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);
53 int ptlrpc_pinger_suppress_pings(void)
55 return suppress_pings;
57 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
59 struct ptlrpc_request *
60 ptlrpc_prep_ping(struct obd_import *imp)
62 struct ptlrpc_request *req;
64 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
65 LUSTRE_OBD_VERSION, OBD_PING);
67 ptlrpc_request_set_replen(req);
68 req->rq_no_resend = req->rq_no_delay = 1;
73 int ptlrpc_obd_ping(struct obd_device *obd)
76 struct ptlrpc_request *req;
80 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
84 req->rq_send_state = LUSTRE_IMP_FULL;
86 rc = ptlrpc_queue_wait(req);
88 ptlrpc_req_finished(req);
92 EXPORT_SYMBOL(ptlrpc_obd_ping);
94 static bool ptlrpc_check_import_is_idle(struct obd_import *imp)
96 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
99 if (!imp->imp_idle_timeout)
102 if (atomic_read(&imp->imp_reqs) > 0)
105 /* any lock increases ns_bref being a resource holder */
106 if (ns && atomic_read(&ns->ns_bref) > 0)
109 now = ktime_get_real_seconds();
110 if (now - imp->imp_last_reply_time < imp->imp_idle_timeout)
116 static int ptlrpc_ping(struct obd_import *imp)
118 struct ptlrpc_request *req;
122 if (ptlrpc_check_import_is_idle(imp))
123 RETURN(ptlrpc_disconnect_and_idle_import(imp));
125 req = ptlrpc_prep_ping(imp);
127 CERROR("OOM trying to ping %s->%s\n",
128 imp->imp_obd->obd_uuid.uuid,
129 obd2cli_tgt(imp->imp_obd));
133 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
134 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
135 ptlrpcd_add_req(req);
140 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
142 #ifdef CONFIG_LUSTRE_FS_PINGER
143 time64_t time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
145 if (imp->imp_state == LUSTRE_IMP_DISCON) {
146 time64_t dtime = max_t(time64_t, CONNECTION_SWITCH_MIN,
148 at_get(&imp->imp_at.iat_net_latency));
149 time = min(time, dtime);
151 imp->imp_next_ping = ktime_get_seconds() + time;
152 #endif /* CONFIG_LUSTRE_FS_PINGER */
155 void ptlrpc_ping_import_soon(struct obd_import *imp)
157 imp->imp_next_ping = ktime_get_seconds();
160 static inline int imp_is_deactive(struct obd_import *imp)
162 return imp->imp_deactive ||
163 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE);
166 static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp)
168 if (imp->imp_server_timeout)
169 return ktime_get_seconds() + (obd_timeout >> 1);
171 return ktime_get_seconds() + obd_timeout;
174 static time64_t pinger_check_timeout(time64_t time)
176 time64_t timeout = PING_INTERVAL;
178 return time + timeout - ktime_get_seconds();
183 void ptlrpc_pinger_ir_up(void)
185 CDEBUG(D_HA, "IR up\n");
188 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
190 void ptlrpc_pinger_ir_down(void)
192 CDEBUG(D_HA, "IR down\n");
195 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
197 static void ptlrpc_pinger_process_import(struct obd_import *imp,
205 spin_lock(&imp->imp_lock);
207 level = imp->imp_state;
208 force = imp->imp_force_verify;
209 force_next = imp->imp_force_next_verify;
211 * This will be used below only if the import is "FULL".
213 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
215 imp->imp_force_verify = 0;
217 if (imp->imp_next_ping - 5 >= this_ping && !force) {
218 spin_unlock(&imp->imp_lock);
222 imp->imp_force_next_verify = 0;
224 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
225 "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n",
226 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
227 ptlrpc_import_state_name(level), level, force, force_next,
228 imp->imp_deactive, imp->imp_pingable, suppress);
230 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
231 /* wait for a while before trying recovery again */
232 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
233 spin_unlock(&imp->imp_lock);
234 if (!imp->imp_no_pinger_recover ||
235 imp->imp_connect_error == -EAGAIN)
236 ptlrpc_initiate_recovery(imp);
237 } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov ||
238 imp_is_deactive(imp)) {
240 "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
241 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
242 ptlrpc_import_state_name(level));
244 imp->imp_force_verify = 1;
245 spin_unlock(&imp->imp_lock);
246 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
247 spin_unlock(&imp->imp_lock);
250 spin_unlock(&imp->imp_lock);
254 static struct workqueue_struct *pinger_wq;
255 static void ptlrpc_pinger_main(struct work_struct *ws);
256 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
258 static void ptlrpc_pinger_main(struct work_struct *ws)
260 time64_t this_ping, time_after_ping, time_to_next_wake;
261 struct obd_import *imp;
262 struct list_head *iter;
265 this_ping = ktime_get_seconds();
267 mutex_lock(&pinger_mutex);
269 list_for_each(iter, &pinger_imports) {
270 imp = list_entry(iter, struct obd_import,
273 ptlrpc_pinger_process_import(imp, this_ping);
274 /* obd_timeout might have changed */
275 if (imp->imp_pingable && imp->imp_next_ping &&
276 imp->imp_next_ping > this_ping + PING_INTERVAL)
277 ptlrpc_update_next_ping(imp, 0);
279 mutex_unlock(&pinger_mutex);
281 time_after_ping = ktime_get_seconds();
282 /* update memory usage info */
283 obd_update_maxusage();
285 if ((ktime_get_seconds() - this_ping - 3) > PING_INTERVAL)
286 CDEBUG(D_HA, "long time to ping: %lld, %lld, %lld\n",
287 this_ping, time_after_ping, ktime_get_seconds());
289 /* Wait until the next ping time, or until we're stopped. */
290 time_to_next_wake = pinger_check_timeout(this_ping);
292 * The ping sent by ptlrpc_send_rpc may get sent out
293 * say .01 second after this.
294 * ptlrpc_pinger_sending_on_import will then set the
295 * next ping time to next_ping + .01 sec, which means
296 * we will SKIP the next ping at next_ping, and the
297 * ping will get sent 2 timeouts from now! Beware.
299 CDEBUG(D_INFO, "next wakeup in %lld (%lld)\n",
300 time_to_next_wake, this_ping + PING_INTERVAL);
301 } while (time_to_next_wake <= 0);
303 queue_delayed_work(pinger_wq, &ping_work,
304 cfs_time_seconds(max(time_to_next_wake, 1LL)));
307 int ptlrpc_start_pinger(void)
309 #ifdef CONFIG_LUSTRE_FS_PINGER
313 pinger_wq = cfs_cpt_bind_workqueue("ptlrpc_pinger", cfs_cpt_tab,
315 if (IS_ERR(pinger_wq)) {
316 CERROR("cannot start pinger workqueue\n");
317 return PTR_ERR(pinger_wq);
320 queue_delayed_work(pinger_wq, &ping_work, 0);
323 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");
328 int ptlrpc_stop_pinger(void)
330 #ifdef CONFIG_LUSTRE_FS_PINGER
334 cancel_delayed_work_sync(&ping_work);
335 destroy_workqueue(pinger_wq);
341 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
343 ptlrpc_update_next_ping(imp, 0);
346 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
348 ptlrpc_update_next_ping(imp, 1);
349 assert_spin_locked(&imp->imp_lock);
351 * Avoid reading stale imp_connect_data. When not sure if pings are
352 * expected or not on next connection, we assume they are not and force
353 * one anyway to guarantee the chance of updating
354 * imp_peer_committed_transno.
356 if (imp->imp_state != LUSTRE_IMP_FULL ||
357 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
358 imp->imp_force_next_verify = 1;
361 int ptlrpc_pinger_add_import(struct obd_import *imp)
364 if (!list_empty(&imp->imp_pinger_chain))
367 mutex_lock(&pinger_mutex);
368 CDEBUG(D_HA, "adding pingable import %s->%s\n",
369 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
370 /* if we add to pinger we want recovery on this import */
371 imp->imp_obd->obd_no_recov = 0;
372 ptlrpc_update_next_ping(imp, 0);
373 /* XXX sort, blah blah */
374 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
375 class_import_get(imp);
377 ptlrpc_pinger_wake_up();
378 mutex_unlock(&pinger_mutex);
382 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
384 int ptlrpc_pinger_del_import(struct obd_import *imp)
388 if (list_empty(&imp->imp_pinger_chain))
391 mutex_lock(&pinger_mutex);
392 list_del_init(&imp->imp_pinger_chain);
393 CDEBUG(D_HA, "removing pingable import %s->%s\n",
394 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
395 /* if we remove from pinger we don't want recovery on this import */
396 imp->imp_obd->obd_no_recov = 1;
397 class_import_put(imp);
398 mutex_unlock(&pinger_mutex);
401 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
403 void ptlrpc_pinger_wake_up(void)
405 #ifdef CONFIG_LUSTRE_FS_PINGER
406 mod_delayed_work(pinger_wq, &ping_work, 0);
410 /* Ping evictor thread */
412 #define PET_TERMINATE 2
414 static int pet_refcount;
415 static int pet_state;
416 static wait_queue_head_t pet_waitq;
417 static LIST_HEAD(pet_list);
418 static DEFINE_SPINLOCK(pet_lock);
420 int ping_evictor_wake(struct obd_export *exp)
422 struct obd_device *obd;
424 spin_lock(&pet_lock);
425 if (pet_state != PET_READY) {
426 /* eventually the new obd will call here again. */
427 spin_unlock(&pet_lock);
431 obd = class_exp2obd(exp);
432 if (list_empty(&obd->obd_evict_list)) {
433 class_incref(obd, "evictor", obd);
434 list_add(&obd->obd_evict_list, &pet_list);
436 spin_unlock(&pet_lock);
442 static int ping_evictor_main(void *arg)
444 struct obd_device *obd;
445 struct obd_export *exp;
446 time64_t expire_time;
449 CDEBUG(D_HA, "Starting Ping Evictor\n");
450 pet_state = PET_READY;
452 wait_event_idle(pet_waitq,
453 (!list_empty(&pet_list)) ||
454 (pet_state == PET_TERMINATE));
456 /* loop until all obd's will be removed */
457 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
461 * we only get here if pet_exp != NULL, and the end of this
462 * loop is the only place which sets it NULL again, so lock
463 * is not strictly necessary.
465 spin_lock(&pet_lock);
466 obd = list_entry(pet_list.next, struct obd_device,
468 spin_unlock(&pet_lock);
470 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
472 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
473 obd->obd_name, expire_time);
476 * Exports can't be deleted out of the list while we hold
477 * the obd lock (class_unlink_export), which means we can't
478 * lose the last ref on the export. If they've already been
479 * removed from the list, we won't find them here.
481 spin_lock(&obd->obd_dev_lock);
482 while (!list_empty(&obd->obd_exports_timed)) {
483 exp = list_entry(obd->obd_exports_timed.next,
485 exp_obd_chain_timed);
486 if (expire_time > exp->exp_last_request_time) {
487 struct obd_uuid *client_uuid;
489 class_export_get(exp);
490 client_uuid = &exp->exp_client_uuid;
491 spin_unlock(&obd->obd_dev_lock);
492 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",
494 obd_uuid2str(client_uuid),
495 obd_export_nid2str(exp),
496 ktime_get_real_seconds() -
497 exp->exp_last_request_time,
498 exp, ktime_get_real_seconds(),
500 exp->exp_last_request_time);
501 CDEBUG(D_HA, "Last request was at %lld\n",
502 exp->exp_last_request_time);
503 class_fail_export(exp);
504 class_export_put(exp);
505 spin_lock(&obd->obd_dev_lock);
507 /* List is sorted, so everyone below is ok */
511 spin_unlock(&obd->obd_dev_lock);
513 spin_lock(&pet_lock);
514 list_del_init(&obd->obd_evict_list);
515 spin_unlock(&pet_lock);
517 class_decref(obd, "evictor", obd);
519 CDEBUG(D_HA, "Exiting Ping Evictor\n");
524 void ping_evictor_start(void)
526 struct task_struct *task;
528 if (++pet_refcount > 1)
531 init_waitqueue_head(&pet_waitq);
533 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
536 CERROR("Cannot start ping evictor thread: %ld\n",
540 EXPORT_SYMBOL(ping_evictor_start);
542 void ping_evictor_stop(void)
544 if (--pet_refcount > 0)
547 pet_state = PET_TERMINATE;
550 EXPORT_SYMBOL(ping_evictor_stop);