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, time_after_ping, time_to_next_wake;
275 struct timeout_item *item;
276 struct obd_import *imp;
277 struct list_head *iter;
280 this_ping = ktime_get_seconds();
282 mutex_lock(&pinger_mutex);
283 list_for_each_entry(item, &timeout_list, ti_chain)
284 item->ti_cb(item, item->ti_cb_data);
286 list_for_each(iter, &pinger_imports) {
287 imp = list_entry(iter, struct obd_import,
290 ptlrpc_pinger_process_import(imp, this_ping);
291 /* obd_timeout might have changed */
292 if (imp->imp_pingable && imp->imp_next_ping &&
293 imp->imp_next_ping > this_ping + PING_INTERVAL)
294 ptlrpc_update_next_ping(imp, 0);
296 mutex_unlock(&pinger_mutex);
298 time_after_ping = ktime_get_seconds();
299 /* update memory usage info */
300 obd_update_maxusage();
302 if ((ktime_get_seconds() - this_ping - 3) > PING_INTERVAL)
303 CDEBUG(D_HA, "long time to ping: %lld, %lld, %lld\n",
304 this_ping, time_after_ping, ktime_get_seconds());
306 /* Wait until the next ping time, or until we're stopped. */
307 time_to_next_wake = pinger_check_timeout(this_ping);
309 * The ping sent by ptlrpc_send_rpc may get sent out
310 * say .01 second after this.
311 * ptlrpc_pinger_sending_on_import will then set the
312 * next ping time to next_ping + .01 sec, which means
313 * we will SKIP the next ping at next_ping, and the
314 * ping will get sent 2 timeouts from now! Beware.
316 CDEBUG(D_INFO, "next wakeup in %lld (%lld)\n",
317 time_to_next_wake, this_ping + PING_INTERVAL);
318 } while (time_to_next_wake <= 0);
320 queue_delayed_work(pinger_wq, &ping_work,
321 cfs_time_seconds(max(time_to_next_wake, 1LL)));
324 int ptlrpc_start_pinger(void)
326 #ifdef CONFIG_LUSTRE_FS_PINGER
330 pinger_wq = cfs_cpt_bind_workqueue("ptlrpc_pinger", cfs_cpt_tab,
332 if (IS_ERR(pinger_wq)) {
333 CERROR("cannot start pinger workqueue\n");
334 return PTR_ERR(pinger_wq);
337 queue_delayed_work(pinger_wq, &ping_work, 0);
340 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");
345 int ptlrpc_pinger_remove_timeouts(void);
347 int ptlrpc_stop_pinger(void)
349 #ifdef CONFIG_LUSTRE_FS_PINGER
353 ptlrpc_pinger_remove_timeouts();
355 cancel_delayed_work_sync(&ping_work);
356 destroy_workqueue(pinger_wq);
362 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
364 ptlrpc_update_next_ping(imp, 0);
367 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
369 ptlrpc_update_next_ping(imp, 1);
370 assert_spin_locked(&imp->imp_lock);
372 * Avoid reading stale imp_connect_data. When not sure if pings are
373 * expected or not on next connection, we assume they are not and force
374 * one anyway to guarantee the chance of updating
375 * imp_peer_committed_transno.
377 if (imp->imp_state != LUSTRE_IMP_FULL ||
378 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
379 imp->imp_force_next_verify = 1;
382 int ptlrpc_pinger_add_import(struct obd_import *imp)
385 if (!list_empty(&imp->imp_pinger_chain))
388 mutex_lock(&pinger_mutex);
389 CDEBUG(D_HA, "adding pingable import %s->%s\n",
390 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
391 /* if we add to pinger we want recovery on this import */
392 imp->imp_obd->obd_no_recov = 0;
393 ptlrpc_update_next_ping(imp, 0);
394 /* XXX sort, blah blah */
395 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
396 class_import_get(imp);
398 ptlrpc_pinger_wake_up();
399 mutex_unlock(&pinger_mutex);
403 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
405 int ptlrpc_pinger_del_import(struct obd_import *imp)
409 if (list_empty(&imp->imp_pinger_chain))
412 mutex_lock(&pinger_mutex);
413 list_del_init(&imp->imp_pinger_chain);
414 CDEBUG(D_HA, "removing pingable import %s->%s\n",
415 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
416 /* if we remove from pinger we don't want recovery on this import */
417 imp->imp_obd->obd_no_recov = 1;
418 class_import_put(imp);
419 mutex_unlock(&pinger_mutex);
422 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
425 * Register a timeout callback to the pinger list, and the callback will
426 * be called when timeout happens.
428 static struct timeout_item *ptlrpc_new_timeout(time64_t time,
429 enum timeout_event event,
430 timeout_cb_t cb, void *data)
432 struct timeout_item *ti;
438 INIT_LIST_HEAD(&ti->ti_obd_list);
439 INIT_LIST_HEAD(&ti->ti_chain);
440 ti->ti_timeout = time;
441 ti->ti_event = event;
443 ti->ti_cb_data = data;
449 * Register timeout event on the the pinger thread.
450 * Note: the timeout list is an sorted list with increased timeout value.
452 static struct timeout_item*
453 ptlrpc_pinger_register_timeout(time64_t time, enum timeout_event event,
454 timeout_cb_t cb, void *data)
456 struct timeout_item *item, *tmp;
458 LASSERT(mutex_is_locked(&pinger_mutex));
460 list_for_each_entry(item, &timeout_list, ti_chain)
461 if (item->ti_event == event)
464 item = ptlrpc_new_timeout(time, event, cb, data);
466 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
467 if (tmp->ti_timeout < time) {
468 list_add(&item->ti_chain, &tmp->ti_chain);
472 list_add(&item->ti_chain, &timeout_list);
478 /* Add a client_obd to the timeout event list, when timeout(@time)
479 * happens, the callback(@cb) will be called.
481 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
482 timeout_cb_t cb, void *data,
483 struct list_head *obd_list)
485 struct timeout_item *ti;
487 mutex_lock(&pinger_mutex);
488 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
490 mutex_unlock(&pinger_mutex);
493 list_add(obd_list, &ti->ti_obd_list);
494 mutex_unlock(&pinger_mutex);
497 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
499 int ptlrpc_del_timeout_client(struct list_head *obd_list,
500 enum timeout_event event)
502 struct timeout_item *ti = NULL, *item;
504 if (list_empty(obd_list))
506 mutex_lock(&pinger_mutex);
507 list_del_init(obd_list);
509 * If there are no obd attached to the timeout event
510 * list, remove this timeout event from the pinger
512 list_for_each_entry(item, &timeout_list, ti_chain) {
513 if (item->ti_event == event) {
518 LASSERTF(ti, "ti is NULL !\n");
519 if (list_empty(&ti->ti_obd_list)) {
520 list_del(&ti->ti_chain);
523 mutex_unlock(&pinger_mutex);
526 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
528 int ptlrpc_pinger_remove_timeouts(void)
530 struct timeout_item *item, *tmp;
532 mutex_lock(&pinger_mutex);
533 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
534 LASSERT(list_empty(&item->ti_obd_list));
535 list_del(&item->ti_chain);
538 mutex_unlock(&pinger_mutex);
542 void ptlrpc_pinger_wake_up(void)
544 #ifdef CONFIG_LUSTRE_FS_PINGER
545 mod_delayed_work(pinger_wq, &ping_work, 0);
549 /* Ping evictor thread */
551 #define PET_TERMINATE 2
553 static int pet_refcount;
554 static int pet_state;
555 static wait_queue_head_t pet_waitq;
556 static LIST_HEAD(pet_list);
557 static DEFINE_SPINLOCK(pet_lock);
559 int ping_evictor_wake(struct obd_export *exp)
561 struct obd_device *obd;
563 spin_lock(&pet_lock);
564 if (pet_state != PET_READY) {
565 /* eventually the new obd will call here again. */
566 spin_unlock(&pet_lock);
570 obd = class_exp2obd(exp);
571 if (list_empty(&obd->obd_evict_list)) {
572 class_incref(obd, "evictor", obd);
573 list_add(&obd->obd_evict_list, &pet_list);
575 spin_unlock(&pet_lock);
581 static int ping_evictor_main(void *arg)
583 struct obd_device *obd;
584 struct obd_export *exp;
585 time64_t expire_time;
588 CDEBUG(D_HA, "Starting Ping Evictor\n");
589 pet_state = PET_READY;
591 wait_event_idle(pet_waitq,
592 (!list_empty(&pet_list)) ||
593 (pet_state == PET_TERMINATE));
595 /* loop until all obd's will be removed */
596 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
600 * we only get here if pet_exp != NULL, and the end of this
601 * loop is the only place which sets it NULL again, so lock
602 * is not strictly necessary.
604 spin_lock(&pet_lock);
605 obd = list_entry(pet_list.next, struct obd_device,
607 spin_unlock(&pet_lock);
609 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
611 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
612 obd->obd_name, expire_time);
615 * Exports can't be deleted out of the list while we hold
616 * the obd lock (class_unlink_export), which means we can't
617 * lose the last ref on the export. If they've already been
618 * removed from the list, we won't find them here.
620 spin_lock(&obd->obd_dev_lock);
621 while (!list_empty(&obd->obd_exports_timed)) {
622 exp = list_entry(obd->obd_exports_timed.next,
624 exp_obd_chain_timed);
625 if (expire_time > exp->exp_last_request_time) {
626 struct obd_uuid *client_uuid;
628 class_export_get(exp);
629 client_uuid = &exp->exp_client_uuid;
630 spin_unlock(&obd->obd_dev_lock);
631 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",
633 obd_uuid2str(client_uuid),
634 obd_export_nid2str(exp),
635 ktime_get_real_seconds() -
636 exp->exp_last_request_time,
637 exp, ktime_get_real_seconds(),
639 exp->exp_last_request_time);
640 CDEBUG(D_HA, "Last request was at %lld\n",
641 exp->exp_last_request_time);
642 class_fail_export(exp);
643 class_export_put(exp);
644 spin_lock(&obd->obd_dev_lock);
646 /* List is sorted, so everyone below is ok */
650 spin_unlock(&obd->obd_dev_lock);
652 spin_lock(&pet_lock);
653 list_del_init(&obd->obd_evict_list);
654 spin_unlock(&pet_lock);
656 class_decref(obd, "evictor", obd);
658 CDEBUG(D_HA, "Exiting Ping Evictor\n");
663 void ping_evictor_start(void)
665 struct task_struct *task;
667 if (++pet_refcount > 1)
670 init_waitqueue_head(&pet_waitq);
672 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
675 CERROR("Cannot start ping evictor thread: %ld\n",
679 EXPORT_SYMBOL(ping_evictor_start);
681 void ping_evictor_stop(void)
683 if (--pet_refcount > 0)
686 pet_state = PET_TERMINATE;
689 EXPORT_SYMBOL(ping_evictor_stop);