From: Jan Kara <jack@suse.cz>
Subject: ext4: Fixup pages without buffers
References: bsc#1205495
Patch-mainline: Never, upstream has page pinning tracking infrastructure which should be used for this

When application use RDMA or various offload engines they pin pages which they
dirty once they are done with these pages. However the page may be cleaned and
buffers reclaimed before it is dirtied which confuses ext4 writeback code.
Detect such case and restore page buffers as good as we can.

Signed-off-by: Jan Kara <jack@suse.cz>

---
 fs/ext4/inode.c |   35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -2122,6 +2122,37 @@ out_no_pagelock:
 }
 
 /*
+ * Create buffers under a dirty page. This is a hack for the case when someone
+ * pinned the page, the page got cleaned and buffers released, and then the
+ * pinning process dirtied the page
+ */
+static void ext4_restore_page_buffers(struct inode *inode, struct page *page)
+{
+	int bsize = 1 << inode->i_blkbits;
+	struct buffer_head *bh, *head;
+	loff_t isize = i_size_read(inode);
+	unsigned int off, end;
+	ext4_lblk_t iblock = page->index << (PAGE_SHIFT - inode->i_blkbits);
+
+	if (page->index == isize >> PAGE_SHIFT)
+		end = isize & (PAGE_SIZE - 1);
+	else
+		end = PAGE_SIZE;
+
+	create_empty_buffers(page, bsize, 0);
+	bh = head = page_buffers(page);
+	off = 0;
+	do {
+		if (off > end)
+			break;
+		ext4_get_block(inode, iblock, bh, 0);
+		off += bsize;
+		iblock++;
+		bh = bh->b_this_page;
+	} while (bh != head);
+}
+
+/*
  * Note that we don't need to start a transaction unless we're journaling data
  * because we should have holes filled from ext4_page_mkwrite(). We even don't
  * need to file the inode to the transaction's list in ordered mode because if
@@ -2186,6 +2217,8 @@ static int ext4_writepage(struct page *p
 	else
 		len = PAGE_SIZE;
 
+	if (!page_has_buffers(page))
+		ext4_restore_page_buffers(inode, page);
 	page_bufs = page_buffers(page);
 	/*
 	 * We cannot do block allocation or other extent handling in this
@@ -2731,6 +2764,8 @@ static int mpage_prepare_extent_to_map(s
 			wait_on_page_writeback(page);
 			BUG_ON(PageWriteback(page));
 
+			if (!page_has_buffers(page))
+				ext4_restore_page_buffers(mpd->inode, page);
 			if (mpd->map.m_len == 0)
 				mpd->first_page = page->index;
 			mpd->next_page = page->index + 1;