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 module_param(suppress_pings, int, 0644);
49 MODULE_PARM_DESC(suppress_pings, "Suppress pings");
51 struct mutex pinger_mutex;
52 static LIST_HEAD(pinger_imports);
53 static struct list_head timeout_list = LIST_HEAD_INIT(timeout_list);
55 int ptlrpc_pinger_suppress_pings(void)
57 return suppress_pings;
59 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
61 struct ptlrpc_request *
62 ptlrpc_prep_ping(struct obd_import *imp)
64 struct ptlrpc_request *req;
66 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
67 LUSTRE_OBD_VERSION, OBD_PING);
69 ptlrpc_request_set_replen(req);
70 req->rq_no_resend = req->rq_no_delay = 1;
75 int ptlrpc_obd_ping(struct obd_device *obd)
78 struct ptlrpc_request *req;
80 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
84 req->rq_send_state = LUSTRE_IMP_FULL;
86 rc = ptlrpc_queue_wait(req);
88 ptlrpc_req_finished(req);
92 EXPORT_SYMBOL(ptlrpc_obd_ping);
94 int ptlrpc_ping(struct obd_import *imp)
96 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)
115 int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
116 if (imp->imp_state == LUSTRE_IMP_DISCON) {
117 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
119 at_get(&imp->imp_at.iat_net_latency));
120 time = min(time, dtime);
122 imp->imp_next_ping = cfs_time_shift(time);
125 void ptlrpc_ping_import_soon(struct obd_import *imp)
127 imp->imp_next_ping = cfs_time_current();
130 static inline int imp_is_deactive(struct obd_import *imp)
132 return (imp->imp_deactive ||
133 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
136 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
138 if (imp->imp_server_timeout)
139 return cfs_time_shift(obd_timeout / 2);
141 return cfs_time_shift(obd_timeout);
144 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
145 static cfs_time_t suspend_wakeup_time = 0;
147 cfs_duration_t pinger_check_timeout(cfs_time_t time)
149 struct timeout_item *item;
150 cfs_time_t timeout = PING_INTERVAL;
152 /* The timeout list is a increase order sorted list */
153 mutex_lock(&pinger_mutex);
154 list_for_each_entry(item, &timeout_list, ti_chain) {
155 int ti_timeout = item->ti_timeout;
156 if (timeout > ti_timeout)
157 timeout = ti_timeout;
160 mutex_unlock(&pinger_mutex);
162 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
166 static wait_queue_head_t suspend_timeouts_waitq;
168 cfs_time_t ptlrpc_suspend_wakeup_time(void)
170 return suspend_wakeup_time;
173 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
175 /*XXX: disabled for now, will be replaced by adaptive timeouts */
177 if (imp->imp_no_timeout)
179 imp->imp_no_timeout = 1;
180 atomic_inc(&suspend_timeouts);
181 CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
182 atomic_read(&suspend_timeouts));
186 void ptlrpc_activate_timeouts(struct obd_import *imp)
188 /*XXX: disabled for now, will be replaced by adaptive timeouts */
190 if (!imp->imp_no_timeout)
192 imp->imp_no_timeout = 0;
193 LASSERT(atomic_read(&suspend_timeouts) > 0);
194 if (atomic_dec_and_test(&suspend_timeouts)) {
195 suspend_wakeup_time = cfs_time_current();
196 wake_up(&suspend_timeouts_waitq);
198 CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
199 atomic_read(&suspend_timeouts));
203 int ptlrpc_check_suspend(void)
205 if (atomic_read(&suspend_timeouts))
210 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
212 struct l_wait_info lwi;
214 if (atomic_read(&suspend_timeouts)) {
215 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
216 atomic_read(&suspend_timeouts));
217 lwi = LWI_INTR(NULL, NULL);
218 l_wait_event(suspend_timeouts_waitq,
219 atomic_read(&suspend_timeouts) == 0, &lwi);
220 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
229 void ptlrpc_pinger_ir_up(void)
231 CDEBUG(D_HA, "IR up\n");
234 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
236 void ptlrpc_pinger_ir_down(void)
238 CDEBUG(D_HA, "IR down\n");
241 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
243 static void ptlrpc_pinger_process_import(struct obd_import *imp,
244 unsigned long this_ping)
251 spin_lock(&imp->imp_lock);
253 level = imp->imp_state;
254 force = imp->imp_force_verify;
255 force_next = imp->imp_force_next_verify;
257 * This will be used below only if the import is "FULL".
259 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
261 imp->imp_force_verify = 0;
263 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
265 spin_unlock(&imp->imp_lock);
269 imp->imp_force_next_verify = 0;
271 spin_unlock(&imp->imp_lock);
273 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
274 "force %u force_next %u deactive %u pingable %u suppress %u\n",
275 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
276 ptlrpc_import_state_name(level), level, force, force_next,
277 imp->imp_deactive, imp->imp_pingable, suppress);
279 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
280 /* wait for a while before trying recovery again */
281 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
282 if (!imp->imp_no_pinger_recover)
283 ptlrpc_initiate_recovery(imp);
284 } else if (level != LUSTRE_IMP_FULL ||
285 imp->imp_obd->obd_no_recov ||
286 imp_is_deactive(imp)) {
287 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
288 "or recovery disabled: %s)\n",
289 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
290 ptlrpc_import_state_name(level));
291 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
296 static int ptlrpc_pinger_main(void *arg)
298 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
300 /* Record that the thread is running */
301 thread_set_flags(thread, SVC_RUNNING);
302 wake_up(&thread->t_ctl_waitq);
304 /* And now, loop forever, pinging as needed. */
306 cfs_time_t this_ping = cfs_time_current();
307 struct l_wait_info lwi;
308 cfs_duration_t time_to_next_wake;
309 struct timeout_item *item;
310 struct list_head *iter;
312 mutex_lock(&pinger_mutex);
313 list_for_each_entry(item, &timeout_list, ti_chain) {
314 item->ti_cb(item, item->ti_cb_data);
316 list_for_each(iter, &pinger_imports) {
317 struct obd_import *imp =
318 list_entry(iter, struct obd_import,
321 ptlrpc_pinger_process_import(imp, this_ping);
322 /* obd_timeout might have changed */
323 if (imp->imp_pingable && imp->imp_next_ping &&
324 cfs_time_after(imp->imp_next_ping,
325 cfs_time_add(this_ping,
326 cfs_time_seconds(PING_INTERVAL))))
327 ptlrpc_update_next_ping(imp, 0);
329 mutex_unlock(&pinger_mutex);
330 /* update memory usage info */
331 obd_update_maxusage();
333 /* Wait until the next ping time, or until we're stopped. */
334 time_to_next_wake = pinger_check_timeout(this_ping);
335 /* The ping sent by ptlrpc_send_rpc may get sent out
336 say .01 second after this.
337 ptlrpc_pinger_sending_on_import will then set the
338 next ping time to next_ping + .01 sec, which means
339 we will SKIP the next ping at next_ping, and the
340 ping will get sent 2 timeouts from now! Beware. */
341 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
342 CFS_TIME_T")\n", time_to_next_wake,
343 cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
344 if (time_to_next_wake > 0) {
345 lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
347 cfs_time_seconds(1)),
349 l_wait_event(thread->t_ctl_waitq,
350 thread_is_stopping(thread) ||
351 thread_is_event(thread),
353 if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
356 /* woken after adding import to reset timer */
357 thread_test_and_clear_flags(thread, SVC_EVENT);
362 thread_set_flags(thread, SVC_STOPPED);
363 wake_up(&thread->t_ctl_waitq);
365 CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
369 static struct ptlrpc_thread pinger_thread;
371 int ptlrpc_start_pinger(void)
373 struct l_wait_info lwi = { 0 };
376 if (!thread_is_init(&pinger_thread) &&
377 !thread_is_stopped(&pinger_thread))
380 init_waitqueue_head(&pinger_thread.t_ctl_waitq);
381 init_waitqueue_head(&suspend_timeouts_waitq);
383 strcpy(pinger_thread.t_name, "ll_ping");
385 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
386 * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
387 rc = PTR_ERR(kthread_run(ptlrpc_pinger_main, &pinger_thread,
388 "%s", pinger_thread.t_name));
389 if (IS_ERR_VALUE(rc)) {
390 CERROR("cannot start thread: %d\n", rc);
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 };
413 if (thread_is_init(&pinger_thread) ||
414 thread_is_stopped(&pinger_thread))
417 ptlrpc_pinger_remove_timeouts();
418 thread_set_flags(&pinger_thread, SVC_STOPPING);
419 wake_up(&pinger_thread.t_ctl_waitq);
421 l_wait_event(pinger_thread.t_ctl_waitq,
422 thread_is_stopped(&pinger_thread), &lwi);
427 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
429 ptlrpc_update_next_ping(imp, 0);
431 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
433 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
435 ptlrpc_update_next_ping(imp, 1);
436 LASSERT(spin_is_locked(&imp->imp_lock));
438 * Avoid reading stale imp_connect_data. When not sure if pings are
439 * expected or not on next connection, we assume they are not and force
440 * one anyway to guarantee the chance of updating
441 * imp_peer_committed_transno.
443 if (imp->imp_state != LUSTRE_IMP_FULL ||
444 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
445 imp->imp_force_next_verify = 1;
448 int ptlrpc_pinger_add_import(struct obd_import *imp)
450 if (!list_empty(&imp->imp_pinger_chain))
453 mutex_lock(&pinger_mutex);
454 CDEBUG(D_HA, "adding pingable import %s->%s\n",
455 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
456 /* if we add to pinger we want recovery on this import */
457 imp->imp_obd->obd_no_recov = 0;
458 ptlrpc_update_next_ping(imp, 0);
459 /* XXX sort, blah blah */
460 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
461 class_import_get(imp);
463 ptlrpc_pinger_wake_up();
464 mutex_unlock(&pinger_mutex);
468 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
470 int ptlrpc_pinger_del_import(struct obd_import *imp)
472 if (list_empty(&imp->imp_pinger_chain))
475 mutex_lock(&pinger_mutex);
476 list_del_init(&imp->imp_pinger_chain);
477 CDEBUG(D_HA, "removing pingable import %s->%s\n",
478 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
479 /* if we remove from pinger we don't want recovery on this import */
480 imp->imp_obd->obd_no_recov = 1;
481 class_import_put(imp);
482 mutex_unlock(&pinger_mutex);
485 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
488 * Register a timeout callback to the pinger list, and the callback will
489 * be called when timeout happens.
491 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
492 timeout_cb_t cb, void *data)
494 struct timeout_item *ti;
500 INIT_LIST_HEAD(&ti->ti_obd_list);
501 INIT_LIST_HEAD(&ti->ti_chain);
502 ti->ti_timeout = time;
503 ti->ti_event = event;
505 ti->ti_cb_data = data;
511 * Register timeout event on the the pinger thread.
512 * Note: the timeout list is an sorted list with increased timeout value.
514 static struct timeout_item*
515 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
516 timeout_cb_t cb, void *data)
518 struct timeout_item *item, *tmp;
520 LASSERT(mutex_is_locked(&pinger_mutex));
522 list_for_each_entry(item, &timeout_list, ti_chain)
523 if (item->ti_event == event)
526 item = ptlrpc_new_timeout(time, event, cb, data);
528 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
529 if (tmp->ti_timeout < time) {
530 list_add(&item->ti_chain, &tmp->ti_chain);
534 list_add(&item->ti_chain, &timeout_list);
540 /* Add a client_obd to the timeout event list, when timeout(@time)
541 * happens, the callback(@cb) will be called.
543 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
544 timeout_cb_t cb, void *data,
545 struct list_head *obd_list)
547 struct timeout_item *ti;
549 mutex_lock(&pinger_mutex);
550 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
552 mutex_unlock(&pinger_mutex);
555 list_add(obd_list, &ti->ti_obd_list);
556 mutex_unlock(&pinger_mutex);
559 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
561 int ptlrpc_del_timeout_client(struct list_head *obd_list,
562 enum timeout_event event)
564 struct timeout_item *ti = NULL, *item;
566 if (list_empty(obd_list))
568 mutex_lock(&pinger_mutex);
569 list_del_init(obd_list);
571 * If there are no obd attached to the timeout event
572 * list, remove this timeout event from the pinger
574 list_for_each_entry(item, &timeout_list, ti_chain) {
575 if (item->ti_event == event) {
580 LASSERTF(ti != NULL, "ti is NULL !\n");
581 if (list_empty(&ti->ti_obd_list)) {
582 list_del(&ti->ti_chain);
585 mutex_unlock(&pinger_mutex);
588 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
590 int ptlrpc_pinger_remove_timeouts(void)
592 struct timeout_item *item, *tmp;
594 mutex_lock(&pinger_mutex);
595 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
596 LASSERT(list_empty(&item->ti_obd_list));
597 list_del(&item->ti_chain);
600 mutex_unlock(&pinger_mutex);
604 void ptlrpc_pinger_wake_up(void)
606 thread_add_flags(&pinger_thread, SVC_EVENT);
607 wake_up(&pinger_thread.t_ctl_waitq);
610 /* Ping evictor thread */
612 #define PET_TERMINATE 2
614 static int pet_refcount = 0;
615 static int pet_state;
616 static wait_queue_head_t pet_waitq;
618 static DEFINE_SPINLOCK(pet_lock);
620 int ping_evictor_wake(struct obd_export *exp)
622 struct obd_device *obd;
624 spin_lock(&pet_lock);
625 if (pet_state != PET_READY) {
626 /* eventually the new obd will call here again. */
627 spin_unlock(&pet_lock);
631 obd = class_exp2obd(exp);
632 if (list_empty(&obd->obd_evict_list)) {
633 class_incref(obd, "evictor", obd);
634 list_add(&obd->obd_evict_list, &pet_list);
636 spin_unlock(&pet_lock);
642 static int ping_evictor_main(void *arg)
644 struct obd_device *obd;
645 struct obd_export *exp;
646 struct l_wait_info lwi = { 0 };
651 CDEBUG(D_HA, "Starting Ping Evictor\n");
652 pet_state = PET_READY;
654 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
655 (pet_state == PET_TERMINATE), &lwi);
657 /* loop until all obd's will be removed */
658 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
661 /* we only get here if pet_exp != NULL, and the end of this
662 * loop is the only place which sets it NULL again, so lock
663 * is not strictly necessary. */
664 spin_lock(&pet_lock);
665 obd = list_entry(pet_list.next, struct obd_device,
667 spin_unlock(&pet_lock);
669 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
671 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
672 obd->obd_name, expire_time);
674 /* Exports can't be deleted out of the list while we hold
675 * the obd lock (class_unlink_export), which means we can't
676 * lose the last ref on the export. If they've already been
677 * removed from the list, we won't find them here. */
678 spin_lock(&obd->obd_dev_lock);
679 while (!list_empty(&obd->obd_exports_timed)) {
680 exp = list_entry(obd->obd_exports_timed.next,
682 exp_obd_chain_timed);
683 if (expire_time > exp->exp_last_request_time) {
684 class_export_get(exp);
685 spin_unlock(&obd->obd_dev_lock);
686 LCONSOLE_WARN("%s: haven't heard from client %s"
687 " (at %s) in %ld seconds. I think"
688 " it's dead, and I am evicting"
689 " it. exp %p, cur %ld expire %ld"
692 obd_uuid2str(&exp->exp_client_uuid),
693 obd_export_nid2str(exp),
694 (long)(cfs_time_current_sec() -
695 exp->exp_last_request_time),
696 exp, (long)cfs_time_current_sec(),
698 (long)exp->exp_last_request_time);
699 CDEBUG(D_HA, "Last request was at %ld\n",
700 exp->exp_last_request_time);
701 class_fail_export(exp);
702 class_export_put(exp);
703 spin_lock(&obd->obd_dev_lock);
705 /* List is sorted, so everyone below is ok */
709 spin_unlock(&obd->obd_dev_lock);
711 spin_lock(&pet_lock);
712 list_del_init(&obd->obd_evict_list);
713 spin_unlock(&pet_lock);
715 class_decref(obd, "evictor", obd);
717 CDEBUG(D_HA, "Exiting Ping Evictor\n");
722 void ping_evictor_start(void)
724 struct task_struct *task;
726 if (++pet_refcount > 1)
729 init_waitqueue_head(&pet_waitq);
731 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
734 CERROR("Cannot start ping evictor thread: %ld\n",
738 EXPORT_SYMBOL(ping_evictor_start);
740 void ping_evictor_stop(void)
742 if (--pet_refcount > 0)
745 pet_state = PET_TERMINATE;
748 EXPORT_SYMBOL(ping_evictor_stop);