dlm: convert rsb list to rb_tree
[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26 /* If the start for which we're re-enabling locking (seq) has been superseded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         dlm_callback_suspend(ls);
62
63         /*
64          * Free non-master tossed rsb's.  Master rsb's are kept on toss
65          * list and put on root list to be included in resdir recovery.
66          */
67
68         dlm_clear_toss_list(ls);
69
70         /*
71          * This list of root rsb's will be the basis of most of the recovery
72          * routines.
73          */
74
75         dlm_create_root_list(ls);
76
77         /*
78          * Add or remove nodes from the lockspace's ls_nodes list.
79          * Also waits for all nodes to complete dlm_recover_members.
80          */
81
82         error = dlm_recover_members(ls, rv, &neg);
83         if (error) {
84                 log_debug(ls, "recover_members failed %d", error);
85                 goto fail;
86         }
87         start = jiffies;
88
89         /*
90          * Rebuild our own share of the directory by collecting from all other
91          * nodes their master rsb names that hash to us.
92          */
93
94         error = dlm_recover_directory(ls);
95         if (error) {
96                 log_debug(ls, "recover_directory failed %d", error);
97                 goto fail;
98         }
99
100         /*
101          * Wait for all nodes to complete directory rebuild.
102          */
103
104         error = dlm_recover_directory_wait(ls);
105         if (error) {
106                 log_debug(ls, "recover_directory_wait failed %d", error);
107                 goto fail;
108         }
109
110         /*
111          * We may have outstanding operations that are waiting for a reply from
112          * a failed node.  Mark these to be resent after recovery.  Unlock and
113          * cancel ops can just be completed.
114          */
115
116         dlm_recover_waiters_pre(ls);
117
118         error = dlm_recovery_stopped(ls);
119         if (error)
120                 goto fail;
121
122         if (neg || dlm_no_directory(ls)) {
123                 /*
124                  * Clear lkb's for departed nodes.
125                  */
126
127                 dlm_purge_locks(ls);
128
129                 /*
130                  * Get new master nodeid's for rsb's that were mastered on
131                  * departed nodes.
132                  */
133
134                 error = dlm_recover_masters(ls);
135                 if (error) {
136                         log_debug(ls, "recover_masters failed %d", error);
137                         goto fail;
138                 }
139
140                 /*
141                  * Send our locks on remastered rsb's to the new masters.
142                  */
143
144                 error = dlm_recover_locks(ls);
145                 if (error) {
146                         log_debug(ls, "recover_locks failed %d", error);
147                         goto fail;
148                 }
149
150                 error = dlm_recover_locks_wait(ls);
151                 if (error) {
152                         log_debug(ls, "recover_locks_wait failed %d", error);
153                         goto fail;
154                 }
155
156                 /*
157                  * Finalize state in master rsb's now that all locks can be
158                  * checked.  This includes conversion resolution and lvb
159                  * settings.
160                  */
161
162                 dlm_recover_rsbs(ls);
163         } else {
164                 /*
165                  * Other lockspace members may be going through the "neg" steps
166                  * while also adding us to the lockspace, in which case they'll
167                  * be doing the recover_locks (RS_LOCKS) barrier.
168                  */
169                 dlm_set_recover_status(ls, DLM_RS_LOCKS);
170
171                 error = dlm_recover_locks_wait(ls);
172                 if (error) {
173                         log_debug(ls, "recover_locks_wait failed %d", error);
174                         goto fail;
175                 }
176         }
177
178         dlm_release_root_list(ls);
179
180         /*
181          * Purge directory-related requests that are saved in requestqueue.
182          * All dir requests from before recovery are invalid now due to the dir
183          * rebuild and will be resent by the requesting nodes.
184          */
185
186         dlm_purge_requestqueue(ls);
187
188         dlm_set_recover_status(ls, DLM_RS_DONE);
189         error = dlm_recover_done_wait(ls);
190         if (error) {
191                 log_debug(ls, "recover_done_wait failed %d", error);
192                 goto fail;
193         }
194
195         dlm_clear_members_gone(ls);
196
197         dlm_adjust_timeouts(ls);
198
199         dlm_callback_resume(ls);
200
201         error = enable_locking(ls, rv->seq);
202         if (error) {
203                 log_debug(ls, "enable_locking failed %d", error);
204                 goto fail;
205         }
206
207         error = dlm_process_requestqueue(ls);
208         if (error) {
209                 log_debug(ls, "process_requestqueue failed %d", error);
210                 goto fail;
211         }
212
213         error = dlm_recover_waiters_post(ls);
214         if (error) {
215                 log_debug(ls, "recover_waiters_post failed %d", error);
216                 goto fail;
217         }
218
219         dlm_grant_after_purge(ls);
220
221         log_debug(ls, "recover %llx done: %u ms",
222                   (unsigned long long)rv->seq,
223                   jiffies_to_msecs(jiffies - start));
224         mutex_unlock(&ls->ls_recoverd_active);
225
226         return 0;
227
228  fail:
229         dlm_release_root_list(ls);
230         log_debug(ls, "recover %llx error %d",
231                   (unsigned long long)rv->seq, error);
232         mutex_unlock(&ls->ls_recoverd_active);
233         return error;
234 }
235
236 /* The dlm_ls_start() that created the rv we take here may already have been
237    stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
238    flag set. */
239
240 static void do_ls_recovery(struct dlm_ls *ls)
241 {
242         struct dlm_recover *rv = NULL;
243
244         spin_lock(&ls->ls_recover_lock);
245         rv = ls->ls_recover_args;
246         ls->ls_recover_args = NULL;
247         if (rv && ls->ls_recover_seq == rv->seq)
248                 clear_bit(LSFL_RECOVERY_STOP, &ls->ls_flags);
249         spin_unlock(&ls->ls_recover_lock);
250
251         if (rv) {
252                 ls_recover(ls, rv);
253                 kfree(rv->nodeids);
254                 kfree(rv->new);
255                 kfree(rv);
256         }
257 }
258
259 static int dlm_recoverd(void *arg)
260 {
261         struct dlm_ls *ls;
262
263         ls = dlm_find_lockspace_local(arg);
264         if (!ls) {
265                 log_print("dlm_recoverd: no lockspace %p", arg);
266                 return -1;
267         }
268
269         while (!kthread_should_stop()) {
270                 set_current_state(TASK_INTERRUPTIBLE);
271                 if (!test_bit(LSFL_WORK, &ls->ls_flags))
272                         schedule();
273                 set_current_state(TASK_RUNNING);
274
275                 if (test_and_clear_bit(LSFL_WORK, &ls->ls_flags))
276                         do_ls_recovery(ls);
277         }
278
279         dlm_put_lockspace(ls);
280         return 0;
281 }
282
283 void dlm_recoverd_kick(struct dlm_ls *ls)
284 {
285         set_bit(LSFL_WORK, &ls->ls_flags);
286         wake_up_process(ls->ls_recoverd_task);
287 }
288
289 int dlm_recoverd_start(struct dlm_ls *ls)
290 {
291         struct task_struct *p;
292         int error = 0;
293
294         p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
295         if (IS_ERR(p))
296                 error = PTR_ERR(p);
297         else
298                 ls->ls_recoverd_task = p;
299         return error;
300 }
301
302 void dlm_recoverd_stop(struct dlm_ls *ls)
303 {
304         kthread_stop(ls->ls_recoverd_task);
305 }
306
307 void dlm_recoverd_suspend(struct dlm_ls *ls)
308 {
309         wake_up(&ls->ls_wait_general);
310         mutex_lock(&ls->ls_recoverd_active);
311 }
312
313 void dlm_recoverd_resume(struct dlm_ls *ls)
314 {
315         mutex_unlock(&ls->ls_recoverd_active);
316 }
317