SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/dlm/recover.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 "dir.h"
  17 #include "config.h"
  18 #include "ast.h"
  19 #include "memory.h"
  20 #include "rcom.h"
  21 #include "lock.h"
  22 #include "lowcomms.h"
  23 #include "member.h"
  24 #include "recover.h"
  25
  26
  27 /*
  28  * Recovery waiting routines: these functions wait for a particular reply from
  29  * a remote node, or for the remote node to report a certain status.  They need
  30  * to abort if the lockspace is stopped indicating a node has failed (perhaps
  31  * the one being waited for).
  32  */
  33
  34 /*
  35  * Wait until given function returns non-zero or lockspace is stopped
  36  * (LS_RECOVERY_STOP set due to failure of a node in ls_nodes).  When another
  37  * function thinks it could have completed the waited-on task, they should wake
  38  * up ls_wait_general to get an immediate response rather than waiting for the
  39  * timer to detect the result.  A timer wakes us up periodically while waiting
  40  * to see if we should abort due to a node failure.  This should only be called
  41  * by the dlm_recoverd thread.
  42  */
  43
  44 static void dlm_wait_timer_fn(unsigned long data)
  45 {
  46         struct dlm_ls *ls = (struct dlm_ls *) data;
  47         mod_timer(&ls->ls_timer, jiffies + (dlm_config.recover_timer * HZ));
  48         wake_up(&ls->ls_wait_general);
  49 }
  50
  51 int dlm_wait_function(struct dlm_ls *ls, int (*testfn) (struct dlm_ls *ls))
  52 {
  53         int error = 0;
  54
  55         init_timer(&ls->ls_timer);
  56         ls->ls_timer.function = dlm_wait_timer_fn;
  57         ls->ls_timer.data = (long) ls;
  58         ls->ls_timer.expires = jiffies + (dlm_config.recover_timer * HZ);
  59         add_timer(&ls->ls_timer);
  60
  61         wait_event(ls->ls_wait_general, testfn(ls) || dlm_recovery_stopped(ls));
  62         del_timer_sync(&ls->ls_timer);
  63
  64         if (dlm_recovery_stopped(ls)) {
  65                 log_debug(ls, "dlm_wait_function aborted");
  66                 error = -EINTR;
  67         }
  68         return error;
  69 }
  70
  71 /*
  72  * An efficient way for all nodes to wait for all others to have a certain
  73  * status.  The node with the lowest nodeid polls all the others for their
  74  * status (wait_status_all) and all the others poll the node with the low id
  75  * for its accumulated result (wait_status_low).  When all nodes have set
  76  * status flag X, then status flag X_ALL will be set on the low nodeid.
  77  */
  78
  79 uint32_t dlm_recover_status(struct dlm_ls *ls)
  80 {
  81         uint32_t status;
  82         spin_lock(&ls->ls_recover_lock);
  83         status = ls->ls_recover_status;
  84         spin_unlock(&ls->ls_recover_lock);
  85         return status;
  86 }
  87
  88 void dlm_set_recover_status(struct dlm_ls *ls, uint32_t status)
  89 {
  90         spin_lock(&ls->ls_recover_lock);
  91         ls->ls_recover_status |= status;
  92         spin_unlock(&ls->ls_recover_lock);
  93 }
  94
  95 static int wait_status_all(struct dlm_ls *ls, uint32_t wait_status)
  96 {
  97         struct dlm_rcom *rc = (struct dlm_rcom *) ls->ls_recover_buf;
  98         struct dlm_member *memb;
  99         int error = 0, delay;
 100
 101         list_for_each_entry(memb, &ls->ls_nodes, list) {
 102                 delay = 0;
 103                 for (;;) {
 104                         if (dlm_recovery_stopped(ls)) {
 105                                 error = -EINTR;
 106                                 goto out;
 107                         }
 108
 109                         error = dlm_rcom_status(ls, memb->nodeid);
 110                         if (error)
 111                                 goto out;
 112
 113                         if (rc->rc_result & wait_status)
 114                                 break;
 115                         if (delay < 1000)
 116                                 delay += 20;
 117                         msleep(delay);
 118                 }
 119         }
 120  out:
 121         return error;
 122 }
 123
 124 static int wait_status_low(struct dlm_ls *ls, uint32_t wait_status)
 125 {
 126         struct dlm_rcom *rc = (struct dlm_rcom *) ls->ls_recover_buf;
 127         int error = 0, delay = 0, nodeid = ls->ls_low_nodeid;
 128
 129         for (;;) {
 130                 if (dlm_recovery_stopped(ls)) {
 131                         error = -EINTR;
 132                         goto out;
 133                 }
 134
 135                 error = dlm_rcom_status(ls, nodeid);
 136                 if (error)
 137                         break;
 138
 139                 if (rc->rc_result & wait_status)
 140                         break;
 141                 if (delay < 1000)
 142                         delay += 20;
 143                 msleep(delay);
 144         }
 145  out:
 146         return error;
 147 }
 148
 149 static int wait_status(struct dlm_ls *ls, uint32_t status)
 150 {
 151         uint32_t status_all = status << 1;
 152         int error;
 153
 154         if (ls->ls_low_nodeid == dlm_our_nodeid()) {
 155                 error = wait_status_all(ls, status);
 156                 if (!error)
 157                         dlm_set_recover_status(ls, status_all);
 158         } else
 159                 error = wait_status_low(ls, status_all);
 160
 161         return error;
 162 }
 163
 164 int dlm_recover_members_wait(struct dlm_ls *ls)
 165 {
 166         return wait_status(ls, DLM_RS_NODES);
 167 }
 168
 169 int dlm_recover_directory_wait(struct dlm_ls *ls)
 170 {
 171         return wait_status(ls, DLM_RS_DIR);
 172 }
 173
 174 int dlm_recover_locks_wait(struct dlm_ls *ls)
 175 {
 176         return wait_status(ls, DLM_RS_LOCKS);
 177 }
 178
 179 int dlm_recover_done_wait(struct dlm_ls *ls)
 180 {
 181         return wait_status(ls, DLM_RS_DONE);
 182 }
 183
 184 /*
 185  * The recover_list contains all the rsb's for which we've requested the new
 186  * master nodeid.  As replies are returned from the resource directories the
 187  * rsb's are removed from the list.  When the list is empty we're done.
 188  *
 189  * The recover_list is later similarly used for all rsb's for which we've sent
 190  * new lkb's and need to receive new corresponding lkid's.
 191  *
 192  * We use the address of the rsb struct as a simple local identifier for the
 193  * rsb so we can match an rcom reply with the rsb it was sent for.
 194  */
 195
 196 static int recover_list_empty(struct dlm_ls *ls)
 197 {
 198         int empty;
 199
 200         spin_lock(&ls->ls_recover_list_lock);
 201         empty = list_empty(&ls->ls_recover_list);
 202         spin_unlock(&ls->ls_recover_list_lock);
 203
 204         return empty;
 205 }
 206
 207 static void recover_list_add(struct dlm_rsb *r)
 208 {
 209         struct dlm_ls *ls = r->res_ls;
 210
 211         spin_lock(&ls->ls_recover_list_lock);
 212         if (list_empty(&r->res_recover_list)) {
 213                 list_add_tail(&r->res_recover_list, &ls->ls_recover_list);
 214                 ls->ls_recover_list_count++;
 215                 dlm_hold_rsb(r);
 216         }
 217         spin_unlock(&ls->ls_recover_list_lock);
 218 }
 219
 220 static void recover_list_del(struct dlm_rsb *r)
 221 {
 222         struct dlm_ls *ls = r->res_ls;
 223
 224         spin_lock(&ls->ls_recover_list_lock);
 225         list_del_init(&r->res_recover_list);
 226         ls->ls_recover_list_count--;
 227         spin_unlock(&ls->ls_recover_list_lock);
 228
 229         dlm_put_rsb(r);
 230 }
 231
 232 static struct dlm_rsb *recover_list_find(struct dlm_ls *ls, uint64_t id)
 233 {
 234         struct dlm_rsb *r = NULL;
 235
 236         spin_lock(&ls->ls_recover_list_lock);
 237
 238         list_for_each_entry(r, &ls->ls_recover_list, res_recover_list) {
 239                 if (id == (unsigned long) r)
 240                         goto out;
 241         }
 242         r = NULL;
 243  out:
 244         spin_unlock(&ls->ls_recover_list_lock);
 245         return r;
 246 }
 247
 248 static void recover_list_clear(struct dlm_ls *ls)
 249 {
 250         struct dlm_rsb *r, *s;
 251
 252         spin_lock(&ls->ls_recover_list_lock);
 253         list_for_each_entry_safe(r, s, &ls->ls_recover_list, res_recover_list) {
 254                 list_del_init(&r->res_recover_list);
 255                 dlm_put_rsb(r);
 256                 ls->ls_recover_list_count--;
 257         }
 258
 259         if (ls->ls_recover_list_count != 0) {
 260                 log_error(ls, "warning: recover_list_count %d",
 261                           ls->ls_recover_list_count);
 262                 ls->ls_recover_list_count = 0;
 263         }
 264         spin_unlock(&ls->ls_recover_list_lock);
 265 }
 266
 267
 268 /* Master recovery: find new master node for rsb's that were
 269    mastered on nodes that have been removed.
 270
 271    dlm_recover_masters
 272    recover_master
 273    dlm_send_rcom_lookup            ->  receive_rcom_lookup
 274                                        dlm_dir_lookup
 275    receive_rcom_lookup_reply       <-
 276    dlm_recover_master_reply
 277    set_new_master
 278    set_master_lkbs
 279    set_lock_master
 280 */
 281
 282 /*
 283  * Set the lock master for all LKBs in a lock queue
 284  * If we are the new master of the rsb, we may have received new
 285  * MSTCPY locks from other nodes already which we need to ignore
 286  * when setting the new nodeid.
 287  */
 288
 289 static void set_lock_master(struct list_head *queue, int nodeid)
 290 {
 291         struct dlm_lkb *lkb;
 292
 293         list_for_each_entry(lkb, queue, lkb_statequeue)
 294                 if (!(lkb->lkb_flags & DLM_IFL_MSTCPY))
 295                         lkb->lkb_nodeid = nodeid;
 296 }
 297
 298 static void set_master_lkbs(struct dlm_rsb *r)
 299 {
 300         set_lock_master(&r->res_grantqueue, r->res_nodeid);
 301         set_lock_master(&r->res_convertqueue, r->res_nodeid);
 302         set_lock_master(&r->res_waitqueue, r->res_nodeid);
 303 }
 304
 305 /*
 306  * Propogate the new master nodeid to locks
 307  * The NEW_MASTER flag tells dlm_recover_locks() which rsb's to consider.
 308  * The NEW_MASTER2 flag tells recover_lvb() which rsb's to consider.
 309  */
 310
 311 static void set_new_master(struct dlm_rsb *r, int nodeid)
 312 {
 313         lock_rsb(r);
 314         r->res_nodeid = nodeid;
 315         set_master_lkbs(r);
 316         rsb_set_flag(r, RSB_NEW_MASTER);
 317         rsb_set_flag(r, RSB_NEW_MASTER2);
 318         unlock_rsb(r);
 319 }
 320
 321 /*
 322  * We do async lookups on rsb's that need new masters.  The rsb's
 323  * waiting for a lookup reply are kept on the recover_list.
 324  */
 325
 326 static int recover_master(struct dlm_rsb *r)
 327 {
 328         struct dlm_ls *ls = r->res_ls;
 329         int error, dir_nodeid, ret_nodeid, our_nodeid = dlm_our_nodeid();
 330
 331         dir_nodeid = dlm_dir_nodeid(r);
 332
 333         if (dir_nodeid == our_nodeid) {
 334                 error = dlm_dir_lookup(ls, our_nodeid, r->res_name,
 335                                        r->res_length, &ret_nodeid);
 336                 if (error)
 337                         log_error(ls, "recover dir lookup error %d", error);
 338
 339                 if (ret_nodeid == our_nodeid)
 340                         ret_nodeid = 0;
 341                 set_new_master(r, ret_nodeid);
 342         } else {
 343                 recover_list_add(r);
 344                 error = dlm_send_rcom_lookup(r, dir_nodeid);
 345         }
 346
 347         return error;
 348 }
 349
 350 /*
 351  * When not using a directory, most resource names will hash to a new static
 352  * master nodeid and the resource will need to be remastered.
 353  */
 354
 355 static int recover_master_static(struct dlm_rsb *r)
 356 {
 357         int master = dlm_dir_nodeid(r);
 358
 359         if (master == dlm_our_nodeid())
 360                 master = 0;
 361
 362         if (r->res_nodeid != master) {
 363                 if (is_master(r))
 364                         dlm_purge_mstcpy_locks(r);
 365                 set_new_master(r, master);
 366                 return 1;
 367         }
 368         return 0;
 369 }
 370
 371 /*
 372  * Go through local root resources and for each rsb which has a master which
 373  * has departed, get the new master nodeid from the directory.  The dir will
 374  * assign mastery to the first node to look up the new master.  That means
 375  * we'll discover in this lookup if we're the new master of any rsb's.
 376  *
 377  * We fire off all the dir lookup requests individually and asynchronously to
 378  * the correct dir node.
 379  */
 380
 381 int dlm_recover_masters(struct dlm_ls *ls)
 382 {
 383         struct dlm_rsb *r;
 384         int error = 0, count = 0;
 385
 386         log_debug(ls, "dlm_recover_masters");
 387
 388         down_read(&ls->ls_root_sem);
 389         list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
 390                 if (dlm_recovery_stopped(ls)) {
 391                         up_read(&ls->ls_root_sem);
 392                         error = -EINTR;
 393                         goto out;
 394                 }
 395
 396                 if (dlm_no_directory(ls))
 397                         count += recover_master_static(r);
 398                 else if (!is_master(r) && dlm_is_removed(ls, r->res_nodeid)) {
 399                         recover_master(r);
 400                         count++;
 401                 }
 402
 403                 schedule();
 404         }
 405         up_read(&ls->ls_root_sem);
 406
 407         log_debug(ls, "dlm_recover_masters %d resources", count);
 408
 409         error = dlm_wait_function(ls, &recover_list_empty);
 410  out:
 411         if (error)
 412                 recover_list_clear(ls);
 413         return error;
 414 }
 415
 416 int dlm_recover_master_reply(struct dlm_ls *ls, struct dlm_rcom *rc)
 417 {
 418         struct dlm_rsb *r;
 419         int nodeid;
 420
 421         r = recover_list_find(ls, rc->rc_id);
 422         if (!r) {
 423                 log_error(ls, "dlm_recover_master_reply no id %llx",
 424                           rc->rc_id);
 425                 goto out;
 426         }
 427
 428         nodeid = rc->rc_result;
 429         if (nodeid == dlm_our_nodeid())
 430                 nodeid = 0;
 431
 432         set_new_master(r, nodeid);
 433         recover_list_del(r);
 434
 435         if (recover_list_empty(ls))
 436                 wake_up(&ls->ls_wait_general);
 437  out:
 438         return 0;
 439 }
 440
 441
 442 /* Lock recovery: rebuild the process-copy locks we hold on a
 443    remastered rsb on the new rsb master.
 444
 445    dlm_recover_locks
 446    recover_locks
 447    recover_locks_queue
 448    dlm_send_rcom_lock              ->  receive_rcom_lock
 449                                        dlm_recover_master_copy
 450    receive_rcom_lock_reply         <-
 451    dlm_recover_process_copy
 452 */
 453
 454
 455 /*
 456  * keep a count of the number of lkb's we send to the new master; when we get
 457  * an equal number of replies then recovery for the rsb is done
 458  */
 459
 460 static int recover_locks_queue(struct dlm_rsb *r, struct list_head *head)
 461 {
 462         struct dlm_lkb *lkb;
 463         int error = 0;
 464
 465         list_for_each_entry(lkb, head, lkb_statequeue) {
 466                 error = dlm_send_rcom_lock(r, lkb);
 467                 if (error)
 468                         break;
 469                 r->res_recover_locks_count++;
 470         }
 471
 472         return error;
 473 }
 474
 475 static int all_queues_empty(struct dlm_rsb *r)
 476 {
 477         if (!list_empty(&r->res_grantqueue) ||
 478             !list_empty(&r->res_convertqueue) ||
 479             !list_empty(&r->res_waitqueue))
 480                 return 0;
 481         return 1;
 482 }
 483
 484 static int recover_locks(struct dlm_rsb *r)
 485 {
 486         int error = 0;
 487
 488         lock_rsb(r);
 489         if (all_queues_empty(r))
 490                 goto out;
 491
 492         DLM_ASSERT(!r->res_recover_locks_count, dlm_print_rsb(r););
 493
 494         error = recover_locks_queue(r, &r->res_grantqueue);
 495         if (error)
 496                 goto out;
 497         error = recover_locks_queue(r, &r->res_convertqueue);
 498         if (error)
 499                 goto out;
 500         error = recover_locks_queue(r, &r->res_waitqueue);
 501         if (error)
 502                 goto out;
 503
 504         if (r->res_recover_locks_count)
 505                 recover_list_add(r);
 506         else
 507                 rsb_clear_flag(r, RSB_NEW_MASTER);
 508  out:
 509         unlock_rsb(r);
 510         return error;
 511 }
 512
 513 int dlm_recover_locks(struct dlm_ls *ls)
 514 {
 515         struct dlm_rsb *r;
 516         int error, count = 0;
 517
 518         log_debug(ls, "dlm_recover_locks");
 519
 520         down_read(&ls->ls_root_sem);
 521         list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
 522                 if (is_master(r)) {
 523                         rsb_clear_flag(r, RSB_NEW_MASTER);
 524                         continue;
 525                 }
 526
 527                 if (!rsb_flag(r, RSB_NEW_MASTER))
 528                         continue;
 529
 530                 if (dlm_recovery_stopped(ls)) {
 531                         error = -EINTR;
 532                         up_read(&ls->ls_root_sem);
 533                         goto out;
 534                 }
 535
 536                 error = recover_locks(r);
 537                 if (error) {
 538                         up_read(&ls->ls_root_sem);
 539                         goto out;
 540                 }
 541
 542                 count += r->res_recover_locks_count;
 543         }
 544         up_read(&ls->ls_root_sem);
 545
 546         log_debug(ls, "dlm_recover_locks %d locks", count);
 547
 548         error = dlm_wait_function(ls, &recover_list_empty);
 549  out:
 550         if (error)
 551                 recover_list_clear(ls);
 552         else
 553                 dlm_set_recover_status(ls, DLM_RS_LOCKS);
 554         return error;
 555 }
 556
 557 void dlm_recovered_lock(struct dlm_rsb *r)
 558 {
 559         DLM_ASSERT(rsb_flag(r, RSB_NEW_MASTER), dlm_print_rsb(r););
 560
 561         r->res_recover_locks_count--;
 562         if (!r->res_recover_locks_count) {
 563                 rsb_clear_flag(r, RSB_NEW_MASTER);
 564                 recover_list_del(r);
 565         }
 566
 567         if (recover_list_empty(r->res_ls))
 568                 wake_up(&r->res_ls->ls_wait_general);
 569 }
 570
 571 /*
 572  * The lvb needs to be recovered on all master rsb's.  This includes setting
 573  * the VALNOTVALID flag if necessary, and determining the correct lvb contents
 574  * based on the lvb's of the locks held on the rsb.
 575  *
 576  * RSB_VALNOTVALID is set if there are only NL/CR locks on the rsb.  If it
 577  * was already set prior to recovery, it's not cleared, regardless of locks.
 578  *
 579  * The LVB contents are only considered for changing when this is a new master
 580  * of the rsb (NEW_MASTER2).  Then, the rsb's lvb is taken from any lkb with
 581  * mode > CR.  If no lkb's exist with mode above CR, the lvb contents are taken
 582  * from the lkb with the largest lvb sequence number.
 583  */
 584
 585 static void recover_lvb(struct dlm_rsb *r)
 586 {
 587         struct dlm_lkb *lkb, *high_lkb = NULL;
 588         uint32_t high_seq = 0;
 589         int lock_lvb_exists = 0;
 590         int big_lock_exists = 0;
 591         int lvblen = r->res_ls->ls_lvblen;
 592
 593         list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
 594                 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
 595                         continue;
 596
 597                 lock_lvb_exists = 1;
 598
 599                 if (lkb->lkb_grmode > DLM_LOCK_CR) {
 600                         big_lock_exists = 1;
 601                         goto setflag;
 602                 }
 603
 604                 if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
 605                         high_lkb = lkb;
 606                         high_seq = lkb->lkb_lvbseq;
 607                 }
 608         }
 609
 610         list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
 611                 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
 612                         continue;
 613
 614                 lock_lvb_exists = 1;
 615
 616                 if (lkb->lkb_grmode > DLM_LOCK_CR) {
 617                         big_lock_exists = 1;
 618                         goto setflag;
 619                 }
 620
 621                 if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
 622                         high_lkb = lkb;
 623                         high_seq = lkb->lkb_lvbseq;
 624                 }
 625         }
 626
 627  setflag:
 628         if (!lock_lvb_exists)
 629                 goto out;
 630
 631         if (!big_lock_exists)
 632                 rsb_set_flag(r, RSB_VALNOTVALID);
 633
 634         /* don't mess with the lvb unless we're the new master */
 635         if (!rsb_flag(r, RSB_NEW_MASTER2))
 636                 goto out;
 637
 638         if (!r->res_lvbptr) {
 639                 r->res_lvbptr = allocate_lvb(r->res_ls);
 640                 if (!r->res_lvbptr)
 641                         goto out;
 642         }
 643
 644         if (big_lock_exists) {
 645                 r->res_lvbseq = lkb->lkb_lvbseq;
 646                 memcpy(r->res_lvbptr, lkb->lkb_lvbptr, lvblen);
 647         } else if (high_lkb) {
 648                 r->res_lvbseq = high_lkb->lkb_lvbseq;
 649                 memcpy(r->res_lvbptr, high_lkb->lkb_lvbptr, lvblen);
 650         } else {
 651                 r->res_lvbseq = 0;
 652                 memset(r->res_lvbptr, 0, lvblen);
 653         }
 654  out:
 655         return;
 656 }
 657
 658 /* All master rsb's flagged RECOVER_CONVERT need to be looked at.  The locks
 659    converting PR->CW or CW->PR need to have their lkb_grmode set. */
 660
 661 static void recover_conversion(struct dlm_rsb *r)
 662 {
 663         struct dlm_lkb *lkb;
 664         int grmode = -1;
 665
 666         list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
 667                 if (lkb->lkb_grmode == DLM_LOCK_PR ||
 668                     lkb->lkb_grmode == DLM_LOCK_CW) {
 669                         grmode = lkb->lkb_grmode;
 670                         break;
 671                 }
 672         }
 673
 674         list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
 675                 if (lkb->lkb_grmode != DLM_LOCK_IV)
 676                         continue;
 677                 if (grmode == -1)
 678                         lkb->lkb_grmode = lkb->lkb_rqmode;
 679                 else
 680                         lkb->lkb_grmode = grmode;
 681         }
 682 }
 683
 684 void dlm_recover_rsbs(struct dlm_ls *ls)
 685 {
 686         struct dlm_rsb *r;
 687         int count = 0;
 688
 689         log_debug(ls, "dlm_recover_rsbs");
 690
 691         down_read(&ls->ls_root_sem);
 692         list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
 693                 lock_rsb(r);
 694                 if (is_master(r)) {
 695                         if (rsb_flag(r, RSB_RECOVER_CONVERT))
 696                                 recover_conversion(r);
 697                         recover_lvb(r);
 698                         count++;
 699                 }
 700                 rsb_clear_flag(r, RSB_RECOVER_CONVERT);
 701                 unlock_rsb(r);
 702         }
 703         up_read(&ls->ls_root_sem);
 704
 705         log_debug(ls, "dlm_recover_rsbs %d rsbs", count);
 706 }
 707
 708 /* Create a single list of all root rsb's to be used during recovery */
 709
 710 int dlm_create_root_list(struct dlm_ls *ls)
 711 {
 712         struct dlm_rsb *r;
 713         int i, error = 0;
 714
 715         down_write(&ls->ls_root_sem);
 716         if (!list_empty(&ls->ls_root_list)) {
 717                 log_error(ls, "root list not empty");
 718                 error = -EINVAL;
 719                 goto out;
 720         }
 721
 722         for (i = 0; i < ls->ls_rsbtbl_size; i++) {
 723                 read_lock(&ls->ls_rsbtbl[i].lock);
 724                 list_for_each_entry(r, &ls->ls_rsbtbl[i].list, res_hashchain) {
 725                         list_add(&r->res_root_list, &ls->ls_root_list);
 726                         dlm_hold_rsb(r);
 727                 }
 728                 read_unlock(&ls->ls_rsbtbl[i].lock);
 729         }
 730  out:
 731         up_write(&ls->ls_root_sem);
 732         return error;
 733 }
 734
 735 void dlm_release_root_list(struct dlm_ls *ls)
 736 {
 737         struct dlm_rsb *r, *safe;
 738
 739         down_write(&ls->ls_root_sem);
 740         list_for_each_entry_safe(r, safe, &ls->ls_root_list, res_root_list) {
 741                 list_del_init(&r->res_root_list);
 742                 dlm_put_rsb(r);
 743         }
 744         up_write(&ls->ls_root_sem);
 745 }
 746
 747 void dlm_clear_toss_list(struct dlm_ls *ls)
 748 {
 749         struct dlm_rsb *r, *safe;
 750         int i;
 751
 752         for (i = 0; i < ls->ls_rsbtbl_size; i++) {
 753                 write_lock(&ls->ls_rsbtbl[i].lock);
 754                 list_for_each_entry_safe(r, safe, &ls->ls_rsbtbl[i].toss,
 755                                          res_hashchain) {
 756                         list_del(&r->res_hashchain);
 757                         free_rsb(r);
 758                 }
 759                 write_unlock(&ls->ls_rsbtbl[i].lock);
 760         }
 761 }
 762