[PATCH] modes: Force output modes for manually enabled heads

[Aaron Plattner]

On Thu, Oct 08, 2009 at 01:09:43PM -0700, Adam Jackson wrote:
> On Thu, 2009-10-08 at 10:44 -0700, Aaron Plattner wrote:
> > On Thu, Oct 08, 2009 at 10:09:20AM -0700, Keith Packard wrote:
> > > Excerpts from Chris Wilson's message of Thu Oct 08 09:43:16 -0700 2009:
> > > > In a headless configuration, in order to convince the XServer to start
> > > > the user must explicitly enable an output.
> > > 
> > > I see that as the fundemental problem here. I'd suggest that the
> > > server should allow the device to start without any outputs connected
> > > and create a default sized frame buffer (1024x768?).
> > 
> > For boards that actually have output connectors (i.e. not Tesla cards),
> > this is a bad idea.  The number of users who actually want a headless
> > configuration is much smaller than the number of users who would be
> > confused and infuriated when starting X makes all of their displays go
> > blank due to misconfiguration or driver bugs and results in what is
> > essentially a hung machine.
> > 
> > X failing to start with an informative log message is far better than a
> > frozen computer that requires the reset button to fix.  I think making
> > people explicitly configure X to be headless is a reasonable compromise.
> 
> Disagree.  Servers regularly start with no monitor attached and admins
> expect to be able to walk up and plug in a display and have it work.

How do you propose having that happen?  Having X autodetect hotplugged
screens and automatically extend the desktop to them?  Does that work
today?

If you're talking about always driving a VGA port and hoping the monitor
can support the signal it's driving, then that still requires a display
device at startup and manual configuration of the mode because there's no
EDID.

I think we need to choose between the following:

 1. Allow X to start with no displays.  Automatically light them up when
    they're plugged in (from within the server -- don't rely on a client to
    do that).  Make damn sure that display hotplug detection is 100%
    reliable, even for VGA.

 2. Fail to start X unless we're reasonably sure that the user will be able
    to see something.

I don't think any combination of the two is acceptible because it can lead
to what are essentially broken computers.

> Presumably this implies moving to a model where X is started as a
> trigger from the kernel seeing a connected output, and not the current
> "start and hope".

No, in that case, failing to start X if it can't manage to find the display
the kernel detected is *still* the right behavior.