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, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * Message decoding, parsing and finalizing routines
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <lnet/lib-lnet.h>
42 lnet_build_unlink_event(struct lnet_libmd *md, struct lnet_event *ev)
46 memset(ev, 0, sizeof(*ev));
50 ev->type = LNET_EVENT_UNLINK;
51 lnet_md_deconstruct(md, &ev->md);
52 lnet_md2handle(&ev->md_handle, md);
57 * Don't need any lock, must be called after lnet_commit_md
60 lnet_build_msg_event(struct lnet_msg *msg, enum lnet_event_kind ev_type)
62 struct lnet_hdr *hdr = &msg->msg_hdr;
63 struct lnet_event *ev = &msg->msg_ev;
65 LASSERT(!msg->msg_routing);
68 ev->msg_type = msg->msg_type;
70 if (ev_type == LNET_EVENT_SEND) {
71 /* event for active message */
72 ev->target.nid = le64_to_cpu(hdr->dest_nid);
73 ev->target.pid = le32_to_cpu(hdr->dest_pid);
74 ev->initiator.nid = LNET_NID_ANY;
75 ev->initiator.pid = the_lnet.ln_pid;
76 ev->source.nid = LNET_NID_ANY;
77 ev->source.pid = the_lnet.ln_pid;
78 ev->sender = LNET_NID_ANY;
80 /* event for passive message */
81 ev->target.pid = hdr->dest_pid;
82 ev->target.nid = hdr->dest_nid;
83 ev->initiator.pid = hdr->src_pid;
84 /* Multi-Rail: resolve src_nid to "primary" peer NID */
85 ev->initiator.nid = msg->msg_initiator;
86 /* Multi-Rail: track source NID. */
87 ev->source.pid = hdr->src_pid;
88 ev->source.nid = hdr->src_nid;
89 ev->rlength = hdr->payload_length;
90 ev->sender = msg->msg_from;
91 ev->mlength = msg->msg_wanted;
92 ev->offset = msg->msg_offset;
99 case LNET_EVENT_PUT: /* passive PUT */
100 ev->pt_index = hdr->msg.put.ptl_index;
101 ev->match_bits = hdr->msg.put.match_bits;
102 ev->hdr_data = hdr->msg.put.hdr_data;
105 case LNET_EVENT_GET: /* passive GET */
106 ev->pt_index = hdr->msg.get.ptl_index;
107 ev->match_bits = hdr->msg.get.match_bits;
111 case LNET_EVENT_ACK: /* ACK */
112 ev->match_bits = hdr->msg.ack.match_bits;
113 ev->mlength = hdr->msg.ack.mlength;
116 case LNET_EVENT_REPLY: /* REPLY */
119 case LNET_EVENT_SEND: /* active message */
120 if (msg->msg_type == LNET_MSG_PUT) {
121 ev->pt_index = le32_to_cpu(hdr->msg.put.ptl_index);
122 ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
123 ev->offset = le32_to_cpu(hdr->msg.put.offset);
125 ev->rlength = le32_to_cpu(hdr->payload_length);
126 ev->hdr_data = le64_to_cpu(hdr->msg.put.hdr_data);
129 LASSERT(msg->msg_type == LNET_MSG_GET);
130 ev->pt_index = le32_to_cpu(hdr->msg.get.ptl_index);
131 ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
133 ev->rlength = le32_to_cpu(hdr->msg.get.sink_length);
134 ev->offset = le32_to_cpu(hdr->msg.get.src_offset);
142 lnet_msg_commit(struct lnet_msg *msg, int cpt)
144 struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
145 struct lnet_counters *counters = the_lnet.ln_counters[cpt];
148 /* set the message deadline */
149 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
150 msg->msg_deadline = ktime_add_ns(ktime_get(), timeout_ns);
152 /* routed message can be committed for both receiving and sending */
153 LASSERT(!msg->msg_tx_committed);
155 if (msg->msg_sending) {
156 LASSERT(!msg->msg_receiving);
157 msg->msg_tx_cpt = cpt;
158 msg->msg_tx_committed = 1;
159 if (msg->msg_rx_committed) { /* routed message REPLY */
160 LASSERT(msg->msg_onactivelist);
164 LASSERT(!msg->msg_sending);
165 msg->msg_rx_cpt = cpt;
166 msg->msg_rx_committed = 1;
169 LASSERT(!msg->msg_onactivelist);
171 msg->msg_onactivelist = 1;
172 list_add_tail(&msg->msg_activelist, &container->msc_active);
174 counters->msgs_alloc++;
175 if (counters->msgs_alloc > counters->msgs_max)
176 counters->msgs_max = counters->msgs_alloc;
180 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
182 struct lnet_counters *counters;
183 struct lnet_event *ev = &msg->msg_ev;
185 LASSERT(msg->msg_tx_committed);
189 counters = the_lnet.ln_counters[msg->msg_tx_cpt];
191 default: /* routed message */
192 LASSERT(msg->msg_routing);
193 LASSERT(msg->msg_rx_committed);
194 LASSERT(ev->type == 0);
196 counters->route_length += msg->msg_len;
197 counters->route_count++;
201 /* should have been decommitted */
202 LASSERT(!msg->msg_rx_committed);
203 /* overwritten while sending ACK */
204 LASSERT(msg->msg_type == LNET_MSG_ACK);
205 msg->msg_type = LNET_MSG_PUT; /* fix type */
208 case LNET_EVENT_SEND:
209 LASSERT(!msg->msg_rx_committed);
210 if (msg->msg_type == LNET_MSG_PUT)
211 counters->send_length += msg->msg_len;
215 LASSERT(msg->msg_rx_committed);
216 /* overwritten while sending reply, we should never be
217 * here for optimized GET */
218 LASSERT(msg->msg_type == LNET_MSG_REPLY);
219 msg->msg_type = LNET_MSG_GET; /* fix type */
223 counters->send_count++;
227 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
229 LNET_STATS_TYPE_SEND);
231 lnet_incr_stats(&msg->msg_txni->ni_stats,
233 LNET_STATS_TYPE_SEND);
235 lnet_return_tx_credits_locked(msg);
236 msg->msg_tx_committed = 0;
240 lnet_msg_decommit_rx(struct lnet_msg *msg, int status)
242 struct lnet_counters *counters;
243 struct lnet_event *ev = &msg->msg_ev;
245 LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
246 LASSERT(msg->msg_rx_committed);
251 counters = the_lnet.ln_counters[msg->msg_rx_cpt];
254 LASSERT(ev->type == 0);
255 LASSERT(msg->msg_routing);
259 LASSERT(msg->msg_type == LNET_MSG_ACK);
263 /* type is "REPLY" if it's an optimized GET on passive side,
264 * because optimized GET will never be committed for sending,
265 * so message type wouldn't be changed back to "GET" by
266 * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
267 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
268 msg->msg_type == LNET_MSG_GET);
269 counters->send_length += msg->msg_wanted;
273 LASSERT(msg->msg_type == LNET_MSG_PUT);
276 case LNET_EVENT_REPLY:
277 /* type is "GET" if it's an optimized GET on active side,
278 * see details in lnet_create_reply_msg() */
279 LASSERT(msg->msg_type == LNET_MSG_GET ||
280 msg->msg_type == LNET_MSG_REPLY);
284 counters->recv_count++;
288 lnet_incr_stats(&msg->msg_rxpeer->lpni_stats,
290 LNET_STATS_TYPE_RECV);
292 lnet_incr_stats(&msg->msg_rxni->ni_stats,
294 LNET_STATS_TYPE_RECV);
295 if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
296 counters->recv_length += msg->msg_wanted;
299 lnet_return_rx_credits_locked(msg);
300 msg->msg_rx_committed = 0;
304 lnet_msg_decommit(struct lnet_msg *msg, int cpt, int status)
308 LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
309 LASSERT(msg->msg_onactivelist);
311 if (msg->msg_tx_committed) { /* always decommit for sending first */
312 LASSERT(cpt == msg->msg_tx_cpt);
313 lnet_msg_decommit_tx(msg, status);
316 if (msg->msg_rx_committed) {
317 /* forwarding msg committed for both receiving and sending */
318 if (cpt != msg->msg_rx_cpt) {
319 lnet_net_unlock(cpt);
320 cpt2 = msg->msg_rx_cpt;
323 lnet_msg_decommit_rx(msg, status);
326 list_del(&msg->msg_activelist);
327 msg->msg_onactivelist = 0;
329 the_lnet.ln_counters[cpt2]->msgs_alloc--;
332 lnet_net_unlock(cpt2);
338 lnet_msg_attach_md(struct lnet_msg *msg, struct lnet_libmd *md,
339 unsigned int offset, unsigned int mlen)
341 /* NB: @offset and @len are only useful for receiving */
342 /* Here, we attach the MD on lnet_msg and mark it busy and
343 * decrementing its threshold. Come what may, the lnet_msg "owns"
344 * the MD until a call to lnet_msg_detach_md or lnet_finalize()
345 * signals completion. */
346 LASSERT(!msg->msg_routing);
349 if (msg->msg_receiving) { /* committed for receiving */
350 msg->msg_offset = offset;
351 msg->msg_wanted = mlen;
355 if (md->md_threshold != LNET_MD_THRESH_INF) {
356 LASSERT(md->md_threshold > 0);
360 /* build umd in event */
361 lnet_md2handle(&msg->msg_ev.md_handle, md);
362 lnet_md_deconstruct(md, &msg->msg_ev.md);
366 lnet_msg_detach_md(struct lnet_msg *msg, int status)
368 struct lnet_libmd *md = msg->msg_md;
371 /* Now it's safe to drop my caller's ref */
373 LASSERT(md->md_refcount >= 0);
375 unlink = lnet_md_unlinkable(md);
376 if (md->md_eq != NULL) {
377 msg->msg_ev.status = status;
378 msg->msg_ev.unlinked = unlink;
379 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
389 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
391 struct lnet_handle_wire ack_wmd;
393 int status = msg->msg_ev.status;
395 LASSERT(msg->msg_onactivelist);
397 if (status == 0 && msg->msg_ack) {
398 /* Only send an ACK if the PUT completed successfully */
400 lnet_msg_decommit(msg, cpt, 0);
403 lnet_net_unlock(cpt);
405 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
406 LASSERT(!msg->msg_routing);
408 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
410 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.source, 0, 0);
412 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
413 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
414 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
416 /* NB: we probably want to use NID of msg::msg_from as 3rd
417 * parameter (router NID) if it's routed message */
418 rc = lnet_send(msg->msg_ev.target.nid, msg, LNET_NID_ANY);
422 * NB: message is committed for sending, we should return
423 * on success because LND will finalize this message later.
425 * Also, there is possibility that message is committed for
426 * sending and also failed before delivering to LND,
427 * i.e: ENOMEM, in that case we can't fall through either
428 * because CPT for sending can be different with CPT for
429 * receiving, so we should return back to lnet_finalize()
430 * to make sure we are locking the correct partition.
434 } else if (status == 0 && /* OK so far */
435 (msg->msg_routing && !msg->msg_sending)) {
437 LASSERT(!msg->msg_receiving); /* called back recv already */
438 lnet_net_unlock(cpt);
440 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
444 * NB: message is committed for sending, we should return
445 * on success because LND will finalize this message later.
447 * Also, there is possibility that message is committed for
448 * sending and also failed before delivering to LND,
449 * i.e: ENOMEM, in that case we can't fall through either:
450 * - The rule is message must decommit for sending first if
451 * the it's committed for both sending and receiving
452 * - CPT for sending can be different with CPT for receiving,
453 * so we should return back to lnet_finalize() to make
454 * sure we are locking the correct partition.
459 lnet_msg_decommit(msg, cpt, status);
465 lnet_finalize(struct lnet_msg *msg, int status)
467 struct lnet_msg_container *container;
473 LASSERT(!in_interrupt());
478 msg->msg_ev.status = status;
480 if (msg->msg_md != NULL) {
481 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
484 lnet_msg_detach_md(msg, status);
485 lnet_res_unlock(cpt);
490 if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
491 /* not committed to network yet */
492 LASSERT(!msg->msg_onactivelist);
498 * NB: routed message can be committed for both receiving and sending,
499 * we should finalize in LIFO order and keep counters correct.
500 * (finalize sending first then finalize receiving)
502 cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
505 container = the_lnet.ln_msg_containers[cpt];
506 list_add_tail(&msg->msg_list, &container->msc_finalizing);
508 /* Recursion breaker. Don't complete the message here if I am (or
509 * enough other threads are) already completing messages */
512 for (i = 0; i < container->msc_nfinalizers; i++) {
513 if (container->msc_finalizers[i] == current)
516 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
520 if (i < container->msc_nfinalizers || my_slot < 0) {
521 lnet_net_unlock(cpt);
525 container->msc_finalizers[my_slot] = current;
527 while (!list_empty(&container->msc_finalizing)) {
528 msg = list_entry(container->msc_finalizing.next,
529 struct lnet_msg, msg_list);
531 list_del(&msg->msg_list);
533 /* NB drops and regains the lnet lock if it actually does
534 * anything, so my finalizing friends can chomp along too */
535 rc = lnet_complete_msg_locked(msg, cpt);
540 if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
541 lnet_net_unlock(cpt);
542 lnet_delay_rule_check();
546 container->msc_finalizers[my_slot] = NULL;
547 lnet_net_unlock(cpt);
552 EXPORT_SYMBOL(lnet_finalize);
555 lnet_msg_container_cleanup(struct lnet_msg_container *container)
559 if (container->msc_init == 0)
562 while (!list_empty(&container->msc_active)) {
563 struct lnet_msg *msg;
565 msg = list_entry(container->msc_active.next,
566 struct lnet_msg, msg_activelist);
567 LASSERT(msg->msg_onactivelist);
568 msg->msg_onactivelist = 0;
569 list_del(&msg->msg_activelist);
575 CERROR("%d active msg on exit\n", count);
577 if (container->msc_finalizers != NULL) {
578 LIBCFS_FREE(container->msc_finalizers,
579 container->msc_nfinalizers *
580 sizeof(*container->msc_finalizers));
581 container->msc_finalizers = NULL;
583 container->msc_init = 0;
587 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
591 container->msc_init = 1;
593 INIT_LIST_HEAD(&container->msc_active);
594 INIT_LIST_HEAD(&container->msc_finalizing);
597 container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
598 if (container->msc_nfinalizers == 0)
599 container->msc_nfinalizers = 1;
601 LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
602 container->msc_nfinalizers *
603 sizeof(*container->msc_finalizers));
605 if (container->msc_finalizers == NULL) {
606 CERROR("Failed to allocate message finalizers\n");
607 lnet_msg_container_cleanup(container);
615 lnet_msg_containers_destroy(void)
617 struct lnet_msg_container *container;
620 if (the_lnet.ln_msg_containers == NULL)
623 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
624 lnet_msg_container_cleanup(container);
626 cfs_percpt_free(the_lnet.ln_msg_containers);
627 the_lnet.ln_msg_containers = NULL;
631 lnet_msg_containers_create(void)
633 struct lnet_msg_container *container;
637 the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
640 if (the_lnet.ln_msg_containers == NULL) {
641 CERROR("Failed to allocate cpu-partition data for network\n");
645 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
646 rc = lnet_msg_container_setup(container, i);
648 lnet_msg_containers_destroy();