1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/pinger.c
38 * Portal-RPC reconnection and replay operations, for use in recovery.
42 #include <liblustre.h>
44 #define DEBUG_SUBSYSTEM S_RPC
47 #include <obd_support.h>
48 #include <obd_class.h>
49 #include "ptlrpc_internal.h"
51 struct semaphore pinger_sem;
52 static struct list_head pinger_imports = CFS_LIST_HEAD_INIT(pinger_imports);
53 static struct list_head timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
55 struct ptlrpc_request *
56 ptlrpc_prep_ping(struct obd_import *imp)
58 struct ptlrpc_request *req;
60 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION,
61 OBD_PING, 1, NULL, NULL);
63 ptlrpc_req_set_repsize(req, 1, NULL);
64 req->rq_no_resend = req->rq_no_delay = 1;
69 int ptlrpc_obd_ping(struct obd_device *obd)
72 struct ptlrpc_request *req;
75 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
79 req->rq_send_state = LUSTRE_IMP_FULL;
81 rc = ptlrpc_queue_wait(req);
83 ptlrpc_req_finished(req);
87 EXPORT_SYMBOL(ptlrpc_obd_ping);
89 int ptlrpc_ping(struct obd_import *imp)
91 struct ptlrpc_request *req;
95 req = ptlrpc_prep_ping(imp);
97 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
98 imp->imp_obd->obd_uuid.uuid,
99 obd2cli_tgt(imp->imp_obd));
101 /* To quickly detect server failure ping timeouts must be
102 * kept small. Therefore we must override/ignore the server
103 * rpc completion estimate which may be very large since
104 * it includes non-ping service times. The right long term
105 * fix will be to add a per-server (not per-service) thread
106 * in order to reduce the number of pings in the system in
107 * general (see bug 12471). */
109 req->rq_timeout = PING_SVC_TIMEOUT +
110 at_get(&imp->imp_at.iat_net_latency);
111 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
114 ptlrpcd_add_req(req);
116 CERROR("OOM trying to ping %s->%s\n",
117 imp->imp_obd->obd_uuid.uuid,
118 obd2cli_tgt(imp->imp_obd));
124 EXPORT_SYMBOL(ptlrpc_ping);
126 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
129 cfs_time_t delay, dtime, ctime = cfs_time_current();
131 if (imp->imp_state == LUSTRE_IMP_DISCON ||
132 imp->imp_state == LUSTRE_IMP_CONNECTING) {
133 /* In the disconnected case aggressively reconnect, for
134 * this request the AT service timeout will be set to
135 * INITIAL_CONNECT_TIMEOUT. To ensure the request times
136 * out before we send another we add one extra second. */
137 dtime = cfs_time_seconds(max_t(int, CONNECTION_SWITCH_MIN,
138 AT_OFF ? 0 : INITIAL_CONNECT_TIMEOUT + 1 +
139 at_get(&imp->imp_at.iat_net_latency)));
141 /* In the common case we want to cluster the pings at
142 * at regular intervals to minimize system noise. */
143 delay = cfs_time_seconds(soon ? PING_INTERVAL_SHORT :
145 dtime = delay - (ctime % delay);
147 /* May harmlessly race with ptlrpc_update_next_ping() */
148 imp->imp_next_ping = cfs_time_add(ctime, dtime);
150 CDEBUG(D_INFO, "Setting %s next ping to "CFS_TIME_T" ("CFS_TIME_T")\n",
151 obd2cli_tgt(imp->imp_obd), imp->imp_next_ping, dtime);
153 #endif /* ENABLE_PINGER */
156 void ptlrpc_ping_import_soon(struct obd_import *imp)
158 imp->imp_next_ping = cfs_time_current();
161 static inline int imp_is_deactive(struct obd_import *imp)
163 return (imp->imp_deactive ||
164 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
167 cfs_duration_t pinger_check_timeout(cfs_time_t time)
169 struct timeout_item *item;
170 cfs_time_t timeout = PING_INTERVAL;
172 /* The timeout list is a increase order sorted list */
173 mutex_down(&pinger_sem);
174 list_for_each_entry(item, &timeout_list, ti_chain) {
175 int ti_timeout = item->ti_timeout;
176 if (timeout > ti_timeout)
177 timeout = ti_timeout;
180 mutex_up(&pinger_sem);
182 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
187 static int ptlrpc_pinger_main(void *arg)
189 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
190 struct ptlrpc_thread *thread = data->thread;
193 cfs_daemonize(data->name);
195 /* Record that the thread is running */
196 thread->t_flags = SVC_RUNNING;
197 cfs_waitq_signal(&thread->t_ctl_waitq);
199 /* And now, loop forever, pinging as needed. */
201 cfs_time_t this_ping = cfs_time_current();
202 struct l_wait_info lwi;
203 cfs_duration_t time_to_next_wake;
204 cfs_time_t time_of_next_wake;
205 struct timeout_item *item;
206 struct list_head *iter;
208 time_to_next_wake = cfs_time_seconds(PING_INTERVAL);
209 time_of_next_wake = cfs_time_shift(PING_INTERVAL);
211 mutex_down(&pinger_sem);
212 list_for_each_entry(item, &timeout_list, ti_chain) {
213 item->ti_cb(item, item->ti_cb_data);
215 list_for_each(iter, &pinger_imports) {
216 struct obd_import *imp =
217 list_entry(iter, struct obd_import,
221 spin_lock(&imp->imp_lock);
222 level = imp->imp_state;
223 force = imp->imp_force_verify;
224 imp->imp_force_verify = 0;
225 spin_unlock(&imp->imp_lock);
227 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_RPCTRACE,
228 "level %s/%u force %u deactive %u pingable %u\n",
229 ptlrpc_import_state_name(level), level,
230 force, imp->imp_deactive, imp->imp_pingable);
232 /* Include any ping which misses the deadline by up to
233 * 1/10 of a second. The pings are designed to clump
234 * and this helps ensure the entire batch gets sent
235 * promptly, which minimizes system noise from pings */
238 cfs_time_aftereq(this_ping, imp->imp_next_ping -
239 (cfs_time_seconds(1) + 9) / 10)) {
240 if (level == LUSTRE_IMP_DISCON &&
241 !imp_is_deactive(imp)) {
242 ptlrpc_update_next_ping(imp, 0);
243 ptlrpc_initiate_recovery(imp);
244 } else if (level != LUSTRE_IMP_FULL ||
245 imp->imp_obd->obd_no_recov ||
246 imp_is_deactive(imp)) {
247 CDEBUG(D_HA, "not pinging %s "
248 "(in recovery: %s or recovery "
249 "disabled: %u/%u)\n",
250 obd2cli_tgt(imp->imp_obd),
251 ptlrpc_import_state_name(level),
253 imp->imp_obd->obd_no_recov);
254 } else if (imp->imp_pingable || force) {
256 /* ptlrpc_pinger_sending_on_import()
257 * will asynch update imp_next_ping
258 * so it must not be used below to
259 * calculate minimum wait time. */
263 if (!imp->imp_pingable)
266 "don't need to ping %s ("CFS_TIME_T
267 " > "CFS_TIME_T")\n",
268 obd2cli_tgt(imp->imp_obd),
269 imp->imp_next_ping, this_ping);
272 /* Wait time until next ping, or until we stopped. */
273 if (cfs_time_before(imp->imp_next_ping,
274 time_of_next_wake)) {
275 time_of_next_wake = imp->imp_next_ping;
276 time_to_next_wake = max_t(cfs_duration_t,
278 cfs_time_sub(time_of_next_wake,
279 cfs_time_current()));
282 mutex_up(&pinger_sem);
283 obd_update_maxusage();
284 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
285 time_to_next_wake, time_of_next_wake);
287 if (time_to_next_wake > 0) {
288 lwi = LWI_TIMEOUT(time_to_next_wake, NULL, NULL);
289 l_wait_event(thread->t_ctl_waitq,
290 thread->t_flags & (SVC_STOPPING|SVC_EVENT),
292 if (thread->t_flags & SVC_STOPPING) {
293 thread->t_flags &= ~SVC_STOPPING;
296 } else if (thread->t_flags & SVC_EVENT) {
297 /* woken after adding import to reset timer */
298 thread->t_flags &= ~SVC_EVENT;
303 thread->t_flags = SVC_STOPPED;
304 cfs_waitq_signal(&thread->t_ctl_waitq);
306 CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
310 static struct ptlrpc_thread *pinger_thread = NULL;
312 int ptlrpc_start_pinger(void)
314 struct l_wait_info lwi = { 0 };
315 struct ptlrpc_svc_data d;
317 #ifndef ENABLE_PINGER
322 if (pinger_thread != NULL)
325 OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
326 if (pinger_thread == NULL)
328 cfs_waitq_init(&pinger_thread->t_ctl_waitq);
331 d.thread = pinger_thread;
333 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
334 * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
335 rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
337 CERROR("cannot start thread: %d\n", rc);
338 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
339 pinger_thread = NULL;
342 l_wait_event(pinger_thread->t_ctl_waitq,
343 pinger_thread->t_flags & SVC_RUNNING, &lwi);
348 int ptlrpc_pinger_remove_timeouts(void);
350 int ptlrpc_stop_pinger(void)
352 struct l_wait_info lwi = { 0 };
354 #ifndef ENABLE_PINGER
359 if (pinger_thread == NULL)
362 ptlrpc_pinger_remove_timeouts();
363 mutex_down(&pinger_sem);
364 pinger_thread->t_flags = SVC_STOPPING;
365 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
366 mutex_up(&pinger_sem);
368 l_wait_event(pinger_thread->t_ctl_waitq,
369 (pinger_thread->t_flags & SVC_STOPPED), &lwi);
371 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
372 pinger_thread = NULL;
376 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
378 ptlrpc_update_next_ping(imp, 0);
381 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
383 ptlrpc_update_next_ping(imp, 1);
386 int ptlrpc_pinger_add_import(struct obd_import *imp)
389 if (!list_empty(&imp->imp_pinger_chain))
392 mutex_down(&pinger_sem);
393 CDEBUG(D_HA, "adding pingable import %s->%s\n",
394 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
395 /* if we add to pinger we want recovery on this import */
396 imp->imp_obd->obd_no_recov = 0;
398 ptlrpc_update_next_ping(imp, 0);
399 /* XXX sort, blah blah */
400 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
401 class_import_get(imp);
403 ptlrpc_pinger_wake_up();
404 mutex_up(&pinger_sem);
409 int ptlrpc_pinger_del_import(struct obd_import *imp)
412 if (list_empty(&imp->imp_pinger_chain))
415 mutex_down(&pinger_sem);
416 list_del_init(&imp->imp_pinger_chain);
417 CDEBUG(D_HA, "removing pingable import %s->%s\n",
418 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
419 /* if we remove from pinger we don't want recovery on this import */
420 imp->imp_obd->obd_no_recov = 1;
421 class_import_put(imp);
422 mutex_up(&pinger_sem);
427 * Register a timeout callback to the pinger list, and the callback will
428 * be called when timeout happens.
430 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
431 timeout_cb_t cb, void *data)
433 struct timeout_item *ti;
439 CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
440 CFS_INIT_LIST_HEAD(&ti->ti_chain);
441 ti->ti_timeout = time;
442 ti->ti_event = event;
444 ti->ti_cb_data = data;
450 * Register timeout event on the the pinger thread.
451 * Note: the timeout list is an sorted list with increased timeout value.
453 static struct timeout_item*
454 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
455 timeout_cb_t cb, void *data)
457 struct timeout_item *item, *tmp;
459 LASSERT_SEM_LOCKED(&pinger_sem);
461 list_for_each_entry(item, &timeout_list, ti_chain)
462 if (item->ti_event == event)
465 item = ptlrpc_new_timeout(time, event, cb, data);
467 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
468 if (tmp->ti_timeout < time) {
469 list_add(&item->ti_chain, &tmp->ti_chain);
473 list_add(&item->ti_chain, &timeout_list);
479 /* Add a client_obd to the timeout event list, when timeout(@time)
480 * happens, the callback(@cb) will be called.
482 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
483 timeout_cb_t cb, void *data,
484 struct list_head *obd_list)
486 struct timeout_item *ti;
488 mutex_down(&pinger_sem);
489 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
491 mutex_up(&pinger_sem);
494 list_add(obd_list, &ti->ti_obd_list);
495 mutex_up(&pinger_sem);
499 int ptlrpc_del_timeout_client(struct list_head *obd_list,
500 enum timeout_event event)
502 struct timeout_item *ti = NULL, *item;
504 if (list_empty(obd_list))
506 mutex_down(&pinger_sem);
507 list_del_init(obd_list);
509 * If there are no obd attached to the timeout event
510 * list, remove this timeout event from the pinger
512 list_for_each_entry(item, &timeout_list, ti_chain) {
513 if (item->ti_event == event) {
518 LASSERTF(ti != NULL, "ti is NULL ! \n");
519 if (list_empty(&ti->ti_obd_list)) {
520 list_del(&ti->ti_chain);
523 mutex_up(&pinger_sem);
527 int ptlrpc_pinger_remove_timeouts(void)
529 struct timeout_item *item, *tmp;
531 mutex_down(&pinger_sem);
532 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
533 LASSERT(list_empty(&item->ti_obd_list));
534 list_del(&item->ti_chain);
537 mutex_up(&pinger_sem);
541 void ptlrpc_pinger_wake_up()
544 pinger_thread->t_flags |= SVC_EVENT;
545 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
549 /* Ping evictor thread */
551 #define PET_TERMINATE 2
553 static int pet_refcount = 0;
554 static int pet_state;
555 static wait_queue_head_t pet_waitq;
556 static struct obd_export *pet_exp = NULL;
557 static spinlock_t pet_lock = SPIN_LOCK_UNLOCKED;
559 int ping_evictor_wake(struct obd_export *exp)
561 spin_lock(&pet_lock);
562 if (pet_exp || (pet_state != PET_READY)) {
563 /* eventually the new obd will call here again. */
564 spin_unlock(&pet_lock);
568 /* We have to make sure the obd isn't destroyed between now and when
569 * the ping evictor runs. We'll take a reference here, and drop it
570 * when we finish in the evictor. We don't really care about this
571 * export in particular; we just need one to keep the obd alive. */
572 pet_exp = class_export_get(exp);
573 spin_unlock(&pet_lock);
579 static int ping_evictor_main(void *arg)
581 struct obd_device *obd;
582 struct obd_export *exp;
583 struct l_wait_info lwi = { 0 };
587 cfs_daemonize_ctxt("ll_evictor");
589 CDEBUG(D_HA, "Starting Ping Evictor\n");
591 pet_state = PET_READY;
593 l_wait_event(pet_waitq, pet_exp ||
594 (pet_state == PET_TERMINATE), &lwi);
595 if (pet_state == PET_TERMINATE)
598 /* we only get here if pet_exp != NULL, and the end of this
599 * loop is the only place which sets it NULL again, so lock
600 * is not strictly necessary. */
601 spin_lock(&pet_lock);
602 obd = pet_exp->exp_obd;
603 spin_unlock(&pet_lock);
605 /* bug 18948: ensure recovery is aborted in a timely fashion */
606 if (target_recovery_check_and_stop(obd) ||
607 obd->obd_recovering /* no evictor during recovery */)
610 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
612 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
613 obd->obd_name, expire_time);
615 /* Exports can't be deleted out of the list while we hold
616 * the obd lock (class_unlink_export), which means we can't
617 * lose the last ref on the export. If they've already been
618 * removed from the list, we won't find them here. */
619 spin_lock(&obd->obd_dev_lock);
620 while (!list_empty(&obd->obd_exports_timed)) {
621 exp = list_entry(obd->obd_exports_timed.next,
622 struct obd_export,exp_obd_chain_timed);
623 if (expire_time > exp->exp_last_request_time) {
624 class_export_get(exp);
625 spin_unlock(&obd->obd_dev_lock);
626 LCONSOLE_WARN("%s: haven't heard from client %s"
627 " (at %s) in %ld seconds. I think"
628 " it's dead, and I am evicting"
629 " it.\n", obd->obd_name,
630 obd_uuid2str(&exp->exp_client_uuid),
631 obd_export_nid2str(exp),
632 (long)(cfs_time_current_sec() -
633 exp->exp_last_request_time));
634 CDEBUG(D_HA, "Last request was at %ld\n",
635 exp->exp_last_request_time);
636 class_fail_export(exp);
637 class_export_put(exp);
638 spin_lock(&obd->obd_dev_lock);
640 /* List is sorted, so everyone below is ok */
644 spin_unlock(&obd->obd_dev_lock);
646 class_export_put(pet_exp);
648 spin_lock(&pet_lock);
650 spin_unlock(&pet_lock);
652 CDEBUG(D_HA, "Exiting Ping Evictor\n");
657 void ping_evictor_start(void)
661 if (++pet_refcount > 1)
664 init_waitqueue_head(&pet_waitq);
666 rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
669 CERROR("Cannot start ping evictor thread: %d\n", rc);
672 EXPORT_SYMBOL(ping_evictor_start);
674 void ping_evictor_stop(void)
676 if (--pet_refcount > 0)
679 pet_state = PET_TERMINATE;
682 EXPORT_SYMBOL(ping_evictor_stop);
683 #else /* !__KERNEL__ */
686 * the current implementation of pinger in liblustre is not optimized
690 static struct pinger_data {
692 cfs_time_t pd_this_ping; /* jiffies */
693 cfs_time_t pd_next_ping; /* jiffies */
694 struct ptlrpc_request_set *pd_set;
697 static int pinger_check_rpcs(void *arg)
699 cfs_time_t curtime = cfs_time_current();
700 struct ptlrpc_request *req;
701 struct ptlrpc_request_set *set;
702 struct list_head *iter;
703 struct obd_import *imp;
704 struct pinger_data *pd = &pinger_args;
707 /* prevent recursion */
708 if (pd->pd_recursion++) {
709 CDEBUG(D_HA, "pinger: recursion! quit\n");
715 /* have we reached ping point? */
716 if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
721 /* if we have rpc_set already, continue processing it */
723 LASSERT(pd->pd_this_ping);
728 pd->pd_this_ping = curtime;
729 pd->pd_set = ptlrpc_prep_set();
730 if (pd->pd_set == NULL)
734 /* add rpcs into set */
735 mutex_down(&pinger_sem);
736 list_for_each(iter, &pinger_imports) {
737 struct obd_import *imp =
738 list_entry(iter, struct obd_import, imp_pinger_chain);
739 int generation, level;
741 /* Include any ping within 1/10 of a second of the deadline */
742 if (cfs_time_aftereq(pd->pd_this_ping, imp->imp_next_ping -
743 (cfs_time_seconds(1) + 9) / 10)) {
745 spin_lock(&imp->imp_lock);
746 generation = imp->imp_generation;
747 level = imp->imp_state;
748 spin_unlock(&imp->imp_lock);
750 if (level != LUSTRE_IMP_FULL) {
752 "not pinging %s (in recovery)\n",
753 obd2cli_tgt(imp->imp_obd));
757 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING,
760 CERROR("out of memory\n");
763 req->rq_no_resend = 1;
764 ptlrpc_req_set_repsize(req, 1, NULL);
765 req->rq_send_state = LUSTRE_IMP_FULL;
766 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
767 req->rq_import_generation = generation;
768 ptlrpc_set_add_req(set, req);
770 CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
771 " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
772 imp->imp_next_ping, pd->pd_this_ping);
775 pd->pd_this_ping = curtime;
776 mutex_up(&pinger_sem);
778 /* Might be empty, that's OK. */
779 if (set->set_remaining == 0)
780 CDEBUG(D_RPCTRACE, "nothing to ping\n");
782 list_for_each(iter, &set->set_requests) {
783 struct ptlrpc_request *req =
784 list_entry(iter, struct ptlrpc_request,
786 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
787 req->rq_import->imp_obd->obd_uuid.uuid,
788 obd2cli_tgt(req->rq_import->imp_obd));
789 (void)ptl_send_rpc(req, 0);
793 rc = ptlrpc_check_set(set);
795 /* not finished, and we are not expired, simply return */
796 if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
797 cfs_time_seconds(PING_INTERVAL)))) {
798 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
803 /* Expire all the requests that didn't come back. */
804 mutex_down(&pinger_sem);
805 list_for_each(iter, &set->set_requests) {
806 req = list_entry(iter, struct ptlrpc_request,
809 if (req->rq_phase == RQ_PHASE_COMPLETE)
812 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
815 /* This will also unregister reply. */
816 ptlrpc_expire_one_request(req, 0);
818 /* We're done with this req, let's finally move it to complete
819 * phase and take care of inflights. */
820 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
821 imp = req->rq_import;
822 spin_lock(&imp->imp_lock);
823 if (!list_empty(&req->rq_list)) {
824 list_del_init(&req->rq_list);
825 atomic_dec(&imp->imp_inflight);
827 spin_unlock(&imp->imp_lock);
828 set->set_remaining--;
830 mutex_up(&pinger_sem);
832 ptlrpc_set_destroy(set);
836 pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
837 cfs_time_seconds(PING_INTERVAL));
838 pd->pd_this_ping = 0; /* XXX for debug */
840 CDEBUG(D_INFO, "finished a round ping\n");
845 static void *pinger_callback = NULL;
846 #endif /* ENABLE_PINGER */
848 int ptlrpc_start_pinger(void)
851 memset(&pinger_args, 0, sizeof(pinger_args));
852 pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
859 int ptlrpc_stop_pinger(void)
863 liblustre_deregister_wait_callback(pinger_callback);
868 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
871 mutex_down(&pinger_sem);
872 ptlrpc_update_next_ping(imp, 0);
873 if (pinger_args.pd_set == NULL &&
874 time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
875 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
876 imp->imp_next_ping, cfs_time_current());
877 pinger_args.pd_next_ping = imp->imp_next_ping;
879 mutex_up(&pinger_sem);
883 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
884 timeout_cb_t cb, void *data,
885 struct list_head *obd_list)
890 int ptlrpc_del_timeout_client(struct list_head *obd_list,
891 enum timeout_event event)
896 int ptlrpc_pinger_add_import(struct obd_import *imp)
899 if (!list_empty(&imp->imp_pinger_chain))
902 CDEBUG(D_HA, "adding pingable import %s->%s\n",
903 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
904 ptlrpc_pinger_sending_on_import(imp);
906 mutex_down(&pinger_sem);
907 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
908 class_import_get(imp);
909 mutex_up(&pinger_sem);
914 int ptlrpc_pinger_del_import(struct obd_import *imp)
917 if (list_empty(&imp->imp_pinger_chain))
920 mutex_down(&pinger_sem);
921 list_del_init(&imp->imp_pinger_chain);
922 CDEBUG(D_HA, "removing pingable import %s->%s\n",
923 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
924 class_import_put(imp);
925 mutex_up(&pinger_sem);
929 void ptlrpc_pinger_wake_up()
933 /* XXX force pinger to run, if needed */
934 struct obd_import *imp;
935 list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
936 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
937 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
938 #ifdef ENABLE_LIBLUSTRE_RECOVERY
939 if (imp->imp_state == LUSTRE_IMP_DISCON &&
940 !imp_is_deactive(imp))
942 /*XXX only recover for the initial connection */
943 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
944 imp->imp_state == LUSTRE_IMP_DISCON &&
945 !imp_is_deactive(imp))
947 ptlrpc_initiate_recovery(imp);
948 else if (imp->imp_state != LUSTRE_IMP_FULL)
949 CDEBUG(D_HA, "Refused to recover import %s->%s "
950 "state %d, deactive %d\n",
951 imp->imp_obd->obd_uuid.uuid,
952 obd2cli_tgt(imp->imp_obd), imp->imp_state,
953 imp_is_deactive(imp));
958 #endif /* !__KERNEL__ */