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) 2004-2005 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  26 /* If the start for which we're re-enabling locking (seq) has been superseded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", 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_error(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_error(ls, "recover_directory failed %d", error);
  93                 goto fail;
  94         }
  95
  96         /*
  97          * Purge directory-related requests that are saved in requestqueue.
  98          * All dir requests from before recovery are invalid now due to the dir
  99          * rebuild and will be resent by the requesting nodes.
 100          */
 101
 102         dlm_purge_requestqueue(ls);
 103
 104         /*
 105          * Wait for all nodes to complete directory rebuild.
 106          */
 107
 108         error = dlm_recover_directory_wait(ls);
 109         if (error) {
 110                 log_error(ls, "recover_directory_wait failed %d", error);
 111                 goto fail;
 112         }
 113
 114         /*
 115          * We may have outstanding operations that are waiting for a reply from
 116          * a failed node.  Mark these to be resent after recovery.  Unlock and
 117          * cancel ops can just be completed.
 118          */
 119
 120         dlm_recover_waiters_pre(ls);
 121
 122         error = dlm_recovery_stopped(ls);
 123         if (error)
 124                 goto fail;
 125
 126         if (neg || dlm_no_directory(ls)) {
 127                 /*
 128                  * Clear lkb's for departed nodes.
 129                  */
 130
 131                 dlm_purge_locks(ls);
 132
 133                 /*
 134                  * Get new master nodeid's for rsb's that were mastered on
 135                  * departed nodes.
 136                  */
 137
 138                 error = dlm_recover_masters(ls);
 139                 if (error) {
 140                         log_error(ls, "recover_masters failed %d", error);
 141                         goto fail;
 142                 }
 143
 144                 /*
 145                  * Send our locks on remastered rsb's to the new masters.
 146                  */
 147
 148                 error = dlm_recover_locks(ls);
 149                 if (error) {
 150                         log_error(ls, "recover_locks failed %d", error);
 151                         goto fail;
 152                 }
 153
 154                 error = dlm_recover_locks_wait(ls);
 155                 if (error) {
 156                         log_error(ls, "recover_locks_wait failed %d", error);
 157                         goto fail;
 158                 }
 159
 160                 /*
 161                  * Finalize state in master rsb's now that all locks can be
 162                  * checked.  This includes conversion resolution and lvb
 163                  * settings.
 164                  */
 165
 166                 dlm_recover_rsbs(ls);
 167         }
 168
 169         dlm_release_root_list(ls);
 170
 171         dlm_set_recover_status(ls, DLM_RS_DONE);
 172         error = dlm_recover_done_wait(ls);
 173         if (error) {
 174                 log_error(ls, "recover_done_wait failed %d", error);
 175                 goto fail;
 176         }
 177
 178         dlm_clear_members_gone(ls);
 179
 180         error = enable_locking(ls, rv->seq);
 181         if (error) {
 182                 log_error(ls, "enable_locking failed %d", error);
 183                 goto fail;
 184         }
 185
 186         error = dlm_process_requestqueue(ls);
 187         if (error) {
 188                 log_error(ls, "process_requestqueue failed %d", error);
 189                 goto fail;
 190         }
 191
 192         error = dlm_recover_waiters_post(ls);
 193         if (error) {
 194                 log_error(ls, "recover_waiters_post failed %d", error);
 195                 goto fail;
 196         }
 197
 198         dlm_grant_after_purge(ls);
 199
 200         dlm_astd_wake();
 201
 202         log_debug(ls, "recover %llx done: %u ms", rv->seq,
 203                   jiffies_to_msecs(jiffies - start));
 204         mutex_unlock(&ls->ls_recoverd_active);
 205
 206         return 0;
 207
 208  fail:
 209         dlm_release_root_list(ls);
 210         log_debug(ls, "recover %llx error %d", rv->seq, error);
 211         mutex_unlock(&ls->ls_recoverd_active);
 212         return error;
 213 }
 214
 215 static void do_ls_recovery(struct dlm_ls *ls)
 216 {
 217         struct dlm_recover *rv = NULL;
 218
 219         spin_lock(&ls->ls_recover_lock);
 220         rv = ls->ls_recover_args;
 221         ls->ls_recover_args = NULL;
 222         clear_bit(LSFL_RECOVERY_STOP, &ls->ls_flags);
 223         spin_unlock(&ls->ls_recover_lock);
 224
 225         if (rv) {
 226                 ls_recover(ls, rv);
 227                 kfree(rv->nodeids);
 228                 kfree(rv);
 229         }
 230 }
 231
 232 static int dlm_recoverd(void *arg)
 233 {
 234         struct dlm_ls *ls;
 235
 236         ls = dlm_find_lockspace_local(arg);
 237         if (!ls) {
 238                 log_print("dlm_recoverd: no lockspace %p", arg);
 239                 return -1;
 240         }
 241
 242         while (!kthread_should_stop()) {
 243                 set_current_state(TASK_INTERRUPTIBLE);
 244                 if (!test_bit(LSFL_WORK, &ls->ls_flags))
 245                         schedule();
 246                 set_current_state(TASK_RUNNING);
 247
 248                 if (test_and_clear_bit(LSFL_WORK, &ls->ls_flags))
 249                         do_ls_recovery(ls);
 250         }
 251
 252         dlm_put_lockspace(ls);
 253         return 0;
 254 }
 255
 256 void dlm_recoverd_kick(struct dlm_ls *ls)
 257 {
 258         set_bit(LSFL_WORK, &ls->ls_flags);
 259         wake_up_process(ls->ls_recoverd_task);
 260 }
 261
 262 int dlm_recoverd_start(struct dlm_ls *ls)
 263 {
 264         struct task_struct *p;
 265         int error = 0;
 266
 267         p = kthread_run(dlm_recoverd, ls, "dlm_recoverd");
 268         if (IS_ERR(p))
 269                 error = PTR_ERR(p);
 270         else
 271                 ls->ls_recoverd_task = p;
 272         return error;
 273 }
 274
 275 void dlm_recoverd_stop(struct dlm_ls *ls)
 276 {
 277         kthread_stop(ls->ls_recoverd_task);
 278 }
 279
 280 void dlm_recoverd_suspend(struct dlm_ls *ls)
 281 {
 282         wake_up(&ls->ls_wait_general);
 283         mutex_lock(&ls->ls_recoverd_active);
 284 }
 285
 286 void dlm_recoverd_resume(struct dlm_ls *ls)
 287 {
 288         mutex_unlock(&ls->ls_recoverd_active);
 289 }
 290