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)
445 int h = atomic_read(healthv);
447 if (h < lnet_health_sensitivity) {
448 atomic_set(healthv, 0);
450 h -= lnet_health_sensitivity;
451 atomic_set(healthv, h);
456 lnet_handle_local_failure(struct lnet_msg *msg)
458 struct lnet_ni *local_ni;
460 local_ni = msg->msg_txni;
463 * the lnet_net_lock(0) is used to protect the addref on the ni
464 * and the recovery queue.
467 /* the mt could've shutdown and cleaned up the queues */
468 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
473 lnet_dec_healthv_locked(&local_ni->ni_healthv);
475 * add the NI to the recovery queue if it's not already there
476 * and it's health value is actually below the maximum. It's
477 * possible that the sensitivity might be set to 0, and the health
478 * value will not be reduced. In this case, there is no reason to
481 if (list_empty(&local_ni->ni_recovery) &&
482 atomic_read(&local_ni->ni_healthv) < LNET_MAX_HEALTH_VALUE) {
483 CERROR("ni %s added to recovery queue. Health = %d\n",
484 libcfs_nid2str(local_ni->ni_nid),
485 atomic_read(&local_ni->ni_healthv));
486 list_add_tail(&local_ni->ni_recovery,
487 &the_lnet.ln_mt_localNIRecovq);
488 lnet_ni_addref_locked(local_ni, 0);
494 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
496 /* lpni could be NULL if we're in the LOLND case */
500 lnet_dec_healthv_locked(&lpni->lpni_healthv);
502 * add the peer NI to the recovery queue if it's not already there
503 * and it's health value is actually below the maximum. It's
504 * possible that the sensitivity might be set to 0, and the health
505 * value will not be reduced. In this case, there is no reason to
508 lnet_peer_ni_add_to_recoveryq_locked(lpni);
512 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
514 /* lpni could be NULL if we're in the LOLND case */
519 lnet_handle_remote_failure_locked(lpni);
524 lnet_incr_hstats(struct lnet_msg *msg, enum lnet_msg_hstatus hstatus)
526 struct lnet_ni *ni = msg->msg_txni;
527 struct lnet_peer_ni *lpni = msg->msg_txpeer;
528 struct lnet_counters_health *health;
530 health = &the_lnet.ln_counters[0]->lct_health;
533 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
534 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
535 health->lch_local_interrupt_count++;
537 case LNET_MSG_STATUS_LOCAL_DROPPED:
538 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
539 health->lch_local_dropped_count++;
541 case LNET_MSG_STATUS_LOCAL_ABORTED:
542 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
543 health->lch_local_aborted_count++;
545 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
546 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
547 health->lch_local_no_route_count++;
549 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
550 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
551 health->lch_local_timeout_count++;
553 case LNET_MSG_STATUS_LOCAL_ERROR:
554 atomic_inc(&ni->ni_hstats.hlt_local_error);
555 health->lch_local_error_count++;
557 case LNET_MSG_STATUS_REMOTE_DROPPED:
559 atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
560 health->lch_remote_dropped_count++;
562 case LNET_MSG_STATUS_REMOTE_ERROR:
564 atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
565 health->lch_remote_error_count++;
567 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
569 atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
570 health->lch_remote_timeout_count++;
572 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
574 atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
575 health->lch_network_timeout_count++;
577 case LNET_MSG_STATUS_OK:
585 * Do a health check on the message:
586 * return -1 if we're not going to handle the error or
587 * if we've reached the maximum number of retries.
588 * success case will return -1 as well
589 * return 0 if it the message is requeued for send
592 lnet_health_check(struct lnet_msg *msg)
594 enum lnet_msg_hstatus hstatus = msg->msg_health_status;
597 /* if we're shutting down no point in handling health. */
598 if (the_lnet.ln_state != LNET_STATE_RUNNING)
601 LASSERT(msg->msg_txni);
604 * if we're sending to the LOLND then the msg_txpeer will not be
605 * set. So no need to sanity check it.
607 if (LNET_NETTYP(LNET_NIDNET(msg->msg_txni->ni_nid)) != LOLND)
608 LASSERT(msg->msg_txpeer);
612 if (hstatus != LNET_MSG_STATUS_OK &&
613 ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
617 * stats are only incremented for errors so avoid wasting time
618 * incrementing statistics if there is no error.
620 if (hstatus != LNET_MSG_STATUS_OK) {
622 lnet_incr_hstats(msg, hstatus);
626 CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
627 libcfs_nid2str(msg->msg_txni->ni_nid),
628 (lo) ? "self" : libcfs_nid2str(msg->msg_txpeer->lpni_nid),
629 lnet_msgtyp2str(msg->msg_type),
630 lnet_health_error2str(hstatus));
633 case LNET_MSG_STATUS_OK:
634 lnet_inc_healthv(&msg->msg_txni->ni_healthv);
636 * It's possible msg_txpeer is NULL in the LOLND
640 lnet_inc_healthv(&msg->msg_txpeer->lpni_healthv);
642 /* we can finalize this message */
644 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
645 case LNET_MSG_STATUS_LOCAL_DROPPED:
646 case LNET_MSG_STATUS_LOCAL_ABORTED:
647 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
648 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
649 lnet_handle_local_failure(msg);
650 /* add to the re-send queue */
654 * These errors will not trigger a resend so simply
655 * finalize the message
657 case LNET_MSG_STATUS_LOCAL_ERROR:
658 lnet_handle_local_failure(msg);
662 * TODO: since the remote dropped the message we can
663 * attempt a resend safely.
665 case LNET_MSG_STATUS_REMOTE_DROPPED:
666 lnet_handle_remote_failure(msg->msg_txpeer);
669 case LNET_MSG_STATUS_REMOTE_ERROR:
670 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
671 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
672 lnet_handle_remote_failure(msg->msg_txpeer);
679 /* don't resend recovery messages */
680 if (msg->msg_recovery) {
681 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
682 libcfs_nid2str(msg->msg_from),
683 libcfs_nid2str(msg->msg_target.nid),
684 msg->msg_retry_count);
689 * if we explicitly indicated we don't want to resend then just
692 if (msg->msg_no_resend) {
693 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
694 libcfs_nid2str(msg->msg_from),
695 libcfs_nid2str(msg->msg_target.nid),
696 msg->msg_retry_count);
700 /* check if the message has exceeded the number of retries */
701 if (msg->msg_retry_count >= lnet_retry_count) {
702 CNETERR("msg %s->%s exceeded retry count %d\n",
703 libcfs_nid2str(msg->msg_from),
704 libcfs_nid2str(msg->msg_target.nid),
705 msg->msg_retry_count);
708 msg->msg_retry_count++;
710 lnet_net_lock(msg->msg_tx_cpt);
713 * remove message from the active list and reset it in preparation
714 * for a resend. Two exception to this
716 * 1. the router case, whe a message is committed for rx when
717 * received, then tx when it is sent. When committed to both tx and
718 * rx we don't want to remove it from the active list.
720 * 2. The REPLY case since it uses the same msg block for the GET
723 if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
724 list_del_init(&msg->msg_activelist);
725 msg->msg_onactivelist = 0;
728 * The msg_target.nid which was originally set
729 * when calling LNetGet() or LNetPut() might've
730 * been overwritten if we're routing this message.
731 * Call lnet_return_tx_credits_locked() to return
732 * the credit this message consumed. The message will
733 * consume another credit when it gets resent.
735 msg->msg_target.nid = msg->msg_hdr.dest_nid;
736 lnet_msg_decommit_tx(msg, -EAGAIN);
737 msg->msg_sending = 0;
738 msg->msg_receiving = 0;
739 msg->msg_target_is_router = 0;
741 CDEBUG(D_NET, "%s->%s:%s:%s - queuing for resend\n",
742 libcfs_nid2str(msg->msg_hdr.src_nid),
743 libcfs_nid2str(msg->msg_hdr.dest_nid),
744 lnet_msgtyp2str(msg->msg_type),
745 lnet_health_error2str(hstatus));
747 list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
748 lnet_net_unlock(msg->msg_tx_cpt);
750 wake_up(&the_lnet.ln_mt_waitq);
755 lnet_msg_detach_md(struct lnet_msg *msg, int cpt, int status)
757 struct lnet_libmd *md = msg->msg_md;
760 /* Now it's safe to drop my caller's ref */
762 LASSERT(md->md_refcount >= 0);
764 unlink = lnet_md_unlinkable(md);
765 if (md->md_eq != NULL) {
766 msg->msg_ev.status = status;
767 msg->msg_ev.unlinked = unlink;
768 lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
773 * if this is an ACK or a REPLY then make sure to remove the
776 if (msg->msg_ev.type == LNET_EVENT_REPLY ||
777 msg->msg_ev.type == LNET_EVENT_ACK)
778 lnet_detach_rsp_tracker(msg->msg_md, cpt);
786 lnet_detach_md(struct lnet_msg *msg, int status)
788 int cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
791 lnet_msg_detach_md(msg, cpt, status);
792 lnet_res_unlock(cpt);
796 lnet_is_health_check(struct lnet_msg *msg)
799 int status = msg->msg_ev.status;
802 * perform a health check for any message committed for transmit
804 hc = msg->msg_tx_committed;
806 /* Check for status inconsistencies */
808 ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
809 (status && msg->msg_health_status == LNET_MSG_STATUS_OK))) {
810 CERROR("Msg is in inconsistent state, don't perform health "
811 "checking (%d, %d)\n", status, msg->msg_health_status);
815 CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
816 hc, status, msg->msg_health_status);
822 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
825 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
826 return "LOCAL_INTERRUPT";
827 case LNET_MSG_STATUS_LOCAL_DROPPED:
828 return "LOCAL_DROPPED";
829 case LNET_MSG_STATUS_LOCAL_ABORTED:
830 return "LOCAL_ABORTED";
831 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
832 return "LOCAL_NO_ROUTE";
833 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
834 return "LOCAL_TIMEOUT";
835 case LNET_MSG_STATUS_LOCAL_ERROR:
836 return "LOCAL_ERROR";
837 case LNET_MSG_STATUS_REMOTE_DROPPED:
838 return "REMOTE_DROPPED";
839 case LNET_MSG_STATUS_REMOTE_ERROR:
840 return "REMOTE_ERROR";
841 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
842 return "REMOTE_TIMEOUT";
843 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
844 return "NETWORK_TIMEOUT";
845 case LNET_MSG_STATUS_OK:
853 lnet_send_error_simulation(struct lnet_msg *msg,
854 enum lnet_msg_hstatus *hstatus)
859 if (list_empty(&the_lnet.ln_drop_rules))
862 /* match only health rules */
863 if (!lnet_drop_rule_match(&msg->msg_hdr, hstatus))
866 CDEBUG(D_NET, "src %s, dst %s: %s simulate health error: %s\n",
867 libcfs_nid2str(msg->msg_hdr.src_nid),
868 libcfs_nid2str(msg->msg_hdr.dest_nid),
869 lnet_msgtyp2str(msg->msg_type),
870 lnet_health_error2str(*hstatus));
874 EXPORT_SYMBOL(lnet_send_error_simulation);
877 lnet_finalize(struct lnet_msg *msg, int status)
879 struct lnet_msg_container *container;
886 LASSERT(!in_interrupt());
891 msg->msg_ev.status = status;
893 /* if the message is successfully sent, no need to keep the MD around */
894 if (msg->msg_md != NULL && !status)
895 lnet_detach_md(msg, status);
898 hc = lnet_is_health_check(msg);
901 * the MD would've been detached from the message if it was
902 * successfully sent. However, if it wasn't successfully sent the
903 * MD would be around. And since we recalculate whether to
904 * health check or not, it's possible that we change our minds and
905 * we don't want to health check this message. In this case also
908 * If the message is successful we're going to
909 * go through the lnet_health_check() function, but that'll just
910 * increment the appropriate health value and return.
912 if (msg->msg_md != NULL && !hc)
913 lnet_detach_md(msg, status);
916 if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
917 /* not committed to network yet */
918 LASSERT(!msg->msg_onactivelist);
925 * Check the health status of the message. If it has one
926 * of the errors that we're supposed to handle, and it has
927 * not timed out, then
928 * 1. Decrement the appropriate health_value
929 * 2. queue the message on the resend queue
931 * if the message send is success, timed out or failed in the
932 * health check for any reason then we'll just finalize the
933 * message. Otherwise just return since the message has been
934 * put on the resend queue.
936 if (!lnet_health_check(msg))
940 * if we get here then we need to clean up the md because we're
941 * finalizing the message.
943 if (msg->msg_md != NULL)
944 lnet_detach_md(msg, status);
948 * NB: routed message can be committed for both receiving and sending,
949 * we should finalize in LIFO order and keep counters correct.
950 * (finalize sending first then finalize receiving)
952 cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
955 container = the_lnet.ln_msg_containers[cpt];
956 list_add_tail(&msg->msg_list, &container->msc_finalizing);
958 /* Recursion breaker. Don't complete the message here if I am (or
959 * enough other threads are) already completing messages */
962 for (i = 0; i < container->msc_nfinalizers; i++) {
963 if (container->msc_finalizers[i] == current)
966 if (my_slot < 0 && container->msc_finalizers[i] == NULL)
970 if (i < container->msc_nfinalizers || my_slot < 0) {
971 lnet_net_unlock(cpt);
975 container->msc_finalizers[my_slot] = current;
977 while (!list_empty(&container->msc_finalizing)) {
978 msg = list_entry(container->msc_finalizing.next,
979 struct lnet_msg, msg_list);
981 list_del_init(&msg->msg_list);
983 /* NB drops and regains the lnet lock if it actually does
984 * anything, so my finalizing friends can chomp along too */
985 rc = lnet_complete_msg_locked(msg, cpt);
990 if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
991 lnet_net_unlock(cpt);
992 lnet_delay_rule_check();
996 container->msc_finalizers[my_slot] = NULL;
997 lnet_net_unlock(cpt);
1002 EXPORT_SYMBOL(lnet_finalize);
1005 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1009 if (container->msc_init == 0)
1012 while (!list_empty(&container->msc_active)) {
1013 struct lnet_msg *msg;
1015 msg = list_entry(container->msc_active.next,
1016 struct lnet_msg, msg_activelist);
1017 LASSERT(msg->msg_onactivelist);
1018 msg->msg_onactivelist = 0;
1019 list_del_init(&msg->msg_activelist);
1025 CERROR("%d active msg on exit\n", count);
1027 if (container->msc_finalizers != NULL) {
1028 LIBCFS_FREE(container->msc_finalizers,
1029 container->msc_nfinalizers *
1030 sizeof(*container->msc_finalizers));
1031 container->msc_finalizers = NULL;
1033 container->msc_init = 0;
1037 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1041 container->msc_init = 1;
1043 INIT_LIST_HEAD(&container->msc_active);
1044 INIT_LIST_HEAD(&container->msc_finalizing);
1046 /* number of CPUs */
1047 container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1048 if (container->msc_nfinalizers == 0)
1049 container->msc_nfinalizers = 1;
1051 LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1052 container->msc_nfinalizers *
1053 sizeof(*container->msc_finalizers));
1055 if (container->msc_finalizers == NULL) {
1056 CERROR("Failed to allocate message finalizers\n");
1057 lnet_msg_container_cleanup(container);
1065 lnet_msg_containers_destroy(void)
1067 struct lnet_msg_container *container;
1070 if (the_lnet.ln_msg_containers == NULL)
1073 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1074 lnet_msg_container_cleanup(container);
1076 cfs_percpt_free(the_lnet.ln_msg_containers);
1077 the_lnet.ln_msg_containers = NULL;
1081 lnet_msg_containers_create(void)
1083 struct lnet_msg_container *container;
1087 the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1088 sizeof(*container));
1090 if (the_lnet.ln_msg_containers == NULL) {
1091 CERROR("Failed to allocate cpu-partition data for network\n");
1095 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1096 rc = lnet_msg_container_setup(container, i);
1098 lnet_msg_containers_destroy();