Re: [RFC PATCH] cxl/region: Allow out of order assembly of autodiscovered regions

[Dan Williams <dan.j.williams@intel.com>]

Alison Schofield wrote:
> > I can see it being an alternate way, but not sure how to weigh one
> > approach vs the other. Is the dynamic update getting in the way of a
> > test case you are thinking about? Otherwise it seemed easy to reason
> > that the registry is always on, but only takes effect when
> > registry_reset_disable is set, and cxl_acpi rebinds the test device.
> 
> The constant work of updating the registry caught my attention, but
> no impact that I know of. The dynamic approach is more intrusive and
> impacts the normal path needlessly. A snapshot approach limits much
> of the impact to users of the new feature.
> 

It does mean more infrastructure to walk the entire topology and save
off all of the decoders. However, this is straightforward because
decoders are devices on the CXL bus.

I am not worried about the intrusiveness as much as how easy it is to
conceptualize new tests. I do think it is easier to conceptualize that
the flow is

"get the decoders the way you want them, snapshot them, ..., rebind",

...rather than the current:

"get the decoders the way you want them, disable-reset, rebind"

Notice the "..." in the first flow where it is a bit more forgiving if
you do some decoder operations between the snapshot step and the rebind
step. Where the latter flow requires that you know the side effects of
how rebind tries to reset the decoder state.

It still ends up with the all same complexity on the read side, but
should be easier to craft new static configurations without needing to
worry about the write side of the registry.