[Intel-gfx] [PATCH] drm/fb-helper: Restore first connection behaviour on deferred setup

[Daniel Vetter]

On Mon, Jul 03, 2017 at 09:44:50AM +0100, Liviu Dudau wrote:
> On Fri, Jun 30, 2017 at 08:13:58PM +0200, Daniel Vetter wrote:
> > On Fri, Jun 30, 2017 at 6:51 PM, Liviu Dudau <Liviu.Dudau at arm.com> wrote:
> > > Prior to commit b0aa06e9a7fd ("drm/fb-helper: Support deferred setup"),
> > > if no output is connected at framebuffer setup time, we get a default
> > > 1024x768 mode that is going to be used when we first connect a monitor.
> > > After the commit, on first connection after deferred setup, we probe
> > > the monitor and get the preferred resolution, but no mode get set
> > > because the drm_fb_helper_hotplug_event() function returns early
> > > when the setup has been deferred. That is different from what happens
> > > on a second re-connect of the monitor, when the native mode get set.
> > >
> > > Create a more consistent behaviour by checking in the
> > > drm_fb_helper_hotplug_event() function if the deferred setup is still
> > > active. If not, that means we now have a valid framebuffer that can be
> > > used for setting the correct mode.
> > >
> > > Fixes: b0aa06e9a7fd ("drm/fb-helper: Support deferred setup")
> > > Signed-off-by: Liviu Dudau <Liviu.Dudau at arm.com>
> > > Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
> > 
> > I thought the analysis over irc was that fbcon picked a different
> > driver as it's console, and that's why nothing shows up on the malidp
> > output in the deferred case? That's mildly annoying, but iirc fbcon
> > has always been rather erratic in multi-gpu setups. Although I thought
> > that it would by default bind all fbdev drivers as consoles (and then
> > you need to rebind the right console driver, if the right Kconfig is
> > enabled, through sysfs).
> 
> That's right, fbcon picks a different driver and binds to it. But
> before the patch, on the first connection with a monitor that happened
> after setup (i.e. when the fbdev emulation mode set the default 1024x768
> mode), a full atomic modeset was trigger for the driver that had the
> connection and the "card" output was enabled. Now with your patch that
> does not happen on the first connection, but it does after I turn off and
> back on the monitor. I know that is how most computer problems are solved,
> but still ... :)
> 
> > 
> > Either way if the register_framebuffer() call in initial_config isn't
> > good enough, then we need to add the set_par in initial_config
> > unconditionally, not just in the deferred probe case. Just disable
> > fbcon entirely for testing, in that case even without deferred probing
> > nothing will show up.
> 
> Actually the un-ballanced call to set_par happens with your patch (it
> gets called for all hotplug events that are *not* deferred. I was trying
> to balance things and call set_par also for deferred setups where the
> __drm_fb_helper_initial_config() call succeeded. And the fact that we do
> call set_par in the hotplug path even when the setup is not deferred
> tells me that register_framebuffer() is probably not enough.

But then you end up doing it twice if that's the fbdev fbcon will bind
too.

And of course we must call set_par to enable the newly hotplugged screen,
it'd be black otherwise. It's just that with just 1 gpu, the
register_framebuffer takes care of the initial config stuff (if you have
fbcon enabled).

If you boot with just 1 gpu, and hotplug that one, it will work correctly.
Multi-gpu + fbcon just don't work in a predictable fashion.

> > I'd say if this is still needed in the single gpu case then we need to
> > investigate more, but for multi-gpu it is what it is (aka fbcon is not
> > great).
> 
> I think single gpu case hides the problem because fbcon forces the right
> sequence of calls. And I think the old behaviour was correct, as I was
> getting output of the second "card" on hotplug, now I have to turn the
> monitor off and on again to get some signal sync.

So the problem is that we rely upon register_framebuffer->fbcon to create
that initial set_par. We could add an unconditional set_par to fix this
bug (the condition really is pre-existing, if you'd boot with 2 gpus and
both connected at start-up the 2nd one would still be off). But that comes
at the cost of forcing everyone to do a 2nd, unecessary modeset right
after boot-up, potentially destroying neat tricks like fastboot.

So we can't just do that either, because that would regress boot-up time.
Now if you have an idea how to fix this without rewriting fbcon+fbdev, we
could try, but I'm not sure that's feasible really ...

After all if you entirely disable fbdev emulation, your screen also stays
off. And maybe you event _want_ it to stay off, even with fbdev emulation
enabled.

So really no idea what exactly we should do here, but adding an
unconditional set_par in the initial_config path isn't the solution I
think.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch