that update of the timestamp and i_size needs to be moved to an I/O
completion handler.  We do this already to convert unwritten requests
to be written in fs/ext4/page_io.c.  See ext4_put_io_end_defer() in
fs/ext4/page_io.c; if we need to convert unwritten extents the
EXT4_IO_END_UNWRITTEN flag is set, and ext4_add_complete_io() tacks
the io_end queue onto a workqueue.  This infrastructure could be made
more general so that it can do other work after the I/O has been
completed, including the i_size update.

That's what we do for the i_size update in XFS.