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  26 /* If the start for which we're re-enabling locking (seq) has been superseded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, "dlm_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          */
  80
  81         error = dlm_recover_members(ls, rv, &neg);
  82         if (error) {
  83                 log_debug(ls, "dlm_recover_members error %d", error);
  84                 goto fail;
  85         }
  86
  87         dlm_set_recover_status(ls, DLM_RS_NODES);
  88
  89         error = dlm_recover_members_wait(ls);
  90         if (error) {
  91                 log_debug(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_debug(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_debug(ls, "dlm_recover_directory_wait error %d", error);
 113                 goto fail;
 114         }
 115
 116         /*
 117          * We may have outstanding operations that are waiting for a reply from
 118          * a failed node.  Mark these to be resent after recovery.  Unlock and
 119          * cancel ops can just be completed.
 120          */
 121
 122         dlm_recover_waiters_pre(ls);
 123
 124         error = dlm_recovery_stopped(ls);
 125         if (error)
 126                 goto fail;
 127
 128         if (neg || dlm_no_directory(ls)) {
 129                 /*
 130                  * Clear lkb's for departed nodes.
 131                  */
 132
 133                 dlm_purge_locks(ls);
 134
 135                 /*
 136                  * Get new master nodeid's for rsb's that were mastered on
 137                  * departed nodes.
 138                  */
 139
 140                 error = dlm_recover_masters(ls);
 141                 if (error) {
 142                         log_debug(ls, "dlm_recover_masters error %d", error);
 143                         goto fail;
 144                 }
 145
 146                 /*
 147                  * Send our locks on remastered rsb's to the new masters.
 148                  */
 149
 150                 error = dlm_recover_locks(ls);
 151                 if (error) {
 152                         log_debug(ls, "dlm_recover_locks error %d", error);
 153                         goto fail;
 154                 }
 155
 156                 dlm_set_recover_status(ls, DLM_RS_LOCKS);
 157
 158                 error = dlm_recover_locks_wait(ls);
 159                 if (error) {
 160                         log_debug(ls, "dlm_recover_locks_wait error %d", error);
 161                         goto fail;
 162                 }
 163
 164                 /*
 165                  * Finalize state in master rsb's now that all locks can be
 166                  * checked.  This includes conversion resolution and lvb
 167                  * settings.
 168                  */
 169
 170                 dlm_recover_rsbs(ls);
 171         } else {
 172                 /*
 173                  * Other lockspace members may be going through the "neg" steps
 174                  * while also adding us to the lockspace, in which case they'll
 175                  * be doing the recover_locks (RS_LOCKS) barrier.
 176                  */
 177                 dlm_set_recover_status(ls, DLM_RS_LOCKS);
 178
 179                 error = dlm_recover_locks_wait(ls);
 180                 if (error) {
 181                         log_debug(ls, "dlm_recover_locks_wait error %d", error);
 182                         goto fail;
 183                 }
 184         }
 185
 186         dlm_release_root_list(ls);
 187
 188         /*
 189          * Purge directory-related requests that are saved in requestqueue.
 190          * All dir requests from before recovery are invalid now due to the dir
 191          * rebuild and will be resent by the requesting nodes.
 192          */
 193
 194         dlm_purge_requestqueue(ls);
 195
 196         dlm_set_recover_status(ls, DLM_RS_DONE);
 197
 198         error = dlm_recover_done_wait(ls);
 199         if (error) {
 200                 log_debug(ls, "dlm_recover_done_wait error %d", error);
 201                 goto fail;
 202         }
 203
 204         dlm_clear_members_gone(ls);
 205
 206         dlm_adjust_timeouts(ls);
 207
 208         dlm_callback_resume(ls);
 209
 210         error = enable_locking(ls, rv->seq);
 211         if (error) {
 212                 log_debug(ls, "enable_locking error %d", error);
 213                 goto fail;
 214         }
 215
 216         error = dlm_process_requestqueue(ls);
 217         if (error) {
 218                 log_debug(ls, "dlm_process_requestqueue error %d", error);
 219                 goto fail;
 220         }
 221
 222         error = dlm_recover_waiters_post(ls);
 223         if (error) {
 224                 log_debug(ls, "dlm_recover_waiters_post error %d", error);
 225                 goto fail;
 226         }
 227
 228         dlm_grant_after_purge(ls);
 229
 230         log_debug(ls, "dlm_recover %llx done: %u ms",
 231                   (unsigned long long)rv->seq,
 232                   jiffies_to_msecs(jiffies - start));
 233         mutex_unlock(&ls->ls_recoverd_active);
 234
 235         return 0;
 236
 237  fail:
 238         dlm_release_root_list(ls);
 239         log_debug(ls, "dlm_recover %llx error %d",
 240                   (unsigned long long)rv->seq, error);
 241         mutex_unlock(&ls->ls_recoverd_active);
 242         return error;
 243 }
 244
 245 /* The dlm_ls_start() that created the rv we take here may already have been
 246    stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
 247    flag set. */
 248
 249 static void do_ls_recovery(struct dlm_ls *ls)
 250 {
 251         struct dlm_recover *rv = NULL;
 252
 253         spin_lock(&ls->ls_recover_lock);
 254         rv = ls->ls_recover_args;
 255         ls->ls_recover_args = NULL;
 256         if (rv && ls->ls_recover_seq == rv->seq)
 257                 clear_bit(LSFL_RECOVERY_STOP, &ls->ls_flags);
 258         spin_unlock(&ls->ls_recover_lock);
 259
 260         if (rv) {
 261                 ls_recover(ls, rv);
 262                 kfree(rv->nodeids);
 263                 kfree(rv->new);
 264                 kfree(rv);
 265         }
 266 }
 267
 268 static int dlm_recoverd(void *arg)
 269 {
 270         struct dlm_ls *ls;
 271
 272         ls = dlm_find_lockspace_local(arg);
 273         if (!ls) {
 274                 log_print("dlm_recoverd: no lockspace %p", arg);
 275                 return -1;
 276         }
 277
 278         while (!kthread_should_stop()) {
 279                 set_current_state(TASK_INTERRUPTIBLE);
 280                 if (!test_bit(LSFL_WORK, &ls->ls_flags))
 281                         schedule();
 282                 set_current_state(TASK_RUNNING);
 283
 284                 if (test_and_clear_bit(LSFL_WORK, &ls->ls_flags))
 285                         do_ls_recovery(ls);
 286         }
 287
 288         dlm_put_lockspace(ls);
 289         return 0;
 290 }
 291
 292 void dlm_recoverd_kick(struct dlm_ls *ls)
 293 {
 294         set_bit(LSFL_WORK, &ls->ls_flags);
 295         wake_up_process(ls->ls_recoverd_task);
 296 }
 297
 298 int dlm_recoverd_start(struct dlm_ls *ls)
 299 {
 300         struct task_struct *p;
 301         int error = 0;
 302
 303         p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
 304         if (IS_ERR(p))
 305                 error = PTR_ERR(p);
 306         else
 307                 ls->ls_recoverd_task = p;
 308         return error;
 309 }
 310
 311 void dlm_recoverd_stop(struct dlm_ls *ls)
 312 {
 313         kthread_stop(ls->ls_recoverd_task);
 314 }
 315
 316 void dlm_recoverd_suspend(struct dlm_ls *ls)
 317 {
 318         wake_up(&ls->ls_wait_general);
 319         mutex_lock(&ls->ls_recoverd_active);
 320 }
 321
 322 void dlm_recoverd_resume(struct dlm_ls *ls)
 323 {
 324         mutex_unlock(&ls->ls_recoverd_active);
 325 }
 326