jbd2: avoid long hold times of j_state_lock while committing a transaction

We can hold j_state_lock for writing at the beginning of
jbd2_journal_commit_transaction() for a rather long time (reportedly for
30 ms) due cleaning revoke bits of all revoked buffers under it. The
handling of revoke tables as well as cleaning of t_reserved_list, and
checkpoint lists does not need j_state_lock for anything. It is only
needed to prevent new handles from joining the transaction. Generally
T_LOCKED transaction state prevents new handles from joining the
transaction - except for reserved handles which have to allowed to join
while we wait for other handles to complete.

To prevent reserved handles from joining the transaction while cleaning
up lists, add new transaction state T_SWITCH and watch for it when
starting reserved handles. With this we can just drop the lock for
operations that don't need it.

Reported-and-tested-by: Adrian Hunter <adrian.hunter@intel.com>
Suggested-by: "Theodore Y. Ts'o" <tytso@mit.edu>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

1 files changed, 38 insertions, 5 deletions

diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
index c0b66a7a795b..116d8251fbff 100644
--- a/fs/jbd2/transaction.c
+++ b/fs/jbd2/transaction.c
@@ -138,9 +138,9 @@ static inline void update_t_max_wait(transaction_t *transaction,
 }
 
 /*
- * Wait until running transaction passes T_LOCKED state. Also starts the commit
- * if needed. The function expects running transaction to exist and releases
- * j_state_lock.
+ * Wait until running transaction passes to T_FLUSH state and new transaction
+ * can thus be started. Also starts the commit if needed. The function expects
+ * running transaction to exist and releases j_state_lock.
  */
 static void wait_transaction_locked(journal_t *journal)
 	__releases(journal->j_state_lock)
@@ -160,6 +160,32 @@ static void wait_transaction_locked(journal_t *journal)
 	finish_wait(&journal->j_wait_transaction_locked, &wait);
 }
 
+/*
+ * Wait until running transaction transitions from T_SWITCH to T_FLUSH
+ * state and new transaction can thus be started. The function releases
+ * j_state_lock.
+ */
+static void wait_transaction_switching(journal_t *journal)
+	__releases(journal->j_state_lock)
+{
+	DEFINE_WAIT(wait);
+
+	if (WARN_ON(!journal->j_running_transaction ||
+		    journal->j_running_transaction->t_state != T_SWITCH))
+		return;
+	prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
+			TASK_UNINTERRUPTIBLE);
+	read_unlock(&journal->j_state_lock);
+	/*
+	 * We don't call jbd2_might_wait_for_commit() here as there's no
+	 * waiting for outstanding handles happening anymore in T_SWITCH state
+	 * and handling of reserved handles actually relies on that for
+	 * correctness.
+	 */
+	schedule();
+	finish_wait(&journal->j_wait_transaction_locked, &wait);
+}
+
 static void sub_reserved_credits(journal_t *journal, int blocks)
 {
 	atomic_sub(blocks, &journal->j_reserved_credits);
@@ -183,7 +209,8 @@ static int add_transaction_credits(journal_t *journal, int blocks,
 	 * If the current transaction is locked down for commit, wait
 	 * for the lock to be released.
 	 */
-	if (t->t_state == T_LOCKED) {
+	if (t->t_state != T_RUNNING) {
+		WARN_ON_ONCE(t->t_state >= T_FLUSH);
 		wait_transaction_locked(journal);
 		return 1;
 	}
@@ -360,8 +387,14 @@ repeat:
 		/*
 		 * We have handle reserved so we are allowed to join T_LOCKED
 		 * transaction and we don't have to check for transaction size
-		 * and journal space.
+		 * and journal space. But we still have to wait while running
+		 * transaction is being switched to a committing one as it
+		 * won't wait for any handles anymore.
 		 */
+		if (transaction->t_state == T_SWITCH) {
+			wait_transaction_switching(journal);
+			goto repeat;
+		}
 		sub_reserved_credits(journal, blocks);
 		handle->h_reserved = 0;
 	}