On 5/27/22 8:45AM, Matthew Wilcox wrote:
On Thu, May 26, 2022 at 10:41:31PM -0700, Christoph Hellwig wrote:
I suspect this might be where your problems lies:
blockdev_direct_IO calls __blockdev_direct_IO with DIO_SKIP_HOLES set.
DIO_SKIP_HOLES causes get_more_blocks to never set the create bit
to get_block except for writes beyond i_size. If we want to replicate
that behavior with iomap, ->iomap_begin needs to return -ENOTBLK
when it encounters a hole for writing. To properly supporting writing
to holes we'd need unwritten extents, which jfs does not support.
gfs2 might be a place to look for how to implement this.
I think JFS does support unwritten extents,
fs/jfs/jfs_xtree.h:#define XAD_NOTRECORDED 0x08 /* allocated but not recorded */
However, we always pass 'false' to extAlloc() today, so I think it
hasn't been tested in a while? I'm not sure I want to be the one to
start using new features on JFS for something that's supposed to be
a relatively quick cleanup.
If I remember correctly, there was an intention to implement unwritten
extents in the future, but it never got implemented. We tried to
anticipate the unwritten extents in the existing code as much as possible.
Shaggy
