The iomap page fault mechanism currently dirties the associated page
after the full block range of the page has been allocated. This
leaves the page susceptible to delayed allocations without ever
being set dirty on sub-page block sized filesystems.

For example, consider a page fault on a page with one preexisting
real (non-delalloc) block allocated in the middle of the page. The
first iomap_apply() iteration performs delayed allocation on the
range up to the preexisting block, the next iteration finds the
preexisting block, and the last iteration attempts to perform
delayed allocation on the range after the prexisting block to the
end of the page. If the first allocation succeeds and the final
allocation fails with -ENOSPC, iomap_apply() returns the error and
iomap_page_mkwrite() fails to dirty the page having already
performed partial delayed allocation. This eventually results in the
page being invalidated without ever converting the delayed
allocation to real blocks.

This problem is reliably reproduced by generic/083 on XFS on ppc64
systems (64k page size, 4k block size). It results in leaked
delalloc blocks on inode reclaim, which triggers an assert failure
in xfs_fs_destroy_inode() and filesystem accounting inconsistency.

Move the set_page_dirty() call from iomap_page_mkwrite() to the
actor callback, similar to how the buffer head implementation works.
The actor callback is called iff ->iomap_begin() returns success, so
ensures the page is dirtied as soon as possible after an allocation.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>