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)
106 * - client_obd_setup() - hashed import
107 * - ptlrpcd_alloc_work()
108 * - ptlrpcd_alloc_work()
109 * - ptlrpc_pinger_add_import
111 if (atomic_read(&imp->imp_refcount) > 4)
114 /* any lock increases ns_bref being a resource holder */
115 if (ns && atomic_read(&ns->ns_bref) > 0)
118 now = ktime_get_real_seconds();
119 if (now - imp->imp_last_reply_time < imp->imp_idle_timeout)
125 static int ptlrpc_ping(struct obd_import *imp)
127 struct ptlrpc_request *req;
131 if (ptlrpc_check_import_is_idle(imp))
132 RETURN(ptlrpc_disconnect_and_idle_import(imp));
134 req = ptlrpc_prep_ping(imp);
136 CERROR("OOM trying to ping %s->%s\n",
137 imp->imp_obd->obd_uuid.uuid,
138 obd2cli_tgt(imp->imp_obd));
142 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
143 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
144 ptlrpcd_add_req(req);
149 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
152 time64_t time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
154 if (imp->imp_state == LUSTRE_IMP_DISCON) {
155 time64_t dtime = max_t(time64_t, CONNECTION_SWITCH_MIN,
157 at_get(&imp->imp_at.iat_net_latency));
158 time = min(time, dtime);
160 imp->imp_next_ping = ktime_get_seconds() + time;
161 #endif /* ENABLE_PINGER */
164 void ptlrpc_ping_import_soon(struct obd_import *imp)
166 imp->imp_next_ping = ktime_get_seconds();
169 static inline int imp_is_deactive(struct obd_import *imp)
171 return imp->imp_deactive ||
172 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE);
175 static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp)
177 if (imp->imp_server_timeout)
178 return ktime_get_seconds() + (obd_timeout >> 1);
180 return ktime_get_seconds() + obd_timeout;
183 static time64_t pinger_check_timeout(time64_t time)
185 struct timeout_item *item;
186 time64_t timeout = PING_INTERVAL;
188 /* This list is sorted in increasing timeout order */
189 mutex_lock(&pinger_mutex);
190 list_for_each_entry(item, &timeout_list, ti_chain) {
191 time64_t ti_timeout = item->ti_timeout;
193 if (timeout > ti_timeout)
194 timeout = ti_timeout;
197 mutex_unlock(&pinger_mutex);
199 return time + timeout - ktime_get_seconds();
204 void ptlrpc_pinger_ir_up(void)
206 CDEBUG(D_HA, "IR up\n");
209 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
211 void ptlrpc_pinger_ir_down(void)
213 CDEBUG(D_HA, "IR down\n");
216 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
218 static void ptlrpc_pinger_process_import(struct obd_import *imp,
226 spin_lock(&imp->imp_lock);
228 level = imp->imp_state;
229 force = imp->imp_force_verify;
230 force_next = imp->imp_force_next_verify;
232 * This will be used below only if the import is "FULL".
234 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
236 imp->imp_force_verify = 0;
238 if (imp->imp_next_ping - 5 >= this_ping && !force) {
239 spin_unlock(&imp->imp_lock);
243 imp->imp_force_next_verify = 0;
245 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
246 "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n",
247 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
248 ptlrpc_import_state_name(level), level, force, force_next,
249 imp->imp_deactive, imp->imp_pingable, suppress);
251 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
252 /* wait for a while before trying recovery again */
253 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
254 spin_unlock(&imp->imp_lock);
255 if (!imp->imp_no_pinger_recover ||
256 imp->imp_connect_error == -EAGAIN)
257 ptlrpc_initiate_recovery(imp);
258 } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov ||
259 imp_is_deactive(imp)) {
261 "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
262 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
263 ptlrpc_import_state_name(level));
265 imp->imp_force_verify = 1;
266 spin_unlock(&imp->imp_lock);
267 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
268 spin_unlock(&imp->imp_lock);
271 spin_unlock(&imp->imp_lock);
275 static struct workqueue_struct *pinger_wq;
276 static void ptlrpc_pinger_main(struct work_struct *ws);
277 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
279 static void ptlrpc_pinger_main(struct work_struct *ws)
281 time64_t this_ping = ktime_get_seconds();
282 time64_t time_to_next_wake;
283 struct timeout_item *item;
284 struct obd_import *imp;
285 struct list_head *iter;
288 mutex_lock(&pinger_mutex);
289 list_for_each_entry(item, &timeout_list, ti_chain)
290 item->ti_cb(item, item->ti_cb_data);
292 list_for_each(iter, &pinger_imports) {
293 imp = list_entry(iter, struct obd_import,
296 ptlrpc_pinger_process_import(imp, this_ping);
297 /* obd_timeout might have changed */
298 if (imp->imp_pingable && imp->imp_next_ping &&
299 imp->imp_next_ping > this_ping + PING_INTERVAL)
300 ptlrpc_update_next_ping(imp, 0);
302 mutex_unlock(&pinger_mutex);
303 /* update memory usage info */
304 obd_update_maxusage();
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)
330 pinger_wq = alloc_workqueue("ptlrpc_pinger", 0, 1);
332 CERROR("cannot start pinger workqueue\n");
336 queue_delayed_work(pinger_wq, &ping_work, 0);
339 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");
344 int ptlrpc_pinger_remove_timeouts(void);
346 int ptlrpc_stop_pinger(void)
352 ptlrpc_pinger_remove_timeouts();
354 cancel_delayed_work_sync(&ping_work);
355 destroy_workqueue(pinger_wq);
361 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
363 ptlrpc_update_next_ping(imp, 0);
366 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
368 ptlrpc_update_next_ping(imp, 1);
369 assert_spin_locked(&imp->imp_lock);
371 * Avoid reading stale imp_connect_data. When not sure if pings are
372 * expected or not on next connection, we assume they are not and force
373 * one anyway to guarantee the chance of updating
374 * imp_peer_committed_transno.
376 if (imp->imp_state != LUSTRE_IMP_FULL ||
377 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
378 imp->imp_force_next_verify = 1;
381 int ptlrpc_pinger_add_import(struct obd_import *imp)
384 if (!list_empty(&imp->imp_pinger_chain))
387 mutex_lock(&pinger_mutex);
388 CDEBUG(D_HA, "adding pingable import %s->%s\n",
389 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
390 /* if we add to pinger we want recovery on this import */
391 imp->imp_obd->obd_no_recov = 0;
392 ptlrpc_update_next_ping(imp, 0);
393 /* XXX sort, blah blah */
394 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
395 class_import_get(imp);
397 ptlrpc_pinger_wake_up();
398 mutex_unlock(&pinger_mutex);
402 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
404 int ptlrpc_pinger_del_import(struct obd_import *imp)
408 if (list_empty(&imp->imp_pinger_chain))
411 mutex_lock(&pinger_mutex);
412 list_del_init(&imp->imp_pinger_chain);
413 CDEBUG(D_HA, "removing pingable import %s->%s\n",
414 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
415 /* if we remove from pinger we don't want recovery on this import */
416 imp->imp_obd->obd_no_recov = 1;
417 class_import_put(imp);
418 mutex_unlock(&pinger_mutex);
421 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
424 * Register a timeout callback to the pinger list, and the callback will
425 * be called when timeout happens.
427 static struct timeout_item *ptlrpc_new_timeout(time64_t time,
428 enum timeout_event event,
429 timeout_cb_t cb, void *data)
431 struct timeout_item *ti;
437 INIT_LIST_HEAD(&ti->ti_obd_list);
438 INIT_LIST_HEAD(&ti->ti_chain);
439 ti->ti_timeout = time;
440 ti->ti_event = event;
442 ti->ti_cb_data = data;
448 * Register timeout event on the the pinger thread.
449 * Note: the timeout list is an sorted list with increased timeout value.
451 static struct timeout_item*
452 ptlrpc_pinger_register_timeout(time64_t time, enum timeout_event event,
453 timeout_cb_t cb, void *data)
455 struct timeout_item *item, *tmp;
457 LASSERT(mutex_is_locked(&pinger_mutex));
459 list_for_each_entry(item, &timeout_list, ti_chain)
460 if (item->ti_event == event)
463 item = ptlrpc_new_timeout(time, event, cb, data);
465 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
466 if (tmp->ti_timeout < time) {
467 list_add(&item->ti_chain, &tmp->ti_chain);
471 list_add(&item->ti_chain, &timeout_list);
477 /* Add a client_obd to the timeout event list, when timeout(@time)
478 * happens, the callback(@cb) will be called.
480 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
481 timeout_cb_t cb, void *data,
482 struct list_head *obd_list)
484 struct timeout_item *ti;
486 mutex_lock(&pinger_mutex);
487 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
489 mutex_unlock(&pinger_mutex);
492 list_add(obd_list, &ti->ti_obd_list);
493 mutex_unlock(&pinger_mutex);
496 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
498 int ptlrpc_del_timeout_client(struct list_head *obd_list,
499 enum timeout_event event)
501 struct timeout_item *ti = NULL, *item;
503 if (list_empty(obd_list))
505 mutex_lock(&pinger_mutex);
506 list_del_init(obd_list);
508 * If there are no obd attached to the timeout event
509 * list, remove this timeout event from the pinger
511 list_for_each_entry(item, &timeout_list, ti_chain) {
512 if (item->ti_event == event) {
517 LASSERTF(ti, "ti is NULL !\n");
518 if (list_empty(&ti->ti_obd_list)) {
519 list_del(&ti->ti_chain);
522 mutex_unlock(&pinger_mutex);
525 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
527 int ptlrpc_pinger_remove_timeouts(void)
529 struct timeout_item *item, *tmp;
531 mutex_lock(&pinger_mutex);
532 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
533 LASSERT(list_empty(&item->ti_obd_list));
534 list_del(&item->ti_chain);
537 mutex_unlock(&pinger_mutex);
541 void ptlrpc_pinger_wake_up(void)
544 mod_delayed_work(pinger_wq, &ping_work, 0);
548 /* Ping evictor thread */
550 #define PET_TERMINATE 2
552 static int pet_refcount;
553 static int pet_state;
554 static wait_queue_head_t pet_waitq;
555 static struct list_head pet_list;
556 static DEFINE_SPINLOCK(pet_lock);
558 int ping_evictor_wake(struct obd_export *exp)
560 struct obd_device *obd;
562 spin_lock(&pet_lock);
563 if (pet_state != PET_READY) {
564 /* eventually the new obd will call here again. */
565 spin_unlock(&pet_lock);
569 obd = class_exp2obd(exp);
570 if (list_empty(&obd->obd_evict_list)) {
571 class_incref(obd, "evictor", obd);
572 list_add(&obd->obd_evict_list, &pet_list);
574 spin_unlock(&pet_lock);
580 static int ping_evictor_main(void *arg)
582 struct obd_device *obd;
583 struct obd_export *exp;
584 time64_t expire_time;
589 CDEBUG(D_HA, "Starting Ping Evictor\n");
590 pet_state = PET_READY;
592 wait_event_idle(pet_waitq,
593 (!list_empty(&pet_list)) ||
594 (pet_state == PET_TERMINATE));
596 /* loop until all obd's will be removed */
597 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
601 * we only get here if pet_exp != NULL, and the end of this
602 * loop is the only place which sets it NULL again, so lock
603 * is not strictly necessary.
605 spin_lock(&pet_lock);
606 obd = list_entry(pet_list.next, struct obd_device,
608 spin_unlock(&pet_lock);
610 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
612 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
613 obd->obd_name, expire_time);
616 * Exports can't be deleted out of the list while we hold
617 * the obd lock (class_unlink_export), which means we can't
618 * lose the last ref on the export. If they've already been
619 * removed from the list, we won't find them here.
621 spin_lock(&obd->obd_dev_lock);
622 while (!list_empty(&obd->obd_exports_timed)) {
623 exp = list_entry(obd->obd_exports_timed.next,
625 exp_obd_chain_timed);
626 if (expire_time > exp->exp_last_request_time) {
627 struct obd_uuid *client_uuid;
629 class_export_get(exp);
630 client_uuid = &exp->exp_client_uuid;
631 spin_unlock(&obd->obd_dev_lock);
632 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",
634 obd_uuid2str(client_uuid),
635 obd_export_nid2str(exp),
636 ktime_get_real_seconds() -
637 exp->exp_last_request_time,
638 exp, ktime_get_real_seconds(),
640 exp->exp_last_request_time);
641 CDEBUG(D_HA, "Last request was at %lld\n",
642 exp->exp_last_request_time);
643 class_fail_export(exp);
644 class_export_put(exp);
645 spin_lock(&obd->obd_dev_lock);
647 /* List is sorted, so everyone below is ok */
651 spin_unlock(&obd->obd_dev_lock);
653 spin_lock(&pet_lock);
654 list_del_init(&obd->obd_evict_list);
655 spin_unlock(&pet_lock);
657 class_decref(obd, "evictor", obd);
659 CDEBUG(D_HA, "Exiting Ping Evictor\n");
664 void ping_evictor_start(void)
666 struct task_struct *task;
668 if (++pet_refcount > 1)
671 INIT_LIST_HEAD(&pet_list);
672 init_waitqueue_head(&pet_waitq);
674 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
677 CERROR("Cannot start ping evictor thread: %ld\n",
681 EXPORT_SYMBOL(ping_evictor_start);
683 void ping_evictor_stop(void)
685 if (--pet_refcount > 0)
688 pet_state = PET_TERMINATE;
691 EXPORT_SYMBOL(ping_evictor_stop);