On Thu, Jul 12, 2018 at 03:14:57PM +0200, Marek Vasut wrote:

On 07/12/2018 03:03 PM, Leonard Crestez wrote:

quoted

On Thu, 2018-07-12 at 11:21 +0200, Stefan Agner wrote:

quoted

On 10.07.2018 11:11, Marek Vasut wrote:

quoted

On 07/10/2018 11:06 AM, Stefan Agner wrote:

quoted

This is one of the situation where states quo is kinda the worst
situation.

Currently imx_v6_v7_defconfig and mxs_defconfig actually still uses
CONFIG_FB_MXS.

I understand that you'd rather prefer to move forward. I suggest we do
it in steps.

In 4.19:

- Change DRM driver.name to mxsfb-drm so we avoid conflicts for now

But this will break mesa if it depends on mxsfb name for ie. etnaviv
binding.

Does it? grep -r -e mxsfb in libdrm and mesa master returns nothing.

There is also .name in struct drm_driver, which is already set to
mxsfb-drm... Is that the one exposed to user space?

Running etnaviv+x11 with a renamed mxsfb driver works fine on imx6sx-
sdb.

Tools like modetest already need -M mxsfb-drm, the drm_driver.name
seems to be what matters.

quoted

- Remove CONFIG_FB_MXS from imx_v6_v7_defconfig/mxs_defconfig now, and
only enable CONFIG_DRM_MXSFB=y

If one of the drivers is renamed then they can coexist: since the
bindings are distinct one driver will return a probe error and the
other will bind successfully. This can even be adjusted so that it
doesn't even print ugly scary errors.

This can last until somebody implements support for old bindings in the
drm driver and then FB_MXS can just be deleted.

So why don't we just convert the DT bindings on boards supported
upstream and zap the old driver ? What is the problem with that?

+1 on zapping drivers :-)

Realistically, how many MXS boards in the field use old DT and new kernel ?

Yeah I think occasionally that entire "DT is API, can't ever change it"
song is overblown. The rule with regressions isn't that you're never
allowed to break anything, but that you're only allowed to break stuff no
one will notice and report.

If there's realistically no users, go ahead and break (instead of huge and
drawn-out compat plan).

If there's a solid (and automatic enough) fallback like the drm driver,
go ahead and break.

We're doing this all the time in graphics with userspace ABI, I don't
think DT is any different. Every once in a while there's a bit of regrets
and a revert because we missed something, but overal it's much less effort
than always trying to do a perfect job with backwards compat.

Is it really a regression if no one reports it? No.

Cheers, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch