tdb: workaround starvation problem in locking entire database.

We saw tdb_lockall() take 71 seconds under heavy load; this is because Linux
(at least) doesn't prevent new small locks being obtained while we're waiting
for a big log.

The workaround is to do divide and conquer using non-blocking chainlocks: if
we get down to a single chain we block.  Using a simple test program where
children did "hold lock for 100ms, sleep for 1 second" the time to do
tdb_lockall() dropped signifiantly.  There are ln(hashsize) locks taken in
the contended case, but that's slow anyway.

More analysis is given in my blog at http://rusty.ozlabs.org/?p=120

This may also help transactions, though in that case it's the initial
read lock which uses this gradual locking routine; the update-to-write-lock
code is separate and still tries to update in one go.

Even though ABI doesn't change, minor version bumped so behavior change
can be easily detected.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

1 files changed, 69 insertions, 17 deletions

diff --git a/lib/tdb/common/lock.c b/lib/tdb/common/lock.c
index 285b7a34c3..803feeecbb 100644
--- a/lib/tdb/common/lock.c
+++ b/lib/tdb/common/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);
@@ -486,11 +478,9 @@ 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) {
@@ -520,11 +510,73 @@ 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;
 	}