL_LAZY
);
885 lnet_ptl_unlock(ptl
);
886 lnet_res_unlock(LNET_LOCK_EX
);
890 EXPORT_SYMBOL(LNetSetLazyPortal
);
893 * Turn off the lazy portal attribute. Delayed requests on the portal,
894 * if any, will be all dropped when this function returns.
896 * \param portal Index of the portal to disable the lazy attribute on.
898 * \retval 0 On success.
899 * \retval -EINVAL If \a portal is not a valid index.
902 LNetClearLazyPortal(int portal
)
904 struct lnet_portal
*ptl
;
907 if (portal
< 0 || portal
>= the_lnet
.ln_nportals
)
910 ptl
= the_lnet
.ln_portals
[portal
];
912 lnet_res_lock(LNET_LOCK_EX
);
915 if (!lnet_ptl_is_lazy(ptl
)) {
916 lnet_ptl_unlock(ptl
);
917 lnet_res_unlock(LNET_LOCK_EX
);
921 if (the_lnet
.ln_shutdown
)
922 CWARN("Active lazy portal %d on exit\n", portal
);
924 CDEBUG(D_NET
, "clearing portal %d lazy\n", portal
);
926 /* grab all the blocked messages atomically */
927 list_splice_init(&ptl
->ptl_msg_delayed
, &zombies
);
929 lnet_ptl_unsetopt(ptl
, LNET_PTL_LAZY
);
931 lnet_ptl_unlock(ptl
);
932 lnet_res_unlock(LNET_LOCK_EX
);
934 lnet_drop_delayed_msg_list(&zombies
, "Clearing lazy portal attr");
938 EXPORT_SYMBOL(LNetClearLazyPortal
);