X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Flock.c;h=2a6f1230c9d7308b83592771ec03de5d7f5a95f1;hp=117596a3f203814daa9e4611d96b1216de050643;hb=729fc5b931c8eb26f192399d8aa7064f9af058df;hpb=83a1386b50966eb9667478c7cef9154dbd622541

diff --git a/ccan/tdb/lock.c b/ccan/tdb/lock.c
index 117596a3..2a6f1230 100644
--- a/ccan/tdb/lock.c
+++ b/ccan/tdb/lock.c
@@ -27,13 +27,104 @@
 
 #include "tdb_private.h"
 
-#define TDB_MARK_LOCK 0x80000000
-
 void tdb_setalarm_sigptr(struct tdb_context *tdb, volatile sig_atomic_t *ptr)
 {
 	tdb->interrupt_sig_ptr = ptr;
 }
 
+static int fcntl_lock(struct tdb_context *tdb,
+		      int rw, off_t off, off_t len, bool waitflag)
+{
+	struct flock fl;
+
+	fl.l_type = rw;
+	fl.l_whence = SEEK_SET;
+	fl.l_start = off;
+	fl.l_len = len;
+	fl.l_pid = 0;
+
+	if (waitflag)
+		return fcntl(tdb->fd, F_SETLKW, &fl);
+	else
+		return fcntl(tdb->fd, F_SETLK, &fl);
+}
+
+static int fcntl_unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
+{
+	struct flock fl;
+#if 0 /* Check they matched up locks and unlocks correctly. */
+	char line[80];
+	FILE *locks;
+	bool found = false;
+
+	locks = fopen("/proc/locks", "r");
+
+	while (fgets(line, 80, locks)) {
+		char *p;
+		int type, start, l;
+
+		/* eg. 1: FLOCK  ADVISORY  WRITE 2440 08:01:2180826 0 EOF */
+		p = strchr(line, ':') + 1;
+		if (strncmp(p, " POSIX  ADVISORY  ", strlen(" POSIX  ADVISORY  ")))
+			continue;
+		p += strlen(" FLOCK  ADVISORY  ");
+		if (strncmp(p, "READ  ", strlen("READ  ")) == 0)
+			type = F_RDLCK;
+		else if (strncmp(p, "WRITE ", strlen("WRITE ")) == 0)
+			type = F_WRLCK;
+		else
+			abort();
+		p += 6;
+		if (atoi(p) != getpid())
+			continue;
+		p = strchr(strchr(p, ' ') + 1, ' ') + 1;
+		start = atoi(p);
+		p = strchr(p, ' ') + 1;
+		if (strncmp(p, "EOF", 3) == 0)
+			l = 0;
+		else
+			l = atoi(p) - start + 1;
+
+		if (off == start) {
+			if (len != l) {
+				fprintf(stderr, "Len %u should be %u: %s",
+					(int)len, l, line);
+				abort();
+			}
+			if (type != rw) {
+				fprintf(stderr, "Type %s wrong: %s",
+					rw == F_RDLCK ? "READ" : "WRITE", line);
+				abort();
+			}
+			found = true;
+			break;
+		}
+	}
+
+	if (!found) {
+		fprintf(stderr, "Unlock on %u@%u not found!\n",
+			(int)off, (int)len);
+		abort();
+	}
+
+	fclose(locks);
+#endif
+
+	fl.l_type = F_UNLCK;
+	fl.l_whence = SEEK_SET;
+	fl.l_start = off;
+	fl.l_len = len;
+	fl.l_pid = 0;
+
+	return fcntl(tdb->fd, F_SETLKW, &fl);
+}
+
+/* list -1 is the alloc list, otherwise a hash chain. */
+static tdb_off_t lock_offset(int list)
+{
+	return FREELIST_TOP + 4*list;
+}
+
 /* a byte range locking function - return 0 on success
    this functions locks/unlocks 1 byte at the specified offset.
 
@@ -42,30 +133,36 @@ void tdb_setalarm_sigptr(struct tdb_context *tdb, volatile sig_atomic_t *ptr)
 
    note that a len of zero means lock to end of file
 */
-int tdb_brlock(struct tdb_context *tdb, tdb_off_t offset, 
-	       int rw_type, int lck_type, int probe, size_t len)
+int tdb_brlock(struct tdb_context *tdb,
+	       int rw_type, tdb_off_t offset, size_t len,
+	       enum tdb_lock_flags flags)
 {
-	struct flock fl;
 	int ret;
 
 	if (tdb->flags & TDB_NOLOCK) {
 		return 0;
 	}
 
+	if (flags & TDB_LOCK_MARK_ONLY) {
+		return 0;
+	}
+
 	if ((rw_type == F_WRLCK) && (tdb->read_only || tdb->traverse_read)) {
 		tdb->ecode = TDB_ERR_RDONLY;
 		return -1;
 	}
 
-	fl.l_type = rw_type;
-	fl.l_whence = SEEK_SET;
-	fl.l_start = offset;
-	fl.l_len = len;
-	fl.l_pid = 0;
+	/* Sanity check */
+	if (tdb->transaction && offset >= lock_offset(-1) && len != 0) {
+		tdb->ecode = TDB_ERR_RDONLY;
+		TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_brlock attempted in transaction at offset %d rw_type=%d flags=%d len=%d\n",
+			 offset, rw_type, flags, (int)len));
+		return -1;
+	}
 
 	do {
-		ret = fcntl(tdb->fd,lck_type,&fl);
-
+		ret = fcntl_lock(tdb, rw_type, offset, len,
+				 flags & TDB_LOCK_WAIT);
 		/* Check for a sigalarm break. */
 		if (ret == -1 && errno == EINTR &&
 				tdb->interrupt_sig_ptr &&
@@ -75,20 +172,38 @@ int tdb_brlock(struct tdb_context *tdb, tdb_off_t offset,
 	} while (ret == -1 && errno == EINTR);
 
 	if (ret == -1) {
+		tdb->ecode = TDB_ERR_LOCK;
 		/* Generic lock error. errno set by fcntl.
 		 * EAGAIN is an expected return from non-blocking
 		 * locks. */
-		if (!probe && lck_type != F_SETLK) {
-			/* Ensure error code is set for log fun to examine. */
-			tdb->ecode = TDB_ERR_LOCK;
-			TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_brlock failed (fd=%d) at offset %d rw_type=%d lck_type=%d len=%d\n", 
-				 tdb->fd, offset, rw_type, lck_type, (int)len));
+		if (!(flags & TDB_LOCK_PROBE) && errno != EAGAIN) {
+			TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_brlock failed (fd=%d) at offset %d rw_type=%d flags=%d len=%d\n",
+				 tdb->fd, offset, rw_type, flags, (int)len));
 		}
-		return TDB_ERRCODE(TDB_ERR_LOCK, -1);
+		return -1;
 	}
 	return 0;
 }
 
+int tdb_brunlock(struct tdb_context *tdb,
+		 int rw_type, tdb_off_t offset, size_t len)
+{
+	int ret;
+
+	if (tdb->flags & TDB_NOLOCK) {
+		return 0;
+	}
+
+	do {
+		ret = fcntl_unlock(tdb, rw_type, offset, len);
+	} while (ret == -1 && errno == EINTR);
+
+	if (ret == -1) {
+		TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_brunlock failed (fd=%d) at offset %d rw_type=%d len=%d\n",
+			 tdb->fd, offset, rw_type, (int)len));
+	}
+	return ret;
+}
 
 /*
   upgrade a read lock to a write lock. This needs to be handled in a
@@ -96,12 +211,29 @@ int tdb_brlock(struct tdb_context *tdb, tdb_off_t offset,
   deadlock detection and claim a deadlock when progress can be
   made. For those OSes we may loop for a while.  
 */
-int tdb_brlock_upgrade(struct tdb_context *tdb, tdb_off_t offset, size_t len)
+int tdb_allrecord_upgrade(struct tdb_context *tdb)
 {
 	int count = 1000;
+
+	if (tdb->allrecord_lock.count != 1) {
+		TDB_LOG((tdb, TDB_DEBUG_ERROR,
+			 "tdb_allrecord_upgrade failed: count %u too high\n",
+			 tdb->allrecord_lock.count));
+		return -1;
+	}
+
+	if (tdb->allrecord_lock.off != 1) {
+		TDB_LOG((tdb, TDB_DEBUG_ERROR,
+			 "tdb_allrecord_upgrade failed: already upgraded?\n"));
+		return -1;
+	}
+
 	while (count--) {
 		struct timeval tv;
-		if (tdb_brlock(tdb, offset, F_WRLCK, F_SETLKW, 1, len) == 0) {
+		if (tdb_brlock(tdb, F_WRLCK, FREELIST_TOP, 0,
+			       TDB_LOCK_WAIT|TDB_LOCK_PROBE) == 0) {
+			tdb->allrecord_lock.ltype = F_WRLCK;
+			tdb->allrecord_lock.off = 0;
 			return 0;
 		}
 		if (errno != EDEADLK) {
@@ -112,55 +244,46 @@ int tdb_brlock_upgrade(struct tdb_context *tdb, tdb_off_t offset, size_t len)
 		tv.tv_usec = 1;
 		select(0, NULL, NULL, NULL, &tv);
 	}
-	TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_brlock_upgrade failed at offset %d\n", offset));
+	TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_allrecord_upgrade failed\n"));
 	return -1;
 }
 
-
-/* lock a list in the database. list -1 is the alloc list */
-static int _tdb_lock(struct tdb_context *tdb, int list, int ltype, int op)
+static struct tdb_lock_type *find_nestlock(struct tdb_context *tdb,
+					   tdb_off_t offset)
 {
-	struct tdb_lock_type *new_lck;
-	int i;
-	bool mark_lock = ((ltype & TDB_MARK_LOCK) == TDB_MARK_LOCK);
-
-	ltype &= ~TDB_MARK_LOCK;
+	unsigned int i;
 
-	/* a global lock allows us to avoid per chain locks */
-	if (tdb->global_lock.count && 
-	    (ltype == tdb->global_lock.ltype || ltype == F_RDLCK)) {
-		return 0;
+	for (i=0; i<tdb->num_lockrecs; i++) {
+		if (tdb->lockrecs[i].off == offset) {
+			return &tdb->lockrecs[i];
+		}
 	}
+	return NULL;
+}
 
-	if (tdb->global_lock.count) {
-		return TDB_ERRCODE(TDB_ERR_LOCK, -1);
-	}
+/* lock an offset in the database. */
+int tdb_nest_lock(struct tdb_context *tdb, uint32_t offset, int ltype,
+		  enum tdb_lock_flags flags)
+{
+	struct tdb_lock_type *new_lck;
 
-	if (list < -1 || list >= (int)tdb->header.hash_size) {
-		TDB_LOG((tdb, TDB_DEBUG_ERROR,"tdb_lock: invalid list %d for ltype=%d\n", 
-			   list, ltype));
+	if (offset >= lock_offset(tdb->header.hash_size)) {
+		tdb->ecode = TDB_ERR_LOCK;
+		TDB_LOG((tdb, TDB_DEBUG_ERROR,"tdb_lock: invalid offset %u for ltype=%d\n",
+			 offset, ltype));
 		return -1;
 	}
 	if (tdb->flags & TDB_NOLOCK)
 		return 0;
 
-	for (i=0; i<tdb->num_lockrecs; i++) {
-		if (tdb->lockrecs[i].list == list) {
-			if (tdb->lockrecs[i].count == 0) {
-				/*
-				 * Can't happen, see tdb_unlock(). It should
-				 * be an assert.
-				 */
-				TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lock: "
-					 "lck->count == 0 for list %d", list));
-			}
-			/*
-			 * Just increment the in-memory struct, posix locks
-			 * don't stack.
-			 */
-			tdb->lockrecs[i].count++;
-			return 0;
-		}
+	new_lck = find_nestlock(tdb, offset);
+	if (new_lck) {
+		/*
+		 * Just increment the in-memory struct, posix locks
+		 * don't stack.
+		 */
+		new_lck->count++;
+		return 0;
 	}
 
 	new_lck = (struct tdb_lock_type *)realloc(
@@ -174,27 +297,89 @@ static int _tdb_lock(struct tdb_context *tdb, int list, int ltype, int op)
 
 	/* Since fcntl locks don't nest, we do a lock for the first one,
 	   and simply bump the count for future ones */
-	if (!mark_lock &&
-	    tdb->methods->tdb_brlock(tdb,FREELIST_TOP+4*list, ltype, op,
-				     0, 1)) {
+	if (tdb_brlock(tdb, ltype, offset, 1, flags)) {
 		return -1;
 	}
 
-	tdb->num_locks++;
-
-	tdb->lockrecs[tdb->num_lockrecs].list = list;
+	tdb->lockrecs[tdb->num_lockrecs].off = offset;
 	tdb->lockrecs[tdb->num_lockrecs].count = 1;
 	tdb->lockrecs[tdb->num_lockrecs].ltype = ltype;
-	tdb->num_lockrecs += 1;
+	tdb->num_lockrecs++;
 
 	return 0;
 }
 
+static int tdb_lock_and_recover(struct tdb_context *tdb)
+{
+	int ret;
+
+	/* We need to match locking order in transaction commit. */
+	if (tdb_brlock(tdb, F_WRLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT)) {
+		return -1;
+	}
+
+	if (tdb_brlock(tdb, F_WRLCK, OPEN_LOCK, 1, TDB_LOCK_WAIT)) {
+		tdb_brunlock(tdb, F_WRLCK, FREELIST_TOP, 0);
+		return -1;
+	}
+
+	ret = tdb_transaction_recover(tdb);
+
+	tdb_brunlock(tdb, F_WRLCK, OPEN_LOCK, 1);
+	tdb_brunlock(tdb, F_WRLCK, FREELIST_TOP, 0);
+
+	return ret;
+}
+
+static bool have_data_locks(const struct tdb_context *tdb)
+{
+	unsigned int i;
+
+	for (i = 0; i < tdb->num_lockrecs; i++) {
+		if (tdb->lockrecs[i].off >= lock_offset(-1))
+			return true;
+	}
+	return false;
+}
+
+static int tdb_lock_list(struct tdb_context *tdb, int list, int ltype,
+			 enum tdb_lock_flags waitflag)
+{
+	int ret;
+	bool check = false;
+
+	/* a allrecord lock allows us to avoid per chain locks */
+	if (tdb->allrecord_lock.count &&
+	    (ltype == tdb->allrecord_lock.ltype || ltype == F_RDLCK)) {
+		return 0;
+	}
+
+	if (tdb->allrecord_lock.count) {
+		tdb->ecode = TDB_ERR_LOCK;
+		ret = -1;
+	} else {
+		/* Only check when we grab first data lock. */
+		check = !have_data_locks(tdb);
+		ret = tdb_nest_lock(tdb, lock_offset(list), ltype, waitflag);
+
+		if (ret == 0 && check && tdb_needs_recovery(tdb)) {
+			tdb_nest_unlock(tdb, lock_offset(list), ltype, false);
+
+			if (tdb_lock_and_recover(tdb) == -1) {
+				return -1;
+			}
+			return tdb_lock_list(tdb, list, ltype, waitflag);
+		}
+	}
+	return ret;
+}
+
 /* lock a list in the database. list -1 is the alloc list */
 int tdb_lock(struct tdb_context *tdb, int list, int ltype)
 {
 	int ret;
-	ret = _tdb_lock(tdb, list, ltype, F_SETLKW);
+
+	ret = tdb_lock_list(tdb, list, ltype, TDB_LOCK_WAIT);
 	if (ret) {
 		TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lock failed on list %d "
 			 "ltype=%d (%s)\n",  list, ltype, strerror(errno)));
@@ -205,48 +390,26 @@ int tdb_lock(struct tdb_context *tdb, int list, int ltype)
 /* lock a list in the database. list -1 is the alloc list. non-blocking lock */
 int tdb_lock_nonblock(struct tdb_context *tdb, int list, int ltype)
 {
-	return _tdb_lock(tdb, list, ltype, F_SETLK);
+	return tdb_lock_list(tdb, list, ltype, TDB_LOCK_NOWAIT);
 }
 
 
-/* unlock the database: returns void because it's too late for errors. */
-	/* changed to return int it may be interesting to know there
-	   has been an error  --simo */
-int tdb_unlock(struct tdb_context *tdb, int list, int ltype)
+int tdb_nest_unlock(struct tdb_context *tdb, uint32_t offset, int ltype,
+		    bool mark_lock)
 {
 	int ret = -1;
-	int i;
-	struct tdb_lock_type *lck = NULL;
-	bool mark_lock = ((ltype & TDB_MARK_LOCK) == TDB_MARK_LOCK);
-
-	ltype &= ~TDB_MARK_LOCK;
-
-	/* a global lock allows us to avoid per chain locks */
-	if (tdb->global_lock.count && 
-	    (ltype == tdb->global_lock.ltype || ltype == F_RDLCK)) {
-		return 0;
-	}
-
-	if (tdb->global_lock.count) {
-		return TDB_ERRCODE(TDB_ERR_LOCK, -1);
-	}
+	struct tdb_lock_type *lck;
 
 	if (tdb->flags & TDB_NOLOCK)
 		return 0;
 
 	/* Sanity checks */
-	if (list < -1 || list >= (int)tdb->header.hash_size) {
-		TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlock: list %d invalid (%d)\n", list, tdb->header.hash_size));
+	if (offset >= lock_offset(tdb->header.hash_size)) {
+		TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlock: offset %u invalid (%d)\n", offset, tdb->header.hash_size));
 		return ret;
 	}
 
-	for (i=0; i<tdb->num_lockrecs; i++) {
-		if (tdb->lockrecs[i].list == list) {
-			lck = &tdb->lockrecs[i];
-			break;
-		}
-	}
-
+	lck = find_nestlock(tdb, offset);
 	if ((lck == NULL) || (lck->count == 0)) {
 		TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlock: count is 0\n"));
 		return -1;
@@ -267,20 +430,14 @@ int tdb_unlock(struct tdb_context *tdb, int list, int ltype)
 	if (mark_lock) {
 		ret = 0;
 	} else {
-		ret = tdb->methods->tdb_brlock(tdb, FREELIST_TOP+4*list, F_UNLCK,
-					       F_SETLKW, 0, 1);
+		ret = tdb_brunlock(tdb, ltype, offset, 1);
 	}
-	tdb->num_locks--;
 
 	/*
 	 * Shrink the array by overwriting the element just unlocked with the
 	 * last array element.
 	 */
-
-	if (tdb->num_lockrecs > 1) {
-		*lck = tdb->lockrecs[tdb->num_lockrecs-1];
-	}
-	tdb->num_lockrecs -= 1;
+	*lck = tdb->lockrecs[--tdb->num_lockrecs];
 
 	/*
 	 * We don't bother with realloc when the array shrinks, but if we have
@@ -296,76 +453,100 @@ int tdb_unlock(struct tdb_context *tdb, int list, int ltype)
 	return ret;
 }
 
-/*
-  get the transaction lock
- */
-int tdb_transaction_lock(struct tdb_context *tdb, int ltype)
+int tdb_unlock(struct tdb_context *tdb, int list, int ltype)
 {
-	if (tdb->have_transaction_lock || tdb->global_lock.count) {
-		tdb->have_transaction_lock++;
+	/* a global lock allows us to avoid per chain locks */
+	if (tdb->allrecord_lock.count &&
+	    (ltype == tdb->allrecord_lock.ltype || ltype == F_RDLCK)) {
 		return 0;
 	}
-	if (tdb->methods->tdb_brlock(tdb, TRANSACTION_LOCK, ltype, 
-				     F_SETLKW, 0, 1) == -1) {
-		TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_lock: failed to get transaction lock\n"));
+
+	if (tdb->allrecord_lock.count) {
 		tdb->ecode = TDB_ERR_LOCK;
 		return -1;
 	}
-	tdb->have_transaction_lock++;
-	return 0;
+
+	return tdb_nest_unlock(tdb, lock_offset(list), ltype, false);
 }
 
 /*
-  release the transaction lock
+  get the transaction lock
  */
-int tdb_transaction_unlock(struct tdb_context *tdb)
+int tdb_transaction_lock(struct tdb_context *tdb, int ltype)
 {
-	if (--tdb->have_transaction_lock) {
-		return 0;
-	}
-	return tdb->methods->tdb_brlock(tdb, TRANSACTION_LOCK, F_UNLCK, F_SETLKW, 0, 1);
+	return tdb_nest_lock(tdb, TRANSACTION_LOCK, ltype, TDB_LOCK_WAIT);
 }
 
-
-
-
-/* lock/unlock entire database */
-static int _tdb_lockall(struct tdb_context *tdb, int ltype, int op)
+/*
+  release the transaction lock
+ */
+int tdb_transaction_unlock(struct tdb_context *tdb, int ltype)
 {
-	bool mark_lock = ((ltype & TDB_MARK_LOCK) == TDB_MARK_LOCK);
+	return tdb_nest_unlock(tdb, TRANSACTION_LOCK, ltype, false);
+}
 
-	ltype &= ~TDB_MARK_LOCK;
 
+/* lock/unlock entire database.  It can only be upgradable if you have some
+ * other way of guaranteeing exclusivity (ie. transaction write lock). */
+int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
+		       enum tdb_lock_flags flags, bool upgradable)
+{
 	/* There are no locks on read-only dbs */
-	if (tdb->read_only || tdb->traverse_read)
-		return TDB_ERRCODE(TDB_ERR_LOCK, -1);
+	if (tdb->read_only || tdb->traverse_read) {
+		tdb->ecode = TDB_ERR_LOCK;
+		return -1;
+	}
 
-	if (tdb->global_lock.count && tdb->global_lock.ltype == ltype) {
-		tdb->global_lock.count++;
+	if (tdb->allrecord_lock.count && tdb->allrecord_lock.ltype == ltype) {
+		tdb->allrecord_lock.count++;
 		return 0;
 	}
 
-	if (tdb->global_lock.count) {
+	if (tdb->allrecord_lock.count) {
 		/* a global lock of a different type exists */
-		return TDB_ERRCODE(TDB_ERR_LOCK, -1);
+		tdb->ecode = TDB_ERR_LOCK;
+		return -1;
 	}
-	
-	if (tdb->num_locks != 0) {
+
+	if (tdb_have_extra_locks(tdb)) {
 		/* can't combine global and chain locks */
-		return TDB_ERRCODE(TDB_ERR_LOCK, -1);
+		tdb->ecode = TDB_ERR_LOCK;
+		return -1;
 	}
 
-	if (!mark_lock &&
-	    tdb->methods->tdb_brlock(tdb, FREELIST_TOP, ltype, op,
-				     0, 4*tdb->header.hash_size)) {
-		if (op == F_SETLKW) {
+	if (upgradable && ltype != F_RDLCK) {
+		/* tdb error: you can't upgrade a write lock! */
+		tdb->ecode = TDB_ERR_LOCK;
+		return -1;
+	}
+
+	if (tdb_brlock(tdb, ltype, FREELIST_TOP, 0, flags)) {
+		if (flags & TDB_LOCK_WAIT) {
 			TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lockall failed (%s)\n", strerror(errno)));
 		}
 		return -1;
 	}
 
-	tdb->global_lock.count = 1;
-	tdb->global_lock.ltype = ltype;
+	tdb->allrecord_lock.count = 1;
+	/* If it's upgradable, it's actually exclusive so we can treat
+	 * it as a write lock. */
+	tdb->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
+	tdb->allrecord_lock.off = upgradable;
+
+	if (tdb_needs_recovery(tdb)) {
+		bool mark = flags & TDB_LOCK_MARK_ONLY;
+		tdb_allrecord_unlock(tdb, ltype, mark);
+		if (mark) {
+			tdb->ecode = TDB_ERR_LOCK;
+			TDB_LOG((tdb, TDB_DEBUG_ERROR,
+				 "tdb_lockall_mark cannot do recovery\n"));
+			return -1;
+		}
+		if (tdb_lock_and_recover(tdb) == -1) {
+			return -1;
+		}
+		return tdb_allrecord_lock(tdb, ltype, flags, upgradable);
+	}
 
 	return 0;
 }
@@ -373,35 +554,38 @@ static int _tdb_lockall(struct tdb_context *tdb, int ltype, int op)
 
 
 /* unlock entire db */
-static int _tdb_unlockall(struct tdb_context *tdb, int ltype)
+int tdb_allrecord_unlock(struct tdb_context *tdb, int ltype, bool mark_lock)
 {
-	bool mark_lock = ((ltype & TDB_MARK_LOCK) == TDB_MARK_LOCK);
-
-	ltype &= ~TDB_MARK_LOCK;
-
 	/* There are no locks on read-only dbs */
 	if (tdb->read_only || tdb->traverse_read) {
-		return TDB_ERRCODE(TDB_ERR_LOCK, -1);
+		tdb->ecode = TDB_ERR_LOCK;
+		return -1;
 	}
 
-	if (tdb->global_lock.ltype != ltype || tdb->global_lock.count == 0) {
-		return TDB_ERRCODE(TDB_ERR_LOCK, -1);
+	if (tdb->allrecord_lock.count == 0) {
+		tdb->ecode = TDB_ERR_LOCK;
+		return -1;
+	}
+
+	/* Upgradable locks are marked as write locks. */
+	if (tdb->allrecord_lock.ltype != ltype
+	    && (!tdb->allrecord_lock.off || ltype != F_RDLCK)) {
+		tdb->ecode = TDB_ERR_LOCK;
+		return -1;
 	}
 
-	if (tdb->global_lock.count > 1) {
-		tdb->global_lock.count--;
+	if (tdb->allrecord_lock.count > 1) {
+		tdb->allrecord_lock.count--;
 		return 0;
 	}
 
-	if (!mark_lock &&
-	    tdb->methods->tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 
-				     0, 4*tdb->header.hash_size)) {
+	if (!mark_lock && tdb_brunlock(tdb, ltype, FREELIST_TOP, 0)) {
 		TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlockall failed (%s)\n", strerror(errno)));
 		return -1;
 	}
 
-	tdb->global_lock.count = 0;
-	tdb->global_lock.ltype = 0;
+	tdb->allrecord_lock.count = 0;
+	tdb->allrecord_lock.ltype = 0;
 
 	return 0;
 }
@@ -410,27 +594,27 @@ static int _tdb_unlockall(struct tdb_context *tdb, int ltype)
 int tdb_lockall(struct tdb_context *tdb)
 {
 	tdb_trace(tdb, "tdb_lockall");
-	return _tdb_lockall(tdb, F_WRLCK, F_SETLKW);
+	return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false);
 }
 
 /* lock entire database with write lock - mark only */
 int tdb_lockall_mark(struct tdb_context *tdb)
 {
 	tdb_trace(tdb, "tdb_lockall_mark");
-	return _tdb_lockall(tdb, F_WRLCK | TDB_MARK_LOCK, F_SETLKW);
+	return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_MARK_ONLY, false);
 }
 
 /* unlock entire database with write lock - unmark only */
 int tdb_lockall_unmark(struct tdb_context *tdb)
 {
 	tdb_trace(tdb, "tdb_lockall_unmark");
-	return _tdb_unlockall(tdb, F_WRLCK | TDB_MARK_LOCK);
+	return tdb_allrecord_unlock(tdb, F_WRLCK, true);
 }
 
 /* lock entire database with write lock - nonblocking varient */
 int tdb_lockall_nonblock(struct tdb_context *tdb)
 {
-	int ret = _tdb_lockall(tdb, F_WRLCK, F_SETLK);
+	int ret = tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_NOWAIT, false);
 	tdb_trace_ret(tdb, "tdb_lockall_nonblock", ret);
 	return ret;
 }
@@ -439,20 +623,20 @@ int tdb_lockall_nonblock(struct tdb_context *tdb)
 int tdb_unlockall(struct tdb_context *tdb)
 {
 	tdb_trace(tdb, "tdb_unlockall");
-	return _tdb_unlockall(tdb, F_WRLCK);
+	return tdb_allrecord_unlock(tdb, F_WRLCK, false);
 }
 
 /* lock entire database with read lock */
 int tdb_lockall_read(struct tdb_context *tdb)
 {
 	tdb_trace(tdb, "tdb_lockall_read");
-	return _tdb_lockall(tdb, F_RDLCK, F_SETLKW);
+	return tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false);
 }
 
 /* lock entire database with read lock - nonblock varient */
 int tdb_lockall_read_nonblock(struct tdb_context *tdb)
 {
-	int ret = _tdb_lockall(tdb, F_RDLCK, F_SETLK);
+	int ret = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_NOWAIT, false);
 	tdb_trace_ret(tdb, "tdb_lockall_read_nonblock", ret);
 	return ret;
 }
@@ -461,7 +645,7 @@ int tdb_lockall_read_nonblock(struct tdb_context *tdb)
 int tdb_unlockall_read(struct tdb_context *tdb)
 {
 	tdb_trace(tdb, "tdb_unlockall_read");
-	return _tdb_unlockall(tdb, F_RDLCK);
+	return tdb_allrecord_unlock(tdb, F_RDLCK, false);
 }
 
 /* lock/unlock one hash chain. This is meant to be used to reduce
@@ -486,7 +670,8 @@ int tdb_chainlock_nonblock(struct tdb_context *tdb, TDB_DATA key)
 /* mark a chain as locked without actually locking it. Warning! use with great caution! */
 int tdb_chainlock_mark(struct tdb_context *tdb, TDB_DATA key)
 {
-	int ret = tdb_lock(tdb, BUCKET(tdb->hash_fn(&key)), F_WRLCK | TDB_MARK_LOCK);
+	int ret = tdb_nest_lock(tdb, lock_offset(BUCKET(tdb->hash_fn(&key))),
+				F_WRLCK, TDB_LOCK_MARK_ONLY);
 	tdb_trace_1rec(tdb, "tdb_chainlock_mark", key);
 	return ret;
 }
@@ -495,7 +680,8 @@ int tdb_chainlock_mark(struct tdb_context *tdb, TDB_DATA key)
 int tdb_chainlock_unmark(struct tdb_context *tdb, TDB_DATA key)
 {
 	tdb_trace_1rec(tdb, "tdb_chainlock_unmark", key);
-	return tdb_unlock(tdb, BUCKET(tdb->hash_fn(&key)), F_WRLCK | TDB_MARK_LOCK);
+	return tdb_nest_unlock(tdb, lock_offset(BUCKET(tdb->hash_fn(&key))),
+			       F_WRLCK, true);
 }
 
 int tdb_chainunlock(struct tdb_context *tdb, TDB_DATA key)
@@ -518,13 +704,15 @@ int tdb_chainunlock_read(struct tdb_context *tdb, TDB_DATA key)
 	return tdb_unlock(tdb, BUCKET(tdb->hash_fn(&key)), F_RDLCK);
 }
 
+
+
 /* record lock stops delete underneath */
 int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off)
 {
-	if (tdb->global_lock.count) {
+	if (tdb->allrecord_lock.count) {
 		return 0;
 	}
-	return off ? tdb->methods->tdb_brlock(tdb, off, F_RDLCK, F_SETLKW, 0, 1) : 0;
+	return off ? tdb_brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0;
 }
 
 /*
@@ -538,16 +726,21 @@ int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off)
 	for (i = &tdb->travlocks; i; i = i->next)
 		if (i->off == off)
 			return -1;
-	return tdb->methods->tdb_brlock(tdb, off, F_WRLCK, F_SETLK, 1, 1);
+	if (tdb->allrecord_lock.count) {
+		if (tdb->allrecord_lock.ltype == F_WRLCK) {
+			return 0;
+		}
+		return -1;
+	}
+	return tdb_brlock(tdb, F_WRLCK, off, 1, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
 }
 
-/*
-  Note this is meant to be F_SETLK, *not* F_SETLKW, as it's not
-  an error to fail to get the lock here.
-*/
 int tdb_write_unlock_record(struct tdb_context *tdb, tdb_off_t off)
 {
-	return tdb->methods->tdb_brlock(tdb, off, F_UNLCK, F_SETLK, 0, 1);
+	if (tdb->allrecord_lock.count) {
+		return 0;
+	}
+	return tdb_brunlock(tdb, F_WRLCK, off, 1);
 }
 
 /* fcntl locks don't stack: avoid unlocking someone else's */
@@ -556,7 +749,7 @@ int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off)
 	struct tdb_traverse_lock *i;
 	uint32_t count = 0;
 
-	if (tdb->global_lock.count) {
+	if (tdb->allrecord_lock.count) {
 		return 0;
 	}
 
@@ -565,5 +758,56 @@ int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off)
 	for (i = &tdb->travlocks; i; i = i->next)
 		if (i->off == off)
 			count++;
-	return (count == 1 ? tdb->methods->tdb_brlock(tdb, off, F_UNLCK, F_SETLKW, 0, 1) : 0);
+	return (count == 1 ? tdb_brunlock(tdb, F_RDLCK, off, 1) : 0);
+}
+
+bool tdb_have_extra_locks(struct tdb_context *tdb)
+{
+	unsigned int extra = tdb->num_lockrecs;
+
+	/* A transaction holds the lock for all records. */
+	if (!tdb->transaction && tdb->allrecord_lock.count) {
+		return true;
+	}
+
+	/* We always hold the active lock if CLEAR_IF_FIRST. */
+	if (find_nestlock(tdb, ACTIVE_LOCK)) {
+		extra--;
+	}
+
+	/* In a transaction, we expect to hold the transaction lock */
+	if (tdb->transaction && find_nestlock(tdb, TRANSACTION_LOCK)) {
+		extra--;
+	}
+
+	return extra;
+}
+
+/* The transaction code uses this to remove all locks.  Note that this
+   may include OPEN_LOCK. */
+void tdb_release_extra_locks(struct tdb_context *tdb)
+{
+	unsigned int i, extra = 0;
+
+	if (tdb->allrecord_lock.count != 0) {
+		tdb_brunlock(tdb, tdb->allrecord_lock.ltype, FREELIST_TOP, 0);
+		tdb->allrecord_lock.count = 0;
+	}
+
+	for (i=0;i<tdb->num_lockrecs;i++) {
+		struct tdb_lock_type *lck = &tdb->lockrecs[i];
+
+		/* Don't release transaction or active locks! */
+		if (tdb->transaction && lck->off == TRANSACTION_LOCK) {
+			tdb->lockrecs[extra++] = *lck;
+		} else if (lck->off == ACTIVE_LOCK) {
+			tdb->lockrecs[extra++] = *lck;
+		} else {
+			tdb_brunlock(tdb, lck->ltype, lck->off, 1);
+		}
+	}
+	tdb->num_lockrecs = extra;
+	if (tdb->num_lockrecs == 0) {
+		SAFE_FREE(tdb->lockrecs);
+	}
 }