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.
41 #define DEBUG_SUBSYSTEM S_RPC
43 #include <obd_support.h>
44 #include <obd_class.h>
45 #include "ptlrpc_internal.h"
47 static int suppress_pings;
48 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
50 struct mutex pinger_mutex;
51 static LIST_HEAD(pinger_imports);
52 static struct list_head timeout_list = LIST_HEAD_INIT(timeout_list);
54 int ptlrpc_pinger_suppress_pings(void)
56 return suppress_pings;
58 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
60 struct ptlrpc_request *
61 ptlrpc_prep_ping(struct obd_import *imp)
63 struct ptlrpc_request *req;
65 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
66 LUSTRE_OBD_VERSION, OBD_PING);
68 ptlrpc_request_set_replen(req);
69 req->rq_no_resend = req->rq_no_delay = 1;
74 int ptlrpc_obd_ping(struct obd_device *obd)
77 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;
97 req = ptlrpc_prep_ping(imp);
99 CERROR("OOM trying to ping %s->%s\n",
100 imp->imp_obd->obd_uuid.uuid,
101 obd2cli_tgt(imp->imp_obd));
105 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
106 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
107 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
112 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);
124 void ptlrpc_ping_import_soon(struct obd_import *imp)
126 imp->imp_next_ping = cfs_time_current();
129 static inline int imp_is_deactive(struct obd_import *imp)
131 return (imp->imp_deactive ||
132 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
135 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
137 if (imp->imp_server_timeout)
138 return cfs_time_shift(obd_timeout / 2);
140 return cfs_time_shift(obd_timeout);
143 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
144 static cfs_time_t suspend_wakeup_time = 0;
146 cfs_duration_t pinger_check_timeout(cfs_time_t time)
148 struct timeout_item *item;
149 cfs_time_t timeout = PING_INTERVAL;
151 /* The timeout list is a increase order sorted list */
152 mutex_lock(&pinger_mutex);
153 list_for_each_entry(item, &timeout_list, ti_chain) {
154 int ti_timeout = item->ti_timeout;
155 if (timeout > ti_timeout)
156 timeout = ti_timeout;
159 mutex_unlock(&pinger_mutex);
161 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
165 static wait_queue_head_t suspend_timeouts_waitq;
167 cfs_time_t ptlrpc_suspend_wakeup_time(void)
169 return suspend_wakeup_time;
172 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
174 /*XXX: disabled for now, will be replaced by adaptive timeouts */
176 if (imp->imp_no_timeout)
178 imp->imp_no_timeout = 1;
179 atomic_inc(&suspend_timeouts);
180 CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
181 atomic_read(&suspend_timeouts));
185 void ptlrpc_activate_timeouts(struct obd_import *imp)
187 /*XXX: disabled for now, will be replaced by adaptive timeouts */
189 if (!imp->imp_no_timeout)
191 imp->imp_no_timeout = 0;
192 LASSERT(atomic_read(&suspend_timeouts) > 0);
193 if (atomic_dec_and_test(&suspend_timeouts)) {
194 suspend_wakeup_time = cfs_time_current();
195 wake_up(&suspend_timeouts_waitq);
197 CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
198 atomic_read(&suspend_timeouts));
202 int ptlrpc_check_suspend(void)
204 if (atomic_read(&suspend_timeouts))
209 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
211 struct l_wait_info lwi;
213 if (atomic_read(&suspend_timeouts)) {
214 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
215 atomic_read(&suspend_timeouts));
216 lwi = LWI_INTR(NULL, NULL);
217 l_wait_event(suspend_timeouts_waitq,
218 atomic_read(&suspend_timeouts) == 0, &lwi);
219 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
228 void ptlrpc_pinger_ir_up(void)
230 CDEBUG(D_HA, "IR up\n");
233 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
235 void ptlrpc_pinger_ir_down(void)
237 CDEBUG(D_HA, "IR down\n");
240 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
242 static void ptlrpc_pinger_process_import(struct obd_import *imp,
243 unsigned long this_ping)
250 spin_lock(&imp->imp_lock);
252 level = imp->imp_state;
253 force = imp->imp_force_verify;
254 force_next = imp->imp_force_next_verify;
256 * This will be used below only if the import is "FULL".
258 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
260 imp->imp_force_verify = 0;
262 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
264 spin_unlock(&imp->imp_lock);
268 imp->imp_force_next_verify = 0;
270 spin_unlock(&imp->imp_lock);
272 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
273 "force %u force_next %u deactive %u pingable %u suppress %u\n",
274 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
275 ptlrpc_import_state_name(level), level, force, force_next,
276 imp->imp_deactive, imp->imp_pingable, suppress);
278 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
279 /* wait for a while before trying recovery again */
280 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
281 if (!imp->imp_no_pinger_recover)
282 ptlrpc_initiate_recovery(imp);
283 } else if (level != LUSTRE_IMP_FULL ||
284 imp->imp_obd->obd_no_recov ||
285 imp_is_deactive(imp)) {
286 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
287 "or recovery disabled: %s)\n",
288 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
289 ptlrpc_import_state_name(level));
290 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
295 static int ptlrpc_pinger_main(void *arg)
297 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
299 /* Record that the thread is running */
300 thread_set_flags(thread, SVC_RUNNING);
301 wake_up(&thread->t_ctl_waitq);
303 /* And now, loop forever, pinging as needed. */
305 cfs_time_t this_ping = cfs_time_current();
306 struct l_wait_info lwi;
307 cfs_duration_t time_to_next_wake;
308 struct timeout_item *item;
309 struct list_head *iter;
311 mutex_lock(&pinger_mutex);
312 list_for_each_entry(item, &timeout_list, ti_chain) {
313 item->ti_cb(item, item->ti_cb_data);
315 list_for_each(iter, &pinger_imports) {
316 struct obd_import *imp =
317 list_entry(iter, struct obd_import,
320 ptlrpc_pinger_process_import(imp, this_ping);
321 /* obd_timeout might have changed */
322 if (imp->imp_pingable && imp->imp_next_ping &&
323 cfs_time_after(imp->imp_next_ping,
324 cfs_time_add(this_ping,
325 cfs_time_seconds(PING_INTERVAL))))
326 ptlrpc_update_next_ping(imp, 0);
328 mutex_unlock(&pinger_mutex);
329 /* update memory usage info */
330 obd_update_maxusage();
332 /* Wait until the next ping time, or until we're stopped. */
333 time_to_next_wake = pinger_check_timeout(this_ping);
334 /* The ping sent by ptlrpc_send_rpc may get sent out
335 say .01 second after this.
336 ptlrpc_pinger_sending_on_import will then set the
337 next ping time to next_ping + .01 sec, which means
338 we will SKIP the next ping at next_ping, and the
339 ping will get sent 2 timeouts from now! Beware. */
340 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
341 CFS_TIME_T")\n", time_to_next_wake,
342 cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
343 if (time_to_next_wake > 0) {
344 lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
346 cfs_time_seconds(1)),
348 l_wait_event(thread->t_ctl_waitq,
349 thread_is_stopping(thread) ||
350 thread_is_event(thread),
352 if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
355 /* woken after adding import to reset timer */
356 thread_test_and_clear_flags(thread, SVC_EVENT);
361 thread_set_flags(thread, SVC_STOPPED);
362 wake_up(&thread->t_ctl_waitq);
364 CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
368 static struct ptlrpc_thread pinger_thread;
370 int ptlrpc_start_pinger(void)
372 struct l_wait_info lwi = { 0 };
375 if (!thread_is_init(&pinger_thread) &&
376 !thread_is_stopped(&pinger_thread))
379 init_waitqueue_head(&pinger_thread.t_ctl_waitq);
380 init_waitqueue_head(&suspend_timeouts_waitq);
382 strcpy(pinger_thread.t_name, "ll_ping");
384 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
385 * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
386 rc = PTR_ERR(kthread_run(ptlrpc_pinger_main, &pinger_thread,
387 "%s", pinger_thread.t_name));
388 if (IS_ERR_VALUE(rc)) {
389 CERROR("cannot start thread: %d\n", rc);
392 l_wait_event(pinger_thread.t_ctl_waitq,
393 thread_is_running(&pinger_thread), &lwi);
396 CWARN("Pings will be suppressed at the request of the "
397 "administrator. The configuration shall meet the "
398 "additional requirements described in the manual. "
399 "(Search for the \"suppress_pings\" kernel module "
405 int ptlrpc_pinger_remove_timeouts(void);
407 int ptlrpc_stop_pinger(void)
409 struct l_wait_info lwi = { 0 };
412 if (!thread_is_init(&pinger_thread) &&
413 !thread_is_stopped(&pinger_thread))
416 ptlrpc_pinger_remove_timeouts();
417 thread_set_flags(&pinger_thread, SVC_STOPPING);
418 wake_up(&pinger_thread.t_ctl_waitq);
420 l_wait_event(pinger_thread.t_ctl_waitq,
421 thread_is_stopped(&pinger_thread), &lwi);
426 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
428 ptlrpc_update_next_ping(imp, 0);
430 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
432 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
434 ptlrpc_update_next_ping(imp, 1);
435 LASSERT(spin_is_locked(&imp->imp_lock));
437 * Avoid reading stale imp_connect_data. When not sure if pings are
438 * expected or not on next connection, we assume they are not and force
439 * one anyway to guarantee the chance of updating
440 * imp_peer_committed_transno.
442 if (imp->imp_state != LUSTRE_IMP_FULL ||
443 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
444 imp->imp_force_next_verify = 1;
447 int ptlrpc_pinger_add_import(struct obd_import *imp)
449 if (!list_empty(&imp->imp_pinger_chain))
452 mutex_lock(&pinger_mutex);
453 CDEBUG(D_HA, "adding pingable import %s->%s\n",
454 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
455 /* if we add to pinger we want recovery on this import */
456 imp->imp_obd->obd_no_recov = 0;
457 ptlrpc_update_next_ping(imp, 0);
458 /* XXX sort, blah blah */
459 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
460 class_import_get(imp);
462 ptlrpc_pinger_wake_up();
463 mutex_unlock(&pinger_mutex);
467 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
469 int ptlrpc_pinger_del_import(struct obd_import *imp)
471 if (list_empty(&imp->imp_pinger_chain))
474 mutex_lock(&pinger_mutex);
475 list_del_init(&imp->imp_pinger_chain);
476 CDEBUG(D_HA, "removing pingable import %s->%s\n",
477 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
478 /* if we remove from pinger we don't want recovery on this import */
479 imp->imp_obd->obd_no_recov = 1;
480 class_import_put(imp);
481 mutex_unlock(&pinger_mutex);
484 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
487 * Register a timeout callback to the pinger list, and the callback will
488 * be called when timeout happens.
490 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
491 timeout_cb_t cb, void *data)
493 struct timeout_item *ti;
499 INIT_LIST_HEAD(&ti->ti_obd_list);
500 INIT_LIST_HEAD(&ti->ti_chain);
501 ti->ti_timeout = time;
502 ti->ti_event = event;
504 ti->ti_cb_data = data;
510 * Register timeout event on the the pinger thread.
511 * Note: the timeout list is an sorted list with increased timeout value.
513 static struct timeout_item*
514 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
515 timeout_cb_t cb, void *data)
517 struct timeout_item *item, *tmp;
519 LASSERT(mutex_is_locked(&pinger_mutex));
521 list_for_each_entry(item, &timeout_list, ti_chain)
522 if (item->ti_event == event)
525 item = ptlrpc_new_timeout(time, event, cb, data);
527 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
528 if (tmp->ti_timeout < time) {
529 list_add(&item->ti_chain, &tmp->ti_chain);
533 list_add(&item->ti_chain, &timeout_list);
539 /* Add a client_obd to the timeout event list, when timeout(@time)
540 * happens, the callback(@cb) will be called.
542 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
543 timeout_cb_t cb, void *data,
544 struct list_head *obd_list)
546 struct timeout_item *ti;
548 mutex_lock(&pinger_mutex);
549 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
551 mutex_unlock(&pinger_mutex);
554 list_add(obd_list, &ti->ti_obd_list);
555 mutex_unlock(&pinger_mutex);
558 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
560 int ptlrpc_del_timeout_client(struct list_head *obd_list,
561 enum timeout_event event)
563 struct timeout_item *ti = NULL, *item;
565 if (list_empty(obd_list))
567 mutex_lock(&pinger_mutex);
568 list_del_init(obd_list);
570 * If there are no obd attached to the timeout event
571 * list, remove this timeout event from the pinger
573 list_for_each_entry(item, &timeout_list, ti_chain) {
574 if (item->ti_event == event) {
579 LASSERTF(ti != NULL, "ti is NULL ! \n");
580 if (list_empty(&ti->ti_obd_list)) {
581 list_del(&ti->ti_chain);
584 mutex_unlock(&pinger_mutex);
587 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
589 int ptlrpc_pinger_remove_timeouts(void)
591 struct timeout_item *item, *tmp;
593 mutex_lock(&pinger_mutex);
594 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
595 LASSERT(list_empty(&item->ti_obd_list));
596 list_del(&item->ti_chain);
599 mutex_unlock(&pinger_mutex);
603 void ptlrpc_pinger_wake_up(void)
605 thread_add_flags(&pinger_thread, SVC_EVENT);
606 wake_up(&pinger_thread.t_ctl_waitq);
609 /* Ping evictor thread */
611 #define PET_TERMINATE 2
613 static int pet_refcount = 0;
614 static int pet_state;
615 static wait_queue_head_t pet_waitq;
617 static DEFINE_SPINLOCK(pet_lock);
619 int ping_evictor_wake(struct obd_export *exp)
621 struct obd_device *obd;
623 spin_lock(&pet_lock);
624 if (pet_state != PET_READY) {
625 /* eventually the new obd will call here again. */
626 spin_unlock(&pet_lock);
630 obd = class_exp2obd(exp);
631 if (list_empty(&obd->obd_evict_list)) {
632 class_incref(obd, "evictor", obd);
633 list_add(&obd->obd_evict_list, &pet_list);
635 spin_unlock(&pet_lock);
641 static int ping_evictor_main(void *arg)
643 struct obd_device *obd;
644 struct obd_export *exp;
645 struct l_wait_info lwi = { 0 };
650 CDEBUG(D_HA, "Starting Ping Evictor\n");
651 pet_state = PET_READY;
653 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
654 (pet_state == PET_TERMINATE), &lwi);
656 /* loop until all obd's will be removed */
657 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
660 /* we only get here if pet_exp != NULL, and the end of this
661 * loop is the only place which sets it NULL again, so lock
662 * is not strictly necessary. */
663 spin_lock(&pet_lock);
664 obd = list_entry(pet_list.next, struct obd_device,
666 spin_unlock(&pet_lock);
668 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
670 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
671 obd->obd_name, expire_time);
673 /* Exports can't be deleted out of the list while we hold
674 * the obd lock (class_unlink_export), which means we can't
675 * lose the last ref on the export. If they've already been
676 * removed from the list, we won't find them here. */
677 spin_lock(&obd->obd_dev_lock);
678 while (!list_empty(&obd->obd_exports_timed)) {
679 exp = list_entry(obd->obd_exports_timed.next,
681 exp_obd_chain_timed);
682 if (expire_time > exp->exp_last_request_time) {
683 class_export_get(exp);
684 spin_unlock(&obd->obd_dev_lock);
685 LCONSOLE_WARN("%s: haven't heard from client %s"
686 " (at %s) in %ld seconds. I think"
687 " it's dead, and I am evicting"
688 " it. exp %p, cur %ld expire %ld"
691 obd_uuid2str(&exp->exp_client_uuid),
692 obd_export_nid2str(exp),
693 (long)(cfs_time_current_sec() -
694 exp->exp_last_request_time),
695 exp, (long)cfs_time_current_sec(),
697 (long)exp->exp_last_request_time);
698 CDEBUG(D_HA, "Last request was at %ld\n",
699 exp->exp_last_request_time);
700 class_fail_export(exp);
701 class_export_put(exp);
702 spin_lock(&obd->obd_dev_lock);
704 /* List is sorted, so everyone below is ok */
708 spin_unlock(&obd->obd_dev_lock);
710 spin_lock(&pet_lock);
711 list_del_init(&obd->obd_evict_list);
712 spin_unlock(&pet_lock);
714 class_decref(obd, "evictor", obd);
716 CDEBUG(D_HA, "Exiting Ping Evictor\n");
721 void ping_evictor_start(void)
723 struct task_struct *task;
725 if (++pet_refcount > 1)
728 init_waitqueue_head(&pet_waitq);
730 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
733 CERROR("Cannot start ping evictor thread: %ld\n",
737 EXPORT_SYMBOL(ping_evictor_start);
739 void ping_evictor_stop(void)
741 if (--pet_refcount > 0)
744 pet_state = PET_TERMINATE;
747 EXPORT_SYMBOL(ping_evictor_stop);