1 /*
2 * GPL HEADER START
3 *
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
19 *
20 * GPL HEADER END
21 */
22 /*
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
25 *
26 * Copyright (c) 2012, Intel Corporation.
27 */
28 /*
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
31 *
32 * lnet/lnet/lib-msg.c
33 *
34 * Message decoding, parsing and finalizing routines
35 */
36
37 #define DEBUG_SUBSYSTEM S_LNET
38
39 #include "../../include/linux/lnet/lib-lnet.h"
40
41 void
lnet_build_unlink_event(lnet_libmd_t * md,lnet_event_t * ev)42 lnet_build_unlink_event(lnet_libmd_t *md, lnet_event_t *ev)
43 {
44 memset(ev, 0, sizeof(*ev));
45
46 ev->status = 0;
47 ev->unlinked = 1;
48 ev->type = LNET_EVENT_UNLINK;
49 lnet_md_deconstruct(md, &ev->md);
50 lnet_md2handle(&ev->md_handle, md);
51 }
52
53 /*
54 * Don't need any lock, must be called after lnet_commit_md
55 */
56 void
lnet_build_msg_event(lnet_msg_t * msg,lnet_event_kind_t ev_type)57 lnet_build_msg_event(lnet_msg_t *msg, lnet_event_kind_t ev_type)
58 {
59 lnet_hdr_t *hdr = &msg->msg_hdr;
60 lnet_event_t *ev = &msg->msg_ev;
61
62 LASSERT(!msg->msg_routing);
63
64 ev->type = ev_type;
65
66 if (ev_type == LNET_EVENT_SEND) {
67 /* event for active message */
68 ev->target.nid = le64_to_cpu(hdr->dest_nid);
69 ev->target.pid = le32_to_cpu(hdr->dest_pid);
70 ev->initiator.nid = LNET_NID_ANY;
71 ev->initiator.pid = the_lnet.ln_pid;
72 ev->sender = LNET_NID_ANY;
73 } else {
74 /* event for passive message */
75 ev->target.pid = hdr->dest_pid;
76 ev->target.nid = hdr->dest_nid;
77 ev->initiator.pid = hdr->src_pid;
78 ev->initiator.nid = hdr->src_nid;
79 ev->rlength = hdr->payload_length;
80 ev->sender = msg->msg_from;
81 ev->mlength = msg->msg_wanted;
82 ev->offset = msg->msg_offset;
83 }
84
85 switch (ev_type) {
86 default:
87 LBUG();
88
89 case LNET_EVENT_PUT: /* passive PUT */
90 ev->pt_index = hdr->msg.put.ptl_index;
91 ev->match_bits = hdr->msg.put.match_bits;
92 ev->hdr_data = hdr->msg.put.hdr_data;
93 return;
94
95 case LNET_EVENT_GET: /* passive GET */
96 ev->pt_index = hdr->msg.get.ptl_index;
97 ev->match_bits = hdr->msg.get.match_bits;
98 ev->hdr_data = 0;
99 return;
100
101 case LNET_EVENT_ACK: /* ACK */
102 ev->match_bits = hdr->msg.ack.match_bits;
103 ev->mlength = hdr->msg.ack.mlength;
104 return;
105
106 case LNET_EVENT_REPLY: /* REPLY */
107 return;
108
109 case LNET_EVENT_SEND: /* active message */
110 if (msg->msg_type == LNET_MSG_PUT) {
111 ev->pt_index = le32_to_cpu(hdr->msg.put.ptl_index);
112 ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
113 ev->offset = le32_to_cpu(hdr->msg.put.offset);
114 ev->mlength =
115 ev->rlength = le32_to_cpu(hdr->payload_length);
116 ev->hdr_data = le64_to_cpu(hdr->msg.put.hdr_data);
117
118 } else {
119 LASSERT(msg->msg_type == LNET_MSG_GET);
120 ev->pt_index = le32_to_cpu(hdr->msg.get.ptl_index);
121 ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
122 ev->mlength =
123 ev->rlength = le32_to_cpu(hdr->msg.get.sink_length);
124 ev->offset = le32_to_cpu(hdr->msg.get.src_offset);
125 ev->hdr_data = 0;
126 }
127 return;
128 }
129 }
130
131 void
lnet_msg_commit(lnet_msg_t * msg,int cpt)132 lnet_msg_commit(lnet_msg_t *msg, int cpt)
133 {
134 struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
135 lnet_counters_t *counters = the_lnet.ln_counters[cpt];
136
137 /* routed message can be committed for both receiving and sending */
138 LASSERT(!msg->msg_tx_committed);
139
140 if (msg->msg_sending) {
141 LASSERT(!msg->msg_receiving);
142
143 msg->msg_tx_cpt = cpt;
144 msg->msg_tx_committed = 1;
145 if (msg->msg_rx_committed) { /* routed message REPLY */
146 LASSERT(msg->msg_onactivelist);
147 return;
148 }
149 } else {
150 LASSERT(!msg->msg_sending);
151 msg->msg_rx_cpt = cpt;
152 msg->msg_rx_committed = 1;
153 }
154
155 LASSERT(!msg->msg_onactivelist);
156 msg->msg_onactivelist = 1;
157 list_add(&msg->msg_activelist, &container->msc_active);
158
159 counters->msgs_alloc++;
160 if (counters->msgs_alloc > counters->msgs_max)
161 counters->msgs_max = counters->msgs_alloc;
162 }
163
164 static void
lnet_msg_decommit_tx(lnet_msg_t * msg,int status)165 lnet_msg_decommit_tx(lnet_msg_t *msg, int status)
166 {
167 lnet_counters_t *counters;
168 lnet_event_t *ev = &msg->msg_ev;
169
170 LASSERT(msg->msg_tx_committed);
171 if (status)
172 goto out;
173
174 counters = the_lnet.ln_counters[msg->msg_tx_cpt];
175 switch (ev->type) {
176 default: /* routed message */
177 LASSERT(msg->msg_routing);
178 LASSERT(msg->msg_rx_committed);
179 LASSERT(!ev->type);
180
181 counters->route_length += msg->msg_len;
182 counters->route_count++;
183 goto out;
184
185 case LNET_EVENT_PUT:
186 /* should have been decommitted */
187 LASSERT(!msg->msg_rx_committed);
188 /* overwritten while sending ACK */
189 LASSERT(msg->msg_type == LNET_MSG_ACK);
190 msg->msg_type = LNET_MSG_PUT; /* fix type */
191 break;
192
193 case LNET_EVENT_SEND:
194 LASSERT(!msg->msg_rx_committed);
195 if (msg->msg_type == LNET_MSG_PUT)
196 counters->send_length += msg->msg_len;
197 break;
198
199 case LNET_EVENT_GET:
200 LASSERT(msg->msg_rx_committed);
201 /*
202 * overwritten while sending reply, we should never be
203 * here for optimized GET
204 */
205 LASSERT(msg->msg_type == LNET_MSG_REPLY);
206 msg->msg_type = LNET_MSG_GET; /* fix type */
207 break;
208 }
209
210 counters->send_count++;
211 out:
212 lnet_return_tx_credits_locked(msg);
213 msg->msg_tx_committed = 0;
214 }
215
216 static void
lnet_msg_decommit_rx(lnet_msg_t * msg,int status)217 lnet_msg_decommit_rx(lnet_msg_t *msg, int status)
218 {
219 lnet_counters_t *counters;
220 lnet_event_t *ev = &msg->msg_ev;
221
222 LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
223 LASSERT(msg->msg_rx_committed);
224
225 if (status)
226 goto out;
227
228 counters = the_lnet.ln_counters[msg->msg_rx_cpt];
229 switch (ev->type) {
230 default:
231 LASSERT(!ev->type);
232 LASSERT(msg->msg_routing);
233 goto out;
234
235 case LNET_EVENT_ACK:
236 LASSERT(msg->msg_type == LNET_MSG_ACK);
237 break;
238
239 case LNET_EVENT_GET:
240 /*
241 * type is "REPLY" if it's an optimized GET on passive side,
242 * because optimized GET will never be committed for sending,
243 * so message type wouldn't be changed back to "GET" by
244 * lnet_msg_decommit_tx(), see details in lnet_parse_get()
245 */
246 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
247 msg->msg_type == LNET_MSG_GET);
248 counters->send_length += msg->msg_wanted;
249 break;
250
251 case LNET_EVENT_PUT:
252 LASSERT(msg->msg_type == LNET_MSG_PUT);
253 break;
254
255 case LNET_EVENT_REPLY:
256 /*
257 * type is "GET" if it's an optimized GET on active side,
258 * see details in lnet_create_reply_msg()
259 */
260 LASSERT(msg->msg_type == LNET_MSG_GET ||
261 msg->msg_type == LNET_MSG_REPLY);
262 break;
263 }
264
265 counters->recv_count++;
266 if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
267 counters->recv_length += msg->msg_wanted;
268
269 out:
270 lnet_return_rx_credits_locked(msg);
271 msg->msg_rx_committed = 0;
272 }
273
274 void
lnet_msg_decommit(lnet_msg_t * msg,int cpt,int status)275 lnet_msg_decommit(lnet_msg_t *msg, int cpt, int status)
276 {
277 int cpt2 = cpt;
278
279 LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
280 LASSERT(msg->msg_onactivelist);
281
282 if (msg->msg_tx_committed) { /* always decommit for sending first */
283 LASSERT(cpt == msg->msg_tx_cpt);
284 lnet_msg_decommit_tx(msg, status);
285 }
286
287 if (msg->msg_rx_committed) {
288 /* forwarding msg committed for both receiving and sending */
289 if (cpt != msg->msg_rx_cpt) {
290 lnet_net_unlock(cpt);
291 cpt2 = msg->msg_rx_cpt;
292 lnet_net_lock(cpt2);
293 }
294 lnet_msg_decommit_rx(msg, status);
295 }
296
297 list_del(&msg->msg_activelist);
298 msg->msg_onactivelist = 0;
299
300 the_lnet.ln_counters[cpt2]->msgs_alloc--;
301
302 if (cpt2 != cpt) {
303 lnet_net_unlock(cpt2);
304 lnet_net_lock(cpt);
305 }
306 }
307
308 void
lnet_msg_attach_md(lnet_msg_t * msg,lnet_libmd_t * md,unsigned int offset,unsigned int mlen)309 lnet_msg_attach_md(lnet_msg_t *msg, lnet_libmd_t *md,
310 unsigned int offset, unsigned int mlen)
311 {
312 /* NB: @offset and @len are only useful for receiving */
313 /*
314 * Here, we attach the MD on lnet_msg and mark it busy and
315 * decrementing its threshold. Come what may, the lnet_msg "owns"
316 * the MD until a call to lnet_msg_detach_md or lnet_finalize()
317 * signals completion.
318 */
319 LASSERT(!msg->msg_routing);
320
321 msg->msg_md = md;
322 if (msg->msg_receiving) { /* committed for receiving */
323 msg->msg_offset = offset;
324 msg->msg_wanted = mlen;
325 }
326
327 md->md_refcount++;
328 if (md->md_threshold != LNET_MD_THRESH_INF) {
329 LASSERT(md->md_threshold > 0);
330 md->md_threshold--;
331 }
332
333 /* build umd in event */
334 lnet_md2handle(&msg->msg_ev.md_handle, md);
335 lnet_md_deconstruct(md, &msg->msg_ev.md);
336 }
337
338 void
lnet_msg_detach_md(lnet_msg_t * msg,int status)339 lnet_msg_detach_md(lnet_msg_t *msg, int status)
340 {
341 lnet_libmd_t *md = msg->msg_md;
342 int unlink;
343
344 /* Now it's safe to drop my caller's ref */
345 md->md_refcount--;
346 LASSERT(md->md_refcount >= 0);
347
348 unlink = lnet_md_unlinkable(md);
349 if (md->md_eq) {
350 msg->msg_ev.status = status;
351 msg->msg_ev.unlinked = unlink;
352 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
353 }
354
355 if (unlink)
356 lnet_md_unlink(md);
357
358 msg->msg_md = NULL;
359 }
360
361 static int
lnet_complete_msg_locked(lnet_msg_t * msg,int cpt)362 lnet_complete_msg_locked(lnet_msg_t *msg, int cpt)
363 {
364 lnet_handle_wire_t ack_wmd;
365 int rc;
366 int status = msg->msg_ev.status;
367
368 LASSERT(msg->msg_onactivelist);
369
370 if (!status && msg->msg_ack) {
371 /* Only send an ACK if the PUT completed successfully */
372
373 lnet_msg_decommit(msg, cpt, 0);
374
375 msg->msg_ack = 0;
376 lnet_net_unlock(cpt);
377
378 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
379 LASSERT(!msg->msg_routing);
380
381 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
382
383 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.initiator, 0, 0);
384
385 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
386 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
387 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
388
389 /*
390 * NB: we probably want to use NID of msg::msg_from as 3rd
391 * parameter (router NID) if it's routed message
392 */
393 rc = lnet_send(msg->msg_ev.target.nid, msg, LNET_NID_ANY);
394
395 lnet_net_lock(cpt);
396 /*
397 * NB: message is committed for sending, we should return
398 * on success because LND will finalize this message later.
399 *
400 * Also, there is possibility that message is committed for
401 * sending and also failed before delivering to LND,
402 * i.e: ENOMEM, in that case we can't fall through either
403 * because CPT for sending can be different with CPT for
404 * receiving, so we should return back to lnet_finalize()
405 * to make sure we are locking the correct partition.
406 */
407 return rc;
408
409 } else if (!status && /* OK so far */
410 (msg->msg_routing && !msg->msg_sending)) {
411 /* not forwarded */
412 LASSERT(!msg->msg_receiving); /* called back recv already */
413 lnet_net_unlock(cpt);
414
415 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
416
417 lnet_net_lock(cpt);
418 /*
419 * NB: message is committed for sending, we should return
420 * on success because LND will finalize this message later.
421 *
422 * Also, there is possibility that message is committed for
423 * sending and also failed before delivering to LND,
424 * i.e: ENOMEM, in that case we can't fall through either:
425 * - The rule is message must decommit for sending first if
426 * the it's committed for both sending and receiving
427 * - CPT for sending can be different with CPT for receiving,
428 * so we should return back to lnet_finalize() to make
429 * sure we are locking the correct partition.
430 */
431 return rc;
432 }
433
434 lnet_msg_decommit(msg, cpt, status);
435 lnet_msg_free(msg);
436 return 0;
437 }
438
439 void
lnet_finalize(lnet_ni_t * ni,lnet_msg_t * msg,int status)440 lnet_finalize(lnet_ni_t *ni, lnet_msg_t *msg, int status)
441 {
442 struct lnet_msg_container *container;
443 int my_slot;
444 int cpt;
445 int rc;
446 int i;
447
448 LASSERT(!in_interrupt());
449
450 if (!msg)
451 return;
452
453 msg->msg_ev.status = status;
454
455 if (msg->msg_md) {
456 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
457
458 lnet_res_lock(cpt);
459 lnet_msg_detach_md(msg, status);
460 lnet_res_unlock(cpt);
461 }
462
463 again:
464 rc = 0;
465 if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
466 /* not committed to network yet */
467 LASSERT(!msg->msg_onactivelist);
468 lnet_msg_free(msg);
469 return;
470 }
471
472 /*
473 * NB: routed message can be committed for both receiving and sending,
474 * we should finalize in LIFO order and keep counters correct.
475 * (finalize sending first then finalize receiving)
476 */
477 cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
478 lnet_net_lock(cpt);
479
480 container = the_lnet.ln_msg_containers[cpt];
481 list_add_tail(&msg->msg_list, &container->msc_finalizing);
482
483 /*
484 * Recursion breaker. Don't complete the message here if I am (or
485 * enough other threads are) already completing messages
486 */
487 my_slot = -1;
488 for (i = 0; i < container->msc_nfinalizers; i++) {
489 if (container->msc_finalizers[i] == current)
490 break;
491
492 if (my_slot < 0 && !container->msc_finalizers[i])
493 my_slot = i;
494 }
495
496 if (i < container->msc_nfinalizers || my_slot < 0) {
497 lnet_net_unlock(cpt);
498 return;
499 }
500
501 container->msc_finalizers[my_slot] = current;
502
503 while (!list_empty(&container->msc_finalizing)) {
504 msg = list_entry(container->msc_finalizing.next,
505 lnet_msg_t, msg_list);
506
507 list_del(&msg->msg_list);
508
509 /*
510 * NB drops and regains the lnet lock if it actually does
511 * anything, so my finalizing friends can chomp along too
512 */
513 rc = lnet_complete_msg_locked(msg, cpt);
514 if (rc)
515 break;
516 }
517
518 if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
519 lnet_net_unlock(cpt);
520 lnet_delay_rule_check();
521 lnet_net_lock(cpt);
522 }
523
524 container->msc_finalizers[my_slot] = NULL;
525 lnet_net_unlock(cpt);
526
527 if (rc)
528 goto again;
529 }
530 EXPORT_SYMBOL(lnet_finalize);
531
532 void
lnet_msg_container_cleanup(struct lnet_msg_container * container)533 lnet_msg_container_cleanup(struct lnet_msg_container *container)
534 {
535 int count = 0;
536
537 if (!container->msc_init)
538 return;
539
540 while (!list_empty(&container->msc_active)) {
541 lnet_msg_t *msg = list_entry(container->msc_active.next,
542 lnet_msg_t, msg_activelist);
543
544 LASSERT(msg->msg_onactivelist);
545 msg->msg_onactivelist = 0;
546 list_del(&msg->msg_activelist);
547 lnet_msg_free(msg);
548 count++;
549 }
550
551 if (count > 0)
552 CERROR("%d active msg on exit\n", count);
553
554 if (container->msc_finalizers) {
555 LIBCFS_FREE(container->msc_finalizers,
556 container->msc_nfinalizers *
557 sizeof(*container->msc_finalizers));
558 container->msc_finalizers = NULL;
559 }
560 container->msc_init = 0;
561 }
562
563 int
lnet_msg_container_setup(struct lnet_msg_container * container,int cpt)564 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
565 {
566 container->msc_init = 1;
567
568 INIT_LIST_HEAD(&container->msc_active);
569 INIT_LIST_HEAD(&container->msc_finalizing);
570
571 /* number of CPUs */
572 container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
573
574 LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
575 container->msc_nfinalizers *
576 sizeof(*container->msc_finalizers));
577
578 if (!container->msc_finalizers) {
579 CERROR("Failed to allocate message finalizers\n");
580 lnet_msg_container_cleanup(container);
581 return -ENOMEM;
582 }
583
584 return 0;
585 }
586
587 void
lnet_msg_containers_destroy(void)588 lnet_msg_containers_destroy(void)
589 {
590 struct lnet_msg_container *container;
591 int i;
592
593 if (!the_lnet.ln_msg_containers)
594 return;
595
596 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
597 lnet_msg_container_cleanup(container);
598
599 cfs_percpt_free(the_lnet.ln_msg_containers);
600 the_lnet.ln_msg_containers = NULL;
601 }
602
603 int
lnet_msg_containers_create(void)604 lnet_msg_containers_create(void)
605 {
606 struct lnet_msg_container *container;
607 int rc;
608 int i;
609
610 the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
611 sizeof(*container));
612
613 if (!the_lnet.ln_msg_containers) {
614 CERROR("Failed to allocate cpu-partition data for network\n");
615 return -ENOMEM;
616 }
617
618 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
619 rc = lnet_msg_container_setup(container, i);
620 if (rc) {
621 lnet_msg_containers_destroy();
622 return rc;
623 }
624 }
625
626 return 0;
627 }
628