git.oblomov.eu Git - linux-2.6/blob - 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  26 /* If the start for which we're re-enabling locking (seq) has been superseded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 static int enable_locking(struct dlm_ls *ls, uint64_t seq)
  30 {
  31         int error = -EINTR;
  32
  33         spin_lock(&ls->ls_recover_lock);
  34         if (ls->ls_recover_seq == seq) {
  35                 set_bit(LSFL_RUNNING, &ls->ls_flags);
  36                 up_write(&ls->ls_in_recovery);
  37                 error = 0;
  38         }
  39         spin_unlock(&ls->ls_recover_lock);
  40         return error;
  41 }
  42
  43 static int ls_recover(struct dlm_ls *ls, struct dlm_recover *rv)
  44 {
  45         unsigned long start;
  46         int error, neg = 0;
  47
  48         log_debug(ls, "recover %llx", (unsigned long long)rv->seq);
  49
  50         mutex_lock(&ls->ls_recoverd_active);
  51
  52         /*
  53          * Suspending and resuming dlm_astd ensures that no lkb's from this ls
  54          * will be processed by dlm_astd during recovery.
  55          */
  56
  57         dlm_astd_suspend();
  58         dlm_astd_resume();
  59
  60         /*
  61          * This list of root rsb's will be the basis of most of the recovery
  62          * routines.
  63          */
  64
  65         dlm_create_root_list(ls);
  66
  67         /*
  68          * Free all the tossed rsb's so we don't have to recover them.
  69          */
  70
  71         dlm_clear_toss_list(ls);
  72
  73         /*
  74          * Add or remove nodes from the lockspace's ls_nodes list.
  75          * Also waits for all nodes to complete dlm_recover_members.
  76          */
  77
  78         error = dlm_recover_members(ls, rv, &neg);
  79         if (error) {
  80                 log_debug(ls, "recover_members failed %d", error);
  81                 goto fail;
  82         }
  83         start = jiffies;
  84
  85         /*
  86          * Rebuild our own share of the directory by collecting from all other
  87          * nodes their master rsb names that hash to us.
  88          */
  89
  90         error = dlm_recover_directory(ls);
  91         if (error) {
  92                 log_debug(ls, "recover_directory failed %d", error);
  93                 goto fail;
  94         }
  95
  96         /*
  97          * Wait for all nodes to complete directory rebuild.
  98          */
  99
 100         error = dlm_recover_directory_wait(ls);
 101         if (error) {
 102                 log_debug(ls, "recover_directory_wait failed %d", error);
 103                 goto fail;
 104         }
 105
 106         /*
 107          * We may have outstanding operations that are waiting for a reply from
 108          * a failed node.  Mark these to be resent after recovery.  Unlock and
 109          * cancel ops can just be completed.
 110          */
 111
 112         dlm_recover_waiters_pre(ls);
 113
 114         error = dlm_recovery_stopped(ls);
 115         if (error)
 116                 goto fail;
 117
 118         if (neg || dlm_no_directory(ls)) {
 119                 /*
 120                  * Clear lkb's for departed nodes.
 121                  */
 122
 123                 dlm_purge_locks(ls);
 124
 125                 /*
 126                  * Get new master nodeid's for rsb's that were mastered on
 127                  * departed nodes.
 128                  */
 129
 130                 error = dlm_recover_masters(ls);
 131                 if (error) {
 132                         log_debug(ls, "recover_masters failed %d", error);
 133                         goto fail;
 134                 }
 135
 136                 /*
 137                  * Send our locks on remastered rsb's to the new masters.
 138                  */
 139
 140                 error = dlm_recover_locks(ls);
 141                 if (error) {
 142                         log_debug(ls, "recover_locks failed %d", error);
 143                         goto fail;
 144                 }
 145
 146                 error = dlm_recover_locks_wait(ls);
 147                 if (error) {
 148                         log_debug(ls, "recover_locks_wait failed %d", error);
 149                         goto fail;
 150                 }
 151
 152                 /*
 153                  * Finalize state in master rsb's now that all locks can be
 154                  * checked.  This includes conversion resolution and lvb
 155                  * settings.
 156                  */
 157
 158                 dlm_recover_rsbs(ls);
 159         } else {
 160                 /*
 161                  * Other lockspace members may be going through the "neg" steps
 162                  * while also adding us to the lockspace, in which case they'll
 163                  * be doing the recover_locks (RS_LOCKS) barrier.
 164                  */
 165                 dlm_set_recover_status(ls, DLM_RS_LOCKS);
 166
 167                 error = dlm_recover_locks_wait(ls);
 168                 if (error) {
 169                         log_debug(ls, "recover_locks_wait failed %d", error);
 170                         goto fail;
 171                 }
 172         }
 173
 174         dlm_release_root_list(ls);
 175
 176         /*
 177          * Purge directory-related requests that are saved in requestqueue.
 178          * All dir requests from before recovery are invalid now due to the dir
 179          * rebuild and will be resent by the requesting nodes.
 180          */
 181
 182         dlm_purge_requestqueue(ls);
 183
 184         dlm_set_recover_status(ls, DLM_RS_DONE);
 185         error = dlm_recover_done_wait(ls);
 186         if (error) {
 187                 log_debug(ls, "recover_done_wait failed %d", error);
 188                 goto fail;
 189         }
 190
 191         dlm_clear_members_gone(ls);
 192
 193         dlm_adjust_timeouts(ls);
 194
 195         error = enable_locking(ls, rv->seq);
 196         if (error) {
 197                 log_debug(ls, "enable_locking failed %d", error);
 198                 goto fail;
 199         }
 200
 201         error = dlm_process_requestqueue(ls);
 202         if (error) {
 203                 log_debug(ls, "process_requestqueue failed %d", error);
 204                 goto fail;
 205         }
 206
 207         error = dlm_recover_waiters_post(ls);
 208         if (error) {
 209                 log_debug(ls, "recover_waiters_post failed %d", error);
 210                 goto fail;
 211         }
 212
 213         dlm_grant_after_purge(ls);
 214
 215         dlm_astd_wake();
 216
 217         log_debug(ls, "recover %llx done: %u ms",
 218                   (unsigned long long)rv->seq,
 219                   jiffies_to_msecs(jiffies - start));
 220         mutex_unlock(&ls->ls_recoverd_active);
 221
 222         return 0;
 223
 224  fail:
 225         dlm_release_root_list(ls);
 226         log_debug(ls, "recover %llx error %d",
 227                   (unsigned long long)rv->seq, error);
 228         mutex_unlock(&ls->ls_recoverd_active);
 229         return error;
 230 }
 231
 232 /* The dlm_ls_start() that created the rv we take here may already have been
 233    stopped via dlm_ls_stop(); in that case we need to leave the RECOVERY_STOP
 234    flag set. */
 235
 236 static void do_ls_recovery(struct dlm_ls *ls)
 237 {
 238         struct dlm_recover *rv = NULL;
 239
 240         spin_lock(&ls->ls_recover_lock);
 241         rv = ls->ls_recover_args;
 242         ls->ls_recover_args = NULL;
 243         if (rv && ls->ls_recover_seq == rv->seq)
 244                 clear_bit(LSFL_RECOVERY_STOP, &ls->ls_flags);
 245         spin_unlock(&ls->ls_recover_lock);
 246
 247         if (rv) {
 248                 ls_recover(ls, rv);
 249                 kfree(rv->nodeids);
 250                 kfree(rv);
 251         }
 252 }
 253
 254 static int dlm_recoverd(void *arg)
 255 {
 256         struct dlm_ls *ls;
 257
 258         ls = dlm_find_lockspace_local(arg);
 259         if (!ls) {
 260                 log_print("dlm_recoverd: no lockspace %p", arg);
 261                 return -1;
 262         }
 263
 264         while (!kthread_should_stop()) {
 265                 set_current_state(TASK_INTERRUPTIBLE);
 266                 if (!test_bit(LSFL_WORK, &ls->ls_flags))
 267                         schedule();
 268                 set_current_state(TASK_RUNNING);
 269
 270                 if (test_and_clear_bit(LSFL_WORK, &ls->ls_flags))
 271                         do_ls_recovery(ls);
 272         }
 273
 274         dlm_put_lockspace(ls);
 275         return 0;
 276 }
 277
 278 void dlm_recoverd_kick(struct dlm_ls *ls)
 279 {
 280         set_bit(LSFL_WORK, &ls->ls_flags);
 281         wake_up_process(ls->ls_recoverd_task);
 282 }
 283
 284 int dlm_recoverd_start(struct dlm_ls *ls)
 285 {
 286         struct task_struct *p;
 287         int error = 0;
 288
 289         p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
 290         if (IS_ERR(p))
 291                 error = PTR_ERR(p);
 292         else
 293                 ls->ls_recoverd_task = p;
 294         return error;
 295 }
 296
 297 void dlm_recoverd_stop(struct dlm_ls *ls)
 298 {
 299         kthread_stop(ls->ls_recoverd_task);
 300 }
 301
 302 void dlm_recoverd_suspend(struct dlm_ls *ls)
 303 {
 304         wake_up(&ls->ls_wait_general);
 305         mutex_lock(&ls->ls_recoverd_active);
 306 }
 307
 308 void dlm_recoverd_resume(struct dlm_ls *ls)
 309 {
 310         mutex_unlock(&ls->ls_recoverd_active);
 311 }
 312