"HWPOISON: report sticky EIO for poisoned file" still has a corner case
where we have possibilities of data lost. This is because in this fix
AS_HWPOISON is cleared when the inode cache is dropped.

For example, consider an application in which a process periodically
(every 10 minutes) writes some logs on a file (and closes it after
each writes,) and at the end of each day some batch programs run using
the log file. If a memory error hits on dirty pagecache of this log file
just after periodic write/close and the inode cache is cleared before the
next write, then this application is not aware of the error and the batch
programs will work wrongly.

To avoid this, this patch makes us pin the hwpoisoned inode on memory
until we remove or completely truncate the hwpoisoned file.

Signed-off-by: Naoya Horiguchi <redacted>
---
 fs/inode.c              | 12 ++++++++++++
 include/linux/pagemap.h | 11 +++++++++++
 mm/memory-failure.c     |  2 +-
 mm/truncate.c           |  2 ++
 4 files changed, 26 insertions(+), 1 deletion(-)

diff --git v3.6-rc1.orig/fs/inode.c v3.6-rc1/fs/inode.c
index ac8d904..8742397 100644
--- v3.6-rc1.orig/fs/inode.c
+++ v3.6-rc1/fs/inode.c

@@ -717,6 +717,15 @@ void prune_icache_sb(struct super_block *sb, int nr_to_scan)
 		}
 
 		/*
+		 * Keep inode caches on memory for user processes to certainly
+		 * be aware of memory errors.
+		 */
+		if (unlikely(mapping_hwpoison(inode->i_mapping))) {
+			spin_unlock(&inode->i_lock);
+			continue;
+		}
+
+		/*
 		 * Referenced or dirty inodes are still in use. Give them
 		 * another pass through the LRU as we canot reclaim them now.
 		 */

diff --git v3.6-rc1.orig/mm/truncate.c v3.6-rc1/mm/truncate.c
index 75801ac..82a994f 100644
--- v3.6-rc1.orig/mm/truncate.c
+++ v3.6-rc1/mm/truncate.c

@@ -574,6 +574,8 @@ void truncate_setsize(struct inode *inode, loff_t newsize)
 
 	oldsize = inode->i_size;
 	i_size_write(inode, newsize);
+	if (unlikely(mapping_hwpoison(inode->i_mapping) && !newsize))
+		mapping_clear_hwpoison(inode->i_mapping);