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 (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/ptlrpc/pinger.c
41 * Portal-RPC reconnection and replay operations, for use in recovery.
45 #include <liblustre.h>
47 #define DEBUG_SUBSYSTEM S_RPC
50 #include <obd_support.h>
51 #include <obd_class.h>
52 #include "ptlrpc_internal.h"
54 cfs_semaphore_t pinger_sem;
55 static CFS_LIST_HEAD(pinger_imports);
56 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
57 struct ptlrpc_request *
58 ptlrpc_prep_ping(struct obd_import *imp)
60 struct ptlrpc_request *req;
62 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
63 LUSTRE_OBD_VERSION, OBD_PING);
65 ptlrpc_request_set_replen(req);
66 req->rq_no_resend = req->rq_no_delay = 1;
71 int ptlrpc_obd_ping(struct obd_device *obd)
74 struct ptlrpc_request *req;
77 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
81 req->rq_send_state = LUSTRE_IMP_FULL;
83 rc = ptlrpc_queue_wait(req);
85 ptlrpc_req_finished(req);
89 EXPORT_SYMBOL(ptlrpc_obd_ping);
91 int ptlrpc_ping(struct obd_import *imp)
93 struct ptlrpc_request *req;
96 req = ptlrpc_prep_ping(imp);
98 CERROR("OOM trying to ping %s->%s\n",
99 imp->imp_obd->obd_uuid.uuid,
100 obd2cli_tgt(imp->imp_obd));
104 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
105 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
106 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
111 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
114 int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
115 if (imp->imp_state == LUSTRE_IMP_DISCON) {
116 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
118 at_get(&imp->imp_at.iat_net_latency));
119 time = min(time, dtime);
121 imp->imp_next_ping = cfs_time_shift(time);
122 #endif /* ENABLE_PINGER */
125 void ptlrpc_ping_import_soon(struct obd_import *imp)
127 imp->imp_next_ping = cfs_time_current();
130 static inline int imp_is_deactive(struct obd_import *imp)
132 return (imp->imp_deactive ||
133 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
136 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
138 if (imp->imp_server_timeout)
139 return cfs_time_shift(obd_timeout / 2);
141 return cfs_time_shift(obd_timeout);
144 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
145 static cfs_time_t suspend_wakeup_time = 0;
147 cfs_duration_t pinger_check_timeout(cfs_time_t time)
149 struct timeout_item *item;
150 cfs_time_t timeout = PING_INTERVAL;
152 /* The timeout list is a increase order sorted list */
153 cfs_mutex_down(&pinger_sem);
154 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
155 int ti_timeout = item->ti_timeout;
156 if (timeout > ti_timeout)
157 timeout = ti_timeout;
160 cfs_mutex_up(&pinger_sem);
162 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
166 static cfs_waitq_t suspend_timeouts_waitq;
168 cfs_time_t ptlrpc_suspend_wakeup_time(void)
170 return suspend_wakeup_time;
173 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
175 /*XXX: disabled for now, will be replaced by adaptive timeouts */
177 if (imp->imp_no_timeout)
179 imp->imp_no_timeout = 1;
180 cfs_atomic_inc(&suspend_timeouts);
181 CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
182 cfs_atomic_read(&suspend_timeouts));
186 void ptlrpc_activate_timeouts(struct obd_import *imp)
188 /*XXX: disabled for now, will be replaced by adaptive timeouts */
190 if (!imp->imp_no_timeout)
192 imp->imp_no_timeout = 0;
193 LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
194 if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
195 suspend_wakeup_time = cfs_time_current();
196 cfs_waitq_signal(&suspend_timeouts_waitq);
198 CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
199 cfs_atomic_read(&suspend_timeouts));
203 int ptlrpc_check_suspend(void)
205 if (cfs_atomic_read(&suspend_timeouts))
210 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
212 struct l_wait_info lwi;
214 if (cfs_atomic_read(&suspend_timeouts)) {
215 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
216 cfs_atomic_read(&suspend_timeouts));
217 lwi = LWI_INTR(NULL, NULL);
218 l_wait_event(suspend_timeouts_waitq,
219 cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
220 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
228 static void ptlrpc_pinger_process_import(struct obd_import *imp,
229 unsigned long this_ping)
233 cfs_spin_lock(&imp->imp_lock);
234 level = imp->imp_state;
235 force = imp->imp_force_verify;
237 imp->imp_force_verify = 0;
238 cfs_spin_unlock(&imp->imp_lock);
240 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
241 "level %s/%u force %u deactive %u pingable %u\n",
242 ptlrpc_import_state_name(level), level,
243 force, imp->imp_deactive, imp->imp_pingable);
245 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
246 this_ping) && force == 0)
249 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
250 /* wait for a while before trying recovery again */
251 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
252 if (!imp->imp_no_pinger_recover)
253 ptlrpc_initiate_recovery(imp);
254 } else if (level != LUSTRE_IMP_FULL ||
255 imp->imp_obd->obd_no_recov ||
256 imp_is_deactive(imp)) {
257 CDEBUG(D_HA, "not pinging %s (in recovery "
258 " or recovery disabled: %s)\n",
259 obd2cli_tgt(imp->imp_obd),
260 ptlrpc_import_state_name(level));
261 } else if (imp->imp_pingable || force) {
266 static int ptlrpc_pinger_main(void *arg)
268 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
269 struct ptlrpc_thread *thread = data->thread;
272 cfs_daemonize(data->name);
274 /* Record that the thread is running */
275 thread_set_flags(thread, SVC_RUNNING);
276 cfs_waitq_signal(&thread->t_ctl_waitq);
278 /* And now, loop forever, pinging as needed. */
280 cfs_time_t this_ping = cfs_time_current();
281 struct l_wait_info lwi;
282 cfs_duration_t time_to_next_wake;
283 struct timeout_item *item;
286 cfs_mutex_down(&pinger_sem);
287 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
288 item->ti_cb(item, item->ti_cb_data);
290 cfs_list_for_each(iter, &pinger_imports) {
291 struct obd_import *imp =
292 cfs_list_entry(iter, struct obd_import,
295 ptlrpc_pinger_process_import(imp, this_ping);
296 /* obd_timeout might have changed */
297 if (imp->imp_pingable && imp->imp_next_ping &&
298 cfs_time_after(imp->imp_next_ping,
299 cfs_time_add(this_ping,
300 cfs_time_seconds(PING_INTERVAL))))
301 ptlrpc_update_next_ping(imp, 0);
303 cfs_mutex_up(&pinger_sem);
304 /* update memory usage info */
305 obd_update_maxusage();
307 /* Wait until the next ping time, or until we're stopped. */
308 time_to_next_wake = pinger_check_timeout(this_ping);
309 /* The ping sent by ptlrpc_send_rpc may get sent out
310 say .01 second after this.
311 ptlrpc_pinger_sending_on_import will then set the
312 next ping time to next_ping + .01 sec, which means
313 we will SKIP the next ping at next_ping, and the
314 ping will get sent 2 timeouts from now! Beware. */
315 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
316 CFS_TIME_T")\n", time_to_next_wake,
317 cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
318 if (time_to_next_wake > 0) {
319 lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
321 cfs_time_seconds(1)),
323 l_wait_event(thread->t_ctl_waitq,
324 thread_is_stopping(thread) ||
325 thread_is_event(thread),
327 if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
331 /* woken after adding import to reset timer */
332 thread_test_and_clear_flags(thread, SVC_EVENT);
337 thread_set_flags(thread, SVC_STOPPED);
338 cfs_waitq_signal(&thread->t_ctl_waitq);
340 CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
344 static struct ptlrpc_thread *pinger_thread = NULL;
346 int ptlrpc_start_pinger(void)
348 struct l_wait_info lwi = { 0 };
349 struct ptlrpc_svc_data d;
351 #ifndef ENABLE_PINGER
356 if (pinger_thread != NULL)
359 OBD_ALLOC_PTR(pinger_thread);
360 if (pinger_thread == NULL)
362 cfs_waitq_init(&pinger_thread->t_ctl_waitq);
363 cfs_waitq_init(&suspend_timeouts_waitq);
366 d.thread = pinger_thread;
368 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
369 * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
370 rc = cfs_create_thread(ptlrpc_pinger_main, &d, CFS_DAEMON_FLAGS);
372 CERROR("cannot start thread: %d\n", rc);
373 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
374 pinger_thread = NULL;
377 l_wait_event(pinger_thread->t_ctl_waitq,
378 thread_is_running(pinger_thread), &lwi);
383 int ptlrpc_pinger_remove_timeouts(void);
385 int ptlrpc_stop_pinger(void)
387 struct l_wait_info lwi = { 0 };
389 #ifndef ENABLE_PINGER
394 if (pinger_thread == NULL)
397 ptlrpc_pinger_remove_timeouts();
398 cfs_mutex_down(&pinger_sem);
399 thread_set_flags(pinger_thread, SVC_STOPPING);
400 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
401 cfs_mutex_up(&pinger_sem);
403 l_wait_event(pinger_thread->t_ctl_waitq,
404 thread_is_stopped(pinger_thread), &lwi);
406 OBD_FREE_PTR(pinger_thread);
407 pinger_thread = NULL;
411 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
413 ptlrpc_update_next_ping(imp, 0);
416 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
418 ptlrpc_update_next_ping(imp, 1);
421 int ptlrpc_pinger_add_import(struct obd_import *imp)
424 if (!cfs_list_empty(&imp->imp_pinger_chain))
427 cfs_mutex_down(&pinger_sem);
428 CDEBUG(D_HA, "adding pingable import %s->%s\n",
429 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
430 /* if we add to pinger we want recovery on this import */
431 imp->imp_obd->obd_no_recov = 0;
432 ptlrpc_update_next_ping(imp, 0);
433 /* XXX sort, blah blah */
434 cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
435 class_import_get(imp);
437 ptlrpc_pinger_wake_up();
438 cfs_mutex_up(&pinger_sem);
443 int ptlrpc_pinger_del_import(struct obd_import *imp)
446 if (cfs_list_empty(&imp->imp_pinger_chain))
449 cfs_mutex_down(&pinger_sem);
450 cfs_list_del_init(&imp->imp_pinger_chain);
451 CDEBUG(D_HA, "removing pingable import %s->%s\n",
452 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
453 /* if we remove from pinger we don't want recovery on this import */
454 imp->imp_obd->obd_no_recov = 1;
455 class_import_put(imp);
456 cfs_mutex_up(&pinger_sem);
461 * Register a timeout callback to the pinger list, and the callback will
462 * be called when timeout happens.
464 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
465 timeout_cb_t cb, void *data)
467 struct timeout_item *ti;
473 CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
474 CFS_INIT_LIST_HEAD(&ti->ti_chain);
475 ti->ti_timeout = time;
476 ti->ti_event = event;
478 ti->ti_cb_data = data;
484 * Register timeout event on the the pinger thread.
485 * Note: the timeout list is an sorted list with increased timeout value.
487 static struct timeout_item*
488 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
489 timeout_cb_t cb, void *data)
491 struct timeout_item *item, *tmp;
493 LASSERT_SEM_LOCKED(&pinger_sem);
495 cfs_list_for_each_entry(item, &timeout_list, ti_chain)
496 if (item->ti_event == event)
499 item = ptlrpc_new_timeout(time, event, cb, data);
501 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
502 if (tmp->ti_timeout < time) {
503 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
507 cfs_list_add(&item->ti_chain, &timeout_list);
513 /* Add a client_obd to the timeout event list, when timeout(@time)
514 * happens, the callback(@cb) will be called.
516 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
517 timeout_cb_t cb, void *data,
518 cfs_list_t *obd_list)
520 struct timeout_item *ti;
522 cfs_mutex_down(&pinger_sem);
523 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
525 cfs_mutex_up(&pinger_sem);
528 cfs_list_add(obd_list, &ti->ti_obd_list);
529 cfs_mutex_up(&pinger_sem);
533 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
534 enum timeout_event event)
536 struct timeout_item *ti = NULL, *item;
538 if (cfs_list_empty(obd_list))
540 cfs_mutex_down(&pinger_sem);
541 cfs_list_del_init(obd_list);
543 * If there are no obd attached to the timeout event
544 * list, remove this timeout event from the pinger
546 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
547 if (item->ti_event == event) {
552 LASSERTF(ti != NULL, "ti is NULL ! \n");
553 if (cfs_list_empty(&ti->ti_obd_list)) {
554 cfs_list_del(&ti->ti_chain);
557 cfs_mutex_up(&pinger_sem);
561 int ptlrpc_pinger_remove_timeouts(void)
563 struct timeout_item *item, *tmp;
565 cfs_mutex_down(&pinger_sem);
566 cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
567 LASSERT(cfs_list_empty(&item->ti_obd_list));
568 cfs_list_del(&item->ti_chain);
571 cfs_mutex_up(&pinger_sem);
575 void ptlrpc_pinger_wake_up()
578 thread_add_flags(pinger_thread, SVC_EVENT);
579 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
583 /* Ping evictor thread */
585 #define PET_TERMINATE 2
587 static int pet_refcount = 0;
588 static int pet_state;
589 static cfs_waitq_t pet_waitq;
590 CFS_LIST_HEAD(pet_list);
591 static cfs_spinlock_t pet_lock = CFS_SPIN_LOCK_UNLOCKED;
593 int ping_evictor_wake(struct obd_export *exp)
595 struct obd_device *obd;
597 cfs_spin_lock(&pet_lock);
598 if (pet_state != PET_READY) {
599 /* eventually the new obd will call here again. */
600 cfs_spin_unlock(&pet_lock);
604 obd = class_exp2obd(exp);
605 if (cfs_list_empty(&obd->obd_evict_list)) {
606 class_incref(obd, "evictor", obd);
607 cfs_list_add(&obd->obd_evict_list, &pet_list);
609 cfs_spin_unlock(&pet_lock);
611 cfs_waitq_signal(&pet_waitq);
615 static int ping_evictor_main(void *arg)
617 struct obd_device *obd;
618 struct obd_export *exp;
619 struct l_wait_info lwi = { 0 };
623 cfs_daemonize_ctxt("ll_evictor");
625 CDEBUG(D_HA, "Starting Ping Evictor\n");
626 pet_state = PET_READY;
628 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
629 (pet_state == PET_TERMINATE), &lwi);
631 /* loop until all obd's will be removed */
632 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
635 /* we only get here if pet_exp != NULL, and the end of this
636 * loop is the only place which sets it NULL again, so lock
637 * is not strictly necessary. */
638 cfs_spin_lock(&pet_lock);
639 obd = cfs_list_entry(pet_list.next, struct obd_device,
641 cfs_spin_unlock(&pet_lock);
643 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
645 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
646 obd->obd_name, expire_time);
648 /* Exports can't be deleted out of the list while we hold
649 * the obd lock (class_unlink_export), which means we can't
650 * lose the last ref on the export. If they've already been
651 * removed from the list, we won't find them here. */
652 cfs_spin_lock(&obd->obd_dev_lock);
653 while (!cfs_list_empty(&obd->obd_exports_timed)) {
654 exp = cfs_list_entry(obd->obd_exports_timed.next,
656 exp_obd_chain_timed);
657 if (expire_time > exp->exp_last_request_time) {
658 class_export_get(exp);
659 cfs_spin_unlock(&obd->obd_dev_lock);
660 LCONSOLE_WARN("%s: haven't heard from client %s"
661 " (at %s) in %ld seconds. I think"
662 " it's dead, and I am evicting"
663 " it. exp %p, cur %ld expire %ld"
666 obd_uuid2str(&exp->exp_client_uuid),
667 obd_export_nid2str(exp),
668 (long)(cfs_time_current_sec() -
669 exp->exp_last_request_time),
670 exp, (long)cfs_time_current_sec(),
672 (long)exp->exp_last_request_time);
673 CDEBUG(D_HA, "Last request was at %ld\n",
674 exp->exp_last_request_time);
675 class_fail_export(exp);
676 class_export_put(exp);
677 cfs_spin_lock(&obd->obd_dev_lock);
679 /* List is sorted, so everyone below is ok */
683 cfs_spin_unlock(&obd->obd_dev_lock);
685 cfs_spin_lock(&pet_lock);
686 cfs_list_del_init(&obd->obd_evict_list);
687 cfs_spin_unlock(&pet_lock);
689 class_decref(obd, "evictor", obd);
691 CDEBUG(D_HA, "Exiting Ping Evictor\n");
696 void ping_evictor_start(void)
700 if (++pet_refcount > 1)
703 cfs_waitq_init(&pet_waitq);
705 rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
708 CERROR("Cannot start ping evictor thread: %d\n", rc);
711 EXPORT_SYMBOL(ping_evictor_start);
713 void ping_evictor_stop(void)
715 if (--pet_refcount > 0)
718 pet_state = PET_TERMINATE;
719 cfs_waitq_signal(&pet_waitq);
721 EXPORT_SYMBOL(ping_evictor_stop);
722 #else /* !__KERNEL__ */
725 * the current implementation of pinger in liblustre is not optimized
729 static struct pinger_data {
731 cfs_time_t pd_this_ping; /* jiffies */
732 cfs_time_t pd_next_ping; /* jiffies */
733 struct ptlrpc_request_set *pd_set;
736 static int pinger_check_rpcs(void *arg)
738 cfs_time_t curtime = cfs_time_current();
739 struct ptlrpc_request *req;
740 struct ptlrpc_request_set *set;
742 struct obd_import *imp;
743 struct pinger_data *pd = &pinger_args;
746 /* prevent recursion */
747 if (pd->pd_recursion++) {
748 CDEBUG(D_HA, "pinger: recursion! quit\n");
754 /* have we reached ping point? */
755 if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
760 /* if we have rpc_set already, continue processing it */
762 LASSERT(pd->pd_this_ping);
767 pd->pd_this_ping = curtime;
768 pd->pd_set = ptlrpc_prep_set();
769 if (pd->pd_set == NULL)
773 /* add rpcs into set */
774 cfs_mutex_down(&pinger_sem);
775 cfs_list_for_each(iter, &pinger_imports) {
776 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
778 int generation, level;
780 if (cfs_time_aftereq(pd->pd_this_ping,
781 imp->imp_next_ping - 5 * CFS_TICK)) {
783 cfs_spin_lock(&imp->imp_lock);
784 generation = imp->imp_generation;
785 level = imp->imp_state;
786 cfs_spin_unlock(&imp->imp_lock);
788 if (level != LUSTRE_IMP_FULL) {
790 "not pinging %s (in recovery)\n",
791 obd2cli_tgt(imp->imp_obd));
795 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
799 CERROR("OOM trying to ping %s->%s\n",
800 imp->imp_obd->obd_uuid.uuid,
801 obd2cli_tgt(imp->imp_obd));
805 req->rq_no_resend = 1;
806 ptlrpc_request_set_replen(req);
807 req->rq_send_state = LUSTRE_IMP_FULL;
808 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
809 req->rq_import_generation = generation;
810 ptlrpc_set_add_req(set, req);
812 CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
813 " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
814 imp->imp_next_ping, pd->pd_this_ping);
817 pd->pd_this_ping = curtime;
818 cfs_mutex_up(&pinger_sem);
820 /* Might be empty, that's OK. */
821 if (cfs_atomic_read(&set->set_remaining) == 0)
822 CDEBUG(D_RPCTRACE, "nothing to ping\n");
824 cfs_list_for_each(iter, &set->set_requests) {
825 struct ptlrpc_request *req =
826 cfs_list_entry(iter, struct ptlrpc_request,
828 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
829 req->rq_import->imp_obd->obd_uuid.uuid,
830 obd2cli_tgt(req->rq_import->imp_obd));
831 (void)ptl_send_rpc(req, 0);
835 rc = ptlrpc_check_set(NULL, set);
837 /* not finished, and we are not expired, simply return */
838 if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
839 cfs_time_seconds(PING_INTERVAL)))) {
840 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
845 /* Expire all the requests that didn't come back. */
846 cfs_mutex_down(&pinger_sem);
847 cfs_list_for_each(iter, &set->set_requests) {
848 req = cfs_list_entry(iter, struct ptlrpc_request,
851 if (req->rq_phase == RQ_PHASE_COMPLETE)
854 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
857 /* This will also unregister reply. */
858 ptlrpc_expire_one_request(req, 0);
860 /* We're done with this req, let's finally move it to complete
861 * phase and take care of inflights. */
862 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
863 imp = req->rq_import;
864 cfs_spin_lock(&imp->imp_lock);
865 if (!cfs_list_empty(&req->rq_list)) {
866 cfs_list_del_init(&req->rq_list);
867 cfs_atomic_dec(&imp->imp_inflight);
869 cfs_spin_unlock(&imp->imp_lock);
870 cfs_atomic_dec(&set->set_remaining);
872 cfs_mutex_up(&pinger_sem);
874 ptlrpc_set_destroy(set);
878 pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
879 cfs_time_seconds(PING_INTERVAL));
880 pd->pd_this_ping = 0; /* XXX for debug */
882 CDEBUG(D_INFO, "finished a round ping\n");
887 static void *pinger_callback = NULL;
888 #endif /* ENABLE_PINGER */
890 int ptlrpc_start_pinger(void)
893 memset(&pinger_args, 0, sizeof(pinger_args));
894 pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
901 int ptlrpc_stop_pinger(void)
905 liblustre_deregister_wait_callback(pinger_callback);
910 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
913 cfs_mutex_down(&pinger_sem);
914 ptlrpc_update_next_ping(imp, 0);
915 if (pinger_args.pd_set == NULL &&
916 cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
917 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
918 imp->imp_next_ping, cfs_time_current());
919 pinger_args.pd_next_ping = imp->imp_next_ping;
921 cfs_mutex_up(&pinger_sem);
925 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
928 cfs_mutex_down(&pinger_sem);
929 ptlrpc_update_next_ping(imp, 1);
930 if (pinger_args.pd_set == NULL &&
931 cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
932 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
933 imp->imp_next_ping, cfs_time_current());
934 pinger_args.pd_next_ping = imp->imp_next_ping;
936 cfs_mutex_up(&pinger_sem);
940 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
941 timeout_cb_t cb, void *data,
942 cfs_list_t *obd_list)
947 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
948 enum timeout_event event)
953 int ptlrpc_pinger_add_import(struct obd_import *imp)
956 if (!cfs_list_empty(&imp->imp_pinger_chain))
959 CDEBUG(D_HA, "adding pingable import %s->%s\n",
960 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
961 ptlrpc_pinger_sending_on_import(imp);
963 cfs_mutex_down(&pinger_sem);
964 cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
965 class_import_get(imp);
966 cfs_mutex_up(&pinger_sem);
971 int ptlrpc_pinger_del_import(struct obd_import *imp)
974 if (cfs_list_empty(&imp->imp_pinger_chain))
977 cfs_mutex_down(&pinger_sem);
978 cfs_list_del_init(&imp->imp_pinger_chain);
979 CDEBUG(D_HA, "removing pingable import %s->%s\n",
980 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
981 class_import_put(imp);
982 cfs_mutex_up(&pinger_sem);
986 void ptlrpc_pinger_wake_up()
989 /* XXX force pinger to run, if needed */
990 struct obd_import *imp;
992 cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
993 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
994 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
995 #ifdef ENABLE_LIBLUSTRE_RECOVERY
996 if (imp->imp_state == LUSTRE_IMP_DISCON &&
997 !imp_is_deactive(imp))
999 /*XXX only recover for the initial connection */
1000 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1001 imp->imp_state == LUSTRE_IMP_DISCON &&
1002 !imp_is_deactive(imp))
1004 ptlrpc_initiate_recovery(imp);
1005 else if (imp->imp_state != LUSTRE_IMP_FULL)
1006 CDEBUG(D_HA, "Refused to recover import %s->%s "
1007 "state %d, deactive %d\n",
1008 imp->imp_obd->obd_uuid.uuid,
1009 obd2cli_tgt(imp->imp_obd), imp->imp_state,
1010 imp_is_deactive(imp));
1015 #endif /* !__KERNEL__ */