jbd2: fix sending of data flush on journal commit

In data=ordered mode, it's theoretically possible (however rare) that
an inode is filed to transaction's t_inode_list and a flusher thread
writes all the data and inode is reclaimed before the transaction
starts to commit.  In such a case, we could erroneously omit sending a
flush to file system device when it is different from the journal
device (because data can still be in disk cache only).

Fix the problem by setting a flag in a transaction when some inode is added
to it and then send disk flush in the commit code when the flag is set.

Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>

2 files changed, 8 insertions, 2 deletions

diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c
index 78c299218681..2d5095ecc25f 100644
--- a/fs/jbd2/commit.c
+++ b/fs/jbd2/commit.c
@@ -219,7 +219,6 @@ static int journal_submit_data_buffers(journal_t *journal,
                         ret = err;
                 spin_lock(&journal->j_list_lock);
                 J_ASSERT(jinode->i_transaction == commit_transaction);
-                commit_transaction->t_flushed_data_blocks = 1;
                 clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
                 smp_mb__after_clear_bit();
                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
@@ -683,7 +682,7 @@ start_journal_io:
          * then we must flush the file system device before we issue
          * the commit record
          */
-        if (commit_transaction->t_flushed_data_blocks &&
+        if (commit_transaction->t_need_data_flush &&
             (journal->j_fs_dev != journal->j_dev) &&
             (journal->j_flags & JBD2_BARRIER))
                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
index 85a055ef93fe..20065c9f2479 100644
--- a/fs/jbd2/transaction.c
+++ b/fs/jbd2/transaction.c
@@ -2147,6 +2147,13 @@ int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
             jinode->i_next_transaction == transaction)
                 goto done;
 
+        /*
+         * We only ever set this variable to 1 so the test is safe. Since
+         * t_need_data_flush is likely to be set, we do the test to save some
+         * cacheline bouncing
+         */
+        if (!transaction->t_need_data_flush)
+                transaction->t_need_data_flush = 1;
         /* On some different transaction's list - should be
          * the committing one */
         if (jinode->i_transaction) {