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_common *common;
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 common = &the_lnet.ln_counters[cpt]->lct_common;
175 common->lcc_msgs_alloc++;
176 if (common->lcc_msgs_alloc > common->lcc_msgs_max)
177 common->lcc_msgs_max = common->lcc_msgs_alloc;
181 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
183 struct lnet_counters_common *common;
184 struct lnet_event *ev = &msg->msg_ev;
186 LASSERT(msg->msg_tx_committed);
190 common = &(the_lnet.ln_counters[msg->msg_tx_cpt]->lct_common);
192 default: /* routed message */
193 LASSERT(msg->msg_routing);
194 LASSERT(msg->msg_rx_committed);
195 LASSERT(ev->type == 0);
197 common->lcc_route_length += msg->msg_len;
198 common->lcc_route_count++;
202 /* should have been decommitted */
203 LASSERT(!msg->msg_rx_committed);
204 /* overwritten while sending ACK */
205 LASSERT(msg->msg_type == LNET_MSG_ACK);
206 msg->msg_type = LNET_MSG_PUT; /* fix type */
209 case LNET_EVENT_SEND:
210 LASSERT(!msg->msg_rx_committed);
211 if (msg->msg_type == LNET_MSG_PUT)
212 common->lcc_send_length += msg->msg_len;
216 LASSERT(msg->msg_rx_committed);
217 /* overwritten while sending reply, we should never be
218 * here for optimized GET */
219 LASSERT(msg->msg_type == LNET_MSG_REPLY);
220 msg->msg_type = LNET_MSG_GET; /* fix type */
224 common->lcc_send_count++;
228 lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
230 LNET_STATS_TYPE_SEND);
232 lnet_incr_stats(&msg->msg_txni->ni_stats,
234 LNET_STATS_TYPE_SEND);
236 lnet_return_tx_credits_locked(msg);
237 msg->msg_tx_committed = 0;
241 lnet_msg_decommit_rx(struct lnet_msg *msg, int status)
243 struct lnet_counters_common *common;
244 struct lnet_event *ev = &msg->msg_ev;
246 LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
247 LASSERT(msg->msg_rx_committed);
252 common = &(the_lnet.ln_counters[msg->msg_rx_cpt]->lct_common);
255 LASSERT(ev->type == 0);
256 LASSERT(msg->msg_routing);
260 LASSERT(msg->msg_type == LNET_MSG_ACK);
264 /* type is "REPLY" if it's an optimized GET on passive side,
265 * because optimized GET will never be committed for sending,
266 * so message type wouldn't be changed back to "GET" by
267 * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
268 LASSERT(msg->msg_type == LNET_MSG_REPLY ||
269 msg->msg_type == LNET_MSG_GET);
270 common->lcc_send_length += msg->msg_wanted;
274 LASSERT(msg->msg_type == LNET_MSG_PUT);
277 case LNET_EVENT_REPLY:
278 /* type is "GET" if it's an optimized GET on active side,
279 * see details in lnet_create_reply_msg() */
280 LASSERT(msg->msg_type == LNET_MSG_GET ||
281 msg->msg_type == LNET_MSG_REPLY);
285 common->lcc_recv_count++;
289 lnet_incr_stats(&msg->msg_rxpeer->lpni_stats,
291 LNET_STATS_TYPE_RECV);
293 lnet_incr_stats(&msg->msg_rxni->ni_stats,
295 LNET_STATS_TYPE_RECV);
296 if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
297 common->lcc_recv_length += msg->msg_wanted;
300 lnet_return_rx_credits_locked(msg);
301 msg->msg_rx_committed = 0;
305 lnet_msg_decommit(struct lnet_msg *msg, int cpt, int status)
309 LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
310 LASSERT(msg->msg_onactivelist);
312 if (msg->msg_tx_committed) { /* always decommit for sending first */
313 LASSERT(cpt == msg->msg_tx_cpt);
314 lnet_msg_decommit_tx(msg, status);
317 if (msg->msg_rx_committed) {
318 /* forwarding msg committed for both receiving and sending */
319 if (cpt != msg->msg_rx_cpt) {
320 lnet_net_unlock(cpt);
321 cpt2 = msg->msg_rx_cpt;
324 lnet_msg_decommit_rx(msg, status);
327 list_del(&msg->msg_activelist);
328 msg->msg_onactivelist = 0;
330 the_lnet.ln_counters[cpt2]->lct_common.lcc_msgs_alloc--;
333 lnet_net_unlock(cpt2);
339 lnet_msg_attach_md(struct lnet_msg *msg, struct lnet_libmd *md,
340 unsigned int offset, unsigned int mlen)
342 /* NB: @offset and @len are only useful for receiving */
343 /* Here, we attach the MD on lnet_msg and mark it busy and
344 * decrementing its threshold. Come what may, the lnet_msg "owns"
345 * the MD until a call to lnet_msg_detach_md or lnet_finalize()
346 * signals completion. */
347 LASSERT(!msg->msg_routing);
350 if (msg->msg_receiving) { /* committed for receiving */
351 msg->msg_offset = offset;
352 msg->msg_wanted = mlen;
356 if (md->md_threshold != LNET_MD_THRESH_INF) {
357 LASSERT(md->md_threshold > 0);
361 /* build umd in event */
362 lnet_md2handle(&msg->msg_ev.md_handle, md);
363 lnet_md_deconstruct(md, &msg->msg_ev.md);
367 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
369 struct lnet_handle_wire ack_wmd;
371 int status = msg->msg_ev.status;
373 LASSERT(msg->msg_onactivelist);
375 if (status == 0 && msg->msg_ack) {
376 /* Only send an ACK if the PUT completed successfully */
378 lnet_msg_decommit(msg, cpt, 0);
381 lnet_net_unlock(cpt);
383 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
384 LASSERT(!msg->msg_routing);
386 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
388 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.source, 0, 0);
390 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
391 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
392 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
394 /* NB: we probably want to use NID of msg::msg_from as 3rd
395 * parameter (router NID) if it's routed message */
396 rc = lnet_send(msg->msg_ev.target.nid, msg, msg->msg_from);
400 * NB: message is committed for sending, we should return
401 * on success because LND will finalize this message later.
403 * Also, there is possibility that message is committed for
404 * sending and also failed before delivering to LND,
405 * i.e: ENOMEM, in that case we can't fall through either
406 * because CPT for sending can be different with CPT for
407 * receiving, so we should return back to lnet_finalize()
408 * to make sure we are locking the correct partition.
412 } else if (status == 0 && /* OK so far */
413 (msg->msg_routing && !msg->msg_sending)) {
415 LASSERT(!msg->msg_receiving); /* called back recv already */
416 lnet_net_unlock(cpt);
418 rc = lnet_send(LNET_NID_ANY, 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 * - The rule is message must decommit for sending first if
429 * the it's committed for both sending and receiving
430 * - CPT for sending can be different with CPT for receiving,
431 * so we should return back to lnet_finalize() to make
432 * sure we are locking the correct partition.
437 lnet_msg_decommit(msg, cpt, status);
443 lnet_dec_healthv_locked(atomic_t *healthv, int sensitivity)
445 int h = atomic_read(healthv);
447 if (h < sensitivity) {
448 atomic_set(healthv, 0);
451 atomic_set(healthv, h);
456 lnet_handle_local_failure(struct lnet_ni *local_ni)
459 * the lnet_net_lock(0) is used to protect the addref on the ni
460 * and the recovery queue.
463 /* the mt could've shutdown and cleaned up the queues */
464 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
469 lnet_dec_healthv_locked(&local_ni->ni_healthv, lnet_health_sensitivity);
471 * add the NI to the recovery queue if it's not already there
472 * and it's health value is actually below the maximum. It's
473 * possible that the sensitivity might be set to 0, and the health
474 * value will not be reduced. In this case, there is no reason to
477 if (list_empty(&local_ni->ni_recovery) &&
478 atomic_read(&local_ni->ni_healthv) < LNET_MAX_HEALTH_VALUE) {
479 CERROR("ni %s added to recovery queue. Health = %d\n",
480 libcfs_nid2str(local_ni->ni_nid),
481 atomic_read(&local_ni->ni_healthv));
482 list_add_tail(&local_ni->ni_recovery,
483 &the_lnet.ln_mt_localNIRecovq);
484 lnet_ni_addref_locked(local_ni, 0);
490 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
492 __u32 sensitivity = lnet_health_sensitivity;
493 __u32 lp_sensitivity;
495 /* lpni could be NULL if we're in the LOLND case */
500 * If there is a health sensitivity in the peer then use that
501 * instead of the globally set one.
503 lp_sensitivity = lpni->lpni_peer_net->lpn_peer->lp_health_sensitivity;
505 sensitivity = lp_sensitivity;
507 lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
509 * add the peer NI to the recovery queue if it's not already there
510 * and it's health value is actually below the maximum. It's
511 * possible that the sensitivity might be set to 0, and the health
512 * value will not be reduced. In this case, there is no reason to
515 lnet_peer_ni_add_to_recoveryq_locked(lpni);
519 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
521 /* lpni could be NULL if we're in the LOLND case */
526 /* the mt could've shutdown and cleaned up the queues */
527 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
531 lnet_handle_remote_failure_locked(lpni);
536 lnet_incr_hstats(struct lnet_msg *msg, enum lnet_msg_hstatus hstatus)
538 struct lnet_ni *ni = msg->msg_txni;
539 struct lnet_peer_ni *lpni = msg->msg_txpeer;
540 struct lnet_counters_health *health;
542 health = &the_lnet.ln_counters[0]->lct_health;
545 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
546 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
547 health->lch_local_interrupt_count++;
549 case LNET_MSG_STATUS_LOCAL_DROPPED:
550 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
551 health->lch_local_dropped_count++;
553 case LNET_MSG_STATUS_LOCAL_ABORTED:
554 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
555 health->lch_local_aborted_count++;
557 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
558 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
559 health->lch_local_no_route_count++;
561 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
562 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
563 health->lch_local_timeout_count++;
565 case LNET_MSG_STATUS_LOCAL_ERROR:
566 atomic_inc(&ni->ni_hstats.hlt_local_error);
567 health->lch_local_error_count++;
569 case LNET_MSG_STATUS_REMOTE_DROPPED:
571 atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
572 health->lch_remote_dropped_count++;
574 case LNET_MSG_STATUS_REMOTE_ERROR:
576 atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
577 health->lch_remote_error_count++;
579 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
581 atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
582 health->lch_remote_timeout_count++;
584 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
586 atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
587 health->lch_network_timeout_count++;
589 case LNET_MSG_STATUS_OK:
597 * Do a health check on the message:
598 * return -1 if we're not going to handle the error or
599 * if we've reached the maximum number of retries.
600 * success case will return -1 as well
601 * return 0 if it the message is requeued for send
604 lnet_health_check(struct lnet_msg *msg)
606 enum lnet_msg_hstatus hstatus = msg->msg_health_status;
609 struct lnet_peer_ni *lpni;
611 /* if we're shutting down no point in handling health. */
612 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
615 LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
618 * if we're sending to the LOLND then the msg_txpeer will not be
619 * set. So no need to sanity check it.
621 if (msg->msg_tx_committed &&
622 LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) != LOLND)
623 LASSERT(msg->msg_txpeer);
624 else if (msg->msg_tx_committed &&
625 LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) == LOLND)
628 if (hstatus != LNET_MSG_STATUS_OK &&
629 ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
633 * stats are only incremented for errors so avoid wasting time
634 * incrementing statistics if there is no error.
636 if (hstatus != LNET_MSG_STATUS_OK) {
638 lnet_incr_hstats(msg, hstatus);
643 * always prefer txni/txpeer if they message is committed for both
646 if (msg->msg_tx_committed) {
648 lpni = msg->msg_txpeer;
651 lpni = msg->msg_rxpeer;
659 CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
660 libcfs_nid2str(ni->ni_nid),
661 (lo) ? "self" : libcfs_nid2str(lpni->lpni_nid),
662 lnet_msgtyp2str(msg->msg_type),
663 lnet_health_error2str(hstatus));
666 case LNET_MSG_STATUS_OK:
668 * increment the local ni health weather we successfully
669 * received or sent a message on it.
671 lnet_inc_healthv(&ni->ni_healthv);
673 * It's possible msg_txpeer is NULL in the LOLND
674 * case. Only increment the peer's health if we're
675 * receiving a message from it. It's the only sure way to
676 * know that a remote interface is up.
677 * If this interface is part of a router, then take that
678 * as indication that the router is fully healthy.
680 if (lpni && msg->msg_rx_committed) {
682 * If we're receiving a message from the router or
683 * I'm a router, then set that lpni's health to
684 * maximum so we can commence communication
686 if (lnet_isrouter(lpni) || the_lnet.ln_routing)
687 lnet_set_healthv(&lpni->lpni_healthv,
688 LNET_MAX_HEALTH_VALUE);
690 lnet_inc_healthv(&lpni->lpni_healthv);
693 /* we can finalize this message */
695 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
696 case LNET_MSG_STATUS_LOCAL_DROPPED:
697 case LNET_MSG_STATUS_LOCAL_ABORTED:
698 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
699 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
700 lnet_handle_local_failure(ni);
701 if (msg->msg_tx_committed)
702 /* add to the re-send queue */
707 * These errors will not trigger a resend so simply
708 * finalize the message
710 case LNET_MSG_STATUS_LOCAL_ERROR:
711 lnet_handle_local_failure(ni);
715 * TODO: since the remote dropped the message we can
716 * attempt a resend safely.
718 case LNET_MSG_STATUS_REMOTE_DROPPED:
719 lnet_handle_remote_failure(lpni);
720 if (msg->msg_tx_committed)
724 case LNET_MSG_STATUS_REMOTE_ERROR:
725 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
726 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
727 lnet_handle_remote_failure(lpni);
734 /* we can only resend tx_committed messages */
735 LASSERT(msg->msg_tx_committed);
737 /* don't resend recovery messages */
738 if (msg->msg_recovery) {
739 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
740 libcfs_nid2str(msg->msg_from),
741 libcfs_nid2str(msg->msg_target.nid),
742 msg->msg_retry_count);
747 * if we explicitly indicated we don't want to resend then just
750 if (msg->msg_no_resend) {
751 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
752 libcfs_nid2str(msg->msg_from),
753 libcfs_nid2str(msg->msg_target.nid),
754 msg->msg_retry_count);
758 /* check if the message has exceeded the number of retries */
759 if (msg->msg_retry_count >= lnet_retry_count) {
760 CNETERR("msg %s->%s exceeded retry count %d\n",
761 libcfs_nid2str(msg->msg_from),
762 libcfs_nid2str(msg->msg_target.nid),
763 msg->msg_retry_count);
766 msg->msg_retry_count++;
768 lnet_net_lock(msg->msg_tx_cpt);
770 /* check again under lock */
771 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
772 lnet_net_unlock(msg->msg_tx_cpt);
777 * remove message from the active list and reset it in preparation
778 * for a resend. Two exception to this
780 * 1. the router case, whe a message is committed for rx when
781 * received, then tx when it is sent. When committed to both tx and
782 * rx we don't want to remove it from the active list.
784 * 2. The REPLY case since it uses the same msg block for the GET
787 if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
788 list_del_init(&msg->msg_activelist);
789 msg->msg_onactivelist = 0;
792 * The msg_target.nid which was originally set
793 * when calling LNetGet() or LNetPut() might've
794 * been overwritten if we're routing this message.
795 * Call lnet_return_tx_credits_locked() to return
796 * the credit this message consumed. The message will
797 * consume another credit when it gets resent.
799 msg->msg_target.nid = msg->msg_hdr.dest_nid;
800 lnet_msg_decommit_tx(msg, -EAGAIN);
801 msg->msg_sending = 0;
802 msg->msg_receiving = 0;
803 msg->msg_target_is_router = 0;
805 CDEBUG(D_NET, "%s->%s:%s:%s - queuing for resend\n",
806 libcfs_nid2str(msg->msg_hdr.src_nid),
807 libcfs_nid2str(msg->msg_hdr.dest_nid),
808 lnet_msgtyp2str(msg->msg_type),
809 lnet_health_error2str(hstatus));
811 list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
812 lnet_net_unlock(msg->msg_tx_cpt);
814 wake_up(&the_lnet.ln_mt_waitq);
819 lnet_msg_detach_md(struct lnet_msg *msg, int cpt, int status)
821 struct lnet_libmd *md = msg->msg_md;
824 /* Now it's safe to drop my caller's ref */
826 LASSERT(md->md_refcount >= 0);
828 unlink = lnet_md_unlinkable(md);
829 if (md->md_eq != NULL) {
830 msg->msg_ev.status = status;
831 msg->msg_ev.unlinked = unlink;
832 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
836 lnet_detach_rsp_tracker(md, cpt);
844 lnet_is_health_check(struct lnet_msg *msg)
847 int status = msg->msg_ev.status;
849 if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
850 !msg->msg_onactivelist) {
851 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
856 if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
857 (msg->msg_rx_committed && !msg->msg_rxpeer)) {
858 CDEBUG(D_NET, "msg %p failed too early to retry and send\n",
863 /* Check for status inconsistencies */
864 if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
865 (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
866 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
867 "checking (%d, %d)\n", msg, status,
868 msg->msg_health_status);
872 CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
873 hc, status, msg->msg_health_status);
879 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
882 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
883 return "LOCAL_INTERRUPT";
884 case LNET_MSG_STATUS_LOCAL_DROPPED:
885 return "LOCAL_DROPPED";
886 case LNET_MSG_STATUS_LOCAL_ABORTED:
887 return "LOCAL_ABORTED";
888 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
889 return "LOCAL_NO_ROUTE";
890 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
891 return "LOCAL_TIMEOUT";
892 case LNET_MSG_STATUS_LOCAL_ERROR:
893 return "LOCAL_ERROR";
894 case LNET_MSG_STATUS_REMOTE_DROPPED:
895 return "REMOTE_DROPPED";
896 case LNET_MSG_STATUS_REMOTE_ERROR:
897 return "REMOTE_ERROR";
898 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
899 return "REMOTE_TIMEOUT";
900 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
901 return "NETWORK_TIMEOUT";
902 case LNET_MSG_STATUS_OK:
910 lnet_send_error_simulation(struct lnet_msg *msg,
911 enum lnet_msg_hstatus *hstatus)
916 if (list_empty(&the_lnet.ln_drop_rules))
919 /* match only health rules */
920 if (!lnet_drop_rule_match(&msg->msg_hdr, LNET_NID_ANY,
924 CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
925 libcfs_nid2str(msg->msg_hdr.src_nid),
926 libcfs_nid2str(msg->msg_txni->ni_nid),
927 libcfs_nid2str(msg->msg_hdr.dest_nid),
928 lnet_msgtyp2str(msg->msg_type),
929 lnet_health_error2str(*hstatus));
933 EXPORT_SYMBOL(lnet_send_error_simulation);
936 lnet_finalize(struct lnet_msg *msg, int status)
938 struct lnet_msg_container *container;
944 LASSERT(!in_interrupt());
949 msg->msg_ev.status = status;
951 if (lnet_is_health_check(msg)) {
953 * Check the health status of the message. If it has one
954 * of the errors that we're supposed to handle, and it has
955 * not timed out, then
956 * 1. Decrement the appropriate health_value
957 * 2. queue the message on the resend queue
959 * if the message send is success, timed out or failed in the
960 * health check for any reason then we'll just finalize the
961 * message. Otherwise just return since the message has been
962 * put on the resend queue.
964 if (!lnet_health_check(msg))
965 /* Message is queued for resend */
970 * We're not going to resend this message so detach its MD and invoke
971 * the appropriate callbacks
973 if (msg->msg_md != NULL) {
974 cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
976 lnet_msg_detach_md(msg, cpt, status);
977 lnet_res_unlock(cpt);
981 if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
982 /* not committed to network yet */
983 LASSERT(!msg->msg_onactivelist);
989 * NB: routed message can be committed for both receiving and sending,
990 * we should finalize in LIFO order and keep counters correct.
991 * (finalize sending first then finalize receiving)
993 cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
996 container = the_lnet.ln_msg_containers[cpt];
997 list_add_tail(&msg->msg_list, &container->msc_finalizing);
999 /* Recursion breaker. Don't complete the message here if I am (or
1000 * enough other threads are) already completing messages */
1003 for (i = 0; i < container->msc_nfinalizers; i++) {
1004 if (container->msc_finalizers[i] == current)
1007 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
1011 if (i < container->msc_nfinalizers || my_slot < 0) {
1012 lnet_net_unlock(cpt);
1016 container->msc_finalizers[my_slot] = current;
1019 while (!list_empty(&container->msc_finalizing)) {
1020 msg = list_entry(container->msc_finalizing.next,
1021 struct lnet_msg, msg_list);
1023 list_del_init(&msg->msg_list);
1025 /* NB drops and regains the lnet lock if it actually does
1026 * anything, so my finalizing friends can chomp along too */
1027 rc = lnet_complete_msg_locked(msg, cpt);
1032 if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1033 lnet_net_unlock(cpt);
1034 lnet_delay_rule_check();
1038 container->msc_finalizers[my_slot] = NULL;
1039 lnet_net_unlock(cpt);
1044 EXPORT_SYMBOL(lnet_finalize);
1047 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1051 if (container->msc_init == 0)
1054 while (!list_empty(&container->msc_active)) {
1055 struct lnet_msg *msg;
1057 msg = list_entry(container->msc_active.next,
1058 struct lnet_msg, msg_activelist);
1059 LASSERT(msg->msg_onactivelist);
1060 msg->msg_onactivelist = 0;
1061 list_del_init(&msg->msg_activelist);
1067 CERROR("%d active msg on exit\n", count);
1069 if (container->msc_finalizers != NULL) {
1070 LIBCFS_FREE(container->msc_finalizers,
1071 container->msc_nfinalizers *
1072 sizeof(*container->msc_finalizers));
1073 container->msc_finalizers = NULL;
1075 container->msc_init = 0;
1079 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1083 container->msc_init = 1;
1085 INIT_LIST_HEAD(&container->msc_active);
1086 INIT_LIST_HEAD(&container->msc_finalizing);
1088 /* number of CPUs */
1089 container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1090 if (container->msc_nfinalizers == 0)
1091 container->msc_nfinalizers = 1;
1093 LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1094 container->msc_nfinalizers *
1095 sizeof(*container->msc_finalizers));
1097 if (container->msc_finalizers == NULL) {
1098 CERROR("Failed to allocate message finalizers\n");
1099 lnet_msg_container_cleanup(container);
1107 lnet_msg_containers_destroy(void)
1109 struct lnet_msg_container *container;
1112 if (the_lnet.ln_msg_containers == NULL)
1115 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1116 lnet_msg_container_cleanup(container);
1118 cfs_percpt_free(the_lnet.ln_msg_containers);
1119 the_lnet.ln_msg_containers = NULL;
1123 lnet_msg_containers_create(void)
1125 struct lnet_msg_container *container;
1129 the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1130 sizeof(*container));
1132 if (the_lnet.ln_msg_containers == NULL) {
1133 CERROR("Failed to allocate cpu-partition data for network\n");
1137 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1138 rc = lnet_msg_container_setup(container, i);
1140 lnet_msg_containers_destroy();