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 <linux/workqueue.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include "ptlrpc_internal.h"
45 static int suppress_pings;
46 module_param(suppress_pings, int, 0644);
47 MODULE_PARM_DESC(suppress_pings, "Suppress pings");
49 struct mutex pinger_mutex;
50 static struct list_head pinger_imports =
51 LIST_HEAD_INIT(pinger_imports);
52 static struct list_head timeout_list =
53 LIST_HEAD_INIT(timeout_list);
55 int ptlrpc_pinger_suppress_pings()
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;
81 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
85 req->rq_send_state = LUSTRE_IMP_FULL;
87 rc = ptlrpc_queue_wait(req);
89 ptlrpc_req_finished(req);
93 EXPORT_SYMBOL(ptlrpc_obd_ping);
95 static bool ptlrpc_check_import_is_idle(struct obd_import *imp)
97 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
100 if (!imp->imp_idle_timeout)
104 * - client_obd_setup() - hashed import
105 * - ptlrpcd_alloc_work()
106 * - ptlrpcd_alloc_work()
107 * - ptlrpc_pinger_add_import
109 if (atomic_read(&imp->imp_refcount) > 4)
112 /* any lock increases ns_bref being a resource holder */
113 if (ns && atomic_read(&ns->ns_bref) > 0)
116 now = ktime_get_real_seconds();
117 if (now - imp->imp_last_reply_time < imp->imp_idle_timeout)
123 static int ptlrpc_ping(struct obd_import *imp)
125 struct ptlrpc_request *req;
128 if (ptlrpc_check_import_is_idle(imp))
129 RETURN(ptlrpc_disconnect_and_idle_import(imp));
131 req = ptlrpc_prep_ping(imp);
133 CERROR("OOM trying to ping %s->%s\n",
134 imp->imp_obd->obd_uuid.uuid,
135 obd2cli_tgt(imp->imp_obd));
139 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
140 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
141 ptlrpcd_add_req(req);
146 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
149 time64_t time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
151 if (imp->imp_state == LUSTRE_IMP_DISCON) {
152 time64_t dtime = max_t(time64_t, CONNECTION_SWITCH_MIN,
154 at_get(&imp->imp_at.iat_net_latency));
155 time = min(time, dtime);
157 imp->imp_next_ping = ktime_get_seconds() + time;
158 #endif /* ENABLE_PINGER */
161 void ptlrpc_ping_import_soon(struct obd_import *imp)
163 imp->imp_next_ping = ktime_get_seconds();
166 static inline int imp_is_deactive(struct obd_import *imp)
168 return (imp->imp_deactive ||
169 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
172 static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp)
174 if (imp->imp_server_timeout)
175 return ktime_get_seconds() + (obd_timeout >> 1);
177 return ktime_get_seconds() + obd_timeout;
180 static time64_t pinger_check_timeout(time64_t time)
182 struct timeout_item *item;
183 time64_t timeout = PING_INTERVAL;
185 /* This list is sorted in increasing timeout order */
186 mutex_lock(&pinger_mutex);
187 list_for_each_entry(item, &timeout_list, ti_chain) {
188 time64_t ti_timeout = item->ti_timeout;
190 if (timeout > ti_timeout)
191 timeout = ti_timeout;
194 mutex_unlock(&pinger_mutex);
196 return time + timeout - ktime_get_seconds();
201 void ptlrpc_pinger_ir_up(void)
203 CDEBUG(D_HA, "IR up\n");
206 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
208 void ptlrpc_pinger_ir_down(void)
210 CDEBUG(D_HA, "IR down\n");
213 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
215 static void ptlrpc_pinger_process_import(struct obd_import *imp,
223 spin_lock(&imp->imp_lock);
225 level = imp->imp_state;
226 force = imp->imp_force_verify;
227 force_next = imp->imp_force_next_verify;
229 * This will be used below only if the import is "FULL".
231 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
233 imp->imp_force_verify = 0;
235 if (imp->imp_next_ping - 5 >= this_ping && !force) {
236 spin_unlock(&imp->imp_lock);
240 imp->imp_force_next_verify = 0;
242 spin_unlock(&imp->imp_lock);
244 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
245 "force %u force_next %u deactive %u pingable %u suppress %u\n",
246 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
247 ptlrpc_import_state_name(level), level, force, force_next,
248 imp->imp_deactive, imp->imp_pingable, suppress);
250 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
251 /* wait for a while before trying recovery again */
252 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
253 if (!imp->imp_no_pinger_recover)
254 ptlrpc_initiate_recovery(imp);
255 } else if (level != LUSTRE_IMP_FULL ||
256 imp->imp_obd->obd_no_recov ||
257 imp_is_deactive(imp)) {
258 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
259 "or recovery disabled: %s)\n",
260 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
261 ptlrpc_import_state_name(level));
263 spin_lock(&imp->imp_lock);
264 imp->imp_force_verify = 1;
265 spin_unlock(&imp->imp_lock);
267 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
272 static struct workqueue_struct *pinger_wq;
273 static void ptlrpc_pinger_main(struct work_struct *ws);
274 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
276 static void ptlrpc_pinger_main(struct work_struct *ws)
278 time64_t this_ping = ktime_get_seconds();
279 time64_t time_to_next_wake;
280 struct timeout_item *item;
281 struct obd_import *imp;
282 struct list_head *iter;
285 mutex_lock(&pinger_mutex);
286 list_for_each_entry(item, &timeout_list, ti_chain)
287 item->ti_cb(item, item->ti_cb_data);
289 list_for_each(iter, &pinger_imports) {
290 imp = list_entry(iter, struct obd_import,
293 ptlrpc_pinger_process_import(imp, this_ping);
294 /* obd_timeout might have changed */
295 if (imp->imp_pingable && imp->imp_next_ping &&
296 imp->imp_next_ping > this_ping + PING_INTERVAL)
297 ptlrpc_update_next_ping(imp, 0);
299 mutex_unlock(&pinger_mutex);
300 /* update memory usage info */
301 obd_update_maxusage();
303 /* Wait until the next ping time, or until we're stopped. */
304 time_to_next_wake = pinger_check_timeout(this_ping);
305 /* The ping sent by ptlrpc_send_rpc may get sent out
306 * say .01 second after this.
307 * ptlrpc_pinger_sending_on_import will then set the
308 * next ping time to next_ping + .01 sec, which means
309 * we will SKIP the next ping at next_ping, and the
310 * ping will get sent 2 timeouts from now! Beware. */
311 CDEBUG(D_INFO, "next wakeup in %lld (%lld)\n",
312 time_to_next_wake, this_ping + PING_INTERVAL);
313 } while (time_to_next_wake <= 0);
315 queue_delayed_work(pinger_wq, &ping_work,
316 cfs_time_seconds(max(time_to_next_wake, 1LL)));
319 int ptlrpc_start_pinger(void)
325 pinger_wq = alloc_workqueue("ptlrpc_pinger", 0, 1);
327 CERROR("cannot start pinger workqueue\n");
331 queue_delayed_work(pinger_wq, &ping_work, 0);
334 CWARN("Pings will be suppressed at the request of the "
335 "administrator. The configuration shall meet the "
336 "additional requirements described in the manual. "
337 "(Search for the \"suppress_pings\" kernel module "
343 int ptlrpc_pinger_remove_timeouts(void);
345 int ptlrpc_stop_pinger(void)
351 ptlrpc_pinger_remove_timeouts();
353 cancel_delayed_work_sync(&ping_work);
354 destroy_workqueue(pinger_wq);
360 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
362 ptlrpc_update_next_ping(imp, 0);
365 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
367 ptlrpc_update_next_ping(imp, 1);
368 assert_spin_locked(&imp->imp_lock);
370 * Avoid reading stale imp_connect_data. When not sure if pings are
371 * expected or not on next connection, we assume they are not and force
372 * one anyway to guarantee the chance of updating
373 * imp_peer_committed_transno.
375 if (imp->imp_state != LUSTRE_IMP_FULL ||
376 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
377 imp->imp_force_next_verify = 1;
380 int ptlrpc_pinger_add_import(struct obd_import *imp)
383 if (!list_empty(&imp->imp_pinger_chain))
386 mutex_lock(&pinger_mutex);
387 CDEBUG(D_HA, "adding pingable import %s->%s\n",
388 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
389 /* if we add to pinger we want recovery on this import */
390 imp->imp_obd->obd_no_recov = 0;
391 ptlrpc_update_next_ping(imp, 0);
392 /* XXX sort, blah blah */
393 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
394 class_import_get(imp);
396 ptlrpc_pinger_wake_up();
397 mutex_unlock(&pinger_mutex);
401 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
403 int ptlrpc_pinger_del_import(struct obd_import *imp)
407 if (list_empty(&imp->imp_pinger_chain))
410 mutex_lock(&pinger_mutex);
411 list_del_init(&imp->imp_pinger_chain);
412 CDEBUG(D_HA, "removing pingable import %s->%s\n",
413 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
414 /* if we remove from pinger we don't want recovery on this import */
415 imp->imp_obd->obd_no_recov = 1;
416 class_import_put(imp);
417 mutex_unlock(&pinger_mutex);
420 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
423 * Register a timeout callback to the pinger list, and the callback will
424 * be called when timeout happens.
426 static struct timeout_item *ptlrpc_new_timeout(time64_t time,
427 enum timeout_event event,
428 timeout_cb_t cb, void *data)
430 struct timeout_item *ti;
436 INIT_LIST_HEAD(&ti->ti_obd_list);
437 INIT_LIST_HEAD(&ti->ti_chain);
438 ti->ti_timeout = time;
439 ti->ti_event = event;
441 ti->ti_cb_data = data;
447 * Register timeout event on the the pinger thread.
448 * Note: the timeout list is an sorted list with increased timeout value.
450 static struct timeout_item*
451 ptlrpc_pinger_register_timeout(time64_t time, enum timeout_event event,
452 timeout_cb_t cb, void *data)
454 struct timeout_item *item, *tmp;
456 LASSERT(mutex_is_locked(&pinger_mutex));
458 list_for_each_entry(item, &timeout_list, ti_chain)
459 if (item->ti_event == event)
462 item = ptlrpc_new_timeout(time, event, cb, data);
464 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
465 if (tmp->ti_timeout < time) {
466 list_add(&item->ti_chain, &tmp->ti_chain);
470 list_add(&item->ti_chain, &timeout_list);
476 /* Add a client_obd to the timeout event list, when timeout(@time)
477 * happens, the callback(@cb) will be called.
479 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
480 timeout_cb_t cb, void *data,
481 struct list_head *obd_list)
483 struct timeout_item *ti;
485 mutex_lock(&pinger_mutex);
486 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
488 mutex_unlock(&pinger_mutex);
491 list_add(obd_list, &ti->ti_obd_list);
492 mutex_unlock(&pinger_mutex);
495 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
497 int ptlrpc_del_timeout_client(struct list_head *obd_list,
498 enum timeout_event event)
500 struct timeout_item *ti = NULL, *item;
502 if (list_empty(obd_list))
504 mutex_lock(&pinger_mutex);
505 list_del_init(obd_list);
507 * If there are no obd attached to the timeout event
508 * list, remove this timeout event from the pinger
510 list_for_each_entry(item, &timeout_list, ti_chain) {
511 if (item->ti_event == event) {
516 LASSERTF(ti != NULL, "ti is NULL !\n");
517 if (list_empty(&ti->ti_obd_list)) {
518 list_del(&ti->ti_chain);
521 mutex_unlock(&pinger_mutex);
524 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
526 int ptlrpc_pinger_remove_timeouts(void)
528 struct timeout_item *item, *tmp;
530 mutex_lock(&pinger_mutex);
531 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
532 LASSERT(list_empty(&item->ti_obd_list));
533 list_del(&item->ti_chain);
536 mutex_unlock(&pinger_mutex);
540 void ptlrpc_pinger_wake_up()
543 mod_delayed_work(pinger_wq, &ping_work, 0);
547 /* Ping evictor thread */
549 #define PET_TERMINATE 2
551 static int pet_refcount = 0;
552 static int pet_state;
553 static wait_queue_head_t pet_waitq;
554 static struct list_head pet_list;
555 static DEFINE_SPINLOCK(pet_lock);
557 int ping_evictor_wake(struct obd_export *exp)
559 struct obd_device *obd;
561 spin_lock(&pet_lock);
562 if (pet_state != PET_READY) {
563 /* eventually the new obd will call here again. */
564 spin_unlock(&pet_lock);
568 obd = class_exp2obd(exp);
569 if (list_empty(&obd->obd_evict_list)) {
570 class_incref(obd, "evictor", obd);
571 list_add(&obd->obd_evict_list, &pet_list);
573 spin_unlock(&pet_lock);
579 static int ping_evictor_main(void *arg)
581 struct obd_device *obd;
582 struct obd_export *exp;
583 struct l_wait_info lwi = { 0 };
584 time64_t expire_time;
589 CDEBUG(D_HA, "Starting Ping Evictor\n");
590 pet_state = PET_READY;
592 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
593 (pet_state == PET_TERMINATE), &lwi);
595 /* loop until all obd's will be removed */
596 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
599 /* we only get here if pet_exp != NULL, and the end of this
600 * loop is the only place which sets it NULL again, so lock
601 * is not strictly necessary. */
602 spin_lock(&pet_lock);
603 obd = list_entry(pet_list.next, struct obd_device,
605 spin_unlock(&pet_lock);
607 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
609 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
610 obd->obd_name, expire_time);
612 /* Exports can't be deleted out of the list while we hold
613 * the obd lock (class_unlink_export), which means we can't
614 * lose the last ref on the export. If they've already been
615 * removed from the list, we won't find them here. */
616 spin_lock(&obd->obd_dev_lock);
617 while (!list_empty(&obd->obd_exports_timed)) {
618 exp = list_entry(obd->obd_exports_timed.next,
620 exp_obd_chain_timed);
621 if (expire_time > exp->exp_last_request_time) {
622 class_export_get(exp);
623 spin_unlock(&obd->obd_dev_lock);
624 LCONSOLE_WARN("%s: haven't heard from client %s"
625 " (at %s) in %lld seconds. I think"
626 " it's dead, and I am evicting"
627 " it. exp %p, cur %lld expire %lld"
630 obd_uuid2str(&exp->exp_client_uuid),
631 obd_export_nid2str(exp),
632 ktime_get_real_seconds() -
633 exp->exp_last_request_time,
634 exp, ktime_get_real_seconds(),
636 exp->exp_last_request_time);
637 CDEBUG(D_HA, "Last request was at %lld\n",
638 exp->exp_last_request_time);
639 class_fail_export(exp);
640 class_export_put(exp);
641 spin_lock(&obd->obd_dev_lock);
643 /* List is sorted, so everyone below is ok */
647 spin_unlock(&obd->obd_dev_lock);
649 spin_lock(&pet_lock);
650 list_del_init(&obd->obd_evict_list);
651 spin_unlock(&pet_lock);
653 class_decref(obd, "evictor", obd);
655 CDEBUG(D_HA, "Exiting Ping Evictor\n");
660 void ping_evictor_start(void)
662 struct task_struct *task;
664 if (++pet_refcount > 1)
667 INIT_LIST_HEAD(&pet_list);
668 init_waitqueue_head(&pet_waitq);
670 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
673 CERROR("Cannot start ping evictor thread: %ld\n",
677 EXPORT_SYMBOL(ping_evictor_start);
679 void ping_evictor_stop(void)
681 if (--pet_refcount > 0)
684 pet_state = PET_TERMINATE;
687 EXPORT_SYMBOL(ping_evictor_stop);