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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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 cfs_mutex_t pinger_mutex;
52 static CFS_LIST_HEAD(pinger_imports);
53 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
54 struct ptlrpc_request *
55 ptlrpc_prep_ping(struct obd_import *imp)
57 struct ptlrpc_request *req;
59 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
60 LUSTRE_OBD_VERSION, OBD_PING);
62 ptlrpc_request_set_replen(req);
63 req->rq_no_resend = req->rq_no_delay = 1;
68 int ptlrpc_obd_ping(struct obd_device *obd)
71 struct ptlrpc_request *req;
74 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
78 req->rq_send_state = LUSTRE_IMP_FULL;
80 rc = ptlrpc_queue_wait(req);
82 ptlrpc_req_finished(req);
86 EXPORT_SYMBOL(ptlrpc_obd_ping);
88 int ptlrpc_ping(struct obd_import *imp)
90 struct ptlrpc_request *req;
93 req = ptlrpc_prep_ping(imp);
95 CERROR("OOM trying to ping %s->%s\n",
96 imp->imp_obd->obd_uuid.uuid,
97 obd2cli_tgt(imp->imp_obd));
101 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
102 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
103 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
108 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
111 int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
112 if (imp->imp_state == LUSTRE_IMP_DISCON) {
113 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
115 at_get(&imp->imp_at.iat_net_latency));
116 time = min(time, dtime);
118 imp->imp_next_ping = cfs_time_shift(time);
119 #endif /* ENABLE_PINGER */
122 void ptlrpc_ping_import_soon(struct obd_import *imp)
124 imp->imp_next_ping = cfs_time_current();
127 static inline int imp_is_deactive(struct obd_import *imp)
129 return (imp->imp_deactive ||
130 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
133 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
135 if (imp->imp_server_timeout)
136 return cfs_time_shift(obd_timeout / 2);
138 return cfs_time_shift(obd_timeout);
141 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
142 static cfs_time_t suspend_wakeup_time = 0;
144 cfs_duration_t pinger_check_timeout(cfs_time_t time)
146 struct timeout_item *item;
147 cfs_time_t timeout = PING_INTERVAL;
149 /* The timeout list is a increase order sorted list */
150 cfs_mutex_lock(&pinger_mutex);
151 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
152 int ti_timeout = item->ti_timeout;
153 if (timeout > ti_timeout)
154 timeout = ti_timeout;
157 cfs_mutex_unlock(&pinger_mutex);
159 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
163 static cfs_waitq_t suspend_timeouts_waitq;
165 cfs_time_t ptlrpc_suspend_wakeup_time(void)
167 return suspend_wakeup_time;
170 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
172 /*XXX: disabled for now, will be replaced by adaptive timeouts */
174 if (imp->imp_no_timeout)
176 imp->imp_no_timeout = 1;
177 cfs_atomic_inc(&suspend_timeouts);
178 CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
179 cfs_atomic_read(&suspend_timeouts));
183 void ptlrpc_activate_timeouts(struct obd_import *imp)
185 /*XXX: disabled for now, will be replaced by adaptive timeouts */
187 if (!imp->imp_no_timeout)
189 imp->imp_no_timeout = 0;
190 LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
191 if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
192 suspend_wakeup_time = cfs_time_current();
193 cfs_waitq_signal(&suspend_timeouts_waitq);
195 CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
196 cfs_atomic_read(&suspend_timeouts));
200 int ptlrpc_check_suspend(void)
202 if (cfs_atomic_read(&suspend_timeouts))
207 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
209 struct l_wait_info lwi;
211 if (cfs_atomic_read(&suspend_timeouts)) {
212 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
213 cfs_atomic_read(&suspend_timeouts));
214 lwi = LWI_INTR(NULL, NULL);
215 l_wait_event(suspend_timeouts_waitq,
216 cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
217 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
225 static void ptlrpc_pinger_process_import(struct obd_import *imp,
226 unsigned long this_ping)
230 cfs_spin_lock(&imp->imp_lock);
231 level = imp->imp_state;
232 force = imp->imp_force_verify;
234 imp->imp_force_verify = 0;
235 cfs_spin_unlock(&imp->imp_lock);
237 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
238 "level %s/%u force %u deactive %u pingable %u\n",
239 ptlrpc_import_state_name(level), level,
240 force, imp->imp_deactive, imp->imp_pingable);
242 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
243 this_ping) && force == 0)
246 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
247 /* wait for a while before trying recovery again */
248 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
249 if (!imp->imp_no_pinger_recover)
250 ptlrpc_initiate_recovery(imp);
251 } else if (level != LUSTRE_IMP_FULL ||
252 imp->imp_obd->obd_no_recov ||
253 imp_is_deactive(imp)) {
254 CDEBUG(D_HA, "not pinging %s (in recovery "
255 " or recovery disabled: %s)\n",
256 obd2cli_tgt(imp->imp_obd),
257 ptlrpc_import_state_name(level));
258 } else if (imp->imp_pingable || force) {
263 static int ptlrpc_pinger_main(void *arg)
265 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
266 struct ptlrpc_thread *thread = data->thread;
269 cfs_daemonize(data->name);
271 /* Record that the thread is running */
272 thread_set_flags(thread, SVC_RUNNING);
273 cfs_waitq_signal(&thread->t_ctl_waitq);
275 /* And now, loop forever, pinging as needed. */
277 cfs_time_t this_ping = cfs_time_current();
278 struct l_wait_info lwi;
279 cfs_duration_t time_to_next_wake;
280 struct timeout_item *item;
283 cfs_mutex_lock(&pinger_mutex);
284 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
285 item->ti_cb(item, item->ti_cb_data);
287 cfs_list_for_each(iter, &pinger_imports) {
288 struct obd_import *imp =
289 cfs_list_entry(iter, struct obd_import,
292 ptlrpc_pinger_process_import(imp, this_ping);
293 /* obd_timeout might have changed */
294 if (imp->imp_pingable && imp->imp_next_ping &&
295 cfs_time_after(imp->imp_next_ping,
296 cfs_time_add(this_ping,
297 cfs_time_seconds(PING_INTERVAL))))
298 ptlrpc_update_next_ping(imp, 0);
300 cfs_mutex_unlock(&pinger_mutex);
301 /* update memory usage info */
302 obd_update_maxusage();
304 /* Wait until the next ping time, or until we're stopped. */
305 time_to_next_wake = pinger_check_timeout(this_ping);
306 /* The ping sent by ptlrpc_send_rpc may get sent out
307 say .01 second after this.
308 ptlrpc_pinger_sending_on_import will then set the
309 next ping time to next_ping + .01 sec, which means
310 we will SKIP the next ping at next_ping, and the
311 ping will get sent 2 timeouts from now! Beware. */
312 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
313 CFS_TIME_T")\n", time_to_next_wake,
314 cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
315 if (time_to_next_wake > 0) {
316 lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
318 cfs_time_seconds(1)),
320 l_wait_event(thread->t_ctl_waitq,
321 thread_is_stopping(thread) ||
322 thread_is_event(thread),
324 if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
328 /* woken after adding import to reset timer */
329 thread_test_and_clear_flags(thread, SVC_EVENT);
334 thread_set_flags(thread, SVC_STOPPED);
335 cfs_waitq_signal(&thread->t_ctl_waitq);
337 CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
341 static struct ptlrpc_thread *pinger_thread = NULL;
343 int ptlrpc_start_pinger(void)
345 struct l_wait_info lwi = { 0 };
346 struct ptlrpc_svc_data d;
348 #ifndef ENABLE_PINGER
353 if (pinger_thread != NULL)
356 OBD_ALLOC_PTR(pinger_thread);
357 if (pinger_thread == NULL)
359 cfs_waitq_init(&pinger_thread->t_ctl_waitq);
360 cfs_waitq_init(&suspend_timeouts_waitq);
363 d.thread = pinger_thread;
365 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
366 * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
367 rc = cfs_create_thread(ptlrpc_pinger_main, &d, CFS_DAEMON_FLAGS);
369 CERROR("cannot start thread: %d\n", rc);
370 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
371 pinger_thread = NULL;
374 l_wait_event(pinger_thread->t_ctl_waitq,
375 thread_is_running(pinger_thread), &lwi);
380 int ptlrpc_pinger_remove_timeouts(void);
382 int ptlrpc_stop_pinger(void)
384 struct l_wait_info lwi = { 0 };
386 #ifndef ENABLE_PINGER
391 if (pinger_thread == NULL)
394 ptlrpc_pinger_remove_timeouts();
395 cfs_mutex_lock(&pinger_mutex);
396 thread_set_flags(pinger_thread, SVC_STOPPING);
397 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
398 cfs_mutex_unlock(&pinger_mutex);
400 l_wait_event(pinger_thread->t_ctl_waitq,
401 thread_is_stopped(pinger_thread), &lwi);
403 OBD_FREE_PTR(pinger_thread);
404 pinger_thread = NULL;
408 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
410 ptlrpc_update_next_ping(imp, 0);
413 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
415 ptlrpc_update_next_ping(imp, 1);
418 int ptlrpc_pinger_add_import(struct obd_import *imp)
421 if (!cfs_list_empty(&imp->imp_pinger_chain))
424 cfs_mutex_lock(&pinger_mutex);
425 CDEBUG(D_HA, "adding pingable import %s->%s\n",
426 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
427 /* if we add to pinger we want recovery on this import */
428 imp->imp_obd->obd_no_recov = 0;
429 ptlrpc_update_next_ping(imp, 0);
430 /* XXX sort, blah blah */
431 cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
432 class_import_get(imp);
434 ptlrpc_pinger_wake_up();
435 cfs_mutex_unlock(&pinger_mutex);
440 int ptlrpc_pinger_del_import(struct obd_import *imp)
443 if (cfs_list_empty(&imp->imp_pinger_chain))
446 cfs_mutex_lock(&pinger_mutex);
447 cfs_list_del_init(&imp->imp_pinger_chain);
448 CDEBUG(D_HA, "removing pingable import %s->%s\n",
449 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
450 /* if we remove from pinger we don't want recovery on this import */
451 imp->imp_obd->obd_no_recov = 1;
452 class_import_put(imp);
453 cfs_mutex_unlock(&pinger_mutex);
458 * Register a timeout callback to the pinger list, and the callback will
459 * be called when timeout happens.
461 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
462 timeout_cb_t cb, void *data)
464 struct timeout_item *ti;
470 CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
471 CFS_INIT_LIST_HEAD(&ti->ti_chain);
472 ti->ti_timeout = time;
473 ti->ti_event = event;
475 ti->ti_cb_data = data;
481 * Register timeout event on the the pinger thread.
482 * Note: the timeout list is an sorted list with increased timeout value.
484 static struct timeout_item*
485 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
486 timeout_cb_t cb, void *data)
488 struct timeout_item *item, *tmp;
490 LASSERT_MUTEX_LOCKED(&pinger_mutex);
492 cfs_list_for_each_entry(item, &timeout_list, ti_chain)
493 if (item->ti_event == event)
496 item = ptlrpc_new_timeout(time, event, cb, data);
498 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
499 if (tmp->ti_timeout < time) {
500 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
504 cfs_list_add(&item->ti_chain, &timeout_list);
510 /* Add a client_obd to the timeout event list, when timeout(@time)
511 * happens, the callback(@cb) will be called.
513 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
514 timeout_cb_t cb, void *data,
515 cfs_list_t *obd_list)
517 struct timeout_item *ti;
519 cfs_mutex_lock(&pinger_mutex);
520 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
522 cfs_mutex_unlock(&pinger_mutex);
525 cfs_list_add(obd_list, &ti->ti_obd_list);
526 cfs_mutex_unlock(&pinger_mutex);
530 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
531 enum timeout_event event)
533 struct timeout_item *ti = NULL, *item;
535 if (cfs_list_empty(obd_list))
537 cfs_mutex_lock(&pinger_mutex);
538 cfs_list_del_init(obd_list);
540 * If there are no obd attached to the timeout event
541 * list, remove this timeout event from the pinger
543 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
544 if (item->ti_event == event) {
549 LASSERTF(ti != NULL, "ti is NULL ! \n");
550 if (cfs_list_empty(&ti->ti_obd_list)) {
551 cfs_list_del(&ti->ti_chain);
554 cfs_mutex_unlock(&pinger_mutex);
558 int ptlrpc_pinger_remove_timeouts(void)
560 struct timeout_item *item, *tmp;
562 cfs_mutex_lock(&pinger_mutex);
563 cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
564 LASSERT(cfs_list_empty(&item->ti_obd_list));
565 cfs_list_del(&item->ti_chain);
568 cfs_mutex_unlock(&pinger_mutex);
572 void ptlrpc_pinger_wake_up()
575 thread_add_flags(pinger_thread, SVC_EVENT);
576 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
580 /* Ping evictor thread */
582 #define PET_TERMINATE 2
584 static int pet_refcount = 0;
585 static int pet_state;
586 static cfs_waitq_t pet_waitq;
587 CFS_LIST_HEAD(pet_list);
588 static cfs_spinlock_t pet_lock = CFS_SPIN_LOCK_UNLOCKED;
590 int ping_evictor_wake(struct obd_export *exp)
592 struct obd_device *obd;
594 cfs_spin_lock(&pet_lock);
595 if (pet_state != PET_READY) {
596 /* eventually the new obd will call here again. */
597 cfs_spin_unlock(&pet_lock);
601 obd = class_exp2obd(exp);
602 if (cfs_list_empty(&obd->obd_evict_list)) {
603 class_incref(obd, "evictor", obd);
604 cfs_list_add(&obd->obd_evict_list, &pet_list);
606 cfs_spin_unlock(&pet_lock);
608 cfs_waitq_signal(&pet_waitq);
612 static int ping_evictor_main(void *arg)
614 struct obd_device *obd;
615 struct obd_export *exp;
616 struct l_wait_info lwi = { 0 };
620 cfs_daemonize_ctxt("ll_evictor");
622 CDEBUG(D_HA, "Starting Ping Evictor\n");
623 pet_state = PET_READY;
625 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
626 (pet_state == PET_TERMINATE), &lwi);
628 /* loop until all obd's will be removed */
629 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
632 /* we only get here if pet_exp != NULL, and the end of this
633 * loop is the only place which sets it NULL again, so lock
634 * is not strictly necessary. */
635 cfs_spin_lock(&pet_lock);
636 obd = cfs_list_entry(pet_list.next, struct obd_device,
638 cfs_spin_unlock(&pet_lock);
640 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
642 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
643 obd->obd_name, expire_time);
645 /* Exports can't be deleted out of the list while we hold
646 * the obd lock (class_unlink_export), which means we can't
647 * lose the last ref on the export. If they've already been
648 * removed from the list, we won't find them here. */
649 cfs_spin_lock(&obd->obd_dev_lock);
650 while (!cfs_list_empty(&obd->obd_exports_timed)) {
651 exp = cfs_list_entry(obd->obd_exports_timed.next,
653 exp_obd_chain_timed);
654 if (expire_time > exp->exp_last_request_time) {
655 class_export_get(exp);
656 cfs_spin_unlock(&obd->obd_dev_lock);
657 LCONSOLE_WARN("%s: haven't heard from client %s"
658 " (at %s) in %ld seconds. I think"
659 " it's dead, and I am evicting"
660 " it. exp %p, cur %ld expire %ld"
663 obd_uuid2str(&exp->exp_client_uuid),
664 obd_export_nid2str(exp),
665 (long)(cfs_time_current_sec() -
666 exp->exp_last_request_time),
667 exp, (long)cfs_time_current_sec(),
669 (long)exp->exp_last_request_time);
670 CDEBUG(D_HA, "Last request was at %ld\n",
671 exp->exp_last_request_time);
672 class_fail_export(exp);
673 class_export_put(exp);
674 cfs_spin_lock(&obd->obd_dev_lock);
676 /* List is sorted, so everyone below is ok */
680 cfs_spin_unlock(&obd->obd_dev_lock);
682 cfs_spin_lock(&pet_lock);
683 cfs_list_del_init(&obd->obd_evict_list);
684 cfs_spin_unlock(&pet_lock);
686 class_decref(obd, "evictor", obd);
688 CDEBUG(D_HA, "Exiting Ping Evictor\n");
693 void ping_evictor_start(void)
697 if (++pet_refcount > 1)
700 cfs_waitq_init(&pet_waitq);
702 rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
705 CERROR("Cannot start ping evictor thread: %d\n", rc);
708 EXPORT_SYMBOL(ping_evictor_start);
710 void ping_evictor_stop(void)
712 if (--pet_refcount > 0)
715 pet_state = PET_TERMINATE;
716 cfs_waitq_signal(&pet_waitq);
718 EXPORT_SYMBOL(ping_evictor_stop);
719 #else /* !__KERNEL__ */
722 * the current implementation of pinger in liblustre is not optimized
726 static struct pinger_data {
728 cfs_time_t pd_this_ping; /* jiffies */
729 cfs_time_t pd_next_ping; /* jiffies */
730 struct ptlrpc_request_set *pd_set;
733 static int pinger_check_rpcs(void *arg)
735 cfs_time_t curtime = cfs_time_current();
736 struct ptlrpc_request *req;
737 struct ptlrpc_request_set *set;
739 struct obd_import *imp;
740 struct pinger_data *pd = &pinger_args;
743 /* prevent recursion */
744 if (pd->pd_recursion++) {
745 CDEBUG(D_HA, "pinger: recursion! quit\n");
751 /* have we reached ping point? */
752 if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
757 /* if we have rpc_set already, continue processing it */
759 LASSERT(pd->pd_this_ping);
764 pd->pd_this_ping = curtime;
765 pd->pd_set = ptlrpc_prep_set();
766 if (pd->pd_set == NULL)
770 /* add rpcs into set */
771 cfs_mutex_lock(&pinger_mutex);
772 cfs_list_for_each(iter, &pinger_imports) {
773 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
775 int generation, level;
777 if (cfs_time_aftereq(pd->pd_this_ping,
778 imp->imp_next_ping - 5 * CFS_TICK)) {
780 cfs_spin_lock(&imp->imp_lock);
781 generation = imp->imp_generation;
782 level = imp->imp_state;
783 cfs_spin_unlock(&imp->imp_lock);
785 if (level != LUSTRE_IMP_FULL) {
787 "not pinging %s (in recovery)\n",
788 obd2cli_tgt(imp->imp_obd));
792 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
796 CERROR("OOM trying to ping %s->%s\n",
797 imp->imp_obd->obd_uuid.uuid,
798 obd2cli_tgt(imp->imp_obd));
802 req->rq_no_resend = 1;
803 ptlrpc_request_set_replen(req);
804 req->rq_send_state = LUSTRE_IMP_FULL;
805 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
806 req->rq_import_generation = generation;
807 ptlrpc_set_add_req(set, req);
809 CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
810 " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
811 imp->imp_next_ping, pd->pd_this_ping);
814 pd->pd_this_ping = curtime;
815 cfs_mutex_unlock(&pinger_mutex);
817 /* Might be empty, that's OK. */
818 if (cfs_atomic_read(&set->set_remaining) == 0)
819 CDEBUG(D_RPCTRACE, "nothing to ping\n");
821 cfs_list_for_each(iter, &set->set_requests) {
822 struct ptlrpc_request *req =
823 cfs_list_entry(iter, struct ptlrpc_request,
825 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
826 req->rq_import->imp_obd->obd_uuid.uuid,
827 obd2cli_tgt(req->rq_import->imp_obd));
828 (void)ptl_send_rpc(req, 0);
832 rc = ptlrpc_check_set(NULL, set);
834 /* not finished, and we are not expired, simply return */
835 if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
836 cfs_time_seconds(PING_INTERVAL)))) {
837 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
842 /* Expire all the requests that didn't come back. */
843 cfs_mutex_lock(&pinger_mutex);
844 cfs_list_for_each(iter, &set->set_requests) {
845 req = cfs_list_entry(iter, struct ptlrpc_request,
848 if (req->rq_phase == RQ_PHASE_COMPLETE)
851 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
854 /* This will also unregister reply. */
855 ptlrpc_expire_one_request(req, 0);
857 /* We're done with this req, let's finally move it to complete
858 * phase and take care of inflights. */
859 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
860 imp = req->rq_import;
861 cfs_spin_lock(&imp->imp_lock);
862 if (!cfs_list_empty(&req->rq_list)) {
863 cfs_list_del_init(&req->rq_list);
864 cfs_atomic_dec(&imp->imp_inflight);
866 cfs_spin_unlock(&imp->imp_lock);
867 cfs_atomic_dec(&set->set_remaining);
869 cfs_mutex_unlock(&pinger_mutex);
871 ptlrpc_set_destroy(set);
875 pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
876 cfs_time_seconds(PING_INTERVAL));
877 pd->pd_this_ping = 0; /* XXX for debug */
879 CDEBUG(D_INFO, "finished a round ping\n");
884 static void *pinger_callback = NULL;
885 #endif /* ENABLE_PINGER */
887 int ptlrpc_start_pinger(void)
890 memset(&pinger_args, 0, sizeof(pinger_args));
891 pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
898 int ptlrpc_stop_pinger(void)
902 liblustre_deregister_wait_callback(pinger_callback);
907 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
910 cfs_mutex_lock(&pinger_mutex);
911 ptlrpc_update_next_ping(imp, 0);
912 if (pinger_args.pd_set == NULL &&
913 cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
914 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
915 imp->imp_next_ping, cfs_time_current());
916 pinger_args.pd_next_ping = imp->imp_next_ping;
918 cfs_mutex_unlock(&pinger_mutex);
922 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
925 cfs_mutex_lock(&pinger_mutex);
926 ptlrpc_update_next_ping(imp, 1);
927 if (pinger_args.pd_set == NULL &&
928 cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
929 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
930 imp->imp_next_ping, cfs_time_current());
931 pinger_args.pd_next_ping = imp->imp_next_ping;
933 cfs_mutex_unlock(&pinger_mutex);
937 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
938 timeout_cb_t cb, void *data,
939 cfs_list_t *obd_list)
944 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
945 enum timeout_event event)
950 int ptlrpc_pinger_add_import(struct obd_import *imp)
953 if (!cfs_list_empty(&imp->imp_pinger_chain))
956 CDEBUG(D_HA, "adding pingable import %s->%s\n",
957 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
958 ptlrpc_pinger_sending_on_import(imp);
960 cfs_mutex_lock(&pinger_mutex);
961 cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
962 class_import_get(imp);
963 cfs_mutex_unlock(&pinger_mutex);
968 int ptlrpc_pinger_del_import(struct obd_import *imp)
971 if (cfs_list_empty(&imp->imp_pinger_chain))
974 cfs_mutex_lock(&pinger_mutex);
975 cfs_list_del_init(&imp->imp_pinger_chain);
976 CDEBUG(D_HA, "removing pingable import %s->%s\n",
977 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
978 class_import_put(imp);
979 cfs_mutex_unlock(&pinger_mutex);
983 void ptlrpc_pinger_wake_up()
986 /* XXX force pinger to run, if needed */
987 struct obd_import *imp;
989 cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
990 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
991 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
992 #ifdef ENABLE_LIBLUSTRE_RECOVERY
993 if (imp->imp_state == LUSTRE_IMP_DISCON &&
994 !imp_is_deactive(imp))
996 /*XXX only recover for the initial connection */
997 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
998 imp->imp_state == LUSTRE_IMP_DISCON &&
999 !imp_is_deactive(imp))
1001 ptlrpc_initiate_recovery(imp);
1002 else if (imp->imp_state != LUSTRE_IMP_FULL)
1003 CDEBUG(D_HA, "Refused to recover import %s->%s "
1004 "state %d, deactive %d\n",
1005 imp->imp_obd->obd_uuid.uuid,
1006 obd2cli_tgt(imp->imp_obd), imp->imp_state,
1007 imp_is_deactive(imp));
1012 #endif /* !__KERNEL__ */