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;
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);
273 static struct workqueue_struct *pinger_wq;
274 static void ptlrpc_pinger_main(struct work_struct *ws);
275 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
277 static void ptlrpc_pinger_main(struct work_struct *ws)
279 time64_t this_ping = ktime_get_seconds();
280 time64_t time_to_next_wake;
281 struct timeout_item *item;
282 struct obd_import *imp;
283 struct list_head *iter;
286 mutex_lock(&pinger_mutex);
287 list_for_each_entry(item, &timeout_list, ti_chain)
288 item->ti_cb(item, item->ti_cb_data);
290 list_for_each(iter, &pinger_imports) {
291 imp = list_entry(iter, struct obd_import,
294 ptlrpc_pinger_process_import(imp, this_ping);
295 /* obd_timeout might have changed */
296 if (imp->imp_pingable && imp->imp_next_ping &&
297 imp->imp_next_ping > this_ping + PING_INTERVAL)
298 ptlrpc_update_next_ping(imp, 0);
300 mutex_unlock(&pinger_mutex);
301 /* update memory usage info */
302 obd_update_maxusage();
304 /* Wait until the next ping time, or until we're stopped. */
305 time_to_next_wake = pinger_check_timeout(this_ping);
307 * The ping sent by ptlrpc_send_rpc may get sent out
308 * say .01 second after this.
309 * ptlrpc_pinger_sending_on_import will then set the
310 * next ping time to next_ping + .01 sec, which means
311 * we will SKIP the next ping at next_ping, and the
312 * ping will get sent 2 timeouts from now! Beware.
314 CDEBUG(D_INFO, "next wakeup in %lld (%lld)\n",
315 time_to_next_wake, this_ping + PING_INTERVAL);
316 } while (time_to_next_wake <= 0);
318 queue_delayed_work(pinger_wq, &ping_work,
319 cfs_time_seconds(max(time_to_next_wake, 1LL)));
322 int ptlrpc_start_pinger(void)
328 pinger_wq = alloc_workqueue("ptlrpc_pinger", 0, 1);
330 CERROR("cannot start pinger workqueue\n");
334 queue_delayed_work(pinger_wq, &ping_work, 0);
337 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");
342 int ptlrpc_pinger_remove_timeouts(void);
344 int ptlrpc_stop_pinger(void)
350 ptlrpc_pinger_remove_timeouts();
352 cancel_delayed_work_sync(&ping_work);
353 destroy_workqueue(pinger_wq);
359 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
361 ptlrpc_update_next_ping(imp, 0);
364 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
366 ptlrpc_update_next_ping(imp, 1);
367 assert_spin_locked(&imp->imp_lock);
369 * Avoid reading stale imp_connect_data. When not sure if pings are
370 * expected or not on next connection, we assume they are not and force
371 * one anyway to guarantee the chance of updating
372 * imp_peer_committed_transno.
374 if (imp->imp_state != LUSTRE_IMP_FULL ||
375 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
376 imp->imp_force_next_verify = 1;
379 int ptlrpc_pinger_add_import(struct obd_import *imp)
382 if (!list_empty(&imp->imp_pinger_chain))
385 mutex_lock(&pinger_mutex);
386 CDEBUG(D_HA, "adding pingable import %s->%s\n",
387 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
388 /* if we add to pinger we want recovery on this import */
389 imp->imp_obd->obd_no_recov = 0;
390 ptlrpc_update_next_ping(imp, 0);
391 /* XXX sort, blah blah */
392 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
393 class_import_get(imp);
395 ptlrpc_pinger_wake_up();
396 mutex_unlock(&pinger_mutex);
400 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
402 int ptlrpc_pinger_del_import(struct obd_import *imp)
406 if (list_empty(&imp->imp_pinger_chain))
409 mutex_lock(&pinger_mutex);
410 list_del_init(&imp->imp_pinger_chain);
411 CDEBUG(D_HA, "removing pingable import %s->%s\n",
412 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
413 /* if we remove from pinger we don't want recovery on this import */
414 imp->imp_obd->obd_no_recov = 1;
415 class_import_put(imp);
416 mutex_unlock(&pinger_mutex);
419 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
422 * Register a timeout callback to the pinger list, and the callback will
423 * be called when timeout happens.
425 static struct timeout_item *ptlrpc_new_timeout(time64_t time,
426 enum timeout_event event,
427 timeout_cb_t cb, void *data)
429 struct timeout_item *ti;
435 INIT_LIST_HEAD(&ti->ti_obd_list);
436 INIT_LIST_HEAD(&ti->ti_chain);
437 ti->ti_timeout = time;
438 ti->ti_event = event;
440 ti->ti_cb_data = data;
446 * Register timeout event on the the pinger thread.
447 * Note: the timeout list is an sorted list with increased timeout value.
449 static struct timeout_item*
450 ptlrpc_pinger_register_timeout(time64_t time, enum timeout_event event,
451 timeout_cb_t cb, void *data)
453 struct timeout_item *item, *tmp;
455 LASSERT(mutex_is_locked(&pinger_mutex));
457 list_for_each_entry(item, &timeout_list, ti_chain)
458 if (item->ti_event == event)
461 item = ptlrpc_new_timeout(time, event, cb, data);
463 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
464 if (tmp->ti_timeout < time) {
465 list_add(&item->ti_chain, &tmp->ti_chain);
469 list_add(&item->ti_chain, &timeout_list);
475 /* Add a client_obd to the timeout event list, when timeout(@time)
476 * happens, the callback(@cb) will be called.
478 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
479 timeout_cb_t cb, void *data,
480 struct list_head *obd_list)
482 struct timeout_item *ti;
484 mutex_lock(&pinger_mutex);
485 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
487 mutex_unlock(&pinger_mutex);
490 list_add(obd_list, &ti->ti_obd_list);
491 mutex_unlock(&pinger_mutex);
494 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
496 int ptlrpc_del_timeout_client(struct list_head *obd_list,
497 enum timeout_event event)
499 struct timeout_item *ti = NULL, *item;
501 if (list_empty(obd_list))
503 mutex_lock(&pinger_mutex);
504 list_del_init(obd_list);
506 * If there are no obd attached to the timeout event
507 * list, remove this timeout event from the pinger
509 list_for_each_entry(item, &timeout_list, ti_chain) {
510 if (item->ti_event == event) {
515 LASSERTF(ti, "ti is NULL !\n");
516 if (list_empty(&ti->ti_obd_list)) {
517 list_del(&ti->ti_chain);
520 mutex_unlock(&pinger_mutex);
523 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
525 int ptlrpc_pinger_remove_timeouts(void)
527 struct timeout_item *item, *tmp;
529 mutex_lock(&pinger_mutex);
530 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
531 LASSERT(list_empty(&item->ti_obd_list));
532 list_del(&item->ti_chain);
535 mutex_unlock(&pinger_mutex);
539 void ptlrpc_pinger_wake_up()
542 mod_delayed_work(pinger_wq, &ping_work, 0);
546 /* Ping evictor thread */
548 #define PET_TERMINATE 2
550 static int pet_refcount;
551 static int pet_state;
552 static wait_queue_head_t pet_waitq;
553 static struct list_head pet_list;
554 static DEFINE_SPINLOCK(pet_lock);
556 int ping_evictor_wake(struct obd_export *exp)
558 struct obd_device *obd;
560 spin_lock(&pet_lock);
561 if (pet_state != PET_READY) {
562 /* eventually the new obd will call here again. */
563 spin_unlock(&pet_lock);
567 obd = class_exp2obd(exp);
568 if (list_empty(&obd->obd_evict_list)) {
569 class_incref(obd, "evictor", obd);
570 list_add(&obd->obd_evict_list, &pet_list);
572 spin_unlock(&pet_lock);
578 static int ping_evictor_main(void *arg)
580 struct obd_device *obd;
581 struct obd_export *exp;
582 struct l_wait_info lwi = { 0 };
583 time64_t expire_time;
588 CDEBUG(D_HA, "Starting Ping Evictor\n");
589 pet_state = PET_READY;
591 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
592 (pet_state == PET_TERMINATE), &lwi);
594 /* loop until all obd's will be removed */
595 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.
603 spin_lock(&pet_lock);
604 obd = list_entry(pet_list.next, struct obd_device,
606 spin_unlock(&pet_lock);
608 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
610 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
611 obd->obd_name, expire_time);
614 * Exports can't be deleted out of the list while we hold
615 * the obd lock (class_unlink_export), which means we can't
616 * lose the last ref on the export. If they've already been
617 * removed from the list, we won't find them here.
619 spin_lock(&obd->obd_dev_lock);
620 while (!list_empty(&obd->obd_exports_timed)) {
621 exp = list_entry(obd->obd_exports_timed.next,
623 exp_obd_chain_timed);
624 if (expire_time > exp->exp_last_request_time) {
625 struct obd_uuid *client_uuid;
627 class_export_get(exp);
628 client_uuid = &exp->exp_client_uuid;
629 spin_unlock(&obd->obd_dev_lock);
630 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",
632 obd_uuid2str(client_uuid),
633 obd_export_nid2str(exp),
634 ktime_get_real_seconds() -
635 exp->exp_last_request_time,
636 exp, ktime_get_real_seconds(),
638 exp->exp_last_request_time);
639 CDEBUG(D_HA, "Last request was at %lld\n",
640 exp->exp_last_request_time);
641 class_fail_export(exp);
642 class_export_put(exp);
643 spin_lock(&obd->obd_dev_lock);
645 /* List is sorted, so everyone below is ok */
649 spin_unlock(&obd->obd_dev_lock);
651 spin_lock(&pet_lock);
652 list_del_init(&obd->obd_evict_list);
653 spin_unlock(&pet_lock);
655 class_decref(obd, "evictor", obd);
657 CDEBUG(D_HA, "Exiting Ping Evictor\n");
662 void ping_evictor_start(void)
664 struct task_struct *task;
666 if (++pet_refcount > 1)
669 INIT_LIST_HEAD(&pet_list);
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);