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 <lustre_net.h>
50 #include "ptlrpc_internal.h"
52 struct semaphore pinger_sem;
53 static struct list_head pinger_imports = CFS_LIST_HEAD_INIT(pinger_imports);
54 static struct list_head timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
56 struct ptlrpc_request *
57 ptlrpc_prep_ping(struct obd_import *imp)
59 struct ptlrpc_request *req;
61 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION,
62 OBD_PING, 1, NULL, NULL);
64 ptlrpc_req_set_repsize(req, 1, NULL);
65 req->rq_no_resend = req->rq_no_delay = 1;
70 int ptlrpc_obd_ping(struct obd_device *obd)
73 struct ptlrpc_request *req;
76 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
80 req->rq_send_state = LUSTRE_IMP_FULL;
82 rc = ptlrpc_queue_wait(req);
84 ptlrpc_req_finished(req);
88 EXPORT_SYMBOL(ptlrpc_obd_ping);
90 int ptlrpc_ping(struct obd_import *imp)
92 struct ptlrpc_request *req;
96 req = ptlrpc_prep_ping(imp);
98 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
99 imp->imp_obd->obd_uuid.uuid,
100 obd2cli_tgt(imp->imp_obd));
102 /* To quickly detect server failure ping timeouts must be
103 * kept small. Therefore we must override/ignore the server
104 * rpc completion estimate which may be very large since
105 * it includes non-ping service times. The right long term
106 * fix will be to add a per-server (not per-service) thread
107 * in order to reduce the number of pings in the system in
108 * general (see bug 12471). */
110 req->rq_timeout = PING_SVC_TIMEOUT +
111 at_get(&imp->imp_at.iat_net_latency);
112 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
115 ptlrpcd_add_req(req);
117 CERROR("OOM trying to ping %s->%s\n",
118 imp->imp_obd->obd_uuid.uuid,
119 obd2cli_tgt(imp->imp_obd));
125 EXPORT_SYMBOL(ptlrpc_ping);
127 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
130 cfs_time_t delay, dtime, ctime = cfs_time_current();
132 if (imp->imp_state == LUSTRE_IMP_DISCON ||
133 imp->imp_state == LUSTRE_IMP_CONNECTING) {
134 /* In the disconnected case aggressively reconnect, for
135 * this request the AT service timeout will be set to
136 * INITIAL_CONNECT_TIMEOUT. To ensure the request times
137 * out before we send another we add one extra second. */
138 dtime = cfs_time_seconds(max_t(int, CONNECTION_SWITCH_MIN,
139 AT_OFF ? 0 : INITIAL_CONNECT_TIMEOUT + 1 +
140 at_get(&imp->imp_at.iat_net_latency)));
142 /* In the common case we want to cluster the pings at
143 * at regular intervals to minimize system noise. */
144 delay = cfs_time_seconds(soon ? PING_INTERVAL_SHORT :
146 dtime = delay - (ctime % delay);
148 /* May harmlessly race with ptlrpc_update_next_ping() */
149 imp->imp_next_ping = cfs_time_add(ctime, dtime);
151 CDEBUG(D_INFO, "Setting %s next ping to "CFS_TIME_T" ("CFS_TIME_T")\n",
152 obd2cli_tgt(imp->imp_obd), imp->imp_next_ping, dtime);
154 #endif /* ENABLE_PINGER */
157 void ptlrpc_ping_import_soon(struct obd_import *imp)
159 imp->imp_next_ping = cfs_time_current();
162 static inline int imp_is_deactive(struct obd_import *imp)
164 return (imp->imp_deactive ||
165 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
168 cfs_duration_t pinger_check_timeout(cfs_time_t time)
170 struct timeout_item *item;
171 cfs_time_t timeout = PING_INTERVAL;
173 /* The timeout list is a increase order sorted list */
174 mutex_down(&pinger_sem);
175 list_for_each_entry(item, &timeout_list, ti_chain) {
176 int ti_timeout = item->ti_timeout;
177 if (timeout > ti_timeout)
178 timeout = ti_timeout;
181 mutex_up(&pinger_sem);
183 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
188 static int ptlrpc_pinger_main(void *arg)
190 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
191 struct ptlrpc_thread *thread = data->thread;
194 cfs_daemonize(data->name);
196 /* Record that the thread is running */
197 thread->t_flags = SVC_RUNNING;
198 cfs_waitq_signal(&thread->t_ctl_waitq);
200 /* And now, loop forever, pinging as needed. */
202 cfs_time_t this_ping = cfs_time_current();
203 struct l_wait_info lwi;
204 cfs_duration_t time_to_next_wake;
205 cfs_time_t time_of_next_wake;
206 struct timeout_item *item;
207 struct list_head *iter;
209 time_to_next_wake = cfs_time_seconds(PING_INTERVAL);
210 time_of_next_wake = cfs_time_shift(PING_INTERVAL);
212 mutex_down(&pinger_sem);
213 list_for_each_entry(item, &timeout_list, ti_chain) {
214 item->ti_cb(item, item->ti_cb_data);
216 list_for_each(iter, &pinger_imports) {
217 struct obd_import *imp =
218 list_entry(iter, struct obd_import,
222 spin_lock(&imp->imp_lock);
223 level = imp->imp_state;
224 force = imp->imp_force_verify;
225 imp->imp_force_verify = 0;
226 spin_unlock(&imp->imp_lock);
228 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_RPCTRACE,
229 "level %s/%u force %u deactive %u pingable %u\n",
230 ptlrpc_import_state_name(level), level,
231 force, imp->imp_deactive, imp->imp_pingable);
233 /* Include any ping which misses the deadline by up to
234 * 1/10 of a second. The pings are designed to clump
235 * and this helps ensure the entire batch gets sent
236 * promptly, which minimizes system noise from pings */
239 cfs_time_aftereq(this_ping, imp->imp_next_ping -
240 (cfs_time_seconds(1) + 9) / 10)) {
241 if (level == LUSTRE_IMP_DISCON &&
242 !imp_is_deactive(imp)) {
243 ptlrpc_update_next_ping(imp, 0);
244 ptlrpc_initiate_recovery(imp);
245 } else if (level != LUSTRE_IMP_FULL ||
246 imp->imp_obd->obd_no_recov ||
247 imp_is_deactive(imp)) {
248 CDEBUG(D_HA, "not pinging %s "
249 "(in recovery: %s or recovery "
250 "disabled: %u/%u)\n",
251 obd2cli_tgt(imp->imp_obd),
252 ptlrpc_import_state_name(level),
254 imp->imp_obd->obd_no_recov);
255 } else if (imp->imp_pingable || force) {
257 /* ptlrpc_pinger_sending_on_import()
258 * will asynch update imp_next_ping
259 * so it must not be used below to
260 * calculate minimum wait time. */
264 if (!imp->imp_pingable)
267 "don't need to ping %s ("CFS_TIME_T
268 " > "CFS_TIME_T")\n",
269 obd2cli_tgt(imp->imp_obd),
270 imp->imp_next_ping, this_ping);
273 /* Wait time until next ping, or until we stopped. */
274 if (cfs_time_before(imp->imp_next_ping,
275 time_of_next_wake)) {
276 time_of_next_wake = imp->imp_next_ping;
277 time_to_next_wake = max_t(cfs_duration_t,
279 cfs_time_sub(time_of_next_wake,
280 cfs_time_current()));
283 mutex_up(&pinger_sem);
284 obd_update_maxusage();
285 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
286 time_to_next_wake, time_of_next_wake);
288 if (time_to_next_wake > 0) {
289 lwi = LWI_TIMEOUT(time_to_next_wake, NULL, NULL);
290 l_wait_event(thread->t_ctl_waitq,
291 thread->t_flags & (SVC_STOPPING|SVC_EVENT),
293 if (thread->t_flags & SVC_STOPPING) {
294 thread->t_flags &= ~SVC_STOPPING;
297 } else if (thread->t_flags & SVC_EVENT) {
298 /* woken after adding import to reset timer */
299 thread->t_flags &= ~SVC_EVENT;
304 thread->t_flags = SVC_STOPPED;
305 cfs_waitq_signal(&thread->t_ctl_waitq);
307 CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
311 static struct ptlrpc_thread *pinger_thread = NULL;
313 int ptlrpc_start_pinger(void)
315 struct l_wait_info lwi = { 0 };
316 struct ptlrpc_svc_data d;
318 #ifndef ENABLE_PINGER
323 if (pinger_thread != NULL)
326 OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
327 if (pinger_thread == NULL)
329 cfs_waitq_init(&pinger_thread->t_ctl_waitq);
332 d.thread = pinger_thread;
334 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
335 * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
336 rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
338 CERROR("cannot start thread: %d\n", rc);
339 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
340 pinger_thread = NULL;
343 l_wait_event(pinger_thread->t_ctl_waitq,
344 pinger_thread->t_flags & SVC_RUNNING, &lwi);
349 int ptlrpc_pinger_remove_timeouts(void);
351 int ptlrpc_stop_pinger(void)
353 struct l_wait_info lwi = { 0 };
355 #ifndef ENABLE_PINGER
360 if (pinger_thread == NULL)
363 ptlrpc_pinger_remove_timeouts();
364 mutex_down(&pinger_sem);
365 pinger_thread->t_flags = SVC_STOPPING;
366 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
367 mutex_up(&pinger_sem);
369 l_wait_event(pinger_thread->t_ctl_waitq,
370 (pinger_thread->t_flags & SVC_STOPPED), &lwi);
372 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
373 pinger_thread = NULL;
377 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
379 ptlrpc_update_next_ping(imp, 0);
382 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
384 ptlrpc_update_next_ping(imp, 1);
387 int ptlrpc_pinger_add_import(struct obd_import *imp)
390 if (!list_empty(&imp->imp_pinger_chain))
393 mutex_down(&pinger_sem);
394 CDEBUG(D_HA, "adding pingable import %s->%s\n",
395 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
396 /* if we add to pinger we want recovery on this import */
397 imp->imp_obd->obd_no_recov = 0;
399 ptlrpc_update_next_ping(imp, 0);
400 /* XXX sort, blah blah */
401 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
402 class_import_get(imp);
404 ptlrpc_pinger_wake_up();
405 mutex_up(&pinger_sem);
410 int ptlrpc_pinger_del_import(struct obd_import *imp)
413 if (list_empty(&imp->imp_pinger_chain))
416 mutex_down(&pinger_sem);
417 list_del_init(&imp->imp_pinger_chain);
418 CDEBUG(D_HA, "removing pingable import %s->%s\n",
419 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
420 /* if we remove from pinger we don't want recovery on this import */
421 imp->imp_obd->obd_no_recov = 1;
422 class_import_put(imp);
423 mutex_up(&pinger_sem);
428 * Register a timeout callback to the pinger list, and the callback will
429 * be called when timeout happens.
431 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
432 timeout_cb_t cb, void *data)
434 struct timeout_item *ti;
440 CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
441 CFS_INIT_LIST_HEAD(&ti->ti_chain);
442 ti->ti_timeout = time;
443 ti->ti_event = event;
445 ti->ti_cb_data = data;
451 * Register timeout event on the the pinger thread.
452 * Note: the timeout list is an sorted list with increased timeout value.
454 static struct timeout_item*
455 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
456 timeout_cb_t cb, void *data)
458 struct timeout_item *item, *tmp;
460 LASSERT_SEM_LOCKED(&pinger_sem);
462 list_for_each_entry(item, &timeout_list, ti_chain)
463 if (item->ti_event == event)
466 item = ptlrpc_new_timeout(time, event, cb, data);
468 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
469 if (tmp->ti_timeout < time) {
470 list_add(&item->ti_chain, &tmp->ti_chain);
474 list_add(&item->ti_chain, &timeout_list);
480 /* Add a client_obd to the timeout event list, when timeout(@time)
481 * happens, the callback(@cb) will be called.
483 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
484 timeout_cb_t cb, void *data,
485 struct list_head *obd_list)
487 struct timeout_item *ti;
489 mutex_down(&pinger_sem);
490 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
492 mutex_up(&pinger_sem);
495 list_add(obd_list, &ti->ti_obd_list);
496 mutex_up(&pinger_sem);
500 int ptlrpc_del_timeout_client(struct list_head *obd_list,
501 enum timeout_event event)
503 struct timeout_item *ti = NULL, *item;
505 if (list_empty(obd_list))
507 mutex_down(&pinger_sem);
508 list_del_init(obd_list);
510 * If there are no obd attached to the timeout event
511 * list, remove this timeout event from the pinger
513 list_for_each_entry(item, &timeout_list, ti_chain) {
514 if (item->ti_event == event) {
519 LASSERTF(ti != NULL, "ti is NULL ! \n");
520 if (list_empty(&ti->ti_obd_list)) {
521 list_del(&ti->ti_chain);
524 mutex_up(&pinger_sem);
528 int ptlrpc_pinger_remove_timeouts(void)
530 struct timeout_item *item, *tmp;
532 mutex_down(&pinger_sem);
533 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
534 LASSERT(list_empty(&item->ti_obd_list));
535 list_del(&item->ti_chain);
538 mutex_up(&pinger_sem);
542 void ptlrpc_pinger_wake_up()
545 pinger_thread->t_flags |= SVC_EVENT;
546 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
550 /* Ping evictor thread */
552 #define PET_TERMINATE 2
554 static int pet_refcount = 0;
555 static int pet_state;
556 static wait_queue_head_t pet_waitq;
557 CFS_LIST_HEAD(pet_list);
558 static spinlock_t pet_lock = SPIN_LOCK_UNLOCKED;
560 int ping_evictor_wake(struct obd_export *exp)
562 struct obd_device *obd;
564 spin_lock(&pet_lock);
565 if (pet_state != PET_READY) {
566 /* eventually the new obd will call here again. */
567 spin_unlock(&pet_lock);
571 obd = class_exp2obd(exp);
572 if (list_empty(&obd->obd_evict_list)) {
574 list_add(&obd->obd_evict_list, &pet_list);
576 spin_unlock(&pet_lock);
582 static int ping_evictor_main(void *arg)
584 struct obd_device *obd;
585 struct obd_export *exp;
586 struct l_wait_info lwi = { 0 };
590 cfs_daemonize_ctxt("ll_evictor");
592 CDEBUG(D_HA, "Starting Ping Evictor\n");
593 pet_state = PET_READY;
595 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
596 (pet_state == PET_TERMINATE), &lwi);
598 /* loop until all obd's will be removed */
599 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
602 /* we only get here if pet_exp != NULL, and the end of this
603 * loop is the only place which sets it NULL again, so lock
604 * is not strictly necessary. */
605 spin_lock(&pet_lock);
606 obd = list_entry(pet_list.next, struct obd_device,
608 spin_unlock(&pet_lock);
610 /* bug 18948: ensure recovery is aborted in a timely fashion */
611 if (target_recovery_check_and_stop(obd) ||
612 obd->obd_recovering /* no evictor during recovery */)
615 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
617 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
618 obd->obd_name, expire_time);
620 /* Exports can't be deleted out of the list while we hold
621 * the obd lock (class_unlink_export), which means we can't
622 * lose the last ref on the export. If they've already been
623 * removed from the list, we won't find them here. */
624 spin_lock(&obd->obd_dev_lock);
625 while (!list_empty(&obd->obd_exports_timed)) {
626 exp = list_entry(obd->obd_exports_timed.next,
627 struct obd_export,exp_obd_chain_timed);
628 if (expire_time > exp->exp_last_request_time) {
629 class_export_get(exp);
630 spin_unlock(&obd->obd_dev_lock);
631 LCONSOLE_WARN("%s: haven't heard from client %s"
632 " (at %s) in %ld seconds. I think"
633 " it's dead, and I am evicting"
634 " it.\n", obd->obd_name,
635 obd_uuid2str(&exp->exp_client_uuid),
636 obd_export_nid2str(exp),
637 (long)(cfs_time_current_sec() -
638 exp->exp_last_request_time));
639 CDEBUG(D_HA, "Last request was at %ld\n",
640 exp->exp_last_request_time);
641 class_fail_export(exp);
642 class_export_put(exp);
643 spin_lock(&obd->obd_dev_lock);
645 /* List is sorted, so everyone below is ok */
649 spin_unlock(&obd->obd_dev_lock);
651 spin_lock(&pet_lock);
652 list_del_init(&obd->obd_evict_list);
653 spin_unlock(&pet_lock);
658 CDEBUG(D_HA, "Exiting Ping Evictor\n");
663 void ping_evictor_start(void)
667 if (++pet_refcount > 1)
670 init_waitqueue_head(&pet_waitq);
672 rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
675 CERROR("Cannot start ping evictor thread: %d\n", rc);
678 EXPORT_SYMBOL(ping_evictor_start);
680 void ping_evictor_stop(void)
682 if (--pet_refcount > 0)
685 pet_state = PET_TERMINATE;
688 EXPORT_SYMBOL(ping_evictor_stop);
689 #else /* !__KERNEL__ */
692 * the current implementation of pinger in liblustre is not optimized
696 static struct pinger_data {
698 cfs_time_t pd_this_ping; /* jiffies */
699 cfs_time_t pd_next_ping; /* jiffies */
700 struct ptlrpc_request_set *pd_set;
703 static int pinger_check_rpcs(void *arg)
705 cfs_time_t curtime = cfs_time_current();
706 struct ptlrpc_request *req;
707 struct ptlrpc_request_set *set;
708 struct list_head *iter;
709 struct obd_import *imp;
710 struct pinger_data *pd = &pinger_args;
713 /* prevent recursion */
714 if (pd->pd_recursion++) {
715 CDEBUG(D_HA, "pinger: recursion! quit\n");
721 /* have we reached ping point? */
722 if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
727 /* if we have rpc_set already, continue processing it */
729 LASSERT(pd->pd_this_ping);
734 pd->pd_this_ping = curtime;
735 pd->pd_set = ptlrpc_prep_set();
736 if (pd->pd_set == NULL)
740 /* add rpcs into set */
741 mutex_down(&pinger_sem);
742 list_for_each(iter, &pinger_imports) {
743 struct obd_import *imp =
744 list_entry(iter, struct obd_import, imp_pinger_chain);
745 int generation, level;
747 /* Include any ping within 1/10 of a second of the deadline */
748 if (cfs_time_aftereq(pd->pd_this_ping, imp->imp_next_ping -
749 (cfs_time_seconds(1) + 9) / 10)) {
751 spin_lock(&imp->imp_lock);
752 generation = imp->imp_generation;
753 level = imp->imp_state;
754 spin_unlock(&imp->imp_lock);
756 if (level != LUSTRE_IMP_FULL) {
758 "not pinging %s (in recovery)\n",
759 obd2cli_tgt(imp->imp_obd));
763 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING,
766 CERROR("out of memory\n");
769 req->rq_no_resend = 1;
770 ptlrpc_req_set_repsize(req, 1, NULL);
771 req->rq_send_state = LUSTRE_IMP_FULL;
772 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
773 req->rq_import_generation = generation;
774 ptlrpc_set_add_req(set, req);
776 CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
777 " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
778 imp->imp_next_ping, pd->pd_this_ping);
781 pd->pd_this_ping = curtime;
782 mutex_up(&pinger_sem);
784 /* Might be empty, that's OK. */
785 if (set->set_remaining == 0)
786 CDEBUG(D_RPCTRACE, "nothing to ping\n");
788 list_for_each(iter, &set->set_requests) {
789 struct ptlrpc_request *req =
790 list_entry(iter, struct ptlrpc_request,
792 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
793 req->rq_import->imp_obd->obd_uuid.uuid,
794 obd2cli_tgt(req->rq_import->imp_obd));
795 (void)ptl_send_rpc(req, 0);
799 rc = ptlrpc_check_set(set);
801 /* not finished, and we are not expired, simply return */
802 if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
803 cfs_time_seconds(PING_INTERVAL)))) {
804 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
809 /* Expire all the requests that didn't come back. */
810 mutex_down(&pinger_sem);
811 list_for_each(iter, &set->set_requests) {
812 req = list_entry(iter, struct ptlrpc_request,
815 if (req->rq_phase == RQ_PHASE_COMPLETE)
818 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
821 /* This will also unregister reply. */
822 ptlrpc_expire_one_request(req, 0);
824 /* We're done with this req, let's finally move it to complete
825 * phase and take care of inflights. */
826 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
827 imp = req->rq_import;
828 spin_lock(&imp->imp_lock);
829 if (!list_empty(&req->rq_list)) {
830 list_del_init(&req->rq_list);
831 atomic_dec(&imp->imp_inflight);
833 spin_unlock(&imp->imp_lock);
834 set->set_remaining--;
836 mutex_up(&pinger_sem);
838 ptlrpc_set_destroy(set);
842 pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
843 cfs_time_seconds(PING_INTERVAL));
844 pd->pd_this_ping = 0; /* XXX for debug */
846 CDEBUG(D_INFO, "finished a round ping\n");
851 static void *pinger_callback = NULL;
852 #endif /* ENABLE_PINGER */
854 int ptlrpc_start_pinger(void)
857 memset(&pinger_args, 0, sizeof(pinger_args));
858 pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
865 int ptlrpc_stop_pinger(void)
869 liblustre_deregister_wait_callback(pinger_callback);
874 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
877 mutex_down(&pinger_sem);
878 ptlrpc_update_next_ping(imp, 0);
879 if (pinger_args.pd_set == NULL &&
880 time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
881 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
882 imp->imp_next_ping, cfs_time_current());
883 pinger_args.pd_next_ping = imp->imp_next_ping;
885 mutex_up(&pinger_sem);
889 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
890 timeout_cb_t cb, void *data,
891 struct list_head *obd_list)
896 int ptlrpc_del_timeout_client(struct list_head *obd_list,
897 enum timeout_event event)
902 int ptlrpc_pinger_add_import(struct obd_import *imp)
905 if (!list_empty(&imp->imp_pinger_chain))
908 CDEBUG(D_HA, "adding pingable import %s->%s\n",
909 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
910 ptlrpc_pinger_sending_on_import(imp);
912 mutex_down(&pinger_sem);
913 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
914 class_import_get(imp);
915 mutex_up(&pinger_sem);
920 int ptlrpc_pinger_del_import(struct obd_import *imp)
923 if (list_empty(&imp->imp_pinger_chain))
926 mutex_down(&pinger_sem);
927 list_del_init(&imp->imp_pinger_chain);
928 CDEBUG(D_HA, "removing pingable import %s->%s\n",
929 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
930 class_import_put(imp);
931 mutex_up(&pinger_sem);
935 void ptlrpc_pinger_wake_up()
939 /* XXX force pinger to run, if needed */
940 struct obd_import *imp;
941 list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
942 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
943 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
944 #ifdef ENABLE_LIBLUSTRE_RECOVERY
945 if (imp->imp_state == LUSTRE_IMP_DISCON &&
946 !imp_is_deactive(imp))
948 /*XXX only recover for the initial connection */
949 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
950 imp->imp_state == LUSTRE_IMP_DISCON &&
951 !imp_is_deactive(imp))
953 ptlrpc_initiate_recovery(imp);
954 else if (imp->imp_state != LUSTRE_IMP_FULL)
955 CDEBUG(D_HA, "Refused to recover import %s->%s "
956 "state %d, deactive %d\n",
957 imp->imp_obd->obd_uuid.uuid,
958 obd2cli_tgt(imp->imp_obd), imp->imp_state,
959 imp_is_deactive(imp));
964 #endif /* !__KERNEL__ */