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, 2014, 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(lnet_libmd_t *md, lnet_event_t *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(lnet_msg_t *msg, lnet_event_kind_t ev_type)
62 lnet_hdr_t *hdr = &msg->msg_hdr;
63 lnet_event_t *ev = &msg->msg_ev;
65 LASSERT(!msg->msg_routing);
69 if (ev_type == LNET_EVENT_SEND) {
70 /* event for active message */
71 ev->target.nid = le64_to_cpu(hdr->dest_nid);
72 ev->target.pid = le32_to_cpu(hdr->dest_pid);
73 ev->initiator.nid = LNET_NID_ANY;
74 ev->initiator.pid = the_lnet.ln_pid;
75 ev->sender = LNET_NID_ANY;
78 /* event for passive message */
79 ev->target.pid = hdr->dest_pid;
80 ev->target.nid = hdr->dest_nid;
81 ev->initiator.pid = hdr->src_pid;
82 ev->initiator.nid = hdr->src_nid;
83 ev->rlength = hdr->payload_length;
84 ev->sender = msg->msg_from;
85 ev->mlength = msg->msg_wanted;
86 ev->offset = msg->msg_offset;
93 case LNET_EVENT_PUT: /* passive PUT */
94 ev->pt_index = hdr->msg.put.ptl_index;
95 ev->match_bits = hdr->msg.put.match_bits;
96 ev->hdr_data = hdr->msg.put.hdr_data;
99 case LNET_EVENT_GET: /* passive GET */
100 ev->pt_index = hdr->msg.get.ptl_index;
101 ev->match_bits = hdr->msg.get.match_bits;
105 case LNET_EVENT_ACK: /* ACK */
106 ev->match_bits = hdr->msg.ack.match_bits;
107 ev->mlength = hdr->msg.ack.mlength;
110 case LNET_EVENT_REPLY: /* REPLY */
113 case LNET_EVENT_SEND: /* active message */
114 if (msg->msg_type == LNET_MSG_PUT) {
115 ev->pt_index = le32_to_cpu(hdr->msg.put.ptl_index);
116 ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
117 ev->offset = le32_to_cpu(hdr->msg.put.offset);
119 ev->rlength = le32_to_cpu(hdr->payload_length);
120 ev->hdr_data = le64_to_cpu(hdr->msg.put.hdr_data);
123 LASSERT(msg->msg_type == LNET_MSG_GET);
124 ev->pt_index = le32_to_cpu(hdr->msg.get.ptl_index);
125 ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
127 ev->rlength = le32_to_cpu(hdr->msg.get.sink_length);
128 ev->offset = le32_to_cpu(hdr->msg.get.src_offset);
136 lnet_msg_commit(lnet_msg_t *msg, int cpt)
138 struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
139 lnet_counters_t *counters = the_lnet.ln_counters[cpt];
141 /* routed message can be committed for both receiving and sending */
142 LASSERT(!msg->msg_tx_committed);
144 if (msg->msg_sending) {
145 LASSERT(!msg->msg_receiving);
147 msg->msg_tx_cpt = cpt;
148 msg->msg_tx_committed = 1;
149 if (msg->msg_rx_committed) { /* routed message REPLY */
150 LASSERT(msg->msg_onactivelist);
154 LASSERT(!msg->msg_sending);
155 msg->msg_rx_cpt = cpt;
156 msg->msg_rx_committed = 1;
159 LASSERT(!msg->msg_onactivelist);
160 msg->msg_onactivelist = 1;
161 list_add(&msg->msg_activelist, &container->msc_active);
163 counters->msgs_alloc++;
164 if (counters->msgs_alloc > counters->msgs_max)
165 counters->msgs_max = counters->msgs_alloc;
169 lnet_msg_decommit_tx(lnet_msg_t *msg, int status)
171 lnet_counters_t *counters;
172 lnet_event_t *ev = &msg->msg_ev;
174 LASSERT(msg->msg_tx_committed);
178 counters = the_lnet.ln_counters[msg->msg_tx_cpt];
180 default: /* routed message */
181 LASSERT(msg->msg_routing);
182 LASSERT(msg->msg_rx_committed);
183 LASSERT(ev->type == 0);
185 counters->route_length += msg->msg_len;
186 counters->route_count++;
190 /* should have been decommitted */
191 LASSERT(!msg->msg_rx_committed);
192 /* overwritten while sending ACK */
193 LASSERT(msg->msg_type == LNET_MSG_ACK);
194 msg->msg_type = LNET_MSG_PUT; /* fix type */
197 case LNET_EVENT_SEND:
198 LASSERT(!msg->msg_rx_committed);
199 if (msg->msg_type == LNET_MSG_PUT)
200 counters->send_length += msg->msg_len;
204 LASSERT(msg->msg_rx_committed);
205 /* overwritten while sending reply, we should never be
206 * here for optimized GET */
207 LASSERT(msg->msg_type == LNET_MSG_REPLY);
208 msg->msg_type = LNET_MSG_GET; /* fix type */
212 counters->send_count++;
214 lnet_return_tx_credits_locked(msg);
215 msg->msg_tx_committed = 0;
219 lnet_msg_decommit_rx(lnet_msg_t *msg, int status)
221 lnet_counters_t *counters;
222 lnet_event_t *ev = &msg->msg_ev;
224 LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
225 LASSERT(msg->msg_rx_committed);
230 counters = the_lnet.ln_counters[msg->msg_rx_cpt];
233 LASSERT(ev->type == 0);
234 LASSERT(msg->msg_routing);
238 LASSERT(msg->msg_type == LNET_MSG_ACK);
242 /* type is "REPLY" if it's an optimized GET on passive side,
243 * because optimized GET will never be committed for sending,
244 * so message type wouldn't be changed back to "GET" by
245 * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
246 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
247 msg->msg_type == LNET_MSG_GET);
248 counters->send_length += msg->msg_wanted;
252 LASSERT(msg->msg_type == LNET_MSG_PUT);
255 case LNET_EVENT_REPLY:
256 /* type is "GET" if it's an optimized GET on active side,
257 * see details in lnet_create_reply_msg() */
258 LASSERT(msg->msg_type == LNET_MSG_GET ||
259 msg->msg_type == LNET_MSG_REPLY);
263 counters->recv_count++;
264 if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
265 counters->recv_length += msg->msg_wanted;
268 lnet_return_rx_credits_locked(msg);
269 msg->msg_rx_committed = 0;
273 lnet_msg_decommit(lnet_msg_t *msg, int cpt, int status)
277 LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
278 LASSERT(msg->msg_onactivelist);
280 if (msg->msg_tx_committed) { /* always decommit for sending first */
281 LASSERT(cpt == msg->msg_tx_cpt);
282 lnet_msg_decommit_tx(msg, status);
285 if (msg->msg_rx_committed) {
286 /* forwarding msg committed for both receiving and sending */
287 if (cpt != msg->msg_rx_cpt) {
288 lnet_net_unlock(cpt);
289 cpt2 = msg->msg_rx_cpt;
292 lnet_msg_decommit_rx(msg, status);
295 list_del(&msg->msg_activelist);
296 msg->msg_onactivelist = 0;
298 the_lnet.ln_counters[cpt2]->msgs_alloc--;
301 lnet_net_unlock(cpt2);
307 lnet_msg_attach_md(lnet_msg_t *msg, lnet_libmd_t *md,
308 unsigned int offset, unsigned int mlen)
310 /* NB: @offset and @len are only useful for receiving */
311 /* Here, we attach the MD on lnet_msg and mark it busy and
312 * decrementing its threshold. Come what may, the lnet_msg "owns"
313 * the MD until a call to lnet_msg_detach_md or lnet_finalize()
314 * signals completion. */
315 LASSERT(!msg->msg_routing);
318 if (msg->msg_receiving) { /* committed for receiving */
319 msg->msg_offset = offset;
320 msg->msg_wanted = mlen;
324 if (md->md_threshold != LNET_MD_THRESH_INF) {
325 LASSERT(md->md_threshold > 0);
329 /* build umd in event */
330 lnet_md2handle(&msg->msg_ev.md_handle, md);
331 lnet_md_deconstruct(md, &msg->msg_ev.md);
335 lnet_msg_detach_md(lnet_msg_t *msg, int status)
337 lnet_libmd_t *md = msg->msg_md;
340 /* Now it's safe to drop my caller's ref */
342 LASSERT(md->md_refcount >= 0);
344 unlink = lnet_md_unlinkable(md);
345 if (md->md_eq != NULL) {
346 msg->msg_ev.status = status;
347 msg->msg_ev.unlinked = unlink;
348 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
358 lnet_complete_msg_locked(lnet_msg_t *msg, int cpt)
360 lnet_handle_wire_t ack_wmd;
362 int status = msg->msg_ev.status;
364 LASSERT(msg->msg_onactivelist);
366 if (status == 0 && msg->msg_ack) {
367 /* Only send an ACK if the PUT completed successfully */
369 lnet_msg_decommit(msg, cpt, 0);
372 lnet_net_unlock(cpt);
374 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
375 LASSERT(!msg->msg_routing);
377 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
379 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.initiator, 0, 0);
381 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
382 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
383 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
385 /* NB: we probably want to use NID of msg::msg_from as 3rd
386 * parameter (router NID) if it's routed message */
387 rc = lnet_send(msg->msg_ev.target.nid, msg, LNET_NID_ANY);
391 * NB: message is committed for sending, we should return
392 * on success because LND will finalize this message later.
394 * Also, there is possibility that message is committed for
395 * sending and also failed before delivering to LND,
396 * i.e: ENOMEM, in that case we can't fall through either
397 * because CPT for sending can be different with CPT for
398 * receiving, so we should return back to lnet_finalize()
399 * to make sure we are locking the correct partition.
403 } else if (status == 0 && /* OK so far */
404 (msg->msg_routing && !msg->msg_sending)) {
406 LASSERT(!msg->msg_receiving); /* called back recv already */
407 lnet_net_unlock(cpt);
409 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
413 * NB: message is committed for sending, we should return
414 * on success because LND will finalize this message later.
416 * Also, there is possibility that message is committed for
417 * sending and also failed before delivering to LND,
418 * i.e: ENOMEM, in that case we can't fall through either:
419 * - The rule is message must decommit for sending first if
420 * the it's committed for both sending and receiving
421 * - CPT for sending can be different with CPT for receiving,
422 * so we should return back to lnet_finalize() to make
423 * sure we are locking the correct partition.
428 lnet_msg_decommit(msg, cpt, status);
434 lnet_finalize(lnet_ni_t *ni, lnet_msg_t *msg, int status)
436 struct lnet_msg_container *container;
442 LASSERT(!in_interrupt());
447 msg->msg_ev.status = status;
449 if (msg->msg_md != NULL) {
450 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
453 lnet_msg_detach_md(msg, status);
454 lnet_res_unlock(cpt);
459 if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
460 /* not committed to network yet */
461 LASSERT(!msg->msg_onactivelist);
467 * NB: routed message can be committed for both receiving and sending,
468 * we should finalize in LIFO order and keep counters correct.
469 * (finalize sending first then finalize receiving)
471 cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
474 container = the_lnet.ln_msg_containers[cpt];
475 list_add_tail(&msg->msg_list, &container->msc_finalizing);
477 /* Recursion breaker. Don't complete the message here if I am (or
478 * enough other threads are) already completing messages */
481 for (i = 0; i < container->msc_nfinalizers; i++) {
482 if (container->msc_finalizers[i] == current)
485 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
489 if (i < container->msc_nfinalizers || my_slot < 0) {
490 lnet_net_unlock(cpt);
494 container->msc_finalizers[my_slot] = current;
496 while (!list_empty(&container->msc_finalizing)) {
497 msg = list_entry(container->msc_finalizing.next,
498 lnet_msg_t, msg_list);
500 list_del(&msg->msg_list);
502 /* NB drops and regains the lnet lock if it actually does
503 * anything, so my finalizing friends can chomp along too */
504 rc = lnet_complete_msg_locked(msg, cpt);
509 if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
510 lnet_net_unlock(cpt);
511 lnet_delay_rule_check();
515 container->msc_finalizers[my_slot] = NULL;
516 lnet_net_unlock(cpt);
521 EXPORT_SYMBOL(lnet_finalize);
524 lnet_msg_container_cleanup(struct lnet_msg_container *container)
528 if (container->msc_init == 0)
531 while (!list_empty(&container->msc_active)) {
532 lnet_msg_t *msg = list_entry(container->msc_active.next,
533 lnet_msg_t, msg_activelist);
535 LASSERT(msg->msg_onactivelist);
536 msg->msg_onactivelist = 0;
537 list_del(&msg->msg_activelist);
543 CERROR("%d active msg on exit\n", count);
545 if (container->msc_finalizers != NULL) {
546 LIBCFS_FREE(container->msc_finalizers,
547 container->msc_nfinalizers *
548 sizeof(*container->msc_finalizers));
549 container->msc_finalizers = NULL;
551 container->msc_init = 0;
555 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
559 container->msc_init = 1;
561 INIT_LIST_HEAD(&container->msc_active);
562 INIT_LIST_HEAD(&container->msc_finalizing);
566 container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
568 LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
569 container->msc_nfinalizers *
570 sizeof(*container->msc_finalizers));
572 if (container->msc_finalizers == NULL) {
573 CERROR("Failed to allocate message finalizers\n");
574 lnet_msg_container_cleanup(container);
582 lnet_msg_containers_destroy(void)
584 struct lnet_msg_container *container;
587 if (the_lnet.ln_msg_containers == NULL)
590 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
591 lnet_msg_container_cleanup(container);
593 cfs_percpt_free(the_lnet.ln_msg_containers);
594 the_lnet.ln_msg_containers = NULL;
598 lnet_msg_containers_create(void)
600 struct lnet_msg_container *container;
604 the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
607 if (the_lnet.ln_msg_containers == NULL) {
608 CERROR("Failed to allocate cpu-partition data for network\n");
612 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
613 rc = lnet_msg_container_setup(container, i);
615 lnet_msg_containers_destroy();