fs/dlm/recoverd.c

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  24 /* If the start for which we're re-enabling locking (seq) has been superseded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