Bug 206211

Hello.

We encounted a BUG_ON() case when we use a JAVA's persistent caching library Ehcache on btrfs.
And we tried to find out the cause of BUG_ON() case and we also tried to fix this bug.
Details are following.

[Bug]
BUG_ON() assertion fails in __do_readpage() fs/btrfs/extent_io.c.

static int __do_readpage(...) {
	...
	while (cur <= end) {
		...
		em = __get_extent_map(inode, page, pg_offset, cur, 
				end - cur + 1, get_extent, em_cached);
		...
		BUG_ON(extent_map_end(em) <= cur); // <-here
		...
	}
	...
}

kern.log is attached.

[How to reproduce]
Keep attached program running on btrfs. This program is a simplified JAVA program which uses Ehcache. The bug can be reproduced in 30 minutes to 2 hours.
We tried to find out easier reproduction code (it may be a user space C program), but we have not yet.

[Cause]
Extent map struct is over-written by another thread.
It might be a race condition.
In try_merge_map(), the extent map is over-written even though another thread is still reading. 
(We figured out previous facts by some experiments. See experiments.txt for more details.)

[Improvement]
To eliminate the race condition, it is necessary to wait for the reference count of the extent map to decrease.
By adapting this improvement, no bugs occurred for more than 72 hours running previous program.

Signed-off-by: ryusuke1925 <st13s20@gm.ibaraki-ct.ac.jp>
---
 extent_map.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/extent_map.c b/extent_map.c
index 6f417ff..6b0fd84 100644
--- a/extent_map.c
+++ b/extent_map.c
@@ -227,6 +227,8 @@ static int mergable_maps(struct extent_map *prev, struct extent_map *next)
 	      prev->block_start == EXTENT_MAP_INLINE) ||
 	     (next->block_start < EXTENT_MAP_LAST_BYTE - 1 &&
 	      next->block_start == extent_map_block_end(prev)))) {
+		while (refcount_read(&next->refs) > 2) {
+		}
 		return 1;
 	}
 	return 0;
-- 
2.17.1

We'd like to fix this bug, because the bug still occured on our production environment.
Could you advice me next actions (sending patch? or studying more about BTRFS?) .

Thanks,

Ryusuke