dlm: keep cached master rsbs during recovery
[linux-2.6.git] / fs / dlm / recoverd.c
1 /******************************************************************************
2 *******************************************************************************
3 **
4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
5 **  Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
6 **
7 **  This copyrighted material is made available to anyone wishing to use,
8 **  modify, copy, or redistribute it subject to the terms and conditions
9 **  of the GNU General Public License v.2.
10 **
11 *******************************************************************************
12 ******************************************************************************/
13
14 #include "dlm_internal.h"
15 #include "lockspace.h"
16 #include "member.h"
17 #include "dir.h"
18 #include "ast.h"
19 #include "recover.h"
20 #include "lowcomms.h"
21 #include "lock.h"
22 #include "requestqueue.h"
23 #include "recoverd.h"
24
25
26 /* If the start for which we're re-enabling locking (seq) has been superseded
27    by a newer stop (ls_recover_seq), we need to leave locking disabled.
28
29    We suspend dlm_recv threads here to avoid the race where dlm_recv a) sees
30    locking stopped and b) adds a message to the requestqueue, but dlm_recoverd
31    enables locking and clears the requestqueue between a and b. */
32
33 static int enable_locking(struct dlm_ls *ls, uint64_t seq)
34 {
35         int error = -EINTR;
36
37         down_write(&ls->ls_recv_active);
38
39         spin_lock(&ls->ls_recover_lock);
40         if (ls->ls_recover_seq == seq) {
41                 set_bit(LSFL_RUNNING, &ls->ls_flags);
42                 /* unblocks processes waiting to enter the dlm */
43                 up_write(&ls->ls_in_recovery);
44                 error = 0;
45         }
46         spin_unlock(&ls->ls_recover_lock);
47
48         up_write(&ls->ls_recv_active);
49         return error;
50 }
51
52 static int ls_recover(struct dlm_ls *ls, struct dlm_recover *rv)
53 {
54         unsigned long start;
55         int error, neg = 0;
56
57         log_debug(ls, "recover %llx", (unsigned long long)rv->seq);
58
59         mutex_lock(&ls->ls_recoverd_active);
60
61         /*
62          * Suspending and resuming dlm_astd ensures that no lkb's from this ls
63          * will be processed by dlm_astd during recovery.
64          */
65
66         dlm_astd_suspend();
67         dlm_astd_resume();
68
69         /*
70          * Free non-master tossed rsb's.  Master rsb's are kept on toss
71          * list and put on root list to be included in resdir recovery.
72          */
73
74         dlm_clear_toss_list(ls);
75
76         /*
77          * This list of root rsb's will be the basis of most of the recovery
78          * routines.
79          */
80
81         dlm_create_root_list(ls);
82
83         /*
84          * Add or remove nodes from the lockspace's ls_nodes list.
85          * Also waits for all nodes to complete dlm_recover_members.
86          */
87
88         error = dlm_recover_members(ls, rv, &neg);
89         if (error) {
90                 log_debug(ls, "recover_members failed %d", error);
91                 goto fail;
92         }
93         start = jiffies;
94
95         /*
96          * Rebuild our own share of the directory by collecting from all other
97          * nodes their master rsb names that hash to us.
98          */
99
100         error = dlm_recover_directory(ls);
101         if (error) {
102                 log_debug(ls, "recover_directory failed %d", error);
103                 goto fail;
104         }
105
106         /*
107          * Wait for all nodes to complete directory rebuild.
108          */
109
110         error = dlm_recover_directory_wait(ls);
111         if (error) {
112                 log_debug(ls, "recover_directory_wait failed %d", error);
113                 goto fail;
114         }
115
116         /*
117          * We may have outstanding operations that are waiting for a reply from
118          * a failed node.  Mark these to be resent after recovery.  Unlock and
119          * cancel ops can just be completed.
120          */
121
122         dlm_recover_waiters_pre(ls);
123
124         error = dlm_recovery_stopped(ls);
125         if (error)
126                 goto fail;
127
128         if (neg || dlm_no_directory(ls)) {
129                 /*
130                  * Clear lkb's for departed nodes.
131                  */
132
133                 dlm_purge_locks(ls);
134
135                 /*
136                  * Get new master nodeid's for rsb's that were mastered on
137                  * departed nodes.
138                  */
139
140                 error = dlm_recover_masters(ls);
141                 if (error) {
142                         log_debug(ls, "recover_masters failed %d", error);
143                         goto fail;
144                 }
145
146                 /*
147                  * Send our locks on remastered rsb's to the new masters.
148                  */
149
150                 error = dlm_recover_locks(ls);
151                 if (error) {
152                         log_debug(ls, "recover_locks failed %d", error);
153                         goto fail;
154                 }
155
156                 error = dlm_recover_locks_wait(ls);
157                 if (error) {
158                         log_debug(ls, "recover_locks_wait failed %d", error);
159                         goto fail;
160                 }
161
162                 /*
163                  * Finalize state in master rsb's now that all locks can be
164                  * checked.  This includes conversion resolution and lvb
165                  * settings.
166                  */
167
168                 dlm_recover_rsbs(ls);
169         } else {
170                 /*
171                  * Other lockspace members may be going through the "neg" steps
172                  * while also adding us to the lockspace, in which case they'll
173                  * be doing the recover_locks (RS_LOCKS) barrier.
174                  */
175                 dlm_set_recover_status(ls, DLM_RS_LOCKS);
176
177                 error = dlm_recover_locks_wait(ls);
178                 if (error) {
179                         log_debug(ls, "recover_locks_wait failed %d", error);
180                         goto fail;
181                 }
182         }
183
184         dlm_release_root_list(ls);
185
186         /*
187          * Purge directory-related requests that are saved in requestqueue.
188          * All dir requests from before recovery are invalid now due to the dir
189          * rebuild and will be resent by the requesting nodes.
190          */
191
192         dlm_purge_requestqueue(ls);
193
194         dlm_set_recover_status(ls, DLM_RS_DONE);
195         error = dlm_recover_done_wait(ls);
196         if (error) {
197                 log_debug(ls, "recover_done_wait failed %d", error);
198                 goto fail;
199         }
200
201         dlm_clear_members_gone(ls);
202
203         dlm_adjust_timeouts(ls);
204
205         error = enable_locking(ls, rv->seq);
206         if (error) {
207                 log_debug(ls, "enable_locking failed %d", error);
208                 goto fail;
209         }
210
211         error = dlm_process_requestqueue(ls);
212         if (error) {
213                 log_debug(ls, "process_requestqueue failed %d", error);
214                 goto fail;
215         }
216
217         error = dlm_recover_waiters_post(ls);
218         if (error) {
219                 log_debug(ls, "recover_waiters_post failed %d", error);
220                 goto fail;
221         }
222
223         dlm_grant_after_purge(ls);
224
225         dlm_astd_wake();
226
227         log_debug(ls, "recover %llx done: %u ms",
228                   (unsigned long long)rv->seq,
229                   jiffies_to_msecs(jiffies - start));
230         mutex_unlock(&ls->ls_recoverd_active);
231
232         return 0;
233
234  fail:
235         dlm_release_root_list(ls);
236         log_debug(ls, "recover %llx error %d",
237                   (unsigned long long)rv->seq, error);
238         mutex_unlock(&ls->ls_recoverd_active);
239         return error;
240 }
241
242 /* The dlm_ls_start() that created the rv we take here may already have been
243    stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
244    flag set. */
245
246 static void do_ls_recovery(struct dlm_ls *ls)
247 {
248         struct dlm_recover *rv = NULL;
249
250         spin_lock(&ls->ls_recover_lock);
251         rv = ls->ls_recover_args;
252         ls->ls_recover_args = NULL;
253         if (rv && ls->ls_recover_seq == rv->seq)
254                 clear_bit(LSFL_RECOVERY_STOP, &ls->ls_flags);
255         spin_unlock(&ls->ls_recover_lock);
256
257         if (rv) {
258                 ls_recover(ls, rv);
259                 kfree(rv->nodeids);
260                 kfree(rv);
261         }
262 }
263
264 static int dlm_recoverd(void *arg)
265 {
266         struct dlm_ls *ls;
267
268         ls = dlm_find_lockspace_local(arg);
269         if (!ls) {
270                 log_print("dlm_recoverd: no lockspace %p", arg);
271                 return -1;
272         }
273
274         while (!kthread_should_stop()) {
275                 set_current_state(TASK_INTERRUPTIBLE);
276                 if (!test_bit(LSFL_WORK, &ls->ls_flags))
277                         schedule();
278                 set_current_state(TASK_RUNNING);
279
280                 if (test_and_clear_bit(LSFL_WORK, &ls->ls_flags))
281                         do_ls_recovery(ls);
282         }
283
284         dlm_put_lockspace(ls);
285         return 0;
286 }
287
288 void dlm_recoverd_kick(struct dlm_ls *ls)
289 {
290         set_bit(LSFL_WORK, &ls->ls_flags);
291         wake_up_process(ls->ls_recoverd_task);
292 }
293
294 int dlm_recoverd_start(struct dlm_ls *ls)
295 {
296         struct task_struct *p;
297         int error = 0;
298
299         p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
300         if (IS_ERR(p))
301                 error = PTR_ERR(p);
302         else
303                 ls->ls_recoverd_task = p;
304         return error;
305 }
306
307 void dlm_recoverd_stop(struct dlm_ls *ls)
308 {
309         kthread_stop(ls->ls_recoverd_task);
310 }
311
312 void dlm_recoverd_suspend(struct dlm_ls *ls)
313 {
314         wake_up(&ls->ls_wait_general);
315         mutex_lock(&ls->ls_recoverd_active);
316 }
317
318 void dlm_recoverd_resume(struct dlm_ls *ls)
319 {
320         mutex_unlock(&ls->ls_recoverd_active);
321 }
322