X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb%2Flock.c;h=dfcac7a540adc30482dcd86e0c299b36bfb52e0b;hp=fd451e58cd314b54167b476e8a74890aa8e23581;hb=448b9194f76a138eef05047b1aaa5d211f72e97f;hpb=497e23b21fe7dc8256c23f79f14cdb4c3e86cc9e;ds=sidebyside

diff --git a/ccan/tdb/lock.c b/ccan/tdb/lock.c
index fd451e58..dfcac7a5 100644
--- a/ccan/tdb/lock.c
+++ b/ccan/tdb/lock.c
@@ -152,14 +152,6 @@ int tdb_brlock(struct tdb_context *tdb,
 		return -1;
 	}
 
-	/* 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_lock(tdb, rw_type, offset, len,
 				 flags & TDB_LOCK_WAIT);
@@ -309,10 +301,44 @@ int tdb_nest_lock(struct tdb_context *tdb, uint32_t offset, int ltype,
 	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 &&
@@ -324,7 +350,18 @@ static int tdb_lock_list(struct tdb_context *tdb, int list, int ltype,
 		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;
 }
@@ -440,11 +477,9 @@ int tdb_transaction_unlock(struct tdb_context *tdb, int ltype)
 	return tdb_nest_unlock(tdb, TRANSACTION_LOCK, ltype, false);
 }
 
-
-/* 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)
+/* 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) {
@@ -474,11 +509,73 @@ int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
 		tdb->ecode = TDB_ERR_LOCK;
 		return -1;
 	}
+	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)));
-		}
+/* 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;
 	}
 
@@ -488,6 +585,21 @@ int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
 	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;
 }