1 // SPDX-License-Identifier: GPL-2.0-only
2 /******************************************************************************
3 *******************************************************************************
4 **
5 ** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
6 ** Copyright (C) 2004-2011 Red Hat, Inc. All rights reserved.
7 **
8 **
9 *******************************************************************************
10 ******************************************************************************/
11
12 #include "dlm_internal.h"
13 #include "lockspace.h"
14 #include "member.h"
15 #include "dir.h"
16 #include "ast.h"
17 #include "recover.h"
18 #include "lowcomms.h"
19 #include "lock.h"
20 #include "requestqueue.h"
21 #include "recoverd.h"
22
23
24 /* If the start for which we're re-enabling locking (seq) has been superseded
25 by a newer stop (ls_recover_seq), we need to leave locking disabled.
26
27 We suspend dlm_recv threads here to avoid the race where dlm_recv a) sees
28 locking stopped and b) adds a message to the requestqueue, but dlm_recoverd
29 enables locking and clears the requestqueue between a and b. */
30
enable_locking(struct dlm_ls * ls,uint64_t seq)31 static int enable_locking(struct dlm_ls *ls, uint64_t seq)
32 {
33 int error = -EINTR;
34
35 down_write(&ls->ls_recv_active);
36
37 spin_lock(&ls->ls_recover_lock);
38 if (ls->ls_recover_seq == seq) {
39 set_bit(LSFL_RUNNING, &ls->ls_flags);
40 /* unblocks processes waiting to enter the dlm */
41 up_write(&ls->ls_in_recovery);
42 clear_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
43 error = 0;
44 }
45 spin_unlock(&ls->ls_recover_lock);
46
47 up_write(&ls->ls_recv_active);
48 return error;
49 }
50
ls_recover(struct dlm_ls * ls,struct dlm_recover * rv)51 static int ls_recover(struct dlm_ls *ls, struct dlm_recover *rv)
52 {
53 unsigned long start;
54 int error, neg = 0;
55
56 log_rinfo(ls, "dlm_recover %llu", (unsigned long long)rv->seq);
57
58 mutex_lock(&ls->ls_recoverd_active);
59
60 dlm_callback_suspend(ls);
61
62 dlm_clear_toss(ls);
63
64 /*
65 * This list of root rsb's will be the basis of most of the recovery
66 * routines.
67 */
68
69 dlm_create_root_list(ls);
70
71 /*
72 * Add or remove nodes from the lockspace's ls_nodes list.
73 *
74 * Due to the fact that we must report all membership changes to lsops
75 * or midcomms layer, it is not permitted to abort ls_recover() until
76 * this is done.
77 */
78
79 error = dlm_recover_members(ls, rv, &neg);
80 if (error) {
81 log_rinfo(ls, "dlm_recover_members error %d", error);
82 goto fail;
83 }
84
85 dlm_recover_dir_nodeid(ls);
86
87 ls->ls_recover_dir_sent_res = 0;
88 ls->ls_recover_dir_sent_msg = 0;
89 ls->ls_recover_locks_in = 0;
90
91 dlm_set_recover_status(ls, DLM_RS_NODES);
92
93 error = dlm_recover_members_wait(ls);
94 if (error) {
95 log_rinfo(ls, "dlm_recover_members_wait error %d", error);
96 goto fail;
97 }
98
99 start = jiffies;
100
101 /*
102 * Rebuild our own share of the directory by collecting from all other
103 * nodes their master rsb names that hash to us.
104 */
105
106 error = dlm_recover_directory(ls);
107 if (error) {
108 log_rinfo(ls, "dlm_recover_directory error %d", error);
109 goto fail;
110 }
111
112 dlm_set_recover_status(ls, DLM_RS_DIR);
113
114 error = dlm_recover_directory_wait(ls);
115 if (error) {
116 log_rinfo(ls, "dlm_recover_directory_wait error %d", error);
117 goto fail;
118 }
119
120 log_rinfo(ls, "dlm_recover_directory %u out %u messages",
121 ls->ls_recover_dir_sent_res, ls->ls_recover_dir_sent_msg);
122
123 /*
124 * We may have outstanding operations that are waiting for a reply from
125 * a failed node. Mark these to be resent after recovery. Unlock and
126 * cancel ops can just be completed.
127 */
128
129 dlm_recover_waiters_pre(ls);
130
131 if (dlm_recovery_stopped(ls)) {
132 error = -EINTR;
133 goto fail;
134 }
135
136 if (neg || dlm_no_directory(ls)) {
137 /*
138 * Clear lkb's for departed nodes.
139 */
140
141 dlm_recover_purge(ls);
142
143 /*
144 * Get new master nodeid's for rsb's that were mastered on
145 * departed nodes.
146 */
147
148 error = dlm_recover_masters(ls);
149 if (error) {
150 log_rinfo(ls, "dlm_recover_masters error %d", error);
151 goto fail;
152 }
153
154 /*
155 * Send our locks on remastered rsb's to the new masters.
156 */
157
158 error = dlm_recover_locks(ls);
159 if (error) {
160 log_rinfo(ls, "dlm_recover_locks error %d", error);
161 goto fail;
162 }
163
164 dlm_set_recover_status(ls, DLM_RS_LOCKS);
165
166 error = dlm_recover_locks_wait(ls);
167 if (error) {
168 log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
169 goto fail;
170 }
171
172 log_rinfo(ls, "dlm_recover_locks %u in",
173 ls->ls_recover_locks_in);
174
175 /*
176 * Finalize state in master rsb's now that all locks can be
177 * checked. This includes conversion resolution and lvb
178 * settings.
179 */
180
181 dlm_recover_rsbs(ls);
182 } else {
183 /*
184 * Other lockspace members may be going through the "neg" steps
185 * while also adding us to the lockspace, in which case they'll
186 * be doing the recover_locks (RS_LOCKS) barrier.
187 */
188 dlm_set_recover_status(ls, DLM_RS_LOCKS);
189
190 error = dlm_recover_locks_wait(ls);
191 if (error) {
192 log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
193 goto fail;
194 }
195 }
196
197 dlm_release_root_list(ls);
198
199 /*
200 * Purge directory-related requests that are saved in requestqueue.
201 * All dir requests from before recovery are invalid now due to the dir
202 * rebuild and will be resent by the requesting nodes.
203 */
204
205 dlm_purge_requestqueue(ls);
206
207 dlm_set_recover_status(ls, DLM_RS_DONE);
208
209 error = dlm_recover_done_wait(ls);
210 if (error) {
211 log_rinfo(ls, "dlm_recover_done_wait error %d", error);
212 goto fail;
213 }
214
215 dlm_clear_members_gone(ls);
216
217 dlm_adjust_timeouts(ls);
218
219 dlm_callback_resume(ls);
220
221 error = enable_locking(ls, rv->seq);
222 if (error) {
223 log_rinfo(ls, "enable_locking error %d", error);
224 goto fail;
225 }
226
227 error = dlm_process_requestqueue(ls);
228 if (error) {
229 log_rinfo(ls, "dlm_process_requestqueue error %d", error);
230 goto fail;
231 }
232
233 error = dlm_recover_waiters_post(ls);
234 if (error) {
235 log_rinfo(ls, "dlm_recover_waiters_post error %d", error);
236 goto fail;
237 }
238
239 dlm_recover_grant(ls);
240
241 log_rinfo(ls, "dlm_recover %llu generation %u done: %u ms",
242 (unsigned long long)rv->seq, ls->ls_generation,
243 jiffies_to_msecs(jiffies - start));
244 mutex_unlock(&ls->ls_recoverd_active);
245
246 return 0;
247
248 fail:
249 dlm_release_root_list(ls);
250 mutex_unlock(&ls->ls_recoverd_active);
251
252 return error;
253 }
254
255 /* The dlm_ls_start() that created the rv we take here may already have been
256 stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
257 flag set. */
258
do_ls_recovery(struct dlm_ls * ls)259 static void do_ls_recovery(struct dlm_ls *ls)
260 {
261 struct dlm_recover *rv = NULL;
262 int error;
263
264 spin_lock(&ls->ls_recover_lock);
265 rv = ls->ls_recover_args;
266 ls->ls_recover_args = NULL;
267 if (rv && ls->ls_recover_seq == rv->seq)
268 clear_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
269 spin_unlock(&ls->ls_recover_lock);
270
271 if (rv) {
272 error = ls_recover(ls, rv);
273 switch (error) {
274 case 0:
275 ls->ls_recovery_result = 0;
276 complete(&ls->ls_recovery_done);
277
278 dlm_lsop_recover_done(ls);
279 break;
280 case -EINTR:
281 /* if recovery was interrupted -EINTR we wait for the next
282 * ls_recover() iteration until it hopefully succeeds.
283 */
284 log_rinfo(ls, "%s %llu interrupted and should be queued to run again",
285 __func__, (unsigned long long)rv->seq);
286 break;
287 default:
288 log_rinfo(ls, "%s %llu error %d", __func__,
289 (unsigned long long)rv->seq, error);
290
291 /* let new_lockspace() get aware of critical error */
292 ls->ls_recovery_result = error;
293 complete(&ls->ls_recovery_done);
294 break;
295 }
296
297 kfree(rv->nodes);
298 kfree(rv);
299 }
300 }
301
dlm_recoverd(void * arg)302 static int dlm_recoverd(void *arg)
303 {
304 struct dlm_ls *ls;
305
306 ls = dlm_find_lockspace_local(arg);
307 if (!ls) {
308 log_print("dlm_recoverd: no lockspace %p", arg);
309 return -1;
310 }
311
312 down_write(&ls->ls_in_recovery);
313 set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
314 wake_up(&ls->ls_recover_lock_wait);
315
316 while (1) {
317 /*
318 * We call kthread_should_stop() after set_current_state().
319 * This is because it works correctly if kthread_stop() is
320 * called just before set_current_state().
321 */
322 set_current_state(TASK_INTERRUPTIBLE);
323 if (kthread_should_stop()) {
324 set_current_state(TASK_RUNNING);
325 break;
326 }
327 if (!test_bit(LSFL_RECOVER_WORK, &ls->ls_flags) &&
328 !test_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
329 if (kthread_should_stop())
330 break;
331 schedule();
332 }
333 set_current_state(TASK_RUNNING);
334
335 if (test_and_clear_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
336 down_write(&ls->ls_in_recovery);
337 set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
338 wake_up(&ls->ls_recover_lock_wait);
339 }
340
341 if (test_and_clear_bit(LSFL_RECOVER_WORK, &ls->ls_flags))
342 do_ls_recovery(ls);
343 }
344
345 if (test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags))
346 up_write(&ls->ls_in_recovery);
347
348 dlm_put_lockspace(ls);
349 return 0;
350 }
351
dlm_recoverd_start(struct dlm_ls * ls)352 int dlm_recoverd_start(struct dlm_ls *ls)
353 {
354 struct task_struct *p;
355 int error = 0;
356
357 p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
358 if (IS_ERR(p))
359 error = PTR_ERR(p);
360 else
361 ls->ls_recoverd_task = p;
362 return error;
363 }
364
dlm_recoverd_stop(struct dlm_ls * ls)365 void dlm_recoverd_stop(struct dlm_ls *ls)
366 {
367 kthread_stop(ls->ls_recoverd_task);
368 }
369
dlm_recoverd_suspend(struct dlm_ls * ls)370 void dlm_recoverd_suspend(struct dlm_ls *ls)
371 {
372 wake_up(&ls->ls_wait_general);
373 mutex_lock(&ls->ls_recoverd_active);
374 }
375
dlm_recoverd_resume(struct dlm_ls * ls)376 void dlm_recoverd_resume(struct dlm_ls *ls)
377 {
378 mutex_unlock(&ls->ls_recoverd_active);
379 }
380
381