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 static void ptlrpc_update_next_ping(struct obd_import *imp)
69 imp->imp_next_ping = cfs_time_shift(
70 (imp->imp_state == LUSTRE_IMP_DISCON ?
71 RECONNECT_INTERVAL : PING_INTERVAL));
74 void ptlrpc_ping_import_soon(struct obd_import *imp)
76 imp->imp_next_ping = cfs_time_current();
79 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
81 if (imp->imp_server_timeout)
82 return cfs_time_shift(obd_timeout / 2);
84 return cfs_time_shift(obd_timeout);
87 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
88 static cfs_time_t suspend_wakeup_time = 0;
89 static wait_queue_head_t suspend_timeouts_waitq;
91 cfs_time_t ptlrpc_suspend_wakeup_time(void)
93 return suspend_wakeup_time;
96 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
98 /*XXX: disabled for now, will be replaced by adaptive timeouts */
100 if (imp->imp_no_timeout)
102 imp->imp_no_timeout = 1;
103 atomic_inc(&suspend_timeouts);
104 CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n", atomic_read(&suspend_timeouts));
108 void ptlrpc_activate_timeouts(struct obd_import *imp)
110 /*XXX: disabled for now, will be replaced by adaptive timeouts */
112 if (!imp->imp_no_timeout)
114 imp->imp_no_timeout = 0;
115 LASSERT(atomic_read(&suspend_timeouts) > 0);
116 if (atomic_dec_and_test(&suspend_timeouts)) {
117 suspend_wakeup_time = cfs_time_current();
118 wake_up(&suspend_timeouts_waitq);
120 CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n", atomic_read(&suspend_timeouts));
124 int ptlrpc_check_suspend(void)
126 if (atomic_read(&suspend_timeouts))
131 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
133 struct l_wait_info lwi;
135 if (atomic_read(&suspend_timeouts)) {
136 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
137 atomic_read(&suspend_timeouts));
138 lwi = LWI_INTR(NULL, NULL);
139 l_wait_event(suspend_timeouts_waitq,
140 atomic_read(&suspend_timeouts) == 0, &lwi);
141 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
149 static void ptlrpc_pinger_process_import(struct obd_import *imp,
150 unsigned long this_ping)
154 spin_lock(&imp->imp_lock);
155 level = imp->imp_state;
156 force = imp->imp_force_verify;
158 imp->imp_force_verify = 0;
159 spin_unlock(&imp->imp_lock);
161 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
162 "level %s/%u force %u deactive %u pingable %u\n",
163 ptlrpc_import_state_name(level), level,
164 force, imp->imp_deactive, imp->imp_pingable);
166 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
167 this_ping) && force == 0)
170 if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
171 /* wait at least a timeout before trying recovery again */
172 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
173 ptlrpc_initiate_recovery(imp);
174 } else if (level != LUSTRE_IMP_FULL ||
175 imp->imp_obd->obd_no_recov ||
177 CDEBUG(D_HA, "not pinging %s (in recovery "
178 " or recovery disabled: %s)\n",
179 obd2cli_tgt(imp->imp_obd),
180 ptlrpc_import_state_name(level));
181 } else if (imp->imp_pingable || force) {
186 static int ptlrpc_pinger_main(void *arg)
188 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
189 struct ptlrpc_thread *thread = data->thread;
192 cfs_daemonize(data->name);
194 /* Record that the thread is running */
195 thread->t_flags = SVC_RUNNING;
196 cfs_waitq_signal(&thread->t_ctl_waitq);
198 /* And now, loop forever, pinging as needed. */
200 cfs_time_t this_ping = cfs_time_current();
201 struct l_wait_info lwi;
202 cfs_duration_t time_to_next_ping;
203 struct list_head *iter;
205 mutex_down(&pinger_sem);
206 list_for_each(iter, &pinger_imports) {
207 struct obd_import *imp =
208 list_entry(iter, struct obd_import,
211 ptlrpc_pinger_process_import(imp, this_ping);
212 /* obd_timeout might have changed */
213 if (imp->imp_pingable && imp->imp_next_ping &&
214 cfs_time_after(imp->imp_next_ping,
215 cfs_time_add(this_ping,
216 cfs_time_seconds(PING_INTERVAL))))
217 ptlrpc_update_next_ping(imp);
219 mutex_up(&pinger_sem);
221 /* Wait until the next ping time, or until we're stopped. */
222 time_to_next_ping = cfs_time_sub(cfs_time_add(this_ping,
223 cfs_time_seconds(PING_INTERVAL)),
226 /* The ping sent by ptlrpc_send_rpc may get sent out
227 say .01 second after this.
228 ptlrpc_pinger_eending_on_import will then set the
229 next ping time to next_ping + .01 sec, which means
230 we will SKIP the next ping at next_ping, and the
231 ping will get sent 2 timeouts from now! Beware. */
232 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
234 cfs_time_add(this_ping, cfs_time_seconds(PING_INTERVAL)));
235 if (time_to_next_ping > 0) {
236 lwi = LWI_TIMEOUT(max_t(cfs_duration_t, time_to_next_ping, cfs_time_seconds(1)),
238 l_wait_event(thread->t_ctl_waitq,
239 thread->t_flags & (SVC_STOPPING|SVC_EVENT),
241 if (thread->t_flags & SVC_STOPPING) {
242 thread->t_flags &= ~SVC_STOPPING;
245 } else if (thread->t_flags & SVC_EVENT) {
246 /* woken after adding import to reset timer */
247 thread->t_flags &= ~SVC_EVENT;
252 thread->t_flags = SVC_STOPPED;
253 cfs_waitq_signal(&thread->t_ctl_waitq);
255 CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
259 static struct ptlrpc_thread *pinger_thread = NULL;
261 int ptlrpc_start_pinger(void)
263 struct l_wait_info lwi = { 0 };
264 struct ptlrpc_svc_data d;
266 #ifndef ENABLE_PINGER
271 if (pinger_thread != NULL)
274 OBD_ALLOC_PTR(pinger_thread);
275 if (pinger_thread == NULL)
277 cfs_waitq_init(&pinger_thread->t_ctl_waitq);
278 cfs_waitq_init(&suspend_timeouts_waitq);
281 d.thread = pinger_thread;
283 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
284 * just drop the VM and FILES in ptlrpc_daemonize() right away. */
285 rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
287 CERROR("cannot start thread: %d\n", rc);
288 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
289 pinger_thread = NULL;
292 l_wait_event(pinger_thread->t_ctl_waitq,
293 pinger_thread->t_flags & SVC_RUNNING, &lwi);
298 int ptlrpc_stop_pinger(void)
300 struct l_wait_info lwi = { 0 };
302 #ifndef ENABLE_PINGER
307 if (pinger_thread == NULL)
309 mutex_down(&pinger_sem);
310 pinger_thread->t_flags = SVC_STOPPING;
311 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
312 mutex_up(&pinger_sem);
314 l_wait_event(pinger_thread->t_ctl_waitq,
315 (pinger_thread->t_flags & SVC_STOPPED), &lwi);
317 OBD_FREE_PTR(pinger_thread);
318 pinger_thread = NULL;
322 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
324 ptlrpc_update_next_ping(imp);
327 int ptlrpc_pinger_add_import(struct obd_import *imp)
330 if (!list_empty(&imp->imp_pinger_chain))
333 mutex_down(&pinger_sem);
334 CDEBUG(D_HA, "adding pingable import %s->%s\n",
335 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
336 /* if we add to pinger we want recovery on this import */
337 imp->imp_obd->obd_no_recov = 0;
338 ptlrpc_update_next_ping(imp);
339 /* XXX sort, blah blah */
340 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
341 class_import_get(imp);
343 ptlrpc_pinger_wake_up();
344 mutex_up(&pinger_sem);
349 int ptlrpc_pinger_del_import(struct obd_import *imp)
352 if (list_empty(&imp->imp_pinger_chain))
355 mutex_down(&pinger_sem);
356 list_del_init(&imp->imp_pinger_chain);
357 CDEBUG(D_HA, "removing pingable import %s->%s\n",
358 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
359 /* if we remove from pinger we don't want recovery on this import */
360 imp->imp_obd->obd_no_recov = 1;
361 class_import_put(imp);
362 mutex_up(&pinger_sem);
366 void ptlrpc_pinger_wake_up()
369 pinger_thread->t_flags |= SVC_EVENT;
370 cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
374 /* Ping evictor thread */
376 #define PET_TERMINATE 2
378 static int pet_refcount = 0;
379 static int pet_state;
380 static wait_queue_head_t pet_waitq;
381 static struct obd_export *pet_exp = NULL;
382 static spinlock_t pet_lock = SPIN_LOCK_UNLOCKED;
384 int ping_evictor_wake(struct obd_export *exp)
386 spin_lock(&pet_lock);
388 /* eventually the new obd will call here again. */
389 spin_unlock(&pet_lock);
393 /* We have to make sure the obd isn't destroyed between now and when
394 * the ping evictor runs. We'll take a reference here, and drop it
395 * when we finish in the evictor. We don't really care about this
396 * export in particular; we just need one to keep the obd alive. */
397 pet_exp = class_export_get(exp);
398 spin_unlock(&pet_lock);
404 static int ping_evictor_main(void *arg)
406 struct obd_device *obd;
407 struct obd_export *exp;
408 struct l_wait_info lwi = { 0 };
412 ptlrpc_daemonize("ll_evictor");
414 CDEBUG(D_HA, "Starting Ping Evictor\n");
416 pet_state = PET_READY;
418 l_wait_event(pet_waitq, pet_exp ||
419 (pet_state == PET_TERMINATE), &lwi);
420 if (pet_state == PET_TERMINATE)
423 /* we only get here if pet_exp != NULL, and the end of this
424 * loop is the only place which sets it NULL again, so lock
425 * is not strictly necessary. */
426 spin_lock(&pet_lock);
427 obd = pet_exp->exp_obd;
428 spin_unlock(&pet_lock);
430 expire_time = CURRENT_SECONDS - (3 * obd_timeout / 2);
432 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
433 obd->obd_name, expire_time);
435 /* Exports can't be deleted out of the list while we hold
436 * the obd lock (class_unlink_export), which means we can't
437 * lose the last ref on the export. If they've already been
438 * removed from the list, we won't find them here. */
439 spin_lock(&obd->obd_dev_lock);
440 while (!list_empty(&obd->obd_exports_timed)) {
441 exp = list_entry(obd->obd_exports_timed.next,
442 struct obd_export,exp_obd_chain_timed);
443 if (expire_time > exp->exp_last_request_time) {
444 class_export_get(exp);
445 spin_unlock(&obd->obd_dev_lock);
446 LCONSOLE_WARN("%s: haven't heard from client %s"
447 " (at %s) in %ld seconds. I think"
448 " it's dead, and I am evicting"
449 " it. exp %p, cur %ld expire %ld"
452 obd_uuid2str(&exp->exp_client_uuid),
453 obd_export_nid2str(exp),
454 (long)(CURRENT_SECONDS -
455 exp->exp_last_request_time),
456 exp, (long)CURRENT_SECONDS,
458 (long)exp->exp_last_request_time);
459 CDEBUG(D_HA, "Last request was at %ld\n",
460 exp->exp_last_request_time);
461 class_fail_export(exp);
462 class_export_put(exp);
463 spin_lock(&obd->obd_dev_lock);
465 /* List is sorted, so everyone below is ok */
469 spin_unlock(&obd->obd_dev_lock);
471 class_export_put(pet_exp);
473 spin_lock(&pet_lock);
475 spin_unlock(&pet_lock);
477 CDEBUG(D_HA, "Exiting Ping Evictor\n");
482 void ping_evictor_start(void)
486 if (++pet_refcount > 1)
489 init_waitqueue_head(&pet_waitq);
491 rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
494 CERROR("Cannot start ping evictor thread: %d\n", rc);
497 EXPORT_SYMBOL(ping_evictor_start);
499 void ping_evictor_stop(void)
501 if (--pet_refcount > 0)
504 pet_state = PET_TERMINATE;
507 EXPORT_SYMBOL(ping_evictor_stop);
508 #else /* !__KERNEL__ */
511 * the current implementation of pinger in liblustre is not optimized
515 static struct pinger_data {
517 cfs_time_t pd_this_ping; /* jiffies */
518 cfs_time_t pd_next_ping; /* jiffies */
519 struct ptlrpc_request_set *pd_set;
522 static int pinger_check_rpcs(void *arg)
524 cfs_time_t curtime = cfs_time_current();
525 struct ptlrpc_request *req;
526 struct ptlrpc_request_set *set;
527 struct list_head *iter;
528 struct pinger_data *pd = &pinger_args;
531 /* prevent recursion */
532 if (pd->pd_recursion++) {
533 CDEBUG(D_HA, "pinger: recursion! quit\n");
539 /* have we reached ping point? */
540 if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
545 /* if we have rpc_set already, continue processing it */
547 LASSERT(pd->pd_this_ping);
552 pd->pd_this_ping = curtime;
553 pd->pd_set = ptlrpc_prep_set();
554 if (pd->pd_set == NULL)
558 /* add rpcs into set */
559 mutex_down(&pinger_sem);
560 list_for_each(iter, &pinger_imports) {
561 struct obd_import *imp =
562 list_entry(iter, struct obd_import, imp_pinger_chain);
563 int generation, level;
565 if (cfs_time_aftereq(pd->pd_this_ping,
566 imp->imp_next_ping - 5 * CFS_TICK)) {
568 spin_lock(&imp->imp_lock);
569 generation = imp->imp_generation;
570 level = imp->imp_state;
571 spin_unlock(&imp->imp_lock);
573 if (level != LUSTRE_IMP_FULL) {
575 "not pinging %s (in recovery)\n",
576 obd2cli_tgt(imp->imp_obd));
580 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING,
583 CERROR("out of memory\n");
586 req->rq_no_resend = 1;
587 ptlrpc_req_set_repsize(req, 1, NULL);
588 req->rq_send_state = LUSTRE_IMP_FULL;
589 req->rq_phase = RQ_PHASE_RPC;
590 req->rq_import_generation = generation;
591 ptlrpc_set_add_req(set, req);
593 CDEBUG(D_HA, "don't need to ping %s ("CFS_TIME_T" > "
594 CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
595 imp->imp_next_ping, pd->pd_this_ping);
598 pd->pd_this_ping = curtime;
599 mutex_up(&pinger_sem);
601 /* Might be empty, that's OK. */
602 if (set->set_remaining == 0)
603 CDEBUG(D_HA, "nothing to ping\n");
605 list_for_each(iter, &set->set_requests) {
606 struct ptlrpc_request *req =
607 list_entry(iter, struct ptlrpc_request,
609 DEBUG_REQ(D_HA, req, "pinging %s->%s",
610 req->rq_import->imp_obd->obd_uuid.uuid,
611 obd2cli_tgt(req->rq_import->imp_obd));
612 (void)ptl_send_rpc(req, 0);
616 rc = ptlrpc_check_set(set);
618 /* not finished, and we are not expired, simply return */
619 if (!rc && cfs_time_before(curtime,
620 cfs_time_add(pd->pd_this_ping,
621 cfs_time_seconds(PING_INTERVAL)))) {
622 CDEBUG(D_HA, "not finished, but also not expired\n");
627 /* Expire all the requests that didn't come back. */
628 mutex_down(&pinger_sem);
629 list_for_each(iter, &set->set_requests) {
630 req = list_entry(iter, struct ptlrpc_request,
633 if (req->rq_phase == RQ_PHASE_COMPLETE)
636 req->rq_phase = RQ_PHASE_COMPLETE;
637 atomic_dec(&req->rq_import->imp_inflight);
638 set->set_remaining--;
639 /* If it was disconnected, don't sweat it. */
640 if (list_empty(&req->rq_import->imp_pinger_chain)) {
641 ptlrpc_unregister_reply(req);
645 CDEBUG(D_HA, "pinger initiate expire_one_request\n");
646 ptlrpc_expire_one_request(req);
648 mutex_up(&pinger_sem);
650 ptlrpc_set_destroy(set);
654 pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
655 cfs_time_seconds(PING_INTERVAL));
656 pd->pd_this_ping = 0; /* XXX for debug */
658 CDEBUG(D_HA, "finished a round ping\n");
663 static void *pinger_callback = NULL;
664 #endif /* ENABLE_PINGER */
666 int ptlrpc_start_pinger(void)
669 memset(&pinger_args, 0, sizeof(pinger_args));
670 pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
677 int ptlrpc_stop_pinger(void)
681 liblustre_deregister_wait_callback(pinger_callback);
686 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
689 mutex_down(&pinger_sem);
690 ptlrpc_update_next_ping(imp);
691 if (pinger_args.pd_set == NULL &&
692 time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
693 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
694 imp->imp_next_ping, cfs_time_current());
695 pinger_args.pd_next_ping = imp->imp_next_ping;
697 mutex_up(&pinger_sem);
701 int ptlrpc_pinger_add_import(struct obd_import *imp)
704 if (!list_empty(&imp->imp_pinger_chain))
707 CDEBUG(D_HA, "adding pingable import %s->%s\n",
708 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
709 ptlrpc_pinger_sending_on_import(imp);
711 mutex_down(&pinger_sem);
712 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
713 class_import_get(imp);
714 mutex_up(&pinger_sem);
719 int ptlrpc_pinger_del_import(struct obd_import *imp)
722 if (list_empty(&imp->imp_pinger_chain))
725 mutex_down(&pinger_sem);
726 list_del_init(&imp->imp_pinger_chain);
727 CDEBUG(D_HA, "removing pingable import %s->%s\n",
728 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
729 class_import_put(imp);
730 mutex_up(&pinger_sem);
734 void ptlrpc_pinger_wake_up()
737 /* XXX force pinger to run, if needed */
738 struct obd_import *imp;
740 list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
741 CDEBUG(D_HA, "Checking that we need to do anything about import"
742 " %s->%s\n", imp->imp_obd->obd_uuid.uuid,
743 obd2cli_tgt(imp->imp_obd));
744 #ifdef ENABLE_LIBLUSTRE_RECOVERY
745 if (imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
747 /*XXX only recover for the initial connection */
748 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
749 imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
751 ptlrpc_initiate_recovery(imp);
752 else if (imp->imp_state != LUSTRE_IMP_FULL)
753 CDEBUG(D_HA, "Refused to recover import %s->%s "
754 "state %d, deactive %d\n",
755 imp->imp_obd->obd_uuid.uuid,
756 obd2cli_tgt(imp->imp_obd), imp->imp_state,
762 #endif /* !__KERNEL__ */