[Intel-gfx] RMW considered harmful (was: Re: [PATCH 2/2] drm/i915/icl: Enable TRANSCODER PORT SYNC for tiled displays across separate ports)

[Jani Nikula]

On Fri, 22 Mar 2019, Ville Syrjälä <ville.syrjala at linux.intel.com> wrote:
> On Fri, Mar 22, 2019 at 11:44:21AM -0700, Manasi Navare wrote:
>> On Fri, Mar 22, 2019 at 08:09:50PM +0200, Ville Syrjälä wrote:
>> > In that case there is no point in doing a rmw.
>> 
>> But isnt it always a good idea to do rmw? I mean what if the master
>> select was set to something else earlier?
>
> RMW is the root of many evils. It should be avoided unless there is a
> really compelling reason to use it.

Hear, hear!

We have the software state that we want to write to the hardware. If we
use RMW to do this, it might all work by coincidence due to the old
values in the registers, or it might just as well break by coincidence
due to some garbage in the registers.

In most cases, there should only be one place that writes a particular
display register during modeset. Sometimes this isn't possible, and RMW
is required.

Some registers also have reserved bits potentially used by the hardware
that must not be changed, and RMW is required. These are documented in
bspec.

BR,
Jani.


-- 
Jani Nikula, Intel Open Source Graphics Center