[RFC PATCH 05/11] ext4: data=journal: prevent journalled writeback deadlock in __ext4_journalled_writepage()

[Mauricio Faria de Oliveira <mfo@canonical.com>]

This patch checks for and prevents the deadlock with a page in
journalled writeback in __ext4_journalled_writepage().

It turns out that __ext4_journalled_writepage() may race with itself
in another task (e.g., two tasks set the same mmap'ed page dirty and
called msync(), one after the other), so we have to check/handle it.

Signed-off-by: Mauricio Faria de Oliveira <mfo@canonical.com>
---
 fs/ext4/inode.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 574a062b8bcd..401313be8a5b 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -1910,6 +1910,7 @@ static noinline int __ext4_journalled_writepage(struct page *page,
 	 */
 	ejwp = kmem_cache_alloc(ext4_journalled_wb_page_cachep,	GFP_KERNEL);
 
+retry_journal:
 	handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
 				    ext4_writepage_trans_blocks(inode));
 	if (IS_ERR(handle)) {
@@ -1948,7 +1949,20 @@ static noinline int __ext4_journalled_writepage(struct page *page,
 			goto out;
 		}
 
-		/* Data integrity writeback; have to wait and do it. */
+		/*
+		 * Data integrity writeback; have to wait and do it.
+		 *
+		 * Check for deadlock with page in journalled writeback
+		 * (i.e. another task running/ran this function as well)
+		 */
+		if (ext4_check_journalled_writeback(handle, page)) {
+			get_page(page);
+			unlock_page(page);
+			ext4_journal_stop(handle);
+			ext4_start_commit_datasync(inode);
+			wait_on_page_writeback(page);
+			goto retry_journal;
+		}
 		wait_on_page_writeback(page);
 	}
 
-- 
2.20.1