[RFC PATCH xserver] modesetting: re-set the crtc's mode when link-status goes BAD
Daniel Vetter
daniel at ffwll.ch
Tue Feb 28 08:42:12 UTC 2017
On Tue, Feb 28, 2017 at 04:07:02AM +0000, Navare, Manasi D wrote:
>
> On Fri, Feb 24, 2017 at 12:09:51PM -0800, Manasi Navare wrote:
> > Hi Daniel,
> >
> > We have ACKs on the userspace design from both Adams and Eric.
> > Is this enough to merge the kernel patches?
> >
> > I spoke to Eric briefly about this and he gave me a verbal r-b as well.
> > He said the userspace patches cant get merged unless DRM patches are merged.
> >
> > So what should be some of our next steps here?
>
> Still needs review on the kernel patches themselves, iirc Jani still had some opens on that. But I was out of the loop for 2 weeks :-) -Daniel
>
> Thanks for merging the 1st patch in the kernel series. Are there any opens on the 2nd patch (the i915 patch)?
> I wanted to actually respin the 2nd patch to add reset for intel_dp->lane_count on the link training failure so that it doesn't accidently retrain the link with the stale/failed values.
Didn't look like that, and we can do this as a follow-up. I only asked
where we should merge it for best results. Let's continue that discussion
in the other thread.
-Daniel
>
> Regards
> Manasi
>
> >
> > Regards
> > Manasi
> >
> >
> > On Mon, Feb 13, 2017 at 01:05:17PM -0800, Eric Anholt wrote:
> > > Martin Peres <martin.peres at linux.intel.com> writes:
> > >
> > > > On 06/02/17 17:50, Martin Peres wrote:
> > > >> On 03/02/17 10:04, Daniel Vetter wrote:
> > > >>> On Fri, Feb 03, 2017 at 01:30:14AM +0200, Martin Peres wrote:
> > > >>>> On 01/02/17 22:05, Manasi Navare wrote:
> > > >>>>> On Wed, Feb 01, 2017 at 11:58:16AM -0800, Eric Anholt wrote:
> > > >>>>>> Jani Nikula <jani.nikula at linux.intel.com> writes:
> > > >>>>>>
> > > >>>>>>> On Tue, 31 Jan 2017, Eric Anholt <eric at anholt.net> wrote:
> > > >>>>>>>> Martin Peres <martin.peres at linux.intel.com> writes:
> > > >>>>>>>>
> > > >>>>>>>>> Despite all the careful planing of the kernel, a link may
> > > >>>>>>>>> become insufficient to handle the currently-set mode. At
> > > >>>>>>>>> this point, the kernel should mark this particular
> > > >>>>>>>>> configuration as being broken and potentially prune the
> > > >>>>>>>>> mode before setting the offending connector's link-status
> > > >>>>>>>>> to BAD and send the userspace a hotplug event. This may
> > > >>>>>>>>> happen right after a modeset or later on.
> > > >>>>>>>>>
> > > >>>>>>>>> When available, we should use the link-status information
> > > >>>>>>>>> to reset the wanted mode.
> > > >>>>>>>>>
> > > >>>>>>>>> Signed-off-by: Martin Peres <martin.peres at linux.intel.com>
> > > >>>>>>>> If I understand this right, there are two failure modes
> > > >>>>>>>> being
> > > >>>>>>>> handled:
> > > >>>>>>>>
> > > >>>>>>>> 1) A mode that won't actually work because the link isn't
> > > >>>>>>>> good enough.
> > > >>>>>>>>
> > > >>>>>>>> 2) A mode that should work, but link parameters were too
> > > >>>>>>>> optimistic and if we just ask the kernel to set the mode
> > > >>>>>>>> again it'll use more conservative parameters that work.
> > > >>>>>>>>
> > > >>>>>>>> This patch seems good for 2). For 1), the
> > > >>>>>>>> drmmode_set_mode_major is going to set our old mode back.
> > > >>>>>>>> Won't the modeset then fail to link train again, and bring
> > > >>>>>>>> us back into this loop? The only escape that I see would
> > > >>>>>>>> be some other userspace responding to the advertised mode
> > > >>>>>>>> list changing, and then asking X to modeset to something
> > > >>>>>>>> new.
> > > >>>>>>>>
> > > >>>>>>>> To avoid that failure busy loop, should we re-fetching
> > > >>>>>>>> modes at this point, and only re-setting if our mode still exists?
> > > >>>>>>> Disclaimer: I don't know anything about the internals of the
> > > >>>>>>> modesetting driver.
> > > >>>>>>>
> > > >>>>>>> Perhaps we can identify the two cases now, but I'd put this
> > > >>>>>>> more
> > > >>>>>>> generally: if the link status has gone bad, it's an
> > > >>>>>>> indicator to userspace that the circumstances may have
> > > >>>>>>> changed, and userspace should query the kernel for the list
> > > >>>>>>> of available modes again. It should no longer trust
> > > >>>>>>> information obtained prior to getting the bad link status,
> > > >>>>>>> including the current mode.
> > > >>>>>>>
> > > >>>>>>> But specifically, I think you're right, and AFAICT asking
> > > >>>>>>> for the list of modes again is the only way for the
> > > >>>>>>> userspace to distinguish between the two cases. I don't
> > > >>>>>>> think there's a shortcut for deciding the current mode is
> > > >>>>>>> still valid.
> > > >>>>>> To avoid the busy-loop problem, I think I'd like this patch
> > > >>>>>> to
> > > >>>>>> re-query
> > > >>>>>> the kernel to ask about the current mode list, and only try to re-set
> > > >>>>>> the mode if our mode is still there.
> > > >>>>>>
> > > >>>>>> If the mode isn't there, then it's up to the DE to take action in
> > > >>>>>> response to the notification of new modes. If you don't have a DE to
> > > >>>>>> take appropriate action, you're kind of out of luck.
> > > >>>>>>
> > > >>>>>> As far as the ABI goes, this seems fine to me. The only concern I had
> > > >>>>>> about ABI was having to walk all the connectors on every uevent to see
> > > >>>>>> if any had gone bad -- couldn't we have a flag of some sort about what
> > > >>>>>> the uevent indicates? But uevents should be super rare, so I'd say
> > > >>>>>> the
> > > >>>>>> kernel could go ahead with the current plan.
> > > >>>>> Yes I agree. The kernel sets the link status BAD as soona s link
> > > >>>>> training fails
> > > >>>>> but does not prune the modes until a new modelist is requested by
> > > >>>>> the userspace.
> > > >>>>> So this patch should re query the mode list as soon as it sees the
> > > >>>>> link status
> > > >>>>> BAD in order for the kernel to validate the modes again based on new
> > > >>>>> link
> > > >>>>> paarmeters and send a new mode list back.
> > > >>>> Seems like a bad behaviour from the kernel, isn't it? It should return
> > > >>>> immediatly
> > > >>>> if the mode is gonna be pruned :s
> > > >>> The mode list pruning isn't relevant for modeesets, the kernel doesn't
> > > >>> validate requested modes against that. The mode list is purely for
> > > >>> userspace's information. Which means updating it only when userspace asks
> > > >>> for it is fine.
> > > >>
> > > >> Hmm, ok. Fair enough!
> > > >>
> > > >>> I also thought some more about the loop when we're stuck on BAD, and I
> > > >>> think it shouldn't happen: When the link goes BAD we update the link
> > > >>> parameter values to the new limits, and the kernel will reject any mode
> > > >>> that won't fit anymore. Of course you might be unlucky, and the new link
> > > >>> limits are also not working, but eventually you're stuck at the "you're
> > > >>> link is broken, sry" stage, where the kernel rejects everything :-)
> > > >>>
> > > >>> So I think the busy-loop has strictly a limited amount of time until it
> > > >>> runs out of steam.
> > > >>
> > > >> OK, I have given it more thoughts and discussed with Ville and Chris IRL or
> > > >> on IRC.
> > > >>
> > > >> My current proposal is based on the idea that the kernel should reject a
> > > >> mode
> > > >> it knows it cannot set. This is already largely tested in real life: Try
> > > >> setting 4K at 60Hz on an HDMI 1.x monitor, it should immediately fail on
> > > >> prepare(). For this proposal to work, we would need to put in the ABI
> > > >> that a
> > > >> driver that sets the link-status to BAD should also make sure that whatever
> > > >> the userspace does, no infinite loop should be created (by changing the
> > > >> maximum link characteristics before setting the link-status property).
> > > >>
> > > >> Re-probing the list of modes and checking if the mode is still in there is
> > > >> inherently racy, as a new screen may be plugged between the moment the list
> > > >> is sent to the userspace and the moment when we try setting the mode. Or
> > > >> the
> > > >> DE may also have changed the mode in the mean time and the kernel would
> > > >> have reduced the limits even more. So, the userspace cannot be expected to
> > > >> always do the right thing here :s.
> > > >>
> > > >> From this point of view, I really do not like the idea of re-probing, since
> > > >> it increases the delay before the DE can handle the change and it does not
> > > >> bring any guarantee of working properly.
> > > >>
> > > >> Did I miss anything? Any opinions?
> > > >
> > > > Any comments on this, Eric? Does it sound logical to you or did I miss
> > > > something?
> > > >
> > > > The kernel patches are stuck until this patch gets in. So far, you look
> > > > like the only person who would be willing to review this patch (Ajax
> > > > acked the patch, but that's the extent he was willing to go).
> > >
> > > I was just trying to provide review to get the kernel unstuck. The
> > > kernel should not be blocked until the patch gets lands (this obviously
> > > isn't the case with ioctls, which *don't* land in userspace until kernel
> > > does), you just need userspace published and generally agreed that the
> > > ABI works.
> >
> >
>
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
More information about the xorg-devel
mailing list