recover.c 22.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38
/******************************************************************************
*******************************************************************************
**
**  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
**  Copyright (C) 2004-2005 Red Hat, Inc.  All rights reserved.
**
**  This copyrighted material is made available to anyone wishing to use,
**  modify, copy, or redistribute it subject to the terms and conditions
**  of the GNU General Public License v.2.
**
*******************************************************************************
******************************************************************************/

#include "dlm_internal.h"
#include "lockspace.h"
#include "dir.h"
#include "config.h"
#include "ast.h"
#include "memory.h"
#include "rcom.h"
#include "lock.h"
#include "lowcomms.h"
#include "member.h"
#include "recover.h"


/*
 * Recovery waiting routines: these functions wait for a particular reply from
 * a remote node, or for the remote node to report a certain status.  They need
 * to abort if the lockspace is stopped indicating a node has failed (perhaps
 * the one being waited for).
 */

/*
 * Wait until given function returns non-zero or lockspace is stopped
 * (LS_RECOVERY_STOP set due to failure of a node in ls_nodes).  When another
 * function thinks it could have completed the waited-on task, they should wake
 * up ls_wait_general to get an immediate response rather than waiting for the
D
David Teigland 已提交
39 40 41
 * timeout.  This uses a timeout so it can check periodically if the wait
 * should abort due to node failure (which doesn't cause a wake_up).
 * This should only be called by the dlm_recoverd thread.
42 43 44 45 46
 */

int dlm_wait_function(struct dlm_ls *ls, int (*testfn) (struct dlm_ls *ls))
{
	int error = 0;
D
David Teigland 已提交
47
	int rv;
48

D
David Teigland 已提交
49 50 51 52 53 54 55
	while (1) {
		rv = wait_event_timeout(ls->ls_wait_general,
					testfn(ls) || dlm_recovery_stopped(ls),
					dlm_config.ci_recover_timer * HZ);
		if (rv)
			break;
	}
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80

	if (dlm_recovery_stopped(ls)) {
		log_debug(ls, "dlm_wait_function aborted");
		error = -EINTR;
	}
	return error;
}

/*
 * An efficient way for all nodes to wait for all others to have a certain
 * status.  The node with the lowest nodeid polls all the others for their
 * status (wait_status_all) and all the others poll the node with the low id
 * for its accumulated result (wait_status_low).  When all nodes have set
 * status flag X, then status flag X_ALL will be set on the low nodeid.
 */

uint32_t dlm_recover_status(struct dlm_ls *ls)
{
	uint32_t status;
	spin_lock(&ls->ls_recover_lock);
	status = ls->ls_recover_status;
	spin_unlock(&ls->ls_recover_lock);
	return status;
}

81 82 83 84 85
static void _set_recover_status(struct dlm_ls *ls, uint32_t status)
{
	ls->ls_recover_status |= status;
}

86 87 88
void dlm_set_recover_status(struct dlm_ls *ls, uint32_t status)
{
	spin_lock(&ls->ls_recover_lock);
89
	_set_recover_status(ls, status);
90 91 92
	spin_unlock(&ls->ls_recover_lock);
}

93 94
static int wait_status_all(struct dlm_ls *ls, uint32_t wait_status,
			   int save_slots)
95
{
96
	struct dlm_rcom *rc = ls->ls_recover_buf;
97 98 99 100 101 102 103 104 105 106 107
	struct dlm_member *memb;
	int error = 0, delay;

	list_for_each_entry(memb, &ls->ls_nodes, list) {
		delay = 0;
		for (;;) {
			if (dlm_recovery_stopped(ls)) {
				error = -EINTR;
				goto out;
			}

108
			error = dlm_rcom_status(ls, memb->nodeid, 0);
109 110 111
			if (error)
				goto out;

112 113 114
			if (save_slots)
				dlm_slot_save(ls, rc, memb);

115 116 117 118 119 120 121 122 123 124 125
			if (rc->rc_result & wait_status)
				break;
			if (delay < 1000)
				delay += 20;
			msleep(delay);
		}
	}
 out:
	return error;
}

126 127
static int wait_status_low(struct dlm_ls *ls, uint32_t wait_status,
			   uint32_t status_flags)
128
{
129
	struct dlm_rcom *rc = ls->ls_recover_buf;
130 131 132 133 134 135 136 137
	int error = 0, delay = 0, nodeid = ls->ls_low_nodeid;

	for (;;) {
		if (dlm_recovery_stopped(ls)) {
			error = -EINTR;
			goto out;
		}

138
		error = dlm_rcom_status(ls, nodeid, status_flags);
139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157
		if (error)
			break;

		if (rc->rc_result & wait_status)
			break;
		if (delay < 1000)
			delay += 20;
		msleep(delay);
	}
 out:
	return error;
}

static int wait_status(struct dlm_ls *ls, uint32_t status)
{
	uint32_t status_all = status << 1;
	int error;

	if (ls->ls_low_nodeid == dlm_our_nodeid()) {
158
		error = wait_status_all(ls, status, 0);
159 160 161
		if (!error)
			dlm_set_recover_status(ls, status_all);
	} else
162
		error = wait_status_low(ls, status_all, 0);
163 164 165 166 167 168

	return error;
}

int dlm_recover_members_wait(struct dlm_ls *ls)
{
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207
	struct dlm_member *memb;
	struct dlm_slot *slots;
	int num_slots, slots_size;
	int error, rv;
	uint32_t gen;

	list_for_each_entry(memb, &ls->ls_nodes, list) {
		memb->slot = -1;
		memb->generation = 0;
	}

	if (ls->ls_low_nodeid == dlm_our_nodeid()) {
		error = wait_status_all(ls, DLM_RS_NODES, 1);
		if (error)
			goto out;

		/* slots array is sparse, slots_size may be > num_slots */

		rv = dlm_slots_assign(ls, &num_slots, &slots_size, &slots, &gen);
		if (!rv) {
			spin_lock(&ls->ls_recover_lock);
			_set_recover_status(ls, DLM_RS_NODES_ALL);
			ls->ls_num_slots = num_slots;
			ls->ls_slots_size = slots_size;
			ls->ls_slots = slots;
			ls->ls_generation = gen;
			spin_unlock(&ls->ls_recover_lock);
		} else {
			dlm_set_recover_status(ls, DLM_RS_NODES_ALL);
		}
	} else {
		error = wait_status_low(ls, DLM_RS_NODES_ALL, DLM_RSF_NEED_SLOTS);
		if (error)
			goto out;

		dlm_slots_copy_in(ls);
	}
 out:
	return error;
208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279
}

int dlm_recover_directory_wait(struct dlm_ls *ls)
{
	return wait_status(ls, DLM_RS_DIR);
}

int dlm_recover_locks_wait(struct dlm_ls *ls)
{
	return wait_status(ls, DLM_RS_LOCKS);
}

int dlm_recover_done_wait(struct dlm_ls *ls)
{
	return wait_status(ls, DLM_RS_DONE);
}

/*
 * The recover_list contains all the rsb's for which we've requested the new
 * master nodeid.  As replies are returned from the resource directories the
 * rsb's are removed from the list.  When the list is empty we're done.
 *
 * The recover_list is later similarly used for all rsb's for which we've sent
 * new lkb's and need to receive new corresponding lkid's.
 *
 * We use the address of the rsb struct as a simple local identifier for the
 * rsb so we can match an rcom reply with the rsb it was sent for.
 */

static int recover_list_empty(struct dlm_ls *ls)
{
	int empty;

	spin_lock(&ls->ls_recover_list_lock);
	empty = list_empty(&ls->ls_recover_list);
	spin_unlock(&ls->ls_recover_list_lock);

	return empty;
}

static void recover_list_add(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;

	spin_lock(&ls->ls_recover_list_lock);
	if (list_empty(&r->res_recover_list)) {
		list_add_tail(&r->res_recover_list, &ls->ls_recover_list);
		ls->ls_recover_list_count++;
		dlm_hold_rsb(r);
	}
	spin_unlock(&ls->ls_recover_list_lock);
}

static void recover_list_del(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;

	spin_lock(&ls->ls_recover_list_lock);
	list_del_init(&r->res_recover_list);
	ls->ls_recover_list_count--;
	spin_unlock(&ls->ls_recover_list_lock);

	dlm_put_rsb(r);
}

static void recover_list_clear(struct dlm_ls *ls)
{
	struct dlm_rsb *r, *s;

	spin_lock(&ls->ls_recover_list_lock);
	list_for_each_entry_safe(r, s, &ls->ls_recover_list, res_recover_list) {
		list_del_init(&r->res_recover_list);
280
		r->res_recover_locks_count = 0;
281 282 283 284 285 286 287 288 289 290 291 292
		dlm_put_rsb(r);
		ls->ls_recover_list_count--;
	}

	if (ls->ls_recover_list_count != 0) {
		log_error(ls, "warning: recover_list_count %d",
			  ls->ls_recover_list_count);
		ls->ls_recover_list_count = 0;
	}
	spin_unlock(&ls->ls_recover_list_lock);
}

293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353
static int recover_idr_empty(struct dlm_ls *ls)
{
	int empty = 1;

	spin_lock(&ls->ls_recover_idr_lock);
	if (ls->ls_recover_list_count)
		empty = 0;
	spin_unlock(&ls->ls_recover_idr_lock);

	return empty;
}

static int recover_idr_add(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;
	int rv, id;

	rv = idr_pre_get(&ls->ls_recover_idr, GFP_NOFS);
	if (!rv)
		return -ENOMEM;

	spin_lock(&ls->ls_recover_idr_lock);
	if (r->res_id) {
		spin_unlock(&ls->ls_recover_idr_lock);
		return -1;
	}
	rv = idr_get_new_above(&ls->ls_recover_idr, r, 1, &id);
	if (rv) {
		spin_unlock(&ls->ls_recover_idr_lock);
		return rv;
	}
	r->res_id = id;
	ls->ls_recover_list_count++;
	dlm_hold_rsb(r);
	spin_unlock(&ls->ls_recover_idr_lock);
	return 0;
}

static void recover_idr_del(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;

	spin_lock(&ls->ls_recover_idr_lock);
	idr_remove(&ls->ls_recover_idr, r->res_id);
	r->res_id = 0;
	ls->ls_recover_list_count--;
	spin_unlock(&ls->ls_recover_idr_lock);

	dlm_put_rsb(r);
}

static struct dlm_rsb *recover_idr_find(struct dlm_ls *ls, uint64_t id)
{
	struct dlm_rsb *r;

	spin_lock(&ls->ls_recover_idr_lock);
	r = idr_find(&ls->ls_recover_idr, (int)id);
	spin_unlock(&ls->ls_recover_idr_lock);
	return r;
}

354
static void recover_idr_clear(struct dlm_ls *ls)
355
{
356 357
	struct dlm_rsb *r;
	int id;
358

359
	spin_lock(&ls->ls_recover_idr_lock);
360

361
	idr_for_each_entry(&ls->ls_recover_idr, r, id) {
T
Tejun Heo 已提交
362
		idr_remove(&ls->ls_recover_idr, id);
363 364 365
		r->res_id = 0;
		r->res_recover_locks_count = 0;
		ls->ls_recover_list_count--;
366

367 368
		dlm_put_rsb(r);
	}
369 370 371 372 373 374 375 376 377

	if (ls->ls_recover_list_count != 0) {
		log_error(ls, "warning: recover_list_count %d",
			  ls->ls_recover_list_count);
		ls->ls_recover_list_count = 0;
	}
	spin_unlock(&ls->ls_recover_idr_lock);
}

378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403

/* Master recovery: find new master node for rsb's that were
   mastered on nodes that have been removed.

   dlm_recover_masters
   recover_master
   dlm_send_rcom_lookup            ->  receive_rcom_lookup
                                       dlm_dir_lookup
   receive_rcom_lookup_reply       <-
   dlm_recover_master_reply
   set_new_master
   set_master_lkbs
   set_lock_master
*/

/*
 * Set the lock master for all LKBs in a lock queue
 * If we are the new master of the rsb, we may have received new
 * MSTCPY locks from other nodes already which we need to ignore
 * when setting the new nodeid.
 */

static void set_lock_master(struct list_head *queue, int nodeid)
{
	struct dlm_lkb *lkb;

D
David Teigland 已提交
404 405
	list_for_each_entry(lkb, queue, lkb_statequeue) {
		if (!(lkb->lkb_flags & DLM_IFL_MSTCPY)) {
406
			lkb->lkb_nodeid = nodeid;
D
David Teigland 已提交
407 408 409
			lkb->lkb_remid = 0;
		}
	}
410 411 412 413 414 415 416 417 418 419
}

static void set_master_lkbs(struct dlm_rsb *r)
{
	set_lock_master(&r->res_grantqueue, r->res_nodeid);
	set_lock_master(&r->res_convertqueue, r->res_nodeid);
	set_lock_master(&r->res_waitqueue, r->res_nodeid);
}

/*
L
Lucas De Marchi 已提交
420
 * Propagate the new master nodeid to locks
421
 * The NEW_MASTER flag tells dlm_recover_locks() which rsb's to consider.
D
David Teigland 已提交
422
 * The NEW_MASTER2 flag tells recover_lvb() and recover_grant() which
D
David Teigland 已提交
423
 * rsb's to consider.
424 425
 */

426
static void set_new_master(struct dlm_rsb *r)
427 428 429 430 431 432 433 434 435
{
	set_master_lkbs(r);
	rsb_set_flag(r, RSB_NEW_MASTER);
	rsb_set_flag(r, RSB_NEW_MASTER2);
}

/*
 * We do async lookups on rsb's that need new masters.  The rsb's
 * waiting for a lookup reply are kept on the recover_list.
436 437 438 439 440
 *
 * Another node recovering the master may have sent us a rcom lookup,
 * and our dlm_master_lookup() set it as the new master, along with
 * NEW_MASTER so that we'll recover it here (this implies dir_nodeid
 * equals our_nodeid below).
441 442
 */

443
static int recover_master(struct dlm_rsb *r, unsigned int *count)
444 445
{
	struct dlm_ls *ls = r->res_ls;
446 447 448 449 450 451 452 453 454 455 456 457 458 459
	int our_nodeid, dir_nodeid;
	int is_removed = 0;
	int error;

	if (is_master(r))
		return 0;

	is_removed = dlm_is_removed(ls, r->res_nodeid);

	if (!is_removed && !rsb_flag(r, RSB_NEW_MASTER))
		return 0;

	our_nodeid = dlm_our_nodeid();
	dir_nodeid = dlm_dir_nodeid(r);
460 461

	if (dir_nodeid == our_nodeid) {
462 463 464 465
		if (is_removed) {
			r->res_master_nodeid = our_nodeid;
			r->res_nodeid = 0;
		}
466

467 468 469 470 471
		/* set master of lkbs to ourself when is_removed, or to
		   another new master which we set along with NEW_MASTER
		   in dlm_master_lookup */
		set_new_master(r);
		error = 0;
472
	} else {
473
		recover_idr_add(r);
474 475 476
		error = dlm_send_rcom_lookup(r, dir_nodeid);
	}

477
	(*count)++;
478 479 480 481
	return error;
}

/*
D
David Teigland 已提交
482 483 484 485 486 487 488 489 490 491 492 493
 * All MSTCPY locks are purged and rebuilt, even if the master stayed the same.
 * This is necessary because recovery can be started, aborted and restarted,
 * causing the master nodeid to briefly change during the aborted recovery, and
 * change back to the original value in the second recovery.  The MSTCPY locks
 * may or may not have been purged during the aborted recovery.  Another node
 * with an outstanding request in waiters list and a request reply saved in the
 * requestqueue, cannot know whether it should ignore the reply and resend the
 * request, or accept the reply and complete the request.  It must do the
 * former if the remote node purged MSTCPY locks, and it must do the later if
 * the remote node did not.  This is solved by always purging MSTCPY locks, in
 * which case, the request reply would always be ignored and the request
 * resent.
494 495
 */

496
static int recover_master_static(struct dlm_rsb *r, unsigned int *count)
497
{
D
David Teigland 已提交
498 499
	int dir_nodeid = dlm_dir_nodeid(r);
	int new_master = dir_nodeid;
500

D
David Teigland 已提交
501 502
	if (dir_nodeid == dlm_our_nodeid())
		new_master = 0;
503

D
David Teigland 已提交
504
	dlm_purge_mstcpy_locks(r);
505 506 507 508 509
	r->res_master_nodeid = dir_nodeid;
	r->res_nodeid = new_master;
	set_new_master(r);
	(*count)++;
	return 0;
510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
}

/*
 * Go through local root resources and for each rsb which has a master which
 * has departed, get the new master nodeid from the directory.  The dir will
 * assign mastery to the first node to look up the new master.  That means
 * we'll discover in this lookup if we're the new master of any rsb's.
 *
 * We fire off all the dir lookup requests individually and asynchronously to
 * the correct dir node.
 */

int dlm_recover_masters(struct dlm_ls *ls)
{
	struct dlm_rsb *r;
525 526 527 528
	unsigned int total = 0;
	unsigned int count = 0;
	int nodir = dlm_no_directory(ls);
	int error;
529 530 531 532 533 534 535 536 537 538 539

	log_debug(ls, "dlm_recover_masters");

	down_read(&ls->ls_root_sem);
	list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
		if (dlm_recovery_stopped(ls)) {
			up_read(&ls->ls_root_sem);
			error = -EINTR;
			goto out;
		}

540 541 542 543 544 545 546 547
		lock_rsb(r);
		if (nodir)
			error = recover_master_static(r, &count);
		else
			error = recover_master(r, &count);
		unlock_rsb(r);
		cond_resched();
		total++;
548

549 550 551 552
		if (error) {
			up_read(&ls->ls_root_sem);
			goto out;
		}
553 554 555
	}
	up_read(&ls->ls_root_sem);

556
	log_debug(ls, "dlm_recover_masters %u of %u", count, total);
557

558
	error = dlm_wait_function(ls, &recover_idr_empty);
559 560
 out:
	if (error)
561
		recover_idr_clear(ls);
562 563 564 565 566 567
	return error;
}

int dlm_recover_master_reply(struct dlm_ls *ls, struct dlm_rcom *rc)
{
	struct dlm_rsb *r;
568
	int ret_nodeid, new_master;
569

570
	r = recover_idr_find(ls, rc->rc_id);
571
	if (!r) {
572
		log_error(ls, "dlm_recover_master_reply no id %llx",
573
			  (unsigned long long)rc->rc_id);
574 575 576
		goto out;
	}

577 578 579 580 581 582
	ret_nodeid = rc->rc_result;

	if (ret_nodeid == dlm_our_nodeid())
		new_master = 0;
	else
		new_master = ret_nodeid;
583

D
David Teigland 已提交
584
	lock_rsb(r);
585 586 587
	r->res_master_nodeid = ret_nodeid;
	r->res_nodeid = new_master;
	set_new_master(r);
D
David Teigland 已提交
588
	unlock_rsb(r);
589
	recover_idr_del(r);
590

591
	if (recover_idr_empty(ls))
592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636
		wake_up(&ls->ls_wait_general);
 out:
	return 0;
}


/* Lock recovery: rebuild the process-copy locks we hold on a
   remastered rsb on the new rsb master.

   dlm_recover_locks
   recover_locks
   recover_locks_queue
   dlm_send_rcom_lock              ->  receive_rcom_lock
                                       dlm_recover_master_copy
   receive_rcom_lock_reply         <-
   dlm_recover_process_copy
*/


/*
 * keep a count of the number of lkb's we send to the new master; when we get
 * an equal number of replies then recovery for the rsb is done
 */

static int recover_locks_queue(struct dlm_rsb *r, struct list_head *head)
{
	struct dlm_lkb *lkb;
	int error = 0;

	list_for_each_entry(lkb, head, lkb_statequeue) {
	   	error = dlm_send_rcom_lock(r, lkb);
		if (error)
			break;
		r->res_recover_locks_count++;
	}

	return error;
}

static int recover_locks(struct dlm_rsb *r)
{
	int error = 0;

	lock_rsb(r);

637
	DLM_ASSERT(!r->res_recover_locks_count, dlm_dump_rsb(r););
638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688

	error = recover_locks_queue(r, &r->res_grantqueue);
	if (error)
		goto out;
	error = recover_locks_queue(r, &r->res_convertqueue);
	if (error)
		goto out;
	error = recover_locks_queue(r, &r->res_waitqueue);
	if (error)
		goto out;

	if (r->res_recover_locks_count)
		recover_list_add(r);
	else
		rsb_clear_flag(r, RSB_NEW_MASTER);
 out:
	unlock_rsb(r);
	return error;
}

int dlm_recover_locks(struct dlm_ls *ls)
{
	struct dlm_rsb *r;
	int error, count = 0;

	down_read(&ls->ls_root_sem);
	list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
		if (is_master(r)) {
			rsb_clear_flag(r, RSB_NEW_MASTER);
			continue;
		}

		if (!rsb_flag(r, RSB_NEW_MASTER))
			continue;

		if (dlm_recovery_stopped(ls)) {
			error = -EINTR;
			up_read(&ls->ls_root_sem);
			goto out;
		}

		error = recover_locks(r);
		if (error) {
			up_read(&ls->ls_root_sem);
			goto out;
		}

		count += r->res_recover_locks_count;
	}
	up_read(&ls->ls_root_sem);

D
David Teigland 已提交
689
	log_debug(ls, "dlm_recover_locks %d out", count);
690 691 692 693 694 695 696 697 698 699

	error = dlm_wait_function(ls, &recover_list_empty);
 out:
	if (error)
		recover_list_clear(ls);
	return error;
}

void dlm_recovered_lock(struct dlm_rsb *r)
{
700
	DLM_ASSERT(rsb_flag(r, RSB_NEW_MASTER), dlm_dump_rsb(r););
701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716

	r->res_recover_locks_count--;
	if (!r->res_recover_locks_count) {
		rsb_clear_flag(r, RSB_NEW_MASTER);
		recover_list_del(r);
	}

	if (recover_list_empty(r->res_ls))
		wake_up(&r->res_ls->ls_wait_general);
}

/*
 * The lvb needs to be recovered on all master rsb's.  This includes setting
 * the VALNOTVALID flag if necessary, and determining the correct lvb contents
 * based on the lvb's of the locks held on the rsb.
 *
717 718 719 720 721 722 723 724
 * RSB_VALNOTVALID is set in two cases:
 *
 * 1. we are master, but not new, and we purged an EX/PW lock held by a
 * failed node (in dlm_recover_purge which set RSB_RECOVER_LVB_INVAL)
 *
 * 2. we are a new master, and there are only NL/CR locks left.
 * (We could probably improve this by only invaliding in this way when
 * the previous master left uncleanly.  VMS docs mention that.)
725 726 727 728 729 730 731 732 733 734 735
 *
 * The LVB contents are only considered for changing when this is a new master
 * of the rsb (NEW_MASTER2).  Then, the rsb's lvb is taken from any lkb with
 * mode > CR.  If no lkb's exist with mode above CR, the lvb contents are taken
 * from the lkb with the largest lvb sequence number.
 */

static void recover_lvb(struct dlm_rsb *r)
{
	struct dlm_lkb *lkb, *high_lkb = NULL;
	uint32_t high_seq = 0;
736 737
	int lock_lvb_exists = 0;
	int big_lock_exists = 0;
738 739
	int lvblen = r->res_ls->ls_lvblen;

740 741 742 743 744 745 746 747 748 749 750 751 752
	if (!rsb_flag(r, RSB_NEW_MASTER2) &&
	    rsb_flag(r, RSB_RECOVER_LVB_INVAL)) {
		/* case 1 above */
		rsb_set_flag(r, RSB_VALNOTVALID);
		return;
	}

	if (!rsb_flag(r, RSB_NEW_MASTER2))
		return;

	/* we are the new master, so figure out if VALNOTVALID should
	   be set, and set the rsb lvb from the best lkb available. */

753 754 755 756
	list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
		if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
			continue;

757
		lock_lvb_exists = 1;
758 759

		if (lkb->lkb_grmode > DLM_LOCK_CR) {
760
			big_lock_exists = 1;
761 762 763 764 765 766 767 768 769 770 771 772 773
			goto setflag;
		}

		if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
			high_lkb = lkb;
			high_seq = lkb->lkb_lvbseq;
		}
	}

	list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
		if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
			continue;

774
		lock_lvb_exists = 1;
775 776

		if (lkb->lkb_grmode > DLM_LOCK_CR) {
777
			big_lock_exists = 1;
778 779 780 781 782 783 784 785 786 787 788 789 790
			goto setflag;
		}

		if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
			high_lkb = lkb;
			high_seq = lkb->lkb_lvbseq;
		}
	}

 setflag:
	if (!lock_lvb_exists)
		goto out;

791
	/* lvb is invalidated if only NL/CR locks remain */
792 793 794 795
	if (!big_lock_exists)
		rsb_set_flag(r, RSB_VALNOTVALID);

	if (!r->res_lvbptr) {
796
		r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819
		if (!r->res_lvbptr)
			goto out;
	}

	if (big_lock_exists) {
		r->res_lvbseq = lkb->lkb_lvbseq;
		memcpy(r->res_lvbptr, lkb->lkb_lvbptr, lvblen);
	} else if (high_lkb) {
		r->res_lvbseq = high_lkb->lkb_lvbseq;
		memcpy(r->res_lvbptr, high_lkb->lkb_lvbptr, lvblen);
	} else {
		r->res_lvbseq = 0;
		memset(r->res_lvbptr, 0, lvblen);
	}
 out:
	return;
}

/* All master rsb's flagged RECOVER_CONVERT need to be looked at.  The locks
   converting PR->CW or CW->PR need to have their lkb_grmode set. */

static void recover_conversion(struct dlm_rsb *r)
{
820
	struct dlm_ls *ls = r->res_ls;
821 822 823 824 825 826 827 828 829 830 831 832 833 834
	struct dlm_lkb *lkb;
	int grmode = -1;

	list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
		if (lkb->lkb_grmode == DLM_LOCK_PR ||
		    lkb->lkb_grmode == DLM_LOCK_CW) {
			grmode = lkb->lkb_grmode;
			break;
		}
	}

	list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
		if (lkb->lkb_grmode != DLM_LOCK_IV)
			continue;
835 836 837
		if (grmode == -1) {
			log_debug(ls, "recover_conversion %x set gr to rq %d",
				  lkb->lkb_id, lkb->lkb_rqmode);
838
			lkb->lkb_grmode = lkb->lkb_rqmode;
839 840 841
		} else {
			log_debug(ls, "recover_conversion %x set gr %d",
				  lkb->lkb_id, grmode);
842
			lkb->lkb_grmode = grmode;
843
		}
844 845 846
	}
}

D
David Teigland 已提交
847
/* We've become the new master for this rsb and waiting/converting locks may
D
David Teigland 已提交
848
   need to be granted in dlm_recover_grant() due to locks that may have
D
David Teigland 已提交
849 850
   existed from a removed node. */

D
David Teigland 已提交
851
static void recover_grant(struct dlm_rsb *r)
D
David Teigland 已提交
852 853
{
	if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
D
David Teigland 已提交
854
		rsb_set_flag(r, RSB_RECOVER_GRANT);
D
David Teigland 已提交
855 856
}

857 858 859
void dlm_recover_rsbs(struct dlm_ls *ls)
{
	struct dlm_rsb *r;
D
David Teigland 已提交
860
	unsigned int count = 0;
861 862 863 864 865 866 867

	down_read(&ls->ls_root_sem);
	list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
		lock_rsb(r);
		if (is_master(r)) {
			if (rsb_flag(r, RSB_RECOVER_CONVERT))
				recover_conversion(r);
868 869 870 871 872

			/* recover lvb before granting locks so the updated
			   lvb/VALNOTVALID is presented in the completion */
			recover_lvb(r);

D
David Teigland 已提交
873
			if (rsb_flag(r, RSB_NEW_MASTER2))
D
David Teigland 已提交
874
				recover_grant(r);
875
			count++;
876 877
		} else {
			rsb_clear_flag(r, RSB_VALNOTVALID);
878 879
		}
		rsb_clear_flag(r, RSB_RECOVER_CONVERT);
880
		rsb_clear_flag(r, RSB_RECOVER_LVB_INVAL);
D
David Teigland 已提交
881
		rsb_clear_flag(r, RSB_NEW_MASTER2);
882 883 884 885
		unlock_rsb(r);
	}
	up_read(&ls->ls_root_sem);

D
David Teigland 已提交
886 887
	if (count)
		log_debug(ls, "dlm_recover_rsbs %d done", count);
888 889 890 891 892 893
}

/* Create a single list of all root rsb's to be used during recovery */

int dlm_create_root_list(struct dlm_ls *ls)
{
B
Bob Peterson 已提交
894
	struct rb_node *n;
895 896 897 898 899 900 901 902 903 904 905
	struct dlm_rsb *r;
	int i, error = 0;

	down_write(&ls->ls_root_sem);
	if (!list_empty(&ls->ls_root_list)) {
		log_error(ls, "root list not empty");
		error = -EINVAL;
		goto out;
	}

	for (i = 0; i < ls->ls_rsbtbl_size; i++) {
906
		spin_lock(&ls->ls_rsbtbl[i].lock);
B
Bob Peterson 已提交
907 908
		for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
			r = rb_entry(n, struct dlm_rsb, res_hashnode);
909 910 911
			list_add(&r->res_root_list, &ls->ls_root_list);
			dlm_hold_rsb(r);
		}
912

913 914
		if (!RB_EMPTY_ROOT(&ls->ls_rsbtbl[i].toss))
			log_error(ls, "dlm_create_root_list toss not empty");
915
		spin_unlock(&ls->ls_rsbtbl[i].lock);
916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933
	}
 out:
	up_write(&ls->ls_root_sem);
	return error;
}

void dlm_release_root_list(struct dlm_ls *ls)
{
	struct dlm_rsb *r, *safe;

	down_write(&ls->ls_root_sem);
	list_for_each_entry_safe(r, safe, &ls->ls_root_list, res_root_list) {
		list_del_init(&r->res_root_list);
		dlm_put_rsb(r);
	}
	up_write(&ls->ls_root_sem);
}

934
void dlm_clear_toss(struct dlm_ls *ls)
935
{
B
Bob Peterson 已提交
936
	struct rb_node *n, *next;
937 938
	struct dlm_rsb *r;
	unsigned int count = 0;
939 940 941
	int i;

	for (i = 0; i < ls->ls_rsbtbl_size; i++) {
942
		spin_lock(&ls->ls_rsbtbl[i].lock);
B
Bob Peterson 已提交
943
		for (n = rb_first(&ls->ls_rsbtbl[i].toss); n; n = next) {
944 945 946 947 948
			next = rb_next(n);
			r = rb_entry(n, struct dlm_rsb, res_hashnode);
			rb_erase(n, &ls->ls_rsbtbl[i].toss);
			dlm_free_rsb(r);
			count++;
949
		}
950
		spin_unlock(&ls->ls_rsbtbl[i].lock);
951
	}
952 953 954

	if (count)
		log_debug(ls, "dlm_clear_toss %u done", count);
955 956
}