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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ptlrpc/pinger.c
34 * Portal-RPC reconnection and replay operations, for use in recovery.
37 #define DEBUG_SUBSYSTEM S_RPC
39 #include <linux/kthread.h>
40 #include <obd_support.h>
41 #include <obd_class.h>
42 #include "ptlrpc_internal.h"
44 static int suppress_pings;
45 module_param(suppress_pings, int, 0644);
46 MODULE_PARM_DESC(suppress_pings, "Suppress pings");
48 struct mutex pinger_mutex;
49 static struct list_head pinger_imports =
50 LIST_HEAD_INIT(pinger_imports);
51 static struct list_head timeout_list =
52 LIST_HEAD_INIT(timeout_list);
54 int ptlrpc_pinger_suppress_pings()
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;
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 static int ptlrpc_ping(struct obd_import *imp)
96 struct ptlrpc_request *req;
99 req = ptlrpc_prep_ping(imp);
101 CERROR("OOM trying to ping %s->%s\n",
102 imp->imp_obd->obd_uuid.uuid,
103 obd2cli_tgt(imp->imp_obd));
107 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
108 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
109 ptlrpcd_add_req(req);
114 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
117 int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
118 if (imp->imp_state == LUSTRE_IMP_DISCON) {
119 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
121 at_get(&imp->imp_at.iat_net_latency));
122 time = min(time, dtime);
124 imp->imp_next_ping = cfs_time_shift(time);
125 #endif /* ENABLE_PINGER */
128 void ptlrpc_ping_import_soon(struct obd_import *imp)
130 imp->imp_next_ping = cfs_time_current();
133 static inline int imp_is_deactive(struct obd_import *imp)
135 return (imp->imp_deactive ||
136 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
139 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
141 if (imp->imp_server_timeout)
142 return cfs_time_shift(obd_timeout / 2);
144 return cfs_time_shift(obd_timeout);
147 static cfs_duration_t pinger_check_timeout(cfs_time_t time)
149 struct timeout_item *item;
150 cfs_time_t timeout = PING_INTERVAL;
152 /* This list is sorted in increasing timeout order */
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)),
169 void ptlrpc_pinger_ir_up(void)
171 CDEBUG(D_HA, "IR up\n");
174 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
176 void ptlrpc_pinger_ir_down(void)
178 CDEBUG(D_HA, "IR down\n");
181 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
183 static void ptlrpc_pinger_process_import(struct obd_import *imp,
184 unsigned long this_ping)
191 spin_lock(&imp->imp_lock);
193 level = imp->imp_state;
194 force = imp->imp_force_verify;
195 force_next = imp->imp_force_next_verify;
197 * This will be used below only if the import is "FULL".
199 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
201 imp->imp_force_verify = 0;
203 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
205 spin_unlock(&imp->imp_lock);
209 imp->imp_force_next_verify = 0;
211 spin_unlock(&imp->imp_lock);
213 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
214 "force %u force_next %u deactive %u pingable %u suppress %u\n",
215 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
216 ptlrpc_import_state_name(level), level, force, force_next,
217 imp->imp_deactive, imp->imp_pingable, suppress);
219 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
220 /* wait for a while before trying recovery again */
221 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
222 if (!imp->imp_no_pinger_recover)
223 ptlrpc_initiate_recovery(imp);
224 } else if (level != LUSTRE_IMP_FULL ||
225 imp->imp_obd->obd_no_recov ||
226 imp_is_deactive(imp)) {
227 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
228 "or recovery disabled: %s)\n",
229 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
230 ptlrpc_import_state_name(level));
232 spin_lock(&imp->imp_lock);
233 imp->imp_force_verify = 1;
234 spin_unlock(&imp->imp_lock);
236 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
241 static int ptlrpc_pinger_main(void *arg)
243 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
246 /* Record that the thread is running */
247 thread_set_flags(thread, SVC_RUNNING);
248 wake_up(&thread->t_ctl_waitq);
250 /* And now, loop forever, pinging as needed. */
252 cfs_time_t this_ping = cfs_time_current();
253 struct l_wait_info lwi;
254 cfs_duration_t time_to_next_wake;
255 struct timeout_item *item;
256 struct list_head *iter;
258 mutex_lock(&pinger_mutex);
259 list_for_each_entry(item, &timeout_list, ti_chain)
260 item->ti_cb(item, item->ti_cb_data);
262 list_for_each(iter, &pinger_imports) {
263 struct obd_import *imp = list_entry(iter,
267 ptlrpc_pinger_process_import(imp, this_ping);
268 /* obd_timeout might have changed */
269 if (imp->imp_pingable && imp->imp_next_ping &&
270 cfs_time_after(imp->imp_next_ping,
271 cfs_time_add(this_ping,
272 cfs_time_seconds(PING_INTERVAL))))
273 ptlrpc_update_next_ping(imp, 0);
275 mutex_unlock(&pinger_mutex);
276 /* update memory usage info */
277 obd_update_maxusage();
279 /* Wait until the next ping time, or until we're stopped. */
280 time_to_next_wake = pinger_check_timeout(this_ping);
281 /* The ping sent by ptlrpc_send_rpc may get sent out
282 say .01 second after this.
283 ptlrpc_pinger_sending_on_import will then set the
284 next ping time to next_ping + .01 sec, which means
285 we will SKIP the next ping at next_ping, and the
286 ping will get sent 2 timeouts from now! Beware. */
287 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
288 CFS_TIME_T")\n", time_to_next_wake,
289 cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
290 if (time_to_next_wake > 0) {
291 lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
293 cfs_time_seconds(1)),
295 l_wait_event(thread->t_ctl_waitq,
296 thread_is_stopping(thread) ||
297 thread_is_event(thread),
299 if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
303 /* woken after adding import to reset timer */
304 thread_test_and_clear_flags(thread, SVC_EVENT);
309 thread_set_flags(thread, SVC_STOPPED);
310 wake_up(&thread->t_ctl_waitq);
312 CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
316 static struct ptlrpc_thread pinger_thread;
318 int ptlrpc_start_pinger(void)
320 struct l_wait_info lwi = { 0 };
321 struct task_struct *task;
323 #ifndef ENABLE_PINGER
328 if (!thread_is_init(&pinger_thread) &&
329 !thread_is_stopped(&pinger_thread))
332 init_waitqueue_head(&pinger_thread.t_ctl_waitq);
334 strcpy(pinger_thread.t_name, "ll_ping");
336 task = kthread_run(ptlrpc_pinger_main, &pinger_thread,
337 pinger_thread.t_name);
340 CERROR("cannot start pinger thread: rc = %d\n", rc);
344 l_wait_event(pinger_thread.t_ctl_waitq,
345 thread_is_running(&pinger_thread), &lwi);
348 CWARN("Pings will be suppressed at the request of the "
349 "administrator. The configuration shall meet the "
350 "additional requirements described in the manual. "
351 "(Search for the \"suppress_pings\" kernel module "
357 int ptlrpc_pinger_remove_timeouts(void);
359 int ptlrpc_stop_pinger(void)
361 struct l_wait_info lwi = { 0 };
362 #ifndef ENABLE_PINGER
367 if (thread_is_init(&pinger_thread) ||
368 thread_is_stopped(&pinger_thread))
371 ptlrpc_pinger_remove_timeouts();
373 thread_set_flags(&pinger_thread, SVC_STOPPING);
374 wake_up(&pinger_thread.t_ctl_waitq);
376 l_wait_event(pinger_thread.t_ctl_waitq,
377 thread_is_stopped(&pinger_thread), &lwi);
381 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
383 ptlrpc_update_next_ping(imp, 0);
386 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
388 ptlrpc_update_next_ping(imp, 1);
389 assert_spin_locked(&imp->imp_lock);
391 * Avoid reading stale imp_connect_data. When not sure if pings are
392 * expected or not on next connection, we assume they are not and force
393 * one anyway to guarantee the chance of updating
394 * imp_peer_committed_transno.
396 if (imp->imp_state != LUSTRE_IMP_FULL ||
397 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
398 imp->imp_force_next_verify = 1;
401 int ptlrpc_pinger_add_import(struct obd_import *imp)
404 if (!list_empty(&imp->imp_pinger_chain))
407 mutex_lock(&pinger_mutex);
408 CDEBUG(D_HA, "adding pingable import %s->%s\n",
409 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
410 /* if we add to pinger we want recovery on this import */
411 imp->imp_obd->obd_no_recov = 0;
412 ptlrpc_update_next_ping(imp, 0);
413 /* XXX sort, blah blah */
414 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
415 class_import_get(imp);
417 ptlrpc_pinger_wake_up();
418 mutex_unlock(&pinger_mutex);
422 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
424 int ptlrpc_pinger_del_import(struct obd_import *imp)
428 if (list_empty(&imp->imp_pinger_chain))
431 mutex_lock(&pinger_mutex);
432 list_del_init(&imp->imp_pinger_chain);
433 CDEBUG(D_HA, "removing pingable import %s->%s\n",
434 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
435 /* if we remove from pinger we don't want recovery on this import */
436 imp->imp_obd->obd_no_recov = 1;
437 class_import_put(imp);
438 mutex_unlock(&pinger_mutex);
441 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
444 * Register a timeout callback to the pinger list, and the callback will
445 * be called when timeout happens.
447 static struct timeout_item *ptlrpc_new_timeout(int time,
448 enum timeout_event event,
449 timeout_cb_t cb, void *data)
451 struct timeout_item *ti;
457 INIT_LIST_HEAD(&ti->ti_obd_list);
458 INIT_LIST_HEAD(&ti->ti_chain);
459 ti->ti_timeout = time;
460 ti->ti_event = event;
462 ti->ti_cb_data = data;
468 * Register timeout event on the the pinger thread.
469 * Note: the timeout list is an sorted list with increased timeout value.
471 static struct timeout_item*
472 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
473 timeout_cb_t cb, void *data)
475 struct timeout_item *item, *tmp;
477 LASSERT(mutex_is_locked(&pinger_mutex));
479 list_for_each_entry(item, &timeout_list, ti_chain)
480 if (item->ti_event == event)
483 item = ptlrpc_new_timeout(time, event, cb, data);
485 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
486 if (tmp->ti_timeout < time) {
487 list_add(&item->ti_chain, &tmp->ti_chain);
491 list_add(&item->ti_chain, &timeout_list);
497 /* Add a client_obd to the timeout event list, when timeout(@time)
498 * happens, the callback(@cb) will be called.
500 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
501 timeout_cb_t cb, void *data,
502 struct list_head *obd_list)
504 struct timeout_item *ti;
506 mutex_lock(&pinger_mutex);
507 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
509 mutex_unlock(&pinger_mutex);
512 list_add(obd_list, &ti->ti_obd_list);
513 mutex_unlock(&pinger_mutex);
516 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
518 int ptlrpc_del_timeout_client(struct list_head *obd_list,
519 enum timeout_event event)
521 struct timeout_item *ti = NULL, *item;
523 if (list_empty(obd_list))
525 mutex_lock(&pinger_mutex);
526 list_del_init(obd_list);
528 * If there are no obd attached to the timeout event
529 * list, remove this timeout event from the pinger
531 list_for_each_entry(item, &timeout_list, ti_chain) {
532 if (item->ti_event == event) {
537 LASSERTF(ti != NULL, "ti is NULL !\n");
538 if (list_empty(&ti->ti_obd_list)) {
539 list_del(&ti->ti_chain);
542 mutex_unlock(&pinger_mutex);
545 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
547 int ptlrpc_pinger_remove_timeouts(void)
549 struct timeout_item *item, *tmp;
551 mutex_lock(&pinger_mutex);
552 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
553 LASSERT(list_empty(&item->ti_obd_list));
554 list_del(&item->ti_chain);
557 mutex_unlock(&pinger_mutex);
561 void ptlrpc_pinger_wake_up()
564 thread_add_flags(&pinger_thread, SVC_EVENT);
565 wake_up(&pinger_thread.t_ctl_waitq);
569 /* Ping evictor thread */
571 #define PET_TERMINATE 2
573 static int pet_refcount = 0;
574 static int pet_state;
575 static wait_queue_head_t pet_waitq;
576 static struct list_head pet_list;
577 static DEFINE_SPINLOCK(pet_lock);
579 int ping_evictor_wake(struct obd_export *exp)
581 struct obd_device *obd;
583 spin_lock(&pet_lock);
584 if (pet_state != PET_READY) {
585 /* eventually the new obd will call here again. */
586 spin_unlock(&pet_lock);
590 obd = class_exp2obd(exp);
591 if (list_empty(&obd->obd_evict_list)) {
592 class_incref(obd, "evictor", obd);
593 list_add(&obd->obd_evict_list, &pet_list);
595 spin_unlock(&pet_lock);
601 static int ping_evictor_main(void *arg)
603 struct obd_device *obd;
604 struct obd_export *exp;
605 struct l_wait_info lwi = { 0 };
611 CDEBUG(D_HA, "Starting Ping Evictor\n");
612 pet_state = PET_READY;
614 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
615 (pet_state == PET_TERMINATE), &lwi);
617 /* loop until all obd's will be removed */
618 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
621 /* we only get here if pet_exp != NULL, and the end of this
622 * loop is the only place which sets it NULL again, so lock
623 * is not strictly necessary. */
624 spin_lock(&pet_lock);
625 obd = list_entry(pet_list.next, struct obd_device,
627 spin_unlock(&pet_lock);
629 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
631 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
632 obd->obd_name, expire_time);
634 /* Exports can't be deleted out of the list while we hold
635 * the obd lock (class_unlink_export), which means we can't
636 * lose the last ref on the export. If they've already been
637 * removed from the list, we won't find them here. */
638 spin_lock(&obd->obd_dev_lock);
639 while (!list_empty(&obd->obd_exports_timed)) {
640 exp = list_entry(obd->obd_exports_timed.next,
642 exp_obd_chain_timed);
643 if (expire_time > exp->exp_last_request_time) {
644 class_export_get(exp);
645 spin_unlock(&obd->obd_dev_lock);
646 LCONSOLE_WARN("%s: haven't heard from client %s"
647 " (at %s) in %ld seconds. I think"
648 " it's dead, and I am evicting"
649 " it. exp %p, cur %ld expire %ld"
652 obd_uuid2str(&exp->exp_client_uuid),
653 obd_export_nid2str(exp),
654 (long)(cfs_time_current_sec() -
655 exp->exp_last_request_time),
656 exp, (long)cfs_time_current_sec(),
658 (long)exp->exp_last_request_time);
659 CDEBUG(D_HA, "Last request was at %ld\n",
660 exp->exp_last_request_time);
661 class_fail_export(exp);
662 class_export_put(exp);
663 spin_lock(&obd->obd_dev_lock);
665 /* List is sorted, so everyone below is ok */
669 spin_unlock(&obd->obd_dev_lock);
671 spin_lock(&pet_lock);
672 list_del_init(&obd->obd_evict_list);
673 spin_unlock(&pet_lock);
675 class_decref(obd, "evictor", obd);
677 CDEBUG(D_HA, "Exiting Ping Evictor\n");
682 void ping_evictor_start(void)
684 struct task_struct *task;
686 if (++pet_refcount > 1)
689 INIT_LIST_HEAD(&pet_list);
690 init_waitqueue_head(&pet_waitq);
692 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
695 CERROR("Cannot start ping evictor thread: %ld\n",
699 EXPORT_SYMBOL(ping_evictor_start);
701 void ping_evictor_stop(void)
703 if (--pet_refcount > 0)
706 pet_state = PET_TERMINATE;
709 EXPORT_SYMBOL(ping_evictor_stop);