f2fs: fix BUG_ON during f2fs_evict_inode(dir)
authorJaegeuk Kim <jaegeuk.kim@samsung.com>
Wed, 15 May 2013 07:40:02 +0000 (16:40 +0900)
committerJaegeuk Kim <jaegeuk.kim@samsung.com>
Tue, 28 May 2013 06:03:01 +0000 (15:03 +0900)
commit74d0b917ef7789097e12d60fc054efa427ce9171
tree1114993f6b17dfb268688ec2cdcbf80fbbdcf075
parent8c26d7d5717adf7f06d98c4416852d09566edd7c
f2fs: fix BUG_ON during f2fs_evict_inode(dir)

During the dentry recovery routine, recover_inode() triggers __f2fs_add_link
with its directory inode.

In the following scenario, a bug is captured.
 1. dir = f2fs_iget(pino)
 2. __f2fs_add_link(dir, name)
 3. iput(dir)
  -> f2fs_evict_inode() faces with BUG_ON(atomic_read(fi->dirty_dents))

Kernel BUG at ffffffffa01c0676 [verbose debug info unavailable]
[<ffffffffa01c0676>] f2fs_evict_inode+0x276/0x300 [f2fs]
Call Trace:
 [<ffffffff8118ea00>] evict+0xb0/0x1b0
 [<ffffffff8118f1c5>] iput+0x105/0x190
 [<ffffffffa01d2dac>] recover_fsync_data+0x3bc/0x1070 [f2fs]
 [<ffffffff81692e8a>] ? io_schedule+0xaa/0xd0
 [<ffffffff81690acb>] ? __wait_on_bit_lock+0x7b/0xc0
 [<ffffffff8111a0e7>] ? __lock_page+0x67/0x70
 [<ffffffff81165e21>] ? kmem_cache_alloc+0x31/0x140
 [<ffffffff8118a502>] ? __d_instantiate+0x92/0xf0
 [<ffffffff812a949b>] ? security_d_instantiate+0x1b/0x30
 [<ffffffff8118a5b4>] ? d_instantiate+0x54/0x70

This means that we should flush all the dentry pages between iget and iput().
But, during the recovery routine, it is unallowed due to consistency, so we
have to wait the whole recovery process.
And then, write_checkpoint flushes all the dirty dentry blocks, and nicely we
can put the stale dir inodes from the dirty_dir_inode_list.

Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>
fs/f2fs/checkpoint.c
fs/f2fs/f2fs.h
fs/f2fs/recovery.c