Re: [PATCH] Revert "drm: Stop resetting connector state to unknown"

[Lyude]

On Tue, 2015-12-15 at 13:42 -0500, Rob Clark wrote:
> On Tue, Dec 15, 2015 at 1:33 PM, Ville SyrjÃlÃ
> <ville.syrjala@xxxxxxxxxxxxxxx> wrote:
> > On Tue, Dec 15, 2015 at 01:23:06PM -0500, Rob Clark wrote:
> > > On Tue, Dec 15, 2015 at 1:07 PM, Lyude <cpaul@xxxxxxxxxx> wrote:
> > > > This reverts commit 5677d67ae394 ("drm: Stop resetting connector state
> > > > to
> > > > unknown")
> > > > 
> > > > Unfortunately, not resetting the connector status to unknown actually
> > > > breaks reprobing on suspend/resume in i915, which is important to have
> > > > working since it means a user docking their laptop in suspend won't have
> > > > their monitors work after resume. This commit was originally pushed to
> > > > fix
> > > > a bug with systemd[1], however said bug has already been fixed in
> > > > userspace.
> > > > 
> > > > Since "unknown" is technically a valid option to return to userspace for
> > > > a
> > > > connector's status, I would think that this sort of behavior should
> > > > probably be expected from userspace. Some good examples of this are the
> > > > radeon driver reporting "unknown" for connectors that have done
> > > > something
> > > > wonky during a hotplug event (e.g. part of the initialization of the
> > > > connector failed), and the omapdrm driver returns "unknown" for certain
> > > > connector types by default.
> > > > 
> > > > [1]: https://bugzilla.kernel.org/show_bug.cgi?id=100641
> > > > 
> > > > Signed-off-by: Lyude <cpaul@xxxxxxxxxx>
> > > > ---
> > > > Âdrivers/gpu/drm/drm_crtc.c | 7 ++++---
> > > > Â1 file changed, 4 insertions(+), 3 deletions(-)
> > > > 
> > > > diff --git a/drivers/gpu/drm/drm_crtc.c b/drivers/gpu/drm/drm_crtc.c
> > > > index 24c5434..474e636 100644
> > > > --- a/drivers/gpu/drm/drm_crtc.c
> > > > +++ b/drivers/gpu/drm/drm_crtc.c
> > > > @@ -5312,11 +5312,12 @@ void drm_mode_config_reset(struct drm_device
> > > > *dev)
> > > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂif (encoder->funcs->reset)
> > > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂencoder->funcs->reset(encoder);
> > > > 
> > > > -ÂÂÂÂÂÂÂmutex_lock(&dev->mode_config.mutex);
> > > > -ÂÂÂÂÂÂÂdrm_for_each_connector(connector, dev)
> > > > +ÂÂÂÂÂÂÂlist_for_each_entry(connector, &dev->mode_config.connector_list, 
> > > > head) {
> > > > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂconnector->status = connector_status_unknown;
> > > > +
> > > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂif (connector->funcs->reset)
> > > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂconnector->funcs->reset(connector);
> > > > -ÂÂÂÂÂÂÂmutex_unlock(&dev->mode_config.mutex);
> > > > +ÂÂÂÂÂÂÂ}
> > > > Â}
> > > 
> > > looks like git-revert might have been a bit over-ambitious and
> > > clobbered a couple subsequent changes.. but that is easy enough to fix
> > > once we figure out what the right thing to do is.
> > > 
> > > Beyond that.. I'm not really sure how to apply the "do not break
> > > userspace" rule here.. since prior to
> > > c484f02d0f02fbbfc6decc945a69aae011041a27 userspace could see "unknown"
> > > for certain hardware.ÂÂBut after that commit it could start seeing
> > > "unknown" for drivers/connectors that never would have returned
> > > "unknown" before.ÂÂIf userspace had a problem with "unknown", it
> > > sounds like a userspace bug that was just unnoticed because no one
> > > tested on the right hardware.
> > > 
> > > But anyways, one idea to revert things to original behavior prior to
> > > c484f02d0f02fbbfc6decc945a69aae011041a27 (so at least userspace
> > > doesn't see 'unknown' for drivers/connectors that never used to report
> > > 'unknown') would be to do something roughly like this in
> > > status_show():
> > > 
> > > Â if (status == unknown)
> > > ÂÂÂÂÂstatus = connector->funcs->detect(connector)
> > > 
> > > So I could go with either just reverting this commit, or reverting
> > > commit plus above change.ÂÂMy $0.02 anyways..
> > 
> > Hmm. Or maybe leave the states alone, and just fire off the uevent
> > unconditionally and let userspace initiate the probe. That way we could
> > skip all ->detect() calls during resume for a bit of extra speed.
The current purpose of all of the detect() calls is to check whether or not it
would be appropriate to send the hotplug event (assuming you mean
drm_kms_helper_hotplug_event()), just so that we can save reprobing a display if
nothing changed. Anyway, if speed were an issue here we could just fire off a
worker so that the reprobing doesn't hold up the resume process.

While I'd be fine with just firing off a probe, I don't think it makes sense to
avoid having an unknown state here when numerous drivers already use unknown
states regularly, unless there was a reason other then the one mentioned in the
bugzilla for doing this.

> That is, iirc, something that Lyude suggested when we discussed in
> person (but I asked him to send revert patch just to start broader
*her please :)

> discussion on what the right solution would be).ÂÂDepending on
> userspace to reprobe seems like at least the wrong thing for fbcon..
> which as much as I wish it would go away, as long as it is enabled in
> kernel config, I don't think we can ignore.
> 
> BR,
> -R
> 
> > Not 100% if that would be safe. Maybe something internal depends on the
> > ->detect() having been called? DP stuff perhaps? Which makes me wonder
> > how i915 copes with a DP monitor geting unplugged/change while
> > suspended. I should probably try that.
> > 
> > --
> > Ville SyrjÃlÃ
> > Intel OTC
-- 
Cheers,
	Lyude

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/