lock_deadlock.c   [plain text]

/*-
 * See the file LICENSE for redistribution information.
 *
 * Copyright (c) 1996-2003
 *	Sleepycat Software.  All rights reserved.
 */

#include "db_config.h"

#ifndef lint
static const char revid[] = "$Id: lock_deadlock.c,v 1.2 2004/03/30 01:23:43 jtownsen Exp $";
#endif /* not lint */

#ifndef NO_SYSTEM_INCLUDES
#include <sys/types.h>

#include <string.h>
#endif

#include "db_int.h"
#include "dbinc/db_shash.h"
#include "dbinc/lock.h"
#include "dbinc/log.h"
#include "dbinc/txn.h"

#define	ISSET_MAP(M, N)	((M)[(N) / 32] & (1 << (N) % 32))

#define	CLEAR_MAP(M, N) {						\
	u_int32_t __i;							\
	for (__i = 0; __i < (N); __i++)					\
		(M)[__i] = 0;						\
}

#define	SET_MAP(M, B)	((M)[(B) / 32] |= (1 << ((B) % 32)))
#define	CLR_MAP(M, B)	((M)[(B) / 32] &= ~(1 << ((B) % 32)))

#define	OR_MAP(D, S, N)	{						\
	u_int32_t __i;							\
	for (__i = 0; __i < (N); __i++)					\
		D[__i] |= S[__i];					\
}
#define	BAD_KILLID	0xffffffff

typedef struct {
	int		valid;
	int		self_wait;
	int		in_abort;
	u_int32_t	count;
	u_int32_t	id;
	u_int32_t	last_lock;
	ssize_t		last_obj;
	u_int32_t	last_locker_id;
	db_pgno_t	pgno;
} locker_info;

static int __dd_abort __P((DB_ENV *, locker_info *));
static int __dd_build __P((DB_ENV *,
	    u_int32_t, u_int32_t **, u_int32_t *, u_int32_t *, locker_info **));
static int __dd_find __P((DB_ENV *,
	    u_int32_t *, locker_info *, u_int32_t, u_int32_t, u_int32_t ***));
static int __dd_isolder __P((u_int32_t, u_int32_t, u_int32_t, u_int32_t));
static int __dd_verify __P((locker_info *, u_int32_t *, u_int32_t *,
	    u_int32_t *, u_int32_t, u_int32_t, u_int32_t));

#ifdef DIAGNOSTIC
static void __dd_debug
	    __P((DB_ENV *, locker_info *, u_int32_t *, u_int32_t, u_int32_t));
#endif

/*
 * __lock_detect_pp --
 *	DB_ENV->lock_detect pre/post processing.
 *
 * PUBLIC: int __lock_detect_pp __P((DB_ENV *, u_int32_t, u_int32_t, int *));
 */
int
__lock_detect_pp(dbenv, flags, atype, abortp)
	DB_ENV *dbenv;
	u_int32_t flags, atype;
	int *abortp;
{
	int ret, rep_check;

	PANIC_CHECK(dbenv);
	ENV_REQUIRES_CONFIG(dbenv,
	    dbenv->lk_handle, "DB_ENV->lock_detect", DB_INIT_LOCK);

	/* Validate arguments. */
	if ((ret = __db_fchk(dbenv, "DB_ENV->lock_detect", flags, 0)) != 0)
		return (ret);
	switch (atype) {
	case DB_LOCK_DEFAULT:
	case DB_LOCK_EXPIRE:
	case DB_LOCK_MAXLOCKS:
	case DB_LOCK_MINLOCKS:
	case DB_LOCK_MINWRITE:
	case DB_LOCK_OLDEST:
	case DB_LOCK_RANDOM:
	case DB_LOCK_YOUNGEST:
		break;
	default:
		__db_err(dbenv,
	    "DB_ENV->lock_detect: unknown deadlock detection mode specified");
		return (EINVAL);
	}

	rep_check = IS_ENV_REPLICATED(dbenv) ? 1 : 0;
	if (rep_check)
		__env_rep_enter(dbenv);
	ret = __lock_detect(dbenv, atype, abortp);
	if (rep_check)
		__env_rep_exit(dbenv);
	return (ret);
}

/*
 * __lock_detect --
 *	DB_ENV->lock_detect.
 *
 * PUBLIC: int __lock_detect __P((DB_ENV *, u_int32_t, int *));
 */
int
__lock_detect(dbenv, atype, abortp)
	DB_ENV *dbenv;
	u_int32_t atype;
	int *abortp;
{
	DB_LOCKREGION *region;
	DB_LOCKTAB *lt;
	DB_TXNMGR *tmgr;
	db_timeval_t now;
	locker_info *idmap;
	u_int32_t *bitmap, *copymap, **deadp, **free_me, *tmpmap;
	u_int32_t i, keeper, killid, limit, nalloc, nlockers;
	u_int32_t lock_max, txn_max;
	int ret;

	/*
	 * If this environment is a replication client, then we must use the
	 * MINWRITE detection discipline.
	 */
	if (__rep_is_client(dbenv))
		atype = DB_LOCK_MINWRITE;

	free_me = NULL;

	lt = dbenv->lk_handle;
	if (abortp != NULL)
		*abortp = 0;

	/* Check if a detector run is necessary. */
	LOCKREGION(dbenv, lt);

	/* Make a pass only if auto-detect would run. */
	region = lt->reginfo.primary;

	LOCK_SET_TIME_INVALID(&now);
	if (region->need_dd == 0 &&
	     (!LOCK_TIME_ISVALID(&region->next_timeout) ||
	     !__lock_expired(dbenv, &now, &region->next_timeout))) {
		UNLOCKREGION(dbenv, lt);
		return (0);
	}
	if (region->need_dd == 0)
		atype = DB_LOCK_EXPIRE;

	/* Reset need_dd, so we know we've run the detector. */
	region->need_dd = 0;

	/* Build the waits-for bitmap. */
	ret = __dd_build(dbenv, atype, &bitmap, &nlockers, &nalloc, &idmap);
	lock_max = region->stat.st_cur_maxid;
	UNLOCKREGION(dbenv, lt);

	/*
	 * We need the cur_maxid from the txn region as well.  In order
	 * to avoid tricky synchronization between the lock and txn
	 * regions, we simply unlock the lock region and then lock the
	 * txn region.  This introduces a small window during which the
	 * transaction system could then wrap.  We're willing to return
	 * the wrong answer for "oldest" or "youngest" in those rare
	 * circumstances.
	 */
	tmgr = dbenv->tx_handle;
	if (tmgr != NULL) {
		R_LOCK(dbenv, &tmgr->reginfo);
		txn_max = ((DB_TXNREGION *)tmgr->reginfo.primary)->cur_maxid;
		R_UNLOCK(dbenv, &tmgr->reginfo);
	} else
		txn_max = TXN_MAXIMUM;
	if (ret != 0 || atype == DB_LOCK_EXPIRE)
		return (ret);

	if (nlockers == 0)
		return (0);
#ifdef DIAGNOSTIC
	if (FLD_ISSET(dbenv->verbose, DB_VERB_WAITSFOR))
		__dd_debug(dbenv, idmap, bitmap, nlockers, nalloc);
#endif
	/* Now duplicate the bitmaps so we can verify deadlock participants. */
	if ((ret = __os_calloc(dbenv, (size_t)nlockers,
	    sizeof(u_int32_t) * nalloc, &copymap)) != 0)
		goto err;
	memcpy(copymap, bitmap, nlockers * sizeof(u_int32_t) * nalloc);

	if ((ret = __os_calloc(dbenv, sizeof(u_int32_t), nalloc, &tmpmap)) != 0)
		goto err1;

	/* Find a deadlock. */
	if ((ret =
	    __dd_find(dbenv, bitmap, idmap, nlockers, nalloc, &deadp)) != 0)
		return (ret);

	killid = BAD_KILLID;
	free_me = deadp;
	for (; *deadp != NULL; deadp++) {
		if (abortp != NULL)
			++*abortp;
		killid = (u_int32_t)((*deadp - bitmap) / nalloc);
		limit = killid;
		keeper = BAD_KILLID;

		if (atype == DB_LOCK_DEFAULT || atype == DB_LOCK_RANDOM)
			goto dokill;
		/*
		 * It's conceivable that under XA, the locker could
		 * have gone away.
		 */
		if (killid == BAD_KILLID)
			break;

		/*
		 * Start with the id that we know is deadlocked
		 * and then examine all other set bits and see
		 * if any are a better candidate for abortion
		 * and that they are genuinely part of the
		 * deadlock.  The definition of "best":
		 * OLDEST: smallest id
		 * YOUNGEST: largest id
		 * MAXLOCKS: maximum count
		 * MINLOCKS: minimum count
		 * MINWRITE: minimum count
		 */

		for (i = (killid + 1) % nlockers;
		    i != limit;
		    i = (i + 1) % nlockers) {
			if (!ISSET_MAP(*deadp, i) || idmap[i].in_abort)
				continue;
			switch (atype) {
			case DB_LOCK_OLDEST:
				if (__dd_isolder(idmap[killid].id,
				    idmap[i].id, lock_max, txn_max))
					continue;
				keeper = i;
				break;
			case DB_LOCK_YOUNGEST:
				if (__dd_isolder(idmap[i].id,
				    idmap[killid].id, lock_max, txn_max))
					continue;
				keeper = i;
				break;
			case DB_LOCK_MAXLOCKS:
				if (idmap[i].count < idmap[killid].count)
					continue;
				keeper = i;
				break;
			case DB_LOCK_MINLOCKS:
			case DB_LOCK_MINWRITE:
				if (idmap[i].count > idmap[killid].count)
					continue;
				keeper = i;
				break;
			default:
				killid = BAD_KILLID;
				ret = EINVAL;
				goto dokill;
			}
			if (__dd_verify(idmap, *deadp,
			    tmpmap, copymap, nlockers, nalloc, i))
				killid = i;
		}

dokill:		if (killid == BAD_KILLID)
			continue;

		/*
		 * There are cases in which our general algorithm will
		 * fail.  Returning 1 from verify indicates that the
		 * particular locker is not only involved in a deadlock,
		 * but that killing him will allow others to make forward
		 * progress.  Unfortunately, there are cases where we need
		 * to abort someone, but killing them will not necessarily
		 * ensure forward progress (imagine N readers all trying to
		 * acquire a write lock).  In such a scenario, we'll have
		 * gotten all the way through the loop, we will have found
		 * someone to keep (keeper will be valid), but killid will
		 * still be the initial deadlocker.  In this case, if the
		 * initial killid satisfies __dd_verify, kill it, else abort
		 * keeper and indicate that we need to run deadlock detection
		 * again.
		 */

		if (keeper != BAD_KILLID && killid == limit &&
		    __dd_verify(idmap, *deadp,
		    tmpmap, copymap, nlockers, nalloc, killid) == 0) {
			LOCKREGION(dbenv, lt);
			region->need_dd = 1;
			UNLOCKREGION(dbenv, lt);
			killid = keeper;
		}

		/* Kill the locker with lockid idmap[killid]. */
		if ((ret = __dd_abort(dbenv, &idmap[killid])) != 0) {
			/*
			 * It's possible that the lock was already aborted;
			 * this isn't necessarily a problem, so do not treat
			 * it as an error.
			 */
			if (ret == DB_ALREADY_ABORTED)
				ret = 0;
			else
				__db_err(dbenv,
				    "warning: unable to abort locker %lx",
				    (u_long)idmap[killid].id);
		} else if (FLD_ISSET(dbenv->verbose, DB_VERB_DEADLOCK))
			__db_err(dbenv,
			    "Aborting locker %lx", (u_long)idmap[killid].id);
	}
	__os_free(dbenv, tmpmap);
err1:	__os_free(dbenv, copymap);

err:	if (free_me != NULL)
		__os_free(dbenv, free_me);
	__os_free(dbenv, bitmap);
	__os_free(dbenv, idmap);

	return (ret);
}

/*
 * ========================================================================
 * Utilities
 */

# define DD_INVALID_ID	((u_int32_t) -1)

static int
__dd_build(dbenv, atype, bmp, nlockers, allocp, idmap)
	DB_ENV *dbenv;
	u_int32_t atype, **bmp, *nlockers, *allocp;
	locker_info **idmap;
{
	struct __db_lock *lp;
	DB_LOCKER *lip, *lockerp, *child;
	DB_LOCKOBJ *op, *lo;
	DB_LOCKREGION *region;
	DB_LOCKTAB *lt;
	locker_info *id_array;
	db_timeval_t now, min_timeout;
	u_int32_t *bitmap, count, dd, *entryp, id, ndx, nentries, *tmpmap;
	u_int8_t *pptr;
	int expire_only, is_first, ret;

	lt = dbenv->lk_handle;
	region = lt->reginfo.primary;
	LOCK_SET_TIME_INVALID(&now);
	LOCK_SET_TIME_MAX(&min_timeout);
	expire_only = atype == DB_LOCK_EXPIRE;

	/*
	 * While we always check for expired timeouts, if we are called
	 * with DB_LOCK_EXPIRE, then we are only checking for timeouts
	 * (i.e., not doing deadlock detection at all).  If we aren't
	 * doing real deadlock detection, then we can skip a significant,
	 * amount of the processing.  In particular we do not build
	 * the conflict array and our caller needs to expect this.
	 */
	if (expire_only) {
		count = 0;
		nentries = 0;
		goto obj_loop;
	}

	/*
	 * We'll check how many lockers there are, add a few more in for
	 * good measure and then allocate all the structures.  Then we'll
	 * verify that we have enough room when we go back in and get the
	 * mutex the second time.
	 */
retry:	count = region->stat.st_nlockers;

	if (count == 0) {
		*nlockers = 0;
		return (0);
	}

	if (FLD_ISSET(dbenv->verbose, DB_VERB_DEADLOCK))
		__db_err(dbenv, "%lu lockers", (u_long)count);

	count += 20;
	nentries = ALIGN(count, 32) / 32;

	/*
	 * Allocate enough space for a count by count bitmap matrix.
	 *
	 * XXX
	 * We can probably save the malloc's between iterations just
	 * reallocing if necessary because count grew by too much.
	 */
	if ((ret = __os_calloc(dbenv, (size_t)count,
	    sizeof(u_int32_t) * nentries, &bitmap)) != 0)
		return (ret);

	if ((ret = __os_calloc(dbenv,
	    sizeof(u_int32_t), nentries, &tmpmap)) != 0) {
		__os_free(dbenv, bitmap);
		return (ret);
	}

	if ((ret = __os_calloc(dbenv,
	    (size_t)count, sizeof(locker_info), &id_array)) != 0) {
		__os_free(dbenv, bitmap);
		__os_free(dbenv, tmpmap);
		return (ret);
	}

	/*
	 * Now go back in and actually fill in the matrix.
	 */
	if (region->stat.st_nlockers > count) {
		__os_free(dbenv, bitmap);
		__os_free(dbenv, tmpmap);
		__os_free(dbenv, id_array);
		goto retry;
	}

	/*
	 * First we go through and assign each locker a deadlock detector id.
	 */
	for (id = 0, lip = SH_TAILQ_FIRST(&region->lockers, __db_locker);
	    lip != NULL;
	    lip = SH_TAILQ_NEXT(lip, ulinks, __db_locker)) {
		if (lip->master_locker == INVALID_ROFF) {
			lip->dd_id = id++;
			id_array[lip->dd_id].id = lip->id;
			if (atype == DB_LOCK_MINLOCKS ||
			    atype == DB_LOCK_MAXLOCKS)
				id_array[lip->dd_id].count = lip->nlocks;
			if (atype == DB_LOCK_MINWRITE)
				id_array[lip->dd_id].count = lip->nwrites;
			if (F_ISSET(lip, DB_LOCKER_INABORT))
				id_array[lip->dd_id].in_abort = 1;
		} else
			lip->dd_id = DD_INVALID_ID;

	}

	/*
	 * We only need consider objects that have waiters, so we use
	 * the list of objects with waiters (dd_objs) instead of traversing
	 * the entire hash table.  For each object, we traverse the waiters
	 * list and add an entry in the waitsfor matrix for each waiter/holder
	 * combination.
	 */
obj_loop:
	for (op = SH_TAILQ_FIRST(&region->dd_objs, __db_lockobj);
	    op != NULL; op = SH_TAILQ_NEXT(op, dd_links, __db_lockobj)) {
		if (expire_only)
			goto look_waiters;
		CLEAR_MAP(tmpmap, nentries);

		/*
		 * First we go through and create a bit map that
		 * represents all the holders of this object.
		 */
		for (lp = SH_TAILQ_FIRST(&op->holders, __db_lock);
		    lp != NULL;
		    lp = SH_TAILQ_NEXT(lp, links, __db_lock)) {
			LOCKER_LOCK(lt, region, lp->holder, ndx);
			if ((ret = __lock_getlocker(lt,
			    lp->holder, ndx, 0, &lockerp)) != 0)
				continue;

			if (lockerp->dd_id == DD_INVALID_ID) {
				dd = ((DB_LOCKER *)R_ADDR(&lt->reginfo,
				    lockerp->master_locker))->dd_id;
				lockerp->dd_id = dd;
				if (atype == DB_LOCK_MINLOCKS ||
				    atype == DB_LOCK_MAXLOCKS)
					id_array[dd].count += lockerp->nlocks;
				if (atype == DB_LOCK_MINWRITE)
					id_array[dd].count += lockerp->nwrites;
				if (F_ISSET(lockerp, DB_LOCKER_INABORT))
					id_array[dd].in_abort = 1;

			} else
				dd = lockerp->dd_id;
			id_array[dd].valid = 1;

			/*
			 * If the holder has already been aborted, then
			 * we should ignore it for now.
			 */
			if (lp->status == DB_LSTAT_HELD)
				SET_MAP(tmpmap, dd);
		}

		/*
		 * Next, for each waiter, we set its row in the matrix
		 * equal to the map of holders we set up above.
		 */
look_waiters:
		for (is_first = 1,
		    lp = SH_TAILQ_FIRST(&op->waiters, __db_lock);
		    lp != NULL;
		    is_first = 0,
		    lp = SH_TAILQ_NEXT(lp, links, __db_lock)) {
			LOCKER_LOCK(lt, region, lp->holder, ndx);
			if ((ret = __lock_getlocker(lt,
			    lp->holder, ndx, 0, &lockerp)) != 0)
				continue;
			if (lp->status == DB_LSTAT_WAITING) {
				if (__lock_expired(dbenv,
				    &now, &lockerp->lk_expire)) {
					lp->status = DB_LSTAT_EXPIRED;
					MUTEX_UNLOCK(dbenv, &lp->mutex);
					continue;
				}
				if (LOCK_TIME_GREATER(
				    &min_timeout, &lockerp->lk_expire))
					min_timeout = lockerp->lk_expire;

			}

			if (expire_only)
				continue;

			if (lockerp->dd_id == DD_INVALID_ID) {
				dd = ((DB_LOCKER *)R_ADDR(&lt->reginfo,
				    lockerp->master_locker))->dd_id;
				lockerp->dd_id = dd;
				if (atype == DB_LOCK_MINLOCKS ||
				    atype == DB_LOCK_MAXLOCKS)
					id_array[dd].count += lockerp->nlocks;
				if (atype == DB_LOCK_MINWRITE)
					id_array[dd].count += lockerp->nwrites;
			} else
				dd = lockerp->dd_id;
			id_array[dd].valid = 1;

			/*
			 * If the transaction is pending abortion, then
			 * ignore it on this iteration.
			 */
			if (lp->status != DB_LSTAT_WAITING)
				continue;

			entryp = bitmap + (nentries * dd);
			OR_MAP(entryp, tmpmap, nentries);
			/*
			 * If this is the first waiter on the queue,
			 * then we remove the waitsfor relationship
			 * with oneself.  However, if it's anywhere
			 * else on the queue, then we have to keep
			 * it and we have an automatic deadlock.
			 */
			if (is_first) {
				if (ISSET_MAP(entryp, dd))
					id_array[dd].self_wait = 1;
				CLR_MAP(entryp, dd);
			}
		}
	}

	if (LOCK_TIME_ISVALID(&region->next_timeout)) {
		if (LOCK_TIME_ISMAX(&min_timeout))
			LOCK_SET_TIME_INVALID(&region->next_timeout);
		else
			region->next_timeout = min_timeout;
	}
	if (expire_only)
		return (0);

	/* Now for each locker; record its last lock. */
	for (id = 0; id < count; id++) {
		if (!id_array[id].valid)
			continue;
		LOCKER_LOCK(lt, region, id_array[id].id, ndx);
		if ((ret = __lock_getlocker(lt,
		    id_array[id].id, ndx, 0, &lockerp)) != 0) {
			__db_err(dbenv,
			    "No locks for locker %lu", (u_long)id_array[id].id);
			continue;
		}

		/*
		 * If this is a master transaction, try to
		 * find one of its children's locks first,
		 * as they are probably more recent.
		 */
		child = SH_LIST_FIRST(&lockerp->child_locker, __db_locker);
		if (child != NULL) {
			do {
				lp = SH_LIST_FIRST(&child->heldby, __db_lock);
				if (lp != NULL &&
				    lp->status == DB_LSTAT_WAITING) {
					id_array[id].last_locker_id = child->id;
					goto get_lock;
				}
				child = SH_LIST_NEXT(
				    child, child_link, __db_locker);
			} while (child != NULL);
		}
		lp = SH_LIST_FIRST(&lockerp->heldby, __db_lock);
		if (lp != NULL) {
			id_array[id].last_locker_id = lockerp->id;
	get_lock:	id_array[id].last_lock = R_OFFSET(&lt->reginfo, lp);
			id_array[id].last_obj = lp->obj;
			lo = (DB_LOCKOBJ *)((u_int8_t *)lp + lp->obj);
			pptr = SH_DBT_PTR(&lo->lockobj);
			if (lo->lockobj.size >= sizeof(db_pgno_t))
				memcpy(&id_array[id].pgno,
				    pptr, sizeof(db_pgno_t));
			else
				id_array[id].pgno = 0;
		}
	}

	/*
	 * Pass complete, reset the deadlock detector bit.
	 */
	region->need_dd = 0;

	/*
	 * Now we can release everything except the bitmap matrix that we
	 * created.
	 */
	*nlockers = id;
	*idmap = id_array;
	*bmp = bitmap;
	*allocp = nentries;
	__os_free(dbenv, tmpmap);
	return (0);
}

static int
__dd_find(dbenv, bmp, idmap, nlockers, nalloc, deadp)
	DB_ENV *dbenv;
	u_int32_t *bmp, nlockers, nalloc;
	locker_info *idmap;
	u_int32_t ***deadp;
{
	u_int32_t i, j, k, *mymap, *tmpmap;
	u_int32_t **retp;
	int ndead, ndeadalloc, ret;

#undef	INITIAL_DEAD_ALLOC
#define	INITIAL_DEAD_ALLOC	8

	ndeadalloc = INITIAL_DEAD_ALLOC;
	ndead = 0;
	if ((ret = __os_malloc(dbenv,
	    ndeadalloc * sizeof(u_int32_t *), &retp)) != 0)
		return (ret);

	/*
	 * For each locker, OR in the bits from the lockers on which that
	 * locker is waiting.
	 */
	for (mymap = bmp, i = 0; i < nlockers; i++, mymap += nalloc) {
		if (!idmap[i].valid || idmap[i].in_abort)
			continue;
		for (j = 0; j < nlockers; j++) {
			if (!ISSET_MAP(mymap, j))
				continue;

			/* Find the map for this bit. */
			tmpmap = bmp + (nalloc * j);
			OR_MAP(mymap, tmpmap, nalloc);
			if (!ISSET_MAP(mymap, i))
				continue;

			/* Make sure we leave room for NULL. */
			if (ndead + 2 >= ndeadalloc) {
				ndeadalloc <<= 1;
				/*
				 * If the alloc fails, then simply return the
				 * deadlocks that we already have.
				 */
				if (__os_realloc(dbenv,
				    ndeadalloc * sizeof(u_int32_t),
				    &retp) != 0) {
					retp[ndead] = NULL;
					*deadp = retp;
					return (0);
				}
			}
			retp[ndead++] = mymap;

			/* Mark all participants in this deadlock invalid. */
			for (k = 0; k < nlockers; k++)
				if (ISSET_MAP(mymap, k))
					idmap[k].valid = 0;
			break;
		}
	}
	retp[ndead] = NULL;
	*deadp = retp;
	return (0);
}

static int
__dd_abort(dbenv, info)
	DB_ENV *dbenv;
	locker_info *info;
{
	struct __db_lock *lockp;
	DB_LOCKER *lockerp;
	DB_LOCKOBJ *sh_obj;
	DB_LOCKREGION *region;
	DB_LOCKTAB *lt;
	u_int32_t ndx;
	int ret;

	lt = dbenv->lk_handle;
	region = lt->reginfo.primary;

	LOCKREGION(dbenv, lt);

	/*
	 * Get the locker.  If its gone or was aborted while
	 * we were detecting return that.
	 */
	LOCKER_LOCK(lt, region, info->last_locker_id, ndx);
	if ((ret = __lock_getlocker(lt,
	    info->last_locker_id, ndx, 0, &lockerp)) != 0 ||
	    lockerp == NULL || F_ISSET(lockerp, DB_LOCKER_INABORT)) {
		if (ret == 0)
			ret = DB_ALREADY_ABORTED;
		goto out;
	}

	/*
	 * Find the locker's last lock.
	 * It is possible for this lock to have been freed,
	 * either though a timeout or another detector run.
	 */
	if ((lockp = SH_LIST_FIRST(&lockerp->heldby, __db_lock)) == NULL) {
		ret = DB_ALREADY_ABORTED;
		goto out;
	}
	if (R_OFFSET(&lt->reginfo, lockp) != info->last_lock ||
	    lockp->holder != lockerp->id ||
	    lockp->obj != info->last_obj || lockp->status != DB_LSTAT_WAITING) {
		ret = DB_ALREADY_ABORTED;
		goto out;
	}

	sh_obj = (DB_LOCKOBJ *)((u_int8_t *)lockp + lockp->obj);

	/* Abort lock, take it off list, and wake up this lock. */
	SHOBJECT_LOCK(lt, region, sh_obj, ndx);
	lockp->status = DB_LSTAT_ABORTED;
	SH_TAILQ_REMOVE(&sh_obj->waiters, lockp, links, __db_lock);

	/*
	 * Either the waiters list is now empty, in which case we remove
	 * it from dd_objs, or it is not empty, in which case we need to
	 * do promotion.
	 */
	if (SH_TAILQ_FIRST(&sh_obj->waiters, __db_lock) == NULL)
		SH_TAILQ_REMOVE(&region->dd_objs,
		    sh_obj, dd_links, __db_lockobj);
	else
		ret = __lock_promote(lt, sh_obj, 0);
	MUTEX_UNLOCK(dbenv, &lockp->mutex);

	region->stat.st_ndeadlocks++;
	UNLOCKREGION(dbenv, lt);

	return (0);

out:	UNLOCKREGION(dbenv, lt);
	return (ret);
}

#ifdef DIAGNOSTIC
static void
__dd_debug(dbenv, idmap, bitmap, nlockers, nalloc)
	DB_ENV *dbenv;
	locker_info *idmap;
	u_int32_t *bitmap, nlockers, nalloc;
{
	u_int32_t i, j, *mymap;
	char *msgbuf;

	__db_err(dbenv, "Waitsfor array\nWaiter:\tWaiting on:");

	/* Allocate space to print 10 bytes per item waited on. */
#undef	MSGBUF_LEN
#define	MSGBUF_LEN ((nlockers + 1) * 10 + 64)
	if (__os_malloc(dbenv, MSGBUF_LEN, &msgbuf) != 0)
		return;

	for (mymap = bitmap, i = 0; i < nlockers; i++, mymap += nalloc) {
		if (!idmap[i].valid)
			continue;
		sprintf(msgbuf,					/* Waiter. */
		    "%lx/%lu:\t", (u_long)idmap[i].id, (u_long)idmap[i].pgno);
		for (j = 0; j < nlockers; j++)
			if (ISSET_MAP(mymap, j))
				sprintf(msgbuf, "%s %lx", msgbuf,
				    (u_long)idmap[j].id);
		(void)sprintf(msgbuf,
		    "%s %lu", msgbuf, (u_long)idmap[i].last_lock);
		__db_err(dbenv, msgbuf);
	}

	__os_free(dbenv, msgbuf);
}
#endif

/*
 * Given a bitmap that contains a deadlock, verify that the bit
 * specified in the which parameter indicates a transaction that
 * is actually deadlocked.  Return 1 if really deadlocked, 0 otherwise.
 * deadmap is the array that identified the deadlock.
 * tmpmap is a copy of the initial bitmaps from the dd_build phase
 * origmap is a temporary bit map into which we can OR things
 * nlockers is the number of actual lockers under consideration
 * nalloc is the number of words allocated for the bitmap
 * which is the locker in question
 */
static int
__dd_verify(idmap, deadmap, tmpmap, origmap, nlockers, nalloc, which)
	locker_info *idmap;
	u_int32_t *deadmap, *tmpmap, *origmap;
	u_int32_t nlockers, nalloc, which;
{
	u_int32_t *tmap;
	u_int32_t j;
	int count;

	memset(tmpmap, 0, sizeof(u_int32_t) * nalloc);

	/*
	 * In order for "which" to be actively involved in
	 * the deadlock, removing him from the evaluation
	 * must remove the deadlock.  So, we OR together everyone
	 * except which; if all the participants still have their
	 * bits set, then the deadlock persists and which does
	 * not participate.  If the deadlock does not persist
	 * then "which" does participate.
	 */
	count = 0;
	for (j = 0; j < nlockers; j++) {
		if (!ISSET_MAP(deadmap, j) || j == which)
			continue;

		/* Find the map for this bit. */
		tmap = origmap + (nalloc * j);

		/*
		 * We special case the first waiter who is also a holder, so
		 * we don't automatically call that a deadlock.  However, if
		 * it really is a deadlock, we need the bit set now so that
		 * we treat the first waiter like other waiters.
		 */
		if (idmap[j].self_wait)
			SET_MAP(tmap, j);
		OR_MAP(tmpmap, tmap, nalloc);
		count++;
	}

	if (count == 1)
		return (1);

	/*
	 * Now check the resulting map and see whether
	 * all participants still have their bit set.
	 */
	for (j = 0; j < nlockers; j++) {
		if (!ISSET_MAP(deadmap, j) || j == which)
			continue;
		if (!ISSET_MAP(tmpmap, j))
			return (1);
	}
	return (0);
}

/*
 * __dd_isolder --
 *
 * Figure out the relative age of two lockers.  We make all lockers
 * older than all transactions, because that's how it's worked
 * historically (because lockers are lower ids).
 */
static int
__dd_isolder(a, b, lock_max, txn_max)
	u_int32_t	a, b;
	u_int32_t	lock_max, txn_max;
{
	u_int32_t max;

	/* Check for comparing lock-id and txnid. */
	if (a <= DB_LOCK_MAXID && b > DB_LOCK_MAXID)
		return (1);
	if (b <= DB_LOCK_MAXID && a > DB_LOCK_MAXID)
		return (0);

	/* In the same space; figure out which one. */
	max = txn_max;
	if (a <= DB_LOCK_MAXID)
		max = lock_max;

	/*
	 * We can't get a 100% correct ordering, because we don't know
	 * where the current interval started and if there were older
	 * lockers outside the interval.  We do the best we can.
	 */

	/*
	 * Check for a wrapped case with ids above max.
	 */
	if (a > max && b < max)
		return (1);
	if (b > max && a < max)
		return (0);

	return (a < b);
}