[Intel-gfx] [PATCH] drm/i915: Resume DP MST before doing any kind of modesetting

Wed Mar 2 09:27:15 UTC 2016

On Tue, Mar 01, 2016 at 07:33:12AM +1000, Dave Airlie wrote:
> On 1 March 2016 at 02:12, Daniel Vetter <daniel at ffwll.ch> wrote:
> > On Wed, Feb 24, 2016 at 08:03:04AM +0530, Thulasimani, Sivakumar wrote:
> >>
> >>
> >> On 2/24/2016 3:41 AM, Lyude wrote:
> >> >As it turns out, resuming DP MST is racey since we don't make sure MST
> >> >is ready before we start modesetting, it just usually happens to be
> >> >ready in time. This isn't the case on all systems, particularly a
> >> >ThinkPad T560 with displays connected through the dock. On these
> >> >systems, resuming the laptop while connected to the dock usually results
> >> >in blank monitors. Making sure MST is ready before doing any kind of
> >> >modesetting fixes this issue.
> >> basic question since i haven't worked on MST much. MST should work like any
> >> other digital panel on resume. i.e detect followed by modeset. in the
> >> modified
> >> commit mentioned below is it failing to detect the panel or failing at the
> >> modeset ?
> >> if we are depending on the intel_display_resume, how about moving the
> >> intel_dp_mst_resume just above intel_display_resume?
> >>
> >>
> >> Generic question to others in mail list on i915_drm_resume
> >> we are doing a modeset and then doing the detect/hpd init.
> >> shouldn't this be the other way round ? almost all displays
> >> will pass a modeset even if display is not connected so we
> >> are spending time on modeset even for displays that were
> >> removed during the suspend state. if this is to simply
> >> drm_state being saved and restored,
> >
> > We must restore anyway, we're not really allowed to shut down a display
> > without userspace's consent. DP mst breaks this, and it's not fun really.
> > So for hotunplug the flow should always be:
> > 1. kernel notices something has changed in the output config.
> > 2. kernel sends out uevent
> > 3. userspace figures out what changed, and what needs to be done
> > 4. userspace asks the kernel to change display configuration through
> > setCrtc and Atomic ioctl calls.
> >
> > Irrespective of hotunplug handling, the kernel also _must_ restore the
> > entire display configuration before userspace resumes. We can do that
> > asynchronously (and there's plans for that), but even then we must stall
> > userspace on the first KMS ioclt to keep up the illusion that a system s/r
> > is transparent.
> >
> > In short, even if we do hpd processing before resuming the display,
> > nothing will be faster - we must wait for userspace to make up its mind,
> > and that can only happen once we've restored the display config.
> >
> > And again, mst is kinda breaking this, since and mst unplug results in a
> > force-disable. Which can upset userspace and in general results in the
> > need for lots of fragile error handling all over.
> 
> >
> >> >We originally changed the resume order in
> >> >
> >> >     commit e7d6f7d70829 ("drm/i915: resume MST after reading back hw state")
> >> >
> >> >to fix a ton of WARN_ON's after resume, but this doesn't seem to be an
> >> >issue now anyhow.
> >> >
> >> >CC: stable at vger.kernel.org
> >> >Signed-off-by: Lyude <cpaul at redhat.com>
> >
> > Wrt the patch itself: I think only in 4.6 we've actually fixed up all
> > these mst WARN_ON. Cc: stable seems extremely risky on this one. Also, the
> > modeset state readout for mst is still not entirely correct (it still
> > relies on software state), so I'm not sure we've actually managed to shut
> > up all the WARNINGs. Iirc the way to repro them is to hot-unplug something
> > while suspended. In short the get_hw_state functions for mst should not
> > rely on any mst software state, but instead just look at hw registers and
> > dp aux registers to figure out what's going on. But not sure whether this
> > will result on WARNINGs on resume, since generally the bios doesn't enable
> > anything in that case.
> >
> > Furthermore MST still does a force-modeset when processing a hotunplug.
> > Doing that before we've resumed the display is likely a very bad idea.
> > What we need to fix that part is to separate the dp mst connector
> > hotplug/unplugging from actually updating the modeset change. This needs
> > reference-counting on drm_connector (so that we can lazily free
> > drm_connector structs after hot-unplug), and is a major change.
> 
> I agree on the dropping stable on this, just put it in 4.6 for now, actually I'm
> happy to send in for 4.5 at this stage as it fixes some laptops I have as well.
> 
> However I'm not sure how you think we can keep hw state programmed when
> the hw is gone away until userspace comes along to do something about it,
> the hw state is mostly on the other side of the MST link which has also gone
> away. My experience with leaving the crtcs programmed was and I might not
> be 100% sure was hw hangs, I even managed to get some MCEs over time.
> 
> Though I might be tempted to go investigate this a bit more.

Hm ... The issue is that stopping the pipe behind userspace back is kinda
really unfriendly. Means you need to sprinkle error handling checks and
fallback code all over, and ofc that code is little-exercise and hence
likely buggy. If it exists at all. That's kinda why I think keeping the
pipe up&running would be the ideal ABI to present to userspace, if we can
do it. Also consistent with what we do for non-mst.

But yeah if this MCEs then not much luck. Otoh I can't see how this could
MCE if we really don't touch anything of the source-side state and just
keep feeding the exact same MST config to the sinks, whether they're still
there or not. That /should/ work, otherwise you could just die when you
unplug. Ofc we need to make any unplugged mst connector as dead, so that
the only thing userspace is allowed to do is disable it. And at that point
it'll get garbage collected.

But hey, that's just my long-term mst dream land ;-) For reality, whatever
floats. But I'd like to figure out what exactly is going on here first,
and Lyude is still investigating (and now also pulling in our intel hw
architects to figure this out).
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch