1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Portal-RPC reconnection and replay operations, for use in recovery.
6 * Copyright (c) 2003 Cluster File Systems, Inc.
7 * Authors: Phil Schwan <phil@clusterfs.com>
8 * Mike Shaver <shaver@clusterfs.com>
10 * This file is part of the Lustre file system, http://www.lustre.org
11 * Lustre is a trademark of Cluster File Systems, Inc.
13 * You may have signed or agreed to another license before downloading
14 * this software. If so, you are bound by the terms and conditions
15 * of that agreement, and the following does not apply to you. See the
16 * LICENSE file included with this distribution for more information.
18 * If you did not agree to a different license, then this copy of Lustre
19 * is open source software; you can redistribute it and/or modify it
20 * under the terms of version 2 of the GNU General Public License as
21 * published by the Free Software Foundation.
23 * In either case, Lustre is distributed in the hope that it will be
24 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
25 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
26 * license text for more details.
30 #include <liblustre.h>
32 #define DEBUG_SUBSYSTEM S_RPC
35 #include <obd_support.h>
36 #include <obd_class.h>
37 #include "ptlrpc_internal.h"
39 struct semaphore pinger_sem;
40 static struct list_head pinger_imports = CFS_LIST_HEAD_INIT(pinger_imports);
42 int ptlrpc_ping(struct obd_import *imp)
44 struct ptlrpc_request *req;
48 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 1, NULL, NULL);
50 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
51 imp->imp_obd->obd_uuid.uuid,
52 obd2cli_tgt(imp->imp_obd));
53 req->rq_no_resend = req->rq_no_delay = 1;
54 ptlrpc_req_set_repsize(req, 1, NULL);
55 req->rq_timeout = PING_INTERVAL;
58 CERROR("OOM trying to ping %s->%s\n",
59 imp->imp_obd->obd_uuid.uuid,
60 obd2cli_tgt(imp->imp_obd));
67 void ptlrpc_update_next_ping(struct obd_import *imp)
70 imp->imp_next_ping = cfs_time_shift(
71 (imp->imp_state == LUSTRE_IMP_DISCON ?
72 RECONNECT_INTERVAL : PING_INTERVAL));
73 #endif /* ENABLE_PINGER */
76 void ptlrpc_ping_import_soon(struct obd_import *imp)
78 imp->imp_next_ping = cfs_time_current();
81 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
83 if (imp->imp_server_timeout)
84 return cfs_time_shift(obd_timeout / 2);
86 return cfs_time_shift(obd_timeout);
89 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
90 static cfs_time_t suspend_wakeup_time = 0;
93 static wait_queue_head_t suspend_timeouts_waitq;
96 cfs_time_t ptlrpc_suspend_wakeup_time(void)
98 return suspend_wakeup_time;
101 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
103 /*XXX: disabled for now, will be replaced by adaptive timeouts */
105 if (imp->imp_no_timeout)
107 imp->imp_no_timeout = 1;
108 atomic_inc(&suspend_timeouts);
109 CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n", atomic_read(&suspend_timeouts));
113 void ptlrpc_activate_timeouts(struct obd_import *imp)
115 /*XXX: disabled for now, will be replaced by adaptive timeouts */
117 if (!imp->imp_no_timeout)
119 imp->imp_no_timeout = 0;
120 LASSERT(atomic_read(&suspend_timeouts) > 0);
121 if (atomic_dec_and_test(&suspend_timeouts)) {
122 suspend_wakeup_time = cfs_time_current();
123 wake_up(&suspend_timeouts_waitq);
125 CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n", atomic_read(&suspend_timeouts));
129 int ptlrpc_check_suspend(void)
131 if (atomic_read(&suspend_timeouts))
136 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
138 struct l_wait_info lwi;
140 if (atomic_read(&suspend_timeouts)) {
141 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
142 atomic_read(&suspend_timeouts));
143 lwi = LWI_INTR(NULL, NULL);
144 l_wait_event(suspend_timeouts_waitq,
145 atomic_read(&suspend_timeouts) == 0, &lwi);
146 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
154 static void ptlrpc_pinger_process_import(struct obd_import *imp,
155 unsigned long this_ping)
159 spin_lock(&imp->imp_lock);
160 level = imp->imp_state;
161 force = imp->imp_force_verify;
163 imp->imp_force_verify = 0;
164 spin_unlock(&imp->imp_lock);
166 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
167 "level %s/%u force %u deactive %u pingable %u\n",
168 ptlrpc_import_state_name(level), level,
169 force, imp->imp_deactive, imp->imp_pingable);
171 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
172 this_ping) && force == 0)
175 if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
176 /* wait at least a timeout before trying recovery again */
177 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
178 ptlrpc_initiate_recovery(imp);
179 } else if (level != LUSTRE_IMP_FULL ||
180 imp->imp_obd->obd_no_recov ||
182 CDEBUG(D_HA, "not pinging %s (in recovery "
183 " or recovery disabled: %s)\n",
184 obd2cli_tgt(imp->imp_obd),
185 ptlrpc_import_state_name(level));
186 } else if (imp->imp_pingable || force) {
191 static int ptlrpc_pinger_main(void *arg)
193 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
194 struct ptlrpc_thread *thread = data->thread;
197 cfs_daemonize(data->name);
199 /* Record that the thread is running */
200 thread->t_flags = SVC_RUNNING;
201 cfs_waitq_signal(&thread->t_ctl_waitq);
203 /* And now, loop forever, pinging as needed. */
205 cfs_time_t this_ping = cfs_time_current();
206 struct l_wait_info lwi;
207 cfs_duration_t time_to_next_ping;
208 struct list_head *iter;
210 mutex_down(&pinger_sem);
211 list_for_each(iter, &pinger_imports) {
212 struct obd_import *imp =
213 list_entry(iter, struct obd_import,
216 ptlrpc_pinger_process_import(imp, this_ping);
217 /* obd_timeout might have changed */
218 if (imp->imp_pingable && imp->imp_next_ping &&
219 cfs_time_after(imp->imp_next_ping,
220 cfs_time_add(this_ping,
221 cfs_time_seconds(PING_INTERVAL))))
222 ptlrpc_update_next_ping(imp);
224 mutex_up(&pinger_sem);
225 /* update memory usage info */
226 obd_update_maxusage();
228 /* Wait until the next ping time, or until we're stopped. */
229 time_to_next_ping = cfs_time_sub(cfs_time_add(this_ping,
230 cfs_time_seconds(PING_INTERVAL)),
233 /* The ping sent by ptlrpc_send_rpc may get sent out
234 say .01 second after this.
235 ptlrpc_pinger_eending_on_import will then set the
236 next ping time to next_ping + .01 sec, which means
237 we will SKIP the next ping at next_ping, and the
238 ping will get sent 2 timeouts from now! Beware. */
239 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
241 cfs_time_add(this_ping, cfs_time_seconds(PING_INTERVAL)));
242 if (time_to_next_ping > 0) {
243 lwi = LWI_TIMEOUT(max_t(cfs_duration_t, time_to_next_ping, cfs_time_seconds(1)),
245 l_wait_event(thread->t_ctl_waitq,
246 thread->t_flags & (SVC_STOPPING|SVC_EVENT),
248 if (thread->t_flags & SVC_STOPPING) {
249 thread->t_flags &= ~SVC_STOPPING;
252 } else if (thread->t_flags & SVC_EVENT) {
253 /* woken after adding import to reset timer */
254 thread->t_flags &= ~SVC_EVENT;
259 thread->t_flags = SVC_STOPPED;
260 cfs_waitq_signal(&thread->t_ctl_waitq);
262 CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
266 static struct ptlrpc_thread *pinger_thread = NULL;
268 int ptlrpc_start_pinger(void)
270 struct l_wait_info lwi = { 0 };
271 struct ptlrpc_svc_data d;
273 #ifndef ENABLE_PINGER
278 if (pinger_thread != NULL)
281 OBD_ALLOC_PTR(pinger_thread);
282 if (pinger_thread == NULL)
284 cfs_waitq_init(&pinger_thread->t_ctl_waitq);
285 cfs_waitq_init(&suspend_timeouts_waitq);
288 d.thread = pinger_thread;
290 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
291 * just drop the VM and FILES in ptlrpc_daemonize() right away. */
292 rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
294 CERROR("cannot start thread: %d\n", rc);
295 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
296 pinger_thread = NULL;
299 l_wait_event(pinger_thread->t_ctl_waitq,
300 pinger_thread->t_flags & SVC_RUNNING, &lwi);
305 int ptlrpc_stop_pinger(void)
307 struct l_wait_info lwi = { 0 };
309 #ifndef ENABLE_PINGER
314 if (pinger_thread == NULL)
316 mutex_down(&pinger_sem);
317 pinger_thread->t_flags = SVC_STOPPING;
318 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
319 mutex_up(&pinger_sem);
321 l_wait_event(pinger_thread->t_ctl_waitq,
322 (pinger_thread->t_flags & SVC_STOPPED), &lwi);
324 OBD_FREE_PTR(pinger_thread);
325 pinger_thread = NULL;
329 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
331 ptlrpc_update_next_ping(imp);
334 int ptlrpc_pinger_add_import(struct obd_import *imp)
337 if (!list_empty(&imp->imp_pinger_chain))
340 mutex_down(&pinger_sem);
341 CDEBUG(D_HA, "adding pingable import %s->%s\n",
342 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
343 /* if we add to pinger we want recovery on this import */
344 imp->imp_obd->obd_no_recov = 0;
345 ptlrpc_update_next_ping(imp);
346 /* XXX sort, blah blah */
347 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
348 class_import_get(imp);
350 ptlrpc_pinger_wake_up();
351 mutex_up(&pinger_sem);
356 int ptlrpc_pinger_del_import(struct obd_import *imp)
359 if (list_empty(&imp->imp_pinger_chain))
362 mutex_down(&pinger_sem);
363 list_del_init(&imp->imp_pinger_chain);
364 CDEBUG(D_HA, "removing pingable import %s->%s\n",
365 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
366 /* if we remove from pinger we don't want recovery on this import */
367 imp->imp_obd->obd_no_recov = 1;
368 class_import_put(imp);
369 mutex_up(&pinger_sem);
373 void ptlrpc_pinger_wake_up()
376 pinger_thread->t_flags |= SVC_EVENT;
377 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
381 /* Ping evictor thread */
383 #define PET_TERMINATE 2
385 static int pet_refcount = 0;
386 static int pet_state;
387 static wait_queue_head_t pet_waitq;
388 static struct obd_export *pet_exp = NULL;
389 static spinlock_t pet_lock = SPIN_LOCK_UNLOCKED;
391 int ping_evictor_wake(struct obd_export *exp)
393 spin_lock(&pet_lock);
394 if (pet_exp || (pet_state != PET_READY)) {
395 /* eventually the new obd will call here again. */
396 spin_unlock(&pet_lock);
400 /* We have to make sure the obd isn't destroyed between now and when
401 * the ping evictor runs. We'll take a reference here, and drop it
402 * when we finish in the evictor. We don't really care about this
403 * export in particular; we just need one to keep the obd alive. */
404 pet_exp = class_export_get(exp);
405 spin_unlock(&pet_lock);
411 static int ping_evictor_main(void *arg)
413 struct obd_device *obd;
414 struct obd_export *exp;
415 struct l_wait_info lwi = { 0 };
419 ptlrpc_daemonize("ll_evictor");
421 CDEBUG(D_HA, "Starting Ping Evictor\n");
423 pet_state = PET_READY;
425 l_wait_event(pet_waitq, pet_exp ||
426 (pet_state == PET_TERMINATE), &lwi);
427 if (pet_state == PET_TERMINATE)
430 /* we only get here if pet_exp != NULL, and the end of this
431 * loop is the only place which sets it NULL again, so lock
432 * is not strictly necessary. */
433 spin_lock(&pet_lock);
434 obd = pet_exp->exp_obd;
435 spin_unlock(&pet_lock);
437 expire_time = CURRENT_SECONDS - (3 * obd_timeout / 2);
439 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
440 obd->obd_name, expire_time);
442 /* Exports can't be deleted out of the list while we hold
443 * the obd lock (class_unlink_export), which means we can't
444 * lose the last ref on the export. If they've already been
445 * removed from the list, we won't find them here. */
446 spin_lock(&obd->obd_dev_lock);
447 while (!list_empty(&obd->obd_exports_timed)) {
448 exp = list_entry(obd->obd_exports_timed.next,
449 struct obd_export,exp_obd_chain_timed);
450 if (expire_time > exp->exp_last_request_time) {
451 class_export_get(exp);
452 spin_unlock(&obd->obd_dev_lock);
453 LCONSOLE_WARN("%s: haven't heard from client %s"
454 " (at %s) in %ld seconds. I think"
455 " it's dead, and I am evicting"
456 " it. exp %p, cur %ld expire %ld"
459 obd_uuid2str(&exp->exp_client_uuid),
460 obd_export_nid2str(exp),
461 (long)(CURRENT_SECONDS -
462 exp->exp_last_request_time),
463 exp, (long)CURRENT_SECONDS,
465 (long)exp->exp_last_request_time);
466 CDEBUG(D_HA, "Last request was at %ld\n",
467 exp->exp_last_request_time);
468 class_fail_export(exp);
469 class_export_put(exp);
470 spin_lock(&obd->obd_dev_lock);
472 /* List is sorted, so everyone below is ok */
476 spin_unlock(&obd->obd_dev_lock);
478 class_export_put(pet_exp);
480 spin_lock(&pet_lock);
482 spin_unlock(&pet_lock);
484 CDEBUG(D_HA, "Exiting Ping Evictor\n");
489 void ping_evictor_start(void)
493 if (++pet_refcount > 1)
496 init_waitqueue_head(&pet_waitq);
498 rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
501 CERROR("Cannot start ping evictor thread: %d\n", rc);
504 EXPORT_SYMBOL(ping_evictor_start);
506 void ping_evictor_stop(void)
508 if (--pet_refcount > 0)
511 pet_state = PET_TERMINATE;
514 EXPORT_SYMBOL(ping_evictor_stop);
515 #else /* !__KERNEL__ */
518 * the current implementation of pinger in liblustre is not optimized
522 static struct pinger_data {
524 cfs_time_t pd_this_ping; /* jiffies */
525 cfs_time_t pd_next_ping; /* jiffies */
526 struct ptlrpc_request_set *pd_set;
529 static int pinger_check_rpcs(void *arg)
531 cfs_time_t curtime = cfs_time_current();
532 struct ptlrpc_request *req;
533 struct ptlrpc_request_set *set;
534 struct list_head *iter;
535 struct pinger_data *pd = &pinger_args;
538 /* prevent recursion */
539 if (pd->pd_recursion++) {
540 CDEBUG(D_HA, "pinger: recursion! quit\n");
546 /* have we reached ping point? */
547 if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
552 /* if we have rpc_set already, continue processing it */
554 LASSERT(pd->pd_this_ping);
559 pd->pd_this_ping = curtime;
560 pd->pd_set = ptlrpc_prep_set();
561 if (pd->pd_set == NULL)
565 /* add rpcs into set */
566 mutex_down(&pinger_sem);
567 list_for_each(iter, &pinger_imports) {
568 struct obd_import *imp =
569 list_entry(iter, struct obd_import, imp_pinger_chain);
570 int generation, level;
572 if (cfs_time_aftereq(pd->pd_this_ping,
573 imp->imp_next_ping - 5 * CFS_TICK)) {
575 spin_lock(&imp->imp_lock);
576 generation = imp->imp_generation;
577 level = imp->imp_state;
578 spin_unlock(&imp->imp_lock);
580 if (level != LUSTRE_IMP_FULL) {
582 "not pinging %s (in recovery)\n",
583 obd2cli_tgt(imp->imp_obd));
587 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING,
590 CERROR("out of memory\n");
593 req->rq_no_resend = 1;
594 ptlrpc_req_set_repsize(req, 1, NULL);
595 req->rq_send_state = LUSTRE_IMP_FULL;
596 req->rq_phase = RQ_PHASE_RPC;
597 req->rq_import_generation = generation;
598 ptlrpc_set_add_req(set, req);
600 CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
601 " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
602 imp->imp_next_ping, pd->pd_this_ping);
605 pd->pd_this_ping = curtime;
606 mutex_up(&pinger_sem);
608 /* Might be empty, that's OK. */
609 if (set->set_remaining == 0)
610 CDEBUG(D_RPCTRACE, "nothing to ping\n");
612 list_for_each(iter, &set->set_requests) {
613 struct ptlrpc_request *req =
614 list_entry(iter, struct ptlrpc_request,
616 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
617 req->rq_import->imp_obd->obd_uuid.uuid,
618 obd2cli_tgt(req->rq_import->imp_obd));
619 (void)ptl_send_rpc(req, 0);
623 rc = ptlrpc_check_set(set);
625 /* not finished, and we are not expired, simply return */
626 if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
627 cfs_time_seconds(PING_INTERVAL)))) {
628 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
633 /* Expire all the requests that didn't come back. */
634 mutex_down(&pinger_sem);
635 list_for_each(iter, &set->set_requests) {
636 req = list_entry(iter, struct ptlrpc_request,
639 if (req->rq_phase == RQ_PHASE_COMPLETE)
642 req->rq_phase = RQ_PHASE_COMPLETE;
643 atomic_dec(&req->rq_import->imp_inflight);
644 set->set_remaining--;
645 /* If it was disconnected, don't sweat it. */
646 if (list_empty(&req->rq_import->imp_pinger_chain)) {
647 ptlrpc_unregister_reply(req);
651 CDEBUG(D_RPCTRACE, "pinger initiate expire_one_request\n");
652 ptlrpc_expire_one_request(req);
654 mutex_up(&pinger_sem);
656 ptlrpc_set_destroy(set);
660 pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
661 cfs_time_seconds(PING_INTERVAL));
662 pd->pd_this_ping = 0; /* XXX for debug */
664 CDEBUG(D_INFO, "finished a round ping\n");
669 static void *pinger_callback = NULL;
670 #endif /* ENABLE_PINGER */
672 int ptlrpc_start_pinger(void)
675 memset(&pinger_args, 0, sizeof(pinger_args));
676 pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
683 int ptlrpc_stop_pinger(void)
687 liblustre_deregister_wait_callback(pinger_callback);
692 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
695 mutex_down(&pinger_sem);
696 ptlrpc_update_next_ping(imp);
697 if (pinger_args.pd_set == NULL &&
698 time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
699 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
700 imp->imp_next_ping, cfs_time_current());
701 pinger_args.pd_next_ping = imp->imp_next_ping;
703 mutex_up(&pinger_sem);
707 int ptlrpc_pinger_add_import(struct obd_import *imp)
710 if (!list_empty(&imp->imp_pinger_chain))
713 CDEBUG(D_HA, "adding pingable import %s->%s\n",
714 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
715 ptlrpc_pinger_sending_on_import(imp);
717 mutex_down(&pinger_sem);
718 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
719 class_import_get(imp);
720 mutex_up(&pinger_sem);
725 int ptlrpc_pinger_del_import(struct obd_import *imp)
728 if (list_empty(&imp->imp_pinger_chain))
731 mutex_down(&pinger_sem);
732 list_del_init(&imp->imp_pinger_chain);
733 CDEBUG(D_HA, "removing pingable import %s->%s\n",
734 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
735 class_import_put(imp);
736 mutex_up(&pinger_sem);
740 void ptlrpc_pinger_wake_up()
743 /* XXX force pinger to run, if needed */
744 struct obd_import *imp;
746 list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
747 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
748 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
749 #ifdef ENABLE_LIBLUSTRE_RECOVERY
750 if (imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
752 /*XXX only recover for the initial connection */
753 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
754 imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
756 ptlrpc_initiate_recovery(imp);
757 else if (imp->imp_state != LUSTRE_IMP_FULL)
758 CDEBUG(D_HA, "Refused to recover import %s->%s "
759 "state %d, deactive %d\n",
760 imp->imp_obd->obd_uuid.uuid,
761 obd2cli_tgt(imp->imp_obd), imp->imp_state,
767 #endif /* !__KERNEL__ */