On Mon, Nov 07, 2011 at 05:30:37AM -0500, Christoph Hellwig wrote:
> With Dmitris fsstress updates I can hit the following assert fairly
> [11904.943956] XFS: Assertion failed: ((iflags & BMV_IF_DELALLOC) != 0) ||
> (map[i].br_startblock != DELAYSTARTBLOCK), file:
> /home/hch/work/linux-2.6/fs/xfs/xfs_bmap.c, line: 5604
> which means we get an delalloc extent back from FIEMAP/GETBMAP despite
> asking for a flush beforehand. While we hold the iolock over the call
> and thus exclude new buffered writers from appearing that doesn't
> prevent shared writeable mmaps from creating new delalloc extents.
Possibly. However, we might actually be tripping over speculative
delalloc regions beyond EOF - flushing will not convert those at
all. There's already a comment earlier on like this:
* even after flushing the inode, there can still be delalloc
* blocks on the inode beyond EOF due to speculative
* preallocation. These are not removed until the release
* function is called or the inode is inactivated. Hence we
* cannot assert here that ip->i_delayed_blks == 0.
Can you see if that is the case that is being triggered?
> I don't think this actually is a real issue, and a workaround would
> be extremly hard. For now I've just remove the assert in my tree:
Probably the right thing to do, anyway.