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.ci_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.ci_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 static void _set_recover_status(struct dlm_ls *ls, uint32_t status)
  89 {
  90         ls->ls_recover_status |= status;
  91 }
  92
  93 void dlm_set_recover_status(struct dlm_ls *ls, uint32_t status)
  94 {
  95         spin_lock(&ls->ls_recover_lock);
  96         _set_recover_status(ls, status);
  97         spin_unlock(&ls->ls_recover_lock);
  98 }
  99
 100 static int wait_status_all(struct dlm_ls *ls, uint32_t wait_status,
 101                            int save_slots)
 102 {
 103         struct dlm_rcom *rc = ls->ls_recover_buf;
 104         struct dlm_member *memb;
 105         int error = 0, delay;
 106
 107         list_for_each_entry(memb, &ls->ls_nodes, list) {
 108                 delay = 0;
 109                 for (;;) {
 110                         if (dlm_recovery_stopped(ls)) {
 111                                 error = -EINTR;
 112                                 goto out;
 113                         }
 114
 115                         error = dlm_rcom_status(ls, memb->nodeid, 0);
 116                         if (error)
 117                                 goto out;
 118
 119                         if (save_slots)
 120                                 dlm_slot_save(ls, rc, memb);
 121
 122                         if (rc->rc_result & wait_status)
 123                                 break;
 124                         if (delay < 1000)
 125                                 delay += 20;
 126                         msleep(delay);
 127                 }
 128         }
 129  out:
 130         return error;
 131 }
 132
 133 static int wait_status_low(struct dlm_ls *ls, uint32_t wait_status,
 134                            uint32_t status_flags)
 135 {
 136         struct dlm_rcom *rc = ls->ls_recover_buf;
 137         int error = 0, delay = 0, nodeid = ls->ls_low_nodeid;
 138
 139         for (;;) {
 140                 if (dlm_recovery_stopped(ls)) {
 141                         error = -EINTR;
 142                         goto out;
 143                 }
 144
 145                 error = dlm_rcom_status(ls, nodeid, status_flags);
 146                 if (error)
 147                         break;
 148
 149                 if (rc->rc_result & wait_status)
 150                         break;
 151                 if (delay < 1000)
 152                         delay += 20;
 153                 msleep(delay);
 154         }
 155  out:
 156         return error;
 157 }
 158
 159 static int wait_status(struct dlm_ls *ls, uint32_t status)
 160 {
 161         uint32_t status_all = status << 1;
 162         int error;
 163
 164         if (ls->ls_low_nodeid == dlm_our_nodeid()) {
 165                 error = wait_status_all(ls, status, 0);
 166                 if (!error)
 167                         dlm_set_recover_status(ls, status_all);
 168         } else
 169                 error = wait_status_low(ls, status_all, 0);
 170
 171         return error;
 172 }
 173
 174 int dlm_recover_members_wait(struct dlm_ls *ls)
 175 {
 176         struct dlm_member *memb;
 177         struct dlm_slot *slots;
 178         int num_slots, slots_size;
 179         int error, rv;
 180         uint32_t gen;
 181
 182         list_for_each_entry(memb, &ls->ls_nodes, list) {
 183                 memb->slot = -1;
 184                 memb->generation = 0;
 185         }
 186
 187         if (ls->ls_low_nodeid == dlm_our_nodeid()) {
 188                 error = wait_status_all(ls, DLM_RS_NODES, 1);
 189                 if (error)
 190                         goto out;
 191
 192                 /* slots array is sparse, slots_size may be > num_slots */
 193
 194                 rv = dlm_slots_assign(ls, &num_slots, &slots_size, &slots, &gen);
 195                 if (!rv) {
 196                         spin_lock(&ls->ls_recover_lock);
 197                         _set_recover_status(ls, DLM_RS_NODES_ALL);
 198                         ls->ls_num_slots = num_slots;
 199                         ls->ls_slots_size = slots_size;
 200                         ls->ls_slots = slots;
 201                         ls->ls_generation = gen;
 202                         spin_unlock(&ls->ls_recover_lock);
 203                 } else {
 204                         dlm_set_recover_status(ls, DLM_RS_NODES_ALL);
 205                 }
 206         } else {
 207                 error = wait_status_low(ls, DLM_RS_NODES_ALL, DLM_RSF_NEED_SLOTS);
 208                 if (error)
 209                         goto out;
 210
 211                 dlm_slots_copy_in(ls);
 212         }
 213  out:
 214         return error;
 215 }
 216
 217 int dlm_recover_directory_wait(struct dlm_ls *ls)
 218 {
 219         return wait_status(ls, DLM_RS_DIR);
 220 }
 221
 222 int dlm_recover_locks_wait(struct dlm_ls *ls)
 223 {
 224         return wait_status(ls, DLM_RS_LOCKS);
 225 }
 226
 227 int dlm_recover_done_wait(struct dlm_ls *ls)
 228 {
 229         return wait_status(ls, DLM_RS_DONE);
 230 }
 231
 232 /*
 233  * The recover_list contains all the rsb's for which we've requested the new
 234  * master nodeid.  As replies are returned from the resource directories the
 235  * rsb's are removed from the list.  When the list is empty we're done.
 236  *
 237  * The recover_list is later similarly used for all rsb's for which we've sent
 238  * new lkb's and need to receive new corresponding lkid's.
 239  *
 240  * We use the address of the rsb struct as a simple local identifier for the
 241  * rsb so we can match an rcom reply with the rsb it was sent for.
 242  */
 243
 244 static int recover_list_empty(struct dlm_ls *ls)
 245 {
 246         int empty;
 247
 248         spin_lock(&ls->ls_recover_list_lock);
 249         empty = list_empty(&ls->ls_recover_list);
 250         spin_unlock(&ls->ls_recover_list_lock);
 251
 252         return empty;
 253 }
 254
 255 static void recover_list_add(struct dlm_rsb *r)
 256 {
 257         struct dlm_ls *ls = r->res_ls;
 258
 259         spin_lock(&ls->ls_recover_list_lock);
 260         if (list_empty(&r->res_recover_list)) {
 261                 list_add_tail(&r->res_recover_list, &ls->ls_recover_list);
 262                 ls->ls_recover_list_count++;
 263                 dlm_hold_rsb(r);
 264         }
 265         spin_unlock(&ls->ls_recover_list_lock);
 266 }
 267
 268 static void recover_list_del(struct dlm_rsb *r)
 269 {
 270         struct dlm_ls *ls = r->res_ls;
 271
 272         spin_lock(&ls->ls_recover_list_lock);
 273         list_del_init(&r->res_recover_list);
 274         ls->ls_recover_list_count--;
 275         spin_unlock(&ls->ls_recover_list_lock);
 276
 277         dlm_put_rsb(r);
 278 }
 279
 280 static struct dlm_rsb *recover_list_find(struct dlm_ls *ls, uint64_t id)
 281 {
 282         struct dlm_rsb *r = NULL;
 283
 284         spin_lock(&ls->ls_recover_list_lock);
 285
 286         list_for_each_entry(r, &ls->ls_recover_list, res_recover_list) {
 287                 if (id == (unsigned long) r)
 288                         goto out;
 289         }
 290         r = NULL;
 291  out:
 292         spin_unlock(&ls->ls_recover_list_lock);
 293         return r;
 294 }
 295
 296 static void recover_list_clear(struct dlm_ls *ls)
 297 {
 298         struct dlm_rsb *r, *s;
 299
 300         spin_lock(&ls->ls_recover_list_lock);
 301         list_for_each_entry_safe(r, s, &ls->ls_recover_list, res_recover_list) {
 302                 list_del_init(&r->res_recover_list);
 303                 r->res_recover_locks_count = 0;
 304                 dlm_put_rsb(r);
 305                 ls->ls_recover_list_count--;
 306         }
 307
 308         if (ls->ls_recover_list_count != 0) {
 309                 log_error(ls, "warning: recover_list_count %d",
 310                           ls->ls_recover_list_count);
 311                 ls->ls_recover_list_count = 0;
 312         }
 313         spin_unlock(&ls->ls_recover_list_lock);
 314 }
 315
 316
 317 /* Master recovery: find new master node for rsb's that were
 318    mastered on nodes that have been removed.
 319
 320    dlm_recover_masters
 321    recover_master
 322    dlm_send_rcom_lookup            ->  receive_rcom_lookup
 323                                        dlm_dir_lookup
 324    receive_rcom_lookup_reply       <-
 325    dlm_recover_master_reply
 326    set_new_master
 327    set_master_lkbs
 328    set_lock_master
 329 */
 330
 331 /*
 332  * Set the lock master for all LKBs in a lock queue
 333  * If we are the new master of the rsb, we may have received new
 334  * MSTCPY locks from other nodes already which we need to ignore
 335  * when setting the new nodeid.
 336  */
 337
 338 static void set_lock_master(struct list_head *queue, int nodeid)
 339 {
 340         struct dlm_lkb *lkb;
 341
 342         list_for_each_entry(lkb, queue, lkb_statequeue) {
 343                 if (!(lkb->lkb_flags & DLM_IFL_MSTCPY)) {
 344                         lkb->lkb_nodeid = nodeid;
 345                         lkb->lkb_remid = 0;
 346                 }
 347         }
 348 }
 349
 350 static void set_master_lkbs(struct dlm_rsb *r)
 351 {
 352         set_lock_master(&r->res_grantqueue, r->res_nodeid);
 353         set_lock_master(&r->res_convertqueue, r->res_nodeid);
 354         set_lock_master(&r->res_waitqueue, r->res_nodeid);
 355 }
 356
 357 /*
 358  * Propagate the new master nodeid to locks
 359  * The NEW_MASTER flag tells dlm_recover_locks() which rsb's to consider.
 360  * The NEW_MASTER2 flag tells recover_lvb() and recover_grant() which
 361  * rsb's to consider.
 362  */
 363
 364 static void set_new_master(struct dlm_rsb *r, int nodeid)
 365 {
 366         r->res_nodeid = nodeid;
 367         set_master_lkbs(r);
 368         rsb_set_flag(r, RSB_NEW_MASTER);
 369         rsb_set_flag(r, RSB_NEW_MASTER2);
 370 }
 371
 372 /*
 373  * We do async lookups on rsb's that need new masters.  The rsb's
 374  * waiting for a lookup reply are kept on the recover_list.
 375  */
 376
 377 static int recover_master(struct dlm_rsb *r)
 378 {
 379         struct dlm_ls *ls = r->res_ls;
 380         int error, ret_nodeid;
 381         int our_nodeid = dlm_our_nodeid();
 382         int dir_nodeid = dlm_dir_nodeid(r);
 383
 384         if (dir_nodeid == our_nodeid) {
 385                 error = dlm_dir_lookup(ls, our_nodeid, r->res_name,
 386                                        r->res_length, &ret_nodeid);
 387                 if (error)
 388                         log_error(ls, "recover dir lookup error %d", error);
 389
 390                 if (ret_nodeid == our_nodeid)
 391                         ret_nodeid = 0;
 392                 lock_rsb(r);
 393                 set_new_master(r, ret_nodeid);
 394                 unlock_rsb(r);
 395         } else {
 396                 recover_list_add(r);
 397                 error = dlm_send_rcom_lookup(r, dir_nodeid);
 398         }
 399
 400         return error;
 401 }
 402
 403 /*
 404  * All MSTCPY locks are purged and rebuilt, even if the master stayed the same.
 405  * This is necessary because recovery can be started, aborted and restarted,
 406  * causing the master nodeid to briefly change during the aborted recovery, and
 407  * change back to the original value in the second recovery.  The MSTCPY locks
 408  * may or may not have been purged during the aborted recovery.  Another node
 409  * with an outstanding request in waiters list and a request reply saved in the
 410  * requestqueue, cannot know whether it should ignore the reply and resend the
 411  * request, or accept the reply and complete the request.  It must do the
 412  * former if the remote node purged MSTCPY locks, and it must do the later if
 413  * the remote node did not.  This is solved by always purging MSTCPY locks, in
 414  * which case, the request reply would always be ignored and the request
 415  * resent.
 416  */
 417
 418 static int recover_master_static(struct dlm_rsb *r)
 419 {
 420         int dir_nodeid = dlm_dir_nodeid(r);
 421         int new_master = dir_nodeid;
 422
 423         if (dir_nodeid == dlm_our_nodeid())
 424                 new_master = 0;
 425
 426         lock_rsb(r);
 427         dlm_purge_mstcpy_locks(r);
 428         set_new_master(r, new_master);
 429         unlock_rsb(r);
 430         return 1;
 431 }
 432
 433 /*
 434  * Go through local root resources and for each rsb which has a master which
 435  * has departed, get the new master nodeid from the directory.  The dir will
 436  * assign mastery to the first node to look up the new master.  That means
 437  * we'll discover in this lookup if we're the new master of any rsb's.
 438  *
 439  * We fire off all the dir lookup requests individually and asynchronously to
 440  * the correct dir node.
 441  */
 442
 443 int dlm_recover_masters(struct dlm_ls *ls)
 444 {
 445         struct dlm_rsb *r;
 446         int error = 0, count = 0;
 447
 448         log_debug(ls, "dlm_recover_masters");
 449
 450         down_read(&ls->ls_root_sem);
 451         list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
 452                 if (dlm_recovery_stopped(ls)) {
 453                         up_read(&ls->ls_root_sem);
 454                         error = -EINTR;
 455                         goto out;
 456                 }
 457
 458                 if (dlm_no_directory(ls))
 459                         count += recover_master_static(r);
 460                 else if (!is_master(r) &&
 461                          (dlm_is_removed(ls, r->res_nodeid) ||
 462                           rsb_flag(r, RSB_NEW_MASTER))) {
 463                         recover_master(r);
 464                         count++;
 465                 }
 466
 467                 schedule();
 468         }
 469         up_read(&ls->ls_root_sem);
 470
 471         log_debug(ls, "dlm_recover_masters %d resources", count);
 472
 473         error = dlm_wait_function(ls, &recover_list_empty);
 474  out:
 475         if (error)
 476                 recover_list_clear(ls);
 477         return error;
 478 }
 479
 480 int dlm_recover_master_reply(struct dlm_ls *ls, struct dlm_rcom *rc)
 481 {
 482         struct dlm_rsb *r;
 483         int nodeid;
 484
 485         r = recover_list_find(ls, rc->rc_id);
 486         if (!r) {
 487                 log_error(ls, "dlm_recover_master_reply no id %llx",
 488                           (unsigned long long)rc->rc_id);
 489                 goto out;
 490         }
 491
 492         nodeid = rc->rc_result;
 493         if (nodeid == dlm_our_nodeid())
 494                 nodeid = 0;
 495
 496         lock_rsb(r);
 497         set_new_master(r, nodeid);
 498         unlock_rsb(r);
 499         recover_list_del(r);
 500
 501         if (recover_list_empty(ls))
 502                 wake_up(&ls->ls_wait_general);
 503  out:
 504         return 0;
 505 }
 506
 507
 508 /* Lock recovery: rebuild the process-copy locks we hold on a
 509    remastered rsb on the new rsb master.
 510
 511    dlm_recover_locks
 512    recover_locks
 513    recover_locks_queue
 514    dlm_send_rcom_lock              ->  receive_rcom_lock
 515                                        dlm_recover_master_copy
 516    receive_rcom_lock_reply         <-
 517    dlm_recover_process_copy
 518 */
 519
 520
 521 /*
 522  * keep a count of the number of lkb's we send to the new master; when we get
 523  * an equal number of replies then recovery for the rsb is done
 524  */
 525
 526 static int recover_locks_queue(struct dlm_rsb *r, struct list_head *head)
 527 {
 528         struct dlm_lkb *lkb;
 529         int error = 0;
 530
 531         list_for_each_entry(lkb, head, lkb_statequeue) {
 532                 error = dlm_send_rcom_lock(r, lkb);
 533                 if (error)
 534                         break;
 535                 r->res_recover_locks_count++;
 536         }
 537
 538         return error;
 539 }
 540
 541 static int recover_locks(struct dlm_rsb *r)
 542 {
 543         int error = 0;
 544
 545         lock_rsb(r);
 546
 547         DLM_ASSERT(!r->res_recover_locks_count, dlm_dump_rsb(r););
 548
 549         error = recover_locks_queue(r, &r->res_grantqueue);
 550         if (error)
 551                 goto out;
 552         error = recover_locks_queue(r, &r->res_convertqueue);
 553         if (error)
 554                 goto out;
 555         error = recover_locks_queue(r, &r->res_waitqueue);
 556         if (error)
 557                 goto out;
 558
 559         if (r->res_recover_locks_count)
 560                 recover_list_add(r);
 561         else
 562                 rsb_clear_flag(r, RSB_NEW_MASTER);
 563  out:
 564         unlock_rsb(r);
 565         return error;
 566 }
 567
 568 int dlm_recover_locks(struct dlm_ls *ls)
 569 {
 570         struct dlm_rsb *r;
 571         int error, count = 0;
 572
 573         down_read(&ls->ls_root_sem);
 574         list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
 575                 if (is_master(r)) {
 576                         rsb_clear_flag(r, RSB_NEW_MASTER);
 577                         continue;
 578                 }
 579
 580                 if (!rsb_flag(r, RSB_NEW_MASTER))
 581                         continue;
 582
 583                 if (dlm_recovery_stopped(ls)) {
 584                         error = -EINTR;
 585                         up_read(&ls->ls_root_sem);
 586                         goto out;
 587                 }
 588
 589                 error = recover_locks(r);
 590                 if (error) {
 591                         up_read(&ls->ls_root_sem);
 592                         goto out;
 593                 }
 594
 595                 count += r->res_recover_locks_count;
 596         }
 597         up_read(&ls->ls_root_sem);
 598
 599         log_debug(ls, "dlm_recover_locks %d out", count);
 600
 601         error = dlm_wait_function(ls, &recover_list_empty);
 602  out:
 603         if (error)
 604                 recover_list_clear(ls);
 605         return error;
 606 }
 607
 608 void dlm_recovered_lock(struct dlm_rsb *r)
 609 {
 610         DLM_ASSERT(rsb_flag(r, RSB_NEW_MASTER), dlm_dump_rsb(r););
 611
 612         r->res_recover_locks_count--;
 613         if (!r->res_recover_locks_count) {
 614                 rsb_clear_flag(r, RSB_NEW_MASTER);
 615                 recover_list_del(r);
 616         }
 617
 618         if (recover_list_empty(r->res_ls))
 619                 wake_up(&r->res_ls->ls_wait_general);
 620 }
 621
 622 /*
 623  * The lvb needs to be recovered on all master rsb's.  This includes setting
 624  * the VALNOTVALID flag if necessary, and determining the correct lvb contents
 625  * based on the lvb's of the locks held on the rsb.
 626  *
 627  * RSB_VALNOTVALID is set if there are only NL/CR locks on the rsb.  If it
 628  * was already set prior to recovery, it's not cleared, regardless of locks.
 629  *
 630  * The LVB contents are only considered for changing when this is a new master
 631  * of the rsb (NEW_MASTER2).  Then, the rsb's lvb is taken from any lkb with
 632  * mode > CR.  If no lkb's exist with mode above CR, the lvb contents are taken
 633  * from the lkb with the largest lvb sequence number.
 634  */
 635
 636 static void recover_lvb(struct dlm_rsb *r)
 637 {
 638         struct dlm_lkb *lkb, *high_lkb = NULL;
 639         uint32_t high_seq = 0;
 640         int lock_lvb_exists = 0;
 641         int big_lock_exists = 0;
 642         int lvblen = r->res_ls->ls_lvblen;
 643
 644         list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
 645                 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
 646                         continue;
 647
 648                 lock_lvb_exists = 1;
 649
 650                 if (lkb->lkb_grmode > DLM_LOCK_CR) {
 651                         big_lock_exists = 1;
 652                         goto setflag;
 653                 }
 654
 655                 if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
 656                         high_lkb = lkb;
 657                         high_seq = lkb->lkb_lvbseq;
 658                 }
 659         }
 660
 661         list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
 662                 if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
 663                         continue;
 664
 665                 lock_lvb_exists = 1;
 666
 667                 if (lkb->lkb_grmode > DLM_LOCK_CR) {
 668                         big_lock_exists = 1;
 669                         goto setflag;
 670                 }
 671
 672                 if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
 673                         high_lkb = lkb;
 674                         high_seq = lkb->lkb_lvbseq;
 675                 }
 676         }
 677
 678  setflag:
 679         if (!lock_lvb_exists)
 680                 goto out;
 681
 682         if (!big_lock_exists)
 683                 rsb_set_flag(r, RSB_VALNOTVALID);
 684
 685         /* don't mess with the lvb unless we're the new master */
 686         if (!rsb_flag(r, RSB_NEW_MASTER2))
 687                 goto out;
 688
 689         if (!r->res_lvbptr) {
 690                 r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
 691                 if (!r->res_lvbptr)
 692                         goto out;
 693         }
 694
 695         if (big_lock_exists) {
 696                 r->res_lvbseq = lkb->lkb_lvbseq;
 697                 memcpy(r->res_lvbptr, lkb->lkb_lvbptr, lvblen);
 698         } else if (high_lkb) {
 699                 r->res_lvbseq = high_lkb->lkb_lvbseq;
 700                 memcpy(r->res_lvbptr, high_lkb->lkb_lvbptr, lvblen);
 701         } else {
 702                 r->res_lvbseq = 0;
 703                 memset(r->res_lvbptr, 0, lvblen);
 704         }
 705  out:
 706         return;
 707 }
 708
 709 /* All master rsb's flagged RECOVER_CONVERT need to be looked at.  The locks
 710    converting PR->CW or CW->PR need to have their lkb_grmode set. */
 711
 712 static void recover_conversion(struct dlm_rsb *r)
 713 {
 714         struct dlm_lkb *lkb;
 715         int grmode = -1;
 716
 717         list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
 718                 if (lkb->lkb_grmode == DLM_LOCK_PR ||
 719                     lkb->lkb_grmode == DLM_LOCK_CW) {
 720                         grmode = lkb->lkb_grmode;
 721                         break;
 722                 }
 723         }
 724
 725         list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
 726                 if (lkb->lkb_grmode != DLM_LOCK_IV)
 727                         continue;
 728                 if (grmode == -1)
 729                         lkb->lkb_grmode = lkb->lkb_rqmode;
 730                 else
 731                         lkb->lkb_grmode = grmode;
 732         }
 733 }
 734
 735 /* We've become the new master for this rsb and waiting/converting locks may
 736    need to be granted in dlm_recover_grant() due to locks that may have
 737    existed from a removed node. */
 738
 739 static void recover_grant(struct dlm_rsb *r)
 740 {
 741         if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
 742                 rsb_set_flag(r, RSB_RECOVER_GRANT);
 743 }
 744
 745 void dlm_recover_rsbs(struct dlm_ls *ls)
 746 {
 747         struct dlm_rsb *r;
 748         unsigned int count = 0;
 749
 750         down_read(&ls->ls_root_sem);
 751         list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
 752                 lock_rsb(r);
 753                 if (is_master(r)) {
 754                         if (rsb_flag(r, RSB_RECOVER_CONVERT))
 755                                 recover_conversion(r);
 756                         if (rsb_flag(r, RSB_NEW_MASTER2))
 757                                 recover_grant(r);
 758                         recover_lvb(r);
 759                         count++;
 760                 }
 761                 rsb_clear_flag(r, RSB_RECOVER_CONVERT);
 762                 rsb_clear_flag(r, RSB_NEW_MASTER2);
 763                 unlock_rsb(r);
 764         }
 765         up_read(&ls->ls_root_sem);
 766
 767         if (count)
 768                 log_debug(ls, "dlm_recover_rsbs %d done", count);
 769 }
 770
 771 /* Create a single list of all root rsb's to be used during recovery */
 772
 773 int dlm_create_root_list(struct dlm_ls *ls)
 774 {
 775         struct rb_node *n;
 776         struct dlm_rsb *r;
 777         int i, error = 0;
 778
 779         down_write(&ls->ls_root_sem);
 780         if (!list_empty(&ls->ls_root_list)) {
 781                 log_error(ls, "root list not empty");
 782                 error = -EINVAL;
 783                 goto out;
 784         }
 785
 786         for (i = 0; i < ls->ls_rsbtbl_size; i++) {
 787                 spin_lock(&ls->ls_rsbtbl[i].lock);
 788                 for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
 789                         r = rb_entry(n, struct dlm_rsb, res_hashnode);
 790                         list_add(&r->res_root_list, &ls->ls_root_list);
 791                         dlm_hold_rsb(r);
 792                 }
 793
 794                 /* If we're using a directory, add tossed rsbs to the root
 795                    list; they'll have entries created in the new directory,
 796                    but no other recovery steps should do anything with them. */
 797
 798                 if (dlm_no_directory(ls)) {
 799                         spin_unlock(&ls->ls_rsbtbl[i].lock);
 800                         continue;
 801                 }
 802
 803                 for (n = rb_first(&ls->ls_rsbtbl[i].toss); n; n = rb_next(n)) {
 804                         r = rb_entry(n, struct dlm_rsb, res_hashnode);
 805                         list_add(&r->res_root_list, &ls->ls_root_list);
 806                         dlm_hold_rsb(r);
 807                 }
 808                 spin_unlock(&ls->ls_rsbtbl[i].lock);
 809         }
 810  out:
 811         up_write(&ls->ls_root_sem);
 812         return error;
 813 }
 814
 815 void dlm_release_root_list(struct dlm_ls *ls)
 816 {
 817         struct dlm_rsb *r, *safe;
 818
 819         down_write(&ls->ls_root_sem);
 820         list_for_each_entry_safe(r, safe, &ls->ls_root_list, res_root_list) {
 821                 list_del_init(&r->res_root_list);
 822                 dlm_put_rsb(r);
 823         }
 824         up_write(&ls->ls_root_sem);
 825 }
 826
 827 /* If not using a directory, clear the entire toss list, there's no benefit to
 828    caching the master value since it's fixed.  If we are using a dir, keep the
 829    rsb's we're the master of.  Recovery will add them to the root list and from
 830    there they'll be entered in the rebuilt directory. */
 831
 832 void dlm_clear_toss_list(struct dlm_ls *ls)
 833 {
 834         struct rb_node *n, *next;
 835         struct dlm_rsb *rsb;
 836         int i;
 837
 838         for (i = 0; i < ls->ls_rsbtbl_size; i++) {
 839                 spin_lock(&ls->ls_rsbtbl[i].lock);
 840                 for (n = rb_first(&ls->ls_rsbtbl[i].toss); n; n = next) {
 841                         next = rb_next(n);;
 842                         rsb = rb_entry(n, struct dlm_rsb, res_hashnode);
 843                         if (dlm_no_directory(ls) || !is_master(rsb)) {
 844                                 rb_erase(n, &ls->ls_rsbtbl[i].toss);
 845                                 dlm_free_rsb(rsb);
 846                         }
 847                 }
 848                 spin_unlock(&ls->ls_rsbtbl[i].lock);
 849         }
 850 }
 851