Whamcloud - gitweb
LU-3032 ptlrpc: race in pinger (use-after-free situation)
[fs/lustre-release.git] / lustre / ptlrpc / pinger.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2013, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/ptlrpc/pinger.c
37  *
38  * Portal-RPC reconnection and replay operations, for use in recovery.
39  */
40
41 #ifndef __KERNEL__
42 #include <liblustre.h>
43 #else
44 #define DEBUG_SUBSYSTEM S_RPC
45 #endif
46
47 #include <obd_support.h>
48 #include <obd_class.h>
49 #include "ptlrpc_internal.h"
50
51 static int suppress_pings;
52 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
53
54 struct mutex pinger_mutex;
55 static CFS_LIST_HEAD(pinger_imports);
56 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
57
58 int ptlrpc_pinger_suppress_pings()
59 {
60         return suppress_pings;
61 }
62 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
63
64 struct ptlrpc_request *
65 ptlrpc_prep_ping(struct obd_import *imp)
66 {
67         struct ptlrpc_request *req;
68
69         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
70                                         LUSTRE_OBD_VERSION, OBD_PING);
71         if (req) {
72                 ptlrpc_request_set_replen(req);
73                 req->rq_no_resend = req->rq_no_delay = 1;
74         }
75         return req;
76 }
77
78 int ptlrpc_obd_ping(struct obd_device *obd)
79 {
80         int rc;
81         struct ptlrpc_request *req;
82         ENTRY;
83
84         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
85         if (req == NULL)
86                 RETURN(-ENOMEM);
87
88         req->rq_send_state = LUSTRE_IMP_FULL;
89
90         rc = ptlrpc_queue_wait(req);
91
92         ptlrpc_req_finished(req);
93
94         RETURN(rc);
95 }
96 EXPORT_SYMBOL(ptlrpc_obd_ping);
97
98 int ptlrpc_ping(struct obd_import *imp)
99 {
100         struct ptlrpc_request *req;
101         ENTRY;
102
103         req = ptlrpc_prep_ping(imp);
104         if (req == NULL) {
105                 CERROR("OOM trying to ping %s->%s\n",
106                        imp->imp_obd->obd_uuid.uuid,
107                        obd2cli_tgt(imp->imp_obd));
108                 RETURN(-ENOMEM);
109         }
110
111         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
112                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
113         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
114
115         RETURN(0);
116 }
117
118 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
119 {
120 #ifdef ENABLE_PINGER
121         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
122         if (imp->imp_state == LUSTRE_IMP_DISCON) {
123                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
124                                   AT_OFF ? 0 :
125                                   at_get(&imp->imp_at.iat_net_latency));
126                 time = min(time, dtime);
127         }
128         imp->imp_next_ping = cfs_time_shift(time);
129 #endif /* ENABLE_PINGER */
130 }
131
132 void ptlrpc_ping_import_soon(struct obd_import *imp)
133 {
134         imp->imp_next_ping = cfs_time_current();
135 }
136
137 static inline int imp_is_deactive(struct obd_import *imp)
138 {
139         return (imp->imp_deactive ||
140                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
141 }
142
143 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
144 {
145         if (imp->imp_server_timeout)
146                 return cfs_time_shift(obd_timeout / 2);
147         else
148                 return cfs_time_shift(obd_timeout);
149 }
150
151 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
152 static cfs_time_t suspend_wakeup_time = 0;
153
154 cfs_duration_t pinger_check_timeout(cfs_time_t time)
155 {
156         struct timeout_item *item;
157         cfs_time_t timeout = PING_INTERVAL;
158
159         /* The timeout list is a increase order sorted list */
160         mutex_lock(&pinger_mutex);
161         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
162                 int ti_timeout = item->ti_timeout;
163                 if (timeout > ti_timeout)
164                         timeout = ti_timeout;
165                 break;
166         }
167         mutex_unlock(&pinger_mutex);
168
169         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
170                                          cfs_time_current());
171 }
172
173 static cfs_waitq_t suspend_timeouts_waitq;
174
175 cfs_time_t ptlrpc_suspend_wakeup_time(void)
176 {
177         return suspend_wakeup_time;
178 }
179
180 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
181 {
182         /*XXX: disabled for now, will be replaced by adaptive timeouts */
183 #if 0
184         if (imp->imp_no_timeout)
185                 return;
186         imp->imp_no_timeout = 1;
187         cfs_atomic_inc(&suspend_timeouts);
188         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
189                cfs_atomic_read(&suspend_timeouts));
190 #endif
191 }
192
193 void ptlrpc_activate_timeouts(struct obd_import *imp)
194 {
195         /*XXX: disabled for now, will be replaced by adaptive timeouts */
196 #if 0
197         if (!imp->imp_no_timeout)
198                 return;
199         imp->imp_no_timeout = 0;
200         LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
201         if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
202                 suspend_wakeup_time = cfs_time_current();
203                 cfs_waitq_signal(&suspend_timeouts_waitq);
204         }
205         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
206                cfs_atomic_read(&suspend_timeouts));
207 #endif
208 }
209
210 int ptlrpc_check_suspend(void)
211 {
212         if (cfs_atomic_read(&suspend_timeouts))
213                 return 1;
214         return 0;
215 }
216
217 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
218 {
219         struct l_wait_info lwi;
220
221         if (cfs_atomic_read(&suspend_timeouts)) {
222                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
223                           cfs_atomic_read(&suspend_timeouts));
224                 lwi = LWI_INTR(NULL, NULL);
225                 l_wait_event(suspend_timeouts_waitq,
226                              cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
227                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
228                 return 1;
229         }
230         return 0;
231 }
232
233 #ifdef __KERNEL__
234
235 static bool ir_up;
236
237 void ptlrpc_pinger_ir_up(void)
238 {
239         CDEBUG(D_HA, "IR up\n");
240         ir_up = true;
241 }
242 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
243
244 void ptlrpc_pinger_ir_down(void)
245 {
246         CDEBUG(D_HA, "IR down\n");
247         ir_up = false;
248 }
249 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
250
251 static void ptlrpc_pinger_process_import(struct obd_import *imp,
252                                          unsigned long this_ping)
253 {
254         int level;
255         int force;
256         int force_next;
257         int suppress;
258
259         spin_lock(&imp->imp_lock);
260
261         level = imp->imp_state;
262         force = imp->imp_force_verify;
263         force_next = imp->imp_force_next_verify;
264         /*
265          * This will be used below only if the import is "FULL".
266          */
267         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
268
269         imp->imp_force_verify = 0;
270
271         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
272             !force) {
273                 spin_unlock(&imp->imp_lock);
274                 return;
275         }
276
277         imp->imp_force_next_verify = 0;
278
279         spin_unlock(&imp->imp_lock);
280
281         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
282                "force %u force_next %u deactive %u pingable %u suppress %u\n",
283                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
284                ptlrpc_import_state_name(level), level, force, force_next,
285                imp->imp_deactive, imp->imp_pingable, suppress);
286
287         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
288                 /* wait for a while before trying recovery again */
289                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
290                 if (!imp->imp_no_pinger_recover)
291                         ptlrpc_initiate_recovery(imp);
292         } else if (level != LUSTRE_IMP_FULL ||
293                    imp->imp_obd->obd_no_recov ||
294                    imp_is_deactive(imp)) {
295                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
296                        "or recovery disabled: %s)\n",
297                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
298                        ptlrpc_import_state_name(level));
299         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
300                 ptlrpc_ping(imp);
301         }
302 }
303
304 static int ptlrpc_pinger_main(void *arg)
305 {
306         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
307         ENTRY;
308
309         cfs_daemonize(thread->t_name);
310
311         /* Record that the thread is running */
312         thread_set_flags(thread, SVC_RUNNING);
313         cfs_waitq_signal(&thread->t_ctl_waitq);
314
315         /* And now, loop forever, pinging as needed. */
316         while (1) {
317                 cfs_time_t this_ping = cfs_time_current();
318                 struct l_wait_info lwi;
319                 cfs_duration_t time_to_next_wake;
320                 struct timeout_item *item;
321                 cfs_list_t *iter;
322
323                 mutex_lock(&pinger_mutex);
324                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
325                         item->ti_cb(item, item->ti_cb_data);
326                 }
327                 cfs_list_for_each(iter, &pinger_imports) {
328                         struct obd_import *imp =
329                                 cfs_list_entry(iter, struct obd_import,
330                                                imp_pinger_chain);
331
332                         ptlrpc_pinger_process_import(imp, this_ping);
333                         /* obd_timeout might have changed */
334                         if (imp->imp_pingable && imp->imp_next_ping &&
335                             cfs_time_after(imp->imp_next_ping,
336                                            cfs_time_add(this_ping,
337                                                         cfs_time_seconds(PING_INTERVAL))))
338                                 ptlrpc_update_next_ping(imp, 0);
339                 }
340                 mutex_unlock(&pinger_mutex);
341                 /* update memory usage info */
342                 obd_update_maxusage();
343
344                 /* Wait until the next ping time, or until we're stopped. */
345                 time_to_next_wake = pinger_check_timeout(this_ping);
346                 /* The ping sent by ptlrpc_send_rpc may get sent out
347                    say .01 second after this.
348                    ptlrpc_pinger_sending_on_import will then set the
349                    next ping time to next_ping + .01 sec, which means
350                    we will SKIP the next ping at next_ping, and the
351                    ping will get sent 2 timeouts from now!  Beware. */
352                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
353                        CFS_TIME_T")\n", time_to_next_wake,
354                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
355                 if (time_to_next_wake > 0) {
356                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
357                                                 time_to_next_wake,
358                                                 cfs_time_seconds(1)),
359                                           NULL, NULL);
360                         l_wait_event(thread->t_ctl_waitq,
361                                      thread_is_stopping(thread) ||
362                                      thread_is_event(thread),
363                                      &lwi);
364                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
365                                 EXIT;
366                                 break;
367                         } else {
368                                 /* woken after adding import to reset timer */
369                                 thread_test_and_clear_flags(thread, SVC_EVENT);
370                         }
371                 }
372         }
373
374         thread_set_flags(thread, SVC_STOPPED);
375         cfs_waitq_signal(&thread->t_ctl_waitq);
376
377         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
378         return 0;
379 }
380
381 static struct ptlrpc_thread pinger_thread;
382
383 int ptlrpc_start_pinger(void)
384 {
385         struct l_wait_info lwi = { 0 };
386         int rc;
387 #ifndef ENABLE_PINGER
388         return 0;
389 #endif
390         ENTRY;
391
392         if (!thread_is_init(&pinger_thread) &&
393             !thread_is_stopped(&pinger_thread))
394                 RETURN(-EALREADY);
395
396         cfs_waitq_init(&pinger_thread.t_ctl_waitq);
397         cfs_waitq_init(&suspend_timeouts_waitq);
398
399         strcpy(pinger_thread.t_name, "ll_ping");
400
401         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
402          * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
403         rc = cfs_create_thread(ptlrpc_pinger_main,
404                                &pinger_thread, CFS_DAEMON_FLAGS);
405         if (rc < 0) {
406                 CERROR("cannot start thread: %d\n", rc);
407                 RETURN(rc);
408         }
409         l_wait_event(pinger_thread.t_ctl_waitq,
410                      thread_is_running(&pinger_thread), &lwi);
411
412         if (suppress_pings)
413                 CWARN("Pings will be suppressed at the request of the "
414                       "administrator.  The configuration shall meet the "
415                       "additional requirements described in the manual.  "
416                       "(Search for the \"suppress_pings\" kernel module "
417                       "parameter.)\n");
418
419         RETURN(0);
420 }
421
422 int ptlrpc_pinger_remove_timeouts(void);
423
424 int ptlrpc_stop_pinger(void)
425 {
426         struct l_wait_info lwi = { 0 };
427 #ifndef ENABLE_PINGER
428         return 0;
429 #endif
430         ENTRY;
431
432         if (thread_is_init(&pinger_thread) ||
433             thread_is_stopped(&pinger_thread))
434                 RETURN(-EALREADY);
435
436         ptlrpc_pinger_remove_timeouts();
437
438         thread_set_flags(&pinger_thread, SVC_STOPPING);
439         cfs_waitq_signal(&pinger_thread.t_ctl_waitq);
440
441         l_wait_event(pinger_thread.t_ctl_waitq,
442                      thread_is_stopped(&pinger_thread), &lwi);
443         RETURN(0);
444 }
445
446 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
447 {
448         ptlrpc_update_next_ping(imp, 0);
449 }
450 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
451
452 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
453 {
454         ptlrpc_update_next_ping(imp, 1);
455         LASSERT_SPIN_LOCKED(&imp->imp_lock);
456         /*
457          * Avoid reading stale imp_connect_data.  When not sure if pings are
458          * expected or not on next connection, we assume they are not and force
459          * one anyway to guarantee the chance of updating
460          * imp_peer_committed_transno.
461          */
462         if (imp->imp_state != LUSTRE_IMP_FULL ||
463             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
464                 imp->imp_force_next_verify = 1;
465 }
466
467 int ptlrpc_pinger_add_import(struct obd_import *imp)
468 {
469         ENTRY;
470         if (!cfs_list_empty(&imp->imp_pinger_chain))
471                 RETURN(-EALREADY);
472
473         mutex_lock(&pinger_mutex);
474         CDEBUG(D_HA, "adding pingable import %s->%s\n",
475                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
476         /* if we add to pinger we want recovery on this import */
477         imp->imp_obd->obd_no_recov = 0;
478         ptlrpc_update_next_ping(imp, 0);
479         /* XXX sort, blah blah */
480         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
481         class_import_get(imp);
482
483         ptlrpc_pinger_wake_up();
484         mutex_unlock(&pinger_mutex);
485
486         RETURN(0);
487 }
488 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
489
490 int ptlrpc_pinger_del_import(struct obd_import *imp)
491 {
492         ENTRY;
493         if (cfs_list_empty(&imp->imp_pinger_chain))
494                 RETURN(-ENOENT);
495
496         mutex_lock(&pinger_mutex);
497         cfs_list_del_init(&imp->imp_pinger_chain);
498         CDEBUG(D_HA, "removing pingable import %s->%s\n",
499                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
500         /* if we remove from pinger we don't want recovery on this import */
501         imp->imp_obd->obd_no_recov = 1;
502         class_import_put(imp);
503         mutex_unlock(&pinger_mutex);
504         RETURN(0);
505 }
506 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
507
508 /**
509  * Register a timeout callback to the pinger list, and the callback will
510  * be called when timeout happens.
511  */
512 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
513                                         timeout_cb_t cb, void *data)
514 {
515         struct timeout_item *ti;
516
517         OBD_ALLOC_PTR(ti);
518         if (!ti)
519                 return(NULL);
520
521         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
522         CFS_INIT_LIST_HEAD(&ti->ti_chain);
523         ti->ti_timeout = time;
524         ti->ti_event = event;
525         ti->ti_cb = cb;
526         ti->ti_cb_data = data;
527
528         return ti;
529 }
530
531 /**
532  * Register timeout event on the the pinger thread.
533  * Note: the timeout list is an sorted list with increased timeout value.
534  */
535 static struct timeout_item*
536 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
537                                timeout_cb_t cb, void *data)
538 {
539         struct timeout_item *item, *tmp;
540
541         LASSERT_MUTEX_LOCKED(&pinger_mutex);
542
543         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
544                 if (item->ti_event == event)
545                         goto out;
546
547         item = ptlrpc_new_timeout(time, event, cb, data);
548         if (item) {
549                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
550                         if (tmp->ti_timeout < time) {
551                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
552                                 goto out;
553                         }
554                 }
555                 cfs_list_add(&item->ti_chain, &timeout_list);
556         }
557 out:
558         return item;
559 }
560
561 /* Add a client_obd to the timeout event list, when timeout(@time)
562  * happens, the callback(@cb) will be called.
563  */
564 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
565                               timeout_cb_t cb, void *data,
566                               cfs_list_t *obd_list)
567 {
568         struct timeout_item *ti;
569
570         mutex_lock(&pinger_mutex);
571         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
572         if (!ti) {
573                 mutex_unlock(&pinger_mutex);
574                 return (-EINVAL);
575         }
576         cfs_list_add(obd_list, &ti->ti_obd_list);
577         mutex_unlock(&pinger_mutex);
578         return 0;
579 }
580 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
581
582 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
583                               enum timeout_event event)
584 {
585         struct timeout_item *ti = NULL, *item;
586
587         if (cfs_list_empty(obd_list))
588                 return 0;
589         mutex_lock(&pinger_mutex);
590         cfs_list_del_init(obd_list);
591         /**
592          * If there are no obd attached to the timeout event
593          * list, remove this timeout event from the pinger
594          */
595         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
596                 if (item->ti_event == event) {
597                         ti = item;
598                         break;
599                 }
600         }
601         LASSERTF(ti != NULL, "ti is NULL ! \n");
602         if (cfs_list_empty(&ti->ti_obd_list)) {
603                 cfs_list_del(&ti->ti_chain);
604                 OBD_FREE_PTR(ti);
605         }
606         mutex_unlock(&pinger_mutex);
607         return 0;
608 }
609 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
610
611 int ptlrpc_pinger_remove_timeouts(void)
612 {
613         struct timeout_item *item, *tmp;
614
615         mutex_lock(&pinger_mutex);
616         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
617                 LASSERT(cfs_list_empty(&item->ti_obd_list));
618                 cfs_list_del(&item->ti_chain);
619                 OBD_FREE_PTR(item);
620         }
621         mutex_unlock(&pinger_mutex);
622         return 0;
623 }
624
625 void ptlrpc_pinger_wake_up()
626 {
627 #ifdef ENABLE_PINGER
628         thread_add_flags(&pinger_thread, SVC_EVENT);
629         cfs_waitq_signal(&pinger_thread.t_ctl_waitq);
630 #endif
631 }
632
633 /* Ping evictor thread */
634 #define PET_READY     1
635 #define PET_TERMINATE 2
636
637 static int               pet_refcount = 0;
638 static int               pet_state;
639 static cfs_waitq_t       pet_waitq;
640 CFS_LIST_HEAD(pet_list);
641 static DEFINE_SPINLOCK(pet_lock);
642
643 int ping_evictor_wake(struct obd_export *exp)
644 {
645         struct obd_device *obd;
646
647         spin_lock(&pet_lock);
648         if (pet_state != PET_READY) {
649                 /* eventually the new obd will call here again. */
650                 spin_unlock(&pet_lock);
651                 return 1;
652         }
653
654         obd = class_exp2obd(exp);
655         if (cfs_list_empty(&obd->obd_evict_list)) {
656                 class_incref(obd, "evictor", obd);
657                 cfs_list_add(&obd->obd_evict_list, &pet_list);
658         }
659         spin_unlock(&pet_lock);
660
661         cfs_waitq_signal(&pet_waitq);
662         return 0;
663 }
664
665 static int ping_evictor_main(void *arg)
666 {
667         struct obd_device *obd;
668         struct obd_export *exp;
669         struct l_wait_info lwi = { 0 };
670         time_t expire_time;
671         ENTRY;
672
673         cfs_daemonize_ctxt("ll_evictor");
674
675         CDEBUG(D_HA, "Starting Ping Evictor\n");
676         pet_state = PET_READY;
677         while (1) {
678                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
679                              (pet_state == PET_TERMINATE), &lwi);
680
681                 /* loop until all obd's will be removed */
682                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
683                         break;
684
685                 /* we only get here if pet_exp != NULL, and the end of this
686                  * loop is the only place which sets it NULL again, so lock
687                  * is not strictly necessary. */
688                 spin_lock(&pet_lock);
689                 obd = cfs_list_entry(pet_list.next, struct obd_device,
690                                      obd_evict_list);
691                 spin_unlock(&pet_lock);
692
693                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
694
695                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
696                        obd->obd_name, expire_time);
697
698                 /* Exports can't be deleted out of the list while we hold
699                  * the obd lock (class_unlink_export), which means we can't
700                  * lose the last ref on the export.  If they've already been
701                  * removed from the list, we won't find them here. */
702                 spin_lock(&obd->obd_dev_lock);
703                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
704                         exp = cfs_list_entry(obd->obd_exports_timed.next,
705                                              struct obd_export,
706                                              exp_obd_chain_timed);
707                         if (expire_time > exp->exp_last_request_time) {
708                                 class_export_get(exp);
709                                 spin_unlock(&obd->obd_dev_lock);
710                                 LCONSOLE_WARN("%s: haven't heard from client %s"
711                                               " (at %s) in %ld seconds. I think"
712                                               " it's dead, and I am evicting"
713                                               " it. exp %p, cur %ld expire %ld"
714                                               " last %ld\n",
715                                               obd->obd_name,
716                                               obd_uuid2str(&exp->exp_client_uuid),
717                                               obd_export_nid2str(exp),
718                                               (long)(cfs_time_current_sec() -
719                                                      exp->exp_last_request_time),
720                                               exp, (long)cfs_time_current_sec(),
721                                               (long)expire_time,
722                                               (long)exp->exp_last_request_time);
723                                 CDEBUG(D_HA, "Last request was at %ld\n",
724                                        exp->exp_last_request_time);
725                                 class_fail_export(exp);
726                                 class_export_put(exp);
727                                 spin_lock(&obd->obd_dev_lock);
728                         } else {
729                                 /* List is sorted, so everyone below is ok */
730                                 break;
731                         }
732                 }
733                 spin_unlock(&obd->obd_dev_lock);
734
735                 spin_lock(&pet_lock);
736                 cfs_list_del_init(&obd->obd_evict_list);
737                 spin_unlock(&pet_lock);
738
739                 class_decref(obd, "evictor", obd);
740         }
741         CDEBUG(D_HA, "Exiting Ping Evictor\n");
742
743         RETURN(0);
744 }
745
746 void ping_evictor_start(void)
747 {
748         int rc;
749
750         if (++pet_refcount > 1)
751                 return;
752
753         cfs_waitq_init(&pet_waitq);
754
755         rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
756         if (rc < 0) {
757                 pet_refcount--;
758                 CERROR("Cannot start ping evictor thread: %d\n", rc);
759         }
760 }
761 EXPORT_SYMBOL(ping_evictor_start);
762
763 void ping_evictor_stop(void)
764 {
765         if (--pet_refcount > 0)
766                 return;
767
768         pet_state = PET_TERMINATE;
769         cfs_waitq_signal(&pet_waitq);
770 }
771 EXPORT_SYMBOL(ping_evictor_stop);
772 #else /* !__KERNEL__ */
773
774 /* XXX
775  * the current implementation of pinger in liblustre is not optimized
776  */
777
778 #ifdef ENABLE_PINGER
779 static struct pinger_data {
780         int             pd_recursion;
781         cfs_time_t      pd_this_ping;   /* jiffies */
782         cfs_time_t      pd_next_ping;   /* jiffies */
783         struct ptlrpc_request_set *pd_set;
784 } pinger_args;
785
786 static int pinger_check_rpcs(void *arg)
787 {
788         cfs_time_t curtime = cfs_time_current();
789         struct ptlrpc_request *req;
790         struct ptlrpc_request_set *set;
791         cfs_list_t *iter;
792         struct obd_import *imp;
793         struct pinger_data *pd = &pinger_args;
794         int rc;
795
796         /* prevent recursion */
797         if (pd->pd_recursion++) {
798                 CDEBUG(D_HA, "pinger: recursion! quit\n");
799                 LASSERT(pd->pd_set);
800                 pd->pd_recursion--;
801                 return 0;
802         }
803
804         /* have we reached ping point? */
805         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
806                 pd->pd_recursion--;
807                 return 0;
808         }
809
810         /* if we have rpc_set already, continue processing it */
811         if (pd->pd_set) {
812                 LASSERT(pd->pd_this_ping);
813                 set = pd->pd_set;
814                 goto do_check_set;
815         }
816
817         pd->pd_this_ping = curtime;
818         pd->pd_set = ptlrpc_prep_set();
819         if (pd->pd_set == NULL)
820                 goto out;
821         set = pd->pd_set;
822
823         /* add rpcs into set */
824         mutex_lock(&pinger_mutex);
825         cfs_list_for_each(iter, &pinger_imports) {
826                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
827                                                         imp_pinger_chain);
828                 int generation, level;
829
830                 if (cfs_time_aftereq(pd->pd_this_ping,
831                                      imp->imp_next_ping - 5 * CFS_TICK)) {
832                         /* Add a ping. */
833                         spin_lock(&imp->imp_lock);
834                         generation = imp->imp_generation;
835                         level = imp->imp_state;
836                         spin_unlock(&imp->imp_lock);
837
838                         if (level != LUSTRE_IMP_FULL) {
839                                 CDEBUG(D_HA,
840                                        "not pinging %s (in recovery)\n",
841                                        obd2cli_tgt(imp->imp_obd));
842                                 continue;
843                         }
844
845                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
846                                                         LUSTRE_OBD_VERSION,
847                                                         OBD_PING);
848                         if (req == NULL) {
849                                 CERROR("OOM trying to ping %s->%s\n",
850                                        imp->imp_obd->obd_uuid.uuid,
851                                        obd2cli_tgt(imp->imp_obd));
852                                 break;
853                         }
854
855                         req->rq_no_resend = 1;
856                         ptlrpc_request_set_replen(req);
857                         req->rq_send_state = LUSTRE_IMP_FULL;
858                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
859                         req->rq_import_generation = generation;
860                         ptlrpc_set_add_req(set, req);
861                 } else {
862                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
863                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
864                                imp->imp_next_ping, pd->pd_this_ping);
865                 }
866         }
867         pd->pd_this_ping = curtime;
868         mutex_unlock(&pinger_mutex);
869
870         /* Might be empty, that's OK. */
871         if (cfs_atomic_read(&set->set_remaining) == 0)
872                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
873
874         cfs_list_for_each(iter, &set->set_requests) {
875                 struct ptlrpc_request *req =
876                         cfs_list_entry(iter, struct ptlrpc_request,
877                                        rq_set_chain);
878                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
879                           req->rq_import->imp_obd->obd_uuid.uuid,
880                           obd2cli_tgt(req->rq_import->imp_obd));
881                 (void)ptl_send_rpc(req, 0);
882         }
883
884 do_check_set:
885         rc = ptlrpc_check_set(NULL, set);
886
887         /* not finished, and we are not expired, simply return */
888         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
889                                             cfs_time_seconds(PING_INTERVAL)))) {
890                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
891                 pd->pd_recursion--;
892                 return 0;
893         }
894
895         /* Expire all the requests that didn't come back. */
896         mutex_lock(&pinger_mutex);
897         cfs_list_for_each(iter, &set->set_requests) {
898                 req = cfs_list_entry(iter, struct ptlrpc_request,
899                                      rq_set_chain);
900
901                 if (req->rq_phase == RQ_PHASE_COMPLETE)
902                         continue;
903
904                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
905                        req);
906
907                 /* This will also unregister reply. */
908                 ptlrpc_expire_one_request(req, 0);
909
910                 /* We're done with this req, let's finally move it to complete
911                  * phase and take care of inflights. */
912                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
913                 imp = req->rq_import;
914                 spin_lock(&imp->imp_lock);
915                 if (!cfs_list_empty(&req->rq_list)) {
916                         cfs_list_del_init(&req->rq_list);
917                         cfs_atomic_dec(&imp->imp_inflight);
918                 }
919                 spin_unlock(&imp->imp_lock);
920                 cfs_atomic_dec(&set->set_remaining);
921         }
922         mutex_unlock(&pinger_mutex);
923
924         ptlrpc_set_destroy(set);
925         pd->pd_set = NULL;
926
927 out:
928         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
929                                         cfs_time_seconds(PING_INTERVAL));
930         pd->pd_this_ping = 0; /* XXX for debug */
931
932         CDEBUG(D_INFO, "finished a round ping\n");
933         pd->pd_recursion--;
934         return 0;
935 }
936
937 static void *pinger_callback = NULL;
938 #endif /* ENABLE_PINGER */
939
940 int ptlrpc_start_pinger(void)
941 {
942 #ifdef ENABLE_PINGER
943         memset(&pinger_args, 0, sizeof(pinger_args));
944         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
945                                                            &pinger_check_rpcs,
946                                                            &pinger_args);
947 #endif
948         return 0;
949 }
950
951 int ptlrpc_stop_pinger(void)
952 {
953 #ifdef ENABLE_PINGER
954         if (pinger_callback)
955                 liblustre_deregister_wait_callback(pinger_callback);
956 #endif
957         return 0;
958 }
959
960 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
961 {
962 #ifdef ENABLE_PINGER
963         mutex_lock(&pinger_mutex);
964         ptlrpc_update_next_ping(imp, 0);
965         if (pinger_args.pd_set == NULL &&
966             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
967                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
968                         imp->imp_next_ping, cfs_time_current());
969                 pinger_args.pd_next_ping = imp->imp_next_ping;
970         }
971         mutex_unlock(&pinger_mutex);
972 #endif
973 }
974
975 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
976 {
977 #ifdef ENABLE_PINGER
978         mutex_lock(&pinger_mutex);
979         ptlrpc_update_next_ping(imp, 1);
980         if (pinger_args.pd_set == NULL &&
981             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
982                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
983                         imp->imp_next_ping, cfs_time_current());
984                 pinger_args.pd_next_ping = imp->imp_next_ping;
985         }
986         mutex_unlock(&pinger_mutex);
987 #endif
988 }
989
990 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
991                               timeout_cb_t cb, void *data,
992                               cfs_list_t *obd_list)
993 {
994         return 0;
995 }
996
997 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
998                               enum timeout_event event)
999 {
1000         return 0;
1001 }
1002
1003 int ptlrpc_pinger_add_import(struct obd_import *imp)
1004 {
1005         ENTRY;
1006         if (!cfs_list_empty(&imp->imp_pinger_chain))
1007                 RETURN(-EALREADY);
1008
1009         CDEBUG(D_HA, "adding pingable import %s->%s\n",
1010                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1011         ptlrpc_pinger_sending_on_import(imp);
1012
1013         mutex_lock(&pinger_mutex);
1014         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
1015         class_import_get(imp);
1016         mutex_unlock(&pinger_mutex);
1017
1018         RETURN(0);
1019 }
1020
1021 int ptlrpc_pinger_del_import(struct obd_import *imp)
1022 {
1023         ENTRY;
1024         if (cfs_list_empty(&imp->imp_pinger_chain))
1025                 RETURN(-ENOENT);
1026
1027         mutex_lock(&pinger_mutex);
1028         cfs_list_del_init(&imp->imp_pinger_chain);
1029         CDEBUG(D_HA, "removing pingable import %s->%s\n",
1030                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1031         class_import_put(imp);
1032         mutex_unlock(&pinger_mutex);
1033         RETURN(0);
1034 }
1035
1036 void ptlrpc_pinger_wake_up()
1037 {
1038 #ifdef ENABLE_PINGER
1039         /* XXX force pinger to run, if needed */
1040         struct obd_import *imp;
1041         ENTRY;
1042         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
1043                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
1044                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1045 #ifdef ENABLE_LIBLUSTRE_RECOVERY
1046                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
1047                     !imp_is_deactive(imp))
1048 #else
1049                 /*XXX only recover for the initial connection */
1050                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1051                     imp->imp_state == LUSTRE_IMP_DISCON &&
1052                     !imp_is_deactive(imp))
1053 #endif
1054                         ptlrpc_initiate_recovery(imp);
1055                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1056                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1057                                      "state %d, deactive %d\n",
1058                                      imp->imp_obd->obd_uuid.uuid,
1059                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1060                                      imp_is_deactive(imp));
1061         }
1062         EXIT;
1063 #endif
1064 }
1065 #endif /* !__KERNEL__ */