Btrfs: fix deadlock due to unsubmitted
authorMiao Xie <miaox@cn.fujitsu.com>
Thu, 7 Feb 2013 10:12:07 +0000 (10:12 +0000)
committerJosef Bacik <jbacik@fusionio.com>
Wed, 20 Feb 2013 17:59:45 +0000 (12:59 -0500)
commit0934856d4697e63c14056375e26e3bd6e8ebd34b
tree5b47949ee3edf90519e901c657b862fa24c1adcc
parent4a7d0f6854c4a4ad1dba00a3b128a32d39b9a742
Btrfs: fix deadlock due to unsubmitted

The deadlock problem happened when running fsstress(a test program in LTP).

Steps to reproduce:
 # mkfs.btrfs -b 100M <partition>
 # mount <partition> <mnt>
 # <Path>/fsstress -p 3 -n 10000000 -d <mnt>

The reason is:
btrfs_direct_IO()
 |->do_direct_IO()
     |->get_page()
     |->get_blocks()
     |  |->btrfs_delalloc_resereve_space()
     |  |->btrfs_add_ordered_extent() ------- Add a new ordered extent
     |->dio_send_cur_page(page0) -------------- We didn't submit bio here
     |->get_page()
     |->get_blocks()
 |->btrfs_delalloc_resereve_space()
     |->flush_space()
 |->btrfs_start_ordered_extent()
     |->wait_event() ---------- Wait the completion of
the ordered extent that is
mentioned above

But because we didn't submit the bio that is mentioned above, the ordered
extent can not complete, we would wait for its completion forever.

There are two methods which can fix this deadlock problem:
1. submit the bio before we invoke get_blocks()
2. reserve the space before we do dio

Though the 1st is the simplest way, we need modify the code of VFS, and it
is likely to break contiguous requests, and introduce performance regression
for the other filesystems.

So we have to choose the 2nd way.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Cc: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
fs/btrfs/extent-tree.c
fs/btrfs/inode.c