[Intel-gfx] [PATCH 3/4] drm/i915: enable VT switchless resume v3

[Jesse Barnes]

On Fri, 16 May 2014 23:42:23 +0100
Chris Wilson <chris at chris-wilson.co.uk> wrote:

> On Fri, May 16, 2014 at 11:38:07PM +0100, Chris Wilson wrote:
> > On Fri, May 16, 2014 at 03:20:47PM -0700, Jesse Barnes wrote:
> > > On Mon, 21 Apr 2014 18:37:31 +0200
> > > Knut Petersen <Knut_Petersen at t-online.de> wrote:
> > > > > +	/* This driver doesn't need a VT switch to restore the mode on resume */
> > > > > +	info->skip_vt_switch = true;
> > > > > +
> > > > >   	drm_fb_helper_fill_fix(info, fb->pitches[0], fb->depth);
> > > > >   	drm_fb_helper_fill_var(info, &ifbdev->helper, sizes->fb_width, sizes->fb_height);
> > > 
> > > Is it sufficient to just revert this part?  Or are the other bits
> > > needed too?
> > > 
> > > Sorry for the delay on this, I've been traveling a lot the past month
> > > and buried in other stuff so out of touch with much of my email.
> > 
> > The key step here is that X is restarted after resume. The slow down was
> > due to X not finding any connected outputs and so disabling them all,
> > setting up a dummy 1024x768 fb which really confused KDE. (KDE queries
> > the config causing a forced reprobe of all outputs, setups the display
> > for the native 1280x1024, but screws up KDE's own bookkeeping.)
> > 
> > The impact has been fixed by handling the connector->status more
> > robusting during initial output probing in X. What remains is the
> > question whether connector->status can be set appropriately upon resume?
> > It requires a detection cycle to be sure that the outputs are still
> > there, which is arguably better deferred to userspace. To be consistent
> > the BIOS take over code should mark connector->status as unknown for the
> > CRTCs it takes over without doing a detection cycle (where we just
> > presume that the CRTC/output being enabled means something is on the
> > other end of the pipe and is still valid).
> 
> Hmm. Why didn't fbcon respond to the hotplug event on resume and perform
> a detection cycle before Knut was able to type startx on the console?
> That I think is the bug.

Well, fbcon resume is delayed, but we do perform an fb re-probe via
intel_resume_hotplug() that should have done this.  Not sure why that's
sufficient... but I agree userspace should probably re-probe on
resume.  We're supposed to generate a uevent on resume, but the fb
layer may suppress that if it detects that no change has occurred.

Maybe X is racing with our resume_hotplug somehow?  It doesn't look
like that should happen...

-- 
Jesse Barnes, Intel Open Source Technology Center