X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Flock.c;h=dfcac7a540adc30482dcd86e0c299b36bfb52e0b;hp=920ba95fe0515a7773ddefd61098510974c18f23;hb=448b9194f76a138eef05047b1aaa5d211f72e97f;hpb=bc1c29d566924ebb7a303dfeb732a5303a59a9b3

diff --git a/ccan/tdb/lock.c b/ccan/tdb/lock.c
index 920ba95f..dfcac7a5 100644
--- a/ccan/tdb/lock.c
+++ b/ccan/tdb/lock.c
@@ -119,6 +119,12 @@ static int fcntl_unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
 	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.
 
@@ -197,13 +203,29 @@ int tdb_brunlock(struct tdb_context *tdb,
   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, F_WRLCK, offset, len,
+		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) {
@@ -214,16 +236,10 @@ 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;
 }
 
-/* list -1 is the alloc list, otherwise a hash chain. */
-static tdb_off_t lock_offset(int list)
-{
-	return FREELIST_TOP + 4*list;
-}
-
 static struct tdb_lock_type *find_nestlock(struct tdb_context *tdb,
 					   tdb_off_t offset)
 {
@@ -273,25 +289,57 @@ int tdb_nest_lock(struct tdb_context *tdb, uint32_t offset, int ltype,
 
 	/* Since fcntl locks don't nest, we do a lock for the first one,
 	   and simply bump the count for future ones */
-	if (tdb->methods->brlock(tdb, ltype, offset, 1, flags)) {
+	if (tdb_brlock(tdb, ltype, offset, 1, flags)) {
 		return -1;
 	}
 
-	tdb->num_locks++;
-
 	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;
 }
 
-/* lock a list in the database. list -1 is the alloc list */
-int tdb_lock(struct tdb_context *tdb, int list, int ltype)
+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)) {
@@ -302,9 +350,28 @@ int tdb_lock(struct tdb_context *tdb, int list, int ltype)
 		tdb->ecode = TDB_ERR_LOCK;
 		ret = -1;
 	} else {
-		ret = tdb_nest_lock(tdb, lock_offset(list), ltype,
-				    TDB_LOCK_WAIT);
+		/* 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_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)));
@@ -315,18 +382,7 @@ 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)
 {
-	/* 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;
-		return -1;
-	}
-
-	return tdb_nest_lock(tdb, lock_offset(list), ltype, TDB_LOCK_NOWAIT);
+	return tdb_lock_list(tdb, list, ltype, TDB_LOCK_NOWAIT);
 }
 
 
@@ -366,19 +422,14 @@ int tdb_nest_unlock(struct tdb_context *tdb, uint32_t offset, int ltype,
 	if (mark_lock) {
 		ret = 0;
 	} else {
-		ret = tdb->methods->brunlock(tdb, ltype, offset, 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
@@ -415,18 +466,7 @@ int tdb_unlock(struct tdb_context *tdb, int list, int ltype)
  */
 int tdb_transaction_lock(struct tdb_context *tdb, int ltype)
 {
-	if (tdb->transaction_lock_count > 0) {
-		tdb->transaction_lock_count++;
-		return 0;
-	}
-
-	if (tdb->methods->brlock(tdb, ltype, TRANSACTION_LOCK, 1, TDB_LOCK_WAIT) == -1) {
-		TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_lock: failed to get transaction lock\n"));
-		tdb->ecode = TDB_ERR_LOCK;
-		return -1;
-	}
-	tdb->transaction_lock_count++;
-	return 0;
+	return tdb_nest_lock(tdb, TRANSACTION_LOCK, ltype, TDB_LOCK_WAIT);
 }
 
 /*
@@ -434,24 +474,12 @@ int tdb_transaction_lock(struct tdb_context *tdb, int ltype)
  */
 int tdb_transaction_unlock(struct tdb_context *tdb, int ltype)
 {
-	int ret;
-	if (tdb->transaction_lock_count > 1) {
-		tdb->transaction_lock_count--;
-		return 0;
-	}
-	ret = tdb->methods->brunlock(tdb, ltype, TRANSACTION_LOCK, 1);
-	if (ret == 0) {
-		tdb->transaction_lock_count = 0;
-	}
-	return ret;
+	return tdb_nest_unlock(tdb, TRANSACTION_LOCK, ltype, false);
 }
 
-
-
-
-/* lock/unlock entire database */
-static int _tdb_lockall(struct tdb_context *tdb, int ltype,
-			enum tdb_lock_flags flags)
+/* Returns 0 if all done, -1 if error, 1 if ok. */
+static int tdb_allrecord_check(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) {
@@ -476,17 +504,101 @@ static int _tdb_lockall(struct tdb_context *tdb, int ltype,
 		return -1;
 	}
 
-	if (tdb->methods->brlock(tdb, ltype,
-				 FREELIST_TOP, 4*tdb->header.hash_size,
-				 flags)) {
-		if (flags & TDB_LOCK_WAIT) {
-			TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lockall failed (%s)\n", strerror(errno)));
-		}
+	if (upgradable && ltype != F_RDLCK) {
+		/* tdb error: you can't upgrade a write lock! */
+		tdb->ecode = TDB_ERR_LOCK;
+		return -1;
+	}
+	return 1;
+}
+
+/* We only need to lock individual bytes, but Linux merges consecutive locks
+ * so we lock in contiguous ranges. */
+static int tdb_chainlock_gradual(struct tdb_context *tdb,
+				 int ltype, enum tdb_lock_flags flags,
+				 size_t off, size_t len)
+{
+	int ret;
+	enum tdb_lock_flags nb_flags = (flags & ~TDB_LOCK_WAIT);
+
+	if (len <= 4) {
+		/* Single record.  Just do blocking lock. */
+		return tdb_brlock(tdb, ltype, off, len, flags);
+	}
+
+	/* First we try non-blocking. */
+	ret = tdb_brlock(tdb, ltype, off, len, nb_flags);
+	if (ret == 0) {
+		return 0;
+	}
+
+	/* Try locking first half, then second. */
+	ret = tdb_chainlock_gradual(tdb, ltype, flags, off, len / 2);
+	if (ret == -1)
+		return -1;
+
+	ret = tdb_chainlock_gradual(tdb, ltype, flags,
+				    off + len / 2, len - len / 2);
+	if (ret == -1) {
+		tdb_brunlock(tdb, ltype, off, len / 2);
+		return -1;
+	}
+	return 0;
+}
+
+/* lock/unlock entire database.  It can only be upgradable if you have some
+ * other way of guaranteeing exclusivity (ie. transaction write lock).
+ * We do the locking gradually to avoid being starved by smaller locks. */
+int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
+		       enum tdb_lock_flags flags, bool upgradable)
+{
+	switch (tdb_allrecord_check(tdb, ltype, flags, upgradable)) {
+	case -1:
+		return -1;
+	case 0:
+		return 0;
+	}
+
+	/* We cover two kinds of locks:
+	 * 1) Normal chain locks.  Taken for almost all operations.
+	 * 3) Individual records locks.  Taken after normal or free
+	 *    chain locks.
+	 *
+	 * It is (1) which cause the starvation problem, so we're only
+	 * gradual for that. */
+	if (tdb_chainlock_gradual(tdb, ltype, flags, FREELIST_TOP,
+				  tdb->header.hash_size * 4) == -1) {
+		return -1;
+	}
+
+	/* Grab individual record locks. */
+	if (tdb_brlock(tdb, ltype, lock_offset(tdb->header.hash_size), 0,
+		       flags) == -1) {
+		tdb_brunlock(tdb, ltype, FREELIST_TOP,
+			     tdb->header.hash_size * 4);
 		return -1;
 	}
 
 	tdb->allrecord_lock.count = 1;
-	tdb->allrecord_lock.ltype = ltype;
+	/* 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;
 }
@@ -494,7 +606,7 @@ static int _tdb_lockall(struct tdb_context *tdb, int ltype,
 
 
 /* unlock entire db */
-static int _tdb_unlockall(struct tdb_context *tdb, int ltype, bool mark_lock)
+int tdb_allrecord_unlock(struct tdb_context *tdb, int ltype, bool mark_lock)
 {
 	/* There are no locks on read-only dbs */
 	if (tdb->read_only || tdb->traverse_read) {
@@ -502,7 +614,14 @@ static int _tdb_unlockall(struct tdb_context *tdb, int ltype, bool mark_lock)
 		return -1;
 	}
 
-	if (tdb->allrecord_lock.ltype != ltype || tdb->allrecord_lock.count == 0) {
+	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;
 	}
@@ -512,9 +631,7 @@ static int _tdb_unlockall(struct tdb_context *tdb, int ltype, bool mark_lock)
 		return 0;
 	}
 
-	if (!mark_lock &&
-	    tdb->methods->brunlock(tdb, ltype,
-				   FREELIST_TOP, 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;
 	}
@@ -529,27 +646,27 @@ static int _tdb_unlockall(struct tdb_context *tdb, int ltype, bool mark_lock)
 int tdb_lockall(struct tdb_context *tdb)
 {
 	tdb_trace(tdb, "tdb_lockall");
-	return _tdb_lockall(tdb, F_WRLCK, TDB_LOCK_WAIT);
+	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_LOCK_MARK_ONLY);
+	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, true);
+	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, TDB_LOCK_NOWAIT);
+	int ret = tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_NOWAIT, false);
 	tdb_trace_ret(tdb, "tdb_lockall_nonblock", ret);
 	return ret;
 }
@@ -558,20 +675,20 @@ 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, false);
+	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, TDB_LOCK_WAIT);
+	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, TDB_LOCK_NOWAIT);
+	int ret = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_NOWAIT, false);
 	tdb_trace_ret(tdb, "tdb_lockall_read_nonblock", ret);
 	return ret;
 }
@@ -580,7 +697,7 @@ 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, false);
+	return tdb_allrecord_unlock(tdb, F_RDLCK, false);
 }
 
 /* lock/unlock one hash chain. This is meant to be used to reduce
@@ -647,7 +764,7 @@ int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off)
 	if (tdb->allrecord_lock.count) {
 		return 0;
 	}
-	return off ? tdb->methods->brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0;
+	return off ? tdb_brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0;
 }
 
 /*
@@ -661,12 +778,21 @@ 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->brlock(tdb, F_WRLCK, off, 1, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
+	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);
 }
 
 int tdb_write_unlock_record(struct tdb_context *tdb, tdb_off_t off)
 {
-	return tdb->methods->brunlock(tdb, F_WRLCK, off, 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 */
@@ -684,36 +810,56 @@ 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->brunlock(tdb, F_RDLCK, off, 1) : 0);
+	return (count == 1 ? tdb_brunlock(tdb, F_RDLCK, off, 1) : 0);
 }
 
 bool tdb_have_extra_locks(struct tdb_context *tdb)
 {
-	if (tdb->allrecord_lock.count) {
+	unsigned int extra = tdb->num_lockrecs;
+
+	/* A transaction holds the lock for all records. */
+	if (!tdb->transaction && tdb->allrecord_lock.count) {
 		return true;
 	}
-	if (tdb->num_lockrecs) {
-		return true;
+
+	/* We always hold the active lock if CLEAR_IF_FIRST. */
+	if (find_nestlock(tdb, ACTIVE_LOCK)) {
+		extra--;
 	}
-	return false;
+
+	/* 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. */
+/* 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;
+	unsigned int i, extra = 0;
 
 	if (tdb->allrecord_lock.count != 0) {
-		tdb_brunlock(tdb, tdb->allrecord_lock.ltype,
-			     FREELIST_TOP, 4*tdb->header.hash_size);
+		tdb_brunlock(tdb, tdb->allrecord_lock.ltype, FREELIST_TOP, 0);
 		tdb->allrecord_lock.count = 0;
 	}
 
 	for (i=0;i<tdb->num_lockrecs;i++) {
-		tdb_brunlock(tdb, tdb->lockrecs[i].ltype,
-			     tdb->lockrecs[i].off, 1);
+		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);
 	}
-	tdb->num_locks = 0;
-	tdb->num_lockrecs = 0;
-	SAFE_FREE(tdb->lockrecs);
 }