1 // SPDX-License-Identifier: GPL-2.0-only
2 /******************************************************************************
3 *******************************************************************************
5 ** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
6 ** Copyright (C) 2004-2011 Red Hat, Inc. All rights reserved.
9 *******************************************************************************
10 ******************************************************************************/
12 #include "dlm_internal.h"
13 #include "lockspace.h"
20 #include "requestqueue.h"
24 /* If the start for which we're re-enabling locking (seq) has been superseded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.
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. */
31 static int enable_locking(struct dlm_ls *ls, uint64_t seq)
35 down_write(&ls->ls_recv_active);
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);
45 spin_unlock(&ls->ls_recover_lock);
47 up_write(&ls->ls_recv_active);
51 static int ls_recover(struct dlm_ls *ls, struct dlm_recover *rv)
56 log_rinfo(ls, "dlm_recover %llu", (unsigned long long)rv->seq);
58 mutex_lock(&ls->ls_recoverd_active);
60 dlm_callback_suspend(ls);
65 * This list of root rsb's will be the basis of most of the recovery
69 dlm_create_root_list(ls);
72 * Add or remove nodes from the lockspace's ls_nodes list.
75 error = dlm_recover_members(ls, rv, &neg);
77 log_rinfo(ls, "dlm_recover_members error %d", error);
81 dlm_recover_dir_nodeid(ls);
83 ls->ls_recover_dir_sent_res = 0;
84 ls->ls_recover_dir_sent_msg = 0;
85 ls->ls_recover_locks_in = 0;
87 dlm_set_recover_status(ls, DLM_RS_NODES);
89 error = dlm_recover_members_wait(ls);
91 log_rinfo(ls, "dlm_recover_members_wait error %d", error);
98 * Rebuild our own share of the directory by collecting from all other
99 * nodes their master rsb names that hash to us.
102 error = dlm_recover_directory(ls);
104 log_rinfo(ls, "dlm_recover_directory error %d", error);
108 dlm_set_recover_status(ls, DLM_RS_DIR);
110 error = dlm_recover_directory_wait(ls);
112 log_rinfo(ls, "dlm_recover_directory_wait error %d", error);
116 log_rinfo(ls, "dlm_recover_directory %u out %u messages",
117 ls->ls_recover_dir_sent_res, ls->ls_recover_dir_sent_msg);
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.
125 dlm_recover_waiters_pre(ls);
127 if (dlm_recovery_stopped(ls)) {
132 if (neg || dlm_no_directory(ls)) {
134 * Clear lkb's for departed nodes.
137 dlm_recover_purge(ls);
140 * Get new master nodeid's for rsb's that were mastered on
144 error = dlm_recover_masters(ls);
146 log_rinfo(ls, "dlm_recover_masters error %d", error);
151 * Send our locks on remastered rsb's to the new masters.
154 error = dlm_recover_locks(ls);
156 log_rinfo(ls, "dlm_recover_locks error %d", error);
160 dlm_set_recover_status(ls, DLM_RS_LOCKS);
162 error = dlm_recover_locks_wait(ls);
164 log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
168 log_rinfo(ls, "dlm_recover_locks %u in",
169 ls->ls_recover_locks_in);
172 * Finalize state in master rsb's now that all locks can be
173 * checked. This includes conversion resolution and lvb
177 dlm_recover_rsbs(ls);
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.
184 dlm_set_recover_status(ls, DLM_RS_LOCKS);
186 error = dlm_recover_locks_wait(ls);
188 log_rinfo(ls, "dlm_recover_locks_wait error %d", error);
193 dlm_release_root_list(ls);
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.
201 dlm_purge_requestqueue(ls);
203 dlm_set_recover_status(ls, DLM_RS_DONE);
205 error = dlm_recover_done_wait(ls);
207 log_rinfo(ls, "dlm_recover_done_wait error %d", error);
211 dlm_clear_members_gone(ls);
213 dlm_adjust_timeouts(ls);
215 dlm_callback_resume(ls);
217 error = enable_locking(ls, rv->seq);
219 log_rinfo(ls, "enable_locking error %d", error);
223 error = dlm_process_requestqueue(ls);
225 log_rinfo(ls, "dlm_process_requestqueue error %d", error);
229 error = dlm_recover_waiters_post(ls);
231 log_rinfo(ls, "dlm_recover_waiters_post error %d", error);
235 dlm_recover_grant(ls);
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);
242 dlm_lsop_recover_done(ls);
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);
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
257 static void do_ls_recovery(struct dlm_ls *ls)
259 struct dlm_recover *rv = NULL;
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);
275 static int dlm_recoverd(void *arg)
279 ls = dlm_find_lockspace_local(arg);
281 log_print("dlm_recoverd: no lockspace %p", arg);
285 down_write(&ls->ls_in_recovery);
286 set_bit(LSFL_RECOVER_LOCK, &ls->ls_flags);
287 wake_up(&ls->ls_recover_lock_wait);
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().
295 set_current_state(TASK_INTERRUPTIBLE);
296 if (kthread_should_stop()) {
297 set_current_state(TASK_RUNNING);
300 if (!test_bit(LSFL_RECOVER_WORK, &ls->ls_flags) &&
301 !test_bit(LSFL_RECOVER_DOWN, &ls->ls_flags)) {
302 if (kthread_should_stop())
306 set_current_state(TASK_RUNNING);
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);
314 if (test_and_clear_bit(LSFL_RECOVER_WORK, &ls->ls_flags))
318 if (test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags))
319 up_write(&ls->ls_in_recovery);
321 dlm_put_lockspace(ls);
325 int dlm_recoverd_start(struct dlm_ls *ls)
327 struct task_struct *p;
330 p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
334 ls->ls_recoverd_task = p;
338 void dlm_recoverd_stop(struct dlm_ls *ls)
340 kthread_stop(ls->ls_recoverd_task);
343 void dlm_recoverd_suspend(struct dlm_ls *ls)
345 wake_up(&ls->ls_wait_general);
346 mutex_lock(&ls->ls_recoverd_active);
349 void dlm_recoverd_resume(struct dlm_ls *ls)
351 mutex_unlock(&ls->ls_recoverd_active);