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(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;
82 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
86 req->rq_send_state = LUSTRE_IMP_FULL;
88 rc = ptlrpc_queue_wait(req);
90 ptlrpc_req_finished(req);
94 EXPORT_SYMBOL(ptlrpc_obd_ping);
96 static bool ptlrpc_check_import_is_idle(struct obd_import *imp)
98 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
101 if (!imp->imp_idle_timeout)
104 if (atomic_read(&imp->imp_reqs) > 0)
107 /* any lock increases ns_bref being a resource holder */
108 if (ns && atomic_read(&ns->ns_bref) > 0)
111 now = ktime_get_real_seconds();
112 if (now - imp->imp_last_reply_time < imp->imp_idle_timeout)
118 static int ptlrpc_ping(struct obd_import *imp)
120 struct ptlrpc_request *req;
124 if (ptlrpc_check_import_is_idle(imp))
125 RETURN(ptlrpc_disconnect_and_idle_import(imp));
127 req = ptlrpc_prep_ping(imp);
129 CERROR("OOM trying to ping %s->%s\n",
130 imp->imp_obd->obd_uuid.uuid,
131 obd2cli_tgt(imp->imp_obd));
135 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
136 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
137 ptlrpcd_add_req(req);
142 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
144 #ifdef CONFIG_LUSTRE_FS_PINGER
145 time64_t time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
147 if (imp->imp_state == LUSTRE_IMP_DISCON) {
148 time64_t dtime = max_t(time64_t, CONNECTION_SWITCH_MIN,
150 at_get(&imp->imp_at.iat_net_latency));
151 time = min(time, dtime);
153 imp->imp_next_ping = ktime_get_seconds() + time;
154 #endif /* CONFIG_LUSTRE_FS_PINGER */
157 void ptlrpc_ping_import_soon(struct obd_import *imp)
159 imp->imp_next_ping = ktime_get_seconds();
162 static inline int imp_is_deactive(struct obd_import *imp)
164 return imp->imp_deactive ||
165 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE);
168 static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp)
170 if (imp->imp_server_timeout)
171 return ktime_get_seconds() + (obd_timeout >> 1);
173 return ktime_get_seconds() + obd_timeout;
176 static time64_t pinger_check_timeout(time64_t time)
178 struct timeout_item *item;
179 time64_t timeout = PING_INTERVAL;
181 /* This list is sorted in increasing timeout order */
182 mutex_lock(&pinger_mutex);
183 list_for_each_entry(item, &timeout_list, ti_chain) {
184 time64_t ti_timeout = item->ti_timeout;
186 if (timeout > ti_timeout)
187 timeout = ti_timeout;
190 mutex_unlock(&pinger_mutex);
192 return time + timeout - ktime_get_seconds();
197 void ptlrpc_pinger_ir_up(void)
199 CDEBUG(D_HA, "IR up\n");
202 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
204 void ptlrpc_pinger_ir_down(void)
206 CDEBUG(D_HA, "IR down\n");
209 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
211 static void ptlrpc_pinger_process_import(struct obd_import *imp,
219 spin_lock(&imp->imp_lock);
221 level = imp->imp_state;
222 force = imp->imp_force_verify;
223 force_next = imp->imp_force_next_verify;
225 * This will be used below only if the import is "FULL".
227 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
229 imp->imp_force_verify = 0;
231 if (imp->imp_next_ping - 5 >= this_ping && !force) {
232 spin_unlock(&imp->imp_lock);
236 imp->imp_force_next_verify = 0;
238 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
239 "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n",
240 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
241 ptlrpc_import_state_name(level), level, force, force_next,
242 imp->imp_deactive, imp->imp_pingable, suppress);
244 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
245 /* wait for a while before trying recovery again */
246 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
247 spin_unlock(&imp->imp_lock);
248 if (!imp->imp_no_pinger_recover ||
249 imp->imp_connect_error == -EAGAIN)
250 ptlrpc_initiate_recovery(imp);
251 } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov ||
252 imp_is_deactive(imp)) {
254 "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
255 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
256 ptlrpc_import_state_name(level));
258 imp->imp_force_verify = 1;
259 spin_unlock(&imp->imp_lock);
260 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
261 spin_unlock(&imp->imp_lock);
264 spin_unlock(&imp->imp_lock);
268 static struct workqueue_struct *pinger_wq;
269 static void ptlrpc_pinger_main(struct work_struct *ws);
270 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
272 static void ptlrpc_pinger_main(struct work_struct *ws)
274 time64_t this_ping = ktime_get_seconds();
275 time64_t time_to_next_wake;
276 struct timeout_item *item;
277 struct obd_import *imp;
278 struct list_head *iter;
281 mutex_lock(&pinger_mutex);
282 list_for_each_entry(item, &timeout_list, ti_chain)
283 item->ti_cb(item, item->ti_cb_data);
285 list_for_each(iter, &pinger_imports) {
286 imp = list_entry(iter, struct obd_import,
289 ptlrpc_pinger_process_import(imp, this_ping);
290 /* obd_timeout might have changed */
291 if (imp->imp_pingable && imp->imp_next_ping &&
292 imp->imp_next_ping > this_ping + PING_INTERVAL)
293 ptlrpc_update_next_ping(imp, 0);
295 mutex_unlock(&pinger_mutex);
296 /* update memory usage info */
297 obd_update_maxusage();
299 /* Wait until the next ping time, or until we're stopped. */
300 time_to_next_wake = pinger_check_timeout(this_ping);
302 * The ping sent by ptlrpc_send_rpc may get sent out
303 * say .01 second after this.
304 * ptlrpc_pinger_sending_on_import will then set the
305 * next ping time to next_ping + .01 sec, which means
306 * we will SKIP the next ping at next_ping, and the
307 * ping will get sent 2 timeouts from now! Beware.
309 CDEBUG(D_INFO, "next wakeup in %lld (%lld)\n",
310 time_to_next_wake, this_ping + PING_INTERVAL);
311 } while (time_to_next_wake <= 0);
313 queue_delayed_work(pinger_wq, &ping_work,
314 cfs_time_seconds(max(time_to_next_wake, 1LL)));
317 int ptlrpc_start_pinger(void)
319 #ifdef CONFIG_LUSTRE_FS_PINGER
323 pinger_wq = cfs_cpt_bind_workqueue("ptlrpc_pinger", cfs_cpt_tab,
325 if (IS_ERR(pinger_wq)) {
326 CERROR("cannot start pinger workqueue\n");
327 return PTR_ERR(pinger_wq);
330 queue_delayed_work(pinger_wq, &ping_work, 0);
333 CWARN("Pings will be suppressed at the request of the administrator. The configuration shall meet the additional requirements described in the manual. (Search for the \"suppress_pings\" kernel module parameter.)\n");
338 int ptlrpc_pinger_remove_timeouts(void);
340 int ptlrpc_stop_pinger(void)
342 #ifdef CONFIG_LUSTRE_FS_PINGER
346 ptlrpc_pinger_remove_timeouts();
348 cancel_delayed_work_sync(&ping_work);
349 destroy_workqueue(pinger_wq);
355 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
357 ptlrpc_update_next_ping(imp, 0);
360 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
362 ptlrpc_update_next_ping(imp, 1);
363 assert_spin_locked(&imp->imp_lock);
365 * Avoid reading stale imp_connect_data. When not sure if pings are
366 * expected or not on next connection, we assume they are not and force
367 * one anyway to guarantee the chance of updating
368 * imp_peer_committed_transno.
370 if (imp->imp_state != LUSTRE_IMP_FULL ||
371 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
372 imp->imp_force_next_verify = 1;
375 int ptlrpc_pinger_add_import(struct obd_import *imp)
378 if (!list_empty(&imp->imp_pinger_chain))
381 mutex_lock(&pinger_mutex);
382 CDEBUG(D_HA, "adding pingable import %s->%s\n",
383 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
384 /* if we add to pinger we want recovery on this import */
385 imp->imp_obd->obd_no_recov = 0;
386 ptlrpc_update_next_ping(imp, 0);
387 /* XXX sort, blah blah */
388 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
389 class_import_get(imp);
391 ptlrpc_pinger_wake_up();
392 mutex_unlock(&pinger_mutex);
396 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
398 int ptlrpc_pinger_del_import(struct obd_import *imp)
402 if (list_empty(&imp->imp_pinger_chain))
405 mutex_lock(&pinger_mutex);
406 list_del_init(&imp->imp_pinger_chain);
407 CDEBUG(D_HA, "removing pingable import %s->%s\n",
408 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
409 /* if we remove from pinger we don't want recovery on this import */
410 imp->imp_obd->obd_no_recov = 1;
411 class_import_put(imp);
412 mutex_unlock(&pinger_mutex);
415 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
418 * Register a timeout callback to the pinger list, and the callback will
419 * be called when timeout happens.
421 static struct timeout_item *ptlrpc_new_timeout(time64_t time,
422 enum timeout_event event,
423 timeout_cb_t cb, void *data)
425 struct timeout_item *ti;
431 INIT_LIST_HEAD(&ti->ti_obd_list);
432 INIT_LIST_HEAD(&ti->ti_chain);
433 ti->ti_timeout = time;
434 ti->ti_event = event;
436 ti->ti_cb_data = data;
442 * Register timeout event on the the pinger thread.
443 * Note: the timeout list is an sorted list with increased timeout value.
445 static struct timeout_item*
446 ptlrpc_pinger_register_timeout(time64_t time, enum timeout_event event,
447 timeout_cb_t cb, void *data)
449 struct timeout_item *item, *tmp;
451 LASSERT(mutex_is_locked(&pinger_mutex));
453 list_for_each_entry(item, &timeout_list, ti_chain)
454 if (item->ti_event == event)
457 item = ptlrpc_new_timeout(time, event, cb, data);
459 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
460 if (tmp->ti_timeout < time) {
461 list_add(&item->ti_chain, &tmp->ti_chain);
465 list_add(&item->ti_chain, &timeout_list);
471 /* Add a client_obd to the timeout event list, when timeout(@time)
472 * happens, the callback(@cb) will be called.
474 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
475 timeout_cb_t cb, void *data,
476 struct list_head *obd_list)
478 struct timeout_item *ti;
480 mutex_lock(&pinger_mutex);
481 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
483 mutex_unlock(&pinger_mutex);
486 list_add(obd_list, &ti->ti_obd_list);
487 mutex_unlock(&pinger_mutex);
490 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
492 int ptlrpc_del_timeout_client(struct list_head *obd_list,
493 enum timeout_event event)
495 struct timeout_item *ti = NULL, *item;
497 if (list_empty(obd_list))
499 mutex_lock(&pinger_mutex);
500 list_del_init(obd_list);
502 * If there are no obd attached to the timeout event
503 * list, remove this timeout event from the pinger
505 list_for_each_entry(item, &timeout_list, ti_chain) {
506 if (item->ti_event == event) {
511 LASSERTF(ti, "ti is NULL !\n");
512 if (list_empty(&ti->ti_obd_list)) {
513 list_del(&ti->ti_chain);
516 mutex_unlock(&pinger_mutex);
519 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
521 int ptlrpc_pinger_remove_timeouts(void)
523 struct timeout_item *item, *tmp;
525 mutex_lock(&pinger_mutex);
526 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
527 LASSERT(list_empty(&item->ti_obd_list));
528 list_del(&item->ti_chain);
531 mutex_unlock(&pinger_mutex);
535 void ptlrpc_pinger_wake_up(void)
537 #ifdef CONFIG_LUSTRE_FS_PINGER
538 mod_delayed_work(pinger_wq, &ping_work, 0);
542 /* Ping evictor thread */
544 #define PET_TERMINATE 2
546 static int pet_refcount;
547 static int pet_state;
548 static wait_queue_head_t pet_waitq;
549 static LIST_HEAD(pet_list);
550 static DEFINE_SPINLOCK(pet_lock);
552 int ping_evictor_wake(struct obd_export *exp)
554 struct obd_device *obd;
556 spin_lock(&pet_lock);
557 if (pet_state != PET_READY) {
558 /* eventually the new obd will call here again. */
559 spin_unlock(&pet_lock);
563 obd = class_exp2obd(exp);
564 if (list_empty(&obd->obd_evict_list)) {
565 class_incref(obd, "evictor", obd);
566 list_add(&obd->obd_evict_list, &pet_list);
568 spin_unlock(&pet_lock);
574 static int ping_evictor_main(void *arg)
576 struct obd_device *obd;
577 struct obd_export *exp;
578 time64_t expire_time;
583 CDEBUG(D_HA, "Starting Ping Evictor\n");
584 pet_state = PET_READY;
586 wait_event_idle(pet_waitq,
587 (!list_empty(&pet_list)) ||
588 (pet_state == PET_TERMINATE));
590 /* loop until all obd's will be removed */
591 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
595 * we only get here if pet_exp != NULL, and the end of this
596 * loop is the only place which sets it NULL again, so lock
597 * is not strictly necessary.
599 spin_lock(&pet_lock);
600 obd = list_entry(pet_list.next, struct obd_device,
602 spin_unlock(&pet_lock);
604 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
606 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
607 obd->obd_name, expire_time);
610 * Exports can't be deleted out of the list while we hold
611 * the obd lock (class_unlink_export), which means we can't
612 * lose the last ref on the export. If they've already been
613 * removed from the list, we won't find them here.
615 spin_lock(&obd->obd_dev_lock);
616 while (!list_empty(&obd->obd_exports_timed)) {
617 exp = list_entry(obd->obd_exports_timed.next,
619 exp_obd_chain_timed);
620 if (expire_time > exp->exp_last_request_time) {
621 struct obd_uuid *client_uuid;
623 class_export_get(exp);
624 client_uuid = &exp->exp_client_uuid;
625 spin_unlock(&obd->obd_dev_lock);
626 LCONSOLE_WARN("%s: haven't heard from client %s (at %s) in %lld seconds. I think it's dead, and I am evicting it. exp %p, cur %lld expire %lld last %lld\n",
628 obd_uuid2str(client_uuid),
629 obd_export_nid2str(exp),
630 ktime_get_real_seconds() -
631 exp->exp_last_request_time,
632 exp, ktime_get_real_seconds(),
634 exp->exp_last_request_time);
635 CDEBUG(D_HA, "Last request was at %lld\n",
636 exp->exp_last_request_time);
637 class_fail_export(exp);
638 class_export_put(exp);
639 spin_lock(&obd->obd_dev_lock);
641 /* List is sorted, so everyone below is ok */
645 spin_unlock(&obd->obd_dev_lock);
647 spin_lock(&pet_lock);
648 list_del_init(&obd->obd_evict_list);
649 spin_unlock(&pet_lock);
651 class_decref(obd, "evictor", obd);
653 CDEBUG(D_HA, "Exiting Ping Evictor\n");
658 void ping_evictor_start(void)
660 struct task_struct *task;
662 if (++pet_refcount > 1)
665 init_waitqueue_head(&pet_waitq);
667 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
670 CERROR("Cannot start ping evictor thread: %ld\n",
674 EXPORT_SYMBOL(ping_evictor_start);
676 void ping_evictor_stop(void)
678 if (--pet_refcount > 0)
681 pet_state = PET_TERMINATE;
684 EXPORT_SYMBOL(ping_evictor_stop);