xref: /linux/fs/dlm/recoverd.c (revision 1c4b5ecb7ea190fa3e9f9d6891e6c90b60e04f24)
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24 /* If the start for which we're re-enabling locking (seq) has been superseded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 
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 
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 
75 	error = dlm_recover_members(ls, rv, &neg);
76 	if (error) {
77 		log_rinfo(ls, "dlm_recover_members error %d", error);
78 		goto fail;
79 	}
80 
81 	dlm_recover_dir_nodeid(ls);
82 
83 	ls->ls_recover_dir_sent_res = 0;
84 	ls->ls_recover_dir_sent_msg = 0;
85 	ls->ls_recover_locks_in = 0;
86 
87 	dlm_set_recover_status(ls, DLM_RS_NODES);
88 
89 	error = dlm_recover_members_wait(ls);
90 	if (error) {
91 		log_rinfo(ls, "dlm_recover_members_wait error %d", error);
92 		goto fail;
93 	}
94 
95 	start = jiffies;
96 
97 	/*
98 	 * Rebuild our own share of the directory by collecting from all other
99 	 * nodes their master rsb names that hash to us.
100 	 */
101 
102 	error = dlm_recover_directory(ls);
103 	if (error) {
104 		log_rinfo(ls, "dlm_recover_directory error %d", error);
105 		goto fail;
106 	}
107 
108 	dlm_set_recover_status(ls, DLM_RS_DIR);
109 
110 	error = dlm_recover_directory_wait(ls);
111 	if (error) {
112 		log_rinfo(ls, "dlm_recover_directory_wait error %d", error);
113 		goto fail;
114 	}
115 
116 	log_rinfo(ls, "dlm_recover_directory %u out %u messages",
117 		  ls->ls_recover_dir_sent_res, ls->ls_recover_dir_sent_msg);
118 
119 	/*
120 	 * We may have outstanding operations that are waiting for a reply from
121 	 * a failed node.  Mark these to be resent after recovery.  Unlock and
122 	 * cancel ops can just be completed.
123 	 */
124 
125 	dlm_recover_waiters_pre(ls);
126 
127 	if (dlm_recovery_stopped(ls)) {
128 		error = -EINTR;
129 		goto fail;
130 	}
131 
132 	if (neg || dlm_no_directory(ls)) {
133 		/*
134 		 * Clear lkb's for departed nodes.
135 		 */
136 
137 		dlm_recover_purge(ls);
138 
139 		/*
140 		 * Get new master nodeid's for rsb's that were mastered on
141 		 * departed nodes.
142 		 */
143 
144 		error = dlm_recover_masters(ls);
145 		if (error) {
146 			log_rinfo(ls, "dlm_recover_masters error %d", error);
147 			goto fail;
148 		}
149 
150 		/*
151 		 * Send our locks on remastered rsb's to the new masters.
152 		 */
153 
154 		error = dlm_recover_locks(ls);
155 		if (error) {
156 			log_rinfo(ls, "dlm_recover_locks error %d", error);
157 			goto fail;
158 		}
159 
160 		dlm_set_recover_status(ls, DLM_RS_LOCKS);
161 
162 		error = dlm_recover_locks_wait(ls);
163 		if (error) {
164 			log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
165 			goto fail;
166 		}
167 
168 		log_rinfo(ls, "dlm_recover_locks %u in",
169 			  ls->ls_recover_locks_in);
170 
171 		/*
172 		 * Finalize state in master rsb's now that all locks can be
173 		 * checked.  This includes conversion resolution and lvb
174 		 * settings.
175 		 */
176 
177 		dlm_recover_rsbs(ls);
178 	} else {
179 		/*
180 		 * Other lockspace members may be going through the "neg" steps
181 		 * while also adding us to the lockspace, in which case they'll
182 		 * be doing the recover_locks (RS_LOCKS) barrier.
183 		 */
184 		dlm_set_recover_status(ls, DLM_RS_LOCKS);
185 
186 		error = dlm_recover_locks_wait(ls);
187 		if (error) {
188 			log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
189 			goto fail;
190 		}
191 	}
192 
193 	dlm_release_root_list(ls);
194 
195 	/*
196 	 * Purge directory-related requests that are saved in requestqueue.
197 	 * All dir requests from before recovery are invalid now due to the dir
198 	 * rebuild and will be resent by the requesting nodes.
199 	 */
200 
201 	dlm_purge_requestqueue(ls);
202 
203 	dlm_set_recover_status(ls, DLM_RS_DONE);
204 
205 	error = dlm_recover_done_wait(ls);
206 	if (error) {
207 		log_rinfo(ls, "dlm_recover_done_wait error %d", error);
208 		goto fail;
209 	}
210 
211 	dlm_clear_members_gone(ls);
212 
213 	dlm_adjust_timeouts(ls);
214 
215 	dlm_callback_resume(ls);
216 
217 	error = enable_locking(ls, rv->seq);
218 	if (error) {
219 		log_rinfo(ls, "enable_locking error %d", error);
220 		goto fail;
221 	}
222 
223 	error = dlm_process_requestqueue(ls);
224 	if (error) {
225 		log_rinfo(ls, "dlm_process_requestqueue error %d", error);
226 		goto fail;
227 	}
228 
229 	error = dlm_recover_waiters_post(ls);
230 	if (error) {
231 		log_rinfo(ls, "dlm_recover_waiters_post error %d", error);
232 		goto fail;
233 	}
234 
235 	dlm_recover_grant(ls);
236 
237 	log_rinfo(ls, "dlm_recover %llu generation %u done: %u ms",
238 		  (unsigned long long)rv->seq, ls->ls_generation,
239 		  jiffies_to_msecs(jiffies - start));
240 	mutex_unlock(&ls->ls_recoverd_active);
241 
242 	dlm_lsop_recover_done(ls);
243 	return 0;
244 
245  fail:
246 	dlm_release_root_list(ls);
247 	log_rinfo(ls, "dlm_recover %llu error %d",
248 		  (unsigned long long)rv->seq, error);
249 	mutex_unlock(&ls->ls_recoverd_active);
250 	return error;
251 }
252 
253 /* The dlm_ls_start() that created the rv we take here may already have been
254    stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
255    flag set. */
256 
257 static void do_ls_recovery(struct dlm_ls *ls)
258 {
259 	struct dlm_recover *rv = NULL;
260 
261 	spin_lock(&ls->ls_recover_lock);
262 	rv = ls->ls_recover_args;
263 	ls->ls_recover_args = NULL;
264 	if (rv && ls->ls_recover_seq == rv->seq)
265 		clear_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
266 	spin_unlock(&ls->ls_recover_lock);
267 
268 	if (rv) {
269 		ls_recover(ls, rv);
270 		kfree(rv->nodes);
271 		kfree(rv);
272 	}
273 }
274 
275 static int dlm_recoverd(void *arg)
276 {
277 	struct dlm_ls *ls;
278 
279 	ls = dlm_find_lockspace_local(arg);
280 	if (!ls) {
281 		log_print("dlm_recoverd: no lockspace %p", arg);
282 		return -1;
283 	}
284 
285 	down_write(&ls->ls_in_recovery);
286 	set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
287 	wake_up(&ls->ls_recover_lock_wait);
288 
289 	while (1) {
290 		/*
291 		 * We call kthread_should_stop() after set_current_state().
292 		 * This is because it works correctly if kthread_stop() is
293 		 * called just before set_current_state().
294 		 */
295 		set_current_state(TASK_INTERRUPTIBLE);
296 		if (kthread_should_stop()) {
297 			set_current_state(TASK_RUNNING);
298 			break;
299 		}
300 		if (!test_bit(LSFL_RECOVER_WORK, &ls->ls_flags) &&
301 		    !test_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
302 			if (kthread_should_stop())
303 				break;
304 			schedule();
305 		}
306 		set_current_state(TASK_RUNNING);
307 
308 		if (test_and_clear_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
309 			down_write(&ls->ls_in_recovery);
310 			set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
311 			wake_up(&ls->ls_recover_lock_wait);
312 		}
313 
314 		if (test_and_clear_bit(LSFL_RECOVER_WORK, &ls->ls_flags))
315 			do_ls_recovery(ls);
316 	}
317 
318 	if (test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags))
319 		up_write(&ls->ls_in_recovery);
320 
321 	dlm_put_lockspace(ls);
322 	return 0;
323 }
324 
325 int dlm_recoverd_start(struct dlm_ls *ls)
326 {
327 	struct task_struct *p;
328 	int error = 0;
329 
330 	p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
331 	if (IS_ERR(p))
332 		error = PTR_ERR(p);
333 	else
334                 ls->ls_recoverd_task = p;
335 	return error;
336 }
337 
338 void dlm_recoverd_stop(struct dlm_ls *ls)
339 {
340 	kthread_stop(ls->ls_recoverd_task);
341 }
342 
343 void dlm_recoverd_suspend(struct dlm_ls *ls)
344 {
345 	wake_up(&ls->ls_wait_general);
346 	mutex_lock(&ls->ls_recoverd_active);
347 }
348 
349 void dlm_recoverd_resume(struct dlm_ls *ls)
350 {
351 	mutex_unlock(&ls->ls_recoverd_active);
352 }
353 
354