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/
33 * Message decoding, parsing and finalizing routines
36 #define DEBUG_SUBSYSTEM S_LNET
38 #include <lnet/lib-lnet.h>
41 lnet_build_unlink_event(struct lnet_libmd *md, struct lnet_event *ev)
45 memset(ev, 0, sizeof(*ev));
49 ev->type = LNET_EVENT_UNLINK;
50 lnet_md_deconstruct(md, ev);
51 lnet_md2handle(&ev->md_handle, md);
56 * Don't need any lock, must be called after lnet_commit_md
59 lnet_build_msg_event(struct lnet_msg *msg, enum lnet_event_kind ev_type)
61 struct lnet_hdr *hdr = &msg->msg_hdr;
62 struct lnet_event *ev = &msg->msg_ev;
64 LASSERT(!msg->msg_routing);
67 ev->msg_type = msg->msg_type;
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->source.nid = LNET_NID_ANY;
76 ev->source.pid = the_lnet.ln_pid;
77 ev->sender = LNET_NID_ANY;
79 /* event for passive message */
80 ev->target.pid = hdr->dest_pid;
81 ev->target.nid = hdr->dest_nid;
82 ev->initiator.pid = hdr->src_pid;
83 /* Multi-Rail: resolve src_nid to "primary" peer NID */
84 ev->initiator.nid = lnet_nid_to_nid4(&msg->msg_initiator);
85 /* Multi-Rail: track source NID. */
86 ev->source.pid = hdr->src_pid;
87 ev->source.nid = hdr->src_nid;
88 ev->rlength = hdr->payload_length;
89 ev->sender = lnet_nid_to_nid4(&msg->msg_from);
90 ev->mlength = msg->msg_wanted;
91 ev->offset = msg->msg_offset;
98 case LNET_EVENT_PUT: /* passive PUT */
99 ev->pt_index = hdr->msg.put.ptl_index;
100 ev->match_bits = hdr->msg.put.match_bits;
101 ev->hdr_data = hdr->msg.put.hdr_data;
104 case LNET_EVENT_GET: /* passive GET */
105 ev->pt_index = hdr->msg.get.ptl_index;
106 ev->match_bits = hdr->msg.get.match_bits;
110 case LNET_EVENT_ACK: /* ACK */
111 ev->match_bits = hdr->msg.ack.match_bits;
112 ev->mlength = hdr->msg.ack.mlength;
115 case LNET_EVENT_REPLY: /* REPLY */
118 case LNET_EVENT_SEND: /* active message */
119 if (msg->msg_type == LNET_MSG_PUT) {
120 ev->pt_index = le32_to_cpu(hdr->msg.put.ptl_index);
121 ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
122 ev->offset = le32_to_cpu(hdr->msg.put.offset);
124 ev->rlength = le32_to_cpu(hdr->payload_length);
125 ev->hdr_data = le64_to_cpu(hdr->msg.put.hdr_data);
128 LASSERT(msg->msg_type == LNET_MSG_GET);
129 ev->pt_index = le32_to_cpu(hdr->msg.get.ptl_index);
130 ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
132 ev->rlength = le32_to_cpu(hdr->msg.get.sink_length);
133 ev->offset = le32_to_cpu(hdr->msg.get.src_offset);
141 lnet_msg_commit(struct lnet_msg *msg, int cpt)
143 struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
144 struct lnet_counters_common *common;
147 /* set the message deadline */
148 timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
149 msg->msg_deadline = ktime_add_ns(ktime_get(), timeout_ns);
151 /* routed message can be committed for both receiving and sending */
152 LASSERT(!msg->msg_tx_committed);
154 if (msg->msg_sending) {
155 LASSERT(!msg->msg_receiving);
156 msg->msg_tx_cpt = cpt;
157 msg->msg_tx_committed = 1;
158 if (msg->msg_rx_committed) { /* routed message REPLY */
159 LASSERT(msg->msg_onactivelist);
163 LASSERT(!msg->msg_sending);
164 msg->msg_rx_cpt = cpt;
165 msg->msg_rx_committed = 1;
168 LASSERT(!msg->msg_onactivelist);
170 msg->msg_onactivelist = 1;
171 list_add_tail(&msg->msg_activelist, &container->msc_active);
173 common = &the_lnet.ln_counters[cpt]->lct_common;
174 common->lcc_msgs_alloc++;
175 if (common->lcc_msgs_alloc > common->lcc_msgs_max)
176 common->lcc_msgs_max = common->lcc_msgs_alloc;
180 lnet_msg_decommit_tx(struct lnet_msg *msg, int status)
182 struct lnet_counters_common *common;
183 struct lnet_event *ev = &msg->msg_ev;
185 LASSERT(msg->msg_tx_committed);
189 common = &(the_lnet.ln_counters[msg->msg_tx_cpt]->lct_common);
191 default: /* routed message */
192 LASSERT(msg->msg_routing);
193 LASSERT(msg->msg_rx_committed);
194 LASSERT(ev->type == 0);
196 common->lcc_route_length += msg->msg_len;
197 common->lcc_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 common->lcc_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 common->lcc_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_common *common;
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 common = &(the_lnet.ln_counters[msg->msg_rx_cpt]->lct_common);
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 common->lcc_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 common->lcc_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 common->lcc_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]->lct_common.lcc_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);
366 lnet_complete_msg_locked(struct lnet_msg *msg, int cpt)
368 struct lnet_handle_wire ack_wmd;
370 int status = msg->msg_ev.status;
372 LASSERT(msg->msg_onactivelist);
374 if (status == 0 && msg->msg_ack) {
375 /* Only send an ACK if the PUT completed successfully */
377 lnet_msg_decommit(msg, cpt, 0);
380 lnet_net_unlock(cpt);
382 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
383 LASSERT(!msg->msg_routing);
385 ack_wmd = msg->msg_hdr.msg.put.ack_wmd;
387 lnet_prep_send(msg, LNET_MSG_ACK, msg->msg_ev.source, 0, 0);
389 msg->msg_hdr.msg.ack.dst_wmd = ack_wmd;
390 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
391 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
393 rc = lnet_send(msg->msg_ev.target.nid, msg,
394 lnet_nid_to_nid4(&msg->msg_from));
398 * NB: message is committed for sending, we should return
399 * on success because LND will finalize this message later.
401 * Also, there is possibility that message is committed for
402 * sending and also failed before delivering to LND,
403 * i.e: ENOMEM, in that case we can't fall through either
404 * because CPT for sending can be different with CPT for
405 * receiving, so we should return back to lnet_finalize()
406 * to make sure we are locking the correct partition.
410 } else if (status == 0 && /* OK so far */
411 (msg->msg_routing && !msg->msg_sending)) {
413 LASSERT(!msg->msg_receiving); /* called back recv already */
414 lnet_net_unlock(cpt);
416 rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
420 * NB: message is committed for sending, we should return
421 * on success because LND will finalize this message later.
423 * Also, there is possibility that message is committed for
424 * sending and also failed before delivering to LND,
425 * i.e: ENOMEM, in that case we can't fall through either:
426 * - The rule is message must decommit for sending first if
427 * the it's committed for both sending and receiving
428 * - CPT for sending can be different with CPT for receiving,
429 * so we should return back to lnet_finalize() to make
430 * sure we are locking the correct partition.
435 lnet_msg_decommit(msg, cpt, status);
441 lnet_dec_healthv_locked(atomic_t *healthv, int sensitivity)
443 int h = atomic_read(healthv);
445 if (h < sensitivity) {
446 atomic_set(healthv, 0);
449 atomic_set(healthv, h);
453 /* must hold net_lock/0 */
455 lnet_ni_add_to_recoveryq_locked(struct lnet_ni *ni,
456 struct list_head *recovery_queue, time64_t now)
458 if (!list_empty(&ni->ni_recovery))
461 if (atomic_read(&ni->ni_healthv) == LNET_MAX_HEALTH_VALUE)
464 /* This NI is going on the recovery queue, so take a ref on it */
465 lnet_ni_addref_locked(ni, 0);
467 lnet_ni_set_next_ping(ni, now);
469 CDEBUG(D_NET, "%s added to recovery queue. ping count: %u next ping: %lld health :%d\n",
470 libcfs_nidstr(&ni->ni_nid),
473 atomic_read(&ni->ni_healthv));
475 list_add_tail(&ni->ni_recovery, recovery_queue);
479 lnet_handle_local_failure(struct lnet_ni *local_ni)
482 * the lnet_net_lock(0) is used to protect the addref on the ni
483 * and the recovery queue.
486 /* the mt could've shutdown and cleaned up the queues */
487 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
492 lnet_dec_healthv_locked(&local_ni->ni_healthv, lnet_health_sensitivity);
493 lnet_ni_add_to_recoveryq_locked(local_ni, &the_lnet.ln_mt_localNIRecovq,
494 ktime_get_seconds());
498 /* must hold net_lock/0 */
500 lnet_handle_remote_failure_locked(struct lnet_peer_ni *lpni)
502 __u32 sensitivity = lnet_health_sensitivity;
503 __u32 lp_sensitivity;
506 * If there is a health sensitivity in the peer then use that
507 * instead of the globally set one.
509 lp_sensitivity = lpni->lpni_peer_net->lpn_peer->lp_health_sensitivity;
511 sensitivity = lp_sensitivity;
513 lnet_dec_healthv_locked(&lpni->lpni_healthv, sensitivity);
515 /* update the peer_net's health value */
516 lnet_update_peer_net_healthv(lpni);
519 * add the peer NI to the recovery queue if it's not already there
520 * and it's health value is actually below the maximum. It's
521 * possible that the sensitivity might be set to 0, and the health
522 * value will not be reduced. In this case, there is no reason to
525 lnet_peer_ni_add_to_recoveryq_locked(lpni,
526 &the_lnet.ln_mt_peerNIRecovq,
527 ktime_get_seconds());
531 lnet_handle_remote_failure(struct lnet_peer_ni *lpni)
533 /* lpni could be NULL if we're in the LOLND case */
538 /* the mt could've shutdown and cleaned up the queues */
539 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
543 lnet_handle_remote_failure_locked(lpni);
548 lnet_incr_hstats(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
549 enum lnet_msg_hstatus hstatus)
551 struct lnet_counters_health *health;
553 health = &the_lnet.ln_counters[0]->lct_health;
556 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
557 atomic_inc(&ni->ni_hstats.hlt_local_interrupt);
558 health->lch_local_interrupt_count++;
560 case LNET_MSG_STATUS_LOCAL_DROPPED:
561 atomic_inc(&ni->ni_hstats.hlt_local_dropped);
562 health->lch_local_dropped_count++;
564 case LNET_MSG_STATUS_LOCAL_ABORTED:
565 atomic_inc(&ni->ni_hstats.hlt_local_aborted);
566 health->lch_local_aborted_count++;
568 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
569 atomic_inc(&ni->ni_hstats.hlt_local_no_route);
570 health->lch_local_no_route_count++;
572 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
573 atomic_inc(&ni->ni_hstats.hlt_local_timeout);
574 health->lch_local_timeout_count++;
576 case LNET_MSG_STATUS_LOCAL_ERROR:
577 atomic_inc(&ni->ni_hstats.hlt_local_error);
578 health->lch_local_error_count++;
580 case LNET_MSG_STATUS_REMOTE_DROPPED:
582 atomic_inc(&lpni->lpni_hstats.hlt_remote_dropped);
583 health->lch_remote_dropped_count++;
585 case LNET_MSG_STATUS_REMOTE_ERROR:
587 atomic_inc(&lpni->lpni_hstats.hlt_remote_error);
588 health->lch_remote_error_count++;
590 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
592 atomic_inc(&lpni->lpni_hstats.hlt_remote_timeout);
593 health->lch_remote_timeout_count++;
595 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
597 atomic_inc(&lpni->lpni_hstats.hlt_network_timeout);
598 health->lch_network_timeout_count++;
600 case LNET_MSG_STATUS_OK:
608 lnet_resend_msg_locked(struct lnet_msg *msg)
610 msg->msg_retry_count++;
613 * remove message from the active list and reset it to prepare
614 * for a resend. Two exceptions to this
616 * 1. the router case. When a message is being routed it is
617 * committed for rx when received and committed for tx when
618 * forwarded. We don't want to remove it from the active list, since
619 * code which handles receiving expects it to remain on the active
622 * 2. The REPLY case. Reply messages use the same message
623 * structure for the GET that was received.
625 if (!msg->msg_routing && msg->msg_type != LNET_MSG_REPLY) {
626 list_del_init(&msg->msg_activelist);
627 msg->msg_onactivelist = 0;
630 * The msg_target.nid which was originally set
631 * when calling LNetGet() or LNetPut() might've
632 * been overwritten if we're routing this message.
633 * Call lnet_msg_decommit_tx() to return the credit
634 * this message consumed. The message will
635 * consume another credit when it gets resent.
637 lnet_nid4_to_nid(msg->msg_hdr.dest_nid, &msg->msg_target.nid);
638 lnet_msg_decommit_tx(msg, -EAGAIN);
639 msg->msg_sending = 0;
640 msg->msg_receiving = 0;
641 msg->msg_target_is_router = 0;
643 CDEBUG(D_NET, "%s->%s:%s:%s - queuing msg (%p) for resend\n",
644 libcfs_nid2str(msg->msg_hdr.src_nid),
645 libcfs_nid2str(msg->msg_hdr.dest_nid),
646 lnet_msgtyp2str(msg->msg_type),
647 lnet_health_error2str(msg->msg_health_status), msg);
649 list_add_tail(&msg->msg_list, the_lnet.ln_mt_resendqs[msg->msg_tx_cpt]);
651 complete(&the_lnet.ln_mt_wait_complete);
655 lnet_check_finalize_recursion_locked(struct lnet_msg *msg,
656 struct list_head *containerq,
657 int nworkers, void **workers)
662 list_add_tail(&msg->msg_list, containerq);
664 for (i = 0; i < nworkers; i++) {
665 if (workers[i] == current)
668 if (my_slot < 0 && workers[i] == NULL)
672 if (i < nworkers || my_slot < 0)
675 workers[my_slot] = current;
681 lnet_attempt_msg_resend(struct lnet_msg *msg)
683 struct lnet_msg_container *container;
687 /* we can only resend tx_committed messages */
688 LASSERT(msg->msg_tx_committed);
690 /* don't resend recovery messages */
691 if (msg->msg_recovery) {
692 CDEBUG(D_NET, "msg %s->%s is a recovery ping. retry# %d\n",
693 libcfs_nidstr(&msg->msg_from),
694 libcfs_nidstr(&msg->msg_target.nid),
695 msg->msg_retry_count);
696 return -ENOTRECOVERABLE;
700 * if we explicitly indicated we don't want to resend then just
703 if (msg->msg_no_resend) {
704 CDEBUG(D_NET, "msg %s->%s requested no resend. retry# %d\n",
705 libcfs_nidstr(&msg->msg_from),
706 libcfs_nidstr(&msg->msg_target.nid),
707 msg->msg_retry_count);
708 return -ENOTRECOVERABLE;
711 /* check if the message has exceeded the number of retries */
712 if (msg->msg_retry_count >= lnet_retry_count) {
713 CNETERR("msg %s->%s exceeded retry count %d\n",
714 libcfs_nidstr(&msg->msg_from),
715 libcfs_nidstr(&msg->msg_target.nid),
716 msg->msg_retry_count);
717 return -ENOTRECOVERABLE;
720 cpt = msg->msg_tx_cpt;
723 /* check again under lock */
724 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
725 lnet_net_unlock(cpt);
729 container = the_lnet.ln_msg_containers[cpt];
731 lnet_check_finalize_recursion_locked(msg,
732 &container->msc_resending,
733 container->msc_nfinalizers,
734 container->msc_resenders);
736 /* enough threads are resending */
738 lnet_net_unlock(cpt);
742 while (!list_empty(&container->msc_resending)) {
743 msg = list_entry(container->msc_resending.next,
744 struct lnet_msg, msg_list);
745 list_del(&msg->msg_list);
748 * resending the message will require us to call
749 * lnet_msg_decommit_tx() which will return the credit
750 * which this message holds. This could trigger another
751 * queued message to be sent. If that message fails and
752 * requires a resend we will recurse.
753 * But since at this point the slot is taken, the message
754 * will be queued in the container and dealt with
755 * later. This breaks the recursion.
757 lnet_resend_msg_locked(msg);
761 * msc_resenders is an array of process pointers. Each entry holds
762 * a pointer to the current process operating on the message. An
763 * array entry is created per CPT. If the array slot is already
764 * set, then it means that there is a thread on the CPT currently
765 * resending a message.
766 * Once the thread finishes clear the slot to enable the thread to
767 * take on more resend work.
769 container->msc_resenders[my_slot] = NULL;
770 lnet_net_unlock(cpt);
776 * Do a health check on the message:
777 * return -1 if we're not going to handle the error or
778 * if we've reached the maximum number of retries.
779 * success case will return -1 as well
780 * return 0 if it the message is requeued for send
783 lnet_health_check(struct lnet_msg *msg)
785 enum lnet_msg_hstatus hstatus = msg->msg_health_status;
786 struct lnet_peer_ni *lpni;
789 bool attempt_local_resend;
790 bool attempt_remote_resend;
791 bool handle_local_health;
792 bool handle_remote_health;
794 /* if we're shutting down no point in handling health. */
795 if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
798 LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
801 * if we're sending to the LOLND then the msg_txpeer will not be
802 * set. So no need to sanity check it.
804 if (msg->msg_tx_committed &&
805 !nid_is_lo0(&msg->msg_txni->ni_nid))
806 LASSERT(msg->msg_txpeer);
807 else if (msg->msg_tx_committed &&
808 nid_is_lo0(&msg->msg_txni->ni_nid))
811 if (hstatus != LNET_MSG_STATUS_OK &&
812 ktime_compare(ktime_get(), msg->msg_deadline) >= 0)
816 * always prefer txni/txpeer if they message is committed for both
819 if (msg->msg_tx_committed) {
821 lpni = msg->msg_txpeer;
822 attempt_local_resend = attempt_remote_resend = true;
825 lpni = msg->msg_rxpeer;
826 attempt_local_resend = attempt_remote_resend = false;
834 CDEBUG(D_NET, "health check: %s->%s: %s: %s\n",
835 libcfs_nidstr(&ni->ni_nid),
836 (lo) ? "self" : libcfs_nidstr(&lpni->lpni_nid),
837 lnet_msgtyp2str(msg->msg_type),
838 lnet_health_error2str(hstatus));
841 * stats are only incremented for errors so avoid wasting time
842 * incrementing statistics if there is no error. Similarly, whether to
843 * update health values or perform resends is only applicable for
844 * messages with a health status != OK.
846 if (hstatus != LNET_MSG_STATUS_OK) {
847 /* Don't further decrement the health value if a recovery
850 if (msg->msg_recovery)
851 handle_local_health = handle_remote_health = false;
853 handle_local_health = handle_remote_health = true;
855 /* For local failures, health/recovery/resends are not needed if
856 * I only have a single (non-lolnd) interface. NB: pb_nnis
857 * includes the lolnd interface, so a single-rail node would
860 if (the_lnet.ln_ping_target->pb_nnis <= 2) {
861 handle_local_health = false;
862 attempt_local_resend = false;
866 lnet_incr_hstats(ni, lpni, hstatus);
867 /* For remote failures, health/recovery/resends are not needed
868 * if the peer only has a single interface. Special case for
869 * routers where we rely on health feature to manage route
870 * aliveness. NB: unlike pb_nnis above, lp_nnis does _not_
871 * include the lolnd, so a single-rail node would have
874 if (lpni && lpni->lpni_peer_net &&
875 lpni->lpni_peer_net->lpn_peer &&
876 lpni->lpni_peer_net->lpn_peer->lp_nnis <= 1) {
877 attempt_remote_resend = false;
878 if (!lnet_isrouter(lpni))
879 handle_remote_health = false;
885 case LNET_MSG_STATUS_OK:
887 * increment the local ni health whether we successfully
888 * received or sent a message on it.
890 * Ping counts are reset to 0 as appropriate to allow for
893 lnet_inc_healthv(&ni->ni_healthv, lnet_health_sensitivity);
895 * It's possible msg_txpeer is NULL in the LOLND
896 * case. Only increment the peer's health if we're
897 * receiving a message from it. It's the only sure way to
898 * know that a remote interface is up.
899 * If this interface is part of a router, then take that
900 * as indication that the router is fully healthy.
902 if (lpni && msg->msg_rx_committed) {
904 lpni->lpni_ping_count = 0;
905 ni->ni_ping_count = 0;
907 * If we're receiving a message from the router or
908 * I'm a router, then set that lpni's health to
909 * maximum so we can commence communication
911 if (lnet_isrouter(lpni) || the_lnet.ln_routing) {
912 lnet_set_lpni_healthv_locked(lpni,
913 LNET_MAX_HEALTH_VALUE);
915 __u32 sensitivity = lpni->lpni_peer_net->
916 lpn_peer->lp_health_sensitivity;
918 lnet_inc_lpni_healthv_locked(lpni,
919 (sensitivity) ? sensitivity :
920 lnet_health_sensitivity);
921 /* This peer NI may have previously aged out
922 * of recovery. Now that we've received a
923 * message from it, we can continue recovery
924 * if its health value is still below the
927 lnet_peer_ni_add_to_recoveryq_locked(lpni,
928 &the_lnet.ln_mt_peerNIRecovq,
929 ktime_get_seconds());
934 /* we can finalize this message */
936 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
937 case LNET_MSG_STATUS_LOCAL_DROPPED:
938 case LNET_MSG_STATUS_LOCAL_ABORTED:
939 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
940 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
941 if (handle_local_health)
942 lnet_handle_local_failure(ni);
943 if (attempt_local_resend)
944 return lnet_attempt_msg_resend(msg);
946 case LNET_MSG_STATUS_LOCAL_ERROR:
947 if (handle_local_health)
948 lnet_handle_local_failure(ni);
950 case LNET_MSG_STATUS_REMOTE_DROPPED:
951 if (handle_remote_health)
952 lnet_handle_remote_failure(lpni);
953 if (attempt_remote_resend)
954 return lnet_attempt_msg_resend(msg);
956 case LNET_MSG_STATUS_REMOTE_ERROR:
957 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
958 if (handle_remote_health)
959 lnet_handle_remote_failure(lpni);
961 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
962 if (handle_remote_health)
963 lnet_handle_remote_failure(lpni);
964 if (handle_local_health)
965 lnet_handle_local_failure(ni);
971 /* no resend is needed */
976 lnet_msg_detach_md(struct lnet_msg *msg, int status)
978 struct lnet_libmd *md = msg->msg_md;
979 lnet_handler_t handler = NULL;
980 int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
984 while (md->md_flags & LNET_MD_FLAG_HANDLING)
985 /* An event handler is running - wait for it to
986 * complete to avoid races.
988 lnet_md_wait_handling(md, cpt);
990 /* Now it's safe to drop my caller's ref */
992 LASSERT(md->md_refcount >= 0);
994 unlink = lnet_md_unlinkable(md);
995 if (md->md_handler) {
996 if ((md->md_flags & LNET_MD_FLAG_ABORTED) && !status) {
997 msg->msg_ev.status = -ETIMEDOUT;
998 CDEBUG(D_NET, "md 0x%p already unlinked\n", md);
1000 msg->msg_ev.status = status;
1002 msg->msg_ev.unlinked = unlink;
1003 handler = md->md_handler;
1005 md->md_flags |= LNET_MD_FLAG_HANDLING;
1008 if (unlink || (md->md_refcount == 0 &&
1009 md->md_threshold == LNET_MD_THRESH_INF))
1010 lnet_detach_rsp_tracker(md, cpt);
1016 lnet_res_unlock(cpt);
1019 handler(&msg->msg_ev);
1022 md->md_flags &= ~LNET_MD_FLAG_HANDLING;
1024 lnet_res_unlock(cpt);
1030 lnet_is_health_check(struct lnet_msg *msg)
1033 int status = msg->msg_ev.status;
1035 if ((!msg->msg_tx_committed && !msg->msg_rx_committed) ||
1036 !msg->msg_onactivelist) {
1037 CDEBUG(D_NET, "msg %p not committed for send or receive\n",
1042 if ((msg->msg_tx_committed && !msg->msg_txpeer) ||
1043 (msg->msg_rx_committed && !msg->msg_rxpeer)) {
1044 /* The optimized GET case does not set msg_rxpeer, but status
1045 * could be zero. Only print the error message if we have a
1049 CDEBUG(D_NET, "msg %p status %d cannot retry\n", msg,
1054 /* Check for status inconsistencies */
1055 if ((!status && msg->msg_health_status != LNET_MSG_STATUS_OK) ||
1056 (status && msg->msg_health_status == LNET_MSG_STATUS_OK)) {
1057 CDEBUG(D_NET, "Msg %p is in inconsistent state, don't perform health "
1058 "checking (%d, %d)\n", msg, status,
1059 msg->msg_health_status);
1063 CDEBUG(D_NET, "health check = %d, status = %d, hstatus = %d\n",
1064 hc, status, msg->msg_health_status);
1070 lnet_health_error2str(enum lnet_msg_hstatus hstatus)
1073 case LNET_MSG_STATUS_LOCAL_INTERRUPT:
1074 return "LOCAL_INTERRUPT";
1075 case LNET_MSG_STATUS_LOCAL_DROPPED:
1076 return "LOCAL_DROPPED";
1077 case LNET_MSG_STATUS_LOCAL_ABORTED:
1078 return "LOCAL_ABORTED";
1079 case LNET_MSG_STATUS_LOCAL_NO_ROUTE:
1080 return "LOCAL_NO_ROUTE";
1081 case LNET_MSG_STATUS_LOCAL_TIMEOUT:
1082 return "LOCAL_TIMEOUT";
1083 case LNET_MSG_STATUS_LOCAL_ERROR:
1084 return "LOCAL_ERROR";
1085 case LNET_MSG_STATUS_REMOTE_DROPPED:
1086 return "REMOTE_DROPPED";
1087 case LNET_MSG_STATUS_REMOTE_ERROR:
1088 return "REMOTE_ERROR";
1089 case LNET_MSG_STATUS_REMOTE_TIMEOUT:
1090 return "REMOTE_TIMEOUT";
1091 case LNET_MSG_STATUS_NETWORK_TIMEOUT:
1092 return "NETWORK_TIMEOUT";
1093 case LNET_MSG_STATUS_OK:
1101 lnet_send_error_simulation(struct lnet_msg *msg,
1102 enum lnet_msg_hstatus *hstatus)
1107 if (list_empty(&the_lnet.ln_drop_rules))
1110 /* match only health rules */
1111 if (!lnet_drop_rule_match(&msg->msg_hdr, LNET_NID_ANY,
1115 CDEBUG(D_NET, "src %s(%s)->dst %s: %s simulate health error: %s\n",
1116 libcfs_nid2str(msg->msg_hdr.src_nid),
1117 libcfs_nidstr(&msg->msg_txni->ni_nid),
1118 libcfs_nid2str(msg->msg_hdr.dest_nid),
1119 lnet_msgtyp2str(msg->msg_type),
1120 lnet_health_error2str(*hstatus));
1124 EXPORT_SYMBOL(lnet_send_error_simulation);
1127 lnet_finalize(struct lnet_msg *msg, int status)
1129 struct lnet_msg_container *container;
1134 LASSERT(!in_interrupt());
1139 msg->msg_ev.status = status;
1141 if (lnet_is_health_check(msg)) {
1143 * Check the health status of the message. If it has one
1144 * of the errors that we're supposed to handle, and it has
1145 * not timed out, then
1146 * 1. Decrement the appropriate health_value
1147 * 2. queue the message on the resend queue
1149 * if the message send is success, timed out or failed in the
1150 * health check for any reason then we'll just finalize the
1151 * message. Otherwise just return since the message has been
1152 * put on the resend queue.
1154 if (!lnet_health_check(msg))
1159 * We're not going to resend this message so detach its MD and invoke
1160 * the appropriate callbacks
1162 if (msg->msg_md != NULL)
1163 lnet_msg_detach_md(msg, status);
1166 if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
1167 /* not committed to network yet */
1168 LASSERT(!msg->msg_onactivelist);
1174 * NB: routed message can be committed for both receiving and sending,
1175 * we should finalize in LIFO order and keep counters correct.
1176 * (finalize sending first then finalize receiving)
1178 cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
1181 container = the_lnet.ln_msg_containers[cpt];
1183 /* Recursion breaker. Don't complete the message here if I am (or
1184 * enough other threads are) already completing messages */
1185 my_slot = lnet_check_finalize_recursion_locked(msg,
1186 &container->msc_finalizing,
1187 container->msc_nfinalizers,
1188 container->msc_finalizers);
1190 /* enough threads are resending */
1191 if (my_slot == -1) {
1192 lnet_net_unlock(cpt);
1197 while (!list_empty(&container->msc_finalizing)) {
1198 msg = list_entry(container->msc_finalizing.next,
1199 struct lnet_msg, msg_list);
1201 list_del_init(&msg->msg_list);
1203 /* NB drops and regains the lnet lock if it actually does
1204 * anything, so my finalizing friends can chomp along too */
1205 rc = lnet_complete_msg_locked(msg, cpt);
1210 if (unlikely(!list_empty(&the_lnet.ln_delay_rules))) {
1211 lnet_net_unlock(cpt);
1212 lnet_delay_rule_check();
1216 container->msc_finalizers[my_slot] = NULL;
1217 lnet_net_unlock(cpt);
1222 EXPORT_SYMBOL(lnet_finalize);
1225 lnet_msg_container_cleanup(struct lnet_msg_container *container)
1229 if (container->msc_init == 0)
1232 while (!list_empty(&container->msc_active)) {
1233 struct lnet_msg *msg;
1235 msg = list_entry(container->msc_active.next,
1236 struct lnet_msg, msg_activelist);
1237 LASSERT(msg->msg_onactivelist);
1238 msg->msg_onactivelist = 0;
1239 list_del_init(&msg->msg_activelist);
1245 CERROR("%d active msg on exit\n", count);
1247 if (container->msc_finalizers != NULL) {
1248 CFS_FREE_PTR_ARRAY(container->msc_finalizers,
1249 container->msc_nfinalizers);
1250 container->msc_finalizers = NULL;
1253 if (container->msc_resenders != NULL) {
1254 CFS_FREE_PTR_ARRAY(container->msc_resenders,
1255 container->msc_nfinalizers);
1256 container->msc_resenders = NULL;
1258 container->msc_init = 0;
1262 lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
1266 container->msc_init = 1;
1268 INIT_LIST_HEAD(&container->msc_active);
1269 INIT_LIST_HEAD(&container->msc_finalizing);
1270 INIT_LIST_HEAD(&container->msc_resending);
1272 /* number of CPUs */
1273 container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
1274 if (container->msc_nfinalizers == 0)
1275 container->msc_nfinalizers = 1;
1277 LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
1278 container->msc_nfinalizers *
1279 sizeof(*container->msc_finalizers));
1281 if (container->msc_finalizers == NULL) {
1282 CERROR("Failed to allocate message finalizers\n");
1283 lnet_msg_container_cleanup(container);
1287 LIBCFS_CPT_ALLOC(container->msc_resenders, lnet_cpt_table(), cpt,
1288 container->msc_nfinalizers *
1289 sizeof(*container->msc_resenders));
1291 if (container->msc_resenders == NULL) {
1292 CERROR("Failed to allocate message resenders\n");
1293 lnet_msg_container_cleanup(container);
1301 lnet_msg_containers_destroy(void)
1303 struct lnet_msg_container *container;
1306 if (the_lnet.ln_msg_containers == NULL)
1309 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
1310 lnet_msg_container_cleanup(container);
1312 cfs_percpt_free(the_lnet.ln_msg_containers);
1313 the_lnet.ln_msg_containers = NULL;
1317 lnet_msg_containers_create(void)
1319 struct lnet_msg_container *container;
1323 the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
1324 sizeof(*container));
1326 if (the_lnet.ln_msg_containers == NULL) {
1327 CERROR("Failed to allocate cpu-partition data for network\n");
1331 cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
1332 rc = lnet_msg_container_setup(container, i);
1334 lnet_msg_containers_destroy();