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, Intel Corporation.
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"
52 EXPORT_SYMBOL(suppress_pings);
53 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
55 struct mutex pinger_mutex;
56 static CFS_LIST_HEAD(pinger_imports);
57 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
59 struct ptlrpc_request *
60 ptlrpc_prep_ping(struct obd_import *imp)
62 struct ptlrpc_request *req;
64 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
65 LUSTRE_OBD_VERSION, OBD_PING);
67 ptlrpc_request_set_replen(req);
68 req->rq_no_resend = req->rq_no_delay = 1;
73 int ptlrpc_obd_ping(struct obd_device *obd)
76 struct ptlrpc_request *req;
79 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
83 req->rq_send_state = LUSTRE_IMP_FULL;
85 rc = ptlrpc_queue_wait(req);
87 ptlrpc_req_finished(req);
91 EXPORT_SYMBOL(ptlrpc_obd_ping);
93 int ptlrpc_ping(struct obd_import *imp)
95 struct ptlrpc_request *req;
98 req = ptlrpc_prep_ping(imp);
100 CERROR("OOM trying to ping %s->%s\n",
101 imp->imp_obd->obd_uuid.uuid,
102 obd2cli_tgt(imp->imp_obd));
106 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
107 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
108 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
113 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
116 int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
117 if (imp->imp_state == LUSTRE_IMP_DISCON) {
118 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
120 at_get(&imp->imp_at.iat_net_latency));
121 time = min(time, dtime);
123 imp->imp_next_ping = cfs_time_shift(time);
124 #endif /* ENABLE_PINGER */
127 void ptlrpc_ping_import_soon(struct obd_import *imp)
129 imp->imp_next_ping = cfs_time_current();
132 static inline int imp_is_deactive(struct obd_import *imp)
134 return (imp->imp_deactive ||
135 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
138 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
140 if (imp->imp_server_timeout)
141 return cfs_time_shift(obd_timeout / 2);
143 return cfs_time_shift(obd_timeout);
146 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
147 static cfs_time_t suspend_wakeup_time = 0;
149 cfs_duration_t pinger_check_timeout(cfs_time_t time)
151 struct timeout_item *item;
152 cfs_time_t timeout = PING_INTERVAL;
154 /* The timeout list is a increase order sorted list */
155 mutex_lock(&pinger_mutex);
156 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
157 int ti_timeout = item->ti_timeout;
158 if (timeout > ti_timeout)
159 timeout = ti_timeout;
162 mutex_unlock(&pinger_mutex);
164 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
168 static cfs_waitq_t suspend_timeouts_waitq;
170 cfs_time_t ptlrpc_suspend_wakeup_time(void)
172 return suspend_wakeup_time;
175 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
177 /*XXX: disabled for now, will be replaced by adaptive timeouts */
179 if (imp->imp_no_timeout)
181 imp->imp_no_timeout = 1;
182 cfs_atomic_inc(&suspend_timeouts);
183 CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
184 cfs_atomic_read(&suspend_timeouts));
188 void ptlrpc_activate_timeouts(struct obd_import *imp)
190 /*XXX: disabled for now, will be replaced by adaptive timeouts */
192 if (!imp->imp_no_timeout)
194 imp->imp_no_timeout = 0;
195 LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
196 if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
197 suspend_wakeup_time = cfs_time_current();
198 cfs_waitq_signal(&suspend_timeouts_waitq);
200 CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
201 cfs_atomic_read(&suspend_timeouts));
205 int ptlrpc_check_suspend(void)
207 if (cfs_atomic_read(&suspend_timeouts))
212 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
214 struct l_wait_info lwi;
216 if (cfs_atomic_read(&suspend_timeouts)) {
217 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
218 cfs_atomic_read(&suspend_timeouts));
219 lwi = LWI_INTR(NULL, NULL);
220 l_wait_event(suspend_timeouts_waitq,
221 cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
222 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
230 static void ptlrpc_pinger_process_import(struct obd_import *imp,
231 unsigned long this_ping)
238 spin_lock(&imp->imp_lock);
240 level = imp->imp_state;
241 force = imp->imp_force_verify;
242 force_next = imp->imp_force_next_verify;
244 * This will be used below only if the import is "FULL".
246 suppress = !!(imp->imp_connect_data.ocd_connect_flags &
247 OBD_CONNECT_PINGLESS);
249 imp->imp_force_verify = 0;
251 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
253 spin_unlock(&imp->imp_lock);
257 imp->imp_force_next_verify = 0;
259 spin_unlock(&imp->imp_lock);
261 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
262 "force %u force_next %u deactive %u pingable %u suppress %u\n",
263 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
264 ptlrpc_import_state_name(level), level, force, force_next,
265 imp->imp_deactive, imp->imp_pingable, suppress);
267 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
268 /* wait for a while before trying recovery again */
269 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
270 if (!imp->imp_no_pinger_recover)
271 ptlrpc_initiate_recovery(imp);
272 } else if (level != LUSTRE_IMP_FULL ||
273 imp->imp_obd->obd_no_recov ||
274 imp_is_deactive(imp)) {
275 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
276 "or recovery disabled: %s)\n",
277 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
278 ptlrpc_import_state_name(level));
279 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
284 static int ptlrpc_pinger_main(void *arg)
286 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
289 cfs_daemonize(thread->t_name);
291 /* Record that the thread is running */
292 thread_set_flags(thread, SVC_RUNNING);
293 cfs_waitq_signal(&thread->t_ctl_waitq);
295 /* And now, loop forever, pinging as needed. */
297 cfs_time_t this_ping = cfs_time_current();
298 struct l_wait_info lwi;
299 cfs_duration_t time_to_next_wake;
300 struct timeout_item *item;
303 mutex_lock(&pinger_mutex);
304 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
305 item->ti_cb(item, item->ti_cb_data);
307 cfs_list_for_each(iter, &pinger_imports) {
308 struct obd_import *imp =
309 cfs_list_entry(iter, struct obd_import,
312 ptlrpc_pinger_process_import(imp, this_ping);
313 /* obd_timeout might have changed */
314 if (imp->imp_pingable && imp->imp_next_ping &&
315 cfs_time_after(imp->imp_next_ping,
316 cfs_time_add(this_ping,
317 cfs_time_seconds(PING_INTERVAL))))
318 ptlrpc_update_next_ping(imp, 0);
320 mutex_unlock(&pinger_mutex);
321 /* update memory usage info */
322 obd_update_maxusage();
324 /* Wait until the next ping time, or until we're stopped. */
325 time_to_next_wake = pinger_check_timeout(this_ping);
326 /* The ping sent by ptlrpc_send_rpc may get sent out
327 say .01 second after this.
328 ptlrpc_pinger_sending_on_import will then set the
329 next ping time to next_ping + .01 sec, which means
330 we will SKIP the next ping at next_ping, and the
331 ping will get sent 2 timeouts from now! Beware. */
332 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
333 CFS_TIME_T")\n", time_to_next_wake,
334 cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
335 if (time_to_next_wake > 0) {
336 lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
338 cfs_time_seconds(1)),
340 l_wait_event(thread->t_ctl_waitq,
341 thread_is_stopping(thread) ||
342 thread_is_event(thread),
344 if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
348 /* woken after adding import to reset timer */
349 thread_test_and_clear_flags(thread, SVC_EVENT);
354 thread_set_flags(thread, SVC_STOPPED);
355 cfs_waitq_signal(&thread->t_ctl_waitq);
357 CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
361 static struct ptlrpc_thread *pinger_thread = NULL;
363 int ptlrpc_start_pinger(void)
365 struct l_wait_info lwi = { 0 };
367 #ifndef ENABLE_PINGER
372 if (pinger_thread != NULL)
375 OBD_ALLOC_PTR(pinger_thread);
376 if (pinger_thread == NULL)
378 cfs_waitq_init(&pinger_thread->t_ctl_waitq);
379 cfs_waitq_init(&suspend_timeouts_waitq);
381 strcpy(pinger_thread->t_name, "ll_ping");
383 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
384 * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
385 rc = cfs_create_thread(ptlrpc_pinger_main,
386 pinger_thread, CFS_DAEMON_FLAGS);
388 CERROR("cannot start thread: %d\n", rc);
389 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
390 pinger_thread = NULL;
393 l_wait_event(pinger_thread->t_ctl_waitq,
394 thread_is_running(pinger_thread), &lwi);
397 CWARN("Pings will be suppressed at the request of the "
398 "administrator. The configuration shall meet the "
399 "additional requirements described in the manual. "
400 "(Search for the \"suppress_pings\" kernel module "
406 int ptlrpc_pinger_remove_timeouts(void);
408 int ptlrpc_stop_pinger(void)
410 struct l_wait_info lwi = { 0 };
412 #ifndef ENABLE_PINGER
417 if (pinger_thread == NULL)
420 ptlrpc_pinger_remove_timeouts();
421 mutex_lock(&pinger_mutex);
422 thread_set_flags(pinger_thread, SVC_STOPPING);
423 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
424 mutex_unlock(&pinger_mutex);
426 l_wait_event(pinger_thread->t_ctl_waitq,
427 thread_is_stopped(pinger_thread), &lwi);
429 OBD_FREE_PTR(pinger_thread);
430 pinger_thread = NULL;
434 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
436 ptlrpc_update_next_ping(imp, 0);
438 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
440 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
442 ptlrpc_update_next_ping(imp, 1);
443 LASSERT_SPIN_LOCKED(&imp->imp_lock);
445 * Avoid reading stale imp_connect_data. When not sure if pings are
446 * expected or not on next connection, we assume they are not and force
447 * one anyway to guarantee the chance of updating
448 * imp_peer_committed_transno.
450 if (imp->imp_state != LUSTRE_IMP_FULL ||
451 imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_PINGLESS)
452 imp->imp_force_next_verify = 1;
455 int ptlrpc_pinger_add_import(struct obd_import *imp)
458 if (!cfs_list_empty(&imp->imp_pinger_chain))
461 mutex_lock(&pinger_mutex);
462 CDEBUG(D_HA, "adding pingable import %s->%s\n",
463 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
464 /* if we add to pinger we want recovery on this import */
465 imp->imp_obd->obd_no_recov = 0;
466 ptlrpc_update_next_ping(imp, 0);
467 /* XXX sort, blah blah */
468 cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
469 class_import_get(imp);
471 ptlrpc_pinger_wake_up();
472 mutex_unlock(&pinger_mutex);
476 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
478 int ptlrpc_pinger_del_import(struct obd_import *imp)
481 if (cfs_list_empty(&imp->imp_pinger_chain))
484 mutex_lock(&pinger_mutex);
485 cfs_list_del_init(&imp->imp_pinger_chain);
486 CDEBUG(D_HA, "removing pingable import %s->%s\n",
487 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
488 /* if we remove from pinger we don't want recovery on this import */
489 imp->imp_obd->obd_no_recov = 1;
490 class_import_put(imp);
491 mutex_unlock(&pinger_mutex);
494 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
497 * Register a timeout callback to the pinger list, and the callback will
498 * be called when timeout happens.
500 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
501 timeout_cb_t cb, void *data)
503 struct timeout_item *ti;
509 CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
510 CFS_INIT_LIST_HEAD(&ti->ti_chain);
511 ti->ti_timeout = time;
512 ti->ti_event = event;
514 ti->ti_cb_data = data;
520 * Register timeout event on the the pinger thread.
521 * Note: the timeout list is an sorted list with increased timeout value.
523 static struct timeout_item*
524 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
525 timeout_cb_t cb, void *data)
527 struct timeout_item *item, *tmp;
529 LASSERT_MUTEX_LOCKED(&pinger_mutex);
531 cfs_list_for_each_entry(item, &timeout_list, ti_chain)
532 if (item->ti_event == event)
535 item = ptlrpc_new_timeout(time, event, cb, data);
537 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
538 if (tmp->ti_timeout < time) {
539 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
543 cfs_list_add(&item->ti_chain, &timeout_list);
549 /* Add a client_obd to the timeout event list, when timeout(@time)
550 * happens, the callback(@cb) will be called.
552 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
553 timeout_cb_t cb, void *data,
554 cfs_list_t *obd_list)
556 struct timeout_item *ti;
558 mutex_lock(&pinger_mutex);
559 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
561 mutex_unlock(&pinger_mutex);
564 cfs_list_add(obd_list, &ti->ti_obd_list);
565 mutex_unlock(&pinger_mutex);
568 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
570 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
571 enum timeout_event event)
573 struct timeout_item *ti = NULL, *item;
575 if (cfs_list_empty(obd_list))
577 mutex_lock(&pinger_mutex);
578 cfs_list_del_init(obd_list);
580 * If there are no obd attached to the timeout event
581 * list, remove this timeout event from the pinger
583 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
584 if (item->ti_event == event) {
589 LASSERTF(ti != NULL, "ti is NULL ! \n");
590 if (cfs_list_empty(&ti->ti_obd_list)) {
591 cfs_list_del(&ti->ti_chain);
594 mutex_unlock(&pinger_mutex);
597 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
599 int ptlrpc_pinger_remove_timeouts(void)
601 struct timeout_item *item, *tmp;
603 mutex_lock(&pinger_mutex);
604 cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
605 LASSERT(cfs_list_empty(&item->ti_obd_list));
606 cfs_list_del(&item->ti_chain);
609 mutex_unlock(&pinger_mutex);
613 void ptlrpc_pinger_wake_up()
616 thread_add_flags(pinger_thread, SVC_EVENT);
617 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
621 /* Ping evictor thread */
623 #define PET_TERMINATE 2
625 static int pet_refcount = 0;
626 static int pet_state;
627 static cfs_waitq_t pet_waitq;
628 CFS_LIST_HEAD(pet_list);
629 static DEFINE_SPINLOCK(pet_lock);
631 int ping_evictor_wake(struct obd_export *exp)
633 struct obd_device *obd;
635 spin_lock(&pet_lock);
636 if (pet_state != PET_READY) {
637 /* eventually the new obd will call here again. */
638 spin_unlock(&pet_lock);
642 obd = class_exp2obd(exp);
643 if (cfs_list_empty(&obd->obd_evict_list)) {
644 class_incref(obd, "evictor", obd);
645 cfs_list_add(&obd->obd_evict_list, &pet_list);
647 spin_unlock(&pet_lock);
649 cfs_waitq_signal(&pet_waitq);
653 static int ping_evictor_main(void *arg)
655 struct obd_device *obd;
656 struct obd_export *exp;
657 struct l_wait_info lwi = { 0 };
661 cfs_daemonize_ctxt("ll_evictor");
663 CDEBUG(D_HA, "Starting Ping Evictor\n");
664 pet_state = PET_READY;
666 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
667 (pet_state == PET_TERMINATE), &lwi);
669 /* loop until all obd's will be removed */
670 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
673 /* we only get here if pet_exp != NULL, and the end of this
674 * loop is the only place which sets it NULL again, so lock
675 * is not strictly necessary. */
676 spin_lock(&pet_lock);
677 obd = cfs_list_entry(pet_list.next, struct obd_device,
679 spin_unlock(&pet_lock);
681 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
683 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
684 obd->obd_name, expire_time);
686 /* Exports can't be deleted out of the list while we hold
687 * the obd lock (class_unlink_export), which means we can't
688 * lose the last ref on the export. If they've already been
689 * removed from the list, we won't find them here. */
690 spin_lock(&obd->obd_dev_lock);
691 while (!cfs_list_empty(&obd->obd_exports_timed)) {
692 exp = cfs_list_entry(obd->obd_exports_timed.next,
694 exp_obd_chain_timed);
695 if (expire_time > exp->exp_last_request_time) {
696 class_export_get(exp);
697 spin_unlock(&obd->obd_dev_lock);
698 LCONSOLE_WARN("%s: haven't heard from client %s"
699 " (at %s) in %ld seconds. I think"
700 " it's dead, and I am evicting"
701 " it. exp %p, cur %ld expire %ld"
704 obd_uuid2str(&exp->exp_client_uuid),
705 obd_export_nid2str(exp),
706 (long)(cfs_time_current_sec() -
707 exp->exp_last_request_time),
708 exp, (long)cfs_time_current_sec(),
710 (long)exp->exp_last_request_time);
711 CDEBUG(D_HA, "Last request was at %ld\n",
712 exp->exp_last_request_time);
713 class_fail_export(exp);
714 class_export_put(exp);
715 spin_lock(&obd->obd_dev_lock);
717 /* List is sorted, so everyone below is ok */
721 spin_unlock(&obd->obd_dev_lock);
723 spin_lock(&pet_lock);
724 cfs_list_del_init(&obd->obd_evict_list);
725 spin_unlock(&pet_lock);
727 class_decref(obd, "evictor", obd);
729 CDEBUG(D_HA, "Exiting Ping Evictor\n");
734 void ping_evictor_start(void)
738 if (++pet_refcount > 1)
741 cfs_waitq_init(&pet_waitq);
743 rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
746 CERROR("Cannot start ping evictor thread: %d\n", rc);
749 EXPORT_SYMBOL(ping_evictor_start);
751 void ping_evictor_stop(void)
753 if (--pet_refcount > 0)
756 pet_state = PET_TERMINATE;
757 cfs_waitq_signal(&pet_waitq);
759 EXPORT_SYMBOL(ping_evictor_stop);
760 #else /* !__KERNEL__ */
763 * the current implementation of pinger in liblustre is not optimized
767 static struct pinger_data {
769 cfs_time_t pd_this_ping; /* jiffies */
770 cfs_time_t pd_next_ping; /* jiffies */
771 struct ptlrpc_request_set *pd_set;
774 static int pinger_check_rpcs(void *arg)
776 cfs_time_t curtime = cfs_time_current();
777 struct ptlrpc_request *req;
778 struct ptlrpc_request_set *set;
780 struct obd_import *imp;
781 struct pinger_data *pd = &pinger_args;
784 /* prevent recursion */
785 if (pd->pd_recursion++) {
786 CDEBUG(D_HA, "pinger: recursion! quit\n");
792 /* have we reached ping point? */
793 if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
798 /* if we have rpc_set already, continue processing it */
800 LASSERT(pd->pd_this_ping);
805 pd->pd_this_ping = curtime;
806 pd->pd_set = ptlrpc_prep_set();
807 if (pd->pd_set == NULL)
811 /* add rpcs into set */
812 mutex_lock(&pinger_mutex);
813 cfs_list_for_each(iter, &pinger_imports) {
814 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
816 int generation, level;
818 if (cfs_time_aftereq(pd->pd_this_ping,
819 imp->imp_next_ping - 5 * CFS_TICK)) {
821 spin_lock(&imp->imp_lock);
822 generation = imp->imp_generation;
823 level = imp->imp_state;
824 spin_unlock(&imp->imp_lock);
826 if (level != LUSTRE_IMP_FULL) {
828 "not pinging %s (in recovery)\n",
829 obd2cli_tgt(imp->imp_obd));
833 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
837 CERROR("OOM trying to ping %s->%s\n",
838 imp->imp_obd->obd_uuid.uuid,
839 obd2cli_tgt(imp->imp_obd));
843 req->rq_no_resend = 1;
844 ptlrpc_request_set_replen(req);
845 req->rq_send_state = LUSTRE_IMP_FULL;
846 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
847 req->rq_import_generation = generation;
848 ptlrpc_set_add_req(set, req);
850 CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
851 " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
852 imp->imp_next_ping, pd->pd_this_ping);
855 pd->pd_this_ping = curtime;
856 mutex_unlock(&pinger_mutex);
858 /* Might be empty, that's OK. */
859 if (cfs_atomic_read(&set->set_remaining) == 0)
860 CDEBUG(D_RPCTRACE, "nothing to ping\n");
862 cfs_list_for_each(iter, &set->set_requests) {
863 struct ptlrpc_request *req =
864 cfs_list_entry(iter, struct ptlrpc_request,
866 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
867 req->rq_import->imp_obd->obd_uuid.uuid,
868 obd2cli_tgt(req->rq_import->imp_obd));
869 (void)ptl_send_rpc(req, 0);
873 rc = ptlrpc_check_set(NULL, set);
875 /* not finished, and we are not expired, simply return */
876 if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
877 cfs_time_seconds(PING_INTERVAL)))) {
878 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
883 /* Expire all the requests that didn't come back. */
884 mutex_lock(&pinger_mutex);
885 cfs_list_for_each(iter, &set->set_requests) {
886 req = cfs_list_entry(iter, struct ptlrpc_request,
889 if (req->rq_phase == RQ_PHASE_COMPLETE)
892 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
895 /* This will also unregister reply. */
896 ptlrpc_expire_one_request(req, 0);
898 /* We're done with this req, let's finally move it to complete
899 * phase and take care of inflights. */
900 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
901 imp = req->rq_import;
902 spin_lock(&imp->imp_lock);
903 if (!cfs_list_empty(&req->rq_list)) {
904 cfs_list_del_init(&req->rq_list);
905 cfs_atomic_dec(&imp->imp_inflight);
907 spin_unlock(&imp->imp_lock);
908 cfs_atomic_dec(&set->set_remaining);
910 mutex_unlock(&pinger_mutex);
912 ptlrpc_set_destroy(set);
916 pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
917 cfs_time_seconds(PING_INTERVAL));
918 pd->pd_this_ping = 0; /* XXX for debug */
920 CDEBUG(D_INFO, "finished a round ping\n");
925 static void *pinger_callback = NULL;
926 #endif /* ENABLE_PINGER */
928 int ptlrpc_start_pinger(void)
931 memset(&pinger_args, 0, sizeof(pinger_args));
932 pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
939 int ptlrpc_stop_pinger(void)
943 liblustre_deregister_wait_callback(pinger_callback);
948 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
951 mutex_lock(&pinger_mutex);
952 ptlrpc_update_next_ping(imp, 0);
953 if (pinger_args.pd_set == NULL &&
954 cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
955 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
956 imp->imp_next_ping, cfs_time_current());
957 pinger_args.pd_next_ping = imp->imp_next_ping;
959 mutex_unlock(&pinger_mutex);
963 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
966 mutex_lock(&pinger_mutex);
967 ptlrpc_update_next_ping(imp, 1);
968 if (pinger_args.pd_set == NULL &&
969 cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
970 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
971 imp->imp_next_ping, cfs_time_current());
972 pinger_args.pd_next_ping = imp->imp_next_ping;
974 mutex_unlock(&pinger_mutex);
978 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
979 timeout_cb_t cb, void *data,
980 cfs_list_t *obd_list)
985 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
986 enum timeout_event event)
991 int ptlrpc_pinger_add_import(struct obd_import *imp)
994 if (!cfs_list_empty(&imp->imp_pinger_chain))
997 CDEBUG(D_HA, "adding pingable import %s->%s\n",
998 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
999 ptlrpc_pinger_sending_on_import(imp);
1001 mutex_lock(&pinger_mutex);
1002 cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
1003 class_import_get(imp);
1004 mutex_unlock(&pinger_mutex);
1009 int ptlrpc_pinger_del_import(struct obd_import *imp)
1012 if (cfs_list_empty(&imp->imp_pinger_chain))
1015 mutex_lock(&pinger_mutex);
1016 cfs_list_del_init(&imp->imp_pinger_chain);
1017 CDEBUG(D_HA, "removing pingable import %s->%s\n",
1018 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1019 class_import_put(imp);
1020 mutex_unlock(&pinger_mutex);
1024 void ptlrpc_pinger_wake_up()
1026 #ifdef ENABLE_PINGER
1027 /* XXX force pinger to run, if needed */
1028 struct obd_import *imp;
1030 cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
1031 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
1032 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1033 #ifdef ENABLE_LIBLUSTRE_RECOVERY
1034 if (imp->imp_state == LUSTRE_IMP_DISCON &&
1035 !imp_is_deactive(imp))
1037 /*XXX only recover for the initial connection */
1038 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1039 imp->imp_state == LUSTRE_IMP_DISCON &&
1040 !imp_is_deactive(imp))
1042 ptlrpc_initiate_recovery(imp);
1043 else if (imp->imp_state != LUSTRE_IMP_FULL)
1044 CDEBUG(D_HA, "Refused to recover import %s->%s "
1045 "state %d, deactive %d\n",
1046 imp->imp_obd->obd_uuid.uuid,
1047 obd2cli_tgt(imp->imp_obd), imp->imp_state,
1048 imp_is_deactive(imp));
1053 #endif /* !__KERNEL__ */