Re: XBF_DONE semantics

[Christoph Hellwig <hch@lst.de>]

On Wed, Nov 29, 2023 at 08:08:37AM +1100, Dave Chinner wrote:
> > But places that use buf_get and manually fill in data only use it in a
> > few cases. 
> 
> Yes. the caller of buf_get always needs to set XBF_DONE if it is
> initialising a new buffer ready for it to be written. It should be
> done before the caller drops the buffer lock so that no other lookup
> can see the buffer in the state of "contains valid data but does not
> have XBF_DONE set".

That makes sense, but we do have a whole bunch of weird things going
on as well:

 - xfs_buf_ioend_handle_error sets XBF_DONE when retrying or failing
 - xfs_buf_ioend sets XBF_DONE on successful write completion as well
 - xfs_buf_ioend_fail drops XBF_DONE for any I/O failure
 - xfs_do_force_shutdown sets XBF_DONE on the super block buffer on
   a foced shutdown
 - xfs_trans_get_buf_map sets XBF_DONE on a forced shutdown

So there's definitively a bunch of weird things not fully in line
with the straight forward answer.