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, 2015, 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 <linux/kthread.h>
44 #include <obd_support.h>
45 #include <obd_class.h>
46 #include "ptlrpc_internal.h"
48 static int suppress_pings;
49 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
51 struct mutex pinger_mutex;
52 static struct list_head pinger_imports =
53 LIST_HEAD_INIT(pinger_imports);
54 static struct list_head timeout_list =
55 LIST_HEAD_INIT(timeout_list);
57 int ptlrpc_pinger_suppress_pings()
59 return suppress_pings;
61 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
63 struct ptlrpc_request *
64 ptlrpc_prep_ping(struct obd_import *imp)
66 struct ptlrpc_request *req;
68 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
69 LUSTRE_OBD_VERSION, OBD_PING);
71 ptlrpc_request_set_replen(req);
72 req->rq_no_resend = req->rq_no_delay = 1;
77 int ptlrpc_obd_ping(struct obd_device *obd)
80 struct ptlrpc_request *req;
83 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
87 req->rq_send_state = LUSTRE_IMP_FULL;
89 rc = ptlrpc_queue_wait(req);
91 ptlrpc_req_finished(req);
95 EXPORT_SYMBOL(ptlrpc_obd_ping);
97 static int ptlrpc_ping(struct obd_import *imp)
99 struct ptlrpc_request *req;
102 req = ptlrpc_prep_ping(imp);
104 CERROR("OOM trying to ping %s->%s\n",
105 imp->imp_obd->obd_uuid.uuid,
106 obd2cli_tgt(imp->imp_obd));
110 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
111 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
112 ptlrpcd_add_req(req);
117 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
120 int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
121 if (imp->imp_state == LUSTRE_IMP_DISCON) {
122 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
124 at_get(&imp->imp_at.iat_net_latency));
125 time = min(time, dtime);
127 imp->imp_next_ping = cfs_time_shift(time);
128 #endif /* ENABLE_PINGER */
131 void ptlrpc_ping_import_soon(struct obd_import *imp)
133 imp->imp_next_ping = cfs_time_current();
136 static inline int imp_is_deactive(struct obd_import *imp)
138 return (imp->imp_deactive ||
139 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
142 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
144 if (imp->imp_server_timeout)
145 return cfs_time_shift(obd_timeout / 2);
147 return cfs_time_shift(obd_timeout);
150 static cfs_duration_t pinger_check_timeout(cfs_time_t time)
152 struct timeout_item *item;
153 cfs_time_t timeout = PING_INTERVAL;
155 /* This list is sorted in increasing timeout order */
156 mutex_lock(&pinger_mutex);
157 list_for_each_entry(item, &timeout_list, ti_chain) {
158 int ti_timeout = item->ti_timeout;
159 if (timeout > ti_timeout)
160 timeout = ti_timeout;
163 mutex_unlock(&pinger_mutex);
165 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
172 void ptlrpc_pinger_ir_up(void)
174 CDEBUG(D_HA, "IR up\n");
177 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
179 void ptlrpc_pinger_ir_down(void)
181 CDEBUG(D_HA, "IR down\n");
184 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
186 static void ptlrpc_pinger_process_import(struct obd_import *imp,
187 unsigned long this_ping)
194 spin_lock(&imp->imp_lock);
196 level = imp->imp_state;
197 force = imp->imp_force_verify;
198 force_next = imp->imp_force_next_verify;
200 * This will be used below only if the import is "FULL".
202 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
204 imp->imp_force_verify = 0;
206 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
208 spin_unlock(&imp->imp_lock);
212 imp->imp_force_next_verify = 0;
214 spin_unlock(&imp->imp_lock);
216 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
217 "force %u force_next %u deactive %u pingable %u suppress %u\n",
218 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
219 ptlrpc_import_state_name(level), level, force, force_next,
220 imp->imp_deactive, imp->imp_pingable, suppress);
222 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
223 /* wait for a while before trying recovery again */
224 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
225 if (!imp->imp_no_pinger_recover)
226 ptlrpc_initiate_recovery(imp);
227 } else if (level != LUSTRE_IMP_FULL ||
228 imp->imp_obd->obd_no_recov ||
229 imp_is_deactive(imp)) {
230 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
231 "or recovery disabled: %s)\n",
232 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
233 ptlrpc_import_state_name(level));
235 spin_lock(&imp->imp_lock);
236 imp->imp_force_verify = 1;
237 spin_unlock(&imp->imp_lock);
239 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
244 static int ptlrpc_pinger_main(void *arg)
246 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
249 /* Record that the thread is running */
250 thread_set_flags(thread, SVC_RUNNING);
251 wake_up(&thread->t_ctl_waitq);
253 /* And now, loop forever, pinging as needed. */
255 cfs_time_t this_ping = cfs_time_current();
256 struct l_wait_info lwi;
257 cfs_duration_t time_to_next_wake;
258 struct timeout_item *item;
259 struct list_head *iter;
261 mutex_lock(&pinger_mutex);
262 list_for_each_entry(item, &timeout_list, ti_chain)
263 item->ti_cb(item, item->ti_cb_data);
265 list_for_each(iter, &pinger_imports) {
266 struct obd_import *imp = list_entry(iter,
270 ptlrpc_pinger_process_import(imp, this_ping);
271 /* obd_timeout might have changed */
272 if (imp->imp_pingable && imp->imp_next_ping &&
273 cfs_time_after(imp->imp_next_ping,
274 cfs_time_add(this_ping,
275 cfs_time_seconds(PING_INTERVAL))))
276 ptlrpc_update_next_ping(imp, 0);
278 mutex_unlock(&pinger_mutex);
279 /* update memory usage info */
280 obd_update_maxusage();
282 /* Wait until the next ping time, or until we're stopped. */
283 time_to_next_wake = pinger_check_timeout(this_ping);
284 /* The ping sent by ptlrpc_send_rpc may get sent out
285 say .01 second after this.
286 ptlrpc_pinger_sending_on_import will then set the
287 next ping time to next_ping + .01 sec, which means
288 we will SKIP the next ping at next_ping, and the
289 ping will get sent 2 timeouts from now! Beware. */
290 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
291 CFS_TIME_T")\n", time_to_next_wake,
292 cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
293 if (time_to_next_wake > 0) {
294 lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
296 cfs_time_seconds(1)),
298 l_wait_event(thread->t_ctl_waitq,
299 thread_is_stopping(thread) ||
300 thread_is_event(thread),
302 if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
306 /* woken after adding import to reset timer */
307 thread_test_and_clear_flags(thread, SVC_EVENT);
312 thread_set_flags(thread, SVC_STOPPED);
313 wake_up(&thread->t_ctl_waitq);
315 CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
319 static struct ptlrpc_thread pinger_thread;
321 int ptlrpc_start_pinger(void)
323 struct l_wait_info lwi = { 0 };
324 struct task_struct *task;
326 #ifndef ENABLE_PINGER
331 if (!thread_is_init(&pinger_thread) &&
332 !thread_is_stopped(&pinger_thread))
335 init_waitqueue_head(&pinger_thread.t_ctl_waitq);
337 strcpy(pinger_thread.t_name, "ll_ping");
339 task = kthread_run(ptlrpc_pinger_main, &pinger_thread,
340 pinger_thread.t_name);
343 CERROR("cannot start pinger thread: rc = %d\n", rc);
347 l_wait_event(pinger_thread.t_ctl_waitq,
348 thread_is_running(&pinger_thread), &lwi);
351 CWARN("Pings will be suppressed at the request of the "
352 "administrator. The configuration shall meet the "
353 "additional requirements described in the manual. "
354 "(Search for the \"suppress_pings\" kernel module "
360 int ptlrpc_pinger_remove_timeouts(void);
362 int ptlrpc_stop_pinger(void)
364 struct l_wait_info lwi = { 0 };
365 #ifndef ENABLE_PINGER
370 if (thread_is_init(&pinger_thread) ||
371 thread_is_stopped(&pinger_thread))
374 ptlrpc_pinger_remove_timeouts();
376 thread_set_flags(&pinger_thread, SVC_STOPPING);
377 wake_up(&pinger_thread.t_ctl_waitq);
379 l_wait_event(pinger_thread.t_ctl_waitq,
380 thread_is_stopped(&pinger_thread), &lwi);
384 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
386 ptlrpc_update_next_ping(imp, 0);
389 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
391 ptlrpc_update_next_ping(imp, 1);
392 assert_spin_locked(&imp->imp_lock);
394 * Avoid reading stale imp_connect_data. When not sure if pings are
395 * expected or not on next connection, we assume they are not and force
396 * one anyway to guarantee the chance of updating
397 * imp_peer_committed_transno.
399 if (imp->imp_state != LUSTRE_IMP_FULL ||
400 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
401 imp->imp_force_next_verify = 1;
404 int ptlrpc_pinger_add_import(struct obd_import *imp)
407 if (!list_empty(&imp->imp_pinger_chain))
410 mutex_lock(&pinger_mutex);
411 CDEBUG(D_HA, "adding pingable import %s->%s\n",
412 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
413 /* if we add to pinger we want recovery on this import */
414 imp->imp_obd->obd_no_recov = 0;
415 ptlrpc_update_next_ping(imp, 0);
416 /* XXX sort, blah blah */
417 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
418 class_import_get(imp);
420 ptlrpc_pinger_wake_up();
421 mutex_unlock(&pinger_mutex);
425 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
427 int ptlrpc_pinger_del_import(struct obd_import *imp)
431 if (list_empty(&imp->imp_pinger_chain))
434 mutex_lock(&pinger_mutex);
435 list_del_init(&imp->imp_pinger_chain);
436 CDEBUG(D_HA, "removing pingable import %s->%s\n",
437 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
438 /* if we remove from pinger we don't want recovery on this import */
439 imp->imp_obd->obd_no_recov = 1;
440 class_import_put(imp);
441 mutex_unlock(&pinger_mutex);
444 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
447 * Register a timeout callback to the pinger list, and the callback will
448 * be called when timeout happens.
450 static struct timeout_item *ptlrpc_new_timeout(int time,
451 enum timeout_event event,
452 timeout_cb_t cb, void *data)
454 struct timeout_item *ti;
460 INIT_LIST_HEAD(&ti->ti_obd_list);
461 INIT_LIST_HEAD(&ti->ti_chain);
462 ti->ti_timeout = time;
463 ti->ti_event = event;
465 ti->ti_cb_data = data;
471 * Register timeout event on the the pinger thread.
472 * Note: the timeout list is an sorted list with increased timeout value.
474 static struct timeout_item*
475 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
476 timeout_cb_t cb, void *data)
478 struct timeout_item *item, *tmp;
480 LASSERT(mutex_is_locked(&pinger_mutex));
482 list_for_each_entry(item, &timeout_list, ti_chain)
483 if (item->ti_event == event)
486 item = ptlrpc_new_timeout(time, event, cb, data);
488 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
489 if (tmp->ti_timeout < time) {
490 list_add(&item->ti_chain, &tmp->ti_chain);
494 list_add(&item->ti_chain, &timeout_list);
500 /* Add a client_obd to the timeout event list, when timeout(@time)
501 * happens, the callback(@cb) will be called.
503 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
504 timeout_cb_t cb, void *data,
505 struct list_head *obd_list)
507 struct timeout_item *ti;
509 mutex_lock(&pinger_mutex);
510 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
512 mutex_unlock(&pinger_mutex);
515 list_add(obd_list, &ti->ti_obd_list);
516 mutex_unlock(&pinger_mutex);
519 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
521 int ptlrpc_del_timeout_client(struct list_head *obd_list,
522 enum timeout_event event)
524 struct timeout_item *ti = NULL, *item;
526 if (list_empty(obd_list))
528 mutex_lock(&pinger_mutex);
529 list_del_init(obd_list);
531 * If there are no obd attached to the timeout event
532 * list, remove this timeout event from the pinger
534 list_for_each_entry(item, &timeout_list, ti_chain) {
535 if (item->ti_event == event) {
540 LASSERTF(ti != NULL, "ti is NULL !\n");
541 if (list_empty(&ti->ti_obd_list)) {
542 list_del(&ti->ti_chain);
545 mutex_unlock(&pinger_mutex);
548 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
550 int ptlrpc_pinger_remove_timeouts(void)
552 struct timeout_item *item, *tmp;
554 mutex_lock(&pinger_mutex);
555 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
556 LASSERT(list_empty(&item->ti_obd_list));
557 list_del(&item->ti_chain);
560 mutex_unlock(&pinger_mutex);
564 void ptlrpc_pinger_wake_up()
567 thread_add_flags(&pinger_thread, SVC_EVENT);
568 wake_up(&pinger_thread.t_ctl_waitq);
572 /* Ping evictor thread */
574 #define PET_TERMINATE 2
576 static int pet_refcount = 0;
577 static int pet_state;
578 static wait_queue_head_t pet_waitq;
579 static struct list_head pet_list;
580 static DEFINE_SPINLOCK(pet_lock);
582 int ping_evictor_wake(struct obd_export *exp)
584 struct obd_device *obd;
586 spin_lock(&pet_lock);
587 if (pet_state != PET_READY) {
588 /* eventually the new obd will call here again. */
589 spin_unlock(&pet_lock);
593 obd = class_exp2obd(exp);
594 if (list_empty(&obd->obd_evict_list)) {
595 class_incref(obd, "evictor", obd);
596 list_add(&obd->obd_evict_list, &pet_list);
598 spin_unlock(&pet_lock);
604 static int ping_evictor_main(void *arg)
606 struct obd_device *obd;
607 struct obd_export *exp;
608 struct l_wait_info lwi = { 0 };
614 CDEBUG(D_HA, "Starting Ping Evictor\n");
615 pet_state = PET_READY;
617 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
618 (pet_state == PET_TERMINATE), &lwi);
620 /* loop until all obd's will be removed */
621 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
624 /* we only get here if pet_exp != NULL, and the end of this
625 * loop is the only place which sets it NULL again, so lock
626 * is not strictly necessary. */
627 spin_lock(&pet_lock);
628 obd = list_entry(pet_list.next, struct obd_device,
630 spin_unlock(&pet_lock);
632 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
634 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
635 obd->obd_name, expire_time);
637 /* Exports can't be deleted out of the list while we hold
638 * the obd lock (class_unlink_export), which means we can't
639 * lose the last ref on the export. If they've already been
640 * removed from the list, we won't find them here. */
641 spin_lock(&obd->obd_dev_lock);
642 while (!list_empty(&obd->obd_exports_timed)) {
643 exp = list_entry(obd->obd_exports_timed.next,
645 exp_obd_chain_timed);
646 if (expire_time > exp->exp_last_request_time) {
647 class_export_get(exp);
648 spin_unlock(&obd->obd_dev_lock);
649 LCONSOLE_WARN("%s: haven't heard from client %s"
650 " (at %s) in %ld seconds. I think"
651 " it's dead, and I am evicting"
652 " it. exp %p, cur %ld expire %ld"
655 obd_uuid2str(&exp->exp_client_uuid),
656 obd_export_nid2str(exp),
657 (long)(cfs_time_current_sec() -
658 exp->exp_last_request_time),
659 exp, (long)cfs_time_current_sec(),
661 (long)exp->exp_last_request_time);
662 CDEBUG(D_HA, "Last request was at %ld\n",
663 exp->exp_last_request_time);
664 class_fail_export(exp);
665 class_export_put(exp);
666 spin_lock(&obd->obd_dev_lock);
668 /* List is sorted, so everyone below is ok */
672 spin_unlock(&obd->obd_dev_lock);
674 spin_lock(&pet_lock);
675 list_del_init(&obd->obd_evict_list);
676 spin_unlock(&pet_lock);
678 class_decref(obd, "evictor", obd);
680 CDEBUG(D_HA, "Exiting Ping Evictor\n");
685 void ping_evictor_start(void)
687 struct task_struct *task;
689 if (++pet_refcount > 1)
692 INIT_LIST_HEAD(&pet_list);
693 init_waitqueue_head(&pet_waitq);
695 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
698 CERROR("Cannot start ping evictor thread: %ld\n",
702 EXPORT_SYMBOL(ping_evictor_start);
704 void ping_evictor_stop(void)
706 if (--pet_refcount > 0)
709 pet_state = PET_TERMINATE;
712 EXPORT_SYMBOL(ping_evictor_stop);